Bayesian Bias Modelling for analysis of observational data

点击翻译按钮获取中文摘要

Observational studies are important tools for empirical research in medicine, the social sciences and economics. But the chief difficulty is that study findings are often plagued with biases, such as confounding and measurement error. This research program will investigate a burgeoning area of data analytic techniques called Bayesian bias modelling. Bias modelling extends conventional parametric models to incorporate additional parameters which capture assumptions about bias. Because the observational study contains no information about the magnitude of bias, it means that the resulting model is nonidentifiable. To analyze the data, the investigator supplies user inputs in the form of prior distributions for bias parameters. Bayes theorem then gives a principled synthesis of the data with prior beliefs about bias. Thus Bayesian bias modelling techniques are better able to quantify total uncertainty in observational studies rather than the usual uncertainty due to random error.