UCLA

Background/Aims

Comparative effectiveness research (CER) investigates the effects of treatments and practices, thus requires causal inference. Routine data such as billing, pharmacy or EHR, while often incomplete on important confounding variables, are the usual sources of information for nonexperimental CER. The lack of randomization introduces important considerations regarding uncontrolled confounding, especially in large datasets, which potentially magnify systematic error. Yet, quantitative bias analysis in CER is not common practice. In this paper we formalize and demonstrate easy-to-implement record-level simulation techniques for analysis of uncontrolled confounding in cancer treatment CER.

Methods

We use recent advancements from the causal theory and risk analysis literature, specifically directed acyclic graphs (DAGs), and Monte-Carlo simulation techniques to introduce a novel form of record-level missing variable imputation that can be implemented during the core data analysis stage, making bias analysis more accessible using standard statistical packages. Further, our methods take into account varying levels of uncontrolled confounding by research center, or other clustering variable that may predict the level of unknown information in the dataset, and are specifically designed for implementation in large datasets, or data from multiple sources. We demonstrate these methods with two example sensitivity analyses of uncontrolled confounding in cancer treatment CER.

Results

Our methodology highlights the underlying causal model assumed for the main analysis in CER. Our technique uses the observed data lacking important confounding variables and informed estimates of the unmeasured variables to impute missing variables. The new variables now have a joint distribution with the observed data that would have been the case had they been observed fully under the assumed interrelationships. This technique is intuitively in line with the missing data framework and inference using partially observed distributions.

Conclusions

Sensitivity analysis for uncontrolled confounding is feasible and indispensable for CER. Unlike existing formula-intensive external adjustment techniques, the new technique can be implemented during core data analysis, is not outcome model specific, is at most semi-parametric and requires no esoteric software. Quantitative uncertainty analysis should be routine practice for CER in large observational data sources. Flexible methods accessible to all researchers should be a priority in this growing area of research.