Genetically Informed Research Designs in Perinatal Pharmacoepidemiology: A Methodological Overview

Abstract

Despite the widespread use of medications during pregnancy, ethical and methodological barriers to clinical trials make observational studies necessary for evaluating medication safety in this population. Observational studies are prone to biases that often limit their validity due to the lack of randomization; integrating genetic information through discordant sibling designs, polygenic scores, and Mendelian randomization can address several confounding issues. However, application of these three approaches in perinatal pharmacoepidemiology has been limited. Complementing traditional designs with these genetically informed research designs can tackle common biases and strengthen causal inference. This paper focuses on applying genetically informed research designs to child outcomes in perinatal pharmacoepidemiology by reviewing various methods, discussing their strengths and limitations, and examining their application to date, as well as considerations for implementing them in future research. Such considerations include the availability of genetic data, the complexity of integrating genetic data with existing epidemiological data, and selection of appropriate genetic instruments for analyses. Incorporating causal inference in perinatal pharmacoepidemiology can ultimately contribute to enhancing safe medication use during pregnancy.

Key Points

Clinical trials are often not feasible to study medication efficacy and safety in pregnancy, but observational studies can be biased due to confounding by genetics.

Methods that take genetic information into account, such as sibling comparisons, using genetic risk scores, or Mendelian randomization, can help establish medication effects.

Combining these genetic approaches with traditional study designs can yield more trustworthy evidence, ultimately helping clinicians make safer, more informed decisions about medication use during pregnancy.

Introduction

Over 80% of pregnant women use prescription or over-the-counter (OTC) medications [1] creating a clinical challenge in balancing treatment of maternal conditions and fetal safety. However, due to ethical and methodological barriers, pregnant women are often excluded from randomized controlled trials (RCTs), limiting access to high-quality trial data. Consequently, observational studies are often the primary source used to assess the efficacy and safety of medications during pregnancy, which, while essential, have limitations in establishing causality [2, 3].

Genetic confounding, where maternal genetic factors influence both treatment exposure (e.g., antidepressant use) and child outcomes, is a major source of bias in perinatal pharmacoepidemiology [4]. Because children inherit half their mother’s genes, they also inherit her genetic liability for disorders and treatment responses. This creates horizontal pleiotropy, where genetic variants independently affect multiple traits, potentially biasing associations between maternal treatments and child outcomes simply due to shared inheritance. Genetically informed family designs, such as discordant sibling comparisons, can address this issue [5]. For instance, if a woman’s disease risk or medication use varies across pregnancies, comparing siblings with differing exposures blocks pleiotropic pathways related to shared maternal genetics, thereby reducing confounding [6].

Second, it is crucial to understand both medication risk on fetal development and the risks posed by untreated maternal conditions during pregnancy [7]. However, confounding by indication, where the effects of a condition are difficult to separate from those of its treatment, poses a major challenge in observational studies [8]. If assumptions are met, Mendelian randomization (MR) allows for an unbiased exposure-outcome association, even in the presence of confounding by indication, as genetic variants are randomly allocated and cannot be influenced by health status [9].

Finally, exposure during pregnancy has high susceptibility to information bias [10]. In traditional observational epidemiology, information may be collected through surveys, pregnancy registries, or healthcare databases. When data is gathered retrospectively through surveys, information bias due to forgetfulness or social desirability is a significant concern [11]. For data sourced from registries or healthcare databases, however, there can be uncertainty regarding whether the medication was actually used, as well as questions about the dosage and timing of administration [12]. Both MR and polygenic score (PGS) designs can address information bias, as MR utilizes genetic variants as proxies for exposure, and PGS is derived from genetic data from biological samples. The underlying data for either approach are individual-level genotypes or genome-wide association studies (GWAS) summary statistics [13]. However, information bias in other study variables may still persist, and both MR and PGS approaches require reliable exposure data to identify suitable instrumental variables or genetic variants associated with the exposure of interest.

The scarcity of robust evidence on medication use during pregnancy complicates clinical decision-making regarding disease management. Therefore, advanced methods that limit confounding in perinatal pharmacoepidemiological studies and enhance risk estimation for both treatment and underlying disease are essential [14]. This paper provides an overview of approaches incorporating genetic information to strengthen causal inference in perinatal pharmacoepidemiology. We delve into discordant sibling designs, polygenic scores, and Mendelian randomization, highlighting their strengths, limitations, and applications. Incorporating genetic information into pharmacoepidemiological studies offers a path towards more balanced assessments of treatment risks versus the harms of untreated maternal illness.

Sibling Designs

Method

Discordant sibling designs, hereafter referred to as sibling designs, are the most widely used genetically informed study designs in perinatal pharmacoepidemiology. This design primarily exploits discordance in exposure between siblings [15]. In perinatal pharmacoepidemiology, it is particularly useful for examining intrauterine medication exposures while controlling for time-fixed confounders (Fig. 1). It offers a key advantage over traditional cohort studies in terms of controlling for shared confounding factors that are difficult to measure. While observed covariates typically explain a small fraction of sibling similarity, many shared factors, especially genetic and early environmental factors, are unmeasured in conventional pharmacoepidemiological studies. Sibling studies circumvent these limitations by comparing siblings who are inherently matched on numerous shared confounders, including both genetic and early-life environmental factors, thereby providing a more isolated view of the drug’s pharmacological effect.

Fig. 1.

Sibling study design illustrating discordance in exposure between siblings, enabling evaluation of outcomes while accounting for shared confounders. In an illustrative example, we consider a mother’s exposure to antidepressants during pregnancy and the risk for ADHD in the child as an outcome. The mother (M) would be using antidepressants differentially between the two pregnancies, where she would be exposed to antidepressants during one pregnancy (X₁) and not in the other (X₂). There would be some shared confounding between pregnancies (C₁₂) (genetic liability to ADHD, family history for ADHD, ethnicity, and socioeconomic status). There would also be some non-shared confounding between pregnancies (C₁, C₂). Siblings then might develop ADHD (Y₁) or not (Y₂). The design can only correct for C₁₂ and not for C₁ nor C₂. C₁, C₂ non-shared confounding, C₁₂ shared, M exposed mother, X₁ exposure in first sibling, X₂ exposure in second sibling, Y₁ outcome in first sibling, Y₂ outcome in second sibling, confounding

This within-family design compares the outcome of interest in two exposure-discordant siblings, thereby removing all measured and unmeasured shared confounding, including genetic predisposition, family history, ethnicity, and socioeconomic status, and other time-invariant characteristics across pregnancies [5, 16–19]. In the absence of covariates, only exposure-discordant families are informative for estimating the within-family exposure-outcome association. The inclusion of additional covariates complicates this, as exposure-concordant families may become indirectly informative if they are discordant on at least one covariate [20].

This design typically employs conditional logistic regression for binary outcomes or stratified Cox models for time-to-event analyses [16]. These sibling-based analyses are usually contrasted with full population (cohort) analyses to assess confounding by shared factors, and a smaller within-sibling effect is often interpreted as evidence that familial confounding influenced the cohort result [21]. Additionally, when heritable maternal traits have an indirect genetic effect through the environment, the sibling design adjusts for all genetic variants associated with maternal exposure [2, 17]. Notably, genotyping is not required in this design since it is based on the known genetic relatedness between siblings [2]. For a more in depth description of the design and its technical implementation, several articles can be referenced [16, 20, 22–25].

Sibling studies are more feasible within national registries and large claims databases. These provide large sample sizes needed for studying rare exposures or outcomes [26, 27]. Nevertheless, this is only possible in data sources with reliable family linkage (e.g., Nordic health registries). Such data sources might still lack granularity in the exposure/outcome data. Alternatively, sibling studies can be conducted within bespoke cohorts where families with sibling pairs are recruited and followed over time. These typically yield small samples for discordant siblings, which limit the ability to examine relatively rare exposures or outcomes [28].

Strengths

Sibling designs leverage a natural experiment, comparing siblings who share maternal genetic risk and early environmental factors, but differ in exposure to the variable of interest (e.g., in utero medication). This approach inherently controls for stable maternal and familial factors, effectively reducing confounding without the need for direct measurement of these variables [29]. By adjusting for unmeasured confounders that remain constant across pregnancies, including factors that are difficult or impossible to observe directly, sibling designs may provide more robust causal inference than traditional observational studies.

Limitations

Despite their strengths, sibling designs introduce unique biases due to non-shared, or non-familial, confounders, for which they cannot control [15]. Sibling designs can be more prone to bias due to unmeasured, non-shared confounding than non-sibling designs, as they only analyze exposure-discordant or covariate-discordant siblings. This biases the results towards the confounded results in the general population. Since family-shared factors influence both siblings equally, they cannot explain why one sibling was exposed while the other was not [16].

Furthermore, bias due to exposure or outcome misclassification can be amplified in sibling studies because the design specifically selects for siblings who differ in non-shared exposure causes or non-shared covariates and thus differ in the direction of any random misclassification error [16]. In other words, fewer individuals will be correctly classified on exposure in discordant pairs compared to the general population, leading to an increased attenuation in the outcome-exposure association.

Additionally, sibling designs are vulnerable to carryover effects, where the exposure or outcome in one sibling influences the exposure status or outcome in the next-born sibling [22]. Another key limitation is the loss in statistical power due to reliance on discordant sibling pairs, which often reduces the available sample size substantially [30]. This constraint on statistical power remains the most recognized limitation of sibling designs [31]. Bespoke cohort studies typically include hundreds to a few thousand sibling pairs, while sibling studies using national registries may have a sample size ranging from thousands to hundreds of thousands. Irrespective of the data source, if the exposure or outcome prevalence is low, statistical power remains limited. Consequently, sibling designs are more feasible for broad exposure categories (e.g., all antidepressants) than for narrower ones (e.g., SSRIs, TCAs) or individual drugs [31].

Finally, the consistency assumption of causal inference, that the outcome observed is equal to the potential outcome corresponding to the treatment level assigned through the intervention, is pivotal yet might not be easily attainable in sibling designs for perinatal pharmacoepidemiology [32]. More specifically, the sibling design connects here to the cross-over design, in which it may be challenging to propose a ‘treatment policy’ estimand for the co-sibling’s exposure status as this depends on the exposure status of the first sibling. The consistency assumption is more likely to be met when the exposure is well-defined and this is especially crucial for the designs’ validity which depends on clearly defined discordance [33]. Thus, researchers should carefully and critically think about well-defined interventions when using sibling designs in a perinatal context to avoid challenges in causal interpretation.

Consequences for Interpretation

In sibling design studies assessing the impact of maternal medication use on child health outcomes, it is often a concern that findings may be generalizable only to discordant siblings, limiting their applicability to the broader population [34]. That is because the selection of discordance implies a selection for pairs that also differ in non-shared causes of the exposure. Furthermore, families with discordant siblings may differ systematically from those with concordant siblings (e.g., in maternal health behaviors, or socioeconomic status) [35]. However, the mechanism by which some families become discordant can be empirically evaluated by comparing concordant (i.e., non-informative) and discordant families across measured non-shared confounders (e.g., birth year, maternal age, socioeconomic status). If both populations are similar, the findings are more likely to be generalizable [20]. Additionally, between–within models can further address concerns about generalizability in this design [20]. While sibling designs inherently adjust for shared familial factors, they do not account for non-shared, time-varying confounders, which can bias outcomes. Incorporating these non-shared confounders can help reduce residual confounding and improve the robustness of the findings [16]. To assess potential carryover effects, researchers should examine whether outcomes vary by birth order, maternal age, or exposure sequence. The presence of such patterns may suggest that carryover effects are influencing the findings, necessitating further adjustments to ensure valid interpretation of results [18, 31]. Finally, it is worth noting that estimands resulting from the aforementioned logistic regression and stratified Cox models, estimate the causal effect on the full sibling pair and thus do not correspond to those that would be estimated from an RCT analogue [25].

Application in Perinatal Pharmacoepidemiology

Sibling designs are widely used in perinatal pharmacoepidemiology to control for familial confounding. For example, Martin et al recently investigated the effects of in utero antidepressant exposure on preterm birth and birthweight [36].

In the general population, maternal exposure to antidepressants during pregnancy was associated with increased odds of preterm delivery (adjusted odds ratio [aOR]: 1.26, 95% confidence interval [CI] 1.23–1.30) and small for gestational age (SGA) with an aOR of 1.04 (95% CI 1.02–1.07). In sibling analyses, slightly attenuated but significant associations remained for preterm birth (aOR: 1.19, 95% CI 1.11–1.27), while SGA became non-significant (aOR: 1.01, 95% CI 0.95–1.08). This rules out familial confounding, but the authors mentioned that confounding by severity of indication remains possible [36].

Similarly, Sujan et al examined the association between prenatal antidepressant exposure and attention deficit hyperactivity disorder (ADHD) risk in offspring using a sibling design [37]. They found no association between first-trimester antidepressant exposure and ADHD risk (hazard ratio [HR]: 0.99; 95% CI 0.79–1.25) in discordant sibling pairs. Conversely, a population-level analysis showed an increased ADHD risk (HR: 1.58; 95% CI 1.46–1.71) associated with first-trimester exposure, indicating familial confounding in the associations observed in the general population [37].

These examples highlight the ability of sibling designs to control for familial confounding. However, results from sibling analyses studies should be interpreted with caution, due to the potential biases from non-shared factors between siblings, such as lifestyle differences.

Polygenic Scores

Method

Genetic susceptibility can confound analyses when associated with both maternal medication use (exposure) and child outcomes. PGSs help mitigate residual genetic confounding, thereby strengthening causal inference (Fig. 2) [10, 38, 39]. The simplest PGSs, derived from GWAS, aggregate individual genetic variants—typically single nucleotide polymorphisms (SNPs)—associated with a disease or trait. This is done by weighting each variant based on its effect size and summing these weights into a single score representing an individual’s genetic liability for the disease or trait [40]. As traits are polygenic, i.e., influenced by multiple genetic variants [41], PGSs provide a quantitative measure of genetic risk for these traits and can be used to predict phenotypes [42, 43] associated with a disease or trait [40]. Although the standard approach for calculating the PGS is using the weighted sum of alleles (e.g., clumping and thresholding, LDpred, PRS-CS) on GWAS summary statistics pre-adjusted for population stratification, more sophisticated approaches can also account for population stratification and external biological information [44–46]. In a perinatal context, the mother’s genetic variants associated with a certain exposure or outcome of interest are consolidated into a single PGS. This helps assess how genetic predisposition influences maternal health during pregnancy and child health outcomes. For example, the PGS can be incorporated as a covariate or exposure to test for genetic confounding.

Fig. 2.

Polygenic scores illustrating direct and indirect PGS effects on offspring. Direct PGS represents the inherited genetic variants that influence the propensity of a trait in the offspring. Indirect PGS represents parental variants (whether transmitted to the offspring or not) shaping the parental and offspring environment [111]. In an example of perinatal antidepressant exposure (X_O) and child ADHD (Y_O), maternal (PGS_M-D) and paternal PGS for ADHD (PGS_P-D) directly affects the child’s genetic risk for developing ADHD (PGS_O). Maternal genetics (PGS_M-D) also directly contribute the exposure to antidepressants (X_O) through genetic liability. Furthermore, maternal (PGS_M-I) and paternal (PGS_P-I) genetics will have an indirect effect on the maternal exposure to antidepressants (X_O), the child’s genetic liability for ADHD (PGS_O), and ADHD in the child (Y_O)

As mentioned, the PGS can correct for genetic liability to exposures or outcomes. Maternal, paternal, and offspring PGS each serve different functions when corrected for. Taking the example of antidepressant perinatal use (exposure) and its association to ADHD in the child (outcome), maternal PGS for depression estimates maternal genetic liability for depression (a proxy for antidepressant use) and helps distinguish genetic confounding from causal intrauterine effects [47]. Alternatively, maternal PGS for ADHD reflects potential influences from maternal genetics on child ADHD through the environment, mood or behavior, a term called “genetic nurture” [48]. Paternal PGS can serve complementary roles in the same example: paternal PGS for depression can act as a negative control where an association with child ADHD indicates familial confounding rather than an intrauterine (genetic) pathway [49], while paternal PGS for ADHD serves to test for indirect genetic effects, where paternal genes influence the child via the environment, not inheritance [50]. Finally, child PGS for depression helps estimate the inherited genetic liability to depression, which may increase the chance of exposure; if the child PGS is associated to maternal antidepressant use, then it suggests genetic confounding [51]. Child PGS for ADHD helps to control for genetic confounding in the perinatal antidepressant exposure-ADHD association. This example is summarized in Table 1. Methodologically, outcome-PGS adjustment is generally preferred, as adjusting for the exposure-PGS may introduce bias amplification (Z-bias), whereas outcome-PGS adjustment avoids systematic bias and may improve efficiency [52]. Furthermore, while this framework is most straightforward when exposure and outcome traits are genetically independent, in practice, genetic correlations between maternal exposures and child outcomes are common. Rather than limiting the approach, these correlations underscore the need for careful modeling and interpretation of shared genetic architecture when applying PGS in perinatal pharmacoepidemiology.

Table 1.

A summary of how different PGSs can serve several purposes depending on whether they are constructed for the mother, father, or child, and whether they are constructed for the exposure, or the outcome. Perinatal exposure to antidepressants and child ADHD is used as an example

PGS type	Exposure PGS (antidepressants)	Outcome PGS (ADHD)	Example
Maternal	Genetic liability to antidepressant use; tests for genetic confounding	Genetic nurture via maternal behavior or mood	Depression PGS → antidepressant use → child ADHD diagnosis
Paternal	Negative control; tests for shared familial/genetic effects	Indirect paternal effects via family environment	Paternal depression PGS → child ADHD (non-intrauterine)
Child	Inherited risk for depression linked to exposure	Direct genetic liability for ADHD; controls for confounding	Child ADHD PGS → ADHD diagnosis

ADHD attention deficit hyperactivity disorder, PGS polygenic score

Strengths

Polygenic scores effectively capture the polygenic nature of complex traits by aggregating the effects of multiple genetic variants, enhancing the estimation of individual risks [53]. Additionally, PGS not only enables controlling for confounding based on genetic risk, but also permits the detection of effect modification, which is quite common and particularly valuable in pharmacoepidemiological studies. By being a single estimate for genetic liability, PGS allows controlling for genetic variability, strengthening causal inference.

Limitations

Polygenic scores are calculated from GWAS summary statistics, typically obtained from major biobanks like the UK Biobank [54] and FinnGen [55]. However, their accuracy is limited by genetic ancestry similarities between the GWAS populations and target groups. With about 80% of GWASs conducted in European populations, representing only 16% of the global population, PGS applicability to wider demographics is restricted [56, 57]. Although biobanks are expanding in regions such as Sub-Saharan Africa [58], Mexico [59], China [60], and Taiwan [61], enabling the creation of GWASs also used in bespoke cohorts with similar ancestry profiles, bespoke studies still face significant challenges. The collection of genetic data is costly and logistically complex [62], often relying on non-random subsamples of participants who consented to genetic analysis, which can introduce selection bias and undermine findings [63]. Despite these hurdles, there is a growing trend towards integrating genetic data in bespoke studies.

Furthermore, the predictive power of PGS is limited by the number, frequency, and effect size of genetic variants included [64]. Finally, although PGSs can help adjust for genetic confounding, they typically explain a small fraction of the variance of a trait or disease, typically up to 10% [65]. However, this figure is frequently inflated in the GWAS discovery. Performance further diminishes in smaller local cohorts, making PGS only partially effective for confounding control. Consequently, PGS should be interpreted with caution especially when applied in smaller samples.

Consequences for Interpretation

Importantly, using PGSs to adjust for genetic confounding may, in certain settings, lead to more bias than not using them. If heritable covariates are included in regression models that also include polygenic scores as independent variables, one risks selection bias if some of these covariates depend on unmeasured factors that affect the outcome [66]. This is due to conditioning on a collider, which leads to spurious associations of gene-environment interactions and can create apparent effect modification even under the null. For instance, Rommel et al found that continued antidepressant use during pregnancy was linked to shorter gestation (adjusted β: 1.7–4.5 days, p < 0.001–0.008) compared to unexposed and discontinuation groups, but only in the first and fifth PGS quintiles [67]. Thus, genetic liability to depression was not linearly correlated with these outcomes and modified the effect of antidepressants differently [68–71]. These observed effects, however, are likely distorted due to collider bias.

In perinatal pharmacoepidemiology, it is important to distinguish between direct and indirect effects of parental PGS. Direct effects stem from variants transmitted to offspring and directly affect offspring traits, while indirect effects arise from parental variants—transmitted or not – that shape the environment affecting offspring traits [72]. Standard GWAS-derived PGS calculations assume unrelated individuals and no assortative mating—assumptions often violated in family-based studies. This can inflate PGS or bias associations since related individuals share more alleles than expected by chance. In family triads (parents and offspring) or dyads (mother and offspring), such calculations tend to overestimate direct effects and overlook indirect effects [73]. For example, a maternal PGS may influence prenatal behaviors or treatment choices, indirectly shaping the intrauterine environment and consequently affecting offspring outcomes. To address this, analyses should account for relatedness (e.g., using kinship matrices) when using PGS calculated for related individuals [73–75].

Application in Perinatal Pharmacoepidemiology

In 2024, Liu et al tested whether the PGSs for major depression, bipolar disorder, and schizophrenia were associated with antidepressant treatment decisions in pregnant women from a Danish cohort [76]. They identified four treatment trajectories across pregnancy and the postpartum period: continuers, early discontinuers, late discontinuers, and interrupters. They found no association between any PGS and the treatment trajectories (relative risk ratios for continuers vs early discontinuers was 0.93 [95% CI 0.81–1.06], 0.98 [0.84–1.13], and 1.09 [0.95–1.27] for per 1-SD increase in PGS for major depression, bipolar disorder, and schizophrenia, respectively) [76].

Another example is the study by Ratanatharathorn et al who assessed whether maternal PGS for mental illness is associated with perinatal risk factors for offspring mental illness [77]. They found that maternal PGS was associated with several known risk factors, namely, smoking during pregnancy, breastfeeding for less than 1 month, partner violence in the year before the birth, and pregestational overweight or obesity. They concluded that genetic risk may partly account for the association between perinatal risk factors and mental illness in offspring.

Similarly, Kinge-Rasmussen et al examined whether PGS for epilepsy, ADHD, and autism spectrum disorder (ASD) influence neurodevelopmental outcomes in children prenatally exposed to antiseizure medications (ASMs). The study found that PGS explained a greater proportion of variance in neurodevelopmental traits among ASM-exposed children compared to unexposed controls, indicating that genetic susceptibility may modify the effects of prenatal drug exposure. However, the generalizability of these findings is limited by the predominantly European ancestry of GWAS participants used to calculate the PGSs. These findings highlight the importance of incorporating genetic factors in future medication-outcome studies, particularly to address potential residual confounding through genetics [78, 79].

Mendelian Randomization

Method

Mendelian randomization employs selected SNPs that are robustly associated with an exposure as instrumental variables to assess for potential causal relationships between that exposure and an outcome. An instrumental variable is a variable associated with the exposure, but not directly with the outcome [80]. Mendelian randomization is increasingly being utilized in perinatal pharmacoepidemiology because maternal genetic variants typically do not directly correlate with child outcomes once genetic transmission, or the children's own genotypes, have been accounted for [81, 82]. In this context, maternal genotype serves as a proxy for medication exposure during pregnancy [81]. Mendelian randomization relies on three key assumptions: the relevance criterion, the exchangeability criterion, and the exclusion restriction criterion. In the context of perinatal MR studies, maternal instrumental variables must be robustly associated with intrauterine exposure to meet the relevance criterion. The exchangeability criterion requires that there be no shared risk factors between the instrumental variables and the outcome. Lastly, the exclusion restriction criterion is satisfied when the maternal instrumental variable influences the offspring outcome solely through the exposure of interest [15, 83–85]. While the exclusion restriction cannot be directly assessed, sensitivity analyses like MR-Egger regression can assess potential pleiotropy [81, 86].

A typical MR analysis uses a two-stage regression. First, the exposure is regressed on the instrumental variable and covariates to estimate the instrumental variable’s effect on the exposure. Then, predicted exposure values are regressed on the outcome thereby estimating the potential causal effect of the exposure on the outcome [87]. Instrument strength is assessed using the F-statistic; values below 10 suggest a weak instrumental variable.

Additionally, most MR studies use a two-sample approach, where the instrumental variable-exposure and instrumental variable-outcome associations are assessed in separate cohorts using the ratio method (instrumental variable-outcome ÷ instrumental variable-exposure) to estimate causal effects [88]. In perinatal pharmacoepidemiology, depending on the outcome, instrumental variable-outcome associations should come from a birth cohort (e.g., MoBa [89], HUNT [90]) to ensure outcome availability. The instrumental variable-exposure association should be derived from a GWAS on a women-only sample to ensure validity [91, 92]. However, if there is no suspected sex-specific effect modification, then including males can increase precision, given the larger sample size [93]. The choice also depends on the exposure—such as the number of available instruments and the impact of including a few that are potentially invalid.

Strengths

Mendelian randomization mimics RCTs by leveraging the random allocation of genetic variants at conception [94], thereby reducing confounding and strengthening causal inference. Maternal genetic variants influencing exposure to medications during pregnancy are unlikely to be confounded by social, behavioral, or environmental factors that occur after the mother’s birth, as these cannot retroactively influence the genetic makeup [13]. Furthermore, MR avoids reverse causation, as maternal instrumental variables are not influenced by the outcome in the offspring [29]. With large, publicly available datasets like those curated in MR-Base, it is becoming increasingly feasible to analyze maternal instrumental variables with the outcomes of interest. This supports assessment of potential genetic overlap between the mother and offspring, helping to evaluate the exclusion restriction assumption [85]. When the assumptions are met, MR allows for unbiased causal estimates even in the presence of unmeasured confounding that typically biases results in non-randomized studies [95].

Limitations

Mendelian randomization faces several limitations that can limit causal inferences. A major challenge is identifying appropriate genetic variants that can serve as instrumental variables [15]. Weak instrumental variables when the F-statistic < 10, can lead to weak instrument bias [83, 96]. Despite growing GWAS discoveries, they still underrepresent many populations and valid genetic proxies for medication exposures remain scarce [13]. Additionally, unmeasured factors like dynastic effects—where offspring traits reflect both their own and inherited parental influences—can confound MR analyses by inadvertently capturing the impact of parental genetics [94].

Linkage disequilibrium (LD), where genetic loci are correlated, may also confound causal relationships via horizontal pleiotropy [13]. Methods like colocalization can address this bias [97]. When employing a genetic proxy to evaluate drug exposure in MR, it is crucial to verify that the chosen genetic variant truly functions as a biological proxy for the drug itself and any possible teratogenic effects. This reflects the gene‑environment equivalence principle (i.e., that the genetic variant should have the same impact on the outcome as the equivalent change of the environmental exposure) [98]. However, the underlying mechanism, including the duration, size and timing of the effect, often differs [98].

In perinatal pharmacoepidemiology, maternal instrumental variables may violate the exclusion restriction if they influence offspring outcomes through shared offspring genotypes. Assessing overlap between maternal and offspring genetic variants is essential. Adjusting for offspring genetic variants may mitigate this, but can introduce collider bias (Fig. 3), where a path is created from the maternal genetic variant to the offspring outcome via paternal genetic variants [85, 94]. Including paternal genotypes may reduce this bias, but such data are often unavailable [85].

Fig. 3.

We illustrate potential collider bias using a directed acyclic graph (DAG) to examine the effect of maternal exposure to antidepressants during pregnancy on the risk of developing ADHD in offspring. The black arrows demonstrate the instrumental variable’s (IV) association to maternal use of antidepressants (e.g., a genetic variant associated with the metabolism of selective serotonin reuptake inhibitors). The small, dashed arrows indicate potential violations of the exclusion restriction assumption, detailing pathways from the IV to the offspring genetic variants (G₁) and from G₁ to the outcome (Y), specifically ADHD in the offspring. The large, dashed arrows illustrate the path from the maternal IV through paternal genetic variants (G₂) and associated paternal traits (C) including paternal history of ADHD and phenotypes related to ADHD, which may violate the exclusion restriction assumption when accounting for offspring genetic variants. Figure based on Lawlor et al [85]

The maternal exposure-offspring outcome association may also be confounded by associations of maternal instrumental variables with other variables. Therefore, adjusting for both offspring and paternal genetic variants is important. Additionally, since confounding paths may involve both maternal and offspring characteristics, assess associations with relevant confounders for both is crucial [85].

Another potential issue arises from assortative mating, where partners select each other based on heritable traits, which can create associations between maternal and paternal genotypes. Adjusting for offspring genotype may induce a negative association between parental genotypes, potentially offsetting this bias [85].

Furthermore, MR typically cannot assess time-varying exposures, which is essential in perinatal pharmacoepidemiology research where trimester-specific effects may influence outcomes differently [82, 99]. While RCTs evaluate intervention impacts during the study time-frame, MR estimates effects across a lifetime, potentially leading to substantial differences in effect estimates due to varying evaluation periods [100]. This limitation is particularly relevant since MR often targets lifetime effects with genetic instruments not specific to time periods, resulting in possible misinterpretations. Although there are some methods for estimating period-specific effects, the brief timeframe of etiologically relevant time windows makes it challenging to identify genetic variants linked to exposures in specific time windows [101, 102].

Additionally, assessing outcomes in liveborn offspring may introduce selection bias if the exposure affects fertility or pregnancy completion since individuals included in the study might have different genetic risk compared to those excluded, violating the exclusion restriction criterion. Lastly, because genetic variants typically have small, lifelong impacts, MR is better suited for identifying potential clinical effects rather than quantifying their magnitude, especially compared to short-term pharmacologic exposures [103].

Consequences for Interpretation

Proper MR analysis requires that all three core assumptions be met; violations can lead to substantial bias [82, 104]. Falsification tests and sensitivity analyses help evaluate potential violation of the assumptions. One common approach is MR-Egger regression, which is used to test and correct for horizontal pleiotropy by examining the intercept term [105]. A non-zero intercept suggests the presence of horizontal pleiotropy.

Additionally, perinatal MR studies often do not report their estimand of interest. Selecting a specific condition alters the relevant population for the estimated effect, leading to differences in average causal effects between subgroups and the overall population [82, 99, 100]. The three instrumental variable conditions test the null hypothesis and determine the presence of a causal effect but do not provide a point estimate for the effect size. This necessitates an additional condition known as a point-estimate identifying condition. Two common assumptions are: (1) homogeneity, where either the effect of exposure is consistent for all individuals or it does not depend on the instrument’s value, and (2) monotonicity, which asserts that genetic variants consistently influence exposure direction. While the relevance of these assumptions varies by estimation method, monotonicity is often most crucial for MR [100].

Moreover, MR findings should be interpreted cautiously due to canalization and developmental stability [82, 99]. Canalization occurs when compensatory developmental processes buffer against environmental or genetic disruptions during development—often due to buffering by maternal genetics. Developmental stability describes how a genotype expressed during fetal development may buffer the effect of the genotype under study (developmental adaptation) [106]. For instance, a fetus genetically predisposed to efficient nutrient utilization, may optimize metabolic pathways for fat storage and energy use to compensate for nutrient scarcity [107]. Both effects lead to underestimating the observed exposure-outcome associations, suggesting that the true causal effect might be stronger, potentially leading to false negative findings.

Application in Perinatal Pharmacoepidemiology

In perinatal pharmacoepidemiology, MR has been applied to assess the safety and effectiveness of antihypertensive medications during pregnancy [38, 103]. Fitton et al conducted a review of observational studies linking perinatal beta-blocker use to low birth weight [108]. Ardissino et al later confirmed this using MR, finding that genetically proxied exposure to beta-blockers was associated with lower birth weight. Specifically, for every 10-mmHg reduction in systolic blood pressure, birth weight decreased by 0.27 standard deviations from the average (p < 0.001). In contrast, genetically proxied calcium-channel blockers showed no effect (0.02 beta per 10 mmHg reduction in systolic blood pressure 95% CI − 0.04 to 0.07). These findings suggest calcium-channel blockers may have a more favorable fetal growth profile. However, pleiotropy may have biased these findings, as genetic variants may affect other traits unrelated to using calcium-channel blockers, potentially violating the exclusion restriction assumption [103].

Barry et al addressed pleiotropy and collider bias by incorporating offspring and paternal genotypes in their analysis. Their findings indicated that genetically proxied beta-blockers may reduce birthweight via offspring genotype, and showed little evidence of these variants influencing maternal conditions like gestational hypertension, a leading indication for the exposure [38]. Given that Barry et al took potential pleiotropic effects into account, their results are likely more robust. Fitton et al additionally noted mixed findings for calcium-channel blockers; studies with higher methodological quality found no impact on birth weight, aligning with the studies that used MR. Possibly, the other studies were hampered by residual confounding.

Conclusion and Future Implications

Genetically informed and conventional observational studies may yield conflicting results in perinatal medication safety research. These discrepancies can stem from genetic confounding—addressed in part by sibling designs and PGS—or due to (residual) confounding, which MR designs can account for. Each approach has distinct benefits and inherent limitations. Table 2 summarizes these designs.

Table 2.

Summary of currently used genetically informed study designs for assessing medication safety in pregnancy, with strengths, limitations, and considerations

Method	Strengths		Limitations		Considerations
Sibling design	Accounts for unmeasured confounding by controlling for shared familial environmental and genetic confounders		Inflation of misclassification bias Loss in statistical power due to reliance on discordant sibling pairs only which reduces sample size		Potential carryover effects should be examined by assessing whether outcomes vary by birth order, maternal age, or exposure sequence [22]
PGS	Can be used to account for confounding		Imperfect adjustment for genetic confounding due to limited explained variance of the PGS Currently limited use in non-European populations		Needs special calculations when used in family dyads or triads to distinguish between direct and indirect genetic effects [112]
Mendelian randomization	Use of genetic instruments that are randomly allocated at conception, reducing confounding and strengthening causal inference		Selection bias if the exposure affects fertility or pregnancy completion and the outcome is measured in the liveborn Possibility of violating the instrumental variable assumptions No time-varying exposures Potential for canalization and collider bias		Pleiotropy may arise and can be assessed for using MR-Egger regression Ideally need to have maternal, paternal, and offspring genotypes for maternal instrumental variables associated with outcome in the offspring through the offspring’s genotype

Depending on the research question and design, maternal genetic data are used (for MR and PGS) or is accounted for (as in sibling studies). Sibling designs control for shared genetic and environmental factors but may be biased by non-shared factors—an issue less likely in studies with unrelated individuals.

Traditional observational studies are prone to confounding, while the effects of genetic variants that are selected for in MR are unlikely to be confounded. However, MR cannot manage time-varying exposures, which observational studies can address—an important aspect in perinatal research. Mendelian randomization also relies on additional assumptions and availability of instruments, whereas PGS typically cannot explain a large proportion of phenotypic variance. Both MR and PGS rely on GWAS studies that determine associations between genetics and disease traits [109]. A comparison between MR and PGS and their applications can be referred to for a deeper overview [110]. Although still limited, genetically informed designs—sibling studies, MR, and PGS—are increasingly used to study medication effects during pregnancy on the offspring. While sibling designs are more common in behavioral research, their application in perinatal pharmacoepidemiology is evolving. These methods address genetic and familial confounding difficult to control in traditional designs. Sibling studies control for unmeasured shared familial confounders, MR employs genetic variants as instrumental variables, and PGSs accounts for genetic susceptibility.

Despite their strengths, these designs have limitations: sibling studies may be biased by non-shared confounders, MR can have violations in the exclusion restriction, and PGSs can have low predictive power. Triangulation—combining multiple methods—can strengthen causal inference by assessing consistency across approaches. Concordant findings increase confidence in causality, while discrepancies highlight potential biases and areas for improvement. In perinatal pharmacoepidemiology, these methods are still emerging and should be integrated with traditional designs to enhance reliability and impact. Mendelian randomization uses a few specific SNPs as instrumental variables to assess causal relationships between exposures and outcomes, while PGSs aggregate the effects of thousands of SNPs to estimate an individual's risk for a trait or disease.

Funding

This study is supported by funding from the UIOLifeScience convergence environment “UiORealArt”. EY is supported by The European Union (GeoGen #101045526 and ESSGN # 101073237) and the Research Council of Norway (#288083, #336078, and #331640).

Declarations

Conflict of Interest

AC, MZ, EWO, KG, TG, IS, EY, HN, and MvG declare no conflict of interest.

Ethical Approval

As this article constitutes a methodological overview of study designs and does not involve any primary data analysis or interaction with human subjects, ethical committee approval was not applicable.

Consent to Participate

Not applicable.

Consent for Publication

Not applicable.

Availability of Data and Material

Data sharing is not applicable to this article as no datasets were generated or analysed during the current study.

Code Availability

Not applicable.

Author Contributions

AC: conceptualization, writing—draft, review, and editing, visualization; MZ: conceptualization, writing—draft, review, and editing, visualization; EWO: review and editing of manuscript; KG: review and editing of manuscript; TG: review and editing of manuscript; IS: conceptualization, review and editing of manuscript; EY: review and editing of manuscript; HN: conceptualization, writing—draft, review, and editing, visualization, supervision; MvG: conceptualization, writing—draft, review, and editing, visualization, supervision, project administration. All authors read and approved the final version.