Effect Modification vs. Confounding ● Effect modification is a biological phenomenon in which the exposure has a different impact in different circumstances
● Confounding is a form of “bias” Confounding ● Obesity/mastitis example ○ Overall, obesity associated with mastitis (crude OR= 1.89) Obese Normal Mastitis + 50 30 Mastitis - 150 170
● Effect modification (interaction) ○ Exposure-disease association differs at different levels of the effect modifier ○ Example: case-control study ■ E = physical activity ■ ≥ 2,500 kcals/< 2,500 kcals ■ D = myocardial infarction ■ Potential effect modifier = alcohol consumption (yes/no) ●Is the association between physical activity and MI “modified” by background alcohol consumption?
Let’s think in line graphs ● How does “risk” change as the level of exposure to the confounder changes ● Predicted probability of MI decreases as physical activity increases in persons who do not consume alcohol. ● Predicted probability of MI increases as physical activity increases in persons who do consume alcohol.
● Contrast – confounding: age is a confounder older cows have a higher probability of mastitis, but the association between obesity and mastitis remains the same at different levels of age. Confounding & EM ● Parallel lines indicate confounding ○ Confounder is associated with disease ■ Old cows are more likely to have mastitis ■ Line for old cows is higher on the graph ○ Effect of exposure on disease is the same at different levels of the confounder ■ Stratum-specific ORs or RRs are equal ■ Results in parallel lines ● Assuming the association is linear… ● Parallel lines indicate confounding ○ Stratum-specific ORs/RRs = 1 ■ Lines are horizontal ○ Stratum-specific ORs/RRs > 1 ■ Lines have an upward slope ○ Stratum-specific ORs/RRs < 1 ■ Lines have a downward slope ● Non-parallel lines indicate EM ○ E – D association differs at different levels of the effect modifier ■ Physical activity & MI study ■ Negative association for alcohol = no ■ Downward sloping ■ Positive association for alcohol = yes ■ Upward sloping
Stratified analysis ● Take a “weighted average” of the stratum specific estimates, and we’ve done away with the confounding! ● The Mantel-Haenszel method provides a pooled odds ratio across the strata of fourfold tables Mantel-Haenszel Estimator ● Crude OR/RR falls outside range of stratum-specific ORs/RRs ○ Then we shouldn’t use that “crude” number in our study ● Mantel-Haenszel Estimator ○ Measure of the average effect of exposure across all strata (accounts for confounding and EM) ●
You've reached the end of your free preview.
Want to read all 15 pages?
- Fall '19
- Mastitis, Confounding, Case-control study