1. When do causes work together? Epidemiology matters: a new introduction to methodological foundations Chapter 11

2. Seven steps 1.Define the population of interest 2.Conceptualize and create measures of exposures and health indicators 3.Take a sample of the population 4.Estimate measures of association between exposures and health indicators of interest 5.Rigorously evaluate whether the association observed suggests a causal association 6.Assess the evidence for causes working together 7.Assess the extent to which the result matters, is externally valid, to other populations Epidemiology Matters – Chapter 12

3. Component causes of disease rarely act in isolation Epidemiologic exposures are typically one of a set of component causes that have to work together in order for a change to occur in the health indicator Interaction: when multiple component causes work together to produce a particular health indicator 3Epidemiology Matters – Chapter 11

4. 1.Interaction, conceptual 2.Assessing interaction in data 3.Interaction across scales 4.Additivity, multiplicativity, and interaction 5.Additive interaction with ratios 6.Random variation 7.Summary Epidemiology Matters – Chapter 114

5. 1.Interaction, conceptual 2.Assessing interaction in data 3.Interaction across scales 4.Additivity, multiplicativity, and interaction 5.Additive interaction with ratios 6.Random variation 7.Summary Epidemiology Matters – Chapter 115

6. Epidemiology Matters – Chapter 86 Non-diseased Diseased Non-exposed Exposed

7. Interaction, conceptual Causes interact when they work together as part of the same sufficient cause, i.e., marble set Causes that interact are causes in which both factors are necessary to cause disease in at least one sufficient cause For example, what can ‘cause’ a sprinter to work a 100 meter dash  Only trains for years Does not win  Only has tied running shoes Does not win  Only reacts promptly to the starter’s pistolDoes not win  Trains for years, tied shoes, prompt reactionSprinter wins 7Epidemiology Matters – Chapter 11

8. Causes of Epititis 8Epidemiology Matters – Chapter 11 Family history Exposure to toxins in utero 20 pack-years of smoking Neighborhood poverty Male sex Stressful experiences in adulthood

9. Interaction, conceptual: Epititis Male sex and family history are both component causes, they are components of different sufficient causes and do not interact Two components interact if they need to work together within the same sufficient cause 9Epidemiology Matters – Chapter 11

10. Comparability and interaction  Family history and in utero exposure are part of same set of marbles that cause Epititis  To develop Epititis as a result of sufficient cause 1 must always have both family history of Epititis and exposure to toxins in utero  No variation in relation between either component cause (marbles) and the outcome (Epititis) when one or the other is present  Family history and toxins interact to produce disease  Therefore, family history is part of mechanism through which in utero exposure to toxins works - does not create non-comparability between exposed and unexposed 10Epidemiology Matters – Chapter 11

11. 1.Interaction, conceptual 2.Assessing interaction in data 3.Interaction across scales 4.Additivity, multiplicativity, and interaction 5.Additive interaction with ratios 6.Random variation 7.Summary Epidemiology Matters – Chapter 1111

12. Interaction in theory  We could determine with certainty who would get disease if we could measure every component cause in a sufficient cause  Those exposed to all component causes would inevitably get disease  Those who do not have all the component causes, would never get disease  However, this is never the case, i.e., we can never know what all the component causes are and we therefore have to assess for causes that work together (i.e., interact) in our data 12Epidemiology Matters – Chapter 11

13. Assessing interaction, core concept We can observe interaction when measure of association for exposure and outcome varies across levels of third variable 13Epidemiology Matters – Chapter 11

14. Interaction example alcohol consumption Question: Is consuming alcohol before driving associated with risk of dying in a motor vehicle crash? Other factors that can contribute to risk of dying in a motor vehicle crash include time of day, wearing a seatbelt Key questions of interest here are  Does alcohol consumption cause a greater risk of dying in a motor vehicle crash?  Does alcohol consumption interact with either (or both) time of day and seatbelt use in its causing motor vehicle crashes? How would we answer these questions? 14Epidemiology Matters – Chapter 11

15. Interaction example alcohol consumption, data  Amount of alcohol consumed before driving  Subsequent death in a motor vehicle crash  Time of day that driving occurs  Driver wearing a seatbelt 15Epidemiology Matters – Chapter 11

16. Alcohol consumption and death seatbelt use 16Epidemiology Matters – Chapter 11 Seatbelt use Risk of death in exposed: 5% Risk of death in unexposed: 1% No seatbelt use Risk of death in exposed: 10% Risk of death in unexposed: 6%

17. Alcohol consumption and death seatbelt use Alcohol use is always associated with greater risk of death Seat belt and alcohol use  Among those who did not wear a seatbelt, the risk of dying in crash was 10% among those who consumed alcohol prior to driving and 6% among those who did not consume alcohol prior to driving  Risk difference (RD) = 0.10 - 0.06 = 0.04 (95% CI 0.0162, 0.0637)  Among those who did wear a seatbelt, the risk of dying in crash was 5% among those who consumed alcohol prior to driving and 1% among those who did not consume alcohol prior to driving  Risk difference (RD) = 0.05 – 0.01 = 0.04 (95% CI 0.0238, 0.0541) Therefore there is no difference in risk difference between those who do and do not use a seatbelt. Seat belt use and alcohol use are part of different ‘marble sets’ and do not operate jointly to cause crash death. This indicates no interaction. 17

18. Alcohol consumption and death time of day 18Epidemiology Matters – Chapter 11 Daytime Risk of death in exposed: 5% Risk of death in unexposed: 1% Nighttime Risk of death in exposed: 15% Risk of death in unexposed: 6%

19. Alcohol consumption and death time of day Alcohol use is always associated with greater risk of death Time of day and alcohol use  Among those who drove at night, the risk of dying in crash was 15% among those who consumed alcohol prior to driving and 6% among those who did not consume alcohol prior to driving  Risk difference (RD) = 0.15 – 0.06 = 0.09 (95% CI 0.0634, 0.1165)  Among those who drove during the day, the risk of dying in crash was 5% among those who consumed alcohol prior to driving and 1% among those who did not consume alcohol prior to driving  Risk difference (RD) = 0.05 – 0.01 = 0.04 (95% CI 0.0238, 0.0541) Therefore there is a difference in risk differences associated with alcohol consumption for nighttime drivers and for daytime drivers; this indicates the presence of interaction 19Epidemiology Matters – Chapter 11

20. Looking for interaction in data  Examine the evidence for interaction in data by comparing magnitude of association between exposure and disease across a third variable  If measure of association differs across levels of the third variable, there is evidence of interaction for that measure  If measure of association does not differ across levels of third variable - is not evidence of interaction 20Epidemiology Matters – Chapter 11

21. 1.Interaction, conceptual 2.Assessing interaction in data 3.Interaction across scales 4.Additivity, multiplicativity, and interaction 5.Additive interaction with ratios 6.Random variation 7.Summary Epidemiology Matters – Chapter 1121

22. Interaction across scales The presence of interaction depends on the measure of association we are examining 22Epidemiology Matters – Chapter 11

23. Interaction across scales, example Question: Is consumption of green tea associated with reduced risk of stomach cancer? Does the relationship vary by whether individuals have diets that are rich in smoked and cured food? Purposive sample of 4000 individuals without stomach cancer  1000 drink green tea and do not eat smoked/cured foods  1000 drink green tea and eat smoked/cured foods  1000 do not drink green tea but eat smoked/cured foods  1000 do not drink green tea and eat smoked/cured foods  All follow forward for twenty years to determine which individuals develop stomach cancer 23Epidemiology Matters – Chapter 11

24. Green tea and cancer no smoked/cured food 24Epidemiology Matters – Chapter 11 Interpretation: Among those who do not eat smoked/cured foods, green tea consumption is associated with 0.5 times the odds of stomach cancer compared with those who do not consume green tea.

25. Green tea and cancer smoked/cured food 25Epidemiology Matters – Chapter 11 Interpretation: Among those who consume smoked/cured foods, green tea consumption is associated with 0.8 times the odds of stomach cancer compared with those who do not consume green tea.

26. Green tea and cancer interaction scale Based on the risk ratio and the odds ratio, green tea consumption has a stronger protective effect among those who do not consume smoked/cured meats than among those who do consume such food. Therefore, there is evidence of interaction between green tea and smoked/cured foods However, based on risk differences across the two strata indicates that green tea consumption is associated with 5 fewer cases of stomach cancer for every 1,000 individuals who consume green tea, regardless of whether an individual consumes smoked/cured foods or not, i.e., no evidence of interaction between green tea and smoked/cured foods Interaction is dependent on whether we use relative measures or difference measure 26Epidemiology Matters – Chapter 11

27. 1.Interaction, conceptual 2.Assessing interaction in data 3.Interaction across scales 4.Additivity, multiplicativity, and interaction 5.Additive interaction with ratios 6.Random variation 7.Summary Epidemiology Matters – Chapter 1127

28. Interaction is scale dependent Additive: if two exposures do not interact, the risk of disease among exposed to both exposures = sum of risk of disease given exposure to one factor + risk of disease given exposure to the other factor Multiplicative: If two exposures do not interact, the risk of disease among those exposed to both = product of risk of disease given exposure to one factor * risk of disease given exposure to the other factor 28Epidemiology Matters – Chapter 11

29. Interaction is scale dependent, example A Risk among those exposed to both X and Y: 10% Risk among those exposed to X but not Y: 6% Risk among those exposed to Y but not X: 5% Risk among those exposed to neither X nor Y: 1% There is no evidence of additive interaction. The risk of disease among those exposed to both X and Y is = sum of the risk associated with exposure to X alone, plus Y alone, minus the exposure associated with neither exposure (10=6+5-1) This is evidence of multiplicative interaction. The risk of disease among those exposed to both X and Y to be 30% if there were no multiplicative interaction, because 6x5=30 - observed risk is 10% < 30% 29Epidemiology Matters – Chapter 11

30. Interaction is scale dependent, example B Risk among those exposed to both X and Y: 30% Risk among those exposed to X but not Y: 6% Risk among those exposed to Y but not X: 5% Risk among those exposed to neither X nor Y: 1% There is no evidence of multiplicative interaction. The risk of disease among those exposed to both X and Y = to product of the risk associated with exposure to X alone, times Y alone (30=6*5) There is evidence of additive interaction. 30% is greater than the sum of risks for those exposed to X but not Y (6%) and Y but not X (5%) (minus the risk among those exposed to neither, 1%) 30Epidemiology Matters – Chapter 11

31. Interaction, use of additive scale When two factors are causal partners in the same sufficient cause, the resulting measures of association will depart from additivity, but not necessarily from multiplicativity The general recommendation is that interaction, or the search for factors that co-occur in the same sufficient cause, should be assessed on an additive scale 31Epidemiology Matters – Chapter 11

32. 1.Interaction, conceptual 2.Assessing interaction in data 3.Interaction across scales 4.Additivity, multiplicativity, and interaction 5.Additive interaction with ratios 6.Random variation 7.Summary Epidemiology Matters – Chapter 1132

33. Additive interaction with ratio Interaction arises when there are two (or more) component causes of the same sufficient cause influencing outcome of interest Evidence of interaction in our data comes when we asses measure of association between exposure and outcome differs across levels of third variable Evidence for interaction will be dependent on measure of association used (additive interaction scale best) 33Epidemiology Matters – Chapter 11

34. Additive interaction with ratio What if we are unable to estimate risk or rate differences? The odds ratio is an appropriate measure of association for some study designs We can therefore estimate interaction with ratio measures (odds ratio, risk ratio, or rate ratio) 34Epidemiology Matters – Chapter 11

35. Additive interaction, with ratio, example  We are interested in the association between consumption of aspartame and stroke  Purposive sample - 200 cases of stroke newly diagnosed at hospitals and 600 individuals who have never had a stroke from communities of hospitals  Hypothesize that individuals with a family history of stroke are vulnerable to effects of aspartame, i.e., that aspartame and family history are causal partners in a sufficient cause for stroke 35Epidemiology Matters – Chapter 11

36. Aspartame and stroke 36Epidemiology Matters – Chapter 11 No family history of strokeFamily history of stroke This does not give us information about presence of additive interaction between aspartame use and family history - we are examining variation in the odds ratio - a multiplicative measure

37. Aspartame and stroke To assess whether additive interaction is present, divide the sample into 1.Family history of stroke and regular aspartame user (A+F+) 2.Regular aspartame user with no family history of stroke (A+F-) 3.Family history of stroke but not an aspartame user (A-F+) 4.No family history and no aspartame use (A-F-) Then estimate three odds ratios and compare each to the fourth category 37Epidemiology Matters – Chapter 11

38. Aspartame and stroke 38Epidemiology Matters – Chapter 11 Aspartame+ Family+ to Aspartame- Family- Aspartame+ Family- to Aspartame- Family- Aspartame- Family+ to Aspartame- Family- Estimate magnitude of interaction between family history and aspartame Interaction contrast ratio (ICR): ICR=OR++ - OR+- - OR-+ + 1 Hypothetical study ICR= OR++ - OR+- - OR-+ + 1 ICR = 2.15 - 1.03 - 1.04 + 1 = 1.08 This suggests some, if not much, additive interaction between aspartame and family history

39. 1.Interaction, conceptual 2.Assessing interaction in data 3.Interaction across scales 4.Additivity, multiplicativity, and interaction 5.Additive interaction with ratios 6.Random variation 7.Summary Epidemiology Matters – Chapter 1139

40. Random variation Appearance of interaction can arise due to chance in sampling process We may collect a sample in which there were, by chance, a large proportion of individuals with disease in a certain subgroup Therefore confidence intervals around interaction measures are important 40Epidemiology Matters – Chapter 11

41. 1.Interaction, conceptual 2.Assessing interaction in data 3.Interaction across scales 4.Additivity, multiplicativity, and interaction 5.Additive interaction with ratios 6.Random variation 7.Summary Epidemiology Matters – Chapter 1141

42. Interaction summary  Interaction occurs when two causes are both components of the same sufficient cause  When two causes interact this means that at least some individuals become diseased through a certain sufficient cause that includes both component causes  We can observe interaction when measure of association for exposure and outcome varies across levels of third variable  Different measures of association will evidence difference variation over a third variable depending on the scale (additive or multiplicative)  Epidemiology we are principally concerned with additive interaction 42Epidemiology Matters – Chapter 11

43. Seven steps 1.Define the population of interest 2.Conceptualize and create measures of exposures and health indicators 3.Take a sample of the population 4.Estimate measures of association between exposures and health indicators of interest 5.Rigorously evaluate whether the association observed suggests a causal association 6.Assess the evidence for causes working together 7.Assess the extent to which the result matters, is externally valid, to other populations Epidemiology Matters – Chapter 143

44. epidemiologymatters.org 44Epidemiology Matters – Chapter 1