Week 10 - Confounding FULL.pptx PDF

Confounding Week 10 Instructor: Dr. Kelly Anderson Learning Objectives: To define confounding in epidemiologic studies To discuss approaches to understanding and assessing confounding Identify ways of controlling confounding in both the design and analysis...

Confounding Week 10 Instructor: Dr. Kelly Anderson Learning Objectives: To define confounding in epidemiologic studies To discuss approaches to understanding and assessing confounding Identify ways of controlling confounding in both the design and analysis stages  How Do Non-Causal Associations Arise? Error Random Precisio Error/Chance Systematic Error Validity n Confounding Selection Bias Bias Information Bias Understanding confounding is important!!! Confounding A variable that distorts the estimate of the effect of an exposure on the outcome due to Confounde its association with exposure and r its influence on the occurrence of the outcome Outcom Exposure e Causality is the central concern of epidemiology, and confounding is the central concern with establishing causality Source: Google Dictionary Confounding as a “Mixing of Effects” Birth Order and the Risk of Down’s Syndrome “Confounding is confusion, or mixing, of effects; the effect of the exposure is mixed together with the effect of another variable, leading to bias” Latin: “confundere” is to mix together Rothman (2014) Epidemiology. An Introductio Mixing of Effects: Water Pipes Analogy Confounder Exposure and disease share a common cause (‘parent’) Exposure Outcome Mixing of effects – cannot separate the effect of exposure from that of confounder Jewell NP (2003) Statistics for Epidemiolo Mixing of Effects: Water Pipes Analogy Confounder If the common cause (‘parent’) is blocked, then the exposure – disease association becomes clearer Exposure Outcome Successful “control” of confounding (adjustment) Jewell NP (2003) Statistics for Epidemiolo Approaches to Understanding Confounding Counterfactual and non-comparability approaches Classical approach based on a priori criteria Collapsibility and data-based criteria Counterfactual Approach Counterfactual Comparison Outcome (CIexp)are identical in “Initial conditions” Exposed cohort the exposed and unexposed groups – because they are the same population! Outcome (CIunexp) Counterfactual, unexposed cohort RRcausal = CIexp / CIunexp “A causal contrast compares disease frequency under two exposure distributions, but in one target population during one etiologic time period” Maldonado & Greenland (2002) Int J Epi. Slide courtesy of M. Pai (Mc  In Reality… Outcome Exposed cohort (CI exp) counterfactual state is not observed (latent) Outcome (CIunexp) Counterfactual, unexposed cohort Outcome (CIsub) Substitute, unexposed cohort A substitute will usually be a population other than the target population during the etiologic time period - INITIAL CONDITIONS Slide courtesy of M. Pai (McGill MAY BE DIFFERENT In Reality… RRcausal = CIexp / CIunexpIDEAL RRassoc = CIexp / CIsubstituteACTUAL Chances are… RRcausal =/= RRassoc Slide courtesy of M. Pai (McGill Counterfactual Definition of Confounding “Confounding is present if the substitute population imperfectly represents what the target would have been like under the counterfactual condition” “An association measure is confounded (or biased due to confounding) for a causal contrast if it does not equal that causal contrast because of such an RRcausal imperfect substitution” =/= RRassoc Maldonado & Greenland (2002) Int J Epi;31:422-2 Recall: Assessing Causality We can’t directly observe counterfactuals We can make assumptions about causal effects if several conditions are met:  Consistency  Exchangeability Confounding  Positivity Exchangeability Exposed cohort Substitute, unexposed cohort We are not able to observe the counterfactual condition We try and get as close to the counterfactual condition as possible by ensuring our exposed and unexposed groups are exchangeable, that is as similar as possible in every way exceptWebb, theBain & Page (2017) Essential Positivity It must be possible for all people to get all levels of the exposure Exposed and unexposed participants must be present at every combination of the values of the observed confounder in the population under study Example:  If ONLY old people get aspirin, and NO young people get aspirin, then we can’t separate the effects of age and aspirin on the risk of heart attack, or we do not have positivity Porta (2016) Dictionary of Epidemiology Small Group Discussion… Are these groups comparable? Why or why not? In studying the effect of prenatal lead exposure on child IQ, researchers compare children born to women who live in a town dominated by a large battery plant with children born to women who live in a seaside resort town. Small Group Discussion… Are these groups comparable? Why or why not? In a study of exercise and bone density, a group of joggers is compared to a group of sedentary people. Small Group Discussion… Are these groups comparable? Why or why not? Researchers studying whether cognitive stimulation protects against dementia randomly assign nursing home residents who are not cognitively impaired to either a daily hour-long Scrabble session or a daily hour of group time socializing in the lounge. Classical Approach “Classical” Approach to Confounding A factor is a confounder if 3 a priori criteria are met: i. The confounder must be associated with the exposure ii. The confounder must be a causal risk factor (or a surrogate measure of a cause) for the outcome iii. The confounder must not be an intermediate step in the causal pathway between the exposure and the outcome Confounding Schematic Confounder Exposure Outcome Intermediate Cause Exposure Confounder Outcome Confounding Schematic: Example Confounding Schematic: Example Maternal Age Down’s Birth Syndrom Order e Are Confounding Criteria Met? SES as a confounder of the association between HRT and heart disease Socioecono mic Status Heart HRT Use Disease Are Confounding Criteria Met? Oral contraceptives use as a confounder of the association between a genetic factor and risk of breast cancer? Oral Contracepti X ves BRCA1 Breast Gene Cancer Are Confounding Criteria Met? HPV virus seropositivity as a confounder between high risk sexual behaviours and cervical cancer HPV High Risk Sexual Cervical Behaviour Cancer s Underlying Causal Mechanism? HPV virus seropositivity as a confounder between high risk sexual behaviours and cervical cancer High Risk Sexual Cervical HPV Behaviour Cancer s Are Confounding Criteria Met? Eye colour as a confounder in the association between fair skin tone and melanoma (skin cancer) Eye Colour X Melanom Fair Skin a Are Confounding Criteria Met? Hypertension as a confounder in the association between obesity and mortality Hypertensio n Obesity Mortality Underlying Causal Mechanism? Hypertension as a confounder in the association between obesity and mortality Hypertensio Obesity Mortality n  Mediator: Direct versus Indirect Effects A variable that occurs on a causal pathway from exposure to outcome. It causes variation in the Hypertension as a confounder in the association between obesity outcome andand variable itself is caused to vary by mortality the exposure. Indirect ~ Porta (2014) Effect Hypertensio Obesity Mortality n Direct Effect Direct effect is portion of the total effect that does not act via an intermediate ca Collapsibility Approach Collapsibility Approach to Confounding Collapsibility is equality of stratum-specific measures of effect with the crude (collapsed), un-stratified measure According to this definition, a factor is a confounding variable if a. The effect measure is homogeneous across the strata defined by the confounder and b. The crude and common stratum-specific (adjusted) effect measures are unequal (this is called “lack of collapsibility”) ORs and IRRs do not have the property of collapsibility, but ORs can be used to assess for confounding if the rare disease assumption holds (and therefore approximates the RR) Collapsibility Approach to Confounding Crude Estimate: does not take into account the effect of the confounding variable Adjusted Estimate: accounts for the confounding variable (generated using Mantel-Haenszel estimator or multivariate analysis) Confounding is likely when: RRcrude RRadjusted PRcrude PRadjusted *ORcrude ORadjusted *Under rare disease assumption Collapsibility Approach to Confounding Association between Depressive Episodes and Mortality: Patten et al (2019) J Affective Disorders:242;165-71 Collapsibility Approach to Confounding Association between HRT and Coronary Heart Disease: ollapsibility Approach to Confounding Crude 2 x 2 table RRCrude Calculate Crude RR Stratify by Confounder Stratum 1 Stratum 2 Calculate RR’s for each stratum RR1 RR2 If stratum-specific RR’s are similar, calculate adjusted RR If Crude RR Adjusted RR, If Crude RR = Adjusted confounding is likely RR, confounding is *Can also use PR, or OR with rare outcome assumption unlikely Slide courtesy of M. Pai (McGill Collapsibility Approach to Confounding: Example Cigarette smoking and Depression Rate of depression higher among cigarette smokers than among non-smokers Hypothesized that smoking can impact neurotransmitters in the brain that impact negative mood and emotion How could sex be a potential source of non-comparability in this association? – Men are more likely than women to be smokers – Men are less likely to experience depression compared with women Keyes & Galea (2014) Epidemiology Collapsibility Approach to Confounding: Example STRATIFY Sex Depress Smoking ion Does sex meet the a priori criteria for confounding? Keyes & Galea (2014) Epidemiology Collapsibility Approach to Confounding: Example Population of interest is adults in general population Sample of 80 individuals with no history of depression Assess smoking status at baseline Follow over 5 years to see how many develop depression Assume no people were lost to follow-up Keyes & Galea (2014) Epidemiology Collapsibility Approach to Confounding: Example Male smoker Female smoker Male non-smoker Female non-smoker Keyes & Galea (2014) Epidemiology Collapsibility Approach to Confounding: Example RRcrude =0.381/0.368 = 1.04 Keyes & Galea (2014) Epidemiology Collapsibility Approach to Confounding: Example Smoking and Sex: 73% of men are smokers 38.3% of women are smokers Men are more likely than women to be smokers Depression and Sex: 15% of men are depressed 53.2% of women are depressed Men are less likely to have depression Keyes & Galea (2014) Epidemiology Collapsibility Approach to Confounding: Example RRmen =0.167/0.111 = 1.50 Among men, those who smoke have 1.50 times the risk of depression compared to those who do not smoke, over 5 years. Keyes & Galea (2014) Epidemiology Collapsibility Approach to Confounding: Example RRwomen =0.667/0.448 = 1.49 Among women, those who smoke have 1.49 times the risk of depression compared to those who do not smoke, over 5 years. Keyes & Galea (2014) Epidemiology Collapsibility Approach to Confounding: Example Smoking was not associated with depression in the crude analysis (RRcrude = 1.04) After stratifying by sex, smoking is associated with the development of depression (RRadjusted = 1.49) RRcrude RRadjusted Therefore, sex is a confounder that obscured the association between smoking and depression  negative confounding Keyes & Galea (2014) Epidemiology Collapsibility Approach to Confounding: Example Case-Control Study on Vitamin E and Coronary Heart Disease Can we assess for confounding using the OR? (Population Prevalence CHD = 6%) ORcrude = 0.59 (95%CI = 0.39, 0.89) Are there potential confounders that can explain the crude OR? Does smoking meet the a priori criteria for confounding? Fitzmaurice, 2004 Collapsibility Approach to Confounding: Example STRATIFY Smoking Vitamin CHD E Does smoking meet the a priori criteria for confounding? Fitzmaurice, 2004 Collapsibility Approach to Confounding: Example Stratum 1: ORsmokers = (11)(200)/(40)(49) = 1.12 95% CI = 0.48, 2.43 Stratum 2: ORnon-smokers = (39)(184)/(461)(16) = 0.97 95% CI = 0.51, 1.91 Fitzmaurice, 2004 Collapsibility Approach to Confounding: Example Interpretation Before adjusting for smoking, Vitamin E had a strong protective effect (ORcrude = 0.59) After stratification, there is little evidence of an association between vitamin E and CHD after controlling for smoking (ORsmoker = 1.12; ORnon- smoker = 0.97; ORadjusted = 1.03) ORcrude ORadjusted (+ rare disease assumption) Vitamin E group contains considerably fewer smokers (9.3% versus 55.5%), thereby decreasing this group's apparent risk of CHD  positive confounding Fitzmaurice, 2004 Confounding: Summary Causality is the central concern of epidemiology, and confounding is the central concern with establishing causality Confounding can be understood using several overlapping approaches i. Counterfactual approach ii. Classical approach iii. Collapsibility approach Recap Questions: Confounding The collapsibility approach to confounding suggests that confounding may be present when: A. The stratum specific effect estimates are approximately equal, but different from the crude effect estimate B. The stratum specific effect estimates are different from each other, and different from the crude effect estimate C. The adjusted effect estimate does not equal the crude effect estimate D. All of the above are correct E. Both A and C are correct Recap Questions: Confounding Maternal smoking is a well-established risk factor for low birth weight. Does birth weight meet the a priori criteria for confounding of the relationship between maternal smoking and infant mortality? A = Yes Birth B = No Weight Materna Infant l Mortalit Smokin y g  Control of Confounding Control of Confounding A. Design Stage B. Analysis Stage Randomization Standardization (Covered in Restriction Week 3) Stratification Matching Multivariable Adjustment (Biostats 9521B) Advanced Techniques (not covered in this class) - Graphical approaches (DAGs) - Propensity score methods - Instrumental variables - Marginal structural models Confounding Control at the Design Stage: Randomization Reduces potential for confounding by generating groups that are fairly comparable with respect to potential confounders Attempts to simulate a counterfactual contrast With a large enough sample, yields groups that are similar on both measured and unmeasured confounding factors – only method for confounding control that can achieve this Goal is to obtain comparison groups that differ only on exposure and are the “same” on all other important covariates Confounding Control at the Design Stage: Randomization There is the same proportion of confounders in both groups Exposed with Exposed without confounder confounder Unexposed with Unexposed without confounder confounder Keyes & Galea (2014) Epidemiology Randomization resulted in highly comparable distribution of potential confounders!  Confounding Control at the Design Stage: Randomization Confounde Randomization “breaks” r X any link between exposure and confounder Outcom Exposure e Confounding Control at the Design Stage: Randomization Limitations to Randomization For ethical reasons, only possible for intervention studies Confounding may still be introduced when there is high attrition/drop-out Covariates may remain imbalanced with insufficient sample size - Can use other methods for confounding control when this occurs RCT on Meditation for Stress among Post-Secondary Students: Small sample size means confounding factors are not well balanced across the groups Which imbalances might be of particular concern? Oman et al (2008) J Am Colleg Health Confounding Control at the Design Stage: Restriction Confounding cannot occur if the distribution of the potential confounding factors do not vary across exposure or disease categories Restriction on a confounding factor will eliminate variation in the confounder Advantages: straightforward, convenient, inexpensive Limitations: - Limits external validity of findings - May lead to recruitment difficulties - Unable to evaluate effect of factors that have been restricted - Confounding may remain if restriction not narrow enough (ex. age 40 to 60 years) Confounding Control at the Design Stage: Matching Involves selection of a comparison group that is forced to resemble the index group with respect to the distribution of one or more potential confounders In a cohort study, exposed individuals are matched to ≥ 1 unexposed individuals on ≥ 1 factor(s) of interest In a case-control study, diseased individuals are matched to a sample of disease free individuals The use of matching usually requires special analysis techniques (e.g. matched pair analyses and conditional logistic regression) Confounding Control at the Design Stage: Matching Exposed with Exposed with no confounder confounder Unexposed with Unexposed with no confounder confounder Each pair is identical with respect to the matched confounding factor Keyes & Galea (2014) Epidemiology Confounding Control at the Design Stage: Matching Advantages: Disadvantages: Increased precision May lose precision if match on weak factors Direct control of confounders May be difficult and expensive to find Practical aspects for control suitable match selection (ex. friend controls) Loss of information for cases where matched control cannot be found Overmatching: Unable to control confounding by factors May occur if the matching other than the one used for matching variable is closely connected with the mechanism by Effect of confounder cannot be estimated which the exposure affects Complexity of data analysis the outcome. ~ Porta (2014) Overmatching Confounding Control at the Design Stage: Matching General rule: paired data must be treated as such in the analysis – don’t break up pairs! The proper form of analysis is needed in matched studies : - Matched odds ratio - Conditional logistic regression Keyes & Galea (2014) Epidemiology Confounding Control at the Design Stage: Matching Matched OR = f/g The pairs in cells e and h do not contribute information because they are concordant Keyes & Galea (2014) Epidemiology Confounding Control at the Analysis Stage Unlike selection and information bias, confounding can be (to some extent!) adjusted in the analysis Options at the analysis stage: – Standardization (Direct and Indirect) – Stratification – Multivariate methods Ability to control confounding in analysis stage is only as good as the quality of our measures – need data on confounding factors to control for them! Keyes & Galea (2014) Epidemiology Confounding Control at the Analysis Stage: Stratification Objective of stratified analysis is to “fix” the level of the confounding variable and produce groups within which the confounder does not vary Then evaluate the exposure-disease association within each stratum of the confounder Within each stratum, the confounder cannot confound because it does not vary  Cases of Down syndrom by birth order and mother's age Cases per 100000 1000 900 800 700 600 500 400 300 200 100 0 1 2 3 4 5 Birth order Stratified Analysis Crude Crude 2 x 2 table RRCrude Calculate Crude RR Stratify by Confounder Stratum 1 Stratum 2 Calculate RR’s for each stratum RR1 RR2 If stratum-specific RR’s are similar, calculate adjusted RR If Crude RR Adjusted RR, If Crude RR = confounding is likely Adjusted RR, confounding isSlide courtesy of M. Pai (McGill Mantel-Haenszel Estimators of Adjusted Effects from Stratified Data Odds Ratio: (ad / n)i ORMH = (bc / n)i Mantel-Haenszl Estimate: A type of weighted average of the Risk Ratio: individual odds ratios, derived from dividing a sample into series of data. [a (c + d) / n]i Ideally, the strata would be internally RRMH homogeneous with respect to = [c(a + b) / n]i confounding factors. Can also be extended to risk ratios and incidence Incidence Rate Ratio: rate ratios. ~ Porta (2014) Dictionary of Epidemiology [a(PY0) / n] i IRRMH [c(PY1) / n]i = Where i is the stratum of the confounding variable and n is the total sample size of the stratum Stratification: Example Non-Smokers: Smokers: 31 RRnon-smoker = 0.167 / 0.063 = RRsmoker = 67.7 / 25.9 = 2.61 2.65 RRcrude = 0.595 / 0.186 = 3.20 RRMH = [1 * 16/22 + 21 * 27/58] [1 * 6/22 + 7 * 31/58] = 2.62 Interpretation: The risk of the outcome among people who are exposed is 2.62 times greater than in the unexposed, after adjusted for the confounding effects of smoking  Stratification: Limitations Crude ORCrude Stratum 1 Stratum 2 Confounder 1 OR1 OR2 Stratum 1-A Stratum 1-B Stratum 2-A Stratum 2-B Confounder 2 “Sparseness of data” problem Computational and interpretational difficulties Slide courtesy of M. Pai (McGill Confounding Control at the Analysis Stage: Multivariable Adjustment Stratification works best when there are few strata (i.e. if only 1 or 2 confounders have to be controlled) If the number of potential confounders is large, multivariable analyses are needed – can handle large number of confounders simultaneously Based on statistical regression models and covered in depth in multivariable methods next term Confounding Control at the Analysis Stage: Multivariable Adjustment Example: Association between preterm labor and low birth weight N = 189 subjects Outcome: low birth weight [low] (yes=1; no=0) Main exposure of interest: Preterm labor [ptl] (yes=1; no=0) Covariates: – Smoking status of mother [smoke] (yes=1; no=0) – Age (mother’s age in years) – Race (0=white; 1=black/others) Confounding Control at the Analysis Stage: Multivariable Adjustment Stratified Analysis (smoking) ORcrude = 4.32 ORsmokers = 4.22 ORnon-smokers = 3.48 ORMH = 3.89 Slide courtesy of M. Pai (McGill  Confounding Control at the Analysis Stage: Multivariable Adjustment Model 1: Logistic regression with no covariates Same OR as calculated from 2 x 2 table Slide courtesy of M. Pai (McGill  Confounding Control at the Analysis Stage: Multivariable Adjustment Model 2: Logistic regression adjusted for smoking Similar OR as calculated from Mantel-Haenszel estimator (adjusted) Slide courtesy of M. Pai (McGill  Confounding Control at the Analysis Stage: Multivariable Adjustment Model 3: Logistic regression with all covariates (fully adjusted) OR adjusted for age, race, and smoking Slide courtesy of M. Pai (McGill Residual Confounding also: Uncontrolled Confounding Confounding that persists after adjustment for the measured confounders May be due to several reasons: – Not all confounders were adjusted for - unmeasured confounding – Mis-specified models - some variables were actually not confounders – Confounders were measured with error - misclassification of confounders – Categories of the confounding variable are improperly defined - e.g. age categories were too broad Recap Question: Which of the following is NOT a strategy for controlling confounding at the design stage? A. Restriction B. Stratification C. Randomization D. Matching E. All of the above are strategies for controlling confounding at the design stage Recap Questions: Confounding Maternal smoking has been shown to increase the risk of infant mortality, possibly through the mediating effects of low birth weight. If we were to adjust for birthweight in our analysis, what would we be estimating? A. The direct effect of maternal smoking on infant mortality B. The indirect effect of maternal smoking on infant mortality C. The total effect of maternal smoking on infant mortality D. All of the above E. A and C only Materna Birth Infant l Weight Mortalit Smokin y g Recap Questions: Confounding Which of the following statements is false regarding restriction as a method for controlling confounding? A. Confounding may remain if the restriction is not narrow enough B. Confounding cannot occur if the distribution of the potential confounding factor does not vary across exposure or disease categories C. It may increase the external validity of the findings D. It is not a good option if you are interested in assessing the effects of the factor that has been restricted E. All of the above are correct Summary Always worry about confounding in your research! Confounding can be controlled at the design stage through randomization, restriction, and matching Confounding can be controlled at the analysis stage through standardization, stratification, and multivariate adjustment (in addition to more advanced techniques) Adjustment for confounding may always be inadequate due to residual confounding