R. Ferri García, J. Beaumont, K. Bosa, J. Charlebois, K. Chu
Selection bias is widespread in nonprobability surveys, and propensity score adjustments are often considered to mitigate this bias. In many situations, especially in large-scale social surveys, the number of variables of interest can be large. Choosing covariates independently for each variable of interest would lead to a different set of adjustments for each variable. In order to have a single set of adjustments, a single set of covariates must be determined, which is achieved by modelling the propensity to participate. However, some covariates can be strongly associated with the propensity to participate but only weakly associated with the variables of interest. Such covariates tend to be useless for reducing the bias and increase the variance of the adjusted estimates. Weight smoothing techniques can be used to address this issue. This study evaluates the performance of weight smoothing in a nonprobability survey context when there are multiple variables of interest.
Keywords: Nonprobability samples, Propensity score adjustment, Tree-based inverse propensity-weighted estimator, Weight smoothing
Scheduled
Invited Session Data Analysis and Social Science. Data integration probabilistic, non-probabilistic surveys and big data
June 7, 2022 3:30 PM
A11
