Do sales prices overstate underlying house prices in market downturns? Evidence from the Canadian house price crash of 1991

Canadian Journal of Administrative Sciences, Dec 2002 by Marion Steele, Richard Goy

Abstract

Over the last decade there has been a mounting realization that the quality-adjusted price of properties that sell may be quite different from the quality-adjusted price of all properties (Gatzlaff & Haurin, 1998). The difference is apt to be especially marked during downturns. Dependence on price indexes based on transactions data could result in overoptimistic appraisals by mortgage lenders during the early quarters of downturns. This paper provides evidence on this issue by examining the residential real estate crash at the start of the 1990s in Canadian cities, especially Toronto. The plunge in prices estimated here for Toronto is 17%. We estimate the drop in the transactions price using MLS data, which are not quality adjusted, supplemented by Royal LePage data, which are. We estimate the drop in the price of houses in the stock using hedonic indexes based on home owners' valuations; outlier observations are cut from the sample on the basis of the DFFITS criterion. The hypothesis of equality of the drop shown by the two measures is strongly rejected, for Toronto, and is also rejected for several other cities.

In the wake of the great price booms of the 1980s and busts of the 1990s, there has been an increasing realization that the quality-adjusted price of properties which sell may be quite different from the quality-- adjusted price of all properties (Haurin & Hendershott, 1991). The difference is apt to be especially marked during downturns. Gatzlaff and Haurin (1998) provide evidence that hedonic-based price indexes using transac- tions data tend to smooth price changes in the stock; for example, transactions prices understate price reductions.

This smoothing is potentially important because it could result in over-optimistic appraisals by mortgage lenders during the early quarters of downturns, too-slow response by banks and other lenders to market change, and understatement of the risk of housing assets and that inherent in the wealth portfolio of the average household. Hendershott and Kane (1995), in fact, suggest that, early in downturns, transactions prices may provide upward-biased estimates of returns on a constant-- quality portfolio of commercial properties. Goetzmann's (1996) analysis of the effects of various rules, used by vendors for setting their reserve price, shows that plausible rules can result in changes in transactions prices greatly outpacing changes in underlying values. Further, the smoothing in transactions prices relative to prices in the stock adds to the smoothing found for appraisal prices relative to transaction prices (Clayton & Hamilton, 1999), so that it reinforces concern about appraisal values.

This paper provides evidence on this issue by examining an important event, the residential real estate crash of 1991, in Toronto and elsewhere. The drop in prices of far more than 5% in Toronto far outweighs the largest drop estimated for Miami by Gatzlaff and Haurin (1998).1 Such a large change in prices underlines the potential importance of differences between transactions and stock-based house price estimates. As data for transactions prices, we use the average prices for houses sold under the Multiple Listing Service (MLS), and for prices of the housing stock, a quality-adjusted index based on homeowners' estimates of the value of their homes. Our investigation indicates that in 1991 the prices of the stock fell sharply in Toronto and in some other Canadian Census Metropolitan Areas (CMAs) while MLS averages fell only slightly and in almost all CMAs there was some drop in prices of the stock, despite the fact that MLS averages fell in only a minority. This is striking evidence of the difference between the change in transactions price and the change in the price of the stock.

This result is subject to the caveat that Statistics Canada's Household Income, Facilities, and Equipment (HIFE) datafile-the data set used to determine qualityadjusted prices of the stock-and MLS averages-used to represent transactions prices-are not ideal for our purposes. Rental housing is absent in the HIFE data and very low-valued housing is under-represented in the edited HIFE sample we use. If rental and very low-valued housing are more affected by economic downturns than the rest of the stock, the under-representation of these housing types in HIFE would tend to yield a downturn in the HIFE value that is less than the downturn in the MLS value. This is the opposite of what we find. Further, we find independent evidence supporting the price change picture given by these data. First, calculations in Tables 4 and 5, based on Royal LePage data, which give prices of a well-specified house, support the HIFE- based estimates of price changes in the stock. Second, in the sixth section of this paper, we give evidence for two cities where some quality-adjusted transaction prices are available, and these support the use of MLS averages.

Because of compelling evidence that homeowners overestimate the value of their homes (DiPasquale & Somerville, 1995; Goodman & Inner, 1992), we use stock values only in index form. We are concerned with the change in values, not in the level. But concern with the effects of homeowners' errors is not totally allayed by this, because Goodman and Ittner find that the error variance is high and, in a nontrivial number of cases, homeowners' errors are extremely large. We deal with such outliers systematically. In the housing and real estate literature outliers are generally ignored or dealt with arbitrarily. Goodman and Ittner and DiPasquale and Somerville are exceptions in their careful analysis of the outlier problem, but their treatment is specific rather than general, although quite general, systematic procedures do exist. In this paper we use one of these. We identify outliers using a 5% DFFITS criterion (Belsley, Kuh, & Welsch, 1980). Davidson and MacKinnon (1993) argue that a measure closely related to this should be routinely used to uncover influential observations, especially those due to data errors.