Originally posted by BlueSeats:
My laymen's sensibility was that whatever muddying variables which might exist in any given year/team/situation would be mitigated by the number of years/teams/situations charted -- and I still believe that to be valid today.
That's not necessarily true for what we are talking about, i.e. the role of pace in distorting per game numbers. Let me try to explain in an analogy. (Sorry if this is a bit long winded, but it's an important statistical point and I believe you can benefit from understanding this better.)
Suppose there's a packaging factory where only two workers check in to work each day. These workers are in a competition of sorts, where whoever produces the greatest number of packages in a given day gets a monetary bonus. Further suppose that, for various reasons, the number of hours in the work day changes from day to day, so that maybe on one day the workers work 8 hours, another day maybe 9.5 hours, etc. The length of the workday can sometimes be affected by the workers themselves, such that on average, some workers tend to work longer hours than others.
Now suppose we want to compare how effective and competetive different workers are at 'beating' their competition and collecting their bonuses. What are we going to look at? Do we compare them by the number of packages they make on average? No, that would not be the best measure. Why? Because the length of the workday varies and that will artificially affect stats on how many packages they make daily. If Jim works 4 hour days on average and makes 4 packages per day, and Bob works 10 hours and makes 5 packages per day, who's really better?
It's a better idea to measure their effectiveness by their
rate of production. Jim makes one package per hour while Bob only makes a half a package per hour, so Jim is better. It's very important now to recall that we want to measure how good these guys are at outproducing their daily competition (NOT the raw number of packages they make), and that the number of hours worked by two competitors is always equal. So for instance, if we wanted to pit Jim and Bob in a given workday, Jim's production would always easily beat Bob's because they'll always be working the same number of hours, and Jim's production per hour is much better. This is the case even though Bob averages more packages made per day.
Suppose we want to compare workers by number of packages produced per day anyway, even though we know it's not the best measure. Suppose we reason that, even though the number of hours may vary between workers, these differences will 'even out' if we look at a large enough number of days. Is this necessarily the case?
No! It *would* even out IF the length of the workday varied randomly for each worker on each workday. But this is not how things work; recall that the workers themselves can exert influence on how long their workdays are. This means that workday length is not assigned completely randomly, but is to some extent dependent on individual workers themselves, and so even if we look at a large number of samples, we may find that some workers still systematically work longer days than others, and so our per day data will still be skewed, no matter how many days we look at.
Hopefully the meaning of this analogy is clear. The workers are NBA players/teams; packages produced correspond to any boxscore stat; and length of the workday corresponds to team pace. Just as the workers in this analogy influence the length of their workdays, it is plausible that NBA point guards influence the pace at which their teams play. If this effect exists and is relatively stable, then we should see the teams of certain point guards playing at paces that systematically differ from the pace of teams guided by other PGs. It is by no means a given that, if we look across enough games and enough teams, that differences in team pace for each point guard will tend to 'even out' and become negligible.
Perhaps it's asking too much to account in-depth for changes in team quality over time in your analysis, as that would be a difficult thing to do. At a bare minimum, though, you should really use stats that are adjusted to take pace into account. They are just better than per game stats, and in some cases discrepencies between per game and per possession numbers are significant to such an extent that they can badly skew conclusions.
For instance, last season John Weisbrod (sp), GM of the Magic, constantly complained about his team's poor defense and even traded an important cog in their offense, Cuttino Mobley, for an aging defensive specialist, Doug Christie. The Magic were last in the league in PPG allowed. But in fact, they were not anywhere near the worst *defensive* team in the league. They played at a very high pace, which artificially inflated the per game stats of both their own team and their competition. In fact, had Weisbrod adjusted for pace, he would have found that his team's defense was actually 15th best-- could have used some improvement, but it wasn't terrible either. Weisbrod's faulty analysis led him to make a desperation trade that wound up hurting his team in the long run.
In any case, I have gone ahead and looked up the pace adjusted numbers for a couple of key seasons.
New Jersey Nets
season team apg team app point guard
01 19.5 21.5 Marbury
02 24.3 26.4 Kidd
percent increase in assists per game: 25%
percent increase in assists per possession: 23%
Phoenix Suns
season team apg team app point guard
03 21.0 22.9 Marbury
05 23.1 24.5 Nash
percent increase in assists per game: 10%
percent increase in assists per possession: 7%
(I used the 03 Suns season because Marbury was traded midway through 04.)
So, for all my long windedness, it turns out that pace adjustments did not make all that much of a difference here. We see that the Nets got a 23% increase in team assist rate from 01 to 02, and the Suns got a 7% increase in team assist rate from 03 to 05. So the overall effect is there in these two instances. But let me take a quick look at quality of teammates.
With the Kidd's Nets, the five best per game scorers (besides Kidd) were Kenyon Martin, Kieth Van Horn, Kerry Kittles, Todd MacCulloch, and Richard Jefferson. With Marbury's Nets, the five best (besides Steph) were Van Horn, Martin, Johnny Newman, Aaron Williams, and Lucious Harris. So although Kidd's Nets saw a sizeable percentage increase in team assist rate, Kidd's Nets were also easily a superior offensive team. Without getting into a deep statistical analysis-- how much do we fault Steph for the fact that one of the league's great playmakers of all time was able to contribute to a greater team assist rate with a superior offensive team?
By comparison, Phoenix's percent increase in team assist rate from Marbury to Nash was a modest 7%. Steve Nash's top five scoring teammates were Stoudemire, Marion, Johnson, Q, and Jim Jackson; Marbury had Stoudemire and Johnson still very much in the developmental phases of their careers, Marion, Penny Hardaway and Casey Jacobson. Given that Nash won MVP last season largely for his superior passing and playmaking (presumably improving team passing as well) and also had a demonstrably better cast of weapons surrounding him, we might even ask if a 7% increase in team assist rate from Marbury to Nash is actually rather modest, or perhaps even *lower* than what we would normally expect if it's really true that Nash is a supreme playmaker and Marbury, not so much. This leads me to be skeptical of the utility of using team assist rates as a measure of how much a PG helps facilitate overall team passing and offensive flow.
help treat disease with your spare computing power : http://www.worldcommunitygrid.org/
I must have missed where he said that 