Global Poverty Monitoring Technical Note 28 Brazil 2021 data update: Methodological adjustments to the World Bank’s poverty and inequality estimates Gabriel Lara Ibarra and Ricardo Campante Vale April 2023 Keywords: Brazil; poverty; inequality; COVID-19; emergency transfers; auxilio emergencial; non-sampling bias Development Data Group Development Research Group Poverty and Equity Global Practice Group GLOBAL POVERTY MONITORING TECHNICAL NOTE 28 Abstract There is evidence of significant under-coverage of the “Auxilio Emergencial” program in Brazil’s Continuous National Household Sample Survey (PNADC in Portuguese) 2021. The program, originally launched as an emergency measure by the Government of Brazil to support families during the pandemic, covered over 68 million individuals in 2020. The previous PNADC 2020 only observed about 20 million, and an approach to impute AE beneficiary status was applied to complement the observed AE status and better capture households’ welfare. A new version of the AE program was launched in 2021. AE in 2021 covered 36 million, while the PNADC 2021 only observed 8 million. An adjustment that incorporates the eligibility criteria of AE 2021 is implemented in the calculation of households’ income and poverty rates. The imputation method described here is included in the World Bank’s poverty and inequality estimates for Brazil 2021 (published in March 2023). The poverty estimates in 2021 are 28.4 percent at the US$6.85 poverty line and 5.8 percent at the US$2.15 line. The Gini coefficient is estimated at 0.526. A set of robustness checks shows qualitatively similar results. The authors are with the World Bank’s Brazil Poverty and Equity team. Corresponding author: Gabriel Lara Ibarra (glaraibarra@worldbank.org). Valuable inputs were provided by David Megill and Otavio Conceicao. The team received useful comments from Francisco Ferreira, and colleagues at the World Bank’s LAC Stats Team, and the World Bank’s Global Poverty Working Group. We are thankful to the colleagues at IPEA, at CLEAR-FGV, and INSPER for the thoughtful comments. The Global Poverty Monitoring Technical Note Series publishes short papers that document methodological aspects of the World Bank’s global poverty estimates. The papers carry the names of the authors and should be cited accordingly. The findings, interpretations, and conclusions expressed in this paper are entirely those of the authors. They do not necessarily represent the views of the International Bank for Reconstruction and Development/World Bank and its affiliated organizations, or those of the Executive Directors of the World Bank or the governments they represent. Global Poverty Monitoring Technical Notes are available at http://iresearch.worldbank.org/PovcalNet/. 1. Introduction An accurate measurement of poverty is essential for governments, civil society, the international development community, and other stakeholders to understand the needs of the most vulnerable and the ability of policies to improve the livelihoods of those most deprived. The onset of the COVID-19 pandemic brought, among many negative consequences, a new challenge to the calculation of poverty rates given the unprecedented surge of large-scale emergency cash transfer programs in several countries. In many cases, the novelty of these programs coupled with the relatively sizable values of the transfers created problems for national household surveys to capture their effects on households’ welfare accurately. A notable example is that of Brazil. The enactment of a massive government social assistance program called Auxílio Emergencial (AE) in 2020 and its continuation in 2021 profoundly changed the Brazilian social protection system landscape during the COVID-19 pandemic. AE was by far the largest emergency cash transfer program in Brazil in both years. According to administrative records, AE reached 68.2 million people in 2020 and 36.4 million in 2021, with benefits amounting to monthly BRL 600 per household in 2020 and BRL 250 in 2021. Compared to the long-standing flagship program Bolsa Família (PBF), created in 2003, AE was much larger and more generous in terms of the benefit values. PBF reached 14.4 million families in 2019, the last pre-pandemic year, and provided an average benefit of BRL 188. Despite the program’s size, the main Brazilian household survey did not capture AE well. In fact, the World Bank’s methodology to estimate the poverty rate for Brazil in 2020 incorporated an imputation method to properly account for the effects of the AE program and provide an improved measure of household’s livelihoods (Lara Ibarra and Vale, 2022). This note builds on this earlier work and presents the methodology used for the estimation of poverty rates for Brazil in 2021, when the same problem of under-coverage continued to exist. In motivating this approach, the note provides further evidence on the under-coverage of several cash transfer programs in Latin America and the Caribbean (LAC) and how this under-coverage tends to be larger for smaller programs (in terms of beneficiaries). Nonetheless, not all programs warrant an imputation methodology. Many of them are relatively small and any correction would require strong assumptions. The AE’s under-coverage at 61% in 2020 and 58% in 2021 appears as a clear outlier when compared to other programs for which we have data. Furthermore, this note argues that AE’s characteristics could warrant an adjustment. First, the AE budget was BRL 293 billion in 2020 and BRL 56 billion in 2021, equivalent to 3.85% of GDP in 2020 and 0.64% in 2021. Second, the number of people covered was 68.2 million in 2020 and 36.4 million in 2021 - the equivalent of 44% of the country's adult population in 2020 and 24% in 2021. Together, AE’s characteristics suggest that not incorporating a correction may lead to large errors of welfare- related indicators in Brazil. 1 The remainder of this note is organized as follows. Section 2 brings detailed information about the AE program in 2021, and Section 3 presents the details of the methodology that we used to impute the AE beneficiary status to those PNAD-C respondents that reported income consistent with the program’s transfers under the label of “Other social programs”, as well as to surveyed individuals whose characteristics indicate that they would be eligible for the program even if they did not report it. Section 4 explains how we calculate the amount of AE income that is imputed for each individual who is flagged as beneficiary. Finally, Section 5 concludes. 2. The Auxilio Emergencial program in 2021 The AE 2021 program was a revised version of the previous year’s program. It provided a monthly payment of BRL 250, except for single-person households who received BRL 150 and single mothers who received BRL 375. Beneficiaries received the transfers during the seven months from April to October.1 The benefit value in 2021 was lower than that of AE 2020, the number of recipients in each family was limited from a maximum of two to a maximum of one individual, and the conditionality on income was strengthened by the establishment of combined thresholds of total income and per capita income. The AE program was conceived as an emergency measure to mitigate the economic toll induced by the COVID-19 pandemic and consisted of monthly monetary transfers for vulnerable families. Even though other government aid programs with similar formats and goals were also at play between 2020 and 2021 in Brazil at the state and municipal level, they were much smaller in terms of the number of beneficiaries relatively to the Brazilian population. The fiscal cost of the AE program was also significant. It was equivalent to 3.85% of the Brazilian GDP in 2020, and 0.64% in 2021, while the fiscal cost of PBF has been typically smaller at about 0.45% between 2013- 2019, on average.2 Other well-known Brazilian transfer programs are the Benefício de Prestação Continuada (BPC) and the unemployment insurance. These had a fiscal cost equivalent to 0.76% and 0.47% of GDP in 2020, while their costs were 0.69% and 0.35% of GDP in 2021, respectively.3 In 2021, the typical AE benefit was BRL 250, while that of BPC amounted to one minimum wage (BRL 1,212). The number of beneficiaries, in contrast, was seven times higher in the AE program.4 1 In November, a rebranded version of the Brazilian cash transfer program Bolsa Familia was launched: Auxilio Brasil. In that month, the structure of Auxilio Brasil’s benefits was characterized by variable benefits that depended mostly on the number of children less than 4 years old and other minors (less than 21 years old) in the household. In December, a minimal benefit of BRL 400 was set. Regardless of family size, gender or age of the members, every family unit started to get at least BRL 400 if they did not reach more than that by the standard program’s rules. In practice, that meant an almost flat benefit design, as more than 95 percent of the families received the BRL 400 floor. Our imputation is restricted to the AE transfer and during the period it was implemented. 2 Bolsa Familia was also in place during a few months in both 2020 and 2021. The costs of the program were 0.42% and 0.31% of GDP, respectively. 3 BPC is a social pension in Brazil that benefits either poor elderly or persons with disabilities. 4 To calculate such ratio, we considered the number of AE (36.4 million) and BPC (5.06 million) beneficiaries in 2021. 2 Table 1. Comparison of different implementation periods of Auxilio Emergencial Auxílio Emergencial Auxílio Emergencial Auxílio Emergencial Residual 2021 April - August 2020 September - April - October 2021 December 2020 Transfers 5 x BRL 600 4 x BRL 300 7 x BRL 250 Per capita family Per capita family Per capita family income up to 1/2 MW income up to 1/2 MW income up to 1/2 MW Income criteria OR total family OR total family income AND total family income up to 3 MWs up to 3 MWs income up to 3 MWs 2 benefits (possibility 2 benefits (family had 1 benefit (reduction of of single mother to opt-in by two BRL 100 for single- double value + Family quota recipients or double person household and another family value for single extra BRL 125 for member standard mother) single mothers) value) Coverage 68.2 million 56.8 million 36.4 million Notes: Table based on Lara de Arruda et al., 2021. Sources: Portal da Transparência, Lei 13982/2020, Decreto 10316/2020, Edict 351/2020, Decreto 10412, MP 1000/2020, Decreto 10488/2020, MP1039/2021, Decreto 10661/2021, Edict 620/2021, Decreto 10740. MW stands for minimum wage. The questionnaires of the Brazilian household survey (Pesquisa Nacional por Amostra de Domicílios Contínua, PNAD-C) 2020 and 2021 were not adapted to ask respondents specifically about their AE beneficiary status. This poses an important challenge in the identification of AE beneficiaries using the household survey data. We argue that if no imputation is performed, the poverty rate would be overestimated in 2021 given the significance of the program, especially among vulnerable families. We document several facts that suggest a significant bias in the PNADC 2021 estimates of the program’s coverage. That is, the estimated coverage rate of the AE program based on the PNAD-C 2021 is much lower than one would expect according to administrative records.5 There is suggestive evidence of the following empirical regularity of household surveys: the under- coverage tends to be greater for smaller programs. Figure 1 shows this stylized fact using data from a subset of countries in LAC. AE in 2020 and in 2021 stand out. The estimated coverage rates are substantially below the rates obtained in other countries’ surveys – given similar coverage rates based on administrative data. In 2021, AE seems to be an outlier given its high number of beneficiaries as a share of the country’s population and its high under-coverage, when compared for example to that of the Uruguayan flagship program Asignaciones Familiares, which reaches approximately 20% of the country’s population. 5 The coverage rate is equal to the share of the population that is receiving the benefit. 3 Figure 1. Relationship between program under-coverage and number of beneficiaries as a share of the country’s population, circa 2015 10% PER URY BOL 0% ARG 0% 10% 20% 30% 40% 50% 60% -10% MEX PAN Under-coverage (%) ECU COL BRA DOM -20% SLV -30% PRY -40% -50% -60% AE 2021 AE 2020 -70% Number of beneficiaries as a share of the country's population (%) Source: CEPAL (2018) for all countries, except for AE 2020 and 2021 which are authors’ calculations. Notes: The values for the vertical axis indicate the average of country-specific program under-coverage in three distinct years: 2011, 2014 and 2015.The under-coverage (UC) for each program was calculated using the following formula: UC = ((Ns/Na)-1) * 100, where Ns is the number of beneficiaries of the corresponding program according to the national household survey and Na is the number of program beneficiaries according to administrative records. The trendline represents the linear regression fit of the UCs on the shares of beneficiaries. The list of countries and their respective programs is as follows: Argentina, Asignación Universal por Hij; Bolivia, Bono Juancito Pinto; Brazil, Bolsa Família; Colombia, Más Familias en Acción; Ecuador, Bono de Desarrollo Humano; Mexico, Prospera; Panama, Red de Oportunidades; Peru, Juntos; Paraguay, Teokoporá; Dominican Republic, Progreando con Solidaridad; El Salvador, Comunidades Solidarias, and Uruguay, Asignaciones Familiares. In the case of the Auxílio Emergencial (AE) program in 2020 and 2021, we used the variable V5003A of PNAD-C 2019 to determine the AE beneficiary status of respondents (“Did you receive money from government social programs other than BPC or Bolsa Família in the month of reference?”). See more details in Section 3. Although the Bolsa Família program is a family-level program, CEPAL (2018) calculated the number of PBF beneficiaries in 2011, 2014 and 2015 as the number of individuals in PBF beneficiary families so that every member of a beneficiary family was considered a PBF beneficiary. The data suggest that this under-coverage is not coming from sampling errors. In fact, the sampling errors of the PNAD-C estimates of the total number of households receiving AE benefits are small, given the relatively large sample size for this survey. This indicates that the undercount of AE beneficiaries in the survey data is mostly due to bias related to nonsampling error, such as the under-reporting of AE transfers by the respondents. It is thus important to consider an imputation methodology for reducing the impact of this nonsampling bias on the measurement of poverty. Given the large size of the AE program compared to other types of cash transfers, the under- reporting of AE in the PNAD-C has a large effect on the poverty estimation. Notably, there is suggestive evidence that other types of cash transfers that are explicitly asked in the PNAD-C questionnaire are less frequent and suffer from even greater under-reporting. For example, Table 2 show the estimated undercount of beneficiaries in the PNAD-C data by year for Bolsa Familia, 4 Benefício de Prestação Continuada (BPC), and Unemployment Insurance (UI) or Seguro-defeso, from comparing the survey estimates of the total number of beneficiaries to the corresponding administrative data. Nevertheless, the potential improvements from implementing a correction on such programs are much lower. Table 2. Estimated undercount of transfers that are collected in PNAD-C, by year A. Bolsa Familia beneficiary households Total households receiving % households in Bolsa Familia Brazil receiving PNADC Administrative % difference Bolsa Familia Year estimate count (undercount) (Administrative) 2017 9,795,568 13,469,672 -27.3% 19.7% 2018 10,070,997 13,954,172 -27.8% 20.0% 2019 10,104,609 13,783,108 -26.7% 19.4% B.Individual Benefício de Prestação Continuada (BPC) beneficiaries Total population receiving BPC % population in Brazil receiving PNADC Administrative % difference BPC Year estimate count (undercount) (Administrative) 2017 2,280,169 4,483,296 -49.1% 2.2% 2018 2,549,874 4,603,745 -44.6% 2.2% 2019 2,689,350 4,627,122 -41.9% 2.2% C. Individual Unemployment Insurance (UI) or Seguro-defeso (SD) beneficiaries Total population receiving UI % population in & SD Brazil receiving PNADC Administrative % difference UI & SD Year estimate count (undercount) (Administrative) 2017 1,305,026 2,453,236 -46.8% 1.2% 2018 1,227,371 2,299,830 -46.6% 1.1% 2019 1,205,732 2,283,779 -47.2% 1.1% Source: own calculations using PNAD-C data. Notes: The sampling errors and 95% confidence intervals for the PNAD-C estimates of the total number of beneficiaries for these types of cash transfer are available upon request. One possible source of the under-reporting bias may be the common use of proxy respondents (Hypólito and Silva, 2021) in the household that are interviewed for the PNAD-C to provide information about the other household members who are the actual recipients of different cash 5 transfers. Some proxy respondents may not know about the transfers received by other household members, or they may be reluctant to provide this information. In the case of the AE transfers, there can only be one beneficiary per household, and this person may not be the survey respondent. In addition, the PNAD-C questionnaire did not contain a specific question about the Auxilio Emergencial program, likely exacerbating the under-reporting among both respondents who were receiving the program and proxy respondents who were not receiving the program themselves. It can be seen in the tables above that among programs explicitly collected in the questionnaire, the survey’s undercoverage tends to correlate negatively with the size of the program (i.e. the beneficiaries share of the population). In the case of Bolsa Familia, the transfers cover approximately one-fifth of households, and the estimated average undercount of beneficiary households in the PNAD-C data is about 27%. However, the BPC and UI/SD transfers only cover about 2.2% and 1.1% of the population, respectively, and the corresponding average undercount of these beneficiaries in the PNAD-C data is 45.2% and 46.9%, respectively. Given that the PNAD-C interviews rely on proxy respondents, it is less likely that the recipients of a less frequent type of cash transfer will be available for the direct interview. At the same time, the undercount of beneficiaries of less frequent types of cash transfers will have a smaller effect on the measurement of poverty compared to the undercount for the more predominant AE transfers. 3. Methodology for imputing Auxilio Emergencial beneficiary status Based on the eligibility rules of the program described above (see Table 1), we build upon the methodology developed in Lara Ibarra and Vale (2022) and impute AE beneficiary status as follows: 1. All the individuals that report receiving income from “Other social programs” that are consistent with AE 2021 benefits (R$150, R$250, and R$375) are considered as potential AE beneficiaries.6 We use the methodology developed by Hecksher (2020) to identify the month of interview in the public data of PNADC and consider all the potential beneficiaries whose household was interviewed between May and November as recipient of the AE cash transfers.7 The reason to use May to November and not April to October is because PNADC asks about the income of the last thirty days.8 6 The PNADC 2021 questionnaire explicitly asks whether the individual received the Continued Benefit program (Benefício de Prestação Continuada or BPC in Portuguese), Bolsa Familia, Unemployment Insurance, Pensions, and Scholarships. Thus, “other social programs” reflect government transfers that are not otherwise captured, such as AE. 7 The identification of the interview month is based on the age of the individuals, their birthdate, and the fact that all households in the same Primary Sampling Unit (PSU) are interviewed during the same month. It is possible to compare individuals’ age and birthdate to check if that date had passed or not on the interview day. From taking all the sampled individuals within the same PSU for analysis, one can restrict the possible dates and detect the month of interview for 95% of the individuals in the annual 2021 data. 8 We ran a robustness check to verify the impacts of this choice. Results can be found in the appendix. 6 2. All individuals who report receiving income from the Bolsa Familia Program (“Programa Bolsa Família”, PBF) when the emergency program was in place are assumed to be potential AE recipients. For those reporting a PBF benefit lower than the potential AE benefit, we impute the status of AE recipient. For individuals reporting a higher benefit, we maintain their PBF beneficiary status. The reason is that the system to assign program benefits had an automatic trigger to “move” families into PBF or AE based on whichever program would make the family better off (i.e. receive a higher monetary benefit). 3. Out of the remaining persons, those who met all the program’s criteria were considered potential recipients of AE benefits: - Individuals aged between 18 and 64. That is people legally underage and those that would have the age to qualify for BPC benefits were not considered. Regarding the last, they should be automatically directed to BPC, at least in theory. Among our robustness checks, we relax that hypothesis (Appendix). - Individuals who were not in formal employment, or who were self-employed but satisfied the income criteria. Formal employment is defined by those under the CLT (Consolidated Labor Laws) regime, civil servants, military personnel, employers and self-employed individuals who pay social contributions.9 - Individuals who were not receiving other government payments. Therefore, individuals who were getting transfers from BPC, Unemployment Insurance, or Pensions are regarded as ineligible. - To account for AE’s income eligibility criteria, we adopted the following thresholds: i. To determine a household's income, we calculated the total income from all sources, both monetary and non-monetary, for all members of the household. This included both formal and informal labor income. Although only formal labor income can be verified, we found evidence that most people would be reporting their full income. ii. When identifying the members of a household, we do not consider non-relatives who do not share costs of living, lodgers, and domestic employees and their relatives. This group makes up a small portion of the population (0.18%), so it does not have a significant impact on the analysis. iii. An individual passes the income criteria if her total income (from part i) was under R$3300 and her household income per capita (from ii) was under R$550.10 9 In theory, people self-reporting themselves as formal employers could also meet the eligibility criteria. For example, people under the formal microentrepreneurs regime (MEI) employing one person. We decided to not include the employers as a conservative approach to prevent the overestimation of AE beneficiaries. If we allow for employers to be eligible, the number of eligible individuals would only increase by about 43,000. A sensitivity analysis relaxing this constraint is run and results are shown in the Appendix. After making all the adjustments to determine effective beneficiaries in that specification (i.e. for instance, capping family beneficiaries at one), the number of recipients turns out to not be statistically different from the one excluding formal employers. 10 Estimates from administrative data of families registered in the social registry (CadUnico) and the social programs payroll suggest that the probability of a family to be an AE beneficiary (or keep its BF benefits should they be more advantageous) among families who complied with all the AE income conditionalities was 71.3 percent. While there is anecdotal evidence of AE non-compliance (i.e. people that didn’t meet both income criteria and still received AE), 7 iv. An individual was not considered a potential beneficiary if her taxable income was higher than R$2379 per month. This restriction was put in place to align with the 2018 pre-pandemic taxable income requirement, which states that individuals earning more than R$28,559.70 in annual taxable income (equivalent to R$2379 per month) during the previous year were not eligible. However, as we do not have information on individuals’ taxable income of 2018, this threshold of R$2379 per month was used as a proxy. Taxable income aggregate is defined from the sum of formal labor market income plus non-labor income (pensions, government transfers, financial income, alimony). - A maximum of 1 individual per household is imputed beneficiary status. - Households that were surveyed in a time that was not between the May to November period were not considered as AE eligible, since the program was not operating. Considering all the eligibility rules above, we account for 15.4 million AE beneficiaries in the PNADC 2021. This combines those self-reporting, as well as those estimated to be eligible (Table 3). This number implies an under-coverage of 26% when compared to administrative data. Table 3. Frequency of AE eligible individuals by quarter Eligible Ineligible Beneficiary Keep BF Q Total Reported 18+ 18+; AE Reported Current OSP Current Minors both BF ineligible imputed AE > AE BF > AE and AE 1 14,261,870 41,939,946 - - - - - 56,201,816 2 13,375,672 35,360,354 2,289,744 2,342,732 85,185 222,817 287,556 53,964,060 3 12,793,358 32,266,979 2,625,105 3,432,295 99,648 214,578 401,488 51,833,451 4 12,601,945 33,291,366 1,836,266 2,058,387 60,655 145,239 278,720 50,272,578 Total 53,032,845 142,858,645 6,751,115 7,833,414 245,488 582,634 967,764 212,271,905 Notes: Domestic workers residing in the household, lodgers and housemates not sharing expenses are excluded from the analysis. Frequencies are estimated using sampling weights provided in the survey dataset. PBF recipients whose benefit was higher than the alternative AE benefit are considered to remain in the PBF program even if they would qualify for AE. Individuals who self- reported “Other Social Programs” (OSP) income above AE in quarters 2 to 4 are considered to be beneficiaries who accumulate other transfers (i.e. social programs from subnational governments). Q refers to the 2021 quarter. Source: own calculations using PNADC 2021 data. estimates from administrative data suggest this inclusion error is close to 12.5 percent. Thus, relaxing the income conditionalities by imputing AE beneficiary status if an individual meets at least one of the income criteria we run the risk of incorporating a much larger share of ‘false positives’ than by applying both income criteria. In one sensitivity analysis (see Appendix) we find that by relaxing these criteria we would predict an “over-coverage” of about 7 percent of the AE program. Thus our preferred approach is to be conservative. 8 Table 4: Consolidated AE direct beneficiaries in PNAD-C 2021 Beneficiary Freq. Percent Cum. No 196,859,254 92.74 92.74 Yes 15,412,651 7.26 100.00 Total 212,271,905 100.00 Notes: Domestic workers residing in the household, lodgers and housemates not sharing expenses are excluded from the analysis. Our methodology relies on two main assumptions11: Assumption 1. Households interviewed out of the period of program duration will not be imputed as AE beneficiaries. The PNADC annual data is made up of a combination of the interviews collected during the quarterly surveys. PNADC sampling works in a (1(2)5) rotation scheme, in which a household is included in the sample for a month, is excluded from the sample for the two subsequent months, and then is incorporated again, up to a maximum of five interviews. During each quarter, some households are participating in their first interview, others in their second, and so forth.12 This means that any statistic estimated from the annual survey is actually using the cross-sectional data collected from surveyed households, which are combined after the whole year has passed. As a corollary of this assumption, we estimate the under-coverage rate of the social programs in the survey correcting the actual total number of beneficiaries by a factor that represents the total population represented in the survey during the period that the program was in operation. For instance, in 2021, a sample representing only 57 percent of the Brazilian population was interviewed between May and November. Thus, from the 36.4 million individuals that actually received the benefit, we estimate that just 20.8 million (57 percent) could be found in the survey data. No imputations are considered for individuals interviewed from January to April nor in December, although they could be reporting AE benefits already received after the program onset (individuals interviewed in April) or late payments (individuals interviewed in December). As mentioned before, we relax this condition in one of our robustness checks. Assumption 2. The information collected from the household at the time of the interview is assumed to reflect the permanent welfare situation for that year. This means that for any household receiving AE income, it is assumed that it is a monthly payment for the entire year. 11 These assumptions follow closely the approach in Lara Ibarra and Vale (2022). 12 Poverty rates in Brazil have typically been based on the annual release of the PNAD-C, which is based on information from households’ first interviews. This changed with the COVID-19 pandemic. Due to the lower response rates for first interviews during 2020 and 2021, the Brazilian National Statistics Office (IBGE 2022b) decided to publish its annual release based on the fifth interview (see Appendix A). 9 Together, assumptions 1 and 2 imply that households interviewed out of the AE program period (2021Q1), and that eventually received AE’s benefits when in place, are assumed to not receive any benefit. This assumption follows the standard approach for estimating statistics from annual household surveys, in which the data gathered at the time of the interview and, particularly incomes, are extrapolated for the rest of the year. This approach is important in understanding the measurement of poverty: the poverty rate will represent an “average” situation in the country for one year. 4. Assignment of Auxilio Emergencial benefits for annual income calculation To determine the value of the AE benefit, individuals are separated into two groups. The first group consists of individuals who reported AE, either with PBF, or on its own.13 For this group, we accept the reported value as true and make no further adjustments. It is worth noting that individuals who reported an income higher than the maximum AE benefit of R$375 in OSP are also classified as AE beneficiary and included in this first group in our approach. The second group of individuals were assigned beneficiary status either by meeting the criteria for the program or by reporting that they received a transfer from PBF that was lower than the potential AE benefit. For individuals in this second group, we impute the benefit of R$250. In case they are characterized as single mothers or residents of single-person households, the benefits are respectively adjusted to R$375 and R$150. Next, the income variables at the individual and household levels are adjusted accordingly. The variable recording the amounts reported as “Other social programs” income in PNADC and its derivates are adjusted to reflect the imputed benefit.14 The variable recording income received from Conditional Cash Transfers Program and its derivates are also adjusted to consider the situation in which individuals were moved from these programs to the AE program.15 In the cases that the AE benefit was considered as self-reported, we do not make any adjustment to it. In the end, all the constructed income variables that are used to estimate the per capita household income (in 2017 PPP terms) are adjusted according to SEDLAC’s standards. These values are deflated using quarterly inflation estimates. However, as in PNADC 2020, the PNADC 2021 data does not include information about the characteristics of the dwelling, making it impossible to 13 Here and below we use “reported AE” as referring to an individual who “reported AE-consistent amounts” in the survey and during the known AE implementation period. Amounts reported were considered AE consistent if the value was equal to the program’s statutory benefits. 14 We only revise variables that are constructed for the harmonized version of PNADC as part of the SEDLAC project (i.e. itrane_ns = “Income from unspecified public transfers”). 15 For Brazil, SEDLAC’s “Income from conditional cash transfer programs ” only includes the amount reported as PBF benefit, which is deflated and adjusted to rural purchasing power according to the general harmonizing procedures. 10 estimate the rent for owner-occupiers.16 Instead, we follow previous methods, which use the PNADC 2019 data to obtain an estimate of “imputed rent” throughout the income distribution in PNADC 2021.17 Additionally, the incomes of people living in rural areas are increased by 15% to account for the higher cost of living in these areas. Our imputation of AE benefits leads to nonnegligible changes in poverty and inequality estimates (Table 5). The poverty rate at the upper middle-income poverty line (US$6.85 2017 PPP) decreases by 1.52 percentage points, while the effect is negative 0.88 pp at the international poverty line. Relative to the exercise conducted for the 2020 survey, the effect of the imputation on poverty rates is small. The reason is that the benefit in 2020 was more than a double the 2021 benefit and covered up to 88 percent more individuals. The numbers highlight that, even after a correction for the AE coverage, the poverty rate surged between 2020 and 2021 due to a slow-paced labor market recovery and high inflation. Table 5. The effects of the imputation on poverty and inequality SEDLAC 2021 data + SEDLAC 2021 imputation Poverty rates 6.85 (PPP17) 29.88% 28.36% 3.65 (PPP17) 12.60% 11.26% 2.15 (PPP17) 6.70% 5.82% GINI 0.535 0.529 Average daily household income per capita, by decile 1 1.73 1.95 2 4.06 4.37 3 5.97 6.27 4 7.96 8.19 5 10.23 10.33 6 12.84 12.86 7 16.25 16.32 8 21.17 21.32 9 30.41 30.53 10 79.30 79.66 Overall mean 18.99 19.18 B40 mean 4.93 5.15 Notes: The columns labelled “PNADC 2021 data” refer to the welfare aggregate obtained by applying the SEDLAC harmonization methodology to the 2021 data without imputing AE beneficiary status. B40 refers to the bottom two quintiles of the per capita income distribution. 16 The response rates of the households’ first interviews of PNADC 2021 were considered too low by IBGE (2021a) sampling specialists and thus data from the first interview was not published. The first interview collected information on dwelling characteristics. 17 Different methodological choices in the extrapolation of 2019 data and imputation of rents may affect households’ estimated income and the estimated poverty rate. In the Appendix, we allow for some alternatives for imputing rents. Our sensitivity analyses suggest that such alternatives lead to qualitatively similar estimates of poverty rates and Gini indexes. 11 5. Concluding remarks The World Bank’s poverty and inequality estimates for Brazil are based on a revised version of the PNADC 2021 data. The decision to apply such a revision follows a similar reasoning as in the calculation of the 2020 Brazil poverty rates: a cash transfer program that provided significant benefits to a large share of the Brazilian population was not directly asked for in the questionnaire and thus not properly captured in the household income survey that is used to measure poverty. The correction in 2021 imputes the Auxilio Emergencial income for individuals who are highly likely to have benefited from the program during the period they were surveyed. Several pieces of evidence support the decision to impute. Notably, the under-coverage of the program was much larger than expected for a program of such a size, considering either programs captured by the PNADC or similar programs in the LAC region. In addition, the program is larger in terms of number of beneficiaries and transfer amounts when compared to other non-emergency cash transfers in the country. The imputation leads to a correction of the under-coverage of the program. Taking the household survey data at face value, we could observe approximately 8.6 million AE beneficiary individuals, while estimates suggests that about 20.8 million individuals (out of 36.4 million in administrative records) should have been ‘observable’ by the survey during the period that the PNAD-C was in the field. After the imputation, we find a total of 15.4 million beneficiaries: a decrease of the under- coverage rate from 58 percent to 26 percent. The imputation methodology followed closely the de jure rules of the program. The poverty rate estimated at the upper middle-income country line (US$6.85/day in PPP2017) is 28.4 percent, 1.5 percentage points below the estimate in the raw data. Whenever the application of the law was not straightforward given the information we can observe in the survey, we tested alternative specifications as a sensitivity analysis. Poverty and inequality estimates are robust to these different methodological choices. The work conducted for the Brazil 2021 data is part of the World Bank’s effort to monitor poverty and inequality and provide the best possible estimates. As mentioned before, this effort is not the first one for Brazil. Adjustments were done in 2020, and an earlier example includes the modeling of imputed rent for the 2012-2015 PNADC (see Atamanov et al., 2020). These efforts are also typically discussed with the national statistical offices, though sometimes they do not coincide with the official estimates. 12 6. References Atamanov,Aziz; Castaneda Aguilar,Raul Andres; Corral Rodas,Paul Andres; Dewina,Reno; Diaz-Bonilla,Carolina; Jolliffe,Dean Mitchell; Lakner,Christoph; Lee,Kihoon; Montes,Jose; Moreno Herrera,Laura Liliana; Mungai,Rose; Newhouse,David Locke; Nguyen,Minh Cong; Prydz,Espen Beer; Sangraula,Prem; Yang,Judy. 2019. March 2019 PovcalNet Update : What's New (English). Global Poverty Monitoring Technical Note, no. 7 Washington, D.C. : World Bank Group. CEPAL. 2018. Cuál es el alcance de las transferencias no contributivas en América Latina? Discrepancias entre encuestas y registros. Serie Estudios Estadísticos N. 96. Hecksher, Marcos. 2020. “Valor impreciso por mês exato: microdados e indicadores mensais baseados na Pnad Contínua”. Nota Técnica IPEA, no. 62. Diretoria de Estudos e Políticas Sociais, Instituto de Pesquisa Econômica Aplicada (IPEA). Hypólito, Elizabeth; Silva, Denise Britz do Nascimento. 2021. “Análise do uso do informante proxy na PNAD contínua”. Nota Técnica do Boletim de Mercado de Trabalho IPEA, no. 72. Instituto de Pesquisa Econômica Aplicada (IPEA). IBGE. 2021a. Nota técnica 02/2021. PNAD Contínua. Sobre o processo de ponderação da PNAD Contínua. IBGE. 2021b. Nota técnica 03/2021. PNAD Contínua. Sobre a divulgação da Reponderação da PNAD Contínua em 2021. IBGE. 2022a. Nota técnica 03/2022. Sobre os módulos anuais de Características dos domicílios e de Características adicionais do mercado de trabalho em 2020 e 2021 IBGE. 2022b. Nota técnica 04/2022. PNAD Contínua. Sobre as Características gerais dos moradores em 2020 e 2021 Lara de Arruda, Pedro; Lyrio de Oliveira, Gabriel; Lazarotto de Andrade, Marina; Falcao Silva, Tiago; Teixeira Barbosa, Diana; Morgandi, Matteo. 2022. Coverage profile of Brazil’s Auxílio Emergencial and special design features for protecting women and other vulnerable groups. World Bank Latin American and Caribbean Studies. Washington, DC: World Bank. Lara Ibarra, Gabriel; Vale, Ricardo Campante. 2022. Brazil 2020 Data Update : Methodological Adjustments to the World Bank’s Poverty and Inequality Estimates. Global Poverty Monitoring Technical Note;21. World Bank, Washington, DC. © World Bank. https://openknowledge.worldbank.org/handle/10986/37513 License: CC BY 3.0 IGO. 13 Moore, R. C., & Hancock, J. T. 2020. Older adults, social technologies, and the coronavirus pandemic: Challenges, strengths, and strategies for support. Social Media+ Society, 6(3), 2056305120948162. 14 Appendix This appendix addresses concerns about the measurement of poverty and inequality indicators using household survey data collected during the pandemic, as well as a discussion of some of the assumptions used in the imputation method applied in the estimation of the 2021 poverty rates. Appendix A discusses the PNADC 2021 data. Appendix B presents sensitivity analyses using variations of the imputation approach described in this note. Appendix C presents the effects of the imputation over the spatial distribution of beneficiaries. Appendix D presents the methodology to identify the implicit month of the interviews of the PNAD-C. A. A brief review of the PNADC 2021 data Similarly to 2020, IBGE faced some challenges to implement the PNADC 2021 survey due to health restrictions imposed during the COVID-19 pandemic. The expectation that Auxilio Emergencial program would be ended in December 2020 was mistaken. Social distancing measures had to be reintroduced to stem the next wave of the disease, further disrupting the economic opportunities of many Brazilian households and thus increasing the demand for government support. After a three-month hiatus, AE was resumed in April 2021. Likely because of this uncertainty, an explicit question about AE was not included in PNADC’s questionnaire. Because of the problems of data collection, IBGE decided to keep the same strategy as in the former year, doing phone surveys instead of in-person field work until the end of 2021 second quarter (IBGE 2021a, IBGE 2022b). The enumeration via phones and the unusual pandemic conditions made it hard for surveyors to reach some households, especially those that had never been interviewed before and had not provided a phone number. The relatively low response rates for first interviews in 2021 mimicked the situation of 2020, and led IBGE to replace, once more, the publication of first interviews as the official microdata for the estimation of annual statistics and instead disclosing only the data from the fifth interviews. Concerns due to non-response The conditions under which the PNADC data collection occurred were very similar to 2020, when IBGE (2021b) analyzed and mitigated the concerns. Lara Ibarra and Campante Vale (2022) describe the sensitivity analyses performed by IBGE. In short, the potential bias induced by lower response rates was tested by simulating the response rates of 2020 and 2021 using the 2019 data. The conclusion was that there were no bias over the main indicators at any major geographic area. It is also worth mentioning that alternative models to correct the non-response bias were tested against the usual approaches and showed no improvements. Differences in telephone ownership were also examined in detail. One corrective measure taken by IBGE was the anticipation of the new gender-age group calibration method to produce sampling weights. Originally planned to be done after the Population Census, but the Census postponement to 2022 led IBGE to publish a revision of PNAD- 15 C sampling weights on November 30, 2021 (IBGE, 2021c). The new weights using the new calibration method were produced for the 2020 data and included a revision of all the data going back to 2012. Sampling weights for the 2021 data are calibrated using the same approach. IBGE stated that one of the most important advantages of the new method is to more accurately correct the response bias due to elderly people having a relatively higher response rate. This issue was particularly relevant in the case that interviews were conducted through the phone, increasing the potential bias. IBGE (2022b) also released a technical note in which it clarifies the public disclosure of the 2020 and 2021 microdata, making public the success rates of interviews.18 As mentioned above, PNADC follows a rotating sampling scheme in which each selected household is interviewed up to five times in different quarters. With this sampling design, one can either use the cross-sectional population in a given quarter or aggregate the interviews of the same order conducted in a given year (i.e. first interviews). Usually, the first interviews had the highest success rates of the survey. But this was no longer the case during the pandemic. Since the PNADC was launched in 2012 (and until 2019), the first interview’s microdata was released as the dataset representative of a given year. But due to the low response rates of the pandemic years, IBGE decided to disclose only the fifth interviews for 2020 and 2021. Table A1. Success rates of PNADC interviews per order of interview and year Interview/Year 2016 2017 2018 2019 2020 2021 1st 89.8 90.0 90.1 89.3 47.4 60.4 2nd 89.4 89.8 89.7 89.2 60.3 63.0 3th 88.9 89.4 88.9 88.7 66.6 65.4 4th 88.4 88.9 88.5 88.2 71.2 66.6 5th 88.0 88.2 87.9 87.7 72.7 69.9 An important caveat of the fifth interviews is that it does contain neither the housing module nor the complementary labor characteristics module (IBGE 2022a), precluding us from running analyses typically performed as part of the poverty rate calculations. For welfare indicators, the housing module is especially important for estimating non-monetary poverty. Regarding the measurement of monetary poverty, it is also essential for the rent imputation for owner-occupiers. 18 More than the simple response to an interview, IBGE considers as a successful interview those that did not skip main questions and that do not present notably inconsistent answers throughout the questionnaire. 16 B. Sensitivity Analysis of AE imputation in the Brazil 2021 data We test for the sensitivity of the poverty and inequality estimates to the number of individuals that receive the AE transfers and hence to our imputation methodology. As it is not possible to match administrative records to survey data, our imputation has to rely on some assumptions. The assumptions we take in our preferred approach are guided by two principles: i) follow the legal determinants of eligibility; and ii) take a conservative approach on the number of beneficiaries that will get AE benefits imputed. To assess the impact of the assumptions in our preferred methodology, we conducted robustness checks in which we change the choices made in the main approach. The robustness checks can be divided into three groups. In the first, which comprises a single alternative exercise, we relax the restriction on the timing of the interviews, allowing early recipients in the month of April and late recipients interviewed in December to be included as beneficiaries. In the second group, we loosen conditions related to the enforcement of the program eligibility rules. Sequentially and then simultaneously, we a) admit that people could hide informal income from the program’s means-testing, b) that employers could be receiving the transfers, and c) that elderly people could join AE regardless of BPC eligibility.19 Then we relax the joint conditionality of the household and per capita income criteria, identifying as AE beneficiaries individuals who satisfy at least one criterion. The last check of this group assumes that no conditionality could be enforced and that anyone could be applying for the AE. We bootstrap over a hundred random samples of imputed beneficiaries preserving the under-coverage rate of our preferred approach, to estimate what would be the poverty and inequality statistics distribution in that case. The last couple of checks relate to the harmonization methodology performed under the SEDLAC project, in particular the rent imputation. To estimate the rent, SEDLAC uses a hedonic regression model, which is mainly based upon the PNADC housing module. While the typical imputation approach already incorporates certain assumptions in the model specification, additional assumptions were necessary in 2020 and 2021 to overcome the unavailability of the housing module. The imputation procedure required extrapolating imputed rent values based on the PNADC 2019 data into the survey year (2021). Thus we run two robustness checks that test: i) what would happen in case the rent imputation model was run before the AE imputation – which 19 The BPC program aims to cover poor elderly people and overlaps with the AE eligibility rules, because it establishes that people aged 65 or more and fulfilling the income criteria qualify for a one minimum wage benefit monthly. AE’s legal framework stipulates that people receiving BPC are not allowed to receive the emergency transfer, but has no explicit rules regarding people that would qualify for the BPC at the moment of their application to AE. Social workers at the Centers of Reference to Social Assistance (CRAS) should refer elderly individuals to BPC, and according to program’s rules they would be automatically accepted. However, in practice this may not have been the case. It is important to note that individuals were able to apply for AE benefits using a simple application on their mobile phones (Lara de Arruda et al., 2021). While this age group typically shows low digital literacy rates (Moore & Hancock, 2020), we cannot rule out completely that all elderly were referred to BPC. We test the implications of this assumption in one of our robustness checks. 17 is implicitly assuming that the AE income does not affect the probability of a family to live in their own house; ii) what would happen if we altered the specification of the regression model predicting which households will be imputed rent in 2021. A description of each sensitivity test and the respective effects on households’ welfare can be found in the Tables A.2 and A.3. Table A.2: Description of the alternative conditions for sensitivity analysis CHECKS DESCRIPTION S1 Considers people interviewed in April and December as eligible. S2 Includes the possibility that eligible people report only formal income. S3 Includes employers as potential beneficiaries. S4 Allows people aged 65+ to receive AE. S5 Includes the possibility of employers to receive, the misreporting of income and allow 65+ to enter the program. S6 Relax the income conditionality: total family income is below 3 minimum wages OR 1/2 minimum wage per capita S7 Disregards all the conditionalities and randomly draws individuals that are not already in AE or BF to be beneficiaries, while matching the under-coverage that we find in our preferred approach. S8 Runs the rent model before the AE imputation. S9 Changes the rent imputation model: Includes a dummy for families receiving rent from other real estate properties and a dummy of capital of state to predict the families who own their house. Source: own compilation 18 Table A.3: Results of the sensitivity analysis SEDLAC SEDLA 2021 + S1 S2 S3 S4 S5 S6 S7 S8 S9 C 2021 imputation Beneficiaries 8,661,536 15,412,651 17,877,969 19,514,070 15,399,602 15,550,468 19,721,114 22,242,855 15,415,168 15,395,964 15,395,964 Under- -58% -26% -14% -6% -26% -25% -5% 7% -26% -26% -26% coverage Poverty rate 29.9% 28.4% 27.8% 28.3% 28.4% 28.3% 28.1% 28.1% 29.3% 28.6% 28.3% at $6.85 SE 0.3% 0.3% 0.3% 0.3% 0.3% 0.3% 0.3% 0.0% 0.3% 0.3% Poverty rate 12.6% 11.3% 10.9% 11.3% 11.3% 11.2% 11.2% 11.3% 12.2% 11.6% 11.3% at $3.65 SE 0.2% 0.2% 0.2% 0.2% 0.2% 0.2% 0.2% 0.0% 0.2% 0.2% Poverty rate 6.7% 5.8% 5.4% 5.8% 5.8% 5.8% 5.8% 5.8% 6.3% 5.9% 5.8% at $2.15 SE 0.1% 0.1% 0.1% 0.1% 0.1% 0.1% 0.1% 0.0% 0.1% 0.1% Gini 0.535 0.529 0.527 0.528 0.529 0.529 0.528 0.527 0.533 0.530 0.529 B40 average 4.93 5.15 5.24 5.19 5.16 5.17 5.19 5.23 5.04 5.15 5.19 income Poverty Gap 12.8% 11.7% 11.2% 11.7% 11.7% 11.6% 11.6% 11.6% 12.4% 11.8% 11.7% Ratio ($6.85) Source: own calculations using PNADC 2021 data. Notes: S1, S2, S3, S4, S5, S6, S7, S8 and S9 refer to results from sensitivity analyses by changing the original assumptions of column [2] approach. Standard errors of poverty rates shown in parenthesis. All poverty lines are expressed in USD 2017 PPP per person per day. 19 The poverty rate at $6.85/day is the one that varies the most across the sensibility exercises. It ranges from 27.8 percent to 29.3 percent, compared with 28.4 percent in our preferred approach. The interval in the Gini coefficient between the estimated minimum and maximum is no more than 0.006. The results reinforce the robustness of our approach and also that the poverty rates are the highest since the launch of the PNAD-C in 2012. C. Concerns related to the spatial distribution of beneficiaries One potential concern of applying an imputation method is about its effect at the subnational level. As PNADC is representative for analysis at the state and metropolitan areas level, the spatial distribution of beneficiaries could be affected, distorting regional analysis. Table A.4 shows the frequency of AE beneficiaries by states and by whether they were identified based on their responses in PNADC, or if AE beneficiary status was imputed. The distribution of the two subpopulations appears to be quite similar. 20 Table A.4 Distribution of AE beneficiaries by AE beneficiary status and by state AE beneficiaries Imputed AE State Freq. Percent Freq. Percent Rondônia 57,217 0.7% 55,208 0.8% Acre 33,018 0.4% 35,464 0.5% Amazonas 158,512 1.8% 181,100 2.7% Roraima 22,642 0.3% 29,705 0.4% Pará 520,453 6.0% 348,005 5.2% Amapá 50,901 0.6% 35,659 0.5% Tocantins 62,042 0.7% 76,789 1.1% Maranhão 370,169 4.3% 371,380 5.5% Piauí 200,979 2.3% 104,085 1.5% Ceará 600,804 6.9% 374,193 5.5% Rio Grande do Norte 232,494 2.7% 140,712 2.1% Paraíba 283,441 3.3% 132,073 2.0% Pernambuco 659,555 7.6% 516,372 7.6% Alagoas 187,574 2.2% 148,267 2.2% Sergipe 142,998 1.7% 123,879 1.8% Bahia 1,000,392 11.5% 660,639 9.8% Minas Gerais 829,498 9.6% 565,757 8.4% Espírito Santo 164,378 1.9% 116,620 1.7% Rio de Janeiro 522,955 6.0% 586,654 8.7% São Paulo 1,272,699 14.7% 1,066,366 15.8% Paraná 302,913 3.5% 263,204 3.9% Santa Catarina 113,331 1.3% 129,608 1.9% Rio Grande do Sul 299,985 3.5% 230,475 3.4% Mato Grosso do Sul 88,237 1.0% 68,059 1.0% Mato Grosso 110,756 1.3% 93,968 1.4% Goiás 293,901 3.4% 229,694 3.4% Distrito Federal 79,692 0.9% 67,180 1.0% Source: Own estimations using PNADC 2021 data. Notes: Frequencies are based on extrapolated numbers using the survey sampling weights. Percentages refer to the share of the population that resides in a given state. 21 D. The identification of months in the public version of PNADC and the reported benefits An essential feature of the 2021 imputation of AE is the identification of the months in which the program was in place. This is very important as the AE 2021 was halted in the middle of a quarter (end of October) and the only explicit data on the timing of the interviews in the PNADC public data is the quarter when it occurred. Two concerns arise from the month identification. First, are they appropriately identified? And second, are social programs really represented in the data during the time they were officially operative? First, it is important to stress that the method is very straightforward and was tested extensively (Hecksher, 2020). Complementing the robustness of the method, the empirical analysis highlights how well the survey represents the social programs in operation at the time. The chart below sequentially depicts the number of individuals by identified month: a) reporting any other social program income, b) any value consistent with AE 2020 benefits (R$600, R$300, R$1,200) reported as other social program income, c) any value consistent with AE 2021 benefits (R$250, R$375, R$150 ) reported as other social program income, d) any income reported as Bolsa Familia, e) any value consistent with the Auxilio Brasil minimal benefit established in December of 2021 (R$400) reported either as other social programs or Bolsa Familia. The dynamics in the data are very consistent with the schedule of these social programs. Figure A.1: Social programs recipients identified by inferred month 22