TOGETHER And Bayesian Outpatient Decisions

Why Bayesian Decisions Mattered in TOGETHER

Outpatient treatment trials during COVID-19 faced a practical problem: many candidate treatments had to be assessed while evidence and clinical practice were changing. TOGETHER used a platform structure to evaluate interventions in high-risk outpatients in Brazil. Its Bayesian monitoring allowed accumulating outcome data to inform decisions about continuing, stopping, or prioritizing treatment arms. The important statistical question was not simply whether an estimated treatment effect looked favorable. It was how much posterior evidence supported benefit, how that evidence compared with a prespecified decision boundary, and what clinical outcome the comparison represented.

The fluvoxamine comparison illustrates the distinction between a Bayesian probability and a clinical conclusion. The trial’s primary outcome was a composite involving COVID-19-related transfer to a tertiary hospital or prolonged observation in an emergency setting. In the published comparison, 79 of 741 participants assigned fluvoxamine and 119 of 756 assigned placebo experienced the primary outcome. The reported posterior probability of superiority was 99.8%; the report also gave a probability of efficacy of 97.6%. These are statements about the specified Bayesian analysis and outcome, not proof that every patient benefits or that the treatment works in every setting.

In this tutorial, \(p_T\) denotes the probability of the adverse outcome under treatment and \(p_C\) the corresponding probability under control. A smaller \(p_T\) is favorable. A common posterior summary is \(P(p_T<p_C\mid\text{data})\), the posterior probability that the treatment event risk is lower. A trial may instead make its decision using an odds ratio, a relative risk, or a clinically meaningful threshold. The endpoint, parameter, prior, and decision boundary must therefore be read together.

A Simple Model for Understanding Posterior Evidence

For a transparent calculation, suppose each participant has a binary outcome and that, within each arm, outcomes are independent with a common event probability. If \(x_T\) events occur among \(n_T\) treated participants, then \(x_T\) has a binomial distribution with event probability \(p_T\); define the control quantities analogously. This model assumes the outcome is classified consistently, participants contribute independent observations, and the event risk is adequately represented by one probability per arm. It does not account automatically for covariates, changing risk over calendar time, or differences in follow-up.

For illustration, use independent Jeffreys priors \(p_T\sim\mathrm{Beta}(0.5,0.5)\) and \(p_C\sim\mathrm{Beta}(0.5,0.5)\). With a beta prior \(\mathrm{Beta}(a,b)\) and binomial data of \(x\) events in \(n\) participants, the posterior is \(\mathrm{Beta}(a+x,b+n-x)\). This follows because the binomial likelihood is proportional to \(p^x(1-p)^{n-x}\), so multiplying it by the beta density gives a beta density with the event and non-event counts added to its two parameters.

Independent arm-specific priors and likelihoods produce independent posteriors in this simple model. The posterior distribution of the risk difference \(p_C-p_T\) can then be calculated numerically, for example by simulation from each beta posterior. A normal approximation based on the difference in posterior means and the sum of posterior variances is useful for hand calculations, but it is an approximation, particularly when event probabilities are near zero or sample sizes are small.

How Efficacy and Futility Boundaries Work

An efficacy boundary is a prespecified requirement for sufficiently strong posterior evidence in favor of a treatment effect. A futility boundary identifies an arm for which the posterior chance of reaching a worthwhile benefit is sufficiently small that continuing is not justified under the design. These rules are asymmetric in purpose: an efficacy boundary supports a positive decision, while a futility boundary can stop an arm for lack of promise. A result that crosses neither boundary is inconclusive for that interim analysis.

A decision rule must specify the effect being judged. For example, \(P(p_T<p_C\mid\text{data})\) concerns any reduction in risk, however small. A rule about a clinically relevant reduction instead concerns an event such as \(P(p_T<0.8p_C\mid\text{data})\). High probability of some benefit does not necessarily imply high probability of a benefit large enough to matter. Likewise, failing an efficacy boundary is not equivalent to demonstrating no effect.

TOGETHER’s fluvoxamine report described both a 99.8% posterior probability of superiority and a 97.6% probability of efficacy; those reported probabilities should be interpreted using the trial’s own model and definitions. The study concluded that fluvoxamine met its efficacy criterion for the primary outcome. This is a decision within the trial’s prespecified framework, not a general rule that a posterior probability near 98% always establishes efficacy. Other interventions evaluated in the platform, including hydroxychloroquine, lopinavir-ritonavir, and ivermectin, did not yield the same efficacy conclusion; some arms were stopped after interim assessment for futility or lack of persuasive benefit. The exact interpretation of each decision depends on its arm-specific data and criteria.

Derivation, Assumptions, and Diagnostics

The conjugate update gives a compact justification for the calculations: a beta prior contributes \(a\) event-equivalents and \(b\) non-event-equivalents to the likelihood. After observing \(x\) events and \(n-x\) non-events, the posterior parameters are \(a+x\) and \(b+n-x\). For independent arms, the difference in risks has posterior variance equal to the sum of the arm-specific variances, because the covariance is zero under the independent model. This explains the variance addition used above; it would not be valid if the posterior risks were dependent.

The simple calculation is deliberately limited. A full platform analysis can model odds ratios or other effect measures, adjust for prognostic variables, accommodate concurrent controls, and reflect the planned interim structure. It may also use priors other than Jeffreys priors. If allocation probabilities change over time, participants accrued at different stages may face different allocation patterns; calendar-time changes in care or disease prevalence can then complicate an unadjusted arm comparison. The simple beta-binomial example assumes away these features and must not be mistaken for the official TOGETHER model.

Diagnostics should focus on whether the decision model still represents the accumulating trial. Check outcome definitions and missing follow-up, event counts by arm and calendar period, baseline imbalance, allocation probabilities, and whether the assumed treatment effect is compatible with observed subgroup or site patterns. A sudden change in standard care or circulating disease can make a single constant event probability questionable. Sensitivity analyses using alternative reasonable priors and model specifications help show whether a boundary crossing depends heavily on one assumption. With sparse data, posterior probabilities and normal approximations can be especially sensitive; calculate the posterior probability numerically rather than relying on a normal approximation when precision matters.

Comparison, Limitations, and Common Pitfalls

A fixed frequentist analysis typically specifies a test and sampling plan in advance, then controls a long-run type I error rate under repeated use of that plan. A Bayesian analysis instead updates a probability distribution for the parameter conditional on the prior and data, then applies a decision rule. Bayesian monitoring can make accumulating evidence directly interpretable for a decision, but it does not eliminate the need to prespecify boundaries, control operational bias, or evaluate performance of the complete adaptive design. Neither approach repairs a poorly defined endpoint or unrepresentative population.

An efficacy or futility decision is also not the same as a statement about every endpoint or population. TOGETHER’s composite outcome included prolonged emergency observation as well as hospital transfer. A favorable decision on that outcome should be described using that endpoint and the enrolled population. Generalizing to other treatments, variants, health systems, or lower-risk outpatients requires separate evidence.

• Mistake: “A 99.8% posterior probability means there is a 99.8% chance the treatment works for an individual.” Correction: it is a model-conditional probability statement about a defined population parameter or contrast, not an individual guarantee.
• Mistake: “Not crossing the efficacy boundary proves no effect.” Correction: the result may be uncertain, underpowered, or incompatible with the particular threshold; futility is a separate decision criterion.
• Mistake: “A high probability of any benefit proves a worthwhile benefit.” Correction: clinically meaningful benefit requires evaluating a prespecified magnitude, such as a risk reduction of at least a specified size.
• Mistake: “A simplified posterior calculation reproduces the platform’s decision.” Correction: verify the actual endpoint, model, prior, adjustment, and stopping rule before interpreting a reported decision.