Simple sensitivity methods to validly and appropriately address missing cessation outcomes

Missing cessation outcomes are a universal challenge in tobacco cessation studies. Complete case analysis, where cases missing these outcomes are omitted, is well known to be biased. The common strategy applied in tobacco cessation studies to address the missing outcomes is to impute a value of “currently using tobacco” for every missing cessation outcome. The rationale behind this strategy is an assumed conservativeness to this approach. We clearly and simply demonstrate that this assumption is in error, even when response rates do not differ by intervention group.

Generally, we do not use statistical methods in order to assure conservativeness in estimation but rather we use those methods that address and reduce bias and error in estimation. For the situation considered here, there are sophisticated statistical methodologies such as pattern-mixture models, selection models, and multiple imputation methods that attempt to do just this by incorporating the mechanism leading to missing outcomes into the estimation process. These methods are not widely applied in tobacco cessation trials, perhaps because application of these methodologies tends to be complex. However, for randomized intervention studies there are pattern-mixture models that are quite simple to apply. We discuss these simple pattern-mixture models, and related sensitivity methods, and demonstrate their application using existing results from smoking cessation trials.