Graham (2019): House Prices, Investors, and Credit in the Great Housing Bust
Graham studies how household and corporate investors stabilized house prices during the housing bust. The idea is simple: investors may moderate house price declines following a negative shock to housing demand because they may enter the market as prices fall (looking to buy “on the cheap”). The strength of this stabilization channel depends crucially on the elasticity of housing investor demand to house prices. Graham argues (and finds supporting evidence) that corporate investors are more price elastic than household investors. His main empirical result shows that regions with a larger corporate investor presence during the bust experienced smaller house price declines. He builds a KMV-style model to show that when corporate investors are the “marginal” house buyer following a shock that exogenously tightens credit conditions, house prices are more stable, compared to when household investors are the marginal house buyer.
Empirical Analysis
Main specification is the ZIP-level fixed effects regression estimated for the bust period 2007–10:
\[ \begin{align} \Delta \log P_{z,t} = \alpha_{c,t} &+ X_{z,t}'\Gamma \\ &+ \beta \Delta \log M_{z,t} \\ &+ \delta_1(\Delta \log M_{z,t} \times \Delta CorpInvShare_{z,t})\\ &+ \delta_2(\Delta \log M_{z,t} \times \Delta HHInvShare_{z,t}) + \epsilon_{z,t} \end{align} \]
Where \(\Delta \log P_{z,t}\) is annual growth rate of real ZIP-level house prices. \(\Delta \log M_{z,t}\) is the annual growth rate of mortgage originations (number). Graham wants \(\delta_1<\delta_2<<0\), which would indicate that corporate investors dampen house price declines more than HH investors following an (exogenous) decline in mortgage origiantions. Of course, mortgage originations and investor shares are endogenous to house prices (reverse causality is probably the main problem since higher price growth would bring in buyers). He instruments for \(\Delta \log M_{z,t}\) using the ZIP-level “exposure” to non-GSE lending activity over 1998–2000. Specifically, for each ZIP he computes the number of mortgages sold on the secondary market in a given year, then computes the share of those mortgages purchased by non-GSEs (using the HMDA variable Type of Purchaser). The instrument is trying to capture regional exposure to the securitization boom in the vein of Mian-Sufi. (Graham mentions [fn3] that a better instrument would be to use the share of mortgages sold into private label securitization. But these sales are underreported in HMDA, which is why he uses the wider cateogry of non-GSEs.)
Results
His first-stage results (Appendix C) are very strong: ZIPs with greater exposure to non-GSE loan purchases experienced larger declines in mortgage originations during the bust. This isn’t very suprising, since these regions were the “hotter” markets during the boom.
His 2SLS results are the stuff of dreams: \(\delta_1<\delta_2<<0\). Thus, both types of investor activity moderate house price declines following a negative mortgage supply shock. His estimates imply that a simultaneous one S.D decline in mortgage originations and a one S.D increase in the corporate investor share is associated with a 5% decline in house prices. Whereas a one S.D decline in mortgage originations combined with a one S.D increase in the household investor share is associated with a 7% decline in prices. Thus, he concludes that for a similar sized shock to mortgage originations, regions with a higher corp. investor share experience a \((1-5/7)*100=29\%\) smaller decline in house prices, compared to regions with a higher HH investor share. He uses this statistic to inform his model counterfactuals.
Issues
- The obvious problem, like with all Mian-Sufi work, is whether the non-GSE share over 1998–2000 is really exogenous to housing market conditions even 10 years later. Non-GSEs probably targetted “hot” markets in the early days of the boom, and these are likely to be precisely the regions that experienced the biggest busts. Graham and Mian-Sufi would like you to think that non-GSE/securitization activity is what caused these regions to be hot, but I doubt that’s true. Anyway, even if his results aren’t causal, they’re still cool!
- Similarly, find it weird that he instruments for the interaction terms by simply interacting his ZIP-level non-GSE share with the lagged investor share (e.g. he instruments for \(\Delta \log M_{z,t} \times \Delta CorpInvShare_{z,t}\) using \(\lambda_{98-00}^{nonGSE} \times \Delta CorpInvShare_{z,t-1}\)). This seems like a little bit of a cop out, because surely lagged investor shares are still endogenous to current price growth.
Model
Essentially KMV with investors, and simplified along some other dimensions:
- Households can rent, buy owner-occupier housing or investor housing
- Investors earn rental income on their investment properties
- Renters cannot invest (i.e. have to be an owner-occupier first). This seems reasonable.
- To “economize on state variables” one mortgage is secured against the value of owner-occupier and investment properties (i.e. no seperate mortgages for owner-occuper and investment properties). This is obviously a pretty big/unrealisitic simplification! But again, it keeps things simple.
There is a representative corporate investor, which takes the form of an unconstrained risk neutral firm that trades property each period, earns rent on its housing stock and faces a convex cost associated with the number of houses held. The firm maximizes the present discounted value of profits, just like the rental firm in KMV. The difference from KMV is that Graham assumes a functional form for the convex holding cost that allows him to play around with the firm’s elasticity of investment demand (to house prices), \(\epsilon\). As emphasised by Greenwald-Guren, the rental firm in KMV has perfectly elastic demand, \(\epsilon=\infty\) – i.e. it is willing to buy/sell an unlimited amount of housing at the equilibrium price equal to the present value of rents.
Differences from KMV
Besides adding investors into the fray, there are some other small differences from KMV (comfortingly, most of these are simplifications):
- Rental services are a continuous choice every period, with an upper bound. He doesn’t actually say how he calibrates the upper bound.
- In his calibration, there is only one house size. This seems like a reasonable simplification relative to KMV.
- It doesn’t look like he has a “warm glow” of ownership. Not sure why not, or if this is important.
- The lender side is completely un-modelled. Graham assumes an exogenous risk-free rate and mortgage rate. This is comfortingly simple. But I think it would be harder to get away with this if modelling the boom, since other papers have emphasized an endogenous relaxation of credit constraints is probably required to match the evidence over the boom years (i.e. to make sure low-income households aren’t priced-out of the housing market).
Model experiment
The simulation exercise is similar to KMV – he starts the economy in steady state, and then exogenously tightens several parameters that govern credit conditions for a few periods. Then he looks at how prices and quantities evolve following the shock.
Mapping the model to empirical results
In the model, \(\epsilon\) fully governs the degree to which the corporate investor firm stabilizes prices following a negative shock. In his baseline calibration, Graham sets \(\epsilon=0\). This calibration corresponds to perfectly inelastic corporate investor demand (i.e. a vertical demand curve), and represents an economy where corporate investors play no stabilizing role (since the firm simply demands a fixed amount of housing each period, no matter what happens with prices). He compares this baseline to an economy with an \(\epsilon\) such that house prices decline by 30% less on impact in response to the tightening of credit conditions. The 30% figure corresponds to his empirical estimate of how much less house prices decline in response to a one S.D decline in mortgage originations in regions with higher coporate investor shares, relative to regions with higher HH investor shares (see above). He ends up going with \(\epsilon=24\) in this case.
Note that Graham doesn’t actually estimate \(\epsilon\) in the data. Rather, through the lens of his model, the empirical results are “informative” of the magnitude of \(\epsilon\). I think this is very cool.
Results
Main result from the model experiment is that in response to an exogenous tightening of credit conditions, house prices decline by much more when household investors are the marginal buyers of houses compared to when corporate investors are the marginal buyers. In other words, household investors are much less price elastic than corporates, consistent with the empirical evidence.
Other comments and thoughts
We often focus on the amplification role of investors, rather than their stabilizing role, so this paper is very cool in that it draws attention to a potentially overlooked aspect of housing markets and shows that it is quantitatively important. But, I guess like with all quantitative macro models, there is some strange stuff going on.
- He calibrates his steady state to the peak of the boom, which seems odd because the economy and housing market were so out-of-whack then. But he’s interested in stabilization through the bust, so maybe this makes sense.
- As he mentions at the very end of the paper, household investors seem way to price elastic in his model – i.e. they seem way to keen to enter the market when prices fall, to moderate the bust. For example, Table 6 (\(\epsilon = 0\)) shows that the HH investor share of home purchases exceeds its steady state value by 300% 5 years after the shock. This seems crazy, given that his steady state is calibrated to the peak of the boom, when the investor share of purchases was at its highest too!
- I think that the fact that HH investors are too price elastic in his model actually points to missing model elements. For instance, he mentions the availability of other risky assets (e.g. equity) would lower the elasticity of HH investor demand. I think that extrapolative house price expectations are far more important.
