Comparing Nuge’s Oranges and Draisaitl’s Oranges

Since Draisaitl’s contract signing, one thing I have heard a lot of people say is that this spells the end of Nuge [probable] in Edmonton, especially because he’s been something of a disappointment [FALSE! SLANDEROUS!]

It’s easy to see why this latter point is argued though – as far as genos and apples go, it hasn’t been a great year for Toddler Nuge (note: he’s no longer a baby). The evil snake/monkey that sat on Eberle’s back and kept biting 14’s hands all season spent plenty of time making Nugey look like a hunchback too.

 

Now, most people recognize that Nuge is the coach’s “hard minutes” guy: via PuckIQ.com, Nuge plays 42% of his time vs Elite competition, easily the highest percentage for any of the regular players, and so we cut him some minor slack for that as a result.

Is “minor slack” fair, though?

The McDavid Effect

While acknowledging QoC is a start, it’s not enough.

See, there’s a big challenge when it comes to assessing any Oiler not named Connor McDavid. And that challenge’s name starts with C and ends with OnnorMcDavid.

When Connor McDavid is on the ice, he has an outlandishly positive effect. For the Oilers he does … for the other team not so much. Unsurprisingly then, McDavid also has that same outlandishly positive effect on his teammates.

Generally speaking, the more time you spend with Connor McDavid, the better you will do. More goals, more shots, more corsis, more benjamins. (I won’t bother to present the hard stats-based evidence for this … seriously, does anyone want to argue the point?)

So I’ll argue that the implication of this outlandish effect is that, if you want to get a sense of how good a player actually is, you need to look at his on-ice results when he’s without The Franchise.

Nuge < Fayne???

Here’s an interesting thing though – Nuge spends basically no time with McDavid. He spent less time with McDavid this season than Mark Fayne – who played all of half an hour the entire season – did!

This means that Nuge’s numbers get absolutely no McDavid rocket sauce. If you’re going to compare Nuge with someone (like Draisaitl) who gets a sh*t ton of time with McDavid, you have to account for this effect. Have to.

Otherwise you’re comparing the lap times of an NSX with a Civic, and blaming the slower Civic lap times on the driver.

The Russell Effect

But wait! That’s not all.

See, there’s a second effect that takes place on the Oilers, around what has to easily be the most polarizing third pair defenseman in the entire g d league. Y’all know who I’m talking about.

I’m NOT going to spend any time analyzing Russell, a topic that has been covered to Hades and back by many, but simply point out the actual effect he has.

One of the reasons many fans (and presumably his teammates and coach) love Russell is that he’s the hockey equivalent of a soldier who throws himself on a grenade to save his squad. Russell’s like that. He will sacrifice everything and anything to prevent a goal, and you have to admire him for that.

Unfortunately, one of the things that Russell also sacrifices while he’s doing this is his own team’s ability to score a goal. Pretty much every Oiler scores less with Russell on the ice than when he’s off the ice – that’s just plain ol’ facts.

This is true even for McDavid, whose lines score like they’ve been touched by the Hand of God without Russell, but merely at garden variety supernatural levels with him.

Ryis Nugent-Russell

One weird thing though – Russell’s anti-offense effect is particularly acute when he’s on the ice with Nuge. Like, so acute it makes Ariana Grande look frumpy.

How bad is it? Take a look:

• RNH 5v5 CF% and GF% with Kris Russell: 44.8% and 27.8%.
• RNH 5v5 CF% and GF% w/o Kris Russell: 49.4% and 58.3%.

Check that line again: <28% vs >58%. This change in goal scoring rates is so dramatic it doesn’t even seem real.

I’m always super-wary of using GF% as a statistic because of its low sample size and resultant noise/volatility, but even at these sample sizes (pulled from ~326 mins of “with” TOI and ~557 mins of “without” TOI), the Bayesian estimators I use to assess uncertainty indicate that those numbers represent a real difference. (see details at end if you’re a power stats geek)

With Russell, Nuge’s lines score under a goal per hour (0.92), which is about on par with any fourth liner you care to name.

Without Russell, RNH’s lines (still without McDavid!) pop above 3 goals per hour, which is getting into that garden variety supernatural level I mentioned earlier. It is an excellent 5v5 scoring rate, worthy of a first line. (Have I used enough ‘super’s yet?)

Monkey Poo

I noticed this rather extraordinary effect in mid-season and tweeted about it. Which of course brought out the usual array of knowledge-free hacks who assumed this was an attack on Russell and immediately insulted my game viewing habits, my lack of NHL playing experience, and of course, stats in general. (I guarantee someone is going to attack this article as if it’s somehow a criticism of Draisaitl, which it most assuredly is not)

Fortunately, amidst the monkey poo being flung my way, I also had a brief conversation about the topic with ultra-smart Tyler Dellow, who wondered if the effect was because both guys play such a conservative defense-oriented game. He noted that RNH is the Oilers C that comes back deeper and more often than any other (to the surprise of no one).

Add that to Russell’s conservative style and that would leave the team with a lot of guys deep any time they recovered the puck, which means an awful long way to go to get back on the attack.

Seems like a credible explanation. Could be that. Maybe something else. That’s something video study by someone with time and an NHL Centre Ice subscription might be able to suss out (hint hint).

Either way, I cannot speak to why this occurs. Only that it does occur.

And if we’re going to assess a player’s true impact on the game, I would argue that it needs to be accounted for.

So let’s do that.

Backing Out the McDavid and Russell Effect

With David Johson’s invaluable puckalytics.com site gone, doing this type of multi-player WOWY is difficult. Fortunately, I have the enormous database I’ve built that feeds the PuckIQ site at my fingertips … I knew that would come in handy one of these days! [this also means any data errors you find are mine and mine alone … unless they’re the NHL’s]

Here’s what Ryan Nugent-Hopkin’s and Draisaitl’s lines accomplished in 2016-2017 when on the ice without either Connor McDavid or Kris Russell (note the axis does not start at zero to allow for more detail to be seen):

Pretty damn similar, right? If anything, Nuge has numbers a little better overall, though most are within error bars. This data is from ~554 mins for Nuge and ~240 mins for Draisaitl – enough for this data to be reasonably stable and representative.

Of course, a key point here is that Drai is younger, so he’s still got some room to improve. RNH on the other hand likely generated these numbers facing tougher competition. Things to keep in mind either way.

In case you’re wondering, and you’re pretty sharp so you probably are, they only spent ~53 mins together with neither Russell nor McDavid on the ice. So these are pretty fair representations of their ability to independently drive lines.

Bottom line: the actual separation between the two when without the rocket sauce of McDavid and without … whatever you want to call Russell’s sauce … is miniscule.

Ignuzzling the Russizzle

For completeness, I should also add that mixing Russell into the picture may strike you as biasing this analysis by selectively post facto taking away a huge negative impact on Nuge (and that’s a very fair comment. See, I told you you were sharp).

But the broad conclusion – that there isn’t much difference between the two players as far as driving their own line goes – still stands if you just look at time away from McDavid alone: RNH CF/GF% = 48.7%, 46.8%, Draisaitl = 47.5%, 46.7%.

I added Russell because I felt it made the comparison better rather than worse, both because of symmetry of treating the best/worst possession players on the Oilers both as outliers, and of course that massive effect on goal differential.

But you may (rightly) disagree on the topic of Russell, in which case you might want to fall back on the McDavid-only numbers.

Farewell, Sweet Nuge?

At the end of the day, I think TMac is aware that a lot of what I’ve shown above is happening, and it’s a big part of why he gives Nuge the tough minutes. For that reason, I’m guessing that when the time comes, he’ll probably fight to keep him.

But … the first dawn of Oiler cap hell starts next year with McDavid’s new contract kicking in and Russell and Lucic with NMCs. So it’s likely Nuge who will be the one waving goodbye.

I just don’t want anyone to not be aware of how good a player it is that will be walking out the door.

Ya got that? GIVE MY NUGEY HIS DUE!

 

---

POST SCRIPT: For Stats Power Geeks

Using Bayesian statistics to suss out whether a GF% differential is meaningful

---

I mentioned earlier that I am loath to use GF% as a stat because of its low sample size. It’s common also to run into the issue of small sample size even with higher sample statistics like Corsi.

Rather than ignore data (the paucity of GF% data is offset by its crucial importance to the game, no?), or just claim ‘small sample warning!’ and go ahead and run the analysis anyway, it would be good if there was some way to characterize the uncertainty of the data.

Broadly speaking, we are starting to recognize that traditional hypothesis testing and the way in which p values are used therein is flawed, or at least is a limited approach.

And I would argue that in a situation like hockey, it’s particularly inapplicable (hence rarely used) because we typically don’t want to reject or accept a hypothesis, we want to know to what extent a hypothesis / a number / a set of numbers is accurate or reliable or likely to be true.

Well, there is a way, and it’s called Bayesian statistics. (I am of the opinion that the future of hockey statistics is Bayesian. I’m far from an expert but trust me, working hard on it!)

In the case of the GF% situation between Nuge with/without Russell, I modeled it this way:

1. To model GF%, I used a beta distribution, which is the ‘right’ distribution for proportions but also has some really nice properties when calculating posterior distributions.

2. I started with what I call a “modestly informative” 50-goal prior centred at 50% (equivalent to perhaps 1/3rd to 1/2 of a season), which is strong enough to provide stability in the face of volatile small sample data, but weak enough to “let the data speak” as they say.

3. I then calculated posteriors for Nuge’s GF% with and without Russell.

4. So how then do we determine statistically if the two posterior distributions are meaningfully different? If we call the Nuge with distribution μ₁ and Nuge without μ₂, one way is to compute P(μ₁ < μ₂ | data). That is to say, what is the probability that the true value of the lower distribution μ₁ is smaller than larger distribution μ₂ given the data we observed?

5. If the number is ~50%, it’s a coin flip and we shouldn’t read anything into the difference. If it’s ~99%, confidence should be very high. Make sense? (this technique and description is from Bayesian Methods for Hackers by Cameron Davidson-Pilon).

6. We can implement this by drawing samples from the posteriors (sounds gross, eh) – the probability P(μ₁ < μ₂ | data) is then simply the fraction of the count of samples from μ₁ that were less than samples from μ_2.

7. To do this, I drew multiple 10,000 count random samples from each of the two posterior distributions. The proportion for which samples from μ₁ were less than samples from μ₂ was consistently around 89.7%.

8. So despite the small sample sizes, we can be pretty damn (that’s a formal statistics term) confident that the observed GF% difference of 28% vs 58% is meaningful.

9. A chart showing the prior, the two posteriors, the mean and 90% credible intervals, and a histogram of the samples from each posterior is shown below.

10. For clarity, the label ‘w Russell [0.344,0.441,0.541]’ for example shows that the mean of the ‘with Russell’ μ₁ posterior distribution is 44.1%, and the 90% credible interval is (34.4%, 54.1%). Note the effect of the prior as well – even though the calculated GF% are 28% and 58%, this analysis pegs the expected value of each distribution at a much less wide-ranging 44.1% and 54.1% respectively. The ‘wo Russell’ μ₂ distribution is closer to the calculated value and also narrower because its larger sample size means lower uncertainty.

11. One way you can be comfortable in asserting that the difference is meaningful is based on where the means and credible intervals sit.  While the distributions overlap (it would be surprising if they didn’t), nevertheless the mean of distribution μ₁ falls below μ₂‘s 90% credible interval, and (interestingly) the mean of μ₂ sits right at the upper value of μ₁‘s 90% credible interval.

12. Feel free to tweet at me (@OilersNerdAlert) if you have any questions or counter-arguments.