Remove imputed earnings from training set when modeling prior-year earnings

[MaxGhenis]

This paper shows that including earnings question nonrespondents (whose incomes are imputed) roughly doubles the year-to-year earnings volatility.

IPUMS assigns this in the QOINCWAGE variable, which I think we'd filter to 0: No allocation. IPUMS doesn't state what that maps to in the raw ASEC microdata, but I think it's I_ERNVAL and I_SEVAL for employment and self-employment income, respectively.

Thanks to @ricco-budget-lab for this tip, and @JackLandry for sharing this paper with him to pass along.