Swing, Swing by The Small-Sample Rejects
![](https://cdn.statically.io/img/blogs.fangraphs.com/wp-content/uploads/2024/07/USATSI_23724792.jpg)
Have you ever had a friend enthusiastically recommend that you watch a TV show and then say, “It takes a few episodes to get going, and the timeline gets weird at the end, and one or two of the main characters can be kind of annoying, but other than that it’s SO GOOD.” And initially you might be put off, thinking that a truly good show wouldn’t require that many qualifiers. Sometimes you’re right about that, but sometimes it turns out the show is Parks and Recreation and even though the first season is about as appealing as living in a pit, the rest of the show is an absolute treat.
Sometimes small components of a larger body of work do a poor job of representing the work as a whole. The oddities that occur in small samples are likely not a new concept to FanGraphs readers, nor will it shock anyone when I note that what constitutes a small sample depends on what exactly we want to measure. Recently, the fine folks at MLB Advanced Media gifted us with a handful of new metrics that make use of Statcast’s bat tracking technology. Every time we dig into a new metric, we must consider the appropriate serving size to satiate our hunger for knowledge, lest we find ourselves hangrily generating takes that we later regret.
For this article, we’ll attempt to determine appropriate sample thresholds for measuring a hitter’s average bat speed; so that players without bats don’t feel left out, we’ll do the same for sword rate from the pitcher’s perspective. For many metrics, the sample size is measured in pitches or plate appearences, but since both bat speed and sword rate are tied specifically to bat movement, their samples will be composed of swings. To determine reasonable sample sizes, I used the split-half correlation method. The idea is to randomly select two samples of size X from a player’s collection of swings, calculate the player’s average bat speed or sword rate for both samples, lather/rinse/repeat for a bunch of players, then take the full set of two-sample pairs for all players and see how well they correlate. We complete the experiment by repeating the process for progressively larger sample sizes. And just to be super thorough, we’ll re-run the experiment several times and average the correlation values. Read the rest of this entry »