Housing and the Business Cycle

Morris Davis ReturnBuy and Federal Reserve Board Jonathan Heathcote Duke University and Stern School, NYU

November 5, 2001

Abstract In the United States, the percentage standard deviation of residential investment is more than twice that of non-residential investment. GDP, consumption, and both types of investment all co-move positively. At the industry level, output and hours worked in construction are more than three times as volatile as in services, and output and hours co-move positively across sectors. We reproduce all these facts in a multi-sector growth model with the following characteristics: dierent nal goods are produced using dierent proportions of the same set of intermediate inputs, construction is relatively labor intensive, residential investment is relatively construction intensive, and housing depreciates much more slowly than business capital. Previous empirical work exploring the determinants of residential investment is re-examined in light of the models equilibrium relationship between residential investment, house prices, and the rental rate on capital. Keywords: Residential investment; House prices; Multi-sector models JEL classication: E2; E3; R3
Corresponding author: New York University, Stern / Economics, 44 W. 4th Str. 7-180, New York NY 10012. Email: heathcote@econ.duke.edu. We thank seminar participants at Duke University, the Board of Governors of the Federal Reserve, NYU, North Carolina State University, the Stockholm School of Economics, the Society for Economic Dynamics Meetings in Costa Rica 2000, the Macro / Labor Conference at the University of Essex 2000, the Bank of Canada Conference on Structural Macro-Models 2000, and the NBER / Cleveland Fed Economic Fluctuations and Growth Workshop in Philadelphia 2000. Heathcote thanks the Economics Program of the National Science Foundation for nancial support.

1. Introduction
The size of the housing stock is large: similar in value to private non-residential structures and equipment combined, similar in value to annual GDP, and three times as large as the total stock of all other consumer durables. Although housing is typically considered part of the economys capital stock in one-sector models (see, for example, Cooley and Prescott in Cooley 1995), there are good reasons for distinguishing between housing on the one hand and non-residential structures and equipment on the other. A conceptual reason is that to the extent that housing is used for production, it is for production at home that for the most part is not marketed and not measured in national accounts. A practical reason for making the distinction is to attempt to account for dierences between the business cycle dynamics of residential and non-residential investment. In particular, residential investment is much more volatile than business xed investment, and strongly leads the cycle whereas non-residential investment lags. The primary goal of this paper is to examine the extent to which a neoclassical multi-sector stochastic growth model can account for the dynamics of residential investment. A model economy with explicit microfoundations is calibrated using industry-level data. The production structure is such that data on the empirical counterpart to each variable in the model is available.1 There are two nal goods sectors. One produces the consumption / investment good, while the second produces the residential investment good. Final goods rms use three intermediate inputs produced in the construction, manufacturing and services sectors. These intermediate inputs are in turn produced using capital and labor rented from a representative household. Productivity is stochastic as a result of exogenous sector-specic labor-augmenting technology shocks. The representative household maximizes expected discounted utility over percapita consumption, housing services and leisure. Each period it decides how much to work and consume, and how to divide savings between physical capital and housing, both of which are perfectly divisible. There is a government which levies stochastic taxes on capital and labor. In this framework, changes in the
This is not the case in the home production literature in which inputs to and productivity within the home sector are imperfectly observed.
1

tax rate on capital income will aect the relative returns to saving in the forms of capital and housing, since the implicit rents from owner-occupied housing are untaxed. We use the model to ask three main questions. First, to what extent can a simple multi-sector model driven primarily by technology shocks account for the business cycle facts relating to residential investment, non-residential investment and house prices? Second, given the methodology for identifying productivity shocks, to what extent can the model account for United States economic history over the post War period? Third, can we use structural equilibrium relationships implied by the model to reinterpret previous empirical work addressing the determinants of residential investment? Relation to literature The model is specically designed to explore the unusual business cycle dynamics of residential investment. In most previous multi-sector real business cycle models it is not possible to focus squarely on residential investment or house prices since housing is typically not distinguished from other consumer durables (see, for example, Baxter 1996 or Hornstein and Praschnik 1997).2 Failure to model housing explicitly may be important because there are important respects in which housing diers from cars or televisions. First, housing is a much better store of value since houses depreciate at a rate of only 1.6 percent per year, compared to 21.4 percent for other durables. Second, the production technology for producing houses is more construction intensive and thus more labor intensive than the technology for producing consumer durables. We shall nd that these two features are crucial in accounting for the dynamics of residential investment. Multi-sector models typically have trouble replicating the strong positive comovement between consumption of durable goods and residential investment on the one hand with consumption of nondurables and business xed investment on the other. Furthermore labor supply tends to co-move negatively across sectors in models, while hours worked co-move positively in dierent sectors of the U.S.
One exception is an exploratory paper by Storesletten 1993 who found that the process for sector-specic shocks can not account for the fact that residential investment leads the cycle. In a recent paper, Edge 2000 considers the dierential eects of monetary shocks on residential and structures investment in a multi-sector model with sticky prices.
2

economy. The reasons for these failures are simple. First, the models incorporate a strong incentive to switch production between sectors in response to sector specic productivity shocks. Second, even if shocks are perfectly correlated across sectors, there is an incentive to increase output of new capital prior to expanding anywhere else, since additional capital is a pre-requisite for eciently expanding output in other sectors.3 Previous authors have also had trouble accounting for the fact that residential investment (or residential investment plus spending on consumer durables) is more volatile than business investment.4 We shall also be interested in assessing the extent to which our novel production structure and calibration methodology can address both the co-movement and relative volatility puzzles. An alternative to our approach of introducing the housing stock directly in the utility function would have been to assume that housing and non-market time are combined according to a home production technology to give a non-marketed consumption good. Greenwood, Rogerson and Wright (1995 p.161) show that these alternative modelling approaches are closely related. In particular, given
Various xes have been proposed to solve the co-movement problem. Fisher 1997 assumes a non-linear function for transforming output into non-durable consumption goods, new consumer durables, and new physical capital. Since in the limit dierent goods must be produced in xed proportions it is easy to see how this approach can resolve the co-movement problem. Baxter 1996 estimates a high correlation between productivity growth across sectors and also introduces sectoral adjustment costs for investment, which dampens the incentive to increase investment in the sector producing capital while reducing investment elsewhere. Chang 2000 combines adjustment costs with substitutability between time and durable goods; thus when households work more in periods of high productivity they also demand more durables. Gomme, Kydland and Rupert 2001 introduce time-to-build in the sector producing new market capital, which has a similar eect to introducing adjustment costs in that it dampens the investment boom in the capital-producing sector and allows investment to rise in all sectors simultaneously. Boldrin, Christiano and Fisher 2001 nd that a combination of limited labor mobility across sectors and a habit in consumption can generate co-movement in hours worked across sectors. 4 Fisher 1997 nds that none of his specications give household investment more volatile than business investment. In Baxters 1996 model, consumption of durables (which includes residential investment) is too smooth and is less volatile than business xed investment in either sector. For all but one of the parameterizations they consider, Gomme, Kydland and Rupert 2001 nd market investment to be more volatile than home investment, contrary to the pattern in the data.
3

(i) a Cobb Douglas technology for producing the home good from capital and labor, and (ii) log-separable preferences over leisure, market consumption and home consumption, the home production model has a reduced form in which only market consumption, market hours and the stock of home capital enter the utility function. The benchmark calibration adopted by Greenwood and Hercowitz (1991) satises these functional-form restrictions, which suggests that our model is closely related to theirs. The main dierence is that contrary to Greenwood and Hercowitz, we do not assume a single (market) production technology. In the results section we systematically compare and contrast the two economies. All the papers described above are based on representative agent economies. We remain within the representative agent framework since our focus is on understanding business cycle dynamics. Models with incomplete markets environments typically focus on steady states (see for example Platania and Schlagenhauf 2000 or Fernandez-Villaverde and Krueger 2001).5 While frictions such as poorly functioning rental and mortgage markets are likely important in accounting for cross-sectional issues (such as life-cycle consumption / savings patterns or heterogeneity in asset holding portfolio choices) it is not obvious that they are important for housing dynamics at the aggregate level.6 In any case it would appear to be sensible to ask whether a representative agent model with complete asset markets is broadly able to capture observed aggregate dynamics before turning to richer environments. The predominant theory of residential investment in the real estate literature (see, for example, Topel and Rosen 1998 or Poterba 1984 and 1991) may be summarized as follows. There is a price for residential housing that clears the market given the existing stock of homes. Developers observe this price and build
Diaz-Gimenez, Prescott, Fitzgerald and Alvarez 1992 and Ortalo-Magne and Rady 2001 both consider the eects aggregate shocks in model economies which incorporate indivisibilities in housing, credit constraints, and life cycle dynamics. However the focus on Diaz-Gimenez et. al. is on monetary policy rather than the business cycle dynamics, while Ortalo-Magne and Rady assume a constant stock of housing, and therefore have nothing to say about residential investment. 6 Krusell and Smith 1998 and Rios-Rull 1994 study the aggregate dynamics of economies in which households face large amounts of uninsurable idiosyncratic risk. They nd that they are virtually identical to those observed when markets are complete.
5

more houses the larger is the gap between this price and the cost of construction, which is often proxied by the real interest rate. Changes in prices are primarily driven by demand-shocks which aect future expected rental rates. Regressions of residential investment on house prices and the real interest rate have typically generated positive and strongly signicant estimates for the coecient on the house price term. This has been presented as evidence in support of the demandshock driven view of residential investment. The rst order conditions of the model developed imply a particular equilibrium relationship between residential investment and new house prices. This relationship indicates a negative relationship between residential investment and house prices, holding other variables constant. However, the equation implied by the model also includes the marginal product of capital and terms involving sectoral capital stocks and the existing stock of houses. Using simulated data from the model we measure the size of omitted variable bias on the house price term when estimating residential investment equations of the type used in previous empirical work. Findings The calibrated model economy is found to account for the following facts: (i) the percentage standard deviation of residential investment (at business cycle frequencies) is almost three times that of non-residential investment, (ii) hours worked and output are most volatile in the construction sector and least volatile in the services sector, (iii) consumption, non-residential investment, residential investment and GDP all co-move positively, (iv) hours worked and output in the construction, manufacturing, and services sectors all co-move positively, (v) house prices are procyclical and positively correlated with residential investment, (vi) residential investment (weakly) leads the cycle, and (vi) the percentage standard deviation of GDP is around two. Among the moments we compute, there are two respects in which the model performs poorly. First the model cannot account for the observed volatility in house prices. Second the model does not reproduce the observation that non-residential investment lags the cycle. Note that the model resolves both the co-movement and relative volatility puzzles. These successes are attributable to certain characteristics of the calibrated

production technologies. First, while our Solow residual estimates suggest only moderate co-movement in productivity shocks across intermediate goods sectors, co-movement in eective productivity across nal goods sectors is amplied by the fact that both nal goods sectors use all three intermediate inputs, albeit in dierent proportions. A second feature we emphasize is that construction and hence residential investment are relatively labor intensive. Residential investment volatility rises when capitals share in construction is reduced because following an increase in productivity less additional capital (which takes time to accumulate) is required to eciently increase the scale of production. A third important aspect of the calibration is that the depreciation rate for housing is much slower than that for capital. This increases the relative volatility of residential investment, since it increases the incentive to concentrate production of new houses in periods of high productivity. When we ask the model how well it can account for observed U.S. history, we nd that it does an good job in accounting for the observed time paths for GDP, consumption and both non-residential and residential investment. Estimating the correct and mis-specied equations for residential investment on simulated model output, we nd that for the mis-specied equation (loosely based on previous empirical work) the point estimate for the house price coecient is positive. Recall that this co-ecient is negative in the correctly specied equation, reecting the fact that residential investment is driven by supply-side productivity shocks. Thus we argue that care should be taken in drawing inferences from previous estimation results as to the appropriate way to model residential investment.

2. The Model
In each period t the economy experiences one event et E. We denote by et the history of events up to and including date t. The probability at date 0 of any particular history et is given by (et ). The population grows at a constant rate . In what follows all variables are in per-capita terms. A representative household supplies homogenous labor and rents homogenous capital to perfectly competitive intermediate-goods-producing rms. These rms 6

allocate capital and labor frictionlessly across three dierent technologies. Each technology produces a dierent good which we identify as construction, manufactures and services, and which we index by the subscripts b, m and s respectively. The quantities of each good produced after history et are denoted xi (et ), i {b, m, s} . Output of intermediate good i is a Cobb-Douglas function of the quantity of capital ki (et ) and labor ni (et ) allocated to technology i : xi (et ) = ki (et )i zi (et )ni (et )
1i

(2.1)

Note that the three production technologies dier in two respects. First, the shares of output claimed by capital and labor, determined by capitals share i , dier across sectors. For example, our calibration will impose b < m , reecting the fact that construction is less capital intensive than manufacturing. Second, each sector is subject to exogenous sector-specic labor-augmenting productivity shocks. We let zi (et ) denote labor productivity in sector i after history et . Let pi (et ) denote the price of good i in units of the nal consumption good delivered after history et . Let w(et ) and r(et ) be the wage and rental rate on capital measured in the same units. The intermediate rms static maximization problem after history et is
{ki (et ),ni (e )}i{b,m,s}

max t

Xn
i

pi (et )xi (et ) r(et )k(et ) w(et )n(et )

subject to eq. 2.1 and to the constraints7 kb (et ) + km (et ) + ks (et ) k(et ), nb (et ) + nm (et ) + ns (et ) n(et ),
n o

ki (et ), ni (et )

i{b,m,s}

0.

The law of motion for intermediate rms productivities has a deterministic and a stochastic component. We assume a constant trend growth rate for each
Note that while capital in sector i at date t is chosen at date t, aggregate capital in place at date t is chosen at t 1, in the standard way. Thus in equilibrium k(et ) does not depend on the shocks realized at date t.
7

technology, but permit the rate to vary across sectors. Let gzi denote the constant gross trend growth rate of labor productivity associated with technology i. The goods produced by intermediate goods rms are used as inputs by rms producing two nal goods: a consumption / capital investment good and a residential investment good. Final goods rms are perfectly competitive and allocate the intermediate goods freely across two Cobb-Douglas technologies. We use the subscript c to index the consumption / capital investment good and h to index residential investment (RESI). Let yj (et ), j {c, h} denote the quantity of nal good j produced after history et using quantities bj (et ), mj (et ) and sj (et ) of the three intermediate inputs. Thus yj (et ) = bj (et )Bj mj (et )Mj sj (et )Sj j {c, h} (2.2)

where Bj , Mj and Sj = 1 Bj Mj denote the shares of construction, manufactures and services respectively in sector j. The technology used to produce the consumption good diers from that used to produce RESI with respect to the relative shares of the three intermediate inputs,. Thus, for example, our calibration leads us to set Bh > Bc , reecting the fact that the housing sector is relatively construction-intensive. We normalize the price of the consumption good after any history to 1, and let pr (et ) denote the price of RESI. The nal goods rms static prot maximization problem after history et is max t yc (et ) + pr (et )yh (et )
X n

{bj (et ),mj (e ),sj (et )}j{c,h}

pi (et )xi (et )

i{b,m,s}

subject to eq. 2.2 and to the constraints bh (et ) + bc (et ) b(et ), mh (et ) + mc (et ) m(et ),
n

sh (et ) + sc (et ) s(et ), bj (et ), mj (et ), sj (et )

o
j{c,h}

0.

There is a government which raises revenue by taxing labor income at rate n (et ) and capital income (less a depreciation allowance) at rate k (et ). Tax revenues are divided between non-valued government spending on the consumption / investment good denoted g(et ), and lump-sum transfers to households denoted (et ). Government consumption is assumed to be a xed fraction of output of the consumption / investment sector. Tax rates, however, are stochastic, and follow an autoregressive process jointly with sector-specic productivity shocks:
0

b z (et+1 ) =

b Tildas are used to indicate that each element of z (et ) records the deviation from t , for example, trend value at e e log zb (et ) = log zb (et ) t log gzb log zb,0 .

b = B z (et ) + b(et+1 ),

e e e e e log zb (et+1 ), log zm (et+1 ), log zm (et+1 ), k (et+1 ), n (et+1 )

et+1 consistent with et .

The 5 5 matrix B captures the deterministic aspect of how shocks are trans mitted through time, and b(et+1 ) is a 5 1 vector of shocks drawn independently through time from a multivariate normal distribution with mean zero and variance-covariance matrix V. The representative household derives utility each period from per-capita household consumption c(et ), from per-capita housing owned h(et ), and from leisure. The size of the household grows at the population growth rate . The amount of per-household-member labor supplied plus leisure cannot exceed the period endowment of time, which is normalized to 1. Period utility per household member after history et is given by c(et )c h(et )h (1 n(et ))1c h U (c(e ), h(e ), (1 n(e )) = 1
t t t

where c and h determine the relative weights in utility on consumption, housing and leisure. At date 0, the expected discounted sum of future period utilities for the representative household is given by
X X

t=0 et E t

(et ) t t U (c(et ), h(et ), (1 n(et )) 9

(2.3)

where < 1 is the discount factor.8 Note that the ow of utility that households receive from occupying housing they own will constitute an implicit rent that is untaxed. Households divide income between consumption, spending on new capital that will be rented out next period and has price pn (et ), and spending on new housing k that will be occupied next period and has price pn (et ). Once rents have been paid h and utility from housing delivered, households sell old capital and housing to a new type of rm, which we label the adjustment rm, at prices po (et ) and po (et ).9 k h The depreciation rate for capital is given by k (houses depreciate at rate h ). Thus the household budget constraint is10 c(et ) + pn (et )k(et+1 ) + pn (et )h(et+1 ) po (et )k(et ) po (et )h(et ) (2.4) k h k h = Adjustment rms combine old capital and housing with the capital investment good and the residential investment good (purchased from nal goods rms) to produce new capital and new housing. Because the capital investment good and the consumption good are perfect substitutes in production, they must have the same price in equilibrium. Consumption is the numeraire good, and this price is therefore unity. The static maximization problem of an adjustment rm is pn (et )k(et+1 ) + pn (et )h(et+1 ) max k h ih (et ),ik (et ),k(et ),h(et ) pr (et )ih (et ) ik (et ) po (et )k(et ) po (et )h(et ) k h subject to the production technologies k(e
8 9

1 n (et ) w(et )n(et ) + r(et )k(et ) k (et ) r(et ) k k(et ) + (et ).

t+1

) = (1 k )k(e ) + k

ik (et ) k(et ), k(et )

(2.5)

Note that the household weights per-household-member utility by the size of the household. Note that this is slightly dierent than the standard formulation, in which households purchase investment rather than new capital. However, equilibrium allocations here are identical to those under the standard description; we adopt this particular decentralization because it is convenient for pricing capital and housing at dierent moments within the period. 10 The population growth rate multiplies variables dated t + 1 because all variables are in per-capita terms.

10

h(e

t+1

) = (1 h )h(e ) + h

ih (et ) h(et ). h(et )

(2.6)

The production technology for new capital and new houses incorporates a penalty for deviating from steady state investment rates. The functions k and h are increasing, concave, and satisfy the following properties in steady state: k

ik k

ik = and h k ik k
ik k

ih h

ih , h

(2.7)

0 k

k =

00 k 0 k

ik k

=
ik k

0 h

ih h

= 1,

ih h ih h ih h

(2.8)

The rst two properties imply zero steady state adjustment costs, while the third implies that the elasticity of the price of new capital with respect to the investment rate (ik /k) is equal across sectors. The representative household chooses k(et+1 ), h(et+1 ), c(et ) and n(et ) for all et and for all t 0 to maximize expected discounted utility (eq. 2.3) subject to a sequence of budget constraints (eq. 2.4) and a set of inequality constraints c(et ), n(et ), h(et ), k(et ) 0 and n(et ) 1. The household takes as given a complete set of history dependent prices, tax rates and transfers po (et ), po (et ), pn (et ), pn (et ), k h k h t ), w(et ), (et ), (et ), (et ), unconditional probabilities over histories given r(e n k t ), and the initial stocks of capital and housing. by (e 2.1. Denition of equilibrium An equilibrium is a set of prices, taxes and transfers pi (et )i{b,m,s} , po (et ), po (et ), k h pn (et ), pn (et ), pr (et ), r(et ), w(et ), k (et ), n (et ), (et ) for all et and for all t 0 k h such that when households solve their problems and rms prot maximize taking these prices as given all markets clear and the governments budget constraint is satised. Market clearing for nal goods implies that c(et ) + ik (et ) + g(et ) = yc (et ), 11

= h =

00 h 0 h

(2.9)

ih (et ) = yh (et ). Market clearing for intermediate goods implies that bh (et ) + bc (et ) = xb (et ), mh (et ) + mc (et ) = xm (et ), sh (et ) + sc (et ) = xs (et ). Market clearing for capital and labor implies that kb (et ) + km (et ) + ks (et ) = k(et ), nb (et ) + nm (et ) + ns (et ) = n(et ). Since the government cannot issue debt, the government budget constraint is satised when (et ) + g(et ) = n (et )w(et )n(et ) + k (et ) r(et ) k(et ). 2.2. Equilibrium prices The rst order conditions for the intermediate goods rms problem are as follows (suppressing history dependence). With respect to capital by sector r = pi i ki
(i 1)

(2.10) (2.11) (2.12)

(zi ni )1i

i {b, m, s} .

(2.13)

With respect to labor by sector

w = zi pi (1 i )ki i (zi ni )i

i {b, m, s} .

(2.14)

The rst order conditions for the nal goods rms problem are as follows. With respect to construction goods, manufactures and services by sector pb = pm = Bc yc B y pr = h h , bc bh Mc yc M y pr = h h , mc mh 12 (2.15) (2.16)

Sc yc Sh yh pr = . (2.17) sc sh The rst order conditions for the adjustment rms problem are as follows. With respect to non-residential and residential investment ps = pn k = 1/0 k

ik k

, ih . h

(2.18) (2.19)

pn (et ) = pr (et )/0 h h po ik k n = (1 k ) + k pk k

With respect to old capital and old housing

0 k

ik k

ik , k

(2.20)

ih ih ih po h 0 . (2.21) h n = (1 h ) + h ph h h h It is straightforward to show that all rms, including adjustment rms, make zero prots in every state. From eqs. 2.14 to 2.17 we derive the following expression for the price of residential investment log pr = {1 + (Bc Bh )(1 b ) log zb + (Mc Mh ) (1 m ) log zm (2.22) + (Sc Sh ) (1 s ) log zs } + [(Bc Bh )b + (Mc Mh )m + (Sc Sh ) s ] (log km log nm ) This expression indicates that a positive productivity shock in sector i tends to reduce the relative price of residential investment if residential investment is relatively intensive in input i. The size of the relative price change is increasing in the dierence in factor intensities across the two nal goods technologies, and is increasing in the labor intensity of sector i.11 From eq. 2.19 it is clear that the price of new housing is closely related to the price of residential investment. However, the two prices are not identical in the presence of adjustment costs. For example, if the residential investment rate is above average, an additional unit of new housing is more expensive than an additional unit of residential investment.
To the extent that a productivity shock aects equilibrium sectoral capital-output ratios, there is a second eect on the relative price of residential investment via the last term in eq. 2.22. This last eect disappears if b = m = s .
11

13

2.3. Rental markets, mortgage markets, and house prices In the description of the model economy above we abstract from many aspects of housing that have attracted attention, such as the existence of rental markets, the market for mortgages, and the deductability of mortgage interest payments. Markets in our model are complete, however, so it is straightforward to imagine rental or mortgage markets, and to price whatever is traded in these markets. For example, one could imagine that each household rents out some or all of the housing they own to its neighbor, thereby breaking the link between ownership and occupation and establishing a rental market. Given equilibrium allocations (which are independent of the size of this hypothetical rental market) the rental rate for housing, denoted q, is such that households are indierent to renting a marginal unit of housing: q= Uh (c, h, (1 n)) . Uc (c, h, (1 n))

If rental income is taxed at a positive rent, households will strictly prefer owner-occupation to owner-renting since the implicit rents from owner-occupation are untaxed. If rental income is not taxed, the size of the rental sector is indeterminate. Suppose next that rather than buying housing out of income, households have the option of borrowing on a mortgage market, where interest payments on these loans are tax deductible. It is straightforward to see that if the rate at which households can deduct mortgage interest payments against tax is exactly equal to the tax rate on capital income, then households will be indierent between paying cash for housing versus taking out a mortgage. If the rate at which households can deduct mortgage interest is less than this, households strictly prefer to pay cash. The intuition is simply that the equilibrium net after tax rate of interest on a mortgage loan in the economy is (r k )(1 ) where is the fraction of interest payments that may be deducted against tax. The marginal benet of taking out a mortgage loan is the return on the extra dollar of savings that can then be saved, with return (r k )(1 k ). Only if = k will households be indierent between alternative ways of nancing house purchases.12
12

Gervais 2001 conducts a richer analysis of the interaction between housing and the tax code.

14

What is the appropriate measure of the price of housing? Thus far we have dened two house prices: pn , the price of a unit of new housing to be delivered at h the start of the next period, and po , the price of a unit of used housing delivered h at the end of the current period. For comparison to the data, we choose to dene the price of housing as the price that would arise in a market for housing delivered at the start of the current period (once shocks have been observed). Households must be indierent between buying a house on this market versus renting a house for the current period, and buying a unit of used housing to be delivered at the end of the period. Thus we dene ph = q + po h In the National Income and Product Accounts, private consumption includes an imputed value for rents from owner-occupied housing. We choose to dene private consumption and GDP consistently with the NIPA. Thus private consumption expenditures is given by P CE = c + qh. Analogously GDP is given by GDP = yc + pr yh + qh. Note lastly that during a simulation of the economy, prices are changing, both because sector specic trends in productivity, and because of sector specic shocks around these trends. We dene real private consumption and real GDP using balanced growth path prices, so that our measures of real quantities capture trends in relative prices, but not short-run changes in relative prices. 2.4. Solution method Our goal is to simulate a calibrated version of the model economy. The rst step towards characterizing equilibrium dynamics is to solve for the models balanced growth path. We have a multi-sector model in which the trend growth rate of labor productivity varies across sectors. A balanced growth path exists since

15

preferences and all production functions have a Cobb-Douglas form. The gross trend growth rates of dierent variables are described in table 1. Several properties of these growth rates may be noted. The trend growth rates of yc , ph yh and pi xi for i {b, m, s} are all equal to gk , the trend growth rate of the capital stock and consumption. This growth rate is a weighted product of productivity growth in the three intermediate goods sectors. For example, B Mc Sc if capitals share is the same across sectors then gk = gzbc gzm gzs . The trend growth rates of intermediate goods prices exactly oset the eects of dierences in productivity growth across sectors such that (i) interest rates are trendless in all sectors (see eq. 2.13) and (ii) wages in units of consumption grow at the same rates across sectors (see eq. 2.14). Given trend growth rates for variables, the next step is to use these growth rates to take transformations of all the variables in the economy such that the transformed variables exhibit no trends. We do this because for computational purposes it is convenient to work with stationary variables. The new stationary variables are dened as follows, where x denotes a generic old variable, gx is the gross trend growth rate of the variable, and xt is the stationary transformation: xt = xt t gx

The penultimate step in the solution method is to linearize a set of equations in stationary variables that jointly characterize equilibrium around the balanced growth path, which corresponds to a vector containing the mean values of the transformed variables in the system. We solve the system of linear dierence equations using a Generalized Schur decomposition (see Klein 2000). 2.5. Data and calibration The model period is one year.13 This is designed to approximately capture the length of time between starting to plan new investment and the resulting increase
For more detail on all data sources and calibration procedures, see the data appendix of this paper which is available as Duke Economics Working Paper Number 00-09 and is also at www.econ.duke.edu/~heathcote/research.htm
13

16

in the capital stock being in place.14 Edge (2000) reports that for non-residential structures the average time to plan is around 6 months, while time to build from commencement of construction to completion is around 14 months. For residential investment the corresponding gures are 3 months and 7 months.15 Parameter values are reported in tables 2 and 3. The population growth rate, , is set to 1.8 percent per year, the average rate of growth of hours worked in private sector between 1949 and 1998, which is the sample period used for calibration purposes. Data from the National Income and Product Accounts Tables, the Fixed Reproducible Tangible Wealth tables, the Gross Product by Industry tables, and the Benchmark Input-Output Accounts of the United States, 1992 (all published by the Department of Commerce) are used to calibrate most remaining model parameters. The empirical analogue of the model capital stock is the stock of private xed capital (excluding the stocks of residential capital and consumer durables) plus the stock of government non-defense capital. The depreciation rate for capital, k , is set to 5.2 percent, which is the average annual depreciation rate for appropriately measured capital between 1949 and 1998. The empirical analogue of the model housing stock is the stock of residential xed private capital. The average annual depreciation rate for housing, which identies h , is 1.6 percent.
A yearly model is also convenient because data on inputs and output by intermediate industry and tax rate estimates are only available on an annual basis. In the rst draft of this paper, however, we set the period length to a quarter, and used an interpolation procedure to estimate sector-specic shocks (see the data appendix). Business cycle statistics are substantively the same for both period lengths. 15 Gomme Kydland and Rupert 2001 argue that faster time to build for residential structures can help account for the fact that non-residential investment lags the cycle, and that residential and non-residential investment are positively correlated contemporaneously. In their calibration they set the time to build for residential investment to 1 quarter, and the time for non-residential investment to 4 quarters. This dierence is probably too large, both because time to build for residential investment is likely longer than a quarter, and also because private non-residential structures only accounts for 28% of total non-residential investment over the 1949 to 1998 period; the majority of non-residential investment is accounted for by investment in equipment and software which can presumably be put in place more quickly. While there is probably still some role for dierential time to build, we abstract from it in this analysis to examine alternative mechanisms for generating realistic dynamics for investment over the business cycle.
14

17

The coecient of relative risk aversion, , is set equal to 2. All other preference parameters are endogenous. The shares of consumption and housing in utility (c and h ) are chosen so that in steady state, households spend 30 percent of their time endowment working, and so that the value of the capital stock is equal to annual GDP, which is the case, on average, for the sample period. The discount factor, , is set so that the annual after tax real interest rate in the model is 6 percent. The implied value for in the benchmark model is 0.967. Industry-specic data (industries are dened according to the 1987 2-digit SIC) are used to calibrate capital shares for the three intermediate sectors of the model: construction (b), manufacturing (m), and services (s). For the model construction sector, we use SIC construction industry data. For manufacturing, we use all NIPA classied goods-producing industries except construction: agriculture, forestry, and shing (AFF), mining, and manufacturing. For services we use all services-producing industries except FIRE: transportation and public utilities, wholesale trade, retail trade, and services.16 For each model sector i, the sectoral capital share in year t, i,t , is dened as i,t = P {V Aj,t IBTj,t P ROj,t }
j

P
j

COMPj,t (2.23)

where the j subscript denotes specic SIC industries included in sector i, and COMPj,t , V Aj,t , IBTj,t and P ROj,t denote, respectively, nominal compensation of employees, nominal value added, nominal indirect business tax and non-tax liabilities, and nominal proprietors income for industry j in year t. The average value of the construction sector capital share over the period is 0.13, for manufacturing it is 0.31, and for services it is 0.24 (see table 3). The logarithm of the (non-stationary) annual Solow residual in intermediate sector i is given by log (zi,t ) =
16

1 [log (xi,t ) i log (ki,t ) (1 i ) log (ni,t )] . 1 i

(2.24)

The FIRE (nance, insurance, and real estate) industry is omitted when calculating the capital share of the service sector of the model because much of FIRE value added is imputed income from owner occupied housing.

18

where xi,t is real output of intermediate sector i in year t, ki,t is real sectoral capital, and ni,t is sectoral hours worked. The residual of a regression of log (zi,t ) on a constant and a time trend over the sample period denes the logarithm of the detrended annual Solow residual e for industry i, denoted log (zi,t ). The annual growth rates of the non-stationary Solow residuals are 0.25 percent in construction, 2.77 percent in manufacturing, and 1.68 percent in services, identifying the quarterly growth rates gzb , gzm , and gzs . Government consumption is set equal to 18.1 percent of GDP, the period average.17 The mean tax rates on capital and labor income, k and n , are set so that along the balanced growth path the model matches two features of the data over the period: the capital stock averages 1.52 times annual output, and government transfers average 8.2 percent of GDP.18 To construct a series of shocks to tax rates, we take the series estimated by McGrattan, Rogerson and Wright (1997), who extend a methodology developed by Joines (1981). These rates are reported from 1947 to 1992. We extend the series for the years 1993 to 1997 using the method developed by Mendoza, Razin and Tesar (1994) and data from the OECD Revenue Statistics and National Accounts.19 In the model, logged detrended sectoral productivities and detrended tax rates are assumed to follow a joint autoregressive process. The estimates of the parameters dening this process are in table 4 at the end of the paper. A few
Government consumption in the data is dened as NIPA government consumption expenditures plus NIPA government defense investment expenditure. 18 Over the sample period, capital tax rates have trended downwards, while labor tax rates have increased. When we solve and simulate the model, however, we assume that there are no trends in tax rates. We do this because (i) this would considerably complicate computation of the models balanced growth path, and (ii) in the long run tax rates are bounded; thus the upward trend in labor tax rates is not sustainable. 19 The estimates for the last ve years are thus constructed using a somewhat dierent methodology than the 1947 to 1992 gures, and generate dierent average tax rates. However, for the period for which estimates based on the both methodologies are available, year to year changes in the estimated rates track each other closely. We therefore identify a series for tax shocks by (i) rescaling appropriately so that the two dierent sets of estimates for capital and labor tax rates coincide in 1992, (ii) estimating linear trends in the two tax rate series by OLS, and (iii) computing deviations in the series from this linear trend.
17

19

features of these estimates are worth mentioning. First, there is little evidence that technology shocks spill-over across intermediate goods sectors. Second productivity shocks appear to be rather more persistent that shocks to tax rates. Third productivity shocks in the construction and manufacturing sectors appear to be considerably more volatile than those in services. Productivity shocks are weakly correlated across sectors, and in particular shocks to the construction sector are essentially uncorrelated with those in manufacturing. Capital tax rates are considerably more volatile than labor tax rates. An important aspect of the calibration procedure concerns the estimates for {Bc , Mc , Sc } and {Bh , Mh , Sh }, the shares of construction, manufacturing and services in production of the consumption-investment good (subscript c) and the residential investment good (subscript h). These parameters determine the extent to which residential investment is produced with a dierent mix of inputs than other goods. At this point, we employ the Use table of 1992 Benchmark NIPA Input-Output (IO) tables. The IO Use table contains two sub-tables. In the rst, total spending on components of nal aggregate demand (personal consumption, private xed investment, etc.) is decomposed into sales purchased from all intermediate industries. In the second, total sales for each private industry (and for the government) are attributed to value-added by that industry, and sales purchased from other industries. Thus, for example, nal sales of the construction industry include value added from construction and sales purchased from the manufacturing and services sectors. One possible approach would be to assume that the distribution of valueadded across intermediate sectors for each component of nal demand is equal to the distribution of sales purchased from the dierent sectors (see table 5). Rather than doing this, we use the second IO table to track down where value was originally created in each intermediate industrys sales. For example, some portion of construction sales is attributed to purchases from manufacturing, which in turn can implicitly be divided into manufacturing value added plus sales to manufacturing from construction and services. Since this trail is never ending, dividing the nal sales of a particular industry into fractions of value-added by each intermediate industry requires an innite recursion. Once we have this breakdown, we use the rst IO Use table to compute, 20

for example, the fraction of value added in residential investment from the construction industry (which will identify the parameter Bh ). The results are given in table 6. The shares of value added by construction, manufacturing and services in the consumption-investment sector are respectively Bc = 0.0307, Mc = 0.2696, and Sc = 0.6998. For residential investment, the corresponding shares are Bh = 0.4697, Mh = 0.2382, and Sh = 0.2921. Comparing tables 5 and 6 it is clear that there are large dierences between the distribution of value added and the distribution of sales. For example, although we attribute residential investment entirely to sales from construction, these construction sales implicitly contain large quantities of value originally created in the manufacturing and service industries, such that only 47 percent of the value of residential investment is ultimately attributable to the construction industry.20 21 The parameters determining the size of adjustment costs in residential and non-residential investment are set such that adjustment costs are identical for the two types of investment, and so that the average percentage standard deviation of residential investment across simulations of the model is 5.11 times that of GDP, the average empirical ratio over the sample period. One measure of the size of adjustment costs is the elasticity of the price of new capital, pn , with respect to k the investment rate, ik /k. In the model this elasticity (equal, by assumption for both types of investment) is 0.137. This represents much smaller adjustment costs than are often used.22 2.6. Questions There are four sets of issues we use the model to address. First, we ask how successful is the calibration procedure in terms of matching rst moments, such
More details concerning the data and the matrix algebra used to construct table 6 are in the data appendix. 21 Hornstein and Praschnik 1997 describe a model in which a non-durable intermediate input is used in durable goods production, but they do not use Input-Output data in their calibration procedure. 22 An alternative measure of the size of adjustment costs is the average value for ik (et ) k ik (et )/k(et ) k(et ) /ik (et ), which is the fraction of resources invested that do not tranlate into additional new capital because of adjustment costs. Under the baseline calibration this value is 0.037 percent.
20

21

as the average fraction of GDP accounted for by residential investment. Second, we simulate the model and compare second moments of simulated model output to the data. This is the standard exercise in the real business cycle tradition. To gain some intuition about how the model works, we systematically compare our model to Greenwood and Hercowitz (1991). Third, we feed in the actual productivity and tax shocks suggested by our calibration procedure, and examine the extent to which the model can account for the observed history of a set of macroeconomic aggregates from 1948 to 1997. Lastly, we contrast our model with the typical framework for thinking about residential investment, and argue that previous empirical work does not necessarily support the view that residential investment dynamics are driven by shocks on the demand side rather than the supply side.

3. Results
3.1. Balanced growth path Table 7 indicates that the model is very successful in terms of matching rst moments. For example, the shares along the balanced growth path of the various components of aggregate demand are virtually identical in the model and the data. In particular, note that the model reproduces the observed shares of nonresidential and residential investment in GDP. This suggests that we are using the correct depreciation rates for capital and housing, and appropriate growth rates for productivity and population.23 In terms of the shares of gross private domestic income accounted for by the three intermediate goods sectors, the calibration delivers the correct average size of the construction industry, but delivers a manufacturing share that is too small relative to the sample average in the data. The reason is that intermediate goods shares in nal goods production were computed
This is interesting in light of the fact that a depreciation rate of around 10 percent on an annual basis is typically assumed for capital, while the annual rate in our calibration is only 5.2 percent. We attribute this dierence to the facts that (1) we exclude consumer durables from our measure of the capital stock, and (2) we explicitly account for both productivity growth and population growth, both of which imply a relatively high investment rate along the balanced growth path even with little depreciation.
23

22

using the 1992 Input - Output tables, and manufacturings share of the economy has declined over the post War period. The average tax rates generated by the calibration procedure are k = 41.8 percent, and n = 28.4 percent. These are extremely close to standard estimates in the taxation literature (see, for example, Domeij and Heathcote 2001). 3.2. Business cycle frequency uctuations We simulate our model economy to determine whether it is capable of accounting for some of the facts regarding the behavior of housing over the business cycle in the United States. A large set of business cycle moments are presented in table 8. We nd that our benchmark model can account for many of the features of the data that we document in the introduction. In particular, we reproduce almost exactly the volatilities of both non-residential investment and residential investment relative to GDP, and residential investment is positively correlated with consumption, non-residential investment and GDP.24 House prices are procyclical and positively correlated with residential investment. Labor supplies and outputs are correlated across intermediate goods sectors, and output and employment are most volatile in the construction industry and least volatile in the services industry. Thus the model can account for both the relative volatility and the co-movement puzzles. There are, however, two respects in which the model performs poorly. First, house prices (at the aggregate level) are slightly more volatile than GDP in the U.S., while in the model they are only half as volatile. Second, a striking feature of residential investment noted in the introduction is that it strongly leads the cycle; the correlation between GDP and residential investment the previous year is larger than the contemporaneous correlation between the two (see table 8). In the model the strongest correlation is the contemporaneous one, and thus the model fails to reproduce this feature of the data. The model can claim a more limited success, however, in that the correlation between residential investment
An implication of the rst nding is that our parameter values and simulation results would have been virtually identical had we calibrated the adjustment cost parameter to match the volatility of non-residential investment relative to GDP (recall that we instead chose to match the observed relative volatility of residential investment).
24

23

at a lead of a year with GDP is larger than the correlation when residential investment is lagged by a year.25 Comparison to Greenwood and Hercowitz (1991) To understand which features of the model allow us to reproduce particular features of the data, we consider several alternative parameterizations (see table 9). For each alternative parameterization we consider, we adjust a set of six preference and scal parameters (, c , h , g, k and n ) so that the balanced growth path ratios to GDP of government spending, transfers, capital and housing are all unchanged, labor supply is 30 percent of the time endowment, and the after tax interest rate is 6 percent. One way to think of this exercise is that for each alternative parameterization we recalibrate the model so that it does a reasonable job in terms of matching certain rst moments of the data, and then simulate to assess how second moments vary across parameterizations. Our rst alternative parameterization is essentially the benchmark model in Greenwood and Hercowitz (1991). The model eectively has only one production technology, capital and housing depreciate at the same rate, and there are no adjustment costs or tax shocks. Thus this set up is a standard one-sector RBC model, except that some fraction of the capital stock enters the utility function rather than the production function. Utility is log separable in consumption, housing and leisure, and thus the model can be reinterpreted as a reduced form of an economy with home production (see the introduction). Simulation results are in column GH of table 10. Since the dierent intermediate goods are produced using identical technologies and enter symmetrically in production of the two nal goods, the model has nothing to say regarding the relative volatilities and cross-sectoral correlations of construction, manufacturing and services. In other respects the model performs poorly. For example, the model predicts very little volatility in house prices, primarily because there is never a productivity differential between production of consumption versus residential investment. Labor supply and residential investment are much less volatile in the model than in the data.
Our paper does about as well in terms of replicating observed lead-lag patterns as Gomme et. al. 2001, who focus on dierential time to build across sectors.
25

24

The main dierences between the GH parameterization and the economy labelled A in tables 9 and 10 is that the coecient of relative risk aversion is increased to 2, (the value in our baseline calibration), and productivity shocks are stationary (but persistent) rather than unit root. These changes only worsen the performance of the model. Non-residential investment becomes much more volatile than residential investment, and the two types of investment co-move negatively contemporaneously. Moreover, non-residential investment strongly leads GDP, while residential investment strongly lags, exactly the opposite of what is observed in the data. The explanation for this poor performance is that following a good productivity shock, households want to immediately allocate a larger fraction of output towards increasing the capital stock used in production, where productivity is high, rather than towards increasing the stock of capital used in utility production, where productivity is unchanged. We now proceed to add increasing realism, layer by layer, to the straw-man models described above. The rst thing we add are adjustment costs, where adjustment costs are at the same level as in our benchmark parameterization. We nd that adjustment costs have large eects on the behavior of the model (see column B of table 10). In particular, the volatilities of both types of investment fall signicantly, and the two types of investment are now (perfectly) positively correlated. The intuition is that with large adjustment costs it is optimal to increase the business capital stock more slowly following a good shock, and this means more new capital is available for adding to the housing stock. Introducing adjustment costs also increases house price volatility, since the price of housing now depends on the ratio of residential investment to the housing stock. There are still many respects, however, in which the gap between the model and the data remains large. One is the relative volatility puzzle; business investment is more volatile than residential investment, contrary to the pattern in the data. We next introduce sector specic productivity shocks (see column C). The shock process is parameterized following a procedure analogous to that described in the calibration section, except that tax rates are not included in the system. Since productivity shocks are estimated to be more volatile in construction and manufacturing than services, this change has the eect of generating relatively more volatile output (and employment) in these sectors. However, since all nal 25

goods are still produced using the same technology, this change has little eect on the business cycle dynamics of any macro aggregates. The next feature we add is sector specic capital shares for intermediate goods rms (column D). The fact that construction is relatively labor intensive means that output of the construction sector becomes more volatile than in the previous case, while the fact that manufacturing is relatively capital intensive reduces the volatility of manufacturing output. The intuition is simply that following a good productivity shock, it is easier to expand output rapidly the more important is labor in production, since holding capital constant, the marginal product of labor declines more slowly. Next, in column E, we introduce dierence depreciation rates for capital and housing. The largest eect of this change is to increase the volatility of residential investment, so that non-residential investment and residential investment are now equally volatile. The reason reducing the depreciation rate for housing increases the volatility of residential investment is that slower depreciation increases opportunities to concentrate residential investment in periods of high productivity; conversely during a prolonged period of low productivity, it is possible to build few or new no homes without bringing about a large fall in the stock. Column F introduces dierent nal goods production technologies for the consumption / business investment versus the residential investment sectors. This is essentially the benchmark model, except that tax rates are still assumed constant. This change has large eects. Because intermediate input intensities dier across the two dierent nal goods, sector specic productivity shocks change the eective relative cost of building new houses versus other goods. This accounts for the decline in the correlation between residential and non-residential investment. Allowing for two nal goods technologies more than doubles the percentage standard deviation of residential investment, such that residential investment is now much more volatile than non-residential investment, as in the data. The explanation for this result hinges on the fact that construction is a much more important input for residential investment than for the rest of the economy. Recall that productivity shocks in the construction industry have a larger variance than those in the services industry, and that construction is also relatively labor intensive. These characteristics of the construction technology tend to increase 26

the relative volatility of construction-intensive residential investment. There is an additional eect in the opposite direction, however, in that output and employment in the construction sector are now more volatile than in the equivalent version of the model with a single nal goods sector (compare columns E and F ). With a separate residential investment sector, a fall in the relative price of construction inputs associated with an increase in construction productivity translates into a fall in the price of residential investment (see eq. 2.22). Since demand for residential investment is very price sensitive this in turn generates a large boom in residential investment and in the demand for construction inputs. Thus, part of the reason the construction industry is so volatile is that a large fraction of the demand for its output is for residential investment. Lastly consider the eects of introducing stochastic tax rates by comparing column F with column G (our benchmark economy). The main eect of this change is to increase the volatility of labor supply, which arises because temporary changes in the labor income tax rate change the relative returns to working at dierent dates. The percentage standard deviation of hours in the services sector (relative to GDP) is now 0.47, compared to 0.23 in the same economy without tax shocks. More volatile labor supply translates into more volatile output; the percentage standard deviation of GDP rises from 1.63 to 1.93. Thus the model now comes to reproducing the observed volatility of output even though all productivity shocks are sector-specic. The volatilities of all other variables also rise relative to the economies without tax shocks, but not proportionately; the volatility of consumption rises slightly relative to GDP, while the volatilities of both types of investment relative to GDP fall. The correlation between house prices and residential investment increases from 0.18 to 0.36. All these changes improve the overall success of the model in terms of replicating observed business cycle dynamics. Our initial expectation was that introducing stochastic capital income taxes would increase the volatility of investment by increasing time variation in the after-tax return to capital, and in the relative expected after-tax returns to saving in the form of taxed capital versus untaxed labor. One reason investment volatility does not rise much is that it is expectations over future capital tax rates that aect

27

investment decisions, and capital tax shocks are not very persistent.26 Co-movement and leading residential investment What accounts for the ability of the benchmark model to reproduce the empirical positive correlation between house prices and residential investment, and the positive correlations between consumption and either type of investment? One important component, as discussed above, is the presence of adjustment costs. Without adjustment costs, increased relative productivity in construction would tend to reduce house prices but increase production of new housing. At the same time output of the relatively expensive consumption / investment good would tend to fall. When adjustment costs are introduced, additional residential investment drives up house prices, which restores a positive residential investment / house price correlation. Furthermore, a smaller increase in residential investment leaves more resources for producing the consumption / investment good, helping restore a positive correlation between dierent components of nal demand. However, adjustment costs are small under the baseline calibration. Moreover, the model continues to generate positive correlations between residential investment on the one hand and GDP, non-residential investment and house prices on the other even when the elasticity of new house (capital) prices with respect to the investment rate is reduced (for both types of investment) from the baseline value of 0.137 to a value of 0.065.27 One reason for this is that while the under-lying productivity shocks are only weakly positively correlated across intermediate goods sectors, the correlation is eectively magnied at the nal goods level, since both nal goods sectors use all three intermediate goods as inputs (albeit in dierent proportions).28
We also experimented with setting all shock innovations to either the capital tax rate or the labor tax rate to zero. With no capital tax rate shocks, the main dierences relative to economy F are that GDP is slightly more volatile (1.80 versus 1.63) and the percentage standard deviation of labor supply (relative to GDP) increases from 0.30 to 0.53. With no labor tax rate shocks, the volatilities of GDP and labor supply are 1.79 and 0.42, and, relative to GDP, residential investment is less volatile than in economy F (5.12 versus 5.43). 27 When the elasticity is 0.065, residential investment is (counter-factually) eight times as volatile as GDP. 28 There is a sense in which the low depreciation rate for housing and the low capital share in construction also contribute to resolving the co-movement puzzle. In particular, suppose
26

28

Part of the reason the residential investment weakly leads GDP in the model is that output of the residential investment sector may be increased relatively eciently without waiting for additional capital to become available. This is because the residential investment technology is construction and therefore ultimately labor intensive. When all nal goods are produced according to the same technology (and are therefore equally labor intensive) it is non-residential investment rather than residential investment that weakly leads the cycle (see column E in table 10). 3.3. Accounting for U.S. history The aim of this section of the paper is to compare the observed timepaths for a set of macro variables over the post-War period to those predicted by our model, given the estimated series of productivity and tax shocks. In particular, we assume that the U.S. economy was on its balanced growth path until the start of 1949, and that from 1949 until 1997 the shocks that generated deviations from the balanced growth path were equal to the residuals generated from the autoregressive estimation procedure described in the calibration section. This is a more ambitious exercise than the simulation exercise conducted above in that we are now assessing the performance of the model at all frequencies. In particular, we shall address the extent to which the model can account for observed long run trends, as well as for business cycle frequency uctuations. Moreover, we look at sectoral output and house prices, in addition to the standard macro aggregates. The series in gures 1 and 2 are all scaled so that they take the value one in 1949. Since constant quality house price data is only available from 1963 we normalize
that we deviate slightly from the baseline calibration by setting h = k = 0.052, and assume b = m = s = 0.25. Suppose, in addition, that we reset the adjustment cost parameter so that, as before, the model reproduces the observed relative volatility of residential investment (the implied elasticity turns out to be 0.057). The values for corr(non RESI, RESI) and corr(Ph , RESI) in this economy are counter-factually negative (0.01 and 0.24 respectively), while the percentage standard deviation of non-residential investment is larger than in the data (2.75 times that of output). This suggests a connection between the co-movement and relative volatility puzzles. In particular, the parameter values that are important for reproducing observed relative volatiliy lead us to adopt, within the calibration procedure, investment adjustment costs large enough to resolve the co-movement puzzle.

29

house prices to be one in 1963. Across the sample period U.S. private consumption and GDP both increased by approximately a factor of ve. At the same time, non-residential investment increased by a factor of seven whereas residential investment grew only half as much. The model does an excellent job of matching trend growth in hours worked, and a very good of capturing long run growth in consumption and output. Since the focus of the paper is primarily on residential investment, it is reassuring that the model correctly predicts residential investment to be the slowest growing component of nal demand. The reason is simply that since residential investment is construction intensive, and negative trend growth in construction productivity means that the relative prices of residential investment and of housing tend to be rising over time. This reduces growth in the demand for new housing; given Cobb Douglas preferences expenditure shares on consumption, leisure and housing are constant along the balanced growth path. The biggest failures of the model at low frequencies are in replicating the growth of non-residential investment and of manufacturing output. Both discrepancies are readily accounted for. The model overpredicts manufacturing growth, since manufacturings share of trend nominal output is constant in the model as a consequence of assuming Cobb Douglas technologies and preferences. However, manufacturings share of nominal output has fallen dramatically over the sample period, as the service sector has grown. The fact that the model underpredicts growth in non-residential investment is likely in part a consequence of assuming a constant depreciation for capital; when we estimate depreciation rates using NIPA nominal depreciation gures, we nd the rate of depreciation to be rising over time. A second reason the model underpredicts non-residential investment growth is that we assume a common production technology for consumption and business investment, whereas in reality non-residential investment is more manufacturing intensive than consumption (see table 6). Thus the model generates too little trend decline in the relative price of business capital (generated by manufacturing productivity growth) and too little real growth in business investment. Consider next the ability of the model to account for U.S. macroeconomic history at business cycle frequencies. To better assess the models cyclical performance, gure 3 describes percentage deviations from a Hodrick Prescott trend 30

for various macro aggregates. Note rst that the model closely reproduces the histories of deviations from trend in GDP and consumption. Slightly poorer performance at the end points of the sample might reect poor endpoint properties of the Hodrick Prescott lter. The poor t in the rst few years of the sample might alternatively be due to the U.S. economy being o its balanced growth path prior 1949, contrary to the assumption made here. The model also does a good job in accounting for historical output uctuations at the sectoral level (see gure 2). Deviations from trend in the data are much larger for non-residential investment than for GDP or consumption, and larger again for residential investment. Comparing the predictions of the model with the data, the t for both types of investment is generally good, suggesting that productivity and tax shocks can largely account for observed investment dynamics. There a few caveats to this conclusion, however. First, swings in non-residential investment are generally somewhat smaller in the model than in the data. Second, non-residential investment in the data appears to slightly lag non-residential investment in the model (which ts with the fact that non-residential investment lags the cycle empirically) whereas the model delivers no strong lead-lag patterns. Third, residential investment in the data appears to slightly lead residential investment in the model during the early part of the sample (which again is consistent with residential investment leading the cycle empirically). Comparing the last two major recessions, the model does a very good in accounting for the depth of the recession in the early 1980s, including a dramatic fall in residential investment which was 38 percent below trend in 1982. However, the model underpredicts the depth of the recession in the early 1990s, a failing which is also clear from gure 1.29 One important respect in which the model performs poorly is in accounting for house price dynamics. The model does at least correctly predict an upward trend in the relative price of housing. This upward trend in the data looks more dramatic if house prices are measured relative to the GDP deator relative to the CPI.30 The reason for the upward trend in the model is simply that productivity
Hansen and Prescott 1993 also investigate the 1990-1991 recession in a multi-sector model. We compare the price of housing relative to the consumption / investment good in the model. In the data the price of output (GDP deator) incorporates changes in the price of
30 29

31

in the construction industry has been declining over time relative to productivity in other industries. In terms of cyclical volatility in house prices, the model does not account for a large fraction of observed price dynamics, and in particular it fails to capture the large boom in house prices that peaked in 1979. We return to this issue in the conclusion. 3.4. Relation to empirical literature on residential investment Our model is not the rst to address the determination of residential investment and house prices. A simple version of what we label the traditional model may be visualized as follows (see Kearl 1979, Topel and Rosen 1998, or Poterba 1984 and 1991). There is an upward sloping supply curve for residential investment, since higher prices encourage developers to build more houses. The demand curve for new houses is innitely elastic, since houses are viewed as nancial assets that must pay the market rate of return. Demand shocks aect future expected rents (dividends on the housing asset) and shift the demand curve up and down. Thus equilibrium price / investment pairs are traced out along the residential investment supply curve. Topel and Rosen (p.737) conclude that the large price elasticity of supply of new houses estimated here must be an important consideration for understanding the great variability in housing investment. Our model shares many features of the traditional view; for example, housing may be viewed as an asset which in equilibrium oers the same expected return as capital, provided implicit rents from owner-occupation are included in the return calculation. However, there are some important dierences between the two frameworks, reecting the fact that we are explicit about all the production technologies in the economy and are therefore able to derive exact equilibrium expressions for all prices. One key dierence is that the dynamics of both residential investment and house prices in our model are primarily driven by supply-side productivity shocks.31 In contrast, the typical assumption in the traditional model is
residential investment. At the same time the CPI incorporates a shelter component. When we constructed a model equivalent of the CPI, we found slightly slower trend growth in the relative price of housing. 31 Contrary to the traditional view, the demand for new housing is not perfectly elastic, since additional housing implies less new capital, which drives up future interest rates, and increases

32

that the current price of housing is largely independent of construction costs, and changes in house prices mostly reect changes in the demand for housing relative to other goods. When residential investment is regressed on house prices and the real interest rate, the coecient on the house price term typically turns out to be positive and strongly signicant. This appears to conrm the traditional demand-shock driven view of residential investment. At the same time, a positive house price co-ecient is prima facie inconsistent with a productivity shock driven theory of residential investment, according to which one might expect an increase in residential investment to be associated with higher productivity in the construction sector and thus lower prices for construction output and housing. We shall argue, however, that care should be taken in interpreting these regression results. In the appendix we show that the rst order conditions of our model imply the following equilibrium relationship between residential investment and new house prices: (1 h )log (yh ) = log(r) log(pn ) h log(h) + log h h

where is a term involving sector capital stocks, and is the balanced growth h ih path value for h . This relationship indicates a negative relationship between residential investment and house prices, holding other variables constant. However, in addition to the rental rate on capital, the equation implied by the model also includes terms involving sectoral capital stocks and the existing stock of houses.32 This relationship cannot be interpreted as a supply curve, since the variables in the equation are all determined endogenously in equilibrium and depend on both preference and technology parameters. Equation 3.1 is not the residential investment equation that has typically been estimated in the literature. In table 11 we consider various alternative specications, in order to assess the potential importance of mis-specication. We estimate each statistical model using simulated output from the model, where the model simulation uses the identied historical shocks (see the previous section).
the opportunity cost of saving in the form of housing. 32 If there are no adjustment costs in the model ( h = 0) then the housing stock term drops out, but the term involving sectoral capital stocks remains.

(3.1)

33

In the rst column we estimate the correct model, using ordinary least squares, as a check on our algebra. Given equation 3.1, the coecients on the price of new housing and the marginal product of capital should be minus one and one respectively, and the regression should t perfectly. The reason the estimated coecients do not exactly match the predicted coecients is that given the numerical solution method, simulation allocations are not exactly equilibrium allocations away from the balanced growth path. In alternatives (1) and (2) we re-estimate the residential investment equation omitting the housing and sectoral capital stock terms. The estimated co-ecient on the house price term now becomes positive (and large and signicant if no time trend is included). Thus omitting variables could lead one to interpret this articial economy as being driven by shocks to the demand for housing, whereas in fact (by construction) it is driven by supply-side productivity shocks. The problem of omitted variables is not the only way in which previous estimation equations have been mis-specied relative to the equilibrium relation that obtains in our model. Two other problems are that in equation 3.1 it is the pre-tax rental rate for capital that enters the equation, rather than the interest rate on a risk-free bond (typically lagged) that has been used in empirical work. In addition the house price term is the price of a new house to be delivered at the start of the next period, rather than the current start of period market price of housing. In columns (3) and (4) we explore the importance of mis-specication in these dimensions. We compute risk free interest rates in the model using the standard pricing formula for risk and tax free one period bonds. We nd that the eect of using the risk free rate rather than the marginal product capital is to further increase the estimated coecient on the house price term, and to weaken the coecient on the interest rate term. However, exactly which measure of house prices we use turns out not to be very important. To summarize, previous empirical work estimating a residential investment supply curve is only loosely grounded in theory, and the equations estimated are mis-specied relative to the structural equilibrium relationship implied by the model developed here. Bias due to mis-specication is potentially large, and thus previous ndings do not necessarily support the view that uctuations in residential investment are primarily driven by demand-side shocks to house 34

prices.

4. Conclusion
This paper suggests that many of the aggregate stylized facts relating to housing, such as the high volatility of residential investment, are natural consequences of the way houses and other goods are produced. Some key features of the calibrated model economys production technologies are the importance of construction as an input to residential investment, the importance of labor as an input in the construction industry, and the fact that housing depreciates much more slowly than capital. An important part of the calibration procedure involves using Input Output data to x the relative importance of dierent intermediate inputs in various components of nal demand. Our focus was on residential investment, and we therefore chose to emphasize the construction industry at the intermediate industry level, and to abstract from dierences in the production technologies for consumption versus non-residential investment at the nal goods level. However, using the same calibration methodology it would be relatively straightforward to consider a ner disaggregation of the components of nal demand, or to increase the number of intermediate sectors. For example, one could use the input mix estimates in table 6 to introduce an explicit non-residential investment sector. The paper leaves several open issues for exploration. First, what can account for the strong lead of residential investment over the cycle? Both this paper and Gomme, Kydland and Rupert (2001) make some progress on this dimension, but neither succeed in reproducing the fact that the correlation between GDP at date t and residential investment (or residential investment plus purchases of consumer durables) at t 1 is larger than the contemporaneous correlation. Second, the model does poorly in accounting for long swings in house prices such as the large run up in house prices that occurred in the 1970s. House prices in the model are pinned down in the medium to long term by relative sectoral productivities.33 This means that there is little prospect of being able to account
This is a consequence of assuming that all markets are competitive and that new and (appropriately discounted) old houses are perfect substitutes. House prices would only reect relative
33

35

for U.S. house price history by extending the model to capture demand side factors that have received attention in the literature, such as the demographics of the baby boom and bust, or changes through time in the eective size of the tax advantage conferred by mortgage interest deductability. Introducing land as a xed factor, however, could allow a larger role for demand side shocks.34 For example, a surge in immigration would tend to drive up land and thus house prices in both the short and long run, though it is not clear the eect would be quantitatively important.

References
[1] Benhabib, J., R. Rogerson, and R. Wright, 1991, Homework in Macroeconomics: Household Production and Aggregate Fluctuations, Journal of Political Economy 99, 1166-1187. [2] Baxter, M., 1996, Are Consumer Durables Important for Business Cycles?, Review of Economics and Statistics 78 (1), 147-155. [3] Boldrin, M., L. Christiano, and J. Fisher, 2001, Habit Persistence, Asset Returns and the Business Cycle, American Economic Review, forthcoming. [4] Chang, Y., 2000, Comovement, Excess Volatility and Home Production, Journal of Monetary Economics 46, 385-396. [5] Cooley, T. and E. Prescott, 1995, Economic Growth and Business Cycles, in T. Cooley ed., Frontiers of Business Cycle Research, Princeton University Press, Princeton, 1-38. [6] Diaz-Gimenez, J., E. Prescott, T. Fitzgerald, and F. Alvarez, Banking in Computable General Equilibrium Economies, Research Department Sta Report 153, Federal Reserve Bank of Minneapolis.
productivities even in the short run if there were no adjustment costs and all sectors were equally capital intensive (see eqs. 2.19 and 2.22). 34 See Poterba 1991 for some evidence on the importance of land prices and for a review of alternative theories of house prices.

36

[7] Domeij, D. and J. Heathcote, 2001, Factor Taxation with Heterogeneous Agents, working paper, available at http://www.econ.duke.edu/~heathcote/ [8] Edge, R., 2000, The Eect of Monetary Policy on Residential and Structures Investment under Dierential Project Planning and Completion Times, International Finance Discussion Paper Number 671, Board of Governors of the Federal Reserve System. [9] Fernandez-Villaverde, J. and D. Krueger, 2001, Consumption and Saving over the Life-Cycle: How Important are Consumer Durables?, working paper, Stanford University. [10] Fisher, J., 1997, Relative Prices, Complementarities and Comovement among Components of Aggregate Expenditures, Journal of Monetary Economics 39 (3), 449-474. [11] Gervais, M., 2001, Housing Taxation and Capital Accumulation, Journal of Monetary Economics, forthcoming. [12] Gomme, P., F. Kydland, and P. Rupert, 2001, Home Production Meets Time to Build, Journal of Political Economy 109 (5), 1115-1131. [13] Greenwood, J., R. Rogerson, and R. Wright, 1995, Household Production in Real Business Cycle Theory, in T. Cooley ed., Frontiers of Business Cycle Research, Princeton University Press, Princeton, 157-174. [14] Greenwood, J. and Z. Hercowitz, 1991, The Allocation of Capital and Time over the Business Cycle, Journal of Political Economy 99, 1188-1214. [15] Hansen, G. and E. Prescott, 1993, Did Technology Shocks Cause the 19901991 Recession?, American Economic Review 83 (2) 280-286. [16] Hornstein, A. and J. Praschnik, 1997, Intermediate Inputs and Sectoral Comovement in the Business Cycle, Journal of Monetary Economics 40, 573-595.

37

[17] Joines, D., 1981, Estimates of Eective Marginal Tax Rates on Factor Incomes, Journal of Business 54 (2), 191-226. [18] Kearl, J., 1979, Ination, Mortgage, and Housing, Journal of Political Economy 87 (5), 1115-1138. [19] Klein, P., 2000, Using the Generalized Schur Form to Solve a Multi-variate Linear Rational Expectations Model, Journal of Economic Dynamics and Control 24 (10), 1405-1423. [20] Krusell, P. and A. Smith, 1998, Income and Wealth Heterogeneity in the Macroeconomy, Journal of Political Economy 106, 867-896. [21] McGrattan, E., R. Rogerson, and R. Wright, 1997, An Equilibrium Model of the Business Cycle with Household Production and Fiscal Policy, International Economic Review 38 (2), 267-290. [22] Mendoza, E., A. Razin and L. Tesar, 1994, Eective Tax Rates in Macroeconomics: Cross-Country Estimates of Tax Rates on Factor Incomes and Consumption, Journal of Monetary Economics 34 (3), 297-323. [23] Ortalo-Magne, F. and S. Rady, 2001, Housing Market Dynamics: On the Contribution of Income Shocks and Credit Constraints, CEPR Working Paper 3015. [24] Platania, J. and D. Schlagenhauf, 2000, Housing and Asset Holding in a Dynamic General Equilibrium Model, working paper, Florida State University. [25] Poterba, J., 1984, Tax Subsidies to Owner-occupied Housing: An Asset Market Approach, Quarterly Journal of Economics 99, 729-752. [26] Poterba, J., 1991, House Price Dynamics: The Role of Tax Policy and Demography, Brookings Papers on Economic Activity 2:1991, 143-203. [27] Rios-Rull, J.V., 1994, On the Quantitative Importance of Market Completeness, Journal of Monetary Economics 34, 462-496.

38

[28] Storesletten, K., 1993, Residential Investment over the Business Cycle, working paper, Institute for International Economic Studies, Stockholm University. [29] Topel, R. and S. Rosen, 1988, Housing Investment in the United States, Journal of Political Economy 96 (4), 718-740.

5. Appendix: residential investment and house prices

The total value of intermediate goods output must equal the total value of nal goods output, since nal goods rms make zero prots: pb xb + pm xm + ps xs = yc + ph yh . (5.1)

From the equilibrium expression for the price of manufactures (eq. 2.16) yc = pm mc . Mc

From the equilibrium expressions for the interest rate (eq. 2.13) xi = Thus eq. 5.1 may be rewritten as r

rKi . pi i

Kb Km Ks + + b m s

pm mc + ph yh . Mc

(5.2)

Using the market clearing condition for manufactures (eq. 2.11) and reusing eq. 2.13 and eq. 2.16 we can derive a convenient alternative expression for mc mc = xm mh rKm Mh ph yh = . pm m pm Substituting eq. 5.3 into eq. 5.2 Kb Km Ks r + + b m s

(5.3)

rKm m

Mh ph yh Mc

+ ph yh .

(5.4)

39

Recall that the price of new housing is related to the price of residential investment as follows ph . pn = 0 h h (yh /h) Thus eq. 5.4 may be rewritten as yh = where , pn 0 (yh /h) h h

Kb Km Ks Mc Km + + . = (Mh Mc ) m (Mh Mc ) b m s In logarithms, the equilibrium relation between residential investment, house prices and the real interest rate is log(yh ) = log(r) + log() log(pn ) log(0 (yh /h)). h h

To make further progress we need to specify an explicit functional form for the adjustment cost function that satises the properties dened in eqs. 2.7 through 2.9. One such function is h (x) =

1
(1) h

x .

h where is the steady state value for ih . Note that for this functional form, the h elasticity of the price of new capital or new housing with respect to the appropriate investment rate is given by

k =

00 k 0 k

ik k ik k

The derivative of the adjustment cost function is 0 (yh /h) =

ik k

= h =

00 h 0 h

ih h ih h

ih h

=1

1 yh (h )1 h

We use this expression to derive the following alternative equation relating residential investment to pn , r, , and the housing stock h : h log (yh ) = log(r) log(pn ) (1 )log(h) + log h h

(1)

40

Table 1: Balanced Growth Path Growth Rates (gross growth rates of per-capita variables) nb, nm, ns, n, r kb, km, ks, k, c, ik, g, yc, w bc, bh, xb mc, mh, xm sc, sh, xs h, ih, yh phyh, pbxb, pmxm, psxs 1
B M S g k = g zbc (1 b ) g zmc (1 m ) g zsc (1 s )

1 1 Bc b M c m S c s

g b = g b g 1 b k zb g m = g m g 1 m k zm g s = g s g 1 s k zs
M g h = g bBh g m h g sS h

gk

Table 2: Tax Rates, Depreciation Rates, Adjustment Costs, Preference Parameters Davis Heathcote Mean tax rate on capital income: k Mean tax rate on labor income: n Govt. cons. to GDP Transfers to GDP Depreciation rate for capital: k Depreciation rate for housing: h Elasticity pnk of wrt ik/k: k Elasticity of pnh wrt ih/h: h Population growth rate: Discount factor: Risk aversion: Consumptions share in utility: c Housings share in utility: h Leisures share in utility: 1-c-h 0.4177 0.2843 0.181*1 0.082* 0.052* 0.016* 0.137 0.137 1.8* 0.9672 2.00* 0.3162 0.0381 0.6457 0.078 0.078 0.0 0.0 0.0 0.96 1.00 0.2600 0.0962 0.6438 Grenwood Hercowitz (GH) 0.50 0.25 0.0

Starred parameter vales are chosen independently of the model.

Table 3: Production Technologies Con. Input shares cons/inv production: Bc, M c, Sc Input shares housing production: Bh, M h, Sh Capitals share by sector: b, m, s Trend productivity growth (%): gzb, gzm, gzs Autocorrelation co-efficient: Std. dev. log productivity innovations: () 0.0307 0.4697 0.13 -0.25 Man. 0.2696 0.2382 0.31 2.77 Ser. 0.6997 0.2921 0.24 1.68 0.30 1.00 1.0 0.022 GH

Table 4: Estimation of Exogenous Shock Process

System estimated: z t +1 = A + Bz t + t +1
log z tb tb m log z tm t log z s , = s and ~ N (0,V ) .2 where z t = t t t t tk tk n n t t

Standard deviation of innovations (b) (m) (s) (k) (n) 0.0409 0.0351 0.0176 0.0236 0.0062

Autoregressive coefficients in matrix B (standard errors in parentheses) log zbt+1 log zbt log z log
m t

log zmt+1 -0.036

(0.097)

log zst+1 0.047

(0.049)

kt+1 0.046
(0.065)

nt+1 -0.015
(0.017)

0.636
(0.113)

-0.121
(0.184)

0.714
(0.157)

0.081
(0.079)

0.036
(0.106)

-0.067
(0.028)

zst

0.012
(0.153)

-0.007
(0.131)

0.899
(0.066)

0.012
(0.088)

0.062
(0.023)

kt nt R2

0.212
(0.279)

-0.078
(0.239)

-0.257
(0.120)

0.525
(0.161)

0.104
(0.042)

-1.094
(0.877)

-0.900
(0.752)

0.264
(0.376)

-0.242
(0.506)

0.543
(0.132)

0.58 b

0.69 m 0.059 1.0

0.91 s 0.345 0.579 1.0

0.36 k -0.055 0.048 0.040 1.0

0.82 n -0.302 -0.256 -0.061 0.513

Correlations of innovations b m
s

1.0

All variables are linearly detrended prior to estimating this system.

Table 5: Decomposition of Final Demand into Final Sales From Industries (%) (based on 1992 IO-Use Table) PCE Construction Manufacturing Services 0.0 23.3 76.7 BFI + RESI 43.9 41.3 14.8 RESI3 100.0 0.0 0.0 BFI 22.6 56.9 20.5 G4 33.6 44.2 22.2

Table 6: Decomposition of Final Demand into Value Added by Industry (%) PCE Construction Manufacturing Services 1.4 23.0 75.7 BFI + RESI 21.3 40.6 38.1 RESI 47.0 23.8 29.2 BFI 11.6 46.9 41.5 PCE + BFI + GOVI5 3.1 27.0 70.0

Table 7: Properties of Steady State: Ratios to GDP % Data (1949-1998) Capital stock (K) Housing stock (Ph x H) Private consumption (PCE) Government consumption (G) Non-residential inv (non-RESI) Residential inv (Pr x RESI) Construction (Pb x Yb) Manufacturing (Pm x Ym) Services (Ps x Ys) Real after tax interest rate (%) 152 101 63.6 18.1 13.4 4.7 5.3 33.4 61.3 Model 152 101 63.8 18.1 13.6 4.6 5.2 26.8 68.0 6.0

We attribute all $225.5bn of residential investment in 1992 to sales from the construction industry, since the first I/O use table does not have a residential investment column. We defend this choice in the data appendix. 4 G is government expenditure, which includes government consumption and government investment expenditures. 5 GOVI is government investment. We assume that the value added composition of government investment by intermediate industry is the same as business fixed investment.

Table 8: Business Cycle Properties6 Data (1949-1998) % Standard Deviations (rel. to GDP) GDP PCE Labor (N) Non-RESI RESI Construction output (Yb) Manufacturing output (Ym) Services output (Ys) Construction hours (Nb) Manufacturing hours (Nm) Services hours (Ns) House prices (Ph) (Data: 1963-1998) Correlations PCE, GDP Ph, GDP (Data: 1963-1998) PCE, non-RESI PCE, RESI non-RESI, RESI Nb, Nm Nb, Ns N m, N s Ph, RESI (Data: 1963-1998) Lead lag patterns i=1 corr(Non-RESI t-i, GDPt) corr(RESI t-i, GDPt) 0.27 0.55 i=0 0.75 0.47 i = -1 0.55 -0.17 i=1 0.50 0.38 i=0 0.96 0.68 i = -1 0.50 0.25 0.80 0.53 0.63 0.67 0.24 0.77 0.89 0.78 0.39 0.98 0.88 0.96 0.56 0.45 0.75 0.64 0.99 0.36 2.31 0.77 1.03 2.23 5.11 2.74 1.84 0.85 2.33 1.52 0.64 1.29 1.93 0.62 0.54 2.19 5.11 3.50 1.44 0.97 1.92 0.52 0.47 0.54 Model

Statistics are averages over 500 simulations, each of length 50 periods, the length of our data sample. Prior to computing statistics all variables are (i) transformed from the stationary representation used in the solution procedure back into non-stationary representation incorporating trend growth, (ii) logged, and (iii) Hodrick-Prescott filtered.

Table 9: Alternative Parameterizations Model GH A Greenwood and Hercowitz One sector model, housing in utility (re-parameterized GH) Selected parameter values see tables 2 and 3 = 2, = 0.85, () = 0.022 k = h = 0.052 b = m = s = 0.25 Bh, = B c, Mh, = M c, Sh, = S c k = h = 0.137 see table 4 b = 0.13, m = 31, s = 0.24 h = 0.016 Bh, = 0.47, Mh, = 0.24, Sh, = 0.29

B C D E F G (Benchmark)

A + adjustment costs B + sector-specific shocks C + sector-specific capital shares D + different depreciation rates E + two final goods technologies E + stochastic taxes

Table 10: Alternative Parameterizations: Business Cycle Properties Data GDP (% std dev) 2.31 GH 1.39 A 1.93 B 1.72 C 1.60 D 1.56 E 1.60 F 1.63 G 1.93

Std. dev. relative to GDP PCE N Non-RESI RESI Yb Ym Ys Ph Correlations Non-RESI, RESI Ph, RESI 0.24 0.39 0.88 0.94 -0.10 0.80 1.00 0.98 1.00 0.98 1.00 0.98 0.99 0.99 0.41 0.18 0.45 0.36 0.77 1.03 2.23 5.11 2.74 1.84 0.85 1.29 0.60 0.36 2.73 2.08 1.25 1.25 1.25 0.11 0.39 0.47 4.01 2.92 1.14 1.14 1.14 0.07 0.53 0.34 2.41 1.97 1.13 1.13 1.13 0.35 0.56 0.32 2.36 1.88 1.90 1.80 1.04 0.34 0.56 0.31 2.35 1.87 2.27 1.68 1.08 0.34 0.59 0.29 2.39 2.40 2.20 1.62 1.04 0.38 0.60 0.30 2.31 5.43 3.87 1.59 0.97 0.55 0.62 0.54 2.19 5.11 3.50 1.44 0.97 0.54

Lead-lag pattern: corr(xt-1, GDPt) corr(x t+1, GDP t) x = RESI. x = Non-RESI 0.73 -0.28 -0.10 0.37 -0.94 0.46 0.07 0.10 0.05 0.09 0.05 0.09 -0.00 0.09 0.11 0.05 0.13 -0.01

Table 11: OLS Regressions on Simulated Data, Using Historical Productivity & Tax Shocks Dependent variable is ln(RESIt)7 (standard errors in parentheses) Model Constant Time trend ln(Pnt) ln(MPKt) ln(t) ln(H t) ln(rfrt-1) ln(Pt) -0.96 (0.06) 0.98 (0.01) 1.11 (0.02) -0.29 (0.02) (1) 0.78 (0.38) 3.31 (0.19) 1.01 (0.47) (2) 0.41 (0.45) 0.010 (0.00) 0.46 (0.75) 2.00 (0.48) (3) -1.11 (0.29) 0.004 (0.00) 2.07 (0.76) 0.25 (0.23) (4) -1.21 (0.29) 0.003 (0.00) 0.24 (0.22) 2.29 (0.73) 0.877

R2

0.999

0.872

0.904

0.871

In the various regressions is the adjustment cost parameter ( = 1 - h = 0.863), RESIt is residential investment at date t, Pnt is the price at date t of a new unit of housing delivered at t+1, MPKt is the marginal product of capital at t, t is the term involving sectoral capital stocks (see the appendix for details), Ht is the stock of housing at t, rfrt is the interest rate on a risk-free tax-free bond that is bought at date t and delivers one unit of consumption at t+1, and Pt is the price at date t of a unit of housing that may be enjoyed in date t (we think of this as the model counterpart to the empirical price of housing). The equation estimated in the Model column of the table is an equilibrium relationship. The model implied co-efficients on the house price term and the marginal product of capital are plus and minus one respectively.