Background: Loss of follow-up represents a potential source of bias. Suggested guidelines propose 20% loss of follow-up as acceptable. However, these guidelines have not been established through scientific investigations. The goal of this study was to evaluate how loss of follow-up influences the statistical significance in a trauma database.
Methods: A database of 637 polytrauma patients with an average follow-up of 17.5 years postinjury was used. The functional outcome of workers' compensation patients versus nonworkers' compensation patients was compared using a validated scoring system. A significant difference between the 2 groups was found (P < 0.05). We simulated a gradually increasing loss of follow-up by randomly deleting an increasing number of patients from 2%, 5%, and 10%, and then increasing in increments of 5% until the significance changed. This process was repeated 50 times, each time with a different electronic random generator. For each simulation series, we documented at which simulated loss of follow-up that the results turned from significant (P < 0.05) to nonsignificant (P > 0.05).
Results: Among 50 simulation series, the turning point from significant to nonsignificant varied between 15% and 75% loss of follow-up. A simulated loss of follow-up of 10% did not change the statistical significance in any of the simulation series; a simulated loss of follow-up of 20% changed the statistical significance in 28% of our simulation series.
Conclusions: A loss of follow-up of 20% or less may frequently change the study results. Researchers should establish protocols to minimize loss of follow-up and clearly state the loss of follow-up in manuscript publications.