Assessing the impact of attrition in randomized controlled trials

J Clin Epidemiol. 2010 Nov;63(11):1264-70. doi: 10.1016/j.jclinepi.2010.01.010. Epub 2010 Jun 22.


Objectives: A survey of randomized controlled trials found that almost a quarter of trials had more than 10% of responses missing for the primary outcome. There are a number of ways in which data could be missing: the subject is unable to provide it, or they withdraw, or become lost to follow-up. Such attrition means that balance in baseline characteristics for those randomized may not be maintained in the subsample who has outcome data. For individual trials, if the attrition is systematic and linked to outcome, then this will result in biased estimates of the overall effect. It then follows that if such trials are combined in a meta-analysis, it will result in a biased estimate of the overall effect and be misleading. The aim of this study was to investigate the impact of attrition on baseline imbalance within individual trials and across multiple trials.

Study design and setting: In this article, we used individual patient data from a convenience sample of 10 trials evaluating interventions for the treatment of musculoskeletal disorders. Meta-analyses using the mean difference at baseline between the trial arms were carried out using individual patient data from these trials. The analyses were first carried out using all randomized participants and secondly only including participants with outcome data on the quality-of-life score. Meta-regression was carried out to evaluate whether the level of baseline imbalance was associated with the level of attrition.

Results: The overall attrition rates for the quality-of-life score ranged between 4% and 28% of the total randomized patients. All trials showed some level of differential attrition between the treatment arms, ranging from 1% to 14%. Attrition within the control group ranged from 3% to 25% and within the intervention group, it ranged from 0% to 31%. For individual trials, there was no indication that attrition altered the results in favor of either the treatment or the control. Forest plots highlighted that the attrition had some impact on the baseline imbalance for the primary outcome score as more heterogeneity was introduced (I-squared value of 0.4% for the initial data set vs. I-squared value of 16.9% for the analyzed data set). However, the standardized mean difference increased only slightly (from 0.01 to 0.03 with 95% confidence interval [CI]: -0.05, 0.10). Meta-regression showed little or no evidence of a significant dose-response relationship between the level of attrition and the baseline imbalance (coefficient 0.73, 95% CI: -0.81, 2.28).

Conclusion: Although, in theory, attrition can introduce selection bias in randomized trials, we did not find sufficient evidence to support this claim in our convenience sample of trials. However, the number of trials included was relatively small, which may have led to small but important differences in outcomes being missed. In addition, only 2 of 10 trials included had attrition levels greater than 15% suggesting a low level of potential bias. Meta-analyses and systematic reviews should always consider the impact of attrition on baseline imbalances and where possible any baseline imbalances in the analyzed data set and their impact on the outcomes reported.

MeSH terms

  • Bias
  • Data Collection / statistics & numerical data*
  • Humans
  • Meta-Analysis as Topic
  • Musculoskeletal Diseases / epidemiology*
  • Patient Dropouts
  • Randomized Controlled Trials as Topic / standards*