Were All Recruited Participants Included in the Analysis?

What happened to everyone who entered the study?

A study recruits 800 participants but reports its primary result using 653. Where did the other 147 go?

Some may never provide the outcome. Others may withdraw, become lost to follow-up, violate eligibility criteria, discontinue treatment, lack required covariates, or be deliberately excluded from an analysis. These situations are not interchangeable. To interpret the evidence, trace participants from recruitment or allocation through follow-up and into the specific analysis that produced the result.

Compare recruitment with the analysis-specific population

Trace participants from entry to the final estimate

Recruitment is only the beginning of participant flow

A headline sample size tells you how many people entered some stage of the research. It does not necessarily tell you how many contributed to the result you are interpreting.

STROBE recommends that observational studies report participant numbers at relevant stages, including those potentially eligible, examined for eligibility, confirmed eligible, included, completing follow-up, and analyzed. It also recommends giving reasons for non-participation and considering a flow diagram.

CONSORT 2025 similarly emphasizes documenting participant flow in randomized trials, including allocation, receipt of intervention, follow-up, and analysis.

Do not collapse all participant losses into “dropout”

Several different processes can reduce the number of participants contributing to an analysis.

CONSORT explicitly distinguishes losses to follow-up from investigator-determined exclusions such as ineligibility, treatment withdrawal, and poor adherence because their implications differ.

In randomized trials, understand what intention-to-treat means

The intention-to-treat principle is central to randomized trials. In its strict form, participants are analyzed in the groups to which they were randomized, regardless of subsequent adherence or deviations from the assigned intervention.

CONSORT 2025 explains that preserving randomization involves including all randomized participants and retaining them in their assigned groups. It also acknowledges an important practical difficulty: if outcomes are missing, including every randomized participant in a strict intention-to-treat analysis may require analytical handling of those missing outcomes.

The two may coincide, but you should verify that rather than assume it.

“Intention-to-treat” on the page does not settle the question

Authors sometimes use labels such as “intention-to-treat,” “modified intention-to-treat,” “per protocol,” “complete case,” or “safety population.” Those labels can conceal substantially different inclusion rules.

CONSORT 2025 recommends defining who is included in each analysis and in which group rather than relying on vague labels such as modified intention-to-treat or per protocol. For example, a so-called modified intention-to-treat analysis might exclude randomized participants who never received treatment, lacked a post-baseline measurement, or failed another criterion.

Ask for the rule, not merely the name.

Withdrawal from treatment is not the same as withdrawal from the study

A participant may stop taking a medication, attending sessions, using an educational platform, or following an assigned program while still providing outcome measurements.

This distinction matters because excluding participants solely because they did not adhere can undermine the randomized comparison and shift the question toward outcomes among adherent participants. Depending on the estimand and analytical strategy, continued outcome collection after treatment discontinuation may remain highly relevant.

Modern trial methodology makes this point more explicit through the estimand framework. ICH E9(R1) treats events occurring after treatment initiation, such as treatment discontinuation or use of additional therapy, as intercurrent events whose handling should reflect the clinical question being estimated rather than being addressed automatically by deleting participants.

Missing outcomes are especially important

A participant can remain formally enrolled while contributing no measurement for the outcome of interest. If those missing outcomes are simply omitted, the analysis may contain only participants with observed outcome data.

That loss can reduce precision. More importantly, bias can arise if outcome availability is related to prognosis, treatment, exposure, or the outcome itself in ways not adequately addressed by the analysis.

CONSORT 2025 recommends reporting the extent and reasons for missing data and explaining the analytical approach used. It notes that complete- or available-case analyses and imputation or model-based approaches depend on assumptions about the missing-data process.

Missing covariates can also exclude participants

You can have a perfectly observed outcome and still disappear from an adjusted model.

Suppose 900 participants have the primary outcome, but one covariate used in the fully adjusted regression is available for only 760. A complete-case model requiring every covariate may analyze only those 760 participants.

That is why the question is not merely whether participants completed follow-up. You need to inspect what data were actually analyzed.

Attrition matters because the remaining participants may differ

If losses occur entirely independently of variables relevant to the analysis, their main consequence may be reduced information and precision. If participants who remain differ systematically from those who are missing in ways related to the outcome or comparison, interpretation becomes more difficult.

For example, participants experiencing severe adverse effects may be more likely to discontinue follow-up. Students struggling most with an intervention may be more likely to stop completing assessments. In either case, analyzing only those who remain could produce a misleading picture under some missing-data mechanisms.

The existence of attrition does not tell you its direction or magnitude. You need information about how much occurred, why it occurred, whether it differed across groups, and how the analysis addressed it.

Different outcomes can have different participant counts

A trial may have almost complete data for mortality but substantial missingness for a questionnaire. A cohort may have laboratory measurements for one subgroup and administrative outcomes for nearly everyone.

CONSORT 2025 therefore recommends reporting participant numbers for each primary and secondary outcome and at each relevant time point, rather than implying that one analysis count applies universally.

When you interpret a particular outcome, identify the participants who contributed to that outcome.

Per-protocol analyses answer a different question

A per-protocol analysis generally focuses on participants meeting specified protocol-related criteria, such as sufficient adherence and absence of certain major deviations. It can sometimes address a scientifically relevant question, but it does not preserve the original randomized groups in the same way as an analysis based on randomized assignment.

Simply comparing adherent participants between groups can introduce differences related to why participants adhered. More sophisticated causal methods may be needed for some per-protocol questions.

Therefore, do not ask whether intention-to-treat or per-protocol is universally “better.” Ask which question the analysis is intended to answer and whether its methods support that question.

Exclusions made after seeing the data deserve particular attention

Some exclusions are planned before analysis. Others arise after researchers inspect data, discover unusual observations, encounter protocol problems, or see analytical results.

Post hoc exclusion is not automatically inappropriate, but it can create additional researcher discretion. Determine whether important exclusion rules were prespecified and whether conclusions change when reasonable alternative rules are used.

This connects directly to what was prespecified and what appears to have been decided later.

Participant flow and analysis populations are related but distinct

A flow diagram can tell you how many people were assessed, enrolled, allocated, lost, and analyzed. It may not tell you everything about how the final statistical model treated missing values, repeated observations, protocol deviations, or intercurrent events.

Use participant flow to establish what happened to people. Then use the statistical methods to establish how their data were handled.

When 500 randomized participants become several different analysis populations

Common mistakes when tracing participants into the analysis

Make every important result carry its denominator

A useful appraisal habit is to attach an analysis population to every major estimate. Instead of writing “the trial found a difference,” write down how many participants were randomized, how many contributed to the relevant outcome, and how those participants were analyzed.

The aim is not to demand that every study retain every observation under all circumstances. Real studies encounter missing data, withdrawals, protocol deviations, and incomplete measurements. The important task is to understand what happened and whether the analysis and interpretation remain aligned with the resulting evidence.

Before interpreting the result, account for the participants

Questions about who was included in the analysis

Recruitment counts do not tell you who produced the final estimate

How to Cite This Guide

This guide is intended to be read, shared, and used in research, teaching, and academic work. If you draw on its ideas, explanations, or other content, please acknowledge the source by citing the guide. Doing so gives appropriate credit and helps your readers locate the original resource.