A researcher conducted a case-control study to investigate the association between oral contraceptive use and myocardial infarction (MI) in Indian women aged 30–50 years. Cases were recruited from hospital cardiac units and controls from the same hospitals' outpatient departments. The study found a strong association between OCP use and MI (OR = 4.2, 95% CI 2.8–6.1). However, when the analysis was stratified by smoking status, the association weakened considerably in both smokers and non-smokers (OR ≈ 1.5 in each stratum). Which type of bias or confounding best explains this finding?

Explanation

## Identification of Simpson's Paradox and Confounding **Key Point:** The dramatic weakening of the association when stratified by smoking status is a classic example of confounding, not bias. The crude OR (4.2) is inflated because smoking is a confounder. ### Why Smoking is a Confounder Smoking satisfies all three criteria for confounding: 1. **Associated with the exposure:** Smokers may be less likely to use OCPs (or vice versa, depending on the population) 2. **Associated with the outcome:** Smoking is an independent risk factor for MI 3. **Not on the causal pathway:** Smoking does not cause OCP use; it is an independent risk factor ### Stratification Reveals the True Effect When the analysis is stratified by smoking status, the confounding is removed. The stratum-specific ORs (≈1.5) are much smaller and more consistent, representing the true association between OCPs and MI after controlling for smoking. This is **Simpson's Paradox** — the direction or magnitude of an association can reverse or diminish when a third variable (confounder) is controlled. **High-Yield:** Confounding is a validity threat that can be detected and controlled through stratification or multivariate analysis. The crude association is biased away from the null due to the confounding effect of smoking. ### Distinction from Bias | Feature | Confounding | Bias | |---------|-------------|------| | **Nature** | Mixing of effects; a third variable distorts the true association | Systematic error in measurement or selection | | **Control** | Stratification, matching, multivariate adjustment | Study design (blinding, standardized protocols) | | **Detection** | Stratum-specific estimates differ from crude estimate | Comparison with gold standard or repeat studies | | **Direction** | Can bias toward or away from null | Usually systematic in one direction | **Clinical Pearl:** In observational studies, confounding is ubiquitous. Always consider whether a strong association might be explained by an unmeasured or inadequately controlled confounder.