Confounding and Interaction I

1. Confounding and Interaction I • Confounding: one of the central problems in observational clinical research • What is it? What does it do? What is its origin? • What kind of variables act as confounders? • Which variables are not confounders (colliders and intermediary variables)? • Use of causal diagrams (DAGs) to conceptualize confounding and guide us for what to adjust for • Emphasis on specifying the research question and understanding the underlying biological/clinical/behavioral system • Confounding is a substantive, not statistical issue

2. Matches and Lung Cancer • A tobacco company researcher believes that exposure to matches is a cause of lung cancer • He conducts a large case-control study to test this hypothesis • Exposure odds ratio = (820/180) / (340/660) = disease odds ratio • OR = 8.8 • 95% CI (7.2, 10.9) • Should we remove matches from the environment as a means of preventing lung cancer?

3. Smoking, Matches, and Lung Cancer Crude OR crude Stratified Smokers Non-Smokers OR CF+ = ORsmokers OR CF- = ORnon-smokers • Stratification produces two 2-by-2 tables • In each stratum, all subjects are homogeneous with respect to smoking • We have adjusted,controlled, or conditioned for smoking • ORcrude = 8.8 (7.2, 10.9) • ORsmokers = 1.0 (0.6, 1.5) • ORnon-smoker = 1.0 (0.5, 2.0) • ORadjusted = 1.0 (0.5, 2.0)

4. Confounding: Smoking, Matches, and Lung Cancer • Illustrates how confounding can create an apparent effect even when there is no actual true effect • Can also be opposite: confounding can mask an effect when one is truly present • Proper terminology • In the relationship between matches and lung cancer, smoking is a confounding factor or a confounder • Smoking confounds the relationship between matches and lung cancer • In clinical research, confounding has a very specific meaning

5. Estes continues to be confounding puzzle Ray RATTO SHAWN ESTES seemed loath to analyze his own performance last night, for fear that people would see the first three innings and use them to obscure the last four. But that's what made his outing so perfectly Estes-like -- an ongoing argument with himself that he eventually won. Well, an argument in which he held his own and his teammates won for him in the bottom of the ninth. Ramon Martinez lined a game-tying single with two outs, and Jeff Kent followed two batters later with a bases-loaded walk off Juan Acevedo to give the Giants a 2-1 victory against Colorado and move them to within 4 1/2 games of division leader Arizona. It was in many ways an eye-opening victory for a team that hadn't had one of this type for a while.

6. Submission to NEJM • Finding: “After an initial course of post-exposure prophylactic (PEP) medication following a sexual exposure to HIV infection, gay men reported a decrease in the practice of high-risk behavior over the following year.” • Reviewer: “Perhaps the men simply withheld the real amount of high-risk behavior they had in order to be eligible for future courses of PEP. How do you account for this confounding?”

7. The study is not over! • To be complete, you decide to examine the relationship between smoking and lung cancer independent from the use of matches. Crude OR crude Matches Present Matches Absent Stratified OR CF+ = ORmatches OR CF+ = OR no matches • ORcrude = 21.0 (16.4, 26.9) • ORmatches = 21.0 (10.7, 41.3) • ORno matches = 21.0 (13.1, 33.6)

8. Confounding: Smoking, Matches, and Lung Cancer • Interpretation? • What is the effect of matches on the relationship between smoking and lung cancer? Matches have no effect on the relationship • Effect of matches could have been predicted based on matches — lung cancer relationship • Illustrates one important component in the requirements of a confounder (aka a confounding factor) - A confounder must be causally related to the outcome

9. Confounding: Examples of Magnitude and Direction Crude (unadjusted) OR crude Potential Confounder Absent Stratified (adjusted) Potential Confounder Present OR CF+ OR CF-

10. Nightlights Let there be light!

11. Nightlights and Myopia • Quinn et al. Nature 1999 • Prevalence Ratio =

12. Insert picture with nightlight off Lights are off and the stumbling around begins.

13. Nightlights and Myopia • Two subsequent studies found no association and explained the prior result by confounding • Zadnik et al. and Gwiazda et al. Nature, 2000

14. How might confounding account for this finding? Night Light ? Child’s Myopia

15. Positive or negative confounding? Night Light Positive X Parental Myopia Child’s Myopia

16. Insert picture with nightlight on again Let there be light (again)!

17. AZT to Prevent HIV After Needlesticks • Case-control study of whether post-exposure AZT use can prevent HIV seroconversion after needlestick (NEJM 1997) Crude ORcrude = 0.61 (95% CI: 0.26 - 1.4) Interpretation? Could confounding be present?

18. AZT Use ? HIV

19. AZT Use Positive or negative confounding? Severity of Exposure ? HIV

20. Adjustment for Confounder Crude • Potential confounder: severity of exposure ORcrude =0.61 Stratified Minor Severity Major Severity ORadjusted = 0.30 (95% CI: 0.12 – 0.79) Negative Confounding “Confounding by Indication”

21. Classification Schemes for Error in Clinical Research • Szklo and Nieto • Bias • Selection Bias • Information/Measurement Bias • Confounding • Chance • Other Common Approach • Bias • Selection Bias • Information/Measurement Bias • Confounding • Chance S+N: confounded relationships are not biased but rather simply not causal; can be useful in identifying high-risk populations for secondary prevention

22. Exposed to night lights Unexposed to night lights time Counterfactuals: Conceptualizing Why Confounding Occurs Night lights and myopia • Ideal study: evaluate children exposed to night lights for several years and directly compare them to the SAME children not exposed to night lights • Result (e.g. risk ratio) is called the “causal effect measure” of night lights • Assuming no measurement error, the “causal effect measure” must be true. • However, since time has passed and children are older it is impossible to assess them without night lights • Hence, the ideal is “counterfactual” – contrary to the fact. It is unobservable. It cannot happen. Go back in time

23. men women time Counterfactuals: Conceptualizing Why Confounding Occurs Gender and heart disease • Ideal study: evaluate men for several years for occurrence of heart disease; compare them directly to SAME individuals who are now women • However, you cannot change a man into a woman and you cannot go back in time • The “causal effect measure” is preposterous to consider. It cannot be observed. It is counterfactual.

24. Exposed to night lights Unexposed to night lights time Counterfactuals: Conceptualizing Why Confounding Occurs Nights and Myopia • Because we cannot perform the counterfactual ideal (SAME population studied under 2 conditions), we must evaluate TWO distinct populations (exposed to a night light and unexposed) to study the problem • Result (e.g. risk ratio): a “measure of association” • The TWO distinct populations may be subject to different influences OTHER than just the night light • If these influences cause the disease under study, any difference in the risk ratio between the SAME population study (effect measure) and the TWO population study (measure of association) is what is known as confounding • Confounding occurs because of these other influences, a mixing of effects Other influences

25. Striving for the Counterfactual In the real (observable) world • All of our strategies in analytic studies are striving to simulate the counterfactual • We strive for our TWO distinct populations (exposed and unexposed) to be “exchangeable” • i.e., identical in the other influences upon them • Whenever the TWO distinct populations are “non-exchangeable”, confounding will be present • Our strategies to manage confounding are attempts to make our populations exchangeable

26. Why Strive for the Counterfactual? • Counterfactual ideal would allow certainty in knowing whether a numerical/statistical association between an exposure and outcome is causal • Knowing whether a relationship is causal is the “holy grail” in clinical research • If a relationship is causal, interventions that change the exposure will change the outcome

27. Back to the Observable (Factual) World: Criteria for Confounding • Confounding occurs because of mixing between exposures of interest and unwanted extraneous effects • Extraneous effects are termed confounders • Classical criteria for a confounder • Must be causally associated with the outcome, or be a surrogate for a causally related variable • Must be associated with the exposure under study, but cannot be caused by the exposure • Must not be on the causal pathway under study (i.e., must not be an intermediary variable)

28. Causal Diagrams -- DAGs • DAGs = directed acyclic graphs; aka chain graphs • Consist of nodes (variables) and edges (lines or arrows) • “Directed”: all edges have one-way direction which depict causal relationships • “Acyclic”: there is never a complete circle around any node (i.e. no factor can cause itself) Research Question: Does E cause D? • Hashed edge with ? sometimes used to depict relationship under study • Other edges drawn based on prior causal knowledge of system • All nodes immediately caused by another node called “child” node; the proximal node is “parent” • All nodes along a path with edges in same direction are “descendants”; all proximal nodes are “ancestors” X E Y ? D Z W

29. Causal Diagrams -- DAGs • Better than the classical criteria for confounding when planning studies and analyses • Frontier of epidemiologic theory • Forces investigator to conceptualize system • Forces the issue about what is known vs unknown • Identifies pitfalls of adjusting and not adjusting for certain variables • WARNING: S + N text does not strictly follow contemporary rules of DAG depiction (requirement of uni-directional edges, etc,) but we will in class and in problem sets

30. Confounding in a DAG Confounding occurs if there is a factor C that is a “Common Cause” of both E and D E C ? • C is part of a “backdoor path” between E and D • Adjusting/controlling for C closes the backdoor path; eliminates confounding D

31. RQ: Do matches cause lung cancer? Matches Smoking ? Smoking is a “common cause” of matches and lung cancer. It therefore confounds the relationship (positive CF) Controlling for smoking blocks the backdoor path and unconfounds relationship Lung Cancer

32. Multi-vitamin Use Genetic factor is the “common cause” but cannot be measured or adjusted for History of birth defects ? Threat: negative confounding Genetic Factor (not measured) Birth Defects Adjusting for history of birth defects, which can be measured, blocks the path between genetic factor and MVI use, and prevents confounding Hernan AJE 2002

33. Safety-oriented Personality (not measured) Use of Helmets in Motorcyclists Threat: positive confounding Safe Driving Practices ? Adjusting for safe driving practices, which can (theoretically) be measured, blocks path from safety-oriented personality to head injury SeriousHead Injury

34. Attraction of DAGs • Abstract: The Classical Criteria • Must be causally associated with the outcome, or be a surrogate for a causally related variable • Must be associated with the exposure under study, but cannot be caused by the exposure • Must not be on the causal pathway under study (i.e. must not be an intermediary variable) • More tangible: DAGs • Draw the system • Look for “common causes” of exposure and disease Multi-vitamin Use History of birth defects ? Genetic Factor (not measured) Birth Defects

35. The Challenge • DAGs provide the framework • However, to identify the confounders, you need to be a subject matter expert • Confounding is a substantive rather than statistical issue • Advice: before planning a study, spend several weeks in the library

36. RQ: Does sexual activity cause greater lifespan? Sexual Activity ? Mortality

37. RQ: Does sexual activity cause greater lifespan? Sexual Activity Self-reported General Health ? Unknown biologic factor(s) (not measured) Mortality

38. RQ: Do calcium channel blockers cause GI bleeding? Ca channel Blockers ? GI Bleeding

39. RQ: Do calcium channel blockers cause GI bleeding? Ca channel Blockers Coronary Artery Disease Other Meds (e.g., aspirin) ? GI Bleeding

40. RQ: Does lack of folate cause birth defects? What should we do with stillbirths (spontaneous abortions)? Folate Intake Slone Epidemiology Unit Birth Defects Study • Stillbirths are associated with folate intake, even among infants without birth defects: OR = 0.50 (protective) • Stillbirths are associated with birth detects: OR = 15.22 • Stillbirths are not on the causal pathway between folate and birth defects • In the past, other investigators have commonly adjusted for stillbirths in analyses, or have limited analyses to live births. • Should we adjust for stillbirths? ? Birth Defects Hernan AJE 2002

41. Adjustment for Stillbirths Slone Epidemiology Unit Birth Defects Study Crude ORcrude = 0.65 (95% CI 0.45 – 0.95) Stratified No stillbirth Stillbirth ORadjusted = 0.80 (95% CI: 0.53 – 1.2) Apparent positive confounding Public health implication: No reason for women to supplement diet with folate Hernan AJE 2002

42. Use of DAGs to Identify What is Not Confounding RQ: Does lack of folate intake cause birth defects? Folate Intake ? Stillbirths Birth Defects Undirected edge (interpret as going either direction) Stillbirths are a “common effect” of both the exposure and disease – not a common cause. Common effects are called “colliders” Adjusting for colliders OPENS paths. Will actually result in bias. It is harmful. Hernan AJE 2002

43. DAGs to Identify What is Not Confounding Behavioral factors (not measured) Multi-vitamin use Maternal Weight Gain ? Genetic Factor (not measured) Birth Defects No common causes for exposure and disease Maternal weight gain is a collider Adjusting for colliders will OPEN the path. Will actually result in bias. It is harmful. Hernan AJE 2002

44. DAGs Force Investigators to First Conceptualize the System Study of sunlight exposure & melanoma • A college intern is given a dataset and asked to estimate relationship between sunlight exposure and melanoma – adjusted for “everything” • He analyzes the data and finds that gum chewing is associated with melanoma and associated with sunlight exposure • After adjusting for gum chewing there is an appreciable difference between the crude and adjusted measure of association • Should gum chewing be controlled for? • No. Just by chance alone there can be the appearance of confounding • Based on our a priori understanding of the role of gum chewing (in melanoma), it is more likely that chance – as opposed to truth -- is causing appearance of confounding • Controlling for a variable should only be done if there is a strong subject matter evidence. • i.e. If it is not in your DAG, don’t control for it.

45. Rules for Reading DAGs • A backdoor path between E and D is blocked if • a collider (“common effect”) is present, which has not been adjusted for (by stratification, mathematical regression or other techniques) Or • a non-collider (“common cause”) is adjusted for • To prevent confounding, block all backdoor paths Folate ? Stillbirths Birth defects Nightlights Parental Myopia ? Child’s Myopia

46. Rules for Reading DAGs • A backdoor path between E and D is open if • A collider (“common effect”) is adjusted for Or • a non-collider (“common cause”) is not adjusted for • Open backdoor paths produce bias Folate ? Stillbirths Birth defects Nightlights Parental Myopia ? Child’s Myopia

47. What other variables are NOT Confounders? • “Must not be on the causal pathway under study (i.e. must not be an intermediary variable)” • A variable that you are conceiving as an intermediate step in the causal path under study between the exposure in question and the disease is not a confounding variable. Despite being associated with both exposure and outcome, Factor I is not a confounder It is on the pathway under study. It is an intermediary variable E factor I D

48. CCR5 and HIV Disease Progression • CCR5: the human cellular receptor for HIV –found on CD4 cells • Genetic defects in CCR5 now described - Are defects associated with slower progression to AIDS? • CD4 count potent predictor of time-to-AIDS • How should CD4 count be handled in assessing the association between CCR5 defect status and progression in HIV disease to AIDS? CCR5 (receptor) defect CD4 count ? AIDS

49. CCR5 and HIV Disease Progression • CCR5: the human cellular receptor for HIV –found on CD4 cells • Genetic defects in CCR5 now described • CD4 count potent predictor of time-to-AIDS • How should CD4 count be handled in assessing the association between CCR5 defect status and progression in HIV disease to AIDS? CCR5 (receptor) defect CD4 count AIDS

50. It depends upon the research question #1: Do CCR5 defects reduce progression to AIDS, irrespective of mechanism? CCR5 defect ? [CD4 count] Do not adjust for CD4 count ! AIDS #2: Do CCR5 defects reduce progression to AIDS, independent of CD4 count? CCR5 defect ? [Other mechanisms] CD4 count AIDS Do Adjust ! High CD4 count Low CD4 count