NBER WORKING PAPER SERIES

HOUSING DYNAMICS

Edward L. Glaeser
Joseph Gyourko

Working Paper 12787

http://www.nber.org/papers/w12787

NATIONAL BUREAU OF ECONOMIC RESEARCH

1050 Massachusetts Avenue
Cambridge, MA 02138
December 2006

Glaeser thanks the Taubman Center for State and Local Government at Harvard University, and Gyourko
thanks the Research Sponsors Program of the Zell/Lurie Real Estate Center at The Wharton School,
University of Pennsylvania for financial support. We appreciate the comments of seminar participants
at the University of Chicago, the NBER Summer Institute, and the Federal Home Loan Mortgage Corporation
on previous versions of the paper. Graham Elliot and James Stock provided helpful guidance. Andy
Moore, Charles Nathanson, and Jon Steinnsen provided superb research assistance. The views expressed
herein are those of the author(s) and do not necessarily reflect the views of the National Bureau of
Economic Research.

© 2006 by Edward L. Glaeser and Joseph Gyourko. All rights reserved. Short sections of text, not
to exceed two paragraphs, may be quoted without explicit permission provided that full credit, including
© notice, is given to the source.
Housing Dynamics
Edward L. Glaeser and Joseph Gyourko
NBER Working Paper No. 12787
December 2006
JEL No. A1

ABSTRACT

The key stylized facts of the housing market are positive serial correlation of price changes at one
year frequencies and mean reversion over longer periods, strong persistence in construction, and highly
volatile prices and construction levels within markets. We calibrate a dynamic model of housing in
the spatial equilibrium tradition of Rosen and Roback to see whether such a model can generate these
facts. With reasonable parameter values, this model readily explains the mean reversion of prices
over five year periods, but cannot explain the observed positive serial correlation at higher frequencies.
The model predicts the positive serial correlation of new construction that we see in the data and the
volatility of both prices and quantities in the typical market, but not the volatility of the nation's more
extreme markets. The strong serial correlation in annual house price changes and the high volatility
of prices in coastal markets are the two biggest housing market puzzles. More research is needed to
determine whether measurement error-related data smoothing or market inefficiency can best account
for the persistence of high frequency price changes. The best rational explanations of the volatility
in high cost markets are shocks to interest rates and unobserved income shocks.

Edward L. Glaeser
Department of Economics
315A Littauer Center
Harvard University
Cambridge, MA 02138
and NBER
eglaeser@harvard.edu

Joseph Gyourko
University of Pennsylvania
Wharton School of Business
3620 Locust Walk
1480 Steinberg-Dietrich Hall
Philadelphia, PA 19104-6302
and NBER
gyourko@wharton.upenn.edu
 I. Introduction

Housing constitutes nearly two-thirds of the typical household’s portfolio, and

more than $18 trillion worth of real estate is owned within the household sector.1

Despite the enormous size of this sector, economists’ understanding of many features of

the housing market remains incomplete.2 For example, in the sample of 115

metropolitan areas from 1980 to 2005 for which we have Office of Federal Housing

Enterprise Oversight (OFHEO) constant quality house price series, a $1 increase in real

house prices in one year is associated with a 71 cent increase the next year. A $1

increase in local market prices over the past five years is associated with a 32 cent

decrease over the next five year period. This predictability of price changes seems to

pose a challenge for an efficient markets view (Case and Shiller, 1989; Cutler, Poterba,

and Summers, 1991).

The large amount of inter-temporal volatility in prices within markets is also

puzzling. The standard deviation of three-year real changes in our sample of

metropolitan area average house prices is $26,354 (in 2000 dollars throughout the paper),

which is about one-fifth of the median price level. Over one, three, and five year periods,

the standard deviation of house price changes is at least three times the mean price

change. Can this volatility be the result of real shocks to housing market or must it

reflect bubbles and animal spirits?

1
The portfolio share is from Tracy, Schneider, and Chan (1999). The dollar value figure is for the fourth
quarter of 2005 and is from Table B.100 Balance Sheet of Households and Nonprofit Organizations which
may be downloaded at http://www.federalreserve.gov/RELEASES/z1/Current/data.htm. The Federal
Reserve’s data includes market value estimates for second homes, vacant homes for sale, and vacant land
owned by the household sector.
2
The debate over whether the recent boom was a bubble is only the latest example. See McCarthy and
Peach (2004), Himmelberg, Mayer and Sinai (2005), and Smith and Smith (2006) for recent analyses that
conclude there is no large-scale bubble in housing prices. Shiller (2005, 2006) and Baker (2006) argue to
the contrary that the bubble is both real and very large.

1
 Another more subtle puzzle is that house price appreciation in the 1990s was

negatively correlated with that in the 1980s (as shown in Figure 1), while housing unit

growth was positively serially correlated over the same time periods (see Figure 2).

Basic demand-driven housing models predict that prices and quantities should move

symmetrically. The mismatch of quantity and price movements seems to suggest that

models of housing prices need to more firmly embed supply as well as demand.

Many housing models also put great stock in macroeconomic variables such as

interest rates and national income, but most variation in housing price changes is local,

not national. Less than eight percent of the variation in price levels and barely more than

one-quarter of the variation in price changes across cities can be accounted for by

national year-specific fixed effects. The large amount of local variation and its

relationship with macroeconomic variables is another challenge for a consistent economic

explanation of housing market dynamics.

In this paper, we present a dynamic, rational expectations model of house price

formation to see whether such a framework can explain the salient moments of housing

price and quantity changes. The model follows the urban tradition of Alonso (1962),

Rosen (1979) and Roback (1982) in which housing prices reflect the willingness to pay

for one location versus another. In this approach, housing prices are determined

endogenously by local wages and amenities, so that local heterogeneity is natural. Our

model then extends the Alonso-Rosen-Roback framework by focusing on high frequency

price dynamics and by incorporating endogenous housing supply.

In Section II of this paper, we present the model and four propositions regarding

its implications. The model shows that the predictability of housing price changes is

2
compatible with a no-arbitrage rational expectations equilibrium. Slow construction

responses and mean reverting wage shocks imply that prices will mean revert. And,

positive serial correlation of labor demand shocks at high frequencies can generate

positive serial correlation of housing prices.

The model can also explain the apparent puzzle of mean reverting prices and

persistent quantity changes shown in Figures 1 and 2. Proposition 4 shows that long-term

trends to city productivity or local amenities will create persistence in population and

housing supply changes, but will have a much smaller impact on prices, since those

trends are anticipated and incorporated into initial prices. Price changes are driven by

unexpected high frequency shocks, which mean revert, while quantity changes are driven

by anticipated low frequency trends that persist.

The model also serves as the basis for the calibrations discussed in Sections III

and IV of the paper. Section III presents our estimates of the model’s key parameters: the

real rate of interest, the degree to which construction responds to higher prices, and the

variance and serial correlation of local demand shocks. We assume constant interest

rates for most of the paper, but do turn to time-varying interest rates in Section V. We

estimate construction cost parameters using data on construction and price variance. The

literature on housing demand provides our estimates of the heterogeneity in preferences

for particular locales. And, we use Bureau of Economic Analysis (BEA) income data to

infer the time series properties of local income shocks.

In Section IV, we compare the moments of the real data with the moments

predicted by the model based on the parameter estimates from Section III. We first

investigate the serial correlation properties of prices and quantities. The parameter values

3
described in Section III predict that housing prices will mean revert over five year periods

at almost exactly the same rate that we see in the data. This mean reversion is the result

of new construction satisfying demand and the observed mean reversion of economic

shocks to local productivity. We fit the modest mean reversion of construction quantities

less perfectly, but the patterns in the real data are quite compatible with reasonable

parameter values.

Over one year periods, we predict strong serial correlation of new construction,

but in the data serial correlation of new permits is even greater than the level that our

model predicts. The model does not predict the strong serial correlation of price changes

at one and three year intervals. This serial correlation could be due to the artificial

smoothing of the underlying data or less rational factors. Persistence itself is not enough

to reject a rational expectations model, but the mismatch between data and model at

annual frequencies indicates that Case and Shiller’s (1989) conclusion regarding

inefficiency could be right. Future work needs to deal with the data smoothing problem

to see whether the actual serial correlation still is far too high relative to the model.

Reasonable parameter values predict variances of new construction and price

changes that are quite close to the variances seen in the median metropolitan area in our

sample. We do overestimate the volatility of price changes at annual frequencies, but that

could be the result of data smoothing. The model does not predict ‘too much’ variation

for three and five year changes, where smoothing should be less of an issue.

While the model can fit the median market, it cannot explain the volatility of

either prices or construction in the nation’s more extreme markets. The model does not

fit the price volatility in California which has huge price changes and it does not fit the

4
construction volatility in sunbelt cities such as Atlanta and Houston. The high

construction volatility in the sunbelt areas is most plausibly the result of lumpiness in the

construction process and the possibility that the sunbelt areas have much lower

construction cost parameters than we estimate. However, these explanations cannot help

us explain the high price volatility areas, and we try to understand them in the

penultimate section of the paper.

We consider three added sources of volatility: amenity fluctuation, unmeasured

income volatility, and volatile interest rates. The one high frequency amenity variable

that we have—crime rates—shows little ability to increase predicted demand and price

variability. Using data from New York City, we examine whether the volatility of

incomes for recent home buyers is higher than the volatility for average income, and find

that it is. The variance of income in areas with big price change areas is higher than the

variance of incomes for the average market. These factors may explain the high variation

of prices in the most volatile markets on the east coast, but do little to help us understand

California outside of the bay region, which has less income volatility.

Volatile interest rates will not increase the volatility of prices or construction in

markets with prices close to construction costs (or to the national median price in our

model), but they can increase the predicted variance for places with permanently high

amenities or productivity. For interest rates to generate high levels of volatility, shocks to

interest rates must be extremely high and areas must be innately extremely attractive, but

these conditions may be true for California over the last two decades.

5
 II. A Dynamic Model of Housing Prices

Our dynamic model of housing prices is based on three equilibrium conditions.

Following Rosen (1979) and Roback (1982) we require consumers to be indifferent

across space at all points in time, which requires utility U(W, A, R) to be equal across

space, where W refers to wages, A to amenities, and R to the flow cost of housing. Our

simplifying assumption that this spatial equilibrium must hold in all periods is the

housing equivalent of assuming no financial transaction costs (as in Hansen and

Jagannathan, 1991). Our second equilibrium condition is in the housing markets: we

require the expected returns from making a house (its expected price) to be equal to the

cost of construction. If the city is not growing, this equilibrium condition need not hold

(as in Glaeser and Gyourko, 2005), but we make the simplifying assumption that the city

is always adding new units. Our final equilibrium condition concerns wages which must

equal the marginal product of labor to firms in the city.

We implement the spatial equilibrium condition by assuming that there is a

“reservation locale” that delivers utility of U (t ) in each period “t” and that the cost of

building a home there always equals “C,” which reflects the physical costs of

construction. Since housing can be built in the reservation locale freely at cost C, we

assume that the price of a house there always equals C.3 The reservation locale

represents the many metropolitan areas in the American hinterland with steady growth

and where prices stay close to the physical costs of construction (Glaeser, Gyourko and

Saks, 2005).4 The annual cost of living in the reservation locale equals the difference

3
While it is possible that prices will deviate around this value because of temporary over- or under-
building, we simplify and assume that the price of a house always equals C.
4
Van Neiuwerburgh and Weill (2006) present a similar model in their exploration of long run changes in
the distribution of income (also studied by Gyourko, Mayer and Sinai, 2006). Our paper was produced

6
between the price of the house at time t and the discounted value of the house at time t+1,

or C-C/(1+r) = rC/(1+r), where r is the assumed fixed rate of interest.5 We abstract from

taxes, maintenance costs and allow time-varying interest rates only in Section 5.6

The spatial equilibrium requires all cities at all times to deliver to the marginal

resident the same utility that always is available in the reservation locale. We focus on

the dynamics in a single representative city (which is different from the reservation city).

The utility flow for person i living in that city during period t is W (i, t ) + A(i, t ) , or wages

plus amenities. We assume that there are a fixed number of firms each of which has

output that is quadratic in labor. This assumption ensures that the marginal product at

each firm is linearly decreasing with the number of workers and that wages in the city are

linearly decreasing with the number of workers. These labor demand schedules

generated by firm optimization underpin our assumption that wages at the city level

include an exogenous component and is linearly decreasing in total city population.

We assume that the time-specific and individual-specific effects that make up the

net utility flow from the city are separable so W (i, t ) + A(i, t ) − U (t ) can be written as

D(t ) + θ (i ) . The composite variable D(t) reflects wages and amenities, which in turn

reflect exogenous shocks and city size. We let N(t) denote the housing stock in the city

and assume that the city’s population and labor force equal a constant times the amount

independently of theirs, and our focus on high frequency variation in prices and quantities is quite different
from their focus on changes in the long run distribution of housing prices. More generally, the approach
taken here differs from most research into housing prices, which employs the user cost approach introduced
by Hendershott and Slemrod (1983) and Poterba (1984). That branch of the literature is too voluminous to
describe in detail. The first three papers referenced in footnote 2 employ a user cost framework to examine
the recent housing boom.
5
This difference would also be the rent that a landlord earning zero profits would charge a tenant.
6
If maintenance costs are independent of housing values and constant over space, they will not change the
analysis. If maintenance costs scale with housing and if there are property taxes, then the cost of owning a
house would be higher than the after-tax interest rate. For this reason, we will assume a relatively high real
rate in our simulations. See below for more on that.

7
of housing.7 We further assume that D(t) moves linearly with city population to allow for

the fact that wages and amenities may fall due to congestion or rise because of

agglomeration economies as city size increases. We assume that θ (i ) is a uniformly

distributed taste for living in this particular locale, so that the value of θ (i ) for the

marginal resident at time t (denoted θ (i * (t )) ) is also linearly decreasing in locale size.

The exogenous components of city amenities and wages include a city-specific

component (denoted D ), a city-specific time trend (denoted qt) and a mean zero

stochastic component (denoted x(t)). Thus, the flow of utility for the city’s marginal

resident at time t with index i*(t) relative to the reservation locale, D(t ) + θ (i * (t )) , can

be written D + qt + x(t ) − αN (t ) , where α captures the assumption that wages, amenities

and the taste of the marginal resident for living in the locale can fall linearly with city

size. We further assume that x(t) follows an auto regressive moving average (ARMA) (1,

1) process so that x(t ) = δx(t − 1) + ε (t ) + θε (t − 1) , where 0 < δ < 1 , and the ε (t ) shocks

are independently and identically distributed with mean zero.

The expected cost of housing in the representative locale equals H (t ) minus

Et ( H (t + 1)) /(1 + r ) , where Et (.) denotes the time t expectation operator. The difference

between the cost of housing in the representative city and housing costs in the reservation

locale, rC/(1+r), should be understood as the cost of receiving the extra utility flow

associated with locating in the city. If extra housing costs in the city equals extra utility

delivered by the city then:

E t ( H (t + 1)) rC
(1) H (t ) − − = D + qt + x(t ) − αN (t ) .
1+ r 1+ r
7
Glaeser, Gyourko and Saks (2006) provide evidence showing that population is essentially proportional to
the size of the housing stock.

8
Equation (1) represents a dynamic version of the Rosen-Roback spatial indifference

equation where differences in housing costs equal differences in wages plus differences

in amenities. We assume a transversality condition on housing prices such that

H (t + j )
lim j →∞ = 0 .8 If housing supply was fixed, so N(t)=N (as might be the case in
(1 + r ) j

a declining city as analyzed in Glaeser and Gyourko, 2005) then:

(1 + r )( D − αN + qt ) (1 + r )q (1 + r ) x(t ) + θε (t )
(2) H (t ) = C + + + .
r r2 1+ r −δ

Housing prices are a function of exogenous population and exogenous shocks to wages

and amenities, and the derivative of housing prices with respect to a one dollar permanent

increase in wages will be (1+r)/r. Note that house price changes are predictable in this

framework as long as there are predictable components to changes in urban wages and

amenities. The ARMA(1,1) structure of the shocks makes it possible to have the positive

correlation of changes at high frequencies and the negative correlation at low frequencies

that we see in the data.

The city can grow with new construction so that N (t ) equals N (t − 1) + I (t ) , where

I(t) is the amount of construction in time t. 9 The physical, administrative and land costs

of producing a house are C + c0 t + c1 I (t ) + c 2 N (t − 1) , where c1 > c 2 because current

housing production should have a bigger impact on current construction costs than

housing production many years ago.10 Investment decisions for time t are made based on

time t-1 information, and there is free entry of risk neutral builders. Thus, if there is any

8
This assumption limits the possible role of housing bubbles. While our focus here is on a purely rational
model, we expect that future work will consider dropping this assumption.
9
For simplicity, we do not allow depreciation which may be reasonable for shorter term housing dynamics,
but would not be appropriate for a very long term analysis of city population changes.
10
We deviate from the investment cost assumptions of Topel and Rosen (1988) by assuming that costs are
increasing with the total level of development and not with changes in the level of investment.

9
building, construction costs will equal the time t expected housing price as described in

equation (3):

(3) Et −1 ( H (t )) = C + c0 t + c1 I (t ) + c 2 N (t − 1) .

As mentioned above, we assume that demand for the city is sufficiently robust so that

there is always a positive quantity of new construction and this equation always holds.11

Equations (1) and (3) then together describe housing supply and demand.

These equations give us the steady state values of housing prices, investment and

c 22 (q(1 + r ) + r (rC + D(1 + r ))) +

α (1 + r )(C + c1 (q(1 + r ) − rc0 )) +
αc 2 (1 + r )((c0 + D − q)(1 + r ) + 2rC ) (1 + r )(αc0 + qc 2 )
housing stock: Hˆ (t ) = + t,
(rc 2 + α (1 + r )) 2
rc 2 + α (1 + r )

rc1 (rc 0 − q(1 + r )) + α (1 + r )(c0 + D (1 + r ))

ˆI (t ) = Iˆ = q (1 + r ) − rc0 and Nˆ (t ) = + c 2 (q(1 + r ) + r ( D (1 + r ) − rc 0 ))

2

+ Iˆt .
rc 2 + α (1 + r ) (rc 2 + α (1 + r )) 2

If x(t)=0 for all t, and Nˆ (t ) = N (t ) for some initial period, then these quantities would

fully describe this representative city.12 Secular trends in housing prices can come from

trend in housing demand as long as c 2 > 0 , or the trend in construction costs as long as

α > 0 . If c 2 = 0 so that construction costs don’t increase with total city size, then trends

in wages or amenities will impact city size but not housing prices. If α = 0 and city size

doesn’t decrease wages or amenities, then trends in construction costs will impact city

size but not prices.

11
The model can be extended to allow for the possibility that, in some states of the world, new construction
will be zero. This adds much complication and only a modest amount of insight into our questions.
12
In this case, the assumption that there is always some construction requires that q (1 + r ) > rc 0 .

10
 Proposition 1 describes housing prices and investment when there are shocks to

demand and when Nˆ (t ) ≠ N (t ) . All proofs are in the appendix.

Proposition 1: At time t, housing prices equal

H (t ) = Hˆ (t ) +
φ
φ −δ
x(t ) +
θ
φ −δ
ε (t ) −
α (1 + r )
1+ r −φ
(
N (t ) − Nˆ (t ) )
and investment equals
I (t + 1) = Iˆ +
(1 + r )
c1 (φ − δ )
( )
(δx(t ) + θε (t )) − (1 − φ ) N (t ) − Nˆ (t ) ,
where φ and φ are the two roots of
c1 y 2 − ((2 + r )c1 + (1 + r )α − c 2 ) y + (1 + r )(c1 − c 2 ) = 0 and φ ≥ 1 + r ≥ 1 > φ ≥ 0 .

This proposition describes the movement of housing prices and construction around their

φ +θ
steady state levels. A temporary shock, ε , will increase housing prices by and
φ −δ

(1 + r )(δ + θ )
increase construction by . Higher values of δ ( i.e., more permanent
c1 (φ − δ )

shocks) will make both of these effects stronger. Higher values of c1 mute the

construction response to shocks and increase the price response to a temporary shock (by

reducing the quantity response). These comparative statics provide the intuition that

places which are quantity constrained should have less construction volatility and more

price volatility.

The next proposition provides implications about expected housing price changes.

Proposition 2: At time t, the expected home price change between time t and t + j is

Hˆ (t + j ) − Hˆ (t ) +
α (1 + r )
1+ r −φ
(
− φ j −1 ((1 − φ )c1 − c 2 ) N (t ) − Nˆ (t ) )
1 1 + r δ j −1 ((1 − δ )c1 − c2 ) − φ j −1 ((1 − φ )c1 − c2 )
− x(t ) + − 1 Et ( x(t + 1)) ,
φ − δ c1 φ −δ
the expected change in the city housing stock between time t and t + j is

11
 1+ r φ j −δ j
jIˆ +
c1 (φ − δ ) φ − δ
Et ( x(t + 1)) − (1 − φ j ) N (t ) − Nˆ (t ) ,( )
and expected time t + j construction is
δ (1 − δ ) − φ j −1 (1 − φ )
( )
j −1
1+ r
Iˆ + Et ( x(t + 1)) − φ j −1 (1 − φ ) N (t ) − Nˆ (t ) .
c1 (φ − δ ) φ −δ

Proposition 2 delivers the implication that a rational expectations model of

housing prices is fully compatible with predictability in housing prices. If utility flows in

a city are high today and expected to be low in the future, then housing prices will also be

expected to decline over time. Any predictability of wages and construction means that

predictability in housing price changes will result in a rational expectations model.

The predictability of construction and prices comes in part from the convergence

to steady state values. If x(t ) = ε (t ) = 0 and initial population is above its steady state

levels, then prices and investment are expected converge on their steady state levels from

above. If initial population is below its steady state level and x(t ) = ε (t ) = 0 , then price

and population is expected to converge on their steady state levels from below. The rate

c1 c2
of convergence is determined by r and the ratios and . Higher levels of these
α α

ratios will cause the rate of convergence to slow by reducing the extent that new

construction will respond to changes in demand.

The impact of a shock, x(t), is explored in the next proposition.

Proposition 3: If N (t ) = Nˆ (t ) , x(t − 1) = ε (t − 1) = 0 , θ > 0 , c 2 = 0 , and ε (t ) > 0 , then

investment and housing prices will initially be higher than steady state levels, but there
exists a value j * such that for all j > j * , time t expected values of time t + j
construction and housing prices will lie below steady state levels. The situation is
symmetric when ε (t ) < 0 .

12
 Proposition 3 highlights that this model not only delivers mean reversion, but

overshooting. Figure 3 shows the response of population, construction and prices relative

to their steady state levels in response to a one time shock. Construction and prices

immediately shoot up, but both start to decline from that point. At first, population rises

slowly over time, but as the shock wears off, the heightened construction means that the

city is too large relative to its steady state level. Eventually, both construction and prices

end up below their steady state levels because there is too much housing in the city

relative to its wages and amenities. Places with positive shocks will experience mean

reversion, with a quick boom in prices and construction, followed by a bust.

Finally, we turn to the puzzling empirical fact that, across the 1980s and 1990s,

there was strong mean reversion of prices and strong positive serial correlation in

population levels. We address this by looking at the one period covariance of price and

population changes. We focus on one-period for simplicity, but think of this proposition

as relating to longer time periods. Since mean reversion dominates over long time

periods, we assume θ = 0 to avoid the effects of serial correlation:

Proposition 4: If N (0) = Nˆ (0) , θ = 0 , x(0) = ε (0) , cities differ only in their demand
trends q and their shock terms ε (0) , ε (1) and ε (2) , and the demand trends are
uncorrelated with the demand shocks, then the coefficient estimated when regressing
second period population growth on first period population growth will be positive if and
δ (rc 2 + α (1 + r ) )
2
Var (q )
only if > (1 − δ − φ ) while the coefficient estimated when
Var (ε ) c1 (φ − δ )
regressing second period price growth on first period price growth will be negative if and
2
rc 2 + α (1 + r ) Var (q )
only if Ω > , where
(1 + r )c1c 2 (φ − δ ) Var (ε )
α (1 + r ) 2 δ αδ (1 + r ) 2
Ω= + c1 (1 − δ )φ (1 − δ − φ ) + c1 (1 − δ + δ 2 )φ .
1+ r −φ 1+ r −φ

13
 Proposition 4 tells us that positive correlation of quantities and negative

correlation of prices are quite compatible in the model. The positive correlation of

quantities is driven by the heterogeneous trends in demand across urban areas. As long as

the variance of these trends is high enough relative to the variance of temporary shocks,

then there will be positive serial correlation in quantities as in Figure 2.

The mean reversion of prices is driven by the shocks, and as long as c2 is

sufficiently low, prices will mean revert. As discussed above, when c2 is low, trends will

have little impact on steady state price growth. The positive trends show up mainly in the

level of prices. However, regardless of the value of c2 , unexpected shocks impact prices

and, if these shocks mean revert, then so will prices.

This suggests two requirements for the observed positive correlation of quantities

and negative correlation of prices: city-specific trends must differ significantly and the

impact of city size on construction costs must be small. The extensive heterogeneity in

city-specific trends is much commented on, with the recent papers by Gyourko, Mayer,

and Sinai (2006) and Van Nieuwerburgh and Weill (2006) attempting to explain the

phenomenon. The literature on housing investment suggests that the impact of city size

on construction costs is small (Topel and Rosen, 1988; Gyourko and Saiz, 2006). Thus,

we shouldn’t be surprised to see positive serial correlation in quantity changes and

negative serial correlation in price changes.

III. Key Parameter Values for the Calibration Exercises

We now use the model as a calibration tool to see what moments of the data,

including its serial correlation properties and variances, can and cannot be explained by

14
our framework. We focus on the movements in prices and construction intensity around

steady state levels. The appendix contains the formulae for the predicted values of these

moments.13 The model’s predictions about variances and serial correlations depend on

seven parameters: the real interest rate (r), the degree to which demand declines with city

population ( α ), the degree to which construction responds to higher costs (c1 and c2),

the time series pattern of local economic shocks ( δ and ), and the variation of those
2
shocks ( ). Table 1 reports the value of these parameters which are used in the

calibration exercise, with the remainder of this section discussing how we estimate or

impute them.

The Real Interest Rate (r)

The first row of Table 1 shows that we use a real interest rate (r) of 4 percent in all

calibrations. This value is higher than standard estimates of the real rate because it is also

meant to reflect other facets of the cost of owning, such as taxes or maintenance

expenses, that might scale with housing. Experimentation shows that the simulation

results are robust to a wide range of alternative values of r (e.g., from 2.5-5 percent).

Supply Side Parameters: c1 and c2

The housing cost parameters are both particularly important for our model and

relatively understudied by the literature. The parameter c1 reflects the extent that

construction costs, including land assembly, permitting and physical construction costs,

13
We do not use the high frequency correlations of prices with other variables to pin down parameter
values. Changes in house prices and changes in income accompany each other at longer horizons (e.g.,
over the past twenty years, the correlation of the two changes is over 50 percent), but the correlation is
much weaker at higher frequencies. Higher frequency correlations are difficult to interpret because the real
world information structure may not match that presumed in our model. For example, if income shocks are
known a period earlier, this will not matter much for predicted variances and serial correlations, but it will
dramatically alter the predicted relationship between income and price changes.

15
rise with the level of current construction activity. The c2 parameter measures the

sensitivity of costs to the level of overall development, or market size.

Topel and Rosen (1988) use national data and estimate a supply elasticity ranging

from 1.4 and 2.2. This supply elasticity is the relationship between the logarithm of

investment and the logarithm of price, which in our model would equal H (t ) / c1 I (t ) .

Using the mean values of investment and housing prices across our cities and an elasticity

of 1.8, this generates a range of c1 from 1 to 151. The median value is 18 which seems

too high to us. The range between 25th and 75th percentiles of the distribution is 5 to 28,

which provides us with one range of parameter estimates.

Alternatively, we can estimate c1 from the variance of prices and quantities

within the data. While this violates the typical calibration rule of using parameter

estimates from outside the data to be explained, we think that this exercise is useful in

generating a range of possible construction cost values. We estimate them for each of our

115 metropolitan areas and use this broad range in our simulations.14 Our hope is that by

showing results for a wide range of estimates, we can diffuse worries that naturally arise

from the fact that our estimates were made using the data that we are trying to explain.

We start with the basic supply-side equation

Et −1 ( H (t )) = C + c0 t + c1 I (t ) + c 2 N (t − 1) and let H (t ) = Et −1 ( H (t )) + µ , where µ is the

prediction error. We further assume that the variance of µ equals

14
Physical construction costs vary the least across U.S. markets, but Gyourko and Saiz (2006) report a 20
percent gap across the interquartile range of major metropolitan areas. The differences are driven largely
by the degree of union penetration in the local construction trades and select other local factors. The nature
of local land use regulation varies much more dramatically by market, with little binding constraint on new
development in markets such as Atlanta and Las Vegas, while the Boston and Bay Area markets have very
stringent and expensive regulation that makes it very hard to build even though market prices of homes are
well above physical construction costs in those places. [See Glaeser and Gyourko (2003), Glaeser,
Gyourko, and Saks (2006), and Saks (2006)) for studies on the impacts of differing land use regulation.]

16
κVar ( H (t ) − C − c0 t − c1 Iˆ − c 2 ( N (0) + tIˆ) , where the parameter κ is one minus the R2

from the best possible prediction of next period’s housing price. To determine this

unexplained share of the deviation from actual house prices, we begin by regressing

house prices on year and metropolitan area dummies. The R2 from that fixed effects

specification is 0.935. When we add two lags of house prices and housing permits to the

specification, the R2 increases to 0.993. Thus, κ = 10% or (0.1 ~ 0.07/0.7) based on these

values. We then let c 2 = ωc1 and using the equation

Var ( H (t ) − C − c0 t − c1 Iˆ − c 2 ( N (0) + tIˆ) = Var ( µ ) +

(4) ,
( (
c 2Var (( I (t ) − Iˆ) + ω N (t − 1) − N (0) − tIˆ
1 ))
it follows that,

Var ( H (t ) − C − c 0 t − c1 Iˆ − c 2 ( N (0) + tIˆ)(1 − κ )

(5) c12 = .
( (
Var (( I (t ) − Iˆ) + ω N (t − 1) − N (0) − (t − 1) Iˆ ))

The numerator of this ratio is the variance of the house price prediction error, weighted

by the share of unexplained variation. The denominator contains two components. The

first reflects the variation associated with yearly new construction deviating from its

average level. The second term in the denominator, which is weighted by , captures the

variation associated with the market’s housing stock being off its trend amount.

To empirically use equation (5), we must impute the housing stock (the N(t-1)

term) because the census provides actual counts of the stock only once each decade. For

each metropolitan area, we know the housing stock at the beginning and end of each

decade and the permits issued each year in between. Our estimate of the housing stock at

j −1
Permitst +i
time t+j, is N (t ) + i =0
9
( N (t + 10) − N (t )) , where N(t) and N(t+10) are the
i =0
Permitst +i

17
housing stocks measured during the two closest censuses. The change in housing stock is

portioned across years based on the observed permitting activity. To measure how many

units the market should have had each year (the “ N (0) + (t − 1) Iˆ ” term), we use the actual

count of the housing stock from the decennial census in our initial year of 1980 and

assume that I-hat is the average change in the actual housing stock between 1980 and

2000 (again using census data to measure the stock).

Since costs do not appear to vary much based on city development, we conclude

that c1 > c2, but beyond that, the literature yields little to help us pin down the

relationship between c1 and c2. We consider a range of values for , including 0, 0.25,

and 0.50.15

The bottom two panels of Table 1 report the distribution of estimated values of c1

and c2 for the three different values of omega. If =0.25, there is a wide range of values

for c1 from 1.5 for the 10th percentile metropolitan area to 28.1 for the 90th percentile

area. The mean value of c1 is 13.6 which is twice the median market’s c1 value of 6.4,

15
An alternative method of estimating these parameters suggests a value of 0.25 for . That approach to
estimating the construction cost parameters follows Rosen and Topel (1988) in inverting the construction
cost equation to obtain I(t) = (1/c1)(Et-1[H(t) – C) – (c2/c1)(Nt-1). In empirically implementing this equation,
we used total housing permits in period t to proxy for new construction in period t+1, and actual house
prices to measure expected values. Obviously, the use of actual prices in lieu of expected prices introduces
some bias, but it should be small since the annual time period over which price is measured is relatively
short. We also imputed the housing stock (N(t)) each year as described above. A simple regression of each
market’s resulting c2 value on its c1 value (with no intercept, as suggested by our assumed functional form)
yielded a coefficient of 0.25. The estimated coefficient is 0.21 if we allow for an intercept. The simple
correlation between c1 and c2 values estimated this way is quite high at 0.92. Finally, it is noteworthy that
the absolute level of these c1 values is higher than those reported in Table 1 based on the approach
described in equations (4) and (5). This was to be expected given that this alternative strategy effectively
assumes that price only reflects demand, not supply, shocks. To the extent that prices incorporate
productivity-enhancing changes, the relationship between new construction and values is magnified. More
problematic for this alternative approach is the issue of the potential endogeneity of housing prices with
respect to new construction. We also estimated a singe nationwide regression using the interaction of
initial industrial characteristics and national economic success of the industries as instruments (as in Bartik,
1989). When a linear specification is estimated, the coefficients are quite imprecise. When we follow
Rosen and Topel (1988) and use a log-log specification, we precisely estimate an elasticity of 2, which
implies a range of estimates of c1 and c2 similar to those presented in Table 1.

18
which reflects the skewness of the distribution of costs. A handful of large and

expanding markets such as Atlanta, Charlotte, Houston, and Dallas have c1 values below

one.16 The top ten percent of markets in terms of c1 values are all in Hawaii, along the

coast of California, or in the suburbs of New York City. These high values of c1 appear

to reflect both high labor costs and regulations that constrain construction.17 The

interquartile range of c2 runs from 0.7 to 3.3. The median value is 1.6

Increases in Population and the Marginal Valuation of an Area:

The value of reflects the impact that an increase in the housing stock will have

on the willingness to pay to live in a locale. If population was fixed, equation (2) tells us

that the derivative of housing prices with respect to the housing stock equals - (1 + r )α / r ,

which can be seen as the slope of the housing demand curve. Typically, housing demand

relationships are estimated as elasticities. Consequently, we must transform estimated

demand elasticities into a levels estimate by multiplying by r/(1+r).

While many housing demand elasticity estimates are around one (or slightly

below-in absolute value), there is a wide range in the literature, so we experiment with a

range from 0 to 2. We begin the transformation from an elasticity to a level by

multiplying by the ratio of price to population, which produces a range of estimates for

(1 + r )α / r of from 0 to 3. Multiplying this span by r/(1+r) yields a range from 0 to 0.15.

We will use a parameter estimate of 0.1 in our simulations which implies that for every

10,000 extra homes sold, the marginal purchaser likes living in the area $1,000 less per

16
The Atlanta area has the lowest value at 0.39.
17
The values for the metropolitan areas of Honolulu, Salinas, Santa Cruz, and Napa each are above 70.
While we cannot tell for sure, these magnitudes probably are associated with regulatory costs, as they seem
too high to solely reflect labor.

19
year (see row 5 of Table 1). This estimate seems high to us, but lower estimated values

of α do not significantly change the simulations.

2
Time Series Properties and Variance of Shocks: , , and

The model does not separately address wages and amenities. There is little

evidence on high frequency changes in amenities, except for crime rates which we will

discuss in Section V. Consequently, we assume here that the high frequency movement

in demand is driven by changes in labor demand, not changes in the valuation of

amenities. More specifically, observed wages W (t ) are presumed to equal

w0 − γN (0) + w1t + x(t ) − γ ( N (t ) − N (0)) , where w0 − γN (0) is a component of D , w1 is

a component of q and γ is a component of α . Controlling for a city-specific fixed effect

and trend will eliminate the term w0 − γN (0) + w1t , and the residual component of wages

equals x(t ) − γ ( N (t ) − N (0)) .18

The most difficult part of estimating the x(t) process is our attempt to control the

impact of population changes, but while our procedure is debatable, it has little impact on

the estimated properties of x(t). The parameter represents the impact that an increase in

city size will have on wages, which is proportional to the impact of labor supply on wage

(or the slope of the labor demand function). Customarily, labor demand is estimated as

Labor Force ∂Wage

an elasticity, , and most estimates of this elasticity are
Wage ∂Labor Force

statistically indistinguishable from zero (e.g., Card and Butcher, 1991). Borjas (2003)

finds a higher estimate of -0.3, although this is at the national level. We use this upper-

18
The distinguished literature on regional shocks (e.g. Blanchard and Katz, 1992) does not yield the
parameter estimates that we need to calibrate the model.

20
bound estimate, but note that it has little differential effect on our results compared to

assuming an elasticity of zero.

Just as with housing demand, we must convert this elasticity into an estimate of

γ . We use BEA data on personal income per capita as our measure of wages, and for our

sample of metropolitan areas, the mean of this variable in 1990 (the middle of our sample

period) was $26,965 (in $2,000). Mean employment in 1990 across these metropolitan

areas was 539,215, so our ratio of wage to the labor force is about 0.05

(~26,965/539,215). Based on these numbers, an elasticity of -0.3 suggests that each

worker is associated with 1.5 cents less annual income in the city. In our sample, there

are on average 1.26 workers per home, so each extra home is associated with 1.9 cents

per year less annual income, which serves as our estimate of γ .

The per capita income series is converted into household income by multiplying

by 2.63 (the average ratio of people per household in our sample in 1990). We adjust this

income variable for N(t) using our estimate of 0.019 for γ and by using

j −1
Permitst +i
N (t ) + i =0
9
( N (t + j ) − N (t )) (as above) as our estimate of total housing stock
i =0
Permitst +i

between census years.

With this corrected income series, we can estimate the time series properties of

income shocks at the local level, by fitting an ARMA(1,1) to the wage series that is first

demeaned with city and year fixed effects and then corrected for city size changes as

21
discussed above. As shown in Table 1, this estimation procedure yields estimates of
2
=0.87, =0.17, and = $3,603,463.19

IV. Calibrating the Model and Matching the Data

In this section, we calibrate the model using the parameters values discussed

above. We then compare this calibration to the moments of the real data. We first

discuss the time series coefficients of prices and construction, and then discuss the

volatility of these series. Our “real data” sample is a set of 115 metropolitan areas for

which we have continuously defined price data from 1980-2005. As discussed above,

our simulations assume that the real interest rate (r) equals 0.04, the variance of the local
2
economic shock ( ) equals $3,603,463, the parameters of the ARMA process describing

that shock are 0.87 ( ) and 0.17 ( ), and the valuation of the metropolitan area by the

marginal entrant ( ) is 0.1. We let c1 and c2 equal the fifteen different pairwise

combinations associated with the three different values of omega and report simulation

results for the full set of those values.

Short-Term Momentum and Longer-Term Mean Reversion in Prices, Rents and Permits

The top row of Table 2 provides evidence on momentum and mean reversion in

OFHEO house prices within market over time. We use absolute price changes rather

than changes in the logarithm of prices in order to be compatible with the model, but our

empirical results are not sensitive to such changes in functional form. Since the OFHEO

index only provides price increases relative to a base year, we convert this into an implied

19
Largely because is so small throughout its relevant range, this adjustment to wages does not have a
material impact on our results. If we use a value of 0 for γ , we estimate a value of δ = 0.86, an estimate of
θ = 0.18, and an estimate of 2=$3,408,250 In addition, we attempted joint maximum likelihood
estimation of , , and 2 for given trend effects and metropolitan area fixed effects, but the program would
not converge because the panel was too short relative to the number of markets.

22
price series by using the median housing value in the metropolitan area in 1980 as a base

price in the metropolitan area and then scaling that value by the appreciation in the

OFHEO index each year.20

The results are estimates from a regression of the current change in prices on the

lag change in prices

(6) Pr icet + j − Pr icet = α MSA + γ Year + β (Pr icet − Pr icet − j ) ,

for j equal to one, three and five years. Because fixed effects estimates such as these

which remove market-specific averages can be biased (with spurious mean reversion

produced especially when the number of time periods is relatively low), in the first row of

Table 2, we report Arellano-Bond estimates which use lagged values of the dependent

variable (price changes) as instruments.21

Our one year estimate of price change serial correlation is 0.71, so a $1 increase

in housing prices between time t and t+1 is associated with a 71 cent increase between

time t+1 and t+2. Our estimate is larger than that reported by the pioneering work of

Case and Shiller (1989). It is now well understood that smoothing of the underlying data

series can bias one towards finding short-run momentum. Case and Shiller (1989) were

able to address this problem by splitting their sample, which consisted of extensive micro

data on sales transactions in four markets (Atlanta, Chicago, Dallas, and San Francisco).

They report coefficients ranging from 0.2-0.5, although they use the logarithm, not the

20
This procedure essentially provides the real price for a constant quality house with the quality being that
associated with the median house in 1980. We have experimented with using values from the 1990 and
2000 censuses as the base. All the results reported below are robust to such changes.
21
See Arellano and Bond (1991) for more detail on this estimation procedure. More specifically, we use
the “xtabond” Stata command with year and area fixed effects.

23
level, of prices so the results are not exactly comparable. Because we cannot perform

any comparable procedure with the OFHEO data, our estimate is surely biased upwards.22

Over three years, there is still momentum. The estimate of 0.27 means that a $1

increase in housing prices between time t and t+3 is associated with a 27 cent increase

between time t+3 and t+6. Over five year periods, we estimate a mean reversion

coefficient of -0.32, so a $1 increase between times t and t+5 is associated with a 32 cent

decline between time t+5 and t+10.23 These estimates are not an artifact of the Arellano-

Bond procedure. The analogous ordinary least squares estimates over 1, 3, and 5 year

horizons are 0.74, 0.18 and -0.39, respectively. 24

The mean reversion in prices that we estimate over five-year horizons is quite

similar in magnitude to that observed for financial assets by Fama and French (1988).

Unfortunately, the short time period for which we have constant quality data at less than

decadal frequencies makes it difficult to know whether this mean reversion is a

permanent feature of urban life or whether it represents the impact of shocks that are

specific to the post-1980 time period. Cutler, Poterba and Summers (1991) also find this

22
The OFHEO index includes data on repeat sales or refinancings of the same house. The latter typically
rely on an appraisal, not a market sale price. Undoubtedly, this results in smoothing of the series and biases
upward our estimate of short-run momentum. Even the Case and Shiller (1989) estimates, which rely only
on actual sales, could be upward biased. Working with a split sample, bias can result if, randomly, some
fraction of homes on which a buyer and seller agree on a price have delayed closings that move their
reported sales dates into the next reported period (quarter, year, etc.). Whatever shock there was in period t
that influenced the agreed upon price, some of its measured impact will spill over into period t+1.
Obviously this is potentially more of a problem the shorter the measurement period.
23
As noted in the Introduction, decadal changes also find significant mean reversion across the 1980s and
1990s.
24
We also addressed concerns about spurious mean reversion by estimating specifications without
metropolitan area fixed effects. If we estimate the following equation,
Pr icet +5 − Pr icet = α + γ Year + β (Pr icet − Pr icet −5 ) , the mean reversion coefficient drops to -0.11 and
becomes only marginally significant. However, as soon as we include percent of adults with college
degrees as a control, the coefficient becomes -0.18 with a t-statistic of three. If we estimate the same
change regression using the logarithm of prices instead of the levels, the coefficient is -0.20 (-0.22 with the
college graduate control) and has a t-statistic of four.

24
pattern of short run momentum and longer- run mean reversion for housing and a number

of other asset markets.

Table 3 reports the comparable results for our simulations using the different c1

and c2 values discussed above. All other parameter values are fixed at the values listed in

Table 1. The first three columns show results for annual serial correlation in prices, the

next three columns present the analogous findings over three year periods, with the final

three columns being for five year periods. Within each time horizon, the fifteen cells

correspond to the fifteen c1, c2 pairs reported in Table 1.

The first three columns document the model’s failure to match the positive serial

correlation observed in the annual data. In fact, our parameter estimates suggest a mild

amount of mean reversion even at such a high frequency. Similarly, the results for three

year horizons reported in the middle columns of Table 3 find a mismatch with the data.

Assuming the middle case for omega ( =0.25, column 4), we predict mean reversion

coefficients from -0.18 to -0.28, not the positive persistence we see in the data.

Our model does a much better job of fitting the -0.32 mean reversion seen at five

year intervals (columns 6-9). At five year horizons, if c1 and c2 are at their medians for

the case of =0.25, we come within ten percent of matching the data (see row 3, column

8). And, if c1=2.7 and c2=0.7, the prediction literally is -0.32 (row 2, column 8). Thus,

the predictable mean reversion of prices at five year intervals cannot be seen as a

challenge for a rational expectations model of housing price movements.

In the model, this mean reversion reflects both the tendency of shocks to mean

revert and of new construction to cause future declines in prices. To decompose the

impact of the two forces, we also looked at the cases where there is no new construction

25
impact (i.e., =0) and where there is no mean reversion in the x(t) parameter. When

mean reversion in wages is turned off (i.e., =1), our simulations predict very low levels

of overall mean reversion in house prices. However, we predict higher levels of mean

reversion close to those found in the data when there is no effect allowed from

construction increasing market size. Thus, our analysis suggests that the majority of

mean reversion is coming from the mean reversion of local demand shocks.

One way to check whether short run momentum reflects the dynamics of euphoria

in an asset market is to see if the same phenomenon appears in rents. In the second row

of Table 2, we report the results from rental regressions of the form:

(7) Re nt t + j − Re nt t = α MSA + γ Year + β (Re nt t − Re nt t − j ) .

Rental data on apartments is collected by an industry consultant and data provider, REIS

Inc. Their data covers only a limited number of metropolitan areas (46 in our sample),

and in general, rental units are not comparable to owner-occupied housing.25

Over one- and three-year horizons, there is strong evidence of persistence, with

the Arellano-Bond estimates being 0.27 in both cases. Over five year time horizons, we

estimate a mean reversion parameter of -0.64. The presence of momentum and mean

reversion in rents suggests that these features do not reflect something unique to housing

asset markets, but rather something about the changing demand for cities.26

Table 4 then reports the predicted values of serial correlation from the simulations

of the model. At annual frequencies, we tend to predict very modest persistence, with the

25
Rental units are overwhelmingly in multi-unit buildings, while owner-occupied housing is
overwhelmingly single-family detached housing. These differences in housing types and the problem of
accurately measuring maintenance costs are two reasons why it is extremely difficult to tell whether
housing prices are high or low relative to rents.
26
The ordinary least squares estimates of these coefficients are 0.28, 0.08 and -0.51 for one, three and five
year horizons, respectively.

26
results ranging from -0.05 to 0.08 when =0.25 (column 2). These estimates are well

below the 0.27 seen in the data over one year horizons. Over three year horizons, we

consistently predict mean reversion, while there is still a positive correlation of rents in

the data. For five year intervals, we predict that rent changes should have a mean

reversion coefficient of about -0.30 if we use the median value of 6.4 and 1.6,

respectively, for c1 and c2 when =0.25 (row 3, column 8). Slightly higher mean

reversion is predicted if c1 is lower, but our estimates still are only about one-half of the

observed mean reversion in that case.

We are again unable to explain the strong positive serial correlations at shorter

time horizons. Since there are many reasons to be suspicious about the properties of the

rental data, especially because of artificial smoothing, we do not attach much importance

to the quantitative mismatch with the data here.27 However, the short run momentum and

long run mean reversion of rents, which are predicted by the model, suggest that these

features could reflect something other than irrationality in an asset market.

To examine the dynamics of housing quantity, we look at housing permit data

from the Census of Construction. The final set of results reported in Table 2 use housing

permits estimated in the following regression: Permits tt + j = α MSA + γ Year + βPermits tt− j ,

where Permits tt− j refers to the number of permits issued between time t-j and time t. The

one-three and five year Arellano-Bond coefficient estimates are 0.84, 0.43, and -0.07,

27
For example, smoothing is a greater problem in the rental data. The industry consultant that provides the
rent data does not survey actual renters, but the landlord owners of apartment buildings. Undoubtedly,
averages are being reported.

27
respectively. Thus, construction also displays high frequency momentum, but little or no

persistence or mean reversion at longer horizons.28

The calibration results for this variable are provided in Table 5. For the case

where c1 and c2 are the median values when =0.25, the predicted coefficients are 0.60,

0.29 and 0.06, for one, three, and five year horizons, respectively. These are reasonably

close to the actual parameters, and minor changes in the values of one or both of the

supply side parameters enable us to fit the data more exactly. While the predictions about

the serial correlation of construction are not as accurate as the predictions about the mean

reversion in prices, the moments of the real data cannot be said to reject the model.

Thus, the model does a reasonable job at fitting the time series properties of new

building and an excellent job at fitting the long term mean reversion of rents. It does a

poor job of fitting the high frequency positive serial correlation of price changes. This

failure may be the result of data smoothing causing us to empirically overestimate

momentum, or as Case and Shiller (1989) suggest, some sort of irrationality in the

housing market.

House Price Change and Construction Variances Across Markets

Table 6 reports the variance of price changes and of new construction in our

sample.29 The volatility of both prices and construction varies enormously across cities

and is quite skewed, with the mean variance much higher than the median variance.

Consequently, our approach to documenting both housing price change and construction

variance is to first run a pooled market regression (separately for each variable, of course)

28
As is the case with the other data, this pattern is not an artifact of our estimation procedure. The
analogous ordinary least squares coefficients are 0.82, 0.37, and 0.07, respectively.
29
Since the rent data are smoothed, we don’t put much weight on the variance of rents and exclude them
from this part of the analysis.

28
controlling for year effects, and then to compute the variance of the residuals from this

regression by metropolitan area. This variance gives us the volatility of prices and

construction, respectively, within a metropolitan area controlling for nationwide effects.

The top panel of Table 6 shows that that the variance of one year price changes

equals $14 million in the tenth percentile metropolitan area and $209 million in the 90th

percentile market. The median market has a one year price change variance of $34

million, which is much smaller that the sample mean of $83 million. This skewness is

driven primarily by California markets and Honolulu. The variance of one-year price

changes in Honolulu is $763 million, which is the largest in our sample. Five other

markets—San Jose, San Francisco, Santa Barbara, Santa Ana and Salinas--had variances

that were ten times greater than the sample median.

The second and third columns of this table report the distribution of variances of

three and five year price changes. The distribution of longer horizon price changes is

again quite skewed, with the mean price change substantially exceeding the change for

the median area. The volatility of price changes is very high at longer horizons. The

variance in five-year price changes is $625 million for the median market, with one

quarter of the metropolitan areas having variances of at least $1.1 billion.

Table 7 reports predicted price change variances from our simulations with the

results arrayed in the same manner as in the serial correlation tables above. At annual

frequencies (columns 1-3), we predict a range of price change variances from a low of

$50 million to a high of $181 million. Not surprisingly, price change volatility is lower

the smaller are c1 and c2—markets in which quantity changes a lot with changes in costs.

However, our variance prediction for the lowest (c1, c2) combination still is well above

29
the $34 million variance found in the median market. Data smoothing would bias price

volatility downward over short time periods, and if so, we would expect this problem to

be less severe for longer time periods.

Our ability to match the volatility of price changes does increase with the horizon

over which those changes are measured. For example, the range of predicted variances of

3-year price changes across all 15 (c1, c2) combination runs from $123-$488 million

(columns 3-6, Table 7). This spans the interquartile range of $124-$445 million found in

the data (column 2, top panel of Table 6), but we still need to have a relatively low value

for c1 (2.7) and a c2=0 to closely match the median. This is not the case for the 5-year

price change variance predictions listed in the final three columns of Table 7. Our range

of predictions runs from $163-$718 million, allowing us to capture much of the lower

half of the distribution of actual price change variation reported in the third column of

Table 6 (top panel).

The median (c1, c2) pair of 6.4 and 1.6 when =0.25 is associated with a predicted

variance of $469 million, which still underpredicts the sample median ($625 million), but

higher c1 values allow us to come much closer. However, we are still unable to approach

matching the very high price change volatility found in the top quarter of markets, and

the top ten percent, especially. That is an issue to which we will turn in the next section.

In sum, the general pattern of results in Table 7 shows an overestimate of price volatility

at high frequencies which then disappears at lower frequencies. This trend is consistent

with data smoothing, but it could also reflect a flaw in our model.

The bottom panel of Table 6 reports the variance in units permitted across our

115 metropolitan area sample. As with price changes, there is substantial heterogeneity

30
in the volatility of construction intensity across markets, and this distribution is skewed

by a few outliers. For example, the bottom quartile of markets has a new construction

variance of about 2 million units per annum, while the top quartile is at least five times

more volatile. Moreover, this distribution is skewed by relatively few markets in the

right tail that have variances of at least 38 million units (column 1, bottom panel of Table

6). Six markets—Phoenix, Dallas, Riverside-San Bernardino, Atlanta, Los Angeles, and

Houston---stand out in this regard, having construction intensity variances that are at least

double the next six highest variance markets. There also is great heterogeneity in

construction intensity variance over longer horizons, as the second and third columns of

this part of Table 6 document.

Table 8 reports the construction intensity variance estimates from our standard set

of simulations. Over annual periods, our model does a decent job of matching this data.

Our range of predicted construction variances runs from 300 thousand to 42 million units.

This captures the bulk of the range across markets in the actual data, which is 2 million

units in the 10th percentile market and 38 million units in the 90th percentile market

(Table 6). The median market in our sample has a one year standard deviation of 3

million units, which precisely equals the 3 million unit variance we predict if c1=6.4, the

median value in our sample when =0.25.

Over three year horizons, we can match the median market in the nation as well,

but the fit is not as perfect. For example, the three-year construction variance in the

median market is 26 million units (middle column, bottom panel of Table 6), while we

predict a 21 million unit variance for our ‘median’ (c1, c2) combination of 6.4 and 1.6

(see the middle cell in columns 4-6 of Table 8). Our range of estimates using our ‘lower’

31
(but not ‘lowest’ (c1, c2) values as in row 2 of these columns) spans the 84 million unit

variance associated with the 75th percentile metropolitan area (Table 6). Our ability to

match outcomes weakens for the most volatile markets. Our top variance estimate of 189

million units for three year horizons is only 58 percent of the 328 million unit variance

observed for the 90th percentile market.

Our under-prediction of construction intensity variation is greater at five year

horizons. Our full range of estimates runs from 3-328 million units, versus a 10th-90th

percentile range of 29-760 million in the data. Our ‘median’ estimate of 44 million units

is only three-quarters of the national market median of 59 million. Just as in the case of

prices, we are able to fit the volatility of the median markets with plausible parameter

values, but we completely fail to predict the volatility of the more extreme cases.

The model implies that markets with elastic supplies of housing (i.e., those with

low c1 and c2 values) should have relatively low price change variation and relatively

high construction volatility. Concomitantly, those areas with high c1 and c2 values that

have inelastic supplies will have little variation in quantities and more volatility in prices.

Hence, if our estimation procedure for determining these two supply side parameters

biases us from extreme values, we could be underestimating the impact of supply

conditions on both price and quantity volatility.

One natural explanation for these extreme markets is that we have significantly

overestimated the value of c1 and c2 in sunbelt region markets with elastic supplies of

housing. To address the role that mismeasurement of these parameters can play, we ran

four additional simulations using values for c1 and c2 from the Atlanta, Phoenix, San

Francisco, and Salinas (CA) metropolitan areas. The Atlanta market has the lowest c1

32
value in our sample, and assuming =0.25, its (c1, c2) combination is (0.4, 0.1). Phoenix

also experiences new supply if prices rise even slightly, and its (c1, c2) combination is

(1.0, 0.25). San Francisco and Salinas are at the opposite end of the spectrum, with

Salinas being a particularly extreme outlier. Their (c1, c2) combinations are estimated to

be (40, 10) and (100, 25) respectively.

Lower c1 and c2 values are associated with increases in predicted construction

volatility. Our parameter estimates for Phoenix predict 1, 3, and 5 year construction

variances of 34 million, 186 million, and 357 million units, respectively. The analogous

figures for our Atlanta parameter estimates are 74 million units, 361 million units, and

651 million units, respectively. These numbers still are well below what we observe in

the most highly volatile construction markets (e.g., Atlanta’s five year variance is 4.52

billion units), but they do match the values for the market in the 90th percentile of the

distribution (see Table 6). The most volatile construction markets still remain a puzzle

for our model, but we interpret these results to imply that underestimating how elastic the

supply side is could help account for a meaningful part of our underestimation of the

variation in construction intensity in many sunbelt markets.

Significantly higher c1 and c2 values do not result in commensurate increases in

predicted price change variation, so even if we are underestimating these parameters in

highly volatile coastal markets, that cannot explain the dramatic price change variation

we see in those place. More specifically, the 1, 3, and 5 year predictions assuming supply

side conditions are those found in San Francisco are $181 million, $487 million, and

$714 million, respectively. The Salinas estimates yield similar predictions: $190 million,

$515 million, and $759 million, respectively. These estimates are barely above the

33
maximum values reported in Table 7. Although supply elasticity can explain a

significant amount of the high construction volatility in the sunbelt, supply inelasticity

cannot explain much of the high price volatility in coastal markets. We now turn to other

potential explanations.

V. Explaining the High Volatility of Housing Prices

To see if we can match the high volatility of housing prices observed in certain

markets in the data, we first turn to omitted demand factors such as local tax rates and

amenities. Following that, we examine the impact of potential mismeasurement of

income shocks. Finally, we look at the role of time-varying real interest rates.

Local Tax Rates and Amenities

While we assume that income volatility is the only source of high frequency

changes in demand in our simulations, changes in local tax rates and amenity flows also

could affect volatility. To address volatility related to tax rates, we used data from the

NBER TAXSIM website on the average tax rate on wage income earned in a given state

each year to create an after-tax income measure for each market. Our analysis of after-

tax income showed that controlling for this factor cannot be responsible for more than a

10 percent increase in local demand variability which would translate into a ten percent

increase in price and construction volatility. Appendix II provides the details, but we

conclude that changes in state level tax rates cannot be driving the high price volatility.

We think of most amenities as being relatively permanent characteristics of a

place, (e.g., the weather, local architecture). The demand for these amenities may change

slowly over time as a society becomes richer or more unequal or as new technologies

34
become available, but it is hard to imagine that their value will fluctuate a lot at annual

frequencies. Crime represents one of the few amenities that does change relatively

rapidly and for which there is available data. Since crime represents a possible omitted

amenity, we collected violent crime rates for the largest cities in each of our metropolitan

areas using continuous crime data from 1985-2005.30 We then created an adjusted

income variable that subtracted the negative effect of crime from our BEA real income

measure. As detailed in Appendix III, the results showed almost no impact on the

variability of local demand from controlling for crime. We infer from this that we are

unlikely to be find an amenity with high frequency variation that can explain much of the

observed volatility in prices or construction.

Measurement of Income Shocks

There are two reasons why our estimates of income volatility might be

understating the true magnitude of income shocks in high volatility markets. First, our
2
estimate of is based on all 115 markets in the sample, and if income variability were

systematically higher in the high price change variance markets, then our estimates of

price volatility would be biased downward in those markets. Second, our use of BEA per

capita income makes no allowance for the possibility that the volatility of the marginal

home buyer’s income could be relatively high.

To look at the impact of heterogeneous income variability across markets, we

focus on coastal markets that have particularly volatile housing prices. We re-estimated
2
using the same ARMA procedure described above for a subsample of 31 markets

whose centroids are within 50 miles of the Atlantic or Pacific Oceans. While the AR ( )

30
We emphasize that this measure is for the local political jurisdiction, which we then impute to the
metropolitan area.

35
and MA ( ) components were little changed from those reported for the 115 market
2
national sample, the estimate of is almost 50 percent higher in the 31 coastal markets:

$5.3 million versus $3.6 million. This difference in volatility of local wage shocks is

large, and to our knowledge, has not been well-documented and is not well-understood.

Highly productive coastal areas might specialize in highly volatile idea-intensive

industries, but this is an appropriate topic for future research.

The puzzle of excess price change variance also could be at least partially

explained if our measure of income variance understates the true year-to-year variation in

the returns from living in the city for the marginal homebuyer. For example, if marginal

buyers are young, then this might mean that their incomes are more volatile. In cities

with vibrant economies, buyers on the margin might be people in the cities’ fastest

growing and most volatile industries. In New York City, the marginal homebuyer might

be more likely to be in finance, for example, and have a more volatile income. In San

Francisco, the marginal homebuyer might be more likely to work in the volatile

technology sector.

To investigate this hypothesis, we turned to the New York City Housing and

Vacancy Survey (NYCHVS) which provides extremely detailed information on New

York City residents. This survey was available for the years 1978, 1981, 1984, 1987,

1991, 1993, 1996 and 2002. In each year, we calculated the mean income for the sample

of people who bought a house within the last two years. This two year window is meant

to capture the population of recent buyers, while also providing a decent sample size

(e.g., there are between 100 and 400 observations in each year).

36
 The incomes for this sample could be more volatile than the BEA per capita

figures simply because of smaller numbers, so our approach is not to look at its variance,

but at its correlation with BEA real income in New York. Specifically, we estimated the

following regression:

(8) Recent Buyer Real Income=-32,451+ 1.29*BEA Per Capita Real Income,
(15,102) (0.19)

There are nine observations (one for each survey year), the R2=0.87, and standard errors

are in parentheses. This regression suggests that recent buyer income increases by $1.29

for every $1 increase in BEA-measured income.31 With respect to our simulations, if we

assume that the value of x(t), the income shock, for recent owners is 1.29 times the shock

for the entire population, then the variance of shocks is 1.66 times the variance for the

entire population.32

Including both effects implies predicted variances in high (c1, c2) markets are 2.4

times greater than those reported in Tables 7 and 8 (1.66*1.44~2.4). Because our

evidence certainly is not strong enough to warrant confidence in such a precise result, we
2
report ranges of values for new simulation results assuming the local demand shock ( )

is 50 percent, 100 percent, and 150 percent greater than the $3.6 million figure used in

31
While the small samples and nine observations make it hard to draw too much from this, it is noteworthy
that this procedure does not generate estimates above $1 automatically. When we performed the same
analysis using the sample of renters in the survey, the following resulted:

Renter Real Income= -2,885 + 0.47*BEA Per Capita Real Income

(5,436) (0.07)

The number of observations again is nine, the R2 still is 0.87, and standard errors are in parentheses.
32
In addition, if recent buyer incomes were more cyclical than average incomes, then the variance of prices
might be substantially higher than the variance predicted by the model. A related point is that if the New
York City regressions are to be believed, then the income of renters is far less volatile than the income of
owners. The regression predicts that the variance of their income shocks is one quarter of the variance of
income shocks for the general population. This difference may be one reason why rents are so stable
relative to housing prices, and it also suggests that comparing rents and prices is a very tricky process,
indeed.

37
the baseline simulations reported above. Because the key outstanding price-related

puzzle for our model is the high volatility of lower frequency price changes in the top

quartile of metropolitan areas, the top panel of Table 9 provides new estimated variances

of five-year price changes for markets with the 75th and 90th percentile values of our

supply-side parameters (c1 and c2), assuming more variable local demand shocks.

The first column reproduces our baseline estimates for these parameter values

from Table 7. Recall that those simulation results are well below the $1.17 billion

variance in five-year price changes found in the metropolitan area that is in the 75th

percentile for all such changes, and not even close to the $3.58 billion variance observed

for the 90th percentile market (see column 3, top panel of Table 6 for those values). The

next three columns report predicted variances assuming the three higher estimates of
2
local demand variability. The first row of the third column indicates that needs to be

doubled in order to account for all of the mismatch between predicted and actual variance

at the 75th percentile of the distribution (i.e., $1,190/$1,170~1).

While this represents a large increase in local volatility, based on the discussion

above, it is not implausible. However, the results in the second row of the top panel of

Table 9 indicate that our model still cannot account for the most volatile ten percent of
2
markets. If our baseline estimate is off by 150 percent, the predicted variance of

$1.71 billion is only 48 percent of the $3.58 billion variance seen in the 90th percentile

metropolitan area.

Even the best of models have trouble accounting for the outliers in any

distribution. Still, examining the 12 most volatile markets (the top 10 percent of our 115

metropolitan areas), in terms of price changes over five-year periods, shows that what our

38
model just cannot explain is the volatility of coastal California (plus Honolulu). 33

Literally, there are no east coast markets in this group, with the Nassau-Suffolk and

Bethesda-Gaithersburg-Frederick metropolitan areas having the 15th and 16th biggest

price change volatilities.

The bottom panel of Table 9 reports new estimated variances of construction

intensity assuming supply-side parameters (c1 and c2) consistent with values associated

with the 10th and 25th percentiles of that distribution and assuming a 66 percent increase
2
in . We focus on the impact in markets with more elastic supply sides because the

remaining puzzle about quantities is the very high variation observed in many of the high

growth sunbelt markets. We only allow demand side volatility to increase by 66 percent

because there is no evidence to suggest their intrinsic income variability is higher than the

average we use in the baseline simulations34, but we would expect the volatility of the

marginal buyer’s income to be greater than the average we are using in the baseline

simulations. Even assuming the NYCHVS results are applicable to these markets, the

results show that underestimating local demand variability cannot account for our

underestimate of quantity variation for the most volatile markets in this respect.

Time-Varying Interest Rates

We have so far assumed that interest rates are fixed for reasons of tractability, but

understand that many authors have claimed that the dramatic rise in house prices,

33
The top ten percent of the most volatile metropolitan areas in terms of five-year price changes (in
ascending order from #104-#115) are as follows: Oakland-Fremont-Hayward, Santa Cruz-Watsonville,
San Luis Obispo-Paso Robles, San Diego-Carlsbad-San Marcos, Oxnard-Thousand Oaks-Ventura, Los
Angeles-Long Beach-Glendale, San Jose-Sunnyvale-Santa Clara, Salinas, Santa Ana-Anaheim-Irvine, San
Francisco-San Mateo-Redwood City, Santa Barbara-Santa Maria, and Honolulu.
34
Mathematically, it is somewhat smaller given that coastal market income volatility is higher. We ignore
this effect for simplicity. It does not change any of our conclusions.

39
especially in high cost markets, over the past decade is best understood as a response to

declining interest rates that make housing in those areas more affordable (e.g.,

Himmelberg, Mayer and Sinai, 2005). A full treatment of interest rates would require an

analysis of long period mortgages and prepayment that lies well beyond the scope of this

paper. We can, however, adjust the model modestly to acquire some understanding of

the potential impact of time-varying interest rates.

To do so, we decompose interest rates into permanent and transitory components,

r and ρ (t ) , resprectively, where r (t ) = r + ρ (t ) , and use the approximations

1 1
≈ and r (t )( H (t ) − C ) = r ( H (t ) − C ) + ρ (t )( H − C ) , where H is meant to
1 + r (t ) 1 + r

reflect the average housing price in the city.35 If we adjust equation (2) for time-varying

interest rates using these approximations, equation (2’) results:

rC E ( H (t + 1))
(2’) H (t ) − − t = D + qt + x(t ) − ρ (t )( H − C ) − αN (t )
1+ r 1+ r

If we then make the somewhat unrealistic assumption that ρ (t ) = λρ (t − 1) + η (t ) so the

difference equation remains linear, the model can be solved in a relatively

straightforward fashion.

This results in equation (9)’s description of prices,

H (t ) = Hˆ (t ) +
(c1φ + Ψ ) x(t )
+
c1θε (t )
−
α (1 + r )
c1 (φ − δ ) + Ψ c1 (φ − δ ) + Ψ 1 + r − φ
(
N (t ) − Nˆ (t ) )
(9)
(c1φ + Ψ )( H − C ) ρ (t )
−
c1 (φ − λ ) + Ψ

35
The first approximation is minor and would have been unnecessary if we assumed that the utility flow
was received at the end of the period rather than the beginning of each period. The second approximation
eliminates interactions between transitory changes in value and transitory changes in the interest rate and it
may be more consequential.

40
This differs from the price equation in Proposition 1 because of its last term which

c1φ + Ψ
multiplies times the interest rate shock times H − C (the gap between
c1 (φ − δ ) + Ψ

average housing prices in the area and construction costs). This term reflects the fact that

a decline in interest rates essentially is a positive demand shock for high amenity and

productivity places. The shock makes it cheaper to live in such places, pushing up

demand and prices.

Since our regressions correct for year effects, this interest rate effect can have no

impact on our empirical estimates for the average market which will have prices close to

construction costs. The interest rate effect does, however, have the capacity to generate

increased variance in both price changes and construction levels for places that are

considerably more expensive on average. We consider four different values of ( H − C ) :

$25,000, $50,000, $100,000 and $200,000, which over the past 25 years captures most of

the range of American metropolitan areas.36

We set c1, and c2 equal to 6.4 and 1.6, respectively, in order to focus on interest

rates effects. We assume that =0.90, but experimentation with values as high as 0.95

yield similar results. The variance of interest rates is far more important, and we use a

range of standard deviations for (t) from 0.005 to 0.02 which appears to encompass

most assessments of the amount of real rate variation.37

36
For example, in 1980 the highest price metropolitan area had a median house value that was about
$170,000 greater than in the median market. The real value of median market’s median house price is
barely changed between 1980 and 2000. Except for a handful of markets in the upper tail of the
metropolitan area price distributions, gaps in excess of $200,000 with the median market do not exist.
Finally, we omit runs with a value of zero because they correspond to the simulations from the previous
section.
37
Campbell’s (2000) review of the asset pricing literature notes that the standard deviation on a one period
riskless asset is 1.76 percent, but concludes that “… perhaps half … is due to ex post inflation shocks (p.

41
 Simulations suggest that these changes to our baseline model make little

difference to the amount of predicted mean reversion. Thus, the results in Tables 10 and

11 focus on the impact of interest rate volatility on the variance of price changes and

construction, respectively. Table 10 shows that including interest rates shocks can

generate significant increases in the variance of price changes if the interest rate shock is

quite high and if the market has house prices much greater than the average (and, thus,

much more than construction costs). For example, comparing the predicted variances in

Table 10 for markets $25,000 or $50,000 above the average market with those reported in

Table 7 (for the same (c1, c2) parameters) shows that interest rate volatility does not

increase the predicted volatility of price changes much at all. To generate a predicted

variance above the $209 million observed over annual periods for the 90th percentile

metropolitan area (see Table 6) requires that H-C be $200,000 if the standard deviation of

(t) is one percent.

At three year intervals, the predicted variance without interest rate shocks is $330

million (see the middle cell of Table 7), which is higher than the median market (Table

6), but less than the mean and far less than the 90th percentile metropolitan area.

Including interest rate shocks with a standard deviation of 0.01 and a $100,000 gap

between prices and construction costs increases the predicted variance to $418,000,

which is near the volatility observed in the 75th percentile city according to the figures in

Table 6. To fit the 90th percentile city’s price change variation of $1.38 billion, the gap

between average prices and construction costs needs to be $200,000 dollars and the

1519).” Thus, the lower half of this range may be more plausible. Recent asset pricing papers such as
Bansal, Kiku, and Yaron (2006) assume a standard deviation of 1 percent.

42
standard deviation of interest rates needs to be much greater than 0.01. We think that

both of these assumptions are extreme.

At five year intervals, including interest rate shocks again increases the predicted

variation significantly in the most attractive or productive markets, but the predicted

variation is still far less than is actually observed in the most volatile markets according

to the data in Table 6. A 0.01 standard deviation interest rate shock and a $100,000 gap

between housing prices in the city and construction costs increase the predicted variance

by about one quarter. To get the much higher variances that we seen in the data, the gulf

between prices and construction costs must be over $100,000 and the shock to interest

rates much have a standard deviation greater than 0.01.

The ability of interest rate shocks to explain the high level of variation in

construction is far more limited. For example, Table 11 documents that even if the

standard deviation of shocks is 0.02 and H − C equals $200,000, the model still does not

predict the sample mean variance of construction. At more reasonable parameter values,

such as a standard deviation of interest rate shocks of 0.01 and H − C equal to $100,000,

interest rate shocks predict almost no increase in the variance of construction at one year

intervals, an increase in construction variance of 30 percent at three year intervals and an

increase in construction variance of 25 percent at five year intervals. Thus, interest rate

shocks do increase predicted construction variance, but they do so modestly.

One clear implication of the model is that if interest rate shocks are important,

then the variance of price changes and construction should be higher in high price areas.

Figure 4 graphs the variance of one year price changes for each metropolitan area against

its average price in 1980. The graph shows a strong positive relationship, just as

43
predicted by the role of interest rates. The most volatile places in the country are places

that were most expensive in 1980. Interest rate shocks are one explanation of this

phenomenon. However, another possible explanation is that these places had high costs

because they restricted construction, so that Figure 4 is showing the impact of restricted

construction on volatility. However, a quick comparison of Table 7 and Table 10 shows

that interest rates can generate this high level of price volatility more readily than

restricted construction without interest rate shocks. We suspect both phenomena are at

work in high cost, high volatility areas.

In Figure 5, we graph the variance of one year construction rates on the average

price in the metropolitan area in 1980. In this case, there is no visible relationship,

perhaps because restrictions on construction in high cost areas ensure low levels of

construction volatility. Nonetheless, we think that both Tables 10 and 11 and Figures 4

and 5 suggest that interest rates shocks can plausibly play a role in explaining some of the

observed price volatility in high cost areas. Those results also suggest that interest rates

are unlikely to explain high construction volatility in lower cost, sunbelt areas.

VI. Conclusion

This paper presents a dynamic rational explanations model of housing markets

based on a cross-city spatial equilibrium. The model predicts that housing markets will

be largely local, which they are, and that construction persistence is fully compatible with

price mean reversion. The model is also consistent with price changes being predictable.

The model has successes and failures at fitting the real data. The model can

explain the serial correlation of construction quantities reasonably well and can explain

44
the five year mean reversion of prices almost perfectly. However, the model cannot

explain the high frequency positive serial correlation of price changes. The model can

explain the price and construction volatility of the nation’s typical housing markets, but it

does a poor job of explaining the most volatile markets.

Time-varying interest rates can in principle explain some of the price variation in

high cost markets, but interest rates are unlikely to explain much of the extreme cases of

construction variation. Our simulations showed that interest rate shocks do a better job of

fitting the markets with high levels of price variance than the markets with high levels of

construction variance. Across cities, price volatility is concentrated in high costs areas,

which is a prediction of the model when it includes interest rates. Construction volatility

is concentrated in lower cost markets, which the model suggests should have little

responsiveness to interest rates.

There are two problems with concluding too much from our interest rate findings.

First, on a theoretical level, we have omitted many important features of mortgage

contracts such as a the prepayment option which seem crucial to us in understanding the

impact the interest rates will have on price dynamics. Second, empirically, interest rates

explain only a small portion of price volatility even in high price areas. We hope that

future research will focus more on this important topic.

Finally, the value of this model is as much in what it cannot explain as in what it

can explain. It suggests that housing economists should focus their attention on high

price volatility in coastal markets and on the positive serial correlation of price changes.

The average volatility and longer-term mean reversion of prices should no longer be

viewed as puzzles.

45
 References

Alonso, William (1964). Location and Land Use. Cambridge, MA: Harvard University
Press.

Arellano, Manuel and Stephen Bond (1991). “Some Tests of Specification for Panel
Data: Monte Carlo Evidence and an Application to Employment Equations”, Review
of Economic Studies, Vol. 58, no. 2 (April): 277-97.

Baker, Dean (2006). “The Menace of an Unchecked Housing Bubble”, The Economists’
Voice, Vol. 3, no. 4, Article 1.

Bansal, Ravi, Dana Kiku, and Amir Yaron (2006). “Risks for the Long Run: Estimation
and Inference”, working paper, October 2006.

Blanchard, Olivier and Lawrence Katz (1992) “Regional Evolutions” Brookings Papers
on Economic Activity 1: 1-75.

Borjas, George (2003). “The Labor Demand Curve Is Downward Sloping: Reexamining
the Impact of Immigration on the Labor Market”, Quarterly Journal of Economics,
Vol. 188, no. 4 (November): 1335-1374.

Campbell, John (2000). “Asset Pricing at the Millennium”, Journal of Finance, Vol. LX,
no. 4 (August): 1515-1567.

Card, David and Kristin Butcher (1991). “Immigration and Wages: Evidence from the
1980s”, American Economic Review, Vol. 81, no. 2 (May): 292-296.

Case, Karl E. and Robert Shiller (1989). “The Efficiency of the Market for Single Family
Homes”, American Economic Review, Vol. 79, no. 1: 125-37.

Cutler, David, James Poterba, and Lawrence Summers (1991). “Speculative Dynamics”,
Review of Economic Studies, Vol. 58, no. 3 (May 1991): 529-46.

Fama, Eugene and Kenneth French (1988). “Permanent and Temporary Components of
Stock Prices”, Journal of Political Economy, Vol. 96, no.2 (April 1988): 246-73.

Glaeser, Edward and Joseph Gyourko (2005). “Urban Decline and Durable Housing”,
Journal of Political Economy, Vol. 113, no. 2 (April): 345-375.

Glaeser, Edward, Joseph Gyourko and Raven Saks (2005). “Why Have House Prices
Gone Up?”, American Economic Review, Vol. 95, no. 2 (May): 329-333.

Glaeser, Edward, Joseph Gyourko and Raven Saks (2006). “Urban Growth and Housing
Supply”, Journal of Economic Geography, Vol. 6, no. 1 (January): 71-89.

46
Glaeser, Edward and Jesse Shapiro (2003). “Urban Growth in the 1990s” Is City Living
Back?’, Journal of Regional Science, Vol. 43 (February): 139-165.

Gyourko, Joseph, Christopher Mayer, and Todd Sinai (2006). “Superstar Cities”, NBER
Working Paper 12355, July 2006.

Gyourko, Joseph and Albert Saiz (2006). “Construction Costs and the Supply of Housing
Structure”, Journal of Regional Science, Vol. 46, no. 4 (November 2006):
forthcoming.

Hansen, Lars and Ravi Jagannathan (1991). “Implications of Security Market Data for
Models of Dynamic Economies”, Journal of Political Economy, Vol. 99, no. 2
(April): 225-262.

Hendershott, Patric and Joel Slemrod (1983). “Taxes and the User Cost of Capital for
Owner-Occupied Housing”, Journal of the American Real Estate and Urban
Economics Association, Vol. 10, no. 4 (Winter): 375-93.

Himmelberg, Charles, Christopher Mayer and Todd Sinai (2005). “Assessing High
House Prices: Bubbles, Fundamentals, and Misperceptions”, Journal of Economic
Perspectives, Vol. 19, no. 4 (Fall 2005): 67-92.

McCarthy, Jonathan and Richard Peach (2004). “Are Home Prices the Next Bubble?”,
Economic Policy Review, Vol. 10, no. 3: 1-17.

Poterba, James (1984). “Tax Subsidies to Owner-Occupied Housing: An Asset Market

Approach”, Quarterly Journal of Economics, Vol. 99, no. 4: 729-52.

Roback, Jennifer (1982). “Wages, Rents, and the Quality of Life”, Journal of Political
Economy, Vol. 90, no. 4 (December 1982): 1257-78.

Rosen, Sherwin (1979). “Wage-Based Indexes of Urban Quality of Life”. In Current

Issues in Urban Economics, edited by Peter Mieszkowski and Mahlon Straszheim.
Baltimore: Johns Hopkins Univerity Press, 1979.

Schwartz, Amy Ellen, Scott Susin and Ioan Voicu (2003). “Has Falling Crime Driven
New York’s Real Estate Boom?”, Journal of Housing Research, Vol. 14, no. 1: 101-
135.

Shiller, Robert (2005). Irrational Exuberance (2nd edition). Princeton University Press.

____________ (2006). “Long-Term Perspectives on the Current Boom in Home Prices”,

The Economists’ Voice, Vol. 3, no. 4, Article 4.

Smith, Margaret Hwang and Gary Smith (2006). “Bubble, Bubble, Where’s the Housing
Bubble”, Pomona College mimeo, March 2006.

47
Thaler, Richard (1978). “A Note on Crime Control: Evidence from the Property
Market”, Journal of Urban Economics, Vol 5, no. 1 (January): 137-145.

Topel, Robert and Sherwin Rosen (1988). “Housing Investment in the United States”,
Journal of Political Economy, Vol 96, no. 4: 718-740.

Tracy, Joseph, Henry Schneider and Sewin Chan (1999). “Are Stocks Overtaking Real
Estate in Household Portfolios?”, Current Issues in Economics and Finance, Vol. 5,
no. 5 (April): 1-6

Van Nieuwerburgh, Stijn and Pierre-Olivier Weill (2006). “Why Has House Price
Dispersion Gone Up?”, NBER Working Paper #12538, September.

48
 Table 1: Model Parameters
R 0.04
0.87
0.17
2
$3,603,463
0.1
=0.0 =0.25 =0.50
th
10 percentile value: 1.7 1.5 1.0
25th percentile value: 3.5 2.7 2.0
C1
50th percentile value: 7.0 6.4 4.8
75th percentile value: 14.4 13.3 10.3
90th percentile value: 28.8 28.1 22.9
C2 =0.0 =0.25 =0.50
10th percentile value: 0.0 0.4 0.5
25th percentile value: 0.0 0.7 1.0
50th percentile value: 0.0 1.6 2.4
75th percentile value: 0.0 3.3 5.2
90th percentile value: 0.0 7.0 11.5

49
 Table 2: Variation in Prices and Quantities Within-Market Over Time
Arellano-Bond Estimates of Coefficients on Lagged Dependent Variable
1, 3, & 5 year horizons
Dependent Variable 1-year 3-year changes 5-year changes
changes
House Price Change 0.71 0.27 -0.32
(0.01) (0.04) (0.07)
N=2,819 N=690 N=345
Rent Change 0.27 0.27 -0.64
(0.03) (0.08) (0.17)
N=1,007 N=274 N=91
New Permits 0.84 0.43 -0.07
(0.01) (0.04) (0.06)
N=2,645 N=690 N=460
Notes:
1. Sample for house price, employment, and permit specifications is 115 metropolitan
area sample described in text.
2. Sample for rent specification is 46 metropolitan areas tracked by REIS.

50
 Table 3: Predicted Serial Correlation in House Prices
One-Year Three-Year Five-Year
c2 if c2 if c2 if c2 if c2 if c2 if c2 if c2 if c2 if
=0 = 0.25 = 0.5 =0 = 0.25 = 0.5 =0 = 0.25 = 0.5
lowest
-0.09 -0.11 -0.13 -0.29 -0.28 -0.25 -0.41 -0.35 -0.31
c1
lower
-0.07 -0.09 -0.09 -0.25 -0.25 -0.22 -0.36 -0.32 -0.28
c1
medium
-0.05 -0.07 -0.07 -0.21 -0.21 -0.19 -0.32 -0.29 -0.26
c1
higher
-0.05 -0.05 -0.05 -0.19 -0.19 -0.17 -0.29 -0.27 -0.25
c1
highest
-0.04 -0.05 -0.05 -0.18 -0.18 -0.17 -0.27 -0.26 -0.25
c1

Notes: Within each time horizon over which serial correlation is estimated, 15 simulations were run, corresponding to the 15 (c1, c2)
pairs reported in Table 1. Thus, the (c1, c2) combinations used in the first, fourth, and seventh columns for which =0 are (1.7, 0.0),
(3.5, 0.0), (7.0, 0.0), (14.4, 0.0), and (28.8, 0.0), going from the top to the bottom of the table. The (c1, c2) combinations used in the
second, fifth, and eighth columns for which =0.25 are (1.5, 0.4), (2.7, 0.7), (6.4, 1.6), (13.3, 3.3), and (28.1, 7.0). The (c1, c2)
combinations used in the third, sixth, and ninth columns for which =0.50 are (1.0, 0.5), (2.0, 1.0), (4.8, 2.4), (10.3, 5.2), and (22.9,
11.5).

51
 Table 4: Predicted Serial Correlation in Apartment Rents
One-Year Three-Year Five-Year
c2 if c2 if c2 if c2 if c2 if c2 if c2 if c2 if c2 if
=0 = 0.25 = 0.5 =0 = 0.25 = 0.5 =0 = 0.25 = 0.5
lowest
0.02 -0.05 -0.14 -0.25 -0.31 -0.34 -0.37 -0.39 -0.38
c1
lower
0.05 -0.00 -0.07 -0.21 -0.26 -0.27 -0.32 -0.35 -0.33
c1
medium
0.07 0.04 0.01 -0.17 -0.21 -0.20 -0.29 -0.30 -0.28
c1
higher
0.08 0.07 0.05 -0.15 -0.17 -0.17 -0.26 -0.27 -0.25
c1
highest
0.09 0.08 0.07 -0.14 -0.15 -0.15 -0.24 -0.25 -0.24
c1

Notes: Within each time horizon over which serial correlation is estimated, 15 simulations were run, corresponding to the 15 (c1, c2)
pairs reported in Table 1. Thus, the (c1, c2) combinations used in the first, fourth, and seventh columns for which =0 are (1.7, 0.0),
(3.5, 0.0), (7.0, 0.0), (14.4, 0.0), and (28.8, 0.0), going from the top to the bottom of the table. The (c1, c2) combinations used in the
second, fifth, and eighth columns for which =0.25 are (1.5, 0.4), (2.7, 0.7), (6.4, 1.6), (13.3, 3.3), and (28.1, 7.0). The (c1, c2)
combinations used in the third, sixth, and ninth columns for which =0.50 are (1.0, 0.5), (2.0, 1.0), (4.8, 2.4), (10.3, 5.2), and (22.9,
11.5).

52
 Table 5: Predicted Serial Correlation in Construction
One-Year Three-Year Five-Year
c2 if c2 if c2 if c2 if c2 if c2 if c2 if c2 if c2 if
=0 = 0.25 = 0.5 =0 = 0.25 = 0.5 =0 = 0.25 = 0.5
lowest
0.68 0.52 0.34 0.40 0.19 0.02 0.17 -0.03 -0.15
c1
lower
0.74 0.56 0.36 0.50 0.24 0.04 0.28 0.01 -0.14
c1
medium
0.78 0.60 0.39 0.57 0.29 0.06 0.37 0.06 -0.13
c1
higher
0.81 0.62 0.37 0.63 0.32 0.05 0.45 0.08 -0.13
c1
highest
0.83 0.63 0.40 0.67 0.33 0.07 0.51 0.09 -0.12
c1

Notes: Within each time horizon over which serial correlation is estimated, 15 simulations were run, corresponding to the 15 (c1, c2)
pairs reported in Table 1. Thus, the (c1, c2) combinations used in the first, fourth, and seventh columns for which =0 are (1.7, 0.0),
(3.5, 0.0), (7.0, 0.0), (14.4, 0.0), and (28.8, 0.0), going from the top to the bottom of the table. The (c1, c2) combinations used in the
second, fifth, and eighth columns for which =0.25 are (1.5, 0.4), (2.7, 0.7), (6.4, 1.6), (13.3, 3.3), and (28.1, 7.0). The (c1, c2)
combinations used in the third, sixth, and ninth columns for which =0.50 are (1.0, 0.5), (2.0, 1.0), (4.8, 2.4), (10.3, 5.2), and (22.9,
11.5).

53
 Table 6: Variance in House Price Changes and Construction Intensity
1, 3, and 5 Year Horizons
House Price Change Variance (millions of $2000)
1 year 3 years 5 years
10th percentile market $14 $69 $183
th
25 percentile market $26 $124 $452
th
50 percentile market $34 $185 $625
75th percentile market $70 $445 $1,170
90th percentile market $209 $1,380 $3,580
Sample mean $83 $484 $1,310
Construction Intensity Variance (millions of units)
1 year 3 years 5 years
th
10 percentile market 2 13 29
th
25 percentile market 2 19 41
50th percentile market 3 26 59
th
75 percentile market 11 84 212
90th percentile market 38 328 760
Sample mean 21 160 417

54
 Table 7: Predicted Variance of Price Changes ($millions)
1, 3, and 5 Year Horizons
One-Year Three-Year Five-Year
c2 if c2 if c2 if c2 if c2 if c2 if c2 if c2 if c2 if
=0 = 0.25 = 0.5 =0 = 0.25 = 0.5 =0 = 0.25 = 0.5
lowest
50 66 64 123 154 148 163 203 200
c1
lower
70 90 97 178 221 239 246 303 337
c1
medium
91 128 138 239 330 360 340 469 520
c1
higher
114 155 166 306 411 442 443 595 647
c1
highest
136 175 181 367 468 488 538 685 718
c1

Notes: Within each time horizon over which serial correlation is estimated, 15 simulations were run, corresponding to the 15 (c1, c2)
pairs reported in Table 1. Thus, the (c1, c2) combinations used in the first, fourth, and seventh columns for which =0 are (1.7, 0.0),
(3.5, 0.0), (7.0, 0.0), (14.4, 0.0), and (28.8, 0.0), going from the top to the bottom of the table. The (c1, c2) combinations used in the
second, fifth, and eighth columns for which =0.25 are (1.5, 0.4), (2.7, 0.7), (6.4, 1.6), (13.3, 3.3), and (28.1, 7.0). The (c1, c2)
combinations used in the third, sixth, and ninth columns for which =0.50 are (1.0, 0.5), (2.0, 1.0), (4.8, 2.4), (10.3, 5.2), and (22.9,
11.5).

55
 Table 8: Predicted Variance of Construction Intensity (millions of units)
1, 3, and 5 Year Horizons
One-Year Three-Year Five-Year
c2 if c2 if c2 if c2 if c2 if c2 if c2 if c2 if c2 if
=0 = 0.25 = 0.5 =0 = 0.25 = 0.5 =0 = 0.25 = 0.5
lowest
15 23 42 101 128 189 221 250 328
c1
lower
7 12 19 48 68 87 110 135 153
c1
medium
3 3 5 21 21 25 49 44 44
c1
higher
1 1 1 8 7 7 19 14 12
c1
highest
0.4 0.3 0.3 3 2 2 7 4 3
c1

Notes: Within each time horizon over which serial correlation is estimated, 15 simulations were run, corresponding to the 15 (c1, c2)
pairs reported in Table 1. Thus, the (c1, c2) combinations used in the first, fourth, and seventh columns for which =0 are (1.7, 0.0),
(3.5, 0.0), (7.0, 0.0), (14.4, 0.0), and (28.8, 0.0), going from the top to the bottom of the table. The (c1, c2) combinations used in the
second, fifth, and eighth columns for which =0.25 are (1.5, 0.4), (2.7, 0.7), (6.4, 1.6), (13.3, 3.3), and (28.1, 7.0). The (c1, c2)
combinations used in the third, sixth, and ninth columns for which =0.50 are (1.0, 0.5), (2.0, 1.0), (4.8, 2.4), (10.3, 5.2), and (22.9,
11.5).

56
 Table 9: The Impact of Greater Local Demand Variability
Five-year Price Change Variance ($millions)
Baseline 2 (=$3.6) Baseline 2 * 1.5 Baseline 2 * 2.0 Baseline 2
* 2.5
th
Market with 75 percentile (c1, c2)
595 893 1,190 1,488
values (c1=13.3, c2=3.3)
Market with 90th percentile (c1, c2)
685 1,028 1,370 1,713
values (c1=28.8, c2=7.0)
Five-Year Quantity Change Variance (millions of units)
2
Baseline (=$3.6) Baseline 2 * 1.66
Market with 25th percentile (c1, c2)
135 224
values (c1=2.7, c2=0.7)
Market with 10th percentile (c1, c2)
250 415
values (c1=1.5, c2=0.4)

Notes: Unless noted explicitly in the table, each simulation reported above using common parameter values for the national sample as
reported in Table 1. In each case, =0.25.

57
 Table 10: Predicted Variance of Price Changes ($millions): Interest Rate Volatility
1, 3, and 5 Year Horizons
One-Year Three-Year Five-Year
H −C Stand. Stand. Stand. Stand. Stand. Stand. Stand. Stand. Stand.
Dev. ( ) Dev. ( ) Dev. ( ) = Dev. ( ) Dev. ( ) = Dev. ( ) = Dev. ( ) Dev. ( ) = Dev. ( ) =
= 0.005 = 0.01 0.02 = 0.005 0.01 0.02 = 0.005 0.01 0.02
$25,000 128 130 136 331 335 352 471 477 501
$50,000 130 136 162 335 352 418 477 501 597
$100,000 136 162 266 352 418 683 501 597 980
$200,000 162 266 681 418 683 1743 597 980 2513

Table 11: Predicted Variance of Construction (millions of units): Interest Rate Volatility
1, 3, and 5 Year Horizons
One-Year Three-Year Five-Year
H −C Stand. Stand. Stand. Stand. Stand. Stand. Stand. Stand. Stand.
Dev. ( ) Dev. ( ) Dev. ( ) = Dev. ( ) Dev. ( ) = Dev. ( ) = Dev. ( ) Dev. ( ) = Dev. ( ) =
= 0.005 = 0.01 0.02 = 0.005 0.01 0.02 = 0.005 0.01 0.02
$25,000 4 4 4 21 22 23 44 45 47
$50,000 4 4 4 22 23 28 45 47 57
$100,000 4 4 7 23 28 46 47 57 97
$200,000 5 7 19 28 46 121 57 97 255

58
 .6
Boulder
Real Appreciation in the 1990s, .1=10%
Salt Lak
Fort Col
Portland
.4

Ogden-Cl
Austin-R
Detroit-
Colorado
Warren-F
Flint
Minneapo San Jose
Peoria Canton-M
BoiseMilwauke
Ci Charlest
Louisvil Tacoma
.2

New Orle Omaha-Co Akron Grand Ra Santa Cr

Baton Ro Lansing- Seattle-
Phoenix-
Spokane
Des Moin
Toledo Nashvill
Tucson
Kansas C Birmingh
Clevelan San FranCambridg
Gary Columbus Bellingh
Cincinna
Tulsa Atlanta-
Raleigh- Chicago- Santa Ro
Salinas
Oklahoma Albuquer Charlott
Jacksonv Lake Cou Napa Essex Co
Boston-Q
Lexingto
St. IndianapSarasota
Loui
Wichita Little R
Houston-San Memphis
Pittsbur
Reno-Spa DaytonWinston-
Greensbo
Tampa-St
Oakland-
Beaumont Anto Columbia Nassau-S
Corpus C Dallas-P
Fort Way Fort Lau
0

Fort Wor Harrisbu

Richmond
Orlando-
Deltona-
Las Vega San Dieg Bridgepo New York
Worceste
West Pal Vallejo- Santa Ba Newark-U
Baltimor Edison
Fresno Washingt
Wilmingt San Luis
Stockton
Bethesda Philadel Providen
Santa An
Buffalo-Sacramen
Oxnard-T
RochesteCamden
-.2

Allentow
Bakersfi Modesto
Riversid Los Ange

Honolulu
-.4

-.5 0 .5 1
Real Appreciation in the 1980s, .1=10%
Figure 1: Real House Price Appreciation in the 1980s and 1990s

59
 .6
Las Vega
Growth in 1990s, .1=10%
.4

Boise Ci Raleigh-

Fort Col
Austin-R
Bellingh Charlott Phoenix-
Atlanta- Orlando-
Ogden-Cl Reno-Spa
Portland Boulder
Columbia
Lak Nashvill
SaltAlbuquer TucsonColorado
.2

Lake Cou
Greensbo
Tacoma Jacksonv
Lexingto
Indianap Dallas-P West Pal
Columbus San Anto
Winston-
Little R Richmond Fort
Seattle- Lau
Sacramen Sarasota Fort
Deltona-
Wor
Spokane Warren-F
Des Moin Grand Ra Fresno
Memphis Houston-
WashingtBakersfiCharlestBethesda Riversid
Birmingh Wilmingt
Baton Ro MinneapoModesto
Louisvil
Cincinna
Omaha-Co
Fort Way Stockton Santa RoSan Luis
HonoluluWichita
Kansas
Baltimor Edison
C Santa An Tampa-St Vallejo-
Gary Akron
Milwauke Allentow
Harrisbu
Lansing- Oxnard-T San Dieg
Chicago- Napa Camden Tulsa Oklahoma
Oakland-
Salinas Corpus C
Flint
Canton-M Rocheste San
Santa Jose
Cr
Peoria New YorkDayton
Clevelan St. LouiCoWorceste
Providen
Cambridg
Nassau-S
Toledo
Newark-U
SanPhiladel
Beaumont Essex
Boston-Q
Fran
Bridgepo
Buffalo-
Pittsbur NewLos Ange
Orle Santa Ba
0

Detroit-

0 .2 .4 .6
Growth in 1980s, .1=10%
Figure 2: Housing Unit Growth in the 1980s and 1990s

60
Figure 3: One-Time Shock

2 4 6 8 10

Population: ------------
Construction: –––––––––
Price: –––––––––

( = 0.1, c1 = 3, c2 = 0.1)

61
 800
Honolulu
Variance in 1 year Price Residual(millions $)
600

San Jose
San Fran
400

Santa Ba An
Santa
Salinas
Bridgepo
Oxnard-T
Santa Cr
Los Ange
200

San Luis Santa

SanRo
Dieg
Oakland-
Nassau-S
New York Napa
Cambridg
Newark-U
Essex Co
Edison Bethesda
Boston-Q Sacramen Boulder
Austin-R Vallejo- Washingt
Worceste SaltRiversid
Lak
Ogden-Cl
San Anto
Providen
Oklahoma Portland Seattle-Reno-Spa
Modesto
Bellingh
Stockton
Colorado
Houston-
Dallas-P
Fort
Las Col
Vega
Beaumont
Corpus
Detroit-
Flint Charlest
Tulsa
MemphisCFort
Greensbo
Winston-
Wichita
LittleRocheste
Gary
Nashvill
Indianap
Grand Charlott
Fort
Louisvil RToledo
Lansing-
RaWayDes Boise
Baton
Raleigh-
Spokane
Canton-M
Peoria
Wor
Akron
Dayton
Omaha-Co New Ci
Ro
Warren-F
Albuquer
Clevelan
Sarasota
Moin Orle Fresno
Fort Lau
Phoenix-
Buffalo-
St. Pittsbur
BirminghHarrisbu
ColumbiaColumbus
Lexingto
Atlanta-
KansasCincinna
C
Allentow
Loui Tacoma
Tucson
Bakersfi
West Pal
Milwauke
Minneapo
Jacksonv Camden
Deltona-
Philadel
Orlando-
Tampa-St
Wilmingt
Baltimor
Richmond Lake Cou
Chicago-
0

50000 100000 150000 200000 250000 300000

1980 Real Price
Figure 4: Variance in 1 yr Price Residual Vs. 1980 Price

62
 400
Variance in 1 year Unit Residual(millions of Units)

Houston-
300

Los Ange
200

Atlanta-
Dallas-P Riversid
Phoenix-
100

San Dieg
Las Vega

Fort Wor Chicago-

Tampa-StAustin-R
New York
Portland Washingt Santa An
San AntoOklahomaOrlando-
Charlott West Pal
Edison Seattle-
Sacramen
Indianap
Tulsa Warren-F
Baltimor
Raleigh-
Columbus Minneapo
Fort Lau Oakland-
Corpus
Beaumont St.
Jacksonv
Memphis
Omaha-Co
Flint
Detroit- Nashvill
Kansas
C
Buffalo-
Wichita
Louisvil Loui esC
Providen
Gary
Akron
Wilmingt
Philadel Peoria
Spokane
Lansing-
Birmingh Toledo
Pittsbur
Rocheste
Harrisbu
Worceste
Deltona-
Fort D
Dayton
Way
Winston-
Grand
Little New
Cincinna
Boise
Baton
Sarasota
Canton-MMoin
Boston-Q
Charlest
Ra
RAllentow
Camden Orle
Ci
Ro
Nassau-SMilwauke
Salt
Bellingh
Tucson
Clevelan
Colorado
TacomaModesto
Albuquer
LexingtoBakersfi
Essex Lake
Lak
Fresno
Cambridg
Ogden-Cl
Co
Stockton Fort CouBoulder
Vallejo-
Col Napa Bethesda
Luis Salinas
Reno-Spa
San Santa Santa
Ro Cr SantaSan
Bridgepo
Oxnard-T Ba Jose San Fran Honolulu
Richmond
Greensbo
Columbia Newark-U
0

50000 100000 150000 200000 250000 300000

1980 Real Price
Figure 5: Variance in 1 yr Unit Residual Vs. 1980 Price

63
 Appendix 1: Proofs of Propositions

Proof of Proposition 1: We use the change of variables I (t ) = m(t ) + Iˆ(t ) ,

N (t ) = n(t ) + Nˆ (t ) , and H (t ) = z (t ) + Hˆ (t ) . Substituting in our definitions of Iˆ , N̂ , and
Ĥ , we reduce the core pricing equation
rC E ( H (t + 1))
H (t ) = D + qt + x(t ) − αN (t ) + + t to
1+ r 1+ r
E ( z (t + 1))
z (t ) = x(t ) − αn(t ) + t , (*)
1+ r
the optimality condition for production C + c0 (t + 1) + c1 I (t + 1) + c 2 N (t ) = Et ( H (t + 1))
to
c1 m(t + 1) + c 2 n(t ) = Et ( z (t + 1)) , (**)
and the defining equation I (t + 1) = N (t + 1) − N (t ) to
m(t + 1) = n(t + 1) − n(t ) . (***)
We seek functions n, z, and m that satisfy the starred equations.

α c2
Define u ≡ and v ≡ ; 0 ≤ u and 0 ≤ v < 1 by the conventions in force. Then
c1 c1

φ=
1
2
(
2 + r + (1 + r )u − v − r 2 + v 2 + 2(1 + r )(2 + r )u + (1 + r ) 2 u 2 + 2v(r − (1 + r )u ) )
and
1
φ=
2
(
2 + r + (1 + r )u − v + r 2 + v 2 + 2(1 + r )(2 + r )u + (1 + r ) 2 u 2 + 2v(r − (1 + r )u ) . )
Because 0 ≤ v < 1 , the expression under the radical is positive. Note that
φ + φ = 2 + r + (1 + r )u − v > (1 + r )(1 + u ) > 0 and φφ = (1 + r )(1 − v) > 0 , so φ , φ > 0 .
Also, r 2 + v 2 + 2(1 + r )(2 + r )u + (1 + r ) 2 u 2 + 2v(r − (1 + r )u ) >
2(1 + r )(2 + r )u + (1 + r ) 2 u 2 − 2(1 + r )u > 2(1 + r )(1 + r )u + (1 + r ) 2 u 2 > 1 + r ,
so
1
φ> (1 + r + (1 + r )u + 1 + r ) ≥ 1 + r > 1 ,
2
which in turn gives
(1 + r )(1 − v)
0≤φ = < 1 − v ≤ 1.
φ

Now define n by the difference equation

1+ r
n(t ) − φn(t − 1) = Et −1 ( x(t )) . (1)
c1 (φ − δ )
A unique solution n exists because φ > 1 > δ ensures φ − δ ≠ 0 and because | φ | < 1
allows us to solve for n explicitly as

64
 1+ r ∞
n(t ) = φ i Li Et −1 ( x(t )) (2),
c1 (φ − δ ) i =0
where L denotes the lag operator. Now that we have defined n, we set
1 α (1 + r )
z (t ) ≡ x(t ) + Et ( x(t + 1)) − n(t ) (3)
φ −δ 1+ r −φ
and
1+ r
m(t + 1) ≡ Et ( x(t + 1)) − (1 − φ )n(t ) . (4)
c1 (φ − δ )
With these choices for z and m, (*) reduces to
α
(n(t + 1) − φn(t ) ) = φ − 1 − r Et ( x(t + 1)) , (5)
1+ r −φ (φ − δ )(1 + r )
which by (1) is equivalent to
u (1 + r ) φ −1− r
Et ( x(t + 1)) = Et ( x(t + 1)) ,
(1 + r − φ )(φ − δ ) (φ − δ )(1 + r )
which is true, as one sees from cross-multiplying the coefficients and using the
previously established formulas for the product and sum of φ and φ . (**) reduces to
α (1 + r ) 1+ r −φ
n(t + 1) − (c1 (1 − φ ) − c2 ))n(t ) = φ − 1 − r Et ( x(t + 1)) ,
1+ r −φ α (1 + r ) φ −δ
which is equivalent to (5), and thus true, because
1+ r −φ
φ= (c1 (1 − φ ) − c2 ) ) = (1 + r − φ )(1 − φ − v) ,
α (1 + r ) u (1 + r )
itself evident from cross-multiplying and using the fact that φ satisfies the quadratic
equation y 2 − (2 + r + (1 + r )u − v) y + (1 + r )(1 − v) = 0 . Finally, (***) reduces to (1).
This shows that our choices for n, z, and m solve the starred equations. To recover
Proposition 1, we use I (t ) = m(t ) + Iˆ(t ) , N (t ) = n(t ) + Nˆ (t ) , and H (t ) = z (t ) + Hˆ (t ) .
[The result then follows from Et ( x(t + 1)) = δx(t ) + θε (t ) .]

Proof of Proposition 2: First note that by induction on i ≥ 1 ,

E t ( x(t + i )) = E t (δx(t + i − 1) + θε (t + i − 1)) = δEt ( x(t + i − 1)) = δ i −1 Et ( x(t + 1)) , so from
(2),
1+ r ∞
Et (n(t + j )) = Et φ i Li Et + j −1 ( x(t + j ))
c1 (φ − δ ) i =0
j −1
1+ r ∞
= φ L Et + j −1 ( x(t + j )) +
i i
φ i Et ( x(t + j − i))
c1 (φ − δ ) i= j i=0
j −1
1+ r ∞
= φ j φ i Li Et −1 ( x(t )) + φ iδ j −1− i
Et ( x(t + 1))
c1 (φ − δ ) i =0 i =0

1+ r φ j −δ j
= φ j n(t ) + E t ( x(t + 1)) . (6)
c1 (φ − δ ) φ − δ

65
Using (4) and (6), we next find that
1+ r
Et (m(t + j )) = Et ( x(t + j )) − (1 − φ ) Et (n(t + j − 1))
c1 (φ − δ )
1+ r φ j −1 − δ j −1
= δ j −1
− (1 − φ ) Et ( x(t + 1)) − φ j −1 (1 − φ )n(t )
c1 (φ − δ ) φ −δ
1+ r δ j −1 (1 − δ ) − φ j −1 (1 − φ )
= Et ( x(t + 1)) − φ j −1 (1 − φ )n(t ) . (7)
c1 (φ − δ ) φ −δ
Finally, using (**), (6), and (7), we get
Et ( z (t + j )) = c1 Et (m(t + j )) + c 2 Et (n(t + j − 1))
1+ r δ j −1
(1 − δ ) − φ j −1 (1 − φ ) φ j −1 − δ j −1
= c1 + c2 Et ( x(t + 1))
c1 (φ − δ ) φ −δ φ −δ
− φ j −1 (c1 (1 − φ ) − c2 )n(t )
1+ r δ (1 − v − δ ) − φ j −1 (1 − v − φ )
j −1
= Et ( x(t + 1)) − c1φ j −1 (1 − v − φ )n(t ) . (8)
φ −δ φ −δ
To recover Proposition 2, we use I (t ) = m(t ) + Iˆ(t ) , N (t ) = n(t ) + Nˆ (t ) ,
and H (t ) = z (t ) + Hˆ (t ) with equations (3), (8) , (4), (7), and (6).

Proof of Proposition 3: Given the hypotheses, we have

x(t ) = δx(t − 1) + θε (t − 1) + ε (t ) = ε (t ) > 0 , Et x(t + 1) = δx(t ) + θε (t ) = (δ + θ )ε (t ) > 0 ,
and n(t ) = 0 , so from (3) and (4) we deduce that z (t ) > 0 and m(t + 1) > 0 : prices and
investment will initially be higher than steady state levels. By assumption, v = 0 , so by
(7) and (8), each of expected time t + j construction, Et (m(t + j )) , and expected time
t + j price, Et ( z (t + j )) , is negative if and only if
δ
(1 − δ ) − φ j −1 (1 − φ )
j −1
< 0 . (9)
φ −δ
If φ > δ , then (9) holds if and only if
j −1
1− δ φ
< ,
1−φ δ
which holds for sufficiently large j because φ / δ > 1 . If φ < δ , then (9) holds if and
only if
j −1
1−φ δ
< ,
1− δ φ
which holds for sufficiently large j because δ / φ > 1 . If φ = δ , then we reduce (9) to
j −1 j−2
(φ − δ ) φ iδ j −1−i
− φ iδ j − 2 −i

φ − δ − (φ − δ
j j j −1 j −1
) i =0 i =0
0< = = φ j −2 ( jφ − j + 1) ,
φ −δ φ −δ

66
which holds for sufficiently large j because 0 ≤ φ < 1 . This shows that there exists j *
such that for all j > j * , time t expected values of time t + j construction and housing
prices will lie below steady state levels. When ε (t ) < 0 , we swap > and < to recover the
symmetric case.

Proof of Proposition 4: By assumption, n(0) = 0 , so from (1), we have

(1 + r )δ
n(1) = ε (0) . (10)
c1 (φ − δ )
From (***), m(1) = n(1) , and from (4) and (10),
(1 + r )δ
m ( 2) = ((δ + φ − 1)ε (0) + ε (1)) .
c1 (φ − δ )
q (1 + r ) − rc0
By definition, I (t ) = + m(t ) , so
rc 2 + α (1 + r )
2 2
1+ r (1 + r )δ
Cov(I (2), I (1) ) = Var (q ) + (δ + φ − 1)Var (ε ) ,
rc 2 + α (1 + r ) c1 (φ − δ )
which is positive if and only if
δ (rc 2 + α (1 + r ) )
2
Var (q )
> (1 − δ − φ ) .
Var (ε ) c1 (φ − δ )
From (3) and (10),
φ
z (0) = ε (0) (11)
φ −δ
and
1 u (1 + r ) 2
z (1) = φ− δε (0) + φ ε (1) . (12)
φ −δ 1+ r −φ
To compute z (2) , we use (1) and (10) to get
(1 + r )δ
n ( 2) = ((φ + δ )ε (0) + ε (1)) ,
c1 (φ − δ )
which yields via (3) that
1 u (1 + r ) 2 (φ + δ ) u (1 + r ) 2
z ( 2) = φδ − δε (0) + φ − δε (1) + φ ε (2) . (13)
φ −δ 1+ r −φ 1+ r −φ
(1 + r )(αc0 + qc 2 )
By definition, H (t ) = Hˆ (0) + t + z (t ) , so from (11), (12), and (13),
rc 2 + α (1 + r )
2
(1 + r )c 2
Cov(H (2) − H (1), H (1) − H (0) ) = Var (q )
rc 2 + α (1 + r )
u (1 + r ) 2 δ u (1 + r ) 2 δ
+ (1 − δ )φ (1 − δ − φ ) + (1 − δ + δ 2 )φ
1+ r −φ 1+ r −φ
− Var (ε ) ,
(φ − δ ) 2

67
which is negative if and only if
2
rc 2 + α (1 + r ) Var (q )
Ω > .
(1 + r )c1c 2 (φ − δ ) Var (ε )

68
 Appendix II: The Contribution of Taxes Local Demand Variance

Data on the average tax rate paid each year in each state was matched to our
metropolitan areas using files from the NBER’s TaxSim web page. We then multiplied
our income numbers by one minus the average tax rate, and calculated new values of δ ,
θ and 2 for this adjusted after-tax income measure. The new “after-tax” values of the
three parameters are very similar to those used in our simulations: =0.87, =0.18, and
2
=$3.3 million. The latter is 92 percent of the $3.6 million figure obtained without any
adjustment for taxes. Hence, correcting for taxes creates an eight percent reduction in the
variance and almost no change in the other parameters. Consequently, we conclude that
including state level tax rates does not offer any hope of explaining the particularly high
volatilities.

Appendix III: The Contribution of Crime to Local Demand Variance

We began by drawing on the hedonic literature on the costs of crime. The range
of estimates of the elasticity of property value with respect to the violent crime rate ran
from 0.05 to 0.15.38 To turn these housing price elasticities into estimates of the impact
of crime on the flow of utility measured in dollar units, we multiply the elasticity by the
average housing price per crime to obtain a relationship between the price of housing and
the level of crime. We then followed our model and multiplied this figure by r/(1+r) to
generate an estimate of the impact of crime on the flow of utility measured in dollars.

Using this method, our elasticity range from 0.05 to 0.15 implies that the impact
of violent crime on the flow of well-being ranges from $35 to $105. The upper bound
estimate of $105 dollars implies that, if the violent crime rate in a city increases from 12
violent crimes per 1,000 inhabitants (the national mean) to 24 violent crimes per 1,000
inhabitants, then this is equivalent to an income loss of about $1,260 dollars, which we
believe is a reasonable result.

We then used this upper bound impact to adjust the underlying BEA real income
variable and δ , θ and 2. As with taxes, crime had little impact on the volatility of the
local income shock. Specifically, there is only a 1.4 percent greater shock variance when
controlling for crime.39 While the crime data is far from perfect for our purposes, this
exercise leads us to conclude that variation in local amenities will explain little of the
high variance price change or construction markets.

38
See Thaler (1978) for the lower bound estimate and Schwartz, Susin, and Voicu (2003) for the upper
bound number.
39
We were able to obtain crime data for the major cities of 105 of our 115 metropolitan areas. The ARMA
estimates of and are virtually unchanged depending upon whether income is adjusted for crime in these
105 markets. As noted, the variability of the ‘after-crime’ income shock is marginally higher.

69