Methods to Control for Bias in Observational Studies

Thesis (Master's)--University of Washington, 2016-06

Observational studies often suffer from the problem of confounding, where observed results are biased due to the presence of factors that are strongly associated with both the exposure of interest and the outcome. Typical sources of confounding include factors such as age, sex, and medical comorbidities. The failure to account for confounding in the analytic framework can lead to biased results and ultimately an incorrect inference. Arguably the most common method of accounting for confounding is through the use of regression based approaches, although other methods such as propensity score matching are described. Beyond confounding, an additional source of bias that must be accounted for is the fact that observational data often is sampled from specified groups of individuals. For example, there may be clusters of individuals who are enrolled in the same health plan or are treated at the same hospital. The effect of this sampling framework is that patient outcomes from one health plan, hospital, etc are correlated. The correlation must be accounted for in the model to account in order to make a correct inference. Models that include multiple levels of analysis (such as patient and hospital) are call multilevel or hierarchical. As with the case of confounding discussed above, there are multiple well described methods to account for unmeasured factors that are contained at the cluster level. This thesis contains two observational studies that were completed by the author during her course of study in the Master’s in Public Health Program. Both studies have been accepted for publication by peer-reviewed journals and this this information is copyrighted. These studies will highlight two separate methods to account for confounding, as well as two approaches for hierarchical data analysis.