When could one use reference based imputation for missing data, and what kind should one use?

Michael O'Kelly, Senior Strategic Biostatistics Director, Centre for Statistics in Drug Development, Innovation, Quintiles

Standard approaches to missing data in clinical trials tend to account for the missing data using information from elsewhere in the data set – partly from the same treatment arm, and partly from information shared across the arms. In these approaches, no one treatment has a special role. All such analyses rely heavily on assumptions about the missingness mechanism that cannot be validated from the data itself. Exploration of the sensitivity to these assumptions is required by the regulatory agencies.

 In contrast we introduce two different approaches to exploring a wider set of scenarios, loosely called reference-based imputation, where one arm, the reference arm, has a special role. Both approaches are implemented as SAS (r) macros available at the DIA web page of www.missingdata.org.uk. We distinguish a number of variants of reference-based assumptions.  We also briefly discuss delta-adjustment or “tilting” of assumptions about missing data. We show how in combination these approaches can provide an alternative perspective aimed at supporting the primary analysis.

We point to future extensions including categorical and other non-Normal data scenarios.