I find EY’s main points very convincing and helpful. After reading this and the follow-on thread, my only nit is that using the suspension-of-relations question as one of the examples seems pedagogically odd, because perfectly rational (OK, bounded-rational but still rational) behavior could have led to the observed results in that case.
The rational behavior that could have led to the observed results is that participants in the second group, having been reminded of the “invade Poland” scenario, naturally thought more carefully about the likelihood of such an invasion (and/or the likelihood of such an invasion triggering suspension), and this more careful thinking caused them to assign a higher probability to the invasion-then-suspension scenario (thus also to the invasion-and-suspension scenario) than they would have assigned to the “suspension” scenario if instead asked Question 1 (which mentions only suspension).
Why? For the simple reason that Question 2 tended to provide them with new information (namely, the upshot of the additional careful thought about the Polish invasion scenario) that Question 1 wouldn’t have.
(To caricature this, imagine showing two separate groups of chess beginners the same superficially-even board position with Player A on move, asking Group 1 participants “what’s the probability that A will win,” and separately asking Group 2 participants “what’s the probability that A will make slightly-tricky-advantageous-move-X and win”? Yes, the event Group 2 was asked about is less likely than the event Group 1 was asked about; Group 2′s answers may nevertheless average higher for quite rational reasons.)
I agree. This notion of question 2 providing a plausible cause that might lead to suspension v. question 1 where the test subject has to conceive of their own cause is a bias, but a different type of bias, not a conjunction fallacy. There could be (and possibly have been) ways to construct the test to control for this. For example, there are 3 test groups where 1 and 2 are the same and for the third, the two events are asked independently:
What are the probabilities of each event:
A. That USSR invades Poland, or
B. That US suspends relations
I find EY’s main points very convincing and helpful. After reading this and the follow-on thread, my only nit is that using the suspension-of-relations question as one of the examples seems pedagogically odd, because perfectly rational (OK, bounded-rational but still rational) behavior could have led to the observed results in that case.
The rational behavior that could have led to the observed results is that participants in the second group, having been reminded of the “invade Poland” scenario, naturally thought more carefully about the likelihood of such an invasion (and/or the likelihood of such an invasion triggering suspension), and this more careful thinking caused them to assign a higher probability to the invasion-then-suspension scenario (thus also to the invasion-and-suspension scenario) than they would have assigned to the “suspension” scenario if instead asked Question 1 (which mentions only suspension).
Why? For the simple reason that Question 2 tended to provide them with new information (namely, the upshot of the additional careful thought about the Polish invasion scenario) that Question 1 wouldn’t have.
(To caricature this, imagine showing two separate groups of chess beginners the same superficially-even board position with Player A on move, asking Group 1 participants “what’s the probability that A will win,” and separately asking Group 2 participants “what’s the probability that A will make slightly-tricky-advantageous-move-X and win”? Yes, the event Group 2 was asked about is less likely than the event Group 1 was asked about; Group 2′s answers may nevertheless average higher for quite rational reasons.)
I agree. This notion of question 2 providing a plausible cause that might lead to suspension v. question 1 where the test subject has to conceive of their own cause is a bias, but a different type of bias, not a conjunction fallacy. There could be (and possibly have been) ways to construct the test to control for this. For example, there are 3 test groups where 1 and 2 are the same and for the third, the two events are asked independently: What are the probabilities of each event:
A. That USSR invades Poland, or B. That US suspends relations