Abstract

This thesis explores subgroup analyses within randomized trials, using cardiovascular interventions as a focal point.

Reviewing 67 large trials published between 1980 and 1997, we found that most reported on multiple single-factor subgroups without pre-specification or statistical tests for subgroup-treatment interactions.

Instead of single-factor subgroup analyses, decision-makers can infer absolute benefits by applying the relative effect size from the overall trial to baseline event rates in subgroups. We tested this approach in the SOLVD prevention (4228 patients) and treatment (2569 patients) trials. We confirmed a published treatment interaction with ejection fraction (p = 0.004), but successfully abolished that interaction when patients were divided into multivariate tertiles of baseline prognostic risk.

This lack of significant variation could represent a beta error. We found no published method to perform a post-hoc power test for the most popular interaction test, the Breslow-Day statistic. Adapting a formula suggested by Agresti, we demonstrated that the power of the interaction test in SOLVD was about 75%. Our approach should be helpful to other investigators evaluating non-significant subgroup-treatment interactions.

We then surveyed clinicians using hypothetical trial data to assess their interpretations of subgroup analyses. Among 435 respondents considering a neutral subgroup within a positive trial, 44% would withhold treatment from subgroup-type patients, notwithstanding the risk of beta-error. Similar interpretive differences emerged in other scenarios. Academics and those engaged or trained in research were more cautious than others in interpreting subgroup data. These findings highlight the need for more education and clearer guidelines for clinicians using subgroup data.

Finally, we used reperfusion therapy as a test case for actual application of a proven treatment to patient subgroups in practice. Among 550 ideal candidates with myocardial infarction, the probability of treatment correlated inversely with baseline 6 month mortality rates. The paradoxical relationship held for cardiac (p=0.0002), noncardiac (p=0.0374), and socio-demographic variables (p=0.0031). This 'treatment-yield' paradox reinforces the value of measuring treatment effects across prognostic subgroups to avoid harmful under-utilization of therapies among those most likely to benefit.

We conclude by proposing some general guidelines to improve the design, conduct, and interpretation of subgroup analyses.