The sensitivity analyses that we did to explore assumptions about missing data produced results that were consistent with the main analyses (supplementary tables B to E).
In this longitudinal setting with monotone dropout, if data were MAR, the likelihood-based analysis that ignores the missing data mechanism produces valid inference; hence, MAR was termed as ignorable dropout whereas MNAR was termed as non-ignorable as the explicit model for missing data mechanism was needed.
A sensitivity analysis estimating models 2 and 3 without missing data imputation produced similar results.
Inclusion of individuals with missing data, however, produced similar patterns of association to those presented.
Sensitivity analysis using multiple imputation of missing data also produced similar findings (see online supplementary appendices 5 and 6).
Essentially, MI is a means for representing uncertainty in missing data, by producing a distribution of plausible values for a missing variable in a record, given the values of that record's non-missing covariates.
Despite its good empirical performance, DACTAL provides no statistical guarantees, and it is likely that statistical methods based on models of evolution that treat indels as events, rather than as missing data, would produce more accurate trees (Warnow, 2012).
In intent-to-treat analyses with baseline weight carried forward for missing data, IBI produced significantly larger mean (SD) weight losses than IDEA at 3 months (5.5 kg [4.4] vs. 1.3 kg [2.1]) and 6 months (5.4 kg [5.6] vs. 1.3 kg [4.1]) (P < 0.001).
There did not appear to be any systematic pattern in the missing data that would produce bias in the analysis of this subset of decedents.
Work history information available for grouping manufacturing workers or locations was also incomplete, but we did not detect any systematic pattern in the missing data that would produce biased estimates of effect.
