Distinguishing Pending Outcomes From Dropout

Why Status Classification Matters

At an allocation update, an assigned participant may have a known outcome, an outcome that is still pending, or an outcome that will not be obtained. These situations are not interchangeable. Coding every unresolved record as a nonresponse can distort arm-specific response estimates; coding every unresolved record as pending can conceal loss to follow-up. A reliable update therefore begins with an explicit status rule, not with an assumption about what an unknown outcome would have been.

The key distinction is between the outcome’s ascertainment and the participant’s participation status. Someone may stop taking assigned treatment but still provide outcome data. Someone may withdraw from some study procedures but permit follow-up. Conversely, a participant may remain willing to participate while a record is temporarily delayed. Allocation updates must use the outcome information actually eligible under the protocol, not infer it from a label such as “withdrawn.”

Operational Definitions and Assumptions

“Pending” is a time-specific status: the same record may be pending today and resolved later. “Overdue” describes a missed expected date, not necessarily a final outcome status. It should trigger follow-up or adjudication under prespecified rules. It does not, by itself, establish that the participant has dropped out.

These definitions assume that the endpoint, its ascertainment window, the reporting deadline, permitted data sources, and follow-up procedures are specified before the update. The protocol must also state what data may still be collected or used after a participant withdraws. Applicable consent and data-use requirements take precedence: a previously recorded outcome must not be used if its use is prohibited.

For arm \(j\), let \(m_j\) be the number of participants with an eligible, resolved outcome at the update, and let \(r_j\) be the number of those outcomes classified as responses. The observed response proportion is \(r_j/m_j\), when \(m_j>0\). Pending records and unavailable outcomes are not included in either count. This calculation describes the available outcomes; it does not, by itself, guarantee an unbiased estimate of the response probability.

How Status Enters an Allocation Update

The following small allocation rule is used only to make the examples concrete. Suppose a prespecified two-arm procedure smooths each arm’s observed response proportion as

Here \(q_j\) is the probability assigned to the next eligible participant for arm \(j\). The added constants are part of this illustrative rule; they are not a universal correction. A real trial must use its prespecified allocation algorithm, including any probability bounds or other safeguards. The status rules below apply regardless of the particular algorithm.

A resolved response contributes one to \(r_j\) and one to \(m_j\). A resolved nonresponse contributes one to \(m_j\), but not to \(r_j\). A pending or unavailable outcome contributes to neither. Assignment history remains recorded for every randomized participant, including those without resolved outcomes; it is the outcome summary for the update, not the historical assignment record, that excludes unresolved outcomes.

A Short Justification: Why “Unknown” Is Not “No”

Let \(Y_i\) equal 1 for a response and 0 for a nonresponse, and let \(R_i=1\) when participant \(i\)’s outcome is resolved at the update and 0 otherwise. If every unresolved record is coded as a nonresponse, the resulting proportion is \(\sum_i R_iY_i/n\), where \(n\) is the number assigned. Under the simplifying conditions that participants share response probability \(p\) and resolution is independent of response, with resolution probability \(c\), its expectation is

Unless all outcomes are resolved (\(c=1\)), this expectation is below \(p\). If resolution depends on response, the distortion need not take this simple form or even have the same direction. In contrast, excluding unresolved records avoids assigning them invented outcome values, but the observed proportion \( \sum_i R_iY_i/\sum_i R_i \) can still be biased when the resolution process is informative. Status classification prevents a coding error; it does not solve every missing-data problem.

Withdrawal, Treatment Discontinuation, and Follow-Up

A withdrawal record should say what stopped, when it stopped, and what data collection remains permitted. A participant who discontinues treatment but completes the endpoint assessment has a resolved outcome for the allocation summary. A participant who withdraws from treatment and further contact before the endpoint may have an unavailable outcome. A participant who declines additional visits but permits a prespecified record-based outcome source may still have an ascertainable endpoint, if use of that source is allowed.

Do not infer a binary response from the reason for withdrawal. For example, treatment intolerance is not automatically a nonresponse unless the endpoint definition explicitly classifies it that way. Likewise, withdrawal is not automatically a nonresponse or an exclusion: the endpoint definition and the permitted data determine whether a valid outcome exists.

Overdue Records, Diagnostics, and Failure Modes

An overdue record requires a status decision under the protocol, not an automatic value assignment. The study may still be resolving a laboratory result, contacting a participant, or checking an allowed data source. If specified attempts and deadlines establish that no outcome can be obtained, the record can be classified as unavailable or lost to follow-up according to the study’s terminology. If information later becomes available and its use is permitted, the status history should show the change and the next update can use the resolved outcome.

Useful diagnostics focus on whether statuses are applied consistently and whether unresolved records are accumulating selectively. At each update, reconcile assigned participants against counts of resolved responses, resolved nonresponses, pending, overdue, withdrawn, and unavailable outcomes. These categories should be mutually interpretable under the data system’s rules, and every status change should have a date and reason.

• Review the age of pending and overdue records, including how many have passed the endpoint deadline.
• Compare resolution and loss-to-follow-up rates by arm; marked differences may signal a data-quality issue or informative resolution.
• Check whether withdrawal reasons are being incorrectly translated into response values.
• Audit whether outcome sources and post-withdrawal data use match the consent and protocol rules.
• Verify that each update uses a defined data cutoff and does not silently mix records from different reporting periods.

The practical alternative is not simply “include” versus “exclude.” One option is to update from resolved outcomes only, as in the examples. Another is to delay an update until more outcomes resolve, which may reduce reliance on a sparse or unrepresentative set but slows adaptation. A model-based procedure may use information about unresolved outcomes, but it requires explicit assumptions and validation; it must not be introduced by quietly treating predictions as observed responses. The choice belongs in the prespecified design, not in ad hoc decisions made after seeing the data.

Common Mistakes and Limits

• Wrong: “The participant has not responded yet, so count a nonresponse.” Correction: If the ascertainment window remains open, the outcome is pending. Count a nonresponse only when the endpoint is resolved and meets the protocol’s nonresponse definition.
• Wrong: “Anyone who stops treatment has dropped out.” Correction: Treatment discontinuation, withdrawal from visits, and inability to ascertain an outcome are distinct statuses.
• Wrong: “An overdue outcome is automatically loss to follow-up.” Correction: Apply the specified contact and deadline criteria; overdue is a warning status until adjudicated.
• Wrong: “Excluding missing outcomes guarantees valid response estimates.” Correction: Exclusion avoids fabricating outcomes, but the resolved cases can still be selected. Inspect resolution patterns and acknowledge the limitation.

Status rules improve the integrity and reproducibility of allocation updates, but they cannot recover an outcome that was never observed or ensure that observed outcomes represent all randomized participants. When unresolved outcomes are common or their resolution differs by arm or likely outcome, the allocation rule may respond to an unrepresentative subset. That limitation should be considered when interpreting the resulting assignments and when planning the trial’s data-collection procedures.