Causal Forests in Clinical Research: Estimating Heterogeneous Treatment Effects Without Overclaiming Personalization

Randomized trials and well-designed observational studies usually begin with an average treatment effect: the difference in outcomes under two treatment strategies for a defined population. That average is often the right primary estimand. It can also hide clinically relevant variation. A therapy may help one subgroup more than another, show no meaningful difference in a second subgroup, or cause harm in a third. The challenge is to investigate such variation without converting every exploratory pattern into a personalized treatment rule.

Causal forests are one way to study heterogeneous treatment effects. They combine the logic of causal inference with tree-based machine learning. Instead of asking only whether treatment works on average, the analyst asks whether the expected treatment contrast differs across patients with different baseline covariate profiles. The flexibility is useful, but it also creates a boundary: an algorithm can discover a pattern in a development sample without proving that the pattern is causal, stable, calibrated, or clinically actionable.

This distinction matters in precision-medicine research. A causal forest does not turn observational data into a randomized trial, and it does not turn a conditional average treatment effect into an immutable individual treatment effect. The method should be used inside a transparent causal design with a prespecified treatment contrast, outcome, time horizon, target population, covariate set, uncertainty strategy, and validation plan.

Why average treatment effects are not the whole story

An average treatment effect summarizes the expected contrast between treatment strategies across the study population. It can support a population-level decision, but it may not describe every clinically meaningful pattern. Heterogeneity may arise from baseline severity, comorbidity, disease subtype, prior treatment, biological markers, age, prognosis, adherence, or interactions among several covariates. The presence of plausible effect modification does not establish that the observed variation is real; it defines a question that requires careful analysis.

Traditional subgroup analyses often examine one variable at a time or test a small number of prespecified treatment–covariate interactions. These analyses can be clinically interpretable, especially when supported by prior evidence and adequate sample size. They can also miss complex relationships or generate false positives when many subgroups are explored. Regression-based interaction models remain important comparators because they make the effect-modification structure explicit and allow researchers to encode clinical knowledge directly.

Causal forests offer a data-adaptive alternative. They can search across multiple baseline covariates and nonlinear partitions without requiring the analyst to specify every interaction in advance. That benefit comes with a need for stronger safeguards: the causal estimand must be clear, the treatment assignment mechanism must support identification, the algorithm must be tuned without leaking outcome information, and the learned effect pattern must be assessed in data that were not used to discover it.

What a causal forest estimates

For a binary treatment, a causal forest typically targets a conditional average treatment effect: the expected difference between potential outcomes under treatment and control for patients with a given set of observed baseline covariates. The notation may differ by software and outcome type, but the interpretation depends on the data structure and assumptions. The estimate is conditional on the covariates supplied to the model and on the causal framework that defines treatment, outcome, time zero, follow-up, and censoring.

The word “conditional” is central. A CATE is not a direct measurement of an individual’s unobservable treatment response. It is an estimated average effect for a covariate-defined group or local region of the covariate space. Two patients with similar measured covariates may still have different unmeasured characteristics, and a patient’s observed response cannot reveal both potential outcomes. The estimate should therefore be described as a model-based treatment-effect quantity, not as a biological certainty for a particular person.

Causal identification comes before machine learning. In a randomized trial, treatment assignment can support identification under the trial’s protocol and analysis population, subject to adherence, missing outcomes, censoring, and other issues. In an observational cohort, the analyst must justify assumptions such as consistency, exchangeability conditional on measured covariates, positivity, and appropriate handling of time zero and follow-up. A flexible forest cannot compensate for unmeasured confounding or a treatment definition that changes during follow-up.

Honesty and cross-fitting are not the same

Applied causal-forest work often uses the terms honesty and cross-fitting. They are related but not interchangeable. Honest estimation separates the data used to determine tree structure or adaptive partitions from the data used to estimate treatment effects within those structures. The aim is to reduce adaptive bias by preventing the same outcomes from both searching for effect patterns and estimating their magnitude.

Cross-fitting divides data into folds, fits nuisance components or effect estimates in one part, and evaluates them in another part, rotating the roles across folds. It can reduce overfitting in nuisance estimation and make flexible procedures more compatible with valid inference. The exact implementation depends on the estimator and software. A manuscript should state what was split, what was learned in each split, how folds were created, whether clusters or sites were kept together, and how the final effect estimates were combined.

Randomly splitting individual patients may be inappropriate when patients are clustered within hospitals, families, physicians, trials, or repeated records. If the deployment target is a new hospital, a site-level split may be more informative than a patient-level split. If the causal estimand concerns a later time period, temporal separation may be needed. The split strategy should follow the transportability question, not only the software default.

Calibration of treatment-effect estimates

Prediction models are commonly evaluated with calibration: whether predicted risks agree with observed risks. Causal-forest analyses need an analogous question for treatment effects. If the model assigns patients to strata of predicted benefit, do the observed treatment contrasts in those strata align with the predicted contrasts? Calibration can be assessed through prespecified strata, smooth calibration approaches, or other design-appropriate summaries. The analysis should make clear whether calibration is being assessed for an outcome prediction, a treatment effect, or a treatment-selection rule.

Calibration is especially important because a causal forest can detect heterogeneity while still produce poorly calibrated effect estimates. In the type 2 diabetes application by Venkatasubramaniam and colleagues, causal forest and penalized regression both characterized treatment-effect heterogeneity, but the external validation results did not show universal superiority for the forest; calibration was better for penalized regression in that application. The lesson is not that regression always wins. The lesson is that flexible treatment-effect models must be compared and validated against clinically interpretable alternatives.

Calibration should be interpreted alongside uncertainty. A predicted effect in a small covariate-defined subgroup may be unstable even when the overall model appears well behaved. Confidence intervals, prediction intervals, bootstrap or other resampling strategies, subgroup sample sizes, overlap diagnostics, and the number of events or continuous-outcome information should be reported. A ranked list of patients by predicted benefit is not sufficient evidence for changing treatment allocation.

Compare causal forests with clinical regression models

A causal forest should be compared with a prespecified regression model that reflects the clinical question. The comparator may include treatment–covariate interactions selected from prior knowledge, restricted nonlinear terms, clinically meaningful subgroup categories, or a penalized regression strategy. The comparison should use the same treatment definition, outcome, follow-up, missing-data strategy, validation dataset, and performance target.

Regression has a transparency advantage. The interaction terms, functional forms, and target contrasts are visible in the model specification. A causal forest has a flexibility advantage: it can detect nonlinear and high-order interactions that would be difficult to list manually. Neither advantage is universal. A complex forest may identify a pattern that is not stable; a simple regression may miss meaningful effect modification. The appropriate conclusion depends on the estimand, the data, the validation results, and the clinical consequences of acting on the model.

Variable importance measures should not be interpreted as treatment-effect proof. A covariate can be useful for partitioning effect estimates without being a treatment effect modifier in a mechanistic sense. Importance can also be affected by scale, correlation, missingness, measurement quality, and algorithmic choices. A clinically persuasive interpretation should connect the model output to prespecified effect contrasts, uncertainty, calibration, and independent replication.

Clinical applications and the evidence boundary

Causal forests have been applied in randomized-trial and routine-care settings to study treatment-effect heterogeneity. In the type 2 diabetes study, trial data were used for development and primary-care data were used for external evaluation of predicted treatment benefits. This design is useful because it separates treatment-effect discovery from transportability assessment. It also demonstrates why a method should not be presented as validated simply because it finds a clinically plausible subgroup.

Tutorial work in psychiatry and epidemiology provides accessible explanations of how causal forests can explore heterogeneous effects. These applications support the method’s usefulness as an analytic tool, but they do not establish that any forest-derived subgroup should receive a different treatment in practice. Implementation requires a clinical decision rule, a threshold for benefit or harm, treatment risks, patient preferences, feasibility, and prospective evaluation of the resulting policy.

Randomized trial data strengthen causal interpretation, but they do not automatically solve heterogeneity analysis. A trial may have limited sample size for interactions, restricted eligibility, treatment nonadherence, missing outcomes, or a population that differs from routine care. An observational dataset may be larger but more vulnerable to confounding and selection bias. Causal forest analysis should make these trade-offs visible rather than hide them behind a flexible algorithm.

Actionable Steps: Build a defensible causal-forest analysis

Common analytical and reporting failures

The first failure is presenting a forest as a replacement for causal design. The algorithm does not determine whether a treatment comparison is exchangeable or whether the outcome was measured after a coherent time zero. The second is reporting only a variable-importance plot. Importance can help organize hypotheses, but it does not show the direction, magnitude, uncertainty, or clinical relevance of a treatment effect.

The third failure is calling a CATE an individual treatment effect. The estimated quantity is conditional on measured covariates and represents an average over patients with similar information. The fourth is describing honesty and cross-fitting as if they were one procedure. The manuscript should specify the actual split, learning, nuisance estimation, and effect-estimation steps.

The fifth failure is skipping calibration and external validation. A model can rank predicted effects while misestimating their magnitude. The sixth is comparing a flexible forest with an under-specified regression model and then claiming that machine learning discovered treatment heterogeneity that regression could not. A fair comparison requires a clinically reasonable comparator and the same target outcome, validation data, and performance measures.

The seventh failure is turning exploratory subgroups into treatment recommendations. A treatment-selection rule requires a prespecified decision threshold, consideration of treatment harms and benefits, patient preferences, and prospective validation. Without those elements, the appropriate conclusion is that the analysis generated or evaluated a hypothesis about heterogeneous effects.

Evidence Summary Table

Conclusion

Causal forests are useful when a clinical research question concerns variation in treatment effects across measured patient characteristics and the analysis is embedded in a defensible causal design. Their flexibility can reveal nonlinear and high-order interactions that would be difficult to specify manually. That flexibility is not a guarantee of causal validity, calibration, transportability, or clinical usefulness.

A rigorous workflow defines the estimand, audits identification, separates adaptive learning from effect estimation, compares the forest with a clinically justified regression model, and validates predicted effects in new data when possible. CATEs should be reported as model-based conditional estimates, not as direct individual truths. The method can strengthen precision-medicine research when used with transparency and restraint; it should not be used to overclaim personalization.