Comparing Weighting Methods for Nonresponse in the HKPSSD Survey and the BCSPS

Abstract

Using data collected from the Hong Kong Panel Study of Social Dynamics (HKPSSD) Survey and the Beijing College Students Panel Survey (BCSPS), the current study examined four commonly used weighting adjustments: the logistic regression model, the Response Propensity Stratification (RPS) method, the Generalized Exponential Model (GEM), and the Random Forests Model (RFM) for longitudinal nonresponse. Our results indicated that the logistic regression model, the RPS method, and the GEM produced comparable results, with the GEM and the RPS methods slightly outperforming in many instances, while the results obtained by the RFM were not as reliable as those generated by the other methods. Moreover, high bias and MSE changes were associated with variables or categories that had low occurrences for all the methods considered. In particular, for categorical variables, there were considerable variations on bias across categories while the overall difference compared to the baseline was not significant. In addition, relative bias and MSE changes were correlated to the specification of the nonresponse model, the baseline weights, as well as the intervals between the baseline, and the wave for weighting adjustment.