Controlling for Confounding in Epidemiological Studies
Analytical tools used to:
Key Point
Stratification and multivariate analysis are essential for disentangling confounding from true exposure-outcome relationships
Stratification is informative because:
Case-control study examining:
Crude analysis: Odds Ratio = 1.71 (males at higher risk)
Results after stratification:
| Stratum | OR |
|---|---|
| Outdoor workers | 1.06 |
| Indoor workers | 1.00 |
Interpretation
Stratum-specific ORs are similar to each other (1.06, 1.00) but different from crude OR (1.71). Occupation is a confounder!
Homogeneous stratum-specific ORs indicate:
Example of interaction:
Tip
Consider biological plausibility and pre-established hypotheses
Crude OR = 1.7 (70% higher odds with OC use)
Age-stratified ORs:
| Age Group | OR |
|---|---|
| 25-29 years | 7.2 |
| 30-34 years | 8.9 |
| 35-39 years | 1.5 |
| 40-44 years | 3.7 |
| 45-49 years | 3.9 |
Adjusted OR = 3.97 (more than double crude estimate)
What is Negative Confounding?
When adjustment moves the estimate away from the null (1.0)
In the OC-MI example:
Procedure:
Tip
Answers: “What would the rate be if both groups had the same age distribution?”
Used for age adjustment of mortality/morbidity data
Procedure:
Results in:
Most common method for adjusted OR/RR
Formula for adjusted OR:
\[OR_{MH} = \frac{\sum_i (a_i d_i / N_i)}{\sum_i (b_i c_i / N_i)}\]
Where: - i = stratum - a, b, c, d = cells in 2×2 table - N = total in stratum
Calculation:
\[OR_{MH} = \frac{(53 \times 3)/81 + (35 \times 79)/219}{(10 \times 15)/81 + (52 \times 53)/219} = 1.01\]
Interpretation:
Key Assumption
Mantel-Haenszel assumes no multiplicative interaction between exposure and stratifying variable
When stratum-specific ORs are similar:
When ORs differ substantially:
Three main limitations:
One exposure at a time: Can only assess one exposure-outcome association per analysis
Categorical only: Continuous variables must be categorized (may cause residual confounding)
Sparse data: Too many strata → small numbers → unstable estimates
Advantages over stratification:
Common models:
| Model | Outcome Type | Interpretation |
|---|---|---|
| Linear | Continuous | Change in mean per unit |
| Logistic | Binary | Change in log odds |
| Cox | Time-to-event | Change in log hazard |
| Poisson | Count/Rate | Change in log rate |
Three conditions required:
Two-step analysis:
Tip
Controls for unmeasured confounding!
Definition: Predicted probability of exposure based on relevant characteristics
Purpose: Mimic randomization by making exposed/unexposed groups comparable
Steps:
Occurs when:
Result
Adjusted estimates still confounded despite adjustment attempt
Definition: Inappropriate adjustment that distorts true relationship
Occurs when adjusting for:
Use stratification when:
Use regression when:
Remember
Questions?
Further Reading:
Contact: Prof Kamarul Imran
Public Health Epidemiology