What a Failed Phase II Trial Can—and Cannot—Tell Us
A Phase II trial that misses its efficacy objective creates a development decision, not an automatic verdict on the drug. The result may reflect insufficient pharmacologic activity, inadequate exposure, poor adherence, an unsuitable population, an insensitive endpoint, or ordinary sampling uncertainty. Distinguishing these explanations matters: each points to a different next step, and some do not justify more development at all.
The randomized treatment comparison remains the primary evidence about the effect of assignment to the tested regimen in the enrolled population. Exposure and adherence data can help explain that result, but exposure is measured after treatment assignment and may be related to prognosis, tolerability, or behavior. An exposure-response association therefore needs more cautious interpretation than the randomized comparison.
The case below is entirely hypothetical. A drug is tested for 12 weeks in 120 participants with a chronic inflammatory disease: 60 are randomized to the drug and 60 to placebo. The primary endpoint is change from baseline in a symptom score, in points; more negative change means greater improvement. A treatment-placebo difference of \(-2.0\) points is the prespecified difference regarded as clinically meaningful. The drug group has mean change \(-5.0\) points and the placebo group \(-4.0\) points. The estimated standard error of the treatment difference is \(1.20\) points.
Start With the Randomized Efficacy Result
Define the treatment effect as the mean change in the drug group minus the mean change in the placebo group. Negative values favor the drug on this endpoint. The trial estimate is:
Under the usual large-sample approximation for the difference in randomized group means, a two-sided 95% confidence interval is the estimate plus or minus \(1.96\) standard errors. This approximation assumes that the standard error is appropriately estimated and that the sampling distribution of the estimate is sufficiently close to normal. In this case:
The interval includes zero, so this result does not establish a statistically detectable treatment difference at the two-sided 0.05 level. It also includes the clinically meaningful value of \(-2.0\) points, as well as values near zero and values favoring placebo. The trial is therefore inconclusive about whether the true effect reaches the clinical threshold; it is not evidence that the effect is exactly zero. The interval is a range of values compatible with the data under the analysis, not a 95% probability statement about the fixed treatment effect.
Worked Example: Quantifying the Phase II Result
Use the hypothetical group means and standard error above to calculate the test statistic and assess the primary result. For this example, use a normal approximation and a two-sided test of \(H_0:\Delta=0\) against \(H_a:\Delta\ne0\), with \(\alpha=0.05\).
Estimate. The treatment-placebo difference is \((-5.0)-(-4.0)=-1.0\) point, favoring the drug.
Test statistic. Divide the estimate by its standard error:
The two-sided normal-approximation p-value is \(2P(Z\le -0.833)\approx 0.405\), where \(Z\sim N(0,1)\). Because \(0.405>0.05\), the trial does not reject \(H_0\). This does not mean there is a 40.5% probability that the null hypothesis is true. It means that, if the true treatment difference were zero and the model assumptions held, a test statistic at least this far from zero in either direction would occur about 40.5% of the time.
Decision-relevant reading. The point estimate is smaller in magnitude than the \(-2.0\)-point clinical target, but the confidence interval includes effects more favorable than that target. Neither “the drug works” nor “the drug does not work” is justified by this result alone.
Use Exposure and Adherence to Update the Model
The trial collected adherence records and estimated average drug concentration over the dosing interval, \(C_{\text{avg}}\), in ng/mL. Among the 60 participants assigned to drug, 24 had mean \(C_{\text{avg}}=2.0\) ng/mL and mean adherence of 62%; the other 36 had mean \(C_{\text{avg}}=6.0\) ng/mL and mean adherence of 94%. These categories describe observed exposure, not randomized exposure groups. The mean exposure across the drug group is \((24(2.0)+36(6.0))/60=4.40\) ng/mL.
A prespecified, covariate-adjusted exposure-response analysis among treated participants estimates the treatment effect relative to a modeled comparator at exposure \(C\) as:
The estimated slope is \(-0.75\) points per ng/mL, with standard error \(0.30\) points per ng/mL. The fitted model assumes a linear relationship over the observed exposure range, adequate adjustment for measured prognostic factors, and a sufficiently appropriate exposure and outcome model. These assumptions are substantive; a good numerical fit alone does not establish them.
The slope quantifies how the model’s estimated treatment effect changes with exposure: each 1 ng/mL higher exposure is associated with a further \(0.75\)-point improvement over the modeled range, conditional on the included covariates. A confidence interval for the slope can be derived by taking the estimate plus or minus a critical value times its standard error. Using an approximate 95% critical value of 2.00:
The interval lies below zero, indicating an inverse exposure-response association in this fitted analysis. It does not, on its own, prove that raising the dose will produce the predicted improvement. Participants with higher exposure may differ in adherence, health status, tolerability, or other factors that also affect the outcome. Covariate adjustment can reduce some measured differences but cannot guarantee removal of unmeasured confounding.
Worked Example: Interpreting Predictions Within the Observed Range
Use the fitted model to calculate its point predictions at the two observed mean exposures, 2.0 and 6.0 ng/mL. Negative effects favor the drug.
At \(C=2.0\) ng/mL:
At \(C=6.0\) ng/mL:
Thus the fitted model predicts little or no benefit at the lower exposure and a point estimate exceeding the \(-2.0\)-point clinical target at the higher exposure. The second prediction is within the observed exposure range, which is preferable to extrapolating beyond it. However, it is still a model-based estimate, not a randomized comparison of participants assigned to 2.0 versus 6.0 ng/mL. The uncertainty in the predicted effect also depends on uncertainty and covariance in the model’s intercept and slope; the slope interval alone is not a confidence interval for either prediction.
From Model Update to Development Decision
The case provides two different kinds of evidence. The randomized comparison estimates the effect of assignment to the tested regimen, including the adherence actually achieved in this trial. The exposure-response analysis suggests that lower exposure may have diluted the average treatment effect and that higher observed exposure could be associated with greater benefit. Taken together, these findings justify investigating exposure optimization, but they do not yet establish that a higher dose will work or be safe.
A disciplined decision separates four possible changes:
- Dose or adherence: Investigate whether a tolerable regimen can reliably achieve the exposure range associated with benefit. Poor adherence may call for adherence support or a more practical regimen; increasing the prescribed dose does not correct missed doses automatically.
- Population: Change the target population only when there is credible evidence that treatment effect differs across a clinically justified subgroup. An apparent exposure subgroup is not, by itself, evidence for a new patient subgroup.
- Endpoint: Change the endpoint only if evidence indicates that the chosen measure was unreliable, insensitive, or poorly aligned with the intended clinical benefit. A nonsignificant result alone is not a sound reason to select a more favorable endpoint after seeing the data.
- Stop development: Stop if plausible, safely attainable exposures are not associated with meaningful benefit, if safety constraints prevent reaching useful exposure, or if the exposure-response result is too fragile to support another informative study.
In this hypothetical case, the higher-exposure point prediction is potentially clinically meaningful and falls within observed exposure, while the randomized estimate remains imprecise. Assume that exposure near 6.0 ng/mL is tolerable based on the available safety data. A proportionate decision is to continue with a targeted dose-and-adherence investigation, rather than immediately select a new population, replace the endpoint, or stop. The next study should create an interpretable comparison of regimens that produce different exposures, monitor safety, and measure adherence and exposure prospectively. If acceptable exposure cannot be achieved or the predicted benefit does not replicate, the decision should be revisited.
Quantify the treatment estimate, uncertainty, and relevance to the clinical target before interpreting post-randomization exposure patterns.
Review exposure distributions, adherence measurement, missing samples, timing, and whether the model covers the observed range adequately.
Assess whether the association is robust to reasonable covariate choices and whether the potentially useful exposure is attainable and tolerable.
Study dose or adherence when exposure is a plausible explanation; change population or endpoint only with supporting evidence; stop when meaningful benefit is not plausibly attainable.
Diagnostics, Alternatives, and Common Pitfalls
Before using an exposure-response result to revise a plan, examine the evidence that could make it misleading. Plot outcomes against exposure and inspect whether a linear model is reasonable across the observed range. Check exposure measurement quality, adherence records, influential observations, missing data, and whether the fitted relationship is driven by a few participants. Compare prespecified and clinically justified model specifications, and examine whether relevant baseline prognostic factors or tolerability-related discontinuation could distort the association. A steep relationship supported only by sparse data at the highest exposures should not be treated as a reliable dose recommendation.
The main alternatives answer different questions. The randomized intention-to-treat comparison estimates the effect of assignment to the studied regimen and protects the comparison from baseline confounding through randomization. An exposure-response analysis can help explain variation in outcome and generate dose hypotheses, but exposure itself was not randomized. A per-protocol or adherence-adjusted analysis may address outcomes under different treatment-use assumptions, but it requires additional assumptions and careful handling of factors related to adherence. None should be presented as interchangeable with the primary randomized result.
A common mistake is to reason, “The trial was nonsignificant, but high-exposure participants improved more, so increasing the dose will work.” This overstates both findings. The nonsignificant result does not prove absence of effect, and a within-treatment exposure association does not establish a causal dose effect. The correction is to report the randomized estimate and its uncertainty, treat the exposure model as supportive and assumption-dependent, and seek a design that can test a dose or adherence strategy while monitoring safety.
Check Your Understanding
Answer each question using the case assumptions. Brief solutions follow the questions.
- Why does the 95% confidence interval \([-3.35,1.35]\) not support the conclusion that the drug has no effect?
- What is the exposure-response model’s predicted treatment effect at \(C=5.0\) ng/mL?
- Why is the contrast between participants with mean exposure 2.0 and 6.0 ng/mL not equivalent to a randomized dose comparison?
- If the useful exposure range were not tolerable, which development decision would become more compelling, and why?
- Suppose the symptom endpoint was shown to have poor measurement reliability in the trial. What evidence would be needed before replacing it in a future study?
Solutions. (1) The interval includes zero but also includes effects more favorable than the \(-2.0\)-point clinical target; the data remain compatible with a range of effects, rather than establishing exactly zero. (2) \(\widehat{\Delta}(5.0)=-1.0-0.75(5.0-4.0)=-1.75\) points, a model-based point prediction. (3) Participants were not randomly assigned to those exposure levels, and exposure can be related to adherence, prognosis, tolerability, or other outcome predictors. (4) Stopping becomes more compelling if no safe, attainable exposure can plausibly yield meaningful benefit, because dose escalation would not provide a viable development path. (5) The endpoint change should be supported by evidence that a better-validated measure captures the intended clinical benefit, with the new choice prespecified and justified rather than selected because it gives a more favorable result.