Sensitivity analysis for incomplete continuous data

Abstract

Models for missing data are necessarily based on untestable assumptions whose effect on the conclusions are usually assessed via sensitivity analysis. To avoid the usual normality assumption and/or hard-to-interpret sensitivity parameters proposed by many authors for such purposes, we consider a simple approach for estimating means, standard deviations and correlations. We do not make distributional assumptions and adopt a pattern-mixture model parameterization which has easily interpreted sensitivity parameters. We use the so-called estimated ignorance and uncertainty intervals to summarize the results and illustrate the proposal with a practical example. We present results for both the univariate and the multivariate cases.