Identification of threshold for large (dramatic) effects that would obviate randomized trials is not possible

Abstract

Objective: To analyze distribution of "dramatic", large treatment effects.

Study design & setting: Pareto distribution modeling of previously reported cohorts of 3,486 randomized trials (RCTs) that enrolled 1,532,459 patients and 730 non-randomized studies (NRS) enrolling 1,650,658 patients.

Results: We calculated the Pareto α parameter, which determines the tail of the distribution for various starting points of distribution [odds ratio_min (OR_min)]. In default analysis using all data at OR_min ≥1, Pareto distribution fit well to the treatment effects of RCTs favoring the new treatments (P = 0.21, Kolmogorov-Smirnov test) with best α = 2.32. For NRS, Pareto fit for OR_min ≥2 with best α = 1.91. For RCTs, theoretical 99th percentile OR was 32.7. The actual 99th percentile OR was 25; which converted into relative risk (RR) = 7.1. The maximum observed effect size was OR = 121 (RR = 11.45). For NRS, theoretical 99th percentile was OR = 315. The actual 99th percentile OR was 294 (RR = 13). The maximum observed effect size was OR = 1473 (RR = 66).

Conclusions: The effects sizes observed in RCTs and NRS considerably overlap. Large effects are rare and there is no clear threshold for dramatic effects that would obviate future RCTs.

Keywords: FDA; large or dramatic treatment effects; methodology; observational studies; randomized trials; statistical modeling.