Data integration using propensity score weighting under missing at random

Abstract

We study data integration for the setting in which a probability sample and a nonprobability sample are both available from a common target population, with auxiliary variables observed in both samples and the study variable observed only in the nonprobability sample. Under a parametric propensity score model for the nonprobability sample inclusion mechanism, we develop a conditional maximum likelihood estimator of the inclusion probabilities and construct an inverse probability weighted (IPW) estimator of the population mean. We establish consistency and asymptotic normality of the proposed IPW estimator under suitable regularity conditions. To guard against misspecification of the propensity score model, we further incorporate an outcome regression model through debiased generalized entropy calibration (GEC) yielding a doubly robust estimator that remains consistent when the propensity score model or the outcome regression model is correctly specified. We also develop variance estimation procedures for both estimators. Simulation studies illustrate the finite-sample performance of the proposed methods, and we apply them to data from the Culture and Community in a Time of Crisis survey, where the nonprobability sample substantially over-represents certain demographic groups relative to the U.S.\ adult population.