Comments on: What I'm Proposing

By: vivaelpujols

vivaelpujols — Sun, 31 Oct 2010 18:03:36 +0000

Ok that makes sense, I’m a little fuzzy on the statistics of it anyways. I think that Chris might prefer to do a binary variable anyway, so if it’s able to be manipulated then it should be fine.

By: stevesommer05

stevesommer05 — Sun, 31 Oct 2010 15:31:57 +0000

I believe the absolute co-linearity is if you use all of the sub-components for a dummy variable with multi categories (i.e. using the 4 dummy variables for a 4 category variable, you’d actually use 3).

I think it’s less applicable to having multiple separate variables.

By: stevesommer05

stevesommer05 — Sun, 31 Oct 2010 15:13:51 +0000

Fair enough, you’ve researched it more than I. In the long run I don’t really care one way or the other. 1-10 is fine with me.

I wasn’t suggesting the scout was the SME in how to create the variables re: regression; just that they have a good feel for the levels of the variables. Might as well express the variable in a way that matches the levels the scout already sees so they are apt to fill it out.

By: vivaelpujols

vivaelpujols — Sun, 31 Oct 2010 14:53:06 +0000

The scout is not the subject matter expert regarding which way the expression of the variables is more conducive to running a regression. That’s us.

Again, I would think that running a regression with 5 (or however many categories TPG chooses) dummy variables would not produce useable results. What I’m reading online is that unless you excluded the intercept, running a regression with all dummy variables yields absolute multicollinearity.

By: stevesommer05

stevesommer05 — Sun, 31 Oct 2010 14:03:34 +0000

Yes, would use regression. I think it would be fine with 4-5 binary (categorical style) variables. I wouldn’t be locked into that approach though. If the subject matter expert (a scout) thought that the variables would be better expressed on a different scale then I would be fine with that too.

I like the binary nature mainly for the ease of interpretation of results.

By: vivaelpujols

vivaelpujols — Sun, 31 Oct 2010 06:00:31 +0000

Well I would assume you’d want to run the thing in a regression. How else would you tease out the multi-collinearity and measure the relative importance of each individual flaw? Would a regression work if you regressed 4 or 5 binary variables on wOBA?

I agree with Valentine that this would be simpler and may reduce bias; however, I disagree that the subjectivity in deciding what’s worse: 3-4, or 8-9, is a damning quality. Isn’t it much worse to assume that either a hitter has the flaw or doesn’t? That simplifies everything into a either perfect or terrible.

I would much rather have the 1-10 scale, all things being equal, but I suppose that’s not really what Chris is going for in his analyses.

By: stevesommer05

stevesommer05 — Sun, 31 Oct 2010 01:18:00 +0000

I’d do it this way initially for a couple of reasons. I think it lessens subjectivity some, and thus hopefully reduces bias. Additionally I think it’s easier to interpret the results if the variables are binary (you either have the flaw or don’t. How do you use the information of going from a 4 to 5 in leaky back elbow lowers your wOBA by 0.005 pts? It just seems hard to understand the 1-10 scale. That being said, I think I’d lean toward letting the scout pick the scale.

By: valatan

valatan — Sat, 30 Oct 2010 23:47:17 +0000

That would depend on how accurate the various thing are. An ordinal variable would only really work if the differenc between a 3 and a 4 was the same as a difference between a 8 and a 9 , after all.

By: vivaelpujols

vivaelpujols — Sat, 30 Oct 2010 21:47:01 +0000

It seems like binomial variables would make it hard to run a proper regression, why not just have him grade the degree of flaw on a 1-10 scale or something?