Does winning matter for Hall of Fame Induction (comments)?
From the comments of part four.
Interesting analysis. A few suggestions for your model:
1) Most important: only count HOFers chosen by the writers. That is the standard we care about, and including Vet Comm choices really lowers the standards. That's the main reason so many of your contemporary players are getting unrealistically high probability scores.
2) You need some measure of how brightly the candle burned, not just how long. The same linear weights total over 14 years is a very different player than spread over 19 years. Easiest thing is something like OPS or OPS+.
3) It would also help to capture peak performance better. If you have top-5 or top-10 MVP (not just winner), would probably strengthen the model. And add some variables that measure dominance in key categories, such as "Black Ink" and/or Silver sluggers.
4) Piazza's absurdly low rating makes me think that the position dummies aren't doing as much as you want them to. Better would be position-adjusting each hitter's OPS+. (But that would be a lot of work.)
To follow up, you won't be sure if your team win% and World Series variables really matter until your model better captures players' rate value (as opposed to cumulative value). Otherwise, it's possible those 2 variables are capturing these other elements.
Thinking some more about this, you might want to include OBP+ and SLG+ separately, rather than OPS+, in that HOF voters may value SLG more highly.
Small point: It might be better to use AB/PA or G, rather than seasons. It will help value players who missed a lot of playing time, like Larkin and Larry Walker, more accurately.
Guy makes some very good points.
1. Some part of me feels like what we're really interested in is who gets in, not just who is elected by the writers. On the other hand I agree that it is part of what is causing problems for the recent players. I think the next step is to drop the probit model and examine votes cast rather than the binary induction variable.
2/3. I think 2&3 are related and I think Guy is right. If you look at the career totals of Larry Doyle and Joe Jackson they are very similar. On the other hand Jackson hit .356 and Doyle .290. Doyle scored 70.7% and Jackson 27.1% -- this difference likely comes from their positions – Doyle was a second baseman and Jackson an outfielder. Usually we think of Jackson as a lock for the hall of fame.
Silver sluggers would be the easiest way to measure this since I have the code written to extract gold glove and MVPs. Unfortunately they have only been around since 1980. I think the black ink test might be a better way. Linear weights per year might also be a good way to go. Right now I think the year variable is capturing two effects: longevity and how long a time it took a player to put up his stats.
4. I don’t think the position effect is constant over time. Now we expect an all-star shortstop to hit .320 with 35 HRs and play good defense. 30 years ago an all-star shortstop hit .280 with 10 HRs. Similar shifts have taken place for second and third basemen. I need to figure out a good way to control for these changes in expectations.
A related point, I think, is the effect of expansion. This leads to more variance in the talent level and is not picked up by the runs per game based league factor. As a result a player like Craig Biggio puts up much better stats today than he would have without expansion (he gets half his at bats against pitchers who would be in the minor leagues). I need to find a better way to address this problem.