Unmasking heterogeneous treatment effects in clinical trials: a location-scale quantile meta-analysis framework (IPD-QMA)

Abstract

Background. Meta-analysis of randomized controlled trials is the basis of evidence-based medicine, but it usually aggregates treatment effects as a mean difference under a 'location-shift' assumption — that treatment shifts the outcome distribution uniformly for all patients. In psychiatry, oncology and critical care, interventions can instead create a 'location-scale shift', increasing outcome variance without changing the mean, so standard methods may fail to detect efficacy and discard treatments that benefit high-risk subgroups while harming low-risk ones.

Methods. We introduce Individual Participant Data Quantile Meta-Analysis (IPD-QMA), a two-stage framework that uses quantile regression to estimate treatment effects across the severity distribution (τ = 0.1 to 0.9) within studies and pools the coefficients with a random-effects model. We compared IPD-QMA against standard meta-analysis (mean difference) and the log variance ratio (lnVR) using 2,000 Monte Carlo simulations on both Normal and skewed (exponential) data, and developed a 'slope test' (Wald statistic) to test effect homogeneity.

Results. In 'hidden-utility' scenarios where treatment increases variance (scale × 3) but not the mean, standard meta-analysis rejected the null at close to the nominal 5% level (4.2% on Normal and 6.4% on skewed data) — effectively no power to detect the effect — whereas IPD-QMA achieved 100% power to detect benefit at the 90th percentile. The lnVR detected heterogeneity with high power but lacked directionality; the IPD-QMA quantile profile revealed that the treatment was harmful for mild cases but beneficial for severe cases.

Conclusions. IPD-QMA is a diagnostic tool for precision medicine: whereas lnVR screens for heterogeneity, IPD-QMA characterizes its shape, preventing the loss of potentially life-saving interventions for high-risk populations. An open-source implementation is provided.