Differential attrition is when participants drop out of a study at different rates depending on which group they were assigned to. In a clinical trial comparing a new drug to a placebo, for example, more people in the drug group might quit because of side effects, while more people in the placebo group might quit because they aren’t getting better. This uneven dropout pattern is one of the most serious threats to the reliability of study results, because it can destroy the very thing randomization was designed to protect: two groups that are comparable at the start.
How It Differs From Overall Attrition
Every long-running study loses participants. People move, lose interest, get busy, or simply stop showing up. When dropouts happen at roughly the same rate across all groups and for similar reasons, that’s general attrition. It shrinks your sample size and reduces statistical power, but it doesn’t necessarily skew results in one direction.
Differential attrition is a different problem. Because the dropout rates or reasons differ between groups, the people who remain in each group are no longer equivalent. Randomization initially made the groups comparable on every characteristic, measured or not. Uneven attrition undoes that balance, reintroducing the selection bias that randomization was supposed to eliminate.
Why It Threatens Study Results
The core issue is that the people who drop out are often systematically different from those who stay. If patients withdraw because of disease progression or treatment side effects, their health outcomes would have been, on average, worse than those of patients who completed the study. Losing these participants from one group but not the other changes what the remaining data actually represents.
Researchers classify missing data into three categories based on the pattern of dropout. Data is “missing completely at random” when attrition has no connection to anything in the study, observed or unobserved. It’s “missing at random” when dropout relates to characteristics the researchers have already measured (age, baseline health) and can account for statistically. The most damaging category is “missing not at random,” where dropout is driven by unmeasured factors like how sick a person actually feels or how badly a treatment is working for them. When data is missing not at random, no analytical method can produce unbiased estimates without somehow modeling the unknown reason people left, which is nearly always impossible in practice.
What It Looks Like in Real Studies
A large randomized trial reported in the BMJ illustrates the pattern clearly. The intervention group lost 28% of participants while the control group lost 22%, a statistically significant difference. More telling than the raw numbers were the reasons: the control group disproportionately lost volunteers who had poor health or a history of fractures, meaning the sickest control participants were the ones disappearing from the data. That kind of selective loss can make the control group look healthier than it actually was, which in turn makes the intervention look less effective (or more effective) than it truly is, depending on the direction of the bias.
Side effects create another common pattern. If a medication causes nausea or fatigue, the people most affected tend to quit the treatment group. The remaining treatment group then consists of participants who tolerated the drug well, which are often the people who also responded to it. Comparing this self-selected group of good responders against a less-filtered control group inflates the apparent benefit of the treatment.
How Researchers Measure the Problem
The What Works Clearinghouse, operated by the U.S. Institute of Education Sciences, uses a specific threshold to evaluate whether attrition in a study is acceptable. Their model estimates how much bias the observed dropout pattern would introduce, measured in standard deviations. If the expected bias is less than 0.05 standard deviations, the study has “low attrition” and the results are considered trustworthy. At or above that threshold, the study is flagged as having “high attrition,” meaning the dropout pattern may have meaningfully distorted the findings.
This threshold applies to the combination of overall attrition and the difference between groups. A study could lose a moderate number of participants overall and still pass, as long as the loss is balanced. Conversely, a study with relatively low total dropout can fail if that dropout is concentrated in one group.
How Studies Report Attrition
The CONSORT guidelines, which most medical journals require for reporting randomized trials, mandate a detailed flow diagram tracking every participant from randomization to final analysis. For each group, researchers must report how many people were randomly assigned, how many received the intended treatment, how many were lost to follow-up, and the reasons for every exclusion. This transparency lets readers (and peer reviewers) judge for themselves whether differential attrition may have compromised the findings.
Strategies for Handling Missing Data
The most widely recommended analytical approach is intention-to-treat analysis, which includes every participant in the group they were originally assigned to, regardless of whether they completed the study or even received the treatment. This preserves the original randomization and prevents the selective filtering that makes differential attrition dangerous. The tradeoff is that it can dilute the measured effect of a treatment, since it includes people who may not have actually taken it.
The alternative, per-protocol analysis, only includes participants who completed the study as planned. This tells you more about the effect of actually receiving a treatment, but it’s exactly the approach most vulnerable to differential attrition bias, because the “completers” in each group may no longer be comparable.
When data is missing, researchers also use statistical techniques to fill in the gaps. The simplest method replaces missing values with the last recorded observation for that participant, essentially assuming nothing changed after they dropped out. A more sophisticated approach called multiple imputation creates several complete versions of the dataset, each with slightly different plausible values substituted for the missing data based on patterns in the observed information. The results from all versions are then combined, which accounts for the uncertainty introduced by the missing data rather than pretending it doesn’t exist. In studies where participants died during follow-up, some researchers assign the worst possible score to the missing outcome, a conservative approach that avoids overstating treatment benefits.
None of these methods fully solve the problem when data is missing not at random. They reduce bias and improve transparency, but the most reliable defense against differential attrition remains designing studies that minimize dropout in the first place: keeping follow-up visits manageable, reducing treatment burden, and maintaining contact with participants even when they stop the assigned treatment.