UPDATE: Further Evaluation of the Gap Between ERA and FIP

I am following up on my article from earlier this week based upon a discussion that occurred with a reader at Fangraphs.

The reader suggested that I analyze infield fly ball rate to see if it can explain some of the uncorrelated gap between ERA and FIP. I was skeptical at first. Here was my thinking: very few pitchers, if any all, have a “skill” of inducing infield fly balls or line drives. Some pitchers can induce groundballs, but in general batted balls would fall into the “luck” component of the gap between ERA and FIP.

But that’s not the point. I am trying to explain the gap between ERA and FIP, not predict it. Even if batted balls were part of the “luck” component, then it may be possible to measure that part of luck. I thought it would be worth checking out. So I did.

Infield Fly Ball Rate does have an impact on the gap between ERA and FIP, but it is minimal. The adjusted r-squared values for ERA-FIP modeled against each batted ball type are as follows:

  • Infield Fly Balls: .034
  • Line Drives: .014
  • Fly Balls: -.002
  • Ground Balls: -.002

There appears to have very little correlation between batted ball type and ERA-FIP. And here’s the thing, when any one of the types of batted balls are added to the original model (ERA-FIP ~ Defensive Metric + RE24), the adjusted r-squared of the model is improved.

  • DRS + RE24 + IFFB: .5484
  • DRS + RE24 + FB: .5566
  • DRS + RE24 + GB: .5493
  • DRS + RE24 + LD: .5457

If you add all the batted ball types to the model (DRS + RE24 + IFFB + FB + GB + LD) you get a .56 adjusted r-squared value, but the p-values of the line drive, fly ball, and ground ball variables are greater than .17, indicating that they are improving the model because there are more variables, but not because they add any significance to the model. Infield fly balls have a p-value of .052, indicating they likely have some significant contribution to the model with all batted balls included.

In the end, batted balls add little to the model, which when originally run as DRS + RE24 produced an adjusted r-squared value of .54. The miniscule impact can be seen in the graphs below displaying the different relationships.

That’s a little disappointing, because I was coming around to the idea. However, I leave off where I last ended.