Causal Modeling with TETRAD July, 23-24, 1999 Richard Scheines Dept. of Philosophy Carnegie Mellon University Session 5: Search Algorithms 2 Search for Patterns Adjacency: •X and Y are adjacent if they are dependent conditional on all subsets that don’t include them •X and Y are not adjacent if they are independent conditional on any subset that doesn’t include them Search X1 X3 X2 X4 Independencies entailed??? Search Independencies entailed X1 X3 X2 X4 X1 _||_ X2 X1_||_ X4 | X3 X2_||_ X4 | X3 Search: Adjacency Caus al Graph Independcies X1 X2 X1 X4 {X3} X2 X4 {X3} X1 X3 X4 X2 X1 Begin with: X3 X2 X4 Search: Adjacency Caus al Graph Independcies X1 X2 X1 X4 {X3} X2 X4 {X3} X1 X3 X4 X2 X1 Begin with: X3 X4 X2 From X1 X1 X3 X2 X2 X4 Causal G raph Independcies X1 X1 X3 X4 X2 X2 X1 X4 {X3} X2 X4 {X3} X1 Begin with: X3 X4 X2 From X1 X1 X3 X2 X4 X2 From X1 X4 X1 {X3} X3 X4 X2 From X2 X4 X1 {X3} X3 X2 X4 Search: Orientation Patterns Before OrientationY Unshielded X X Y Z X Z|Y Non-collider Collider X Y Z|Y Z X Y Z X Y Z X Y Z X Y Z Search: Orientation PAGs Y Unshielded X X Y Z X Z|Y Non-collider Collider X Y Z|Y Z X Y Z Search: Orientation Away from Collider Test Conditions X1 X3 * 1) X1 - X2 adjacent, and into X2. 2) X2 - X3 adjacent 3) X1 - X3 not adjacent X2 Test X1 X3 | X2 Yes No X1 * X3 X2 X1 * X3 X2 Search: Orientation After Orientation Phase X1 Pattern X3 X1 X3 X4 X4 X2 X2 X1 || X2 PAG X1 X1 X3 X4 X2 X3 X4 X3 X4 X2 X1 X1 || X4 | X3 X2 || X4 | X3 X1 X3 X2 X4 X2 Search Algorithms in TETRAD 3 PC Algorithm – Input: Independence facts, {time order, required causes, prohibited causes} – Assumes no unmeasured common causes (Causal Sufficiency) – Output: Pattern FCI Algorithm – Input: Independence facts, {time order, required causes, prohibited causes} – Does not assume Causal Sufficiency – Output: PAG Search Algorithms in TETRAD 3: The Build Module User Specified Independence Facts Continuous data Discrete data test 2 test Knowledge: Time order, Causal Suficeincy, etc. Independence Facts Build FCI PAG PC Pattern DEMO Build X1 X2 True DAG X3 X4 Build Create a graph among {X1,X2,X3,X4} Create a SEM model Generate data N=2000 Give data to neighbor Run build twice on data from neighbor – PC – FCI Compare output with neighbor Applications: Regression to select Causes Y = 0 + 1X1 + 2X2 + .....nXn + Causal Interpretation of regression model: Edge from Xi Y just in case i 0. Y = 1X1 + 2X2 + 3X3 + 2= 0 corresponds to: X1 X2 X3 Y Applications: Causal Regression Let the other regressors O = {X1, X2,....,Xi-1, Xi+1,...,Xn} i = 0 if and only if Xi,Y.O = 0 In a multivariate normal distribuion, Xi,Y.O = 0 if and only if Xi || Y | O Applications: Causal Regression Tetrad Adjacency Xi,Y dependent on every subset of other regressors Conditional Independence Orientation Collider etc., Conditional Independence Regression Adjacency X ,Y dependent i the set of all on other regressors Conditional Independence Orientation Prefixed Conditional Independence Applications: Causal Regression T1 X1 X2 T2 X3 Y True Model X1 X2 Y PAG X3 Detecting a Causal Relation 1. From Assuming Z prior to X and Y 2. From Assuming nothing about time order Detecting a Causal Relation 1. From Assuming Z prior to X and Y Z1 Equivalence Class X Y Independence Relations I1) Z1 X I2) Y Z1 | X 1) From time order and I1 - Z1 -- X must be "into X" 2) From Z1--X "into X", and I2, we apply the away from collider orientation rule, and orient X Y Detecting a Causal Relation Equivalence Class Z1 X Y Z2 PAG: Representation of Equivalence Class (Common Causal Features) Z1 X T Y Z2 Z1 X Y Z2 Z1 X Z2 T Y Independence Relations Z1 Y Z2 { Z1 Z2 } | X The Instruments Z1 Equivalence Class X Independence Relations Y Z1 Z2 Y Z1 Instruments X Z1 X Z2 Z1 Z2 X Z2 Z2 { Z1 Z2 } | X The Causal Relation Equivalence Class Z1 X Independence Relations Y Z1 Z2 Y X Z1 X Y Y Z2 Z1 Y|X Z2 Y|X Z2 { Z1 Z2 } | X Detecting a Causal Relation 1. Find a triple Z1, Z2, X s.t. Instruments - Z1_||_ Z2 Z1 - Z1 strongly associated with X X Z2 - Z2 strongly associated with X 2. Find a Y s.t. Z1 - Y strongly associated with X X Z2 - Y _||_ {Z1Z2} | X Y Parallel to Randomized Trials X treatment - Y response We need a Z that is: 1) Into X Z o X 2) No direct connection to Y except through X Parallel to Randomized Trials X treatment - Y response We need a Z that is: 1) Into X : Z o X 2) No direct connection to Z1 X Y except through X: 3) Z _||_ Y | X - no common cause of X - Y. Y ? Sewell and Shaw College Plans 10,318 Wisconsin high school seniors. sex iq = Intelligence Quotient cp = college plans pe = parental encouragement ses = socioeconomic status Variables [male = 0, female = 1] [least = 0, ... highest = 3] [yes = 0, no = 1] [0 = low, 1 = high] [0 = lowest, ... 3 = highest] The Causes of College Plans Questions: 1) Do sex, IQ, socio-economic status, and parental encouragement have any influence on college plans? 2) Does SES influence iq?