Response to Invited Commentary Kramer et al. Respond to ‘‘Causation or ‘noitasuaC’?’’

We appreciate the opportunity to reflect on the thoughtful commentary by Schisterman et al. (1). Directed acyclic graphs, marginal structural models, and structural nested models provide a useful approach for modeling time-varying exposure and time-varying confounding and, thus, for reduc-ing the potential for reverse causality (1–4). However, as the authors point out, such causal modeling requires longitudinal data on the reasons (causes) for changes in exposure, if those reasons have a causal effect on the outcome or are themselves intermediate outcomes. CD4 (T-helper) cells provide a com-mon and useful example in the context of outcomes (e.g., survival) of treatment in patients with acquired immunodefi-ciency syndrome (2). The CD4 count varies over time and can reflect both the status of the disease (i.e., an intermediate out-come) and the efficacy of treatment. Low CD4 counts indicat