Defense Independent Pitching (DIPs vs. ERA)
Voros
I decided, based upon the advice of others, that I would use some of the
Defense Independent Pitching ("DIP") Methodology to find out how well it
correlates with the following year's ERA for pitchers.
First the problems with doing this:
1. Pitchers often remain on the same team with virtually the same
defense behind them and in the same park from year to year. DIP stats
adjust for these things but for it to correlate with ERA the following
year it shouldn't. You see ERA in any year is subject to the biases of
league, park and team defense. If the biases don't change for a pitcher
from year to year, an adjustment that attempts to remove them will
necessarily cause any correlation with ERA to degrade.
2. DIP stats are not a prediction stat. They are a performance stat that
attempts to evaluate what is actually part of the pitcher's performance
and what results from other factors. If the ultimate goal was to
correlate with ERA, we would include past year data. We'd shape some of
the stats and shape each stat to varying degrees. We'd take into account
the pitcher's age and occassionally his workload. Therefore we must
assume that we could devise a better method for correlating with the
following year's ERA than DIP stats.
However there are workarounds here. The latter issue can be dealt with
rather simply. I think any correlation that does exist between DIP stats
and ERA mean more due to this problem. If the stat isn't designed to
correlate with ERA but does so anyway, it certainly indicates that it
has a strong correlation with _ability_ as well.
Problem two involves some work though. As I stated before, DIP stats
make adjustments to a pitcher's stats on the basis of league, park, and
stats that can be influenced by the defense behind him. However many of
these things remain static from year to year for the majority of
pitchers. As such their effects on ERA remain static. However one
effect, the wide degree of randomness in the stats affected by fielding,
affects ERA over and above overall team defense. This is the factor
that causes such a low degree of correlation in a pitchers percentage of
hits given up on balls in play, as an example. This factor, a major
reason why I believe DIP is a strong measure, can be filtered out and
cause an increase in correlation with ERA the following year. Without
filtering this out, you essentially have the same measure as Component
ERA.
So how do we do this? Since we're only gonna adjust for randomness in
stats that the defense has some control over, and we're not adjusting
for park and league factors, we can leave all stats that aren't affected
by defense the same, i.e. HR, BB, SO.
The first study I did used pitching stats from 1993-1998 for all
pitchers that pitched at least 100 IP in back to back years (Split team
performances are ignored for simplicity's sake). I used 100 IP instead
of 162 IP because of the problems 1994 & 1995 might cause. There were
400 such pitchers. I did not have BFP data available to me easily so I
made an estimate which I'll get to in a bit. As I said before BB, HR and
SO remain the same. The rest of the numbers are based upon _team_
averages of H allowed and IP, etc.
The numbers are calculated and then an ER estimate is made based upon
Jim Furtado's Extrapolated Runs formula. We use our ER estiamte and our
new IP estimate for the pitcher and we have our DIP ERA for the study
(remember this figure is not adjusted for park, league or _overall_
quality of his team's defense, only the defense that happened while he
was pitching).
I also calculated Component ERA ("ERC") and strikeouts per 9 Innings
("K9IP") using the original stats. I then made a Correlation comparison
with the following year's ERA with all of those ERA estimates as well as
the current year's ERA. Here's how well they correlated with the next
year's ERA:
ERA= .387
ERC= .422
K9IP=-.362
DIP= .500
These correlations are lower than they would be if 162 IP were the
standard (obviously). As you can see with a sample size of 400 instances
of back to back 100+ IP seasons from 1993, DIP's estimate handily beats
ERA and K9IP and comfortably bests ERC.
I also ran correlations by year. In every year except for 1995-1996, DIP
came out with the highest correlation. In 1995-1996 it finished a close
second to K9IP. In 1995-1996 all the correlations were very low, don't
ask me why.
Finally I ran correlation based on BFP estiamtes. I took the pithcers
with the 100 highest BFP estimates, the next highest, etc. Anyway DIP
again corrleated the highest in all four of these groups. Also, as I
suspected, the lower the BFP totals the greater of an advantage DIP
showed over the others. Among the groups where the first year had
estimates of only 414-620 BFP, the correlations were:
ERA= .178
ERC= .243
K9IP=-.233
DIP= .402
Remember this is not a particularly low number of IP for this set of
pitchers. They're almost all above any totals ANY relief pitcher would
ever post.
ERA in and of itself becomes close to useless. A correlation of .178 is
not real good. Component ERA and SO per Nine Innings fare little better.
But DIP's ERA estimate still does relatively OK. It's correlation for
the lowest workload group is better than the ERA and K9IP correlation
overall.
So what can we start to theorize about here? Well here's some ideas I
have:
-Pitching stats that can be affected by fielding exhibit most of the
characteristics of a stat that correlates poorly with ability and has a
very large random element in it. These stats correlate poorly for
pitchers from year to year. They degrade at an accelerated rate as
sample size decreases. When removed from our calculations, our ability
to evaluate a pitcher's ability actually improves instead of decreases.
And that Hits Allowed is one such stat and is currently a focus of many
of the major sabermetric tools used to evaluate picthing.
-The affects described above can greatly mask a pitcher's true level of
ability over a full season's performance. Filtering out these effects
can cause very wide shifts in our evaluations of pitcher quality.
-ERA for a relief picther in a single season is a virtually meaningless
measure if you have other stats to work with.
-Pitchers don't have the same opportunities for individual
accomplishment that Hitters do. The pitcher is much like a quarterback
in that much of what he does is irrecoverably tied to the performance of
his teammates. Hitters have a greater opportunity to accomplish
individual feats.
That's what I have so far. The key from here is to refine the DIP
measures (assuming now that they are a valuable tool) to be more
accurate. We can assign different "league average" values for various
pitchers depending on whether they are left handed or right handed,
starter or reliever, groundball pitcher or flyball pitcher and other
designations which could cause subtle shifts in the overall evaluations.
Comments?
--
Voros McCracken
vo...@daruma.co.jp
http://www.enteract.com/~mccracke/dips
"As always, victory finds a hundred fathers
but defeat is an Orphan"
-Count Galeazzo Ciano
--
* Keith Woolner, Moderator for rec.sport.baseball.analysis *
* Submissions: rs...@stathead.com *
* Questions/Info/Contact: rsba-r...@stathead.com *
* Charter: http://www.stathead.com/rsba-charter.htm *