Confounding and DAGs (Directed Acyclic Graphs) Hein Stigum Apr-20 H.S. 1 Agenda • Confounder • Collider Causal Knowledge as a Prerequisite for Confounding Evaluation… Hernán et al. AJE, 2002 Apr-20 H.S. 2 CONFOUNDER DEFINITION Apr-20 H.S. 3 Associations • E and D associated E D – E causes D C – E and D have common cause E D C – Both E Apr-20 D • Overall E-D association = spurious effect from C + causal E-D effect H.S. 4 Adjusting is OK CONFOUNDER EXAMPLES Apr-20 H.S. 5 Classic confounder C is the confounder RRED is biased RRED|C is unbiased C E D age + + vitamins birth defects RRED =0.8 positive bias, towards the null Adjust for age: RRED|C=0.5 is unbiased true biased 0 Apr-20 1 H.S. 6 Mediated confounder C U E D obesity + + age + vitamin birth defects Apr-20 U is the confounder C is a confounder mediator Effect on D is mediated RRED is biased RRED|C is unbiased RRED =0.8 positive bias, towards the null Adjust for obesity: RRED|C=0.5 is unbiased H.S. 7 Mediated confounder 2 U C E D gene + + earlier + birth vitamin malform defects Apr-20 U is the confounder C is a confounder mediator Effect on E is mediated RRED is biased RRED|C is unbiased RRED =0.8 positive bias, towards the null Adjust for earlier malformations: RRED|C=0.5 is unbiased H.S. 8 Marker for confounder U E D C U is the confounder C is a confounder marker RRED is biased RRED|C is less biased + gene - folate - neural tube + enzyme RRED =0.3 negative bias, away from the null Adjust for enzyme: RRED|C=0.5 is less biased biasedtrue 0 Apr-20 1 H.S. 9 COLLIDER DEFINITION Apr-20 H.S. 10 Definition • Confounder C – Common cause for E and D E D C • Collider – Common effect of E and D Apr-20 H.S. E D 11 Definition cont. • Classic E D C + diet cancer + weight loss Apr-20 C is a collider, no confounding RRED is unbiased RRED|C is biased • Example RRED =1.0 cancer not diet related Positive collider=negative bias? Adjust for weight loss: RRED|C=0.8 is biased H.S. 12 Adjusting is not OK COLLIDER EXAMPLES Apr-20 H.S. 13 Classic collider E D C folate - preterm + LBW Apr-20 C is a collider, There is no confounding RRED is unbiased RRED|C is biased RRED =0.5 is unbiased Negative collider=positive bias? Adjust for low birth weight (LBW): RRED|C=0.8 is biased H.S. 14 Shared cause U E D C Low age + vitamin - Apr-20 birth + LBW defects C is a collider of U and D C and E has a shared cause RRED is unbiased RRUD|C is biased RRED|C is biased RRED =0.8 is unbiased Adjust for low birth weight (LBW): Positive bias, towards the null? RRED|C=1.0 is biased H.S. 15 Shared cause 2 U E D C C is a collider of U and E C and D has a shared cause RRED is unbiased RREU|C is biased RRED|C is biased gene + vitamin - + birth defects + Apr-20 RRED =0.8 is unbiased maternal Adjust for maternal weight gain: weight Negative bias, away from the null? gain RRED|C=0.5 is biased H.S. 16 Folate-neural tube study • Case Control study – Exposure: – Disease: – ORED=0.65 folate intake neural tube defects Should we restrict analysis to live births? – ORED|C=0.80 Apr-20 H.S. 17 Graph 1, collider folate - neural tube + stillbirth Apr-20 RRED =0.65 is unbiased Adjust for stillbirth: Positive bias, towards the null RRED|C=0.8 is biased H.S. 18 Graph 2, confounder marker + RRED =0.65 is biased stillbirths + folate neural tube Adjust for stillbirth: Positive bias, towards the null + stillbirth RRED|C=0.8 is unbiased Opposite conclusion as graph 1 Apr-20 H.S. 19 SUMMING UP Apr-20 H.S. 20 Confounder versus Collider • Confounder – Common cause for E and D – or some variant thereof – Adjust C E D • Collider – Common effect of E and D – or some variant thereof – Not adjust Apr-20 H.S. C E D 21 Collider • Common effect of: E D C E and D U E D C cause of E and D U E C D E and cause of D Reasons not to adjust for C • C is a collider • C is a weak confounder and – C has missing – C has errors – C is highly correlated with other cofactors Apr-20 H.S. 23 The birth weight “paradox” uncovered, Hernández-Díaz at al. AJE, 2006 COLLIDER EXAMPLE Apr-20 H.S. 24 Birth weight distribution Apr-20 H.S. 25 Infant mortality Apr-20 H.S. 26 To adjust or not adjust • Crude – RRsmoke=1.55 (1.50, 1.59) • Adjusted for birth weight – RRsmoke=1.09 (1.05, 1.12) Should we adjust for birth weight? Apr-20 H.S. 27 A likely DAG U E C D C is a collider of U and E C and D has a shared cause RRED is unbiased RREU|C is biased RRED|C is biased U + + U: malnutrition, malformation smoke + LBW + mort + Apr-20 H.S. 28 DAG simplified U + + smoke + LBW mort RRED is unbiased =1.0 RREU|C is biased <1.0 RRED|C is biased <1.0 Negative bias U + + Same direction of bias smoke + LBW + mort + Apr-20 biased 0 H.S. 1.09 true 1.55 29 Summing up • Confounding smoke yellow fingers – “Yellow fingers”-”lung cancer” association is useful – not causal lung cancer • Collider U + + smoke + LBW + mort + Apr-20 – Among Low Birth Weight children “smokers” do better, less likely that the cause of LBW is malformation – “smoking” does not protect against mortality H.S. 30