Report No. 19584-PNG Papua New Guinea Poverty and Access to Public Services February 18, 2000 FOR OFFICIAL USE ONLY Document of the World Bank This document has a restricted distribution and may be used by recipients only in the performance of their official duties. Its contents may not otherwise be disclosed without World Bank authorization. CURRENCY EQUIVALENTS (As of October 11, 1999) Currency Unit = Kina (K) $1.00 = K3.13 K 1.00 $0.32 ABBREVIATIONS GOPNG Government of Papua New Guinea NCD National Capital District PNG Papua New Guinea NGO Non-Government Organization PSU Primary Sampling Unit CU Census Unit NSO National Statistical Office CARP Cluster Analysis and Regression Package Vice-President: Jean-Michel Severino, EAPVP Country Director: Klaus Rohland, EACNI Sector Manager: Homi Kharas, EASPR Task Manager: Monika Huppi, EASPR FOR OFFICIAL USE ONLY PAPUA NEW GUINEA POVERTY AND ACCESS TO PUBLIC SERVICES CONTENTS Page No. Executive Summary .......................................v Background .......................................v Poverty Profile ......................................v Poverty and access to public services ...................................... vi Social Safety Nets ...................................... ix Information gaps and need for further analysis .......................................x 1. Assessing Poverty in PNG .. A. Background .I B. Per capita consumption, distribution and inequality .3 C. Measuring poverty in PNG .5 2. Poverty And Access To Public Services ..17 A. Introduction .17 B. Health and Nutrition .18 C. Education and Literacy .26 D. Rural Infrastructure .35 E. Living Standards and Access to Public Services .38 F. Decentralization and Provision of Basic Services ..................................... 38 3. Social Safety Nets ..42 A. Introduction .42 B. Informal Social Safety Nets .42 C. Formal Social Safety Nets .46 4. Information Gaps and Future Analysis to Guide Poverty Alleviation Policies.. 50 A. Information Gaps and Additional Analysis .50 B. Lessons from the 1996 Household Survey .52 Bibliographic References .................. 55 This document has a restricted distribution and may be used by recipients only in the performance of their official duties. Its contents may not otherwise be disclosed without World Bank authorization. PAPUA NEW GUINEA: POVERTY AND ACCESS TO PUBLIC SERVICES Annexes I . Measuring the Standard of Living .............................................................. 60 2. Setting A Poverty Line .............................................................. 97 3. Effect of Using Sampling Unit Specific Prices and Urban/Rural Price ..................... 110 Differentials in Poverty Lines 4. Determinants of Child Growth .............................................................. 115 5. Decomposing the Gender Gap in Primary School Enrolment .................................. 119 6. Multivariate Analysis of Poverty in Papua New Guinea ........................................... 126 7. Private Transfers and the Social Safety Net in Papua New Guinea ........................... 141 Tables in Text 1.1 Comparative Social Indicators ..............................................................1 1.2 Relative Importance of Economic Sectors and Labor Force Distribution ...................2 1.3 Consumption Across Income Groups and Regions ......................................................4 1.4 Regional Poverty Lines ..............................................................6 1.5 Poverty Measures by Region ..............................................................8 1.6 Poverty by Household Size, Age and Gender of Household Head ........................... 12 1.7 Poverty by Educational Attainment of Household Head ........................................... 12 1.8 Incidence and Severity of Poverty by Main Income Source of Household Head ...... 14 1.9 Poverty and Income Sources for Working Age Adults .............................................. 16 2.1 Distribution Of Social Indicators Across Expenditure Groups .................................. 17 2.2 Perceptions about Adequacy of Public Services ....................................................... 18 2.3 Comparative Health Indicators .............................................................. 19 2.4 Distribution and Condition of Health Care Facilities across Provinces .................... 20 2.5 Access to Health Facilities and Medical Expenditures by Consumption Quartile .... 21 2.6 Contacts with Health Care Facilities by Consumption Quartile ............................... 21 2.7 Source of Drinking Water and Sanitation by Consumption Level ............................ 23 2.8 The Distribution of Stunting in Young Children ....................................................... 24 2.9 The Distribution of Wasting in Young Children ....................................................... 25 2.10 Anthropometric Indicators for Adults ........................ ...................................... 26 2.11 Adult Literacy Rates and Gender Gaps by Region .................................................... 27 2.12 Distribution of Schooling ..................... 28 2.13 School Enrolment Ratios in PNG and EAP ....................................... 29 2.14 Net Enrollment Rates at Primary and Secondary School Level . . 30 2.15 Distribution of (within-year) School Drop-Outs by Schooling Level .. 31 2.16 Distribution of Reasons for Dropping-Out ........................................ 32 2.17 Average Travelling Time to Schools ............................................ 32 2.18 Access to Public Transportation ..................................................... 36 2.19 Simulated Effect of Certain Changes on Incidence of Poverty . . 39 ii POVERTYAND ACCESS TO PUBLIC SERVICES 3.1 Percentage of Households Giving and Receiving Private Inter-household .............. 43 Cash and in-Kind Transfers in Papua New Guinea 3.2 Average Value of Private Inter-household Cash and in-Kind Transfers as ............. 44 a Percentage of Total Household Expenditure in Papua New Guinea Figures in Text 1.1 Growth in Real GDP, Formal Employment and Labor Force: 1980-1995 ................. 2 1.2 Regional Contribution to National Poverty ............................................................ 9 1.3 Headcount Index and Contribution to Poverty by Agro-ecological Region ............. 11 2.1 Adult Literacy Rates by Sex and Consumption Quartile ....................... ................... 27 2.2 Contribution to National Poverty by School Attainment of Household Head .......... 28 2.3 Educational Attainment Across Regions ........................................................... 28 2.4 Affordability of Education ........................................................... 33 2.5 Consumption and Access to Transportation ........................................................... 36 Boxes in Text 1. Inequality Comparisons .5 2. International Poverty Comparisons .10 3. Summary Profile of Poor Households .14 4. Summary Profile of Poor Households by Region .15 5. Decentralization, Redistribution and the Provision of Social Services .41 6. Benefits Incidence of Public Expenditures on Health and Education .51 in Vietnam and Malaysia iii POVERTYANDACCESS TO PUBLICSERVICES Acknowledgements This report was prepared by Monika Huppi and John Gibson under the direction of Klaus Rohland, Country Director for Papua New Guinea. It has benefited from the comments and inputs of Martin Ravallion, Robin Hide, Tamar Manuelyan-Atinc, Gaurav Datt, Lionel Demery, Bruce Harris, Norbert Schady and Cyrus Talati. The report is largely based on the 1996 Household Survey and a number of background studies-on agriculture, health, education, social safety nets, and NGOs-which were managed by David Klaus. The household survey was the first ever for Papua New Guinea and would not have been possible without the funding that was provided by the Governments of Australia, Japan, New Zealand, and the World Bank. A large number of institutions and individuals were involved in the preparation and implementation of the survey and background studies. These include John Gibson of the University of Waikato, Scott Rozelle of Stanford University, the late Christopher Scott, B.J. Allen and R.M. Bourke Allen of Australian National University, Carol Jenkins of the Institute of Medical Research, Richard Guy, Pani Tawaiyole, John Khambu, Wari lamo, Agogo Mawuli, of the National Research Institute, John Millet of the Institute of National Affairs, Education Development Centre, Michael French Smith, Micael Olsson, and Malcolm Levett and Unisearch of the University of Papua New Guinea. Iv PAPUA NEW GUINEA: POVERTY AND ACCESS TO PUBLIC SERVICES EXECUTIVE SUMMARY BACKGROUND 1. At an average annual income of US$890 per capita (1998), Papua New Guinea (PNG) is classified as a lower middle-income country. Compared to other countries in this group and in the East Asia and Pacific region, PNG scores poorly with respect to basic social indicators. This suggests that an important share of PNG's population may be less well off than the country's income level would imply. 2. This report analyzes the distribution of income, constructs a poverty profile and looks at the extent to which the poor have access to basic services. The analysis is based on data collected during a national household survey in 1996. This was the first and only multipurpose and nationally representative household survey carried out in PNG. Data on a range of social and economic indicators were collected from a nationally representative sample of 1,200 households representing urban and rural households and PNG's five major regions (National Capital District, Papuan/South Coast, Highlands, Momase/North Coast, New Guinea Islands). 3. Distribution of Consumption. The household survey data show that consumption is very unevenly distributed in PNG. The wealthiest 25 percent of the population have a real per capita consumption level over eight times higher than the poorest quartile. There are also marked disparities in consumption levels across regions. Overall, the distribution of consumption in PNG is more unequal than in most countries with comparable income levels. The Gini coefficient derived from the household expenditure data (adjusted for spatial price variations and adult equivalent consumption) is 0.46. POVERTY PROFILE 4. Poverty Line. To measure poverty based on household consumption, a minimum acceptable level of consumption (a poverty line) needs to be determined which separates the poor from the non-poor. PNG does not have an official poverty line. The poverty lines used in this report are based on the cost of a food consumption basket which meets a minimum food -energy requirement of 2,200 calories per adult equivalent per day and reflects the dietary pattern of the lower income groups (food-poverty line). Food expenditures are supplemented by an allowance for non-food expenditures based on the expenditure pattern of those households whose food expenditures just reach the food-poverty line. This results in an average national poverty line of 461 kina per adult equivalent per year. A second, somewhat lower poverty line (399 kina per adult equivalent per year) is based on the same food expenditures, but contains a more restricted allowance for non-food expenditures based on the non-food expenditure share and consumption pattern of those households whose overall expenditures reach the food-poverty line. Because there are significant spatial price variations across PNG, the report calculates separate poverty lines for each one of the five major regions used in the analysis. v POVERTYAND ACCESS TO PUBLIC SER VICES 5. Level and Distribution of Poverty. Based on an average national poverty line of 461 kina per adult equivalent per year (1996 prices), about 37 percent of PNG's population must be considered as poor. The vast majority (93 percent) of those who are poor live in rural areas, where over 41 percent of the population fall below the poverty line. This compares to a headcount index of poverty of just 16 percent for urban areas. Poverty is highest in the Momase/North Coast region (46 percent), but in absolute terms, the highest number of poor households live in the Highlands region. The poor in the Highlands account for 38 percent of PNG's poor. Poverty is lowest in the National Capital District (NCD) (just under 26 percent) and the poor in this region account for less than 4 percent of the country's poor. The report also measures poverty in PNG at an international poverty line of US$ 1/capita/day (1985 prices) converted at the purchasing power parity exchange rate and finds that poverty levels in PNG are high compared to other countries with similar income levels. 6. Who are the poor? Over 93 percent of the poor live in rural areas. Many of them do not earn any cash income and thus derive their livelihood almost entirely from subsistence agriculture. Of those who earn cash income, poverty is highest among those engaged in small scale tree crop production, domestic agriculture and hunting - gathering. Tree crop producers are the most important group in terms of contribution to national poverty and account for 42 percent of PNG's poor. Poverty is significantly lower among households whose head does not depend on agriculture, fishing, hunting or gathering as a main income source. Those households whose head have a formal sector wage job register the lowest incidence of poverty (17 percent), followed by those whose head runs a business (25 percent). Poverty is more widespread among households with older heads and among those who have not attended school. The gender of a household head does not appear to be a good predictor for poverty, as differences in poverty measures between female- and male-headed households are not statistically significant. POVERTY AND ACCESS TO PUBLIC SERVICES 7. Data from the household survey show that lower expenditure groups in PNG fare significantly worse than the upper groups across a wide range of social indicators. Because most of these indicators are substantially influenced by access to basic services such as education, health care, rural infrastructure and utilities, the report reviews distribution of access to such services in more detail. Unequal access to health, education, transport facilities and utilities can further accentuate the effects of unequal income distribution. 8. The household survey shows that there is a clear positive correlation between the level of consumption and a person's satisfaction with his family's access to public services, such as health care, education and transport facilities. This suggests that the upper income groups benefit from better access to basic public services than the poor. It is, however, striking that PNG's population overall shows a very low level of satisfaction with the provision of basic social services. Over half the population consider that their children do not get appropriate access to schooling, almost 60 percent consider their access to health care unsatisfactory and two thirds of the population express dissatisfaction with their access to public transportation. vi POVERTYAND ACCESS TO PUBLIC SER VICES Health Services 9. Although PNG has a relatively well developed health care infrastructure, the health and nutritional status of its population compares unfavorably to that of other countries in the region and has stagnated over the past decade. The low health standards point to unsatisfactory performance and substantial inefficiencies of PNG's health care system. 10. The three tier community based health care system with aid posts serving the rural population, health centers at the next level and hospitals at the tertiary level which PNG inherited at independence has ceased to function properly. The drop in the quality of health services available in rural areas has been particularly severe due to the increasing bias towards urban based curative care. Among the reasons for the increasingly poor performance of PNG's health care system are: (i) growing personnel expenditures which crowd out other operational expenditures; (ii) an erratic flow of funds which makes it impossible to properly plan and execute priority health programs; (iii) low skill levels among health workers and sectoral managers; (iv) a breakdown of the supervision and referral system where higher order facilities provide guidance and supervision to lower end facilities and; (v) an unclear division of responsibilities between local and central Government agencies with respect to management and supervision of the health sector. 11. Access to health facilities by lower income groups is substantially inferior to that of the upper income groups. People from the lowest consumption quartile travel more than twice as long to reach an aid post or a health care center than those among the top consumption quartile. Lower income groups make much less use of formal health care providers than the upper groups and they appeal almost exclusively to lower end facilities (aid posts and health centers). On the other hand, almost one third of all contacts by the top consumption quartile are with a hospital, where the quality of service remains substantially better than in the lower end facilities. 12. Households ranked improved access to health care among the highest priorities for public intervention and NGO assistance. Reform of PNG's health care system is critical to improve the population's health standards and raise the lower income groups' access to proper health care. 13. Water and Sanitation. Availability of clean water and appropriate sanitation has an important bearing on the population's health status. Over 60 percent of the population rely on rivers, lakes, creeks and similar unprotected sources for drinking water. The poor are particularly at risk for unsafe water. Not surprisingly, improvements in access to safe water was ranked as one of the highest priorities by households and the Government should make provision of safe water a main concern in the area of public health. 14. Poverty and Malnutrition. Malnutrition is a considerable problem among PNG's population, particularly among women and children. It is closely linked to poverty. The risks of children and women from the lower consumption quartiles being chronically malnourished are significantly higher than those of children and women from the upper quartiles. vii POVERTYAND ACCESS To PUBLICSERVICES Schooling and Literacy 15. Literacy. Literacy and schooling are key determinants of a person's ability to take advantage of income-earning opportunities. Literacy rates remain low in PNG and there is a marked gender gap in literacy achievement. Based on the household survey, only 63 percent of PNG's men and 44 percent of PNG's women consider that they are literate. Literacy rates and the gender gap vary widely between regions. Literacy rates drop significantly with decreasing income levels. 16. Educational Attainment. Only half of all women aged 15 and above and two thirds of men have ever attended school. Educational attainment in PNG is strongly related to economic welfare. Over half of all poor in PNG live in households whose head has never attended school, although this group only accounts for 38 percent of the population. There are also marked disparities in the regional distribution of educational attainment. The proportion of the adult population who have never attended school ranges from a moderate 15 percent in the urban NCD to a high 57 percent in the Highlands. 17. School Enrolments. School attendance among PNG's children remains low. Net enrolment rates are 51 percent for primary school age children and 17 percent for secondary school age children, according to the household survey. The gender gap in school enrolments increases with decreasing consumption levels. Enrolment rates vary significantly across regions: less than one half of all primary-school age children in the Highlands are in primary school, while the NCD and New Guinea Islands achieve enrolment rates of 76 percent and 68 percent respectively. 18. The lack of access to proper schooling by children of the lower income groups is driven by several factors, including a need to travel long distances to schools, particularly secondary schools; lack of teachers in remote areas; and an inability to meet educational expenditures. 19. GOPNG has recently launched a major education sector reform program. The program aims to achieve universal coverage of basic education, to improve system efficiency and to increase girls' participation in education. The program has made a good start. But its initial success now threatens to be tempered by growing capacity constraints at the central and the provincial levels and limited availability of resources. Overall budgetary constraints make it unlikely that substantial additional resources will be available to finance the continued implementation of the education reform program. Therefore, measures which substantially raise the internal efficiency of PNG's education system, such as a reduction in drop-out rates and more efficient deployment of teachers, will be essential to the continued successful implementation of the program. An intra-sectoral shift in resource allocation away from higher education towards basic education should accompany efficiency improvements. Successful implementation of the program will be essential for poverty alleviation in PNG. viii POVERTYAND ACCESS TO PUBLIC SER VICES Rural Infrastructure 20. Access to transport infrastructure is an important determinant of economic welfare in PNG. The household survey has shown that there is a marked difference in access to transportation infrastructure between income groups. The lowest consumption quartile must travel over twice as long to gain access to the closest mode of transport than the richest quartile. 21. Although PNG compares relatively favorably with other developing countries in terms of meters of road per person and per square kilometer, a vast majority of roads are poorly maintained and inaccessible during and after rains. Together with the poor integration of different modes of transportation (road, water, air), this results in a highly fragmented and unreliable transportation system and high transportation costs. The latter in turn increases marketing costs and limits possibilities of market integration. 22. Investments in rural infrastructure (roads, shipping facilities, markets, water supply) are critical to spur broad-based rural development and alleviate poverty in PNG. First priority should be given to rehabilitate existing roads and to assure subsequent proper maintenance. The country's difficult topography and terrain will make it impossible to provide all areas with easy, year-round access to transportation. Apart from modest investments in rural tracks, investments in new transportation facilities should be carefully considered for their economic viability. Substantial investments in isolated areas with very low population densities may well not turn out to be economically viable. A policy that over time encourages out-migration from such areas may be a more effective means to alleviate poverty. In the short run, alternative interventions aimed at alleviating poverty are needed in such areas. Among them are improved access to basic education and health services and small scale infrastructure investments that help facilitate living conditions. 23. To properly target small scale infrastructure investments and assure effective participation by the local communities, local investment fund arrangements should be considered. Such funds would provide financing for small scale investments to communities willing to participate and contribute to a particular project of importance to them. Operation of such funds could conceivably be modeled on the functioning of Social Investment Funds which have successfully helped alleviate poverty and mobilize local communities in many parts of the world. SOCIAL SAFETY NETS 24. PNG has an extensive informal safety net system. This system allows for income transfers and other support from members of a particular wantok (informal network based on kinship, ethnicity, language or sometimes merely friendship) to needy members of the same wantok system. The household survey has found that inter-household income transfers remain a very important means of assisting households in need across all income groups, although informal income transfers do not appear to improve the income distribution in rural areas. The wantok system has adapted relatively well to changing economic and social environments and ix POVERTYAND ACCESS To PUBLiC SER VICES remains as important in urban as in rural areas. Wantoks may, however, be limited in their effectiveness in communities characterized by very high poverty rates or in times of shock, particularly those brought about by natural disasters. There is therefore a need to supplement these informal safety nets with interventions such as targeted income transfers (e.g., targeted provision of subsidized health and education services) and improved provision of government and NGO funded emergency and relief services. The establishment of self-targeted workfare schemes should also be explored as a means to help alleviate poverty. Such schemes would provide publicly financed unskilled work on demand at a wage rate low enough to guarantee that only those in real need are willing to participate. The labor would be used to build and maintain high priority community infrastructure in poor areas. 25. PNG's wage linked social insurance schemes, although only benefiting a small part of the population, are an important safety net mechanism and will grow in importance as the country will develop further and a larger share of the labor force will enter the formal labor market as wage earners. However, their future effectiveness will largely depend on the Government's ability to implement a series of measures which can strengthen the financial performance of these funds. INFORMATION GAPS AND NEED FOR FURTHER ANALYSIS 26. Further analysis. The data collected during the first nationally representative household survey has made it possible to produce a poverty profile for PNG and to see to what extent the lower income groups have access to basic social services. However, to effectively guide Government interventions in favor of poverty alleviation, additional analysis which was not undertaken in this report because of information gaps will be required. This includes a better understanding of the factors which hinder productivity increases and income diversification of low income agricultural producers; a detailed analysis of the evolution, intra-sectoral allocation and benefits-incidence of public expenditures on health and education, and an evaluation of the impact of macro-economic policies on the evolution of poverty. 27. Poverty monitoring and lessons from the household survey. To assess the progress made with respect to poverty alleviation and the effectiveness of Government policies designed to improve the welfare of the poor, regular analysis of the poverty situation is needed. This requires that household surveys be carried out at regular time intervals. PNG should aim at repeating the household living standard survey within two years of the upcoming Population Census. This would allow to effectively draw on the Census as a sample frame and benefit from the capacity build-up at the National Statistics Office. Because an important purpose of the next household survey will be to see whether poverty has decreased since the last survey, it will be important to maintain comparability of data collection methods and a similar coverage for the consumption aggregate and other indicators of living standards. The ultimate goal of a household survey in PNG should be to have a sample large enough to allow for provincial poverty profiles. As a next step, a sample large enough to at least allow for separation of urban and rural sectors in the five main regions should be used. The next survey should also include x POVERTYANDACCESS To PUBLICSERVICES information on land holdings and utilization, access to agricultural support services, credit and markets and the quality of basic public services. xi PAPuEA NEw GUIEA.- POPERTYA,vD ACCESS TO PUBLiC SERI'ICES 1. ASSESSING POVERTY IN PNG A. BACKGROUND 1.1 Although Papua New Guinea (PNG) is classified as a lower middle income country with an average annual per capita income of about US$890, the living standard of the vast majority of its population is akin to that in low-income countries. PNG scores poorly on most social indicators compared to its income level (Table 1.1). This suggests that the fruits of economic growth may have been unevenly distributed and that poverty remains an important development problem in PNG. Table 1.1: Comparative Social Indicators Lower-Middle East-Asia and Indicator PNG Income Countries Pacific Region Infant mortality (per 1000 life birth) 61 38 57 Life expectancy at birth 58 68 69 Primary school enrolment (gross) 80 103 117 % of population with access to 22 58 29 adequate sanitation GNP/Capita (1998 US$/cap) 890 1710 990 Source: World Bank, World Development Indicators, 1998. 1.2 PNG is a nation rich in natural resources, with gold, copper and agricultural products comprising the most important sources of export earnings. PNG's economic development in the two decades since independence has been driven by a small modem enclave sector, mainly based on mineral resource extraction, commercial logging and tree crop plantations. Government policies have almost exclusively focussed on fostering the development of these activities. Because it is heavily based on natural resource extraction and plantation agriculture, the performance of PNG's economy is substantially driven by that of world market commodity prices. Overall, PNG's enclave economy experienced significant but fluctuating growth in output and exports throughout the last two decades, with little impact on the rest of the economy, particularly the agriculture sector. 1.3 The strong focus on the enclave sector, particularly capital intensive mineral extraction, has done little to create employment opportunities outside subsistence agriculture. While industry and manufacturing account for 40 percent of PNG's value added, they provide 1 POVERTYAND A CCESS To PUBLIC SER vICES employment for only seven percent of the population. Agriculture, accounting for about one quarter of the country's value added, continues to provide a livelihood for almost 80 percent of the population. However, only about 4 percent of the economically active rural population are engaged in modem plantation based agriculture. The remaining 96 percent are village small holders and many of them are engaged in subsistence or semi-subsistence production. Table 1.2: Relative Importance of Economic Sectors and Labor Force Distribution Share in GDP (°) Share in Labor Force (%) Average Annual GDP Growth Rate Share in Labor Absorption 1980-96 (%) 1980-96 (%) Agriculture Industry a Services Agriculture Industry Services Agriculture Industry Services Agriculture IndustTy Service 1980 33 27 40 83 5 12 1996 26 40 33 79 7 14 1.8 5.7 3.1 64 14 22 aIndusty includes manufacturing. Manufacturing accounted for 10 percent and for 8 percent of GDP in 1980 and 1996, respectively. It grew by an annual average rate of 1.7 percent between 1980 and 1996. This underlines the importance of the mining sector in industry. Source: World Development Indicators, 1998. Figure 1.1: Growth in Real GDP, Formal Employment and Labor Force: 1980-1995 1.4 Rigid labor market policies, such as high minimum 4000 wages, have contributed to very 3500 slow employment creation, thus 3000 - requiring a large share of the 2500 - +Real GDP population to remain in near 2000 - -Employment subsistence agriculture. Although 1500 \ Labour Force growth in the agricultural sector 1000 \ was substantially below that in 500 0 the industrial and service sectors 1980 1985 1990 1995 1997 between 1980 and 1996, agriculture absorbed almost two thirds of new entrants in the labor market during this time. Industry (including manufacturing) and services created only 14 percent and 22 percent respectively of new jobs between 1980 and 1996. Sluggish performance of the agricultural sector throughout the 1980s, partly due to policies biased against agriculture, resulted in little employment generation in rural areas. As a result, open and disguised unemployment have continued to increase substantially, with income levels for the majority of the population stagnating. 1.5 Poverty alleviation remains a major development challenge in PNG over the years to come. To guide Government policy in this direction, it is important to gain insights into the extent and distribution of poverty and to learn what characterises the poor. This report draws up a national poverty profile that can support GOPNG's efforts to improve the design of its poverty -2- POVERTY AND ACCESS To PUBLICSERVICES reduction policies. It also looks at how provision of basic public services and overall development policies to date have affected the poor. 1.6 Chapter one describes the level and distribution of household consumption, based on data collected during the 1996 household survey and presents a poverty profile. Chapter two focuses on the access of the poor to basic public services, such as health, education and infrastructure. It also explores what the impact of improved access to public services would be on poverty levels in PNG. Chapter three reviews formal and informal safety nets. Chapter four outlines further work that should be carried out to guide GOPNG's future efforts on poverty alleviation and summarizes the lessons learned from the 1996 household survey. 1.7 The analysis in this report is mainly based on data collected during a national household survey carried out in 1996. The survey collected infornation on a wide range of household characteristics, including demographics, employment, consumption, wealth accumulation, access to basic public services, anthropometrics and perception of quality of life. This was the first survey to draw on a nationally representative sample of 1200 households. The sample covers urban and rural households in PNG's five major regions (National Capital District, Papuan/South Coast, Highlands, Momase/North Coast, New Guinea Islands). Annex 1 provides a more detailed description of the household survey. B. PER CAPITA CONSUMPTION, DISTRIBUTION AND INEQUALITY 1.8 Consumption. Per capita consumption is often used as a basic indicator to measure welfare. To correctly depict living standards, particularly for purposes of comparing living standards across various groups, household expenditure data should be adjusted to capture differences in needs and in prices faced by various households. Typically, the cost of sustaining a child is lower than the cost of sustaining an adult. It is therefore sensible to present consumption data on an adult equivalent basis, meaning that a smaller weight is given to the cost of children than to that of an adult. Analysis of the household survey consumption data for PNG suggests that the cost of supporting a child between ages 0-6 years is in the area of 50 percent of the costs of sustaining an adult, while the costs of a child 7 years and above are close to those of an adult (see Annex 1). Therefore, an adult equivalent adjustment of 0.5 was used for children below 6 years of age, while children above that age were given the same weight as adults. Because prices vary significantly across PNG's regions, household consumption was also adjusted for spatial price variation. This permits a comparison of consumption across households from various parts of the country (see Annex 1). 1.9 Estimates from the PNG household survey show that nominal per capita consumption per adult equivalent amounted to about 912 kina per year in 1996 (US$ 700). There are, however, significant inter -and intra-regional variations. In nominal terms, adult equivalent per capita consumption in the urban National Capital District (NCD) amounts to over three times as much as that in the poorest New Guinea Islands region. Even after adjustment for spatial price variations, marked inter-regional variations in per capita consumption remain. Consumption per adult equivalent in the NCD still amounts to almost twice as much as that in the poorest New -3- POVERTYAND ACCESS To PUBLIC SER VICES Guinea Islands region and to 1.4 times as much as the national average. This points to a marked difference in the level of welfare between the urban capital region and the predominantly rural rest of the country. Other inter-regional differences in mean consumption are less marked and not always statistically significant, because of large intra-regional dispersions around the mean. 1.10 Papua New Guinean households spend a relatively high share of their expenditures on food (63 percent), suggesting an overall low standard of living. Average caloric availability per adult equivalent is almost 3,000 calories per day and thus well above the often used 2,200 minimum caloric requirement. Average protein intake of 55 grams/day/adult equivalent is also well over the often suggested minimum requirement of 45 grams/day. These relatively favourable national averages mask, however, a very strong variation in food intake across income groups. 1.11 Distribution. There are very marked disparities in the distribution of per capita consumption across expenditure groups. The wealthiest 25 percent of the population have a real per capita consumption level over eight times higher than the poorest quartile. Average caloric availability for the poorest 25 percent of the population falls short of the daily requirement and so does the protein intake for the poorest 50 percent of the population. This marked disparity in nutritional intake between income groups reflects a diet of the lower expenditure groups which is dominated by tubers and starchy staples, but poor in grains, animal fats and proteins. Table 1.3: Consumption Across Income Groups and Regions. Consumption per adult Consumption per adult equivalent per year (kina) equivalent per day Nominal Real (1) Food share Calories Protein (g) Consumption quartile I (poorest) 248 258 0.67 1955 27 II 457 464 0.67 2587 41 III 814 781 0.65 3158 60 IV (richest) 2135 2127 0.55 4200 94 Region National Capital District 2401 1226 0.50 2697 82 Papuan/South Coast 1118 902 0.68 3326 70 Highlands 838 860 0.60 2868 48 Momase/North Coast 706 1007 0.67 3101 53 New Guinea Islands 680 642 0.66 2685 54 Total PNG 912 907 0.63 2974 55 Note: (1) Deflated by spatial price deflators, see Annex I Source: PNG Household Survey 1996 1.12 The distribution of consumption in PNG is more unequal than in most countries with comparable income levels. The richest 10 percent of the population account for 36 percent of measured consumption, while the poorest 50 percent account for only 20 percent of consumption. The Gini coefficient derived from the household survey consumption data is a -4- POVERTYAvD ACCESS TO PUBLICSERVICES high 46.1. For comparison purposes the Gini index was also calculated without adjusting consumption for spatial price variations and adult equivalencies. This results in a Gini of 48.4, which is significantly higher than that of other countries in the region and of other countries with similar per capita incomes (Box 1). C. MEASURING POVERTY IN PNG Setting a Poverty Line 1.13 Food Poverty Line. To measure poverty based on household consumption, a minimum acceptable level of consumption (a poverty line) needs to be determined which separates the poor from the non-poor. PNG does not have an official poverty line. The first step in defining a poverty line is to determine the cost of a basket of food which allows an adequate daily energy intake per person. Typically, such energy requirements have been set between 2,000-2,200 calories/day for the East Asia Region. This report takes a minimum caloric requirement of 2,200 calories/day per adult equivalent to maintain comparability with previously used poverty lines in PNG. The second step consists of defining a food basket which reflects the foods consumed by the lower income groups and determining its cost. This then defines the food poverty line. Because dietary patterns and costs vary significantly across PNG's regions, a separate food basket reflecting the dietary pattern of low income groups was used for each of PNG's five regions and each basket was costed at regional prices, resulting in separate food poverty lines for each region (for details see Annex II). The food poverty lines thus give the required nominal value of consumption per year, in each region, for an adult to obtain 2200 calories per day from a diet of similar quality to the diets of other poor people in the same region. Box 1: Inequality Comparisons 1.14 Poverty Line. As a next step, an GNP PPP per Gini Year of capita coefficient Survey allowance needs to be made for non-food US$1996 expenditures. Even households which are PNG 2820 48.4 1996 poor in the sense that they are consuming Vietnam 1570 35.7 1993 less than the recommended daily calorie Pakistan 1600 31.2 1992 requirement still spend some of their SriLanka 2290 30.1 1990 money on non-food items. Two different Bolivia 2860 42.0 1990 approaches, resulting in an upper and a Indonesia 3310 34.2 1995 lower poverty line, have been adopted Morocco 3320 39.2 1994 for this purpose. For the upper poverty Jamaica 3450 41.1 1991 line an allowance is made for Philippines 3550 42.9 1994 Romania 4580 28.7 1994 expenditures on basic non-food items Ecuador 4730 46.6 1994 based on the expenditure pattern of those Thailand 6700 46.2 1992 households whose food expenditures just Note: Gini coefficients are based on household consumption reach the food poverty line. The sum of per capita, except for Bolivia, where it is based on the food poverty line plus this allowance household income per capita, which tends to result in of non-food expenditures results in the higher inequality. upper poverty line, which will be used as Source: World Development Indicators, 1998 PNG Household Survey, 1996. -5- POVERTYANDACCESS TOPUBLI cSERVICES the reference poverty line in this report. A second, lower poverty line is based on the same food expenditure, but contains a more restricted allowance for non-food expenditures. This allowance is based on the non-food expenditure share and consumption pattern of those households whose overall expenditure reach the food poverty line (see Annex II). 1.15 Regional Poverty Lines. With the above approach separate poverty lines have been calculated for the five major regions in PNG (Table 1.4). These poverty lines thus reflect differing consumption patterns and prices faced by the lower income groups in each region. Taking the population-weighted average of these region-specific poverty lines results in a national average poverty line of 461 kina per adult equivalent per year. 1.16 Ideally, separate poverty lines for urban and rural areas within each region should be constructed, because urban households are likely to face somewhat higher prices on essential items than rural households. However, the relatively small sample size of the PNG household survey and the fact that the sample includes only one urban primary sampling unit for most regions preclude the use of separate poverty lines for urban and rural areas within each region. An alternative would have been to create a single poverty line applicable to all non-NCD urban households. However, an analysis of cluster price variations of key items making up the poverty line suggests that urban/rural price differentials within regions are generally less important than inter-regional price variations (see Annex 2). Therefore, this report makes use of a single poverty line for each one of the five regions specified below. Annex 3 explores the effects which this approach may have on the actual poverty measures. Table 1.4: Regional Poverty Lines (kina per adult equivalent per year) NCD Papua/South Highlands Momase/North New Guinea PNG weighted Coast Coast Islands average Upper PL 1016 547 464 314 479 461 Lower PL 779 496 390 280 424 399 Food PL 543 391 288 218 326 302 Note: See Annex 11 for details of poverty line calculations. Measuring and Comparing Poverty 1.17 National Poverty Rates: Based on the above poverty lines, 41 percent of PNG's rural population and 16 percent of the urban population live in households where the real value of consumption per adult equivalent is below the upper poverty line. The national headcount index of poverty based on the upper poverty line is 37.5 percent. The estimated incidence of poverty at the lower poverty lines is 30.2 percent. Slightly over one-sixth of the population have a total consumption valued at less than the cost of the poverty line food basket, meaning that they couldn't meat the basic calorie requirement implied by the typical food consumption basket of the poor even if they spent all their money on food. Depending on which poverty line is used, the rural poor account for between 93 percent (upper poverty line) and 98 percent (food poverty -6- POVERTYANDACCESS TO PUBLICSERVICES line) of PNG's poor, thus indicating that poverty is largely a problem in rural areas. Therefore, antipoverty programs have to be primarily targeted towards rural areas. 1.18 While the above reported headcount index indicates the proportion of the population with a standard of living below the poverty line, it does not indicate how poor the poor are and hence doesn't change if people below the poverty line become poorer. The poverty gap index, which is the average overall people of the gaps between poor people's standard of living and the poverty line, expressed as a ratio to the poverty line, shows the average depth of poverty. Combining the headcount index and poverty gap indices gives the average consumption level of the poor, which is just over two-thirds of the value of the upper poverty line. It would thus be necessary to transfer almost K250 million per year to poor households to raise the value of their consumption to the level of the upper poverty line. 1.19 The poverty severity index is a distributionally sensitive poverty measure which takes into account the distribution of consumption of those falling below the poverty line.' This index shows that poverty is significantly deeper in rural areas of PNG than in urban areas (Table 1.5). This means that the extent by which the average consumption of poor households in rural areas falls below the poverty line is significantly higher than that of poor households in urban areas. 1.20 Regional Pattern of Poverty. Finding out where poor people live is one of the most basic pieces of information for an antipoverty program. Ideally, a household survey should be able to help in placing targeted interventions. However, the diversity of environments in Papua New Guinea makes this an impossible task for a survey of any feasible size. Even the more limited goal of estimating poverty rates by province would require a very much larger household survey than the one conducted in 1996. Instead, the poverty comparisons presented here are for the four major geographical regions of the country, with the NCD counted as a fifth area, separate from the Papuan region. 1.21 The incidence and extent of poverty vary significantly across the above described five major regions: poverty is lowest in the NCD and highest in the Momase/North Coast region. Only about one quarter of the population of the NCD falls below the upper poverty line, while over 45 percent of the population in the Momase/North Coast region fall below the poverty line. The other three regions (Papuan/South Coast, Highlands, and New Guinea Islands) have poverty Table 1.5: Poverty Measures by Region The headcount index, the poverty gap index, and the poverty severity index can all be estimated using the same general equation, through choice of values for a parameter (Foster, Greer and Thorbecke, 1984 - hereafter FGT). The equation is: Pa = -Ki i-) where the poverty line is z, the value of expenditure per capita for thejth person's household is xj and the poverty gap for individualj is gj = z - xj. Total population size is n and q is the number of poor people (those where xj < z). When parameter ca is set to zero, Po is simply the headcount index. When a is set equal to one, PI is the poverty gap index, and when oX is set equal to two, P2 is the poverty severity index -7- POVERTYAND ACCESS TO PUBLIC SERVICES Headcount Index Poverty Gap Index Poverty Severity Share of total population Index Contribution Index Contribution Index Contribution to total (%) to total (%/6) to total (%) UPPER PL National Capital Dist. 25.8 3.8 8.1 3.6 3.3 3.3 5.5 Papuan/South Coast 33.2 13.2 11.9 14.3 5.5 14.7 14.9 Highlands 35.8 38.3 11.7 38.1 5.3 38.0 40.1 Momase/North Coast 45.8 35.5 14.4 34.1 6.6 34.2 29.2 New Guinea Islands 33.6 9.2 11.8 9.9 5.3 9.8 10.3 PNG 37.5 100.0 12.4 100.0 5.6 100.0 100.0 Urban 16.1 6.5 4.3 5.3 1.6 4.2 15.1 Rural 41.3 93.5 13.8 94.7 6.3 95.8 84.9 LOWER PL National Capital Dist. 16.2 3.0 3.8 2.3 1.4 1.9 5.5 Papuan/South Coast 30.3 14.8 9.8 16.1 4.3 16.4 14.9 Highlands 26.0 34.6 8.0 35.1 3.4 34.7 40.1 Momase/North Coast 38.8 37.5 11.2 35.9 5.0 36.9 29.2 New Guinea Islands 29.8 10.2 9.3 10.5 3.8 10.1 10.3 PNG 30.2 100.0 9.1 100.0 3.9 100.0 100.0 Urban 11.4 5.7 2.2 3.7 0.7 2.6 15.1 Rural 33.5 94.3 10.3 96.3 4.5 97.4 84.9 Source: PNG, Household Survey 1996 rates which are clustered slightly below the national average, ranging from 33.2 percent to 35.8 percent. The depth of poverty as measured by the poverty severity index is twice as high in the poorest Momase/North Coast region as in the national capital district. Viewed in combination with the relatively high average per capita consumption in the Momase/North Coast region (see Table 1.2), this suggests a severely skewed distribution of consumption in this region. These conclusions remain largely unchanged when the lower poverty line is considered. 1.22 Contribution to National Poverty. Another way of viewing the distribution of poverty across regions in PNG is in terms of each region's contribution to national poverty (Figure 1.2). The Momase region accounts for almost 36 percent of the poor in PNG but contains just 29 percent of the population. The Highlands, with about 40 percent of the population, contain 38 percent of the poor. Thus, over three quarters of PNG's poor live in these two regions. The same two regions account for over 70 percent of PNG's poverty when poverty is measured with the distributionally sensitive poverty severity index. This suggests that policies which aim -at improving the living standards in these regions could substantially help reduce poverty in PNG. Only 3.8 percent of the poor are found in the NCD, and this proportion falls if either the more austere poverty line is used, or if the poverty gap and poverty severity indices are used. As a comparison, NCD accounts for 5.5 percent of PNG's population. -8- POVERTYAND ACCESS TO PUBLIC SER vICES Figure 1.2: Regional Contribution to National Poverty Figure 1.2: Regional Contribution to National Poverty New Guinea Papuan/South Islands NCD Coast 9% 4% 13% Momase/North 36% _ : ghlads 38% Source: PNG Household Survey 1996 1.23 How robust are these results with respect to changes in the poverty line and allowance for urban-rural price differentials in the regions outside the NCD? While sample size limitations do not allow us to introduce a separate poverty line for urban and rural sectors within each region, separate poverty lines for urban and rural areas were calculated for the Momase region which had the largest number of urban sampling units outside the NCD. Similarly, a sensitivity analysis was carried out to see to what extent regional and sectoral (urban/rural) poverty comparisons would change if price information from each primary sampling unit were used to determine the poverty line for that sampling unit, rather than average prices for each region (see Annex 3 for details). The results of these analyses show that the main conclusions with respect to regional and urban/rural poverty patterns remain unchanged with the use of alternative poverty lines and price indices. Poverty is predominantly a rural problem in PNG, no matter what poverty line is used. Similarly, while the absolute contribution of different regions to national poverty changes somewhat with the use of different prices for each primary sampling unit, the relative ranking of regions does not change. NCD continues to have the lowest poverty rate and to contribute least to national poverty, while the Highlands continue to account for the largest share of PNG's poor (see Annex 3). -9- POVERTYANDACCESS TO PUBLIC SER VICES 1.24 Agro-ecological Zones. Administrative regions may not provide the best groupings for the analysis of poverty, as there may be significant environmental and socio-cultural variation within a given region and similarities across administrative regions. For example, one of the high elevation areas in the Momase region has more in common with settlements in the Highlands than it does with lower lying areas in the Momase region. An alternative therefore is to look at the poverty profile by agro-ecological regions. Such an analysis shows that the dry lowlands are the agroecological zone with the highest poverty rate, followed by the wet lowlands on the mainland and the high altitude highlands. 1.25 International Poverty Comparisons. How do PNG's poverty rates compare to those of other countries in the region or of other countries with similar levels of income? International comparisons often use a poverty line of US$1/capitalday at 1985 prices converted to the national currency at a purchasing power parity exchange rate. Applying this poverty line to PNG results in a headcount index of 31.0. This is very high compared to most other countries in the region and to countries with similar income levels (Box 2). Characteristics of the Poor 1.26 Besides knowing where in the country overty i prevalet, it isBox 2: International Poverty Comparisons country poverty is prevalent, it iS important to know what characterizes poor GNP PPP Headcount Index Year households, so that properly targeted per capita International of poverty alleviation policies can be US$1996 Poverty Line Survey devised. Age, gender and household size US$1/day are easily identifiable characteristics that PNG 2820 31.0 1996 Vietnam 1570 42.2 1993 can potentially be used for targeting Pakistan 1600 11.6 1991 antipoverty interventions. Table 1.6 below SriLanka 2290 4.0 1990 shows that the extent and depth of poverty Egypt 2860 7.6 1991 tends to rise with the age of the household Kazkhstan 3230 <2 1993 Indonesia 3310 7.7 1996 head, although the second oldest age Morocco 3320 <2 1991 group (41-50 years) exhibits the highest China 3330 22.2 1995 poverty rate. The gender of the household Jamaica 3450 4.3 1993 head, on the other hand, does not appear to Guetamala 3820 53.3 1989 be a good predictor of poverty. Although Romania 4580 17.7 1992 the headcount index of poverty for female Thailand 6700 <2 1992 Notes: The International Poverty line is US$ 1/capitallday headed households is higher than that for at 1985 prices, converted at PPP exchange rate. male headed households, the difference is S 2 ~~~~Source: World Development Indicators, 1999. not statistically significant2 and the PNG Household Survey, 1996. poverty severity index is higher for male I headed households than for female headed households. This suggests that the gender of the household head is not a good predictor of whether a household is poor. 2 The headcount index for female headed households is surrounded by a wide standard error and the t-ratio for the difference in the headcount index for female and male headed households is 1.4. -10- POVERTYAND ACCESS TO PUBLIC SER VICES Figure 1.3: Headcount Index and Contribution to Poverty by Agro-ecological Region (based on foot-poverty line) 40 wet island lowlands highlands 30 _dry lowlands 25 - E high altitude O Headcount Index 25 - wet mainland highlands I * Contribution to National Poverty 20 -lowlands 10 - _ _ | small urban 10 5 major urban 0 _ Source: PNG Household Survey 1996 1.27 Poverty increases with household size and with the number of children per household when no allowance is made for economies of size in household consumption (Table 1.6). However, these results must be interpreted with caution, because larger households typically appear to be poorer when no allowance for household size economies is made. Economies of size in household consumption allow the cost per person of reaching a certain standard of living to fall as household size rises. For example, a household with ten people may not need to spend ten times as much as a single-person household to enjoy the same standard of living. However, there is no fully satisfactory method of determining the proper correction factor for size economies in practice. Annex I carries out some sensitivity analysis, by setting the correction factor to various levels and finds that indeed the relationship between household size and poverty becomes less prominent once corrections are made for household size economies. 1.28 The relationship between poverty and education is critical because education is the major human capital investment that individuals face. As Table 1.7 demonstrates, educational attainment of a household's head is also a good predictor for poverty. Over half of all poor people live in a household whose head has never attended school, although they are only 38 percent of total population. The poverty rate among this group is 51 percent, well above the national average. The poverty rate drops by 10 percentage points for households where the head attended school but did not graduate from community school and poverty continues to fall with rising educational attainments of household heads. This is not surprising, as higher education levels provide better opportunities to diversify income and engage in non-traditional income generation. It also points to the importance of providing the poor with increased access to education as a mechanism of poverty alleviation. (see Chapter 2). -11- POVERTYANDACCESS TO PUBLICSERVICES Table 1.6: Poverty by Household Size, Age and Gender of Household Head Headcount IndexI Poverty Severity Index1 Population Share Index Contribution to Index Contribution to (%/-) total poverty (%/o) total poverty (%) Age of Household Head 0-30years 33.1 17.4 5.4 19.0 19.7 31-40 years 32.2 30.4 4.6 28.9 35.4 41-50 years 46.0 27.7 7.0 28.3 22.6 51+ years 41.4 24.5 6.0 23.7 22.2 Gender of Household Head Male 36.8 91.8 5.6 94.5 93.6 Female 48.1 8.2 4.7 5.5 6.4 Household Size2 1-2 persons 17.2 1.3 1.6 0.9 2.9 3-4 persons 27.4 12.5 4.6 14.1 17.1 5-6 persons 32.2 20.5 5.3 22.6 23.9 7-8 persons 40.1 25.6 5.9 25.5 24.0 9-10 persons 45.7 19.8 6.4 18.6 16.3 > 10 persons 48.1 20.2 6.5 18.3 15.7 Number of Children in Household2 No children 22.7 4.1 3.4 4.2 6.8 1-2 children 29.8 24.4 4.5 24.9 30.7 3-4 children 40.4 40.2 6.2 41.6 37.4 5-6 children 40.6 21.0 5.6 19.3 19.5 > 6 children 67.1 10.3 9.8 10.0 5.7 Notes: 1. Based on Upper Poverty Line 2. Poverty measures are for per capita household consumption, without allowance for size economies. These results, therefore, must be interpreted with caution. See Annex I for details. Source: PNG Household Survey 1996 Table 1.7: Poverty by Educational Attainment of Household Head Headcount Poverty Severity Population Share Index Contribution Index Contribution to (%) to total total poverty poverty (%) (/) Education of Household Head No schooling 51.0 51.7 8.0 54.2 38.0 < grade 6 40.6 22.7 6.6 24.7 21.0 Community school (gr. 6) 30.9 13.9 4.1 12.3 16.9 High school (gr. 7-12) 16.0 4.6 1.8 3.5 10.8 Vocational ed. 23.9 7.1 2.7 5.3 11.1 University ed. 0.0 0.0 0.0 0.0 2.1 Source: PNG Household Survey 1996 -12- POVERTYANDACCESS TO PUBLIC SER VICES 1.29 Poverty and Sources of Income. The ability to earn cash is an important determinant of whether a person is poor or not. Even in Papua New Guinea, where much of household consumption is self-produced, cash incomes are needed for essential non-food items such as school fees, kerosene, and garden tools. Cash can also improve the quality of the diet and provide insurance for periods of agricultural stress. Thus it is important to examine the relationship between poverty and the income earning activities of working-age household members. 1.30 People living in households where the household head derives income mainly from minor non-agricultural activities such as hunting, fishing and gathering have the highest poverty rate, at 57 percent (Table 1.8). The next highest poverty rate is for people in households where the head does not have any cash earning activities. The poverty rate is lower, but still above average, for households where the head's main source of income is from either domestic agriculture (betelnut, food crops, livestock) or export tree crop agriculture (mainly coffee, cocoa, copra, oil palm). A closer look at the main components of domestic agriculture - betelnut, food crops, and livestock - suggest that the three income categories bring similar risks of poverty, although there is a small disadvantage when the head earns most of their income from livestock. 1.31 There are wide differences in the risk of poverty according to which export tree crop provides income for the household head. The risk of being poor is greatest for households whose head earns money from growing cocoa. People living in households whose head grows coffee also have above average poverty rates, and this group constitutes over one-quarter of the country's poor. The lowest poverty rate for tree crop producers in 1996 was for oil palm growers. These variations in poverty rates amongst tree crop producing households partly reflect international market conditions, with 1996 being a good year for oil palm prices after several low priced years earlier in the decade3. 1.32 The lowest poverty rates are for households whose head has a wage job or gets income from a formal business. There is a premium to being employed in the public sector, in terms of a lower poverty rate. The incidence, depth, and severity of poverty is lower for public sector wage workers than for private sector wage workers. A household whose head works in the private sector has a two-thirds greater likelihood of being below the upper poverty line, compared with a household headed by a public sector worker. 1.33 In terms of the contribution to national poverty, almost one-half (43 percent) of the poor live in households where tree crop agriculture provides the main source of income for the 3 The sample size of the survey prevents a meaningful comparison of poverty rates of different types of tree crop producers between regions. While it is important to note that poverty is widespread among export tree crop producers across the country, it must be kept in mind that characteristics and conditions of producers of even the same tree crop (say coffee) can vary widely between regions and even within a given region. This means that any meaningful intervention to help raise living standards of export tree crop producers would have to be tailored to local conditions. The same applies to other categories of agricultural producers with high poverty rates. -13- POVERTYAND ACCESS TO PUBLIC SER VICES household head. A further 19 percent live in households where the head gets most of their income from agricultural sales for the local market. Thus, almost two-thirds of the poor are found in households where the head relies upon agriculture for the main source of income, even though this group is only 53 percent of the population. The next major contributor to poverty is households where the head does not earn any cash income and is thus mainly engaged in subsistence agriculture. (17 percent). Although people in hunter-gatherer households have the highest poverty rate, they are only a small group of the population (5.1 percent), so they comprise just eight percent of the total poor. Only 14 percent of the poor live in households where the head has a wage job or formal business, even though this group comprises 29 percent of the population. Table 1.8: Incidence and Severity of Poverty by Main Income Source of Household Head Headcount Index Poverty Severity Index Index Contribution to total Index Contribution to % of poverty (%/o) total poverty (%ND) population No income source (1) 46.9 16.6 9.5 22.6 13.3 Fishing, hunting, gathering 57.3 7.9 7.5 6.9 5.1 Domestic Agriculture 42.7 19.0 5.5 16.6 16.7 Betelnut 38.1 5.4 5.0 4.7 5.3 Food Crops 43.2 6.8 5.2 5.5 5.9 Livestock 46.5 6.8 6.4 6.3 5.5 Cash Crops 44.0 42.5 6.6 42.8 36.3 Coffee 41.5 26.1 6.8 28.8 23.6 Cocoa 56.6 7.9 8.0 7.5 5.2 Copra 50.2 5.8 6.3 4.9 4.3 Oil Palm 32.3 2.6 2.1 1.1 3.0 Running a business 24.6 4.2 2.7 3.1 6.4 Wagejob 16.6 9.8 2.0 8.1 22.2 Public 13.5 4.4 1.6 3.6 12.2 Private 20.3 5.4 2.5 4.5 10.0 Notes: (1) No income source means that the household head does not earn any cash income, but the household may still obtain cash income from other household members. (2) Poverty measures at upper poverty line Source: PNG Household Survey 1996. Box 3: Summary Profile of Poor 1.34 The main income source of Poor Non-Poor the household head is an important variable for classifying households. averageyearsofschooling of household head 2,4 5,0 However, it is also worthwhile % of household head completed community school 24,3 49,5 studying the relationship between % of household head completed secondary/technical school 7,9 23,6 % literate 36,3 64,1 poverty and other (i.e., secondary) % living in rural areas 95,6 82,1 income sources, whether these be % with main income from agriculture 62,2 52,7 % female headed households 10,4 6,7 from alternative activities of the average floor area per capita 6,4 9,8 household head or from the earnings % access to piped water 3,3 15,0 average real expenditure/adult equivalent 277,0 1287,0 Source: PNG Household Survey 1996 -14- POVERTYAND ACCESS TO PUBLICSERVICES of other household members. Almost one-third of all working-age adults do not participate in any cash earning activities. The most common primary income sources for individuals are: coffee, which gives primary employment for 19 percent of the working-age population, food crops (14 percent; of which one-third is solely sweet potato), betelnut (7 percent), and private sector wage jobs (6 percent). Box 4: Summary Profile of Poor Households by Region Poor Non-Poor Poor Non-Poor Poor Non-Poor Poor Non-Poor Poor Non-Poor Poor Non-poor NCD NCD Papuan Papuan Highlands Highlands Momase Momase NGI NGI Household Size 6.59 5.44 10.1 6.2 6.7 5.7 6.5 5.7 6.4 5.0 6.6 4.9 average years of 2.41 5.02 4.7 10.9 2.6 6.4 1.4 3.4 2.7 5.3 4.8 5.0 schooling of household head % ofhouseholdhead 24.3 49.5 41 92 27 65 14 36 26 51 50 51 completed community school % of household head 7.9 23.6 19 73 n.a. 34 5 13 10 24 21 20 completed secondary/technical school % literate 36.3 64.1 77 97 38 72 20 45 43 73 67 80 %living in rural areas 95.6 82.1 0 0 100 89 100 97 96 71 94 91 % with main income 62.2 52.7 5 4 58 43 59 58 71 55 60 64 from agriculture %femaleheaded 10.4 6.7 4 13 2 1 5 6 17 6 21 15 households average floorareaper 6.4 9.8 4.5 15.1 6.5 9.1 5.3 7.6 7.7 11.7 6.0 11.3 capita (m2) % access to piped water 3.3 15.0 86 96 0 5 2 7 2 22 0 0 average real 277 1287 320 1816 248 1126 287 1158 275 1630 269 863 expenditure/adult equivalent (kina) Source: PNG Household Survey, 1996. 1.35 Which primary income sources are important for poor people? Working-age adults in the households covered by the household survey were classified as poor if their household's consumption level was below the upper poverty line (i.e., intra-household equality is assumed). The highest poverty rates were for people engaged in minor activities (hunting and gathering, miscellaneous livestock raising, spices and vanilla, and handicrafts and artifacts). Amongst the more important economic activities, the highest poverty rates were for people primarily earning from cocoa (45 percent), other food crops (43 percent), and those with no cash earning activities (39 percent). The lowest poverty rates were for people with formal businesses and wage jobs. Overall, the poverty profile by income source thus varies little whether the income source of the household head or the income source of all working age adults is considered. -15- POVERTYAND ACCESS To PUBLIC SER VICES Table 1.9: Poverty And Income Sources For Working-Age Adults Main Income Source All Earning Activities Income earning activity Participation Poverty Participation Female % of Rate Rate Ratea participants Coffee 18.5 38.5 22.3 43.0 Cocoa 4.3 45.2 11.2 45.5 Copra 3.7 36.4 6.1 49.4 Betelnut 6.5 29.7 20.2 68.1 Sweet potato 4.2 32.4 19.6 87.4 Other food 9.8 42.6 26.6 82.6 Oil palm 1.6 25.8 2.4 27.0 Pyrethrum, spices, vanilla 0.1 46.2 1.7 81.2 Chickens 0.9 9.5 5.3 45.3 Pigs 2.4 35.9 9.6 36.3 Cattle and other livestock 0.1 48.0 0.3 33.3 Catching and selling fish 1.4 34.7 5.7 51.8 Hunting, selling bush animals 0.6 69.2 4.9 40.1 Gathering, selling firewood 0.3 70.1 1.0 49.4 Making, selling artifacts 1.3 42.3 5.0 83.9 Owning, running a store 0.8 10.2 2.4 49.6 P.M.V. business 0.2 5.1 0.5 22.1 Running another business 1.6 24.0 3.4 39.2 Wage job-public sector 4.4 12.9 5.0 17.4 Wage job - agricultural sector 0.6 13.4 0.8 16.8 Wage job - other private sector 5.6 18.7 6.9 16.3 No cash earning activities 31.0 38.8 31.0 48.4 a Sums to more than 100 because some people earn income from multiple activities. Source: PNG Household Survey 1996 1.36 The conclusion which results from this poverty profile is that the vast majority of poor people in PNG live in rural areas. Many of them do not earn any cash income and thus derive their livelihood almost entirely from subsistence agriculture. Of those who earn cash income, poverty is highest among those engaged in small scale tree crop production, domestic agriculture and hunting - gathering. Poverty is significantly more widespread among households whose head has not attended school. Over 40 percent of all poor households in PNG derive their main income from the production of export tree crops. Poverty is highest in the MomaselNorth Coast region, but in absolute terms, the highest number of poor households live in the Highlands region. Together these two regions account for three quarters of PNG's poor. -16- POVERTYAND ACCESS To PUBLIC SER VICES 2. POVERTY AND ACCESS TO PUBLIC SERVICES A. INTRODUCTION 2.1 Although consumption-based measures of poverty provide a good indication of the distribution of living standards, they do not fully take into account other dimensions of welfare. Access to publicly provided basic services has a direct impact on the welfare of the population. Unequal access to health, education, transport facilities and utilities can further accentuate the effects of an unequal income distribution. 2.2 Data from the household survey show that lower expenditure groups in PNG fare significantly worse than the upper groups across a wide range of social indicators (Table 2.1). Because most of these indicators are substantially influenced by access to basic services such as education, health care, rural infrastructure and utilities, the distribution of access to such services merits closer analysis. Table 2.1: Distribution Of Social Indicators Across Expenditure Groups Poorest Second Third Richest PNG Quartile Quartile Quartile Quartile Literacy % women aged 15+ 31 38 44 57 43 % of men aged 15+ 48 54 61 77 61 Never attended school % of women aged 15+ 60 54 47 37 49 % of menaged 15+ 40 38 29 20 31 Stunting 52 44 42 34 43 % of children 0-5 years Piped water 4 6 11 27 12 % of households Flush toilet 0 3 5 21 7 % of households Electricity 2 3 13 31 12 % of households Source: PNG Household Survey 1996 2.3 There is a clear positive correlation between the level of consumption and a person's satisfaction with his or her family's access to public services, such as health care, education and transport facilities (Table 2.2). This suggests that the upper income groups benefit from better access to basic public services than the poor. However, it is striking that PNG's population overall shows a very low level of satisfaction with the provision of basic social services. Over half the population consider that their children do not get appropriate access to schooling, almost 60 percent consider their access to health care unsatisfactory and two thirds of the population express dissatisfaction with their access to public transportation. The level of satisfaction with -17- POVERTYAND ACCESS TO PUBLIC SER VICES access to public services varies significantly across regions and is generally lowest in the regions where poverty is most severe. Table 2.2: Perceptions About Adequacy of Public Services % of population who believe that they have inadequate access to services compared to their family's needs Health Care Children's Public Transport schooling Consumption Quartile Poorest 64 59 75 Second 63 55 78 Third 60 50 66 Richest 50 44 48 Region NCD 36 28 25 Papuan/South Coast 50 45 81 Highlands 63 61 67 Momase/North Coast 65 57 68 New Guinea Islands 46 30 62 PNG 59 52 67 Source: PNG Household Survey 1996 B. HEALTH AND NUTRITION Health 2.4 Health Indicators. Although PNG has a relatively well developed health care infrastructure, the health and nutritional status of its population compares unfavorably to that of other countries in the region and has stagnated over the past decade (Table 2.3). 2.5 Deteriorating status of health infrastructure. The low health standards, despite a relatively well developed health care infrastructure, point to unsatisfactory performance and substantial inefficiencies in PNG's health care system. The physical status of the infrastructure has deteriorated substantially over the past two decades. As a result, a vast majority of facilities are now in unsatisfactory condition and provide services poorly. According to the 1996 Demographic and Health Survey, only 39 percent of health facilities are in satisfactory condition and only about 70 percent of rural aid posts are operating. There are also marked regional variations in the availability, operational condition and performance of health care facilities. For example, availability of aid posts or health centers ranges from a high of 267 facilities per -18- POVERTYAND ACCESS TO PUBLIC SERVICES 100,000 inhabitants in Manus province, to a low of 35 facilities per 100,000 inhabitants in North Solomons, while the share of facilities in good condition range from a satisfactory 76 percent in the national capital area to a low 3 percent in North Solomons. The poor conditions of facilities in the latter province are at least partly a result of the effects of civil unrest. Table 2.3: Comparative Health Indicators PNG Indonesia Thailand China Vietnam * Health Centers/lOO,000 population 14 3 14 6 17 * Population per hospital bed 260 1743 665 465 389 * Infant mortality (per 1000 live births) 62 49 34 33 38 * Under 5 mortality (per 1000 live births) 85 60 38 39 NA * Matemal mortality. rate (per 100,000 live births) 930 390 200 115 NA * Low-birth weight babies ( % of births) 23 14 13 6 17 * DTP vaccinated (% of children < I year) 55 89 65 89 95 * Measles vaccinated (% of children < I year) 50 92 73 92 94 * Memo: GNP/cap (US$ 1998) 890 680 2200 750 330 Source: World Bank, World Development Indicators, 1998; UNICEF, State of the World's Children, 1998. 2.6 The three tier community-based health care system, with aid posts serving the rural population, health centers at the next level and hospitals at the tertiary level which PNG inherited at independence has ceased to function properly. The drop in the quality of health services available in rural areas has been particularly severe due to the increasing bias towards urban- based curative care. The severe deterioration of PNG's health care system has been brought about by several factors. Growing personnel expenditures crowd out other operational expenditures and thus leave facilities with insufficient means to attend to the health needs of the population. The flow of funds is often erratic, making it impossible for sectoral managers to properly plan and execute priority health programs. Skill levels among health workers are low due to lack of training. Decentralization has left sectoral administration and management at the provincial and district level in the hands of staff who lack the necessary skills. The supervision and referral system where higher order facilities provide guidance and supervision to lower end facilities to assure proper quality of service has broken down, partly as a result of decentralization. Decentralization has left unclear the division of responsibilities between local and central Government agencies with respect to management and supervision of the health sector. 2.7 The unsatisfactory performance of PNG's health care system is also confirmed by the household survey: almost 60 percent of PNG's population declared that they were dissatisfied with the health care system. As a result, the contact rate with health care facilities has declined since the 1980s. with only 2.1 outpatient visits per capita and 140 hospital admissions per 10,000 population registered by the 1996 (see previous page). -19- POVERTYAND ACCESS TO PUBLIC SER VICES Table 2.4: Distribution and Condition of Health Care Facilities across Provinces Aid posts or No. of persons / % buildings in % of Aid posts health centers/ hospital bed good condition open 1 00,000 pop. PNG 84 384 39 72 Western 135 449 51 76 Gulf 155 649 20 82 Central 88 190 43 78 NCD NA 392 76 100 Oro 138 348 9 80 Southem Highlands 73 377 40 88 Enga Province 68 259 36 74 Western Highlands 40 252 39 76 Sirnbu 66 383 47 73 Eastern Highlands 61 304 20 58 Morobe 89 277 41 75 Madang 89 481 66 82 East Sepik 98 70 27 64 Sandaun 104 596 34 80 Manus 267 588 31 93 New Ireland 92 797 43 91 East New Britain 57 453 55 96 West New Britain 86 522 44 84 North Solomons 35 59 3 61 Source: Demographic and Health Survey, 1996 2.8 Access to health care. Access to health facilities by lower income groups is substantially inferior to that of the upper income groups. People from the lowest consumption quartile travel twice as long to reach an aid post and over three times a long to get to a health care center than those from the top consumption quartile (Table 2.5). Poor households' access to health care is further constrained by their limited ability to pay for health services and medicines. Households from the richest consumption quartile spend 18 times more on health care per person per year than households from the poorest quartile (Table 2.5). Case studies carried out in several areas of the country suggest that the inability to pay for user fees, medicines and transportation costs often prevents lower income groups from seeking medical care4. 2.9 Lower income groups make much less use of formal health care providers than the upper groups (Table 2.6). Furthermore, the poorer segments of the population make almost exclusively use of lower end facilities (aid posts and health centers). On the other hand, one third of all contacts by the top consumption quartile is with a hospital, where the quality of service remains substantially better than in the lower end facilities. 4 Jenkins, Carol, "Poverty, Nutrition and Health Care in Papua New Guinea: A Case Study in Four Communities", Background Paper for Poverty Assessment, Goroka, PNG, October 1996. -20- POVERTYANDACCESS TO PUBLICSERVICES Table 2.5: Access to Health Facilities and Medical Expenditures by Consumption Quartile Consumption Quartile Poorest Second Third Richest Average travelling time to medical facility (minutes) Aid Post 82 73 54 42 Health Center 218 121 88 63 Child Health or Nursing Service5 340 168 129 74 Average annual per capita expenditure on health care (kina) Medicines 0.36 0.84 1.47 4.09 Medical Fees 0.42 1.12 3.24 9.78 Total 0.77 1.97 4.70 13.87 Source: PNG Household Survey 1996 Table 2.6: Contacts With Health Care Facilities by Consumption Quartile Consumption Quartile Poorest Second Third Richest PNG Number of visits per person per month to: Aid Post .19 .16 .17 .16 .17 Health Center .36 .28 .46 .40 .38 Hospital .03 .06 .14 .27 .13 Total .59 .50 .77 .84 .68 Source: PNG Household Survey 1996 2.10 Improved Service Delivery in Health Care. Local communities are acutely aware of the unsatisfactory performance of PNG's health care system. Dissatisfaction with available health care is negatively correlated with household consumption and reflects the lack of access to proper health care by the lower income groups. Almost two thirds of the poorest consumption quartile consider that they do not have adequate access to proper health care (Table 2.2). Households rank the provision of better health services among the highest priorities for Government intervention and NGO assistance6. 5This includes an adjustment for the frequency of field stops of mobile clinics. For example, if women in a particular Census Unit usually attended a mobile 'clinic that stopped at least once per month at a location that was two hours walking distance, a travelling time of 120 minutes would be recorded. But if that clinic was held only every two months, at a location that was two hours walking distance, a travelling time of 240 minutes would be recorded. 6 Education Development Center, "Effective NGO Action for the Delivery of Rural Services in PNG", Boroko NCD, PNG, 1996. -21- POVERTYAND ACCESS TO PUBLIC SER VICES 2.11 Health reform is critical to raise the population's health standards and to improve the lower income groups' access to proper health care. Devolution has shifted the responsibility to provide health care to provincial and district governments. However, local governments largely lack the institutional capacity to appropriately manage the health care system and assure proper service delivery7. Therefore, the responsibilities of government agencies involved in health care need to be carefully reviewed. Responsibilities must be clearly designated and the institutions in charge must have acquired the necessary institutional capacity to carry out their job properly. NGOs and churches constitute an important partner in the area of local service provision. The collaboration and coordination between the public sector, the private sector and the NGO sector in this area must be increased. 2.12 There is also an important need to review the allocation of public sector resources in the health sector. This will show whether expenditure priorities are in line with the health needs of PNG's population. Many countries with poor or falling health performance indicators have found that a reallocation of public expenditures in line with the population's health needs could substantially help increase the sector's performance and improve the lower income groups' access to health care. Such expenditure reallocations generally shift resources away from expensive tertiary care which tends to excessively benefit the upper income groups in urban areas, towards basic primary care provided at the local level. Lack of detailed expenditure data on health prevented such an analysis in this report, but it should be made a high priority so that it can help guide decision making with respect to health care reform in PNG. Water and Sanitation 2.13 Availability of clean water and adequate sanitation has an important bearing on the population's health status. The household survey showed that access to safe drinking water is an important issue across PNG, with over 60 percent of the population relying on rivers, lakes, creeks and similar unprotected sources for drinking water (Table 2.7). The poor are particularly at risk from unsafe water: over 80 percent of the lowest consumption quartile get their drinking water from unprotected sources. This ratio is almost twice as high as that for the richest 25 percent of the population. Not surprisingly, households ranked improvements in access to safe water as one of the highest priorities for Government intervention during the household survey. Given the importance of protected water supply to the population's health and welfare, the Government should make provision of safe water a high priority item in the area of public health. Poverty and Malnutrition 2.14 Adequate nutrition is not only important for an individual's physical well-being, it is also crucial for a person's long term development. Chronic malnutrition, particularly during early childhood, can have a negative impact on a child's mental development, thus reducing his or her chances for future income earning opportunities. 7 See for example, D. Campos-Outcall, K. Kewa and J. Thomason, "Decentralization of Health Services in Western Highlands Province, PNG", in Social Science and Medicine, Vol 40, No. 8, pp. 1091-1098, 1995. -22- POVERTYAND ACCESS To PUBLIC SER VICES Table 2.7: Source of Drinking Water and Sanitation by Consumption Level Consumption Quartile Poorest Second Third Richest PNG Source of Drinking Water: Piped water 4.0 6.5 10.7 26.7 12.0 Well 7.0 11.5 8.4 12.8 10.0 RainwaterTank 8.7 11.9 17.8 16.1 13.6 River, lake, creek, spring 80.0 68.7 58.1 42.5 62.3 Other 0.4 1.4 5.0 1.9 2.2 Access to sanitation: Flush toilet 0.5 2.8 5.0 20.7 7.3 Pit or other toilet 74.5 77.3 80.0 70.6 75.6 No toilet facility 25.0 19.9 15.0 8.7 17.1 Source: PNG Household Survey 1996 2.15 Data from the 1996 household survey indicate that malnutrition is a serious problem among PNG's population, particularly among women and children. Malnutrition in PNG is closely linked to poverty. Stunting reflects chronic malnutrition due to extended periods of inadequate food intake and the cumulative effects of past episodes of infection and sickness. Stunting is particularly prevalent among young children in the lower income groups in rural areas. Overall, 43 percent of children aged 0-5 are stunted, which is very high by international standards. The risk of a child from the poorest consumption quartile being stunted is 18 percentage points higher than that of a child from the highest consumption quartile. (Table 2.8). Econometric analysis of the determinants of children's height for age shows that per capita consumption and by inference poverty is strongly correlated with children's height for age scores in PNG8. This suggests that policies that help raise household income and consumption can also bring about an improvement in the nutritional status of PNG's children. 2.16 There is considerable regional variation in the prevalence of stunting. The risk of being stunted is lowest for children in the NCD and New Guinea Islands regions, higher in the Papuan and Momase regions, and highest in the Highlands (Table 2.8). By contrast, the risk of wasting is lowest in the Highlands (Table 2.9). This puzzling finding is in line with results from the 1982/83 National Nutrition Survey, which found that highland (above 1200 m.) children were significantly shorter but also significantly heavier than lowland (below 600 m.) children. There is a lively and long-standing debate about the extent to which these persistent (and seemingly contradictory) differences across regions may be explained by dietary, environmental or genetic factors. 8 See, Gibson. J. 1999, "Child Height, Household Resources and Household Survey Methods ", Mimeo, University of Waikato, New Zealand. -23- POVERTYAND ACCESS TO PUBLIC SER viCES Table 2.8: The Distribution of Stunting in Young Children (% of children with height-for-age z-score below minus two)a Age group (months) 0-11 12-23 24-35 36-47 48-59 Total Consumption Quartile I (poorest) 25.4 61.0 62.0 64.2 61.6 52.3 II 20.5 54.6 50.3 42.8 56.1 43.6 III 32.9 45.8 46.8 44.7 37.2 42.1 IV (richest) 24.4 34.1 30.2 44.0 32.1 34.4 Region NCD 8.6 28.9 24.1 22.4 16.4 20.3 Papuan/South Coast 20.8 35.2 58.7 40.1 39.5 40.5 Highlands 40.1 76.7 46.1 56.8 66.8 55.8 Momase/North Coast 18.2 35.2 46.4 50.8 46.7 38.8 New Guinea Islands 4.8 27.6 28.7 45.7 n.a.b 25.6 Papua New Guinea 25.3 48.4 46.8 48.4 46.0 42.9 a Height is more than two standard deviations below the median height for that age in the reference population used by the National Center for Health Statistics. bNo stunted children found due to insufficient number of observations available (n=7) Source: PNG Household Survey 1996 2.17 Wasting reflects acute malnutrition associated with a failure to gain weight or a loss of weight. Wasting appears to be less prevalent than stunting in Papua New Guinean children (Table 2.9). Across all ages from year zero to year five, 8.1 percent of children are wasted (with a standard error of 1.1 percent). In terms of international comparisons, this can be considered a medium level of wasting. The highest rate of wasting occurs in the second year of life (13.1 percent), which was also the time of greatest risk found by the 1982/83 National Nutrition Survey. The cause appears to be the late introduction, infrequent feeding, and low nutrient density of the foods introduced as supplements to breast milk. The risk of wasting appears to be the same for boys as for girls. The prevalence of wasting is lowest in the Highlands region (which would suggest that low height for age scores may be due to genetic differences rather than chronic malnutrition). The highest rate of wasting was found in the NCD and Momase regions. Some caution must be exercised in interpreting the incidence rates because the infrequent occurrence of wasting means that sampling errors are large. Nevertheless, it is puzzling that children in the NCD have such a high rate of wasting, given the relative affluence of households there and the low rate of stunting. 2.18 Anthropometric data for adults also indicate that there is a close relationship between the average level of the body mass index and the level of household consumption, especially for women (Table 2.10). However, the proportion of the population with a body mass index below the threshold is less closely related to consumption, with the prevalence being greatest for the second poorest quartile of the population. The fact that the average value of the index for women of the poorest quartile is lower than that for women of the second poorest quartile suggests that -24- POVERTYAND ACCESS To PUBLIC SER VICES some women in the poorest quartile must have a very low body mass index. Overall, the risk of chronic energy deficiency as measured by a low body mass index is three times higher for women than for men. Table 2.9: The Distribution of Wasting in Young Children (% of children with weight-for-height z-score below minus two)a Age group (months) 0-11 12-23 24-35 36-47 48-59 Total Consumption Quartile I (poorest) 19.8 13.9 0.0 6.1 6.8 9.8 II 12.1 22.0 4.1 6.2 11.7 10.4 III 8.7 12.1 1.3 2.7 5.6 6.1 IV (richest) 6.6 4.7 4.3 11.0 1.3 5.9 Region NCD 13.3 28.9 20.1 9.4 9.4 17.5 Papuan/South Coast 16.2 6.6 0.0 14.8 0.0 7.9 Highlands 3.9 4.8 3.0 1.9 4.9 3.5 Momase/North Coast 21.0 24.0 2.5 4.0 13.3 12.2 New Guinea Islands 6.2 10.0 0.0 10.0 0.0 6.6 Papua New Guinea 12.0 13.1 2.7 6.5 6.0 8.1 Weight is more than two standard deviations below the median weight for that height in the reference population used by the National Center for Health Statistics. Source: PNG Household Survey 1996 2.19 An analysis of the determinants of child growth in PNG suggests that a mother's level of education and her own height have a significant impact on the long term nutritional status of her children (Annex IV). The higher the level of a mother's education, the less likely her children are to suffer from chronic malnutrition. Controlling for parental height, and household economic resources, the analysis indicates that each additional year of schooling of the mother results in a 2 percent decrease in the likelihood of her child being stunted. This effect is four times higher than that of an additional year of schooling of the father9. 2.20 The marked effects of a mother's height and education on her children's stunting probabilities indicate that there is a long-run, intergenerational effect of women's education on child health. Educating the current generation of girls will reduce the risk that their own children are stunted. Those children are likely to become taller adults which will then reduce the risk of the next generation of children being stunted. This suggests that there is a role for public action 9The associations between maternal education and stature and children's stunting depicted by the PNG Household Survey are consistent with finding of earlier localized studies from various parts of PNG. See, for example, Harvey P.W.J et. al. (1984); Groos. A. and R.L. Hide (1989); Gibson, J. (1996). -25- POVERTYANDACCESS TO PUBLICSERVICES Table 2.10: Anthropometric Indicators for Adults' Women Men Height Weight BMIP') % with Height Weight BMI°b With (m) (kg) BMI< (m) (kg) BMI< 18.5 18.5 Consumption Quartile I (poorest) 150.6 47.5 20.9 9.7 159.9 56.1 21.8 2.5 II 152.3 50.2 21.6 16.1 163.6 56.9 21.3 12.0 III 153.9 54.0 22.7 14.3 162.6 61.0 23.0 0.4 IV (richest) 155.2 56.6 23.5 8.7 165.1 65.9 24.1 0.2 PNG - - 22.1 12.4 - - 22.5 4.1 Notes: a The sample is the parents of children age five years and under who were weighed and measured. A total number of 544 women and 456 men were included. They were distributed across consumption quartiles as follows (in ascending order from poorest to richest quartile): Women: 109, 142, 141, 152; Men: 93, 106, 123, 134 BMI = Weight (kg) / [Height (m)]2 A body mass index of less than 18.5 indicates chronic energy deficiency. Source: PNG Household Survey 1996 because private choices are likely to lead to less women's education than is socially desirable. When parents choose whether to send their daughter to school they are probably unaware of the impact that this choice has on the health of their yet-to-be-born grandchildren. This "external" effect may lead parents to under-invest in the schooling of their daughters. C. EDUCATION AND LITERACY 2.21 Literacy. Literacy and schooling are key determinants of a person's ability to take advantage of income-earning opportunities. Better schooling generally increases a person's chance of being literate and both provide greater opportunities to earn higher incomes. Literacy rates remain low in PNG. The household survey showed that literacy is not only closely related to household consumption, but that there is also a marked gender gap in literacy achievement. Only 61 percent of PNG's men and 43 percent of PNG's women consider that they are literate (Figure 2.1). This figure drops to just 48 percent for men and to less than one third for women among the poorest population quartile. There are large regional disparities in adult literacy rates in PNG (Table 2.1 1). Over three-quarters of the adult population in the National Capital District and New Guinea Islands can read, but just over one-third of the adults in the Highlands can read. The gender gap between male and female literacy rates varies widely between regions but shows the same pattern as the adult literacy rate. Female literacy rates are almost on par with male rates in the National Capital District and New Guinea Islands, but are less than two-thirds as high as rates for males in the Highlands and North Coast, which are the two regions with the lowest overall literacy rates. -26- POVERTYAND ACCESS TO PUBLIC SER VICES Figure 2.1: Adult Literacy Rates by Sex and Consumption Quartile 90 a0 EIMale a)00 U ~~~~~~~~~Female so 450 30 a) 20 Poorest 11 III Richest TOTAL PNG Source: PNG Household Survey 1996. 2.22 School attainment: School attainment Table 2.11: Adult Literacy Rates and Gender among PNG's adult population is strongly . . related to the level of consumption and is G I significantly lower for women than for men. Adult Literacy Gender Over half of all poor in PNG live in households (% of adult whose head has never attended school, population) although this group only accounts for 38 percent of the population (Figure 2.2). Only National Capital District 85.8 5.3 half of all women aged 15 and above, but over Papuan/South Coast 59.0 15.8 two thirds of all men (15+) have ever attended Highlands 346 18 Momase/North Coast 56.3 25.1 school. This figure drops to 40 percent for New Guinea Islands 77.7 5.8 women and 60 percent for men among the PNG 51.9 18.2 lowest consumption quartile. Similarly, only Notes: 1. Literacy is defined as being able to read a 23 percent of women and 34 percent of men newspaper. from the lowest consumption quartile have 2. Gender gap is male literacy rate minus female literacy rate. completed primary school, while this figure is Source: PNG Household Survey 1996. 66 percent and 48 percent for men and women respectively among the richest 25 percent of the population (Table 2.12). -27- POVERTYAND ACCESS TO PUBLIC SER VICES Figure 2.2: Contribution to National Poverty by School Attainment of Household Head (% of poor households falling into each school attainment category) gacW hdisitdN WO 5/0 3 hours (-4.05) (-3.84) Decrease travelling time to road to 2 hours 31.34 n.a. 28.09 for communities where currently > 2 hours (-4.35) (4.13) Decrease combined travelling time to 3 key 31.61 3.49 27.36 social services by 50 percenta (-3.51) (-66.51) (-6.60) Decrease combined travelling time to 3 key 31.32 n.a. 28.07 services to 6 hours if currently > 6 hours (-4.40) (-4.18) Note: Headcount index calculated as the weighted average of the predicted probability that each household is poor, following a simulated change, where the weights are the household sampling weights multiplied by the number of persons in the household. Model used to predict poverty of rural households reported in Table 3 of Annex 5 and model for urban households reported in Table 6 of Annex 5. The percent change from base is calculated from the predicted baseline values. aHealth center, high school and government station. Source: Annex VI. -39- POVERTYAND ACCESS TO PUBLICSERVICES in many cases been given additional responsibilities, without being provided with the necessary financial resources to carry them out. As a result, many important programs, particularly in the social sectors and in the area of infrastructure development and maintenance have remained underfunded. This has led to a marked reduction of service delivery, particularly in poorer areas. Many district and local governments have not been able to recruit and retain adequately trained staff to assume the increased responsibilities which have been placed on them as a result of decentralization. A centrally determined and unified salary scale for public servants has made it virtually impossible for local agencies in remote areas to attract well qualified staff, because the scale does not allow for any premium payments in difficult service areas. As a result, many remote areas have remained without the necessary staff to provide them with basic health and education services. 2.48 Decentralization can improve the provision of basic services to the population if it is properly planned, implemented and funded. The central government must assume a leadership role in this effort. Decentralization should not be launched until detailed implementation guidelines have been prepared for local governments and the central government is ready to provide adequate guidance throughout implementation. It is also the central government's responsibility to see to it that decentralization does not result in a situation where service provision in poorer regions lags substantially behind that in wealthier areas because poorer regions lack the funding to provide basic social services. To improve basic service delivery in PNG, the achievements of and problems afflicting decentralization need to be carefully reviewed so that corrective measures can be taken. Box 5 outlines key aspects which need to be considered when the provision of social services is decentralized. -40- PO VERTY AND ACCESS TO PUBLIC SERVICES Box 5: Decenteralization, Redistribution and the Provision of Social Services Efforts to decentralize various government functions, including the provision of social services, are underway in scores of developing countries. Perhaps the most important appeal of decentralization is that it can result in more efficient resource allocation and more appropriate service delivery if lower tiers of government have better information about household preferences. Because decentralization can make more apparent the connection between taxes collected and services provided, it may increase consumers' willingness to pay for these services. Decentralization can also provide better opportunities for local residents to participate in decision-making, result in greater accountability of public officials, and strengthen democratic processes. On the other hand, decentralization may lead to increases in regional disparities and greater inequity. Moreover, if local governments do not have adequate capacity, some of the expected increases in efficiency could fail to materialize. There is no cookbook recipe for successful decentralization, but the following issues require consideration. The function or service to be decentralized: Decentralization is not an "all-or-nothing" proposition. In many instances, efficiency gains may be possible without increases in inequality if the central government keeps primary responsibility for financing while local governments take over responsibility for spending decisions, inputs, and implementation. Some social services may inherently be more difficult to decentralize than others. Economic theory suggest that redistribution may best be carried out by higher levels of government, because labor mobility will make attempts by lower jurisdictions to change the distribution of income self-defeating as the poor gravitate to areas of higher redistribution, while the rich cluster in areas of low redistribution. Still, even if central governments take primary responsibility for the financing of safety nets, and establish the criteria which determine eligibility for transfers, local governments may have an informational advantage screening applicants. Decentralization of the health sector is also complicated because of the need for effective referral across levels- from health posts which provide basic services, to high-technology hospitals. Unless these inter-linkages are considered carefully, decentralization can result in a deterioration of some aspects of the services provided, as may have happened in the Philippines, Bolivia, and Zambia. The level of the sub-national government to which responsibilities are decentralized: Economists often argue that decentralization should follow the principle of "subsidiary", whereby decisions are made at the lowest level of government consistent with allocative efficiency. This often involves a careful parsing out of responsibilities. In education, for example, national governments are often responsible for setting standards, curriculum development, and textbook production and distribution; and local governments, communities, and parent-teacher associations are responsible for construction and maintenance of school facilities, and the day-to-day running of schools, as happens in a multitude of countries, from the United States to Bhutan. International experience suggests that efficiency gains in the provision of social services frequently materialize when the central government devolves responsibilities to the community or facility level, but rarely when they are devolved to provinces or regions. In every instance, it is important that revenues for social services follow responsibilities for their delivery. The extent of community mobilization and oversight: Increases in allocative efficiency in the delivery of social services can only take place if more accurate local information can reach decision-makers, and if there are mechanisms whereby these decision-makers are held accountable for their performance. In Colombia, accountability to constituents pushed local mayors to concentrate more on training and hiring effective civil servants. In Northeast Brazil, community oversight and fear of job loss helped motivate civil servants. Initial conditions: The initial distribution of income is important. If income is distributed more unevenly within jurisdictions than across them, decentralization could be equalizing if local authorities have the capacity to transfer income to the poor and share the equity objectives of the center. On the other hand, if there are large initial differences in capacity or jurisdictions do not share the same equity objectives, some sub-national governments may not effectively target the poor. Central governments may therefore have to target poverty funds themselves or create stronger incentives for sub-national governments to do so. -41- POVERTYAND ACCESS TO PUBLIC SERVICES 3. SOCIAL SAFETY NETS A. INTRODUCTION 3.1 The basic elements of an effective poverty alleviation strategy are economic growth and investments in human capital. However, regardless of how well such strategies succeed, there will always be some who will not be able to fully participate. First, it may take too long for some, particularly those in very remote areas, to fully participate and certain groups like the old or disabled may never be able to do so. Second, even among those who will benefit from these policies, there will be some who remain acutely vulnerable to adverse events, such as temporary unemployment, natural calamities or death. In many countries, including PNG, there are informal community or family based arrangements which help those in need cope. However, there are limits to what such informal arrangements can do to protect the most needy. In particular, coping arrangements which work well in normal times, may fail to provide the necessary relief during shock situations. The effectiveness of community-based insurance depends strongly on the extent to which local incomes are affected simultaneously. Unanticipated shock resulting from natural disasters, deterioration of terms of trade or famine can leave entire communities unable to cope on their own. 3.2 There is therefore a role for Government to assist vulnerable households and communities in the face of shock and to ensure minimal levels of provision to those who are unable to benefit from generalized poverty alleviation policies. The latter group is generally best protected through some form of income transfer scheme, while those in need in the face of shock are best aided through some sort of insurance scheme. Successful Government policies in both areas will need to assess what informal arrangements exist and how effective they are. They will then attempt to strengthen or supplement such informal arrangements rather than displace them. B. INFORMAL SOCIAL SAFETY NETS 3.3 The most widespread informal social safety net in PNG is the wantok system"6. This system allows for income transfers and other support from members of a particular wantok to needy members of the same wantok. In rural areas wantok systems are generally based on kinship. In urban areas they often have a broader base which may include kinship, place of origin, ethnicity, language, place of residence or merely friendship. Wantoks which span across urban and rural areas are generally also based on kinship. 3.4 The household survey has found that inter-household income transfers remain a very important means of assisting households. Over 90 percent of households received transfers and over 90 percent of households made transfers. Income transfers (both on the receiving and the giving side) were important across all income groups and in all geographic regions. Although the share of households making transfers declined somewhat as consumption decreased, over 80 16 Wantok is Pidgin English for "one talk", i.e. a group of people speaking the same language. -42- POVERTYAND ACCESS TO PUBLIC SERVICES percent of the poorest quartile still made transfers. In-kind transfers were substantially more important than cash transfers, although about one third of all households received cash transfers. Transfers received averaged over 15 percent of total household expenditure, with little differences across consumption groups, while transfers made averaged over 17 percent of household expenditures. Table 3.1: Percentage of Households Giving and Receiving Private Inter-household Cash and in-Kind Transfers in Papua New Guinea Receiving Transfers Giving Transfers Cash In-Kind Total Cash In-Kind Total Consumption Quartile I (poorest) 25.3 88.6 90.1 22.7 79.4 81.2 II 28.3 89.0 90.0 30.5 90.1 92.2 III 38.8 92.9 93.1 38.5 91.8 95.0 IV (richest) 46.7 92.2 94.5 57.8 92.6 96.8 Region National Capital District 45.6 84.5 89.1 71.9 82.2 90.9 Papuan/South Coast 40.8 94.3 96.0 29.9 93.7 93.7 Highlands 43.3 93.4 94.2 46.6 89.0 92.2 Momase/North Coast 20.9 84.8 86.6 25.3 87.0 90.5 New Guinea Islands 38.4 96.3 96.3 44.6 89.9 91.4 Papua New Guinea 35.6 90.9 92.1 38.6 89.4 91.8 Note: In-kind transfers include self-produced items (e.g., foods) and purchased items given away or received during the interview period (ca. 14 days), plus receipts and donations of less frequently purchased items where the reference period was the previous 12 months. Cash transfers reported are those of K50 or more received from people outside the household or given to people outside the household in the previous 12 months (including family members who live away from the household). Source: PNG Household Survey 1996. 3.5 How effective are wantoks in protecting those in need? Studies of wantoks in urban and rural areas confirm that they remain the most important income transfer mechanisms and have adapted relatively well to changing economic and social environments"7. Their prevalence not withstanding, however, it must be noted that wantoks are not necessarily transfers of charity. Wantok related transfers are part of a system of reciprocity and personal obligations and there is no guarantee that the recipients of transfers are always the most needy. 17 See Papua New Guinea National Research Institute, "Formal and Informal Social Safety Nets in Papua New Guinea", draft report, 1997. -43- POVERTY AND ACCESS TO PUBLIC SERviCES Table 3.2: Average Value of Private Inter-household Cash and in-Kind Transfers as a Percentage of Total Household Expenditure in Papua New Guinea Transfers Received Transfers Given Cash In-Kind Total Cash In-Kind Total Consumption Quartile I (poorest) 2.4 12.8 15.2 3.4 13.9 17.4 II 3.1 12.3 15.3 3.7 12.7 16.4 III 2.5 13.2 15.8 3.1 17.5 20.5 IV (richest) 2.8 12.6 15.4 3.3 15.0 18.4 Region National Capital District 2.0 7.7 9.6 3.8 6.9 10.7 Papuan/South Coast 2.3 11.8 14.1 1.5 20.0 21.5 Highlands 3.3 14.4 17.6 5.0 16.1 21.1 Momase/North Coast 2.2 10.8 13.0 1.7 11.2 12.9 New Guinea Islands 3.0 15.7 18.7 4.4 10.9 15.4 Papua New Guinea 2.7 12.7 15.4 3.4 14.1 17.5 Notes: See Table 3.1 for details on how the data were collected. It appears that respondents either forget some transfers received or put a higher value on the transfers they make compared with the transfers they receive because the reported value of transfers given exceeds the reported value of transfers received. Source: PNG Household Survey, 1996. 3.6 Statistical analysis of the pattern of transfer receipts in the household survey data suggests that informal transfers in the rural sector of PNG leave the distribution of income unchanged - in other words, they are not directed mainly from rich households to poorer households (see Annex VII for details). This is also apparent in the aggregate data (which is dominated by the rural sector) in Table 3.2 which shows that transfer receipts have an almost constant ratio to household expenditures across consumption quartiles. In contrast, transfer receipts in the urban sector do indicate some aversion to inequality on the part of donors, so the distribution of income in the urban sector is made more equal by the wantok system. 3.7 In both the urban and rural sectors, transfer receipts are higher for female-headed households and this may explain why the household survey found that female-headed households were no poorer than other households. In the rural sector, transfer receipts are also higher (and transfer outlays lower) for households where a baby was born in the past year and for households where the head has no sources of cash income. However, these patterns were not apparent in the urban sector. Somewhat surprisingly, transfer receipts do not seem to be higher for older households or in the case of illness (see Annex VII for details). One factor which might help explain why transfer receipts for older households or households afflicted by illness are not higher is that transfer receipts recorded by the household survey were limited to transfers in cash and in kind. However, particularly in rural settings, transfers of labor, childcare and garden work sharing are very common and are likely to particularly benefit the elderly and the sick. While such transfers would not have been recorded by the household survey, they are an important part of the informal social safetynet. -44- POVERTYANDACCESS TO PUBLIC SER VICES 3.8 Thus, the overall pattern of these results suggests that income transfers made under the wantok system may respond to particular shocks and types of disadvantage in the rural sector, whilst not having any overall effect on the distribution of income. In the urban sector, these transfers are targeted towards the poor, but they don't appear to increase with the age of the household head or in the case of illness. The elderly and the sick may, however, well benefit from other transfers, particularly transfers of labor. 3.9 Another geographically more limited analysis of wantoks in an urban setting showed that voluntary transfers of cash made on a regular basis were targeted towards the poor and that households received more transfers if they suffered from a sudden misfortune in the form of lost income. The study also confirmed that wantoks and thus voluntary inter-household transfers remain important for households who have access to formal credit and insurance facilities". 3.10 There is considerable debate about whether the wantok system operates as an effective safety net and redistributive force in urban squatter settlements. Observers of these squatter areas note that the population may have lost all ties to their place of origin and may have no wantoks living close by who can help them in times of need. But proposals to supplement the informal safety net with formal interventions remain controversial because of concern about the incentives created for the formation of new squatter areas. A detailed statistical analysis of urban household survey data from the 1980s suggests that households in urban squatter areas are more likely to participate in the giving or receiving of private inter-household transfers than are other urban households. Multivariate modelling of the pattern of transfer receipts across households suggested that transfers in urban squatter settlements were not quite as well targeted towards the poor as they were in the non-settlement urban areas but they still acted to reduce inequality. Moreover, there was no difference between squatter and non-squatter areas in the net flow of private transfers towards households headed by the elderly or those in ill-health or unemployment. Households in squatter areas also appeared more willing than non-settlement households to raise their outlays on transfers to wantoks as their own incomes rose.'9 Thus, the overall picture from this statistical analysis is not one of the wantok system being less effective in urban squatter areas than it is in other urban areas. 3.11 The importance of wantoks not withstanding, there are limits to what these systems can achieve. In communities which are characterized by very high poverty rates, the possibilities for inter-household transfers remain limited by low household incomes. In times of shock, particularly those brought about by natural disasters, community support arrangements often fail, unless the community includes a substantial number of wantok members outside the affected area (such as urban dwellers). There is thus a need to supplement these informal safety net arrangements with limited formal interventions. 18 See Gibson, J. G. Boe-Gibson and F. Scrimgeour, (1998), "Are Voluntary Transfers and Effective Safety Net in Urban Papua New Guinea?", Mimeo. 19 Gibson, J, (1999 d), "Are There Holes in the Safety Net? Private Transfers in Squatter Settlements and Other Urban Areas of Papua New Guinea", Mimeo. -45- POVERTYAND ACCESS TO PUBLIC SER VICES 3.12 What is the most effective way to supplement informal safety nets? An analysis of social safety net arrangements in PNG carried out by the National Research Institute, identified three types of measures could that effectively complement and support PNG's informal social safety net structures.20 The three interventions identified by this study include: (i) targeted income transfers in the form of free or substantially subsidized health and education services to those who are most at risk and disadvantaged (single parents, families with many children, unemployed); (ii) improved provision of government and NGO funded emergency and relief services, which would assure that relief efforts in response to natural disasters effectively reach those in need, and (iii) investments in basic social services and rural infrastructure to spur overall economic growth and improve welfare, particularly in areas with substantial poverty. This in turn would provide existing wantoks with increased means to assist those in need. Finally, the study underlined that no attempts should be made to directly strengthen wantok systems per se, as such interventions might result in their disintegration. Rather, conditions should be established for wantoks to effectively operate in an overall improved economic environment. C. FORMAL SOCIAL SAFETY NETS 3.13 PNG has several formal safety net arrangements which operate with varying effectiveness. Among them are: (i) wage-linked social security/insurance schemes; (ii) price support schemes; (iii) social development schemes; and (iv) emergency relief actions. 3.14 The wage-linked social security schemes include the National Provident Fund, the Public Officers Superannuation Fund and the Defense Force Retiree Benefit Fund. These schemes provide an effective safety net for wage earners and their families in the formal and the public sector. They provide unemployment, old age, disabled and survivor benefits to their participants. While effective safety net mechanisms, these funds are limited to formal and public sector wage earners, mainly in urban areas, and do thus not benefit the vast majority of those who either are poor or are at risk of being poor. 3.15 Furthermore, all three funds have experienced a significant deterioration in their financial performance and coverage over the past few years. These difficulties are due to a number of factors, including the lack of prudential supervision and regulation, poor investments, Government use of pension funds for capital investments of questionable quality, political interference as reflected by the frequent replacement and appointment of board members and managing directors for political reasons rather than for their qualifications and expertise, and large Government arrears to the fund for public sector employees. The future effectiveness of these funds as safety net mechanisms for those employed by the formal sector thus requires a series of reforms. Among them are the urgent need to establish a body to regulate and supervise 20 See Papua New Guinea National Research Institute, "Formal and Informal Social Safety Nets in Papua New Guinea", draft report, 1997. -46- POVERTYAND ACCESS To PUBLIC SER VICES these funds, the adoption of superannuation legislation, measures to improve governance and increase accountability and revoking the power of the Government to issue investment guidelines to these funds.21 3.16 Price Support Schemes. PNG has had a long history of price stabilization and/or price support schemes for its major export crops, coffee, copra, cocoa and palm oil. The idea behind these schemes was that they would help protect producers of these crops from fluctuations in world market prices and thus smooth their income. The schemes were to pay growers a bonus during times of low world market prices. This bounty was to be financed through levies collected during times of high commodity prices. Through the end of the 1980s these schemes were essentially self-financing. However, with the collapse of world market commodity prices in the late 1980s, the stabilization funds became essentially insolvent and were replaced with Government-assisted export price support programs. These Government financed schemes were extremely costly and ill targeted. The average annual expenditures during the first half of the 1990s amounted to over two thirds of the Department of Agriculture and Livestock's overall budget in support of agricultural development.22 The fiscal burden of maintaining these schemes proved to be unsustainable. Furthermore, while these schemes did help smooth income of some export crop producers, the bulk of the benefits accrued to owners of large estates rather than small holders. Overall, agricultural price support schemes were not an effective means to help prevent poverty. As the household survey shows, poverty remains widespread among small scale tree crop producers. Given resource constraints, the funds allocated to such schemes would be used more effectively to help spur agricultural and rural development through investments in rural infrastructure and basic social and agricultural support services. 3.17 Social Development Schemes: The most prominent scheme is the Rural Action Program, also referred to as Electoral Development Fund. The scheme started in the late 1970s as the Village Development Fund and was meant to provide grant and loan based assistance to villagers with viable project proposals and willing to make a 30 percent contribution of their own. The scheme, however, soon got transferred into a fund for Parliamentarians who distributed resources to their electorate at will. More recently the fund has been put under the administration of the District Planning Committee, of which the local member of Parliament is the chairman. The fund has no obligations to regularly report on its use of funds. Some allocations have been made to those in need in the form of income subsidies, assistance with education or medical expenses and assistance to disaster-stricken families. Other uses include contributions to community development projects. Overall, however, this fund is hardly an effective social safety net mechanism. There are no clear criteria to define target beneficiaries nor are there clear rules about how to obtain these funds and what they can finance. To become an effective poverty alleviation mechanism, this fund should either be transferred into a true social assistance scheme 21 For further details, see World Bank, (1999b), Papua New Guinea: Country Economic Memorandum, draft, June, 1999. 22 See, World Bank, (1997), Papua New Guinea: Accelerating Agricultural Growth - An Action Plan, Washington DC. -47- POVERTYAND ACCESS TO PUBLICSERVICES which is administered based on well defined, transparent rules and criteria or it should be turned into a local development or public employment schemes fund (see para. 3.20). 3.18 Disaster Relief Mechanisms. Both Government and NGOs have mechanisms to provide relief to communities struck by natural disaster. The Government's program is carried out by the National Disaster and Emergency Service, while that of the NGOs is carried out by a number of NGOs, including the Red Cross, Salvation Army and St. John Ambulance. While these services have certainly provided much-needed relief to many communities and families touched by natural calamities, their operations overall often suffer from limited technical know-how, lack of well developed emergency relief operational plans and underfunding. 3.19 PNG has a wide-spread informal social safety net which covers a vast majority of the population. Although transfers under this system to not appear to improve rural income distribution, they do seem to provide protection against certain types of shock. In urban areas they also appear to be targeted towards the poor. Low income levels, particularly in the more isolated communities, and unforeseeable events, however, put limits on how much the informal safety net can protect those in need. There is thus a need for the Government to supplement the informal safety net with well-targeted interventions. In particular, the possibility of providing poor families with many children with assistance to meet basic education and health expenditures should be considered. However, effective targeting, particularly when only certain groups (as opposed to entire regions) are selected, involves significant administrative efforts and costs, and the best way of selecting beneficiaries for such benefits in PNG needs to be evaluated carefully. It is highly unlikely that formal means tests could effectively be employed for this purpose in PNG. Therefore, alternative means, such as household characteristics will have to be employed. The data from the household survey have shown that the level of schooling of the household head and the main income source of the household may be possible selection indicators for targeting of such benefits. 3.20 An alternative way to provide those in need with support would be through self-targeted work-fare schemes. Such programs would provide unskilled low-wage work on demand at a wage rate low enough to guarantee that only those in real need are willing to participate. The labor would be used to build and maintain high priority community infrastructure in poor areas. Communities in designated target areas would be given the opportunity to propose small-scale infrastructure works for funding under the program. If a proposed project met a set of basic selection criteria, the work-fare program would pay for labor and in poorer areas possibly some material costs of executing the scheme. Participation would be open to all who are willing to work for the wage paid by the program. Participation would need to be limited to a maximum number of days per worker, so as to assure that the available financial resources are sufficient to provide work to all those who are willing to work under the scheme at the prevailing low wage rate. If properly designed and funded, work-fare schemes are an effective safety net mechanism to help alleviate chronic and transient poverty (e.g. poverty resulting from natural disaster, economic shocks etc). 3.21 Assistance for disaster-struck areas will remain critical in PNG. The agencies to provide such assistance would benefit from support to strengthen their technical and planning capacities. -48- POVERTYAND ACCESS TO PUBLIC SER VICES Schemes such as agricultural price supports or the Rural Action Program are not effective poverty alleviation mechanisms, nor good safety nets. These are costly operations which are not well targeted and should thus be replaced by more appropriate mechanisms either in support of overall rural development or to provide well targeted social assistance benefits. The wage-linked social insurance schemes, although only benefiting a small part of the population, are an important safety net mechanism and will grow in importance as the country develops further and a larger share of the labor force will enter the formal labor market as wage earners. However, their future effectiveness will largely depend on the Government's ability to implement a series of measures to strengthen the financial performance of these funds. -49- POVERTYAND ACCESS TO PUBLICSER VICES 4. INFORMATION GAPS AND FURTHER ANALYSIS TO GUIDE POVERTY ALLEVIATION POLICIES A. INFORMATION GAPS AND ADDITIONAL ANALYSIS 4.1 The data collected during the first nationally representative household survey has allowed to produce a poverty profile for PNG and to see to what extent the lower income groups have access to basic social services. The data has shown that about 37 percent of PNG's population are poor if an average national poverty line of 461 kina per adult equivalent per year (1996 prices) is used. Compared to countries with similar income levels, poverty in PNG is high. Poverty alleviation thus remains an important development challenge in PNG. Broad based economic growth23 and improved provision of public services, including rural infrastructure, will be key ingredients of a poverty alleviation strategy in PNG. To further guide Government interventions in favour of poverty alleviation, additional analysis will be required on several fronts. 4.2 Agricultural production: The vast majority of PNG's poor are either subsistence or small- scale tree crop producers. Therefore, a successful anti-poverty strategy will need to help these producers increase their productivity, integrate into the market economy and diversify income sources. To formulate proper interventions which can facilitate this process, an analysis of access to and use of land, rural credit and production issues faced by small scale producers in various agro-ecological zones is necessary. This will require that future household surveys include data on land holdings and utilization, agricultural production, access to agricultural support services and markets. In addition, more needs to be understood about the technology base and options for long term productivity increases for many of the crops produced by PNG's small-scale farmers. This in turn requires a detailed analysis of the production systems in various agro- ecological zones of PNG, particularly the Highlands and Momase/North Coast which account for over 75 percent of poverty in PNG. Such analysis will need to be paralleled with proper agricultural research to build up the technology base for productivity increases. 4.3 Effectiveness of Public Spending on Health and Education. The household survey has clearly shown that the lower income groups do not have proper access to basic education and health services and that PNG's population is dissatisfied with its access to public services. There is significant inter-regional variation in the performance of the health and education systems. The poor performance of the health and education systems are at least partly due to inadequate levels and erratic flows of public funding. It is likely that the regional variation in performance is also influenced by differences in funding levels and in the effectiveness of sectoral resource allocation between provinces. Many countries with poor or falling health performance indicators have found that a reallocation of public expenditures in line with the population's health needs could substantially help increase the sector's performance and improve the lower income 23 As the past has shown, economic growth will only markedly benefit the lower income groups if it is not accompanied by a worsening in the income distribution. -50- POVERTYAND ACCESS TO PUBLIC SERVICES access to health care. The same applies to education. It is therefore of great importance that the use of public expenditures for health and education in PNG be closely monitored. 4.4 Public expenditures for the health and education sectors should be analyzed from three angles. First, the evolution of the level and relative importance of health and education sector expenditures should be reviewed. The lack of reliable data on health and education expenditures by the provinces currently make it impossible to reliably assess the adequacy of the overall level and regional distribution of such expenditures. A large share of health and education expenditures are now met by local governments. A detailed analysis of social sector public expenditures per capita by province could help guide central Government decisions on the need for corrective measures, such as the provision of earmarked equalizing grants to poorer provinces. The analysis of sectoral expenditures per capita in conjunction with local performance indicators (e.g. school attainment or enrolment rates, infant mortality or vaccination rates) can furthermore provide important insights about variations in the efficiency with which sectoral resources are used by different provinces. 4.5 Second, it is of critical importance to regularly analyze intra-sectoral allocation of public expenditures on health and education. It is widely believed that the breakdown of the local health care system in PNG is partly due to insufficient levels of funding for primary care in rural areas. Box 6: Benefit Incidence Analysis of Public Expenditures on Health and Education In Vietnam and Malaysia Distribution of Public Sector Health Expenditures by Consumption Quintiles in Vietnam 30 20 _____1 : ____ ____ ____ __ l IODO d.nOleotit 100 Niu Pereun,t of p.blel health Poeot t 2nd 31d 4th Fr.ie The upper income groups benefit more fum public sector health expenditures in Vieotar than the lower groups because the country's health expenditures are skewed in favor of hospital care which is mainly used by the upper ncome groups Imr-oving the poor's access to health care will require a reallocation of public expenditures tovard prinary care services which reach the poor more effectively Sourco The World Ban, Viosran, Po.ert Assc-seuni Distribution of Public Education Expenditures by Consumption Quintile in Malaysia (% of total) an 20 a P.-rt 2nd 31d 4th FWent Malaysia has made great efforts to provide basic education serricos to its people. Aggressire pohcies to ensure unirersal access to basic education and espand access to secondary education has resulted in substantial increases in primary and secondary school participatIon rates by the lower income groups and raised educational attaintoents at all levels. An imporant mans to achieve these results wsas the increasingly higher allocations topnmary and secondary schools Sourcr von do Vallc D and Kl N-d. "Public Spnding md tic Pur -51- POVERTYAND ACCESS To PUBLIC SER VICES Similarly, there are indications that an excessively high share of public sector expenditures are going towards tertiary education, leaving insufficient funds for basic education. There are also indications that high wage bills are crowding out other sectoral operating expenditures, thus leaving the health and education sectors with insufficient funds to properly provide the necessary services. To see whether this is really the case and how intra-sectoral resource allocation could be improved, health and education sector expenditures (both central and provincial) must be analyzed by expenditure category and expenditure level. 4.6 As a third step, the allocation of public expenditures on health and education should be reviewed to see who benefits from these expenditures (see Box 5). The household survey provides information on user rates by level of service for both health and education. This information can be combined with information on the amount of public sector expenditures per user at each level of service. This then shows whether the upper income groups benefit disproportionately from public sector subsidies or not and can help guide future resource allocation in view of poverty alleviation. Reaching the poor effectively means giving priority in allocating public resources to those health and education programs which the poor are most likely to use (i.e. primary health care and basic education). 4.7 Effect of macro-economic policies on poverty. Macro-economic and sectoral policies invariably have an effect on poverty. Past economic policies in PNG have mainly focussed on the development of the capital-intensive mining sector and plantation agriculture. As a result, employment growth in the formal sector was neither sufficient to absorb new entrants into the labor force nor to draw redundant labor from the informal sector, particularly subsistence agriculture, towards production with higher returns to labor. This past trend clearly needs to be reversed if the poor are to actively benefit from economic growth in the years to come. While necessary, economic growth alone will not suffice to substantially reduce poverty in PNG. It will have to be accompanied by an improvement in income distribution which can only be brought about if the lower income groups will be able to fully participate and benefit from growth. 4.8 To assess the progress made with respect to poverty alleviation and the effectiveness of Government policies designed to improve the welfare of the poor, regular analysis of the poverty situation is needed. This requires that household surveys be carried out at regular time intervals (say every 5 years), so that the level, extent and geographic distribution of poverty can be compared over time. PNG should therefore aim at repeating the household living standard survey at regular intervals. B. LESSONS FROM THE 1996 HOUSEHOLD SURVEY 4.9 The 1996 Household Survey was the first multi-purpose survey of living standards that was fielded in Papua New Guinea with a nation-wide scale. Some important lessons for future surveys of a similar nature can be leamed from it. One of the most important lessons from the survey was the need for household surveys to be comparable - and ideally, identical - in methods, if comparisons of poverty from two separate surveys are to be reliable. The only temporal comparison of poverty that was possible in PNG was for Port Moresby where a -52- POVERTYAND ACCESS TO PUBLICSER VICES household survey had been carried out in 1986. The 1996 survey clearly showed that this comparison was affected by changes in survey methods. Specifically, poverty estimates from 1986 were based on data collected in expenditure diaries and when these were compared with 1996 estimates calculated from data collected by verbal recall, there was a large overstatement in the increase in poverty. An important purpose of the next household survey will be to see whether poverty decreased. Therefore, it will be important to maintain comparability of data collection methods and a similar coverage for the consumption aggregate and other indicators of living standards. This implies that changes to the survey design used in 1996 need to be incremental, and ideally, tested with controlled experiments so that the effect of methodological change can be removed from any temporal poverty comparisons. 4.10 Although the 1996 survey was fielded only six years after the 1990 Census (which was the source of the sampling frame), the listings done in each selected Primary Sampling Unit (PSU) showed considerable variation in household numbers compared with the enumeration in 1990. This is partly a function of the great mobility of households in some areas of PNG and likely also reflects under-enumeration by the Census in some remote parts of the country. Although sampling weights can be introduced to deal with the change in the estimated size of the PSUs between the time when the sample frame is formed and the time of the survey, when sampling weights become too variable they reduce statistical efficiency. Therefore, it would be useful for the next survey to follow more closely after the upcoming Census so that there is less variation in sampling weights. A further advantage of following closely after the Population Census would be that the household survey could benefit from the interviewing and data management capacity built up in the National Statistics Office by the Census. PNG should therefore aim at repeating the household living standards survey within two years of the 2000 Population Census. 4.11 The 1996 survey was designed to cover 1200 households and to allow disaggregated information to be presented at a regional level (five regions). Given the increasing importance of provincial governments for the provision of basic services, including many infrastructure investments, and the need for the central Government to supplement provincial budgets with grants from the central budget, knowing how different provinces (or groups of provinces) fare with respect to poverty is important. More detailed regional or provincial poverty profiles could help guide resource allocation decisions by the central Government and help devise geographically targeted interventions. Geographic targeting can be an effective means of channeling resources to areas with a high prevalence of poverty at a fraction of the cost of untargeted programs. The ultimate goal of a PNG household survey should therefore be to allow provincial-level estimates of consumption and poverty. This is, however, a challenging task when there are 20 provinces of widely varying population, unless each province is treated as a separate strata so that sampling rates vary between provinces. Even then, a considerable increase in sample size will be needed and this leads to the increased risk of non-sample errors due to the reduced control that survey organizers have over interviewer teams. A more feasible goal for the next household survey may to be allow disaggregated information by sector (rural and urban) within each region using a sample of approximately 200 PSUs (as opposed to the 120 used in the 1996 survey). The ultimate goal, however, should be to eventually carry out living standard -53- POVERTYAND ACCESS TO PUBLIC SER VICES measurement surveys which allow to determine poverty rates by province or at least for regions which are substantially smaller than the currently used 5 regions. 4.12 Given the importance of agriculture as a means of livelihood for the poor in PNG, the next household survey should include information on land holdings and utilization, agricultural productivity, access to agricultural support services (including formal and informal credit) and markets. One of the reasons why the 1996 survey did not include such information was that a Rural Household Survey that was pre-tested by the National Statistical Office in 1984 was frustrated in its attempt to measure the multiple scattered plots of semi-subsistence farmers. The attempt to measure the area of land in crop production was found to be too time-consuming, even in the topographically favorable areas where the pre-tests took place. Measuring land utilization would be much more difficult in other areas, where some gardens are more than two hours walk from the household and from each other. Therefore, rather than burden an entire survey with the effort of measuring land utilization, an alternative approach might be to select a random subset of selected PSUs and adopt a more intensive fieldwork schedule for the households in these PSUs by employing extra interview labor for the measurement of land holdings. 4.13 The Community Questionnaire used in 1996 asked about access to public services (travelling time by the usual means of people in the community). However, the quality of public facilities is at least as relevant as the ease of access. For example, there is no point in knowing that an Aid Post is only 30 minutes away if it never has the required medicines and sick people are forced to walk two hours to the nearest health centre. Information on the quality of public services would also be helpful in helping to determine the allocation of social investment between construction and maintenance/refurbishment. Therefore, a future survey should include questions on the quality of public services. 4.14 Finally, the collection of prices is crucial for the formation of poverty lines and regional deflators, especially as the number of points where poverty lines are formed is increased. A major problem in PNG is the absence of markets, due to the low population density and high reliance on own-production. This affects the ability to measure prevailing prices - especially of items like fresh produce and firewood. Future surveys need to find methods of measuring prices even when markets are not available. -54- POVERTYAND ACCESS TO PUBLIC SER VICES BIBLIOGRAPHY Allen, J.B. and R.M. Rourke, (1997), "Poverty and Agriculture in Papua New Guinea: An In- Depth Description of Six Agricultural Systems", Mimeo, Canberra, Australia. Bakker, M.L., (1996), The Provincial Populations of Papua New Guinea: Profiles Based on 1990 Census Data, UNFPA/ILO, Port Moresby. Campos-Outcall, D. , K. Kewa and J. Thomason, (1995), "Decentralization of Health Services in Western Highlands Province, PNG", Social Science and Medicine, 40 (8): 1091- 1098. Dahanayake, P.A.S., (1991), "Housing Market Abnormalities in Papua New Guinea" Islands/Australia Working Paper 91/6, National Centre for Development Studies, Canberra. Datt, G. and D. Jolliffe, (1998), "Simulating Poverty Measures from Regression Models of Household Consumption", Mimeo, International Food Policy Research Institute, Washington D.C. Deaton, A., (1997), The Analysis of Household Surveys: A Microeconometric Approach to Development Policy, Johns Hopkins, Baltimore. Deaton, A. and C. Paxson, (1996), "Economies of Scale, Household Size, and the Demand for Food", Mimeo. Deaton, A., J. Ruiz-Castillo and D. Thomas, (1989), "The Influence of Household Composition on Household Expenditure Patterns: Theory and Spanish Evidence", Journal of Political Economy 97(1): 179-200. Deaton, A. and J. Muellbauer, (1986), "On Measuring Child Costs: With Applications to Poor Countries" Journal of Political Economy 96(4): 720-743. Foster, J., J. Greer, and E. Thorbecke, (1984), "A Class of Decomposable Poverty Measures" Econometrica 52(4): 761-765. Education Development Center, (1996), "Effective NGO Action for Delivery of Rural Services in Papua New Guinea ", Background Report for Poverty Assessment. Even, W. and D. Macpherson, 1993. The Decline of Private Sector Unionism and The Gender Wage Gap. Journal of Human Resources, 28(2): 279-296. -55- POVERTYAND ACCESS TO PUBLICSERVICES Gaiha, R. (1988), "On Measuring the Risk of Rural Poverty in India". In Rural Poverty in South Asia, ed. T. N. Srinivasan and P. K. Bardhan, 219-261. irvington, N.Y., U.S.A.: Columbia University Press. Gibson, J., (1999a), "Literacy and Intrahousehold Externality", World Development, forthcoming. Gibson, J., (1999b), "Multivariate Analysis of Poverty in PNG", Mimeo, prepared for Poverty Assessment. Gibson, J., (1999c), "Is There a Malenesian Moral Economy? Private Transfers and the Social Safety Net in Papua New Guinea", Mimeo, prepared for Poverty Assessment. Gibson, J., (1999d), "Are there Holes in the Safety Net? Private Transfers in Squatter Settlements and Other Urban Areas of PNG", Mimeo, University of Waikato, New Zealand. Gibson, J., (1999e), "Child Height, Household Resources and Household Survey Methods", Mimeo, University of Waikato, New Zealand. Gibson, J., (1999f), "Decomposing The Gender Gap in Primary School Enrolments in Papua New Guinea", Mimeo, prepared for Poverty Assessment Gibson, J., (1998a). "Identifying the Poor for Efficient Targeting: Results for Papua New Guinea", New Zealand Economic Papers 32(1): 1-18. Gibson, J., (1998b), "Indirect Tax Reform and the Poor in Papua New Guinea", Mimeo, University of Waikato, New Zealand. Gibson, J., (1996). "Women's Education and Child Growth in Urban Papua New Guinea", Papua New Guinea Poverty Assessment Working Paper Series, Working Paper No. 2, Port Moresby, Institute of National Affairs. Gibson, J. and S. Rozelle, (1998), Results of the Household Survey Component for the 1996 Poverty Assessment for Papua New Guinea, Background Report for Poverty Assessment. Gibson, J., G. Be-Gibson, F. Scrimgeour, (1998), "Are Voluntary Transfers An Effective Safety Net in Urban Papua New Guinea?", Mimeo, University of Waikato, New Zealand. Gould, D. (1995), "Human Capital, Trade, and Economic Growth", Weltwirtschaftliches Archiv, 131(3): 425-445. -56- POVERTY AND A CCESS TO PUBLIC SER VICES Groos, A. and R.L. Hide, (1989), 1987/88 Nutrition Surveys of Karimui and Gumine Districts, Simbu Province, Madang, PNG Institute of Medical Research. Grootaert, C.(1997), "Determinants of Poverty in C6te d'Ivoire in the 1980s", Journal of African Economies 6 (2): 169-196. Harvey, P.W. J.; R.L. Hide, J. Shield, J. Tulloch, H. Vrbova and J. Barker, (1984), "Nutritional Status of Children", in R.L. Hide, ed. South Simbu: Studies in Demography, Nutrition and Subsistence, Port Moresby, IASER: pp.163-206. Heywood. P., N. Singleton and J. Ross (1988), "Nutritional Status of Young Children: The 1982/83 National Nutrition Survey", Papua New Guines Medical Journal 31:91-101. Howes, S. and J.Olson Lanjouw, (1998), "Does Sample Design Matter for Poverty Comparisons?", Review of Income and Wealth 44 (1): 99-109. Jenkins, C. (1996), "Poverty, Nutrition and Health Care in Papua New Guinea: A Case Study in Four Communities", Background Paper for Poverty Assessment. Lanjouw, P., and M. Ravallion, (1995), "Poverty and Household Size", The Economic Journal 105 (November): 1415-1434. Papua New Guinea, Department of Health, (1975), Nutrition for Papua New Guinea, Port Moresby, Papua New Guinea. Papua New Guinea National Research Institute (1997), Formal and Informal Social Safety Nets In Papua New Guinea, (Draft), Papua New Guinea. Papua New Guinea National Research Institute, (1996), Education and the Quality of Life in Papua New Guinea: Five Village Case Studies ", Background Paper for Poverty Assessment. Papua New Guinea National Statistical Office, (1996), Papua New Guinea Demographic and Health Survey 1996: National Report, Port Moresby, Papua New Guinea. Papua New Guinea National Statistical Office, (1980), Lowest Food Cost; A Statistical Tool. Port Moresby, Papua New Guinea. Ravallion, M. (1998), "Poor Areas" in A. Ullah and D. Giles (ed.) Handbook of Applied Economic Statistics, Marcel Dekker, New York, pp. 63-91. Ravallion, M. (1994), Poverty Comparisons. Chur, Switzerland, Harwood Academic Publishers. -57- POVERTYAND A CCESS To PUBLICSER VICES Ravallion, M. and Wodon, Q., 1999. "Does Child Labour Displace Schooling? Evidence on Behavioural Responses to an Enrolment Subsidy", Economic Journal, forthcoming. Ravallion, M. and G. Datt, (1991), "Growth and Redistribution Components of Changes in Poverty Measures" LSMS Working Paper No. 83, The World Bank. Ravallion, M. and D. van de Walle, (1991), "Urban-Rural Cost-of-Living Differentials in a Developing Economy", Journal of Urban Economics 29: 113-127. Ravallion, M. and L. Dearden, (1988), "Social Security in a 'Moral Economy': An Empirical Analysis for Java", Review of Economics and Statistics 70(1): 3 6-44. Rivers, D. and Vuong, A. 1988. Limited Information Estimators and Exogeneity Tests For Simultaneous Probit Models. Journal of Econometrics, 39(3): 347-366. Smith, T., J. Earland, K. Bhatia, P. Heywood, and N. Singleton, (1993), "Linear Gowth of Cildren in Papua New Guinea in Reation to Dietary, Environmental and Genetic Factors", Ecology of Food and Nutrition 31(1): 1-25. UNICEF (1998), The State of the World's Children, New York. Van de Valle, D. and K. Nead, (eds.), (1995), Public Spending and the Poor. Theory and Evidence, Johns Hopkins University Press, Baltimore. WHO (1983) Measuring Change in Nutritional Status: Guidelines for Assessing the Nutritional Impact of Supplementary Feeding Programmes for Vulnerable Groups, World Health Organisation, Geneva. World Bank (1999a), World Development Indicators, Washington D.C. World Bank (1 999b), Papua New Guinea: Country Economic Memorandum, Poverty Reduction and Economic Management Unit, East Asia and Pacific Regional Office, Draft Report, Washington DC. World Bank (1998), World Development Indicators, Washington D.C. World Bank (1998), Papua New Guinea: Education Sector Study, (Draft), Education Sector Unit, East Asia and Pacific Regional Office, Washington D.C. World Bank (1997), Papua New Guinea: Accelerating Agricultural Growth, Report No. 16737-PNG, Washington D.C. World Bank (1996), Philippines: A Strategy to Fight Poverty, Washington D.C. -58- POVERTYANDACCESS TO PUBLIC SER VICES World Bank (1995), Papua New Guinea: Delivering Public Services, Report No. 14414-PNG, Washington D.C. World Bank (1995), Vietnam: Poverty Assessment and Strategy, Report. No. 11 13442-VN, Washington D.C. World Bank (1992), Poverty Reduction Handbook, Washington D.C. World Bank (1990), Poverty, World Development Report, Washington D.C. -59- POVERTYAVD ACCESS TO PUBLICSER VICES ANNEX I. MEASURING THE STANDARD OF LIVING 1. This annex provides a summary on the design of the 1996 Household Survey which is the basis for the poverty profile in this report. It also describes how total household consumption was determined based on the data collected, what adjustments were made for spatial and temporal price variations and to adjust for differences in consumption between adults and children. The annex draws on a report by John Gibson and Scott Rozelle entitled "Results of the Household Survey Component of the 1996 Poverty Assessment for Papua New Guinea". The 1996 Household Survey of Papua New Guinea 2. The poverty analysis in this report is based on the 1996 Household Survey of Papua New Guinea. The survey was carried out as a joint exercise between the national university's research consultancy company (Unisearch) and an independent research institute (the Institute of National Affairs). Most of the survey interviewers were recent graduates from the university. Funding was provided by the Governments of Japan, Australia, and New Zealand, and the World Bank. The survey began in January and continued through all months of 1996 (interviews for 12 households in Port Moresby took place in early 1997). 3. The survey was comprehensive and multipurpose. In addition to estimating the value of household consumption, the survey also collected data on demographics, education levels and costs, income sources, the use of health facilities, birth details for young children, diet, housing conditions, agricultural assets, agricultural input spending, and participants' perception of their quality of life. Young children and their parents were weighed and measured. The various sections of the household questionnaires are reported in Table 1. Table 1: Sections of the Household Survey Questionnaires First Interview Second Interview Household Roster and Information Food Purchases Education Other Frequent Purchases Income Sources Own-Production of Food Health Gifts Received of Food and Other Goods Foods in the Diet (24 hour recall) Annual Expenses Housing Conditions Inventory of Durable Goods Agricultural Assets, Inputs and Services Inward Transfers of Money Anthropometrics Outward Transfers of Money Food Stocks (second interview also) Prices Daily Attendance (second interview also) Repeat of Anthropometric Measurements Quality of Life 4. The survey used a closed interval recall method, with households interviewed twice so that the start of the consumption recall period was signalled by the first interview. These two -60- POVERTYAND ACCESS TO PUBLICSERVICES interviews were designed to be two weeks apart, which is the usual length of the pay period in Papua New Guinea. Data were collected on all food consumption (36 categories) and other frequent expenditures (20 categories) during the recall period. An unbounded recall covered 31 categories of annual expenses, and an inventory of 16 durable assets collected information needed to estimate the value of the flow of services. 5. In addition to the household interviews, a community questionnaire was used to collect information on access to education, health, transport and communication services. Price surveys were also carried out, with prices collected for 35 items. Sample Design 6. Sample selection was based on a stratified two-stage random sample design. The 1990 Census of Population provided the area sampling frame. The sample was divided into two main strata: (i) the National Capital District (NCD), and (ii) all other areas of the country (excluding North Solomons province where the Bougainville crisis prevented surveying, and in fact had also prevented the 1990 Census). The sample was designed so that 240 households (one-fifth) would come from the NCD, and 960 households would come from the rest of the country. The main reason for this stratification was to allow different sampling fractions, with a five times heavier sampling rate in the NCD. This heavier sampling was designed to create a big enough sample in the NCD to allow comparison of poverty rates in 1996 and 1985-87; the estimates of poverty for the earlier period coming from the Urban Household Survey. The different sampling fractions for the two strata are controlled for by using weights (expansion factors) when forming national estimates. 7. A further stratification was applied to the non-NCD sample, mainly to improve the precision of sample estimates. All primary sampling units ("Census Units") were divided into one of 18 strata, based on three attributes: elevation, rainfall, and a proxy for "economic accessibility". Elevation was chosen because results from the 1982-83 National Nutrition Survey (NNS) suggest that malnutrition of young children is greatest in the highlands fringe zone of elevation 600-1200m. Only a small proportion (six percent) of the national population lives in this elevation range, so stratification and systematic sampling (selection at a fixed interval with a random start) was chosen to ensure that the right proportion of sampled households came from the highlands fringe. Rainfall was chosen because households in low rainfall areas may have more variable consumption patterns over the course of a year. Even within the same elevation and rainfall class, there is great variability, so "economic accessibility" was used as the third stratifying variable, to ensure the right proportion of both remote and accessible villages in the sample. 8. The proxy for "economic accessibility" was based on data from two sources. The first was estimates of mean cash income per household from agricultural activities, for the "agricultural system" in which the Census Unit is located. This came from a national agricultural mapping project. The second was estimates of the proportion of households who ate rice the -61- POVERTYANDACCESS To PUBLICSERVICES previous day, coming from the 1982-83 NNS sample of approximately 1000 villages, aggregated up to a Census Division average. The reason for using the two sources was that neither gave complete national coverage. The agricultural mapping project was not complete at the time the sample was drawn and the NNS sample villages were located in only 336 out of 536 Census Divisions (1980 boundaries). However, there was a significant correlation between agricultural income and rice eating, reflecting the fact that rice is perhaps the most common and widespread cash purchase of Papua New Guinean households. The equation: % eating rice = 0.050 * ag income + 0.095 * (ag income * Papuan region dummy) was estimated from the 240 Census Divisions where both types of data were available. This equation was used to predict the proportion of households eating rice in 162 Census Divisions where agricultural income estimates were available but NNS sample estimates were not available. Remaining areas were allocated into rice-eating categories based on expert knowledge. All urban Census Units were put into the category having the highest proportion of households eating rice. 9. The breakdown of the non-NCD sample into the 18 strata is reported in Table .2. The proportion in each stratum differs from the underlying household population proportions by less than one percentage point. Two strata had no sample villages selected because only a small proportion of households (0.5 percent) live in these areas (low and high rainfall highlands fringe, with high rice consumption). Table 2: Percent Distribution of the Sample Census Units by Stratum: Non-NCD Sample Lowlands Midlands Highlands Rice Rainfall (0-600m) (600-1200m) (>1200m) Consumption Low 2.5 1.3 2.5 Low (0-5%) (0-2500mm) 8.8 1.3 11.3 Medium (5-33%) 7.5 0.0 5.0 High (>33%) High 3.8 2.5 5.0 Low (0-5%) (>2500mm) 10.0 1.3 20.0 Medium (5-33%) 15.0 0.0 2.5 High (>33%) 10. Within each of the two main strata, Census Units (CUs) were selected with probability proportional to size. This means that villages with a greater number of households had a greater probability of being selected into the sample. Prior to selection the CUs were sorted by stratum and then by province (for the non-NCD sample), then by Census Division and Census Unit. Following this, CUs were selected with probability proportional to size using systematic selection. This meant that even though there was not explicit stratification by province or region for the non-NCD sample, the regional distribution of the selected CUs was roughly correct. The only important deviations were for the Papuan/South Coast and New Guinea Islands regions, whose small populations made rounding problems more important (e.g., proportionality called for 12.5 villages to be selected from the Papuan region, rather than the 13 actually selected). -62- POVERTYA,VDACCESS To PuBLiC SER ViCES 11. A subset of one-sixth of the selected CUs were chosen as a longitudinal sub-sample. Households were interviewed in these CUs at two different times of the year, roughly eight months apart, so that the correlation between expenditures in different months of the year for the same household could be measured. This correlation is needed for unbiased estimation of the between-households variance of annual expenditures when the data come from short period recall expenditures. The usual method of calculating the variance of annual expenditures, which is based on extrapolating annual expenditure from short period recall data, overstates the true variance. This can bias estimates of the incidence of poverty because the overstated variance increases the proportion of households whose expenditures fall below a certain absolute cut-off, e.g., a poverty line. 12. The revisit interviews for the longitudinal sub-sample also provide an opportunity to make the sample more representative of an entire year. The survey had initially been planned to take place only in the early months of 1996 because an October deadline had been set for availability of the results, to coincide with preparations for the annual PNG Budget. This deadline was subsequently relaxed after fieldwork had started but by then a greater number of interviews had been carried out in January and February than would have taken place if interviewing was spread evenly over the 12 months of the year. However, by using revisit interviews, which took place from August onwards, to replace first visit interviews, the over- representation of January and February in the sample is reduced, and interviews are spread more evenly over the year. This replacement is possible because the CUs to be revisited were chosen from a randomly selected group and the complete survey was repeated when the household was revisited. In fact, the interviews are possibly more accurate during the revisits than the initial visits because of the greater experience of the interview teams. 13. At the second stage of sampling, households were randomly selected from within the selected CUs. Selection was based on a listing of all households within the Census Unit. This listing was carried out by the survey team just after arrival in each CU. A fixed number of households were selected: six per Census Unit in the NCD sample, and 12 in the non-NCD sample. The reason for making this difference is that fieldwork costs (e.g., transportation) are lower in an urban area like the NCD, so the sampling efficiency advantages of a smaller cluster size outweigh the budgetary advantages of a larger cluster size. If the size of the CU, in terms of 1990 census households reported, were always equal to the 1996 household count, the selection of a fixed number of households in every CU would lead to a self-weighting sample. In practice though, these two numbers are likely to differ in most CUs, whether due to error in the census, error in the household count, or real population change. Thus self-weighting cannot be assumed and weights will be needed. See below. 14. A set of "reserve" households were also selected from the household listing in each CU, by the same method that was used to select the initial sample. These reserve households were used if any of the originally selected ones were unavailable or were unwilling to co-operate with the survey. Reserves were used only for households that were unavailable for the first interview. If the first interview was completed for a particular household but that household then became -63- POVERTYANDACCESS To PUBLIC SER VICES unavailable at the time of the consumption recall interview, no replacement was made. The interviewers just repeated their attempts to make contact with the household on the remaining days that they were present in the Census Unit, but in many cases this was unsuccessful because the residents had left for an extended period of time or sometimes indicated that they did not want to cooperate with the second visit of the survey. Hence there are different numbers of households in the sample, and different sampling weights, for data collected in the first interview compared with the second interview (Table 3). Table 3: Sample Size For First and Second Interviews of the 1996 Household Survey Initial visits Revisits (longitudinal (full sample) sub-sample) Household Consumption Household Consumption Region data data data data Total Number of households with complete data National Capital District 249 238 23 20 258 Papuan/South Coast' 154 146 21 20 166 Highlands 406 390 63 62 452 Momase/North Coast 286 279 64 57 336 New Guinea Islands2 108 104 22 20 124 Total 1203 1157 193 179 1336 Notes: Household data were collected in the first interview, consumption data in the second interview. Row totals are for households with complete consumption data. 'Includes Oro province. 2Excludes North Solomons province. 15. The data in Table 3 show that approximately four percent of households who had first interviews completed were unavailable for the second, consumption recall, interview. Some of the data collected during the first interview were used to test whether these absent households differ from the rest of the sample. On average there were 0.3 more persons resident in these households than in households with second interviews completed (and the confidence level that this difference is not just due to chance is 56 percent, which is too low for the difference to be statistically significant). The households without second interviews had an average floor area of 10.6 square metres per resident, compared with an average of 9.2 for the other households (confidence level is 78 percent).Self-selection out of the sample did not appear to be related to literacy levels; the proportion of residents in the households without second interviews who could read a newspaper was 0.43, while in the other households it was 0.41 (confidence level is 45 percent). There was no statistically significant difference between the two groups of households in the prevalence of stunting amongst children aged 0-5 years; the proportion with height-for-age z-scores (HAZ) less than -2 was 0.36 for the households without second interviews and 0.34 for the households with second interviews (confidence level is 20 percent). These small, and usually statistically insignificant differences, make it unlikely that self-selection out of the sample led to a bias in the consumption estimates collected from the remaining 96 percent of sampled households. -64- POVERTYAND A CCESS T0 PUBLIC SERVICES 16. The attrition of originally selected households was also apparent in the longitudinal sub- sample. Only 89 percent of the originally interviewed households in the 20 CUs of the longitudinal sub-sample were able to be re-interviewed, approximately eight months after the first interviews. Some of this attrition reflected the mobility of households and some reflected an unwillingness to further participate. This unwillingness was due partly to an unmet demand for monetary compensation to replace the gifts that were given to participating households. 17. Three of the selected CUs in the non-NCD sample were too small to provide a sample of 12 households. Two were combined with neighbouring Units and listing and household selection was carried out on the pseudo-Census Unit created from the joining of the two neighbouring Units. The third. was in an area that was too remote and sparsely populated to allow a pseudo- Census Unit to be created, so only 11 households were selected in this instance. Five other non- NCD CUs had first and second interview data for only 11 households rather than the planned number of twelve. In some cases this arose because the number of households absent or refusing exceeded the number of reserve households provided; in two further cases questionnaires were lost after the interviews were completed. For all of these CUs, sampling weights for the remaining households are modified so that the CUs are not under-represented (the weight is raised by a factor of 1.09=12/1 1). 18. The first cell in Table 3 shows that more households were surveyed in the National Capital District than the target of 240. The cause was the experiment of comparing diaries versus recall (described below) which required three households in each CU to be surveyed by the diary method and three by the recall method. Some interviewers became confused and issued diaries to all six households, so a further three households had to be chosen from the reserve list for the recall sub-sample. Rather than waste the information provided by the 'excess' households, they are included in the sample and the sampling weights are multiplied by a factor of 0.67 (=6/9) for households in the affected CUs. 19. The sample in the NCD was also distinguished by selecting some CUs in growth areas that had not existed at the time of the 1990 Census. The National Statistical Office recognised and mapped approximately 50 new CUs in the NCD, as part of their work for a Demographic and Health Survey. Four of these new growth areas (two in settlements and two in formal housing schemes) were selected as part of the sample of 40 CUs, using the same fixed interval sampling as the rest of the NCD sample. In the non-NCD sample no attempt was made to identify, and separately sample, new villages or new CUs not in the census. This does not necessarily imply significant under-coverage: new household settlements may often create new hamlets which are automatically absorbed in the nearest CU, or individual new households may be added to the CU simply because boundaries are large and vague. In any case the selected CU is always subject to a pre-survey household count and this provides a population up-date which should in most cases be complete. 20. Some rural CUs were too large to fully list. This was especially in rural areas of the Highlands, where the Census still operates by "line-up enumeration". Some of these CUs had -65- POVERTYAND ACCESS TO PUBLIC SER VICES over 500 households and these households were not located in villages, but were scattered widely in small hamlets over a considerable geographical area. Moreover, not every household living in the same area lines up together at the Census because clan lineage and marriage rather than physical location of residency appears to determine where a person lines up for Census enumeration. These large, spread-out CUs were partitioned into segments by survey teams in the field before listing, usually following expert local advice, and the segment to list and select households from was then randomly picked. In some cases, a separate household listing was carried out for the whole Census Unit, after the main survey had finished in December, to get more accurate estimates of the sampling rate in terms of household numbers, which is needed for the sampling weights (discussed below). 21. Surveying was not possible in six of the selected CUs, due to problems of tribal fighting, religious beliefs, and refusal of some communities to accept the survey. These six CUs and the replacements that were selected for them are listed in Table 4. The usual method of replacement was to form a list of CUs in the same province and stratum as the original Census Unit, and then make a random selection from that list, with probability proportional to size (1990 household numbers). In the case of Juguna, in the Baiyer area of Western Highlands, the replacement chosen by this method had itself to be replaced (by Aviamp) because of tribal fighting. The replacement made for Bitokara in West New Britain was done by the survey organiser in the field using expert local knowledge to try to recreate the attributes of the original selection (distance to main towns, access to health and school facilities, extent of cash cropping). This was achieved by combining a village with an adjacent mission (with school and health centre). Table 4: Sample Replacements for the Non-NCD Sample Census Replacement CU Census Reason for Province Original CU Count Count Replacement Oro Tiare 31 Nindewari 33 Religious beliefs Enga Yagonda 193 Papiyuku 264 Tribal fighting Western Highlands Juguna 61 Aviamp 269 Tribal fighting/cults Morobe Bupu 19 Muanu 53 Community refusal West New Britain Tiaruru 232 Sarakolok 323 Oil palm dispute West New Britain Bitokara 180 Gavuvu and 53 Land dispute Malalia mission Sampling Weights * A set of weights, which potentially can vary for each Census Unit in the sample, are needed when estimating statistics from the household data. These weights are needed to correct for the varying selection probabilities of areas and households in different parts of the sample. They can be regarded as made up of three components: * Correction for the differing sampling rates used in the strata at the area stage of sampling. This affects the two principal strata NCD and non-NCD. In addition, since areas are selected with probability proportional to "size", the size variable (census households reported) has to -66- POVERTYANDACCESS ToPUBLICSkRVICES be included in the weighting. Essentially the weights here are obtained as the sampling interval divided by the census size variable. * Correction for varying numbers of households selected in the CU. This is made up of two components: (1) the planned variation between 6 (in NCD) or 12 (in non-NCD), and (2) non- response, when the number of interviewed households within the CU differs from the desired number ( 6 or 12). These selected households have to be blown up to the number existing, in each CU, that is to the number listed during the household listing operation. The weight in each CU is the ratio of the number listed to the number interviewed. Separate weights are needed for the first visit (household data) and the second visit (consumption data). * A third weight is introduced in a few special cases, namely where the CU is subdivided into segments. If si is the number of segments created in the i-th CU, and one of these is selected at random, and if the listing covers only the selected segment, the weight must be si , the listing total Mi' being the segment listing total. However, if the household listing is extended to cover the whole CU with a CU listing total of M*", then the weight should be Mj"/Mi'. In the latter case the Mi' in the preceding term will of course cancel the Mi' in the final term. 22. Summarising, the weights to use for each Census Unit are given by the formula: Whi =[ lh / bhi I . I Mhi'/ Mhi I where the main strata (NCD and non-NCD) are indexed by h and Census Units by i, and Ih is the interval used in the first stage sampling to select CUs (given by the ratio of the number of households in that strata of the sampling frame to the number of CUs selected) bhi is the number of households interviewed in the Census Unit (where this can vary between first interviews (household data) and second interviews (consumption data)) Mhi is the number of households in the CU recorded at the 1990 Census (and this number, in conjunction with the interval 'h and the systematic ordering of the CUs, lead to the selection of the particular Census Unit into the sample) Mhi'is the number of households listed in the Census Unit or segment in 1996 (and this determines the selection probability of any given household in the CU). 23. The sample of CUs for the National Capital District was drawn from two sampling frames that were put together. The first sampling frame was the 1990 census, which had 468 CUs. The second frame was 50 new CUs that had been recently recognised (and mapped) by the NSO as part of the work requirements for the DHS. A systematic sample of 40 CUs was drawn from this combined frame. 24. The count of households done in 1996 by the interview teams was often substantially different from the 1990 Census count, ranging from being over four times larger, to being only one-quarter as large. Many of the CUs with apparently large changes in household numbers were in the Highlands, where the 1990 Census operated by 'line-up' enumeration. To ensure that the -67- POVERTYAND ACCESS TO PUBLICSERVICES apparent size changes were not caused by the method of splitting Census Units into segments, return visits were made to completely list several of the large Highlands CUs. The locations where this re-listing was carried out are noted in Table 1. 25. Even after re-listing, several CUs with apparently large shifts in household numbers, and therefore potentially large or small weights, remained in the sample. Some of the apparent declines in household numbers were able to be explained by factors such as people moving out of areas with severe tribal fighting or crime problems (e.g., a 70 percent decrease in Buk in Western Highlands). However, some of the instances of apparently rapid population growth seem more easily explained as errors of undercoverage in the 1990 Census (Table 5). In five of the six CUs that appear to have household numbers double between 1990 and 1996, the 1990 count of households was lower than the 1980 count, by a substantial amount in Won'em'men'ya (EHP), Wariman (ESP) and Kailge (WHP). It is only Erave Station (SHP) where the rapid growth between 1990 and 1996 is consistent with earlier growth rates. Table 5: Census Units With Apparently Rapid Growth in Household Numbers Estimates of Household Numbers Annual Growth Rates 1980 1990 1996 1990-96 1980-96 Won'em'men'ya (EHP) 28 13 56 0.243 0.043 Wariman (ESP) 64a 25 61 0.149 -0.003 Baisarik (Madang) 30 26 61 0.142 0.044 Semin (SHP) 170 165 380 0.139 0.050 Kailge (WIHP) 161 91 211 0.140 0.017 Erave St. (SHIP) 45 93 204 0.131 0.094 aBased on population, converted into household numbers assuming same average household size as found for the Census Unit in 1990. 26. Although the aim was to select CUs with probability proportional to size, there is no assumption that the measures of size M, are accurate. If they are, the sampling (random) error is likely to be smaller, which is an advantage, but whether they are accurate or not there will be no bias, provided the above weights are used. Therefore the inaccuracy of the 1990 census is not a cause for serious concern. Statisticians refer to the method used here as "probability proportional to estimated size" and it is just as "respectable" as the more ideal case of true PPS. Above all it is important not to try to "correct" the values M, after the sample has been used. The value M; which appears in the denominator of the weighting formula is introduced in order to cancel the varying probabilities which were used in selecting the CUs. This cancelling will only occur if we use identically the same values Mi in the weighting formula as were used in selecting the sample. That the size measures are inaccurate does not matter because they are going to be updated by introducing the values Mi', coming from the very latest count. Sampling Errors 27. Any estimates of averages, totals or ratios that are made from these survey data are subject to sampling error because the estimates come from a sample rather than the full -68- POVERTYANDACCESS TO PUBLiCSERViCES population of households in Papua New Guinea. The extent of sampling error is usually measured by the "standard error", with smaller standard errors indicating that the estimate from the survey is more precise. In calculating the standard errors, allowance must be made for the two-stage sampling, where the selection of the 120 CUs was the first stage, and the selection of households within these CUs was the second stage. This two-stage sampling is less efficient than simple random sampling in statistical terms (i.e., causes larger standard errors). This is because the households within a CU tend to have similar characteristics, so a sample drawn from them reflects less of the population's diversity than would a simple random sample with the same number of households. Therefore, if methods appropriate for calculating standard errors from a simple random sample are used, the standard errors that are reported will underestimate the true magnitude of sampling error. 28. The CENVAR software package, produced by the US Census Bureau, was used for calculating standard errors. This package is based on PC CARP (Cluster Analysis and Regression Package) software, and uses the ultimate clusters sampling model for calculating standard errors from complex survey designs with the algorithm based on Taylor series approximation (or "linearization"). Full details are available in the CENVAR documentation. Measuring the Value of Household Consumption 29. Five components of household consumption were measured by the survey: food consumption; other frequent expenses; annual expenses; consumption of non-housing durables; and consumption of housing services. The main components of consumption that were not measured are consumption of leisure, consumption of environmental services, and consumption of freely provided (or subsidised) public services. These items were excluded because of the conceptual and practical difficulties in measuring them. 30. The reference period for food consumption and other frequent purchases was set by the time between the first and second interviews. This bounded period between the two interviews was designed to be two weeks, but often differed from this, so the dates of each interview were recorded. The reference period for annual expenses was the past 12 months. The estimates of total consumption are made on an annual basis by extrapolating food consumption and other frequent expenses from the bounded recall period up to a full year. The inter-household variance of these estimates of annual consumption will be greater than if consumption was continuously measured over a full 12 month period because some of the shocks that occur in the short reference period would be evened out over the course of a year. The extent of this upward bias in the measures of dispersion is discussed in the sensitivity analyses, using results from the longitudinal sub-sample. 31. One problem in extrapolating short-term consumption from the bounded recall reference period, and then adding it to annual expenses, is that the size of the household may differ between the short and the long time periods. For example, a household may usually have five residents but during the bounded recall period two members leave to visit friends in another household for six days. This will cause the consumption of food and other short-term expenses to fall and when this is extrapolated to the full year it will lower the estimated annual value of the -69- POVERTY AND ACCESS TO PUBLIC SERVICES household's consumption. If total annual consumption was divided by the number of usual residents (five), the result would understate the true level of per capita consumption, because for the rest of the year the two people who were absent during part of the recall period are living in the household and contributing to its consumption. Therefore some adjustment needs to be made for the effect of absent residents, and for the opposite effect of short-term guests (and these corrections need to offset so that there is no double counting). 32. The specific adjustment used was as follows: the potential number of person-days during the recall period was estimated as the number of calendar days between the two interviews multiplied by the number of persons listed in the household roster. The actual number of person- days was measured from a schedule of absences and guests (present for at least seven days). An allowance was made for the smaller impact of absent or additional young children by estimating person-days in adult-equivalent terms, where each child aged six and below had a weight of 0.5 (see below for evidence on this equivalence scale). The ratio of the actual number of person-days to the potential number was formed and used to estimate the annualized equivalent of the bounded recall period consumption as: 365 Potential days (Food + Other Frequent Expenses) x x (Bounded Period) Actual days 33. If someone was listed as a usual resident on the roster but was substantially absent during the bounded recall period (absent at least 10 days), they were not counted when estimating household size, and therefore the above correction factor for absences was not applied to get the annual estimate of the recall period consumption. An exception to this rule was children who were away at school during the bounded recall period but whose school fees and other costs are still being met by the household. These children were considered to be part of the household and they therefore counted when estimating household size. Food Consumption 34. Food consumption was measured by a residual approach using the equation: Purchases + Own-production + Gifts received - Sales - Gifts given - Stock increases = Consumption. 35. This approach was chosen for two reasons. First, it was the method that had been used previously in PNG in the Urban Household Survey and the Household Expenditure Survey. Second, there is substantial interest in some of the components of consumption in their own right, especially the quantity and value of domestic food production and sales. -70- POVERTYAND ACCESS TO PUBLICSERVICES 36. The inclusion of sales of items that were either self-produced or purchased is an important part of the fornula for calculating consumption. Previous evidence from urban areas shows that ignoring this can substantially overstate the consumption of items like beer that are bought to be then resold on the black market, and also overstates the consumption of foods where a substantial amount of production is sold, e.g., betel nut. The importance of small-scale selling activity might be overstated if the purchases and sales of more formal business entities, like trade stores, were recorded as household transactions. Therefore, food purchases that were expenses of a formal business were ignored, as were the corresponding sales by these same household- business units. However, the expenses of an informal business, like flour purchases by a scone seller, were counted as food purchases, and sales (valued at input costs) were then subtracted to give the estimate of consumption. 37. The residual approach to measuring consumption does suffer some drawbacks. If forms of food disappearance that are not included in the formula -- such as plate waste and feeding household food to domestic animals -- are important, apparent food consumption levels will be overstated. The residual approach is also less intuitive than directly asking households what they eat, and there is international evidence that estimates of calorie intake based on what households eat are less extreme (higher for poor households and lower for rich households) than are measures of calorie availability based on an expenditure approach. Because the residual approach depends on six separate components there is also more scope for error. For example, having higher stocks of a food at the end of the recall period than at the start, but no record of purchases, production or gifts received to account for the higher stocks, could be due to error in any one of four possible parts of the interview. 38. The most difficult part of the food consumption equation to obtain reliable measures of was food production. Locally produced foods, which include root crops, bananas, vegetables and sago, are not sold by weight in Papua New Guinea. Instead, they are sold in fixed price bunches, bundles, heaps, or piles. Although some rural people are familiar with kilogram weights from sales of coffee or purchases of flour, rice or stock food, the densities of these items are much different to the densities of root crops. A further difficulty is the high volume of production, due to the bulky nature of the local staples (excluding sago); an average sized rural household can easily produce over 100 kg of root crops per week. In these circumstances it was considered unwise to ask respondents: "how many kilograms of food x did your household produce since my last visit?" Such answers reported in kilogram units would give the appearance of accuracy but this would just disguise the problem of measuring food production and force it onto the interviewers and the respondents. 39. Instead of relying exclusively on kilogram units, the question on food production allowed flexibility in the choice of units. Five units were available. The first was an empty 25 kg rice bag, which had three graduations ("1/4", "21'", ""3/4) marked on the outside. This bag was given to the household at the start of the survey and they were asked to put their produce into it during the recall period. This was the recommended unit for bulky crops. The other units were bunches and heaps; kilograms; singles, which were recommended for items like coconut and betelnut, and -71- POVERTYAND ACCESS TO PUBLIC SER VICES livestock (e.g., one pig, three chickens); and "other". Table 6 reports the frequency with which these units were used for recording the production of the various foods. Table 6: Frequency of Use of Units for Measuring Food Production Rice bags Bunch/heap Kilogram Singles Other Sweet potato 747 12 59 15 0 Cassava 233 11 26 56 0 Taro and Chinese Taro 409 5 36 59 0 Yams 170 2 13 34 0 English potato 115 0 1 1 0 Bananas 349 375 47 2 4 Sago 84 1 22 0 1 Sugar cane 30 30 8 444 2 Other fresh fruit 65 13 25 234 1 Coconuts 22 3 0 393 2 Peanuts 65 32 5 1 0 Aibika 179 226 28 27 55 Other greens, vegetables, nuts 459 194 35 69 48 Rice 1 0 0 0 0 Lamb and mutton 0 0 1 0 0 Chicken 1 0 5 47 1 Pork 5 0 19 62 4 Other meat (incl. Bush meat) 4 0 12 68 4 Fish (fresh, dried, shellfish) 41 4 18 114 3 Eggs I 1 0 63 1 Bete] nut 50 189 10 53 6 40. The 25 kg rice bag was the most frequently used measuring unit for the staple root crops and sago. The other important staple -- bananas - was measured with equal frequency in terms of either rice bags or bunches. Other foods where bunches or heaps were frequent units of measure are aibika (a leafy vegetable), a residual category for 'all other vegetables and nuts', and betelnut. Coconuts, fruit, sugar cane, pork, fish and other meat were all commonly measured by counting the number of single units produced. 41. Three of the measuring units -- rice bags, singles, and kilograms - were considered "reliable" because in principle factors to convert them into kilogram weights could be estimated and applied. The remaining two units - bunches/heaps and other - were considered uninformative as to the actual weight of food produced by the household. The problems that would have been encountered in assuming an average weight for a bunch or heap are best illustrated by the example of bananas: a bunch could refer to a 'hand' of approximately 10 single bananas weighing between one and two kilograms or else it could refer to a 'rope' or a 'branch' that is made up of several hands of bananas and can sometimes weigh 20 kilograms. 42. The first step in transforming bunches/heaps and other unspecified measures into kilogram quantities was to convert the "reliable" measures -- rice bags and singles -- into kilograms. The conversion factors used are reported in Table 7. The unit value was then calculated for each household that produced the food and reported production using one of the "reliable" measures. The median of these production unit values was then estimated for each -72- POVERTYANDACCESS To PUBLICSERVICES food in every Census Unit in the non-NCD sample, with the exception of meat and fish, where the median was calculated over all households in the region. The final step in imputing kilograrn quantities was to take the monetary value that households had assigned to the particular food that was measured with an uninformative quantity unit, and divide that value by the median unit value. If no median unit value had been calculated for that Census Unit - either because no other households produced the food or because the quantity measurement for all households who produced the food was an uninformative one - the imputation was based on the median unit value from the full sample of households in all non-NCD Census Units. In the NCD sample, the monetary values assigned to production of foods with uninformative quantity units were divided by market prices, in order to impute the quantity of production in kilograms. 43. The distinction made above between "reliable" and uninformative quantity units is blurred somewhat by the imprecision of the conversion factors estimated for rice bags and singles. The three foods having the largest number of measurements of both weight and rice bags were sweet potato, taro and bananas. The coefficient of variation (standard deviation relative to the mean) of the weight of a rice bag full of each of these foods was estimated as 0.3, 0.2, and 0.2. A further imprecision is that the relationship between how full the rice bag was and the weight is not linear. Instead a full bag would contain roughly one kilogram more than two half full bags, but the survey form just records the total number of bags produced and not whether that total is made up of very many fractions of bags or just a few full bags. The situation is even worse for the variability in the weight of single items: the average weight of a single sweet potato assumed by the unit value (and calorie) calculations was 0.45 kg, but weights found during the survey ranged from 2.5 kg to 0.1 kg. The estimated coefficients of variation for single sweet potato and cassava were 0.9, for taro 1.0, for yam 1.4, and for banana (single 'fingers') 0.7. Even items where individual units have a relatively regular size, like coconuts and betelnuts, had estimated coefficients of variation of about 0.4. The problems are bigger when assigning an average weight to an individual animal such as a pig, because whole pigs were not able to be weighed during the course of the survey. For items that are aggregations or residual categories (e.g., "other fresh fruit", "other greens, vegetables, and nuts") the average weight used for both rice bags and singles will be a very noisy measure of the actual weight of production of any particular household because we don't know if, for example, the item produced is a pineapple, whose average weight is about 2 kg, or a mango, whose average weight is just 0.3 kg. 44. The imprecision of the conversion factors in Table 7 means that for any given household, the estimated quantity of food production will contain a substantial element of random noise. If this noise increases the measured between-household dispersion, the number of households below a poverty line might be overstated. But sample summary statistics such as means and total should not be affected if the conversion factors are right, on average. This also applies to any measures derived from food quantities, such as per capita calorie availability. -73- POVERTY AND ACCESS TO PUBLIC SERVICES Table 7: Weight Conversion Factors Used for Measuring Food Production ---------- Kilograms per --- 25 kg rice bag Single Bunch/heap/other Sweet potato 19 0.45 2.0 Cassava 16 0.7 1.5 Taro and Chinese Taro 16 0.6 1.8 Yams 16 0.7 1.4 English potato 20 0.1 0.6 Bananas 12 0.16 1.6 Sago 30 n.a. 1.0 Sugar cane 20 2.5 5.0 Other fresh fruit 12 0.8 1.0 Coconuts 18 1.3 2.6 Peanuts 8 0.01 0.2 Aibika 4 0.06 0.2 Other greens, vegetables, nuts 10 0.4 1.0 Rice 20 n.a. 1.0 Lamb and mutton 15 20 n.a. Chicken 15 1.5 n.a. Pork 15 30 n.a. Other meat (incl. bush meat) 15 3.0 n.a. Fish (fresh, dried, shellfish) 15 0.5 1.0 Eggs 15 0.06 0.5 Betel nut 19 0.03 0.15 45. Although considered an uninformative unit of measure, there was one circumstance when weights had to be assumed for bunches/heaps and "others". This occurred when a household either gave away or sold food that it had produced, and the gift or sale was measured in units of bunches or heaps, while the original production had been measured in another unit. The final column of Table 7 contains the conversion factors that were assumed for this purpose. This should not cause large errors because the most common way of reporting gifts and sales was in the same units as production was reported, which made calculating the fraction given or sold easy, regardless of whether production was recorded using uninformative quantity units. 46. The computer programs that calculated food consumption used the median production unit value in three other ways, additional to converting uninformative quantity units into kilograms. These uses were: * Imputing the quantity of gifts received and the quantity of purchases, from the value of gifts received and value of purchases, if the quantities had not been recorded. Although in principle, a Census Unit median could be calculated for purchase unit values and gift unit values, for the locally produced foods there was much more information on production than on gifts received or purchases. * Assigning a value to beginning food stocks, if the household reported no production, purchases or gifts received of the food concerned. Usually there would be either -74- POJVERTYANDACCESS TO PUBLIC SERVICES production or purchases or gifts received, so the ratio of sub-total values and quantities could be used to implicitly give food stock changes the same household-specific unit value as was applied to production, purchases and gifts. Creating an alternative measure of the value of each household's production of each food, by multiplying the household's production quantity by the Census Unit median of the production unit value for that food. This alternative measure can only be created if "reliable" quantity units were used. 47. For all three of these uses, the regional average market price was used for foods that are not produced in PNG (including rice and sheepmeat, whose local production is almost zero). For the food consumption of households in the NCD sample, the average market price was used for all foods instead of the median production unit value. 48. Using the median unit value to adjust the value of production is appropriate if (i) it is assumed that the law of one price holds within CUs, and (ii) it is believed that quantities measured with "reliable" units are more accurate than reports of how much it would cost to buy or sell the quantity of a certain food produced by the household. The reason for doubting the reports of values is that very few households who are producing a food also buy that same food, so there may be a lot of misinformation about prices. Sales activity is also low so many households would not be able to infer purchase values from the prices they receive when selling the food. In these circumstances there may be a lot of random noise in the values attributed to self-produced foods. 49. The unit values for production of sweet potato were examined to see how important this random noise in the values attributed to self-produced foods may be. Attention was restricted just to households who used the 25 kg rice bag as the unit of measure and just to CUs where unit values for the rice bag were available for at least four households (n=694). The within Census Unit coefficient of variation in the value that each household placed on a rice bag full of self- produced sweet potato was 0.45, which is twice as high as the variation in the measured weight of rice bags of sweet potato. These households were ranked within their CUs, according to the value they placed on a rice bag of sweet potato. The ratio of the highest unit value to the lowest unit value was formed, and the median value of this ratio was estimated to be four. Hence, there appears to be large variation in the value that households within the same Census Unit placed on a given volume of production. Some of this variation will be caused by the differing weight of sweet potato that each household's rice bag contained, but this source cannot explain it all. Some more of the variation may be due to differences in the quality of sweet potato that each household was producing. However, at least some of the remaining variation would seem to be due to differences in opinions about prices. 50. The variation in sweet potato unit values found within CUs is likely to occur with other self-produced foods as well. It may seem unreasonable that two households, who produce the same quantity of a food in the same location, can have that production valued differently. A household might fall below the poverty line just by being too pessimistic when valuing their own food production because they think prices are lower than they truly are. Using the adjusted values -75- POVERTYANDACCESS To PUBLICSERVICES of food production, based on the CU median unit value, would seem to be an improvement. However, two factors caution against too much reliance on this adjustment: (i) it cannot be applied to food production measured by uninformative units, and (ii) it places a lot of faith in quantity measurements, which, especially for animal products, are based on little more than values taken from the PNG literature and informed guess work regarding conversion factors. 51. The effect of adjusting the reported value of food production by basing it on the Census Unit median unit value is to cause a slight reduction in measured inequality in the value of food consumption. The Gini index for the value of food consumption is 0.44 when the value placed on self-produced food is not adjusted. The Gini index falls to 0.42 if self-produced foods are valued at the CU median unit values. Therefore, variation in unit values within CUs, which seems to be due at least partly to differing opinions about prices, is not a big contributor to measured inequality. 52. The above discussion has illustrated the imprecision that exists in the food production data. The aim of the discussion was to give readers a sense of the possible non-sampling errors that affect any results using the food production data. It is difficult to know if these errors are any larger than occur in household surveys in other countries where the main foods produced are bulky root crops because such detailed information is usually not published. One test of the reliability of the data on food production is to see how plausible the estimates are when aggregated to the national level (Table 8). However, the available data on aggregate food production (e.g., from the FAO Production Yearbook) is known to be unreliable because they are Table 8: Estimates of Annual (Household) Food Production in Papua New Guinea Quantity Value '000 t std. error kg/person K (mil) std. Error Sweet potato 1286 151 264 290 38 Cassava 124 25 25 32 6 Taro and Chinese Taro 314 52 64 97 13 Yams 143 31 29 47 12 Bananas 413 46 85 150 17 Sago 95 22 19 26 8 Coconut 195 21 40 30 4 Pork 60 11 12 243 47 Chicken 4 1 1 20 7 Other meat (incl. Bush meat) 16 4 3 26 7 Fish (fresh, dried, shellfish) 50 12 10 60 17 Sugar cane 190 19 39 29 4 Other fresh fruit 59 10 12 21 4 Peanuts 21 8 4 56 27 Aibika 40 5 8 18 3 Other greens, vegetables, nuts 264 30 54 71 10 Irish potato 10 4 2 5 2 Betel nut 49 9 10 78 18 -76- POVERTYAND ACCESS TO PUBLIC SER VIcES simply extrapolations from the 1961-62 Survey of Indigenous Agriculture (the standard errors for this survey were over 25 percent. A stronger test of the plausibility of the current estimates will come when final results are available from the national agricultural mapping project. 53. The dominance of sweet potato in local food production is evident from Table 8. The quantity of sweet potato production is three times higher than the next highest food, bananas, and it is also the most valuable food crop. The other important foods in terms of quantity are taro and Chinese taro (which in a mistaken decision early in the survey planning were combined as one recall item to shorten the list of foods), coconut, sugar cane, yams, cassava, and the residual category of vegetables. In terms of value, pig production was second to sweet potato, followed by bananas, taro and Chinese taro, and betelnut. The aggregate value of food production was approximately K1.3 billion (standard error of KI 14 million). If the value of firewood and tobacco production was included, the total value of household production would be almost K1.6 billion per year. Other Frequent Expenses 54. Measuring consumption of non-food frequent expenses (e.g., gambling, P.M.V. fares, kerosene, tobacco) was easier than measuring food consumption because, with the exception of leaf tobacco and firewood, consumption was entirely from purchases and gifts, rather than from own-production. The other simplification was that quantities were not required. 55. The only problem encountered was assigning values to firewood, because in some areas firewood is not bought and sold, so respondents had difficulty assigning a value to the firewood they harvested or received as gifts. A regression equation was used to impute firewood values if they were missing, using three assumptions to guide the specification of the equation. More firewood, and hence a higher value, is needed if: * there is more food to be cooked, * there are more people in the household to cook for and keep warm, and * the outside temperature is colder. The equation used was: Firewood Value = 39.1 (Number in household) (4.0) + 0.017 (Value of Food Purchased & Produced) (2.8) + 61.5 (Number in household * Highlands dummy) (4.6) + 0.027 (Value of Food * Highlands dummy) (4.0) N=821 R2z=0.30 (t-statistics). -77- POVERTYAND ACCESS TO PUBLICSERVICES Annual Expenses 56. Annual expenses were estimated from an unbounded 12 month recall, with consumption derived from three components: Value of Purchases - Value of Gifts Given + Value of Gifts Received. Spending on some of the items on the annual expenses schedule was also recorded in some sections of the first interview questionnaire (e.g., school fees, charges for water) and these data were used to check the consistency of results. Any purchases of an item on the annual expenses schedule during the bounded recall period were also recorded, and these short-term data were used to assess the consistency of the reported values given in the long term recall. 57. One of the categories of annual expense items that had data collected - loan payments and finance fees -- is not included in measured consumption. Many of the data reported for this category appeared to be for repayments of principle as well as payments of interest. The appropriate component for measuring consumption is just the interest payment because this is the cost to the household of bringing forward in time the acquisition of assets. If principal repayments were included, measured consumption would have been overstated. 58. Two other notable items of the annual expenses are spending on bride price ceremonies, and spending on death feasts. Consumption of these two items was defined in the conventional way, so for example, contributing to the bride price expenses of someone who lives outside the household subtracts from the household's measured consumption. It could be argued that in Papua New Guinea contributing to the bride price and death feast expenses of kinsmen is a requirement of participating in the everyday life of the community. Hence, these expenses on people living outside the household might be a legitimate component of consumption and of a consumption-based poverty line (World Bank, 1990). However, that line of argument was not followed when constructing the annual expenses variable. Thus spending on bride price ceremonies and death feasts for people outside the household was treated in the same way as other items purchased and then given away as gifts, by subtracting it from the households measured consumption. Consumption of Durables 59. An inventory of 16 durable assets was used to collect data on the purchase price (or value if it was a gift) and date of acquisition of each asset, and the price that it would realize if sold. Purchase prices were inflated to 1996 terms using the movement in the Consumer Price Index. The straight-line annual depreciation was calculated as: (Purchase Value - Estimated Sales Value) / (1996 - Year of Acquisition). If components of the formula were missing, a default value of annual depreciation was assigned based on the average depreciation for that asset calculated from the sample of households with complete data. Table 9 contains these default depreciation values along with estimates of the ownership rates for each durable in each region and the country as a whole. -78- POVERTYANDACCESS TO PUBLICSERVICES 60. The questions in the inventory of durable assets used ownership rather than use as the criteria for including or not including a particular asset. This may result in underestimates of consumption for people living in households where chattels are provided (e.g., chairs, refrigerators) but it is hoped that the rent paid on the house would include a component due to the provision of these assets. Table 9: Details on Durables Values and Estimated Ownership Rates Default Percentage of Households Owning the Durable in: DURABLE ITEM Deprecn PNG NCD Papua Highlands Momase NGI Chairsandtables 20 18.4 47.5 13.7 15.1 13.6 38.5 Primus or portable stoves 12 20.8 62.8 24.1 20.2 11.5 25.5 Kerosene lamps 3 74.7 43.3 79.6 67.5 80.9 91.7 Refrigerators or freezers 120 5.2 44.3 6.5 1.8 4.3 2.3 Sewing machines 23 18.0 53.6 27.8 14.9 13.5 12.9 Generators 240 2.3 6.7 4.9 1.8 1.0 1.9 Guns 60 6.7 6.0 10.5 9.5 3.5 0.0 Traditional canoes 18 7.6 2.1 21.4 0.0 6.0 22.7 Metal/fibreglass dinghies 280 1.0 4.2 1.1 0.0 1.0 2.9 Outboard motors 310 1.4 4.7 4.2 0.0 1.0 1.9 Bicycles 20 10.4 22.9 15.6 4.7 10.8 18.6 Motorcycles 650 0.5 0.1 0.6 0.9 0.0 0.0 Cars or pickup trucks 1200 4.4 30.4 6.2 3.2 1.7 2.9 Cameras 10 13.8 56.7 17.2 10.5 11.3 9.8 Radios/cassette players/stereo 25 35.0 75.0 42.1 27.8 33.2 40.5 Television/video equipment 95 7.3 59.2 9.6 2.4 7.3 1.4 Note: Default depreciation is the value of annual depreciation imputed for an item if the actual depreciation was not able to be calculated due to missing data on acquisition dates or prices. The Consumption of Housing Services 61. Houses were divided into two classes: "mainly traditional" and "mainly non-traditional" based on the building materials used in the floor, roof, and walls. All houses in the National Capital District sample were allocated to the "mainly non-traditional" class, and a further 115 houses from the sample outside the NCD were also allocated to this class. The remaining 839 "mainly traditional" houses were further divided into those with a traditional roof (n=731) and those with an iron or metal roof (n=108). The "mainly non-traditional" houses were divided into a high/medium-cost group (n-=248) and a low-cost/village/urban settlement group (n=1 16). 62. For owner-occupied dwellings, the economic life of the house is needed so that capital costs can be depreciated over a certain number of years in order to estimate the annual flow of housing services. The life expectancy of a mainly traditional house with a traditional roof was estimated as 3.5 years, based on the age structure of the houses in the sample. The life expectancy of a traditional house with a metal roof was estimated as 4.8 years. An upward -79- POVERTYAND ACCESS TO PUBLICSERVICES adjustment in the estimate of the economic life was made to reflect the fact that some housing materials are salvaged and reused in the next dwelling (e.g., sound roofing iron can be transferred from a dilapidated house to its replacement). The economic life of a traditional house with a traditional roof was set at four years, and a traditional house with a metal roof was set at six years. The economic life of a high-cost non-traditional house was estimated as 20 years and a low-cost non-traditional house was estimated to have a life of 10 years. For all house types, if a dwelling was observed to have a greater age than the expected economic life, the actual age was used when calculating annual depreciation. 63. The capital cost of traditional houses was established from two questions. The first asked how much it cost to build or buy the house. These costs were inflated to 1996 terms, using the movement in the Consumer Price Index since the year of construction of the house. The second question asked about the number of days of unpaid labor that were used to build the house. Ideally, person-days spent on house building would be converted into kina terms using an opportunity cost of labor that was specific to each household. However the survey did not include questions on incomes and earnings, so external data on the opportunity cost of labor were used. Specifically,. estimates of mean annual cash income per household from agricultural activities, for the agricultural system that each census unit was located in, were used. These estimates ranged from K25 per household to K1200 per household, and were divided by 200 to get estimates of the opportunity cost of labor per person per day. For CUs that were not assigned to agricultural systems (urban areas, rural non-village and plantations), the opportunity cost of labor was set at K6 per day. Obviously, these estimates of the opportunity cost of labor are only rough approximations. A further source of underestimation of capital costs is that unpaid for materials used in construction (e.g., bamboo, sago thatch) did not have a value imputed for them (but see the sensitivity analyses). 64. The capital cost of some traditional houses had to be imputed because the data on unpaid labor used constructing the house were not reliable. For example, 20 people may have spent three days building the house, and the data reported are that the house was built with three days of unpaid labor. The answer that was really wanted was 60 person-days of unpaid labor but the survey question was not adequately worded to get this answer in all cases. Unreliable data were identified from scatter plots of the capital cost of the house (which included unpaid labor days valued at opportunity cost) against floor area, plotted separately for each census unit, and also taking account of construction materials. Data for 170 houses were deemed unreliable using this method. A cost function was estimated using the data from the remaining 669 traditional houses: ln (Capital Cost) = 1.29 + 0.024 (Floor Area) - 0.000069 (Floor Area)2 (6.08) (8.66) (6.13) + 0.062 (Number of Rooms) + 0.877 (Iron Roof) (2.41) (8.15) - 0.746 (Temporary Walls) - 0.012 (Dirt Floor) (2.17) (0.14) + cluster dummy variables R2=0.79 -80- POVERTYAND AccEss TO PUBLICSERVICES where In is the natural logarithm and t-statistics are in (). This equation was used to impute the capital cost of the 170 houses where the original data were unreliable. 65. Estimating the capital cost of non-traditional houses was more difficult. Fewer of the occupiers were owners, especially in urban areas, so they did not know details about the age of the dwelling or construction costs. In principle, this should not matter because if these occupiers are paying rent, their rental payments serve as an estimate of the consumption of housing services. However, in Papua New Guinea the rental market is distorted by shortages of housing (due partly to collective clan ownership of land) and employer provision of housing. Survey results reported by Dahanayake (1991) suggest that two-thirds of middle-class urban dwellers live in employer-provided housing and, on average, have 94 percent of their rental costs subsidized. These subsidies can cause enormous variations in measured consumption levels. For example, in one high-cost Census Unit in the NCD sample, the occupiers of one house were paying K1700 per month rent while the occupiers of another house of similar age, size and quality were paying only K40 per month rent. Therefore it is important to have estimates of the capital cost of the house so that the estimated annual depreciation could be compared with the annual rent paid. If the rent paid was much less than the estimated depreciation, it was assumed that a rental subsidy was being provided and the estimated value of annual depreciation was used instead. 66. The capital cost of non-traditional houses was established by two different methods. For low-cost/settlement houses, the rules used for traditional houses (described above) were applied because occupiers had usually built the house themselves or purchased it on an open market, and this made them more informed about costs. But there were 15 houses where cost data were unavailable so imputation was based on the following cost function: In (Capital Cost) = 6.39 + 0.030 (Floor Area) - 0.00016 (Floor Area)2 (17.8) (2.54) (2.04) + 0.100 (Number of Rooms) + 0.677 (Cement walls) (1.23) (3.20) + cluster dummy variables R2=0.33 where t-statistics are in ( ). 67. For medium/high-cost houses where the occupiers were unable to give estimates of costs, an imputed cost was based on the construction costs of a house of similar size, built in a similar era, by the National Housing Commission. The Housing Commission has been the main provider of urban housing and it has used a set of basic designs which are usually identifiable by the floor area of the house. Housing Commission costs were used rather than estimating a cost function from the group of medium/high-cost houses in the sample with complete cost data because there was a much higher ratio of houses with missing data to those with complete data than in the traditional and low-cost non-traditional sub-samples. 68. Specifically, for the medium/high-cost houses, data on the construction costs of standard Housing Commission houses in several different years were obtained and used to estimate a -81- POVERTYAND A CCESS To PUBLIC SERvICES regression of real construction costs, in 1996 terms, on floor area (an intercept and quadratic term were found to be statistically insignificant): (CPI1996\ (Construction Cost)t x ( _) = 816.2 (Floor Area) R2=0.86 (27.4) where the t-statistic is in ( ). This equation was used to impute the construction cost of any medium/high-cost houses that lacked cost data but had estimates of floor area. The cost that is being imputed is for a basic design, so can be considered as a lower bound estimate of the actual construction cost. 69. The annual depreciation for all houses was calculated using a straight-line method, with the capital cost in 1996 terms being divided by which ever was the greater of either the expected economic life for the type of house or the actual life of the house. If the calculated annual depreciation exceeded, by more than 150 percent, any annual rent payments that were reported, the estimated depreciation was used as a proxy for the true economic rental of the house. The annual depreciation was also used as the estimate of the consumption of housing services for all owner-occupied dwellings. Comparing Consumption Between Households 70. To correctly indicate living standards, the consumption estimates require deflation to capture differences in needs and differences in prices faced by households. Spatial price differences are controlled for using a price index that is derived from the regional poverty lines. This price index therefore measures the relative cost of meeting a "basic needs" standard of living in each region. Another price index is constructed to account for temporal variation in nominal price levels. Differences in needs are controlled for using estimates of child costs and estimates of economies of household size to create an adult equivalence scale. Spatial Price Index 71. The ideal way to control for spatial differences in the prices facing households is to calculate a "true cost-of-living index". This true cost-of-living index is based on the expenditure function, c = c(u, p), which gives the minimum cost, c for a household to reach utility level u when facing the set of prices represented by the vector p. For two, otherwise identical households, one living in the base region and facing prices p°, and the other living in another region facing prices pl, the true cost-of-living index is: True cost - of - living index = C(W, p') c(ii, po) which can be interpreted as the relative price in each region of a fixed level of utility. Although this is the ideal spatial price index, it is not commonly calculated, even in developed countries. Holding households at a constant level of the (unobservable) utility, and controlling for their -82- POVERTYAND A CCESS To PUBLICSER VICES varying characteristics, requires econometric estimation of an integrable demand model, which is technically demanding. 72. The usual approach to controlling for spatial price differences is to use a price index formula that approximates the true cost-of-living index. A common choice is the Laspeyre's index, which calculates the relative cost in each region of buying the base region's basket of goods. In other words, it calculates the price of a fixed bundle of goods rather than the price of a fixed level of utility. One problem with the Laspeyre's index is that it overstates the cost-of- living in high price areas because it does not allow households to make economising substitutions away from items in the basket of goods that are more expensive in their home region than they are in the base region. Technically, the Laspeyre's index assumes that the utility function is Leontief. An occasionally used improvement is the superlative price index (e.g., Fisher, T mqvist), which will closely approximately an exact cost-of-living index for any utility function. 73. The Laspeyre's and the superlative price index formulas require a full set of prices for all items in the consumption basket. Household surveys are typically not able to collect prices for all consumption items, and the current survey was no exception. In fact, 28 percent of the average household's consumption basket was not priced (Table 10). One solution would be to make assumptions about the regional pattern of the missing prices. For example, Glewwe (1991) used the price of canned tomato paste in place of the missing prices for all non-food items in the Cote d'Ivoire, because the only reliably collected prices from the Cote d'Ivoire Living Standards Survey were for food items. 74. The solution adopted to the problem of missing prices in this study was to derive the spatial price index from the regional poverty lines. Unlike price index formulas, poverty lines can be properly calculated when there are missing non-food prices (see Part II). The price index was calculated from the poverty lines for five regions of Papua New Guinea: the National Capital District (NCD), Papuan/South Coast (which includes Oro Province), the Highlands, Momase/North Coast, and the New Guinea Islands. The price index combines rural and urban areas within each region because usually there was only one urban Census Unit per region (there were no rural CUs in the sample for the NCD). The following paragraphs describe the price data used in the construction of the poverty lines and spatial price index. 75. During the course of the household survey, price data were collected from each of the 120 CUs that were in the sample. The price survey was also repeated when CUs that were part of the longitudinal sub-sample were revisited. Prices were collected for food and non-food items. Two types of price survey were carried out by the interview teams. The prices of packaged food items (e.g., rice, sugar, tin meat, beverages) and non-food items (e.g., soap, kerosene, cigarettes) were collected from the two main trade stores or supermarkets used by the households in the Census Unit. The prices of locally produced foods were collected from the nearest local market, with the price and weight of up to six different lots of each food being recorded. The market price survey was carried out on two different days, to help improve the precision of average prices, so potentially the weights and prices of 12 lots of each food were recorded for each Census Unit. -83- POVERTYAND ACCESS TO PUBLIC SERVICES However in some areas markets were held only infrequently so the price survey could be carried out only once. In a few cases, interviewers made special return visits to an area to coincide with market day. Table 10: Average Budget Shares of Consumption Items That Were Not Priced Firewood 6.182 Jewellery, watches, clocks, umbrellas, bags 0.173 P.M.V. fares 3.129 Medicines (modem and traditional) 0.162 Wedding expenses and brideprice 2.443 Life, health and personal effects insurance 0.128 Meals consumed away from home 2.094 Entertainment equipment (toys, books,..) 0.127 Depreciation of non-housing durables 1.994 School stationary, text books & uniforms 0.126 Adult's clothing (new and used) 1.468 Charges for sewerage, garbage and water 0.112 School fees (tuition and boarding) 1.428 Other home maintenance products 0.092 Airfares, ship fares, car hire 1.162 Domestic services 0.091 Compensation payments 1.061 Vehicle registration and insurance charges 0.090 Gambling (except lottery tickets) 1.054 Toilet paper 0.084 Children's clothing 0.953 Diesel 0.078 Burial and death feast expenses 0.621 Lottery tickets 0.073 Linen 0.534 Entrance fees for films and videos 0.066 Kitchen utensils 0.466 Telephone charges 0.058 Medical fees 0.348 Other household equipment and services 0.054 Footwear 0.343 Fumishings 0.042 Household and garden tools 0.285 House repairs and maintenance 0.038 Vehicle repairs and maintenance 0.222 Entrance fees to sports matches 0.036 Electricity and gas charges 0.201 Stamps and postage fees 0.029 Newspapers and magazines 0.183 Holiday accommodation and tours 0.022 Other personal care products 0.173 Land rent and land taxes for the section 0.013 Kitchen electrical equipment 0.010 Total budget share 28.191 76. The average price of each item was calculated for each Census Unit. For items from the trade store survey, this usually meant taking an average of two prices, but if the item was only available in one store the single price was used. Sometimes the items available were not of the desired specification (e.g., the specification for tinned meat was a 340g can of "Ox and Palm" brand, but sometimes the price collected may have been for a 200g can, or for a 340g can of a different brand). When the price for the non-specification good was the only one available in a particular Census Unit it was used to predict the missing price of the good of desired specification, if price relativities between that specification and the desired specification were able to be estimated from the rest of the sample. Otherwise, the missing price was predicted from a cross-sectional regression of the price of the desired item on the prices of all other trade store goods, for the sample of CUs where the price was not missing. The logic of this regression is that the spatial pattern of prices for trade store goods reflects transport costs. -84- POVERTYANDACCESS TOPUBLICSERVICES 77. An exception to these procedures was made for the CUs in the NCD sample, where a simple average was taken of the prices for each item collected in all 40 CUs. The reason for this exception was that people who live in one part of the capital city can easily buy in other parts of the city. Therefore the spatial distribution of prices inside Port Moresby is less relevant than is the spatial distribution of prices within the other, larger, regions. 78. If a locally produced food was not able to be priced in a particular Census Unit during the survey period, the price was treated as missing data and was not predicted. There are two reasons for treating the prices of locally grown foods in a different manner to the treatment for packaged food items. First, the reasons for not observing a price are more complex: the Census Unit may lie outside the environmental range of production for the food, or the crop may be available in only certain seasons of the year, or in more remote areas the food could be so widely grown that no market exists because demand is meet by own-production. Second, there is no single factor like transport costs that could be used for predicting the price. 79. The collection of prices spanned 12 months so there will be temporal as well as spatial variation. The temporal variation should not affect the average prices calculated for the four regions outside of the NCD because surveying progressed at approximately the same pace in those four regions. However, the NCD survey did not start until April, three months after the first prices were collected in the other regions. The quarterly Consumer Price Index, calculated by the NSO, was used to correct for the different distribution across time of the samples. Specifically, a weighted average of the CPI was taken for the non-NCD sample, where the weights were the number of CUs surveyed in each quarter. A similar average was taken for the NCD sample and the ratio of the two was estimated as 1.0051. Therefore, all prices collected in the NCD survey were reduced by a factor of 1/1.0051, so as to be directly comparable, timewise, when comparisons were being made with the prices collected in the other regions. 80. Some items were priced at regional level because market prices for these items are not widely available in all Census Units. This included formally marketed items that are sold mainly in urban areas (such as bread, lamb and mutton) and informally marketed items produced in only a few areas but sold more widely throughout urban markets (e.g., potatoes). 81. One important item of consumption whose price is not directly observable, and was therefore not collected during the price survey, is the consumption of housing services. Instead, the implicit price of housing services in each region was computed by regressing the logarithm of annual rents (actual and imputed) on a vector of housing characteristics. The vector of characteristics included: the floor area of the house, the number of rooms, the material of the walls (four dummy variables), the material of the floor (four dummy variables), and the material of the roof (two dummy variables). The equation also included dummy variables for the four regions outside of the NCD. The sample excluded dwellings whose capital cost was imputed (equations in paragraph 65 and 67) and dwellings with subsidized rents and insufficient information to estimate annual depreciation. In total, 918 observations were available (76 percent of the total). The results were: -85- POVERTYAND ACCESS TO PUBLICSERVICES In (Annual Rent) = 5.78 + 0.036 (Floor Area) - 0.0001 (Floor Area)2 (27.66) (9.55) (5.29) + 0.071 (Number of Rooms) - 1.142 (Papuan dummy) (2.33) (7.18) + 0.256 (Highlands dummy) - 0.767 (Momase dummy) (1.50) (4.59) - 0.735 (New Guinea Islands dummy) (4.12) + wall, floor, and roof dummies R2=0.74 82. The exponential of the coefficient on each regional dummy variables gives the price of housing services in that region (with NCD=1), holding housing quality constant. The implied rent price index is: NCD 100; Papuan 32; Highlands 129; Momase 46; New Guinea Islands 48. There appear to be large regional differences in the (constant-quality) price of housing services. The high price of rents in an urban area like the NCD is not surprising, and is consistent with evidence from other countries. The high price of housing services in the Highlands is more surprising, and may be caused by: (i) the need for formal building materials (e.g., nails, roofing iron) to be imported from the coast, (ii) the relative scarcity of timber in dry, grassy, highland regions, and (iii) the need for more sturdy and weatherproof construction due to the cold climate. 83. The average budget shares of the items that were able to be priced (including the implicit price of housing rent) are reported in Table 11. All food items specified in the survey, except meals consumed out of the home, had regional average prices calculated. Table 2 lists the regional average prices. 84. The spatial price index derived from the regional poverty lines is reported in the first column of Table 12. The region with the highest poverty line is the National Capital District, and this was set equal to 100 so as to form the price index for the other regions. The second column of Table 12 is identical to the first, except that the price index has been re-based, so that the national average price level is equal to 100. The third column of the Table reports a price index for food consumption, based on the food poverty lines. The fourth column gives the re-based food price index, making the national average the base. 85. The National Capital District has the highest cost-of-living, at almost twice the national average (Table 12). The lowest cost-of-living is in the Momase/North Coast region. The Highlands also has a below average cost-of-living. Price levels are slightly above average in the New Guinea Islands, and higher still in the Papuan/South Coast region. The same ranking of regions occurs when using the food price index. -86- POVERTYAND ACCESS TO PUBLICSERviCES Table 11: Average Budget Shares of Items With Prices Sweet potato 10.599 Sugar 0.959 Banana (cooking and sweet) 6.136 Aibika 0.956 Taro and Chinese Taro 4.743 Soft drink 0.929 Rice 4.107 Butter, margarine, oil and dripping 0.883 Housing rent (actual and imputed)a 4.069 Soap 0.876 Betel nut, lime and mustard 3.450 Other fresh fruit 0.843 Pork 2.799 Biscuits 0.734 Other greens, vegetables and nuts 2.720 Cigarettes 0.734 Sago 1.970 Flour 0.724 Yams 1.924 Peanutsa 0.696 Chicken 1.909 Batteries 0.687 Tobacco and mutrus 1.786 Tea, coffee, milo 0.480 Tinned fish 1.747 Salt, pepper, spices, saucesb 0.444 Fish (fresh, frozen, dried, shellfish) 1.480 Laundry powder 0.415 Beer 1.480 Milk (liquid, powdered, canned) 0.374 Coconut 1.468 Breadb 0.362 Tinned meat 1.435 Petrol 0.336 Cassava (Tapiok) 1.191 Other diary and cereal products and eggsb 0.244 Sugar cane 1.176 Snack food (Twisties, chewing gum, etc)b 0.233 Meat (except pork, chicken & lamb)b 1.167 Other alcoholb 0.208 Lamb and muttonb 1.095 Potatob 0.167 Kerosene 1.077 Matches 0.138 Total budget share 71.809 almplicit price index computed at regional level. bThese items not available to be priced in all Census Units, so prices collected at regional level. Table 12: Spatial Price Index Derived From Poverty Lines All Consumption Food Consumption National Capital District 100.0 195.3 100.0 180.9 Papuan/South Coast 63.6 124.3 71.9 130.0 Highlands 50.1 97.9 52.9 95.7 MomaselNorth Coast 4 6. 0 70.4 40.0 72.4 New Guinea Islands 54.4 106.3 60.0 108.5 Papua New Guineaa 51.2 100.0 * 55.3 100.0 aPopulation weighted average. 86. It should be stressed that the spatial price index measures the relative price of a "basic needs" standard of living in each region. It is thus weighted towards the poor, rather than towards the average household. This is appropriate when making inter-household comparisons for poverty analysis but there may be some other purposes for which it would be appropriate to use another spatial price index. The regional average budget shares, which in conjunction with regional average prices are needed for the calculation of a spatial price index, are reported in Table 3. -87- POVERTYAND ACCESS TO PUBLIC SER VICES Temporal Price Index 87. Ideally temporal price variation should be controlled for at the same time that spatial price variation is controlled for. However, the composition of the sample changed from month to month, tending in some months to be more remote than in others. Thus, the comparison of average monthly prices may pick up more than just temporal variation. The only temporal comparisons of prices that avoid this problem are for the 20 clusters revisited as part of the longitudinal sub-sample. However, these revisits did not start until after June, so a March quarter to June quarter price comparison is unavailable. 88. Instead of using the prices collected by the survey, variation in nominal price levels over the 12 month period of the field work was controlled for by using the Consumer Price Index as a deflator (Table 3). The deflator was applied uniformly to each region. It should be stressed that the CPI is actually calculated just for urban areas so it may not be an ideal guide to intra-year changes in the cost of living for the whole of PNG in 1996. Table 13: Temporal Price Deflator March June September December Annual Index 344.4 345.9 351.4 350.8 348.1 Deflator 98.9 99.4 100.9 100.8 100.0 Adjusting For Differences in Household Composition 89. A household with small children may not need as much consumption expenditure to reach the same standard of living as a similarly sized household where all of the members are adults. Earlier research in PNG used consumption per adult-equivalent, where children aged less than 15 years count as 0.5 adult-equivalents but this scale was not based on actual measurements of child costs. Recently, Gibson (1996) estimated child costs using data from the Urban Household Survey, and found that a scale of 0.5 was appropriate for young children (0-6 years) but older children had the same cost as adults. 90. Adjusting for differences in needs only matters if the share of children in household numbers varies widely between households (or between population sub-groups). If children's demographic share is roughly constant, the correction factor to convert household numbers into adult equivalents would also be roughly constant and would not lead to much re-ranking of households, compared with a rank formed over per capita expenditure. In fact, there is a lot of variation in the share of children in household numbers. The share of the six-years-and-under age group is 20 percent on average, but ranges from zero to 80 percent. There is also variation between the averages for each region, ranging from a 17 percent share in the NCD to a 22 percent share in the South Coast and Momase regions. Hence, treating young children like adults, and giving them a weight of 1.0 when forming per capita consumption might lower the estimated standard of living for households in the Momase and South Coast regions relative to NCD households, if in fact young children have fewer needs than adults. -88- PO VERTYAND A CC-ESS TO PUBLIC SER VICES 91. here are two main methods of measuring child costs: the Engel method and the Rothbarth method. The Engel method is based on the assumption that a household's standard of living is indicated by its budget share for food. The cost of an additional child can be measured by calculating the amount of compensation that would have to be paid to the parents to maintain the same food share as before the child was born. The Rothbarth method assumes that adults' standard of living is indicated by the value of expenditure on "adult goods" (goods not consumed by children). Expenditures on adult goods should fall when children are added to the household because the child brings additional consumption needs without any offsetting increase in income. Therefore, the cost of an additional child can be measured by calculating the amount of compensation that would have to be paid to parents to restore expenditure on adult goods to the former level. Deaton and Muellbauer (1986) show that the Engel method overestimates the cost of children, while the Rothbarth method underestimates them, so the two methods are used here to set bounds on the adult equivalence scale for children. 92. he food Engel curve used by Deaton and Muellbauer (1986) is of the form: Wf= Cc-/6In(-JY~j nj +8.-v + ( n)+ Jj= where Wf is the food budget share, x is total expenditure, n is the total number of people in the household, nj is the number of persons in category j (=1, . . , J), v is a vector of other control variables, [I is a random error, and D, [, Cl and 6 are parameters. The food Engel curve was estimated using three demographic groups: na, ncl, and nc2, which are the number of adults, the number of older children (age 7 to 14), and the number of young children (O to 6 years). Total household expenditure, x was deflated by the value of the food poverty line (in index form), zf. Dummy variables for each region were included as controls for price and non-price variation that affect the food budget share. Results are reported in column (i) of Table 14. Table 14: Estimates of Food Engel Curves Explanatory variable (i) (ii) In real per capita expenditure [(xlzj/ln] -0.0511 (2.93) In real household expenditure [xIzA ... -0.0523 (2.94) In household size ... 0.0109 (0.31) Number of adults -0.0066 -0.0000 (1.97) (0.00) Number of older children -0.0004 0.0072 (0.07) (0.86) Number of young children -0.0077 -0.0001 (1.20) (0.01) Intercept 0.813 0.841 R2 0.130 0.132 Note: Absolute t-values in ()corrected for clustering, sampling weights and stratification. Regional dummy variables also included in each equation. N= 1144. -89- POVERTYAND A CCESS To PUBLICSERViCES 93. For the specification in column (i) the food budget share for the reference household with two adults and no children is given by: wf = a + flnxx -,Bln2 + y1 .2 where y, is the coefficient on the number of adults, and the regional dummy variables and the food price deflator have been suppressed for convenience. For another household in the same region, but having two adults and an older child, the food budget share is given by: wj = a + ,6lnxl -,/ln3 + y, 2 +72 1, where y2 is the coefficient on the number of older children. Under the Engel assumption, the two households are equally as well off when their food shares are equal. Therefore, setting these two equations equal to one and other and solving for x gives the equivalence scale: In x = 2 - ln. x 0 8 3 94. Household total expenditure has to rise by 49 percent when an older child is added, so an older child appears to cost just as much as an adult. Replacing y2 with y3 (the coefficient on the number of younger children) in the formula indicates that household total expenditure has to rise by 29 percent when a younger child is added, so the adult-equivalence of that child is 0.58. In general, for this Engel curve, the equivalence scale for household h to reach the same welfare level as the reference household 0 is: Eh (n-h)exp J-p)(nh - nO)]. 95. The specification used for the results reported in column (i) of Table 14 imposes a homogeneity assumption, by using per capita expenditure. A more flexible specification, which will also be useful for estimating household size economies, is to allow the logarithm of real household expenditure and the logarithm of household size to enter the equation as separate variables. For this specification, the equivalence scale for adding an older child to the reference household with two adults is given by: x 7 2 1 3 x 0 = - 2 where ri is the coefficient on the logarithm of household size. Results for this model are reported in column (ii) of Table 14. Household total expenditure has to rise by 25 percent when an older child is added to the reference household, so the adult-equivalence of that child is 0.50. For a younger child the adult equivalence is 0.18. 96. The adult-equivalence scale derived from the results in column (ii) shows surprisingly low costs of children in comparison with earlier estimates for PNG and other countries. This is especially because, in theory, the Engel method overestimates the cost of children. Which results should be used? The homogeneity assumption to derive the column (i) specification from the column (ii) specification (r/,=-1) is not rejected, with X2(1) =1.4 (p<0.24). Thus, statistical testing -90- POVERTYAND ACCESS To PuBLic SER VICES cannot help in making the choice. The results in column (i) have the advantage of being linked to previous uses of the method in PNG and other countries. A third option is to use neither, due to the excess sensitivity to minor specification changes. Further support for not using the Engel results comes from Deaton (1997), "the method is unsound and should not be used" (p. 255). Therefore, it is worth turning attention to the results of using the Rothbarth method. 97. The first step in carrying out the Rothbarth method of measuring child costs is to find the set of adult goods. This is done by testing whether the presence in the household of children influences the budget shares for candidate adult goods, controlling for aggregate adult goods spending (Deaton, Ruiz-Castillo, and Thomas, 1989). This procedure suggested that adult clothing, beer, other alcohol, cigarettes, tobacco, gambling, lotteries, and meals consumed away from home were valid adult goods. Following Deaton, Ruiz-Castillo, and Thomas (1989) the Engel curve estimated for the aggregate group of adult goods was: WA =O0.0091 +o0.0190 In('j - 0.0128 In n - 0.0551 (-)- 0.0014 (-) R2=0.I0 (2.67) (1.46) (2.62) (0.06) where t-statistics are in ( ) and regional dummy variables were also included in the regression. For the reference household of two adults, with mean (log) per capita expenditure, the predicted budget share for adult goods is 0.11. This implies annual expenditures on adult goods of K139. The addition of an older child (ncl=l) to the household reduces wA to 0.08 and implied expenditures on adult goods falls to K100. To restore adult goods expenditure to its initial level, household total expenditure would have to increase by 31 percent. This suggests that an older child costs 62 percent of the average cost of an adult. Adding a younger child (nc2=1 ) to the reference household reduces WA to 0.10, and implied adult goods expenditure falls to K1 22 per year. Increasing total household expenditure by 12 percent would restore adult goods expenditure to its initial level, suggesting that a younger child costs 24 percent of what an adult costs. 98. A third method of calculating child costs, which is appropriate when household budgets are mainly allocated to food, is to compare the recommended dietary intakes of children and adults. The results reported in the third column of Table 15 show that in PNG older children have the same calorie requirements as an average adult, while the calorie requirements for younger children are only 62 percent as high. Table 15: Estimates of Adult Equivalence Scales for Younger and Older Children Increase in household expenditure to compensate a couple Recommended calorie for cost of one child intake relative to all adult Engela Rothbarth age/sex groupsb (i (ii) (iii) Child7-14years 0.49 0.31 1.02 Child 0-6 years 0.29 0.12 0.62 -91- POVERTYANDACCESS TO PUBLIC SERVICES aUsing the results in column (iii) of Table 14. bCalculated from requirements in Nutritionfor Papua New Guinea, Department of Health (1975) 99. The two most reliable methods of calculating child costs (Rothbarth and calorie requirements) suggest that the adult-equivalence of a 7-14 year old child is between 0.62 and 1.02 . The adult equivalence of younger children lies between 0.24 and 0.62 according to these same two methods. The most plausible estimates from the Engel method are for the specification used by Deaton and Muellbauer (1986). Those Engel estimates are similar to the results using calorie requirements, with older children costing 98 percent of an adult, and younger children costing 58 percent of an adult. In view of these results, the equivalence scale used here is that children age six years and below count as 0.5 adult-equivalents, and everyone else counts as 1.0. Adjusting for Differences in Household Size 100. Economies of size in household consumption allow the cost per person of reaching a certain standard of living to fall as household size rises. For example, a household with ten people may not need to spend ten times as much as a single-person household to enjoy the same standard of living. Economies of size can affect two types of poverty comparisons. The first is between demographic groups, such as "small" households and "large" ones. When no allowance is made for economies of size, large households typically appear to be poorer (Lanjouw and Ravallion, 1995). This correlation may lead policy makers to use "demographic targeting" to alleviate poverty by making programs specially available to large households. If in fact the lower per capita expenditure by large households just reflects economies of size, such demographically targeted policies are a wasteful way of alleviating poverty. Secondly, unaccounted for economies of size can distort regional poverty profiles because estimates of poverty will be too high for regions with larger households and too low for regions with smaller households. This could affect poverty comparisons between the NCD and other regions, because the average household size in the NCD (6.9 dwellers per household) is 25 percent higher than in other regions. 101. A modified Engel method of measuring size economies has recently been introduced by Lanjouw and Ravallion (1995). Instead of deflating total expenditure by person numbers, n, they deflate by equivalent person numbers, n9 (0<01), where 0 is the size elasticity of the cost of living. Noting that ,l1n(x/n9) can also be expressed as f)lnx-0,8lnn, the Engel curve can be estimated with lnx and lnn as separate variables, and the ratio of the coefficient on lnn to the coefficient on lnx gives an estimate of 0. This estimate of the size elasticity of the cost of living applies only to households of the same composition. When demographic composition is controlled for using the numbers of persons, as in Table 14, the size elasticity is calculated as: O= -17 + J n where rl is the coefficient on Inn, ,B is the coefficient on lnx, yj is the coefficient on the number of people in the household from the jth demographic group, niln is the share for the jth -92- POVERTYAND ACCESS TO PUBLICSERVICES demographic group, and n is household size. Using the coefficient estimates in column (ii) of Table 14, and population averages for nj and n, the calculated size elasticity is 0.34 with a standard error of 0.25. 102. These results have important implications for the measurement of living standards. The hypothesis that per capita expenditure is the correct variable for comparing households of different sizes (i.e., 9=1) is rejected, with X2(l)=6.14 (p<0.02). The hypothesis that no adjustment for size is needed when comparing different households (i.e., 9=0) is not rejected, with X2(l,l .90 (p<0. 17). In other words, the data do not reject the hypothesis that each person in the household, beyond the first resident, costs nothing. Under this hypothesis, all items in the household's budget have non-rival consumption, which is not a plausible description of items like food. Thus, something appears wrong with this method of measuring household size economies. 103. One improvement is to drop the three demographic composition terms (na, ncl, and nc2) from the model because they are not statistically significant (p<0.67). With this change, the value obtained for a is 0.41, with a standard error of 0.22. The assumption of homogeneity (*=1) is still rejected, with X2(,)=6.88 (p<0.0l). The other hypothesis, that no adjustment for number of residents is needed when comparing different households (90), is now rejected only at relaxed levels of statistical significance, with X2(,)-3.36 (p<0.07). 104. The Engel method of measuring household size economies has been criticised by Deaton and Paxson (1996). The criticism is that as household size increases the sharing of public goods within the household should free up resources that will increase the demand for a (poorly substitutable) private good like food. Thus, there should be a positive relationship between household size and food demand, holding outlay per person constant. Instead, the results reported here, the results of Lanjouw and Ravallion for Pakistan, and results in a number of other countries, show the budget share for food (and therefore food consumption, unless pecuniary effects are very strong) falling with household size, at constant per capita outlay. The experiment comparing the performance of the recall survey instrument with the diary survey instrument in the NCD may be of relevance. For households surveyed using the recall instrument, the food share fell with increases in household size, holding outlay per person constant. However, for the matched households who used the diary survey, there is no relationship between the food share and household size, with outlay per person held constant. One interpretation of this is that there are measurement errors in food expenditures that are correlated with household size (Gibson, 1997). If this is the case, the estimate of 0.41 for the size elasticity of the cost of living may overstate the true extent of household size economies. 105. It may be appropriate to use another welfare indicator for estimating size economies, in view of the problems in using the budget share for food. The Rothbarth assumption, that adult goods expenditure indicates (adult) welfare levels may be useful. But in contrast to the case of estimating child costs, total spending on adult goods is no longer an appropriate welfare indicator because some of the additions to the household will be adults, whose presence would reduce per -93- POVERTY AND ACCESS To PUBLIC SERVICES adult consumption of adult goods. Instead, adult goods expenditure per adult is used as the welfare indicator. The equation estimated for the aggregate group of adult goods is PA qA = -193.3 +252.0 lnx -307.7 Inn -201.9 ( -74.1 (nl) R2=0.24 nA n n (4.90) (5.34) (1.90) (0.97) where PA qA is expenditure on adult goods, nA is the number of adults, t-statistics are in (), and the equation also includes regional dummy variables. The size elasticity is given by the ratio of the coefficient on ln n to the coefficient on ln x, which is 1.22 (standard error of 0.1 1). This implies that to maintain a constant value of adult goods consumption per adult, household total expenditure has to increase by at least the same percentage as the increase in household size. If the welfare indicator used was the budget share for adult goods, the hypothesis that OA1 is also not rejected (p<0.30). Thus, in contrast to the results using the food budget share, methods using adult goods as the welfare indicator do not find economies of household size. 106 What can be made of these diverse estimates of the extent of household size economies? Although it is possible that x/n04' is the correct indicator of living standards, the Engel method does not seem sufficiently uncontroversial and its results may be affected by measurement error. Therefore, the main analyses in this report follow usual procedure and do not adjust for size economies. 107. To explore the robustness of the results in Table 1.6 of the main report, poverty measures for varying household size groups were recalculated using two economy of size correction factors: 0.7 which is about equal to the poor's food expenditure share and 0.41 which is the correction factor that was estimated from this data set (Tables 16 and 17). The analysis shows that poverty measures for larger households drop significantly, while those for smaller households increase when household economy correction factors are employed. However, given the controversy surrounding the setting of the correction factor, it is preferable to follow standard practice and not adjust household consumption for size economies for the poverty profile in this report. Nevertheless, the marked difference in the poverty measures for various household size groups when correction factors are employed, suggests that the results in the main report, which show a strong association between household size and poverty, must be must be interpreted with caution. -94- POVERTYAND ACCESS TO PUBLICSERVICES Table 16: Distribution of Poverty by Household Size Group (With Correction for Economies of Household Size) Headcount Poverty gap Poverty severity Share of Index Contribution Index Contribution Index Contribution total pop (%) to total (%) (%) to total (%) (%) tototal (%) (%) Economies of scale parameter O= 0.7 Upper Poverty Line 37.2 100.0 11.6 100.0 5.4 100.0 100.0 1-2 persons 37.0 2.9 11.2 2.8 4.7 2.6 2.9 3-4 persons 33.4 15.4 13.0 19.2 6.7 21.4 17.1 5-6 persons 36.1 23.2 12.7 26.0 6.0 26.9 23.9 7-8 persons 37.7 24.3 12.1 24.9 5.3 23.7 24.0 9-10 persons 42.0 18.4 10.9 15.2 4.9 14.8 16.3 Morethan lOpersons 37.2 15.8 8.7 11.8 3.6 10.5 15.7 Economies of scale parameter 0 0.41 Upper Poverty Line 36.2 100.0 12.3 100.0 5.9 100.0 100.0 1-2 persons 49.0 4.0 21.2 5.1 11.2 5.6 2.9 3-4 persons 46.6 22.0 18.2 25.3 9.7 28.2 17.1 5-6 persons 39.5 26.0 14.5 28.2 7.2 29.0 23.9 7-8 persons 36.8 24.4 11.6 22.7 5.0 20.4 24.0 9-10 persons 32.3 14.5 8.6 11.4 4.0 10.9 16.3 More than 10 persons 20.9 9.1 5.7 7.3 2.2 5.9 15.7 Note: The welfare indicator used is total consumption per effective adult equivalent. The number of effective adult equivalents is given by c x n where 9 is the elasticity of the cost of living with respect to household size and c is the factor that normalises no for a poor household of mean size (c= 1.7 if 0=0.7 and c=2.9 if 0-0.41). -95- PO VER TY AND A CCESS TO PUBLIC SER VICES Table 17: Distribution of Poverty by Number of Children in Household (With Correction for Economies of Household Size) Headcount Poverty gap Poverty seven'ty Share of Index Contribution Index Contribution Index Contribution total pop (%) to total (%) (%) to total (%) (%) to total (%) (%) Economies of scale parameter 0 = 0.7 Upper Poverty Line 37.2 100.0 11.6 100.0 5.4 100.0 100.0 Zero children 32.8 6.0 11.5 6.7 5.3 6.7 6.8 1-2 children 32.7 27.0 11.7 30.8 5.6 32.0 30.7 3-4 children 39.5 39.6 12.4 39.8 5.7 39.6 37.4 5-6 children 37.5 19.6 9.8 16.5 4.2 15.2 19.5 More than 6 children 50.5 7.8 12.6 6.3 6.0 6.5 5.7 Economies of scale parameter O= 0.41 Upper Poverty Line 36.2 100.0 12.3 100.0 5.9 100.0 100.0 Zero children 40.2 7.5 16.4 9.0 8.7 9.9 6.8 1-2 children 39.1 33.0 14.7 36.5 7.4 38.3 30.7 3-4 children 38.2 39.4 12.3 37.2 5.7 35.9 37.4 5-6 children 29.1 15.6 8.2 13.0 3.4 11.3 19.5 More than 6 children 27.9 4.4 9.2 4.3 4.7 4.5 5.7 Note: The welfare indicator used is total consumption per effective adult equivalent. The number of effective adult equivalents is given by c x n9 where 0 is the elasticity of the cost of living with respect to household size and c is the factor that normalises n9 for a poor household of mean size (c= 1.7 if 0-0.7 and c=2.9 if 0=0.4 1). -96- POVERTYAND ACCESS TO PUBLIC SERVICES ANNEX II. SETTING A POVERTY LINE 1. This annex describes how the set of regional poverty lines used in this report has been constructed. It is based on a report by John Gibson and Scott Rozelle entitled "Results of the Household Survey Component of the 1996 Poverty Assessment for Papua New Guinea". Methods of Setting the Food Poverty Line 2. The first step is to estimate the cost of buying a basic diet that meets food-energy requirements, which are set at 2200 calories per adult equivalent per day in this study to maintain comparability with previously defined poverty lines in Papua New Guinea. It would be possible to use linear programming methods to find the cost-minimising diet that achieved this target (possibly in conjunction with protein and micro-nutrient targets), for any given set of local prices. That approach has in fact been used in PNG by the Bureau of Statistics (1980) in their publication Lowest Foodcost: A Statistical Tool. However, the estimates produced were never used for policy because the foods selected by the linear programming algorithm were not commonly eaten by poor people in PNG (e.g., beef liver, brown rice rather than white rice) and the mixture of foods selected by the computer was not very palatable. 3. The more usual approach is to set the food poverty line as the cost of meeting food energy requirements from a diet consisting of the foods that are actually eaten by poor people in the country whose poverty line is being formed. The foods to include in this basket, and their relative importance, can be set by looking at the food budgets of a group of poor households. Ideally, this group should not include households ultimately found to be above the poverty line, so that it is the dietary patterns of the poor but no others that count in forming the basket. Once the list of foods and their relative importance is determined the size of the basket can be scaled up or down (holding budget shares constant within the basket) until it exactly achieves the food energy requirement. The cost of buying this (scaled) food basket can then be estimated separately for each region and sector, giving a set of food poverty lines. 4. A complication arises when dietary patterns differ between regions. The logic of using a basket of foods that reflects local tastes suggests that different baskets should be formed for each region. Separate regional baskets are also closer to the ideal of a true cost-of-living index where the price of a fixed level of utility is computed, rather than the price of a fixed bundle of goods. The concern is that different baskets mean diets of different qualities, so the poverty line for one region may give a superior standard of living to the poverty line for another region. This could certainly occur if the baskets for each region were formed from the same percentiles of the income distribution (e.g., the poorest quintile). Regional differences in living standards would then be directly translated into differences in quality because the poorest quintile in a rich region are likely to eat better than the poorest quintile in a poor region. A solution to this problem is to form the separate baskets for each region from the food budgets of households that fall below a certain cut-off value of real consumption per adult equivalent, where that cut-off is the same for all regions. That way, the food preferences of the non-poor do not influence the poverty line food basket in regions where the proportion who are poor is much lower than average. -97-. POVERTYANDACCESS TO PUBLICSERVICES 5. An additional safeguard against quality differences in the regional food baskets is to test for "revealed preferences". This can be done by seeing whether the cost in regionj of buying the region i basket of foods is less than the cost of buying the regionj basket of foods, at regionj prices. If the cost is less, households in regionj could buy the region i basket of foods but seeing as they do not they must prefer the regionj basket, implying that the region i basket is of lower quality. In this case, some adjustment would need to be made to the food baskets to ensure that the quality of the poverty line diet is equal in region i andj. 6. There is one problem in using a cut-off value of real consumption to find the group of poor households whose food budgets determine the composition of the poverty line basket of foods. When there are large differences in the cost of living between regions - as there are in PNG -- a regional price deflator is needed so that poor households from regions where prices are high have as much chance of influencing the basket of foods as do poor households in regions where prices are low. If prices were available for all consumption items, a regional deflator could be calculated with any of the commonly used price index formulas (e.g., Fisher, Tornqvist) to help with this initial ranking of households. However, as discussed in Part I, prices are not available for all non-food consumption items so any deflator that was constructed would involve untestable assumptions about the regional pattern of these missing prices. Moreover, care needs to be taken when selecting the weights (budget shares or quantities) that such a price index would need. Items important in the budgets of the poor may be less important in the budgets of the "average" households whose consumption patterns usually form the weights in price index formulas. 7. The solution adopted to the problem of the missing deflators in this study is to iterate on the poverty line, until the deflator implied by the poverty line converges. Specifically, we start out with a group of households who are provisionally defined as poor according to the nominal value of consumption per adult equivalent. Regional baskets of foods that supply 2200 calories per adult equivalent per day are formed using the food budgets of this group of households. The cost of each basket gives the initial set of food poverty lines. A non-food allowance is then calculated, using methods described below. (We choose the "austere" poverty line to iterate on because its more restricted definition of essential non-food items is closer to the concept of a basic needs utility level.) The resulting set of regional poverty lines are used as implicit deflators to define real consumption per adult equivalent (putting all values in terms of the region with the highest-priced poverty line). A new group of poor households is found according to whether real consumption per adult equivalent falls below the cut-off value, new baskets of foods are formed, and a new set of poverty lines is calculated. The process is repeated until the regional deflators implied by the poverty lines converge. 8. The process of forming deflators implicitly from the poverty line and then iterating until they converge has two advantages over the alternative of using a deflator calculated in some other manner. First, no assumptions are made other than those already inherent in the formation of a poverty line. There is no need to "guess" the regional pattern of missing non-food prices becausc the poverty line calculations make allowance for regional variation in non-food prices. Thus, the iterative method is internally consistent. Second, the deflator that is formed is calculated -98- PO VERTYAND A CCESS To PUBLIC SER vICES specifically for the poor, rather than for "average" households. There is no reason why regional variation in the cost of living is the same for poor people as it is on average for the population, so if the purpose of a deflator is to find a group of poor households it makes sense to use the deflator that applies specifically to those households. Methods of Setting the Non-Food Component of the Poverty Line 9. The austere allowance for non-food items is based on the typical value of non-food spending by households whose total expenditure just equals the cost of the food poverty line. Consuming these non-food items means that some food needs are ignored, so the non-food items can be considered as essentials (Ravallion, 1994). The average food share for these households (in regionj) is found from the following Engel curve (dropping household subscripts): K J-1 w = a +,8ln Xj+1knk + YjDj +e where w is the food budget share, x is total expenditure, n is the number of persons, zF is the food poverty line for an adult-equivalent in regionj, nk is the number of people in the kth demographic category, and Dj is an intercept dummy for region j. When total expenditure exactly equals the cost of the food poverty line, ln(x/(n -z))= 0,so k= CI ik in + q. gives the average food share k=1 in region j, where iik is the mean of the demographic variables for the households used to form the poverty line basket of foods. The lower poverty line z, is given by the sum of the food and non-food components, zy = ZF+ Z F(l_j) = Z (2-u4). 10. A more generous allowance for non-food items can be found from the typical value of non- food spending by a household whose food spending actually reaches the food poverty line. Finding the food share, w* at this expenditure level requires a numerical solution, characterised by F n Zj = x * W*. This can be substituted into the Engel curve to give: w' = aX +fln(w*) Using w-1 to approximate lnw, an initial solution of w,=(aj+P)/(l+I) can be found. This estimate can be improved upon by iteratively solving the following equation, t times: (wI* ,81n w,j* - aj) W, = W and the upper poverty line is estimated as zu = zj4/w' (Ravallion, 1994). -99- POVERTYANDACCESS TO PUBLIC SERVICES Results of Setting the Food Poverty Line 11. The initial poverty line food baskets used in this study were formed from the food consumption budgets of households where nominal total consumption per adult equivalent was below K570 per year. This group of 437 households represents the poorest 50 percent of the population. It was expected that the size of this group would fall once consumption was valued at the prices in the highest-price region (using the deflator implied by the poverty line). 12. The weighted average consumption quantities were estimated in units of grams per adult equivalent per day for 35 foods in each of five regions (National Capital District, Papuan/South Coast, Highlands, Momase/North Coast, and New Guinea Islands). The average calorie availability per adult equivalent was then calculated from these food quantities using data on the calorie content and edible fraction of each food (Table 1). An allowance was made for calories obtained from meals consumed away from home, by assuming that this source of calories was twice as expensive as the average 'price' of all other calories that the household obtained. These calorie 'prices' (strictly speaking, 'unit values') were calculated separately for each household. A scaling factor was calculated for each region by dividing the average daily calorie availability by 2200. This scaling factor was applied to all quantities in the food basket (values in the case of meals away from home) to ensure that the scaled basket supplied exactly 2200 calories. 13. An analysis of covariance (dummy variable) model was used to statistically test whether regional differences in the composition of the food baskets were statistically significant. The set of regression equations estimated by weighted least squares were of the form: Wi = CC + (3 1(Papuan) + 132 (Highlands) + P3 (Momase) + 4 (New Guinea Islands) where wi is the share of calories provided by the ith food and the right-hand-side variables are of 0- 1 form, indicating which region the observation (i.e., household) came from. Shares are used as the dependent variable so that differences in mean levels of food consumption between regions do not affect the results (the same effect is achieved in the fornation of the food basket by scaling quantities). Also, the regression is in terms of calories rather than food quantities, so that foods without quantity data (e.g., meals consumed away from home) can be included. The hypothesis that 3=02=P3=j4-0 was then tested (because this is a linear hypothesis the Wald test results are not affected by the choice of the NCD as the base region). This equation was estimated separately for each of the 36 foods in the basket because software for estimating systems of equations with clustered standard errors is unavailable. Equation-by-equation estimation is inefficient, because it ignores the correlations between the residuals in each equation, so the hypothesis of equal shares between regions is less likely to be rejected. Nevertheless, the hypothesis of equal shares was rejected for 27 out of the 36 foods (at p=0.05), and these 27 foods account for 97.4 percent of the average calorie supply for households below the cut-off. A major contributor to inequality in regional average shares was sweet potato, which provides 64 percent of the calories in the highlands, only 1.7 percent of the calories in the NCD, and between 10 and 16 percent of the calories in the other regions. It view of these results it seems appropriate to estimate separate poverty line food baskets for each region. -10o- POVERTYAANDACCESS To PUBLC SER ViCES 14. The poverty line food baskets for each region were priced using regional averages of local market prices. The only food in the baskets without price information was meals consumed away from home. An allowance for this cost was made by scaling the sub-total price of all of the other foods in the basket up by 1/((2200-k)/2200), where k is the number of calories in the basket provided by meals consumed away from home. Table 1: Scaled Poverty Line Food Basket Using National Average Quantities (With Associated Data) Mean grams per Kcal Edible Kcal per adult adult equivalent per kg Fraction equivalent per day per day Sweet potato 665 1144 0.84 639 Cassava 70 1295 0.87 78 Taro 183 1117 0.84 171 Yam 58 1140 0.81 53 Banana 302 1165 0.65 228 Sago 119 3313 1.00 395 Coconut 103 3837 0.65 256 Rice 23 3830 1.00 90 Lamb and Mutton 1 3780 0.84 5 Pork 7 3290 0.84 18 Chicken 1 2040 0.72 2 Other meat (incl. bush meat) 13 2480 0.99 17 Fish 5 1398 0.74 5 Sugar cane 131 678 1.00 89 Other fresh fruit 24 433 0.83 8 Peanut 1 5516 0.69 6 Aibika 38 350 0.50 7 Other vegetables and nuts 157 521 0.74 60 Potato 1 750 0.80 0 Betel nut 17 1100 0.40 7 Flour 2 3433 1.00 8 Tinned meat 1 1921 1.00 1 Tinned fish 2 1820 1.00 4 Milk 0 4924 1.00 0 Sugar 2 3935 1.00 7 Bread 1 2370 1.00 3 Biscuits 1 3674 1.00 3 Dripping 1 8741 1.00 7 Other diary, cereals, and eggs 0 2374 0.96 1 Tea, coffee, milo 0 484 1.00 0 Snack food 0 5026 1.00 0 Salt, spices, sauces 1 205 1.00 0 Soft drink 1 466 1.00 0 Beer 0 347 1.00 0 Other alcohol 0 1929 1.00 0 Meals eaten out of home 29 Total 2200 Note: Estimated from budgets of households with real consumption per adult equivalent K570/year. -101- POVERTYAND A CCESS TO PUBLIC SER VICES 15. The region with the highest cost of buying 2200 calories per day was the NCD, where the price was K543 per year. Households in the Highlands and Momase regions faced the lowest cost, at K307 and K250 per year. Note that these cost differences do not seem to imply differences in the quality of the diets supplied. Consumers in the NCD would have to pay K696 per year to buy the basket of foods for the Highlands region and K640 per year to buy the Momase basket. Both of these are more expensive than the actual basket bought by NCD households, and therefore the NCD basket is not revealed to be preferred to the other baskets. 16. An allowance for essential non-food items was calculated and the resulting poverty lines suggested that the highest cost of living for poor people was in the NCD, followed by the Papuan and New Guinea Islands regions. The lowest cost of living was in Momase. Setting the (austere) poverty line in the NCD equal to 100, the poverty lines in the other regions were: Papuan 74.2; Highlands 53.5; Momase 41.5; and New Guinea Islands 58.3. 17. These poverty lines, in index form, were used as deflators to convert nominal consumption into real consumption. This gave a group of 177 households (20 percent of the population) with real consumption per adult equivalent less than the cut-off value of K570 per year. A new set of regional food baskets was formed from the budgets of these households and the resulting set of poverty lines was calculated. These poverty lines, in index form, showed some changes from the results of the previous iteration so the process was repeated. After the fifth iteration, the only poverty line changing was for the Highlands where the index number oscillated between 49.9 and 50.1. These changes were sufficiently small to accept the convergence. The iterations were also started from the values of a Tormqvist regional price index (calculated with the available prices) and the poverty lines still converged to the same levels. 18. The poverty lines, in index form, resulting from the sixth iteration were used as the last set of deflators for defining real consumption so as to find the group of poor households whose food budgets would influence the poverty line food baskets. The resulting food poverty lines are reported in Table 2. The Table also reports the cost of buying each region's basket of foods at the average prices prevailing in every other region. 19. One problem immediately apparent in Table 2 is that two of these food poverty lines fail the revealed preference test. The basket of foods for the Momase and Papuan regions could be bought in the New Guinea Islands region for only 77 percent and 92 percent of the cost of the Islands region food basket. Hence, the poverty line for the New Guinea Islands appears to give a superior standard of living to the poverty lines used in the Momase and Papuan regions. Similarly, the basket of foods for the Momase region can be bought in the Papuan region at only 88 percent of the cost of the Papuan region basket. -102- PO VERTYA ND A CCESS To PUBLIC SER VICES Table 2: Cost of Region i Poverty Line Basket of Foods at Regionj Prices (separate baskets for all five regions) regionj prices region i basket NCD Papuan Highlands Momase NGI (kina per year) National Capital Dist. 543 578 544 510 525 Papuan/South Coast 629 402 318 223 345a Highlands 708 580 288 271 408 Momase/North Coast 549 353a 318 188 288a New Guinea Islands 599 478 420 293 373 a Fails revealed preference test. 20. These two instances of the food basket for the Momase region lying within the budget set of consumers in other regions suggests that this basket of foods is of lower quality. Therefore, poverty lines based on this basket of foods (such as those reported in index form for Momase ) would refer to a standard of living that is lower than the living standards achieved at the poverty lines in other regions. Consequently, the measured rate of poverty in the Momase region would be too low. 21. The reverse problem occurs in the New Guinea Islands region, where poverty lines based on the regional basket of foods would be "too high" and would lead to an overstatement of poverty in this region. Further evidence of the superior quality of the basket of foods for the New Guinea Islands region comes from Figure 2. The price of the separate regional baskets is compared with the cost in each region of the national basket (the weights for this national basket are in Table 1). Usually, the cost is higher (by between 12 and 28 percent) if households in each region are forced to buy the national basket of goods. However, the national basket can be bought in the Islands region at a cheaper price than the regional basket can be bought, which suggests that the regional basket provides a better quality diet. -103- POVERTYAND A CCESS To PUBLICSER VICES Figure 2: Distribution of Food Poverty Lines By Region 700 629 [lNational Basket > 600 _ *Separate Regional Baskets 500 457 z 400 --35537 323 300 I n0f2 _ ~~~~~~~~~~~~~~24018 Cu200 1L 100 0 NCD Papuan Highlands Momase Islands 22. One solution to the apparent differences in quality of the food baskets for the Papuan, Momase and New Guinea Islands regions would be to force consumers in each region to purchase the Papuan basket of foods. This would upgrade the quality of the Momase basket (as K223 exceeds K188 from Table 2) and downgrade the quality of the New Guinea Islands basket (K345 < K373). However it seems strange to force poor households to eat the foods commonly eaten in another region rather than the foods commonly eaten in their own region. At the very least their preferences should count in forming a pooled basket of foods, formed from the food preferences of poor households in all three regions. 23. A pooled basket of foods for the Papuan, Momase and New Guinea Islands regions was in fact formed. This was based on the food budgets of the surveyed households in this combined region whose real total consumption per adult equivalent was below K570 per year. This merging of food preferences in these three regions is more defensible than the use of a single national basket of foods because the dietary patterns in these three regions are the most closely alike. In contrast, diets in the Highlands are dictated partly by environmental constraints (e.g., altitude and climate preclude any significant coconut and sago production), while diets in the National Capital District are conditioned by the arid climate, the poor links to the rest of the economy and the abundance of imported and commercially produced foods. 24. A test of similarity in diets for the Papuan, Momase and New Guinea Islands regions was carried out, using a covariance model. The set of regression equations estimated by weighted least squares were of the form: wi = cx + I3 (Papuan) + P2 (Momase), where wi is the share of calories provided by the ith food and the right-hand-side variables are of 0-1 form, indicating which region the observation (i.e., household) came from. The null hypothesis that the share of calories contributed by a particular food to each regions diet is the same in all three of the regions (i.e., P,1 = -104- POVERTYANDACCESS TO PUBLIC SER VICES ~2 =0) was rejected for only six out of the 36 foods. In contrast, this hypothesis was rejected for 27 out of the 36 foods when the test for a single national diet was carried out. 25. All of the steps and iterations described above were repeated, but with food poverty lines calculated from a basket of foods for the NCD, a basket of foods for the Highlands, and a common basket of foods for the Papuan, Momase, and New Guinea Islands regions. The main difference from the previous set of iterations was that the final value of the poverty line, in index form, for Momase was higher than it had been (36.0 vs 30.9) and the value for the New Guinea Islands was lower (54.4 vs 62.6). These changes are consistent with the movement towards poverty lines that give equal standards of living in these two regions. 26. The values of the food poverty lines in each region are reported in Table 3. Unlike the previous values in Table 2, there are no failures of the revealed preference test. The food quantities for each of the three scaled baskets are reported in Table 4. Table 3: Cost of Region i Poverty Line Basket of Foods at Regionj Prices (combined basket for Papuan, Momase, and Islands regions) regionj prices region i basket NCD Papuan Highlands Momase NGI (kina per year) National Capital Dist. 543 578 544 510 525 Papuan + Momase + Islands 594 391 330 218 326 Highlands 708 580 288 271 408 27. One final concern in using the food poverty lines reported in Table 3 may be that they are based on market prices, when in fact a large share of consumption is from own-production. The value that households reported for their self-produced food may have been lower than market prices. This possible discrepancy between the prices used to value consumption and the prices used to form the poverty line could inflate estimates of the incidence of poverty in areas where subsistence food production was a large component of the value of measured consumption. Results of Estimating the Non-Food Allowance 28 The food Engel curve was estimated with intercept dummy variables for each region so that different raising factors (i.e., 2-a;j) could be calculated for each region. The equation included the same set of demographic variables that were previously used when the Engel curve was estimated for measuring child costs and size economies. Hence, the only difference from the previous food Engel curve is that the level of the food poverty line is used as the deflator for household total -105- POVERTYAND ACCESS TO PUBLICSER VICES Table 4: Scaled Poverty Line Food Baskets for Each Region Papuan + Momase + New NCD Highlands Guinea Islands Mean grams per adult equivalent per day Sweet potato 38 1494 272 Cassava 105 70 51 Taro 1 134 234 Yam 0 35 87 Banana 89 82 371 Sago 7 4 201 Coconut 66 2 143 Rice 257 20 14 Lamb and Mutton 30 3 0 Pork 0 3 5 Chicken 15 2 0 Other meat (incl. bush meat) 32 24 7 Fish 25 0 8 Sugar cane 0 150 117 Other fresh fruit 21 17 42 Peanut 0 4 0 Aibika 11 24 42 Other Greens, vegetables, nuts 20 303 101 Potato 10 2 0 Betel nut 3 6 21 Flour 63 0 2 Tinned meat 19 0 1 Tinned fish 25 2 2 Milk 2 0 0 Sugar 27 1 2 Bread 15 0 1 Biscuits 8 0 1 Dripping 8 2 0 Other dairy, cereals, and eggs 6 0 0 Tea, coffee, milo 2 0 0 Snack food I 0 0 Salt, spices, sauces 3 1 1 Soft drink 29 0 0 Beer 3 0 0 Other alcohol 0 0 0 Meals eaten out of home (Kcal) 23 44 21 Note: Estimated from budgets of households with real consumption per adult equivalent < K570/year. expenditure, x so as to scale the intercept. The weighted least squares results from estimating the equation on the full sample (rather than on just the households used to form the baskets for the food poverty lines) were (t-statistics in ( )): wf 0.732 - 0.052 In(I ZF) - 0.008na - O.01ncl - 0.004nc2 - 0.133 NCD + 0.033 Papuan (2.94) (2.19) (0.22) (0.63) (5.48) (0.89) - 0.054 Highlands + 0.014 Momase R2=0.13. -106- POVERTYAND ACCESS TO PUBLIC SER VICES (2.12) (0.55) The regression results show that the average food share for households who could just afford the food poverty line, if they devoted all expenditure to food, is 0.70 in the base region (the New Guinea Islands), taking account of the effect of the demographic variables. The food share where household food spending typically reaches the food poverty line is 0.68 in the base region. The food shares in the NCD and Highlands are lower than in the rest of the country. Consequently, the non- food allowance will be proportionately larger for these two regions than it is for the rest of the country. The Poverty Lines 29. Table 5 contains estimates of the food poverty line, the lower poverty line and the upper poverty line for each region. The upper poverty lines range from K 1016 per adult equivalent per year in the NCD to K314 in Momase. The weighted average of the upper poverty lines is K461, at national average price levels. The lower poverty lines range from K779 in the NCD to K280 in Momase, with a national average of K399 per adult equivalent per year. The interpretation of these poverty lines bears repeating. The food poverty lines give the required nominal value of consumption per year, in each region, for an adult to obtain 2200 calories per day from a diet of similar quality to the diets of other poor people in the same region. The lower poverty lines include an allowance for the consumption of basic non-food items, which is based on the displacement of essential food spending by poor people in each region. The upper poverty lines include a more generous non-food allowance. Table 5: Poverty Lines by Region (kina per adult equivalent per year) Upper Poverty Lower Poverty Food Poverty Line Line Line National Capital District 1016 779 543 Papuan/South Coast 547 496 391 Highlands 464 390 288 Momase/North Coast 314 280 218 New Guinea Islands 479 424 326 Papua New Guineaa 461 399 302 a Population weighted average, in national average prices. 30. Ideally, separate poverty lines for urban and rural areas within each region should be constructed, because urban households are likely to face somewhat higher prices on essential items than rural households. However, the relatively small sample size of the PNG household survey and the fact that the sample includes only one urban primary sampling unit for most regions, preclude the use of separate poverty lines for urban and rural areas within each region.The prices in the single sampled urban PSU are not likely to be representative for the urban prices throughout a whole region, particularly when the sampled urban PSU is not the largest urban area in that particular region. -107- POVERTYANDACCESS TO PUBLICSERVICES 31. Another alternative would be to calculate a single urban poverty line applicable to all non-rural PSUs outside the NCD. Whether it is sensible depends on whether price variation is greater between regions, or between rural and urban clusters within regions. To answer this question, an analysis of covariance (i.e. dummy variable) model was specified, to see whether a greater contribution to the variance in cluster-level prices of important commodities came from regions or from sectors (urban/rural). This test was done for the non-NCD regions, because there are no rural clusters within the NCD. The regression model for the price of the ith food was: Pricei = cx + f31 Highlands + P2 Momase + P33 New Guinea Islands + 4 Urban, so the reference group was rural clusters in the Papuan region. The test of the hypothesis that regional effects are unimportant was: P1=P2=P3=0, while for the hypothesis that rural/urban effects were unimportant it was: P4=0. The results suggested that regional effects were more important for most items, and especially for the locally produced foods that supply the bulk of the calories in the food poverty lines (Table 6). Even some imported and commercially distributed items, like kerosene, showed more regional price variation than rural/urban price variation. These results may reflect the very limited inter-regional trade in PNG, which is caused by the difficult topography and lack of transport infrastructure. Thus, urban areas in PNG usually have closer links with their rural hinterlands, than do the rural hinterlands with the rural areas of other regions. In light of these results and given sample design constraints, this report does not differentiate between urban and rural poverty lines within a given region, although the sensitivity analysis carried out in Annex 3, explores the possible effects of this omission on the poverty measures for PNG. -108- POVERTYANDACCESS TO PUBLIC SER VICES Table 6: Determinants of Regional and Sectoral Price Variations Regional effects = 0 Urban/rural effects = 0 F-value p-value F-value p-value Sweet potato 23.39 0.00 5.57 0.03 Betelnut 21.16 0.00 1.40 0.25 Sugarcane 6.76 0.00 3.53 0.07 Banana 6.48 0.00 2.72 0.11 Cassava 6.46 0.00 2.56 0.13 Kerosene 6.02 0.00 4.79 0.04 Taro 5.11 0.00 8.37 0.02 Flour 4.60 0.00 35.57 0.00 Rice 4.50 0.00 44.27 0.00 Coconut 4.31 0.02 0.08 0.79 Aibika 4.24 0.00 1.16 0.29 Canned meat 3.73 0.01 7.86 0.01 Petrol 2.55 0.07 1.92 0.17 Yams 2.36 0.10 0.69 0.42 Dripping 1.07 0.37 0.27 0.61 Sugar 2.58 0.06 35.91 0.00 Coarse-cut tobacco 2.28 0.09 20.74 0.00 Canned fish 2.56 0.06 11.63 0.00 Sago 1.22 0.34 1.61 0.23 Tea 0.70 0.56 1.05 0.31 -109- POVERTYANDACCESS TO PUBLICSERVICES ANNEX 11124: EFFECT OF USING SAMPLING UNIT SPECIFIC PRICES AND URBAN/RURAL PRICE DIFFERENTIALS IN POVERTY LINES 1. The poverty measurements in chapter 1 of this report rely on poverty lines that are set for just five regions: the Papuan/South Coast, the Highlands, the Momase/North Coast, the New Guinea Islands, and the National Capital District. Arithmetic averages of food prices in each region were used to calculate the cost of buying the poverty line basket of foods, which anchors the poverty lines. It is likely that these regional average prices overstate the cost of buying the basket of foods in some Census Units within each region, while understating it for others. Measured poverty will be too high in CUs where regional average prices overstate the cost of the basket of foods because these same (high) prices are not used for valuing food consumption. Hence, the value of some households' consumption will be above the poverty line if that line is priced using local (i.e., Census Unit) prices, but below the poverty line if regional average prices are used. This problem occurs even if consumption is valued using local market prices, rather than respondent-Teported values. Bias in the opposite direction (measured poverty too low) will occur in CUs where regional average prices understate the local cost of the poverty line basket of foods. 2. At first glance, it would seem that there is no net effect of using regional average prices because the overstatement of poverty in some CUs within the region is cancelled out by the understatement in others. This would be true if the distribution of food prices within each region was symmetric, with the mean equalling the median (e.g., a Normal distribution). However, there is evidence that the within-region distribution of prices is positively skewed, with the mean exceeding the median. For example, the hypothesis that the distribution of surveyed sweet potato prices across Census Units in the Highlands comes from a Normal distribution is rejected (p48 months), and four regional dummy variables. Dependent variable is binary, taking a value of I if the child is stunted (height-for-age more than two standard deviations below NCHS median) and 0 otherwise. 258 observations=1 and 356 observations=O.Numbers in ( ) are standard errors (corrected for population weights, stratification and clustering). Numbers in [ ] are probability derivatives: aPlaX = P(1 - P)fi which gives the change in the probability of the child being stunted given a unit change in the independent variable. 6. The first result, contained in column (i) of Table 1, is that each extra year of education completed by the parents decreases the risk of the child being stunted by almost four percentage points, holding the child's age, and region (which picks up ethnic effects) constant. In the sample used to estimate the model, approximately 50 percent of children were stunted. (This is higher than the national average of 43 percent because only a subset of observations, with both parents -116- POVERTYAND ACCESS TO PUBLUCSER VICES in the household and having full data available, were used for the model, N=614.) Thus, for children of the same age group in the same region, having parents who had both completed community school (six years of education) would almost halve the risk of being stunted, compared with children whose parents had never attended school. 7. The results in column (ii) show that most of this effect comes from the education of women, rather than of men. An extra year of education for the mother reduces the risk of child stunting by three percentage points. This effect is three times larger than the effect of an extra year of education for the father. This suggests that efforts to especially increase the education gained by women may have big payoffs. The adult women in the sample had two years less education, on average, than the men (2.9 years versus 5.0 years). Closing this schooling gap could reduce the proportion of children who are stunted by six percentage points. Not only is the payoff to men's schooling -- in terms of improved child health -- lower, it is also less precisely estimated. In fact, the hypothesis that father's education has no effect on the risk of stunting would not be rejected at usual confidence levels (p=O.11). 8. Are these findings about the more powerful effect of mother's education on child health robust? The first test is to add a variable measuring the economic resources of the household. The gender gap in adult schooling levels falls as household income rises, so any excluded effect of household economic resources on child stunting may bias the coefficient on mother's education. Results of the test are shown in column (iii), with the logarithm of real consumption per adult equivalent used to measure household economic resources. There is a slight fall in the size and statistical significance of the coefficients on both parents education, but mother's schooling still appears three times more effective than father's schooling (the statistical significance of the difference is also unchanged from the results in column (ii)). 9. The significant effect of mother's education on stunting probabilities, even when household economic resources are controlled for, suggests that most of the effect of education on child growth does not come through the income pathway. Instead, it may work either through the efficiency pathway, with mother's education most effective because they spend more time caring for children than fathers do, or through the power pathway, with children benefiting from the bigger say that educated women have in household decision-making. 10. A second robustness test of robustness is to add variables that measure parental health, because the coefficients on schooling may be picking up excluded effects of, especially, maternal health. Results of this test are shown in column (iv), with parental height used as a proxy for health (height will also pick up genetic and ethnic effects). Parental height is highly relevant to the risk of a child being stunted, but the difference between the effect of mother's height and father's height is not significant (p=0.71). Adding parental height to the model has most impact on the measured effect of household economic resources: the coefficient on consumption per adult-equivalent falls to one-quarter its previous value and the null hypothesis that economic resources have no effect on the risk of stunting is not rejected (p=0.74). The lower coefficients on schooling in column (iv) suggest that the previous results were picking up some excluded effects of maternal health but even -117- POVERTYANDACCESS TO PUBLICSERVICES with that addition to the model, mother's education still exerts a powerful effect on the risk of child stunting. 11. The significant effects of mother's height and mother's education on stunting probabilities indicates that there is a long-run, intergenerational, effect of women's education on child health. Educating the current generation of girls will reduce the risk that their own children are stunted. Those children are likely to become taller adults which will then reduce the risk of the next generation of children being stunted. 12. In summary, the results of this analysis suggest a role for public action because private choices are likely to lead to less women's education than is socially desirable. When parents choose whether to send their daughter to school they are probably unaware of the impact that this choice has on the health of their yet-to-be-born grandchildren. This "external" effect may lead parents to under-invest in the schooling of their daughters. -1 18- POVERTYAND ACCESS To PUBLICSERVICES ANNEX V: DECOMPOSING THE GENDER GAP IN PARIMARY SCHOOL ENROLMENTS IN PAPUA NEW GUINEA26 I. Introduction 1. Results from the educational module of the 1996 Papua New Guinea Household Survey show that the country is a long way from achieving universal primary education. Over one-third of children between age 6 and age 20 years have never been to school and even at the age when enrolments peak (10-12 years), less than 70 percent of children are in school. There is also a large gender gap in school enrolments, with female enrolment rates declining from 68 percent at age 1 I to 45 percent at age 14 years and 13 percent by age 18. In contrast, 60 percent of males are still enrolled at age 14 and 26 percent are enrolled at age 18.27 2. The success at enrolling children in school varies widely between regions. Overall enrolment rates are highest, and the gender gaps are smallest, in the NCD and the New Guinea Islands. The overall enrolment rate is lowest in the Highlands, at less than two-thirds of the rate in the NCD and Islands regions, while the gender gap is largest in the Momase region with girls only two-thirds as likely as boys to be in the appropriate level of school. 3. A regression model of the probability of a child being enrolled in Community School (i.e., Grades 1-6) is used in this paper to gain some insights into the causes of this poor enrolment record.28 This model may help to explain whether the gap in enrolments between boys and girls and between regions is simply a function of variation in observable characteristics. This knowledge can help to determine whether it is more appropriate to use geographical, household or individual (i.e., gender) characteristics for targeting interventions designed to raise enrolment rates. II. Data and Model Specification 4. Data used in this paper come from the 1996 PNG Household Survey. The survey started while students were still on school holidays, so all questions about current enrolments referred to the previous (1995) year. The sample selected for this analysis is based on all members of respondent households aged between 7-17 years in 1995, who had not completed Community School (i.e., Grade 6) before the 1995 school year. This wide age interval is needed because over-age enrolment is a common problem in Papua New Guinea (e.g., 20 percent of the 17 year 26 This annex is based on Gibson, John, (1999f), "Decomposing The Gender Gap in Primary School Enrolments in Papua New Guinea", Mimeo, prepared for the Poverty Assessment. 27 These enrolment rates are age-specific, and show the proportion of children of a given age attending any level of school. 28 The method of analysis is similar to that recently used by Handa (1996) and Ravallion and Wodon (1999). -119- POVERTYAND ACCESS TO PUBLIC SER VICES olds in our sample were still attending Community School). These selection criteria gave a sample of 1528. The model seeks to explain the enrolment status of children, as captured by the question "Did you go to school in 1995?" A total of 755 children in this sample (corresponding to a population-weighted proportion of 0.452) attended school in 1995. 5. A wide range of variables measuring the potential determinants of enrolments are available from the survey and these are described under the following four headings: child characteristics, household characteristics, community characteristics and regional fixed effects. Child characteristics: a linear and quadratic term in the child's age and an indicator variable for whether the child's parent is the head of the household. Household characteristics: (log) household size, the share of young children (age 0-6), older children (age 7-14), and prime-age adults (age 15-50), plus a binary variable for female-headed households are included. It is difficult to identify the parents of many of the children in the sample, because household members were classified only by their relationship to the household head. So in place of the usual variables of mother's and father's education levels, the household averages of the completed school years for adult males and adult females are used. These variables should still capture the potential demand by educated adults for child enrolments because there is evidence of considerable educational spillovers within households in Papua New Guinea (Gibson, 1999). Community characteristics: the travelling time to the nearest community school using the transport method most typically used by students within the PSU. Regional fixed effects: Dummy variables for the regions (n=5) used in descriptive analysis. If these variables remain significant after controlling for the other observables it may suggest that geographical targeting is a sensible basis for programs that seek to raise school enrolment rates. MII. Results and Interpretation 6. Table 1 reports the means and standard deviations of the variables, disaggregated by sex. Table 2 reports probit estimates of the enrolment equations. To test whether estimating the model separately for boys and girls was appropriate, a dummy variable for the sex of the child was interacted with all slope variables and the model was re-estimated on the pooled sample. The test that the male intercept dummy plus all 16 of the interacted slope dummy variables were jointly zero was rejected at the p<0.08 level (F(1789)l1.69). Although one might pool the data on the basis of this result, it is worth emphasising that the hypothesis test used is a very conservative one, with degrees of freedom based on the number of clusters rather than the number of observations.29 Thus, the estimation of separate models for the enrolments of boys and girls is appropriate. 29 If the test is based on the number of observations, while still controlling for sampling weights and clustering, the pooling assumption is rejected at the p1200m) which affects climate and limits the range of crops that can be grown. 11. Regionalfixed effects: Even with the geo-climatic variables included there are likely to be spatial effects whose omission biases the coefficients on the included variables, so a set of fixed effects are included. Although it is possible that the relevant fixed effects are at the PSU level, capturing any unobserved community-level determinants of living standards, the inclusion of a PSU set of intercept dummies would make it impossible to identify any of the community level variables. As a compromise, the fixed effects are defined at regional (n=5) level. 12. The model is estimated separately for urban and rural sectors, for two reasons. First, there is likely to be genuine heterogeneity in the effect of the independent variables on the welfare ratio across sectors. For example, ownership of agricultural capital goods is likely to have less effect on consumption in urban areas, where only a small proportion of households are engaged in agricultural activities. A more practical reason is that not all of the community characteristics can be used for urban households because even though the same community questionnaire was used, the interpretation of variables differs between urban and rural sectors. For example, all urban PSUs are within minimum travelling time to a road and if they lack a particular facility within the PSU it is still readily available in adjoining PSUs (although the survey doesn't record this). The geo-climatic variables are also of less relevance in the urban sector, given that they are included to control for spatial differences in land productivity. Tables I and 2 contain descriptive statistics on the variables in the model for each sector. -128- POVERTYAND ACCESS TO PUBLIC SER VICES Table 1: Descriptive Statistics for the Model of Rural Poverty (N=830) Mean Std Dev. Minimum Maximum Ln (real expenditure per adult equivalent)a 0.429 0.763 -1.608 3.170 Demographics Household size 5.709 2.917 1 18 Number below age 15 2.458 1.853 0 11 Number above age 50 0.408 0.708 0 5 Age of household head (years) 40.410 12.783 18 85 Dummy: Female-headed household 0.079 0.269 0 1 Education Dummy: Household head is literate 0.485 0.500 0 I Average years of schooling of adults 3.107 2.750 0 12 Employment And Occupation Dummy: Head's income from minor sourcesb 0.036 0.187 0 1 Dummy: Head is tree crop farmer 0.433 0.496 0 1 Dummy: Head is in formal sector 0.185 0.388 0 1 % of adults with no cash income sources 0.261 0.322 0 1 Assets Dummy: Owns agricultural capital goods' 0.189 0.392 0 1 Number of pigs owned 2.213 3.387 0 26 Community Characteristics Index of market developmentd 7.050 6.802 0 36 Travelling time to nearest road (hours) 3.215 7.275 0.25 30 Travelling time to key social servicese 10.043 14.792 0.75 106 Geo-Climatic Variables Dummy: Wet zone (>2500mm rain/year) 0.564 0.496 0 1 Dummy: Low elevation (<600m) 0.441 0.497 0 1 Dummy: High elevation (>1200m) 0.472 0.500 0 1 Regional Fixed Effects Dummy: Papuan/South Coast region 0.154 0.361 0 1 Dummy: Highlands region 0.440 0.497 0 1 Dummy: Momase/North Coast region 0.288 0.453 0 1 Note: Means and standard deviations based on household sampling weights. The excluded dummies are male household head, illiterate head, household head's main occupation is food crop production, household owns no major agricultural capital goods, household lives in a PSU in the dry zone, at mid-elevation (600-1200m) and in the New Guinea Islands region. a The adult equivalence scale counts children age 0-6 as 0.5 adults and all others as 1.0. Nominal annual consumption expenditure is normalized by region-specific poverty lines (at national average prices). 'Includes hunting, fishing, firewood selling and making of artifacts. 'Includes trucks, tractors, sprayers, coffee pulpers, cocoa fermentaries and copra driers. ' Combined number of tradestores, public transport (PMV) businesses and fresh produce markets in the PSU. 'Combined travelling time to the nearest health centre, high school and government station (by usual means of travel for the people in the PSU). -129- POVERTYAND ACCESS TO PUBLIC SER VICES Table 2: Descriptive Statistics for the Model of Urban Poverty (N=314) Mean Std Dev. Minimum Maximum In (real expenditure per adult equivalent)a 1.365 0.909 -1.130 3.800 Demographics Household size 6.552 3.466 1 23 Number below age 15 2.675 1.771 0 10 Number above age 50 0.369 0.678 0 3 Age of household head (years) 38.838 12.419 16 72 Dummy: Female-headed household 0.087 0.282 0 1 Education Dummy: Household head is literate 0.960 0.197 0 1 Average years of schooling of adults 7.702 3.463 0 18 Employment and Occupation Dummy: Head's income from minor 0.080 0.271 0 1 sourcesb Dummy: Head is tree crop farmer 0.086 0.281 0 1 Dummy: Head is in formal sector 0.757 0.430 0 1 % of adults with no cash income sources 0.298 0.270 0 1 Assets Dummy: Owns agricultural capital goodsc 0.065 0.247 0 1 Number of pigs owned 0.381 1.436 0 8 Community Characteristics Travelling time to key social servicesd 1.011 0.388 0.75 2 Regional Fixed Effects Dummy: National Capital District (NCD) 0.349 0.477 0 1 Note: Means and standard deviations based on household sampling weights. The excluded dummies are male household head, illiterate head, household head's main occupation is food crop production, household owns no major agricultural capital goods, household lives in an urban area outside of the NCD. a The adult equivalence scale counts children age 0-6 as 0.5 adults and all others as 1.0. Nominal annual consumption expenditure is normalized by region-specific poverty lines (at national average prices). bIncludes hunting, fishing, firewood selling and making of artifacts. 'Includes trucks, tractors, sprayers, coffee pulpers, cocoa fermentaries and copra driers. d Combined travelling time to the nearest health centre, high school and government station (by usual means of travel for the people in the PSU). III. RESULTS AND INTERPRETATION 13. Table 3 and 4 contains results of the basic regression models. When community, environmental and fixed effects were excluded from the model, and urban and rural samples pooled, an F-test suggested that the coefficients on household characteristics were not the same across urban and rural sectors (F(1690)=10.30, p<0.000), supporting the decision to estimate the models separately. Other observations that can be made from the results are: -130- POVERTYAND ACCESS TO PUBLiC SERVICES Table 3: OLS Estimates of the Basic Model of Log Welfare Ratio for Rural Households (N=830) Coefficient Std Error iti p-value Demographics Household size -0.106 0.030 3.51 0.001 Household size, squared 0.005 0.001 3.62 0.001 Number below age 15 -0.070 0.020 3.45 0.001 Number above age 50 -0.015 0.044 0.34 0.738 Age of household head (years) -0.014 0.012 1.21 0.231 Squared age of household head 0.000 0.000 1.21 0.233 Dummy: Female-headed household -0.043 0.097 0.44 0.662 Education, Employment and Occupation Dummy: Household head is literate 0.213 0.056 3.81 0.000 Average years of schooling of adults 0.026 0.012 2.09 0.041 Dummy: Head's income from minor sources -0.348 0.120 2.89 0.005 Dummy: Head is tree crop farmer -0.145 0.088 1.65 0.104 Dummy: Head is in formal sector 0.252 0.102 2.47 0.016 % of adults with no cash income sources -0.142 0.097 1.46 0.148 Assets Dummy: Owns agricultural capital goods 0.200 0.072 2.78 0.007 Number of pigs owned 0.020 0.009 2.33 0.023 Community Characteristics Index of market development 0.013 0.007 1.82 0.073 Travelling time to nearest road (hours) -0.011 0.006 1.80 0.077 Travelling time to key social services -0.005 0.003 1.33 0.189 Geo-Climatic Variables Dummy: Wet zone (>2500mm rain/year) -0.224 0.089 2.50 0.015 Dummy: Low elevation (<600m) 0.104 0.162 0.64 0.526 Dummy: High elevation (>1200m) -0.059 0.148 0.40 0.693 Regional Fixed Effects Dummy: Papuan/South Coast region 0.237 0.146 1.62 0.110 Dummy: Highlands region 0.432 0.198 2.18 0.033 Dummy: Momase/North Coast region 0.095 0.169 0.56 0.575 Intercept 0.955 0.359 2.66 0.010 R2 0.322 Zero-slopes F-testa F(24.3n=10.51 0.000 Note: Results corrected for the effect of clustering, sampling weights and stratification. For notes on definition of variables, and excluded dummies see Table 1. aThis is an adJusted Wald (W)test: (d -k+I/kd)W-F(k,d -k+1), where d is thenumberofclustersminus the number of strata (60), and k is the number of slope variables (StataCorp, 1999). -131- POVERTYANDACCESS TO PUBLIC SERVICES * There are a number of significant demographic effects in both sectors. Chief among them is household size: the larger the household the lower its welfare ratio, especially in the urban sector. * There are significant gains from both extra years of schooling and literacy of the household head in both rural and urban areas, although the impact of education and literacy appears higher in urban areas. * The are significant differences in consumption level associated with occupations of the household head, with involvement in the formal sector bringing the highest gains across both sectors. * The proportion of adults in the household without access to cash incomes in the previous year does not emerge as a significant deterninant of consumption. * Ownership of agricultural capital goods raises household consumption in both urban and rural sectors, while the number of pigs positively affects consumption in rural areas but not the urban areas. * Community characteristics appear to be relevant in both sectors, with consumption in the rural sector rising with market development and as roads become more accessible, while access to social services is an important predictor of consumption in urban areas.32 * High rainfall, but not elevation, emerges as a negative influence on consumption in rural areas while regional fixed effects are relevant in both areas. 14. Datt and Jolliffe (1999) suggest that the marginal effects of household and community characteristics on consumption are not constant across households, so introduce interaction effects into their model to control for this. Although every variable can potentially be interacted with every other variable, multicollinearity is likely to result, with fragile coefficient estimates leading to potentially unreliable poverty simulations. Therefore, Datt and Jolliffe restrict attention to just a few interactions, mainly between schooling and other variables. Table 5 contains results of testing for interaction effects between schooling and some other variables in each model. The presence of interaction effects is not supported in the rural sector (p<0.43) although the coefficient on female headship is suggestive of lower returns to education for female-headed households.33 In the urban sector there does appear to be an interaction between 32 Travelling time to roads was not able to be used for the urban model because all PSUs had roads within the minimum specified time (0-30 minutes) and the index of market development wasn't calculated for urban areas because the link between access to a facility and having it physically located in the PSU does not apply in urban areas - people can easily access facilities in neighbouring PSUs. 33 Even when the other four interaction effects are dropped, the interaction of school years and female headship remains statistically insignificant (p 3 hours (-4.05) (-3.84) Decrease travelling time to road to 2 hours 31.34 n.a. 28.09 for communities where currently > 2 hours (-4.35) (-4.13) Decrease combined travelling time to 3 key 31.61 3.49 27.36 social services by 50 percenta (-3.51) (-66.51) (-6.60) Decrease combined travelling time to 3 key 31.32 n.a. 28.07 services to 6 hours if currently > 6 hours (-4.40) (-4.18) Note: Headcount index calculated as the weighted average of the predicted probability that each household is poor, following a simulated change, where the weights are the household sampling weights multiplied by the number of persons in the household. Model used to predict poverty of rural households reported in Table 3 and model for urban households reported in Table 6 (except that no control for sample stratification in comparison to those two tables. The percent change from base is calculated from the predicted baseline values. a Health centre, high school and government station. -136- POVERTYAND ACCESS To PUBLICSERVICES ANNEX VII: PRIVATE TRANSFERS AND THE SOCIAL SAFETY NET IN PAPUA NEW GUINEA34 I. INTRODUCTION 1. This annex uses a model developed by Ravallion and Dearden (1988) to study the distribution of transfer receipts amongst recipient households surveyed during the 1996 household survey. The objective is to see whether transfer receipts are effectively targeted at the needy and improve income distribution. II. THE RAVALLION AND DEARDEN MODEL 2. If tdr is the transfer made by donor d to recipient r, it is assumed that the values of tdr are chosen by the donor to maximise the function: U (Yi, XI, ..., Yd, Xd,...,Yfl,Xf) (I) where Yi is the income after transfers of household i and Xi is a vector of other attributes of i. The maximization is constrained by: Yr f (tdr, Xr) (2) n Yd = Yd> tdr (3) r=l where Yd is the donor's (fixed) pre-transfer income and it is assumed that aYr /atdr is non- negative.' At an optimum, the donor's marginal utility from giving an extra dollar cannot exceed the donor's marginal utility of own-income (otherwise they will give the extra dollar so the previous state cannot have been optimal), thus: a2u> au - and tdr Ž 0 with complementary slackness. (4) a Yd Dyl td 3.To put explicit, and tractable, functional forms on the three derivatives in equation (4), Ravallion and Dearden assume that: (i) each donor's utility function is of constant elasticity of substitution (CES) in the vector of recipients' incomes, and the elasticity of substitution is constant across donors, (ii) donor's utility functions are separable between own-income and the incomes of others, (iii) donor's marginal utility of own-income is iso-elastic in own-income, with the same elasticity across donors (iv) recipient's incomes are iso-elastic in their total transfer receipts, with a constant elasticity across recipients. 34 Based on Gibson, J. (1999c), "Is there a Malenesian Moral Economy? Private Transfers and the Social Safety Net in Papua New Guinea", Mimeo, Prepared for Poverty Assessment. -137- POVERTYAND ACCESS TO PUBLIC SER VICES Under these assumptions: =g(Xd )Yd (IT > O) Yd au hI(Xr )Y (>o 0) (5) at = k(Xr )Y, ltdr where 7r is the elasticity of the donor's marginal utility of own-income and E is the coefficient of relative inequality aversion, with weight increasingly placed on the lowest income as -* Co (i.e., the Rawlsian case). 4. If a transfer is made, the optimality condition (equation (4)) implies that: tdr =g(Xd)Ih(X,)k(X,)Y,6 Yd. (6) 5.When equation (6) is summed over donors and recipients, the following allocation of transfer receipts is obtained: n Ti = ,tdi =a -h (Xi)k(Xi)Yi forieR d=1 =0 otherwise (7) where Tir is transfer receipts, a = E g(Xd ) 1 Yd and R is the set of all households with positive receipts. Outlays are allocated according to: Tid = tir =b-g(Xj)lYj foriED r=1 = 0 otherwise (8) where Ti.d are transfer outlays, b = E h(Xr ) k (Xr )Yk'-8 and D is the set of all households with positive donations. Taking logs of equation (7) and (8), assuming that the fumctions h, g, and k are log-linear, and introducing stochastic error terms gives the following econometric model: logT/ = 1 +11og17 + Xi7y +±u1i if ieR Tir =0 otherwise (9) d log T = a2 +,62 log Yi + Xiy2 + l2i if i E D d Tid = 0 otherwise (10) where the a's, ifs, and i s are parameters to be estimated and the u's are normally, independently and identically distributed errors. III. DATA 6. Data used in this paper come from the Papua New Guinea Household Survey. The survey obtained comprehensive information on the private transfers between households. The questions -138- POVERTYAND ACCESS To PUBLiC SER VICES about the production and purchase of each food item during the consumption recall period also asked about inward and outward transfers of those food items, with a similar format used for the non-food frequent expenses. The questions about each category of infrequent expenses also asked about the value of in-kind gifts given and received of these items during the previous 12 months. Details were also obtained on each inward and outward monetary transfer exceeding K50 in the previous 12 months, with additional questions asked about the destination (origin) of this transfer, the relationship of the household head to the recipient (donor), and whether any specific purpose was stipulated for the transfer. IV. MODEL SPECIFICATION 7. Table 2 describes the explanatory variables used in the models of transfers. Where possible, the variables used by Ravallion and Dearden (1988) are chosen. One exception was that the survey did not obtain inforrnation on deaths within the household during the previous year. The replacement variable used is whether the household is headed by a female, which may reflect widowhood which is likely to motivate transfer receipts in the same way that a recent death does. Also, because the Ravallion and Dearden study was for just one province in central Java (Yogyakarta), while the current study is for the whole of PNG, regional dummy variables have been added to the model. These dummies should capture regional price differences and also proxy for other excluded locational effects. Table 2: Explanatory variables used in the models of transfers Mean (Std. Dev) Variable Rural Urban Description Expenditures 3736 11299 Annual value of household total expenditure (including (3812) (8785) imputed value of stock changes, own-production, and gifts received) and excluding the value of gifts given. Household size 5.709 6.552 Total number of residents usually present in the (2.92) (3.47) household at the time of the survey. 111-health 0.727 0.592 Average number of times treated at health facilities per (1.33) (1.70) person in the month prior to the survey. Births 0.193 0.225 Dummy variable =1 if anyone in the household was (0.39) (0.42) born in the 12 months prior to the survey. Head's age 40.4 38.8 Years of age of the household head. (12.8) (12.4) Female head 0.079 0.087 Dummy variable =1 if the household head is female. (0.27) (0.28) Unemployment 0.150 0.057 Dummy variable =1 if household head had no source of (0.36) (0.23) cash income in the 12 months prior to the survey. -139- POVERTYAND ACCESS To PUBLICcSER VICES The income term in Ravallion and Dearden's model is a metric of each household's welfare after transfers, and in this regard it differs from other studies which focus on pre-transfer income. The empirical income variable used by Ravallion and Dearden was the predicted value of (log) total household consumption of all goods and services, excluding expenditure on gifts. Predicted values were needed because household consumption is a function of transfer receipts, and this simultaneity violates the independence of the regressor from the model error term. The instruments used for log consumption by Ravallion and Dearden were: labor earnings, income from enterprises, rents received, years of schooling, dwelling characteristics, occupational dummy variables, and the variables in the Xvector in equations x and y. 8. It was not possible to use all of these instruments in the current study because the PNG survey did not calculate cash incomes. However, data on certain assets were available (the value of household durables, ownership of agricultural capital goods like coffee pulpers, copra driers and trucks, and the number of pigs owned), and these should be correlated with cash incomes. Additionally, a cluster-level variable measuring the travelling time to the nearest transport facility (road, airstrip or port) was included for the rural sector, because inaccessibility is a major constraint on cash incomes in PNG. The other instruments included the head's years of schooling, characteristics of the dwelling (floor area and roof type) and a vector of three occupational dummy variables indicating the household head's main source of cash income (agriculture, running a business, wages). These instruments raised the R2 in the first stage equation predicting log expenditures from 0.33 to 0.47 in the rural sector and from 0.20 to 0.60 in the urban sector and F-tests suggested that the instruments were highly significant predictors of log expenditures (see Appendix Table 1). -140- POVERTYAND ACCESS TO PUBLICSERVICES V. RESULTS AND INTERPRETATION 9. The coefficient A1 in equation (9) can be considered the elasticity of transfer receipts with respect to income. Three scenarios for the value of A pose different interpretations on the preferences that the donor has over the resulting income distribution. If A = 1, transfers rise or fall proportionate to recipient households change in income, and therefore transfers are a constant share of recipient household income. A donor giving to two households, one rich and one poor, makes transfers proportionate to each households income so if before transfers the rich household was twice as rich as the poor household, it will also be twice as rich after the transfer. The distribution of income is not altered, implying no aversion to inequality on the part of people donating transfers. If ,Ai < 1, transfers do not rise as fast as income rises, so transfers would contribute a bigger share to the post-transfer consumption of a poor household than of a rich household, and in this way transfers reduce disparities in income and makes the distribution more equal. Therefore, an elasticity of transfer receipts with respect to income less than one implies aversion to inequality on the part of the people donating transfers. Conversely, if A, > 1, transfers increase disparities in income and makes the distribution more unequal than it was before, indicating donors' preference for inequality. 10. Table 3 reports the results for gross and net receipts of transfers. Recall that gross transfers are relevant if it is assumed that donors hold myopic expectations. Specifically, myopic expectations hold if donors decide how much to donate without taking account of any reverse transfers that the recipient may make to them in future and also ignoring any passing on of their donation by the recipient to someone else. If donors hold rational expectations of recipients' behaviour, taking into account return transfers and the passing on of transfers to others, then it is the net transfer that is relevant. Although the information requirements for rational expectations are high, it is plausible that donors who are engaged in long term relationships with recipients, and are often kinsfolk, can correct anticipate recipients behaviour on average. 11. In the Table 3 results, the hypothesis that ,A,=1 (no inequality aversion) is rejected at the p<0.04 level in favour of the alternative hypothesis that ,B,>1 (i.e., transfers are inequality increasing) when using gross receipts. The null of no inequality aversion is not rejected for the rural sector in the model using net receipts (p 0X) 0.926 0.890 0.507 0.487 Note: Reported absolute t-values are corrected for clustering, sampling weights and sample stratification. F- test is an adjusted Wald (W) test: ((d - k + l)/kd) W-F(k, d - k + 1), where d is the number of clusters minus the number of strata (60 for rural and 45 for urban), and k is the number of slope variables. The excluded region in the rural sector is the New Guinea Islands, and in the urban sector it is all urban areas outside of the National Capital District (NCD). 'Predicted value used for In expenditures, with predictions from the model in Appendix Table 1. -143- POVERTYANDA CCESS TO PUBLICSERVICES Table 4: Two-stage Tobit estimates of the receipts equation - day-to-day transfers Gross Receipts Net Receipts Rural Urban Rural Urban 13 itt 1 itt D Itl D itt In expendituresa 1.259 4.08 0.385 0.80 0.819 1.33 0.335 0.34 In household size -0.448 1.84 -0.165 0.40 -0.210 0.40 0.098 0.16 111-health 0.091 1.82 0.107 1.30 0.124 0.98 0.132 1.05 Births (=1) 0.545 2.23 -0.506 0.92 1.816 3.37 -0.531 0.62 Head's age -0.049 1.20 0.017 0.15 0.145 1.61 -0.095 0.50 [Head's age]2 0.000 0.97 0.000 0.22 -0.001 1.23 0.001 0.66 Female head (=1) 0.695 2.20 0.742 2.02 1.167 1.99 2.864 3.04 Unemployment (=1) -0.040 0.15 0.477 0.76 0.123 0.23 -2.071 1.34 Papuan region (=1) -0.683 1.70 -3.115 3.95 Highland region (=1) -0.793 2.64 -1.342 1.79 Momase region (=1) -1.132 2.54 -2.129 2.56 NCD region (=1) 0.202 0.32 1.254 1.48 Constant -2.985 1.29 0.953 0.25 -7.759 1.57 -0.714 0.10 Zero-slopes F-test 5.45 2.94 6.28 4.79 Prob (T > 01X) 0.865 0.815 0.511 0.551 Note: Reported absolute t-values are corrected for clustering, sampling weights and sample stratification. F-test is an adjusted Wald (W) test: ((d - k + I)/kd) W-F(k, d - k + 1), where d is the number of clusters minus the number of strata (60 for rural and 45 for urban), and k is the number of slope variables. The excluded region in the rural sector is the New Guinea Islands, and in the urban sector it is all urban areas outside of the National Capital District (NCD). a'Predicted value used for In expenditures, with predictions from the model in Appendix Table 1. -144- POVERTYANDACCESS TO PUBLICSERVICES Table 5: Two-stage Tobit estimates of the receipts equation - infrequent transfers only Gross Receipts Net Receipts Rural Urban Rural Urban 3 iti 1 ItI D It Iti In expendituresa 2.379 4.46 1.098 1.38 1.126 1.53 -1.104 0.74 Inhousehold size -0.188 0.51 -0.188 0.26 0.359 0.55 1.246 1.38 Ill-health -0.053 0.47 -0.116 0.52 0.064 0.34 -0.146 0.34 Births (=1) 0.841 2.39 0.655 0.60 1.562 3.21 -1.021 0.31 Head's age -0.088 1.39 -0.346 2.06 -0.106 1.09 -0.710 3.15 [Head's age]2 0.001 1.58 0.005 2.57 0.001 1.24 0.008 3.34 Female head (=1) 0.398 0.58 -2.507 1.57 0.893 0.94 -2.216 0.77 Unemployment (=1) 0.408 0.89 1.195 0.85 1.014 1.74 2.407 0.85 Papuan region (=1) -1.191 1.08 -0.148 0.12 Highland region (=1) 0.424 0.52 0.465 0.47 Momase region (=1) -0.166 0.18 0.104 0.09 NCD region (=1) 0.755 1.30 -0.449 0.28 Constant -14.85 3.41 -2.468 0.31 -9.262 1.52 18.260 1.49 Zero-slopes F-test 8.68 5.49 3.86 6.09 Prob (T > ol0x) 0.634 0.559 0.393 0.256 Note: Reported absolute t-values are corrected for clustering, sampling weights and sample stratification. F- test is an adjusted Wald (W) test: ((d - k + I)/kd) W-F(k, d - k + 1), where d is the number of clusters minus the number of strata (60 for rural and 45 for urban), and k is the number of slope variables. The excluded region in the rural sector is the New Guinea Islands, and in the urban sector it is all urban areas outside of the National Capital District (NCD). a Predicted value used for In expenditures, with predictions from the model in Appendix Table 1. -145- POVERTYAND ACCESS TO PUBLIC SERVICES Table 6: Two-stage Tobit estimates of the outlays equation Gross Outlays Net Outlays Rural Urban Rural Urban In expendituresa 1.630 6.97 0.779 2.16 0.284 0.48 1.972 1.56 ln household size -0.901 4.51 -0.339 1.00 -0.190 0.38 -0.346 0.54 Ill-health -0.071 1.28 -0.255 2.83 -0.126 0.76 -0.224 0.81 Births (=1) 0.099 0.50 -0.066 0.13 -2.040 2.80 -0.700 0.64 Head's age -0.014 0.37 -0.135 1.37 -0.167 1.41 0.084 0.34 [Head's age]2 0.000 0.20 0.001 1.48 0.002 1.31 -0.001 0.37 Female head (=1) -0.495 1.18 -1.192 1.89 -1.185 1.18 -2.776 1.79 Unemployment (=1) 0.038 0.11 0.427 0.67 -1.574 2.94 0.641 0.26 Papuan region (=1) 0.368 1.21 3.596 3.60 Highland region (=1) 0.122 0.53 2.108 2.24 Momase region (=1) 0.022 0.07 1.767 1.76 NCD region (=1) 0.212 0.70 -0.618 0.55 Constant -5.847 3.16 2.100 0.48 0.910 0.17 -17.23 1.60 Zero-slopes F-test 8.34 6.58 6.12 4.15 Prob(T>OIX) 0.918 0.916 0.473 0.496 Note: Reported absolute t-values are corrected for clustering, sampling weights and sample stratification. F-test is an adjusted Wald (W) test: ((d - k + )/Ikd) W-F(k, d - k + 1), where d is the number of clusters minus the number of strata (60 for rural and 45 for urban), and k is the number of slope variables. The excluded region in the rural sector is the New Guinea Islands, and in the urban sector it is all urban areas outside of the National Capital District (NCD). a Predicted value used for In expenditures, with predictions from the model in Appendix Table 1. -146- POVERTYAND A CCESS TO PUBLIC SERVICES Appendix Table 1: First stage regressions predicting In (household expenditures) Rural Urban Itl iti Head's years of school 0.031 4.03 0.038 4.67 Floor area of dwelling 0.003 2.23 0.006 7.21 Iron roof on dwelling 0.457 5.45 0.382 1.43 Agricultural capital goods owned 0.155 2.01 0.286 1.86 Number of pigs owned 0.022 2.44 -0.040 1.21 Value of household durables 0.000 6.65 0.000 9.56 Walking time to nearest road etc' -0.001 1.22 ... ... Household head's main cash incomefrom: Agriculture 0.200 1.60 -0.504 3.86 Running a business 0.274 1.71 -0.063 0.21 Wage employment 0.425 3.07 -0.032 0.16 In household size 0.498 8.60 0.098 1.26 Ill-health 0.043 2.62 0.011 0.88 Births (=1) -0.087 0.81 0.263 1.52 Head's age -0.005 0.39 -0.001 0.04 [Head's age]' 0.000 0.53 0.000 0.10 Female head -0.003 0.03 0.014 0.14 Unemployed head (nil income) 0.085 0.55 -0.279 1.29 Regional intercepts Papuan/South Coast 0.376 3.13 Highlands 0.449 3.38 Momase/North Coast -0.183 1.20 National Capital District * - 0.212 1.83 Constant 6.308 17.42 7.728 13.76 R 2 0.47 0.60 Zero-slopes F-test 25.47 185.99 F-test on additional instruments 20.47 77.50 N=830 N=314 Note: Reported absolute t-values are corrected for clustering, sampling weights and sample stratification. F- test is an adjusted Wald (W) test: ((d - k + I)/kd) W-F(k, d - k + 1), where d is the number of clusters minus the number of strata (60 for rural and 45 for urban), and k is the number of slope variables. a Airstrips and berthing points for water transport vessels are used if they are closer than roads. All urban areas have transport facilities accessible within the minimum category for walking time (0-30 minutes). This does not rule out disincentive effects of transfers, which occur if the derivative is less than one, nor positive productivity effects of transfers, which occur if the derivative exceeds one. -147-