Brady West, Ph.D.
Collegiate Research Professor
Survey Methodology Program
Institute for Social Research
University of Michigan-Ann Arbor
Abstract. Methods for integrating probability and non-probability samples have grown in popularity in recent years, in part because it may be infeasible to measure a variable of scientific interest in a large probability sample, but feasible to do so in a non-probability sample. One such method is mass imputation, where a variable that is not measured in the probability sample is fully imputed in that sample using a model fitted to the non-probability sample. A fundamental assumption of this method is that the imputation model fitted to the non-probability sample is transportable to the probability sample, meaning that the same regression function for the variable of interest holds for both samples. If the selection mechanism for the non-probability sample is non-ignorable, given the variable of interest, then there may be significant bias in the imputation model coefficients, resulting in poor imputations and subsequent bias in the mass imputation estimator. This paper leverages recent work on novel measures of selection bias for regression coefficients to propose an adjustment of the imputation model coefficients for selection bias prior to imputation. Via a simulation study and a case study, we demonstrate the ability of this method to improve estimates produced using mass imputation procedures.