The Importance of Rigorous Epidemiologic Methods to Advancing Science in Paediatric and Perinatal Epidemiology

The evolution of innovative study designs and advances in epidemiological methods over the past few decades have enabled robust causal and non-causal inferences in observational research. This special issue of Paediatric and Perinatal Epidemiology featured a collection of peer-reviewed papers that reflect recent methodological advances in the field and their application to address common biases in paediatric and perinatal epidemiology. This issue is organised into seven broad topics: selection biases in epidemiologic research, exposure misclassification, time-varying exposures, causal inference, emulation of target trials, missing data and other methodology-related topics. To encourage scientific dialogue, every paper is accompanied by a thoughtful, balanced commentary (now a recurring theme for all manuscripts published in Paediatric and Perinatal Epidemiology!) that highlights the paper's strengths and identifies gaps for future research. Various issues can lead to selection bias in epidemiologic research, and bias due to left truncation is perhaps the least studied. Kinlaw and colleagues [1] undertook a Monte-Carlo simulation to study the impact of left truncation bias when estimating the causal effect of placental abruption on perinatal mortality in the context of abnormal placentation, with spontaneous abortion as a censoring event. They show that left truncation can bias the association between abruption and perinatal mortality, differentially depending on whether placentation was normal or abnormal. Dib and colleagues [2] conducted a simulation study to compare the direction and magnitude of bias arising from two censoring mechanisms (loss to follow-up and competing events) across three common analytic approaches: inverse probability weighting, instrumental variable estimation, and the sibling-comparison design. Their study demonstrated that the mean bias in all three approaches generally deviated away from the null in the same direction and often to a similar magnitude. LaTour and colleagues [3] also undertook a simulation study to examine bias resulting from censoring versus accounting for healthy live birth as a competing event in estimates of the total effect of prenatal medication use on pregnancy outcomes. Their study shows that censoring healthy live births results in overestimated risks and biased, imprecise total treatment effect estimates. All three papers address selection bias from the perspective of censored events prior to the event in question or contrasting different analytic methods to address censoring. Nearly all studies suffer from varying degrees of exposure misclassification, potentially biasing effect estimates. The extent to which uterine fibroids, a common gynaecological condition that is often underdiagnosed, potentially biases the results in epidemiologic studies, is a unique problem that remains poorly addressed. DiTosto and colleagues [4] address this question through a simulation study and demonstrate that underdiagnosing uterine fibroids can bias results toward the null, particularly with self-report or modest effect sizes, potentially obscuring true effects. These findings have important implications for reproductive health research, and as the authors note, the problem of misclassification of uterine fibroids is indeed an ‘invisible burden’! The second study addresses the impact of exposure misclassification in case–control studies of self-reported medication use (nonsteroidal anti-inflammatory drugs [NSAIDs] in early pregnancy) and the risk of amniotic band syndrome. Analysing data from the National Birth Defects Prevention Study, 1997 to 2011, Adrien and colleagues [5] show that the magnitude of bias depended on the mechanism and the extent of misclassification of the exposure data. Exposures that vary over time present additional complexities to the analysis. We feature two interesting papers that present important applications of time-varying exposures. The contribution by Dayan and colleagues [6] examines how reproductive risk factors affect the risk of long-term chronic conditions (e.g., cardiovascular disease). These authors apply a weighted cumulative exposure (WCE) modelling strategy that integrates exposure data based on the timing, severity, and recurrence. They show that the use of flexible WCE modelling to quantify risks of pregnancy events (e.g., severe maternal morbidity), adjusted for reproductive and non-reproductive cardiovascular risk factors, can account for variations in timing and severity of these events and can capture their cumulative effects across a woman's reproductive trajectory. The second paper by Wang and colleagues [7] examines time-varying associations between physical activity and injury risk in children. Specifically, the authors undertook an empirical analysis to assess how previous activity patterns are associated with injury risk among children. This study shows that when time-varying exposures are modelled using the WCE method, a complex temporal relationship exists between past physical activity history and injury in children. A fundamental tenet in observational epidemiology is to gauge how well treatment effect evaluations remain generalisable using observational data. This is the causal issue that Burgos-Ochoa and colleagues [8] attempt to tackle in their elegant work. Using data from the National Health and Nutrition Examination Survey (NHANES) 2017–2020, the authors assessed the effect of household second-hand smoke exposure, a non-randomised treatment, on blood pressure in school-aged children. The associations were derived using inverse probability of treatment weighting (IPTW), g-computation, targeted maximum likelihood estimation (TMLE), and regression adjustment. They show that the generalisability of the average treatment effect, as estimated from data sampled using common survey designs, to a defined population requires the use of sample weights. The approach to emulating a target trial using observational data brings us closer to evidence from randomised controlled trials for inferring causal treatment effects. We feature two papers on this topic. Simmons and colleagues evaluate the impact of increasing midwifery-led prenatal care on birth outcomes, a study that shows that, based on the g-computation method, with exposure set to a 50% increase in midwifery-led care, the prevalence of caesarean deliveries was reduced by 5.4% (95% confidence interval [CI] −5.7, −5.1) and of maternal infection decreased by 1.3% (95% CI −1.6, −1.0) [9]. Unexpectedly, the corresponding prevalence estimates of postpartum haemorrhage increased. Given that midwife-assisted deliveries remain underutilised, these findings bear important policy implications. There is a strong potential for immortal time and selection biases in studies exploring the benefits of antihypertensive medication for hypertensive diseases during pregnancy. The second study by DiTosto and colleagues [10] explores this through the design of a target trial emulated using observational data to estimate the effect of treatment for non-severe gestational hypertension at 24 to 35 weeks of gestation on the risk of adverse perinatal outcomes. The authors present an eloquent argument of how a target trial emulation can circumvent these aforementioned biases by aligning follow-up with treatment initiation. Missing data in observational studies is common, and appropriate handling is crucial for drawing robust conclusions. A study from Norway by Sakai and colleagues [11] compared the validity of three longitudinal methods (the past pregnancy record, the future pregnancy record, and the closest pregnancy record) with that of the multiple imputation method. They show that the longitudinal method outperformed the multiple imputation method in predicting missing data based on maternal characteristics in a medical birth registry. This issue also features two other important papers that address substantive topics in paediatric and perinatal epidemiology with innovative analytic approaches. The first by Srinivas and colleagues [12] examine the effect of attention-deficit/hyperactivity disorder (ADHD) medication use in pregnancy and risk of preterm birth based on a population-based cohort study with a combined analysis of data from Norway and Sweden. The authors show that women who received two or more filled prescriptions (medications for ADHD) during pregnancy were associated with increased preterm birth risk. This study provides a nice balance of treatment risks versus untreated illness on preterm birth, an important consideration for reproductive-aged women faced with challenges of ADHD. The second paper by Ehresmann and colleagues [13] develops a privacy-preserving process to link the health records of patients shared by two health systems for a perinatal health study, and to assess data enhancements associated with the linkage. The authors show that the proposed matching approach can increase study size, patient diversity, and data completeness in a privacy-preserving manner for perinatal health studies among patients who use multiple healthcare systems. We hope the readership of Paediatric and Perinatal Epidemiology appreciates the important methodological contributions and nuances identified in this collection and how their application to paediatric and perinatal epidemiology, and beyond, can yield insightful inferences. The editors take this opportunity to thank all contributors for their insightful research and peer reviewers for their thoughtful and timely reviews. Cande Ananth and Sonia Grandi conceived and co-wrote this editorial. The authors have nothing to report. The authors declare no conflicts of interest. No data are being used for this editorial.