Marinomics: A Dismal Baseball Science

In case anyone stumbles upon this from a Google search, something Malcolm Gladwell wrote made me update something I wrote a while ago.

This is my updated "Suspicious Seasons" list - players who enjoyed remarkable jumps in power figures (slugging and isolated power) compared to their career, up to that point. This list is a little different - it goes back to 1930. This is the Top 40 sorted on the average of the rank of their Unexpected SLG and Unexpected Isolated Power.

First Last      year uxSLG uxIsoP
Barry Bonds     2001 0.296 0.258
Jeff Bagwell    1994 0.286 0.213
Champ Summers   1979 0.296 0.187
Randy Winn      2005 0.269 0.194
Brady Anderson  1996 0.244 0.197
Luis Gonzalez   2001 0.228 0.185
Bob Bailey      1970 0.226 0.191
Jim Spencer     1979 0.215 0.180
Jose Guillen    2003 0.231 0.155
Ken Caminiti    1996 0.218 0.158
Sammy Sosa      2001 0.214 0.159
John Lowenstein 1982 0.227 0.153
Javy Lopez      2003 0.209 0.163
Cito Gaston     1970 0.240 0.145
Jose Hernandez  1995 0.204 0.173
Barry Bonds     2004 0.210 0.146
Ivan Rodriguez  2000 0.202 0.155
Mark McGwire    1998 0.196 0.158
Mark McGwire    1996 0.208 0.148
Jeff Reed       1997 0.203 0.148
Ralph Kiner     1947 0.209 0.142
Felix Mantilla  1964 0.191 0.155
Barry Bonds     2002 0.214 0.136
Alex Ochoa      2000 0.187 0.148
Mark Bellhorn   2002 0.197 0.136
Richard Hidalgo 2000 0.193 0.138
Henry Rodriguez 1996 0.183 0.150
Troy Glaus      2000 0.190 0.141
Kevin Mitchell  1989 0.178 0.155
Gary Sheffield  1992 0.204 0.132
Larry Walker    1997 0.210 0.129
Rico Petrocelli 1969 0.192 0.136
Ray Boone       1953 0.190 0.139
Jhonny Peralta  2005 0.198 0.135
Willie Mays     1954 0.209 0.129
Al Kaline       1955 0.198 0.132
Reggie Jackson  1969 0.182 0.144
Mark McGwire    1995 0.177 0.152
Adrian Beltre   2004 0.201 0.128
Mike Epstein    1969 0.184 0.136

Persons with 2005 seasons in the Top 40:
Randy Winn (Go M's!), Jhonny Peralta
Almost made the list:
Derrek Lee, Tony Clark, Javier Valentin

But of course, this stuff goes the other way, too. The players who had the very worst 2005s, in terms of power outages, compared to their careers:

First Last      year uxSLG uxIsoP
Sammy Sosa      2005 -0.168 -0.113
Brian Jordan    2005 -0.125 -0.087
Jeff Davanon    2005 -0.117 -0.085
Mike Lowell     2005 -0.117 -0.076
Bernie Williams 2005 -0.120 -0.069
Edgardo Alfonzo 2005 -0.091 -0.080
Magglio Ordonez 2005 -0.089 -0.084
Kevin Millar    2005 -0.093 -0.073
Ryan Klesko     2005 -0.097 -0.063
Jason Kendall   2005 -0.097 -0.062

Acting on something Smokin' Joe Morgan said during the Dodgers game (Beltre is 'adjusting' to AL pitching) - a followup to the league switching thing I did on hitters. The same thing for pitchers (1950 forward with paired seasons of more than 50 innings pitched - sample ~9000 paired seasons with 900 or so involving league switches).

Equation:
ERA+ = League ERA/Pitcher's ERA
Change in ERA+ from Year 1 to Year 2 = Intercept + B1*([ERA+ in Year1] - 1) + B2*Age + B3*(Switched Leagues)

            Coeff  StdErr t-Stat     P
Intercept   -0.070 0.023  -3.067 0.002
ERA+-1      -0.721 0.010 -70.461 0.000
Age          0.004 0.001   5.416 0.000
Switch       0.012 0.011   1.012 0.312

There's even less evidence for a league switching effect here (the t-statistic for the League Switching effect is closer to 0). In fact, switching leagues even has a slightly positive effect. Like before, whatever league switching effect that people may perceive observationally is probably completely overwhelmed by mean reversion. If a pitcher is unbelievably good or bad, you'd be better served by assuming that his performance the next year will just be closer to average, irrespective of what league he's going to pitch in.

I updated the attendance models through Wednesday's game and changed the structure a little bit with a few extra explanatory variables.

Date     Opp       Actual 04Mod  03Mod  3+4Mod Middle Diff      Note
Mon 4/4  Minnesota 46,249 26,544 27,733 27,318 27,318           Opening Day
Tue 4/5  Minnesota 28,373 29,733 26,196 28,958 28,958 (585) 
Wed 4/6  Minnesota 25,580 28,064 25,945 27,941 27,941 (2,361) 
Fri 4/8  Texas     29,652 37,796 35,936 37,181 37,181 (7,529) 
Sat 4/9  Texas     31,501 40,519 37,863 39,466 39,466 (7,965) 
Sun 4/10 Texas     30,434 40,734 38,212 39,750 39,750 (9,316) 
Wed 4/20 Oakland   24,841 31,523 30,647 30,493 30,647 (5,806) 
Thu 4/21 Oakland   22,428 30,103 32,402 31,301 31,301 (8,873) 
Fri 4/22 Cleveland 43,207 35,125 31,232 30,497 31,232 +11,975  Ichiro bobblehead
Sat 4/23 Cleveland 33,564 37,848 33,159 32,782 33,159 +405  
Sun 4/24 Cleveland 32,889 38,062 33,508 33,066 33,508 (619) 
Mon 5/2  LAAngels  24,184 37,278 45,652 39,515 39,515 (15,331)
Tue 5/3  LAAngels  29,917 40,467 44,915 41,154 41,154 (11,237)
Wed 5/4  LAAngels  26,303 38,798 44,664 40,137 40,137 (13,834)

The 2003 version likes the Angels quite a bit, since they were coming off the 2002 World Series.

In re: Michael Lewis' NY Times Magazine piece that mentions Beltre's jump...

Players who have enjoyed jumps in various power categories since 1980.

These lists were updated (mostly so Phil Bradley would get on one more, but also because of some stability patterns I saw), to all be the Top 25, Career AB >= 200, Season AB >= 200 (instead of Season AB >= 175). Added Isolated Power and Extra Bases. Added AB numbers to all of the lists to help temper some of them.

Name            Year UxTB CrAB SeasAB
Brady Anderson  1996 141.5 3271 579
Barry Bonds     2001 141.1 7456 476
Luis Gonzalez   2001 139 5096 609
Sammy Sosa      2001 123.4 5893 577
Adrian Beltre   2004 120 2864 598
Larry Walker    1997 119.4 3132 568
Ken Caminiti    1996 119.2 3967 546
Robin Yount     1982 118.5 4212 635
Kirby Puckett   1986 118.2 1248 680
Sammy Sosa      1998 114.7 4021 643
Jeff Bagwell    1994 114.2 1675 400
Gary Sheffield  1992 113.7 1110 557
Richard Hidalgo 2000 107.5 656 558
Troy Glaus      2000 107.3 716 563
Ellis Burks     1996 105.8 3720 613
Aramis Ramirez  2001 103.7 561 603
Bret Boone      2001 102.9 3911 623
Ryne Sandberg   1984 101.9 1274 636
Mark McGwire    1998 100 4622 509
Henry Rodriguez 1996 97.6 766 532
Rich Aurilia    2001 97.2 1919 636
Kevin Mitchell  1989 96.9 1311 543
Javy Lopez      2003 95.7 3546 457
Paul Konerko    1999 94.8 224 513
Robin Yount     1980 94.7 3224 611
    
Name            Year UxSLG CarAB SeaAB
Barry Bonds     2001 0.296 7456 476
Jeff Bagwell    1994 0.286 1675 400
Brady Anderson  1996 0.244 3271 579
Terry Shumpert  1999 0.236 884 262
Jose Guillen    2003 0.231 2050 315
Luis Gonzalez   2001 0.228 5096 609
John Lowenstein 1982 0.227 2548 322
Ken Caminiti    1996 0.218 3967 546
Barry Bonds     2002 0.214 7932 403
Sammy Sosa      2001 0.214 5893 577
Barry Bonds     2004 0.21 8725 373
Larry Walker    1997 0.21 3132 568
Javy Lopez      2003 0.209 3546 457
Albert Belle    1994 0.209 1881 412
Mark McGwire    1996 0.208 3659 423
Gary Sheffield  1992 0.204 1110 557
Jose Hernandez  1995 0.204 234 245
Jeff Reed       1997 0.203 2101 256
Ivan Rodriguez  2000 0.202 4443 363
Adrian Beltre   2004 0.201 2864 598
Mark Bellhorn   2002 0.197 323 445
Mark McGwire    1998 0.196 4622 509
Richard Hidalgo 2000 0.193 656 558
Hubie Brooks    1986 0.192 2648 306
    
Name            Year UxIsoP CrAB SeaAB
Barry Bonds     2001 0.258 7456 476
Jeff Bagwell    1994 0.213 1675 400
Brady Anderson  1996 0.197 3271 579
Luis Gonzalez   2001 0.185 5096 609
Jose Hernandez  1995 0.173 234 245
Javy Lopez      2003 0.163 3546 457
Sammy Sosa      2001 0.159 5893 577
Ken Caminiti    1996 0.158 3967 546
Mark McGwire    1998 0.158 4622 509
Kevin Mitchell  1989 0.155 1311 543
Ivan Rodriguez  2000 0.155 4443 363
Jose Guillen    2003 0.155 2050 315
John Lowenstein 1982 0.153 2548 322
Mark McGwire    1995 0.152 3342 317
Pedro Feliz     2003 0.15 373 235
Henry Rodriguez 1996 0.15 766 532
Jeff Reed       1997 0.148 2101 256
Alex Ochoa      2000 0.148 1083 244
Mark McGwire    1996 0.148 3659 423
Barry Bonds     2004 0.146 8725 373
Jim Edmonds     1995 0.143 350 558
Phil Bradley    1985 0.142 389 641
Todd Hundley    1996 0.142 1468 540
Gary Sheffield  1994 0.141 2161 322
    
Name            Year UxExB CrAB SeaAB
Barry Bonds     2001 122.8 7456 476
Brady Anderson  1996 114.2 3271 579
Luis Gonzalez   2001 112.4 5096 609
Kirby Puckett   1986 93.5 1248 680
Sammy Sosa      2001 91.7 5893 577
Phil Bradley    1985 90.7 389 641
Ken Caminiti    1996 86.4 3967 546
Jeff Bagwell    1994 85.2 1675 400
Kevin Mitchell  1989 84.3 1311 543
Robin Yount     1982 82.4 4212 635
Sammy Sosa      1998 82.2 4021 643
Robin Yount     1980 80.8 3224 611
Mark McGwire    1998 80.2 4622 509
Henry Rodriguez 1996 79.8 766 532
Jim Edmonds     1995 79.6 350 558
Troy Glaus      2000 79.4 716 563
Steve Finley    1996 78.4 3364 655
Richard Hidalgo 2000 77.1 656 558
Todd Hundley    1996 76.4 1468 540
Adrian Beltre   2004 76.4 2864 598
Jay bell        1999 76.1 5651 589
Javy Lopez      2003 74.3 3546 457
Sammy Sosa      1999 74   4664 625
Gary Sheffield  1992 73.8 1110 557
Larry Walker    1997 73.1 3132 568


UxSLG = Unexpected Slugging Percentage = SLG in Year X - Career SLG
UxTB = Unexpected Total Bases = TB in Year X - Career SLG * AB in Year X
UxIsoP = Unexpected Isolated Power = IsoP in Year X - Career IsoP
UxExB = Unexpected Extra Bases = TB minus Hits in Year X - Career IsoP * AB in Year X

Make sure you write your Congressman.

The spark for this was some notes in the paper about really low attendance figures. I'll compare this season's 6 home dates (7 with opening day) against predictions from models based on the 2004 season, the 2003 season and the 2004 and 2003 seasons together.

Date    Opp  2004M    2003M    CombM    3MAvg   Actual Difference*
Mon 4/4 Min                                     46,249 
Tue 4/5 Min  33,403   28,071   29,342   30,272  28,373 -1,899
Wed 4/6 Min  31,468   27,738   28,080   29,095  25,580 -3,515
Fri 4/8 Tex  38,040   36,658   37,527   37,408  29,652 -7,756
Sat 4/9 Tex  40,783   38,585   39,825   39,731  31,501 -8,230
Sun4/10 Tex  40,998   38,934   40,109   40,014  30,434 -9,580
Wed4/20 Oak  33,123   31,437   32,015   32,192  24,841 -7,351

* - Acutal less 3 Model Avg

(The predicted results for April from the 2004 model are going to be higher because the team's late season attendance was where it nose dived)

Ignoring the opening day result, these figures are remarkably weak - sometimes shockingly so. The -9,580 for the Sunday game against Texas would be in the 1% worst outcomes from the models.

You could come up with a few storylines that have merit for this. Previous year's record makes the public unwilling to show up until a pattern of winning is established - which means that the April results would necessarily be low, but the August and September results could be much higher. The individual teams they are playing are just unattractive crowd draws. Their new park effect is finally gone (after 5 years!). Or you could say that the previous year's levels were just unsustainable.

The 2.5 million that management publicly projected for this year - off from 2.94 million in 2004 and 3.27 in 2003 - would imply an average difference of around minus 6,600 per game from the 3 model average.

Is it possible to uncover umpire biases? Could their be some grain of truth the beer soaked conspiracy theories that arise when you're at the ballpark - that this ump is a fan of the opponent, or he just hates our team?

Going through game log data back to 1970 - around 70some thousand games - I get the seasonal strikeout and walks per PA rate for teams, pitchers, and umpires. Then I apply these rates to the individual games and get arrive at an excess ("unexplained") strikeouts and walks from the team, umpire and pitcher expectations. However, at least one of these (total team, pitcher or umpire strikeouts) should be close enough to zero that it's statistically insigificant. Why just one?

If there are just a lot of strikeouts to begin with, it could be that the team is merely prone to strikeouts.
If the team is striking out more than we'd expect, then the umpire might just call more strikeouts. As long as he is calling more strikes for everyone, I don't see a problem with this.
If the team is striking out more than we'd expect and there's more strikeouts than we'd expect from the umpire, maybe the pitchers that the team faces just throws a lot of strikeouts.
If we're still left with an unexplained amount of strikeouts and walks after considering the respective proclivities of the team, the umpire and the pitchers - then it begins to look kind of funny.

Then start comparing it to the data from the opponent. Again, if these things are happening to both teams, I guess I just don't care that much. Subtracting out the opponent's excess strikeouts and walks removes the suspiciousness if it is actually happening to both teams, and therefore makes it more noticable if it's only happening to one team.

Then I start adding game after game together. If a particular umpire is actually "for" or "against" a team, this pattern should reveal itself when we add all of these games together and the unusual results from umpires that are merely fair should disappear.

So are there some umpire biases? Well, yes, it is possible to tease this out. Restricting it to a 1% probability level and a reasonable number of games umpired (say X>50), there's 5 umpires whose results with a particular team seem unusual, either for or against a team - though only 1 of them is a current umpire. There were nearly 5,000 umpire/team combinations in total (with game totals ranging from 1 to 175).

I'm not going to "out" any umpire, since I don't think it's really fair. I'm a guy with a few databases, some graduate level stats work and a website, most of these guys were MLB umpires when I was learning long division. Further, there's some things that I don't really like about the approach and I recognize are problems.

The biggest issues are that I want more pitcher data and to have any degree of statistical confidence in the umpire measurements, I need many, many games with the umpire/team combo. To get that, the umpire is necessarily going to have been around a while - and since the 1994 strike, a team gets only around 2 or 3 games with a particular umpire behind the plate each season.

I've been slowly putting together the bones of a paper the past few weeks, so this seems appropriate.

The Effects of Labor Strikes on Consumer Demand: A Re-examination of Major League Baseball: Attendance levels adjusted for new stadiums have not returned to the pre-strike levels.
and
Striking Out? The Economic Impact of Major League Baseball Work Stoppages on Host Communities: The economic impact of baseball is (surprise!) inflated by chambers of commerce. Has 2 different methodologies, both pretty interesting - using the expected deviation in personal income during strike years to measure the economic activity that baseball represents and the ratio of incomebetween MLB and non-MLB cities.

A Fall Classic? Assessing the Economic Impact of the World Series: Similar methodology as the strike paper - the post-season isn't an economic windfall.

Location and attendance in major league baseball: Not free. Builds an attendance model that incorporates the distance between teams and new teams. Has a couple of very interesting implications for the WaNats and Orioles - like decreasing the Orioles attendance by around 125,000.

Reputation Effects in Gold Glove Award Voting: Award models are hard to do. Develops different models for each of the infield positions. Suggests that Gold Glove voters use a player's past GG wins when voting.

Convergence and clustering in major league baseball: the haves and have nots?
Also not free. Uses a clustering algorithm to chop up the teams into "haves" and "have nots" to test some ideas from the Blue Ribbon Panel and some earlier research that contends baseball is more competitively balanced. Over the long haul, they say that the clustering hypothesis works but over the recent past, where revenue figures are available, it doesn't.

And an older one that is a very interesting idea:
Using Baseball Card Prices to Measure Star Quality and Monopsony: Also not free. It uses baseball card prices as a proxy for how the public values a player by based on statistical appeal and "star quality". It uses an attendance model that includes "star quality" - which is the sum of the errors from the card price equation (which also uses a white/non-white ethnic dummy variable, which turns out to be insignificant). I've tried using things like salaries and salaries above a threshold to capture that, but never pulled out the stacks of Beckett that were sitting under my bed as a kid.