Anywhere to find Regular Defensive Adjusted Plus Minus?

[Kathoro]

Not RAPM, but just APM for the defensive end. Thanks.

[J.E.]

What do you want it for? There's really no reason to use it over RAPM, ever.

And, if you ever use vanilla APM, please use multiyear studies (2yr+).

There's at least one APM multiyear study on 82games, but it's rather old

[Kathoro]

But I thought that RAPM only provides reasonable results for players that play considerable minutes. I thought for players that don't play many minutes, RAPM usually provides numbers close to zero. I don't want to use something that is heavily affected by playing time like that.

[J.E.]

Hm.. APM sets the rating of everyone who had played less than a certain threshold to zero before doing the regression

How is that not "heavily affected by playing time"?

[Kathoro]

Oh thanks. I didn't know regular APM was heavily affected by playing time as well.

[xkonk]

My understanding was that low-minute players are grouped together into a single 'replacement player', who then gets estimated with a single APM value. That value isn't 0, is it? And APM doesn't need to be calculated that way; each player can be estimated separately. That just leads to other problems, right?

[J.E.]

My understanding was that low-minute players are grouped together into a single 'replacement player', who then gets estimated with a single APM value. That value isn't 0, is it?

I'm relatively sure that it is, but you'd have to ask Aaron from bbv I guess. Ilardi/Rosenbaum might have used a different value.
My point of APM being influenced by minutes played still stands, though.

And APM doesn't need to be calculated that way; each player can be estimated separately.

Not sure what you meant by this, but if you run the regression without fixating all low minute player's ratings you'll get extremely funky results. One (very low minute) player could easily get a rating of +100

[xkonk]

if you run the regression without fixating all low minute player's ratings you'll get extremely funky results. One (very low minute) player could easily get a rating of +100

Right, that's what I meant by other bad things could happen. I think any low-minute-player's rating in any metric has to be viewed with the understanding that it might be very wrong.

[DSMok1]

if you run the regression without fixating all low minute player's ratings you'll get extremely funky results. One (very low minute) player could easily get a rating of +100

Right, that's what I meant by other bad things could happen. I think any low-minute-player's rating in any metric has to be viewed with the understanding that it might be very wrong.

With APM, you have the added advantage of low-minute-players' wrong ratings messing up everyone else's ratings, also!

[mikey1987]

With APM, you have the added advantage of low-minute-players' wrong ratings messing up everyone else's ratings, also!
I wonder if there is a way of doing it.

Hi,

Do you have any suggestions then? I am running into a huge problem, my regression is messed up. I have even gone to my local university and asked a statistician for help, but he was no help. Can you help?

[J.E.]

What exactly is your problem?

[DSMok1]

One way would be to assign, say, 50 minutes of -3.5 play to each player (to stabilize things for the low minutes players), and then at the end scale back up so it's still minutes-weighted summing to 0.

(In other words, like ridge regression only less mathematically rigorous!)

Otherwise, combining all players below, say, 5% of minutes played into 1 players will help things (but you know longer can differentiate them).

[J.E.]

Simply deleting the players with low minutes from the regression works, too.

Probably worse than DSmok1's first suggestion (in terms of accuracy) but easier to compute, I suppose