Today my column on Baseball Prospectus (subscription required) is a sort of part II building on last week's offering (no subscription required) and delves a bit more deeply into the double steal and discusses both the limits of play by play event codes like those found in Retrosheet and double steal attempts from a Win Expectancy (WX) standpoint. The thumbnail version of the discussion on the underlying strategy (in the article I also discuss individual manager's use of the double steal and the runners who were most prolific in this regard) can be summarized in the following table that includes data from 1970 through 2006 and excluding 1999. The table shows the base situation, number of outs, successes, double steal attempts, the percentage of attempts that were successful (defined as both runners being safe but not necessarily awarded stolen bases) and the average break even percentage for such attempts. The break even rates were calculated using a model where the most likely costly failure was used in the calculation which involves the lead runner being thrown out and other runners advancing.
Base Outs Succ Att Percent Avg BE
12x 0 643 1127 57.1 0.587
12x 1 1595 2258 70.6 0.667
1x3 0 21 70 30.0 0.728
1x3 1 147 478 30.8 0.590
x23 0 2 4 50.0 0.717
x23 1 2 51 3.9 0.633
123 0 0 5 0.0 0.544
123 1 5 39 12.8 0.524
Total 2415 4032 59.9 0.635
Attempts with two outs are not included for reasons discussed in the article related to the incompleteness of the scoring system. Considering only attempts with runners on first and second it appears that managers generally do pretty well in their decision making process and are successful 66% of the time with a break even percentage of 64%.
However, with runners on first and third the overall success rate is under 31% while the break even percentage is over twice that indicating that delayed double steals are very risky and typically not a good idea. However, it should be remembered that the definition of a successful delayed double steal also typically includes the scenario where the runner on second is put out with the runner on third scoring. When those are included the success rate rises to 41.2%. To determine whether the strategy is then a good idea we would also need to recalibrate the break even model to where successes of these types were also included. Doing so, however, would actually raise the break even rate. For example, with runners at the corners with nobody out in a tie game in the top of 5th inning, the visiting team has a 58% chance of winning in a run environment of 5.0 runs per game. If the double steal fails with the runner on third getting thrown out at the plate with the runner on first advancing, that probability shrinks to 48%. If the attempt is successful under the model used in the table above (both runners advancing) the probability shoots up to 65% but if the play is successful with the runner on third scoring and the runner on second being thrown out the probability raises to just 59%. In the former case the break even percentage would be 60% while in the latter it is 93%. This is so since the loss of the trail runner and the additional out will always lower the offensive team's probability of winning. Keep in mind that your mileage will vary depending on the inning, score, and run environment but the general rule will hold for all such scenarios.
Of course some percentage of these attempts are also not able to be teased out of the play by play data since the runner on third may retreat to the bag while the runner on second is charged with a caught stealing. There is no way to differentiate between that scenario and one where the runner on third is simply a spectator in normal stolen base attempt by the runner on first. In order to do so the codes would have to reflect which runners were breaking with the pitch. That is one of the limitations we have to live with at present. My guess is that if successes and failures in that scenario were able to be found, they would generally cancel each other out - but that's just a guess.
The situation with runners on second and third and the bases loaded are more interesting since unsuccessful attempts are characterized by one or more runners being caught stealing while other runners advance. With runners on second and third that means that either the runner on second was thrown out at third and the runner on third scored or the runner was caught at the plate and the runner on second advanced. But here successful attempts would not include delayed steals and so the overall success rate of just 9% and the extreme rarity of the play is evidence that it is a poor percentage play.