Stat guys Duncan vs Garnett (bear with me)

[Blackmill]

This thread finally convinced me to make an account after months of lurking. Anyways, I've found this entire discussion very useful and interesting. I also have a question about RAPM that I've been meaning to ask. Looking at Colt's data (I think the stats-for-the-nba site is xRAPM but correct me if I'm wrong) I noticed that the 2002/2003 RAPM is significantly lower than the late 2000s RAPM. Shaq, perhaps one of the greatest offensive players ever, only posted a 2.1 on offense in 2002 and 3.8 in 2003 compared to frequent 5+ offensive RAPM ratings in 2007 and onwards. I think its fair to say that many of the players posting 2+ or even 5+ offensive RAPM weren't more impactful than Shaq. Is this discrepancy due to the large error in RAPM ratings? Is it because 2002 is the first prior-informed year in Colt's data? Or is there some other reason why players generally have better RAPM in the late 2000's?

[colts18]

This thread finally convinced me to make an account after months of lurking. Anyways, I've found this entire discussion very useful and interesting. I also have a question about RAPM that I've been meaning to ask. Looking at Colt's data (I think the stats-for-the-nba site is xRAPM but correct me if I'm wrong) I noticed that the 2002/2003 RAPM is significantly lower than the late 2000s RAPM. Shaq, perhaps one of the greatest offensive players ever, only posted a 2.1 on offense in 2002 and 3.8 in 2003 compared to frequent 5+ offensive RAPM ratings in 2007 and onwards. I think its fair to say that many of the players posting 2+ or even 5+ offensive RAPM weren't more impactful than Shaq. Is this discrepancy due to the large error in RAPM ratings? Is it because 2002 is the first prior-informed year in Colt's data? Or is there some other reason why players generally have better RAPM in the late 2000's?

This.

But 2002 RAPM wasn't a prior informed year. It was actually non prior informed. Plus the RAPM data from 2002-2006 has a lot of the season missing which depresses the value of the players. I wish J.E. would go more in depth with some of the results from that time period and the 2007 season, but I doubt that will ever happen.

[Blackmill]

Plus the RAPM data from 2002-2006 has a lot of the season missing which depresses the value of the players. I wish J.E. would go more in depth with some of the results from that time period and the 2007 season, but I doubt that will ever happen.

Is this because of the particular data set used or does there not even exist complete pbp data prior to 2007?

[colts18]

Plus the RAPM data from 2002-2006 has a lot of the season missing which depresses the value of the players. I wish J.E. would go more in depth with some of the results from that time period and the 2007 season, but I doubt that will ever happen.

Is this because of the particular data set used or does there not even exist complete pbp data prior to 2007?

The dataset that J.E. used came from ESPN.com. At that time, ESPN didn't have the pbp for every game. But, B-R currently does have pbp for all games from 2001-onwards. As a result, J.E. did run a NPI RAPM which you can find on my site. Click on any of the 2001-2006 for a NPI RAPM that has full season data.

[Doctor MJ]

This thread finally convinced me to make an account after months of lurking. Anyways, I've found this entire discussion very useful and interesting. I also have a question about RAPM that I've been meaning to ask. Looking at Colt's data (I think the stats-for-the-nba site is xRAPM but correct me if I'm wrong) I noticed that the 2002/2003 RAPM is significantly lower than the late 2000s RAPM. Shaq, perhaps one of the greatest offensive players ever, only posted a 2.1 on offense in 2002 and 3.8 in 2003 compared to frequent 5+ offensive RAPM ratings in 2007 and onwards. I think its fair to say that many of the players posting 2+ or even 5+ offensive RAPM weren't more impactful than Shaq. Is this discrepancy due to the large error in RAPM ratings? Is it because 2002 is the first prior-informed year in Colt's data? Or is there some other reason why players generally have better RAPM in the late 2000's?

The nature of RAPM is that the ratings become detached from actual basketball points. When you look at pure APM, a +1 rating means that the guy is estimated to have an impact of 1 point per 100 possessions. The machine learning in RAPM by contrast will make the adjustments it needs to get it's closest match to what the data tells us it should get. This is why when I put my spreadsheet of PI RAPM together, I normalize it by taking the standard deviation of each year, and adjusting accordingly.

This is not to say that there can't be further trends making the earlier years more likely to have lower numbers, but speaking most generally, you shouldn't look at it as if the later years are literally saying those guys were more impactful than earlier Shaq was until you do a basic adjustment, because there's no reason to assume you're looking at an apples-to-apples comparison.