Statistical Graphics of Pearson Residuals in Survey Logistic Regression Diagnosis

Survey data logistic regression analysis, as computationally available in SAS SURVEYLOGISTIC procedure, has been widely conducted in survey research practice. A set of diagnostic statistics in the procedure, borrowed from the logistic regression in generalized linear models, is used for model assessment. However, for survey data, the statistical underpinnings of these statistics may need to be reexamined. In practice, we have observed irregular behaviors of these diagnostic statistics, which make their established statistical criterion suspect. Their naïve use can be misleading. This presentation reports our use of Pearson residuals normality graphs as graphical diagnostic statistics, to assess survey data logistic modeling, as in a recent NASS study of sampling frame coverage. Statistical graphics summarization may provide broader scope and more elaborated information than would be available through analytical summarization. The statistical graphs of Pearson residuals showed their diagnostic ability, and careful reading of the residual graphs may reveal delicate diagnostic information on modeling effects. We illustrate the statistical graphical modeling process with our analysis.