Composite variable bias: causal analysis of weight outcomes

Researchers often use composite variables (e.g., BMI and change scores). By combining multiple variables (e.g., height and weight or follow-up weight and baseline weight) into a single variable it becomes challenging to untangle the causal roles of each component variable. Composite variable bias-an issue previously identified for exposure variables that may yield misleading causal inferences-is illustrated as a similar concern for composite outcomes. We explain how this occurs for composite weight outcomes: BMI, 'weight change', their combination 'BMI change', and variations involving relative change.

Data from the National Child Development Study (NCDS) cohort surveys (n = 9223) were analysed to estimate the causal effect of ethnicity, sex, economic status, malaise score, and baseline height/weight at age 23 on weight-related outcomes at age 33. The analyses were informed by a directed acyclic graph (DAG) to demonstrate the extent of composite variable bias for various weight outcomes.

Estimated causal effects differed across different weight outcomes. The analyses of follow-up BMI, 'weight change', 'BMI change', or relative change in body size yielded results that could lead to potentially different inferences for an intervention.

This is the first study to illustrate that causal estimates on composite weight outcomes vary and can lead to potentially misleading inferences. It is recommended that only follow-up weight be analysed while conditioning on baseline weight for meaningful estimates. How conditioning on baseline weight is implemented depends on whether baseline weight precedes or follows the exposure of interest. For the former, conditioning on baseline weight may be achieved by inclusion in the regression model or via a propensity score. For the latter, alternative strategies are necessary to model the joint effects of the exposure and baseline weight-the choice of strategy can be informed by a DAG.