Sensitivity Analysis and Bias Analysis

Methodological shortcomings of studies can lead to bias, in the sense of systematic (non-random) distortion of estimates from the studies. Among well-recognized bias sources are non-random selections of study subjects, failure to measure or adjust for confounding variables, adjustments for non-confounding variables, and measurement errors in or misclassification of variables in the analysis. Sensitivity and bias analyses model these shortcomings of the design and execution of the study as features of the data-generating mechanism. These features include causes of selection, treatment assignment, censoring, missing data, and measurement errors. The modeling process can provide useful extensions of conventional methods for experimental and survey statistics to imperfect experiments and observational studies. This chapter provides some of the basic concepts and methods for this process.