Analysis of the Mismatch between Tanzania Household Budget Survey and National Panel Survey Data in Poverty and Inequality Levels and Trends

This study carries out a thorough investigation of the potential sources of mismatch in poverty and inequality levels and trends between the Tanzania National Panel Survey and Household Budget Survey. The main findings of the study include the following. First, the difference in poverty levels between the Household Budget Survey and the National Panel Survey is essentially explained by the differences in the methods of estimating the poverty line. Second, the discrepancy in poverty trends can be mainly attributed to the difference in inter-year temporal price deflators, and, to a lesser extent, spatial price deflators. The use of the consumer price index for adjusting consumption variation across years would show a decline in poverty during the past five years for the Household Budget Survey and the National Panel Survey. Third, despite noticeable differences in the methods of household consumption data collection, the Household Budget Survey and National Panel Survey show close mean household consumption levels in the last rounds, when using the consumer price index to adjust for inter-year price variations. Mean household consumption levels in the Household Budget Survey 2011/12 and National Panel Survey 2010/11 are comparable, and the mean consumption level in the National Panel Survey 2012/13 is around 10 percent higher. The difference is driven by higher levels of aggregate and food consumption by the better-off groups in the National Panel Survey. Fourth, the mismatch in inequality trends and pro-poor growth patterns between the two surveys could not be resolved and is a subject for further analysis.


Policy Research Working Paper 8361
This paper is a product of the Poverty and Equity Global Practice. It is part of a larger effort by the World Bank to provide open access to its research and make a contribution to development policy discussions around the world. Policy Research Working Papers are also posted on the Web at http://econ.worldbank.org. The authors may be contacted at nbelghith@ worldbank.org.
This study carries out a thorough investigation of the potential sources of mismatch in poverty and inequality levels and trends between the Tanzania National Panel Survey and Household Budget Survey. The main findings of the study include the following. First, the difference in poverty levels between the Household Budget Survey and the National Panel Survey is essentially explained by the differences in the methods of estimating the poverty line. Second, the discrepancy in poverty trends can be mainly attributed to the difference in inter-year temporal price deflators, and, to a lesser extent, spatial price deflators. The use of the consumer price index for adjusting consumption variation across years would show a decline in poverty during the past five years for the Household Budget Survey and the National Panel Survey. Third, despite noticeable differences in the methods of household consumption data collection, the Household Budget Survey and National Panel Survey show close mean household consumption levels in the last rounds, when using the consumer price index to adjust for inter-year price variations. Mean household consumption levels in the Household Budget Survey 2011/12 and National Panel Survey 2010/11 are comparable, and the mean consumption level in the National Panel Survey 2012/13 is around 10 percent higher. The difference is driven by higher levels of aggregate and food consumption by the better-off groups in the National Panel Survey. Fourth, the mismatch in inequality trends and pro-poor growth patterns between the two surveys could not be resolved and is a subject for further analysis.

I. Background and Main Objectives
The official poverty figures announced by the Government of Tanzania in November 2013 revealed a decline in the basic needs poverty rate from around 34 percent to 28.2 percent between 2007 and 2012-this is the first significant decline the country has experienced during the last 20 years. This reduction in poverty has been confirmed in the recently published Poverty Assessment report for Mainland Tanzania. The report examined the recent trends in poverty and inequality and their determinants and explored the responsiveness of poverty reduction to economic growth using the Household Budget Survey (HBS) collected in 2011/12. The report shows that poverty dropped by approximately 1 percentage point per year between 2007 and 2011/12, and that inequality, measured by the Gini coefficient of real per capita monthly consumption, declined from around 39 to 36 during the same period. The report also found emerging signs of pro-poor growth, despite a persistently high number of people living in poverty.
However, the declining trend in poverty revealed by the HBS data is in contrast to the increasing trend that is observed using the National Panel Survey (NPS) data, which show an increase in poverty from 14.6 percent to 18.1 percent and subsequently to 21.2 percent between 2008/09, 2010/11 and 2012/13. The data also show a slight increase in inequality across the three rounds and do not support the pro-poor growth pattern revealed by HBS data. Although HBS is the official source for official poverty numbers, this mismatch in poverty levels and trends between the two surveys is puzzling. The NPS is a national level longitudinal survey designed to collect data from the same households over time in an attempt to better track the progress of the National Strategy for Growth and Reduction of Poverty (MKUKUTA), understand the poverty dynamics and evaluate policy impacts. This study aims to carry out a thorough investigation of the potential sources of mismatch in poverty and inequality levels and trends between the NPS and HBS. The investigation will focus on the key candidate sources for the divergence between the two surveys. These include the methodological differences in the construction of the consumption aggregates and the estimation of the poverty lines, Source: HBS 2007and 2011/12 and NPS 2008/09, 2010/11 and 2012 Discrepancies in poverty incidence and trends are observed at the regional level. As shown in Table   1, the poverty level in the rural areas that was estimated using the NPS data is two times lower than that estimated using the HBS data, while the poverty incidence in urban areas measured using NPS is over three times lower than that estimated using HBS. In addition, HBS suggests a decline in poverty in all regions, while NPS indicates a decline in poverty only in Dar es Salaam (and Zanzibar).   At the regional level, both NPS and HBS reveal that inequality is higher in Dar es Salaam and secondary cities than in rural areas. HBS data suggest that the distribution of consumption equalized over time in all the regions, with the most substantial improvement occurring in the rural areas, as can be seen from the changing shape of the Lorenz curves in Figure 3. Much of the reduction in inequality seems to have been driven by an increase in the welfare share accruing to the poorest segment of the population, as the consumption share of the poorest quintile grew by more than 16 percent between 2007 and 2011/12, with an increase exceeding 20 percent in the rural areas. In contrast, NPS data suggest a deterioration of the distribution of welfare particularly in the rural areas where inequality seems to have significantly increased between 2008 and 2013. As shown in Figure 4, the increase in inequality seems to have been driven by a deterioration of the consumption share accruing to the poorest population groups, which declined by around 12 percent at the national level and by 11 percent in rural areas. The consumption share of the poor seems however to have improved in the urban sector, particularly in Dar es Salaam where it increased by over 20 percent.

C. Distributional pattern of growth
Using changes in household consumption as the measure of growth, this section examines whether both NPS and HBS support the emerging signs of "pro-poor" growth during the recent years. The Growth Incidence Curve (GIC) for HBS 2007-2011/12, which shows the percent change in average consumption for each percentile of the distribution, is downwardly sloped, indicating higher growth among the poorest ( Figure 5). However, the pattern of real consumption growth using HBS differs from that using NPS as indicated by the upwardly sloped GIC for NPS 2008/09-2012/13 which suggests that the richer groups were the main beneficiaries of growth.

D. Comparison of HBS and NPS density functions
To further explore the differences in expenditures between NPS and HBS, we plot the kernel density functions of consumption expenditure of both surveys in Figure 6.a and Figure 6.b. Figure 6  The differences between the HBS and NPS data sets in poverty and inequality levels and trends as well as in the distributional pattern of growth are puzzling and deserve an in-depth investigation of the sources of such differences. The following sections will explore whether these discrepancies are due to differences in survey methods or to differences in the approaches used for the construction of the consumption aggregates, measurement of the poverty lines or estimation of the price deflators.
We start by reviewing the survey methods, focusing on the method of data capture, length of reference period for reporting consumption and the level of commodity details. We then examine the approaches used to estimate poverty and consumption in both surveys and investigate how the estimates vary if the same approach is used in both data sets.

III. Comparison of HBS and NPS Survey Methods
The section reviews the differences between the HBS and NPS data sets in survey characteristics and methods of collecting consumption data.  Table 2 presents some general characteristics of both surveys including the date and duration of the fieldwork, sample size, and total (weighted) estimated population. The HBS has a large sample size and is restricted to the mainland, while NPS has a smaller sample size but covers the Zanzibar archipelago. However, eliminating Zanzibar from the comparative analysis does not induce significant differences in the results. The sample sizes in both the HBS and NPS are considered large enough to give reasonably precise estimates of poverty at the national level and by geographic domain (rural, other urban, and Dar es Salaam). While the increase of the population of approximately 25 percent shown by the NPS data appears to be an overestimate, adjusting the NPS weights using a linear interpolation of population between the 2002 and 2012 census and re-scaling weights accordingly revealed only slight changes in the NPS poverty and inequality levels and no changes in the trends. This points to the need to further investigate this issue.

A. Survey characteristics
Also, there may be concerns about the potential effect of attrition in the panel surveys since nonrandom attrition can cause the survey samples to become unrepresentative of the general population over time. However, the attrition rates of 3 percent in NPS 2010/11 and 4 percent in NPS 2012/13 are too low to significantly bias poverty and inequality estimates and to affect their trend. It should be noted that the NPS used propensity score matching to address attrition by compensating for the lost households.
The HBS and NPS data were collected over a period of approximately 12 months, which excludes seasonality as a potential source of the mismatch in poverty and inequality indicators. While the survey years are relatively close, the advent of the financial crisis in 2008 might have induced substantial changes in household consumption patterns that may have been captured by NPS 2008/09 and NPS 2010/11, which could explain the increase of poverty between the two rounds.  Table 3 compares the methods of consumption data collection between HBS and NPS. The two surveys diverge in several aspects, which induce fundamental differences in survey design, explained below, that are difficult to adjust ex-post and will remain present when comparing the two surveys' results. Households were asked to report the total expenditures for women's education and for men's education. 6 Households reported their expenditures on school fees, books and materials, uniforms, transport, extra tuition, other contributions, and the cost of meals.

B. Consumption data collection methods
The main differences between the HBS and NPS in consumption data collection methods can be summarized as follows:

i. "Recall" versus "Diary" method
Both the HBS 2007 and 2011/12 used a 28-day diary to collect data on food consumption. In 2007 the diary was administered for the whole month but at the analytical step this was adjusted to create expense values for 28 days. Diaries in the HBS 2007 always started at the beginning of the month, while diaries in the HBS 2011/12 were staggered across the months. In both surveys, each household member aged five years and above was asked to fill out a 'booklet' to record his/her daily transactions for consumption purposes, including consumption of own-produced items. Enumerators were then instructed to transcribe the data from these booklets into the main diary form (every other day).
However, in practice this might not have happened in this manner since enumerators were expected to have worked with one household member every other day to fill in the main diary form directly (rather than transcribing the information from the booklets).
In contrast, the NPS uses a seven-day recall method to collect data on food consumption, asking the head of the household or their spouse to recall how much they consumed of various food items in the past seven days. According to a study by Beegle et al. (2012), the diary method produces lower food and total consumption aggregates, higher poverty levels and lower inequality levels, though the variations reported in the study are not as important as those observed between the HBS and NPS data sets.
Food consumed outside the household is captured in the HBS through an additional diary filled in only by adult household members, while it is collected in the NPS by way of recall of the last seven days asked to all household members. The HBS 2007 does include the data on food consumed outside.
Non-food items were collected using both the diary and recall methods in the HBS, while in the NPS they were collected using the recall method only. To avoid the duplication of expenditures under the diary and recall module in the HBS, an effort was made in both years to exclude non-food items already recorded in the diary from the recall module. In 2011/12, interviewers were asked to carefully check potential duplication between non-food items reported in the recall and those recorded in the diary. Potential duplication was also carefully checked at the analytical stage during the evaluation of the welfare aggregates and poverty estimates. In the HBS 2007, interviewers were asked to "request details of irregular purchases of consumer durables and costs of other services during previous twelve months excluding the survey month". Excluding the survey months from the recall most likely also intended to reduce double-counting. Potential duplication was also checked at the analytical stage following the same procedure as in 2011/12.
The HBS 2007 used a 12-month recall period for the collection of non-food items with the exception of rent, while the HBS 2011/12 used recall periods of 1, 3 and 12 months depending on the item (see Table 3). The NPS used a 12-month recall for some non-food items and recall periods of 1 and 4 weeks for others such as transportation, health and education. While changes in the recall period can affect the welfare and poverty estimates, the induced variations would not be expected to be as high as those observed above in Section II (see Beegle et al., 2012).

ii. Evaluation of home-produced food
For many Tanzanians, particularly those living in the rural areas, most of their caloric intake comes from food that they produced themselves. The value of own-produced food is difficult to evaluate as its market price cannot be directly observed. Different methods have been suggested in the literature to estimate own-produced food values, each with its many pros and cons. In the HBS, the value of own-produced food is reported directly by the household. Respondents are asked to report a shilling value for all food consumed, whether it is purchased or produced at home. In the NPS, the valuation of own-produced food is based on the prices paid by the household for similar items in the same geographic stratum.

iii. Degree of commodity details
Another key difference between the two surveys arises from the degree of commodity details. The list of non-food items collected in the HBS is more extensive than in the NPS. This is particularly true for HBS 2011/12 where households were provided a very detailed list of the items to be reported in the recall module. For example, HBS 2011/12 solicited information for over 300 non-food items compared to the 52 non-food items solicited in NPS 2012/13.
According to Beegle et al. (2012), a more detailed commodity list is expected to lead to higher consumption aggregates and lower poverty levels, but the differences between HBS and NPS poverty measures is contrary to what should be expected as HBS poverty indicators are significantly higher than NPS indicators.
Further, it is worth noting that the HBS survey instruments have improved significantly over time, while there were no substantial changes in survey methods between the NPS waves except for a few additions in the questionnaire for the third round. Great attention is generally devoted to the supervision of the NPS. To ensure strict control over data quality during fieldwork, the NPS survey uses a smaller and more closely supervised group of enumerators. The survey uses mobile teams, each consisting of seven people (1 supervisor, 4 enumerators, 1 data entry operator and 1 driver).   Since the decline in poverty between 2007 and 2011/12 using the HBS data may be due to the changes in the survey design, different imputation methods are used to address this issue and check whether the reduction in poverty is a reality. The different prediction approaches supported the decline in poverty between 2007 and 2011/12, although they revealed a slightly lower pace of poverty reduction, suggesting that the improvements of survey methods in the HBS are not the cause of the difference in poverty trends between HBS and NPS data.    Both the HBS and NPS used the Fisher price index to adjust for spatial and intra-year differences in the cost of living. In the HBS, separate food and non-food Fisher price indices are estimated based on unit values (value/quantity) from the survey data. The overall (food and non-food) price deflator was computed using the weighted average of food and non-food indices, where the weights were the average budget shares on total nominal food and non-food consumption. Price indicators were calculated by geographic stratum and quarter. The NPS data used a similar method to adjust for spatial and intra-year price differences, but the Fisher price index was based on food unit values only. Table 6 compares the values of the spatial price deflators by survey quarter and strata and shows no To better understand the potential effect of the differences in (survey based) inflation rates on the consumption trends in HBS and NPS, we also use the CPI to adjust for inter-year temporal price variations.  As most of the difference is coming from the significantly higher consumption level of the richest quintiles, we would expect either higher underreporting in the HBS or significant differences in the sampling between the two surveys. This point will be discussed below.

D. Aggregate consumption
When adjusting for inter-year price variations using the survey price deflators, mean household consumption levels appear to be higher in the HBS 2011/12 than in NPS 2012/13. The difference seems to be due to much lower consumption levels of the poorest quintiles in the NPS than in the HBS. While there is almost no difference between HBS and NPS consumption levels of the richest population groups, the difference seems to be very important for the poorest groups attaining around 40 percent. Also, HBS shows an increase in mean household consumption levels over time, mainly driven by an increase of the consumption of the poorest groups, while NPS shows a decline in mean household consumption levels over time mainly driven by a reduction of the consumption of the poorest quintiles.
As these figures are difficult to compare due to the differences in the inter-year price deflators and base year, we also use the CPI to adjust for inter-year price variations and take HBS 2011/12 as base year for all HBS and NPS rounds. This reduces the discrepancy between HBS and NPS mean household consumption levels and shows a similar upward trend in mean household consumption over time for both surveys. While we continue to observe a larger increase in mean household consumption in the HBS than in the NPS and a larger increase in consumption levels of the better off in the NPS compared to HBS, both surveys now display improvements over time of mean household welfare, particularly for the better-off groups.
The main differences that stand out can be summarized as follows: i) There is a decline in aggregate consumption as well as in food consumption between NPS 2008/09 and NPS 2010/11, no matter the inter-year price deflator used. This can be explained by the advent of the financial crisis and food price hikes in 2008 whose effects may have started to appear after 2009. The decline of food consumption levels while non-food consumption remained stable lends support to this presumption.
ii) There is a decline in overall consumption between NPS 2010/11 and NPS 2012/13 (using the survey inter-year deflators) for the three poorest quintiles of the population, with the decline being more substantial for the poorest quintile. This decline seems to be driven by a reduction of food consumption accompanied by an even greater reduction of non-food consumption.
However, this decline vanishes when the CPI for inter-year price adjustment is used, as we observe an increase in consumption levels of all population groups including for the poorest quintiles even though the improvements remain more substantial for the better-off groups.
iii) In contrast with NPS, the HBS data show a significant increase of the food and total consumption levels of the poorest segments of the population and a slight reduction of the food consumption levels for the richest group. The following sections will explore other potential sources of the mismatch between the two surveys.  Education expenses did not come from the Education section where questions were asked for each household member, they come from Form II which had a very detailed section specific for education expenditures, divided into private, public, formal, informal. Unlike the NPS, the structure of the questions was very similar to the other nonfood items.
Education expenses were included in household expenditures but they were collected separately (Section C). Information is collected for each household member over 5 years old with a recall period of 12 months (question 14 in 2008/09, question 28 in 2010/11 and 2011/13). Total expenses were calculated by the numerator by adding up individual expenses. Notice that these expenses included some clothing (uniforms) and footwear (shoes).
Health expenses did not come from the Health section where questions were asked for each household member. There was a separate section with 15 questions for health expenditures. Unlike the NPS, the structure of the questions was very similar to the other nonfood items.
Health expenditures were included in household total expenditures but they were collected separately. The NPS collected information for each household member over 12 years and old (Section D). Some questions referred to the past four weeks, some to the past 12 months (question 13 onwards) but they seemed properly harmonized in the do files. Health expenditures included expenses related to visits to the health practitioner, health treatments, hospitalization and medications.
There is a section for vehicle and a separate section for public transport expenses. Transport expenses included: Public transport (7 days recall), "petrol or diesel", "motor vehicle service, repair, or parts", "bicycle service, repair, or parts" (30 days recall).
Communications included telephone landlines, mobile phones, personal computers, satellite decoders.
Communication expenses had a 30-day recall period (Section L). They included: "cellphone vouchers" and "phone, internet and postal services". Recreation and spare time section was much more complete, it solicited information about: sport and camping equipment; swimming pools, gym, tennis courts expenditures; tickets to sporting shows, concerts, theater, museums, etc.; lottery tickets, photographic equipment, musical instruments, amusement items, etc.
Recreation expenditures were collected using a recall period of 12 months (Section M). It included: "sports and hobby equipment, musical instruments, toys" and "film, film processing, camera". There were only two questions in this category and the reported values were low compared to the HBS.
Detailed questions about the main dwelling expenditures included questions on: electric power; fixed telephones; mobile phones; TV subscriptions; internet subscriptions; water; common expenditures such as lighting cleaning on primary and secondary buildings; gas, charcoal, kerosene, coal, and firewood.
Miscellaneous non-food expenditures included around 25 detailed questions about furniture and furnishing, tools and appliances for household maintenance, small electric household appliances, dishes, utensils and domestic workers.
-Household expenses had a 30-day recall period and included: milling fees and grain; household cleaning products (dish soap); wages paid to servants; repairs to household and personal items; carpet, rugs, drapes, curtains; linens, towels, sheets, blankets; mats for sleeping or for drying maize flour; mosquito nets, mattresses; repairs to consumer durables. -Miscellaneous non-food expenditures had a 30-day recall period and included: bar soap (body soap or clothes soap); laundry soap (powder); toothpaste, toothbrush; toilet paper; glycerin, Vaseline and skin creams; other personal products (shampoo, razor); insurance (health or auto); other costs not stated anywhere.
There is a section on travel, holidays and hotels outside and inside Tanzania and another one on restaurants.
The non-food consumption did NOT include expenses incurred at restaurants (This is different from HBS). However, the NPS collected detailed information on food consumed outside the home (which included full dinner with a 7-day recall period) in Section F.
Alcohol expenditures included information from the non-food question as well as from Section F (food outside household), which included beer, wine, or hard drinks consumed outside the household in the past 7 days. Table 8 presents average levels of food consumption per adult per month for different sub-groups of food commodities. All values are presented in nominal terms, real terms adjusted for the spatial and seasonal differences in the cost of living, and for inter-year price variations using the CPI and taking HBS 2011/12 and NPS 2010/11 as base periods for HBS and NPS, respectively. The number of food items is much higher in the HBS than in the NPS. The level of consumption expenditure seems to be higher on food items such as meat, milk and cheese, fruits, sugar, and coffee, tea and soft drinks in the NPS than in the HBS. Both surveys show similar trends in consumption on most items, except for bread and cereals, fish, milk, cheese and eggs, and vegetables, with the variations being much more important in the HBS than in the NPS.  Figure 9 presents the shares of food groups in total food consumption. It shows that Tanzanians tend to consume bread and cereals the most, which make up about one-third of food consumption. This is consistent across the HBS and NPS surveys. The second most consumed food group is vegetables for the HBS and meat for the NPS.  Table 9 shows non-food consumption per adult per month as well as non-food consumption separated by groups of goods and services. All values are presented in nominal terms, real terms, and adjusted for inter-year price variation using the CPI. In general, non-food expenditures are larger in the HBS compared to the NPS. Among the sources of these differences is the fact that the first two waves of NPS did not include "clothing and footwear". NPS also does not include "restaurants and hotels". Moreover, it is worth noting that the value of "housing, water, electricity, gas and other fuels"

F. Non-food consumption
in the HBS 2011/12 is almost 3 times larger than the value in the NPS 2010/11 and that the "recreation and culture" expenses are over 25 times larger in the HBS; however, expenses on education and miscellaneous items are around 3 times larger in the NPS.  Both surveys show similar trends in consumption at the sub-aggregate levels, except for education and miscellaneous expenditures, but with the variations being much larger in the HBS. This can be partly explained by the changes in the survey design in the HBS. It is worth mentioning that when addressing the changes in the HBS design through imputation methods, we still observe a significant increase in the consumption of food and non-food items.

IV. Comparison of NPS and HBS Methods for Poverty Line and Poverty Indicators Measurement
This section examines the differences in the measurement of the poverty line and estimation of the poverty indicators between HBS and NPS. It also investigates the potential sources of differences in the inequality indicators. More specifically, the analysis re-evaluates the poverty line of NPS using the methodology of the HBS 2011/12, to separate differences in the poverty indicators resulting from the estimation approach (that can be addressed) and those stemming from the survey methods and that are more difficult to adjust. We also estimate the poverty numbers using the US$1.90 international poverty line to explore how the levels and trends of poverty compare between the two surveys. We finally examine the potential sources of discrepancy between the two surveys in the consumption distribution patterns for NPS data and decomposition methods.

A. Poverty line estimation
The poverty line in the NPS is not directly comparable with the poverty line in the HBS due to the differences in the consumption measurement methods listed in the above sections as well as differences in the reference period, reference population, measurement of the cost per calorie and adjustment for inter-year price variations. Most of these differences affect the food line and are explained more in detail below.

i. Food line
Both where qk is the total quantity of item k consumed in the reference population, p0k is the national median price of item k, qhk is the quantity of item k consumed by household h, and calk is the corresponding caloric conversion factor of each item established by the Tanzania National Bureau of Statistics. Median prices p0k are based on the most frequent unit of consumption for each item, with all units being converted to the most frequent unit when possible. If the household consumed the food item in a unit that does not have a metric conversion to the most frequent unit (e.g. piece to kg) the respective item is dropped. 11 The reference group in the NPS includes the bottom 50 percent of the population ranked in terms of real per adult equivalent consumption as opposed to nominal per adult equivalent consumption. Real consumption is obtained by adjusting the nominal consumption according to temporal and spatial cost-of-living differences. Temporal price differences are associated with seasonal differences (quarters), while spatial differences are associated with the location of a household (geographic stratum: Dar es Salaam, other urban, rural, Zanzibar where pk is the median price of item k in the reference population, is the average per adult equivalent consumption of item k, and • ∑ is the total caloric consumption per adult equivalent by household.  shows the food line and extreme poverty rates resulting from all the adjustments when they are applied simultaneously. Line (e) shows that using the HBS method to estimate the NPS food line leads to close values of both lines; however, the proportion of extreme poor population substantially increases to 11.8 percent in the NPS. When the food line is adjusted for inter-year differences in the cost of living using the CPI, extreme poverty initially increases and then stagnates in 2012/13.

iii. Basic needs poverty line
In both surveys, the non-food component of the basic needs poverty line is based on average nonfood consumption of households whose total consumption is close to the food poverty line.
In the HBS 2011/12, the households in the reference group are those whose total consumption lies within the interval between the food line and 1.     show a slight increase in poverty between the first two rounds. It is worth noting that the estimation of the international poverty rates follows the Povcalnet method which does not adjust the consumption values for spatial cost of living differences and which seems to partly resolve the mismatch in poverty trends between the two surveys.

iv. Comparison of the basic needs poverty lines across the survey rounds
Based on these findings, it seems that despite the differences between the HBS and NPS in survey design and methods of consumption data collection, the discrepancies in poverty levels and trends between both surveys are mainly resulting from the differences in: 1) the methods of calculation of food and basic needs poverty lines; 2) the inter-year price deflators; and 3) to a lesser extent, the spatial price deflators.

V. Comparison of Inequality and Distributional Patterns between HBS and NPS
This section compares the inequality indicators between the HBS and NPS surveys and performs the unconditional quantile decomposition to examine the specific household attributes that contribute to the changes of consumption over time in both the HBS and the NPS.

A. Inequality indicators
As stated in the first section, HBS and NPS show different inequality trends, declining in the HBS and increasing across the NPS waves. The adjustments above are relevant for the poverty estimates only and cannot help in addressing the inequality discrepancies.

B. Unconditional quantile decomposition
This section investigates the basic factors that might explain the discrepancy in inequality (and propoor growth patterns) between the HBS and NPS surveys by performing the unconditional quantile regression decomposition technique. The method decomposes the changes in consumption over time into two components: one component that is due to improvements in personal characteristics or endowments (better education, increased ownership of land and other assets, access to employment opportunities, local infrastructure, and so forth) and one component attributable to changes in the returns to those characteristics (returns to education, land productivity, returns to business, and so forth). These components are then further decomposed to identify the specific attributes that contribute to the changes of consumption. The decomposition is applied at each decile group of the consumption distribution to understand the patterns of the changes for the different welfare groups. 15 We start by examining the factors contributing to the variation of consumption between 2007 and 2011/12 using HBS data. The results are reported in Figure 10 and indicate that the increase of poor households' consumption is due mainly to improvements in households' endowments. Returns also improved but to a lesser extent and only for the 20 percent poorest groups.
One can observe from Figure 10 an improvement of households' endowments for all the population groups, but the improvements are more marked for the 30 percent poorest segments.
The increase of the endowments is driven by a significant expansion of asset ownership, mainly transportation and communication means, and to a lesser extent agricultural land. Educational attainment of household heads has improved as well but less significantly. The access to local infrastructure has deteriorated in general, but access to local roads seems to have slightly improved for the poor. The decomposition indicates also a decline of households' engagement in business activities, particularly among the poorest groups.
The improvements of households' endowments were coupled with an increase of the returns to those endowments, but only for the poorest quintile group. Except for the first two deciles, returns appear to have declined over time. But this decline masks divergent trends across the different attributes.
As observed from the table in Figure 10, the gains from household businesses, essentially nonfarm activity, increased quite significantly between 2007 and 2011/12 particularly for the three bottom deciles. Returns to land seem also to have improved over time, though less significantly for the poor.
The returns to community infrastructure also improved, indicating a higher positive influence of access to local markets and roads on needy households' living standards.
Large household size and number of children seem to be a continuing constraint on household wellbeing, although their negative impact appears to have diminished somewhat, as is apparent from the positive change in the returns to demographic structure.
However, the observed improvements in the returns to some household attributes have been offset by a significant decline of the returns to assets followed by a decline of returns to education, inducing a loss of returns for the moderate poor and better-off households.  We also apply the unconditional quantile regression decomposition technique to analyze the factors contributing to the variation of consumption between 2008/09 and 2012/13 using NPS data.
The results are presented in Figure 11. Similar to the HBS data, NPS reveals a quite significant improvement in household endowments over time; however, in contrast to the HBS results, the improvement of endowments is more marked for the richest population groups. We also observe a quite significant deterioration of returns, particularly for the poorest groups, that have offset the endowments' improvements, inducing a decline in total consumption. It is important to note that in this decomposition procedure, consumption is adjusted by the (inter-year) survey price deflators.
The use of the CPI for adjusting consumption would have shown a less sharp decline in returns and a slight increase in overall consumption over time; however, the variations across the different deciles would have remained unchanged.
As for HBS, the results in Figure 11 indicate that the increase of the endowments is driven by a significant expansion of asset ownership, mainly transportation and communication means.
Educational attainment of household heads has improved as well but to a lesser extent. Access to local markets seems to have improved but only for better-off households. The NPS findings also indicate a potential decline in households' engagement in business activities, but the results are not significant.

44
As for HBS, NPS data indicate a decline in returns to households' endowments, but contrary to HBS findings, the decline seems more marked for the poorest groups. Here again, this decline masks divergent trends across the different attributes. Returns to education and assets seem to have improved, while returns to access to markets appear to have declined. These results deserve further investigation and confirmation.

VI. Some Concluding Remarks
This study attempts to investigate the underlying causes of the mismatch in poverty and inequality levels and trends between the NPS and HBS surveys. The analysis has focused on the key candidates for the divergence between the two surveys. These include the differences in methods of consumption data collection, methodological differences in the construction of the consumption aggregates and estimation of the poverty lines, adjustments for temporal and spatial price variations, and the consistency of within-household spending and asset ownership trends with poverty trends.
The main findings can be summarized as follows: I. Despite noticeable differences in the methods of household consumption data collection, both HBS and NPS show close consumption levels when using comparable inter-year price deflators. The comparison of the levels and distribution of consumption between the two surveys, when adjusted by the CPI, shows that total consumption and food consumption are III. The discrepancy in poverty trends can be mainly attributed to the difference in temporal price deflators and, to a lesser extent, spatial price deflators. The use of the CPI for adjusting consumption variation over time would show a decline in poverty during the last five years for both HBS and NPS. However, the decline in poverty revealed by the HBS data would remain much higher than that observed with the NPS data. Given the greater degree of commodity detail in the consumption module that was added to the HBS 2011/12 questionnaire-which would suggest a better capture of consumption information-it is possible that the HBS under-estimated consumption in 2007 and hence overestimated poverty then and its subsequent decline. Also, the increase in poverty between NPS 2008/09 and NPS 2010/11 continues to be observed and could potentially be explained by the financial crisis and international food price variations.
IV. The mismatch in inequality trends between HBS and NPS could not be resolved. The analysis of the variation of consumption distribution over time using HBS and NPS data shows that both surveys indicate significant improvements in households' endowments over time.
However, while HBS reveals that endowments increased faster for the poorest groups, NPS shows that the richest groups experienced higher improvements in their endowments. Also, while HBS shows a slight increase in returns for the poorest groups, NPS reveals a deterioration. This might be partly driven by the inter-year deflator but would need further investigation and confirmation. All these results point to the importance of examining the sampling design.
Based on these findings, we would suggest the following recommendations: 1) Enhance closer collaboration inside NBS between the teams working on HBS data and those processing NPS surveys and harmonize the methodologies for the evaluation of the poverty lines and price indicators; 2) Attempt as much as possible to harmonize the HBS and NPS design, particularly the methods for household consumption data collection; 3) Examine further the sampling procedure and the potential differences resulting from sampling design. 4) Further explore the underlying causes of the divergence between both surveys in growth and distributional patterns.