We had to throw out 15 years of data.”

Why? Completely ignoring that PEDs existed well before 1990, I get that PEDs can cause your models to be off, but don't you just put the new data into your system and adjust the range of possible outcomes? Especially if you don't know who's using or not, or what they're even using.

I assume he's being offhand about throwing away 15 years of data and does mean something closer to "our models didn't work anymore." I'm not sure I really buy that either but nevermind.

It would be an interesting analysis to do. Was there a discontinuity in the distribution of outcomes? Were massive jumps (relative to context) more common than before? There are always going to be extreme values but was there a cluster of them and was the cluster large enough to really matter? How does it compare to other shift points between extreme eras? All needing to adjust for the fact that the mean and the standard deviation of the distribution were changing as well.

I think the idea is that with PEDs not being tested for that data cannot be reliable used to predict current day when it is tested for.
Probably pre 1990 data isn't that reliable now either (don't have good non-baseball stats, age/heights etc.)

It might not matter, but you can't tell because of unknown who/what/whens.

I think the idea is that with PEDs not being tested for that data cannot be reliable used to predict current day when it is tested for.

But this is unlikely to be true. There was a level shift in offense ... not unlike other big offense eras. So the raw numbers aren't useful but they never have been and it's not hard to deal with that level shift. For this data to be un-useful, you have to show that the trends, the correlations, etc. all shifted dramatically from the past.

As an over-simplified example, based on pre-1994 data you might project a given age curve for a 1B with a 115 OPS+ at age 32 with that age curve showing he was likely to be useless by the time he hit 35-36. As long as such players in the sillyball era were following a similar age curve, there's little issue with using that data.

If the sillyball era was essentially a different population you were sampling from then maybe the easiest solution would be to discard the data with the other alternative being finding and controlling for variables that adjust for the differences. But, prima facie, I see no reason to think that era would be any more disruptive than other extreme eras.