Why Zero Events Still Leave Uncertainty
A trial arm with no observed events can look reassuring, but it does not establish that the underlying event rate is zero. If only a small number of participants were observed, appreciable event rates may still be compatible with the data. Bayesian inference makes this uncertainty explicit: the posterior distribution assigns probabilities to a range of possible event rates, including values above zero.
This tutorial considers a binary endpoint measured over a defined follow-up period. The focus is on what happens when event counts are zero or small, and on interpreting treatment contrasts without confusing an observed rate of zero with certainty about the underlying rate.
The binomial model is suitable when the endpoint definition and observation window are consistent, outcomes are recorded for the participants included in the analysis, and the arm-level event probability is a reasonable summary. Independence can fail, for example, with clustered participants or repeated observations treated as independent. Unequal follow-up, informative loss to follow-up, or substantial variation in risk across participants can also make a simple binomial model inadequate.
What the Beta-Binomial Update Does at Zero
With \(y_j\) observed events among \(n_j\) participants, the likelihood is proportional to \(p_j^{y_j}(1-p_j)^{n_j-y_j}\). Combining it with the beta prior gives the posterior update below. It applies to zero-event arms as well as arms with events.
In particular, if \(y_j=0\), then the posterior is \(\operatorname{Beta}(a_j,b_j+n_j)\). The likelihood contribution is \((1-p_j)^{n_j}\), which downweights larger event probabilities, but it is not evidence that rules out every positive value of \(p_j\). With a proper beta prior, the posterior is continuous on values between zero and one; it has no point mass at zero. The amount of remaining uncertainty depends on the prior and the number of participants with observed outcomes.
For a posterior \(\operatorname{Beta}(a,b)\), the mean and variance are \(a/(a+b)\) and \(ab/\{(a+b)^2(a+b+1)\}\), respectively. These summaries remain defined when the observed count is zero. A posterior credible interval can be obtained from the beta quantiles. Its interpretation is conditional on the model, prior, and observed data: a 95% posterior interval has 95% posterior probability of containing the unknown event rate. It is not a frequentist statement about the interval’s repeated-sampling coverage.
Worked Example: No Events in Either Arm
Suppose, hypothetically, that a trial observes no events among 10 participants assigned to treatment and no events among 10 assigned to control. Use independent \(\operatorname{Beta}(1,1)\) priors in both arms.
The binomial model requires a binary event outcome for each of the 20 participants over the same observation period. For this example, assume outcomes are recorded and conditionally independent within each arm. The equal, independent priors imply independent posteriors with the same distribution.
For each arm, the posterior is \(\operatorname{Beta}(1+0,1+10)=\operatorname{Beta}(1,11)\). Thus the posterior mean event rate is
The posterior variance is
For each arm, an equal-tail 95% posterior interval is approximately \((0.00230,\ 0.2849)\). These limits follow from the \(\operatorname{Beta}(1,11)\) quantiles; its cumulative distribution function is \(1-(1-p)^{11}\). Despite zero events, the interval includes event rates well above zero.
Let the treatment risk difference be \(D=p_T-p_C\). The arm posteriors are identically distributed and independent, so \(E(D\mid\text{data})=0.08333-0.08333=0\), and symmetry gives \(P(D<0\mid\text{data})=0.5\). This does not mean both event rates are known to be small: their marginal posterior distributions remain broad. It means these data, under these priors, do not favor either arm’s rate over the other.
Contrasts When One Arm Has No Events
A treatment contrast is a posterior quantity computed from both arm-specific event rates. For the risk difference \(D=p_T-p_C\), a negative value means the treatment-arm event rate is lower. Given independent arm posteriors, the posterior mean difference is the difference between the posterior means, and its variance is the sum of the arm-specific variances. Those summaries describe the contrast but do not by themselves determine the probability that it is below zero.
For more complicated summaries, such as a risk ratio or odds ratio, compute the contrast from the joint posterior rather than treating the observed zero count as a literal zero underlying rate. A practical approach is to draw values independently from the two arm-specific beta posteriors, calculate the contrast for each pair, and summarize those draws. The fraction of draws with \(p_T<p_C\), for example, estimates the posterior probability of a lower treatment-arm event rate. For ratios, draws also make clear when the posterior is highly skewed or has substantial mass near an extreme value.
Worked Example: Zero Events on Treatment, Three on Control
Suppose, hypothetically, that 0 of 12 treatment participants and 3 of 12 control participants experience the event. Assign independent \(\operatorname{Beta}(1,1)\) priors. Assume the same binary endpoint and follow-up window in each arm, with complete outcome recording and conditionally independent outcomes.
The posterior event rates are
Their posterior means are \(1/14=0.07143\) for treatment and \(4/14=0.2857\) for control. The posterior mean risk difference is therefore
This is a posterior mean difference of about 21.43 percentage points in the direction of a lower event rate on treatment. It is not a guarantee that the treatment rate is lower. To quantify that direction directly, let \(X=p_T\) and \(Y=p_C\). Since \(X\sim\operatorname{Beta}(1,13)\), its cumulative distribution function is \(1-(1-x)^{13}\). Therefore,
The beta-function ratio equals \((10)_{13}/(14)_{13}\), where \((z)_k=z(z+1)\cdots(z+k-1)\). Canceling common factors gives \(10\cdot11\cdot12\cdot13/(23\cdot24\cdot25\cdot26)\approx0.04783\). Thus the posterior probability that the treatment event rate is lower is about 95.22%, conditional on this model, prior, and data. This probability does not say that the risk difference is exactly \(-0.2143\); that value is its posterior mean.
Prior Sensitivity and Sparse Counts
When counts are small, the prior can have a visible effect on posterior summaries. This does not make the analysis invalid; it makes prior choice and sensitivity checks especially important. A prior should be specified for a clinically interpretable event rate, and analysts should examine whether reasonable alternatives change the substantive conclusion.
Worked Example: Checking Prior Sensitivity
Consider a hypothetical trial with 0 events among 5 treatment participants and 1 event among 5 control participants. Compare independent uniform \(\operatorname{Beta}(1,1)\) priors with independent Jeffreys \(\operatorname{Beta}(1/2,1/2)\) priors. Suppose the endpoint and follow-up are comparable across arms and the binomial assumptions are reasonable.
With uniform priors, the posteriors are \(\operatorname{Beta}(1,6)\) for treatment and \(\operatorname{Beta}(2,5)\) for control. Their means are \(1/7=0.1429\) and \(2/7=0.2857\), giving a mean risk difference of \(-1/7=-0.1429\).
With Jeffreys priors, the posteriors are \(\operatorname{Beta}(1/2,11/2)\) for treatment and \(\operatorname{Beta}(3/2,9/2)\) for control. Their means are \((1/2)/6=0.08333\) and \((3/2)/6=0.2500\), giving a mean risk difference of \(-1/6=-0.1667\).
Both analyses give a negative posterior mean difference, but the numerical summaries differ. With only five participants per arm, the data provide limited information, so the prior’s contribution is appreciable. A responsible report would show this sensitivity and consider whether the chosen prior is clinically defensible; it would not hide the difference by reporting only the observed rates.
Diagnostics, Alternatives, and Common Pitfalls
Sparse binary data warrant checks of the data and model before interpretation. Confirm the number randomized and the number with observed outcomes in each arm; verify that the event definition and follow-up window match; and establish whether zero means no events or missing outcome information. Examine whether the binomial assumption is credible. For example, if participants are clustered by site, treating all outcomes as independent can understate uncertainty. If follow-up differs materially across participants, an event probability over a common fixed period may not describe the observed data well.
Prior sensitivity is another key diagnostic. Compare posterior rates and the treatment contrast under plausible priors, especially when the total event count is very small. Also inspect the full posterior contrast distribution, not only its mean. Posterior draws or numerical integration can reveal skewness, wide credible intervals, or a high probability of meaningful benefit alongside substantial uncertainty. A point estimate alone can conceal all three.
- Raw rates: \(y/n\) is simple to report but is zero when \(y=0\), and does not express uncertainty about the underlying rate.
- Bayesian beta-binomial analysis: gives posterior distributions for each arm and contrasts, with prior information stated explicitly. Results can be sensitive to the prior when data are sparse.
- Frequentist exact methods: can be useful for small-sample inference, but their p-values or confidence intervals do not provide posterior probabilities for event rates or treatment contrasts.
A frequent mistake is to reason, “There were zero treatment events, so the treatment event probability is zero and the risk ratio is zero.” The correction is that the observed count is zero, while the posterior event rate generally remains positive and uncertain under a continuous beta prior. Treatment contrasts must be calculated from posterior rates, not by substituting the zero observed rate as if it were known exactly. In particular, do not confuse the ratio of posterior means with the posterior distribution of a ratio.
These methods also have limits. They do not repair unreliable outcome ascertainment, missing outcomes, misspecified follow-up, or unmodeled clustering. Independent arm posteriors rely on the specified independent priors and separate binomial likelihoods; a model that links arms or borrows information across studies will yield different dependence and possibly different contrasts. With very sparse data, communicate uncertainty and prior dependence alongside any directional conclusion.
Check Your Understanding
Answer each question, then compare your reasoning with the solutions below.
- With a \(\operatorname{Beta}(2,3)\) prior and zero events among 8 participants, what is the posterior distribution and its mean?
- Two arms each have zero events among 6 participants and independent uniform priors. What is the posterior mean in each arm, and what is the posterior probability that the treatment rate is lower?
- With independent uniform priors, a trial observes 0 of 10 events on treatment and 2 of 10 on control. Find both posterior distributions and the posterior mean risk difference.
- Name two conditions that could make the simple binomial model unsuitable for a trial with zero recorded events.
Solutions. (1) The update gives \(\operatorname{Beta}(2+0,3+8)=\operatorname{Beta}(2,11)\), with mean \(2/(2+11)=2/13=0.1538\). (2) Each posterior is \(\operatorname{Beta}(1,7)\), with mean \(1/8=0.1250\). The posteriors are independent and identically distributed, so symmetry gives \(P(p_T<p_C\mid\text{data})=0.5\). (3) The treatment posterior is \(\operatorname{Beta}(1,11)\); the control posterior is \(\operatorname{Beta}(3,9)\). Their means are \(1/12\) and \(3/12\), so the mean risk difference is \(1/12-3/12=-2/12=-0.1667\). (4) Examples include clustered participants treated as independent, materially unequal follow-up when a common fixed-period event probability is intended, or incomplete outcome recording. Each can undermine the assumptions or interpretation of the binomial model.