An Application of Propensity Modeling: Comparing Unweighted and Weighted Logistic Regression Models for Nonresponse Adjustments

Using logistic regression models to predict the probability that a unit will respond is one method for adjusting for survey nonresponse. The inverse of the propensity score can be the weight adjustment factor. This method can make use of more predictive variables than in the weighting class method. Having used this method for two previous rounds of a large physician survey, this paper describes the results from the most recent round, round four. The logistic regression models used to estimate the propensity score were unweighted in round four, and the independent variables were expanded in round four to include design variables, basic sampling weights, and higher-order interactions. The predictive power of the propensity models was substantially improved over previous rounds, but also presented some interesting issues. The more effective models produced more extreme adjustment factors. In this paper, we evaluate the impact of using weights as covariates, rather than for weighting the models, on the adjustment factors and various measures of predictive power and goodness of fit.