Bias, Confounding, and Interaction: Lions and Tigers, and Bears, Oh My!
Epidemiologists seek to make a valid inference about the causal effect between an exposure and a disease in a specific population, using representative sample data from a specific population. Clinical researchers likewise seek to make a valid inference about the association between an intervention and outcome(s) in a specific population, based upon their randomly collected, representative sample data. Both do so by using the available data about the sample variable to make a valid estimate about its corresponding or underlying, but unknown population parameter. Random error in an experiment can be due to the natural, periodic fluctuation or variation in the accuracy or precision of virtually any data sampling technique or health measurement tool or scale. In a clinical research study, random error can be due to not only innate human variability but also purely chance. Systematic error in an experiment arises from an innate flaw in the data sampling technique or measurement instrument. In the clinical research setting, systematic error is more commonly referred to as systematic bias. The most commonly encountered types of bias in anesthesia, perioperative, critical care, and pain medicine research include recall bias, observational bias (Hawthorne effect), attrition bias, misclassification or informational bias, and selection bias. A confounding variable is a factor associated with both the exposure of interest and the outcome of interest. A confounding variable (confounding factor or confounder) is a variable that correlates (positively or negatively) with both the exposure and outcome. Confounding is typically not an issue in a randomized trial because the randomized groups are sufficiently balanced on all potential confounding variables, both observed and nonobserved. However, confounding can be a major problem with any observational (nonrandomized) study. Ignoring confounding in an observational study will often result in a “distorted” or incorrect estimate of the association or treatment effect. Interaction among variables, also known as effect modification, exists when the effect of 1 explanatory variable on the outcome depends on the particular level or value of another explanatory variable. Bias and confounding are common potential explanations for statistically significant associations between exposure and outcome when the true relationship is noncausal. Understanding interactions is vital to proper interpretation of treatment effects. These complex concepts should be consistently and appropriately considered whenever one is not only designing but also analyzing and interpreting data from a randomized trial or observational study.