Background Understanding which factors predict or cause individuals’ dropout from weight-loss programmes can provide insight into possible adaptations that minimise dropout. Prior research indicating ‘predictors’ of dropout does not allow causal interpretation, which limits its utility to inform possible adaptations.
Thus, this study examines dropout in weight-loss programmes through the re-analysis of secondary datasets to: (1) explore if, and by how much, insights differ between prediction and causal inference approaches; and (2) identify associated methodological issues in both prediction and causal inference.
Methods We used data from eight systematically identified studies published between 2013–2020 on factors related to dropout from weight-loss programmes. The eight primary analyses attempted to identify variables that predict or are correlated with whether a participant dropped out of a weight-loss intervention, without necessarily recognising that this might conflate the two distinct data science tasks of prediction and causal inference. The same datasets were re-analysed for prediction and causal inference, using approaches appropriate for each task. To identify predictors, a single model containing a full set of variables was developed for each dataset. To estimate causal effects, a directed acyclic graph (DAG) was coproduced with authors from each original study, and multiple models were generated depending on the exposures of interest. The results (and corresponding conclusions) were then compared.
Results In the primary analyses, age and body mass index (BMI) (with inconsistency in the direction of the relationship) were often identified as predictors of dropout, while in the secondary analyses, factors causally associated with dropout were older age, women, higher baseline weight, and more challenging weight-loss targets. None of the primary studies could inform causal interpretation due to methodological issues such as Table 2 Fallacy, reversal paradox, and challenges related to analysing composite variables. With the exception of two studies, all causal analyses suffered positivity violations.
Conclusion Weight-loss programme design cannot be improved by knowing which factors predict dropout; in fact, this may cause more harm than good. There were some differences in the coefficients generated by the prediction and causal inference models (even sign reversal), indicating that the potential for misinterpreting causal relationships could be severe. To identify causal factors related to dropout, a causal inference evaluation, informed by an appropriate DAG, is preferred, but existing study datasets may not meet the criteria for robust inference. Our findings suggest that study designs appropriate for causal analysis should be considered when determining which weight-loss programme adaptations are needed to reduce dropout.
Statistics from Altmetric.com
If you wish to reuse any or all of this article please use the link below which will take you to the Copyright Clearance Center’s RightsLink service. You will be able to get a quick price and instant permission to reuse the content in many different ways.