I admit that starting pitcher length is a tired subject. We all know that it’s gone inexorably down, and we know about how long starting pitchers last in today’s game. We watch games, after all. So to highlight starting pitcher length is generally not to ferret out meaning.
But I got to thinking of the subject in actuarial terms. What I mean by this is, while one is trivial and one is too big to contemplate, the road to a pitcher’s getting knocked out of a game or running out of steam can be compared to the road to one’s death. At a sufficiently advanced age, there are few survivors, just because every unit of time you are around is a time you might die, and because these units of time get increasingly fraught with age. So it is with starting pitchers.
To introduce my statistical framework, I used only home starts, since the losing road team only plays defense for 8 innings unless a game is a walk-off, and I eliminated other games of less than 9 innings. Since it is safe to assume starting pitchers will not pitch into extra innings, I would have liked to have included extra-inning games, but Stathead’s prompts would have made that difficult, so they, too had to go to the cutting room floor. The sample, again, is home starts in 9-inning games.
2023 Survival Rates
Within these parameters, if you compare the number of starting pitchers entering an inning to the number completing that inning, here is the actuarial progression for 2023.
Through 1st: .996
Through 2nd: .968
Through 3rd: .959
Through 4th: .924
Through 5th: .832
Through 6th: .578
Through 7th: .320
Through 8th: .181
Through 9th: .358
It’s not necessarily surprising that the rate decreases each inning, 1 through 8, but it is interesting to see this quantified, and it is interesting to see exactly how steep the descent is.
If you multiply the 9 rates together, you get .0085, which corresponds to the 19 complete games from 2226 attempts. This should also help you understand the meaning of the survival rates, if you didn’t initially. In other words, that .832 does not mean that 83.2% of starters pitched 5+ innings. It means that 83.2% of the ones who pitched 4 made it through the 5th.
Perhaps the most interesting part of the graph is how high the final survival rate was for pitchers who made it through 8 innings. That number seemed ticketed for single digits, looking at the others. And pitchers are surely usually running into non-negotiable pitch limits at that point, right? Yet the number increased, not just from the 8th-inning survival rate (.181), but from the 7th-inning survival rate as well (.320). I don’t know of any other way to read this than that, despite everything we hear, managers do try to get pitchers complete games. The actual numbers were 240 7-inning and 7+ inning starts, 34 8-inning and 8+ inning starts, and 19 complete games. The .358 comes from 19 of 53.
We know that deference is paid to pitchers going for no-hitters, but the frequency of this situation doesn’t seem nearly high enough to explain the complete game survival rate, even before we remember that pitchers who lose their no-hitters can be replaced, and would then fall into the 8+ category. I think managers value the achievement of complete games, and I think on certain days and nights they like the idea of completely resting their bullpen.
You could also say, perhaps, that the sample of pitchers who negotiate 8 innings already consists of he-men (like Sandy Alcantara, 2022) who see themselves as also capable of going 9 innings. The high 9th-inning survival rate then doesn’t so much show that the lure of a complete game affects decisions in general, as that pitchers are categorized as “9 inning possibles” and “7 and less” pitchers. So you have something of a dichotomous distribution at play. If this is what goes on, I would say the distinction isn’t entirely rational, although I’ll keep my mouth shut if it means we get the occasional complete game. I have a feeling most of you like the romance of the complete game as well.
There are some other things to note in the data, but I will mostly hold off until later. I will point out that that .968 2nd-inning survival rate bears the mark of openers, but that it is as high as it is indicates that they were really comparatively rare.1 Openers don’t seem to be particularly affecting completion of the first inning; I guess you could think of the opening as generally synonymous with the first inning, and looking at all pitchers, regular and openers alike, they combined to get through that 99.7% of the time. But then those openers who are out of the game in the second inning affect the rate of pitchers going from the 1st to the 2nd.
Historical Survival-Rate Data
I am afraid it has the appearance of adding a tired approach to a tired subject, but I had my reasons, and after calculating the survival rates for 2023, I then did exactly what you would have expected me to do, extending my work to other years. I chose four of them: 2008 (AL), 1973 (AL), 1941 (MLB), and 1912 (MLB). I kept to the same requirement of home games that lasted exactly 9 innings. The samples for each of the new years are more than 50% lower than the sample for 2023, but they are very close in size to one another, with 1912, 1941, and 2008 ranging from 1044 to 1098 starts, and 1973 including 876 starts.
In the years that didn’t have a DH, pitchers were subject to another cause of death in the need to hit for them in certain circumstances. This means that the data gain an unnatural element, an element beyond considerations of workload, as it can be nearly as urgent to swap out the pitcher for a real hitter in the 7th as the 9th.
Here are the survival rates from inning bridge 0-1 through inning bridge 8-9 for the five years.
2023: .996, .968, .959, .924, .832, .578, .320, .181, .358
2008: .995, .988, .979, .962, .904, .755, .517, .302, .291
1973: .969, .967, .955, .932, .917, .888, .835, .748, .790
1941: .978, .966, .956, .953, .941, .911, .901, .846, .827
1912: .983, .958, .954, .958, .951, .941, .930, .929, .883
Discussion of Last Five Innings, Historical
To deal with the unsurprising parts first, for the last five bridges, the rates are in almost perfect reverse order of year. In other words, look at any of those columns, and 1912 leads the pack, 2023 is last, and the other years fall in accordingly.
The only exception is that the 9th-inning survival rate in 2008 was .291, lower than the .358 of 2023. Just based on this, you might say that 2008 didn’t show a complete game effect, but if you’ll look more closely at the data, you’ll see that you were dead wrong. Given the previous three survival rates of .755, .517, and .302, the 9th-inning survival rate of 2008 had no business being as high as .291. All you can say in the other direction is that the consideration of complete games wasn’t as strong as in 2023. That number was so elevated, some correction to our understanding was likely. With relatively few starts even going through the 8th, sample fluctuation plays a much bigger role than with other survival rates.
Complete Game Rates, Historical
The product of the numbers, which is the same as the complete game rate, returns the same verdict we expected going in: 58.9% for 1912, 46.4% for 1941, 33.6% for 1973, 2.9% for 2008, 0.9% for 2023.2
Discussion of First Four Innings, Historical
The “simple and boring” narrative takes another blow when we assimilate the survival rates of the first four innings. We learn that teams aren’t simply less patient with their starters than they used to be, and aren’t simply more willing to make changes at all times. If there is a general trend conveyed by the data, it is that early hooks have in fact become less common, while of course accompanied by almost automatic eventual hooks. Of the years sampled, this pattern reached its apotheosis in the 2008 AL, which had the highest survival rate of the five years in the 2nd, 3rd, and 4th innings. Then, quite a few first-inning removals could have been added to 2023 (.996) and 2008 (.995) and still left them with a higher survival rate than the first-inning survival rates of the other three years, which ranged from .969 to .983.
If not for openers, it is likely 2023’s 2nd-inning rate would have been right with 2008’s; as it is, it is still higher than the survival rate of the other three years. There are certainly some two-inning openers, but they probably don’t completely explain 2008’s .979 to .959 third-bridge edge over 2023. The 4th-inning benchmark is definitely starting to be something that teams insist upon less; at .924 in 2023, the fourth-inning survival rate was competitive with 1912, 1941, and 1973, but lower than them.
I want to emphasize the significance of 2008 having a higher survival rate for each of the first four innings than any of the three previous years. Remember that one high survival rate does nothing to help the next (the scores are independent). So, if we looked at four-inning starts the way we look at complete games, we would get completion percentages of
2008 .926
1912 .860
1941 .860
2023 .854
1973 .835
2008 is out ahead of the pack.
Among the other years, there’s not much distinguishing them, although you can make a case that the managers of 1973 were particularly impatient. Billy Martin’s impulsivity and showmanship seem to go with quick hooks, and he managed the Tigers for most of that year, was relieved of his job, then took over the Rangers for their final 23 games. Martin himself wasn’t and probably couldn’t have moved these data, but the 1970s was a colorful time in more ways than one, and my hunch is that there were other managers in the Billy Martin mold.
Patterns in Survival Rates: Both General, and Within Individual Years
Returning to the complete data, changes that we see within the data strings can perhaps reflect three different dynamics. 1) A general drop through the innings (the slope is simply negative, although in recent years, the function is certainly not linear). 2) A threshold drop because of pitcher or game considerations. (The drop is around a particular inning or innings. For example, managers take the idea of six-inning pitchers to heart, or maybe seek to get their starter to the primary setup man.) 3) A threshold change with a statistical origin (e.g., the five-inning start and the “win” drives data; the complete game drives data).
I talk about managers in some ways apparently have been less by the book in the old days (orchestrating many complete games, but also sometimes taking their starters out very early), but it is important to note that, if the 9th inning is set aside, every year except 1912 does show straight decline for bridges 1-8s. So this fits into dynamic #1. The early years just started with the survival rates being a bit lower than they are in the modern game, and less severe slopes applied.
I would classify the 7th as a threshold point in 1941. 91.1% of pitches moved from the 5th to the 6th, 90.1% moved from the 6th to the 7th, but only 84.6% moved from the 7th to the 8th. So (and I don’t want to exaggerate this) managers seemed to stop and consider, “Do I want my pitcher to pitch the 8th?"? The 8th inning was not a given. As in most years, there was a special calculus for the 9th, too, where managers had to weigh the starter’s fatigue against the desire to get him a complete game.
In both 2008 and 2023, the survival rate falls below 80% for the first time from the 5th to the 6th. This might register as meaningful and reflect that managers give a sigh of relief after the pitcher gets through 5 and can get his win. After that point, they can finally do what they want without restraint. On the other hand, the survival rate in both 2008 and 2023 fell dramatically over the next bridge as well. So it is unclear how to understand the drop after the 5th inning.
In 1973, you absolutely see the complete game effect, just as you do in 2008 and 2023 (the last four numbers: .888, .835, .748, .790).
The complete game effect was smaller in 1941, but I would still aver it. The 9th-inning survival rate is lower than the 8th-inning survival rate, but the difference between inning 9 and inning 8 is just 1.9%, when the preceding drop was 5.5%. The attrition function would seem to call for a drop of much larger than 5.5%, not a smaller one.
The data for 1912 are in many ways distinct. The threshold for removal hardly seems to drop at all throughout a start, and there is absolutely not a complete game effect. In fact, really the only visible crack came after the pitcher had thrown 8 innings, with the survival rate “only” .883 for the 9th. Even with the attrition that had occurred through the first 8 innings, 9th-inning removals outpaced all other innings by at least 26 (84 pitchers threw only 8 innings or 8+, while, for innings 2-7, the number of pitchers out after each of those innings ranged from 41-58).3
Pitcher Removals: Within, vs. After Innings, Introduction
The other path by which I came to these analyses, aside from being interested in framing pitcher removals in actuarial terms, was that I had noticed that the unforgettable Erick Fedde of my notes had had many 6-inning starts in his career but few of 6.1 or 6.2. The hypothesis I eventually favored to explain this had to do with the actuarial aspect — that it was simply unlikely for Fedde to be pushed past the 6th — rather than that it reflected hesitation to remove pitchers in the middle of the 7th inning or in the middle of innings generally. But I didn’t quite discard this idea. So, in getting the actuarial rates, instead of just aggregating all starts by number of innings and the partial innings above it (the starts of 6, 6 1/3, and 6 2/3, let’s say), I initially compiled the frequency for the whole innings and the partial innings separately (e.g., leaving 6 separate from 6 1/3 and 6 2/3, the latter two of which I combined).
When I did this for 2023, one thing I noticed was that, compared to other inning breakdowns, the ratio of 5-inning starts to starts of 4+ innings was rather high.4 This suggested to me that teams might still care about giving their pitchers an opportunity to get a win. This was interesting because it flew in the face of what I often seem to observe, which is that the gentleman’s agreement that the starter goes 5 innings and qualifies for a potential win no longer exists. Pitchers may still throw their gloves in the dugout when they’re taken out in the 5th, but managers don’t care.
In any event, if I had statistical evidence backing the win as a force keeping pitchers in games in 2023, I guessed it would emerge as very powerful for earlier years. This actually is what drove my choice of 2008 as one of the other years to examine. It seemed to me that 2008 was rather safely before modern analytics, rather safely before everyone seemed to jointly renounce the win.5
Of course, comparison of the 4th-inning, 5th-inning, and 6th-inning survival rates also shed light on this subject. Those data, as I discussed, did not reveal the 5th inning as any kind of threshold.
Whether full versus partial-inning starts are compared, or departure times, I find the interpretation difficult, perhaps because wanting to enable a pitcher to qualify for a win doesn’t necessarily mean that a manager takes him out right after he does so. And sometimes allowing a pitcher to go 4.1 or 4.2 indicates he was given a final chance to make it 5. Maybe it is really removal of pitchers at these points who aren’t in trouble that indicates indifference to the statistical win. But in this macro study, I didn’t investigate in that kind of depth.
Just as a five-inning push can be posited by comparing 5-inning starts to starts of 4+, it seemed to me that the same approach could be taken comparing complete games to 8+ inning starts. In contrast to the normal-looking actuarial data around the 5th that characterized the historical analysis, we know that this approach did support the existence of a robust complete-game push. So I wanted to see if the theory would hold up comparing the complete game count to the 8.1 and 8.2 count. To gauge either the 5-inning pull or the 9th-inning pull, of course, for the sake of comparison, it is necessary to have the data for all innings.
A couple final notes before diving into the data. First, innings game data does not distinguish between final innings for a pitcher that didn’t involve his trying the next inning at all, and those where he did go back out for the next inning (the true “inning+” in the box score) but didn’t get an out. In an ideal world, we would certainly have that data. So, as it is, my percentage of partial-inning starts is severely undercounted. “Zero out” situations, after all, represent one-third of out situations, although I don’t know if they comprise a full third of starts when a pitcher goes part of the final inning.
Second, the exception to this class of systematically missing data comes when starters fail to get a single out in a game (such starts show up as 0 innings). Having the information, I included these as partial inning outings when calculating the comparison of 1.0 innings starts to starts of less than an inning. But there is also a number I will share which is the percentage of full-inning starts, taking out complete games. For this percentage, I counted the “0 inning games” as games of a full (or round) number of innings.
Percentage of Complete-Inning Removals by Inning, 2023
First, here are the data that prompted me to obtain the pattern in full for all five years. Breaking down games by start length, below is the percentage where the starter began the final inning and also finished it (noting the proviso in the last two paragraphs of the preceding section). So, for the 8th inning, the .627 corresponds to 32 8-inning starts vs. 19 starts of 7.1 or 7.2.
1 vs. 0.1/0.2: .808
2 vs. 1.1/1.2: .569
3 vs. 2.1/2.2: .576
4 vs. 3.1/3.2: .615
5 vs. 4.1/4.2: .696
6 vs. 5.1/5.2: .665
7 vs. 6.1/6.2: .674
8 vs. 7.1/7.2: .627
9 vs. 8.1/8.2: .905
The imbalance represented by the .808 of the 1st inning is again a product of the opener effect. Openers, while most often assigned 1 just inning, are rarely jerked before they complete it.
If the 9 vs. 8+ ratio is to be understood as a complete game effect, it rates as beyond robust (those 19 CG were offset by just two starts of 8.1 or 8.2 innings). Again, certainly the 9th-inning data do need to be interpreted in light of the small sample size, and it would be reasonable to expect a lower number this year (not because anything has changed, but just because the only two 8+ inning starts tell us that complete games are pretty unlikely in this day and age).
How much to make of the limited distance between the 5-inning full and the other full-inning percentages is debatable, but the 5th did have the highest score outside of the 1st and 9th.
Timing of Pitcher Removals: Broad Era Differences
When I began to study the historical data, something that emerged that I didn’t necessarily expect going in was that there have been general changes in how often pitchers have been removed at the end of innings or in the middle of them.
In trying to gauge this aspect, I think one has to remove complete games. First, they were 65 times as likely in 1912 as 2023, so will exert an enormous influence on the ratio of full to partial final innings. Second, I don’t think the decline in complete games at all reflects a greater comfort with taking a pitcher out in the middle of an inning; I think it reflects many other factors, such as wanting to keep pitchers healthy, taking advantage of flame-throwing relievers, and wanting to make starters more effective by not asking them to pace themselves.
Without complete games, the percentage of round-number starts in each year was as follows.
2023 66.4
2008 64.3
1973 37.3
1941 57.5
1912 70.4
So 1973 stands out, and the trend was to remove pitchers in the middle of innings. In all of the games in question, there was a pitching change made, so the data don’t speak to relative embrace of relievers. It is theoretically possible that they reflect the degree of focus on individual match-ups and the tendency to change pitchers accordingly, but since relievers were often going multiple innings in 1973, this is a theory that doesn’t seem to fit.
More plausible is that the low 1973 percentage of full-inning starts indicates, broadly, that managers required their pitchers to be knocked out rather than making more elective changes. Sure, unless a pitcher sustains an injury, he doesn’t absolutely have to be taken out at any juncture, but every manager and pitcher does have a threshold for offensive barrage, and as a group, it is very likely that starters taken out mid-inning show uncharacteristically bad statistics.
And you can certainly ask if pitchers were coming out in the middle of innings in 1973 because the designated hitter had been adopted, and pitchers for the first time did not have to be pinch hit for (if a pitcher is removed for a pinch hitter, he will always register for a full-inning outing.) I think this is sound thinking, and it was a conscious decision on my part to survey the first year of the designated hitter.6
However, there are still a couple of things that are curious. One, with the DH in 2008 and 2023, and no pinch hitters, we trended much more to full-inning starts than in 1973. Second, regarding the other part of my theory, the high intra-inning removal rate in 1973 came in the context of a relatively high percentage of overall early removals, if you will remember. I suppose having an occasional quick hook and a preference for giving your guy a last chance on the mound aren’t absolutely incompatible,7 but my point is that it’s not like managers handled their starters passively in general. My sense is that managers were adjusting to the DH rule, and they would adjust further. A thorough study would no doubt involve finding out just how often pinch hitters did spell the end of starts before the DH, and what the overall full-inning rate was for, let’s say, 1972.
Comparison of Innings in Historical “Timing of Removal” Data
I can’t present a table here for all five years’ worth of the “timing of removal” data, so if my discussion of highlights isn’t sufficient for you, please refer to the “full percentage” tab in the attached spreadsheet.8
Although generally slightly lower, the full-inning rates for 2008, inning by inning, look much like 2023. They are much lower for the first two innings, but the low sample size at those points in 2008, and the lack of openers, require us not to pay this much heed. The data for 2008 and 2023 converge and differ from 1941 and 1912 in that the rates are flat or slightly descend from the 5th inning through the 8th inning. However, to my surprise, the 5th-inning rate (.671) was actually lower in 2008 than it was in 2023, indicating no greater push to have starters pitch 5 innings then. And the ratio of 6-inning to 5+ starts (.683 score) in 2008 was actually greater than the ratio of 5 inning to 4+ starts.
Across the years, the full-inning percentages are lower in the first 3 innings than later in the game. First, our reference points are that the average inning returned a .598 rate, the median inning a .602 rate. Collapsing across the years, the averages were .412, .448, and .509 for the 1st through the 3rd, respectively, and .368, .482, and .511 for those innings if we use the median. 2023 again has that .808 for the 1st, driven by the opener invention, but other than that, the highest score for one of the first three innings for the five years is the .600 for the 3rd in 1912. By contrast, 23 of the 30 scores for innings 4-9 are .600 or higher.
I think the takeaway is again that managers tend to keep their pitchers in games early unless they are knocked out. So the higher rate of partial-inning starts in short outings reflects the longer leash that is in place generally early in games (as reflected by the descending survival rates revealed by the innings analysis).
You may have noticed that, not only were the full-inning rates for the first three innings well under the overall average, but that the pattern of increase held when they were considered separately as well (the 3rd had a higher full-inning removal percentage than the 2nd, which in turn was higher than the 1st). If you chart out all of the innings, this pattern continues.
Average Full-inning Removal Rate (2023, 2008, 1973, 1941, 1912)
1st: .412
2nd .448
3rd .509
4th .557
5th .600
6th .642
7th .651
8th .647
9th .917
It’s not visible here, but as you would expect, there is some variation within each inning. Consequently, other than noting the stunning size of the complete game effect, the pattern initially largely escaped me, and I’m still not sure how strongly I should tout it for the late innings.
If a year increased in this statistic for every inning, it would have 8 increases. No year did that. Three years increased 6 times (2008, 1941, 1912), one year increased 5 times (1973), and one year increased 4 times. But as you can see, on average, the pattern just missed running the table.
The one descent on average, from .651 in the 7th to .647 in the 8th, occurred despite a .963 8th-inning score in 1912. This is one of several quirks marking the 1912 data. To illustrate that score, while 8-inning starts outpaced 7-inning ones 77 to 52 in 1912, 7.1 and 7.2 inning starts (3 occurrences) were a small fraction of 6.1 and 6.2 inning starts (22). While the attrition rates cited earlier do not indicate a complete game effect in 1912, the full-inning rates perhaps show a special aversion to taking pitchers out as early as the 8th-inning (the 9th-inning full-inning rate, for its part, was .989 — 636/643).
Based on the full-inning rates, the best cases for a 5th-inning/win effect can be made for 2023 and 2008, and as we saw, it isn’t very strong. The 5th-inning round rates for 1973, 1941, and 1912 clearly argue against a 5th-inning effect existing for those years. When I add these data to the attrition data, I have to say that I have no statistical evidence that a 5th-inning effect has been a durable fact in MLB history.
This is not hard to accept for 1912 and 1941, and maybe not even for 1973. It stands to reason that, in these eras, if you felt inclined to take your pitcher out before the end of the 5th inning, because of sub-par performance he probably wasn’t in position to win the game, anyway.
I put numbers to this by asking Stathead for the percentage of games in each sample that were wins of exactly 5 innings. They made up 5.9% of starts in 2023, 3.8% in 2008, 0.8% in 1973, 0.5% in 1941, and 0.6% in 1912. I found seven 5-inning wins in 1973, and six in both 1941 and 1912. I’m some sort of surprised there were any in 1912.
It really is the low number of starts meeting the innings criterion keeping 5-inning wins at bay. Given only 22 5-inning starts in 1973, 7 wins from that group is pretty impressive. This suggests that the 5-inning effect may be there, but of so little practical significance you can’t see it.
…
The one constant pattern across the full versus partial inning analyses was the complete game effect. One does wonder if I am not taking liberties using that moniker, if the spike in the numbers really means something else and is a statistical confound of some sort, but if so, I’m not sure what the shape of that would be. Remember, because I only used home games, pitchers did not need to be ahead in the game to have their starts included. And, since it was the top of the 9th, pitchers were also never “walked off.” The manager was always free to remove them, and if he did, this would be included in the partial-innings count. But the data show that, when starting pitchers get the 9th inning, they finish that inning much more often than other innings. Finally, my incredulity would run a lot hotter if the attrition approach did not return the same story.
Side Issues
You might be interested in the data I have presented for many reasons, some of which differ from mine, so I wanted to explain what effect the games I omitted had on the survival rates. If you remember, I felt it necessary to omit road games, games of more than 9 innings, and games of less than 9 innings.
Maybe I’m mistaken, but I’m going to trust that you don’t care about the effects of the last decision. In any event, I didn’t look into it.
As for home and road, we know performance is generally better at home. But while I am not an expert, I know this manifests to different degrees depending on what aspect of performance is being examined. For instance, any category pertaining to the strike zone is likely to show a fairly large home field advantage. In contrast, in my post about home and road doubles totals9, because of an offsetting bias (involving plate appearances), I cited years of remarkable parity between home and road home runs.
The question here is starter length. Those 2226 starts of 2023 in 9-inning games taken by home starting pitchers all had their counterparts in the performances of the road starting pitchers. Knowing this, I found how many times road starters made it 6 innings or more, compared to home starters. The difference was a rate of 41.1% through the 6th for home starters, and 38.1% for road starters. If you look at this in terms of the totals, the difference gets your attention: 915 to 849.
I was surprised the disparity was this great, and what it suggested to me was that road starting pitchers may leave games earlier because, driven by more walks, their pitches rack up faster. Unfortunately, the way that FanGraphs has their “splits” set up, I don’t think there is a way to check average number of pitches per inning at home and on the road. In terms of overall performance, home teams only had a slight edge on road teams, winning 52.1%. The home field advantage is pretty small in general these days, but home field winning percentage was 53.3% in 2022, 53.9% in 2021, and 52.9% in 2019. So, if anything, I expect the normal difference between home and road start length is greater than I found (although, again, if pitches-per-inning is the main cause of the disparity, the correlation between overall home field advantage and start-length differential might be small, unless pitches-per-inning also decrease winning percentage).
The final bucket of missing games from my study was extra-inning games. Here, I was convinced I would prove to be leaving out starts of longer than average length, because my monograph in progress is about “closeness,” and from immersing myself in this subject I know that average spread increases as total scoring in a game does. No game is closer than a tied one, so it figures these extra-inning games are low-scoring and well-pitched.
Extra-inning games present a small sample, so I used the last 15 years (2009-2023), querying Stathead for the percentage of home starts with the DH being used that went 6 innings or more. My hypothesis was confirmed: the 6-inning mark was reached 58.4% of the time, compared to a 52.6% rate in non extra-inning games that met the same conditions.10
Road starts are a little shorter than home starts, and games destined for extra innings have starts a little longer than games decided in regulation, but the reason I consider these sample biases unimportant is that they aren’t relevant to the questions I was asking. For instance, if road starts were added to the 2008 sample, presumably the early-inning survival rates would run a little lower, but would still be higher than those in 1973 after the same amendment was made to that sample. Studying the influence of the fifth inning and the complete game on pitching changes also does not depend on having a sample distribution reflecting the whole season’s parameters. (It should also be stated that departures from the real average in any year are very small based on the biases intrinsic to the changes to the sample. The differences between the omitted and retained games are small to begin with, and if I expanded the sample, some 45% of it would overlap with the current sample, anyway.)
Conclusion
I am afraid skipping a conclusion is a more material omission than the license I took with road starts. However, having needed ill-advised late-in-the-day coffee to get this far, my personal survival rate does not permit composing one. I will therefore throw my final pitch down the middle, and hope that, you, reader, are able to make a good play behind me to seal the deal.
Nothing changes if you switch the analysis from the home to the road team. The survival rate of the 2nd for road starters was .967. Road starters did fail to complete the 1st inning 23 times, though, while that count was 10 for home starters.
It would be complacent to take the complete game rate as a sufficient representation of overall length, however. After all, there could be a practice where teams always allowed their starter to go 8, then always removed him for a reliever in the 9th. The complete game rate would be 0. The alternate approach of averaging each of the survival rates circumvents this problem, and does also return the inverse order of the years.
Calculating the average inning survival rate is similar to capturing the average start length, but not the same; among other differences, the sample for each survival rate is different. I know of no reason why the average survival rate is an improvement on the more straightforward measure, but thought it was fun to entertain a different approach.
Some important context is that the record of the teams in the 8-inning starts (leaving out the 8.1- and 8.2-inning starts) was 10-67. It makes sense generally that the complete game meant either neither nothing or next-to-nothing in 1912, and it certainly makes sense that it didn’t mean anything as a statistical category. And it’s really hard to fathom that the public valued “complete game losses.” So, I think it was natural for losing pitchers to be removed after 8, if the manager thought that was best. He presumably wasn’t inhibited at all by the complete game sentiment when his team was losing.
Additionally, to have stayed around for 8 innings in 1912, unlike in later years, you didn’t have to have pitched particularly well. It would be natural to look at these pitchers, losing after 8, and think someone else might be able do better in the 9th. That wasn’t so much true in the later years sampled.
Note that the actuarial breakdown combines starts of an inning length with those of a higher partial-inning length, while the full versus partial analysis, primarily of interest for the 5th and 9th, compares full-inning counts to those half an inning below. The purpose of each analysis was different, so the coding is opposite.
Moneyball was published in 2003 but hadn’t jumped on this particular trope.
Because 2023 had a DH, I wanted the most distant year possible that also had one, for the sake of even comparison. Then I also knew that 1973 is known as a year when starting pitchers pitched many innings, so I thought the season would make for an interesting contrast.
The two pieces come together in the “1st-inning full” percentage. There were only 4 starts of 1.0 inning in the 1973 dataset, but 27 of 0.0, 0.1, or 0.2 innings. That’s a conspicuous .129 full-inning score. The second lowest of the 45 across the data was much higher, .286 for first inning, 2008. And that consisted of just 7 data points: 2 1.0 inning starts, and 5 first-inning knockouts.
You are welcome to browse the other tabs as well, keeping in mind that the spreadsheet was meant to help me write the Substack, not to present my work. Not that I would be capable of something neat and shiny even if I tried.
See section “Home Runs: Home, Road” in the following post (or to finger the excerpt even more precisely, search for 2938). Home, Road HR Totals
I would assume the result simply reflects the small sample, but this did not really hold up in 2023: home starters made it 6 innings or more 42.3% of the time. You’ll remember the reference point for 2023 in home games was 41.1%. The sample was 201 extra-inning games. So 6-inning starts in extra-inning games only outdistanced the expected number by 2.4.