THE BOOK--Playing The Percentages In Baseball

The database will be fun to look at.

I have made adjustments manually in the past (upwards for past performance for players who played through an injury and downwards for future performance for players coming off certain injuries), but it has always been a messy process because I did not have comparable information for all players.

I have not RTFA yet, but this looks like great stuff, and yes, an area that has been untapped so far - how do past injuries affect projections?  There would be 3 avenues I would think.  One, the player has been injured for a while and thus his historical stats are depressed, such that when he is healthy his performance will be better than his Marcel.  Two, he has an injury, usually an acute one, such that when he comes back he is not 100% and his performance will be worse than his Marcel, and the trick will also be to figure out the trajectory of the that performance in terms of getting better and better as he recovers over time. Third, does a certain kind of injury or for certain players does an injury cause a permanent (or very long term) decline in performance, such that his performance will be worse than his Marcel, at least for a while (until his pre-injury stats are not longer relevant).

And I agree that for the pitcher this type of thing might be more significant, but probably more tricky to analyze as well.

Right, you need some sort of “recovery rate”, and “level of recovery” parameter.

So, say you have a pulled hamstring.  You might recover back to 100% say 80% of the time, and the 20% of the time you don’t recover fully, you recover to the 60% level.

And maybe you distinguish between a speedster and not, the past frequencies, etc.  You really have to set up some sort of actuarial table.

A little off-topic, but this quote from Colin in the comments section of the article is important.  Those of you who are not that well-versed in statistical concepts should make sure you understand this.  We see questions all the time regarding, “How large of a sample is large enough to make reliable statistical inferences?”

Here is the answer from Colin:

I don’t like making a binary distinction between significant and not significant when it comes to sample size. Larger sample sizes are obviously more useful than smaller sample sizes, of course.

So you have to look at two things - the size of the sample and the magnitude of the effect. A large magnitude in a small sample can tell you something important, you just can’t treat it as being as important as an effect of that magnitude over a larger sample.

I’ll add one thing:  A small effect over a large sample can be quite significant as well.

I have a friend who knows just a little about statistical concepts. He often falls into the trap of ignoring any empirical result from a very small sample size.  For example, let’s say that in 10 observations, we get a sample mean of .8.  And let’s say that the null hypothesis expects a .5 mean.  My friend would say that the .8 is meaningless, because, “Anything can happen in 10 observations.” I have to remind him that .8 is actually 2 SD away from .5 in 10 observations, which means that it might be quite significant.  I also have to remind him that it is exactly (well, not quite) equivalent to 55% in 500 observations.