Summary

Household surveys are the main source of quantitative information for measuring social and economic indicators, such as poverty, inequality, employment, education, and many others, in Latin America and worldwide. They are used to estimate indicators that guide the design, evaluation, and monitoring of public policies, and they are also a primary source of information for monitoring the Sustainable Development Goals within the framework of the United Nations 2030 Agenda. The Economic Commission for Latin America and the Caribbean (ECLAC) manages microdata repositories, including the Household Survey Data Bank (BADEHOG), which compile, harmonize, and document surveys from more than 18 countries in the region since 2000, precisely because the information produced by these sources is essential for evidence-based decision-making.

However, behind this apparently democratized access to data lies a structural technical issue that often goes unnoticed: most household surveys do not come from simple random samples. On the contrary, they are based on complex sampling designs that include stratification, multiple stages, unequal selection probabilities, and calibration-based weighting adjustments. When an analyst applies traditional statistical methods designed for independent and identically distributed samples to data that do not have those characteristics, the consequences include biased estimates, underestimated standard errors, overly narrow confidence intervals, and hypothesis tests with inflated type I error rates.

As an illustrative example, consider a population with two regions of very different sizes and very different average incomes. A survey that selects the same number of people in each region and analyzes the data without weights will produce a structurally biased estimator of national mean income, because units in the smaller region have greater empirical weight in the sample than their actual weight in the population. This error also propagates to the variance. An analyst who computes the variance while ignoring the cluster structure will treat households in the same census sector as independent observations, when in fact they share common characteristics that reduce the effective information in the sample.

This document seeks to close that gap and provide researchers with the theoretical foundations and computational tools needed to work properly with surveys drawn from complex designs. Throughout the text, practical aspects are addressed in an integrated way, ranging from the definition of the sampling design to the estimation of indicators, such as totals, counts, means, and proportions, the construction of their confidence intervals, the formulation of hypothesis tests, modeling through advanced regression techniques, and the treatment of common imputation problems.