Propensity Models vs. Weighting Cell Approaches to Non-response Adjustment: A Methodological Comparison
Chromy, J. R., Siegel, P. H., & Copello, E. A. (2005, August). Propensity Models vs. Weighting Cell Approaches to Non-response Adjustment: A Methodological Comparison. Presented at Joint Statistical Meetings, Minneapolis, MN.
Statistical adjustment of nonresponse is a deep and pervasive issue for sample surveys. Contemporary statistical methods offer two broad classes of approach to nonresponse adjustment. One is the use of a traditional weighting cell approach. More recently, response propensity modeling, typically using logistic regression, has been developed as another approach to nonresponse adjustment. Additionally, RTIs General Exponential Model (GEM) is a generalization of weight adjustment methods; in addition to nonresponse adjustment, it can optionally include features such as poststratification and weight trimming. We compare the results of the weighting class method, raking, a logistic regression propensity model, and GEM, focusing not only on nonresponse adjustment but on extreme weight adjustment and poststratification. For the one-dimensional case, it can be shown that the three methods will produce the same results. For multiple control dimensions, marginal totals, variances, and weight distributions are compared. We also conduct a limited nonresponse bias analysis and examine the effects each method has on reducing nonresponse bias.