3 thoughts on “5-2 Vincent Ginis (for Cathy Macharis). Investigating the origins of redistributing behavior in non-ergodic contexts.”
Benjamin Skjold
Thank you for a great talk, Vincent. Your experimental setup is super interesting and actually quite closely resembles an idea we have been considering. Our idea was/is to keep the gamble fixed and offer an alternative fixed amount while asking participants to choose between the two options, which I think would give the same qualitative results as your setup.
One aspect that I find particularly interesting, but that you did not mention, is that the results of your experiment indicate that humans approximate time-optimal decisions without the need for repeated gambles.
A few questions:
– Have you considered what would happen in a different dynamic? Eg. the additive where the time average and the ensemble average are equal; does this mean that tau is either (approx) 0 or 1 for all gambles?
– The length of the experiment is rather short, is this simply due to time restrictions, or is there a particular reason for this? And in connection with that, have you run any simulations, where the length is longer?
– One criticism that EE often receives is that when people are presented with two gambles with the same growth rate, EE predicts they are indifferent, while it is argued that they will take the one with lesser variance (especially when the time frame is short). In this framework, I imagine this equates to observed tau being greater than the optimal value, but the opposite is observed. I can see this being an artefact of the very artificial environment an experiment is played in, but don’t know if that is the reason. Do you have any thoughts on this?
Very nice experiment. I love how simple it is. It would be interesting to pin down the connection to the Kelly criterion, optimal leverage etc. Probably you already have some calculations that you didn’t have time to present here?
I also like Benjamin’s suggestion to try a version of this with non-multiplicative dynamics. I’m guessing people would hedge less as their capital grows?
I’m afraid I have some annoying psychological questions, that are not altogether in line with my enthusiasm for this research.
You are presenting participants with a single dial they can tweak or not. The scale is from 0 to 130% if I read correctly. So the concern is that there is a demand characteristic, which is is to say the participant interprets from the experimenter that its somehow expected to turn the dial away from zero. So the sceptic could argue, your main effect of a tau per bet in the realm of {0.1 to 0.4}, is driven by the participant’s expectation that they should turn the knob, but that they wouldn’t other.
Of course you can counter that the chosen tau is tracking optimal tau, but the question is how strong is this effect? if the slope is 0.07, then whats the variance explained by optimal tau? I’m assuming that’s something like 0.5%.
I wonder if this effect is stronger if you select the gambles where tau has a stronger influence on growth. That would be a statistical model in which you model the interaction between optimal-tau and influence-of-tau. There presumably you would see a larger effect, and hopefully that could counter the criticisms above somewhat.
Thank you for a great talk, Vincent. Your experimental setup is super interesting and actually quite closely resembles an idea we have been considering. Our idea was/is to keep the gamble fixed and offer an alternative fixed amount while asking participants to choose between the two options, which I think would give the same qualitative results as your setup.
One aspect that I find particularly interesting, but that you did not mention, is that the results of your experiment indicate that humans approximate time-optimal decisions without the need for repeated gambles.
A few questions:
– Have you considered what would happen in a different dynamic? Eg. the additive where the time average and the ensemble average are equal; does this mean that tau is either (approx) 0 or 1 for all gambles?
– The length of the experiment is rather short, is this simply due to time restrictions, or is there a particular reason for this? And in connection with that, have you run any simulations, where the length is longer?
– One criticism that EE often receives is that when people are presented with two gambles with the same growth rate, EE predicts they are indifferent, while it is argued that they will take the one with lesser variance (especially when the time frame is short). In this framework, I imagine this equates to observed tau being greater than the optimal value, but the opposite is observed. I can see this being an artefact of the very artificial environment an experiment is played in, but don’t know if that is the reason. Do you have any thoughts on this?
Very nice experiment. I love how simple it is. It would be interesting to pin down the connection to the Kelly criterion, optimal leverage etc. Probably you already have some calculations that you didn’t have time to present here?
I also like Benjamin’s suggestion to try a version of this with non-multiplicative dynamics. I’m guessing people would hedge less as their capital grows?
Great talk, very elegant design.
I’m afraid I have some annoying psychological questions, that are not altogether in line with my enthusiasm for this research.
You are presenting participants with a single dial they can tweak or not. The scale is from 0 to 130% if I read correctly. So the concern is that there is a demand characteristic, which is is to say the participant interprets from the experimenter that its somehow expected to turn the dial away from zero. So the sceptic could argue, your main effect of a tau per bet in the realm of {0.1 to 0.4}, is driven by the participant’s expectation that they should turn the knob, but that they wouldn’t other.
Of course you can counter that the chosen tau is tracking optimal tau, but the question is how strong is this effect? if the slope is 0.07, then whats the variance explained by optimal tau? I’m assuming that’s something like 0.5%.
I wonder if this effect is stronger if you select the gambles where tau has a stronger influence on growth. That would be a statistical model in which you model the interaction between optimal-tau and influence-of-tau. There presumably you would see a larger effect, and hopefully that could counter the criticisms above somewhat.
Cheers,
O