Concepts of Interaction Matthew Fox Advanced Epi What is interaction? Interaction? Covar + Covar - D+ D- E+ 600 400 E300 700 E+ 40 960 E20 980 Total Risk 1000 0.4 1000 0.2 1000 0.04 1000 0.02 RR 2 2 OR 3.5 2.04 Interaction? Smokers Non-smokers Asbestos + Asbestos - Asbestos + Asbestos - LC 20 10 3 1 No LC 980 990 997 999 Total 1000 1000 1000 1000 Risk 0.02 0.01 0.003 0.001 RR 2 3 RD 0.010 0.002 Last Session New approaches to confounding Instrumental variables – Propensity scores – – Variable strongly predictive of exposure, no direct link to outcome, no common causes with outcome Summarize confounding with a single variable Useful when have lots of potential comparisons Marginal structural models – – Use weighting rather than stratification to adjust Useful when we have time dependent confounding This session Concepts of interaction – – – Very poorly understood concept Often not clear what a person means when they suggest it exists Often confused with bias Define each concept – – Distinguish between them Which is the most useful 3 Concepts of Interaction Effect Measure Modification – Interdependence – Measure of effect is different in the strata of the modifying variable Risk in the doubly exposed can’t be explained by the independent effects of two single exposures Statistical Interaction – Cross-product term in a regression model not = 0 Point 1: Confounding is a threat to validity. Interaction is a threat to interpretation. Concept 1: Effect Measure Modification Effect measure modification (1) Measures of effect can be either: – – Difference scale (e.g., risk difference) Relative scale (e.g., relative risk) To assess effect measure modification: – – – – Stratify on the potential effect measure modifier Calculate measure of effect in all strata Decide whether measures of effect are different Can use statistical tests to help (only) No EMM corresponds to Difference scale: – – – If RD comparing A+ vs A- among B- = 0.2 and RD comparing B+ vs B- among A- = 0.1, then RD comparing A+,B+ to A-,B- (doubly exposed to doubly unexposed) should be: 0.2 + 0.1 = 0.3 Relative scale: – – – If RR comparing A+ vs A- among B- = 2 and RR comparing B+ vs B- among A- = 3, then RR comparing A+,B+ to A-,B- should be: 2 * 3 = 6 EMM on Relative Scale? Smokers Non-smokers Asbestos + Asbestos - Asbestos + Asbestos - LC 20 10 3 1 No LC 980 990 997 999 Total 1000 1000 1000 1000 Risk 0.02 0.01 0.002 0.001 2 Ref 2 Ref S+, A+ vs S-,A- S+ vs Samong A- A+ vs Aamong S- RR 0.02/0.001 0.01/0.001 0.002/0.001 RR 20 10 2 RR 20 = 10 * 2 EMM on Difference Scale? Smokers Non-smokers Asbestos + Asbestos - Asbestos + Asbestos - LC 20 10 3 1 No LC 980 990 997 999 Total 1000 1000 1000 1000 Risk 0.02 0.01 0.002 0.001 RD 0.01 Ref 0.001 Ref S+, A+ vs S-,A- S+ vs Samong A- A+ vs Aamong S- RD 0.02-0.001 0.01-0.001 0.002-0.001 RD 0.019 0.009 0.001 0.019 ≠ 0.009 + 0.001 = 0.010 Effect measure modification (2) If: – – Then: – – Exposure has an effect in all strata of the modifier Risk is different in unexposed group of each stratum of the modifier (i.e., modifier affects disease) There will always be some effect measure modification on one scale or other (or both) you must to decide if it is important Therefore: – More appropriate to use the terms “effect measure modification on the difference or relative scale” Example 1 (1) All Death Person-years Rate rate difference relative rate CD4 <200 175 25915 68/10000 33/10000 1.97 CD4 ≥200 89 25949 34/10000 0 Example 1 (2) Death Person-years Rate Rate Difference Relative Rate – – Treated CD4 <200 CD4 ≥200 157 81 24531 24545 64/10000 33/10000 31/10000 0 1.94 1 Is there confounding? – Untreated CD4 <200 CD4 ≥200 18 8 1384 1404 130/10000 57/10000 73/10000 0 2.28 1 Does the disease rate depend on treatment in unexposed? Does exposure prevalence depend on treatment in pop? Is the relative rate collapsible? Effect measure modification — difference scale? Effect measure modification — relative scale? But EMM of OR can be misleading Total Risk Covar + E+ 600 400 1000 0.4 RR 2.0 2.0 OR 3.5 2.04 D+ D- E300 700 1000 0.2 Covar E+ 40 960 1000 0.04 E20 980 1000 0.02 A simple test for homogeneity Large sample test – – More sophisticated tests exist (e.g., Breslow-Day) Assumes homogeneity, must show heterogeneous Tests provide guidance, not the answer ˆ RDˆ R D z 1 2 ln RRˆ z 1 SE (RDˆ1 )2 SE (RDˆ 2 )2 ln RRˆ 2 SE (ln RRˆ 1 )2 SE (ln RRˆ 2 )2 SE for difference measures a b SE (IRD) 2 2 PTE PTE a c bd SE (RD) 2 2 NE (NE 1) NE (NE 1) SE for relative measures SE (IRR ) 1 1 a b SE (RR ) 1 1 1 1 a NE b NE SE (OR) 1 1 1 1 a b c d Point 3: Effect measure modification often exists on one scale by definition. Doesn’t imply any interaction between variables. Perspective With modification, concerned only with the outcome of one variable within levels of 2nd – – The second may have no causal interpretation Sex, race, can’t have causal effects, can be modifiers Want to know effect of smoking A by sex M: – Pr(Ya=1=1|M=1) - Pr(Ya=0=1|M=1) = Pr(Ya=1=1|M=0) - Pr(Ya=0=1|M=0) or – Pr(Ya=1=1|M=1) / Pr(Ya=0=1|M=1) = Pr(Ya=1=1|M=0) / Pr(Ya=0=1|M=0) Surrogate modifiers Just because stratification shows different effects doesn’t mean intervening on the modifier will cause a change in outcome Cost of surgery may modify the effect of heart transplant on mortality – More expensive shows a bigger effect Likely a marker of level of proficiency of the surgeon – Changing price will have no impact on the size of the effect Concept 2: Interdependence Interdependence (1) Think of the risk of disease in the doubly exposed as having four components: – – – – Baseline risk (risk in doubly unexposed) Effect of the first exposure (risk difference 1) Effect of the second exposure (risk difference 2) Anything else? Baseline risk Effect of E1 Effect of E2 Anything else Think again about multiplicative scale Additive scale: – – – Risk difference Implies population risk is general risk in the population PLUS risk due to the exposure Assumes no relationship between the two Multiplicative scale: – – – Risk ratio Implies population risk is general risk in the population PLUS risk due to the exposure Further assumes the effect of the exposure is some multiple of the baseline risk Four ways to get disease Cases of D in doubly unexposed Cases of D in those exposed to A Cases of D in those exposed to B Cases of D in double exposed A+ A- B+ B- B+ B- 4 8 8 6 2 6 2 2 100 10/100 100 8/100 4 0.06 2 Total Risk RR RD 100 20/100 2 0.1 100 2/100 A+ A- B+ B- B+ B- 0 8 8 6 2 6 2 2 100 10/100 100 8/100 4 0.06 2 Total Risk RR RD 100 16/100 1.6 0.06 100 2/100 So how to get at interaction? Point 4: It is the absolute scale that tells us about biologic interaction (biologic doesn’t need to be read literally) Point 4a: Since Rothman’s model shows us interdependence is ubiquitous, there is no such thing as “the effect” as it will always depend on the distribution of the complement Interdependence (2) In example, doubly exposed group are low CD4 count who were untreated – Their mortality rate is 130/10,000 Separate this rate into components: – – – – Baseline mortality rate in doubly unexposed (high CD4 count, treated) Effect of low CD4 count instead of high Effect of no treatment instead of treatment Anything else (rate due to interdependence) Interdependence (3) Death Person-years Rate Rate difference Treated CD4 <200 CD4 ≥200 157 81 24531 24545 64/10000 33/10000 31/10000 0 Component 1: – Untreated CD4 <200 CD4 ≥200 18 8 1384 1404 130/10000 57/10000 24/10000 The baseline rate in the doubly unexposed The doubly unexposed = high CD4/treated – Their mortality rate is 33/10,000 Interdependence (4) Death Person-years Rate Rate difference Treated CD4 <200 CD4 ≥200 157 81 24531 24545 64/10000 33/10000 31/10000 0 Component 2: – Untreated CD4 <200 CD4 ≥200 18 8 1384 1404 130/10000 57/10000 24/10000 The effect of exposure 1 (low CD4 vs. high) Calculate as rate difference – – (low - high) in treated stratum Rate difference is 31/10,000 Interdependence (5) Death Person-years Rate Rate difference Treated CD4 <200 CD4 ≥200 157 81 24531 24545 64/10000 33/10000 31/10000 0 Component 3: – Untreated CD4 <200 CD4 ≥200 18 8 1384 1404 130/10000 57/10000 24/10000 Effect of exposure 2 (untreated vs treated) Calculate as rate difference – – (untreated - treated), in unexposed (high CD4) Rate difference is 24/10,000 Interdependence (6) Anything else left over? – Rate in doubly exposed is 130/10,000 – – – Do components add to rate in doubly exposed (low CD4 count, untreated)? component 1 (rate in doubly unexposed): 33/10,000 component 2 (effect of low CD4 vs high): 31/10,000 component 3 (effect of not vs treated): 24/10,000 These 3 components add to 88/10,000 – There must be something else to get to 130/10,000 Interdependence (7) The something else is the “risk (or rate) due to interdependence” between CD4 count and treatment R(E ,C ) R(E ,C ) RD(E ) RD(C ) R(I ) 130 33 31 24 ?[R(I )] 10,000 10,000 10,000 10,000 Interdependence (8) Calculate the rate due to interdependence two ways: R(I ) R (E ,C ) [R (E ,C ) R (E ,C )] [R (E ,C ) R (E ,C )] R (E ,C ) or R(I ) [R (E ,C ) R(E ,C )] [R (E ,C ) R(E ,C )] Component 3 Component 2 Component 1 Interdependence (9) Death Person-years Rate Rate difference Rate Difference Untreated CD4 <200 CD4 ≥200 18 8 1384 1404 130/10000 57/10000 24/10000 73/10000 Treated CD4 <200 CD4 ≥200 157 81 24531 24545 64/10000 33/10000 31/10000 0 31 / 10000 Calculate the rate due to interdependence two ways: 130 24 31 33 42 R(I ) 10,000 10,000 10,000 10,000 10,000 or 73 31 42 R(I ) 10,000 10,000 10,000 Perspective of interdependence With interdependence we care about the joint effect of two actions – – Now we care about: – Action is A+B+, A+B-, A-B+, A-BLeads to four potential outcomes per person Pr(Ya=1,b=1=1) - Pr(Ya=0,b=1=1) = Pr(Ya=1,b=0=1) - Pr(Ya=0,b=0=1) Both actions need to have an effect to have interdependence – Surrogates are not possible Biologic interaction under the CST model: general A study with two binary factors (X & Y), producing four possible combinations: – x=I, y=A; x=R, y=A; x=I, y=B; x=R, y=B Binary outcome (D=1 or 0) – 16 possible susceptibility types (24) Three classes of susceptibility types: – Non-interdependence (like doomed & immune) – Positive interdependence (like causal CST) – Negative interdependence (like preventive CST) Interdependence under the CST model: the non-interdependence class Type (description) 1 (doomed) Risks x=I x=R x=I x=R y=A y=A y=B y=B 1 1 1 1 Effect of x=I within strata of y y=A y=B Difference 0 0 0 4 1 1 0 0 0 0 0 6 1 0 1 0 1 1 0 11 0 1 0 1 -1 -1 0 13 0 0 1 1 0 0 0 16 (immune) 0 0 0 0 0 0 0 Interdependence under the CST model: the non-interdependence class Type (description) 1 (doomed) Risks x=I x=R x=I x=R y=A y=A y=B y=B 1 1 1 1 Effect of x=I within strata of y y=A y=B Difference 0 0 0 4 1 1 0 0 0 0 0 6 1 0 1 0 1 1 0 11 0 1 0 1 -1 -1 0 13 0 0 1 1 0 0 0 16 (immune) 0 0 0 0 0 0 0 The four possible combinations of factors X and Y Interdependence under the CST model: the non-interdependence class Type (description) 1 (doomed) Risks x=I x=R x=I x=R y=A y=A y=B y=B 1 1 1 1 Effect of x=I within strata of y y=A y=B Difference 0 0 0 4 1 1 0 0 0 0 0 6 1 0 1 0 1 1 0 11 0 1 0 1 -1 -1 0 13 0 0 1 1 0 0 0 16 (immune) 0 0 0 0 0 0 0 Strata of Y Interdependence under the CST model: the non-interdependence class Type (description) 1 (doomed) Risks x=I x=R x=I x=R y=A y=A y=B y=B 1 1 1 1 Effect of x=I within strata of y y=A y=B Difference 0 0 0 4 1 1 0 0 0 0 0 6 1 0 1 0 1 1 0 11 0 1 0 1 -1 -1 0 13 0 0 1 1 0 0 0 16 (immune) 0 0 0 0 0 0 0 Indicates whether or not the outcome was experienced For a particular type of subject with that combination of X and Y Interdependence under the CST model: the non-interdependence class Type (description) 1 (doomed) Risks x=I x=R x=I x=R y=A y=A y=B y=B 1 1 1 1 Effect of x=I within strata of y y=A y=B Difference 0 0 0 4 1 1 0 0 0 0 0 6 1 0 1 0 1 1 0 11 0 1 0 1 -1 -1 0 13 0 0 1 1 0 0 0 16 (immune) 0 0 0 0 0 0 0 Example of one susceptibility type Interdependence under the CST model: the non-interdependence class Type (description) 1 (doomed) Risks x=I x=R x=I x=R y=A y=A y=B y=B 1 1 1 1 Effect of x=I within strata of y y=A y=B Difference 0 0 0 4 1 1 0 0 0 0 0 6 1 0 1 0 1 1 0 11 0 1 0 1 -1 -1 0 13 0 0 1 1 0 0 0 16 (immune) 0 0 0 0 0 0 0 Example of one susceptibility type Interdependence under the CST model: the non-interdependence class Type (description) 1 (doomed) Risks x=I x=R x=I x=R y=A y=A y=B y=B 1 1 1 1 Effect of x=I within strata of y y=A y=B Difference 0 0 0 4 1 1 0 0 0 0 0 6 1 0 1 0 1 1 0 11 0 1 0 1 -1 -1 0 13 0 0 1 1 0 0 0 16 (immune) 0 0 0 0 0 0 0 Example of one susceptibility type Interdependence under the CST model: the non-interdependence class Type (description) 1 (doomed) Risks x=I x=R x=I x=R y=A y=A y=B y=B 1 1 1 1 Effect of x=I within strata of y y=A y=B Difference 0 0 0 4 1 1 0 0 0 0 0 6 1 0 1 0 1 1 0 11 0 1 0 1 -1 -1 0 13 0 0 1 1 0 0 0 16 (immune) 0 0 0 0 0 0 0 Example of one susceptibility type Interdependence under the CST model: the non-interdependence class Type (description) 1 (doomed) Risks x=I x=R x=I x=R y=A y=A y=B y=B 1 1 1 1 Effect of x=I within strata of y y=A y=B Difference 0 0 0 4 1 1 0 0 0 0 0 6 1 0 1 0 1 1 0 11 0 1 0 1 -1 -1 0 13 0 0 1 1 0 0 0 16 (immune) 0 0 0 0 0 0 0 Interdependence under the CST model: the positive interdependence class Type (description) 3 Risks Effect of x=I within x=I x=R x=I x=R strata of Y y=A y=A y=B y=B y=A y=B Difference 1 1 0 1 0 -1 1 5 1 0 1 1 1 0 1 7 1 0 0 1 1 -1 2 8 (causal synergism) 1 0 0 0 1 0 1 15 0 0 0 1 0 -1 1 Interdependence under the CST model: the negative interdependence class Type (description) 2 (single plus joint causation by x=1 and y=1) 9 Risks Effect of x=I within x=I x=R x=I x=R strata of y y=A y=A y=B y=B y=A y=B Difference 1 1 1 0 0 1 -1 0 1 1 1 -1 10 0 1 1 0 -1 12 0 1 0 0 -1 14 0 0 1 0 0 0 1 0 1 -1 -2 -1 -1 Assessing biologic interdependence: Are there any cases due to joint occurrence of component causes? CST Types R(E+,C+) = R(I,A) = [p1+p4+p6+p3+p5+p7+p8+p2] The risk in the doubly exposed [R(I,A)] equals: the effect of x=I in y=B R(E+,C-)=R(I,B) = [r1+r6+r13+r5+r2+r9+r10+r14] [R(I,B) - R(R,B] + the effect of y=A in x=R R(E-,C+)=R(R,A) = [q1+q4+q11+q3+q2+q9+q10+q12] [R(R,A) - R(R,B] + the baseline risk R(E-,C-)=R(R,B) = [s1+s11+s13+s3+s5+s7+s15+s9] [R(x=R,y=B)] + the interaction contrast [IC] Solve for IC, assess direction and magnitude IC <?> 0 IC = R(I,A) [R(I,B) - R(R,B)][R(R,A) - R(R,B)][R(R,B)] Require both have 3-way partial exchangeability: IC = (p3+p5+2p7+p8+p15) (p2+p9+2p10+p12+p14) IC is the difference in the sum of positive interdependence CSTs and sum of negative interdependence CSTs Point 5: Remember, lack of additive EMM usually means multiplicative EMM. So interaction in a logistic regression model cannot tell us about additive interaction! Attributable Proportions What proportion of the risk in the doubly exposed can be attributed to each exposure? Baseline Low CD4 Untreated R(I) Total Rate (per 10,000 PY) 33 24 31 42 130 Proportion of total 25.4% 18.5% 23.8% 32.3% 100% Proportion of total Attributable Proportions What proportion of the risk in the doubly exposed can be attributed to each exposure? Baseline Low CD4 Untreated R(I) Total Rate (per 10,000 PY) 33 24 31 42 130 Proportion of total 25.4% 18.5% 23.8% 32.3% 100% Proportion of total 25.4% 50.8% 56.2% 132% Back to Rothman’s Model: Attributable %s don’t need to add to 100% Concept 3: Statistical Interaction Statistical Interaction (1) Most easily understood in regression modeling Write model as effect = baseline + effects of predictor variables (exposure, covariates, and their interaction) outcome intercept b1 exposure b2 covariate b3 exposure covariate Statistical Interaction (2) Where: – – Exposure: 1 = exposed, 0 = unexposed Covariate: 1 = covariate+, 0 = covariate- And: – – – – Intercept is baseline risk or rate b1 is effect of exposure b2 is effect of covariate b3 is the statistical interaction Statistical Interaction (3) Rate or risk model (e.g., linear regression) – Effects are the risk differences b3 is risk due to interdependence 33 rate 10,000 31 exposure 10,000 24 covariate 10,000 42 exposure covariate 10,000 Statistical Interaction (4) Relative risk model (e.g., logistic regression) – – Effects are the log of the relative effects On log scale division becomes addition b3 is departure from MULTIPLICATIVE interaction – – Because deviation from additivity on log scale =relative NO correspondence to R(I) ln(RR) ln(1) ln(1.94) exposure ln(1.73) covariate ln(1.17) exposure covariate Summary: Difference Scale Effect measure modification on the difference scale implies: A non-zero risk due to interdependence [R(I)], because risk due to interdependence equals difference in risk differences A non-zero cross-product term in linear regression models of the risk or rate Nothing about departure from multiplicativity Summary: Relative Scale Effect measure modification on the relative scale implies: A non-zero cross-product term in logistic regression models Nothing about departure from additivity