Invited commentary: hypothetical interventions to define causal effects--afterthought or prerequisite?

‘‘It is philosophy, not science.’’ Physicists are familiar with this criticism of string theory, a theory that provides a unified description of all forces operating in the universe (1). Unlike philosophical arguments, scientific theories or their predictions need to be confirmed empirically. String theory involves an elegant set of mathematical equations; unfortunately, it is unclear whether its predictions are, or will be, testable—not a small obstacle for a ‘‘theory of everything.’’ Some of the string theorists’ tribulations regarding untestable predictions are shared by epidemiologists and other researchers who use counterfactual theory to draw causal inferences from observational data. I do not mean that epidemiologists are limited by the practical impossibility of conducting certain subatomic experiments. Rather, I refer to a fundamental shortcoming of causal inference from observational data in certain settings: the absence of a welldefined causal effect. When the causal effect of interest is ill defined, the counterfactual theory of causal inference from observational data and the elegant statistical methods derived from it lead to predictions that are untestable. Note that this problem is separate from the concept of bias or noncomparability between the exposed and the unexposed— confounding, selection bias, measurement error, random variability—that threatens the validity of the estimates of causal effects in epidemiologic studies. Ideally, comparability could always be achieved by randomization coupled with smart design and adequate funding. The study by Haight et al. (2) in this issue of the Journal highlights the need for a sharp definition of the causal effect of interest in epidemiologic research. The authors analyzed data from 1,655 older California residents followed for some time between 1995 and 2001. The goal of the study was to estimate the joint causal effect of body composition and physical activity—the exposures—on several measures of functional limitation in the elderly. To accomplish this goal, the authors strived to 1) ensure approximate comparability (exchangeability) across levels of exposure and 2) use an analytic method (inverse probability weighting) that preserves comparability. First, Haight et al. (2) collected data on a number of variables that may confound the causal effect of the exposure on the outcome. These variables included presence of chronic conditions, self-rated health status, body mass index, living arrangements, walking speed, and so forth (3). Although exchangeability can never be guaranteed in an observational study, for the purposes of this commentary let us assume that the authors succeeded at achieving approximate exchangeability. Second, the authors noticed that the use of standard statistical methods could lead to loss of the exchangeability so laboriously obtained by careful design and data collection. This (selection) bias may appear in longitudinal studies with time-varying exposures when there are time-varying confounders and these confounders are affected by prior exposure (or share common causes with exposure) (4). Both conditions are met in the study by Haight et al. (2) because time-varying confounders such as health status are also affected by the exposure. To overcome this problem, Haight et al. used inverse probability weighting to estimate the parameters of a marginal structural model. Robins (5) developed methods based on inverse probability weighting in the context of counterfactual theory for causal inference from complex longitudinal data. Under the assumption of no model misspecification, these methods do not introduce selection bias even if the time-dependent confounders are affected by prior exposure. For example, in studies of the effect of antiretroviral therapy on the risk of acquired immunodeficiency syndrome, estimates from marginal structural models are closer to those obtained from randomized trials, while estimates from standard statistical models are greatly attenuated (6, 7). The two articles by Haight et al. and Tager et al. (3) are innovative in that they describe the first actual analysis of marginal structural models applied to