A researcher is investigating the association between oral contraceptive use and venous thromboembolism (VTE) in Indian women. The study finds a strong positive association. However, upon stratification by smoking status, the association disappears in both smokers and non-smokers, though it was present in the overall sample. What is the most appropriate investigation to formally quantify and confirm this phenomenon?

Explanation

## Identifying and Quantifying Confounding in Epidemiological Studies **Key Point:** When an association observed in crude (unstratified) data **disappears** upon stratification by a third variable in **all strata**, that third variable is a **confounder** — not an effect modifier. The **Mantel-Haenszel (MH) test** is the gold-standard statistical investigation to formally quantify and confirm confounding in stratified 2×2 tables. ### Why Mantel-Haenszel Test? The MH test: 1. Calculates a **pooled (weighted) odds ratio** across strata (here: smokers and non-smokers separately) 2. Compares the **crude OR** with the **MH-adjusted OR** — if they differ by >10%, confounding is confirmed 3. Provides a formal **p-value** for the pooled association after controlling for the confounder 4. Is specifically designed for **categorical exposure-outcome data** in stratified 2×2 tables — exactly the scenario described **High-Yield (Park's Textbook of PSM):** The Mantel-Haenszel method is the standard technique for controlling confounding in stratified analysis. A crude OR that collapses to a null MH-adjusted OR confirms that the third variable (smoking) fully explains the apparent association. ### Mechanism of Confounding Here | Feature | Explanation | |---------|-------------| | **Crude association** | OCP use → VTE (appears positive overall) | | **Confounder (Smoking)** | Associated with both OCP use AND VTE risk independently | | **After stratification** | Association disappears in BOTH smokers and non-smokers | | **Conclusion** | Smoking is a **confounder** — it created a spurious association | **Clinical Pearl:** Smoking increases VTE risk independently AND is more prevalent among OCP users in the study population, creating a spurious crude association. The MH test reveals this by showing the pooled stratum-specific OR is null, while the crude OR was elevated — classic confounding. ### Why Not Other Options? - **Chi-square test (B):** Tests independence in a single 2×2 table but does **not** account for stratification or adjust for a third variable — cannot detect or quantify confounding. - **Pearson correlation (C):** Appropriate only for **continuous** variables; not applicable to categorical exposure-outcome pairs like OCP use and VTE. - **Logistic regression with interaction term (D):** An interaction term tests for **effect modification** (whether the OR differs across strata). Here, the association disappears *uniformly* in both strata — this is confounding, not interaction. Logistic regression with an interaction term would be appropriate if the OR *differed* between smokers and non-smokers (i.e., effect modification). For formally quantifying confounding via stratified analysis, MH is the designated tool. **Mnemonic:** **MH = Confounder Hunter** — Mantel-Haenszel detects when a third variable explains away an association across all strata. *Reference: Park's Textbook of Preventive and Social Medicine, 26th ed.; Kelsey JL et al., Methods in Observational Epidemiology.*