Treatment of missing data in follow-up studies of randomised controlled trials: A systematic review of the literature

Background/aims: After completion of a randomised controlled trial, an extended follow-up period may be initiated to learn about longer term impacts of the intervention. Since extended follow-up studies often involve additional eligibility restrictions and consent processes for participation, and a longer duration of follow-up entails a greater risk of participant attrition, missing data can be a considerable threat in this setting. As a potential source of bias, it is critical that missing data are appropriately handled in the statistical analysis, yet little is known about the treatment of missing data in extended follow-up studies. The aims of this review were to summarise the extent of missing data in extended follow-up studies and the use of statistical approaches to address this potentially serious problem.

Methods: We performed a systematic literature search in PubMed to identify extended follow-up studies published from January to June 2015. Studies were eligible for inclusion if the original randomised controlled trial results were also published and if the main objective of extended follow-up was to compare the original randomised groups. We recorded information on the extent of missing data and the approach used to treat missing data in the statistical analysis of the primary outcome of the extended follow-up study.

Results: Of the 81 studies included in the review, 36 (44%) reported additional eligibility restrictions and 24 (30%) consent processes for entry into extended follow-up. Data were collected at a median of 7 years after randomisation. Excluding 28 studies with a time to event primary outcome, 51/53 studies (96%) reported missing data on the primary outcome. The median percentage of randomised participants with complete data on the primary outcome was just 66% in these studies. The most common statistical approach to address missing data was complete case analysis (51% of studies), while likelihood-based analyses were also well represented (25%). Sensitivity analyses around the missing data mechanism were rarely performed (25% of studies), and when they were, they often involved unrealistic assumptions about the mechanism.

Conclusion: Despite missing data being a serious problem in extended follow-up studies, statistical approaches to addressing missing data were often inadequate. We recommend researchers clearly specify all sources of missing data in follow-up studies and use statistical methods that are valid under a plausible assumption about the missing data mechanism. Sensitivity analyses should also be undertaken to assess the robustness of findings to assumptions about the missing data mechanism.