A COMPARISON OF TWO METHODS TO ADJUST WEIGHTS FOR NON-RESPONSE: PROPENSITY MODELING AND WEIGHTING CLASS ADJUSTMENTS

Abstract

The most common method used to adjust sampling weights for non-response involves forming weighting classes or cells of homogeneous sample members. Within each cell, the weights of the respondents are inflated to account for the non- respondents. Some problems inherent in this method are cells with too few respondents, adjustment or inflation factors that are too high, and potentially large differences in these adjustment factors from one cell to the next. While there are ways to deal with some of these problems, they generally involve the risk of increasing the mean square error, either through bias or in the variance of the weights. A relatively new approach involves developing logistic regression models to predict response, using a potentially much broader set of predictive variables than can be used in the weighting class methodology. The inverse of the response propensity resulting from the application of such a model can then be used as the adjustment factor to the weights. We applied both of these methods when computing weights for round two of the Community Tracking Study Household Survey. This paper explores the differences resulting from these two methods.