A causal inference framework for poststratification: a method for improving external validity in epidemiological studies

Abstract

Background Poststratification, a method for improving the representativeness of nonprobability samples, has developed primarily within survey methodology. Meanwhile, g-methods such as inverse probability weighting (IPW) and standardization have been developed in epidemiology for causal inference from observational data. Despite evolving under different terminology in two fields, these methods share similar underlying assumptions and estimation strategies. In this article, we systematically articulate the formal connections between poststratification and g-methods, demonstrating how each field can inform the other. Methods We develop a methodological framework demonstrating how poststratification can be understood through established causal inference principles. We formally map the three core assumptions required for valid poststratification onto the identifiability conditions used in causal inference from observational data. We show that the two principal implementation approaches for poststratification parallel inverse probability weighting and standardization. To illustrate the practical application, we apply poststratification to data from the Korean Genome and Epidemiology Study (KoGES) of 10,030 Korean adults aged 40-69 years in 2001, adjusting for age and sex distributions to estimate population-level smoking prevalence. Results The formal mapping reveals that poststratification and g-methods share analogous assumptions and estimation strategies. This parallel provides principled guidance for auxiliary variable selection, clarifies when poststratification succeeds or fails, and enables application of established diagnostic tools from causal inference to poststratification problems. In the KoGES example, poststratification adjusted the crude smoking prevalence from 25.7% to 26.5%, accounting for oversampling of older participants. Conclusions Understanding poststratification through the lens of causal inference offers a rigorous foundation for making valid population-level inferences from nonprobability samples. This framework facilitates cross-disciplinary learning and enables more principled interpretation of results from convenience samples, cohort studies, and other nonprobability sampling designs. Future work could explore how poststratification methods relate to the broader literature on generalizability and transportability of results from randomized trials.