Scale, Causal Pies and Interaction 1h Hein Stigum Presentation, data and programs at: http://folk.uio.no/heins/ May-16 H.S. 1 Agenda • Concepts – Scale – Causal Pies – Interaction and Effect Modification • Methods – Regression and Scale – Regression and Interaction May-16 H.S. 2 SCALE May-16 H.S. 3 30 The importance of scale 20 Females Multiplicative scale Absolute increase Relative increase Females: 30-20=10 Males: 20-10=10 Females: 30/20=1.5 Males: 20/10=2.0 Conclusion: Same increase for males and females Conclusion: More increase for males y, RD RR (OR, HRR) 0 10 Males Additive scale T1 May-16 T2 H.S. 4 Obesity and death RD: The effect of obesity on death increases with age! RR: The effect of obesity on death decreases with age! RR=1.5 D ep re ss ed RR=2.0 Happy clam Thin asasa anail 20 30 40 50 60 70 Age May-16 H.S. 5 Lessons learned • Scale is important – Use both additive and multiplicative – When reporting RR or RD or similar, always report reference risk May-16 H.S. 6 CAUSAL PIES May-16 H.S. 7 Causal pies • Sufficient cause: Three causes for a disease – 1 to 3 (AIDS) • Component cause: Sufficient Cause 1 Sufficient Cause 2 Sufficient Cause 3 A A A – A to F (A=HIV, B=sex, E=injection) • Necessary cause: – A (HIV) C B D B F E • Interaction – A and B (smoke+radoncancer) • Induction time: – time to accumulate A to C (accumulate mutations cancer) • Attributable fraction (AF) – Sum>100% (remove E:33%, remove B:66%, Remove A:100%) May-16 H.S. 8 Pies and Risk of lung cancer N=1000 Cases Risk U Smoke Radon 10 70 20 1% 7% Smoke + - + 13 % 3% Radon 8% 1% 2% Observable? Risk Difference (RD) Radon Smoke May-16 30 3% H.S. 9 INTERACTION, EFFECT MODIFICATION May-16 H.S. 10 Definitions of interaction • Risk factors A and B A B B A • No additive interaction: B1. B A =0 2. RDAB=RDA+RDB 3. RDA is independent of B (and vice versa) • The 3 definitions are identical May-16 H.S. 11 Comparing definitions of no additive interaction risks U U Radon + Smoke Smoke S + - RDsmoke - S+R+U R+U S S+U U S RDradon R R S+R RDsmoke independent of radon Radon Radon Smoke May-16 RDradon independent of smoke RDsmoke,radon = RDsmoke+RDradon R What happens if radon-smoke interaction in not 0? H.S. So far so good! 12 Interaction and scale 1% U Smoke 7% Radon 2% Radon Smoke 0% Smoke RDsmoke RRsmoke + - + 10 % 3% 7% 3.33 Radon 8% 1% RDradon RRradon 2% 2% 1.25 3.00 RRradon dependent of smoke 10.00 RRsmoke,radon RRsmoke*RRradon 9% 7% 8.00 RRsmoke dependent of radon Lesson learned: No additive interaction multiplicative interaction Interaction is scale dependent May-16 H.S. 13 Interaction versus Effect Modification Interaction • Risk factors (Actions) Effect Modification • Variables (No actions) – Smoking – Asbestos – Sex – Age • Two risk factors acting together • The effect of a risk factor modified by a variable smoking and asbestos may act together to produce lung cancer The effect of smoking on heart disease is different for men and women The two definitions are mathematically equivalent, only the type of variable differs Both concepts are scale dependent! May-16 H.S. 14 REGRESSION AND INTERACTION AND SCALE May-16 H.S. 15 Regression and scale • Linear models (linear-regression, -risk, -survival): additive – No interaction if: RDAB=RDA+RDB or RDA is independent of B • “Other” models (logistic, Poisson, log-risk, Cox): multiplicative • No interaction if: RRAB=RRA*RRB May-16 or RRA is independent of B H.S. 16 Estimating interaction in regression Linear model U A U B U A B Observable? y b0 b1 A b2 B b3 AB Effects y A b1 b3 B is independent of B if b3=0 Test b3 0 Interaction if b30 ConfidenceInterval (easy or technical) May-16 H.S. 17 U 1% Smoke 7% Regression example Smoke 2% Radon + - RDsmoke Radon Smoke Radon + 13 % 8% 3% 1% 10 % 7% 3% Linear risk model (all variables=0/1) 𝐿𝑢𝑛𝑔𝐶𝑎𝑛𝑐𝑒𝑟 = 𝑏0 + 𝑏1 𝑠𝑚𝑜𝑘𝑒 + 𝑏2 𝑟𝑎𝑑𝑜𝑛 + 𝑏3 𝑠𝑚𝑜𝑘𝑒 ∙ 𝑟𝑎𝑑𝑜𝑛 𝐿𝑢𝑛𝑔𝐶𝑎𝑛𝑐𝑒𝑟 = 0.01 + 0.07𝑠𝑚𝑜𝑘𝑒 + 0.02𝑟𝑎𝑑𝑜𝑛 + 0.03𝑠𝑚𝑜𝑘𝑒 ∙ 𝑟𝑎𝑑𝑜𝑛 RDsmoke 0.07 0.03radon RDradon ? 0.07 if radon=0 0.10 if radon=1 Stata: margins, dydx(smoke) at(radon=(0 1)) May-16 H.S. 18 Stratify or use interaction term Two models radon=0 radon=1 smoke 7 % 10 % Co 1 Co 2 Co 3 Co 4 const Model with interaction radon=0 radon=1 smoke 7 % 10 % Co 1 Co 2 Co 3 Co 4 const May-16 • Alt 1 : Two models (stratify on radon) – Easy – No test for interaction – Inefficient (12 estimates) • Alt 2: Model with interaction – Technical (ci) – Test for interaction – Efficient (7 estimates) H.S. 19 Summing up 1 • Scale (additive or multiplicative) is important • Causal Pies (SCC) – Multifactorial, Additive May-16 H.S. 20 Summing up 2 • Interaction/ effect modification – Same concept (action*action / action*immutable) – Scale dependent • Regression – Linear models are additive – “All” other models are multiplicative – In both: estimate interaction as product term May-16 H.S. 21