Attrition bias is the systematic distortion that results when participants who drop out of a study differ, in some way related to the outcome, from participants who stay in it. It is a longitudinal-design problem specifically: the sample you drew at baseline is not the sample you are left with by the final follow-up, and if who leaves is not random with respect to the outcome, the comparison you report is no longer the comparison you set out to make. Losing participants is not, by itself, attrition bias — every real study loses some. Attrition becomes bias only when the loss is differential: unequal in size or in kind across the groups being compared.
Attrition Is Not the Same as Sampling Bias
Attrition bias is a close cousin of sampling bias but occurs at a different stage. Sampling bias distorts who gets into the study in the first place — a flawed sampling frame or a non-random recruitment mechanism. Attrition bias distorts who is left in the study by the time it is analyzed. A study can start with a perfectly representative, unbiased sample and still end up with a badly biased one if dropout during follow-up is not random. The two problems compound rather than substitute for each other: a study can have both, and each requires a different fix. Sampling bias is addressed at recruitment; attrition bias is addressed through retention design, monitoring during the study, and analytic handling of missing outcomes afterward.
Attrition bias also differs from ordinary missing data in the statistical sense. Missing data can arise from a single unanswered survey item with no bearing on retention; attrition specifically means a participant leaves the study before providing the outcome measurement altogether, in a longitudinal or multi-timepoint design. All attrition produces missing data, but not all missing data comes from attrition.
Overall Attrition vs. Differential Attrition
These are the two figures that matter, and they answer different questions.
Overall attrition rate
The proportion of the total baseline sample lost to follow-up by the analysis point, regardless of group. A high overall attrition rate (conventions vary by field, but rates above roughly 20% are widely treated as a threshold warranting explicit scrutiny) threatens the study’s external validity and statistical power even if it is perfectly balanced across groups — the completers may no longer represent the population the study intended to describe, and the effective sample size for the final analysis shrinks.
Differential attrition rate
The difference in attrition rate between comparison groups (for example, treatment vs. control, or exposed vs. unexposed). This is the more dangerous figure, because differential attrition is what actually introduces bias into a group comparison. A study can have low overall attrition (say, 8%) and still be seriously compromised if nearly all of that loss is concentrated in one arm — for instance, participants dropping out of a treatment arm because of side effects, or dropping out of a control arm because they weren’t experiencing perceived benefit. Differential attrition breaks the balance that randomization (in a trial) or matching (in an observational design) was intended to create, in the same way that post-randomization exclusion undermines an intention-to-treat analysis — both re-introduce a selection process after the design already tried to eliminate one.
A useful way to hold both figures in mind: overall attrition tells you how much statistical power and generalizability you’ve lost; differential attrition tells you whether the comparison itself is still valid. A study can survive a lot of the first kind of damage. It cannot fully recover from a large amount of the second.
How to Detect Attrition Bias
Attrition bias cannot be observed directly — you cannot measure the outcome for someone who left the study. What is checked instead is whether dropout looks random with respect to everything that was measured before people left. None of these checks proves the absence of bias; they narrow the range of plausible explanations.
1. Compare completers to dropouts on baseline characteristics
For every baseline variable collected (demographics, disease severity, prior exposure, baseline value of the outcome itself), compare the group of participants who completed the study against the group who dropped out. If dropouts differ meaningfully from completers on a variable that plausibly relates to the outcome, that is direct evidence attrition is not random. This comparison should be reported for the full sample and, separately, within each study arm — a balanced overall comparison can still hide an imbalance that exists only in one group.
2. Compare attrition rates and reasons across groups
Report the number lost to follow-up in each arm and, wherever it was recorded, why. A treatment arm losing participants to adverse events and a control arm losing participants to loss of contact are not the same kind of missingness, even if the raw counts are similar — the first is plausibly related to the outcome (informative, or “not missing at random”), the second is more plausibly unrelated (closer to “missing at random” or “missing completely at random”). The reason for dropout, not just its rate, is often the more informative signal.
3. Test for a group-by-attrition interaction
Where the design allows it, a formal test of whether the association between group assignment and dropout is statistically significant provides a more direct check than eyeballing a table of baseline comparisons, though it shares the same fundamental limit: it can only use variables that were actually measured before dropout occurred.
4. Sensitivity analysis under different missingness assumptions
Because no baseline comparison can rule out bias driven by the unmeasured outcome itself, the standard complementary step is a sensitivity analysis: re-running the primary analysis under alternative assumptions about what the missing outcomes would have been (for example, best-case/worst-case imputation, or multiple imputation under a missing-at-random assumption) to see whether the study’s conclusion holds up across plausible scenarios. If the result flips under a modestly pessimistic assumption about dropouts, the finding is fragile with respect to attrition.
Reporting Requirements: The CONSORT Flow Diagram
For randomized trials, attrition reporting is not optional or left to author discretion — it is a checklist requirement under the CONSORT reporting guideline (originally CONSORT 2010, superseded by CONSORT 2025). The CONSORT flow diagram traces participants through four stages — enrollment, allocation, follow-up, and analysis — and requires, at the follow-up stage specifically, the number lost to follow-up or discontinued in each arm and the reason. This structural requirement is unchanged across the 2010-to-2025 revision.
The point of putting this in a mandatory diagram rather than a sentence of prose is precisely to make differential attrition visible at a glance: a reviewer or reader can see, arm by arm, whether the losses are balanced and what caused them, without having to dig for the numbers in the text. A trial report that gives an overall dropout percentage without an arm-by-arm breakdown and reasons is not meeting the CONSORT standard, even if the number itself is accurately stated.
Observational study guidelines take the same position for the same reason: STROBE-style reporting for cohort studies likewise expects the number of participants followed up and lost to follow-up to be reported at each stage, because loss to follow-up in a cohort study creates the same differential-attrition risk as dropout in a trial.
Mitigating Attrition Bias: Design Through Analysis
Attrition bias is cheaper to prevent than to correct after the fact, but no single stage fully solves it — mitigation is layered across the life of the study.
At the design stage
- Build in retention from the start. Minimize participant burden, plan realistic follow-up intervals, and budget for retention activities (reminder contacts, flexible scheduling, reimbursement for time) rather than treating retention as something to manage only once dropout is already happening.
- Inflate the target sample size for expected attrition. A sample-size calculation should account for the anticipated dropout rate so that the study is still adequately powered on the completers, not just at enrollment, in the same way an effect-size assumption feeds that calculation — this is core planning, not an afterthought.
- Collect baseline data on variables likely to predict dropout. You can only check whether attrition is differential on variables you measured before people left, so baseline measurement should anticipate this need.
During data collection
- Track attrition in real time, by arm. Monitoring cumulative and differential dropout as the study runs (rather than discovering it only at analysis) allows a trial’s data and safety monitoring process to flag a developing imbalance while it can still be investigated.
- Record the reason for every withdrawal. This is the single input that most improves later interpretation — a code for “adverse event,” “moved away,” “withdrew consent, no reason given,” and so on turns a bare dropout count into something that can be reasoned about.
At the analysis stage
- Report both overall and differential attrition, per the CONSORT flow diagram structure above, regardless of whether the trial guideline formally applies to the study design in question.
- Use intention-to-treat as the primary analysis for randomized designs, which keeps participants in their assigned group and forces an explicit decision about how missing outcomes are handled, rather than silently excluding them. See CASRAI’s guide to intention-to-treat analysis for how this interacts with post-randomization exclusions specifically.
- Use principled methods for missing outcomes — multiple imputation or mixed-effects models that use all available data under a stated missingness assumption — rather than a simple complete-case analysis, and state which missingness assumption (missing completely at random, missing at random, or missing not at random) the chosen method depends on.
- Run and report sensitivity analyses so a reader can judge how much the conclusion depends on assumptions about the participants who were lost.
Frequently Asked Questions
What attrition rate is considered acceptable?
There is no single universal threshold, and field conventions vary, but an overall attrition rate above roughly 20% is widely treated as warranting explicit justification and sensitivity analysis. The bigger determinant of risk, however, is not the overall rate but whether attrition is differential across groups — a low overall rate concentrated in one arm can be more damaging than a higher rate evenly spread across arms.
Is attrition bias the same as survivorship bias?
They are related but not identical. Survivorship bias is the broader phenomenon of drawing conclusions only from the subjects that “survived” some selection process, in any context, including outside longitudinal research entirely (the classic example being reasoning only from surviving aircraft, not the ones that were shot down). Attrition bias is the specific instance of that pattern that arises from participant dropout in a study with follow-up over time.
Can attrition bias be fully corrected statistically after the fact?
No analytic method can fully correct for attrition bias driven by data that is missing not at random — that is, where the reason a participant left is itself tied to the unobserved outcome value. Methods like multiple imputation and mixed-effects models can produce less-biased estimates under weaker assumptions (missing at random) than a complete-case analysis would, and sensitivity analyses can show how fragile a conclusion is to plausible missingness scenarios, but none of this substitutes for minimizing attrition, and differential attrition, during data collection.
Does attrition bias apply outside clinical trials?
Yes. Any longitudinal design with repeated measurement over time — cohort studies, longitudinal surveys, panel studies in the social sciences, multi-wave program evaluations — is exposed to attrition bias whenever the people who stop participating differ systematically from the people who continue. The CONSORT flow diagram is specific to randomized trials, but the underlying detection and mitigation logic (compare completers to dropouts, check whether loss is differential, run a sensitivity analysis) applies to any design with follow-up.
