Pregnant women are commonly excluded from pre-approval randomised clinical trials; therefore, post-approval studies often rely on real-world data (RWD). Target trial emulation (TTE) has become a widely adopted conceptual framework for causal inference, and its implementation in pregnancy studies necessitates consideration of certain nuances, as illustrated in an accompanying study by DiTosto et al. 1 In this commentary, we expand the overall description of the framework, including its limitations, and elaborate on specific features for pregnancy studies, with emphasis on sequential emulation. The TTE is a structured approach to study design that helps avoid many common methodological pitfalls in causal analyses of RWD. It consists of two steps. First, investigators articulate a well-defined causal question by specifying the protocol elements of a target trial that would answer. Second, they answer the causal question by emulating those protocol elements through a study based on RWD 2, as shown in the application by DiTosto et al. 1 This two-step process is usually summarised in a table (table 1 in DiTosto et al.). A recent addition to that table is a row with the identifying assumptions, that is, the assumptions required to identify the effect even when the sample size is infinite 3. Such assumptions will depend closely on the type of causal contrast being made. For example, to identify the intention-to-treat effect, exchangeability only needs to be assumed at baseline, provided follow-up is complete (such an assumption is weak in the target trial thanks to randomisation, but in the emulation, it is stronger because it relies on adjusting for baseline confounders). In contrast, to identify the per-protocol effect, additional assumptions related to post-baseline confounding are needed, similar to those in randomised clinical trials (although rarely explicitly stated), because factors that influence adherence to the treatment strategies may also influence the outcome and be affected by prior treatments. The identifying assumptions, together with the estimand (i.e., eligibility criteria, treatment strategies, assignment, outcomes, follow-up, and causal contrasts) and the estimator (i.e., the statistical procedures applied to the data), convey a transparent depiction of the study design 2. The target trial is not a completely idealised clinical trial that we would design in the absence of constraints, but a pragmatic trial that can be realistically emulated within the constraints of available RWD 3. For example, it is generally not possible to emulate a placebo-controlled trial, trials of drugs not on the market (the corresponding treatment strategy will not be represented in the RWD), or trials with double-blinded treatment assignment. Therefore, the specification of the target trial must be grounded in the practical realities of clinical care and the data it generates. In the case of DiTosto et al., inferences apply to the use of methyldopa as an antihypertensive agent because the use of other drugs like nifedipine and hydralazine was barely represented in the data they used (despite nifedipine and hydralazine being pregnancy compatible, they were used much less frequently than methyldopa in the geography and during the time covered by the Tsepamo study). Importantly, TTE does not address the limitations inherent to RWD, such as unmeasured confounding or measurement error, and is subject to the same transportability limitations as randomised clinical trials. In the case of DiTosto et al., one should consider whether the observed effect estimates would apply to populations with a different HIV prevalence or where clinical practice standards differ. Pregnancy losses and gestational age are essential in the specification and emulation of target trials for interventions initiated during pregnancy 4. Pregnancy losses act as competing events (i.e., events that make it impossible for the event of interest to occur or be measured) 5 for birth outcomes; for example, in the present study, examining initiating an antihypertensive treatment at gestational Week 24 or later, stillbirths are competing events for preterm, small for gestational age births and neonatal deaths. Although the literature is extensive on methods to handle and define causal effects in the presence of competing events, within the TTE framework, it has been proposed to report the risk of non-live birth (e.g., via the Kaplan–Meier estimator to account for right censoring), the risk of the outcome among live births, and the risk of live birth with the outcome (e.g., via the product of the probability of live birth times the risk for the outcome among live births) 4, 6. Gestational age is important for two reasons. First, many outcomes are defined by gestational age; for example, by definition, no preterm deliveries can occur after gestational Week 37. This important consideration led to work by Hernández-Díaz et al. 4 proposing to classify pregnancy trials into four categories based on the range of gestational age for the relevant outcome of interest: ‘periconceptional’ (before 12 weeks), ‘early pregnancy’ (before 20 weeks), ‘late pregnancy’ (before 37 weeks) and ‘any-trimester pregnancy’ (at any time during pregnancy) trials. Second, the risk of outcomes changes over gestational age, hence highlighting the importance of balancing gestational age across treatment groups. In the case of DiTosto et al., an equal distribution of gestational age in both treatment groups was achieved via the creation of a sequence of gestational week-specific cohorts. Emulating a target trial requires appropriate determination of time zero for each patient, which may not be straightforward if one of the treatment arms consists of ‘no treatment’. To address this and to ensure comparability of gestational age at time zero, DiTosto et al. conceptualised observational data as a sequence of target trials. In this sequence, at each gestational week, women who met the eligibility criteria and had not yet initiated antihypertensive therapy would be randomly assigned to either initiate antihypertensive therapy or remain untreated throughout follow-up. They would subsequently be followed until the occurrence of the outcome of interest or discharge from the birth hospitalisation. Thus, a new trial starts every week of gestational age, which is the time zero for a new emulated trial, with eligible women being assigned to the antihypertensive group if they initiate treatment during that week or to the control group if they do not. After treatment assignment, women are followed as in the target trial. DiTosto et al. combined the trial-specific estimates to estimate an observational analogue of the intention-to-treat effect of initiating antihypertensive medication any time from gestational Week 24 to Week 36. The intention-to-treat effect in this scenario would be close to the effect under full adherence if the observed adherence to both treatment strategies was complete (i.e., patients in the antihypertensive therapy group do not stop therapy in the absence of adverse events, and patients in the control group do not initiate antihypertensive medication while not progressing to severe gestational hypertension). Sequential emulation can be more efficient than emulating a single trial by choosing a single eligibility time (e.g., at random or the first eligible time) 7. A common alternative design would be to handle treatment initiation as a time-varying exposure; that is, rather than emulating a sequence of target trials, this single-cohort design would start follow-up at 24 weeks of gestation and code antihypertensive treatment as a time-varying variable that switches from 0 (no antihypertensive treatment) to 1 (antihypertensive treatment) on the week of initiation and stays as 1 thereafter. Models that account for exposed person-time can then be used to estimate a time-averaged hazard ratio for initiation versus no treatment. The resulting estimand would have differed from those obtained by DiTosto et al. for several reasons. First, the corresponding target trial is one in which eligible women are recruited at gestational Week 24 and are randomly assigned to initiate antihypertensive therapy at subsequent gestational weeks, as long as they continue to meet the eligibility criteria at those times. Such a target trial corresponds to an estimand that approximates the per-protocol hazard ratio obtained by pooling sequential trials, censoring women when they deviate from their treatment strategies, assuming that time-varying confounders do not exist, and that women remain eligible for all weekly trials 8. This ‘single-trial’ design has drawbacks compared with the ‘sequential target trials’ implemented by DiTosto et al. First, time-varying risk factors associated with treatment initiation after gestational Week 24 act as ‘time-varying confounders’ and must be adjusted for using g methods, such as inverse probability weighting of a marginal structural model 9 or the g-formula 10. In contrast, when sequential target trials are emulated, these factors are time-fixed confounders occurring at each trial's baseline, with the estimation of the intention-to-treat effect only requiring adjustment for baseline confounding. Second, because antihypertensive therapy can be initiated at different gestational weeks, this approach does not naturally lead to the estimation of absolute risks; therefore, effect estimates are based on the time-averaged hazard ratio. Third, women may become ineligible after gestational Week 24 if, for example, they develop severe hypertension. When sequential trials are emulated, persons with severe hypertension are excluded at baseline. In the single-trial approach, however, persons who initiate antihypertensive therapy and those who do not can become increasingly noncomparable after Week 24, especially when such therapy is prescribed after a diagnosis of fetal growth restriction (which entails a higher likelihood of subsequent adverse birth outcomes). The TTE framework is useful for avoiding design-induced biases in pregnancy studies that utilise RWD. Its emphasis on proper alignment of time zero, eligibility, and treatment assignment proves critical in handling issues related to gestational age in pregnancy studies. X.G.A. drafted the commentary, and A.M. reviewed and revised it. Both authors contributed to the final version of the manuscript. The authors have nothing to report. Both authors are employees of RTI Health Solutions. The authors declare no conflicts of interest. The authors have nothing to report.
Garcia‐Albeniz et al. (Sun,) studied this question.