Binary Noninferiority in an Active-Control Trial

Why Binary Noninferiority Needs a Directional Contrast

An active-control trial asks whether a new treatment retains an acceptable amount of the established treatment’s benefit. For a binary efficacy outcome, such as whether a patient responds by a fixed assessment time, the comparison is between two response probabilities. Noninferiority is not established merely because the observed response rates look similar: the trial must address how much lower the new treatment’s response probability could plausibly be.

Let \(p_E\) denote the probability of a favorable outcome under the experimental treatment and \(p_C\) the probability under the active control. Use the risk difference \(\Delta=p_E-p_C\). A positive value favors the experimental treatment; a negative value means its response probability is lower. Let \(M>0\) be the prespecified noninferiority margin, expressed as a probability difference. The noninferiority region is \(\Delta>-M\). For example, \(M=0.08\) allows a loss of up to 8 percentage points in response probability. The margin should be justified clinically and fixed before examining the trial results.

This sign convention depends on the outcome being favorable. If the event is undesirable, such as a complication, either define the contrast so that larger values remain favorable or reverse the inequality. A sign error can turn a correct calculation into the wrong clinical conclusion.

Binomial Model and Posterior Distribution

Suppose \(x_E\) of \(n_E\) randomized participants in the experimental arm have the favorable outcome, and \(x_C\) of \(n_C\) participants in the control arm do. Conditional on each arm’s outcome probability, model the counts as binomial:

This simple model assumes a fixed binary assessment for every participant, independent outcomes within each arm conditional on its probability, and no unmodeled clustering that would materially change the uncertainty. It also assumes the analysis population and handling of missing outcomes have been specified appropriately. In a randomized trial, randomization supports a treatment comparison, but it does not by itself correct outcome misclassification, informative missingness, or major departures from the intended treatment strategy.

For a basic conjugate analysis, assign independent beta priors \(p_E\sim\operatorname{Beta}(a_E,b_E)\) and \(p_C\sim\operatorname{Beta}(a_C,b_C)\). A beta distribution has density proportional to \(p^{a-1}(1-p)^{b-1}\) for \(0<p<1\). Its parameters must be chosen and justified before looking at the trial outcomes. With independent priors and the binomial likelihoods, the posterior distributions are:

The update follows directly by multiplying prior density by likelihood. In one arm, the likelihood is proportional to \(p^x(1-p)^{n-x}\). Multiplication by the beta prior gives a density proportional to \(p^{a+x-1}(1-p)^{b+n-x-1}\), which is the kernel of a \(\operatorname{Beta}(a+x,b+n-x)\) distribution. Under the independent model, the two arm posteriors remain independent.

For a posterior \(\operatorname{Beta}(A,B)\), the mean is \(A/(A+B)\) and the variance is \(AB/[(A+B)^2(A+B+1)]\). Therefore the posterior mean of \(\Delta\) is the experimental-arm posterior mean minus the control-arm posterior mean. Because the arm posteriors are independent, the posterior variance of \(\Delta\) is the sum of their posterior variances. These summaries describe the center and spread of the contrast, but the noninferiority decision concerns a probability relative to the margin, not just the mean contrast.

Calculating the Noninferiority Probability

The exact posterior probability can be obtained from the two beta distributions. Conditional on a value of \(p_E\), the condition \(\Delta>-M\) is equivalent to \(p_C<p_E+M\). Let \(f_E\) be the posterior density of \(p_E\), and let \(F_C\) be the posterior cumulative distribution function of \(p_C\). Then:

This one-dimensional integral can be evaluated by numerical integration. Alternatively, draw paired independent values from the two arm-specific posterior distributions, calculate their differences, and estimate the probability by the proportion exceeding \(-M\). Numerical integration is deterministic up to numerical precision; simulation has Monte Carlo error, which should be checked by using enough draws and, where relevant, repeated runs or an estimated Monte Carlo standard error.

A useful approximation for a quick calculation is to treat the posterior difference as normal with its calculated posterior mean and variance. If the mean is \(\mu_\Delta\) and the posterior standard deviation is \(s_\Delta\), this gives:

where \(\Phi\) is the standard normal cumulative distribution function. This is an approximation to the posterior probability, not a change to the model’s exact beta posteriors. It is most credible when both arm posteriors are away from the boundaries and are reasonably concentrated. With sparse outcomes, very small probabilities, or substantial posterior skewness, use the beta integral or posterior simulation instead.

Checks, Alternatives, and Limits

Before reporting a result, confirm that the outcome direction, contrast, and margin all match the clinical question. For a favorable endpoint, \(\Delta=p_E-p_C\) and the threshold is \(-M\); the probability being reported is the posterior mass above that threshold. Report the margin in percentage points and describe the posterior probability in context. If the decision rule requires, for example, at least 0.95 posterior probability, state that rule explicitly and distinguish it from the calculated probability.

The normal approximation is a diagnostic shortcut, not a substitute for checking the posterior shape. Compare it with numerical integration or posterior simulation if either arm has few events, many outcomes at one boundary, or a visibly skewed posterior. For simulation, inspect Monte Carlo precision: an estimate based on too few draws can fluctuate enough to affect a decision near its probability threshold. Also check the prior’s implications. A \(\operatorname{Beta}(1,1)\) prior is uniform for each arm probability, but it is not automatically noninformative for every contrast or for a clinical decision.

The binomial analysis can be misleading if participants’ outcomes are dependent, assessment times differ, or missing outcomes are handled in a way that could favor one arm. The active control must also have been used in circumstances that make its expected effect credible; inadequate adherence or poor execution can make both arms look alike without demonstrating that the experimental treatment preserves meaningful benefit. Sensitivity analyses for reasonable prior and missing-data assumptions may be important. The basic calculation does not repair flaws in trial conduct or endpoint definition.

A frequentist noninferiority analysis typically assesses a confidence interval relative to the same margin, with its own prespecified testing procedure. The Bayesian analysis instead reports posterior probability under an explicit model and prior. Neither approach makes margin choice automatic, and neither should be described as showing “no difference” just because noninferiority is supported. Noninferiority also differs from superiority: it allows an acceptable loss and does not require evidence that the experimental treatment improves the outcome.