I was in touch with Adam Goldstein who is keen to run multiperiod models. We will follow up on this once we have the experimental data from the larger cohort.
But generally we will put a call out at some point in the future asking if people want to build their own models to explain the data.
I think you missed a constraint which we imposed that mitigates the problem you illustrate so nicely. Which is that the trajectories are filtered such that as stated in the paper: “we set the bounds of the random increments for Passive+ to the central 85th percentile interval of the absolute wealth changes on Day×.”
This means that the graphs you show of the very different trajectories for the additive vs. the multiplicative are filtered out by this constraint. The constraint was in place to try to counter exactly the problem you articulate.
That said, I think the spirit of the criticism is very much valid, there are different properties between the additive and the multiplicative conditions during the passive phase that may confound the interpretation of the difference observed in behavior. Many other criticisms have had this flavour, differences in variance, skew etc. between the passive trajectories.
There are two classes of solution to dealing with this.
1) minimise particular differences via experimental design
2) let them vary and assess via statistical modelling
In our project together we are clearly taking both of these strategies. 🙂
Thanks Benjamin for the update. I recall Ollie Hulme tried to start an adversarial collaboration. Got this initiative ever started?
I was in touch with Adam Goldstein who is keen to run multiperiod models. We will follow up on this once we have the experimental data from the larger cohort.
But generally we will put a call out at some point in the future asking if people want to build their own models to explain the data.
Thanks Benjamin.
I think you missed a constraint which we imposed that mitigates the problem you illustrate so nicely. Which is that the trajectories are filtered such that as stated in the paper: “we set the bounds of the random increments for Passive+ to the central 85th percentile interval of the absolute wealth changes on Day×.”
This means that the graphs you show of the very different trajectories for the additive vs. the multiplicative are filtered out by this constraint. The constraint was in place to try to counter exactly the problem you articulate.
That said, I think the spirit of the criticism is very much valid, there are different properties between the additive and the multiplicative conditions during the passive phase that may confound the interpretation of the difference observed in behavior. Many other criticisms have had this flavour, differences in variance, skew etc. between the passive trajectories.
There are two classes of solution to dealing with this.
1) minimise particular differences via experimental design
2) let them vary and assess via statistical modelling
In our project together we are clearly taking both of these strategies. 🙂