Poverty mapping in Latin America: ECLAC experiences on small area estimation

1.Introduction

Most Latin American and Caribbean countries regularly implement nationally representative household surveys to measure living conditions indicators, including poverty and income inequality. These surveys can generally be disaggregated geographically by urban and rural areas and the first-level administrative division (i.e. departments, provinces or regions). However, when it comes to obtaining a more dissagregated direct estimation of such indicators, results are subject to lower levels of accuracy and precision, which might be below the established quality criteria for their use. Small Area Estimation (SAE) techniques allow obtaining such disaggregated estimates while improving inference quality.

This article presents the recent experience of the Statistics Division of the United Nations Economic Commission for Latin America and the Caribbean (ECLAC) in applying SAE techniques to estimate geographically disaggregated poverty indicators based on household disposable income in seventeen (17) countries in Latin America. The results show a significant improvement in the precision of FGT-family indicators for geographical areas where the surveys do not attain adequate representativeness.

ECLAC regularly produces standardized national estimates of extreme poverty and poverty for Latin American countries, using a methodology that aims to achieve regional comparability. Even though countries in the region publish their official poverty statistics, the diversity of procedures and assumptions used in these estimates prevent direct comparison, possibly leading to erroneous conclusions by not considering their methodological differences.

The ECLAC approach for measuring poverty classifies a person as poor when the per capita income of their household is lower than the poverty line, based on the cost of meeting their food needs and other basic non-food needs [1]. The cost of food needs is estimated through the construction of basic food baskets, which provide the recommended amounts of energy and nutrients while reflecting the consumption habits of the population. The requirements come from current international recommendations to sustain a healthy life. Consumption habits are captured through household income and expenditure surveys and correspond to those of a particular subset of the population (reference population) that satisfies a set of basic needs.

The monthly cost of the basic food basket is referred to as the extreme poverty line. The cost of non-food needs is included in the poverty line by multiplying the extreme poverty line by the Orshansky coefficient (quotient between total expenditure and expenditure on food) of the same population of reference used to define the basic food basket.

The indicators commonly used to measure poverty corresponds to the family of parametric indices proposed by [2]. These indices (denoted FGT) correspond to the following equation:

In this expression, Nd is the total population in the subpopulation of interest d, yd⁢i is the per capita income per household, z is the value of the basic food and non-food basket (i.e. poverty line), and α is a parameter greater than or equal to 0 that determines the properties that the index fulfils. When α=0, the FGT index corresponds to the traditional “headcount index” (denoted by H); that is, the proportion of people below the poverty line. If α=1, the FGT index corresponds to the “poverty gap”, which measures the “depth” of poverty and indicates the average distance of poor people’s income to the poverty line, weighted by the incidence of poverty. If α=2, the coefficient assigns a greater relative weight to the observations whose income is at a greater distance from the poverty line and, therefore, makes the indicator sensitive to the income distribution among people under the poverty line [3].

Household surveys are designed and implemented by national statistical offices to generate representative statistics at a predefined level of aggregation, generally based on large geographic subdivisions, sex, or socioeconomic groups of the population. However, when direct estimations of different indicators are needed in smaller subdivisions than those envisaged initially, the inference resulting from the surveys is not enough precise or accurate. The higher the disaggregation, the less efficient the estimators become, and their reliability declines ostensibly. In the case of some complex indicators, this can even generate bias problems in the direct estimation and its standard error.

The implementation of the ECLAC methodology for poverty maps in each of the 17 countries followed these six stages:

After this introduction, Section 2 summarizes the SDG mandate for disaggregated statistics and reviews the three information sources used in this data integration process. Section 3 introduces the problem of updating intercensal counts as an input for the SAE model. Sections 4 and 5 show the theoretical foundations of the multilevel models based on unit-level SAE models and their corresponding mean square error estimation based on MCMC replicas. Section 6 describes the fundamentals of model checking and the benchmarking methodology. Section 7 delivers some final remarks on the usefulness of poverty maps in the region and presents the resulting poverty maps along with some technical considerations for their use and interpretation.

2.SDGs and sources of information

The 2030 Agenda for Sustainable Development comprises 17 sustainable development goals (SDG) that integrate the different dimensions of development, such as the economic, social, and environmental. The 2030 Agenda focuses on the most vulnerable subgroups of the population. This is why the Leave No One Behind (LNOB) mandate required to disaggregate SDG indicators by income, sex, age, race, ethnicity, migratory status, disability, geographic location, or other characteristics [4].

According to United Nations (2019), almost one-third of the global SDG indicators can be derived from household surveys. Poverty indicators are not an exception, as they are traditionally derived from continuous household surveys in Latin American countries. However, regular surveys are not planned to provide disaggregated poverty estimates for the LNOB mandate subgroups. However, data integration from various sources and small area estimation methods can serve as a proper vehicle to provide society with this kind of estimate.

Small area estimation (SAE) is a set of statistical techniques that serve to obtain disaggregated estimates of population parameters to improve inference quality when the disaggregation of direct household survey estimates does not meet the quality criteria required for publication. ECLAC has developed a production system to monitor poverty in 17 Latin-American countries at the levels that the LNOB mandate requires. Implementing the SAE approach presented in this paper requires having access to three sources of information: household surveys, censuses and satellite imagery.

National household surveys are compiled annually by ECLAC for its Household Survey Data Bank (BADEHOG), a repository of household surveys from 18 Latin American countries maintained by the Statistics Division. While these surveys are of different types – labour force surveys, living conditions surveys, and income and expenditure surveys – they are the sources of information used to produce the national official poverty and inequality statistics in each country. Table 1 shows a comprehensive summary of the household surveys used in this poverty mapping exercise.

Data from national population censuses are obtained from ECLAC’s census data bank, maintained by the Population Division (CELADE). In addition to microdata from the previous and current rounds of censuses, CELADE also provides the software Redatam. This computational solution handles large volumes of census microdata with a hierarchical structure down to the smallest area of the census exercise (blocks). It allows accessing and processing of encrypted census databases at high speed. Table 2 shows the population and housing censuses used in this exercise.

Additional data sources can also improve the accuracy of probabilistic surveys. The use of satellite images for small area estimation originated in 2002 [5] when crop yields were enhanced by employing a post-stratification estimator based on satellite spectral data from Indian satellites. This information made it possible to predict crop yields at the level of small geographic areas in India. Satellite data was obtained from Google Earth Engine, through Javascript and Python programming languages, and recently since 2021, in R with the rgee package [6].

Information based on remote sensing usually has allow more geographic disaggregation that can be difficult to attain by traditional means such as surveys or administrative records. The use of satellite images, particularly night lights, urban cover fraction and crop cover fraction may give countries the option of compensating for the lack of population censuses and detailed surveys [7].

3.Updating intercensal counts

Population and housing censuses in Latin American countries are the primary source of detailed information about their inhabitants and socio-demographic characteristics. In most developing countries, population censuses are carried out only every 10–15 years, which means the auxiliary information may be outdated. It would be expected that the quality of indicators resulting from an SAE procedure would benefit more from other sources of supplementary information.

Household surveys in Latin America are conducted annually, or at least more frequently than censuses. To keep cost at a feasible level, sample sizes are not enough to allow quality statistics to be generated at more disaggregated levels then study domains. This means that the survey cannot be used to adequately answer questions such as “how many indigenous young women in a given locality have completed their primary education?” To improve the quality of the estimates, SAE methods usually combine information from a probabilistic sampling survey with ancillary information, for example, from a population census.

As not every country in Latin America has a recent census, it is desirable to update census population counts by subgroups. For that purpose, we take advantage of the marginal counts provided by a survey to update the census counts. Then, when making predictions, we ensure our prediction is consistent with the updated distribution of the population across groups.

The need to work with up-to-date census tables is not unique to small area estimation. The field of demography offers different methods to update census counts [8] and table updating techniques have also been used to generate synthetic population data.

For simplicity, let us consider a two-way table Xa⁢j of population counts, where the rows of the table represent administrative areas, and age categories are defined in the columns. Each cell in the table contains the number of people corresponding to each combination of area and age groups. In addition, row margins, Xa. and margins X.j gives the population totals by areas and age groups, respectively.

One solution to update census counts is given by the Structure Preserving Estimator (SPREE) in one or more categorical variables of interest according to study domains for post-census years. The SPREE method is popular in the context of small area estimation. The procedure employs the iterative proportional fitting (IPF) procedure [9], also found in the literature as raking ration or multiplicative raking, to adjust the counts of a contingency table based on a set of given margins.

Because SPREE (through IPF) uses direct (and reliable) estimates, usually marginal totals from survey data, this technique is commonly seen as part of synthetic estimators. The SPREE method assumes the allocation structure obtained from an updated survey providing recent and reliable margins (rows and columns). On the other hand, it is assumed that the association structure that the census can provide is more solid; that is, it manages to represent the cells (interactions between rows and columns) in a more appropriate way than the survey data. In this sense, the SPREE method assumes that these interactions remain constant in the post-census years.

4.Small area models

ECLAC has implemented two approaches for poverty mapping in Latin America; depending on the characteristic of the sources of information, it can be possible to disaggregate poverty estimates at the department level or the municipality level. The former case is used to produce annual poverty estimates at the first-level administrative divisions (i.e. departments, provinces or states), disaggregated by a set of covariates from the 2030 Agenda’s Leave No One Behind Mandate. For example, department-level maps of poverty rate by age group and ethnicity. The latter case is the traditional poverty map, where the second-level administrative division (municipalities) is of interest to visualize the poverty distribution across the countries. ECLAC performs both approaches and has automatized them to regularly produce both kinds of estimates.

4.2SAE for municipalities

ECLAC uses a unit-level model with adjustments to the complex sampling design to estimate average income and poverty indicators at the municipality level when a recent census is available. The chosen approach was first proposed by [14], and it induces an approximation of the best empirical predictor (Pseudo-EBP) based on the model with nested errors [15]. Figures 2–4 shows the poverty maps for three latinamerican countries using this approach.

This method assumes that the transformed income variable yd⁢i*=log⁡(yd⁢i+c) follows the model described below (for simplicity, we will denominate the transformed variable as yd⁢i).

yd⁢i*=xd⁢iT⁢𝜷+ud+ed⁢i.

i=1,…,Nd,d=1,…,D,

where 𝜷 is the vector of regression coefficients, ud is the area random effect ud⁢∼𝑖𝑖𝑑N⁢(0,σu2) and ed⁢i⁢∼𝑖𝑖𝑑N⁢(0,σe2) are the errors for individuals in the dth area and are considered independent from the random effects. Notice that we transform the variable of interest to ensure the normality structure of the model with nested errors. The model considers a transformation of the variable percapita income that guarantees an approximately normal distribution. For this purpose, the Box-Cox and Log-Shift transformation families were explored. The latter was chosen to carry out the transformation of income in the models of the three countries, although the parameters associated with each country were different.

Figure 1.

Density of the Log-Shift transformation for per capita income for Colombia (left), Chile (center) and Perú (right). Source: Prepared by the authors.

The procedure consists in creating a new variable f⁢(yd⁢i)=log⁡(yd⁢i+c) for a grid of predefined values of c, which is a constant that ensures the normality of the transformed variable. Also, we make sure that Fisher’s coefficient of skewness is close to zero. The values for the constant c used in each country are expressed in local currency units and are the following: 8,600 (CLP in Chile), 81,958 (COP in Colombia), and 10.96 (PEN in Peru). Figure 1 illustrates the symmetry of the distribution associated with the transformation of per capita income.

Under this model, the vectors 𝒚𝒅 are independent and follow a normal distribution with mean 𝝁𝒅=𝑿𝒅⁢𝜷 and covariance matrix given by:

V𝒅=σu2⁢𝟏Nd⁢𝟏NdT+σe2⁢𝟏Nd⁢𝟏NdT.

According to [16], for those FGT indicators that can be defined as a function of yd – that is, 𝜹𝒅=𝒇⁢(𝒚𝒅) – the best linear predictor is the one that minimizes the Mean Square Error (MSE) and is given by:

δ~dB(𝜽)=E𝒚𝒅⁢𝒓[δd(𝒚𝒅)|𝒚d⁢s;𝜽]

In this expression, the expected value of the response variable for out-of-sample elements within the domain d (denoted yd⁢r) is conditional on the values observed in selected households and the model parameter vector (usually unknown). As [16] states, replacing 𝜽 with a consistent estimator 𝜽^ gives the best empirical linear predictor 𝜹~𝒅𝑬⁢𝑩=𝜹~𝒅𝑩⁢(𝜽^). By decomposing the yd vector and Xd, Vd matrices in two parts, one associated with the sample and the other related to the out-of-sample observations; we have that

𝒚𝒅=(yd⁢syd⁢r),𝑿𝒅=(Xd⁢sXd⁢r),

𝑽𝒅=(Vd⁢sV𝑑𝑠𝑟V𝑑𝑟𝑠Vd⁢r)

Since 𝒚𝒅 follows a normal distribution, the conditional distribution 𝒚𝒅⁢𝒓|𝒚𝒅⁢𝒔 will also be a normal distribution parameterized as follows:

𝒚d⁢r|𝒚d⁢s∼𝑖𝑛𝑑N(𝝁dr|s,𝑽dr|s) 𝑤𝑖𝑡ℎ

d=1,…,D

To avoid the bias induced by ignoring the sampling design in the model, the parameters of the above distribution may consistently be estimated by including the sampling weights (wk⁢d)

𝝁^dr|s=𝑿d⁢r⁢𝜷^+γ^d⁢(y¯d⁢w-𝒙¯d⁢wT⁢𝜷^)⁢𝟏Nd-nd

𝑽^dr|s=(σ^e2+σ^u2⁢(1-γ^d))⁢𝟏Nd-nd⁢𝟏Nd-ndT

Being 𝟏j a vector of ones of length j, y¯d⁢w=∑j∈sdwk⁢d⁢yk⁢d/∑j∈sdwk⁢d, 𝒙¯d⁢w=∑j∈sdwk⁢d⁢𝒙k⁢d/∑j∈sdwk⁢d, γ^d=σ^u2/(σ^u2+σ^e2⁢δd2) y δd2=∑j∈sdwk⁢d2/(∑j∈sdwk⁢d)2. Furthermore, 𝜷^, σ^e2 y σ^u2 are consistent estimators for the regression coefficients, the variance of the error, and the variance of the random effects, respectively. Therefore, the conditional distribution of an out-of-sample household (i∈rd) is given by:

Yd⁢i|𝒚d⁢s∼N(𝝁di|s,𝝈di|s2) 𝑤𝑖𝑡ℎ

d=1,…,D

Moreover, it is estimated by the following expression:

𝝁^di|s=𝒙d⁢iT⁢𝜷^+γ^d⁢(y¯d⁢w-𝒙¯d⁢wT⁢𝜷^)

𝑽^di|s=σ^u2⁢(1-γ^d)+σ^e2

The estimation equation for the FGT poverty indicator of order α(δd=Fα⁢d) is obtained by rewriting the indicator as a function of the response variables yd⁢i. However, since it is not possible to identify and link the units of the survey samples with those of the censuses, then the approach used is a Census-EB type, which assumes that all the elements of the census are associated with out-of-sample observations, as follows:

F~α⁢dB⁢(θ)=1Nd⁢(∑i∈rdF~α,d⁢iB⁢(θ))

It is essential to clarify that the ratio of the number of sample units to the number of people in the country is low in most countries. Therefore the Census-EB predictor will perform quite similarly to Pseudo-EBP. We consider a Monte Carlo simulation procedure to estimate poverty indicators since the expectation that defines the best predictor often cannot be calculated analytically. The steps of the algorithm are as follows:

• • Fit the unit-level model based on the survey data (𝒚𝒔,𝑿𝒔) to obtain consistent estimators 𝜽^=(𝜷^T,σ^u2,σ^e2)T of the vector of unknown parameters θ.

• • Estimate the vector of means μ(di|s) and of variances 𝝈(di|s)2 using the estimate 𝜽^=(𝜷^T,σ^u2,σe2)T from step 1 in the equations defined for the conditional distribution Yd⁢i|yd⁢s∼N(μdi|s,σdi|s2) for d=1,2,…,D.

• • For each small area d, generate A Monte Carlo simulations of the vector of the variable of interest for households outside the sample, 𝒚d⁢r(a).

• • Calculate the indicator of interest: Fα⁢d(a)~=Fα⁢d⁢(𝒚d(a)) for each Monte Carlo simulation a=1,…,A.

Although ECLAC uses its own poverty and extreme poverty lines for each country, they remain comparable. This way, Monte Carlo’s estimate of the Census-EB of the FGT indicator (F^α⁢dEB) is obtained by averaging the indicators for the A simulated censuses, that is,

F^α⁢dEB=1A⁢∑a=1AFα⁢d(a)~

In the estimation algorithm previously described, the Monte Carlo estimate of the FGT indicator depends on the vector of the regression coefficients 𝜷^. Therefore, an essential part of this stage is identifying the auxiliary variables’ predictive capacity. We generate different linear models from several combinations of the covariates (with and without intercept) and compare them. We consider the number of significant variables and goodness of fit measures (AIC or BIC) in the analysis and comparability of the models. Furthermore, we use Ridge and adapted Lasso regressions to obtain a first impression of the feasibility of a set of covariates. These regressions belong to a family of Gaussian models where, if we have observations xi∈ℝp and responses yi∈ℝ,i=1,…,n, we seek to minimize the following expression:

min(β0,β)∈ℝp+112⁢n⁢∑i=1n(yi-β0-xiT⁢β)2

 +λ⁢[(1-α)⁢||β||222+α⁢||β||1]

where λ⩾0 is the complexity parameter and 0⩽α⩽1 is a known parameter. When α=0, the ridge regression is obtained, while with α=1, the Lasso regression is obtained.

5.Prediction on small areas and their corresponding MSE

Note the importance of standardizing the relevant variables by applying homogeneous definitions and categories in both data sources. In this way, possible biases induced by the different measurements in the covariates or errors in the prediction due to different variables with similar names are ruled out. To this end, standardized structures and a dictionary of variables are generated that describes the categories and other necessary specifications. For instance, the variable “years of study” in censuses is created using the criteria adopted for surveys in BADEHOG.

We apply a parametric Bootstrap method to estimate the ECM of the Census-EB predictor. The unit-level models fitted to the survey data are replicated using census microdata. The algorithm considered describes a slight modification to the one published in (16) and has the following steps:

Steps 2 through 7 are repeated B times to complete the Bootstrap simulation process. The mean square error estimator is given by:

In this equation, B is the number of Bootstrap simulations performed, D is the number of small areas (provinces, communes, and municipalities) selected in the surveys of each country, F^a⁢dE⁢B*(b) is the Monte Carlo estimate for each domain in every stage of the process, and F^a⁢d*(b) is the estimated FGT indicator in the pseudo-census. With the estimated mean square error, the coefficient of variation (a measure that allows defining the quality of the estimates) has the following expression:

We exclude from the map any province, commune, and municipality with a coefficient of variation (C⁢V^) greater than 30%, as they do not reach the desired precision.

6.Model assumptions and benchmarking

Another stage of the procedure consists in benchmarking the results using the estimated FGT indicators from the survey at the national, national urban, national rural, and first administrative division levels. This guarantees consistency between the published figures as the aggregations of provinces, communes, and municipalities must be identical to those reported at different levels of disaggregation. This procedure also reduces the bias produced by a model miss-specification and improves the estimates’ quality in the provinces, communes, and municipalities.

In this step, we use the multivariate calibration of ratios described in [17]. The Monte Carlo simulation makes it possible to access the prediction vector for all the households in the pseudo-census. In addition, we have the poverty estimate from the survey. These quantities have the following ratio representation Rq=ty⁢qtx⁢q. The purpose of the algorithm is to find a set of weights dk(k∈U), such that they meet the following restrictions:

For instance, using the survey poverty rates for national rural and national urban levels in each country, the benchmark algorithm would calibrate these ratios Rq(q=1,2) over the entire pseudo-census to find the set of weights dk that satisfy both restrictions at the same time. We execute this algorithm for each of the iterations of the Monte Carlo procedure.

After this stage, it is essential to carry out a strict validation and model assessment process to detect possible violations of the model assumptions. Tests used include the Kolmogorov-Smirnov, for normality (although it is also possible to use graphical diagnostics in the form of density kernels and quantile-quantile graphs) and White and Breusch-Pagan, for heteroskedasticity. Also, we use Cook’s distances as measures of the influence of the observations in the sample. The last stage uses geographic information for each country and shapes of interest (at the province, commune, and municipality level) to generate the maps.

7.Final remarks

This work aims to produce geographically disaggregated poverty indicators and produce maps to visualize the resulting estimates, helping policymakers to have a clear perspective on the incidence of the estimated indicator in different geographical domains, using different shades or colors to represent the magnitude of income and poverty indicators. The appendix presents a poverty map for 17 countries of Latin American, and poverty maps at the municipality-level for three countries: Chile, Perú and Colombia. All the poverty maps were adjusted using four cut-off points and a scale ranging from light green (lower poverty rate) to dark red (higher poverty rate) to illustrate the distribution of poverty by provinces, communes, or municipalities in the countries.

Figure 2 shows the SAE estimates for 17 countries in Latin America at the first administrative level, using similar cut-off points across countries to allow for regional comparison. Figures 3 to 5 refer to the EPB model at the second administrative level for Chile (Fig. 3), Perú (Fig. 4) and Colombia (Fig. 5). In this case, cut-off points are specific to each country, given the marked differences in their national poverty rates. Colombia has the highest poverty rates of the three countries under analysis. In Fig. 5 it can be observed that the poorest municipalities are located on the periphery of the country; specifically, in the Pacific region (Chocó, Cauca and Nariño), the Caribbean region (La Guajira, Magdalena, Bolívar, Sucre, Córdoba and Cesar), border departments (Norte de Santander and Arauca), and the Orinoquía and Amazon region (Vichada, Guainía, Vaupés, Amazonas and Caquetá).

In Peru, the highest poverty rates are in the border areas of the north and south of the country. Specifically, to the north, the highest incidence of poverty is concentrated in Loreto, Amazonas, the southern part of Cajamarca, and the western part of La Libertad. In the border area with Bolivia, another focus of poverty is observed in the regions of Puno and Cuzco.

An important feature of poverty maps is to allow for disaggregation by geographical areas and individual or household characteristics relevant to poverty. Figure 6 shows how Colombia’s poverty affects people of indigenous origin and people with low educational attainment, disaggregated at the department level. This is particularly relevant in the context of the 2030 Agenda and its mandate to leave no one behind.

The use of SAE methodologies can lead to the production of highly disaggregated official statistics that describe the situation of different population groups and geographical areas as an essential input for public policies that aim to improve living conditions for all. Therefore, the methodologies described in this paper and their results are particularly relevant in the context of the 2030 Agenda and its mandate to leave no one behind.

In addition, visualizing this information on a map at disaggregated geographical levels emerges as an effective communication tool to facilitate the interpretation and analysis of spatial relationships. It can be a valuable input to establish priority areas of attention, implement the geographic targeting of public spending, and improve coverage of social programs, among others. In addition, contrasting the poverty maps with complementary geospatial information, such as the availability of roads or electricity, can lead to a better understanding of the interrelation between poverty and living conditions in general.