MED106 Research Methods in Medicine & Essential Medical Statistics PDF

MED106 Research Methods in Medicine & Essential Medical Statistics Introduction to Confounding II: Dealing with confounding and identifying 'independent' risk factors Elena Critselis, MPH PhD Associate Professor in Epidemiology and Public Health Department of Primary Care and Population Health Unive...

MED106 Research Methods in Medicine & Essential Medical Statistics Introduction to Confounding II: Dealing with confounding and identifying 'independent' risk factors Elena Critselis, MPH PhD Associate Professor in Epidemiology and Public Health Department of Primary Care and Population Health University of Nicosia Medical School Learning Objectives Differentiate between crude and adjusted estimates Describe the concept of the 'independent' risk factor Differentiate between a confounder and a mediator Describe the concept of residual confounding and overadjustment Dealing with confounding Exposure-Outcome association Exposure Outcome Confounding Confounding: a third factor (confounder) explains all or part of the association between an exposure and an outcome. A potential confounder needs to fulfil the following criteria: 1. Has to be associated with the outcome of interest 2. Has to be associated with the exposure of interest 3. Should not lie on the causal pathway between exposure and outcome. Confounding Confounding: a third factor (confounder) explains all or part of the association between an exposure and an outcome. Smoking Alcohol consumption Lung cancer Dealing with confounding In order to deal with confounding we need to follow a 3-step procedure: 1. Identify potential confounders (see previous session) 2. Adjust for potential confounders (this session) 3. Compare crude and adjusted estimates (this session) The process of statistical adjustment Statistical adjustment for confounding The process of statistical adjustment aims at reducing (at least) or eliminating (at best) the confounding effect of potential confounders in any exposureoutcome association. In other words, adjustment aims to remove any effects of the potential confounder on the outcome, thus providing a more ‘clean’ estimate of the exposure-outcome association. After adjusting (also called ‘controlling’) our estimates for a potential confounder (i.e. age, smoking status, etc.), these estimates are then said to be adjusted for that confounder (i.e. age-adjusted estimates, smoking-adjusted, etc.). Statistical adjustment for confounding Adjustment is a statistically cumbersome technique, whose details are beyond the scope of our course. Briefly, what statistical adjustment is doing is keeping the potential confounder constant (i.e. making it a non- variable) and re-calculating the estimates for the exposure-outcome association (i.e. Odds Ratios, mean difference, regression coefficient, etc.). In this way, we get more accurate (adjusted) estimates for exposure-outcome associations. Test your understanding! In order to deal with confounding in any exposure-outcome association, we need to follow a 3-step procedure. Please put the steps of this procedure in the right order. A. Compare crude and adjusted estimates B. Identify confounders C. Adjust for potential confounders Test your understanding! In order to deal with confounding in any exposure-outcome association, we need to follow a 3-step procedure. Please put the steps of this procedure in the right order. SOLUTION: A. Identify confounders B. Adjust for potential confounders C. Compare crude and adjusted estimates Statistical adjustment for confounding One way of adjusting for a given confounder is to stratify the analysis based on the categories of the confounder. This process of stratification involves performing the analysis for the exposureoutcome association separately in the categories of the confounder (i.e. for smokers and non-smokers if smoking is the confounder). This will therefore give us 2 different (stratified) estimates (assuming that the confounder had 2 categories) for the exposure-outcome association, one for each category of the confounder variable. Statistical adjustment for confounding These stratified estimates will then need to be combined, in order to get a combined (adjusted) estimate. This adjusted estimate is ‘clear’ of any confounding effect from the specific confounder, since the analysis was performed separately in its categories and the results were then combined. Note that if the exposure-outcome association is substantially different in the categories of the confounder, then the estimates cannot be combined and we present them separately (this is known as effect modification). Comparing crude and adjusted estimates Crude and adjusted estimates for confounding Crude estimates are simply estimates for exposure-outcome associations (Odds Ratio, mean difference, etc.) before applying any adjustment. Adjusted estimates are estimates for exposure-outcome associations after applying statistical adjustment for any potential confounder. Adjusted estimates aim to answer the following question: ‘What would the exposure-outcome association be if everyone was the same (i.e. had the same value) in terms of the potential confounder variable?’ Crude and adjusted estimates for confounding Examples: If age was the potential confounder in an association between smoking and dementia, the age-adjusted estimate would give us the measure of association between smoking and dementia (e.g. Odds Ratio) if everyone had the same age! If smoking was the potential confounder in an association between alcohol consumption and lung cancer, the smoking-adjusted estimate would give us the measure of association between alcohol consumption and lung cancer (e.g. Odds Ratio) if everyone had the same smoking status! Dealing with confounding (Example) Lung cancer Alcohol consumption Yes No Heavy 535 2721 No or low 1182 15562 Crude (unadjusted) Odds Ratio for lung cancer comparing high consumers to no/low consumers: 2.50 (95% CI: 2.30 - 2.90) Dealing with confounding (Example) Dealing with confounding by stratification Smokers Alcohol consumption Lung cancer Non-smokers Alcohol consumption Lung cancer Dealing with confounding (Example) Non-smokers only (59%) Smokers only (41%) Lung cancer Lung cancer Alcohol consumption Heavy No or low Yes 520 980 No 1980 4720 Odds Ratio IN SMOKERS ONLY = 1.26 Alcohol consumption Yes No Heavy 15 741 No or low 202 10842 Odds Ratio IN NON-SMOKERS ONLY = 1.09 Dealing with confounding (Example) Can the two estimates be combined? Smokers only Odds Ratio = 1.26 Non-smokers only Odds Ratio = 1.09 Yes! The two estimates are similar and thus can be combined! Dealing with confounding (Example) The two estimates are combined in a weighted manner, based on the sample size of each category of the confounder variable (i.e. smokers / non-smokers in this case) In current study: smokers 41%, non-smokers 59% Smokers: Odds Ratio: 1.26 x 0.41 Non-smokers: Odds Ratio: 1.09 x 0.59 Combined (adjusted) estimate: (1.26 Χ 0.41) + (1.09 Χ 0.59) = 1.16 Note: You are not expected to manually perform adjustment as displayed in this example. This is only included for getting an idea about the concept of adjustment! Dealing with confounding (Example) Crude Odds Ratio for the association between alcohol consumption and lung cancer = 2.50 Adjusted (for smoking) Odds Ratio for the association between alcohol consumption and lung cancer = 1.16 Smoking is a confounder in the association between alcohol consumption and lung cancer ! Note: You are expected to be able to judge whether a factor is a confounder, by comparing crude and adjusted estimates (as above)! Judging the effect of adjustment on confounding Judging the effect of adjustment Adjustment can have different effects on the estimate of any given exposure-outcome association: 1. The association becomes weaker 2. The association disappears 3. The association appears 4. The association becomes stronger 5. The association is reversed 6. The association remains unaffected Crude and adjusted estimates Example 1a: The association becomes weaker A study found a crude Odds Ratio for the association between alcohol consumption (heavy vs. moderate) and stroke of 3.00 (95%CI: 2.20 - 3.80). After adjustment for age, the age-adjusted Odds Ratio for the association between alcohol consumption and stroke was 2.45 (95%CI: 1.65 - 3.25). Interpretation: The potential confounder explains a small part of the association between exposure and outcome. Crude and adjusted estimates Example 1b: The association becomes weaker A study found a crude Odds Ratio for the association between alcohol consumption (heavy vs. moderate) and stroke of 3.00 (95%CI: 2.20 - 3.80). After adjustment for dietary patterns, the diet-adjusted Odds Ratio for the association between alcohol consumption and stroke was 1.55 (95%CI: 1.05 - 2.05). Interpretation: The potential confounder explains a large part of the association between exposure and outcome. Crude and adjusted estimates Example 2: The association disappears A study found a crude Odds Ratio for the association between tea consumption (consumers vs. non-consumers) and type 2 diabetes of 1.30 (95%CI: 1.05 - 1.55). After adjustment for physical activity, the activity-adjusted Odds Ratio for the association between tea consumption and type 2 diabetes was 1.03 (95%CI: 0.77 - 1.28). Interpretation: The potential confounder explains all of the association between exposure and outcome (i.e. after adjusting for the confounder, there is no association between exposure and outcome!). Crude and adjusted estimates Example 3: The association appears A study found a crude Odds Ratio for the association between educational attainment (high vs. low) and breast cancer of 1.02 (95%CI: 0.87 - 1.17). After adjustment for age, the age-adjusted Odds Ratio for the association between educational attainment and breast was 1.28 (95%CI: 1.13 - 1.43). Interpretation: The potential confounder was masking the association and adjusting for it made the association appear. Crude and adjusted estimates Example 4: The association becomes stronger A study found a crude Odds Ratio for the association between smoking (smokers vs. non-smokers) and Alzheimer’s Disease of 1.65 (95%CI: 1.55-1.75). After adjustment for gender, the gender-adjusted Odds Ratio for the association between smoking and Alzheimer’s Disease was 2.00 (95%CI: 1.05 - 2.65). Interpretation: The potential confounder was diluting the association and adjusting for it made the association stronger. Crude and adjusted estimates Example 5: The association is reversed A study found a crude Odds Ratio for the association between hypercholesterolemia and dementia of 0.85 (95%CI: 0.75 - 0.95). After adjustment for body weight status, the weight- adjusted Odds Ratio for the association between hypercholesterolemia and dementia was 1.45 (95%CI: 1.35 1.55). Interpretation: The potential confounder was reversing the association between exposure and outcome (i.e. after adjusting for the confounder, the association from negative became positive!). Crude and adjusted estimates Example 6: The association remains unaffected A study found a crude Odds Ratio for the association between birth weight (high vs normal) and hyperactivity in childhood of 1.15 (95% CI: 1.04; 1.26). After adjustment for parental social class, the social class- adjusted Odds Ratio for the association between birth weight and hyperactivity was 1.13 (95% CI: 1.02; 1.53). Interpretation: The potential confounder had no effect on the exposure-outcome association (i.e. this was not actually a confounder in the specific association!). Deciding on whether a factor is a confounder Note 1 (previous session): If a factor is associated with the outcome but not with the exposure (or vice versa), then this factor cannot be a confounder in the specific association! Note 2 (previous session): If a factor is associated with both the outcome and the exposure, but lies in the causal pathway between the two, then this factor is not called a confounder but a mediator! Note 3: Adjusting for mediators may have exactly the same effects on an exposure-outcome association as adjusting for a confounder (i.e. association becomes weaker, stronger, etc.). The only difference is in the interpretation! Test your understanding! In a study aiming to determine the association between social isolation and depression, age was treated as a potential confounder and was adjusted for. The Crude Relative Risk for this association was 1.80 (comparing depressed to non-depressed individuals), while the ageadjusted estimate was 1.20. What would you conclude about the role of age in the social isolation – depression association? A. B. C. D. E. Age is a confounder but does not explain any of the association Age is a confounder which explains all of the association Age is a confounder which explains part of the association Age is not a confounder since it is not associated with the exposure Age is not a confounder since it is not associated with the outcomes Test your understanding! In a study aiming to determine the association between social isolation and depression, age was treated as a potential confounder and was adjusted for. The Crude Relative Risk for this association was 1.80 (comparing depressed to non-depressed individuals), while the ageadjusted estimate was 1.20. What would you conclude about the role of age in the social isolation – depression association? A. B. C. D. E. SOLUTION: Age is a confounder but does not explain any of the association Age is a confounder which explains all of the association Age is a confounder which explains part of the association Age is not a confounder since it is not associated with the exposure Age is not a confounder since it is not associated with the outcomes Overadjustment and residual confounding If a factor which lies in the causal pathway between exposure and outcome (i.e. a mediator) is treated like a confounder and adjusted for, this may lead to an underestimation of the true exposure-outcome association, a phenomenon called overadjustment. On the other hand, if the researchers adjusted for some but not all potential confounders, or in case the confounders adjusted for where not accurately measured, we may have an overestimation of the true exposure- outcome association, a phenomenon called residual confounding. Further reading Petrie A. & Sabin C. Medical Statistics at a Glance, 3rd Edition, Chapter 34 [ISBN : 978-1-4051-8051-1] Buring EJ. Epidemiology in Medicine, Chapters 3, 9 [ISBN : 9780316356367] http://www.ncbi.nlm.nih.gov/pubmed/11274518 http://www.healthknowledge.org.uk/node/803 For additional information: [email protected]