confounding variable

DEFINITION: A confounder is a variable that (1) is associated with the exposure, (2) is independently associated with the outcome (a risk factor or cause), and (3) is NOT an intermediate step in the causal pathway (not a mediator). Ignoring it makes the crude association differ from the true association.

CLASSIC EXAMPLES: - Coffee drinking and lung cancer: smoking is the confounder, because smokers drink more coffee and smoking causes lung cancer. - Alcohol and lung cancer: smoking is again the confounder. - Age is a common confounder in many studies, such as exercise and mortality. - Oral contraceptive use and MI: smoking can confound the relationship.

CONFOUNDING VS. OTHER CONCEPTS: - Confounding vs. effect modification (interaction): A confounder is a nuisance bias to be removed. Effect modification is a real biologic phenomenon in which the effect of the exposure differs across levels of a third variable (for example, the effect of a drug differs by sex). Effect modification is reported, not eliminated. Clue: if stratum-specific estimates differ from each other, think effect modification. If they are similar to each other but differ from the crude estimate, think confounding. - Confounding vs. selection bias or information bias: Confounding comes from the real-world distribution of a third variable. Selection and information (measurement) biases come from how subjects are chosen or how data are measured. - Confounding vs. mediator: A mediator lies on the causal path (exposure → mediator → outcome). Adjusting for a mediator removes part of the real effect. - Confounding is not chance: Larger sample size does not fix it, whereas it reduces random error.

HOW TO CONTROL CONFOUNDING: - Design stage: - Randomization is the best method. It balances both known and unknown confounders, especially with a large sample. - Restriction limits enrollment to one level of the confounder (for example, only non-smokers). - Matching pairs subjects by confounders. It is common in case-control studies (for example, age and sex). - Analysis stage: - Stratification (for example, Mantel-Haenszel) gives adjusted estimates. - Multivariable regression (logistic, linear, Cox) adjusts for multiple confounders. - Standardization (age-adjusted rates) and propensity scores are also used. - Limitations: Only measured confounders can be adjusted for in the analysis. Residual confounding persists if a confounder is unmeasured or measured imperfectly. This is why observational studies (cohort, case-control) cannot prove causation as well as RCTs.

CLINICAL/BOARD PEARLS: - Observational studies are most vulnerable to confounding. Well-conducted RCTs are least vulnerable. - Positive confounding exaggerates the association (away from the null). Negative confounding masks or reduces it (toward the null). - A typical vignette gives a crude association between exposure and disease, and then a stratified analysis shows the association disappears or changes. The answer is confounding. - Another common vignette: the stratified relative risks are different in each stratum (for example, RR 5 in men and RR 1 in women). The answer is effect modification. - Confounding by indication: Sicker patients are more likely to receive a treatment, so the drug appears harmful (for example, a drug looks associated with death because it is given to the sickest patients). - Biostatistics Step 1 question: "Which of the following best controls for confounding?" The answer is usually randomization (design) or multivariable regression/stratification (analysis of observational data).