Does low participation in cohort studies induce bias?

Abstract

Background: Participation rates in large cohort studies have decreased during the last 2 decades. The consequences of this trend for relative risk estimation are unknown.

Methods: The impact of a low participation rate (30%) on the Danish National Birth Cohort was examined among 49,751 women from the source population, including 15,373 participants in the cohort study. On the basis of independent data collection, we estimated odds ratios (ORs) in the source population and among participants for 3 exposure-risk associations: (a) in vitro fertilization and preterm birth, (b) smoking during pregnancy and birth of a small-for-gestational-age infant, and (c) prepregnancy body mass index and antepartum stillbirth. The effect of nonparticipation was described by a relative odds ratio (ROR), calculated as the OR(participants)/OR(source population). Two methods for calculation of confidence intervals for the relative odds ratio also were assessed.

Results: The effect of nonparticipation on the selected ORs was small. The relative ORs were close to one and the bias was never larger than 16%, although some of the confidence intervals were wide. The 2 methods for calculation of confidence intervals gave very similar results and a small simulation study showed that the coverage probabilities were close to the 95% nominal level.

Conclusion: For the 3 chosen associations, the ORs were not biased by nonparticipation. The results are reassuring for studies based on the Danish cohort and similar cohorts of pregnant women. The methodology used to compute confidence intervals for the relative odds ratios performed well in the scenarios considered.