No Rates Were Harmed in the Making of This Paper: Response to Critiques.

The debate featured in this issue highlights the complexity of identifying the ‘appropriate’ denominator for estimating gestational-age-specific risk of postnatal endpoints. According to Caughey and Snowden, interpretable estimates for most neonatal endpoints can only be obtained with live births as the denominator. Smith notes how neither live births nor fetuses perfectly capture the population at risk of neonatal death, although analysis based on the former may be less flawed. As some conditions diagnosed in childhood may nevertheless start in utero, Schisterman and Lindsey emphasize how the ‘correct’ approach can only be evaluated given a specific question. Joseph favours fetuses as the denominator in instances involving outcomes with a plausible prenatal origin. Although most researchers agree on fetuses as the denominator for several endpoints (e.g. antepartum stillbirth, preterm birth, and pregnancy complications), risk of outcomes identified after birth is generally measured among live births, despite the limitations imposed by such a choice. These ‘conventional’ estimates are predictive and clinically useful but cannot be relied upon for causal interpretation, given how they often result in overall harmful exposures appearing to be protective at preterm weeks. The extended fetuses-at-risk (FAR) formulation is touted as providing causally interpretable estimates, although what is meant by ‘causal’ is unclear. The aim of my paper was to show how, due to their dependence on the probability of live birth, weekspecific FAR rates of postnatal endpoints can be higher in pregnancies exposed to a factor that reduces length of gestation, regardless of whether such a factor actually increases risk. Given that my demonstration relied entirely on simple algebraic re-formulation of the week-specific FAR rate (as used, e.g., in Joseph et al.), I was surprised by some of the commentators’ reactions. There was no model; no rates were harmed, modified, or burdened with assumptions in the making of my paper. All that was done was to multiply both the numerator and denominator of the week-specific FAR rate by the number of live births at that week – that is, multiplication by 1.0 – an accomplishment that can hardly be characterized as ‘courageous’. Schisterman and Lindsey criticize my formula as being ‘overly simplified’, compared with that of Kramer et al., neglecting that the latter measured the composite outcome of stillbirth and neonatal death. My formula, on the other hand, applies to any postnatal outcome, be it cerebral palsy, which often originates prenatally, or diaper rash, which – presumablydoes not. Regardless, neither can be diagnosed until after live birth. Schisterman and Lindsey’s criticism is puzzling also in light of the fact that Schisterman co-authored a commentary to Kramer et al.’s paper above, remarking on how using composite outcomes for competing events changes the question. Unlike Joseph, I share this perspective and, with others, I believe that an association between an exposure and a childhood condition is interpretable only when estimated among those at risk, although assessment of the exposure effect on competing outcomes would be useful. Knowing the question is thus not only necessary to evaluate whether an answer is correct, but also to make sense of it (or it might as well be ‘forty-two’, the answer to the ultimate question that nobody knew). If we could identify the fetuses that would have developed the endpoint of interest regardless of whether and when they were born, FAR analyses would be appropriate in many situations. In my example, the extended FAR approach applied to a simple scenario failed to yield a causally interpretable answer to a specific question. Could it provide the ‘correct’ answer in a different scenario or to a different question? Possibly, but – Correspondence: Olga Basso, Department of Epidemiology, Biostatistics, and Occupational Health, McGill University, Purvis Hall, 1020 Pine Avenue West, Montreal, QC H3A 1A2, Canada. E-mail: [email protected] bs_bs_banner