"We just need a change in coordinator."
Whether the offending coordinator is Greg Robinson or Greg Davis, the refrain is familiar. There's anecdotal evidence to support the benefits of a coordinator change, most recently with Manny Diaz at Mississippi State or Dana Holgorsen at Oklahoma State. But I wanted to attempt a more systematic look at the effects of coordinator changes. This diary is an update of a post in December and incorporates responses to many thoughtful comments to that earlier version.
This is a loooong diary. There's few steps to get to the end, and I know many tl;dr'ers are impatient. Putting the conclusions first seems a bit like putting the cart before the horse, but…. cart?
In a nutshell: firing a defensive coordinator tends to 'work' (to a degree), but firing an offensive coordinator does not. The number of returning starters turns out to be much more important than whether or not the coordinator was replaced. Because coordinators tend to be fired from poor-performing teams, often what we see as a positive boost from a coordinator change is simply regression to the mean.
As an added bonus, I stumbled upon these crude models to predict a change in a team's FEI rank vs. the prior year, where a positive change in rank is desirable (e.g. moving from a rank of 30 in the prior year to a rank of 10 in the following year is a +20 in rank):
Change in Offensive FEI Rank from Prior Year: +14 positions in rank if the starting QB returns, +3 for each additional returning offensive starter, and -0.5 for each level above last that the team was ranked in the previous year.
Change in Defensive FEI Rank from Prior Year: +3 for each returning defensive starter and -0.3 for each level above last that the team was ranked in the previous year.
These models certainly aren't going to put Football Outsiders out of business, but they're easy to use and somewhat intuitive. More detailed explanations below.
Team performance is based on the Fremeau Efficiency Index (FEI) from FootballOutsiders.com. Their free published data only goes back to 2007. FEI is not perfect, but it's easily available and eliminates some of the noise in scoring or yardage data. Because the focus will be on change in FEI performance, there are three years of data available (2008 vs. 2007, 2009 vs. 2008, and 2010 vs. 2009).
To determine coordinator changes, I used Rivals.com's annual "coordinator carousel." I coded the coordinators into four categories – stayed, promoted, fired/demoted, and left/unknown. A coordinator was classified as 'promoted' if he got a coordinator job at a 'better' school (arbitrarily determined by me, based mainly on the conference of the school) or if he got a head coaching job at any school. A coordinator was classified as fired if he didn't get a new job, took the same job at a 'worse' school, or took a position job at any school. The 'unknowns' are mainly coordinators who went on to take a position in the NFL or the same position at a similar school – in many cases it's hard to determine if that's a promotion or a demotion. Nearly all Michigan fans believe Jim Herrmann's trip to the NFL was encouraged, but for some coaches a job in the NFL could be their desired career path. I did some Googlestalking to try to parse out which was which, but if I could find no definitive sentiment I just coded them into the 'unknown' bucket . The coding was a bit tedious and I would welcome anyone who wants to double-check or validate my coding.
An issue confounding coordinator changes is that they often come with a head coaching change as well. If a head coach comes in with a whole new staff and the FEI metrics improve, is that because of the head coach or the coordinators? So I separated out coordinators that came on board as part of a new coaching regime (e.g. Malzahn) and those that came on with an existing head coach (e.g. GERG).
With 120 FBS teams and 3 seasons, there are 360 total data points (technically 359 since Western Kentucky was new to the FBS in 2008). Here's a breakdown of what happened with the coordinators. Keep in mind that when a head coach changes, the coordinators also go.
And now please say goodbye to the 'unknown' category, because we'll be ignoring them for most of the rest of this piece.
The Exceptions: Best/Worst Firings
To judge the 'best' or 'worst' firings, I calculated the change in the team's Offensive or Defensive FEI rank from the season before the change to the season after the change. This is simplistic, to be sure, but it's also clean and easy to understand. I looked into using the actual FEI metric, but the metric values seem a bit more volatile than the rankings. And the ranks are frankly easier to deal with/explain. I also looked at change in performance 2 years after the change to account for more time for a coordinator's influence to take effect. If anything, the evidence is weaker 2 years down the line, so I focused on 1 year changes.
And here's the best and worst firings of coordinators, judged solely by the unit's movement in FEI in the season after the coordinator change.
For those curious:
Manny Diaz's job in 2010 with Mississippi State reflects the 6th best performance following a DC getting fired (+59 places in FEI rank)
Greg Robinson's first year with Michigan ranks as the 6th worst performance following a DC getting fired (-25 places)
The best change in DFEI rank for a DC who didn't get fired was North Carolina State in 2010, under Mike Archer (with a Jon TAHNOOTA boost?) (+70 places)
The best change in OFEI rank for an OC who didn't get fired was San Diego State in 2010, under hey, that's Al Borges! (+81 places) [Brian's piece on Borges had SDSU as improving only 67 spots in FEI rank from 2009 to 2010, but I think that was using pre-2010 bowl season FEI numbers]
One other curiosity – the top changes in performance after a coordinator was fired all occurred in 2010. This is strange. We know Fremeau Efficiency isn't perfect, but it uses the same methodology in each year. I can think of only two explanations: 1) it's a fluke 2) coaches/athletic directors are getting more astute about when to fire/not fire a coordinator. My guess is it's the former, but I'm open to other interpretations.
The Rule: General Trends with Coordinator Replacement
As we move from looking at particularly good or bad firings to looking at the overall picture, I need to make this point clear: A firing by itself does not cause an improvement in FEI performance. Obviously, the who's matter (Shafer to GERG, anyone?). And the aggregate averages we look at include outliers on either end (and those outliers are potentially the cases where the 'who' really does matter, for good or bad).
A couple hypotheses to test:
H1) Teams that fire their coordinators should improve more dramatically in FEI than teams that stand pat with their coordinators. Head coaches will typically fire a coordinator only if he is perceived to be underperforming with his unit, so a change of coordinator should mean more of a rebound in FEI than in normal circumstances.
H2) Teams that lose a coordinator due to a promotion should decline more in FEI than teams that stand pat with their coordinators. The thinking here is that a coordinator must have been overperforming with his unit to merit a promotion, so his departure will coincide with the unit declining more than normal.
For the same reasons discussed in the previous section, we'll evaluate the hypotheses based on the change in a unit's FEI rank from the previous season to the current season. And looking at our three years of data across 120 FBS teams get this:
Offense first. Here we see little support for H1) – a team that fires its offensive coordinator improves by 2.3 positions in FEI rank, on average, but teams that stand pat improve by 1.1 positions in FEI rank – virtually no difference. The story is the same even if we look at performance two years out (not shown here). We do see some support for H2) – a team that loses its offensive coordinator to a promotion tends to decline in performance by 7.8 positions in FEI rank the following season.
For the defense, the pattern is reversed. We see strong support for H1) – a team that fires its defensive coordinator improves by nearly 12 positions in FEI rank, on average, while a team that stands pat with its coordinator has almost no change in FEI rank. But we don't see much support for H2) – a unit that loses a defensive coordinator to promotion drops by only 3.1 positions in FEI rank.
By now you're thinking three things:
Holy shit, this is too long!
What about the players a coordinator inherits? Did the best players graduate, or did an inexperienced group get more seasoning? You can't look at coordinator performance without considering the players.
What about regression to the mean, or the tendency of units that do incredibly well in one year to slip the next year, or for units that do incredibly poorly in one year to improve the next year – regardless of who is coaching/coordinating?
For 1), you're right, and we're not even halfway! For 2), you're right, but hold on a second. Let's look at 3).
First, we'll look at teams that finished in the top 60 in FEI in the previous season. We'd expect the teams to decline in performance, in aggregate, simply because of regression to the mean. But we can still evaluate our old friends H1) and H2).
And the verdict is… basically no support at all for H1), for either offense or defense. If a coordinator of a top 60 team is fired, the typical team performs about the same as the typical team where a coordinator stayed. We see some support for H2), but only for offenses; top 60 teams that lose an offensive coordinator to promotion tend to fall back even moreso than top 60 teams that stand pat. And oh by the way, top 60 teams that change their entire staffs tend to really drop off the next year.
So let's do the same for the bottom 60 teams in FEI.
For H1), we see strong support for the defensive side but no support for the offensive side. A team that fires its defensive coordinator tends to improve 24.4 positions in FEI rank, while a team that stands pat tends to improve only 9 points in defensive FEI rank [Beating a dead horse note: I'm not saying that firing a defensive coordinator causes the FEI to improve, only that there's an association between the two. The real cause, most likely, is that the unit was underperforming badly vs. the head coach's expectations, which caused the coordinator to get fired]. On the offensive side, a team that fires its coordinator actually performs worse than a team that stands pat, on average.
H2) gets no support for either offense or defense. Sample sizes are small (because not many coordinators from teams in the bottom 60 in FEI get promoted), so mileage varies, but these teams didn't appear to have any difficulty replacing their promoted coordinators.
Double bar charts! What do they all mean?
More or less this:
Among teams that finish in the bottom half of FEI rankings, those that fire a defensive coordinator outperform their stand pat counterparts by about 15 points in FEI rank, on average. This is the only time we see a substantial positive impact from canning a coordinator.
Among teams that finish in the top of half of FEI rankings, those that lose an offensive coordinator to a promotion underperform their stand pat counterparts by about 11 points in FEI rank, on average. So either offensive coordinators get out while the getting is good, or those promoted offensive coordinators really were/are offensive geniuses whose team suffers without them.
Adding Player Quality into the Mix
In the first version of this article I basically punted on trying to quantify the quality of the rosters. Obviously a coordinator inheriting a bunch of bad freshmen will not fare nearly as well as a coordinator inheriting a roster full of good upperclassmen. To quantify the quality of the roster, I'm using Phil Steele's favorite metric, returning starters. Returning starter data comes from Vegas Insider in 2008, Phil Steele's blog post in the Orlando Sentinel in 2009, and Phil Steele's own site for 2010. I decided against using recruiting rankings not just because it was more work (especially splitting out offense vs. defense), but also because teams tend not to vary too much in recruiting year over year (Scout's team per recruit averages correlate at .9 year over year), and when they do vary a lot it tends to coincide with either one or two fluky recruits or with a head coaching change, which I'm already incorporating.
So, do teams that return more starters tend to do better in FEI? Well, Phil Steele would cry if the data didn't support it. I split unit/years into rough thirds based on the number of returning starters, and ...
We see a clear effect, moreso for the offense. A team in the top third of returning offensive starters tends to improve by 13 positions of FEI rank from one year to the next, while a team in the bottom third declines by 14 positions of FEI rank. The same story applies, albeit less dramatically, for defenses.
(Finally!) Putting It All Together
The main question is, once you account for returning starters, does the change in coordinators still matter? If you've made it this far, you're as tired of bar charts as I am. For this analysis I'm going to turn to the simple statistician's best friend, regression analysis. Regression analysis is great for having factors basically fight it out to see which is the better predictor of our target variable.
You can't do a regression without good prior beliefs, so I'll put 'em down here:
More returning starters should lead to a positive change in FEI rank.
Returning QB's likely matter more than other returning offensive starters, so we should look at them separately (just like Phil Steele does).
All other things being equal, it's likely a team at the bottom of the rankings will improve the next year, and likely a team at the top of the rankings to decline the next year. This is the regression to the mean hypothesis, which implies the previous year's rank should have an impact on the change in FEI rank.
If coordinator changes truly have an impact, it needs to show up after we account for returning starters and previous year's FEI rank.
[NOTE: the gurus at FootballOutsiders use both returning starters and recruiting rankings, along with a five year program success score, fluky turnover margins, etc. in their predictive models. Some information about what goes into their projection stew is available here and here.]
Predicting Change in Offensive FEI Rank
If we look at the offense, basically (1), (2), and (3) are confirmed and (4) is blown out of the water. Here's the full model:
A flag/dummy variable was used to test the effects of three of the four categories of coach changes (OC fired, OC promoted, and whole staff swept out – for statistical reasons, the intercept/'default' includes those situations where the coaching staff didn't change at all). I'll wait to interpret the impacts until we get to the reduced model. For now we'll just worry about our statistical confidence – higher is better. The typical rule of thumb is to look for 90% or 95% confidence (which means that in only 1 in 10 or 1 in 20 samples would we see nonzero effects when in fact there were zero effects). We have four variables with very high confidence, so they'll stay in the model. But the variables reflecting a coordinator change have a low confidence, meaning we can't reject the hypothesis that coordinator changes have no effect on change in FEI rank. In cruder terms – coordinator changes don't seem to matter on the offensive side of the ball.
And now the reduced model, dropping the irrelevant coordinator variables. The R-squared (overall measure of fit) is unchanged at 0.32.
Even though you read MGoBlog, you're probably more of a football fan than a stats nerd. So some interpretation:
The R-squared of 0.32 is pretty good, considering we don't have a lot of inputs in the model.
The intercept/default is the predicted change in Offensive FEI rank if all the other variables are at their minimum. In effect, if a team was first in Offensive FEI rank and had no returning starters, its predicted change in FEI rank would be to go from 1st to 59th – a true regression to the mean.
The Offensive FEI Rank coefficient basically means a team is expected to gain a half a point in rank for every point of rank it had in the previous year. If the 120th ranked team has no returning starters, its predicted change in rank is -58.1 + 120*(0.5), or +1.9. In other words, it would be predicted to finish 118th in FEI rank, or just about at rock bottom again.
Returning offensive starters is fairly straightforward – each returning starter is worth +3.2 points in Offensive FEI rank.
Phil Steele thinks a returning quarterback is special, and the data supports his claim. While a typical returning starter is worth +3.2 points, a returning quarterback gets a 10.8 point bonus, for a total impact of 14 ranking points.
To help get a handle on what the model is saying, below are some examples of hypothetical situations.
The model's not perfect – it can't predict any team to be ranked above 13th or below 118th in the following year. And in R-squared terms, the majority of variance in the data is left unexplained. But all effects are statistically significant, and the model makes intuitive sense. You may recognize the last line of the table as Michigan's current situation. If Michigan behaves according to the model, the change in coaching will have no impact and Michigan will regress to the mean a bit, falling from 2nd to 20th. Of course no particular situation, including Michigan's, is guaranteed to perform as the model predicts, but it's a prediction we can test.
Predicting Change in Defensive FEI Rank
QB's don't play defense, and Phil Steele doesn't call out a particular defensive position as critical, so we're testing three hypotheses: that more returning starters = improvement in rank, that worse previous year's rank leads to better next year's rank, and that a coordinator change can have an impact on FEI performance. Table?
The story here is a bit more complicated. As with the offensive side, the previous year's rank has a clear impact, as does the number of returning starters. Also similar to the offensive side, the promotion of a coordinator has no statistically significant effect. But the other two flags have borderline 'significance'. Keeping the head coach but firing the DC leads to a 7.5 position gain in FEI rank, while firing the DC *and* the head coach leads to a 6.4 point decline in FEI rank. Because they are borderline, I'm going to leave them in and just drop the "DC promoted" variable from the final model, which is below.
R-squared of .23 is not as good as with the offensive side of the ball, perhaps because of the absence of a single key returning starter as with the QB on the offensive side. [Note: if we used actual FEI, not FEI rank, the R-squared is much stronger (about double) but with the same general results. However, interpreting FEI changes is not as transparent as interpreting FEI rank changes, so I'm sticking with rank changes here. On the offensive side, performance of the model is about the same whether we use FEI or FEI rank].
Interpretation of the intercept/'default' is the same as with the offense. Assuming all other variables are at minimum (i.e. previous FEI rank was 1st, 0 returning defensive starters, no DC change), then the team is predicted to drop to 40th in FEI the following year.
The intercept is smaller in the defensive model, but so is the 'reward' for having a bad FEI rank the previous year, so it balances out. A team that finished 120th in Defensive FEI in the previous year, has no returning starters and no changes in coordinator, is predicted to change in rank -39.2 + 120*(0.3), or -3.2 points, to a rank of 123. OK, that's not possible (there's only 120 teams), but it's close enough and again shows some intuitive power of the model – take a crappy defense and return no starters, and it will remain crappy.
Remarkably, the worth of a returning defensive starter is virtually the same as the worth of a returning offensive starter.
In situations where the defense is so bad the coordinator is fired, teams tend to get a 7.3 point boost in FEI rank. But if the head coach is also swept out, teams tend to drop 6.6 points in FEI rank.
As before, below are some hypotheticals to help show how to interpret the model.
Once again the last line is Michigan's predicted outcome for 2011 based on the model – a modest improvement to 94th in FEI. Note that if Rodriguez had stayed but Robinson was fired, the model would have predicted a defensive FEI rank of 80th in 2011. For the love of all that is holy this is just presented for interest's sake; we have no idea what would have been, and of course we have no idea of what will be until it actually happens. I for one certainly hope (and believe) Mattison's defensive unit will outperform the model's expectation.
The model is not saying that defensive coordinators should be fired every year in order to get a 7+ point boost in FEI rank. Even though we're using regression, this is a descriptive model, not a normative one – it's only describing what has happened, not prescribing what should happen.
Both models have statistically significant results. But they're nowhere close to perfect predictors of performance – in statistical terms, a lot more variance is left unexplained than explained. So while we can use the model to make rough predictions, your mileage will definitely vary in individual situations.
I don't want to diminish the importance of individual coaches. Manny Diaz may bring golden blitzes of thunder and rage wherever he goes, and GERG may bring doom wherever he goes. It's just that in the aggregate, we don't see much evidence for clear changes in unit performance based on a coordinator switch. If the trends of 2010 continue, however, that may change.
Another way coordinators can impact the team long-term is with recruiting. If a coordinator happens to be a heckuva recruiter along with being a decent coach, that will likely pay dividends longer down the line. That's not investigated here.
The guys at FootballOutsiders do a much better job of prediction than this model. The only issue is that their models are both more complicated and less transparent. And I'm not sure if they've ever tested coordinator changes. In any case, this article from 2008 says their model had a correlation of .8 for predicting next year's FEI – which corresponds to an R-squared about double the models above.
Wrapping It All Up
For me, this analysis has three big surprises:
No matter how we slice it, changing an offensive coordinator can't be tied to a systematic gain in offensive performance. Maybe coordinators are fired more for philosophical or chemistry differences than for performance-related issues.
In the aggregate, firing a defensive coordinator does correspond with a boost in the unit's performance, but it's not huge – roughly equivalent to the benefit of having two more returning starters on defense.
Roster quality, as measured by returning starters, has a clear positive impact on change in FEI rank, and it's almost exactly the same for returning offensive and defensive starters (QB's excluded).
Another surprise is that Al Borges, happily, pops up twice in the good column – after he left Auburn its offense went in the FEI tank, and in his last season at SDSU his offense improved more in FEI rank than any other unit that did not change coordinators in the three years of data I examined.
Things I'll think more about:
Whenever I hear an offensive coordinator is fired, my first reaction will that it's a short term desperation move. My crazy prediction: Mack Brown's removal of Greg Davis is an indicator that Mack's in his last few years of being a head coach.
Fremeau may be flawed. By all accounts Auburn's offense was in decline when Borges was fired in 2007, and yet its offense was ranked 24th – Fremeau may be 100% right, but in some cases the FEI is clearly contrary to conventional wisdom.
Was the spate of 'good' firings in 2010 a one-time fluke, or part of a trend? The only way to tell is with more years of data. It is possible that coaches/AD's are getting better at knowing when to fire/not fire a coordinator, and who to hire as a replacement?
It could be interesting to look at average recruiting rankings of a unit in the two years prior to a coordinator change to the two years after a change. I think the focus would have to be on a per recruit basis, not per class basis, to make up for uneven class sizes and roster needs.
[EDIT: new] Either now, or after a bit more tinkering with the model, we can go back and look at teams/units that overperformed (or underperformed) relative to the model's expectations. For instance, and not surprising, it looks like Michigan's 2010 defense underperformed the model's expectations by 40+ spots, while the offense overperformed the model's expectations by 50+ spots. In fact, GERG's D underperformed expectations 2 years running, which is more statistical evidence of his incompetence. And coaches/coordinators who systematically overperform expectations could be true geniuses/motivators. By the way, if you look at the teams that had a high net underperformance in a certain year, many were either in their first year under a new coach (Washington St. 2008, Kansas 2010) or had a head coaching change the following year (Washington 2008, San Jose State 2009). Need a little more time to think about/look into this one.
This piece is still a work in progress, and there may be blind spots I haven't even considered. At the very least I'll try to update it next year with a new round of data. As ever, comments/feedback, especially of the constructive variety, are welcome. Go Blue!