Skip to content
Back to search results

Poverty Mapping (Small-Area Estimation)

Introduced by
Chris Elbers, Jean O. Lanjouw & Peter Lanjouw
Year
2003
Type
Census-survey small-area poverty estimation method
Family
Poverty mapping and small-area estimation
Font size

What does it do?

ELL poverty mapping, named after Chris Elbers, Jean Lanjouw, and Peter Lanjouw, is a small-area estimation method that produces poverty and inequality estimates for geographic units far smaller than a household survey can support on its own.

It combines two data sources: a detailed household survey that measures consumption but covers too few households per locality, and a population census that covers everyone but does not measure consumption. The method estimates a model of consumption on variables common to both, imputes consumption into the census, and simulates to generate poverty estimates — with statistically valid standard errors — for districts, communes, or even villages, which are then drawn as poverty maps.

The intuition behind the method

Surveys can tell you the poverty rate of a country or a province, but ask about a single district and the survey has too few households there to give a reliable number. A census reaches every household but never asks the long questions needed to measure consumption. ELL bridges the gap. In the survey, where both consumption and a set of ordinary household characteristics are observed, it learns how consumption relates to those characteristics. Because the census records the very same characteristics for everyone, that learned relationship can be used to predict consumption for every census household. Predictions are uncertain, so rather than plug in a single number ELL repeatedly simulates whole consumption distributions, computes the poverty rate each time, and reports the average and its sampling variability for each small area.

History

Elbers, Lanjouw, and Lanjouw set out the method in their 2003 Econometrica paper 'Micro-Level Estimation of Poverty and Inequality', solving the problem of producing reliable welfare estimates for areas too small for direct survey estimation by statistically combining survey and census data. The World Bank adopted and operationalized it through the PovMap software and the supporting volume 'More Than a Pretty Picture' (2007), which documented applications and policy uses across many countries. Poverty maps built with the ELL method became a workhorse of geographic targeting, and subsequent research refined the error modeling and proposed empirical-best-prediction alternatives in the small-area-estimation literature.

When to use it

Use ELL poverty mapping when you need poverty or inequality estimates for small geographic areas that a survey alone cannot deliver, and you have a recent household consumption survey and a population census that share comparable covariates close together in time. It is the standard method behind national poverty maps used to allocate transfers and target programs geographically. It demands strong assumptions — that the survey-estimated consumption model holds in the census, and that covariate definitions match — and it is sensitive to the time gap between survey and census and to the modeling of the error components. Where consumption data are absent, asset-based small-area methods are an alternative; where only area-level auxiliary data exist, Fay-Herriot area-level models are used instead.

Limitations and considerations

  • Relies on the strong assumption that the consumption model estimated in the survey transfers unchanged to the census population.
  • Accuracy degrades as the time gap between the survey and the census grows, since the relationship between covariates and consumption drifts.
  • Estimates are sensitive to the specification of the error components (cluster effects, heteroskedasticity); mis-modeling them understates the true uncertainty.
  • Requires harmonized, comparably measured covariates in both data sources, which limits the usable variables and demands careful preparation.

Source conditions do not replace independent scientific review or validation by an analysis engine.

Strengths

  • Produces poverty and inequality estimates for very small areas — districts, communes, villages — far below the resolution a survey alone supports.
  • Delivers statistically grounded standard errors for each small area by simulating model and idiosyncratic uncertainty, not just point estimates.
  • Exploits the complementary strengths of survey (rich welfare measure) and census (full coverage) without requiring consumption to be measured for everyone.
  • Yields intuitive, policy-ready poverty maps that make geographic targeting and budget allocation concrete and transparent.

Applications

  • National poverty maps used to allocate budgets and target social programs to the poorest districts or communes.
  • Geographic targeting of cash-transfer and infrastructure programs below the level of survey representativeness.
  • Identifying pockets of poverty masked by provincial or national averages.
  • Combining poverty maps with data on access to services or hazards to study spatial deprivation and vulnerability.
  • Informing decentralization and fiscal-transfer formulas that depend on sub-national poverty estimates.

How it works

How it works

1

On the survey data, log per-capita consumption is regressed on a vector of household and location characteristics that are also available in the census: ln y_ch = x_ch'β + u_ch, with the disturbance decomposed into a location (cluster) component η_c and an idiosyncratic household component ε_ch. Modeling this error structure carefully — allowing spatial correlation within clusters and heteroskedasticity in the idiosyncratic part — is essential, because it is the main driver of the precision of the final small-area estimates.

Formula
\ln y_{ch} = x_{ch}^{\top}\beta + u_{ch}, \qquad u_{ch} = \eta_c + \varepsilon_{ch}
2

The covariates used in the model must be defined identically in the survey and the census so that the estimated relationship can be transferred. This means restricting the regressors to variables that are common, comparably measured, and stable between the two sources (often augmented with census-derived means at the cluster level). Careful harmonization of variable definitions and reference periods is what makes the imputation legitimate.

3

Consumption is imputed into the census by Monte Carlo simulation rather than by plugging in point predictions. In each of R replications the method draws parameter values (β) from their estimated sampling distribution and draws the location and household error components from the estimated error model, then computes a simulated consumption for every census household. This propagates both model (parameter) uncertainty and the unexplained idiosyncratic variation into the predictions.

Formula
\hat{y}_{ch}^{(r)} = \exp\!\left(x_{ch}^{\top}\tilde{\beta}^{(r)} + \tilde{\eta}_c^{(r)} + \tilde{\varepsilon}_{ch}^{(r)}\right)
4

For each simulation and each small area, a welfare measure — the headcount poverty rate is the share of people whose simulated consumption falls below the poverty line z, and inequality measures can be computed the same way — is calculated. Averaging across the R simulations gives the point estimate for the area, and the spread across simulations gives its standard error. The estimates are then mapped, yielding a geographically detailed picture of poverty with explicit uncertainty.

Formula
\hat{P} = \frac{1}{R}\sum_{r=1}^{R} \frac{1}{N}\sum_{i=1}^{N} \mathbb{1}\!\left(\hat{y}_i^{(r)} < z\right)

Common pitfalls

  • Reporting ELL point estimates without their simulated standard errors, hiding that small-area estimates can be quite imprecise.
  • Using a census and survey separated by many years, so the imputation model no longer reflects current relationships.
  • Including covariates that are defined or measured differently in the two sources, biasing the imputed consumption.
  • Ignoring the location (cluster) error component, which leads to substantially understated confidence intervals for the area estimates.

Frequently asked questions

How does the ELL method differ from Fay-Herriot small-area estimation?

Both are small-area estimation methods, but they operate at different levels. The ELL method is a unit-level approach: it estimates a model of household consumption and imputes it into individual census households, then aggregates. The Fay-Herriot model is an area-level approach: it combines direct survey estimates for each area with area-level auxiliary covariates through a mixed model, without needing unit-level census records. ELL is used when a matching census exists and consumption can be modeled at the household level; Fay-Herriot is used when only aggregated area-level data are available.

Why does ELL simulate consumption instead of using predicted values directly?

Plugging the regression's point predictions into the census would ignore two sources of uncertainty — the sampling error in the estimated coefficients and the unexplained household and cluster variation — and would systematically distort poverty estimates, which are nonlinear functions of the consumption distribution. By drawing parameters and error components repeatedly and recomputing poverty in each draw, ELL reconstructs the full distribution of consumption and of the poverty estimate, yielding both an unbiased point estimate and a valid standard error for each small area.

How close in time must the survey and census be?

Ideally they should be contemporaneous or only a year or two apart. The method assumes the relationship between household characteristics and consumption estimated in the survey still holds in the census; the larger the gap, the more economic conditions, prices, and that relationship can change, biasing the imputed welfare. A common rule of thumb is to keep the gap to a few years and to be cautious about interpreting maps built from a census and survey separated by a long interval.

Related methods

Sources

  1. Elbers, C., Lanjouw, J. O., & Lanjouw, P. (2003). Micro-Level Estimation of Poverty and Inequality. Econometrica, 71(1), 355-364.

    Open DOI reference
    Copied.
  2. Bedi, T., Coudouel, A., & Simler, K. (Eds.) (2007). More Than a Pretty Picture: Using Poverty Maps to Design Better Policies and Interventions. World Bank, Washington, DC.

    ISBN: 9780821369319
    Copied.