Worked Simulation of a Phase 2 Dose-Ranging Decision

From Dose-Response Assumptions to a Decision

A Phase 2 dose-ranging study must do more than detect a treatment effect. It must provide evidence that helps choose a dose for further development, while accounting for safety and the uncertainty of a finite trial. Simulation makes this decision problem explicit: specify plausible virtual patients, run the planned design and analysis repeatedly, and estimate how often the decision rule selects a suitable dose.

The simulation below is wholly hypothetical. It compares two ways of allocating 240 participants among placebo and three doses. Its purpose is to show how assumptions, allocation, analysis, and the performance measure fit together. The estimated probabilities are properties of this specified simulation scenario, not predictions that a future clinical study will achieve those outcomes.

Suppose the endpoint is a continuous measure of clinical benefit at the primary assessment, with larger values better. Let \(d\) denote dose in mg. The assumed mean benefit relative to placebo follows an Emax model:

For this example, \(E_{\max}=0.90\) benefit units and \(ED_{50}=5\) mg. The candidate doses are 5, 10, and 20 mg; their assumed mean benefits are 0.45, 0.60, and 0.72 units, respectively. Placebo has mean benefit zero. The common residual standard deviation is 1.2 units. Independently, each participant has a binary safety event, with probabilities 0.02 on placebo and 0.03, 0.06, and 0.14 at 5, 10, and 20 mg, respectively. These are simplifying assumptions: they exclude covariate effects, dropout, and dependence between benefit and safety.

For the decision criterion, define 10 mg as the target dose: its assumed mean benefit is at least 0.50 units and its safety-event probability is no more than 0.10. The 20 mg dose has greater mean benefit but exceeds the assumed safety limit. A suitable dose-selection probability will therefore count trials that select 10 mg, not simply trials that select the dose with the largest mean benefit.

Generate Virtual Patients and Specify the Trial

Each simulated trial randomizes 240 virtual participants. In the equal-allocation design, 60 participants receive each of placebo, 5 mg, 10 mg, and 20 mg. In the dose-weighted design, the allocations are 60, 40, 60, and 80, respectively. Both designs retain 60 participants on placebo; the second assigns more participants to the highest dose and fewer to 5 mg.

For each participant, the simulation draws a normally distributed benefit outcome using the assigned dose-specific mean and standard deviation, then draws a safety indicator from a Bernoulli distribution using that dose's safety-event probability. These draws are independent conditional on dose in this teaching model. Randomization is assumed to be followed, outcomes are complete, and no other design features alter the generated data.

The analysis and decision rule must be specified before evaluating simulated performance. In each replicate, estimate each dose-versus-placebo benefit contrast. Test the three contrasts using a one-sided Dunnett procedure controlling the familywise type I error rate at \(\alpha=0.05\). A dose passes the efficacy screen only if its adjusted test is significant. It passes the safety screen only if its observed safety-event proportion is at most 0.10. Among doses passing both screens, select the one with the largest estimated benefit; if no dose passes both, select no dose.

For this example, the performance measure is the proportion of replicates in which 10 mg is selected. Also record how often each other dose is selected and how often no dose is selected. This makes the target of the simulation concrete: it evaluates the decision rule against the assumed dose-response and safety scenario, rather than treating statistical significance alone as success.

Work Through One Simulated Trial

A replicate of the equal-allocation design might produce mean benefits of 0.08, 0.48, 0.70, and 0.82 units for placebo, 5 mg, 10 mg, and 20 mg. Suppose its pooled residual standard deviation is 1.20 units. With 60 participants in each arm, the standard error for each unadjusted dose-placebo contrast is:

The observed contrasts and their unadjusted test statistics are therefore:

These unadjusted statistics do not replace the specified Dunnett tests: the multiplicity-adjusted results determine which doses pass the efficacy screen. Suppose this replicate’s safety-event counts are 4, 4, 7, and 9 in the four arms. The active-dose event proportions are \(4/60=0.0667\), \(7/60=0.1167\), and \(9/60=0.1500\). Under the rule, 5 mg passes the safety screen, but 10 mg and 20 mg do not. If 5 mg does not pass the adjusted efficacy test, no dose is selected in this replicate—even though the observed mean benefit is largest at 20 mg.

The example illustrates why a dose-ranking table and a decision rule are not interchangeable. The rule combines statistical evidence and a safety constraint; a dose with the largest observed benefit need not be eligible for selection.

Repeat the Trial and Estimate Selection Probabilities

Now repeat the complete process 1,000 times for each allocation design. Each replicate gets newly generated patients, is analyzed by the same multiplicity-adjusted efficacy and safety screens, and produces one decision. The table gives hypothetical teaching outputs from such a run. Counts within each row sum to 1,000; the percentages are the corresponding simulated selection probabilities.

In this hypothetical run, equal allocation estimates a higher probability of selecting 10 mg than dose-weighted allocation: 62.0% versus 59.0%. It also selects 20 mg more often and selects no dose less often. These differences are not automatically meaningful: Monte Carlo variability, rounding, and the assumed scenario matter. The output does not establish that equal allocation is generally superior. A design comparison should use the same scenarios and decision rules, examine Monte Carlo error, and consider the full set of relevant performance measures.

Justification, Diagnostics, and Alternatives

The Monte Carlo calculation follows directly from the indicator for success. For replicate \(r\), let \(I_r=1\) if the rule selects 10 mg and \(I_r=0\) otherwise. Under independent, identically generated replicates, each \(I_r\) is Bernoulli with success probability \(p\). Thus the estimated selection probability is the sample mean of these indicators, and its variance is \(p(1-p)/M\). Replacing the unknown \(p\) by \(\hat{p}\) gives the estimated Monte Carlo standard error used above. Increasing the number of replicates reduces this error approximately in proportion to \(1/\sqrt{M}\), but does not repair incorrect assumptions about the trial or patients.

A practical simulation should be checked at several levels. Confirm that generated endpoint means and safety-event proportions approach their specified values across many virtual patients. Verify arm sizes, randomization, and that every replicate applies the same analysis and decision rule. Check that reported decision counts sum to the number of replicates. Repeat the simulation with more replicates to determine whether apparent design differences are larger than Monte Carlo error.

The simple rule here is not the only reasonable choice. A continuous utility score can trade benefit against safety instead of imposing a pass/fail safety screen. For example, suppose a hypothetical observed mean benefit is 0.76 units and 7 of 60 participants have a safety event; define utility as mean benefit minus twice the observed event proportion. Then:

Other alternatives include allocating more participants to doses where precise estimation is most valuable, or using model-based dose-response estimates rather than selecting from observed arm means alone. These approaches can improve a particular objective, but they introduce their own assumptions. An Emax analysis, for example, can borrow information across doses but may be misleading if the assumed curve is wrong. Equal allocation is straightforward and supports comparisons across arms; dose-weighted allocation can increase information at selected doses but leaves fewer participants at other doses. Earlier material on allocation ratios and Emax-based design evaluation provides foundations for those trade-offs; this worked decision emphasizes how the complete analysis rule turns trial data into a selection outcome.

Common Pitfalls and Limits

• Choosing the largest observed mean automatically. That ignores sampling variation, multiplicity, and safety. Apply the prespecified rule, including its eligibility criteria.
• Calling a simulated probability a clinical guarantee. The 62.0% estimate is conditional on the assumed model, design, endpoint, and rule. It is not the probability that a real study will select 10 mg under all circumstances.
• Confusing Monte Carlo error with model uncertainty. More replicates reduce numerical simulation error, but do not make the assumed response curve or event rates more credible.
• Over-interpreting small differences between designs. Compare their Monte Carlo uncertainty and performance across multiple clinically plausible scenarios, not just one set of assumptions.

This simulation is limited by its single scenario, normal continuous endpoint, fixed follow-up, common residual variability, independent benefit and safety, and complete data. A decision analysis for development would usually examine sensitivity to response and safety assumptions, variability, missingness, and relevant treatment heterogeneity. The resulting performance summaries should be interpreted as conditional evidence for choosing a design and rule, not as a substitute for clinical judgment.