Transcript Example

Confounding
Tim Wiemken PhD MPH CIC
Assistant Professor
Division of Infectious Diseases
University of Louisville, Kentucky
Overview
1. Define and Identify Confounding
2. Calculate Risk Ratio and Stratified
Risk Ratio
3. Identify How to Select Confounding
Variables for Multivariate Analysis
Overview
1. Define and Identify Confounding
2. Calculate Risk Ratio and Stratified
Risk Ratio
3. Identify How to Select Confounding
Variables for Multivariate Analysis
Confounding
Definition:
A variable related to the exposure
(predictor) and outcome but not in the
causal pathway
Confounding
Confounding
Why does this happen?
Risk factor that has different prevalence in
two study populations…
e.g. Coffee drinking and lung cancer
Example
Men vs Women Example….
25% Risk of
lung cancer
5% Risk of
Lung Cancer
Example
Men vs Women Example….
25% Risk of
lung cancer
5% Risk of
Lung Cancer
Conclusion: People who drink coffee die more
therefore coffee causes lung cancer
Example
Men vs Women Example….
25% Risk of
lung cancer
5% Risk of
Lung Cancer
Truth: Coffee drinkers are more likely to smoke.
Smoking is associated with a higher risk of lung cancer.
mortality.
Example
Predictor:
Coffee
Outcome:
Lung cancer
Confounder:
Smoking
Example
Predictor:
Coffee
Outcome:
Lung cancer
Confounder:
Smoking
Smoking associated with coffee drinking and lung
cancer. Smoking is not caused by drinking coffee.
Overview
1. Define and Identify Confounding
2. Calculate Risk Ratio and Stratified
Risk Ratio
3. Identify How to Select Confounding
Variables for Multivariate Analysis
Example
Question: Are coffee drinkers more likely to get
lung cancer?
Warning: The upcoming data are made up. Do not
make any decisions based on the outcomes of our
example!
Example Flowchart
178 cancer+
1307 coffee+
1129 cancer2648
Enrolled
79 cancer+
1341 coffee-
3154 subjects
1262 cancer506 Excluded
Example
What Type of Study is That?
Example
What Type of Study is That?
What is the correct measure of association?
Example
What Type of Study is That?
What is the correct measure of association?
OK. Now Calculate the Correct Measure of
Association
Example
Do coffee drinkers get lung cancer
more than non coffee drinkers?
Data
Cancer+
Coffee+
Coffee-
Cancer-
Example Flowchart
178 cancer+
1307 coffee+
1129 cancer2648
Enrolled
79 cancer+
1341 coffee-
3154 Subjects
1262 cancer506 Excluded
Example
Do coffee drinkers get lung cancer more
than non coffee drinkers?
Data
Cancer+
Cancer-
Coffee+
178
1129
Coffee-
79
1262
Example
Do coffee drinkers get lung cancer more
than non coffee drinkers?
Well?
Example
Do coffee drinkers get lung cancer more
than non coffee drinkers?
Yes!
RR: 2.31,
P=<0.001,
95% CI: 1.79 – 2.98
Example
Is this a true relationship or is another variable
confounding that relationship?
Example
Is this a true relationship or is another variable
confounding that relationship?
We noticed a lot of coffee drinkers also smoke,
much more than those patients who didn’t drink
coffee.
Could this be a confounder?
Example: Step 1
Input your data in the 2x2
Cancer+
Cancer-
Coffee+
178
1129
Coffee-
79
1262
This gives you a ‘crude’ odds or risk ratio
Example: Step 2
Stratify on the potential confounder
Stratified data:
Stratified data:
Smoker+
Coffee+/ Cancer+: 168
Coffee -/Cancer+: 34
Coffee+/Cancer-: 880
Coffee-/Cancer-: 177
SmokerCoffee+/ Cancer+: 10
Coffee -/Cancer+: 45
Coffee+/Cancer-: 249
Coffee-/Cancer-: 1085
Example: Step 2
Compute Risk Ratios for Both,
Separately
Smoker+
Cancer+
Cancer-
Cancer+
Cancer-
Coffee+
Coffee-
SmokerCoffee+
Coffee-
Example: Step 2
Calculate the adjusted measure of
association
Stratified data:
Stratified data:
Smoker+
Coffee+/ Cancer+: 168
Coffee -/Cancer+: 34
Coffee+/Cancer-: 880
Coffee-/Cancer-: 177
SmokerCoffee+/ Cancer+: 10
Coffee -/Cancer+: 45
Coffee+/Cancer-: 249
Coffee-/Cancer-: 1085
Example: Step 2
2. Compute Risk Ratios for Both,
Separately
Smoker+
Cancer+
Cancer-
Coffee+
168
880
Coffee-
34
177
Cancer+
Cancer-
Coffee+
10
249
Coffee-
45
1085
Smoker-
Example
What do you see?
Example: Step 3
Ensure that, in the group without the outcome,
the potential confounder is associated with
the predictor
Example: Step 4
Compute the adjusted odds/risk ratios
Compute the percent difference between
the ‘crude’ and adjusted ratios.
Adjusted Ratio Must be >10%
Different than the Crude Ratio
Example
If the criteria are met, you have a
confounder
Issues with Confounding
As in our example, a confounder can
create an apparent association between
the predictor and outcome.
Issues with Confounding
As in our example, a confounder can
create an apparent association between
the predictor and outcome.
A confounder can also mask an
association, so it does not look like there
is an association originally, but when you
stratify, you see there is one.
Overview
1. Define and Identify Confounding
2. Calculate Risk Ratio and Stratified
Risk Ratio
3. Identify How to Select Confounding
Variables for Multivariate Analysis
Multiple Confounding Variables
Regression methods adjust for
multiple confounding variables at once
– less time consuming.
Logistic Regression
Linear Regression
Cox Proportional Hazards Regression
… and many others
Multiple Confounding Variables
1: The way we just did it.
This is probably the most reliable
method with a few more steps.
Multiple Confounding Variables
2. Include all clinically significant
variables or those that are previously
identified as confounders.
Issues:
• May have too many confounders
• Confounding in other studies does
NOT mean it is a confounder in yours.
Multiple Confounding Variables
3: If that variable is significantly
associated with the outcome (chisquared) then include it.
Sun, G. W., Shook, T. L., & Kay, G. L. (1996). Inappropriate use of bivariable analysis to screen risk factors for
use in multivariable analysis. J Clin Epidemiol, 49(8), 907-916.
Multiple Confounding Variables
3: If that variable is significantly
associated with the outcome (chisquared) then include it.
Many issues with this method.
What is significant?
Sun, G. W., Shook, T. L., & Kay, G. L. (1996). Inappropriate use of bivariable analysis to screen risk factors for
use in multivariable analysis. J Clin Epidemiol, 49(8), 907-916.
Multiple Confounding Variables
3: If that variable is significantly
associated with the outcome (chisquared) then include it.
Many issues with this method.
Just because the ‘confounder’ is associated with the predictor
doesn’t mean it is associated with the outcome and not in the
causal pathway!
Sun, G. W., Shook, T. L., & Kay, G. L. (1996). Inappropriate use of bivariable analysis to screen risk factors for
use in multivariable analysis. J Clin Epidemiol, 49(8), 907-916.
Multiple Confounding Variables
4. Automatic Selection Regression
Methods
Many ways to do this, and relatively
reliable with certain methods.
• Forward Selection
• Backward Selection
• Stepwise
Multiple Confounding Variables
Caveats
Need to control for as few confounding
variables as possible.
Multiple Confounding Variables
Caveats
Need to control for as few confounding
variables as possible.
You are limited by the number of cases of
the outcome you have (10:1 Rule)
Multiple Confounding Variables
Caveats
Need to control for as few confounding
variables as possible.
You are limited by the number of cases of
the outcome you have (10:1 Rule)
Some journals just want it done a certain
way.
Multiple Confounding Variables
Overview
1. Define and Identify Confounding
2. Calculate Risk Ratio and Stratified
Risk Ratio
3. Identify How to Select Confounding
Variables for Multivariate Analysis