Does winning matter for Hall of Fame Induction (part 2)?

(see intro in previous post)

Below I present results for three probit models. The first table shows compares goodness-of-fit measures. The row designated “P > .5” shows the correlation coefficient between HOF induction and whether the predicted probability is over .5. The second table lists predicted probabilities for selected players.

Model 1 is based on Professor Bradbury’s specification. The differences in his predictions and model 1 are likely the result of differences in the way certain variables were calculated. Offensive production is measured using career linear weights, adjusted for ballpark factor and league run environment. Additionally I use the number of gold glove awards and the number MVP awards, as well as a dummy indicating whether the individual retired before 1960 (the first gold gloves were awarded in 1957). Finally the regression includes the number of years played and a dummy variable indicating the position at which the player played the most games (all outfielders are grouped together).

Model 2 replaces linear weights with career hits, runs, home runs and stolen bases. Additionally it includes controls for ballpark factor and run environment. I was personally surprised at how poorly it performed relative to the original model. This suggests that even though voters may not understand linear weights they look at more than just a few statistical categories.

Model 3 adds to model 1 the average winning percentage of a players teams as well as a dummy variable indicating whether he won a world series. I also experimented with variables that captured whether a player played in a major market, but these seemed unimportant. This specification might perform somewhat better, but the difference is likely small.

	Model 1	Model 2	Model 3
Pseudo-R2	0.7657	0.7684	0.7989
Log-Likelihood	-152.42	-150.31	-130.82
P > .5	0.7033	0.5102	0.7556
P > .05	0.6064	0.5313	0.6430
P > .01	0.5279	0.4686	0.5623

				Prob. Of Induction
Name	startyr	endyr	position	Model 1	Model 2	Model 3
Pete Rose	1963	1986	OF	100.0%	100.0%	100.0%
Bill Dahlen	1891	1911	SS	95.0%	93.0%	98.0%
Joe Torre	1960	1977	C	78.5%	69.4%	73.7%
George Van Haltren	1887	1903	OF	76.5%	76.9%	63.0%
Sherry Magee	1904	1919	OF	73.1%	42.7%	70.5%
Bob Elliott	1939	1953	3B	71.1%	61.4%	57.5%
Larry Doyle	1907	1920	2B	70.7%	55.4%	75.5%
Stan Hack	1932	1947	3B	70.2%	47.5%	74.0%
Dwight Evans	1972	1991	OF	66.1%	64.9%	69.7%
Rusty Staub	1963	1985	OF	65.4%	33.5%	39.3%
Jimmy Ryan	1885	1903	OF	63.5%	81.5%	49.0%
Bob Johnson	1933	1945	OF	63.3%	72.2%	21.8%
Dick Bartell	1927	1946	SS	57.6%	58.5%	43.5%
Dale Murphy	1976	1993	OF	55.7%	57.7%	14.1%
Keith Hernandez	1974	1990	1B	55.2%	21.3%	47.5%
Phil Cavarretta	1934	1955	1B	54.9%	38.9%	38.8%
Ted Simmons	1968	1988	C	52.9%	60.9%	41.2%
Vern Stephens	1941	1955	SS	46.8%	62.8%	35.8%
Dave Parker	1973	1991	OF	45.4%	65.2%	41.0%
George Burns	1911	1925	OF	42.5%	37.4%	76.3%
Joe Gordon	1938	1950	2B	39.8%	55.2%	74.8%
Joe Jackson	1908	1920	OF	27.1%	8.3%	28.7%
Jim Rice	1974	1989	OF	25.1%	45.1%	23.8%
Steve Garvey	1969	1987	1B	24.4%	45.0%	40.4%

The real difference between models 1 and 3 shows up when comparing players. Model 3 was particularly harsh on Professor Bradbury’s favorite player, Dale Murphy. According to it, Murphy’s chances were hurt by his lack of a World Series ring and the .443 winning percentage of the teams he played on.

I surprised by how well Joe Torre scored and also by how poorly Jim Rice scored. I should note that though predictions were generated for them, Pete Rose and Joe Jackson were not included in the initial regressions.

Next time I’ll look at the players who are enshrined with the lowest predicted probabilities.

2/9/07 Update: while working on the next parts I discovered a couple of minor mistakes in my code. The results did not change much, but the tables have been updated.