Paternal exposures (and other non-maternal factors) around pregnancy could have important effects on offspring health. One challenge is that data on partners are usually from a subgroup of mothers with data, potentially introducing selection bias, limiting generalisability of findings. We aimed to investigate the potential for selection bias in studies using partner data.
We characterise availability of data on father/partner and mother health behaviours (smoking, alcohol, caffeine and physical activity) around pregnancy from three UK cohort studies: the Avon Longitudinal Study of Parents and Children (ALSPAC), Born in Bradford and the Millennium Cohort Study. We assess the extent of sample selection by comparing characteristics of families where fathers/partners do and do not participate. Using the association of parental smoking during pregnancy and child birthweight as an example, we perform simulations to investigate the extent to which missing father/partner data may induce bias in analyses conducted only in families with participating fathers/partners.
In all cohorts, father/partner data were less detailed and collected at fewer timepoints than mothers. Partners with a lower socio-economic position were less likely to participate. In simulations based on ALSPAC data, there was little evidence of selection bias in associations of maternal smoking with birthweight, and bias for father/partner smoking was relatively small. Missing partner data can induce selection bias. In our example analyses of the effect of parental smoking on offspring birthweight, the bias had a relatively small impact. In practice, the impact of selection bias will depend on both the analysis model and the selection mechanism.
