The Shape of the Tax: Transaction Shifting at New York City's 2019 Transfer-Tax Notches and the Mortgage Recording Tax

New York's 2019 budget created seven buyer-paid transfer-tax notches on one date, from USD 2 million to USD 25 million, and left the USD 1 million mansion-tax notch unchanged. The paper compares sale prices at the same thresholds before and after the reform, with the USD 1 million threshold as a control. At each new threshold with enough sales to measure, sales moved to just below it. The shift is larger than at any of sixteen never-taxed round prices below USD 6 million, and a year-by-year event study shows it starting with the reform. It appears in houses and condominiums and does not come from foreclosure transfers. Most of it comes from resales; sponsor sales, whose prices developers set in advance, respond clearly only at USD 3 million. At USD 2 million and USD 3 million, the range of prices emptied above the threshold is two to three times as wide as the tax jump. The response does not grow in proportion to the buyer's rate increase: the sampling bootstrap rejects proportionality, and comparisons with placebo noise point the same way more weakly. Equal responses are not established either. Bunching costs USD 2–3 million a year in avoided tax, 2–3% of what the new taxes raise between USD 2 million and USD 6 million. The paper then applies these results to New York City's mortgage recording tax, which charges borrowers 1.80–1.925% of the loan and charges a purchase without a mortgage nothing. On 172,725 observed loans the charge averages 1.41% of price for houses and 1.28% for condominiums. With transactions held fixed, it raises about USD 244 million a year on residential purchase money between USD 300,000 and USD 6 million in four boroughs, and the paper costs continuous schedules that would raise the same static amount. The paper identifies neither economic incidence nor an optimal tax. Version 2.5 corrects the placebo inference. Size-matched placebo draws are now taken with replacement; version 2.4 drew them without replacement, so they varied less than repeated samples would. The main results are unchanged. Placebo ranks and tail shares are now described as reference comparisons, not exact tests. The recording-tax table's condominium row now uses the 47,702 loans of the public linked dataset, with the same statistics as before, so the replication package reproduces it. Replication package version 1.8 reproduces this version. See revision_notes.md for details.