Why Apply Safety as a Separate Constraint?
Efficacy-based dose selection asks whether a candidate dose achieves a clinically relevant effect. A dose can satisfy that target and still carry an unacceptable risk of adverse events. Conversely, a dose with a reassuring safety profile may not provide enough benefit to justify selection. When the decision has a prespecified efficacy target and a prespecified safety limit, each is a requirement: a dose that fails either one is not eligible.
This is a constrained decision, not a calculation of a single benefit-risk score. A larger efficacy estimate cannot compensate for exceeding the safety limit, and a low adverse-event rate cannot compensate for missing the efficacy target. The decision rule should be set before inspecting dose-specific results, so the requirements are not adjusted after seeing which doses pass.
The quantities must be defined in context before applying the rule. For example, an efficacy target might be a minimum placebo-adjusted mean improvement of 8 points on a specified endpoint, while the safety limit might be an adverse-event probability no greater than 10% in the population under consideration. The endpoint, population, event definition, and observation period must be consistent with the evidence used to evaluate each dose. This tutorial treats those as already specified; it does not determine them.
Writing the Two Gates Precisely
With a Bayesian dose-response analysis, uncertainty can be represented by posterior distributions for \(E(d)\) and \(R(d)\). One common rule requires a high posterior probability that the efficacy target is met and a high posterior probability that the safety risk is below its limit:
Here, \(\Delta_E\) is the minimum efficacy target, \(r_{\max}\) is the maximum acceptable safety risk, and \(\gamma_E\) and \(\gamma_S\) are prespecified probability thresholds. A dose is eligible only when both inequalities hold. The thresholds are decision specifications, not universal constants: their choice depends on the development context and should be justified and documented.
An alternative is to require that an interval bound meet each target—for example, a lower confidence bound for efficacy at or above \(\Delta_E\), and an upper confidence bound for risk at or below \(r_{\max}\). That is a different inferential rule and should be described as such. A point estimate alone does not express uncertainty: observing an event rate below the safety limit does not, by itself, establish that the underlying risk is below it.
The posterior probabilities above are marginal probabilities for each gate. Requiring each to exceed its own threshold does not necessarily require the posterior probability of meeting both requirements simultaneously to exceed that same threshold. If the intended rule is a joint probability requirement, it must be stated directly as \(P(E(d)\geq\Delta_E,\ R(d)\leq r_{\max}\mid\text{data})\geq\gamma\). That remains a constraint-based rule, not a composite utility, but it is not interchangeable with two marginal gates.
Worked Example: Screening Three Doses With Separate Gates
Hypothetical data and rule. Suppose a development team sets an efficacy target of at least 8 points of placebo-adjusted mean improvement and a safety limit of at most 10% risk of a specified adverse event. It prespecifies posterior thresholds \(\gamma_E=0.80\) and \(\gamma_S=0.90\). For illustration, suppose the efficacy model gives a normal posterior at each dose, with the displayed mean and standard deviation, and the safety model uses a uniform prior \(R(d)\sim\mathrm{Beta}(1,1)\) with binomial event data. The dose-level posterior distributions are treated as applicable to the same target population and event definition.
| Dose | Efficacy posterior | Safety events |
|---|---|---|
| 50 mg | \(E(50)\mid\text{data}\sim N(9,2^2)\) | 0 of 40 |
| 75 mg | \(E(75)\mid\text{data}\sim N(10.5,2^2)\) | 1 of 40 |
| 100 mg | \(E(100)\mid\text{data}\sim N(12,2^2)\) | 6 of 40 |
Efficacy calculations. For a normal posterior with standard deviation 2, the probability of exceeding 8 is the standard normal probability above the standardized threshold. At 50 mg, \(z=(8-9)/2=-0.5\), so \(P(E(50)\geq8\mid\text{data})=\Phi(0.5)=0.6915\). At 75 mg, \(z=(8-10.5)/2=-1.25\), giving \(\Phi(1.25)=0.8944\). At 100 mg, \(z=(8-12)/2=-2\), giving \(\Phi(2)=0.9772\). Thus 50 mg fails the 0.80 efficacy gate, while 75 mg and 100 mg pass it.
Safety calculations. With \(x\) events among \(n\) participants and a \(\mathrm{Beta}(1,1)\) prior, the posterior risk is \(\mathrm{Beta}(x+1,n-x+1)\). At 50 mg, zero events among 40 give \(R(50)\mid\text{data}\sim\mathrm{Beta}(1,41)\). Its cumulative probability at 0.10 is \(1-(1-0.10)^{41}=1-0.9^{41}=0.9867\), so this dose passes the safety gate. At 75 mg, one event gives \(\mathrm{Beta}(2,40)\), with \(P(R(75)\leq0.10\mid\text{data})=1-0.9^{41}-4.1(0.9^{40})=0.9261\), which also passes. At 100 mg, six events give \(\mathrm{Beta}(7,35)\); the beta cumulative probability is about 0.110 at 0.10, below 0.90, so the dose fails the safety gate.
Decision. The 50 mg dose fails efficacy, and 100 mg fails safety. The 75 mg dose passes both separate requirements and is eligible under this rule. This conclusion does not establish that 75 mg is uniquely optimal; it says only that it meets the stated gates among the doses evaluated.
Why Uncertainty Is Part of the Safety Constraint
The safety gate concerns the risk parameter, not just the proportion observed in a trial. A small sample with no observed events can still be compatible with a range of underlying risks. The posterior probability used in the gate makes that uncertainty explicit, subject to the chosen model and prior. The same principle applies to an upper confidence bound in a frequentist analysis: a favorable point estimate is not a substitute for adequate precision.
Worked Example: No Observed Events Does Not Guarantee a Safety Pass
Hypothetical data and rule. At a candidate dose, no specified adverse events are observed among 10 participants. Use the same uniform prior and 10% risk limit as above, with a safety threshold of 0.90. Suppose the efficacy posterior is \(E(d)\mid\text{data}\sim N(11,2^2)\), and the efficacy target remains 8 points with threshold 0.80.
Efficacy gate. The standardized threshold is \(z=(8-11)/2=-1.5\), so \(P(E(d)\geq8\mid\text{data})=\Phi(1.5)=0.9332\). Efficacy passes.
Safety gate. Zero events among 10 produce a \(\mathrm{Beta}(1,11)\) posterior for risk. Therefore, \(P(R(d)\leq0.10\mid\text{data})=1-0.9^{11}=0.6862\), which is less than 0.90. Safety fails under the prespecified rule.
Decision. The dose is not eligible, even though no event was observed and the efficacy gate passes. The result is not evidence that the risk is necessarily above 10%; it indicates that the available data do not give the required posterior support that risk is at or below 10%.
Separate Gates, Joint Probability, and Other Decision Rules
Separate gates have a transparent interpretation: each requirement is checked on its own terms, and a failure cannot be offset by a favorable result on the other dimension. But the probability of passing both gates may differ from either marginal probability. For any two events \(A\) and \(B\), their joint probability satisfies \(P(A\cap B)\geq P(A)+P(B)-1\). Thus, if each marginal pass probability is 0.92, the joint pass probability could be as low as 0.84; it is not automatically 0.92.
Worked Example: Interpreting Marginal Pass Probabilities
Hypothetical posterior summary. At a candidate dose, suppose \(P(E(d)\geq\Delta_E\mid\text{data})=0.92\) and \(P(R(d)\leq r_{\max}\mid\text{data})=0.92\). Under separate gates with both thresholds equal to 0.90, the dose passes each gate.
What follows—and what does not. Let \(A\) be the event that efficacy meets its target and \(B\) the event that safety risk meets its limit. The lower bound on the joint probability is \(0.92+0.92-1=0.84\), while the upper bound is the smaller marginal probability, 0.92. Thus the joint probability can lie between 0.84 and 0.92; the two marginal results alone do not establish a 0.90 joint probability.
Decision implication. If the prespecified decision rule is two marginal gates, this dose passes. If the requirement is at least 0.90 posterior probability of meeting both criteria together, the joint posterior must be evaluated and compared with 0.90. Reporting the marginal results as though they proved that joint requirement would overstate what the analysis shows.
A frequentist alternative can apply an efficacy lower confidence bound and a safety upper confidence bound as separate gates. It can be appropriate when the analysis plan specifies those bounds and their operating characteristics. It is not numerically identical to a posterior-probability rule: the inferential interpretations differ, and one should not select whichever produces the more favorable dose after seeing the results.
A composite benefit-risk utility is another kind of approach, but it answers a different question by assigning trade-offs between efficacy and harm. It is deliberately not used here. When the decision specifies a minimum efficacy and a maximum risk, the direct operational consequence is exclusion of a dose that fails either gate—not a weighted score that might allow one result to compensate for the other.
Assumptions, Diagnostics, and Failure Modes
A gate-based conclusion is only as credible as the estimates and definitions entering it. Before interpreting pass probabilities, verify that the efficacy and safety analyses refer to the intended patient population, dose, endpoint, event definition, and relevant observation period. The posterior calculations in the examples assume the stated normal efficacy posterior and binomial safety likelihood with a uniform prior; different data structures or modeling choices require an appropriate model rather than mechanical reuse of those formulas.
- Inspect uncertainty near each boundary. A probability just above a threshold is a pass under the rule, but it is a borderline pass. Report the probability and the target or limit, not only the word “pass.”
- Check the safety data support. Few participants or few events may yield a broad risk posterior. A zero count should not be read as zero underlying risk.
- Check model fit and dose-level predictions. A misspecified dose-response model can give misleading probabilities, especially where data are sparse or the observed pattern conflicts with the model. Examine whether predictions are plausible at each candidate dose and whether uncertainty is represented adequately.
- Keep the decision rule fixed. Changing \(\Delta_E\), \(r_{\max}\), \(\gamma_E\), or \(\gamma_S\) after inspecting results can make the apparent pass status data-driven. If a change is clinically justified, state the rationale and assess its effect transparently.
- Account for evaluating multiple doses. Pointwise posterior probabilities describe each dose under the fitted model. They do not by themselves guarantee a particular repeated-sampling error rate for a procedure that selects among many doses. The decision strategy and its performance should be evaluated for the intended use.
A useful justification records the target and limit, the two thresholds, the model and data supporting each probability, and the pass or fail result at every candidate dose. This makes clear whether a dose was excluded for inadequate efficacy, inadequate safety evidence, or both. It also prevents an efficacy advantage from obscuring a safety failure, or a reassuring safety result from being mistaken for evidence of adequate benefit.
Check Your Understanding
Answer each question using the stated decision rule. Short solutions follow the questions.
- A dose has posterior efficacy pass probability 0.88 and safety pass probability 0.91. The separate thresholds are 0.80 and 0.90. Is the dose eligible?
- With a uniform prior, what is the posterior distribution for an adverse-event risk after 2 events among 30 participants?
- For 0 events among 20 participants and a uniform prior, calculate the posterior probability that risk is at most 10%. Does it meet a safety threshold of 0.90?
- Why do marginal pass probabilities of 0.95 for efficacy and 0.95 for safety not, by themselves, prove a joint pass probability of at least 0.95?
Solutions. (1) Yes, under the stated separate gates: \(0.88\geq0.80\) and \(0.91\geq0.90\). This does not assert a 0.90 probability of passing both jointly. (2) The posterior is \(\mathrm{Beta}(2+1,30-2+1)=\mathrm{Beta}(3,29)\). (3) The posterior is \(\mathrm{Beta}(1,21)\), so \(P(R\leq0.10\mid\text{data})=1-0.9^{21}=0.8906\). Since \(0.8906<0.90\), the dose fails the safety gate. (4) The joint probability is at most the smaller marginal probability, 0.95, and can be lower; its lower bound is \(0.95+0.95-1=0.90\). The joint probability therefore is not established to be at least 0.95 without evaluating the joint posterior or imposing an appropriate joint rule.