DAGs intro, Answers Hein Stigum http://folk.uio.no/heins/ courses 29. mai. H.S. 1 Report variables as equals? Variable OR Comments (surprises) Diabetes 2 2.0 Physical activity 1.2 Protective in other studies? Obesity 1.0 No effect? Bone density 0.8 1. P is a confounder for E→D, but is E a confounder for P→D? No, E is in the causal path P→ E→ D 2. Which effects are reported correctly in the table? P physical activity E Only the effects of E and B! P has indirect effect P→ E→ D O has indirect effects O→ E→ D and O→ B→ D D fractures diabetes 2 O B obesity bone density DAGs dictate one exposure and one outcome Report adjusted effect of exposure May-16 H.S. 2 Stratified analysis (Simpson’s paradox) Population D A 1 0 C=0 D 1 0 sum 210 150 290 350 500 500 1000 RR= 1.4 40% more disease if treated risk 0.4 0.3 A 1 0 C=1 D 1 0 sum risk 10 70 90 330 100 400 500 0.10 0.18 RR= 0.6 40% less disease if treated C A 1 A0 1 0 sum risk 200 80 200 20 400 100 500 0.50 0.80 RR= 0.6 40% less disease if treated C D A C is a collider, not stratify, RR=1.4 D C is a confounder, stratify, RR=0.6 C could also be a mediator, RR=? May-16 H.S. 3 Physical activity and Coronary Heart Disease (CHD) 1. Total effect: adjust for sex and age Unconditional Path 1 ED 2 EC1D 3 EC2D Conditioning on C1 and C2 Path 1 ED 2 EC1]D 3 EC2]D May-16 Type Causal Non-causal Non-causal Status Open Open Open Bias What if sex does not affect physical activity? Type Causal Non-causal Non-causal Status Open Closed Closed H.S. No bias 4 Tea and depression O coffee 1. Paths C 2. Total effect: adjust for O caffeine 3. Direct effect: adjust for C (and O) E tea Path 1 E→D 2 E→C→D 3 E←O→C→D D depression Type Status Causal Open Causal Open Non-causal Open 4. Caffeine is both intermediate and part of a confounder path. direct indirect total Questions: Coffee → tea or tea → coffee? Controlled direct effect! May-16 H.S. 5 Statin and CHD 1. 2. 3. 4. Paths: Total effect: no adjustments Direct effect: not possible C intermediate or collider: intermediate (path 2) and collider (path 3) Unconditional Path 1 ED 2 ECD 3 ECUD Type Causal Causal Non-causal Status Open Open Closed Conditioning on C Path 1 ED 2 EC]D 3 EC]UD Type Causal Causal Non-causal Status Open Closed Open May-16 H.S. no bias bias! Questions: RandomizedControlledTrial? Subgroup analyses? 6 Exercise: nested counterfactuals 1. A set to 1, M set to m D1,m 2. A set to 0, M set to m D0,m 3. A set to 1, M set to M0 D1,M0 4. A set to 0, M set to M0 D0,M0 May-16 H.S. 7 Exercise: Labor marked discrimination 1. Describe the experiment for estimating the controlled direct effect: – Draw names, use standard CV 2. Describe the experiment for estimating the natural direct effect: – CV Job name Draw names, draw CV from sample of “white CVs” (assuming white is “untreated”) May-16 H.S. 8 Hair dye and congenital malformations E* Reported dye ? + E + True dye D malformation E as exposure Path 1 ED 2 EE*D Type Status Causal Open Non-causal Closed E* as exposure Path 1 E*ED 2 E*D Type Status Non-causal Open Non-causal Open True E* No bias Bias, but Bias from E 4) Yes 0 29. mai. H.S. 9 Survival bias S R survival risk E D early exposure later disease Paths: ED Causal E[S]RD Non-causal Open Open Conclusion: Have survival bias Must adjust for R to remove the bias 29. mai. H.S. 10 U RCT E R + - + 85 0 85 15 100 115 D R + - + 43 63 106 57 37 94 D E + - + 32 74 106 53 41 94 N Risk RD 100 100 200 0.85 0.00 0.85 0 N Risk RD 100 100 0.43 0.63 -0.20 0 N Risk RD 85 115 0.38 0.64 -0.27 0 R c E c compliance 0.22 D IVe ITTe 𝐼𝑉 = Negative bias from confounding 29. mai. -0.15 𝐼𝑇𝑇 𝑐 = −0.2 0.85 = −0.24 ITT ITT effect PP PP effect ITT always “weaker” than IV H.S. 11 Exercise: causal pies 1. Causal pies for hospital 2. Selection bias: Sufficient causes for Hospital: 1) or 2) and 3) both Selection bias: Question Variable with only 1 value? May-16 1) 2) 3) H.S. Negative bias Positive bias ? 12 Exercise: Dust and COPD Chronic Obstructive Pulmonary Disease D0 S cur. worker diseases H health COPD risks: Health E0 prior dust E cur. dust good poor Dust low high 5 % 10 % 10 % 20 % RRdust RDdust 2.00 5 % 2.00 10 % D COPD 1. No interaction, can we still have selection bias? 2. Paths: ED E E0D0[S]HD Causal Non-causal Open Open 3. Crude RR=0.7. Adjust for H, Health: True RR=2. 4. Healthy worker bias (selection bias concept 2) 5. No interaction based selection bias, no interaction (on RR scale, but on the RD). May-16 H.S. 13 Dust and COPD 1. 2. 3. 4. Paths: Total effect: no adjustments Direct effect: not possible C intermediate or collider: intermediate (path 2) and collider (path 3) Unconditional Path 1 ED 2 ECD 3 ECUD Type Causal Causal Non-causal Status Open Open Closed Conditioning on C Path 1 ED 2 EC]D 3 EC]UD Type Causal Causal Non-causal Status Open Closed Open May-16 H.S. no bias bias! Questions: RandomicedControlledTrial? Subgroup analyses? 14 Exercise: M-structure A B C E D A B C E 1. Show the paths 2. Should we adjust for C? 3. If the design implies a selection on C, what would you call the resulting bias: selection bias or confounding? 1. Paths: (No adjusting) ED Causal EACBD Non-causal Open Closed No bias 2. Paths: (Conditioning on C) ED Causal EA[C]BD Non-causal Open Open bias 3. We are conditioning on a collider, hence: selection bias! D 29. mai. May-16 H.S. 15 Exercise: Collider stratification H Hospital risk: D 1 E 0 2.0 1 0 0.6 0.3 0.2 0.1 Response= 16 % Population D E 1 0 0 sum 36 164 64 736 100 900 1000 RR= 2.0 May-16 E Hospital D 1 E 3.0 1 0 D 2.0 No hospital D 1 0 sum 22 49 13 74 35 123 157 RR= 1.6 E 1 0 1 0 sum 15 115 51 663 65 777 843 RR= 1.5 Selection bias H.S. Collider stratification bias 16 Oestrogen and endometrial cancer Paths from E (oestrogen) to C (cancer) Unconditional Path 1 EBAC 2 EBC Type Status Non-causal Closed Non-causal Closed Conditioning on B Path Type Status 1 E[B]AC Non-causal Closed 2 E[B]C Non-causal Open No bias collider bias Confounding between mediator and outcome Paths from X to Y: Unconditional Path 1 XY 2 XZY 3 XZUY Type Causal Causal Non-causal Status Open Open Closed Conditioning on Z Path 1 XY 2 XZ]Y 3 XZ]UY Type Causal Causal Non-causal Status Open Closed Open total effect direct effect collider bias Dust and lung disease S C C.worker health E D dust lung disease Paths: ED Causal E[S]CD Non-causal Open Open • Selection bias (collider stratification) • Healthy worker effect, can show protective effects S C C.worker health E D dust lung disease May-16 • Selection bias (interaction based) Good health: Bad health: RRED = 2.0 RRED = 3.0 Interaction • Can not show protective effect H.S. 19 Dust and lung disease S C C.worker health E D dust lung disease • Selection bias (collider stratification) • Healthy worker effect, can show protective effect • Can show untrue effects • Adjust for health to remove bias S C C.worker health E D dust lung disease May-16 • Selection bias (interaction based) Good health: Bad health: RRED = 1.0 RRED = 2.0 Interaction • Can not show protective effect • The E-D arrow is not required • Can only show “true effects” H.S. 20 Exercise: Simpson’s paradox Want the effect of treatment (T) on disease (D) Both T and D affect blood pressure (Bp) 1. Draw the DAG 2. Calculate the population effect of T on D 3. Conclusions? Bp T D Population D T 1 0 Low blood pressure D 1 0 sum 70 230 130 570 200 800 1000 RR= 1.2 risk 0.4 0.3 20% more disease if treated Unbiased May-16 T 1 0 1 0 sum risk 65 180 45 90 110 270 380 0.6 0.7 RR= 0.9 10% less disease if treated Collider bias H.S. High blood pressure D T 1 0 1 0 sum risk 5 50 85 480 90 530 620 0.06 0.09 RR= 0.6 40% less disease if treated Collider bias 21