1 Introduction

Predictive approaches to heterogeneity of treatment effects (HTE) aim at the development of models predicting either individualized effects or which of two (or more) treatments is better for an individual [1]. In prior work, we divided such methods in three broader categories based on the reference class used for defining patient similarity when making individualized predictions or recommendations [2]. Risk-modeling approaches use prediction of baseline risk as the reference; treatment effect modeling approaches also model treatment-covariate interactions, in addition to risk factors; optimal treatment regime approaches focus on developing treatment assignment rules and rely heavily on modeling treatment effect modifiers.

Risk-modeling approaches to predictive HTE analyses provide a viable option in the absence of well-established treatment effect modifier [3,4]. In simulations, modeling treatment-covariate interactions, often led to miscalibrated predictions of absolute benefit, contrary to risk-based methods, despite their weaker discrimination of benefit in the presence of true effect modifiers [5]. Most often, risk-modeling approaches are carried out in two steps: first a risk prediction model is developed externally or internally on the entire RCT population, “blinded” to treatment; then the RCT population is stratified using this prediction model to evaluate risk-based treatment effect variation [6]. This approach identified substantial absolute treatment effect differences between low-risk and high-risk patients in a re-analysis of 32 large trials [7]. However, even though estimates at the risk subgroup level may be accurate, these estimates may not apply to individual patients.

In the current simulation study, we aim to summarize and compare different risk-based models for predicting treatment effects. We simulate different relations between baseline risk and treatment effects and also consider potential harms of treatment. We illustrate the different models by a case study of predicting individualized effects of treatment for acute myocardial infarction (MI) in a large RCT.