Causal inference Hein Stigum Presentation, data and programs at: http://folk.uio.no/heins/talks April 15 H.S. 1 Contents • Background – Error – Bias • Define causal effect – Individual – Average • Identify causal effect – Exchangeability – Positivity – Consistency April 15 H.S. 2 Background April 15 H.S. 3 Error Random error Systematic error • Source: sampling • Source: design • Expressed as: • Expressed as bias: – p-values – Confidence intervals (precision) 1. Selection bias 2. Information bias 3. Confounding • Affect • – All measures Affect: – Frequency measure – Association measure Causality field: Strong focus on bias at the expense of precision Apr-15 H.S. 4 Precision and Bias Precision Bias True value Estimate Precision Bias Causal effect Apr-15 Association H.S. Bias: association causal effect Objective: find effects 5 Define Causal Effects April 15 H.S. 6 Individual causal effect • Counterfactual outcome Treated Not treated Individual causal effect Zeus Died Lived Yes Hera Lived Lived No • Important – Clear definition – Notation mathematical proofs – Notation new methods • Estimate individual effect? – No, but Crossover design April 15 H.S. 7 Individual causal effects Rheia Kronos Demeter Hades Hestia Poseidon Hera Zeus Artemis Apollo Leto Ares Athena Hephaestus Aphrodite Cyclope Persephone Hermes Hebe Dionysus April 15 Treatment No Yes 0 1 1 0 0 0 0 0 0 0 1 0 0 0 0 1 1 1 1 0 0 1 1 1 1 1 0 1 0 1 0 1 1 1 1 0 1 0 1 0 20 subjects 12 individual causal effects H.S. 8 Average causal effect • Counterfactual outcome Treated Population 10/20=0.5 Not treated Average causal effect 10/20=0.5 No • Estimate average effect? – Yes, Randomized controlled trial April 15 H.S. 9 Identify Causal Effects April 15 H.S. 10 Ideal Randomized Trial • Trial – Randomize, Treat, Compare outcomes C • Features E – Exchangeability D • Comparable groups – Positivity • Both treated and untreated – Consistency • Well defined intervention and contrast April 15 H.S. 11 Conditional Randomized Trial • Conditional Trial C – By sex: Randomize, Treat, Compare outcomes E • Features sex – Conditional Exchangeability C • Comparable groups by sex – Conditional Positivity • Both treated and untreated by sex E Males D C – Consistency • Well defined intervention and contrast April 15 D H.S. E Females D 12 Observational study • Observational study Make = conditionally randomized trial C • Need Features – Conditional Exchangeability E D • Comparable groups by all values of C – Conditional Positivity • Both treated and untreated by all values of C – Consistency • Well defined intervention and contrast April 15 H.S. 13 Conditional exchangeability Need to measure all relevant factors Conditional exchangeability = No unmeasured confounding Two ways to remove confounding: Adjust: C E D C Balance: April 15 H.S. E D 14 Balance by Inverse Probability Weights • Weights – Estimate probability of exposed by C = pi • Balance – Weight exposed by 1/ pi – Weight unexposed by 1/(1- pi) , for plot 100/pi , for plot 100/(1-pi) C • Effect April 15 E H.S. D 15 IPW and plots Crude distributions C - smoke Weight: Normal Overweight + E D overweight Blood pressure Effect of E on D: Crude: 0 biased Adjusted: 4 true 50 100 150 Blood pressure mmHg 200 Inverse probability weighted distributions Weight: Normal (mean=113) Overweight (mean=117) Balance the data using IPW Result: all plots of D versus E are adjusted Problem: N gets large 50 April 15 H.S. 100 150 Blood pressure mmHg 200 16 Conditional positivity example All – Estimate dose response for each sex? 20 10 • Positivity problem 0 Response – Dose response is linear 30 40 • Prior knowledge low high Dose Females 30 20 Response 30 20 10 10 0 0 Response 40 40 Males low high Dose low high Dose Conditional positivity E=0 Conditional positivity = exposed and unexposed for all values of C E=1 C positivity 30 40 55 70 E=1 E=0 110 130 Exposure April 15 150 170 250 C>55 150 C=40 to 55 150 90 Parametric assumption: linear “dose response” 200 250 200 250 200 150 C<40 70 D E=1 300 E=0 300 E=1 300 E=0 E 350 350 350 Confounder, C 70 90 110 130 Exposure 150 170 70 90 110 130 Exposure H.S. 150 170 18 Consistency Consistency = Well defined intervention and contrast April 15 H.S. 19 Air pollution Excess mortality from air pollution? Standard method: estimate attributable fraction Implicit contrast: current levels versus zero Implicit intervention: not existent April 15 H.S. 20 Body Mass Index Excess mortality from obesity? Standard method: Implicit contrast: Implicit intervention: April 15 estimate attributable fraction 30 versus <25 Exercise Diet Mortality Smoking H.S. 21 Poorly defined intervention may affect exchangeability • Adjust for lung disease? C C C lung disease lung disease lung disease E D E D E D exercise mortality diet mortality smoking mortality Adjust April 15 Need not adjust H.S. Should not adjust 22 Poorly defined intervention may affect positivity • Confounder status unknown – Can not asses positivity April 15 H.S. 23 Summing up • Defined bias – Objective: find effects • Conditions to find effects – Exchangeability: – Positivity: – Consistency: April 15 comparable E+ and EE+ and E- in all strata well defined intervention and contrast H.S. 24 Litterature • Hernan and Robins, Causal Inference April 15 H.S. 25