Commentary

In this issue, Mendola and colleagues1 present estimates of the controlled direct effect of preeclampsia on several perinatal outcomes. The controlled direct effect is the counterfactually defined exposure effect that would be observed under an intervention that sets a mediator to a fixed value uniformly in the population. When exposure-mediator interactions are present, or when the exposure affects confounders of the mediator outcome relation, controlled direct effects cannot be estimated using standard regression models. Mendola and colleagues1 use marginal structural models (MSMs) to deal with these difficulties and note that, under several assumptions, “marginal structural models provide us with a causal response” to questions about the effects of hypothetical interventions. In this commentary, we discuss the importance of the counterfactual question of interest. It is generally understood that, while they can better handle complex biases that beset standard approaches, methods such as MSMs do not automatically confer a counterfactually defined causal interpretation. Less clear, however, is that sometimes the main problem lies in the research question itself. UNMEASURED CONFOUNDING Several obstacles hamper the estimation of causal effects using observational data. These obstacles are usually described as unverifiable identifiability assumptions that must hold, such as the absence of uncontrolled confounding. In constructing their model, the authors considered most confounders of the association between preeclampsia and the outcomes under study (with the notable exception of smoking), as well as factors that affect both duration of gestation and outcome risk (including those factors affected by preeclampsia). Their analysis shows that, when delivery is set to 37 weeks or later for all women in the population, preeclampsia substantially increased the risk of virtually all the examined outcomes. Overall, preeclampsia is associated with increased perinatal mortality2 and with cerebral palsy.3 However, among preterm births, preeclampsia appears to be associated with a lower risk of both outcomes, while the opposite is true for births at term.3–5 Other risk factors yield similar findings, likely because at least some of the other unmeasured causes of preterm birth directly affect risk of these endpoints.5–7 An alternative question would be whether the apparent protective effect of preeclampsia among preterm births would still be observed after adjusting for measured mediator–outcome confounders. Assessing the effect of preeclampsia under a hypothetical intervention in which labor is induced at a specified week, rather than delayed, would better coincide with available strategies to manage preeclampsia; unfortunately, addressing this question is complicated by a number of problems—particularly the fact that some confounders of the mediator–outcome relation are unknown.8,9 Mendola and colleagues1 carefully address potential unmeasured confounding with detailed sensitivity analyses among infants delivered at term, which suggested that only rather extreme mediator–outcome confounding scenarios would result in estimates substantially different from those presented. However, the unmeasured factors in question almost certainly play a much larger role at preterm weeks, as virtually all preterm births will be the result of some pathologic process.5,6,8 Thus, a similar sensitivity analysis for the direct effect of preeclampsia under an intervention causing delivery before 37 weeks could shed some light on how severe the confounding must be to eliminate the protective effect of preeclampsia. COUNTERFACTUAL CONSISTENCY Threats to validity posed by unmeasured confounding are well recognized. Less well recognized are the difficulties stemming from violations of counterfactual consistency.10,11 Such violations originate from the fact that both the exposure (preeclampsia) and mediator (gestational age at birth) can be manipulated only to a limited extent, if at all. In general, results from observational studies cannot be interpreted causally when researchers do not explicitly specify interventions on an exposure or mediator that can bring about the expected counterfactual outcomes.12 The situation in perinatal epidemiology is complicated by the fact that the ability to delay delivery is limited, and whether and when preeclampsia will occur cannot be reliably predicted. Nevertheless, the authors present results for the relation between preeclampsia and various perinatal outcomes were all women to give birth at or later than 37 completed weeks of gestation. These results do provide insight on the magnitude of the relation between preeclampsia at term and several perinatal outcomes that are not due to biases encountered in traditional approaches to mediation analysis.13 However, as Mendola et al note, it is difficult to conceive of an intervention that will feasibly delay gestation to ≥37 completed weeks and, depending on when preeclampsia manifests itself, such an intervention may well result in worse, not better, maternal and perinatal outcomes. Thus, while the authors emphasize the importance of an underlying intervention that sets the mediator to a specific level, they do not identify one that might bring about such a change, and they recognize that such an intervention may not be appropriate. When the aim of an analysis is to quantify the magnitude of an exposure–outcome relation using novel methods without ascribing a counterfactually-based causal interpretation, specific details of the (possibly hypothetical) intervention take less precedence. However, when using observational data to infer causal effects, a central challenge is to pre-emptively identify the clinically relevant intervention of interest, and then estimate those intervention-based effects. Treating the intervention as an afterthought once an analysis has been done14 can seriously hamper the interpretation of analytic results in a more consequential light.15 TARGETING RELEVANT RESEARCH QUESTIONS Even if strong unmeasured confounding and counterfactual consistency were not at issue, controlled direct effects require that the mediator be set to a fixed value uniformly in the population. This is often an unrealistic target, and particularly so in this case. To address this issue, recently proposed stochastic mediation contrasts16 may be used to assess the relation between preeclampsia and perinatal outcomes under hypothetical interventions that shift the distribution of gestational age, rather than force everyone to be delivered at a single gestational age interval. This latter approach would provide evidence on the magnitude of the relation between preeclampsia and perinatal outcomes that would be observed under more realistic induction strategies. Ultimately, though, the challenge for obstetricians managing preeclampsia presenting early in pregnancy lies in balancing the trade-off between delaying delivery (when waiting may pose risks to both the mother and the infant) and delivering a very premature baby. Finding a causal answer to this question using observational data is complex, and would require a careful analysis in which maternal and neonatal risks are contrasted under different scenarios of labor induction and expectant management. Such an analysis could be carried out using, for example, the parametric g-formula,17 but it would require data on the timing of onset of preeclampsia symptoms, and on whether delivery was spontaneous or not. These data are often hard to come by and, indeed, were unavailable in the study by Mendola et al.1 Given the somewhat contradictory evidence from clinical trials on this question,18,19 using the parametric g-formula or related strategies20 may be a cost-effective means of shedding additional light on this clinically important question. ADDITIONAL CONSIDERATIONS Although Mendola et al do not specify how preeclampsia was defined in this study, the diagnosis was likely based on the presence of both hypertension and proteinuria. A recent report by the American College of Obstetricians and Gynecologists21 suggests that physicians consider other factors in addition to hypertension, such as low platelets and other problems, and that the presence of elevated levels of proteinuria be no longer required to diagnose preeclampsia. Using this approach would have resulted in a broader set of patients being defined as exposed, although it is difficult to predict whether such a change in definition would have changed the findings. Finally, Mendola et al used an autoregressive working correlation matrix to account for correlations between different pregnancies by the same woman. However, Tchetgen Tchetgen et al22 have recently shown that, for inverse probability weighted generalized estimating equations, using anything but an independent working correlation matrix can lead to bias; this is likely to be only a minor issue here, as fewer than 10% of women contributed more than 1 pregnancy. CONCLUSIONS Recent decades have seen a proliferation of causal inference methods, including inverse probability weighted MSMs, to estimate well-defined mathematical quantities, such as the controlled direct effect. However, the estimates they provide cannot always be translated into counterfactual-based causal quantities, especially in the absence of clearly delineated exposure or mediator interventions, or when the target mediator values may be unachievable or inadvisable. ABOUT THE AUTHORS OLGA BASSO is a reproductive and perinatal epidemiologist, with a particular interest in understanding the separate contribution of the different factors acting before and during pregnancy on perinatal and child health. Educated in Italy and currently Associate Professor at McGill University (Canada), she has held positions at the Danish Epidemiology Science Centre (Denmark) and the National Institute of Environmental Health Sciences (Research Triangular Park, NC). ASHLEY I NAIMI is an assistant professor at McGill University (Canada) whose research focuses on causal inference, epidemiologic methods, and the social determinants adverse birth outcomes.

Récupéré en direct depuis OpenAlex et désinversé. Les résumés ne sont pas conservés dans cette base de données : les index inversés représentent 8,6 Go des 9,3 Go de texte de la base, et le serveur dispose de 13 Go libres.