Tag Archives: SImulations

Is There a Hit-by-Pitch Hangover?

One of the things I’ve been curious about recently and have on my list of research questions is what the ramifications of a hit-by-pitch are in terms of injury risk—basically, how much of the value of an HBP does the batter give back through the increased injury risk? Today, though, I’m going to look at something vaguely similar but much simpler: Is an HBP associated with an immediate decrease in player productivity?

To assess this, I looked at how players performed in the plate appearance immediately following their HBP in the same game. (This obviously ignores players who are injured by their HBP and leave the game, but I’m looking for something subtler here.) To evaluate performance, I used wOBA, a rate stat that encapsulates a batter’s overall offensive contributions. There are, however, two obvious effects (and probably other more subtle ones) that mean we can’t only look at the post-HBP wOBA and compare it to league average.

The first is that, ceteris paribus, we expect that a pitcher will do worse the more times he sees a given batter (the so-called “trips through the order penalty”). Since in this context we will never include a batter’s first PA of a game because it couldn’t be preceded by an HBP, we need to adjust for this. The second adjustment is simple selection bias—not every batter has the same likelihood of being hit by a pitch, and if the average batter getting hit by a pitch is better or worse than the overall average batter, we will get a biased estimate of the effect of the HBP. If you don’t care about how I adjusted for this, skip to the next bold text.

I attempted to take those factors into account by computing the expected wOBA as follows. Using Retrosheet play-by-play data for 2004–2012 (the last year I had on hand), for each player with at least 350 PA in a season, I computed their wOBA over all PA that were not that player’s first PA in a given game. (I put the 350 PA condition in to make sure my average wasn’t swayed by low PA players with extreme wOBA values.) I then computed the average wOBA of those players weighted by the number of HBP they had and compared it to the actual post-HBP wOBA put up by this sample of players.

To get a sense of how likely or unlikely any discrepancy would be, I also ran a simulation where I chose random HBPs and then pulled a random plate appearance from the hit batter until I had the same number of post-HBP PA as actually occurred in my nine year sample, then computed the post-HBP wOBA in that simulated world. I ran 1000 simulations and so have some sense of how unlikely the observed post-HBP performance is under the null hypothesis that there’s no difference between post-HBP performance and other performance.

To be honest, though, those adjustments don’t make me super confident that I’ve covered all the necessary bases to find a clean effect—the numbers are still a bit wonky, and this is not such a simple thing to examine that I’m confident I’ve gotten all the noise out. For instance, it doesn’t filter out park or pitcher effects (i.e. selection bias due to facing a worse pitcher, or a pitcher having an off day), both of which play a meaningful role in these performance and probably lead to additional selection biases I don’t control for.

With all those caveats out of the way, what do we see? In the data, we have an expected post-HBP wOBA of .3464 and an actual post-HBP wOBA of .3423, for an observed difference of about 4 points of wOBA, which is a small but non-negligible difference. However, it’s in the 24th percentile of outcomes according to the simulation, which indicates there’s a hefty chance that it’s randomness. (Though league average wOBA changed noticeably over the time period I examined, I did some sensitivities and am fairly confident those changes aren’t covering up a real result.)

The main thing (beyond the aforementioned haziness in this analysis) that makes me believe there might be an effect is that the post-walk effect is actually a 2.7 point (i.e. 0.0027) increase in wOBA. If we think that boost is due to pitcher wildness then we would expect the same thing to pop up for the post-HBP plate appearances, and the absence of such an increase suggests that there is a hangover effect. However, to conclude from that that there is a post-HBP swoon seems to be an unreasonably baroque chain of logic given the rest of the evidence, so I’m content to let it go for now.

The main takeaway from all of this is that there’s an observed decrease in expected performance after an HBP, but it’s not particularly large and doesn’t seem likely to have any predictive value. I’m open to the idea that a more sophisticated simulator that includes pitcher and park effects could help detect this effect, but I expect that even if the post-HBP hangover is a real thing, it doesn’t have a major impact.

Valuing Goalie Shootout Performance (Again)

I wrote this article a few months ago about goalie shootout performance and concluded two things:

  • Goalies are not interchangeable with respect to the shootout, i.e. there is skill involved in goalie performance.
  • An extra percentage point in shootout save percentage is worth about 0.002 standings points per game. This is based on some admittedly sketchy calculations based on long term NHL performance, and not something I think is necessarily super accurate.

I’m bringing this up because a couple of other articles have been written about this recently: one by Tom Tango and one much longer one by Michael Lopez. One of the comments over there, from Eric T., mentioned wanting a better sense of the practical significance of the differences in skill, given that Lopez offers an estimate that the difference between the best and worst goalies is worth about 3 standings points per year.

That’s something I was trying to do in the previous post up above, and the comment prompted me to try to redo it. I made some simple assumptions that align with the one’s Lopez did in his followup post:

  • Each shot has equal probability of being saved (i.e. shooter quality doesn’t matter, only goalie quality). This probably reduces the volatility in my estimates, but since a goalie should end up facing a representative sample of shooters, I’m not too concerned.
  • The goalie’s team has an equal probability of converting each shot. This, again, probably reduces the variance, but it makes modelling things much simpler, and I think it makes it easier to isolate the effect that goalie performance has on shootout winning percentage.

Given these assumptions, we can compute an exact probability that one team wins given team 1’s save percentage p_1 and team 2’s p_2. If you don’t care about the math, skip ahead to the table. Let’s call P_{i,j} the probability that team i scores j times in the first three rounds of the shootout:

P_{i,j} = {3 \choose j} p_i^j(1-p_i)^{3-j}

P(\text{Team 1 Wins } | \text{ } p_1, p_2) = \sum_{j=1}^3 \sum_{k=0}^{j-1} P_{1,j} \cdot P_{2,k} + \left ( \sum_{j=1}^3 P_{1,j}\cdot P_{2,j} \right ) \frac{p_1(1-p_2)} {1-(p_1p_2+(1-p_1)(1-p_2))}

The first term on the right side is just the sum of the probabilities of the ways that team 1 can win the first three rounds, e.g. 2 goals for and 1 allowed or 3 goals for and none allowed. The term on the right is the sum of all the ways they can win if the first three rounds end in a tie, which can be expressed easily as the sum of a geometric series.

Ultimately, we don’t really care about the formula so much as the results, so here’s a table and a plot showing the performance of a goalies who are a given percentage below or above league average when facing a league average goalie:

Calculated Winning

Percentage Points Above/Below League Average Winning Percentage
-20 26.12
-19 27.14
-18 28.18
-17 29.24
-16 30.31
-15 31.41
-14 32.52
-13 33.66
-12 34.81
-11 35.98
-10 37.17
-9 38.37
-8 39.60
-7 40.84
-6 42.10
-5 43.37
-4 44.67
-3 45.98
-2 47.30
-1 48.64
0 50.00
1 51.37
2 52.76
3 54.16
4 55.58
5 57.01
6 58.45
7 59.91
8 61.38
9 62.86
10 64.35
11 65.85
12 67.37
13 68.89
14 70.42
15 71.96
16 73.51
17 75.06
18 76.62
19 78.19
20 79.76

We would expect most of these figures to be close to league average, so if we borrow Tom Tango’s results (see the link above) we figure the most and least talented goalies are going to be roughly 6 percentage points away from the mean. The difference between +0.06 and -0.06 is about 0.16 in the simulation output, meaning the best goalies are likely to win sixteen shootouts per hundred more than the worst goalies assuming both play average competition.

Multiplying this by 13.2%, the past frequency of shootouts, and we get an estimated benefit of only about 0.02 standings points / game from switching from the worst shootout goalie to the best. For a goalie making 50 starts, that’s only about 1 point added to the team, and that’s assuming maximal possible impact.

Similarly, moving up this curve by one percentage point appears to be worth about 1.35 wins per hundred; multiplying that by 13.2% gives a value of 0.0018 standings points / game, which is almost exactly what I got when I did this empirically in the prior post, which leads me to believe that that estimate is a lot stronger than I initially thought.

There’s obviously a lot of assumptions in play here, including the assumptions going into my probabilities and Tango’s estimates of true performance, and I’m open to the idea that one or another of those is suppressing the importance of this skill. Overall, though, I’m largely inclined to hew to my prior conclusions saying that for a difference in shootout performance to be enough to make one goalie better overall than another, it has to be a fairly substantial one, and the difference in actual save percentage has to be correspondingly fairly small.

Uncertainty and Pitching Statistics

One of the things that I occasionally get frustrated by in sports statistics is the focus on estimates without presenting the associated uncertainty. While small sample size is often bandied about as an explanation for unusual results, one of the first things presented in statistics courses is the notion of a confidence interval. The simplest explanation of a confidence interval is that of a margin of error—you take the data and the degree of certainty you want, and it will give you a range covering likely values of the parameter you are interested in. It tacitly includes the sample size and gives you an implicit indication of how trustworthy the results are.

The most common version of this is the 95% confidence interval, which, based on some data, gives a range that will contain the actual value 95% of the time. For instance, say we poll a random sample of 100 people and ask them if they are right-handed. If 90 are right handed, the math gives us a 95% CI of (0.820, 0.948). We can draw additional sample and get more intervals; if we were to continue doing this, 95% of such intervals will contain the true percentage we are looking for. (How the math behind this works is a topic for another time, and indeed, I’m trying to wave away as much of it as possible in this post.)

One big caveat I want to mention before I get into my application of this principle is that there are a lot of assumptions that go into producing these mathematical estimates that don’t hold strictly in baseball. For instance, we assume that our data are a random sample of a single, well-defined population. However, if we use pitcher data from a given year, we know that the batters they face won’t be random, nor will the circumstances they face them under. Furthermore, any extrapolation of this interval is a bit trickier, because confidence intervals are usually employed in estimating parameters that are comparatively stable. In baseball, by contrast, a player’s talent level will change from year to year, and since we usually estimate something using a single year’s worth of data, to interpret our factors we have to take into account not only new random results but also a change in the underlying parameters.

(Hopefully all of that made sense, but if it didn’t and you’re still reading, just try to treat the numbers below as the margin of error on the figures we’re looking at, and realize that some of our interpretations need to be a bit looser than is ideal.)

For this post, I wanted to look at how much margin of error is in FIP, which is one of the more common sabermetric stats to evaluate pitchers. It stands for Fielding Independent Pitching, and is based only on walks, strikeouts, and home runs—all events that don’t depend on the defense (hence the name). It’s also scaled so that the numbers are comparable to ERA. For more on FIP, see the Fangraphs page here.

One of the reasons I was prompted to start with FIP is that a common modification of the stat is to render it as xFIP (x for Expected). xFIP recognizes that FIP can be comparatively volatile because it depends highly on the number of home runs a pitcher gives up, which, as rare events, can bounce around a lot even in a medium size sample with no change in talent. (They also partially depend on park factors.) xFIP replaces the HR component of FIP with the expected number of HR they would have given up if they had allowed the same number of flyballs but had a league average home run to fly ball ratio.

Since xFIP already embeds the idea that FIP is volatile, I was curious as to how volatile FIP actually is, and how much of that volatility is taken care of by xFIP. To do this, I decided to simulate a large number of seasons for a set of pitchers to get an estimate for what an actual distribution of a pitcher’s FIP given an estimated talent level is, then look at how wide a range of results we see in the simulated seasons to get a sense for how volatile FIP is—effectively rerunning seasons with pitchers whose talent level won’t change, but whose luck will.

To provide an example, say we have a pitcher who faces 800 batters, with a line of 20 HR, 250 fly balls (FB), 50 BB, and 250 K. We then assume that, if that pitcher were to face another 800 batters, each has a 250/800 chance of striking out, a 50/800 chance of walking, a 250/800 chance of hitting a fly ball, and a 20/250 chance of each fly ball being a HR. Plugging those into some random numbers, we will get a new line for a player with the same underlying talent—maybe it’ll be 256 K, 45 BB, and 246 FB, of which 24 were HR. From these values, we recompute the FIP. Do this 10,000 times, and we get an idea for how much FIP can bounce around.

For my sample of pitchers to test, I took every pitcher season with at least 50 IP since 2002, the first year for which the number of fly balls was available. I then computed 10,000 FIPs for each pitcher season and took the 97.5th percentile and 2.5th percentile, which give the spread that the middle 95% of the data fall in—in other words, our confidence interval.

(Nitty-gritty aside: One methodological detail that’s mostly important for replication purposes is that pitchers that gave up 0 HR in the relevant season were treated as having given up 0.5 HR; otherwise, there’s not actually any variation on that component. The 0.5 is somewhat arbitrary but, in my experience, is a standard small sample correction for things like odds ratios and chi-squared tests.)

One thing to realize is that these confidence intervals needn’t be symmetric, and in fact they basically never are—the portion of the confidence interval above the pitcher’s actual FIP is almost always larger than the portion below. For instance, in 2011 Bartolo Colon had an actual FIP of 3.83, but his confidence interval is (3.09, 4.64), and the gap from 3.83 to 4.64 is larger than the gap from 3.09 to 3.83. The reasons for this aren’t terribly important without going into details of the binomial distribution, and anyhow, the asymmetry of the interval is rarely very large, so I’m going to use half the length of the interval as my metric for volatility (the margin of error, as it were); for Colon, that’s (4.64 – 3.09) / 2 = 0.775.

So, how big are these intervals? To me, at least, they are surprisingly large. I put some plots below, but even for the pitchers with the most IP, our margin of error is around 0.5 runs, which is pretty substantial (roughly half a standard deviation in FIP, for reference). For pitchers with only about 150 IP, it’s in the 0.8 range, which is about a standard deviation in FIP. A 0.8 gap in FIP is nothing to sneeze at—it’s the difference between 2013 Clayton Kershaw and 2013 Zack Greinke, or between 2013 Zack Greinke and 2013 Scott Feldman. (Side note: Clayton Kershaw is really damned good.)

As a side note, I was concerned when I first got these numbers that the intervals are too wide and overestimate the volatility. Because we can’t repeat seasons, I can’t think of a good way to test volatility, but I did look at how many times a pitcher’s FIP confidence interval contained his actual FIP from the next year. There are some selection issues with this measure (as a pitcher has to post 50 IP in consecutive years to be counted), but about 71% of follow-up season FIPs fall into the previous season’s CI. This may be a bit surprising, as our CI is supposed to include the actual value 95% of the time, but given the amount of volatility in baseball performance due to changes in skill levels, I would expect to see that the intervals diverge from actual values fairly frequently. Though this doesn’t confirm that my estimated intervals aren’t too wide, the magnitude of difference suggests to me it’s unlikely that that is our problem.

Given how sample sizes work, it’s unsurprising that the margin of error decreases substantially as IP increases. Unfortunately, there’s no neat function to get volatility from IP, as it depends strongly on the values of the FIP components as well. If we wanted to, we could construct a model of some sort, but a model whose inputs come from simulations seemed to me to be straying a bit far from the real world.

As I only want to see a rule of thumb, I picked a couple of round IP cutoffs and computed the average margin of error for every pitcher within 15 IP of that cutoff. The 15 IP is arbitrary, but it’s not a huge amount for a starting pitcher (2–3 starts) and ensures we can get a substantial number of pitchers included in each interval. The average FIP margin of error for pitchers within 15 IP of the cutoffs is presented below; beneath that is are scatterplots comparing IP to margin of error.

Mean Margin of Error for Pitchers by Innings Pitched
Approximate IP FIP Margin of Error Number of Pitchers
65 1.16 1747
100 0.99 428
150 0.81 300
200 0.66 532
250 0.54 37

FIP Scatter xFIP Scatter

Note that due to construction I didn’t include anyone with less than 50 IP, and the most innings pitched in my sample is 266, so these cutoffs span the range of the data. I also looked at the median values, and there is no substantive difference.

This post has been fairly exploratory in nature, but I wanted to answer one specific question: given that the purpose of xFIP is to stabilize FIP, how much of FIP’s volatility is removed by using xFIP as an ERA estimator instead?

This can be evaluated a few different ways. First, the mean xFIP margin of error in my sample is about 0.54, while the mean FIP margin of error is 0.97; that difference is highly highly significant. This means there is actually a difference between the two, but looking at the average absolute difference of 0.43 is pretty meaningless—obviously a pitcher with an FIP margin of error of 0.5 can’t have a negative margin of error. Thus, we instead look at the percentage difference, which gives us the figure that 43% of the volatility in FIP is removed when using xFIP instead. (The median number is 45%, for reference.)

Finally, here is the above table showing average margins of error by IP, but this time with xFIP as well; note that the differences are all in the 42-48% range.

Mean Margin of Error for Pitchers by Innings Pitched
Approximate IP FIP Margin of Error xFIP Margin of Error Number of Pitchers
65 1.16 0.67 1747
100 0.99 0.53 428
150 0.81 0.43 300
200 0.66 0.36 532
250 0.54 0.31 37

Thus, we see that about 45% of the FIP volatility is stripped away by using xFIP. I’m sort of burying the lede here, but if you want a firm takeaway from this post, there it is.

I want to conclude this somewhat wonkish piece by clarifying a couple of things. First, these numbers largely apply to season-level data; career FIP stats will be much more stable, though the utility of using a rate stat over an entire career may be limited depending on the situation.

Second, this volatility is not something that is unique to FIP—it could be applied to basically any of the stats that we bandy about on a daily basis. I chose to look at FIP partially for its simplicity and partially because people have already looked into its instability (hence xFIP); in the future, I’d like to apply this to other stats as well; for instance, SIERA comes to mind as something directly comparable to FIP, and since Fangraphs’ WAR is computed using FIP, my estimates in this piece can be applied to those numbers as well.

Third, the diminished volatility of xFIP isn’t necessarily a reason to prefer that particular stat. If a pitcher has an established track record of consistently allowing more/fewer HR on fly balls than the average pitcher, that information is important and should be considered. One alternative is to use the pitcher’s career HR/FB in lieu of league average, which gives some of the benefits of a larger sample size while also considering the pitcher’s true talent, though that’s a bit more involved in terms of aggregating data.

Since I got to rambling and this post is long on caveats relative to substance, here’s the tl;dr:

  • Even if you think FIP estimates a pitcher’s true talent level accurately, random variation means that there’s a lot of volatility in the statistic.
  • If you want a rough estimate for how much volatility there is, see the tables above.
  • Using xFIP instead of FIP shrinks the margin of error by about 45%.
  • This is not an indictment of FIP as a stat, but rather a reminder that a lot of weird stuff can happen in a baseball season, especially for pitchers.