The lay abstract featured today (for Dealing with time-dependent exposures and confounding when defining and estimating attributable fractions—Revisiting estimands and estimators by Johan Steen, Paweł Morzywołek, Wim Van Biesen, Johan Decruyenaere and Stijn Vansteelandt) is from Statistics in Medicine with the full article now available to read here.
Invasive interventions and impaired health during hospitalization put patients at increased risk of infection and other possibly life-threatening complications. Accurate assessment of the burden of such harmful exposures is challenging but important for informing clinical and policy decisions. Methodological challenges arise, for instance, because patients with an unfavorable prognosis are at increased risk of such exposures, which makes it difficult to differentiate the burden of exposure from that related to other prognostic factors (due to residual ‘confounding bias’). In addition, patients may not live long enough to get exposed, such that exposure may not only lead to increased mortality, but (premature) death may also prevent patients from getting exposed. This form of reverse causality may translate into a spurious survival advantage among exposed patients (often referred to as ‘immortal time bias’). Quantification of the burden of hospital exposures in terms of a crude comparison of death rates in unexposed and exposed patients is therefore subject to different forms of bias.
In this article, the authors review and revisit two more refined effect measures (so-called ‘estimands’) that have been proposed in the literature to express the fraction of cases that can be causally attributed to an exposure with a time-dependent onset. They clarify that one of the proposed measures, despite being termed an ‘attributable fraction’, cannot be interpreted as a measure of causal attribution because it likewise captures a subtle notion of reverse causality. This built-in bias may translate into seemingly protective exposure effects whenever, in truth, the exposure is unconfounded and exerts no causal effect. The authors point out that this subtle form of bias has long gone unnoticed by the statistical community and that, along with the application of misguided terminology, this may have contributed to persisting misinterpretation in the medical literature.
This article therefore aims to improve the applied statistician’s understanding of interpretational differences between proposed effect measures, connections and differences between their respective estimation approaches, and the nature of different sources of bias within a formal causal framework. The assumptions under which a causally interpretable measure can be estimated from observable data (without systematic bias) are formalized and recent estimation approaches are discussed and compared to earlier work. As an illustration, the authors present a real-life application to estimate the fraction of ICU deaths that can be causally attributed to ICU-acquired infections.