clock menu more-arrow no yes

Filed under:

Beating save percentage to death: Teams

New, comments

 

I used my Even Strength Save Data to look at the effect of Teams. There is not much there.

 

Using the dense data, I constructed the logits (ln(p/(1-p)) myself. I first corrected the observed save percentages for year.

 

> LinearModel.2 <- lm(CalcLogit ~ Team, data=Dataset, weights=ESA)

> anova(LinearModel.2)

 

Analysis of Variance Table

Response: CalcLogit

          Df   Sum Sq Mean Sq  F value Pr(>F)

Team      29   1871.2 64.524   3.8031  1.240e-10 ***

Residuals 895 15184.5 16.966


At a first glance, that looks promising. However, the sum of squares is less than 1/3 the sum of squares we saw for goalies. And we know that teams and individual goalies are highly confounded.


Looking at teams and goalies (with no interaction term)

> LinearModel.4 <- lm(CalcLogit ~ last.first+Team, data=Dataset, weights=ESA)

> anova(LinearModel.4)


Analysis of Variance Table

Response: CalcLogit

           Df  Sum Sq Mean Sq F value Pr(>F)

last.first 231 6041.3 26.153  1.6677  4.121e-07 ***

Team       29   601.9 20.755  1.3236  0.1209

Residuals 664 10412.6 15.682


Adding in goalies, the significance of the team effect goes away.


Looking at teams and goalies (with interaction)


> LinearModel.4 <- lm(CalcLogit ~ last.first*Team, data=Dataset, weights=ESA)

> anova(LinearModel.3)


Analysis of Variance Table

Response: CalcLogit

                 Df Sum Sq Mean Sq F value Pr(>F)

last.first      231 6041.3 26.153  1.6835  9.678e-07 ***

Team             29  601.9 20.755  1.3361  0.1156

last.first:Team 168 2707.5 16.116  1.0375  0.3772

Residuals       496 7705.0 15.534


So the effect of goaltenders remains highly significant. The team effect is not significant, and in magnitude is only about 1/10 that of the goaltenders. Finally, we can compare the models directly, to see if adding in team and the team*goalie interaction adds significance to the model.


> anova (LinearModel.1, LinearModel.3, test="F")

Analysis of Variance Table


Model 1: CalcLogit ~ last.first

Model 2: CalcLogit ~ last.first * Team

  Res.Df RSS Df   Sum of Sq F      Pr(>F)

1 693    11014

2 496    7705 197 3309.5    1.0814 0.249


And it does not.


Conclusion

Using weighted logistic regression, there is not a statistically significant differences between NHL teams in even strength save percentages over the period 1997-2010 if you control for the effect of individual goaltenders.