Targeted Error Correction: a framework for making cheap data useful
报告人： Mike Baiocchi, Stanford School of Medicine
时间：2016-06-30 14:00 ~ 15:00
Access to cheap, easy-to-collect data sources is increasing (think: electronic health records). These data sources are coming to dominate many corners of health research. Regrettably, they are not of particularly high fidelity. In this talk we detail a framework for addressing key issues of data quality – measurement error and missingness. The framework is quite general but requires a bit of shoe leather from the researcher. The benefits are valid inference, even in situations where the missingness is informative and errors are differential, even in cases where errors are dependent on the outcome of interest. The seminar will introduce a slight generalization of the usual multiple imputation framework. We’ll then motivate the problem using work done in the Veteran’s Affairs to estimate the relationship between HIV status and cancer rates.
About the Speaker:
Mike Baiocchi is an Assistant Professor in the Stanford Prevention Research Center at Stanford School of Medicine. Before that he was a Stein Fellow in the Statistics Department at Stanford. He specializes in creating simple, easy to understand methodologies for causal inference and observational studies. In graduate school he had the privilege of training with Dylan Small. He also worked closely with Paul Rosenbaum and Scott Lorch. He became a statistician to help analyze and improve health care.