Cognitive Effects of Prenatal Exposure to SSRIs in Children: A Systematic Review and Meta-Analysis.

Causal inference requires that the intervention precede the outcome, imposing a necessary temporal ordering. However, in practice, some outcomes may be unobservable because units drop out of the study or become inaccessible to the researcher for other reasons during the period between the intervention and the outcome measurement. This issue can arise in virtually any type of study, even in randomized controlled trials. When outcomes are missing, identification assumptions are necessary to relate causal estimands to observable data quantities. A commonly adopted assumption is the “missing at random” (MAR) mechanism, under which outcome missingness is independent of the value of the outcome itself, conditional on all other observed data. Thus, any association between missingness and the outcome is fully explained by observable characteristics, enabling point identification of causal effects (see, e.g., [–]). If instead missingness is a function of the outcomes themselves, or unobserved variables that determine both missingness and the outcomes, then the outcomes are said to be “missing not at random” (MNAR), and commonly targeted causal effects, such as the average treatment effect, are not in general identifiable. Unfortunately, testing whether MAR holds versus MNAR is not possible, and MAR is often implausible in practice. For example, in randomized trials examining disease progression or mortality, unobserved deterioration in health could directly lead to both study dropout and an adverse outcome. These issues are even more pronounced in observational studies utilizing electronic health records or insurance claims data, which typically feature higher rates of missingness. In this setting, outcomes such as diagnoses or procedures might not be documented if patients seek care elsewhere or forgo treatment altogether. Additionally, the reasons for missingness in such settings are often unknown yet plausibly linked to critical downstream outcomes, such as loss of insurance coverage. Furthermore, when missingness constitutes a competing event, an outcome that prevents the primary outcome from occurring, MAR is violated, and traditional causal estimands such as the average treatment effect may not be interpretable []. In contrast to the substantial literature addressing potential violations of the “ignorability” assumption, which states that treatment assignment is independent of potential outcomes conditional on observed covariates, there has been relatively limited exploration of MAR violations in causal inference. While ignorability holds by design in a randomized controlled trial, it generally does not hold in an observational study. As a result, a substantial body of literature has proposed sensitivity analyses that do not generally point identify causal effects, but instead provide bounds in cases where ignorability may not hold []. However, less work has extended these methods to address violations of MAR when outcomes are missing.