A while ago, I discussed several simple imputation methods in LOCF, BOCF, WOCF, and MVTF. Recently, I noticed another simple imputation method Placebo Mean Imputation (PMI). This simple imputation method of PMI seems to be used Only for the purpose of sensitivity analyses, not for the primary analysis. It is also true that this approach is mainly used in certain therapeutic areas such as analgesic drug (pain medication) and anti-bacterial drug.
In FDA's Clinical Review document for a chronic pain medication in 2011, the Placebo mean imputation is described:
"Placebo mean imputation (PMI): the missing pain measurements for each day after discontinuation were replaced by the mean of all available pain intensity scores for all placebo-treated patients who completed treatment. Therefore if a patient discontinued treatment or recorded their last pain score at Week 8 of the Maintenance Period, the pain intensity score at Week 12 was imputed using the Week 12 mean pain intensity score for all placebo-treated patients who completed treatment. Also a placebo missing pain score at some time-point was imputed by the observed placebo group mean pain intensity at the same time-point"
In FDA's Statistical Review for the same indication, FDA statistician assessed that PMI is not an appropriate approach
"Since the estimated treatment effect is only influenced by the data from patients completing the study, the PMI method is similar to an analysis of completers. Analyzing completers is
problematic since the outcome of patients completing the study may not represent the outcome of patients not completing the study. In the placebo group, patients completing the study are likely to be the less severely afflicted patients; while in the NUCYNTA group, patients completing the study are likely to be the more severely afflicted patients. As a result, the PMI method assigned good scores from the placebo completers to patients dropping out due to adverse events in the treatment group. Based on these reasons, I conclude that the PMI method is not appropriate."
In a briefing document for FDA anti-infective drug advisory committee meeting for a cystic fibrosis drug in 2012, PMI approach was used by FDA statistical reviewer:
“…In this sensitivity analysis, missing values were imputed using the placebo group mean of -0.57% (i.e. the minimum of 0 and the least favorable group mean) as performed in the Reviewer’s primary analysis and other Reviewer analyses.”
The use of these simple imputation approaches is mainly driven by the perception that these approaches provide conservative estimate of the treatment effect (people have shown that this is not always true). This fits into the Intention-to-treat principle to be conservative in estimating the treatment effect in superiority studies. While it is ok to perform the sensitivity analyses using various simple imputation approaches such as LOCF, BOCF, WOCF, and PMI, these imputation methods should not be used as the primary analysis.
- Gual et al (2013) A randomised, double-blind, placebo-controlled, efﬁcacy study of nalmefene, as-needed use, in patients with alcohol dependence
- Carpenter and Kenward (2007) Missing data in randomised controlled trials — a practical guide
- The National Academy Press (2010) The Prevention and Treatment of Missing Data in Clinical Trials