Before we get to more number crunching, you should check out the Bleacher Guy’s take on MSU’s seven-year conference title drought. It’s a critical, but balanced, commentary on MSU’s struggles in conference play in recent years, compared to the improvement over the course of the full season the early Izzo Final Four teams exhibited. As far as this season’s team goes, here’s the key paragraph:
But a lack of toughness doesn’t make you turn the ball over nearly 20 times a game. Toughness won’t help you overcome an offense described as “very predictable” by analyst Tim McCormick (another huge fan of Izzo, by the way), or offset a 21-0 deficit from three-point range in a game against Purdue. Sometimes it seems that the constant search for toughness, or for another leader in the mold of Mateen Cleaves, is a bit of a red herring. I want Michigan State to be a tough basketball team, but I’m not sure I buy “get tougher” as answer to every question that’s raised about the Spartans anymore.
Definitely click through to read the whole post, though. Note that I’ve linked to it twice to emphasize how worthwhile it is.
OK, now to the numbers. After my stab at using linear regression to examine MSU’s turnover woes, there were a couple of good suggestions for revisions to the model from the comment section faithful:
GBBound noted there may be multicollinearity between the opponent’s defensive TO% and defensive efficiency stats. (“Multicollinearity,” by the way, is a great word to throw around to make yourself sounds smarter than you really are. Which is not to say that GBBound isn’t as smart as he sounds.) In other words, defensive efficiency is partially based on defensive TO%, so including both as variables is redundant.
Hubert suggested a dummy variable for Big Ten opponents, on the theory that our conference foes know us better and can, therefore, come in prepared to disrupt our offense.
A third possible variable occurred to me: A time-based variable to see if MSU is improving or getting worse at holding on to the ball once you control for other variables. Taking these three concerns into account, here’s the list of independent variables for the new model:
AWAY: Dummy variable for away games. Playing in hostile environments could cause a team to turn the ball over more. (Note: Texas coded as home game; Missouri/BYU coded as away; 0.5 for UCLA game.)
DefTO%: Opponent’s defensive turnover % (for the season). The more turnovers an opponent tends to create, the more turnovers MSU is likely to make.
DefShoot%: A stat I cobbled together based on the opponent’s effective FG% and free throw rate (for the season). It’s the team equivalent of PPWS but on the eFG% scale. Replaces defensive efficiency as a measure of whether a team forces tough shots–as opposed to being good at creating turnovers. It could be MSU turns it over forcing the action against solid defenses.
PACE: My working theory has been that the slower the pace of a game, the more MSU bogs down in the half-court offense and turns the ball over.
BIGTEN: Dummy variable for whether the opponent is a Big Ten team.
TIME: Variable running from 1 to 24 for the sequence of the games MSU has played to date.
The dependent variable is MSU’s offensive turnover percentage (MSUto%) for a particular game.
Data was pulled today, so it accounts for two additional games. Here’s the equation resulting from the linear regression I ran in Excel:
MSUto% = 38.70 + (2.05 x AWAY) + (0.42 x DefTO%) + (.06 x DefShoot%) + (-0.38 x PACE) + (10.83 x BIGTEN) + (-0.82 x TIME)
This model has two advantages over the previous model:
The R-squared is higher: 0.54 vs. 0.44. So we’re explaining a little more of the variation in MSU’s turnover % from game to game.
The directions of all the coefficients are intuitive. MSU turns the ball over more (1) on the road, (2) against opponents that create more turnovers, (3) against opponents that force tougher shots, and (4) against Big Ten teams. The turn the ball over less (1) when the game is played at a higher pace and (2) as the season progresses.
The big difference versus the previous model is in which variables are statistically significant. Previously, they were the opponent’s defensive TO% and pace. Neither is significant in this model. Rather, the Big Ten dummy variable and the Time variable are now significant–both at the 0.025 level.
Implication: It’s playing Big Ten teams that are familiar with MSU that causes them the most problems in holding on to the ball. A higher percentage of MSU’s Big Ten opponents (1) create more turnovers defensively and (2) play at a slower pace, which accounts for why those two variables showed up as significant in the previous model.
As for the Time variable, this makes some sense in light of this graph:
The last time we looked at this graph–right before conference play started–we were feeling good about the downward trend in turnovers over the nonconference season. Then MSU turned it over 25%+ of their possessions against their first three conference opponents and we didn’t feel so good anymore. Since then, you’ll note there has been a subtle downward trend in turnover % again, albeit not as distinct as the first downward trend. The Big Ten dummy variable accounts for the spike going from nonconference play to conference play.
The regression results above indicate that (1) playing a Big Ten opponent makes MSU’s turnover % go up nearly 11 percentage points per game (!) and (2) after you factor everything else out, our turnover % is going down about 4/5 of a percentage point each game.
Those are some dramatic results. I’m not sure if their magnitude makes me more or less likely to believe this version of statistical reality. But it’s a pretty good reality: It implies MSU is, in fact, improving over the course of the season and can be expected to perform much better once we get out of conference play and into the NCAA Tournament.
I’m a tad skeptical, though, for three reasons:
- What my eyes see when I watch MSU play make it really hard for to believe this team is improving.
- If they are indeed improving, there is a natural limit to how much further improvement can occur. You can’t go down 4/5 of a percentage point of TO% per game indefinitely.
- At some point, won’t it be patently obvious to nonconference opponents how they should play us? Slow down the pace, force us into a half-court game, and do everything possible to disrupt us from running our set plays.
Time permitting, I’ll try to run last season’s data through this model to see if it holds up.
Anyway, here’s hoping the numbers trump my intuition in this case. An NCAA Tournament run would be great. Of course, adjusting our offense to become less predictable against Big Ten opponents would also be great.
Responses to comments:
1) Sample size is clearly an issue. Tough to do a study of a single team in a single season.
2) The interaction of TIME and BIGTEN is a key issue. Could be a fluke of the small sample size. The fact that both are staistically significant is curious, though.
3) There may be an issue in measuring quality of opponent outside the defensive variables.
4) I get a predicted TO% of 17.8% for the IU game.
In a real-world conversation, BoilerBrewer (Official Purdue Fan, Home Brewer, and Economist of the Spartans Weblog) notes that OLS really probably isn’t the best methodology here–since TO% can never go negative. To which I replied: It’s the only methodology I know how to use!
Here are the results running the model on last season’s data:
MSUto% = -2.96 + (1.06 x AWAY) + (1.93 x DefTO%) + (-0.41 x DefShoot%) + (0.19 x PACE) + (2.70 x BIGTEN) + (-0.32 x TIME)
R-squared is lower: 0.34. Coefficients, including the constant, vary pretty extensively from this season’s results.
The significant variables are DefTO% and TIME. BIGTEN has a pretty large coefficient, but doesn’t meet the statistically significant threshold.
I think, at this point, the thing to do is . . . give up. But hopefully this exercise has at least been useful to think through some of the potential issues causing MSU’s turnover woes. And I’ll stick with my intuitive judgement, which these models have generally confirmed:
MSU’s offense has to become less predictable, through some combination of pushing the ball more on offense and allowing more offensive freedom for our guards (Lucas, in particular–but Neitzel and even Walton, as well) when our set plays aren’t working.