Challenges of subgroup analyses in multinational clinical trials: Experiences from the MERIT-HF trial

Background International placebo-controlled survival trials (Metoprolol Controlled-Release Randomised Intervention Trial in Heart Failure [MERIT-HF], Cardiac Insufficiency Bisoprolol Study [CIBIS-II], and Carvedilol Prospective Randomized Cumulative Survival trial [COPERNICUS]) evaluating the effects of P-blockade in patients with heart failure have all demonstrated highly significant positive effects on total mortality as well as total mortality plus all-cause hospitalization. Also, the analysis of the US Carvedilol Program indicated an effect on these end points. Although none of these trials are large enough to provide definitive results in any particular subgroup, it is natural for physicians to examine the consistency of results across various subgroups or risk groups. Our purpose was to examine both predefined and post hoc subgroups in the MERIT-HF trial to provide guidance as to whether any subgroup is of increased risk, despite an overall strongly positive effect, and to discuss the difficulties and limitations in conducting such subgroup analyses. Methods The study was conducted at 313 clinical sites in 16 randomization regions across 14 countries, with a total of 3991 patients. Total mortality (first primary end point) and total mortality plus all-cause hospitalization (second primary end point) were analyzed on a time to first event. The first secondary end point was total mortality plus hospitalization for heart failure. Results Overall, MERIT-HF demonstrated a hazard ratio of 0.66 for total mortality and 0.81 for mortality plus all-cause hospitalization. The hazard ratio of the first secondary end point of mortality plus hospitalization for heart failure was 0.69. The results were remarkably consistent for both primary outcomes and the first secondary outcome across all predefined subgroups as well as for nearly all post hoc subgroups. The results of the post hoc US subgroup showed a mortality hazard ratio of 1.05. However, the US results regarding both the second primary combined outcome of total mortality plus all-cause hospitalization and of the first secondary combined outcome of total mortality plus heart failure hospitalization were in concordance with the overall results of MERIT-HF. Tests of country by treatment interaction (14 countries) revealed a nonsignificant P value of .22 for total mortality. The mortality hazard ratio for US patients in New York Heart Association (NYHA) class III/IV was 0.80, and it was 2.24 for patients in NYHA class II, which is not consistent with causality by biologic gradient. We have not been able to identify any confounding factor in baseline characteristics, baseline treatment, or treatment during follow-up that could account for any treatment by country interaction. Thus we attribute the US subgroup mortality hazard ratio to be due to chance. Conclusions Just as we must be extremely cautious in overinterpreting positive effects in subgroups, even those that are predefined, we must also be cautious in focusing on subgroups with an apparent neutral or negative trend. We should examine subgroups to obtain a general sense of consistency, which is clearly the case in MERIT-HF. We should expect some variation of the treatment effect around the overall estimate as we examine a large number of subgroups because of small sample size in subgroups and chance. Thus the best estimate of the treatment effect on total mortality for any subgroup is the estimate of the hazard ratio for the overall trial.