Quantifying and adjusting for selection biases in the Norwegian Mother, Father and Child Cohort Study using population-wide individual-level registry information.
Selective participation in research studies hampers researchers' ability to draw valid and generalizable inferences from analyses. Quantifying and adjusting for selective participation are desirable but can be challenging given the paucity of data for non-participants.
We used individual-level information from population registers to predict baseline and continued participation in the Norwegian Mother, Father and Child Cohort Study (MoBa). Inverse probability weights were computed from logistic regression models with elastic-net regularization. We predicted selective participation and attrition in 296 987 mothers, of whom 29% returned the first MoBa questionnaire and 12.5% returned a follow-up questionnaire 8 years later. To quantify bias, we computed sample characteristics and exposure-outcome associations in the target population and stratified samples. To compare approaches for adjusting for bias, we computed weighted sample estimates by using three sets of weights.
Unweighted sample estimates were systematically different from target-population values, indicating bias due to selective participation. Baseline participation weights substantially reduced the impact of selection bias on mean values by 93% and associations by 75%. Attrition weights reduced attrition bias on mean values by 78% and associations by 50% but were less effective for reducing participation bias (22% and 14% respectively).
In this sample, characteristics and effect estimates are substantially different from target-population values. Participation weights were relatively effective at reducing bias due to selective participation but attrition weights were not. Estimates from studies that use attrition weights may still contain non-negligible selection bias-particularly if the baseline sample is not representative of the target population. Future studies should prioritize opportunities for deriving sampling weights for participation as well as attrition.
We used individual-level information from population registers to predict baseline and continued participation in the Norwegian Mother, Father and Child Cohort Study (MoBa). Inverse probability weights were computed from logistic regression models with elastic-net regularization. We predicted selective participation and attrition in 296 987 mothers, of whom 29% returned the first MoBa questionnaire and 12.5% returned a follow-up questionnaire 8 years later. To quantify bias, we computed sample characteristics and exposure-outcome associations in the target population and stratified samples. To compare approaches for adjusting for bias, we computed weighted sample estimates by using three sets of weights.
Unweighted sample estimates were systematically different from target-population values, indicating bias due to selective participation. Baseline participation weights substantially reduced the impact of selection bias on mean values by 93% and associations by 75%. Attrition weights reduced attrition bias on mean values by 78% and associations by 50% but were less effective for reducing participation bias (22% and 14% respectively).
In this sample, characteristics and effect estimates are substantially different from target-population values. Participation weights were relatively effective at reducing bias due to selective participation but attrition weights were not. Estimates from studies that use attrition weights may still contain non-negligible selection bias-particularly if the baseline sample is not representative of the target population. Future studies should prioritize opportunities for deriving sampling weights for participation as well as attrition.
Authors
Rayner Rayner, Hannigan Hannigan, Badini Badini, Demange Demange, Ofstad Ofstad, Ystrom Ystrom, McAdams McAdams
View on Pubmed