APBRmetrics

socialpsychologist · Joined: 10 Jul 2009 Posts: 6

Explanations of adjusted plus/minus typically describe the metric as estimating a player's impact on the score controlling for the other players on the court (and sometimes other variables such as home/away, coach, refs, etc.).

In the social sciences, to test for the "main effect" of player A, you would control for the effect of these other variables AND all higher order interactions. Does the computation of adj. +/- do this? If not,why?

The interactions among players are often treated as another step that COULD be taken. But it seems that the interactions are more important than the "main effect" of player. However, this doesn't get much attenion in the public realm. Is there a reason for this?

Just one more question. Are the players on the opposing team controlled for in the analysis?

DraftGuy · Joined: 23 May 2007 Posts: 22

yes, opponents on the court are controlled for.

and yes the interactions are pretty important. i think theyre also a bit unpredictable tho and u have to look for them. they may be important in short term, like a year ro 2. they take a while to even out. mcuban talked about the value being the players value given the roll the coach gives him. change the roll and the value changes. that implies interaction fx. ive tried to estimate the interactions and what they mean but its a big mess.

socialpsychologist · Joined: 10 Jul 2009 Posts: 6

thanks, I have been looking into how to do the calculations myself, and I see that others are doing the same thing you mention.

I am still a little stuck on the interaction point though. Typically, in regression if I wanted to predict variable Y using variables A, B, and C then you would enter:

A
B
C
A x B
A x C
B x C
A x B x C

all as predictors. With adj. +/- the higher order interactions don't seem to be entered. I can see why because for each possession you would have 10 variables (5 home players, 5 away players), which would give you many many interaction terms. This of course, is problematic for many reasons.

So, my questions are (1) am I portraying adj. +/- calculations accurately? and (2) when regressions are run in other domains with large numbers of predictor variabes, are the interactions included or omitted?

DraftGuy · Joined: 23 May 2007 Posts: 22

gabefarkas · Joined: 31 Dec 2004 Posts: 1313 Location: Durham, NC

DraftGuy · Joined: 23 May 2007 Posts: 22

Mike G · Posted: Thu Jul 16, 2009 9:25 am Post subject:

DraftGuy, I find it hard/impossible to read your posts. Please put a few seconds into using real words (rather than u, for example). Note that a non-word (misspelled) gets a red underline in your reply box.

There's no tax on punctuation or capitalization, either.

Also, there's no reason to 'quote' the entire previous reply every time, since that text is already right there above your response. The extra scrolling past repeated text is tiresome. If you're replying to a specific part of the post, you might hit 'quote', and then delete the other parts.

Sorry to nag, but you might connect with a wider audience.
_________________
`
36% of all statistics are wrong

DraftGuy · Joined: 23 May 2007 Posts: 22

Mike G · Posted: Thu Jul 16, 2009 12:05 pm Post subject:

Hey, thanks for humoring me.
I see you have some capitalization now.
In your first post, you didn't quote the entire previous comment, and it works just fine.
It may take a bit more of your time to write better, but it saves everyone time to read it. I get to your ideas more efficiently (and I think you have much to offer).
I don't see any toggle for spell-check. It just happens, for me.

Suppose we make enough mistakes in our writing, even when we try not to. That may be a manageable level, not so severe as to delay comprehension.
_________________
`
36% of all statistics are wrong

Kevin Pelton · Posted: Thu Jul 16, 2009 12:30 pm Post subject:

The spell check underlining is a Firefox feature, I suspect. Somewhat ironically, it currently says I've misspelled spellcheck. [Edit: Apparently it's technically two words. Good catch.]

As for Gabe's description of "priming the pump," he may stop by to explain it but I would guess his point is that if you find interactions that don't make sense in basketball terms, that is still meaningful information in that it suggests that any interactions that do seem logical may be just as coincidental as those that don't.

These being interaction effects between players after accounting for their overall value, I'm not sure the difference between which are logical and which aren't is so obvious myself. Then again, I also think the noise-to-signal ratio on any such effort would probably be pretty enormous.

DraftGuy · Joined: 23 May 2007 Posts: 22

gabefarkas · Joined: 31 Dec 2004 Posts: 1313 Location: Durham, NC

DraftGuy · Joined: 23 May 2007 Posts: 22

[quote="gabefarkasIn the context of this discussion, it means that you'd be only picking interactions that you decide are of interest. This is bad, because each interaction entered into the model -- whether or not the term is found to be significant -- will affect the LS coefficients of every other term in the model. So, who's to say that you didn't just pick and choose which interaction terms you specified in the model until you found a set that showed the results you wanted to find.[/quote]

yes this is bad. you have to pick interactions before hand. One that you think are valid and basketball real. That is harder to do with apm.

gabefarkas · Joined: 31 Dec 2004 Posts: 1313 Location: Durham, NC

Kevin Pelton · Posted: Thu Jul 16, 2009 4:35 pm Post subject: