Logistic Regression With Incomplete Covariate Data in Complex Survey Sampling

Abstract

Weighted survey data with missing data for some covariates presents a substantial challenge for analysis. We addressed this problem by using a reweighting technique in a logistic regression model to estimate parameters. Each survey weight was adjusted by the inverse of the probability that the possibly missing covariate was observed. The reweighted estimating equations procedure was compared with a complete case analysis (after discarding any subjects with missing data) in a simulation study to assess bias reduction. The method was also applied to data obtained from a national health survey (National Health and Nutritional Examination Survey or NHANES). Adjusting the sampling weights by the inverse probability of being completely observed appears to be effective in accounting for missing data and reducing the bias of the complete case estimate of the regression coefficients.