Comparing covariate adjustment in interventional and observational studies Markus Kalisch, Seminar fΓΌr Statistik, ETH ZΓΌrich
What is the total causal effect ? Treatment π Outcome π ο§ If we apply treatment π , how will outcome π change ? ο§ Data collection: - observational study - interventional study (RCT) 1
Outline for the rest of the talk ο§ Total causal effect and covariate adjustment ο§ Issues in observational studies ο§ Issues in interventional studies ο§ Insights from recent theoretical developments 2
Causal Model: How the real world might look like ο§ We use directed acyclic graphs (DAG) β no feedback loops ο§ Example: DAG π» π 5 π 2 π 3 π π 1 ο§ Terminology: π 4 Set of all variables: π = {π 1 , π 2 , β¦ , π 5 , π} Path: π 1 , π 2 , π 3 , π Directed p ath = βcausal - pathβ: π 1 , π 2 , π 3 Not directed path = Non-causal path: (π 4 ,π 3 ,π) Parents pπ π 3 = {π 2 , π 4 } , Children cβ π 1 = π 2 Think of family tree Ancestor ππ , Descendant ππ , Non-descendants ππ 3
More details: Structural Equation Model (SEM) ο§ Example of SEM: π 1 = π 1 Causal π 2 = 4π 1 + π 2 interpretation π 1 , π 2 βΌ π 0,1 πππ ο§ Visualization of causal structure : π 1 π 2 ο§ Difference to arbitrary hierarchical system of equations: Due to causal interpretation, solving for a variable on the RHS is not meaningful in SEM. 4
Quantifying the total causal effect Define intervention distribution by replacing (some) structural equations ο§ do-Operator Reference: Pearl, J. (2009). Causality: Models, Reasoning and Inference. 2nd edition. Cambridge Univ. Press. E.g. Β«intervention on π Β»: ο§ Old SEM: π with equation π = 2 + π 5 + π π ο§ New SEM: α π with equation π = 4 ο§ New SEM generates new distribution: π α π (π) = π π (π|ππ π = 4 ) and in particular P(π|ππ(π = 4)) ο§ Final goal: Estimate intervention distribution given observational data ο§ Oftentimes: Expectation is enough β e.g. πΉ(π|ππ π = 4 ) 5
Covariate adjustment: Adjustment set ο§ Idea: Identify intervention effects by only using conditional probabilities / expectations Β«doΒ» No Β«doΒ» πΆ π π = π§ ππ π = π¦ = ΰ· π π = π§ π = π¦, πΆ = π π(πΆ = π) Adjustment πβπΆ π π set ο§ Practice: Often interested in πΉ π = π§ ππ π = π¦ ο§ Can show for multivariate Gaussian density: = π½ + πΏπ¦ + πΎ π πΉ πΆ πΉ π ππ π = π¦ d ο§ Total Causal Effect: ππ¦ πΉ π ππ π = π¦ = πΏ This is the regression coefficient of π in the regression of π on π and πΆ 6
Outline for the rest of the talk ο§ Total causal effect and covariate adjustment ο§ Issues in observational studies ο§ Issues in interventional studies ο§ Insights from recent theoretical developments 7
Causal Diagram: Example 1 - confounder Lab 1,2,β¦ Treatment Outcome Should we add the lab information as covariate ? 8
Example 1 in numbers π π π ο§ π π βΌ π(0,1) , π π βΌ π(0,1) , π π ~ π(0,1) independent ο§ True causal system: π = π π π = 0.7 β π + π π π = 1 β π + 0.5 β π + π π ο§ True causal effect of π on π : 1 If we increase π by one unit, π will also increase by one unit ο§ Can we estimate the true causal effect with a linear regression ? 9
Example 1 in numbers π π π ο§ True causal effect of π on π : 1 ο§ Simple Regression: ππ(π ~ π) Incorrect ο§ Multiple Regression: ππ(π~π + π) Missing the confounder introduced a bias ! Correct 10
Causal Diagram: Example 2 β selection variable Treatment Outcome Follow-up test Should we add the info of the follow-up test as covariate ? 11
π π Example 2 in numbers π ο§ π π βΌ π(0,1) , π π βΌ π(0,1) , π π ~ π(0,1) independent ο§ True causal system: X = π π Y = 0.7 β π + π π Z = 0.8 β π + 0.5 β π + π π ο§ True causal effect of π on π : 0.7 If we increase π by one unit, π will also increase by 0.7 units ο§ Can we estimate the true causal effect with a linear regression ? 12
π π Example 2 in numbers π ο§ True causal effect of π on π : 0.7 ο§ Simple Regression: ππ(π ~ π) Correct ο§ Multiple Regression: ππ(π~π + π) Including the selection variable Incorrect introduced a bias ! 13
βParent Criterionβ (PC) ο§ Take parents of π as adjustment set (special case of Pearlβs back -door criterion) ο§ Sufficient but not complete ο§ Example 1: π π π PC: π is a valid adjustment set; would {} be a valid adjustment set, too β ??? (perhaps we can not measure π although we know it exists) ο§ Example 2: π π π PC: {} is a valid adjustment set; would π be a valid adjustment set, too β ??? 14
Conclusion 1 In observational studies: Judging if an adjustment set is valid is not trivial 15
Outline for the rest of the talk ο§ Total causal effect and covariate adjustment ο§ Issues in observational studies ο§ Issues in interventional studies ο§ Insights from recent theoretical developments 16
RCT: Evaluation Treatment Treatment Control Control Treatment Control Treatment Control Treatment Control ο§ Cage: Experimental Unit ο§ 5 cages with treatment ( π = 1 ), 5 cages with control ( π = 0 ) ο§ Randomize allocation: In causal diagram think of βdeleting all incoming edges to π β 17
RCT in causal diagram π 1 π 1 π π΅ π π π΅ π RCT π· π· π 2 π 2 πΆ πΆ PC: PC: Valid adjustment set Valid adjustement set is {} is {π 1 , π 2 } PC: {} is always valid adjustment set after randomization 18
RCT: Evaluation Treatment Treatment Control Control Treatment Control Treatment Control Treatment Control ο§ Given a proper design, we can do a two-sample t-test with two groups (i.e. empty adjustment set). ο§ What if we have more covariates (sex, age, intermediate blood test, follow-up information, β¦) ? ο§ Is it always better to add covariates to the analysis ? 19
Messing up the evaluation of a randomized controlled trial (RCT) ο§ You can bias ( βmess upβ ), the analysis by adding the βwrongβ covariates . ο§ RCT : It is always safe to add no covariates to the analysis. ο§ Adding the βrightβ covariates might increase precision. 20
Causal Diagram: Example 1 Treatment Intermediate Blood T est Outcome Should we add the intermediate blood test as covariate ? 21
Example 1 in numbers π π π ο§ π π βΌ π(0,1) , π π βΌ π(0,1) , π π ~ π(0,1) independent ο§ True causal system: π = π π π = 2 β π + π π π = 0.5 β π + π π ο§ True causal effect of π on π : 2 β 0.5 = 1 If we increase π by one unit, π will also increase by one unit ο§ Can we estimate the true causal effect with a linear regression ? 22
Example 1 in numbers π π π ο§ True causal effect of π on π : 2 β 0.5 = 1 ο§ Simple Regression: ππ(π ~ π) Correct ο§ Multiple Regression: ππ(π~π + π) Adding a covariate introduced a bias ! Incorrect 23
Causal Diagram: Example 2 Lab 1,2,β¦ Treatment Outcome Should we add the lab information as covariate ? 24
Example 2 in numbers π π π ο§ π π βΌ π(0,1) , π π βΌ π(0,1) , π π ~ π(0,1) independent ο§ True causal system: π = π π π = π π π = 1 β π + 0.5 β π + π π ο§ True causal effect of π on π : 1 If we increase π by one unit, π will also increase by one unit ο§ Can we estimate the true causal effect with a linear regression ? 25
Example 2 in numbers π π π ο§ True causal effect of π on π : 1 ο§ Simple Regression: ππ(π~π) Correct ο§ Multiple Regression: ππ(π~π + π) β’ Adding a covariate did not introduce a bias Correct β’ Confidence interval with covariate is slightly smaller (0.12 vs 0.14) 26
Summary ο§ Adding the wrong variable will introduce a bias β Wrong variable β: On causal path from π to π or Β«descendantsΒ» of those nodes ( post-intervention ) ο§ Adding the right variables might increase precision π 1 π 2 β Right variable β: Parents of nodes on causal path from π to π ( pre-intervention ) π π΅ π π· ο§ Problem in practice: πΆ Usually donβt know true causal structure! What are βrightβ and βwrongβ variables ? ο§ If in doubt, donβt use covariate ! ο§ Safe variables: Things that clearly βprecededβ π (e.g. gender) 27
Outline for the rest of the talk ο§ Total causal effect and covariate adjustment ο§ Issues in observational studies ο§ Issues in interventional studies ο§ Insights from recent theoretical developments 28
Adjustment Criteria Getting the βright estimateβ: ο§ given causal structure, criterion to check if a set is a valid adjustment set ο§ assuming causal structure is a strong assumption in practice ο§ discussion can shift to discussing reasonable causal structures ο§ Pearlβs back -door criterion ο§ Generalized Adjustment criterion 29
Recommend
More recommend