Multiple Imputation for Data That Are Missing Not at Random: Extending the Fully Conditional Specification Procedure

• Activity Number: 605
• Type: Contributed
• Date/Time: Wednesday, August 12, 2015 : 2:00 PM to 3:50 PM
• Sponsor: Survey Research Methods Section
• Author(s): Finbarr P. Leacy* and Ian R. White
• Companies: MRC Biostatistics Unit/University of Cambridge and MRC Biostatistics Unit
• Keywords: missing data ; sensitivity analysis ; multiple imputation

Abstract

While Fully Conditional Specification (FCS) or chained equations is a popular procedure for performing multiple imputation under the Missing at Random (MAR) assumption, it has lacked a principled approach for handling data that are Missing Not at Random (MNAR). By considering the full conditional distributions implied by a specified joint distribution for the substantive variables and their missingness indicators, we demonstrate how the standard FCS procedure can be extended to impute under general MNAR missingness mechanisms. Assuming that this joint distribution follows a loglinear or conditional Gaussian model, we examine the finite sample properties of the extended procedure in a range of simulation studies. We place our results in context with recent work on the theoretical properties of the standard FCS procedure. Finally, we use the procedure to assess the robustness of two recently published analyses of data from the Avon Longitudinal Study of Parents and Children to departures from MAR: these analyses explored exposure to a parental suicide attempt before age 11 and intelligence measured at age 8 as potential risk factors for self-harming behaviour at age 16, respectively.