xGA/60:
last three seasons:
Subban: -.083, .053, .081
Niskanen: -.004, .076, -.072
Braun: -.281, -.004, -.165
Ok, now what are those numbers telling me?, because again I'm not so good with the advanced stats. It seems like PK is grading out in the middle between Braun & Niskanen?
PK has steadily deteriorated, minus is good, plus is bad.
Niskanen bounced back last year, but is more toward the middle of the pack.
Braun is excellent defensively.
Now some of this has to do with role, if you play a big offensive role, you'll be on the ice for more scoring chances against your team.
But a top defender will also be on the ice against the best lines, which will hurt his numbers.
Which is why 3rd pair defenders often have "inflated" advanced metrics, we saw that with Sanheim and his move to the 1st pair.
Be careful about comparisons with these metrics. While the use of them is correct, the analysis is technically wrong. This is due to not understanding the methodology behind the statistics.
From the way you wrote,
@deadhead , it seems like you’re looking at the marginal gradient- ie; how the player differs from the margin.
Re: “...,we saw that with Sanheim and his move to the 1st pair”
Re: “Niskanen bounced back... toward the
middle of the pack”
Those 2 remarks you made cannot be “true”, when using statistical comparisons- mainly because you’re looking for the aggregate stat and then measuring after deriving the average.
To find the measure you need a good least square measurement, which should obviously be the Best Linear Unbiased Estimator (BLUE). Since the data for this type of analysis needs to be unbiased, uncorrelated, and of equal variance- you can’t make those comparisons without further analysis.
Time-series data doesn’t work that way since it is serially correlated and the variance is determined through specific time periods (or regimes).
Serially Correlated: (TOI/60)
Regime Variane: (ATOI- Shift)
You would have to model using every stat, between the aforementioned players, on every second of their shift and then model for exact TOI differences. The error doing it your way makes the analysis “non trustworthy”.
EDIT: Even after doing this, the variance would not be “stationary”; or equal throughout.
————————
Not trying to be that person and get you on technicalities, but stats don’t tell every story- especially being told that way.