Efficient and Robust Propensity-Score-Based Methods for Finite Population Inference with Non-Probability Epidemiologic Cohorts

• Activity Number: 476 - Innovations in Analytic Approaches for Survey Data
• Type: Contributed
• Date/Time: Thursday, August 6, 2020 : 10:00 AM to 2:00 PM
• Sponsor: Survey Research Methods Section
• Author(s): Lingxiao Wang* and Yan Li and Barry Graubard and Hormuzd Katki
• Companies: University of Maryland, College Park and University of Maryland, College Park and National Cancer Institute and US National Cancer Institute
• Keywords: Non-Probability sample; Complex survey sample; Propensity score weighting; Taylor series linearization variance; Jackknife variance estimation; Weight scaling

Abstract

Finite population inference from non-probability samples by utilizing probability survey samples as external references is studied. Popular propensity-score (PS)-based weighting methods, e.g. inverse PS weighting (IPSW), can reduce bias but is often inefficient. Previously developed PS-based matching method that fractionally distributes survey weights to the nonprobability sample units based on kernel smoothed distance in PS (KW) can improve mean-squared error (MSE) over IPSW in estimating finite population means. We develop a unifying framework to advance pseudo-weighting from both PS-based weighting and matching methods. First, our framework identifies a fundamental Strong Exchangeability Assumption (SEA) underlying the PS-based matching methods. We make a Weak Exchangeability Assumption (WEA) by using survey weighting in PS estimation. Second, we propose IPSW.S and KW.S methods that reduce the variance of PS-based estimators by scaling the survey weights in propensity estimation. In simulations, the KW.S and IPSW.S estimators had smaller MSE. In our data example under WEA, the original KW estimates had large bias, whereas the KW.S estimates had the smallest MSE.