Backboards: 
Posts: 155
Favourites: 1

DTL nerds: I've updated my regression work if you care to see.

There are two big changes.

(1) I brought in height. It's known to have an impact on results and we should be treating a 6'10" center with only 1 block per game differently than a 6'5" center with the same.

(2) I performed backward selection on each of the models, that is, I let R do it. Differences between the full model and the reduced model were universally quite small, so why keep the noise factors?

I did tinker around with interaction terms for a items that are correlated, but they were rarely significant, and sometimes had wild effects on the parameter estimates, so I bagged it.

I would love to use the GPA/Exam value as a predictor, but I can't since it has to serve as an outcome.


Results of my DTL regressions inside. (extensive nerdliness inside)
I have data from S24 for the whole league -- I can't recall who shared this with me but thanks again! -- and used it as the basis for modeling. Scout ratings were converted to numbers on the scale -2.33 (None-) to +2.33 (Exc+), increasing by 1/3 at each step up the scale. For Int, I used a quick estimate of (Exam score-50)/12.5 to achieve a similar range. Stats were converted to Z scores on a by-position basis using means and standard deviations for players at those position. I ran the predictions once for the entire sample, and again separately for each position to check for differences. Shooting: PPG is the dominant factor. C and PG also have a relationship with TO, but in opposite directions. Curious. Defense: Steals and blocks, of course, but the proportion of each varies. Blocks>Steals for C, Steals>Blocks for PF, almost Steals>>Blocks for the rest. Hands: Assists minus turnovers, what did you expect? Also a weak association with points for the smalls but not the bigs, so go ahead and fire it into your guys in the post. It doesn't matter if they drop it. Rebounding: Rebounds. Negative association with blocks which sounds strange except the stats are correlated with each other so you see this phenomenon in regression where the stronger stat is exaggerated and the weaker flips sign. I should probably add in an interaction term for RPG x BPG. Intelligence: Kind of a noisy factor all-around. Some association with assists, except for SF. Some association with points, except for C. The least necessary regression since we get GPA and test score without a scout filter. Athleticism: Seems to affect every position differently. Makes your C better blockers, your PF less prone to TO, your SF better at steals, your SG stronger rebounders, and your PG better passers. It's also associated MORE turnovers in PG, so maybe that's a wash. R^2 stats for the regressions weren't great so there's a lot unaccounted for. Shooting and Hands models (ha! hand models) were OK (~0.4), the others were lower. I might try a couple more interaction terms for associated factors like PPG x APG or APG x TO to see if I can sharpen them up. Don't want to add too many terms, though, since the sample size is limited and that risks overfitting. Feel free to use these results if you wish. I'm going to take them for a spin next season. I think their primary function would be to inform you how accurate the less reliable scouts are. Is that Fair Scout with a +5 really a +5 or a +2? Run the model, see what it thinks.
Post a message   top
Replies are disabled on threads older than 7 days.