Dr. Eric V. Slud, Professor at University of Maryland, and US Census Bureau will give a lecture entitled "Design-based Weight Adjustment for Low-Response-Rate Surveys."

Abstract:

This talk describes new theory and methodology for weighting two very low-response Tracking Surveys about attitudes to the US Census conducted for the US Census Bureau from September 2019 through June 2020. One was a Telephone (RDD) probability survey and the other a Web-based nonprobability opt-in survey. In both, the raw demographics of respondents differed markedly from the general US population, and the base or uniform initial weights moved dramatically when calibrated (raked) to national demographic proportions. Well-established theory (Deville and Sarndal, 1992 JASA) shows that when inverse-inclusion-probability survey weights are calibrated (without any nonresponse) to correct totals in a probability survey, the design-based estimates are consistent and have reduced variances, and the weights move very little. New theory was developed to address the Tracking Survey case, where `true' weight ratios over the initial weights satisfy a parametric (e.g., loglinear) model, with some results also for the case where no such model holds. New variance estimates (differing from those obtainable from Generalized Regression survey estimates as part of Deville & Sarndal's theory) were developed and are illustrated for the Tracking Surveys. The take-away point of this talk is that weight calibration and variance estimation in surveys with either low response or non-probability designs is unavoidably model-based.