Imperfect Data Analysis

Abstract

Trial planning should prevent errors, but imperfections are unavoidable. Transparency is necessary (listing and description of imperfections as well as rules to deal with imperfect data should be given in the trial report). Between purism and laxism, realistic rules are often acceptable.

Decisions on data should be independent of the results: written rules, and independent and blind decisions are necessary. Intent-to-treat analysis avoids biased post hoc exclusions, but a few exceptions are acceptable in extreme cases. Sensitivity analysis enables testing the robustness of results to some decisions on data. Rules are proposed for the most frequent situations: missing data, outliers, cases of protocol violations, treatment discontinuations, poor compliance, and losses to follow-up.