Paternal bias: The impact of not accounting for paternal confounders in reproductive epidemiological studies

OBJECTIVE:

A growing number of researchers have recently discussed paternal factors as independent determinants of offspring health.^1,2 For example, animal studies have shown that sperm carries an epigenetic cargo,³ allowing not only genetic but also epigenetic marks to be passed to offspring, with implications for offspring health. Given that paternal factors are generally associated with maternal exposures, the former may also play an important role as confounders of the associations between maternal exposures and child outcomes, and should be accounted for in epidemiologic studies. For instance, when evaluating the association of maternal body mass index (BMI) on an infant outcome such as birthweight (BW), we may need to account for paternal BMI, which is both associated with maternal BMI and a potential independent risk factor for BW (due to underlying genetics and/or social factors).⁴ In practice, however, most pregnancy cohorts have limited data on paternal characteristics,² leaving room for potential bias due to unmeasured confounding.

STUDY DESIGN:

By using simulations from a common and simple scenario, we provide an example of the potential impact associated with not accounting for paternal confounders. Details on the simulation and R code with results for all evaluated settings are available at https://github.com/andreabellavia/paternalbias. We used the example of paternal BMI (evaluated as a binary indicator of obese vs nonobese) as a potential confounder of maternal BMI and offspring BW, simulating a dataset of n = 10,000 pregnancies in which the true effect of maternal BMI on BW was of 25 g for any 5-unit increase in maternal BMI. We investigated several scenarios by means of the following: (1) varying the true independent effect of paternal BMI on offspring BW (ie, the difference in BW between infants born to obese vs nonobese fathers) from 5 to 50 g; (2) varying the association between maternal BMI and paternal obesity (ie, the difference in maternal BMI between women with obese vs nonobese partners) from 1 to 5 kg/m²; and (3) including an interaction between maternal and paternal BMI in predicting infant BW.

RESULTS:

The Figure presents results in the setting in which the association between paternal obesity and maternal BMI is β = 5 (ie, women with an obese partner have, on average, 5 higher points in BMI), and evaluates the bias at increasing hypothetical coefficients for the independent effect of paternal obesity on offspring BW from 5 to 50 g. Neglecting to account for paternal obesity would severely bias the estimated effect of maternal BMI, with increasingly larger bias as the effect of the paternal risk factor increases. For example, if the difference in offspring BW between infants born to obese vs nonobese fathers was 40 g, but paternal obesity was omitted, the real effect of maternal BMI on offspring BW (25 g; white circle on the Figure) would be overestimated at ~40 g (black circle on the Figure). Overestimates of the true effect were observed in all other evaluated scenarios, with smaller bias when the association between maternal and paternal BMI was weaker, and larger bias when an interaction was present (see details at https://github.com/andreabellavia/paternalbias).

FIGURE.

Potential bias in the effect of maternal exposures on infant birth weight (grams [g]), when omitting a paternal confounder

CONCLUSION:

Results of this simulation demonstrate the importance of incorporating paternal factors, even when seeking to evaluate only associations between maternal exposures and child outcomes. Not accounting for paternal factors as confounders biases the effect of the maternal exposure if not accounted for. Ideally, existing cohorts must be leveraged or new cohorts established to allow a full understanding of the interplay of maternal and paternal factors in predicting child outcomes. When evaluating the effect of maternal exposures or characteristics on offspring health, if paternal exposures are not available, we recommend conducting a set of sensitivity analyses⁵ to evaluate the potential impact of these unmeasured confounders on study results. In this context, a useful tool to quantify the potential bias due to unmeasured confounders, when the true effect is unknown, is the E-value, which quantifies the minimal value of association that the unmeasured confounder should have to explain away the observed association.⁶