Editing categorical response errors in repeated self-report surveys: An example from an STD prevention study

Conflicting answers to survey questions, and inconsistent responses to identical repeated questions such as gender, exist in longitudinal self-report surveys. Deterministic data editing techniques correct these errors but subsequent analysis assumes the edit is correct and does not allow for edit error. We propose models to perform multiple edit in direct analogy with multiple imputation. We multiply edit erroneous data under a model and combine the multiply edited data sets using Rubin's rules for combining multiply imputed data sets. This requires a model for the missing correct data given the clearly incorrect responses. We illustrate this process by considering a Bayesian latent variable model for student reports of being born in the US and how that varies as a function of age and ethnicity. We illustrate a conditional probit model of cell probabilities for contingency tables with what should be a structural zero and apply it to model the probability the student knows about and uses a condom distribution program on campus. The motivating data set consists of a four year sample from a longitudinal intervention study on Los Angeles middle- and high-school students.