Reading Outcomes Across FOCUS4 Treatment Modules

Why FOCUS4 Results Need to Be Read Module by Module

FOCUS4 used a shared screening and randomization platform, but its treatment modules did not answer one interchangeable question. Patients entered modules according to molecular findings and were assigned to different interventions against different comparators. A result from one module therefore describes a particular treatment contrast in a particular population—not automatically an effect that applies to every patient enrolled in FOCUS4.

This matters when comparing reported results. A treatment may improve progression-free survival (PFS) in a molecularly selected group, while a different treatment shows little evidence of benefit in another group. Those findings are not contradictory: they concern different interventions and populations. Nor can they be combined just because the modules share a platform or use the same endpoint name.

Three Modules, Three Treatment Questions

FOCUS4 enrolled patients with advanced colorectal cancer whose disease had not progressed after initial chemotherapy, then used molecular screening to guide assignment to treatment modules. The important comparison is not simply “FOCUS4 treatment versus control.” The module determines who is compared, what is given, and what the result can support.

The N group is best understood as the non-stratified module population, not as a claim that every participant had identical molecular biology. C and D selected patients by different molecular features. Their interventions also differed: aspirin is not capecitabine or adavosertib, and placebo is not the same treatment condition as active monitoring. These details are part of the estimand—the precise treatment effect the comparison is intended to estimate.

What Makes an Effect Comparable?

For a time-to-event endpoint such as PFS, a common summary is a hazard ratio (HR). In a proportional-hazards model, the HR compares the instantaneous event rates between assigned groups over follow-up, conditional on the model. An HR below 1 favors the intervention when the event is progression or death. This interpretation depends on how the endpoint and analysis are defined; an HR is not a ratio of median PFS times.

A shared endpoint label is helpful but insufficient for pooling. The modules must also have sufficiently comparable definitions of progression and death, assessment schedules, follow-up, censoring rules, and analysis populations. Moreover, an overall summary needs a clinical target. For example, it might target an average treatment effect across a specified distribution of module populations. Without defining that target, “the FOCUS4 effect” has no single clear meaning.

One way to see the assumptions behind a simple summary is to consider independent module estimates \(\hat\theta_j=\log(\widehat{HR}_j)\) with estimated variances \(v_j\). Under a fixed-effect model, each module is assumed to estimate one common log-HR \(\theta\), with sampling error but no real between-module difference. The inverse-variance estimate follows by minimizing the weighted squared deviations:

Differentiating and setting the result to zero gives \(\sum_j(\hat\theta_j-\theta)/v_j=0\). Rearranging yields the weighted estimate and its model-based standard error:

Exponentiating \(\hat\theta_{\mathrm{FE}}\) gives a pooled HR. But the calculation does not prove the common-effect assumption. In FOCUS4, different molecular populations, drugs, and control conditions make that assumption a substantive scientific claim, not a routine consequence of using the same platform.

Population, Comparator, and Endpoint Must Travel Together

A treatment effect can vary across populations. Molecular selection may identify patients with a different prognosis, a different response to a drug, or both. The control condition also matters: active monitoring estimates the effect of offering treatment instead of monitoring, while placebo control estimates the effect of the active drug relative to its placebo condition. Background care and treatment burden are part of that contrast.

Endpoints require equally careful reading. PFS is commonly used in these maintenance comparisons, but identical wording does not ensure identical measurement or analysis. PFS depends on the time origin, the rules for defining progression, the timing of assessments, and how incomplete follow-up is handled. Overall survival (OS), response, and adverse events answer different questions and cannot be substituted for PFS. A result on one endpoint does not establish benefit on all others.

Diagnostics, Alternatives, and Limits

Before making a cross-module claim, check whether it is going wrong in any of these ways:

• Population drift: The eligibility criteria or molecular profiles differ, but the conclusion is worded as if it applied to one common patient group.
• Comparator substitution: Active monitoring and placebo are treated as equivalent without considering what participants received and what contrast randomization identifies.
• Endpoint shorthand: A shared label such as PFS is taken to guarantee the same time origin, assessments, event definition, or censoring.
• Significance-by-comparison: One module has a statistically significant result and another does not, so a difference in treatment effects is asserted. A difference between significance labels is not itself a statistical test of the difference.
• Over-reading pooled estimates: A narrow interval from a calculation is taken as evidence that the populations and treatment effects are genuinely interchangeable.

The strongest alternative to indirect comparison is a randomized head-to-head comparison in the population and setting of interest. If a direct comparison is unavailable, a carefully specified synthesis may still be useful, but it needs a target population, harmonized outcome definitions, and assumptions about comparability and effect variation. Depending on the question, investigators may report module-specific effects separately, use a model allowing between-module variation, or standardize effects to a defined patient mix. These approaches do not repair missing comparability by themselves; their assumptions must be stated and examined.

FOCUS4’s modular structure was useful for evaluating different strategies in biologically defined groups. Its limitation for cross-module interpretation is the same feature: different modules were designed to answer different questions. The reported evidence should therefore remain attached to its population, intervention, comparator, and endpoint unless a separate analysis justifies broader transport or synthesis.