Predictive And Prognostic Biomarkers In Treatment Trials

Why the Distinction Matters

A biomarker can identify patients with different expected outcomes without identifying patients who benefit differently from treatment. These are different questions. A prognostic biomarker is associated with outcome; a predictive biomarker indicates that the treatment effect differs according to marker status. Confusing the two can lead a trial team to narrow enrollment to a group with a better prognosis even when treatment works equally well across groups.

The distinction is especially important when a trial uses biomarker information to consider subgroup selection or enrichment. A higher response rate in a marker-positive group is not, by itself, evidence that the treatment works better for marker-positive participants. The relevant evidence compares treatment with control within each marker group, then compares those treatment effects.

Definition: A prognostic biomarker is associated with the outcome under a specified treatment context, such as the control arm, or across treatment arms. A predictive biomarker identifies variation in the treatment effect across biomarker values. Prediction is therefore about a treatment-by-marker difference in effects, not merely a marker-outcome association.

Define the Contrast Before Interpreting the Marker

For a simple example, let \(T=1\) denote assignment to the experimental treatment and \(T=0\) assignment to control. Let \(M=1\) denote biomarker-positive status and \(M=0\) biomarker-negative status. Let \(Y\) be a binary outcome, with \(Y=1\) representing a favorable response. Write \(\mu_{tm}=P(Y=1\mid T=t,M=m)\). On the risk-difference scale, the treatment effect in marker group \(m\) is

\Delta_m=\mu_{1m}-\mu_{0m}.

On this scale, the marker is predictive when the treatment effects differ: \(\Delta_1-\Delta_0\ne 0\), beyond the degree of difference plausibly due to sampling variation. A marker may be prognostic if outcomes differ between marker groups in the control arm, for example if \(\mu_{01}\ne\mu_{00}\). The marker can be both prognostic and predictive, or prognostic without being predictive.

The biomarker itself is generally not randomized. Thus, a difference between marker groups does not establish that the marker causes the outcome difference. In a properly randomized trial, however, randomization can support a causal treatment comparison within each prespecified marker group, provided the trial’s other assumptions hold.

Assumptions for interpreting subgroup treatment contrasts: Marker status is measured before treatment and defined consistently; treatment is randomized within, or remains randomized across, marker groups; both treatments have participants in each group; outcomes and follow-up are measured comparably; and treatment assignment corresponds to the treatment received as defined by the trial estimand. In addition, the analysis must account for missing outcomes and any selection or multiplicity created by examining multiple subgroups.

The effect scale must be stated. Equal risk differences do not imply equal risk ratios, and equal risk ratios do not imply equal risk differences. Consequently, whether a marker appears predictive can depend on the scale used to define treatment effect. The scale should be clinically meaningful and prespecified, rather than chosen after inspecting which scale gives the most favorable subgroup result.

Worked Examples

Worked Example: Prognostic, but Not Predictive on the Risk-Difference Scale

Suppose a hypothetical randomized trial reports these response proportions: in the control arm, 40% of marker-positive participants and 20% of marker-negative participants respond; in the treatment arm, the corresponding proportions are 60% and 40%. Assume the marker was measured at baseline and treatment was randomized within both marker groups.

In the marker-positive group, the estimated treatment effect is

\hat\Delta_1=0.60-0.40=0.20.

This is a 20-percentage-point increase in response. In the marker-negative group,

\hat\Delta_0=0.40-0.20=0.20.

The difference between treatment effects is \(0.20-0.20=0\), so these data show no treatment-effect difference on the risk-difference scale. Yet the marker is prognostic in these data: the response proportion is 20 percentage points higher for marker-positive participants in control (\(0.40-0.20=0.20\)) and also 20 points higher in treatment (\(0.60-0.40=0.20\)). Better outcomes for marker-positive participants do not mean they receive greater treatment benefit on this scale.

The scale qualification matters. The response risk ratios are \(0.60/0.40=1.50\) for marker-positive participants and \(0.40/0.20=2.00\) for marker-negative participants. Those ratios differ. This does not contradict the equal risk differences: it illustrates that effect modification is scale-dependent.

Worked Example: Comparing Treatment Effects Across Marker Groups

In a hypothetical trial, each treatment-by-marker cell contains 100 participants. There are 30 responses among 100 control participants in each marker group; among treated participants, there are 65 responses in the positive group and 35 in the negative group. Assume independent participants, complete outcome ascertainment, and randomized treatment assignment within each marker group.

The estimated risk differences are

\hat\Delta_1=\frac{65}{100}-\frac{30}{100}=0.35, \qquad \hat\Delta_0=\frac{35}{100}-\frac{30}{100}=0.05.

The estimated difference between treatment effects is \(0.35-0.05=0.30\), or 30 percentage points. This is evidence of a potentially predictive marker on the risk-difference scale, subject to uncertainty and the analysis plan.

For a rough uncertainty assessment, the standard error of a difference between independent sample proportions is estimated by the square root of the sum of their estimated variances. For the positive group:

SE(\hat\Delta_1) = \sqrt{\frac{0.65(1-0.65)}{100}+\frac{0.30(1-0.30)}{100}} = \sqrt{0.004375} = 0.06614.

For the negative group, the calculation is

SE(\hat\Delta_0) = \sqrt{\frac{0.35(1-0.35)}{100}+\frac{0.30(1-0.30)}{100}} = \sqrt{0.004375} = 0.06614.

Because the four treatment-by-marker cells contain independent participants, the estimated standard error of the difference between treatment effects is

SE(\hat\Delta_1-\hat\Delta_0) = \sqrt{0.004375+0.004375} = 0.09354.

The approximate Wald statistic is \(0.30/0.09354=3.21\). A two-sided normal approximation gives a p-value of about 0.0013. This is evidence against equal risk differences in this hypothetical setting, not the probability that the marker is predictive. A confirmatory conclusion would also depend on prespecification, the number of subgroup comparisons, and the trial’s inferential plan.

Worked Example: A Marker Can Be Both Prognostic and Predictive

Consider hypothetical response proportions from another randomized trial: control response is 50% for marker-positive participants and 20% for marker-negative participants; treatment response is 80% and 30%, respectively. Assume marker status is measured at baseline, both groups are represented in both arms, and outcomes are assessed in the same way.

The treatment effects on the risk-difference scale are

\hat\Delta_1=0.80-0.50=0.30, \qquad \hat\Delta_0=0.30-0.20=0.10.

The effect contrast is \(0.30-0.10=0.20\), a 20-percentage-point difference. Thus, the results suggest predictive value on this scale. The marker is also prognostic: control response differs by \(0.50-0.20=0.30\), and treatment response differs by \(0.80-0.30=0.50\). The terms describe distinct features of the data; they are not mutually exclusive labels.

Why the Treatment-by-Marker Contrast Is Necessary

The key comparison can be written as a difference in differences. Let \(\mu_{tm}\) be the response probability in treatment group \(t\) and marker group \(m\). Then

\Delta_1-\Delta_0 = (\mu_{11}-\mu_{01})-(\mu_{10}-\mu_{00}).

This subtracts the control-to-treatment response difference for marker-negative participants from the corresponding difference for marker-positive participants. If the marker shifts outcomes similarly in both treatment arms, that shift does not by itself create a treatment-effect difference on this scale. In the first worked example, the marker-positive group has better response in each arm, but the treatment-control contrast is 20 percentage points in both groups.

A comparison of response proportions between marker groups in the treatment arm alone does not make this adjustment. It mixes the marker’s association with outcome and any difference in treatment benefit. The control arm supplies the reference needed to distinguish those patterns.

Diagnostics, Design Choices, and Common Pitfalls

Before interpreting a subgroup pattern, examine whether the data can support it. Check the number of participants and outcomes in every treatment-by-marker cell, confirm that marker status was measured before randomization or treatment, and look for missing or unevaluable marker results. Sparse cells produce imprecise subgroup effects; a large observed contrast from a small cell may be unstable. Also verify that the endpoint, follow-up time, and analysis population are consistent across the groups being compared.

A forest plot of treatment effects can help display estimates and uncertainty by subgroup, but it does not replace a direct comparison of effects. One subgroup having a statistically significant result and another not having one does not establish that their treatment effects differ. The relevant question is whether the estimated treatment-effect contrast is itself sufficiently large and precise under the prespecified analysis. Searching many markers, cutoffs, outcomes, or time points also increases the chance of finding an apparently striking contrast by chance.

Stratified randomization and biomarker enrichment answer different design needs. Stratifying randomization can help balance treatment assignment across marker groups; it does not show that treatment effects differ by marker. Enrichment restricts enrollment to a selected population and can be useful when the scientific and clinical rationale supports that population, but a marker’s prognostic association alone is not evidence that restricting enrollment will improve the treatment contrast. An all-comer trial with prespecified subgroup comparisons can retain information about both marker groups, though it must be designed to estimate the relevant contrasts with adequate precision.

In an adaptive setting, interim subgroup patterns can be tempting grounds for changing eligibility. A noisy early contrast may be exaggerated if the most favorable subgroup is selected for further study. Any adaptation based on marker-specific treatment effects therefore needs prespecified rules and appropriate control of the inferential consequences of selection. The earlier material on subgroup selection and treatment selection is relevant to that design concern; the core scientific distinction remains the same: prognosis describes outcome differences, while prediction concerns treatment-effect differences.

Common mistake: “Marker-positive participants respond more often on treatment, so the marker predicts treatment benefit.” This reasoning compares marker groups within one arm and cannot separate prognosis from treatment-effect modification. The correction is to estimate treatment versus control within each marker group and compare those treatment effects on a stated scale.

Check Your Understanding

For each item, distinguish marker-outcome association from treatment-effect modification. Solutions follow the questions.

In a hypothetical trial, control response is 25% for marker-positive participants and 10% for marker-negative participants. Treatment response is 45% and 30%, respectively. Are the results prognostic, predictive, or both on the risk-difference scale?
Why is a significant treatment result in one marker subgroup and a nonsignificant result in another insufficient to establish a predictive biomarker?
What does randomization support when treatment is compared within a baseline marker group, and what does it not establish about the marker itself?
On the risk-difference scale, treatment effects are 15 percentage points for marker-positive participants and 5 points for marker-negative participants. Calculate the treatment-effect contrast and interpret it.

Solutions. (1) Both. The treatment effects are \(0.45-0.25=0.20\) and \(0.30-0.10=0.20\), so there is no risk-difference effect modification in these proportions. The marker is prognostic because control response differs by \(0.25-0.10=0.15\). (2) Significance in each subgroup depends on its estimate and standard error; the difference in significance is not a test of the difference between treatment effects. The treatment-effect contrast must be assessed directly. (3) With valid randomization and the stated trial assumptions, it supports a causal treatment comparison within the marker group. It does not make marker status randomized or show that the marker causes outcome differences. (4) The contrast is \(0.15-0.05=0.10\), or 10 percentage points. The estimated treatment benefit is 10 points larger in the marker-positive group on the risk-difference scale; uncertainty and the design’s multiplicity plan still matter.