Making Data Count: Estimating a Poverty Trend for Nigeria between 2009 and 2019

Abstract

Monitoring poverty reduction requires frequent microdata on household welfare that can be compared over time. Such data are unavailable in many countries, given limited statistical capacity, shocks that prevent data collection, and regular improvements to survey methodology. This paper demonstrates how jointly deploying backcasting and survey-to-survey imputations can help to overcome this in a setting where estimating a poverty trend is badly needed, given the scale of the poverty-reduction challenge, but where survey-to-survey imputations are more likely to succeed and can be directly tested. In Nigeria, the most recent official survey that can be used to construct an imputation model was collected through the same methodology and in the same year as the target survey. This data landscape could arise in other settings where the methodology for smaller, interstitial surveys is updated more quickly than for larger, official consumption surveys. Naively comparing Nigeria's last two official consumption surveys would suggest that the poverty rate fell by 17 percentage points between 2009 and 2019. Yet the methods presented in this paper both suggest a much smaller reduction in poverty of between 3 and 7 percentage points, echoing Nigeria's performance on nonmonetary welfare indicators over the same period. The paper therefore provides guidance on when and how backcasting and survey-to-survey imputation techniques can be most valuable for monitoring poverty reduction.

1. Introduction

Monitoring progress towards poverty reduction requires rich microdata on household welfare that are collected frequently and that can be compared over time. However, in lower-income countries with limited statistical capacity or in contexts where household survey methodology is constantly being improved, these data are often outdated (Mahler, Castañeda Aguilar, and Newhouse 2022), missing (Dang, Jolliffe, and Carletto 2019), or incomparable with previous survey rounds. This has important implications for monitoring changes in poverty (Deaton 2001, 2005; Beegle et al. 2012) and can affect the policy response needed for poverty reduction—issues that become even more important in times of crisis such as during conflicts or global pandemics (Gentilini et al. 2020; Tandon and Vishwanath 2020).

Assessing the share of the global population covered by household surveys demonstrates the extent of the data-deprivation challenge, especially in poorer countries. The World Bank's global poverty database shows that data deprivations disproportionately affect Sub-Saharan Africa where, in 2019, only about half of the population was covered by a household survey in the previous three years, and the Middle East and North Africa, which has been severely affected by conflict and by limited data access (Ekhator-Mobayode and Hoogeveen 2022; World Bank 2022b). Survey coverage is also especially limited for low-income countries, at just 30 percent in 2019. Coverage for lower-middle-income countries (LMICs) appears to be relatively high, at 90 percent in 2019, yet this is entirely dependent on the presence of imputed data in the World Bank global poverty database. In particular, this includes imputed data for two large countries: (1) India (Newhouse and Vyas 2019; Sinha Roy and Van Der Weide 2022; World Bank 2022b) and (2) Nigeria, which uses the imputed household consumption vector presented in this paper. If imputed data were removed from the database, the coverage for LMICs would plummet to 43 percent.

Countries in Sub-Saharan Africa, in the Middle East and North Africa, and those that are low income are also less likely to have household survey data that are comparable and recent. In only around half of countries in Sub-Saharan Africa is it possible to technically compare the latest two surveys in the World Bank global poverty database; the proportion is similar for low-income countries too (see table 1). Additionally, in the Middle East and North Africa data are generally outdated; there, only around one-third of countries had a household survey after 2015.

This paper demonstrates how to overcome this data-deprivation problem by jointly using two completely distinct techniques—namely, backcasting and survey-to-survey imputations—to estimate poverty trends. It does so in the context of Nigeria, where two most recent official surveys used for poverty measurement in the country—the 2009/10 Harmonised Nigerian Living Standards Survey (HNLSS) and the 2018/19 Nigerian Living Standards Survey (NLSS)—are almost a decade apart and measure household consumption in an incomparable way. Yet the 2018/19 NLSS, which can be used to construct an imputation model (the training survey) was collected through virtually the same methodology as the target surveys into which the analysis imputes (the General Household Surveys (GHSs)). Also, data collection for the NLSS and the GHS (the training and target surveys) overlapped in 2018/19. This type of data landscape could become increasingly common as methodological improvements are applied more quickly to smaller, interstitial surveys compared to larger, official surveys that measure income or consumption (Yoshida et al. 2022).

The scale of Nigeria's poverty-reduction challenge further underlines the importance of addressing data deprivations to estimate poverty trends correctly. Given its vast population, Nigeria is estimated to have the second-largest number of poor people in the world, with around 4 in 10 Nigerians living below the international poverty line of US|${\$}$|1.90 2011 PPP per person per day. Therefore, its poverty-reduction efforts affect not only the country itself, but also the West Africa region, and the entire world (World Bank 2020).

Naively comparing poverty estimates from Nigeria's two most recent official consumption surveys would suggest a dramatic decline in poverty in the 2010s. Comparing, the 2009/10 HNLSS-based estimate and the 2018/19 NLSS-based estimate suggests that poverty fell by as much as 17 percentage points (Castaneda et al. 2022). However, over that period, Nigeria's National Bureau of Statistics (NBS) effected vital improvements to its survey methodology, including refining questionnaires, sampling, and survey implementation protocols. While these changes are key to augment data quality and match the best international standards for poverty measurement, they make it difficult to construct a measure of household welfare—and hence poverty—that can be compared over time (Deaton 2001; Beegle et al. 2012).

Both the techniques applied in this paper—backcasting and survey-to-survey imputations –suggest that Nigeria experienced far more modest poverty reduction in the decade prior to COVID-19. The backcasting approach involves mapping macroeconomic data on sectoral real gross domestic product (GDP) growth rates to microdata from the 2018/19 NLSS and then constructing estimates of the full consumption distribution for each year since 2009, from which poverty statistics can be calculated. The backcasts suggest that the poverty rate was between 42 and 46 percent in 2009–implying a drop of between 3 and 7 percentage points up to 2019–depending on the assumptions made about the pass-through rate from national accounts growth data to household consumption. The backcasts also indicate that poverty dropped slightly at the start of the 2010s, but then stagnated, or even began to increase, following the recession Nigeria experienced in 2016.

The survey-to-survey imputations suggest similarly muted poverty reduction during the 2010s. Specifically, the paper estimates a simple model linking monetary and nonmonetary variables using the 2018/19 NLSS, which is then is used to impute into the GHS, collected in 2010/11,2012/13,2015/16, and 2018/19. These survey-to-survey imputations suggest that the poverty rate was about 44 percent in 2010/11, implying a drop of less than 5 percentage points between 2010/11 and 2018/19. Echoing the backcasts, the imputed estimates also indicate that poverty reduction was slight in the first half of the 2010s, but then stalled and reversed following the 2016 recession.

Using these results, the paper aims to make six specific contributions.

First, it demonstrates the value of simultaneously applying two very different approaches with very different assumptions for producing poverty estimates: the survey-to-survey imputation approach can offset some of the caveats of the backcasting exercise and vice versa (see “Caveats” in section 6). That these two approaches produce such similar results builds confidence in the poverty trends presented in this paper. This shows the benefit of applying two (or more) alternative techniques to triangulate poverty trends when directly comparable consumption or income data are lacking.

Second, the paper is able to apply survey-to-survey imputations techniques with data that are especially conducive to their success, but which also allow some of their underlying assumptions to be tested. While the literature on survey-to-survey imputations is rich and growing (Stifel and Christiaensen 2007; Christiaensen et al. 2012; Douidich et al. 2016; Newhouse and Vyas 2019; Takamatsu et al. 2021; Sinha Roy and Van Der Weide 2022; Takamatsu, Yoshida, and Kotikula 2022), there are situations where such imputation methods fail to predict poverty accurately (Newhouse et al. 2014; Van Der Weide et al. 2022). The data environment in Nigeria gives such imputation methods the best chance to succeed. This is because the nonmonetary variables in the GHS—into which the analysis imputes—were collected through a virtually identical methodology to the 2018/19 NLSS, the official household consumption survey used to construct the imputation model. Moreover, the accuracy of the survey-to-survey imputation approach can be tested by exploiting the overlap of the training and target surveys in 2018/19. In that year, the imputed and official estimates of the poverty headcount rate differ by only around 3 percentage points, a relatively small difference given how prevalent poverty is in Nigeria. This provides a solid foundation for the imputed poverty estimates for the decade prior to 2018/19.

Third, and relatedly, the data landscape for Nigeria in the decade prior to COVID-19 may be relevant for other countries. Often, large, official surveys for measuring consumption or income will be accompanied by smaller, interstitial surveys, which are more limited in scope or smaller in sample size, but which are implemented more frequently—exactly like the NLSS and the GHSs in Nigeria. Given their frequency, improvements in survey methodology may be adopted more quickly by the smaller, interstitial surveys, making changes in their methodology more incremental, and making it more likely that two such surveys can be compared over time. Therefore, finding ways to use such surveys for poverty measurement—even if they do not directly contain suitable information on consumption and income—could be relevant for other countries outside Nigeria. For example, only 54 countries have a household survey available after 2019 in the World Bank's global poverty database (91 percent of which are for countries in Europe and Central Asia, high-income countries, or Latin America), while information available from the COVID-19 World Bank Phone Surveys is available for 85 countries. Other countries with outdated household surveys available in the World Bank global poverty database, could apply an analytical framework similar to the one presented in this study exploiting information from GHS-like surveys available in the Living Standards Measurement Study (https://www.worldbank.org/en/programs/lsms/initiatives/lsms-ISA#2). For example, the latest available official household survey for Ethiopia dates back to 2015, but a LSMS survey is available for 2018/19. Other countries that recently underwent statistical capacity building and regional-survey harmonization exercises (as was done for countries across West Africa in the Harmonized Surveys on Household Living Conditions (PEHCVM) project), and where the latest survey is not comparable to the previous one, could also adopt a similar analytical framework to assess changes in poverty over time.

Fourth, the paper builds on existing evidence showing the value of combining data on GDP growth with consumption or income data from household surveys—as in the backcasts in this paper—to assess poverty dynamics. Much of the existing evidence has considered whether forecasts and nowcasts from older household survey data are able to match the “true” results emerging from newer household surveys (Deaton 2005; Prydz, Jolliffe, and Serajuddin 2021; Prydz et al. 2019; Mahler, Castañeda Aguilar, and Newhouse 2022). When applied to data from multicountry databases like the World Bank global poverty database, these forecasts and nowcasts appear to perform relatively well. This paper lends further credence to the idea that combining growth data with household survey data may be a tenable approach for assessing poverty dynamics when other data are lacking; but it does so focusing on one country, going backwards in time, and by comparing the backcasted results with survey-to-survey imputations. This is important because even in countries facing severe data deprivations, they may have still data on GDP growth and one household survey from which to backcast, nowcast, or forecast.

Fifth, the paper shows the cost of naively using incomparable consumption or income data to estimate poverty trends. Far from the rapid decline in poverty for the decade prior to COVID-19 implied by comparing the 2009/10 HNLSS and 2018/19 NLSS, both the backcasts and survey-to-survey imputations suggest modest decline at best and even some increase in poverty since the 2016 recession. When these trends are misestimated for large countries like Nigeria, regional and global poverty measurement is also affected. Adopting the trend resulting from this analysis (yielding a poverty rate of about 46 percent in 2019) results in a poverty rate for 2009 that is 5 percentage points lower for West Africa, 2 percentage points lower for Sub-Saharan Africa, and 0.2 percentage points lower for the world than the poverty rate obtained when drawing a trend between the 2009/10 HNLSS and the 2018/19 NLSS. This is driven by there being an estimated 16 million fewer poor in Nigeria in 2009, when the backcasts and imputations are applied.

Sixth, the paper provides crucial information to Nigeria's policymakers. While using “snapshots” of data can provide some information on the drivers of poverty and what corrective policies might be needed, understanding poverty dynamics is essential for gauging the right mix of policies for reducing poverty at the country level (World Bank 2022a).

The paper is organized as follows. Section 2 describes the methodology used in the analysis. Section 3 details the data, and section 4 shows the main results. Robustness checks and sensitivity analysis are reported in section 5. Additional discussion of the results with possible appears in section 6. Section 7 concludes. The data availability section follows.

2. Methodology

This section describes the backcasting and survey-to-survey imputation approaches in more detail. This section therefore underscores one key contribution of the paper: showing the value of triangulating similar results using two very different methods.

Backcasting Methodology

The backcasting exercise takes the full consumption vector in the 2018/19 NLSS and then constructs the consumption vector in each previous year by “rolling back” consumption for each household using sectoral real GDP growth rates and population growth rates. The sectors considered are agriculture, industry, and services. The sectoral GDP data are already in real terms, having been adjusted using the GDP deflator: the analysis therefore does not conduct any additional price adjustments to deflate the consumption vector in each year. Real sectoral GDP growth rates are converted to per capita terms by applying population growth “flat” to each sector. This approach does not, therefore, allow for sectoral switching, yet given the slow pace of structural transformation in Nigeria, this may be a tenable assumption over the period of interest (Jenq, Lain, and Vishwanath 2021). The study uses the household head's employment sector to match the 2018/19 NLSS to real GDP growth rates in each sector. Those household heads whose sector could not be distinguished because the household contained multiple enterprises or those who were not working at all were assigned a weighted average of the per capita real GDP growth from agriculture, industry, and services.

The backcasted series starts by assuming that the 2018/19 NLSS effectively corresponds to 2019, for the purposes of mapping it to the macroeconomic data. This seems like a reasonable assumption since the data were collected between September 2018 and October 2019, covering two-thirds of 2019. The analysis applies the growth rate between 2018 and 2019 to the survey estimate to backcast an estimate for 2018, then for 2017, and so on for all the other years until 2009.

The main formula for the backcasts can be written:

where |$g_{t - 1}^{cons,s} = pass \times g_{t - 1}^s - {p}_{t - 1}$|

|${C}^s$| is household consumption for households whose head is employed in sector s; |$pass$| is the assumed pass-through rate value between growth in national accounts and in household consumption;|${g}^s$| is real sectoral GDP growth in sector s; p is population growth. The pass-through rate is initially assumed to be the same across sectors and across richer and poorer Nigerians. The main results assume a pass-through rate of 1: they assume that growth in national accounts (in real, per capita terms) is fully passed onto household consumption. Nevertheless, sensitivity analysis, which applies different pass-through rates to check the robustness of the results, is presented in section 5. The backcasted consumption vector is converted to US|${\$}$| 2011 PPP terms to estimate poverty rates at the international poverty line.

Survey-to-Survey Imputation Methodology

The survey-to-survey imputations use nonmonetary indicators and household consumption data from a “baseline” or “training” survey to impute consumption into a “target” survey(s) that contain(s) the same nonmonetary indicators. The exercise proceeds in three main steps.

First, the analysis selects a set of comparable nonmonetary indicators that are available in both the NLSS and GHS (table 3). It is possible to further verify that the indicators coming from the two surveys are indeed comparable—as would be expected given their similar methodologies—using data from the 2018/19 NLSS and 2018/19 GHS. While the two surveys differed in their data-collection schedule—with the 2018/19 GHS being collected in two visits in July–September 2018 and January–February 2019 and the 2018/19 NLSS being collected over 12 months—they provide information for part of the same year. As column 3 in table 3 shows, the indicators are overall highly comparable between the two surveys, with only a few variables—such as household head's employment variables and consumption frequency of food items—showing any differences, possibly due to seasonal variation.¹ Section 6 tests the robustness of the results to different specifications of the consumption model that do not include these variables.

Second, the study estimates a consumption model using the selected variables and consumption data from 2018/19 NLSS. The variables for the consumption model are selected using stepwise selection with an optimal p-value of 0.01. Variables used in the consumption model include regional dummies, household demographics (dependency ratio), household head characteristics (gender, employment category), dwelling characteristics (main source of cooking fuel, toilet availability), asset ownership (air conditioning, washing machine, cars and other vehicles, generator, microwave, TV set, computer), and consumption frequency dummies (food and nonfood items). The model therefore includes variables that capture short-run variation—such as employment and a set of dummy variables for whether certain food and nonfood items were being consumed—as well as more stable household characteristics (Yoshimura et al. 2022).²

Specifically, the analysis estimates the following regression:

where |${y}_{h,t}$| is the natural logarithm of annual spatially adjusted household consumption expressed in local currency units for household h in time t. |${X}_{i,h,t}$|is a vector of household head's characteristics, |$Zon{e}_t$| are geographical zone-areas dummy variables. The error term is drawn from a normal distribution. The results of this estimation are available in table 3.

Third, using the parameters estimated from this consumption model, it is possible to impute consumption into the target survey and calculate the relevant poverty rates. The imputed consumption vector is estimated using 100 imputations. Figure 1 shows the distribution of the imputed consumption vector and compares it to the distribution of the NLSS-based consumption aggregate.³ The imputed consumption vector is then converted to US|${\$}$| 2011 PPP to estimate poverty rates at different international poverty lines. This step is repeated for the same nonmonetary indicators from each GHS wave to impute consumption in 2010/11,2012/13,2015/16, and 2018/19.

3. Data

All of the analysis in this paper depends on the 2018/19 NLSS; as the base year of microdata for the backcasts and as the training survey for the survey-to-survey imputations. The 2018/19 NLSS was conducted between September 2018 and October 2019 and was designed to provide estimates for a wide range of socioeconomic indicators—including consumption and poverty—for Nigeria's 36 states and the Federal Capital Territory (FCT), Abuja. The sample of around 22,000 households is representative at the national, zone, and state levels, aside from Borno state (which accounts for around 2.5 percent of the Nigerian population). The household questionnaire provides information on demographics, education, health, employment, food and nonfood consumption, and various other welfare indicators. Overall estimates of household consumption were constructed using the modules on food and nonfood consumption and spatially and temporally deflated using a price index constructed from unit prices in the food consumption module: this provides a consistent measure of welfare for the whole of Nigeria. From this, poverty estimates at the international poverty line of US|${\$}$|1.90 2011 PPP per person per day can be constructed by deflating over time using Consumer Price Index (CPI) data and converting to dollars using purchasing power parities (PPP) (Atamanov et al. 2018; Lakner et al. 2018).

The study then use stwo additional data sources that are available in Nigeria to implement the backcasts and survey-to-survey imputations and estimates a poverty trend for the country; first, the backcasting exercise uses national accounts data on GDP. The analysis uses yearly national accounts data on real sectoral GDP growth for Nigeria itself as the main foundation for “rolling back” the 2018/19 consumption vector to 2009 (table 2). Cross-country GDP per capita data from the World Development Indicators (WDIs) are also used to estimate different pass-through rates between growth in the national accounts and growth in household consumption.

Second, the survey-to-survey imputations use household survey data from the GHS, available in four waves: 2010/11,2012/13,2015/16, and 2018/19. These data lend themselves to survey-to-survey imputations as they collect household-level information on nonmonetary indicators in the same way as in the 2018/19 NLSS. The NLSS and GHS not only had identical questionnaires for these nonmonetary indicators, but they were also collected by the same team as part of an ongoing NBS-World Bank collaboration, therefore minimizing discrepancies driven by survey methodology and implementation. Moreover, the timing of data collection for the 2018/19 NLSS and 2018/19 GHS overlapped, making it possible to test whether the imputed consumption and poverty estimates are well aligned with actual consumption and poverty estimates from a similar period. Each GHS wave contains around 5,000 households and is representative at the national, zone, and urban-rural level. Within each wave, data are collected in two distinct visits to the same set of households: the first “post planting” visit takes place some time between August and October and the second “post-harvest” visit takes places some time between January and April of the following year.

While the GHS also collects data on household consumption, these data are subject to a series of limitations that do not allow them to be used directly for poverty measurement purposes. In particular, the early rounds of the GHS imposed standard units (such as grams on kilograms) on quantities in the food-consumption module when nonstandard units may have been more appropriate: this was addressed in later rounds of the GHS. However, this issue does not affect the nonmonetary indicators in the GHS, on which the survey-to-survey imputations rely.

4. Main Results

The results of the backcasting and survey-to-survey imputation yield very similar results. Both show a decline in poverty in the first half of the decade prior to COVID-19 followed by a period of stagnation—and even a slight increase—between the 2016 economic recession and 2019.

Backcasting Results

Figure 2 shows the backcasted trend in poverty headcount rates (panel a) and number of poor (panel b) at the US|${\$}$|1.90 poverty line for the period 2009–2019, assuming a pass-through rate of 1. The backcasts suggest that poverty rates were considerably lower in 2009 (46.1 percent or 73.2 million people) than the estimates obtained using the 2009/10 HNLSS directly (56.4 percent or 89.4 million people). The backcasts therefore suggest a drop in poverty of at most 7 percentage points between 2009 and 2018/19, about 10 percentage points smaller than the drop implied by comparing the 2018/19 NLSS and 2009/10 HNLSS.

Survey-to-Survey Imputations

The analysis first imputes into the 2018/19 GHS—the year when the two surveys overlap—to check that the NLSS and GHS surveys are indeed comparable. Imputing into the 2018/19 GHS produces a poverty rate—at the US|${\$}$|1.90 poverty line—of 41.9 percent, which is within 3 percentage points of the official NLSS-based poverty headcount rate for 2018/19 of 39.1 percent (see fig. 3). The small gap between the imputed GHS-based and actual NLSS-based results may be explained by the differences in some of the nonmonetary variables used in the consumption model (shown in table 3); these differences in turn could be driven by the two surveys having different data-collection schedules. Nevertheless, since the overall differences between the results from imputing into the 2018/19 GHS and those coming directly from the 2018/19 NLSS are relatively small, this result lays reasonable preconditions to impute back into previous GHS rounds.

Turning to the results from imputing into earlier GHSs, the survey-to-survey imputations echo the backcasts. The imputed poverty headcount rate at the US|${\$}$|1.90 poverty line is 43.5 percent in 2010/11, decreasing to 42.5 percent in 2012/13 and to 40.7 percent in 2015/16 (see fig. 3). While the imputed poverty headcounts rates are, on average, 2 percentage points higher than their backcasted equivalents, they yield a poverty trend that runs parallel to the backcasted series (see fig. 3). Figure 3 also shows the confidence intervals around the imputed estimates (see also table 4) and compares these to those around the backcasted estimates. This suggests that the results from the two methods are not statistically different from each other. To test this formally, the analysis once again exploits the overlap between the 2018/19 NLSS and 2018/19 GHS survey and finds that the 2.79 percentage point difference between the two estimates is not statistically different from 0 (t = 1.26). The study also tests the variation of the imputation results by looking at different prediction model specifications and different distributional assumptions (see “Survey-to-Survey Imputation Robustness and Sensitivity Analysis” in section 5).

The imputed consumption vector can also be used to estimate measures of inequality (see table 4). These suggest that inequality has barely changed between 2010/11 and 2018/19, which reinforces the assumption of distribution-neutral pass-through rates adopted in the backcasting exercise. Yet there are many other underlying assumptions for both the backcasts and survey-to-survey imputations, which the study aims to relax in the sensitivity checks below.

The article also examines how the nonmonetary variables used in the consumption model evolved across the four waves of the GHS (see fig. 4). Several of these indicators showed an improvement in household welfare over the first three GHS waves, but then worsened after 2015/16, possibly in reaction to Nigeria's 2016 oil-price-induced recession. For example, ownership of assets—especially television sets, generators, and cars and other vehicles—was increasing up until 2015/16 and but then fell in 2018/19 suggesting that households may have sold assets to reduce the impact of the recession on their welfare. The fact that these patterns emerge for the raw variables demonstrates that the results that follow are not purely a product of the functional form of the consumption model.

5. Robustness and Sensitivity Analysis

Backcasting Robustness and Sensitivity Analysis

To test the robustness of the backcasted poverty estimates, three sensitivity checks are conducted. Specifically, the analysis checks whether the results are sensitive to: (1) assumptions about the overall pass-through rate; (2) the assumption that pass-through is the same for all households; and (3) the mapping from sectoral growth rates to households.

First, the analysis tests the sensitivity of the results to different assumptions about how much of the growth in national accounts is passed onto growth in household consumption. An extensive literature shows that growth in national accounts differs from growth in average household consumption measured in household surveys (Ravallion 2003; Deaton 2005; Pinkovskiy and Sala-i-Martin 2016, Lakner et al. 2022). To account for these differences, the study turns to global data to estimate different values of the pass-through rate using survey data on household consumption to national accounts data from the WDIs. The analysis regresses household welfare (measured as either income or consumption) on GDP per capita using all available country-years, splitting up the sample according to countries’ income group and region. These regressions are run on all available surveys in the World Bank global poverty database. From these regressions, the study obtains an estimated pass-through of 0.87 when using all available surveys, and of 0.75 when using only consumption-based surveys, in line with the literature (Mahler, Castañeda Aguilar, and Newhouse 2022).

Results are reported in fig. 5 for pass-through rates between 0.42 (pass-through estimate for countries in fragile conflict-affected situations, see Corral et al. 2020) and 0.87. Poverty estimates for 2009 at the US|${\$}$|1.90 poverty line range between 42.2 (assuming a pass-through of 0.42) and 45.3 percent (assuming a pass-through of 0.87). These result show that changing the assumption about the pass-through rates has little effect on the backcasted series. Indeed, if anything, using lower pass-through rates implies even slower poverty reduction, indicating a larger gap with the 2009/10 HNLSS poverty estimate.

Second, the assumption is relaxed that growth in the national accounts is passed through to growth in household consumption at the same rate across the entire consumption distribution. To do this, the analysis constructs three separate growth incidence curves (GICs) using imputed consumption data from the survey-to-survey imputations. The GICs are constructed for the following time periods: 2010/11 to 2018/19, 2010/11 to 2015/16, and 2015/16 to 2018/19. Over the entire 2010/11 to 2018/19 period, Nigeria's growth incidence curve (GIC) was sloped slightly downwards, implying that poorer Nigerians benefited slightly more from growth than richer Nigerians: this corresponds to a small drop in the Gini coefficient of 0.6 points over this period. However, this picture is distorted by the effects of the 2016 oil-price-induced recession. The GICs based on imputed data indicate that richer households lost out significantly more than poorer households when the economic shock hit. However, during the first part of the decade—when Nigeria was growing more strongly—richer Nigerians disproportionately enjoyed the gains. Put differently, the consumption of richer Nigerians was more sensitive to Nigeria's overall growth performance than the consumption of poorer Nigerians. This is why it is important to separate out the periods before and after the 2016 recession when constructing the GICs and testing the robustness of the backcasts. Once they have been constructed, the GICs are then used to adjust the pass-through rate applied to each decile of the consumption distribution, holding the overall average pass-through fixed at 1.

Relaxing the assumption of distribution-neutral pass-through alters the backcasted trend, but even under the assumption that more growth was passed through to poorer Nigerians during the decade up to 2018/19, poverty reduction over that period remains in single digits. Figure 5 shows that using the 2010/11–2018/19 GIC—which was slightly pro-poor—in the backcasting exercise results in a larger reduction in poverty over the decade, such that the estimated poverty headcount rate is 47.9 percent in 2009 (1.6 percentage points higher than the main backcasted result). Conversely, using the 2015/16–2018/19 GIC results in a more stagnant backcasted trend over the decade and in an estimated poverty headcount rate of 44.6 percent in 2009; the 2015/16–2018/19 GIC accurately captures the fact that richer Nigerians’ consumption is more sensitive to Nigeria's growth. Overall, therefore, it appears that relaxing the assumption of a flat pass-through rate across the distribution shifts the backcasted estimate for 2009, but not enough to reproduce anything like the 17 percentage point drop implied by using 2009/10 HNLSS poverty estimate directly.

Third, the study checks whether the results are robust to using different methods to map the growth rates in sectoral GDP in the national accounts to the 2018/19 NLSS. In the main results, the mapping of households to sectors is based only on information about the household head's employment sector. To check whether the results are sensitive to this particular macro-micro mapping approach, the analysis instead uses the sector of employment of (1) the oldest working household member or (2) the household member closest to 40 in age as alternative variables to map the household data to the sectoral GDP series. This has virtually no impact on the results, yielding estimates for the poverty headcount rate in 2009 of 46.1 percent and 46.2 percent respectively.

Survey-to-Survey Imputation Robustness and Sensitivity Analysis

The study conducts two sensitivity checks to test the robustness of the survey-to-survey imputations. Specifically, it (1) examines whether the results are sensitive to the inclusion and exclusion of explanatory variables that could be subject to seasonality and (2) tests whether the results are sensitive to making alternative distributional assumptions about the consumption vector.

First, it appears that including variables that could be influenced by seasonality in the consumption model does not substantially affect the imputation results. In particular, employment and consumption of specific items might—both of which were included as explanatory variables in the main consumption model—could vary significantly during the year. This indeed turns out to be the case when comparing the summary statistics from the GHS post-planting and post-harvesting visits (see table 3). The value of these indicators in the GHS might therefore be different from the average value of the same indicators collected over a 12-month period, as in the 2018/19 NLSS—the target survey. Yet excluding these potentially seasonal variables from the estimation of the consumption model appears to make little difference to the results. To check this, the analysis estimates alternative models that exclude the household head's wage-employment (model 1); wage-employment and a dummy variable for there being no toilet facility in the household (model 2); wage-employment, no toilet facility, and the imported rice consumption dummy (model 3); nonfarm employment (model 4); and all employment and food consumption dummy variables (model 5). When using these alternative models, each year's imputed poverty estimates at the US|${\$}$|1.90 poverty line remains within 2 percentage points of the main results (see table 5a, table 5b).

Second, the analysis tests the robustness of the results to different distributional assumptions for the household consumption vector. The survey-to-survey imputations presented so far assume that household consumption is lognormally distributed. To test whether this is a valid assumption, the analysis applies zero-skewedness and Box-Cox transformations to the original 2018/19 NLSS consumption vector and to the imputations into the 2018/19 GHS, exploiting information for the overlapping year. The result is that assuming a lognormal distribution is a reasonable approximation for the data, in line with other literature on survey-to-survey imputations (Takamatsu et al. 2021). Specifically, that study compares whether different moments of the distribution estimated under the two different distributional assumptions are closer to the true distribution from the 2018/19 NLSS. It is found that the poverty headcount rate at the US|${\$}$|1.90 poverty line is 42.4 percent when using a zero-skewedness transformation and 42.6 percent when using a Box-Cox transformation. Both estimates confirm the robustness of the imputation exercise for 2018/19. However, the difference with the 2018/19 NLSS official poverty headcount rate (39.1 percent) is slightly higher than what is obtained with the preferred specification (41.9 percent), suggesting that assuming a lognormal distribution in the imputations is a good fit for the data.

6. Additional Discussion

This section explores the implications of the results and discusses possible caveats.

Sense-Checking with Other Surveys

Nonmonetary indicators from other surveys support the results of the backcasts and survey-to-survey imputations, further suggesting that improvements to household welfare stagnated and even reversed in Nigeria in the 2010s. Looking at markers of education and basic infrastructure in Nigeria's Demographic Health Surveys (DHS) demonstrates, for example, that secondary school enrollment and the share of the population with access to improved sanitation improved dramatically between 2003 and 2008, but showed virtually no change between 2008 and 2018 (see fig. 6). Given the high correlation between monetary and nonmonetary welfare indicators, this directly supports the poverty trends presented in this paper.

Caveats

Notwithstanding the robustness of the results to the checks described above, some standard caveats remain.

The backcasting exercise relies on two particularly strong assumptions; first, the backcasts assume that inflation is fully captured by the GDP deflator. Accelerating inflation in Nigeria in recent years has been driven disproportionately by food prices, and even poor households—many of whom are concentrated in subsistence agriculture—purchase food, so this could have uneven effects on consumption and hence alter the progress of poverty reduction (Joseph-Raji et al. 2021). Second, the backcasts assume that there are no switches between sectors over time. While this is an important assumption, evidence on Nigeria's labor market shows that structural transformation was slow over this period, suggesting that this may not be too big of a concern for the purposes of this analysis (Jenq, Lain, and Vishwanath 2021). While the study does not address these limitations directly in the backcasting exercise, the fact that the survey-to-survey imputations produce such similar results offsets some of these concerns. For example, the assumption of distribution neutrality is relaxed by the survey-to-survey imputation approach because variation in the distribution of nonmonetary indicators would capture changes in the distribution of consumption over time.

The main caveat of the survey-to-survey imputation exercise is the difference in data-collection schedules between the GHS and NLSS. The GHS data are only available during the post-planting and post-harvest seasons, which might bias the results if the indicators used in the consumption model vary significantly throughout the year and are hence different from those collected over a 12-month period in the 2018/19 NLSS. This may explain the difference between the imputed 2018/19 GHS estimate and the 2018/19 NLSS estimate. Yet, if the GHS systematically overestimates poverty rates—as also seems to be the case when comparing estimates in other waves to the backcasted series—this would suggest an even lower “true” poverty rate in 2010/11, even lower poverty reduction in the decade to 2019, and an even larger difference with the 2009/10 HNLSS poverty estimate. Moreover, any concerns that the relationship between monetary and nonmonetary may be too unstable, for survey-to-survey imputations could be at least partially offset by appealing to the backcasts’ results.

Changes to PPP Conversion Factors

Changes to the PPP conversion factors used for international comparisons could alter estimates of the poverty headcount rate at international poverty lines. This paper has used PPP conversion factors based on price data collected in 2011. However, Jolliffe et al. (2022) show that using new PPP conversion factors, created using price data collected in 2017, would have a large impact on estimates of the poverty headcount rate in some countries, including Nigeria. In order to understand these implied differences in the poverty estimates, further analysis is needed to examine how and why the new 2017 PPP data affect the conversion of the welfare vector from local currency units to international U.S. dollars for Nigeria. This is left for future work.

7. Conclusion

This paper proposes a comprehensive approach for estimating poverty trends when the detailed household consumption data needed to measure changes in household welfare are not available for a long period of time and cannot be compared to the previous official household survey.

The paper does this by first using sectoral GDP data to backcast household consumption and hence poverty rates from the latest official household survey (2018/19 NLSS) and second by using a survey-to-survey imputation approach, constructing a model for consumption using data from the 2018/19 NLSS and imputing into several waves of the GHS. Despite having very different foundations, these two approaches produce remarkably similar results. Far from poverty dropping by 17 percentage points—as making the invalid comparison between the 2009/10 HNLSS and 2018/19 NLSS would imply—it appears that poverty dropped by between 3 and 7 percentage points in the decade before the COVID-19 crisis. The results suggest that the 2010s were initially marked by gradual poverty reduction, but this subsequently stagnated and was even reversed following the 2016 recession.

This analysis provides vital evidence on how to estimate a poverty trend in contexts where official household consumption survey data are infrequent and where changes to the data-collection methodology do not allow survey estimates to be compared over time. Rather than proposing a new methodology, the paper shows how different data sources can be used to estimate a trend and how applying different methodologies can help improve the robustness of the results. A similar approach could be replicated in contexts where data on national accounts and/or nonmonetary indicators of household welfare are available, but where household-level consumption data are more limited. The “preconditions” of the data environment in Nigeria were especially conducive to applying and testing the backcasting and survey-to-survey imputation approaches. In particular, the availability of GHS data for an overlapping year with the official NLSS survey was crucial for validating the imputations, before going back throughout the 2010s.

This analysis also shows the importance of regularly collecting comparable data on household consumption. While this work shows that alternative data sources can be useful for estimating long-run poverty trends, it also highlights how many additional assumptions and checks are needed to produce robust evidence. Having direct estimates of monetary welfare, with trends as well as snapshots, would provide more precise and timely information on poverty. This is particularly relevant during economic crises—including the COVID-19 pandemic, rising inflation, and other shocks—when rapidly rolling out countervailing policies to help support households is critical.

Conflict of Interest

None.

Data Availability

Microdata from the 2018/19 NLSS survey and GHS survey are available at the World Bank Microdata library (https://microdata.worldbank.org/index.php/catalog/3827; https://microdata.worldbank.org/index.php/catalog/3557). Users need to register to access the data, and access to the data is regulated by the Microdata library terms of use (https://microdata.worldbank.org/index.php/terms-of-use). PovcalNet/PIP data is publicly available and can be accessed using the povcalnet STATA command (https://github.com/worldbank/povcalnet).

Author Biography

Jonathan Lain (corresponding author is an Economist in the Povery and Equity Global Practice of the World Bamk; his email address is [email protected]. Marta Schoch (corresponding author) is an Economist in the Poverty and Equity Global Practice of the World Bank; her email address is [email protected]. Tara Vishwanth is a Lead Economist with the Poverty & Equity Global Practice of the World Bank; her email address is [email protected]. The authors are extremely grateful for guidance received from Dean Jolliffe, Christoph Lakner, Berk Ozler, Aibek Baibagysh Uulu, Nobuo Yoshida, Paul Corral, Roy Van der Weide, Shinya Takamatsu, and the members of the Global Poverty Working Group. The authors gratefully acknowledge financial support from the UK government through the Data and Evidence for Tackling Extreme Poverty (DEEP) Research Programme funded by UK Foreign, Commonwealth & Development Office.

Footnotes

References