Causal Inference with Complex Survey Designs

Abstract

This article discusses methods that combine survey weighting and propensity score matching to estimate population average treatment effects. Beginning with an overview of causal inference techniques that incorporate data from complex surveys and the usefulness of survey weights, it then considers approaches for incorporating survey weights into three matching algorithms, along with their respective methodologies: nearest-neighbor matching, subclassification matching, and propensity score weighting. It also presents the results of a Monte Carlo simulation study that illustrates the benefits of incorporating survey weights into propensity score matching procedures, as well as the problems that arise when survey weights are ignored. Finally, it explores the differences between population-based inferences and sample-based inferences using real-world data from the 2012 panel of The American Panel Survey (TAPS). The article highlights the impact of social media usage on political participation, when such impact is not actually apparent in the target population.