Estimating Treatment Effects Using Longitudinal Observational Data
Keywords: marginal structural models, inverse probability weighting, local control
We introduce and demonstrate implementations of local control and longitudinal trajectory analyzes for assessing treatment effects in observational data that focus on making comparisons within clusters of most similar patients.
Summary: The availability and use of very large datasets of electronic medical records, health care claims, disease registries or other observational (non-randomized) studies posts new challenges in estimation of real-world treatment effects. These datasets are highly likely to feature treatment selection bias and confounding among observed patient X-characteristics as well as potential for important, unmeasured confounders. Furthermore, interest is likely to be focused more on quantifying the full spectrum of patient differential response to treatment (a distribution of effect sizes) than on just overall means (main effects) and associated p-values. Thus we introduce and illustrate the use of "Local Control" techniques and/or classes of latent, longitudinal trajectories for assessment of treatment effects. For instance, clustering of patients in X-predictor space yields a local, robust (non-parametric) estimate of the counterfactual treatment outcome difference (treatment minus control) for each patient. We then present ways to verify that the observed distribution of these constructed, local Y-outcomes is less biased than the overall comparison and to systematically check its sensitivity to choice of clustering parameters. Finally, any number of well known data mining methods can then be used to predict the most likely Y-outcome for a patient with any given set of X-characteristics.