Figure 2.

History-adjusted marginal structural outcome model. Survival was modeled via a weighted logistic regression model for repeated measures using rolling windows (indicated by hashed box). For each subject month ending at time t, the probability of survival was modeled based on Hgb-Var, level, and temporal trend from the previous 6 mo (t − 6 through t − 1). Adjustment was made for baseline demographics and comorbid diseases, and laboratory measures (Labs t − 9, t − 8, t − 7) and medication exposures (Medications t − 9, t − 8, t − 7) averaged over the 3 mo before hemoglobin assessment. Patient-months were weighted as described in the text. Windows were then rolled forward by 1 mo to generate the next observation (arrow). Variance estimates were corrected for lack of independence of observations within subject.