Does Experience Improve Performance?

In my earlier post, "First Year Performance - An Evaluation Method", I discussed how, over the history of the NFL, from 1940-2008, a quarterback's passing performance, as measured using CMI, and more specifically, the standard deviation from the mean of CMI, changes over time (in that post, see in particular the third graphical illustration labeled "CMI - SDs from Mean - By Qualifying Year").  That illustration pointed out that it takes about 5 years for a quarterback to become "average" (or, better than average).  That's obviously only true for quarterbacks who qualify in at least that many years.  Meaning that since the graphical illustration includes all quarterbacks (who have thrown enough passes to qualify in at least 1 year), that the earlier years could be biased by quarterbacks who only qualify in  those earlier years, and not the later ones.  

This is the issue pointed out by a commenter regarding that post.  He (she) was suggesting that I have a biased sample, in that the later years are only populated by quarterbacks who throw for that many years, and hence, performance should be expected to improve, as the NFL weeds out the under-performers of the earlier years.

Here's the comment:

"This effect could be due to sampling bias.  It may be that there are a large pool of first year QBs.  Some are good and some are bad.  The good ones make it to year four, and the bad ones are cut, released, or benched before they can play four years...."

He (she) then went on to ask the following question:

"...if you restrict the analysis to only QBs that have played >= 4 years, do you get the same result?"

Before I answer the question, let's address the comment.  

First, the commenter does make a very valid (and what would seem to be an obvious) point, in that those QBs that make it to year 5 and beyond, must necessarily be the better ones, as the under-performers drop-off.  This point has several implications though.  The first is that the NFL does not look at quarterbacks (or, more accurately, their passing performance) the way I do.  So we can't simply assume that the point is in fact that obvious.  Second, perhaps the way I look at quarterbacks passing performance, CMI, is in fact, a decent (good?) measure of a their true passing ability, and, the graph validates that fact (this would be rather presumptuous of course, but that is what I am trying to prove after all - that I have, in fact, developed a method that appropriately evaluates a QBs passing performance, and, perhaps even more importantly, that I have developed a method by which one can identify a good passer versus a poor passer early on).

Second, the commenter suggests that my sample is biased because of this fact (i.e. it is biased towards the better players in the later years).  There are two ways we can look at that.  The first is that the sample isn't biased because it uses the entire population.  In other words it is what it is (and perhaps this again validates my measurement technique).  The second point is that the sample is biased (remember, however, the context in which the graphical illustration was shown - in that it was trying to look at all quarterbacks, and how their performances changed over time), and that we can, if we choose to do so, look at it differently.  This is, in fact, what the commenter was suggesting with their question.

So let's take a look.  As a matter of fact, let's take the question a bit further and look at it in four different groupings:  

QBs qualifying in at least 1 year (the original graph) - 345 Quarterbacks

QBs qualifying in at least 3 or more years - 180 Quarterbacks

QBs qualifying in at least 5 or more years - 121 Quarterbacks

QBs qualifying in at least 7 or more years - 84 Quarterbacks

The graphical illustration is followed by a tabular display of the same data.

There's no question about it.  Performance improves as a quarterback gains experience.  All groupings follow the same basic pattern:  Sub-par performance in year 1, followed by several years of gradually improving performance.  When the improvement takes place, and by how much, does vary by group.  For example, after three years, teams still haven't been able to markedly separate the good quarterbacks from the bad ones.  By the fifth year, it's now quite apparent who has "cut it", and who hasn't.  For all quarterbacks qualifying in at least 7 or more years, they appear to show their skill, or differentiate themselves, as early as year two.  So, to prove the obvious point that the commenter in the previous post making, those that don't turn in positive performances early do tend to get weeded out.  The process, however, happens to take a few years.

The question this prompts however, is, why does it take so long to weed out the poor performers?  

Well, there are at least two different answers.  

The first, something I alluded to earlier, is that the NFL doesn't evaluate passers the way I do.  I am a statistical analyst, and what I have created is a an objective methodology to evaluate a player's performance.  Teams in the NFL, aren't inclined to do the type of statistical analysis that I do.  In addition, it's just simply not clear, even using my statistical methods, that you can be sure that you can identify the good and bad future quarterbacks after year 1.  By year 2, however, year 3 the latest, my method does show who the keepers should be.

The second is, that teams pay a lot of supposedly smart people (scouts, GMs, etc) a lot of money to evaluate, and draft good talent.  Subsequently, these teams draft these "talented" players early in the draft, and pay them lofty sums of money, a lot of which is guaranteed.  Now what do you do?  You've got a poor-performing team (generally, a team's draft order is determined by their won-loss record, with the teams with poorer records drafting earlier, and vice-versa) that just drafted the next "Peyton Manning", and have paid them a lot of money, and they have a lousy first year.  You're not likely to blame the young quarterback for that year's performance.  The team still has a lot of weaknesses elsewhere, so you pick early again the next year (perhaps shoring up the defense).  Second year passing performance is better than the first, but still not better than average.  You even ignore the statistical evidence (he can't be that bad).  You're still paying him lots of money, you've seen improvement, team has won perhaps a couple of more games, and this gives you hope that, the third year is THE YEAR.  Then, much to your dismay, the quarterback that you paid millions of dollars for, doesn't pan out as you have another miserable year (the teams is "close" to 0.500).  You get impatient,  fire the head coach, try an aging veteran as your quarterback, he has a decent year, a quarterback controversy brews, ya da ya da ya da....and, you continue to ignore the statistical evidence....

Just off the top of my head, I can think of several recent obvious examples - Alex Smith of the Forty Niners (and, proving the corollary, the approach taken by the Green Bay Packers with Aaron Rodgers), and JaMarcus Russell of the Oakland Raiders.

Here are three quarterbacks that the NFL, teams, and fans, have fallen in love with, but, statistically have yet to prove themselves.

Eli Manning (I can hear the argument - but they won the Superbowl)

Tony Romo

Michael Vick

So there you have it.  My method is the Moneyball approach.  What actually happens in the NFL is quite different.  NFL teams could save themselves a lot of money (and perhaps even win a few more games) if they were to adopt a statistical approach to their business (I think a lot of NFL teams actually believe they take a statistical approach, but I don't think that most actually go through rigorous analyses the way I do).  

My next post will be my preseason rankings for 2009.

The Best and Worst of 2008

This is a simple post.  Really simple.  The list should pretty much speak for itself.  The best and worst of 2008.  I use the same criteria I've been using the past few posts.  CMI, and, more specifically, standard deviation of a player's CMI relative to the mean.  30 players attempted enough passes in 2008 to make the list.  

The Best 3:

3.  Jeff Garcia, Tampa Bay Buccaneers - Why no team has picked him up is a complete mystery to me - the man can still play, even though his durability is certainly a question.  Completed nearly 65% of his passes, and less than 1.6% of his passes were intercepted.

2.  Peyton Manning, Indianapolis Colts - Do we even need to say anything? Coming off off-season knee surgery, Mr. Consistency now holds the highest current active streak of games started consecutively - 176 games. He completed almost 67% of his passes.

1.  Chad Pennington, Miami Dolphins - Let go by the Jets so that they could welcome Brett Favre (that worked out well for the Jets, didn't it?) and subsequently picked up by the Dolphins. Had a terrific season, completing more than 67% of his passes, and less than 1.5% of his passes were intercepted.

In previous posts on my blog, I have spent time discussing each of the above 3 players.

The Worst 3:

28.  Tyler Thigpen, Kansas City Chiefs - Won't show up at the bottom of the list in 2009.  Why?  Matt Cassel (#10 in 2008).  The Chiefs made a deal with the New England Patriots to get Cassel (and Mike Vrabel) for a 2nd round pick in the upcoming draft.  In 2008, Thigpen, replacing Damon Huard, completed fewer than 55% of his passes.

29.  Gus Frerotte, Minnesota Vikings - Replaced an ineffective Tarvaris Jackson in the 3rd game of the season, started 11 games, and then was himself replaced by Jackson against the Lions.  Completed less than 60% of his passes, and at almost 5.0%, was one of only three quarterbacks whose interception percentage exceeded 3.0% - the other two being Ben Roethlisberger of the Steelers (3.2%), and Brett Favre of the Jets (4.2%).

30.  Derek Anderson, Cleveland Browns - Barely completed 50% of his passes, then got hurt.  Was replaced by Brady Quinn.  Is it Quinn's time in 2009?

Here's the entire list.

My next post - a pre-season ranking for 2009

First Year Quarterback Performance - An Evaluation Method

This past weekend, the New England Patriots traded Matt Cassel to the Kansas City Chiefs.  In fact, they traded linebacker Mike Vrabel and Cassel for a 2nd round pick (pick #34) in the upcoming April draft (see here for the details).  Depending on what your news source is, this was labeled as either a good deal for Kansas City, or a bad deal for New England.  The experts discussing this issue don't seem to support their arguments by any objective measures - they just opine.  I am not going to jump into the debate here.  I will offer this however - we will probably not know who will end up on the better end of this deal until many years from now, and even then, we may not know.  

What about the objective evidence?

Matt Cassel is not the only quarterback in 2008 that had a good year as a first-year quarterback.  There was Aaron Rodgers of the Packers, Shaun Hill of the 49'ers, and Matt Ryan of the Falcons.  There were four other quarterbacks who threw enough passes this year to make the list of first-year quarterbacks - Ryan Fitzpatrick of the Bengals, Joe Flacco of the Ravens, JaMarcus Russell of the Raiders and Tyler Thigpen of the Chiefs.  

Notice here that I didn't use the term rookie quarterback, but instead used the term first year.  The way I look at a quarterback, is that I only consider a quarterback who has thrown "enough" passes to qualify in a given year.  The # of passes has varied by year (as opposed to the fixed standard used by NFL.com).  For the 69-year period from 1940-2008, my database contains 1,451 passing seasons, 336 of which were by a quarterback qualifying for the first time.

How did the aforementioned 8 quarterbacks do?  

I use a statistical measure I have previously developed and discussed, called CMI (Completions Minus Interceptions, calculated as Completions/Attempts - 3*Interceptions/Attempts) instead of the NFL Passer Rating system (using the NFL passer rating system gets you essentially the same answers), and then use the standard deviation from the mean as the measuring stick (for the pool of qualifying passers, I calculate the the mean and standard deviations for each year, and relate a particular performance in that year to the mean using that year's standard deviation).  Using the mean and standard deviation this way not only allows me to compare these 8 quarterbacks relative to each other, but also relative to all quarterbacks this year, or any other year.

In a post not too long ago, I showed how, over the past 69 years from 1940-2008, when the standard deviation is used as the measuring stick, the accumulated data form a near-perfect theoretical standard normal distribution.  And I also showed how, this is true regardless of whether one used the NFL passer rating system or CMI as the statistical base.

I can now show you how these first year quarterbacks did compare to the rest of the QBs in 2008, using this measure.  But this picture may not tell us much, if anything.  Here's the graphical illustration:

Perhaps the only thing that may be apparent from this is that first year quarterbacks don't do nearly as well as quarterbacks with more experience.  Even drawing that conclusion from this is a little iffy.  However, it is clear (at least by this measure) that Cassel had the best year of the 8 first year QBs.  Rodgers was a close second.  JaMarcus Russell and Tyler Thigpen had disastrous first years.  In an earlier post regarding Jeff Garcia, I had suggested that Kansas City might be a good place for Garcia.  Scott Pioli, the new Chiefs' GM apparently was thinking the same thing - upgrade at quarterback, and he did - with Cassel.

Going back to the standard normal curve, let's look at just the first year performances.  Recall that, when all 1,451 seasons are aggregated in a histogram, it looks very much like a standard normal curve.

This doesn't quite look "normal", does it?  The first year performances appear skewed to the left, or, compared to all quarterback performances, appear to under-perform.  Since the total database is normally distributed, then performances in years 2 and beyond must obviously be skewed to the right.  This prompts the question: does experience improve performance, and, in particular, how and when?

One way to look at this is to take the entire database, and look at the average standard deviation from the mean for each additional year a player qualified.  Here's what that graph looks like.

Wow!  I didn't realize that it takes about 5 qualifying years to turn into an average quarterback.  And that's just for those playing that long!  Even though this graph goes out to 18 years, there is hardly any data beyond the 15th year (there are only 8 quarterbacks who have had qualifying seasons 16 different years).  According to this illustration, basically, first year quarterbacks, as a group, struggle mightily.  They then gradually improve each year until about year nine.  Quarterbacks, on average, must still endure sub-par years in years 2-4, and for those that survive, their reward is about four more years of improving performance, assuming they stay healthy for that long.

Note: This analysis - the method of breaking down a quarterback's expected career in terms of passing performance as it relates to experience and how it is expected to change over time, is the first I've seen published publicly.  But I digress.

Ok, so we've taken a look at these 8 QBs relative to all quarterbacks in 2008, and we've taken a look at first year quarterbacks in general.  So the next step is to take a look at the best first year seasons.  In my database, there are 336 first year seasons in all (in other words, in the 69-year period from 1940-2008, 336 different quarterbacks threw enough passes in a season to qualify at least once).  The table below shows the 75 best.

Cassel (#48) and Rodgers (#60) are the only 2 of the 8 in 2008 that crack the top 75 all-time.  On the other end of the scale, we get JaMarcus Russell (#264) and Tyler Thigpen (#290).

When Matt Ryan (#128) got off to a solid start, and the Falcons surprised many by getting into the playoffs, there were many debates as to whether his season was one of the best ever by a first year player.  Not according to this measure.  Often, during those discussions, Kurt Warner's first year was brought up.  In looking at the table above, well, he did have an outstanding first year in 1999.  It ranks #6 all-time in terms of first year seasons.  The best ever - Roger Staubach's 1971 season (Staubach was a rookie in 1966, and played in both 1969 and 1970.  However, he didn't attempt enough passes in either year to qualify).  

Notable quarterbacks high on this list: Joe Montana in 1980 (#4), Tom Brady in 2001 (#14) - (see my post earlier comparing Brady's season to Cassel's here), Brett Favre in 1992 (#15), Dan Marino in 1983 (#21), and Johnny Unitas in 1956 (#22).  

Notable absentees:  Steve Young in 1986 (#154), Peyton Manning in 1998 (#208), Joe Namath in 1970 (#202)*, and Troy Aikman in 1989 (#311).  The all-time worst first year performance - Terry Bradshaw in 1970 (#337)*.

* Keep in mind that 1970 was the first year of the NFL following the merger, and 15 quarterbacks showed up in the database as having that year as their first year, although Bradshaw's 1970 season was his actual rookie season.

Are we done?  Hardly.  So after I went through this list, and still trying to objectively value Matt Cassel, I was interested in answering the following question(s):  what does the first-year tell us about a quarterback's future potential?  Do quarterbacks who have good first years, have good careers?  What about quarterbacks who have sub-par first years?  In other words, is there a correlation between year 1 performance and subsequent years?  When in doubt, it's always a good idea to take a look at the data.

Whoa!  The correlation between year 1 performance, and career performance, as measured by CMI standard deviations relative to the mean, is 0.75.  What that really means is that 57% of the variation in a quarterback's career performance can be explained by his year 1 performance.  I find that incredible.  So I looked a little deeper.  It turns out that 112 of the 336 quarterbacks only have 1 qualifying year.  And, for that group, the correlation coefficient is, well, 100%!  If I only looked at those quarterbacks whose career included at least 2 different years in which they threw enough passes to qualify, then the correlation coefficient drops to 64%.  Still quite remarkable.  

The table below breaks down the 336 first year seasons in terms of CMI standard deviations to see if we can glean any additional insight into whether a good first year translates into a good career and vice versa.


Well, the evidence is pretty clear.  Taken as two groups, players that have better than average (remember, the mean and standard deviation is relative to all quarterbacks, not just first year quarterbacks) first years tend to have longer careers and more productive careers than those that have sub-par first years.  And, breaking it down even further (in other words, just looking at the group whose first year was better than average, or the group whose first year was worse than average), you can clearly see that even within those sub-groups, the better the first year, the longer the career (although this correlation isn't nearly as strong), and the more productive the career. 

I haven't spent too much time on my blog discussing individual careers, except in the posts discussing Brett Favre and Jeff Garcia.  If I do this en masse, it would be such a powerful post - I'll basically be giving you my list of greatest passers to ever play the game, that I need to put a lot more thought into it than I have to-date.  In any case, some of the obvious names will show up on that list.  However, you'll find some exceptions as well (and I'm sure many of you will take exceptions to the list).  The point I really want to make here is that, when combining the table above, and the discussion following the list of the top 75 first years, it is quite remarkable how Troy Aikman "escaped" the trend, and turned into a great quarterback (or, more accurately, a great passer) - he truly turned out to be an exception (as for Young and Manning, even though they didn't rank well on the first year list, their first years weren't "that bad" compared to other first year seasons - as a matter of fact, Young's was "above average" compared to a typical first year).  Notice how I don't mention Bradshaw and Namath - it is because, statistically, their careers were sub-par.  When I publish my list of all-time greatest (and worst) passers, we'll get into it in more detail.

So, in trying to look at Matt Cassel's first year performance objectively, we've looked at how quarterbacks perform as their careers progress, how quarterbacks perform in their first year, who had great first years, where Cassel's season ranks on that list, and, how a first year translates into (or, is predictive of) a career.  Now perhaps, we can view Matt Cassel's 2008 year in an objective manner, and, if so inclined, you can at least evaluate the trade armed with some data.

Naturally, while I was putting this post together a question popped into my mind, and that is:  is a quarterback's second year a good indicator of their career?  What about looking at their first and second years?  Check back in a week or two to find out.