1 / 23

SIMPSON’S PARADOX, ACTIONS, DECISIONS, AND FREE WILL

SIMPSON’S PARADOX, ACTIONS, DECISIONS, AND FREE WILL. Judea Pearl UCLA www.cs.ucla.edu/~judea. SIMPSON’S PARADOX (Pearson et al. 1899; Yule 1903; Simpson 1951). Any statistical relationship between two variables may be reversed by including additional factors in the analysis.

steeleg
Download Presentation

SIMPSON’S PARADOX, ACTIONS, DECISIONS, AND FREE WILL

An Image/Link below is provided (as is) to download presentation Download Policy: Content on the Website is provided to you AS IS for your information and personal use and may not be sold / licensed / shared on other websites without getting consent from its author. Content is provided to you AS IS for your information and personal use only. Download presentation by click this link. While downloading, if for some reason you are not able to download a presentation, the publisher may have deleted the file from their server. During download, if you can't get a presentation, the file might be deleted by the publisher.

E N D

Presentation Transcript


  1. SIMPSON’S PARADOX, ACTIONS, DECISIONS, AND FREE WILL Judea Pearl UCLA www.cs.ucla.edu/~judea

  2. SIMPSON’S PARADOX (Pearson et al. 1899; Yule 1903; Simpson 1951) • Any statistical relationship between two variables may be reversed by including additional factors in the analysis. Application: The adjustment problem • Which factors should be included in the analysis.

  3. SIMPSON’S PARADOX (1951 – 1994) M R R T 18 12 30 T 7 3 10 25 15 40 M R R T 2 8 10 T 9 21 30 11 29 40 R R T 20 20 40 T 16 24 40 36 44 80 + = 0.6 < 0.7 0.2 < 0.3 0.5 > 0.4 T – Treated T – Not treated R – Recovered R– Dead M – Males M – Females Easy question (1950-1994) • When / why the reversal? Harder questions (1994) • Is the treatment useful? Which table to consult? Why is Simpson’s reversal a paradox?

  4. SIMPSON’S REVERSAL Group behavior: Pr(recovery | drug, male)<Pr(recovery | no-drug, male) Pr(recovery | drug, female)<Pr(recovery | no-drug, female) Overall behavior: Pr(recovery | drug)>Pr(recovery | no-drug)

  5. TO ADJUST OR NOT TO ADJUST? Gender Treatment Mediating factor Treatment F T T F Recovery Recovery R R Solution: Adjust iff F blocks all back-door paths

  6. Surprise surfaces only when we speak about“efficacy,”not aboutevidencefor recovery. THE INEVITABLE CONCLUSION: THE PARADOX STEMS FROM CAUSAL INTERPRETATION TWO PROOFS: • When two causal models generate thesame statistical dataand in one we decide to use the drug yet in the other not to use it, our decision must be driven by causal and not by statistical considerations. Thus, there is no statistical criterion to warn us against consulting the wrong table. • CanTemporalinformation help? A.No!,see Figure (c). F = Car Type

  7. WHY TEMPORAL INFORMATION DOES NOT HELP Treatment Treatment Treatment T T T T F F F F Gender Blood Pressure Car R R R R Recovery Recovery Recovery (a) (b) (c) (d) In (c),F may occur before or after T, and the correct answer is to consult thecombinedtable. In (d),Fmay occur before or after T, and the correct answer is to consult theF-specifictables.

  8. WHY SIMPSON’S PARADOX EVOKES SURPRISE • People thinkcauses, notproportions. • "Reversal" ispossiblein the calculus ofproportionsbutimpossiblein the calculus ofcauses.

  9. CAUSAL CALCULUS PROHIBITS REVERSAL Group behavior: Pr(recovery | drug, male)<Pr(recovery | no-drug, male) Pr(recovery | drug, female)<Pr(recovery | no-drug, female) do{drug} do{no-drug} do{drug} do{no-drug} Assumption: Pr (male |do{drug}) = Pr (male |do{no-drug}) Entailed overall behavior: Pr(recovery | drug)<Pr(recovery | no-drug) do{drug} do{no-drug}

  10. THE SURE THING PRINCIPLE Theorem 6.1.1 AnactionA that increases the probability of an event E ineach subpopulationmust also increase the probability of E inthe population as a whole, provided that the action does not change the distribution of the subpopulations.

  11. R 0.7 R R R T T T DECISION TREE FOR F = Gender F Gender R Male 0.5 Female 0.5 (a) F T T R 0.6 R 0.7 R 0.2 R 0.3 0.3 0.2 0.35 0.15 0.1 0.4 0.15 0.35 0.3 < 0.35 0.1 < 0.15

  12. R 0.4 R 0.3 R 0.8 T T DECISION TREE FOR F = Blood Pressure F Blood Pressure R T (b) -HBP 0.75 HBP 0.25 -HBP 0.25 HBP 0.75 F F R 0.6 R 0.2 R 0.7 E 0.3 -E 0.7 0.45 0.2 0.05 0.2 0.175 0.075 0.225 0.525 > 0.5 0.4

  13. R 0.5 R 0.6 T T COMPRESSED TREE FOR F = Blood Pressure F Blood Pressure R (b) T R 0.5 R 0.4 0.5 > 0.4

  14. WHAT DECISION TREE SHOULD WE BUILD FOR F = Car ? ??? Treatment T Should F = Car be handled like F = Gender or F = Blood Pressure F Car R Recovery (c) ???

  15. Treatment R 0.7 R R R T T Wrong 0.6 0.3 TEMPORAL ORDER IS DECEPTIVE T F Car R Expensive 0.5 Cheap 0.5 Recovery F (c) T T R 0.7 R 0.2 R 0.3 R 0.6 0.3 0.2 0.35 0.15 0.1 0.4 0.15 0.35 0.3 < 0.35 0.1 < 0.15

  16. Treatment R 0.5 R 0.6 T CORRECT DECISION TREE T F Car R Recovery (c) T R 0.5 R 0.4 0.5 > 0.4

  17. Treatment R 0.7 R R R T T Wrong 0.6 0.3 WHAT WENT WRONG WITH F = Car ? T F Car R Expensive 0.5 Cheap 0.5 Recovery F (c) T T R 0.7 R 0.2 R 0.3 R 0.6 0.3 0.2 0.35 0.15 0.1 0.4 0.15 0.35 0.6 =P(R | T,F) P(R | do(T), F)

  18. R 0.7 R R R T T Right 0.6 0.3 T WHAT WENT RIGHT WITH F = Gender ? F Gender R Male 0.5 Female 0.5 (a) F T T R R 0.7 R 0.2 R 0.3 0.2 0.35 0.15 0.1 0.4 0.15 0.35 0.6 =P(R | T,F) =P(R | do(T), F)

  19. THE PERCEPTION OF FREE WILL Compelled Act Free Act (Explained reaction) (Decision) Treatment Treatment T T F F Car Car R R Recovery Recovery (c) (c)

  20. DISAMBIGUATING THE TWO ROLES OF ACT Whatever evidence an act might provide On motives that cause such acts, Should never be used to help one decide On whether to choose that same act. (p. 109)

  21. God Physics Physics God THE PARADOXOF FREE WILL Compelled Act Free Act (Explained reaction) (Decision) Treatment Treatment T T F F Car Car R R Recovery Recovery (c) (c)

  22. COMPUTATIONAL RESOLUTION OF THE PARADOX OF FREE WILL Free will is an illusion that demands a computational model, to explain: • What mental activity generates the illusion of acting freely? • Why/how is such illusion generated from the activity above? • How can we survive the distortion generated by the illusion? Is the distortion harmless? Is the illusion dispensable? • What is the computational usefulness of the illusion?

  23. CONCLUSIONS • Simpson's Paradox is resolved through the calculus of causation. • People think causally, not probabilistically • Decision trees cannot replace causal models • Free-will is a software illusion to be explained in computational terms.

More Related