What Partial Disclosure Can Reveal
An interim adaptation may require a trial to change its planned sample size, pause enrollment, or review an emerging treatment-effect summary. Communicating that an action occurred is not the same as releasing the data that prompted it. Yet an operational update can still carry information: recipients may know the adaptation rule, its thresholds, or the usual reasons for a pause, and use those facts to infer what the interim results looked like.
The central question is therefore not simply whether a message names treatment groups. It is what can a recipient reasonably infer, given the message and what they already know? A message that omits treatment labels may still be informative, while a message that reports a sample-size increase may reveal little about the direction of a treatment difference if the increase was determined only by a blinded estimate of variability.
For this discussion, a treatment contrast is a specified difference between treatment groups, with its direction tied to the chosen subtraction order. A blinded update does not identify which group received which treatment in the interim results; an unblinded summary preserves that identification. Blinding can reduce direct disclosure, but it does not guarantee that recipients cannot infer a difference.
Three Common Updates, Three Information Levels
A sample-size change is usually an operational quantity: it tells recipients how many participants are now planned, not necessarily which treatment is performing better. Its meaning depends on the rule that produced it. If an update depends only on a pooled variance or a pooled event rate, it may indicate that a nuisance quantity differed from planning assumptions. If it depends on an unblinded treatment contrast, the same announcement may indirectly convey information about that contrast.
An enrollment pause is also an action rather than a statistic. By itself, it reports that enrollment has stopped temporarily. Its reason, duration, timing, and unusualness may add substantial context. A pause for a disclosed supply interruption says something different from a pause announced while a known safety review is under way. If the reason is withheld, recipients may still interpret the pause in light of what they know about the trial and its review process.
A treatment-effect summary is generally the most direct of these three disclosures. An estimate with treatment labels and a stated contrast direction can reveal both direction and magnitude. A confidence interval communicates uncertainty as well as the estimate; a test result may communicate evidence against a null hypothesis. Even when a summary is masked, coded labels or an identifiable ordering can allow recipients to work out which group is which.
| Disclosure | What it directly communicates | What may also be inferred |
|---|---|---|
| Sample-size change | The revised planned enrollment | The type of interim quantity or threshold that triggered the change |
| Enrollment pause | That enrollment has stopped temporarily | Possible reasons, especially if the stated rationale, timing, or duration is informative |
| Treatment-effect summary | An estimate, direction, or measure of uncertainty, if reported | Which treatment appears better and how strong or precise the interim difference may be |
Worked Example: Interpreting a Sample-Size Increase
Suppose a hypothetical two-arm trial plans equal allocation and uses a continuous outcome. Its blinded sample-size rule uses a prespecified target difference \(\delta=5\) points, a two-sided \(\alpha=0.05\), and 90% power. For a normal-approximation calculation with a common standard deviation \(\sigma\), the approximate required number per group is
Here \(z_{1-\alpha/2}=1.96\) and \(z_{1-\beta}=1.282\). If the planning standard deviation is 12 points, then
Rounding up gives 122 participants per group, or 244 in total. If a blinded interim re-estimation instead uses a pooled standard deviation of 14 points, the same rule gives
The revised target is 165 per group, or 330 in total. Under the stated rule, announcing the increase reveals that the re-estimated variability led to a larger target. It does not by itself show which treatment has the better mean: the calculation used a pooled standard deviation and a fixed target difference, not the observed direction of the treatment contrast. This interpretation assumes that the rule was followed and that no other unblinded information affected the decision. If recipients know the rule also uses an unblinded contrast, the same announcement may carry more information.
A Short Justification: Information Depends on the Rule
A useful way to formalize indirect disclosure is to compare how likely the same message would be under different interim states. Let \(C_+\) denote an interim contrast favoring treatment B over treatment A, and \(C_-\) a contrast favoring A over B. Let \(D\) be a message recipients receive. Bayes' rule gives
The first ratio on the right is the likelihood ratio of the disclosure. If the message is equally likely under either direction, that ratio is 1, and it does not change the prior odds between \(C_+\) and \(C_-\). If the message is much more likely under one direction, it shifts those odds. This is a comparison of beliefs about possible interim states, not a claim that the message proves either state.
For example, under a model in which a sample-size change is determined solely by pooled variability, and that variability has the same distribution under \(C_+\) and \(C_-\), the sample-size message has equal likelihood under the two directions. It does not distinguish which treatment is favored under those assumptions. By contrast, a rule that increases enrollment only when an unblinded contrast falls in a specified range can make the announcement informative about that contrast. The conclusion depends on the actual rule and its inputs, not on the words “sample-size change” alone.
Worked Example: Reading an Enrollment-Pause Notice
In a hypothetical trial with a planned total enrollment of 240 participants, enrollment stops temporarily after 120 have enrolled. The notice states that the pause will last about two weeks because a shipment of study materials is delayed. It gives no interim outcomes, arm counts, or treatment comparisons.
Direct disclosure. Recipients learn that recruitment is paused, that half the planned total has enrolled, and that the stated operational reason is a supply delay. They do not receive a treatment-effect estimate.
What the notice does not establish. The pause alone does not show that either treatment is more effective or less safe. If the supply delay is genuinely the only trigger and the notice is accurate, the stated rationale supports an operational explanation rather than a data-driven one. It still does not logically rule out other information unless recipients know the relevant process.
How context changes the inference. If recipients instead know that pauses are issued only after a safety review, the same bare announcement may make them suspect that such a review is under way. That suspicion depends on their prior knowledge of the pause rule. The pause is not itself a numerical estimate of comparative risk, and its presence alone cannot quantify a treatment difference.
What an Effect Summary Reveals
The more detail an effect summary includes, the more directly it may convey interim information. Consider the contrast \(B-A\) for a lower-is-better continuous score. An estimate below zero favors B on that score; an estimate above zero favors A. A standard error and confidence interval indicate the summary's precision under the model used. However, an interim interval or test result must be interpreted in light of the trial's monitoring plan. A conventional confidence interval or \(p\)-value calculated as if there had been only one analysis may not have its usual final-analysis error properties after repeated interim looks.
Worked Example: Interpreting an Interim Treatment Summary
Suppose a hypothetical interim report gives 80 participants per group, with mean scores of 52.0 in A and 47.8 in B. The score is measured in points, and lower is better. The sample standard deviations are 8.0 points in A and 8.5 points in B. Assume independent groups and use a normal approximation for the mean difference.
Define the reported contrast as \(B-A\). Its estimate is
The estimated standard error is
A conventional 95% normal-approximation interval is
The estimate and interval favor B because the reported \(B-A\) contrast is negative and the entire interval is below zero. The interval describes uncertainty in this interim estimate under the calculation's assumptions; it does not mean there is a 95% probability that the fixed treatment contrast lies inside these particular endpoints. Nor should this ordinary interval automatically be treated as a final trial conclusion: the interim monitoring procedure may require adjusted inference. This report reveals substantially more about the interim comparison than either a sample-size target or a pause notice without an effect summary.
Checking a Disclosure Before It Is Communicated
A practical review asks what the recipient can combine with the message. The same number can be innocuous to someone who does not know the adaptation rule and revealing to someone who knows its thresholds and timing. Review the proposed disclosure against the design, the adaptation rule, and information already available to the intended audience.
Write down exactly what is being reported: a revised target, a pause, an interim estimate, or a reason for an action. Do not treat these as interchangeable.
Determine whether the action depends on blinded pooled data, an unblinded treatment contrast, a safety summary, or more than one input. The inputs define what the action can signal.
Ask whether someone who knows the rule could work backward from the message to infer a threshold crossing, a contrast direction, or a likely reason for a pause.
Where the operational purpose permits, avoid unnecessary arm labels, subgroup counts, exact trigger values, or outcome summaries. A limited message should still be accurate and useful.
Warning signs include a sample-size change with a known, effect-dependent trigger; an exact pause duration that maps to a predefined review outcome; and a masked summary that retains an identifiable A/B ordering. A second warning sign is a communication whose meaning changes sharply depending on what recipients already know. These are reasons to examine the disclosure, not proof that recipients have inferred the correct interim result.
Common Pitfall and Limitations
Wrong reasoning: “No treatment labels were released, so nobody learned anything about comparative effects.” This is too strong. A known adaptation rule, a stated reason, or an identifiable coded ordering may make a label-free announcement informative. The correction is to assess the message together with the rule and recipients' prior information, rather than to equate omitted labels with no disclosure.
The reverse mistake is also common: treating every operational change as evidence that one treatment is better or worse. A larger sample size may reflect greater pooled variability, a conservative rule, or a fixed operational constraint; a pause may have a logistical cause. Without knowing the rule and its inputs, the action does not identify a unique interim result.
Partial disclosure is not a substitute for prespecifying the adaptation and analysis procedures. Communicating an interim action does not, by itself, determine whether the trial's statistical error control is preserved; that depends on the design and inferential method. But disclosure can shape how people understand the trial and can expose clues beyond the intended summary. The appropriate aim is therefore not to promise that a message reveals nothing, but to communicate only what is needed, accurately state what it does and does not show, and avoid presenting interim summaries as unqualified final conclusions.
Check Your Understanding
Answer each question, then compare your reasoning with the solutions below.
- A blinded rule increases enrollment only when a pooled standard deviation exceeds a threshold. Does announcing an increase reveal which group has the better mean under the rule as stated? What assumption supports your answer?
- A trial announces a pause but gives no reason. Does this establish that a safety concern caused the pause? Explain what additional information would affect the inference.
- A report gives a negative \(B-A\) estimate for a lower-is-better outcome. Which group does the estimate favor, and what convention must be checked?
- Why might an ordinary 95% interim confidence interval fail to provide the usual final-analysis interpretation in a trial with repeated interim looks?
- Give one example of a disclosure that appears masked but could still be informative.
Solutions. (1) No, not under the rule as stated: the update uses pooled variability, not the direction of the group means. The conclusion assumes the rule has no other unblinded inputs and that the pooled quantity does not distinguish the contrast directions in the model being used. (2) No. A pause may have a safety, efficacy, logistical, or other reason. The recipient's knowledge of the pause criteria and any stated rationale would change how informative the notice is. (3) It favors B, because the contrast is defined as \(B-A\) and lower scores are better; reversing the subtraction order would reverse the sign interpretation. (4) Repeated interim looks affect the behavior of ordinary intervals and tests across the full monitoring process. The trial's monitoring plan may therefore require an adjusted procedure before making a final inferential claim. (5) A report may use coded labels but preserve a known A/B ordering, or announce an adaptation whose trigger is known to depend on an unblinded contrast. In either case, recipients may infer comparative information without seeing explicit treatment names.