OBJECTIVES: The Look AHEAD trial examined the effect of an intensive lifestyle intervention (ILI) for weight loss compared to diabetes support and education (DSE) in adults with type 2 diabetes and overweight or obesity. ILI produced larger reductions in BMI and HbA_1c than DSE. Due to lack of treatment effect in the primary outcome of cardiovascular morbidity or mortality the intervention was discontinued after 12 years. This research investigates how the effect of the ILI varies across participants and if it is possible to predict participant characteristics likely to result in higher treatment response.

METHODS: We use causal forests to test for heterogeneity in outcomes from the Look AHEAD trial including short- to medium-term HbA_1c and BMI, and long-term cardiovascular morbidity and mortality, prescription drug spending, and insulin usage. We present covariates which serve as primary drivers of heterogeneity and differences in estimated treatment effect by data-generated subgroups.

RESULTS: We find heterogeneity in treatment effects of ILI on BMI (P<0.001) and HbA_1c (P<0.001) in Year 1. Low baseline levels of urine albumin and triglycerides were associated with a higher treatment response for BMI and lower treatment response for HbA_1c. Surprisingly, we found participants with large reductions in Year 1 BMI were less likely to have large reductions in HbA_1c and vice versa.

Heterogeneity observed in Year 1 did not persist over time. Being a high responder in Year 1 is not associated with being a high responder in Year 4. Additionally, no heterogeneity was found in Year 4 BMI or HbA_1c, end-of-trial cardiovascular morbidity and mortality, or in post-intervention pharmaceutical spending or insulin use.

CONCLUSIONS: Although early heterogeneity in treatment response was identified, it was found not to persist. Overall, our results demonstrate the power of causal tree methods to uncover detailed patterns of heterogeneous effects not detectable in pre-specified subgroup analyses.