Adjusting Statistics

For those of you who consider yourselves “sabermetricians”, this article may be a bit of a bore, and you are encouraged to choose another of today’s offerings with your morning coffee or lunchtime sandwich. For the rest of you – a camp which sometimes I count myself in and sometimes not – the topic will hopefully be of more interest. Incidentally, for those of you who are really new to all this, a “sabermetrician” is someone who studies baseball analytically (that is, more or less scientifically – usually through statistics of various sorts), and “sabermetrics” is the analytical study of baseball.

I’m going to talk about a very basic (indeed, elementary) aspect of sabermetrics today. I admit that this is a bit of a tease, as it’s a lead-in to a further article that should appear in two (or four) weeks, which is about statistics for college players. But before we get into an extensive project like that, it can be worthwhile to step back and consider the most basic elements of sabermetric analysis. Such as today’s topic: adjusting statistics.

I should be precise about what I mean. By “adjusting statistics” I don’t mean doing something like multiplying earned runs by nine and dividing by innings pitched (which, of course, gives you ERA). Instead, I mean taking statistics – whether raw “counting stats” such as home runs, games played, or sacrifice flies, or compound stats such as batting average, runs created, or OPS – and adjusting them to account for the conditions under which those numbers were compiled.

The easiest way to tell an adjustment from a new compound statistic, is that with an adjusted statistic, once you have done your mathematical calculation to produce the adjustment, you have the same statistic as before. If you are (for example) adjusting a player’s home runs based on the park he played in, what you end up with is still a count of home runs, and not some other thing.

Park adjustments, or “park factors” if you like, are the canonical adjustment, so basic that very few serious analysts would give much consideration to raw statistics without taking the park factors into account. A park adjustment realizes that more runs, or fewer homers, or more strikeouts, happen in some parks than in others, and that players who play their home games in those parks will achieve that result more or less, as a consequence of the park. It’s a very simple idea, but very powerful, and bringing it into sabermetrics has been one of the central achievements of the field. But there are other types of adjustments which can be important, or useful, and I think it’s useful to look at these together.

Era adjustments is something we are using to some degree over at the Hall of Merit on Baseball Primer. Conditions in different eras in baseball history have produced massively different results, despite the game being largely unchanged in substance. For example, there were at least five times as many errors in the 1880s as there are today (due not so much to the skill of the players, but somewhat more to bumpy and uneven fields, a complete lack of artificial light, and no fielder’s gloves whatsoever). If you are measuring fielding percentages or unearned runs across that span of time, you may want or need to adjust for the era in which players played.

Another, simpler adjustment that is often made across eras is a season length adjustment. This one is real simple, obviously. A player’s achievements over a full season will look very different in a 70-game season like many in the 1870s, a 130-game schedule like in 1918, a 154-game schedule, or a modern 162-game schedule.

Another adjustment which is frequently useful is a competition adjustment. If you have two or more pools of players who have competed against different opponents, you will often want to adjust the stats they compiled based on the competition they faced. This is another frequent topic at the Hall of Merit.

In the 1880s, there were two major leagues (the American Association and the National League). The NL was generally superior to the AA, and if you are comparing a player who played in the AA to one who played in the NL, you will want to make an adjustment to their statistics so that the AA player doesn’t look better just because he faced inferior competition. This can also be important where teams play severely unbalanced schedules – as is the case in NCAA play (and, increasingly, in the major leagues).

There are other types of context adjustments that might be made as well. Defensive statistics can be adjusted to take account of various biases. Charlie Saeger’s Context-Adjusted Defense, for example, adjusts defensive plays made by each fielder according to the team’s groundball/flyball ratio and the team’s overall defensive ability.

More important than all these adjustments, though, is the more general question of why we adjust. What purpose does adjusting a statistic serve? This is not only an important question in analyzing statistics, but also in interpreting them – something even casual fans, and especially roto players, do every day.

Very often, what we are doing in looking at statistics is comparing players – sometimes in individual comparisons, sometimes in large groups (say, all the players in the NL), and sometimes against abstract benchmarks (such as the league average, or “replacement level”). Adjustments serve the comparison of players, by putting them on a level playing field with each other. Or they can serve to translate a player’s performance from one context to another.

Essentially, an adjustment is about changing the context in which a performance was done – either by putting it into some sort of different context (useful when trying to project what a traded player might do when switching teams) or some neutral context (useful when trying to compare performances of all players in an entire league, or players in different eras). If an adjustment isn’t putting numbers into a context that you find useful or interesting for your purposes, then you shouldn’t make it or pay attention to it.

But as in every field of human endeavour, context in baseball means a great deal. Dante Bichette can create 130 runs in Coors Field in 1995, and that number is certainly meaningful. Adjusting that number to take its context into account, and to compare it to Frank Robinson’s 96 runs created in Memorial Stadium in 1970 (a vastly superior season) doesn’t mean that the runs Dante Bichette was creating were worth any less than Frank Robinson’s runs.

But they occurred in a context where creating runs was much easier, and if we want to go beneath the statistics and find out actually how good the players were, then we need to make adjustments. We don’t need to make adjustments to find out what happened (i.e. what the statistics are); but we do need to make adjustments to find out the value of what happened.

For some comparisons, as in the study I am doing on college players, the adjustments required are so large that at times they can overwhelm what is actually being measured. But the right adjustments will still yield meaningful results. When someone says that a certain body of statistics is “useless”, they often mean that they haven’t been properly adjusted. No adjustment is going to be perfect; but in any circumstance where players play in different conditions, adjusting their statistics will help you immesurably in evaluating their performance.