Bad stats: A regular series exploring slip-ups, snafus and salutary lessons from the world of statistics

Convergence should be non-negotiable The missing-data problem nobody is talking about Longitudinal studies that track individuals over time face a near-inevitable problem: missing data. As researchers, we design these studies assuming data collection will proceed smoothly. It seldom does. Teams with substantial funding may be able to mitigate the problem through dedicated follow-up staff, participant incentives, and sophisticated tracking systems. But even these measures do not guarantee retention. Participants relocate. They change their contact details or encounter life events that studies are simply not equipped to capture. The result is a loss of follow-up information that matters. And these gaps are more than a technical inconvenience. They influence the evidence that informs scientific and policy decisions. So, what happens when the statistical fixes meant to address them end up introducing new problems of their own? A common response to missing data is to drop incomplete records from the analysis entirely. But discarding participants is rarely ideal. It shrinks sample sizes and risks distorting results, particularly when those who drop out differ from those who remain. Other approaches attempt to fill gaps with single values, which, if done incautiously, trades one problem for another by treating uncertain estimates as certainties and ignoring the variability that missingness introduces. Multiple imputation was developed to avoid both pitfalls. In simple terms, the method works by generating several plausible versions of the missing data based on patterns in the observed information. Each complete data set is analysed separately, and the results are then pooled in a way that preserves the uncertainty introduced by the missingness. The method is now routine as researchers across fields reach for it as a matter of course. Yet far fewer pause to ask whether the algorithm behind it has done what it is supposed to do: converge. Software defaults share some of the blame. Many packages ship with iteration settings optimised for speed over stability. And researchers, pressed for time or unfamiliar with the internals, tend to accept those defaults without further scrutiny.