More Fun With Sample Sizes

Season opener, Rexall Place, October 10: Jordan Eberle, finally playing his first NHL game, takes Jim Vandermeer's wall pass in the neutral zone, bursts into the Calgary zone, puts an absolutely sick move on Ian White in which he toe-drags the puck under the sliding defender's stick while simultaneously (and blindly) jumping over it, retains sufficient balance, possession, and presence of mind to deke Miikka Kiprusoff to the backhand, and goes top cheese while crashing to the ice. It was a moment to remember for long-suffering Oiler fans, and for this one remains The Moment of another long, lousy, losing season. A perfect 10 on 10/10/10.

The cherry on top was that there was one second remaining in a Calgary powerplay at the time, making Eberle's spectacular effort officially a shorthanded goal. Immediately installed on the penalty killing unit, Jordan followed up with a second brilliant shortie against the Sharks a couple weeks later (pictured) and became a fixture on the PK. It seemed like a good idea at the time, but the experiment was destined to end in failure when (some of) the percentages turned against Eberle, and hard.

Eberle has played 45 minutes on the PK, virtually all of it in the first quarter of the season. Given he missed a stretch of games he currently has 44 GP, which leaves his average SH TOI just barely above the one-minute-per-game threshold that I routinely use to identify "regular" PKers over at Gabe Desjardins' outstanding advanced stats resource, BehindtheNet.ca (BtN). Before Eberle slides below that threshold I thought it would be a good time to draw attention to his very remarkable set of PK stats.

Applying a further games played filter to eliminate part-timers, there are currently 171 NHL forwards who meet the twin standards of 30+ GP with 1:00+ SH TOI per. The Corsi page of BtN's 4v5 stats shows a remarkable dichotomy: Jordan Eberle ranks first in the NHL in Sh% ON with a remarkable 66.7%, but 171st and dead last in Sv% ON with an execrable .680. That's right, his shooting percentage number very nearly exceeds his save percentage! That Sv% figure is so bad that there are only two other forwards within 100 basis points of it - Fredrik Modin (.729) and Eberle's usual partner-in-crime, Shawn Horcoff (.732), a victim of similar circumstance. The guy in 168th place (Vancouver stalwart Tanner Glass) is all the way up at .795, everybody else at .800 or better. Eberle and his on-ice PK mates - I'm looking at you, Nikolai - aren't even at .700. Egads.

On the other hand, that shooting percentage figure is out-of-this-world good, the sort of number you rarely see - to say the least - in a player with a sample size big enough to meet any sort of qualifying threshold. Eberle is more than 200 basis points clear of the second place performer, Boston's Brad Marchand, and fully double the guys in third.

Turning to the sum total of the two, PDO, Eberle sports a remarkable figure of 1.347, second in the NHL behind just Marchand's 1.372. A good PDO is generally seen as a player "riding the percentages" to plus/minus results perhaps better than he deserves. Surely Ebs is riding high with that outrageously unsustainable number? 

Not so fast. Turning now to BtN's Scoring and +/- page we have to scroll all the way down the page to find Eberle ranked 171st and last in the NHL at an atrocious -18.7 per 60. Only the bottom ten guys in the league are even half that bad. Second best in PDO, dead last in +/-, that's not a combination one expects to see.

The catalyst of this seeming anomaly is sample size, specifically unbalanced sample size. The elephant in the room is the 50 shots against the Oilers in Eberle's PK time, compared to the mouse that is just 3 shots for. PDO treats both equally, simply adds the two percentages together without regard for weighting them by the 50:3 ratio that is needed to bring them in line with actual results (2 GF, 16 GA in those 45 minutes). Once you get past the shiny glare of those two dazzling shorties, pretty dreadful results, albeit results that can be placed largely on goaltending that was two levels below "Inadequate". When we eliminate the shorties for and look simply at PPGA-against rates, Eberle again ranks dead last in the NHL at a shudderingly brutal -21.37; only 8 of the other 170 regular PKers are even in double digits.

Compare the one guy whose PDO is even higher, Marchand. He has played double Eberle's PK minutes (89) and allowed half the number of goals, just 8, for a goals against rate barely a quarter of the Oiler rookie. Like Eberle, Marchand's on-ice shooting percentage is way out of whack at 45% (5 goals on 11 shots), and his on-ice save percentage is also exceptional, but in a good way at .918. The shots allowed per 60 by the two are very similar at 65.5 for Marchand and 66.8 for Eberle, but Tim Thomas and friends have let in just 8% of those shots, while Nikolai Khabibulin and accomplices have been torched for 32%. That explains the 4X difference in goals against rate right there.

The other elephant in the room which I will not even attempt to address quantitatively is shot quality. Is it possible that Jordan Eberle was such a miserable penalty killer that Oilers were giving up an inordinate number of ten-bell chances while he was out there? My guess is yes, a little, but nowhere near as much as those percentages suggest. To some extent the other guys were making their shots - mid-air deflections off the post and in, that kind of thing, there was an epidemic of that stuff early in the season. Plus, obviously, our goalies were not making their saves. It's an open question as to whether the coach's decision to gong Ebs from the unit was based on any sort of analysis of ineffective play specific to him or simply a knee-jerk reaction to results while his luck was (mostly) running against him.

Conclusions:

1. PDO is a pretty useless stat for special teams due to the heavy imbalance in numbers of shots. (Duh! but it had to be said)
2. PDO's individual components are of greater importance in special teams outcomes, specifically Sh%ON for powerplayers and Sv%ON for penalty killers.   
3. At even strength those individual components are still pretty important for players who have an unbalanced ratio of shots for:against. Two players posting identical shots ratios and PDO will sport very different plus minus outcomes if their component percentages are radically different, as per this simple example of two fictional players:

   	SF/60	SA/60	Sh% ON	Sv% ON	PDO	GF/60	GA/60	SF/60
   Player A	22	31	7.0%	0.930	1.000	1.54	2.17	-0.63
   Player B	22	31	11.0%	0.890	1.000	2.42	3.41	-0.99

4. PDO is great back-of-the-envelope tool, but the risk is it can be as dirty as it is quick. Like all percentages, and pretty much all stats period for that matter, context is key.