On The Interpretation Of Assists

[v-zero]

I'm making this thread (as most of my threads go) to poll the most knowledgeable basketball analysts on the internet about their opinion on assists. After lots of work on evaluating assists I have come to the personal conclusion that assists are not a scoring measure, or a usage measure, I believe they are a measure of offensive involvement. I have come to this conclusion because usage is also a measure of offensive involvement, which would explain why the usage*assist interaction is superior to non-interacted assists as a predictor, but also because it explains another important issue I came across when using a detailed box-score derived from PBP data: taking the credit for an assist from the scorer of the basket (since in PBP you know who assisted whom) has an inferior impact on the predictive ability of a box-score metric relative to simply considering assists as a time-weighted measure and taking credit from each player on the floor who is not the assister equally.

Has anybody else done similar work and found similar results? Does this surprise anybody? Do we need to stop thinking about assists as a sort of second-tier of scoring?

[Mike G]

" ...taking credit from each player on the floor who is not the assister equally."

Don't we know the fraction of a player's points which are unassisted?
Unassisted points are highly correlated with assists, but other stats are also correlated: 3FGA, FTA, TO, and Stl, I've found.

[v-zero]

" ...taking credit from each player on the floor who is not the assister equally."

Don't we know the fraction of a player's points which are unassisted?
Unassisted points are highly correlated with assists, but other stats are also correlated: 3FGA, FTA, TO, and Stl, I've found.

We do, that was part of my point: if you debit the receiver of the assist and not his team mates then the resultant metric will be an inferior predictor to one which simply debits everybody who didn't make the assist.

[PD123]

Do we need to stop thinking about assists as a sort of second-tier of scoring?

I certainly think we need to stop think about all assists as a sort of second-tier scoring. The biggest problem we have when it comes to using/valuing assists is that the box-score just kind of lumps everything that qualifies as an assist together; they're often treated like most other stats are given a relatively uniform value and that's that. But we know (or at least we're pretty sure) that that's not true. The first, and most obvious, way to break down assists is by shot location, with the idea being that, the farther the shooter is from the basket, the greater the impact the shooter has on the scoring play. Any player can hopefully convert an open lay-up at a relatively high rate and so the skill of the player taking the shot isn't nearly as important as being able to successfully completing the pass in the first place. The farther out you go, the more a players shooting skill comes into play and affects whether or not the possession is successful and an assist is awarded; the passer is still adding value by getting it to "the right player", but at a much lower level than at a closer range.

Furthermore, I think that you can break assists down into four general categories:

1. Non-Assist Assists: These are assists that are handed out incorrectly (like the first assist from this clip: http://www.youtube.com/watch?v=y_9eTpCqSfg). We know it happens and so it leads to a distortion of player value. How often it happens, and how large the distortion is, becomes a matter of debate; if assists get given incorrectly at a really low rate, it's not that big of a problem.

2. "Lucky Assists": These are assists where the player who is credited with the assists kind of just gives up the ball to another player, and that player just happens to score within the "assist" framework (be it a contested jumper, or the scorer making a quick drive to the basket). Take the Jennings "assist" on the Ellis buzzer-beater http://www.youtube.com/watch?v=H5SX9Ip_1FE; Jennings didn't really do anything to create value there, he kind of just half-heartedly passed the ball to Monta Ellis who did everything of value.

3. "Passing Assists" (this is where I think you start to get into the discussion of value and distribution of credit): These are assists where the player who is credited with the assist gets it by passing to an already open player. It could be a player who is open because the defense didn't rotate properly, who is hit in stride while cutting to the basket, who is coming off a screen, etc. Somebody other than the passer is "creating" a large portion of the scoring opportunity for the scorer, but the passer is putting the cherry on top by getting the open player the ball.

4. "Creation Assists": These are assists where the player who is credited with the assists actually "creates" a large portion of the scoring opportunity (generally accomplished with dribble penetration or crafty ball-fakes). I'm talking about things like this- http://www.youtube.com/watch?v=GTskofT6QaU&t=0m28s, where Steve Nash drives into the lane, pulls Amar'e's defender away and then passes back to an open Amar'e for an easy dunk.

There certainly is value in having a player who has a propensity to distribute "Creation Assists;" this player is likely to use his skills to enhance the play of his teammates by creating efficient shot opportunities for them. There is less value in having a player with a propensity to accumulate "Passing Assists," but this player, at least, can be relied upon to manage an offense and get the ball where it needs to be without turning it over. There is little to no value in having players who rack up "Lucky Assists" and non-assist assists. The problem we face when assessing assists in this vein is obvious, you can't derive the nature of a particular assist from PBP data and the like; it only becomes evident through visual observation, and evaluating every assist would be a painstaking undertaking (one fraught with all manner of subjective choices). That really only leaves us with the ability to distinguish assists based on shot location and, while not giving nearly as much of the picture as we may like to have, I don't think it is entirely without it's uses (for instance, it tells us that, as has been pointed out in other threads, people should be cautious when looking at Rondo's gaudy assist totals-- he racks up a lot of assists in the around the 20ft range).

[Mike G]

There is little to no value in having players who rack up "Lucky Assists" and non-assist assists.

Well, being in the right places, you'll have better "luck", and this will help your team win. Nobody really makes a living on lucky bounces or continually being unguarded under the basket.

Passing the ball to a high% shooter, whether from 2 ft or 20 ft, is a better pass than one that has a lesser chance of becoming an assist -- a lower% shot, that is. In other words, assists are (in theory) = (completed passes) X FG%

You can strip away "homer" assists by using a team's (Away Ast/FG)/(Home Ast/FG) factor on their home assists. This year, about 19% of Lakers' and Bobcats' home assists may be bogus.

On the matter of distance being the determiner of assist value: Some of us think it's pretty hard to score "among the trees", and shooters have a wide range of comfort zones.
You might even say that assisting a 3FG has 1.5 times the value of assisting a layup. On average.

[v-zero]

In my work on building a predictive player metric from PBP data I have found that three point assists are better than close assists, which are better than long assists, which are better than mid-range assists. I would suggest that long assists are more valuable than mid-range assists because they are more likely to be wide open, and quite possibly some of them are would-be three point assists on which the receiver is just inside the three point line, and hence is at fault for not taking a step back.

[PD123]

[Nobody really makes a living on lucky bounces or continually being unguarded under the basket.

I think that there are several players that make a decent living off of being good at converting relatively open, assisted shots around the basket: DeAndre Jordan, Tyson Chandler, JaVale McGee, Reggie Evans, Omer Asik, etc (now they all have other skills, primarily on defense, but I'd say that make quite a good living converting relatively open, assisted shots at the basket). As for the "lucky bounces," that's not what I'm really talking about, and perhaps I shouldn't have chosen the example I did. When I'm talking about "lucky assists," I don't mean fortuitous rolls or bounces, I mean something like this, Calderon passes the ball to DeRozen who turns and drives to the basket through a group of defenders and gets a layup. The assist that was awarded to Calderon, I think most would agree, has little "value" and did the bare minimum to contribute to the basket (mainly, giving a guy the ball and getting out of his way). Sure, nobody makes a living doing stuff like that, but those kinds of assists can impact how we view a passers assist total and how we view the predictive value of assists; if 10% of Player A's assists fit the "Lucky Assist" mold, it's certainly seems enough to impact perception and prediction.

You can strip away "homer" assists by using a team's (Away Ast/FG)/(Home Ast/FG) factor on their home assists. This year, about 19% of Lakers' and Bobcats' home assists may be bogus.

I agree with the "home-assist adjustment" idea in theory, but I've never really agreed with the way it's been translated into a general measurement. The problem I have with judging it on the home/road split is that there can be a lot of other factors that go into it. Using (Away Ast/FG)/(Home Ast/FG) seems to me to be a somewhat crude way to go about it given that we know offenses generally tend to be more efficient at home than on the road (to the tune of around 4-5%). Yes, players get unwarranted assists at home, but I'm not sure that I buy the idea that anywhere near 19% of the Lakers/Bobcats' home assists are "bogus."

You might even say that assisting a 3FG has 1.5 times the value of assisting a layup. On average.

You might, and others do agree with you (v-zero seems inclined to agree with that and EvanZ seems to as well), but I'm not sure that I do. Yes, I believe that a 3FG is 1.5 times as valuable as a layup, but I don't know that the same can be said of their assists. Taking the stance that an assist on a 3FG is "more valuable" than an assist on a layup relies on the idea that the value of the 3FG is more than layup. But I would argue that such a philosophy is a misguided way at assigning value to assists in that it relies on the same retroactivity that we use to distribute assists in the first place. To me, the question that is more interesting, and more germane, is how assists impact outcomes, i.e. how much more likely is a player to score when an assist will be awarded if he does? There is a study about potential assists on 82games (http://www.82games.com/assisted.htm) that the FG% on potentially assisted 3FG is only somewhat higher (3.5-4%) than on unassisted 3FG. When an assist could be awarded on "close shots," that difference in FG% between potentially assisted shots and unassisted shots grew to about 12.5%. (The conversion rate on potentially assisted dunks was only around 7%, but that I think has more to do with the fact that an attempted dunk already has such a high success rate.)

(It should be noted that, according to the study, most 3FGA happen when a potential assist can be awarded, which could imply that assists are generally "needed" for 3FG, but I'm not sure that such is the case, and I would go back to distinguishing between "Passing Assists" and "Creative Assists;" it seems to me that a good deal of the "assists" that are awarded off of 3FG come in the general course of ball rotations where a shooter more or less decided he's open enough and takes the shot. Sure, the passer recognized that there was a possible opportunity for the 3FGA, but most of the "work" to break down the defense and provide the open opportunity for the shooter was done by somebody else.)

I don't disagree that it is "hard to score 'among the trees,'" and I don't think that a layup is more valuable than a 3FG, but I do think that the assist that results in a shot that is close to the basket is more valuable than the assist that results in a shot farther away from the basket because of the massive disparity between the two shot-types' success rates. Now, there gets to be a point where the added value of a 3FG over a FG outweighs the slight increase in the likelihood of a made shot, but I think that point is saying that, somewhere around 10-15 ft, the added value of a 3FG outweighs the increased likelihood of a successful shot closer to basket and so the value of its assist is larger.

[PD123]

In my work on building a predictive player metric from PBP data I have found that three point assists are better than close assists, which are better than long assists, which are better than mid-range assists. I would suggest that long assists are more valuable than mid-range assists because they are more likely to be wide open, and quite possibly some of them are would-be three point assists on which the receiver is just inside the three point line, and hence is at fault for not taking a step back.

The problem I have with using PBP data to quantify the value of assists is that PBP data doesn't include any sort of "assist failure" beyond bad passes. There isn't any measure of success-vs-failure rates to get an adequate understanding of how the potential act of assisting increases respective shots, nor is there a way measure a passer's decision-making ability via who he passes to and how they shoot (at least, beyond the fact that 100% of the people he assists score). It seems to me that, until we can figure out a method of making a general value statement for these sorts of things (or we can get our hands on the SportsVU data), assists are to remain a somewhat nebulous statistic.

[v-zero]

I once thought that, I no longer do. Assists on their own have no real predictive value, whereas assists coupled with usage have very good predictive value. This stems back to my argument that assists are more of an indicator of offensive involvement, than of anything else.

[Mike G]

Using (Away Ast/FG)/(Home Ast/FG) seems to me to be a somewhat crude way to go about it given that we know offenses generally tend to be more efficient at home than on the road...

That's right, but note the formula is the ratio of 2 ratios: the fraction of FG which are "assisted", home and away.

Charlotte shoots .427 at home and .413 on the road (raw FG%). That's about the league average disparity, nothing surprising.
They average the same FGA at home and away -- 81.8.
Of an average 34.9 FG made at home, 60% have been assisted.
Of avg 33.7 FG on the road, just 50% are counted as assisted.

This is independent of efficiency, home and away. There's little or no correlation between teams' shooting better at home and having a higher Ast/FG ratio at home.

I'd theorize that only 5 of every 6 assists at home would have been counted on the road.
Now, the crude part of this is to assume every player is affected the same. Home scorekeepers may well have "favorites" on the team. Blake Griffin got 50% more assists at home in his rookie year.

..the assist that results in a shot that is close to the basket is more valuable than the assist that results in a shot farther away from the basket because of the massive disparity between the two shot-types' success rates.

It's important to distinguish between a pass -- a potential assist, if you will -- and an actual assist, which only is registered when the shot is made. The made shots after the passes are quantified by the FG%. A mere pass that leads to a 20' shot is only going to be an assist maybe 45% of the time.

That's why I mentioned that Ast = (pass completion%) * FG%. Just as the pass into the area near the basket is less likely to be completed, the outside shot is less likely to be made.
If player A always passes inside, then 20 passes may result in 12 assists. Player B who only passes outside, may only get 9 assists from 20 good passes.
Player A has gotten his just recognition from making "better" passes.

[Mike G]

The average team is 2211/4900 on FG shooting and has 1319 Assists.
They have scored 5848 points, of which 1003 are FT.
That leaves 4845 points from FG, or 2.19 per FG

Since on Assisted shots they are 1319 of 1319, on unassisted shots they're just 892 of 3581 -- 25%
An unassisted FGA has an avg value of .25 * 2.19 = .55 points
An Assisted FGA has avg value of 2.19
Is an avg Assist therefore worth (2.19 - .55) 1.64 points?

I've capitalized 'Assist' to distinguish from a 'pass' or a 'good pass'. It's the fraction of passes that result in a teammate FG.
This analysis assumes 2FG and 3FG are equally assisted -- which they actually are not.

[PD123]

Since on Assisted shots they are 1319 of 1319, on unassisted shots they're just 892 of 3581 -- 25%
An unassisted FGA has an avg value of .25 * 2.19 = .55 points
An Assisted FGA has avg value of 2.19

This is part of my point, and the general problem with the assist. Just saying that a team went 1319 of 1319 on their assisted shots doesn't really tell us anything; it's a truism that holds little practical meaning. Likewise, lumping in all missed shots as "unassisted" is misleading because it ignores the fact that sometimes assisted shots fail to occur, and are instantaneously considered "unassisted", not because of passer has done something wrong but because shooter just missed. What I think is the more important question is how many shots that would have resulted in assists were missed? What is the Point/Shot value of potentially assisted shots (and what's the corresponding new value of unassisted shots)? It's certainly greater than unassisted shots, but by how much?

How many times does the passer do something to increase the likelihood that a shot taken will go in (this would indicate a "high value" pass), how many times does the passer give the shooter the ball without impacting his success rate (there is still some value in getting the ball to the right people), and how many times does the passer give the ball to someone with the idea that they'll eventually make something happen on their own (little to no value in the pass)? I think these three situations have distinct values and yet lumping them all together as assists assigns them to a similar fate.

[Mike G]

By assigning a huge value like 1.64 points to each registered Assist -- and zero value to those passes leading to a missed opportunity -- you're basically assuming each counted Assist is representative of more than 1.0 good plays. It may take 6 timely dribbles and 3 passes to create 1 Assist.

Analogous is the value of the defensive rebound. You could argue that they aren't so valuable, because one or 2 teammates might have gotten it, if you hadn't.
But the flip side of the argument is that each DReb that you got represents another (or fraction thereof) that you could have gotten.

In both cases, a player gets an Ast or a DReb a certain fraction of the time he might have gotten one, due to his good play.

[DSMok1]

This is part of my point, and the general problem with the assist. Just saying that a team went 1319 of 1319 on their assisted shots doesn't really tell us anything; it's a truism that holds little practical meaning. Likewise, lumping in all missed shots as "unassisted" is misleading because it ignores the fact that sometimes assisted shots fail to occur, and are instantaneously considered "unassisted", not because of passer has done something wrong but because shooter just missed. What I think is the more important question is how many shots that would have resulted in assists were missed? What is the Point/Shot value of potentially assisted shots (and what's the corresponding new value of unassisted shots)? It's certainly greater than unassisted shots, but by how much?

How many times does the passer do something to increase the likelihood that a shot taken will go in (this would indicate a "high value" pass), how many times does the passer give the shooter the ball without impacting his success rate (there is still some value in getting the ball to the right people), and how many times does the passer give the ball to someone with the idea that they'll eventually make something happen on their own (little to no value in the pass)? I think these three situations have distinct values and yet lumping them all together as assists assigns them to a similar fate.

I think that reading this old thread from the archive ("Adjusting Assists") will help further this discussion: http://godismyjudgeok.com/DStats/APBRme ... 57%20.html

[Mike G]

Just saying that a team went 1319 of 1319 on their assisted shots doesn't really tell us anything; it's a truism that holds little practical meaning. Likewise, lumping in all missed shots as "unassisted" is misleading because it ignores the fact that sometimes assisted shots fail to occur, and are instantaneously considered "unassisted", not because of passer has done something wrong but because shooter just missed.

Right, well the 1319 was just a number (assists) to subtract from both the numerator and the denominator of FG/FGA, to see what unassisted FGA success rate was.
You're right that it overloads the resultant FG% with missed shots, as every one of the unsuccessful "potentially assisted" shots is in that subset.

Suppose you have a league in which a player is not allowed to shoot unless someone has just passed him/her the ball. No shooting off the dribble, nor putbacks, etc. Then every FGA would be "potentially assisted".

Without putbacks or "garbage baskets" near the hoop, some high-% shots are not attempted, or they're disallowed. Overall FG% goes down. Because of the restriction, the game calls for more passers and fewer dynamic scorers.

It may now be that teams average 30 FG (and 30 Ast) out of 90 attempts.
Each registered Assist is accompanied by 2 "potential assists" -- those which resulted in missed shots.
Since the pass is necessary to every shot, it's reasonable to say the Ast is as valuable as the FG, on average.

Does it really matter whether we know the reason a player missed a shot? It was either an ineffective pass, or he did not take a good shot, or some combination. Good passers need good shooters to get a large number of assists.

Knowing the fraction of 'potentially assisted' shots vs 'unassisted' shots doesn't tell us much, without knowing the quality of each element (pass and shot) of each play.