OBJECTIVES Using randomized clinical trials (RCTs) for clinical decision-making necessitates making decisions for individuals based on average treatment effects.  While many assume important patient variation in treatment effects, identifying patients most likely to benefit is problematic.  Stratifying patients by their risk of the primary outcome was proposed as a method to identify high versus low benefit patients. METHODS From publically available sources, we identified 13 large RCTs with greater than ~1000 enrollees and overall statistically significant results.  We derived Cox or logistic regression models using established risk factors blinded to treatment assignment and stratified the patient population into quartiles of risk for the outcome.  Treatment effect within each risk quartile was estimated on relative and absolute scales.  Heterogeneity of treatment effect (HTE) was evaluated statistically by testing for an interaction between treatment and the linear predictor of risk, and by comparing hazard (or odds) ratios and absolute risk reduction in the extreme risk quartiles. RESULTS Among 19 unique treatment comparisons analyzed, there was no apparent relationship between baseline risk and the hazard (or odds) ratios across trials; only 1 of 19 analyses had a significant interaction between treatment and baseline risk on the proportional scale. The difference in the log hazard ratio between the extreme risk quartiles ranged from -0.89 to 0.60 (median=0.03; inter-quartile range (IQR)=-0.4-0.2).   However, absolute risk reduction was generally higher in high risk strata, ranging from -1.4 to 18.3% (median=4.6%; IQR=0.8-6.1%) in quartile one and from 0.8 to 35.0% (median=11.5%; IQR=3.3-19.8%) in quartile four.  The difference in the absolute risk reduction between the extreme risk quartile ranged from -3.2 to 28.3% (median=7.7%; IQR=0.3-11.3).  CONCLUSIONS Clinically significant HTE is common even in phase 3 “efficacy” trials on the absolute scale. A multivariate risk stratified approach to subgroup analysis is feasible and often clinically informative when assessing treatment efficacy.