Revisions in repeat-sales price indexes: Here today, gone tomorrow?

Price indexes based on the repeat-sales model are revised all the way to th e beginning of the sample every time a new quarter of information becomes a vailable. Revisions can adversely affect practitioners. In this paper we ex amine this revision process both theoretically and empirically. The theory behind the repeat-sales method says that revisions should lower the standar d error of the estimated indexes; we prove that, in fact, the revised index is more efficient than the original one. This implies that large samples s hould make revisions trivial. However, our data, and the Freddie-Fannie dat a, suggest that revisions are large, insensitive to sample size and systema tic; revisions are more likely to be downward than upward. In Los Angeles a nd Fairfax, revisions are usually downward and statistically significant. T his bias in initial repeat-sales estimates is caused by sample selectivity; properties with only one or two years between sales do not appreciate at t he same rate as other properties. We hypothesize that these "flips" are imp roved (possibly cosmetically) between sales. One implication of our analysi s is that flips should be removed or downweighted before calculating repeal -sales indexes. The same model estimated without flips appears free of bias . We find small increases in efficiency from adding up to 4,300 observation s to a base of 1,200.