Does attrition affect estimates of association: A longitudinal study. (March 2019)

• Record Type: Journal Article
• Authors: Saiepour, N.; Najman, J.M.; Ware, R.; Baker, P.; Clavarino, A.M.; Williams, G.M.
• Is Part Of: Journal of psychiatric research. Volume 110(2019)
• Page Start: 127
• Page End: 142
• Publication Date: 2019-03
• Subjects: Loss to follow-up -- Magnitude of bias -- Mental disorder -- Socio-economic -- Disadvantage; Psychiatry -- Periodicals; Mental Disorders -- Periodicals; Maladies mentales -- Périodiques; Psychiatry; Electronic journals; Periodicals; 616.89005
• DOI: 10.1016/j.jpsychires.2018.12.022
• Languages: English
• ISSNs: 0022-3956

Abstract

Abstract: Survey research frequently involves missing cases attributable to refusals to participate, lack of success in accessing all potential respondents or loss to follow-up in longitudinal studies. There is concern that those not recruited or those lost are a select group whose absence from a study may bias the findings of the study. This study provides a test of the extent to which selective loss to follow-up in a longitudinal study may lead to biased findings. The Mater-University Study of Pregnancy collected baseline information for 7718 pregnant women. Follow-ups occurred five years, 14 years, 21 years and 27 years after the birth, for 6753 eligible women. Participants at baseline were partitioned according to follow-up status for each follow-up. We compare baseline (at recruitment) measures of association, with these same measures of association for those retained in the study (Group A) and those lost to follow-up (Group B) at each phase of data. Using univariate logistic regression we compared the strength of association between maternal mental health and various baseline socio-demographic factors for different rates of loss to follow-up. Estimates of association at baseline, and at each follow-up are similar irrespective of the rate of loss to follow-up and whether the comparison is with those retained in the study or those lost to follow-up. There were no statistically significant differences in 90.8% of baseline comparisons with Group A, and 96.9% of comparisons with Group B measures of association. We conclude that differential loss to follow-up rarely affects estimates of association. We suggest that loss to follow-up may produce misleading findings only in circumstances where loss to follow-up is combined with a number of other sources of bias.