Post hoc subgroups in clinical trials: Anathema or analytics?

Abstract

Background:

There is currently much interest in generating more individualized estimates of treatment effects. However, traditional statistical methods are not well suited to this task. Post hoc subgroup analyses of clinical trials are fraught with methodological problems. We suggest that the alternative research paradigm of predictive analytics, widely used in many business contexts, can be adapted to help.

Methods:

We compare the statistical and analytics perspectives and suggest that predictive modeling should often replace subgroup analysis. We then introduce a new approach, cadit modeling, that can be useful to identify and test individualized causal effects.

Results:

The cadit technique is particularly useful in the context of selecting from among a large number of potential predictors. We describe a new variable-selection algorithm that has been applied in conjunction with cadit. The cadit approach is illustrated through a reanalysis of data from the Randomized Aldactone Evaluation Study trial, which studied the efficacy of spironolactone in heart-failure patients. The trial was successful, but a serious adverse effect (hyperkalemia) was subsequently discovered. Our reanalysis suggests that it may be possible to predict the degree of hyperkalemia based on a logistic model and to identify a subgroup in which the effect is negligible.

Conclusion:

Cadit modeling is a promising alternative to subgroup analyses. Cadit regression is relatively straightforward to implement, generates results that are easy to present and explain, and can mesh straightforwardly with many variable-selection algorithms.