HEAM seminar - A User-Friendly Technique for Implementing Survey Weights Using Stata

We develop a command that implements the Imbens & Lancaster (1994) and Hellerstein & Imbens (1999) approach to estimating population weighted regression models for survey data by taking advantage of auxiliary information on moments reflecting the population from which the sample was drawn. Typically, the method is employed to improve regression estimates from a modest size researcher-conducted survey dataset with selected variables in common with administrative data, a census or large-scale survey. In a generalized method of moments framework, the method simultaneously minimizes the score functions of the estimator of interest (e.g., OLS/Logit/Probit) while also matching suitable moments (e.g., means, squares and/or cross products) for the variables common to the survey of interest and the population. As is the case with traditional approaches to generating survey weights, this addresses selection on observed variables that may arise due to issues, such as survey design, attrition, and/or non-response. Using Monte Carlo simulation, we show how this method can sometimes improve estimates compared to unweighted Logit and traditional methods of generating weights. We illustrate this method using nhanes2 data. That the method is easy to use is also beneficial.