You have previously been introduced to the criteria for confounding. To recap, a confounding factor must be associated with the exposure in the source population, associated with the outcome in the absence of the exposure and must not be on the causal path between exposure and the outcome. These criteria must be confirmed whenever a characteristic is suspected of being a confounding factor. Now, I will take you through an example of checking for confounding in a study examining the association between alcohol consumption and oral cancer. On this slide, you can see a two by two table, which allows you to calculate the magnitude of the association between alcohol consumption and oral cancer. Based on this data, the calculated crude odds ratio is 2.26. That is, the risk of Oral Cancer is 2.26 times greater among those who consume alcohol, compared to those who abstain from alcohol. But is this a real association? Alcohol may increase the risk for oral cancer, but could something else, such as smoking, have an impact on the association? Hence could smoking be a confounder in this study? The first thing you consider is the causal definition or knowledge of the subject matter. We know from the literature that smoking is a risk factor for developing oral cancer. Smoking is an extraneous factor that is related to the disease among the unexposed group. That is, smoking has an effect on oral cancer incidence, among those who abstain from alcohol, in this case. Alcohol consumption and smoking are also known to be correlated in the general population. With smokers being more likely to drink and vice versa. Based on the above, we could decide that smoking may cause confounding in this case. Now we will use a different method to explore the statistical definition of confounding. Using data, we will apply the criteria for confounding. The first thing you need to check is if smoking is associated with alcohol consumption in the control group, or the source population. So those without oral cancer in this case. Here, you construct a two by two table which has no outcome and perform a chi-square test. So you can see that 90% of smoker drink, while only 27% of non-smokers drink. The p-value for the chi-square test is statistically significant and there seems to be an association between drinking and smoking. Hence, the first criterion for confounding is met. Next, we want to check the second criterion for confounding. Is smoking associated with oral cancer in the absence of alcohol consumption? Using the figures in this two-by-two table, you calculate the risk of oral cancer among those who do not drink. As you can see, there is a positive association between smoking and oral cancer, with an odds ratio of 3.5. You can also check for confounding by stratifying the results by smoking status. That is, remove the effect of smoking status to examine if the association between alcohol consumption and oral cancer for non-smokers and smokers, are similar in magnitude. And if they differ for the crude odds ratio. So here you can see that the odds ratio for the non-smoker group is 1.25 and 0.97 for smoker group. You should notice that the stratum specific odds ratio are quite similar, but very different than the crude odds ratio. Therefore, this could serve as an indication that there is confounding. Another way to check for confounding using the data, is with a simple logistic regression model. You can see that we have one regression with a crude odds ratio of 2.26. When we add the variable for smoking, the odds ratio changes to 1.14. The difference between the odds ratios here is greater than 15% so again, there is an indication of confounding. Now you have seen an example of checking for each of the criteria for confounding. However, you only need to use one of these methods to decide if a factor is a potential confounder. [MUSIC]