This chapter is under heavy development and may still undergo significant changes.
Once we have identified confounding, what do we do about it? A similar question that is often asked is, ”how do we control for confounding?” There are several things we can do. We will go further into the details of some of these as course moves on. You can think of this as a quick preview.
Also note that G-methods and instrumental variable methods are considered advanced topics and will not be taught in this course.
Study design considerations - Randomization - Matching – including propensity matching - Restriction
Statistical analysis tools - Stratification – including regression - G-methods - Inverse probability weighting - Standardization (G-formula) - G-estimation - Instrumental variable methods
Randomly assigning exposure to ensure equal distribution of confounders in each exposure category
Advantages/Disadvantages - Controls for known/unknown confounders - Data analysis straightforward - May require large sample sizes - Limited to longitudinal studies
Limiting inclusion criteria to individuals within a specified category of the confounder, making groups identical for that variable.
Advantages/Disadvantages - Complete control of known confounders - Limits sample size and generalizability - Residual confounding may occur - Can not evaluate association across levels of factor
Selecting subjects according to the value of suspected confounders to ensure equal distributions among study groups
Advantages/Disadvantages - Efficient, more power smaller sample size - Useful when insufficient number in subgroup - Costly and time-consuming, requiring extensive searching and record keeping to find matches - Factor can no longer be evaluated as a risk factor
Evaluating the association between exposure and disease within homogeneous categories (strata) of the confounding variable; also used to control for confounding
Advantages/Disadvantages - Clear way to present exposure, disease and factor relationship - Direct and logical strategy; computations easy to carry out - Minimal assumptions - Can only examine one factor at a time - Continuous variables must be categorized for analysis
Small number of confounders - Stratification - Mantel-Haenszel - Multivariable Modeling
Many potential confounders - Multivariable regression models