Summary of “Graphical and Causal Modeling in Genetics and Epidemiology” by Vanessa Didelez Nuala Sheehan Hein Stigum http://folk.uio.no/heins/ May-16 H.S. 1 Graphical Models • Pedigrees – Compute carrier probability • Causal Reasoning – Conditional independence – Two main types • Undirected graphs • Directed acyclic graphs • Medelian randomization May-16 H.S. 2 Introduction “Close to the edge” May-16 H.S. 3 Associations • E and D associated if: E D – E causes D C – C is a common cause of E and D E D C E May-16 D Confounder – C is a common effect of E and D and we condition on C Collider H.S. 4 Marginal and conditional dependence EE+ E and D are: May-16 Disease All C10 % 3% 14 % 3% Marginally dependent H.S. C+ 21 % 21 % Conditionally Independent | C 5 Test of H0 • Can test H0 only if – E and D are conditionally independent, given the variables we adjust for (C) C C E D E D Not OK OK Need graphic tools! May-16 H.S. 6 Graphic tools “We have the moral edge!” May-16 H.S. 7 Undirected Graph Definition: A and B are separated by C if all paths from A to B pass thru C 1 4 2 3 5 Are 1 and 5 separated by 2 ? Are 1 and 5 separated by 3 ? Are 1 and 5 separated by 3 and 4 ? Yes No Yes If 1 and 5 are separated by 2, then 1 and 5 are conditionally independent given 2 May-16 H.S. 8 Example: Chewing tobacco and ulcers Chewing tobacco % Ulcers No 10 Yes 14 Age Young Old Chewing tobacco No Yes No Yes T % Ulcers 3 3 21 21 N 500 500 N 300 200 200 300 T and U are marginally dependent T and U are conditionally independent given A U T and U are separated by A A May-16 H.S. 9 Directed Acyclic Graphs, DAGs 1 3 5 Are 1 and 4 separated by 2 ? 2 No 4 Steps: 1. Take ancestral graph of {1,2,4} 2. Moralize the graph 3. Look for separation May-16 H.S. 10 Directed Acyclic Graphs, DAGs 1 3 2 1 1 Are 1 and 4 separated by 2 ? No 4 3 2 5 Take ancestral graph of {1,2,4} 4 3 Moralize the graph 2 4 Are 1 and 4 separated by 2 ? May-16 H.S. No 11 Estrogen and Endometrial cancer U D E U=unknown uterine abnormality D=endometrial cancer E=estrogen C=vaginal bleeding A=ascertained cancer A C Null hypothesis: E and D independent. Can we test H0? Case-control study, condition on A. Are E and D cond. independent given A? Does it help to adjust for C? U D E May-16 A C 1. Take ancestral graph of {E,D,A} 2. Moralize 3. Separation by A? No 4. Separation by C,A? No H.S. 12 Test of H0 • Can test H0 only if – E and D are conditionally independent, given the variables we adjust for (C) C C E D E D Not OK OK Use tools to verify! May-16 H.S. 13 Cause versus Association • Observe association, not necessarily causal: – Confounding – Reverse causation – Selection effect C E D E D E D S – Time trends T E May-16 H.S. D 14 Definitions • Association C – Observing E predicts D E • Causation – Manipulating E predicts D May-16 H.S. D C do(E) D 15 Problem • Association ≠ causation • Intervention on association may be useless • Randomization not always feasible • Need causal information from observational studies May-16 H.S. 16 Mendelian randomization May-16 H.S. 17 Observation versus trial • U Observational study – • E D Randomized trial 1. 2. 3. – • all measured confounders adjusted for Strong effect, compliance Does not exist Does not exist RD if and only if E causes D U 2 R 1 E D 3 Medelian randomization 1. Gene/ exposure association strong, or large N 2. Should not exist, Mendel’s 2. law 3. Must not exist, depends on the function of the gene – GD if and only if E causes D May-16 H.S. U 2 G 1 E D 3 18 Ex: Alcohol and blood pressure • U Observational study – – Alcohol use increases blood pressure Many ”lifestyle” confounders A BP Alcohol use alcohol ml/day • Gene: ALDH2, 2 alleles – 2,2 type suffer nausea, headache after alcohol – low alcohol regardless of lifestyle (U) 40 30 20 10 0 1,1 • Medelian randomization 1. 2. 3. – Gene ALDH2 is highly associated with alcohol Mendel’s 2. law, no ass. to obs. confounders OK, gene function is known Result: 2,2 type BP +7.4 mmHg G A BP 3 May-16 H.S. 2,2 U 2 1 1,2 Genotype 19 Violations of core conditions 1. Gene/exposure association • • Gene rare or weak effect large N Compensation 2. and 3. Gene independent of U and D • • • Pleiotropy Linkage disequilibrium Population stratification U 2 G 1 E D 3 May-16 H.S. 20