Interpreting Interim Efficacy Displays

Why Interim Efficacy Displays Need a Structured Read

An interim efficacy table or plot is a snapshot, not a final verdict. Participants may have different amounts of follow-up, some may not yet have reached the endpoint assessment, and a small number of observations may noticeably affect a summary. Even an accurately calculated percentage can be misleading if its denominator or observation window is unclear.

The first task is therefore not to decide which treatment “looks better.” It is to establish what data the display summarizes, how mature those data are, and whether the displayed summaries agree with one another. These checks help a reviewer describe the pattern faithfully while avoiding conclusions that the interim information cannot support.

Start With the Endpoint, Time Point, and Denominator

Before interpreting a value, identify the endpoint, its measurement time or event definition, the data cut, and the population included in its denominator. In a trial with a response assessed at week 12, for example, a participant randomized shortly before the cut may not yet have had the opportunity to complete that assessment. A week-12 response percentage among participants assessed is not automatically the same quantity as a response percentage among everyone randomized.

Common denominators include the number randomized, the number who started treatment, the number with an assessment in the endpoint window, and the number included in a particular analysis. These are not interchangeable. A display should make clear which is used, and whether unavailable or missing outcome values are excluded or handled by a specified rule. A count of responses without its denominator is insufficient for interpreting a response proportion.

For a binary endpoint, the observed proportion among assessed participants is \(r/n_{\mathrm{assessed}}\), where \(r\) is the number meeting the endpoint and \(n_{\mathrm{assessed}}\) is the number with an evaluable assessment. This is a descriptive summary of those assessed. It does not, without further assumptions or a prespecified missing-data method, establish the response proportion in all randomized participants.

Assess Maturity and Follow-Up Balance

Maturity should be judged against the endpoint’s required observation period and the data cut, not inferred from enrollment totals alone. Useful display information includes the number randomized, time since randomization, number reaching the endpoint window, number assessed, and number still awaiting assessment. For a time-to-event endpoint, relevant context includes the amount and distribution of follow-up, events observed, and participants censored or still event-free at the cut. The median follow-up by itself may not reveal all differences in the follow-up distributions.

Unequal follow-up matters because participants observed for longer have had more opportunity to experience an event. Comparing raw event counts between arms with different follow-up can therefore be uninformative. A crude rate per person-time can provide a descriptive adjustment for unequal exposure time, but it does not automatically account for differences in event timing, changing risk, or informative censoring. A time-to-event curve or model-based summary may use the available event-time information differently; it still requires its own assumptions and sufficient maturity.

Look for Influential Observations and Internal Consistency

A summary can be numerically correct yet depend heavily on one observation. This is especially important for small interim samples, continuous endpoints with extreme values, and subgroup displays. A simple diagnostic is to compare the reported summary with a version recalculated after omitting one observation at a time. This is a sensitivity check, not a license to remove an inconvenient value. An observation should be excluded only if a prespecified rule or a documented data-quality reason supports exclusion.

For a mean difference, the leave-one-out calculation makes the source of influence visible: remove one value, recompute the group mean using the remaining observations, and compare the resulting contrast with the original. If the contrast changes substantially, investigate the observation’s provenance, units, endpoint derivation, visit timing, and any relevant data query. The check does not decide whether the value is valid or what the treatment effect is.

Consistency checks apply across the entire package. Recalculate displayed percentages from their counts and denominators; verify units, endpoint labels, visit windows, treatment-arm labels, and data-cut dates; and compare tables with plots that purport to show the same analysis. Different displays can legitimately use different populations or denominators, but that distinction must be explicit.

Interpret Patterns Without Overreading Them

A practical reading sequence separates data checks from efficacy interpretation:

The main alternatives to a raw interim count or percentage are summaries suited to the endpoint and observation process, such as a time-to-event display or a model-based estimate. Those alternatives can be more informative about timing or adjustment, but they do not remove the need to check maturity, definitions, and assumptions. A raw rate, a curve, and a model estimate may differ for legitimate reasons; their interpretation depends on what each estimates and which data it uses.

Common failure modes include treating “not yet assessed” as “no response” without a prespecified rule, comparing event counts when follow-up differs, relying on a mean without checking extreme values, and assuming a discrepancy must indicate a treatment effect or a data error before checking populations and denominators. The correction is to trace each display back to its endpoint, data cut, analysis set, and calculation. In an adaptive trial, interim efficacy displays are particularly easy to overread because the information is still accumulating; a visually clear separation is not a substitute for checking how much and what kind of information produced it.