Local Economic Conditions and Repeat-Sale Indices Performance: Evidence from a Moderation Effect Specification

Abstract

The paper relates the documented smoothness of housing repeat-sale indices growth to time-varying changes in local economic conditions. Empirically, a metro economic condition index serves as a moderator variable in a dynamic partial equilibrium model that connects price appreciation to innovations in trading volume. The estimated moderation effect substantially dampens the long-run response to shocks in volume. Overall, the empirical findings are consistent with theoretical models that hypothesize that sellers’ reservation prices change counter-cyclically, thus smoothing repeat-sale index price appreciation.

Keywords:

• Economic Conditions
• repeat-sale indices
• sample selection bias

Notes

1 Seasonality, for instance, does affect sample selection but is relatively easy to deal with.

2 Giannetti (2018a) shows that, contrary to common belief, high persistence is not caused by the three-month moving average window used in sampling transactions to calculate the S&P CoreLogic home price indices.

3 Fisher et al. (2003) distinguish between the two intimately related concepts of sample selection bias and variable-liquidity price effects. The former is essentially a quality bias whereas the latter is a liquidity bias resulting from shifts in the business cycle. Econometrically, both problems are jointly addressed.

4 This is as opposed to the existing home sale statistics released by the National Association of Realtors (NAR) which covers transactions above and beyond those utilized to calculate repeat-sale indices.

5 The S&P Case-Shiller indices are computed using the arithmetic repeat-sale methodology introduced by Shiller (1991).

6 Because the index for economic conditions for the Miami-Fort Lauderdale-West Palm Beach, FL metropolitan statistical area (MSA) is poorly estimated, focus is restricted on the remaining nine cities component of the CS10 as the Miami MSA data is excluded.

7 Additionally, Wallace and Meese (1997) test and reject another crucial assumption of repeat-sale indices: the constancy of implicit prices of housing attributes over time. This assumption essentially precludes secular shifts in the characteristics of transacted properties over time (like, for instance, a secular increase of the average square footage of single-family homes).

8 The hedonic index procedure allows the researchers to construct constant-quality indices.

9 The data is publicly available at https://my.spindices.com/index-family/real-estate/sp-corelogic-case-shiller

10 In technical documentation, Standard & Poor’s (February 2015) explains that the three-month moving average window allows for late reported transactions and “keeps sample sizes large enough to create meaningful price change averages.”

11 See Giannetti (2018b) for further estimation details.

12 In March 2014, the sources for sales transaction data were changed to sources provided by CoreLogic from prior sources. This resulted in exogenous, non-market driven variability of the pair counts.

13 Like for fixed-effects in panel data, centering the monthly indicator variables amounts to imposing the constraint${\sum }_{i=1}^{12}{d}_i=0.$ This allows to prevent the so-called “dummy trap” and to anchor the series unconditional mean as the model intercept$\alpha .$

14 See Gascon and Rapach (2014, 2016) for details about the original series and their reporting frequency.

15 For instance, as noted by Gascon and Rapach (2016), even as early as 6 months before the official national peak, some MSAs like Las Vegas had already dipped into recession.

16 The latter reflects the fact that the dotcom bust affected more significantly Silicon Valley than Southern California.

17 Gascon and Rapach (2016) argue that measurement error in the underlying economic time series may be to blame for those extreme oscillations.

18 See Aiken et al. (1991) for details about estimation and interpretation of interaction models.

19 The number of lags for the HAC correction is conditioned by the prior that price appreciation follows an ARMA (1, 4) process (see Giannetti [2018a] for details).

20 Explicitly, 1, 2, 4, and 6 lags (including lag 5) are considered.

21 More specifically, expression (5) specifies a long dynamic panel (i.e., small N and large T). However, also note that the number of instruments does not grow with the time dimension. Hence, weak instrumentation typical of short dynamic panels à la Arellano and Bond (1991), Arellano and Bover (1995) and Blundell and Bond (1998) is not a priori an issue. A discussant is credited for raising this point.

22 Consistent with the partial adjustment model in expression (3), only one lag of the system’s variables is included. Nonetheless, up to six lags of panel variables are used as instruments. A lag-length test shows an increase in both BIC and AIC when two more lags are introduced. See Dröes and Francke (2018) for similar specifications in their price-turnover study.

23 Forward-mean differencing only takes care of the MSAs fixed-effects${ f}_i.$ Love and Zicchino (2006) handle the time fixed-effects $f_t$ by monthly demeaning each variable.

24 The cross-sectional average standard deviation of condominium count growth is 4.95%.

25 For the sake of thoroughness, the statistical fit of the partial adjustment model is substantially degraded by the inclusion of the Miami index of economic conditions into the estimation sample. Indeed, herein unreported results show overall larger Hansen’s J statistics as well as moderation effect point estimates cut in half.

26 The cross-sectional average standard deviation of single-family home count growth is 4.14%.

27 Alternative Choleski orderings elicit mostly similar IRFs. The eigenvalue condition for dynamic stability of the PVAR systems is also satisfied.