Any eight year old baseball fanatic can isolate the handful of players in each league that are likely to win the MVP.  And with a high degree of certainty, the winner being inevitably chosen from that handful of players is justifiably expected.  Popularity is obviously a major indicator of likely MVP consideration.  But I have a deep suspicion in any line of reasoning derived from untested opinion or visceral assumptions, rather than some form of empiricism.  This calls for a regression model, using the voting trends from the last ten years (2000-2009).

Qualitative measures are hard to analyze. Let’s use Carlos Gonzales as an example. Playing with the mid-market profile Rockies, qualitative points are by default removed simply because of the perceived Coor’s Field offensive inflation factor, as well as not being in a major market. His stats may be considered deceiving, and for him to have a realistic chance, even though the raw numbers would indicate evidence to the contrary, the Rockies will probably have to win the NL West. Immediately this brings forth a comparison to Matt Holiday, who also played for the Rockies, Wild-Card winner in 2007, the year Holiday finished 2nd in the MVP, despite leading the league in total bases, RBI, batting average, doubles, and tied for 2nd in the league in OPS as well as a top 5 WAR. The eventual winner, Jimmy Rollins, playing for the major market Philadelphia Phillies, had a underwhelming OPS+ (119), a WAR rated three spots lower than Holiday, and a ballpark that had yet to reach its current consensus status of bandbox.

Not saying Rollins was undeserving, but just creating a proxy for the categorical particulates that may dim the chances of certain players receiving the votes necessary to win.

Certainly there are other intangibles that can manifest themselves spontaneously, and at times erroneously, because ultimately the votes accumulated for the MVP award is attributed to the subjectivity of the writers.

To expound further in that respect, a meta-analysis, which would involve an overall abstract of each player’s popularity level and his respective team’s status in relation to the configured mindset of the writers responsible for the voting process, could be broken down into various components that enables a measure of quantification. But that would take too long.

The goal here is to search for value, in order to accomplish this, creating odds to compare to the actual odds is the best approach. Then evaluating the players from a qualitative standpoint. Which I won’t address here.  But as the season reaches its end I will update the prevailing consensus concerning the likely top MVP candidates, and adjust my odds accordingly.

A multivariate analysis using Stata, a program that operates with a sharp understanding on the benefits of parsimonious exertion, allows for a flexible and facilitating process.

I found the correlations between each crucial stat line and the overall voting points, which is what decides the winner of the MVP. Of course, I separated the American and National League, each having their separate yet similar MVP selection process.

I stripped pitcher’s from the MVP equation, and will revisit the pitchers later when I assess the Cy Young odds.

The variables used were team wins, WAR, batting average, home runs, RBI, runs, stolen bases.  These were chosen after some trial and error, but the inherent value of the variables used can certainly reconcile with logic.  WAR may be considered redundant in terms of information content, as well as redundant in compatibility with the other player variables.  However, WAR has its advantages because of the all-inclusive nature of the statistic: Position adjustments, defense, sophisticated offense metrics.  It is an overall measure of player viability that has yet to hit mainstream and a faction of writers may not even attempt to recognize or consider it as part of their voting practices.  Since the regression coalesces to the tendencies of the voters by way of voting points, the most common statistics must be included, and redundancies are to be expected.  Runs and HRs are not mutually exclusive variables, neither are Home runs and Slugging, or even Home runs and Batting Average.  All the preceding events happen simultaneously.

One more thing before I proceeded with both leagues, I created an arbitrary though sufficient way of invoking the playoff variable.  Teams primed for a playoff position were given an extra weight in wins, for team wins have a positive, albeit stunningly slight, correlation to voting points.  The weight was calculated by adding the square root of projected team wins to itself.  So a team on pace for 100 wins would be credited with 110 if they are likely to make the playoffs, and conversely, a team would not be weighted if they had a projection of 90 wins with no post-season prospects.  Its a sliding scale.  Again, arbitrary, but after messing around with the correlations, this seemed to be a solid method of including the post-season factor.

American League

Here is the correlation matrix for the American League (‘twx’ indicates adjusted team wins , everything else is labeled appropriately):

             |  votepts      war      twx       hr      rbi        r       ba      obp      ops      slg       sb
-------------+---------------------------------------------------------------------------------------------------
     votepts |   1.0000
         war |   0.5355   1.0000
         twx |   0.2800   0.0654   1.0000
          hr |   0.3809   0.2404  -0.0676   1.0000
         rbi |   0.4573   0.2614  -0.0265   0.8230   1.0000
           r |   0.4026   0.5326   0.0392   0.2707   0.3019   1.0000
          ba |   0.3341   0.3763  -0.0767  -0.2653  -0.0843   0.1744   1.0000
         obp |   0.3946   0.5200   0.0610   0.2222   0.1924   0.1843   0.5332   1.0000
         ops |   0.4841   0.4874  -0.0397   0.6781   0.5803   0.2022   0.3519   0.8006   1.0000
         slg |   0.4560   0.3952  -0.0863   0.8119   0.6937   0.1798   0.2019   0.5726   0.9496   1.0000
          sb |  -0.0891   0.1202   0.0456  -0.4882  -0.5237   0.3247   0.0853  -0.1977  -0.4473  -0.5093   1.0000

Stolen bases having a negative correlation may seem surprising, but at this point I feel it may be appropriate to refer to Gould’s ‘bio-mechanical limit’ theory to dignify. However, I won’t rhapsodize further (or even at all), for fear of frivolous in-articulation, as well as a mind-hurtling digression that is completely unnecessary.

The coefficients:

      Source |       SS       df       MS              Number of obs =     212
-------------+------------------------------           F(  5,   206) =   48.97
       Model |  1058497.82     5  211699.563           Prob > F      =  0.0000
    Residual |  890627.995   206  4323.43687           R-squared     =  0.5431
-------------+------------------------------           Adj R-squared =  0.5320
       Total |  1949125.81   211  9237.56309           Root MSE      =  65.753

------------------------------------------------------------------------------
     votepts |      Coef.   Std. Err.      t    P>|t|     [95% Conf. Interval]
-------------+----------------------------------------------------------------
         twx |   2.060374   .3189046     6.46   0.000     1.431638    2.689109
         war |   11.34832   3.032244     3.74   0.000     5.370106    17.32653
          hr |   2.330891    .661193     3.53   0.001     1.027318    3.634464
          ba |   1362.608   226.1278     6.03   0.000     916.7862    1808.429
          rp |   1.278251   .4477948     2.85   0.005     .3954025    2.161099
       _cons |  -782.4358   79.27622    -9.87   0.000    -938.7326   -626.1391
------------------------------------------------------------------------------

The variable ‘rp’ indicates Runs Produced, which is simply . The voting tendency of the AL MVP, being an offensively geared league as compared to the NL, sees a leaning towards an inclusion of runs scored as well as RBI. For ocularity I just inserted Runs produced into the process, and running the two regressions onto voting points result in virtually identical descriptive statistics.

Now with the given data above its easier to construct a formidable scale of probability. Then later the scale can be leveraged to the general nature of the players involved (what team they play for, etc…).

Here are the odds of the AL MVP, using players that registered voting points produced by the above coefficients and the current season stats. Current in this case means current projection/pace. All statistics were purely flat projections, except for Batting Average which was regressed to a .285 over the remaining number of likely ABs for each player (.285 based solely on my preference, applies to NL as well).

Once the real odds are released I’ll re-address the table and strip away any players that are unlikely to win. A brief survey and one can identify about half a dozen that have a very slim chance of being considered. The odds will be adjusted accordingly.

National League

There is a systematic infection of irreducibility that arises after analyzing the NL data corresponding to Barry Bonds. I removed Bonds for obvious reasons, he skews the entire process. After removing Bonds I re-allocated the voting points to the other candidates, with some reason and adequacy in the end, but didn’t put too much thought into it so it could have been better.

The rest of the process was the same as the AL. For the NL, there was a small difference in the variables that were selected as a result of the most optimal regression, and they are apparent by observing the results below. In both leagues, the four major variables (batting average, RBI, HR, wins) provide a much expected symbiotic relationship with the distribution of voting points.

The NL correlation matrix:

             |  votepts      twx      war       hr      rbi        r       ba      ops      obp      slg       sb
-------------+---------------------------------------------------------------------------------------------------
     votepts |   1.0000
         twx |   0.2844   1.0000
         war |   0.5730   0.1194   1.0000
          hr |   0.5342  -0.1629   0.3935   1.0000
         rbi |   0.5187  -0.1536   0.3689   0.8242   1.0000
           r |   0.4161  -0.1116   0.5578   0.4451   0.4117   1.0000
          ba |   0.2885  -0.0937   0.4192  -0.0546   0.0701   0.1110   1.0000
         ops |   0.5379  -0.1334   0.5907   0.6584   0.5781   0.2957   0.6023   1.0000
         obp |   0.3718  -0.0757   0.5785   0.2745   0.2970   0.2247   0.6984   0.8366   1.0000
         slg |   0.5562  -0.1467   0.5269   0.7705   0.6488   0.2956   0.4834   0.9629   0.6579   1.0000
          sb |  -0.0993  -0.0047  -0.0051  -0.3482  -0.3985   0.2915  -0.1109  -0.3776  -0.2489  -0.3972   1.0000

Here are the regression results:

      Source |       SS       df       MS              Number of obs =     216
-------------+------------------------------           F(  6,   209) =   54.96
       Model |  1696461.17     6  282743.528           Prob > F      =  0.0000
    Residual |  1075285.46   209  5144.90649           R-squared     =  0.6121
-------------+------------------------------           Adj R-squared =  0.6009
       Total |  2771746.63   215  12891.8448           Root MSE      =  71.728

------------------------------------------------------------------------------
     votepts |      Coef.   Std. Err.      t    P>|t|     [95% Conf. Interval]
-------------+----------------------------------------------------------------
         twx |   3.310143     .38732     8.55   0.000     2.546588    4.073697
         war |   8.539037   3.239998     2.64   0.009     2.151772     14.9263
          hr |   4.478822   .8344577     5.37   0.000     2.833789    6.123855
          ba |   1313.972   239.9688     5.48   0.000     840.9021    1787.041
         rbi |   1.012889    .395473     2.56   0.011     .2332621    1.792517
          sb |   1.402283   .4044013     3.47   0.001     .6050547    2.199512
       _cons |  -916.8518   91.12105   -10.06   0.000    -1096.486   -737.2176
------------------------------------------------------------------------------

NL MVP Odds:

I gave teams that are in playoff/non-playoff contiguity the benefit of the doubt in both leagues (i.e. Red Sox, Rockies). If value is to be found it can only be found in a scenario that enables the most strict possible calculation of probability.

The spectrum of calculating odds is tethered to the number of players considered as candidates. With time and some employment of elementary baseball logic, comes a higher semblance of accuracy. Once the season progresses I can strive for more precision.

I’ll revisit as players vault in and out of MVP candidacy.