Attrition Bias

Attrition bias is a systematic error that arises when participants who drop out of a study before completion differ in meaningful ways from those who remain, creating differences between groups that are not attributable to the intervention. When the reasons for dropout vary across treatment arms, the remaining participants may no longer be comparable, threatening the validity of between group comparisons.

Attrition bias is one of the most serious threats to the internal validity of randomized controlled trials and longitudinal studies. Even a well executed randomization procedure can be undermined if subsequent participant loss creates systematic differences between groups. Systematic reviewers routinely assess studies for attrition bias; in Cochrane's RoB 2 tool for randomized trials it falls under the domain of bias due to missing outcome data, and a judgment of high risk can lead to downgrading the certainty of evidence in GRADE assessments. Regulatory agencies scrutinize attrition patterns when reviewing drug and device applications, and unexplained differential dropout can raise concerns about the safety or tolerability of an intervention.

Attrition bias occurs through a specific mechanism: participants leave the study for reasons that are related to both the treatment assignment and the outcome. For example, in a trial of a new analgesic, participants in the active drug arm who experience inadequate pain relief may withdraw at higher rates than those in the placebo arm, leaving behind a selected subgroup that responded unusually well to the drug. When only these remaining participants are analyzed, the drug appears more effective than it truly is.

The bias can operate in either direction. If participants in the treatment arm with the most side effects drop out, the treatment may appear safer among completers. If participants in the control arm with deteriorating conditions drop out, the control group may appear healthier than expected, reducing the apparent treatment effect.

Attrition bias is assessed by comparing the characteristics of participants who completed the study with those who dropped out, and by examining whether attrition rates and reasons differ between groups. Key indicators of potential attrition bias include substantial differences in overall attrition rates between arms, differences in the reasons for dropout (e.g., adverse events in one arm vs. inconvenience in another), and differences in baseline characteristics between completers and dropouts.

Several analytic strategies can mitigate attrition bias. The first defense is to keep missing outcome data to a minimum. Intention to treat analysis, which includes all randomized participants regardless of whether they completed the study, is the primary analytic defense, but participants with missing outcomes can be included only if those outcomes are handled by the methods described next. Missing outcome data for dropouts can be handled through methods such as multiple imputation, last observation carried forward (though this approach has limitations), or mixed effects models that use all available data. Sensitivity analyses, such as worst case and best case scenarios for missing data, help quantify how much the conclusions might change under different assumptions about the dropouts.

Assessment of attrition bias is relevant in every trial or longitudinal study that experiences participant loss. Researchers should report attrition rates and reasons by group, compare baseline characteristics of completers and dropouts, and conduct sensitivity analyses to evaluate the potential impact of missing data. Systematic reviewers should assess attrition bias for every included study using standardized tools.

A common error is dismissing differential attrition as unimportant because the overall attrition rate is "low." Even small differences in dropout rates between groups can introduce meaningful bias if the dropouts are systematically different from completers. Another mistake is relying on per protocol or completer analyses without also conducting intention to treat analyses, which preserves the balance created by randomization.

Researchers sometimes assume that last observation carried forward (LOCF) adequately handles missing data from dropouts, but this method makes the strong assumption that the last observed value would have remained constant, which is often implausible, particularly in progressive diseases or deteriorating conditions. More principled approaches such as multiple imputation or pattern mixture models should be considered. Finally, failing to plan for attrition at the design stage by not building in strategies to minimize dropout and not inflating the sample size to account for expected losses is a preventable oversight.

A 24 week randomized trial comparing a new antidepressant with placebo enrolls 300 participants (150 per arm). By week 24, 38 participants (25%) in the drug arm and 22 (15%) in the placebo arm have withdrawn. In the drug arm, 18 withdrawals are due to side effects and 20 are for other reasons. In the placebo arm, only 3 withdrawals are due to side effects. The higher side effect related dropout in the drug arm means that the remaining drug group participants are those who tolerated the medication, biasing the completer analysis toward showing a larger treatment effect. The intention to treat analysis using multiple imputation yields a smaller but still significant treatment effect (mean difference = 2.8 points on the Hamilton Depression Rating Scale, 95% CI: 0.9 to 4.7), compared to the completer analysis (mean difference = 4.1 points).