Why a Lung-MAP Result Needs Its Substudy Context
Lung-MAP organized trials around molecularly selected groups of patients with advanced squamous non-small-cell lung cancer. A result from one substudy is therefore not simply a result about “lung cancer” or even about all patients enrolled in Lung-MAP. It concerns the people who met that substudy’s molecular and clinical eligibility criteria, received its assigned treatments, and were followed for its specified outcomes.
The practical task is to read three features together: patient selection, the control treatment, and the reported outcome. These features determine what a comparison estimates. They also explain why a treatment result in one molecular subgroup cannot, by itself, establish benefit or lack of benefit in another.
Start With the Patients Who Could Enter
In a biomarker-selected substudy, eligibility is not just “has lung cancer.” It depends on the disease setting and on the molecular rule used to assign or match patients to a treatment question. In Lung-MAP, different substudies investigated different molecular alterations. For example, S1400A evaluated the FGFR inhibitor AZD4547 in an FGFR-altered group, while S1400B evaluated taselisib in a group selected for PI3K-pathway alterations. These were different treatment hypotheses in differently selected populations.
Selection may involve several layers: histology, disease stage, prior treatment, fitness for the proposed therapy, availability of a usable tumor sample, and a molecular result meeting the protocol definition. A mutation, amplification, or other alteration is not interchangeable with every other finding in the same pathway. The precise assay and eligibility rule matter: a patient classified as eligible under one rule may not qualify under another.
This has two consequences. First, a trial estimate applies most directly to patients who resemble those enrolled and meet the same selection criteria. Second, a biomarker can be associated with prognosis—outcome regardless of treatment—as well as with treatment response. A favorable outcome in a selected group does not establish that the biomarker predicts a treatment benefit. To support a predictive claim, evidence must address whether the treatment effect differs by biomarker status, typically through an appropriate treatment-by-biomarker comparison.
Read the Control and Outcome Before Comparing Results
A randomized comparison estimates the effect of assignment to one treatment rather than another within the trial’s eligible population, subject to the design and analysis assumptions. It does not automatically estimate the effect relative to every possible standard treatment, nor does it answer whether the experimental treatment is better than a control used in a different period or population.
Some Lung-MAP targeted-treatment questions used docetaxel as a control. In S1400B, taselisib was compared with docetaxel in the PI3K-pathway-selected population; the reported progression-free survival results did not establish superiority of taselisib. This is evidence about that comparison in that selected group. It is not a direct estimate for the FGFR-altered S1400A group, which studied a different drug and molecular selection. In addition, as lung cancer care changed, the relevance of a control could change. A control that was suitable when a study began may not represent the same treatment context as a later control.
Outcomes also answer different questions. Objective response rate is the proportion of patients whose tumors meet specified response criteria. Progression-free survival concerns time until progression or death, according to the protocol’s definition. Overall survival concerns time until death from any cause. A higher response rate does not guarantee longer progression-free or overall survival. Differences in follow-up, censoring, and subsequent treatment can also affect how endpoints are interpreted.
| What to check | Why it changes the interpretation |
|---|---|
| Molecular eligibility rule | Defines which patients the result most directly concerns. |
| Control treatment | Defines the comparison; a treatment effect is relative to this control. |
| Outcome and time point | Determines whether the claim concerns response, progression, or survival. |
| Study design and uncertainty | Indicates how strong a causal comparison is and how precisely it is estimated. |
| Care context | Helps determine whether the control and treatment comparison remain relevant. |
Worked Examples: Interpreting Substudy Evidence
Worked Example: A Response Difference With Wide Uncertainty
Hypothetical teaching data, not Lung-MAP results. Suppose a biomarker-selected randomized substudy observes responses in 9 of 40 patients assigned to an experimental drug and 8 of 42 assigned to its control. Estimate the response-rate difference and interpret it.
The estimated experimental and control response rates are \(9/40=0.225\) and \(8/42=0.1905\). Define the difference as experimental minus control:
The estimated difference is 3.45 percentage points in favor of the experimental drug. For an approximate Wald interval, the standard error is
The approximate 95% interval is \(0.0345\pm1.96(0.0896)\), or \((-0.141, 0.210)\). In percentage points, the data are compatible with an experimental-minus-control response difference from about 14.1 points lower to 21.0 points higher. This interval includes zero, so the data do not establish a response-rate difference. The estimate alone is not evidence that the experimental treatment is superior. The approximation also relies on independent patient outcomes and adequate sample sizes for the normal approximation; with small counts, a method suited to sparse binomial data may be preferable.
Worked Example: A Progression-Free Survival Estimate
Hypothetical teaching data, not Lung-MAP results. Suppose a randomized substudy reports a progression-free survival hazard ratio of 0.78 for experimental treatment versus control, with a 95% confidence interval of 0.56 to 1.08. Assume the estimate comes from a prespecified time-to-event analysis with suitable handling of censoring.
The hazard ratio compares the estimated event hazards over follow-up under the model used for analysis. A value below 1 favors the experimental treatment for the event as defined (progression or death). Here, 0.78 corresponds to an estimated 22% lower hazard, since \(1-0.78=0.22\). It does not mean 22% fewer patients progressed, nor does it mean each patient’s progression time increased by 22%.
The interval runs from 0.56 to 1.08 and includes 1, the null value for a hazard ratio. Thus, the estimate is compatible with a substantial hazard reduction and also with a modest hazard increase. The data do not establish a difference at the conventional two-sided 5% level based on this interval. The conclusion must remain specific to the studied molecularly eligible population and the control used. Before interpreting the hazard ratio as a single summary over time, check whether the proportional-hazards assumption is plausible; markedly crossing or diverging survival curves may make one hazard ratio inadequate.
Worked Example: Why Two Substudy Results Are Not a Head-to-Head Comparison
Hypothetical teaching data, not Lung-MAP results. Substudy A enrolls patients with molecular alteration A and reports a 30% response rate on its experimental treatment. Substudy B enrolls patients with alteration B and reports a 20% response rate on a different experimental treatment. Each has its own control arm, and the studies were conducted in different periods.
The subtraction \(30\%-20\%=10\) percentage points is arithmetically correct, but it is not a randomized estimate of the difference between the two experimental drugs. Patients in A and B were not randomized between those drugs. The molecular groups may differ in prognosis, prior treatments, eligibility, and other features. The controls and care contexts may differ as well.
The defensible interpretation is that each response rate describes its own study population and treatment setting. To compare the drugs causally would require a design and assumptions supporting that comparison—for example, a direct randomized comparison in a population eligible for both treatments, or a carefully justified indirect comparison with adequate adjustment and compatible trials. Similar-looking percentages do not supply those conditions.
A Short Justification: Why Selection Changes Generalization
Let \(S=A\) denote eligibility for a subgroup A substudy, and let \(Y(1)\) and \(Y(0)\) denote the outcome that a patient would have under experimental treatment and control, respectively. A randomized trial in subgroup A can estimate the average treatment contrast for patients with \(S=A\):
The corresponding contrast for subgroup B is \(E\{Y(1)-Y(0)\mid S=B\}\). These two quantities need not be equal: the groups may have different biology or other characteristics that modify treatment effect. Even if both studies estimate their own contrast without bias, that fact alone does not show equality between them. Generalizing the A result to B therefore requires additional evidence or assumptions, not merely a shared cancer type or platform name.
This reasoning also clarifies the role of randomization. Within a substudy, random assignment supports a fair comparison of treatment groups on average. It does not randomize patients into molecular subgroups, and it does not make the patients in separate substudies exchangeable. Randomization protects the within-study contrast; it does not erase differences between studies.
Diagnostics, Alternatives, and Limits
A careful reading can be organized as a sequence of checks. This is not a substitute for evaluating the full protocol and report, but it catches common interpretive errors.
Record the histology, disease setting, prior-treatment requirements, molecular assay, and exact eligibility definition.
Name the experimental and control treatments, whether assignment was randomized, and the analysis population.
Distinguish response from time-to-event outcomes; note the effect measure, confidence interval, follow-up, and number of events.
Look for small samples, early stopping, incomplete follow-up, changes in standard care, and whether the result addresses the claim being made.
For time-to-event results, inspect the number of events and the survival curves as well as the hazard ratio. Heavy censoring, few events, or evidence that hazards vary over time can make a single summary less informative. For response results, check the response criteria, evaluable population, and uncertainty interval. Across substudies, compare assay definitions and eligibility before comparing outcomes. Differences in observed response rates can reflect selection or measurement as well as treatment activity.
The main alternative to comparing separate subgroup reports is a direct randomized comparison when both treatments are appropriate for the same eligible patients. A second option is a formal indirect comparison, but it requires strong conditions: sufficiently similar populations, compatible outcome definitions and follow-up, and measured factors that permit credible adjustment. Simply placing two published estimates side by side is not such an analysis.
Lung-MAP results have additional limits typical of molecularly selected studies. A narrowly defined group may yield a small sample and imprecise estimates. Screening and assay availability affect who can be enrolled, so trial participants may not represent every patient with the disease or even every patient with the alteration. A negative or inconclusive result in one subgroup does not prove that all targeted treatments fail; a positive result in one subgroup does not establish activity in other subgroups. Conclusions should be no broader than the design and evidence permit.
Common Mistake
Wrong reasoning: “The experimental treatment had a higher response rate in the FGFR-selected substudy than in the PI3K-pathway substudy, so it must be the better treatment for Lung-MAP patients.”
Correction: The two results concern different molecularly selected populations and may involve different drugs, controls, eligibility rules, and study periods. They are not a randomized head-to-head comparison. State what each substudy found in its own setting, then require direct or well-supported comparative evidence before ranking treatments across subgroups.
Check Your Understanding
Answer each question, then compare your reasoning with the solutions below.
- Why does a randomized treatment comparison in one molecularly selected Lung-MAP substudy not automatically estimate the treatment effect in another selected subgroup?
- A trial reports an objective response rate of 25% in the experimental arm and 20% in the control arm. What is the estimated response-rate difference, and what does it not establish by itself?
- A hazard ratio for progression or death is 1.00. What does this value indicate in the comparison, and what is its null value?
- Two substudies use different controls and report different overall-survival estimates. Name two reasons their estimates cannot be compared as though patients had been randomized between the experimental treatments.
Solutions. (1) Randomization supports the comparison within the enrolled subgroup; it does not make separate molecular groups exchangeable or establish equal treatment effects across them. (2) The estimated difference is \(25\%-20\%=5\) percentage points in favor of the experimental arm. Without uncertainty and design context, it does not establish a real treatment benefit or generalize beyond the studied population. (3) A hazard ratio of 1.00 indicates equal estimated hazards under the fitted comparison; 1 is the null value. (4) The groups may differ in molecular eligibility, prognosis, prior treatment, or care period, and the control treatments may differ. Neither comparison randomizes patients between the two experimental treatments.