Shared Trial Data Analysis for a
Treatment Effect Heterogeneity Question

RESEARCH QUESTION

What the Secondary Analysis Set Out to Ask

This shared trial data analysis example follows an outcomes research group that wants to know whether the absolute benefit of a cardiovascular drug differs between patients at lower and at higher baseline risk. The data come from one completed sponsor trial, shared through an independent clinical trial data-sharing platform, and the estimates are the hazard ratio and the absolute risk difference at three years, each with its 95% confidence interval.

Platforms of this kind typically ask for a research proposal with a statistical analysis plan, require a qualified statistician on the team, and pass the proposal to an independent panel. After approval, every team member accepts the data use agreement. The need therefore comes in two stages: an analysis plan for the proposal, then the analysis itself, run in the platform's workspace.

STUDY DESIGN AND DATA

One Completed Trial in a Remote Workspace

One completed randomized trial of about 9,000 participants with a median follow-up of around three years, analyzed as randomized. The dataset arrives with a data dictionary and stays inside a remote workspace that has R but no open internet connection, so the packages the analysis needs are listed before the workspace opens. Only outputs that pass the platform's review leave it. Some baseline covariates are missing, and death from a non-cardiovascular cause can end follow-up before the outcome occurs.

METHOD AND ASSUMPTIONS

Why Risk Groups Replace One-Variable Subgroups

Subgroups defined one variable at a time ask a question that fits few patients, who differ on several risk factors at once, and they multiply comparisons while each has little power. The plan follows the PATH Statement instead. A multivariable risk model for the outcome is developed without using treatment arm, and its calibration and discrimination are checked, with a bootstrap correction for optimism because it is fitted and evaluated in the same sample. Patients are then grouped by predicted risk, and the hazard ratio and the three-year absolute risk difference are estimated in each quarter of risk.

Whether the hazard ratio itself varies with predicted risk is tested as a treatment-by-risk interaction and reported with its interval. The two scales answer different questions, because an absolute benefit can grow with baseline risk even when the hazard ratio is the same in every group, so the absolute difference is the headline. Absolute risks are cumulative incidences that account for competing deaths. Missing baseline covariates are imputed several times with the outcome in the imputation model and the results combined. Conventional one-variable subgroups appear only as secondary analyses, with the number of comparisons and their credibility stated.

OUTPUT AND INTERPRETATION

How the Results Are Read

The output has the calibration plot and the discrimination of the risk model, the distribution of predicted risk, and for each quarter of risk the hazard ratio and the three-year absolute risk difference with 95% confidence intervals, together with the interaction estimate and its interval. The number of participants and events sits beside each estimate, so that the precision of every quarter is visible.

Every table and figure is prepared for the platform's output review before anything leaves the workspace, with small counts handled as the agreement requires. The write-up describes where the benefit appears larger or smaller and how uncertain that statement is. It also says that a pattern found in one trial is a hypothesis for confirmation in other data, not a basis for changing treatment by itself.

REPORTING

What the Researcher Received

A written report structured to the PATH Statement for predictive approaches to heterogeneity of treatment effects, with the source trial described using the relevant CONSORT items and the analysis plan written in the form the data request needs. OnlineSPSS works as a member of the research team within the data use agreement the team holds. The data request, the interpretation and the conclusions remain with the research team.