Identified Versus Interesting Causal Effects in Fertility Trials and Other Settings With Competing or Truncation Events

Abstract

Randomized trials are the gold standard for causal inference because assumptions needed to identify total treatment effects, capturing all pathways by which treatment affects the outcome, are guaranteed when the trial is perfectly executed (no loss to follow-up, perfect protocol adherence). However, when the outcome is subject to competing or truncation events—events determining that the outcome cannot occur or is undefined, respectively— total treatment effects may not be of interest (possibly undefined). Motivated by a trial comparing gonadotropin and letrozole, Chiu et al1 reviewed existing definitions of causal effect in competing or truncation event settings proposed over decades, including more recent proposals.2–9 They applied these definitions to studying effects of fertility treatments on offspring events, e.g., neonatal intensive care unit (NICU) admission. Both the total effect on live births and the total effect on NICU admission are defined and can be identified in this case, provided the trial was perfectly executed. However, even when identified, the total effect on NICU admission is not always of primary interest, which is why so many alternatives have been proposed. It is particularly challenging to interpret this total effect when an analysis suggests that, say, gonadotropin, on average, is more effective for producing live births but also increases the risk of NICU admission compared with letrozole. It is possible that the increase in NICU admissions is entirely explained by greater effectiveness of gonadotropin for producing live births. Alternatively, it could be partly explained by harmful actions of gonadotropin on a developing fetus. Many will care whether gonadotropin affects NICU admissions only by causing more live births. To address this, scientists need to consider alternative effects quantifying causal mechanisms. Chiu et al1 reviewed existing alternatives to the total effect, taking a largely agnostic view to their relative value. In a critical response to Chiu et al,1 Snowden et al10 advocate against the total effect as a useful causal question (also rejecting all alternatives). We agree with one of the messages suggested by Snowden et al10 that identifiable causal effects should not be confused with interesting causal effects. However, we disagree with other aspects of their comment. Despite a title suggesting the dangers of recommending an analytic approach without articulating a question, Snowden et al10 do just this. They make judgments on “appropriate denominators” and “flawed approaches” yet suggest no alternative effect definition, ultimately recommending “A more logical approach is to estimate multiple conditional probabilities ... ”.