rbgramacy
commented
9 years ago I guess I should try owning this collaborative environment.
Responding to Matt's query for me, yes I think we should keep this paper simple. Although I do think we should talk about fully Bayesian calculation in the introduction/review. I'll be working on that over the next several days and try to have something by the end of next week.
I thought we were already including the post-season. At least, the data and scripts we used for the blog used all goals data except penalty shots. We also included an indicator for post-season in the model.
Here is a question. The editors provided us with a simple latex template from CRC. I'm working on getting that going today. Should we create a new tex directory on the repo? I'll leave that up to you, Matt, because its your space.
-B
On Wed, Sep 2, 2015 at 7:04 PM mataddy notifications@github.com wrote:
hi @rbgramacy https://github.com/rbgramacy and @sentian https://github.com/sentian,
I've updated this repo (note that it has moved from mataddy to taddylab, which makes it easier for me to manage collaboration).
The new reorg has the old material in code/blog, and the new code files design.R will build you either the regression info for goals or for shots. performance.R then fits the results and compares to salary; the relevant output tables are the full table of betas, PPM, and PM in https://github.com/TaddyLab/hockey/blob/master/results/performance.csv and the correlations with salary in https://github.com/TaddyLab/hockey/blob/master/results/salarycorr.csv. I think that performance.R is pretty easy to parse but please hit me up with any questions here.
The currently checked-in results study the goals data using the up-to-201314 data that I grabbed from bobby's directory.
The analysis all uses standard lasso L1 penalization, and AICc selection. @rbgramacy https://github.com/rbgramacy : I think that perhaps we can stick with this to make the current paper simple? We then differ from the older analysis in a few ways:
- we control for team-season specific effects as well as special teams effects. these effects are not penalized.
- I include player*season interactions, so that there is potentially a different beta for each player for each season.
- we do the translation from betas to partial plus minus.
Finally, we compare to salary and find that the correlation with PPM increases after the 2005 lockout.
Purpose of this issue is to decide what else we need, if anything. Adding in post-season? Putting salary straight into the original regression? Considering other terms like corsi in the regression?
— Reply to this email directly or view it on GitHub https://github.com/TaddyLab/hockey/issues/2.
mataddy
hi @rbgramacy and @sentian,
I've updated this repo (note that it has moved from mataddy to taddylab, which makes it easier for me to manage collaboration).
The new reorg has the old material in code/blog. In the new code, design.R will build you either the regression info for goals or for shots, while performance.R fits the regression and compares to salary; the relevant output tables are the full table of betas, PPM, and PM in https://github.com/TaddyLab/hockey/blob/master/results/performance.csv and the correlations with salary in https://github.com/TaddyLab/hockey/blob/master/results/salarycorr.csv. I think that performance.R is pretty easy to parse but please hit me up with any questions here.
The currently checked-in results study the goals data using the up-to-201314 data that I grabbed from bobby's directory.
The analysis all uses standard lasso L1 penalization, and AICc selection. @rbgramacy : I think that perhaps we can stick with this to make the current paper simple? We then differ from the older analysis in a few ways:
Finally, we compare to salary and find that the correlation with PPM increases after the 2005 lockout.
Purpose of this issue is to decide what else we need, if anything. Adding in post-season? Putting salary straight into the original regression? Considering other terms like corsi in the regression?