Propensity Models Versus Weighting Cell Approaches to Nonresponse Adjustment: A Methodological Comparison
Siegel, P. H., Chromy, J. R., & Copello, E. A. (2005, May). Propensity Models Versus Weighting Cell Approaches to Nonresponse Adjustment: A Methodological Comparison. Presented at American Association for Public Opinion Research 60th Annual Conference, Miami Beach, FL.
Statistical adjustment of nonresponse is a deep and pervasiveissue for sample surveys. Contemporary statistical methods offer twobroad classes of approach to nonresponse adjustment. One is the use ofa traditional weighting cell approach. More recently, responsepropensity modeling, using, typically, logistic regression, has beendeveloped as a further approach to nonresponse adjustment.Additionally, RTI’s General Exponential Model (GEM) is a generalizationof raking type weight adjustments, and in addition to nonresponseadjustment it can optionally include features such aspoststratification and weight trimming. We use the data from theEducation Longitudinal Study of 2002 (ELS:2002) to compare the resultsof the weighting class method, raking, a logistic regression propensitymodel, and GEM. We focus on nonresponse adjustment but also look atextreme weight adjustment and poststratification. For theone-dimensional case where each unit is in one unique cell, it can beshown that the three methods will produce the same results. Expandingto many variables and multiple dimensions, marginal totals, variances,and weight distributions can be compared, i.e., we also conduct alimited nonresponse bias analysis and examine the effects each methodhas on reducing nonresponse bias.