Download Local Risk, Local Factors, and Asset Prices∗

Local Risk, Local Factors,
and Asset Prices
Selale Tuzely
Miao Ben Zhangz
October 2015
Abstract
This article provides a new link between …rm location and stock returns. We show that
the industrial composition of a local economy, in particular, how cyclical its industries
are, a¤ects …rm risk. We propose a metric of this cyclicality, labeled “local beta”, and
demonstrate that local factor prices such as wages and real estate prices are more sensitive
to aggregate shocks in areas with high local beta. While procyclical wages provide a natural
hedge against aggregate shocks and lead to lower …rm risk, procyclical prices of real estate,
which is part of …rm’s assets, should increase …rm risk. We con…rm that …rms located in
higher beta areas have lower industry-adjusted returns and conditional betas, and show
that the e¤ect is stronger among …rms with low real estate holdings. A production-based
equilibrium model explains these empirical …ndings.
JEL classi…cation: D24, G12, J21, R30
Keywords: Industry Composition, Local Factor Prices, Local Beta, Cross Section of Returns
We have bene…ted from conversations with Ken Ahern, Frederico Belo (discussant), Jonathan Berk, Sandra
Black, Daniel Carvalho, Riccardo Colacito, Ed Coulson, Ian Dew-Becker (discussant), Andres Donangelo, Wayne
Ferson, Mike Gallmeyer, Mark Grinblatt, Daniel Hirschleifer, Ayse Imrohoroglu, Chris Jones, Chang-Mo Kang,
Leonid Kogan, Erica Li, Debbie Lucas, Hanno Lustig (discussant), Oguzhan Ozbas, Vincenzo Quadrini, Clemens
Sialm, Sheridan Titman, Hakon Tretvoll (discussant), John Wei (discussant), Jiro Yoshido, two anonymous referees, Ken Singleton (the editor), and seminar participants at McGill, UC Irvine, Oregon, Penn State, CKGSB,
UVA McIntire, USC, UT Austin, Jackson Hole Finance Meetings, Western Finance Association Meetings, University of Washington Summer Finance Conference, CAPR Workshop at BI, USC-UCLA-UCI Finance Day, SFM
Conference, and the Econometric Society Meetings. We thank Joseph Engelberg for sharing his data on …rm
headquarter location changes and Diego Garcia and Oyvind Norli for providing data on geographic locations of
…rm operations.
y
Department of Finance and Business Economics, Marshall School of Business, University of Southern California. E-mail: [email protected].
z
Department of Finance, McCombs School of Business, University of Texas at Austin.
E-mail:
[email protected]
Most work in the …nance literature treats labor and capital markets as perfectly competitive
and homogenous at the aggregate level, where wages and rental rates equalize across di¤erent
locations. In reality, workers face frictions when moving from place to place (e.g., transaction
costs in housing, job search frictions, family coordination issues, etc.). Moreover, signi…cant
parts of physical capital, such as land and structures, are completely immobile. Such production
factors that are subject to geographical immobility, namely local factors, account for a large
part of the economic output.1 Fluctuations in factor prices due to local economic conditions
can have important e¤ects on the …rms using them.
In this paper, we show that local factor prices respond to aggregate shocks di¤erently across
localities based on the types of industries that dominate the areas. Conceptually, if local factors
are immobile across areas, the market for these factors clears within each area. Local factor
prices, e.g., wages and real estate rents, consequently aggregate the shocks to the …rms in the
area since all …rms tap into the same local labor pool and real estate market. If the major
industries that drive the economy of an area co-vary highly with the aggregate economy, local
factor prices in that area are likely to be pro-cyclical. For example, accommodation, a fairly
cyclical industry, is the main industry in three metropolitan areas: Las Vegas, Norwich-New
London, and Atlantic City. Its signi…cant weight in these economies makes the areas have
some of the most cyclical economies in the U.S. During the 2009 economic downturn, real
wages in these three metropolitan areas declined more than twice the national average. More
importantly, the excess wage decline in these areas relative to the national average manifests
not only in the accommodation sector but also in unrelated sectors such as plastics, nonmetallic
mineral, and transportation equipment manufacturing. Similarly, these areas also experienced
much larger declines in rents and real estate prices compared to other metropolitan areas in
2009. Larger ‡uctuations in input prices during economic cycles have implications for the risk
and returns of …rms located in these areas. We explore these implications in two steps.
We propose a metric of local risk, which we label “local beta,” to measure how cyclical
the local economy is. We compute local beta of a metropolitan statistical area (MSA) as the
average of industry betas, weighted by the industry shares in the local market. The industry
beta is the beta of industry output on the aggregate GDP. We con…rm that the sensitivity of
1
The estimates for the output share of labor range from 60% (Cooley and Prescott, 1995) to 75% (Imrohoroglu
and Tuzel, 2012). Campbell (1996) uses 2/3. The output share of land and structures is roughly 15% (Tuzel,
2010). The two local factors jointly claim more than 75% of the total economic output.
1
wage growth (within industry or occupation) to aggregate shocks is higher in MSAs with high
local beta. Similarly, we …nd more pro-cyclical house prices, commercial real estate prices, and
rents in MSAs with high local beta, re‡ecting a more cyclical demand for real estate in these
areas. These …ndings are evidence that industry composition is an important driver of the
heterogeneous ‡uctuations of local factor prices.
We then explore the implications of the heterogeneous ‡uctuations in local factor prices
for …rm risk by comparing …rms that are in the same industry but located in di¤erent areas.
Intuitively, if all …rms in the same industry are subject to the same aggregate productivity
shocks, their locations will a¤ect their risk and equity returns only through the local production
factors. Two competing channels are at work here. On one hand, more cyclical wages absorb
part of the aggregate shocks. This provides a natural hedge to …rms in high beta areas and
lowers their risk relative to their industry peers located in low beta areas. On the other hand,
real estate values also respond more signi…cantly to aggregate shocks in high beta areas than in
low beta areas. Since …rm value is partly derived from the value of its capital, which includes
real estate, this mechanism implies higher equity risk for …rms in high beta areas than in low
beta areas. Therefore, for …rms that hold real estate, the two channels have opposite e¤ects on
…rm risk.
We examine the empirical relation between local beta and …rm risk and highlight the importance of both channels. We …nd that local beta negatively predicts conditional equity betas
and future equity returns and also con…rm that this predictability is particularly strong for
the subsample of …rms with few real estate holdings and gets weaker and insigni…cant for the
subsample of …rms with high real estate holdings. To obtain these results, we use three di¤erent
econometric speci…cations: Panel regressions with time-industry …xed e¤ects, Fama-MacBeth
regressions, and portfolio sorts. Our …ndings suggest that both the labor and real estate channels are at work while the labor hedging channel dominates for the average …rm.
In order to formalize these ideas, we develop a production-based equilibrium model with
local markets. In this model, …rms belong to either a low beta or high beta industry where the
industry beta is determined by the sensitivity of industry’s output to aggregate productivity
shocks. Local markets have di¤erent compositions of low beta and high beta industries. We also
introduce heterogeneity to the real estate intensity of the industries to capture their varying real
estate needs. All …rms produce a homogenous good, receive aggregate (economy-wide) and …rm2
level productivity shocks, and use three factors of production: labor, capital equipment, and
land (i.e., immobile capital and real estate). Labor and land are local factors of production with
limited supply, whereas equipment is not. Firms are ex-ante identical except for their location,
which determines the mix of …rms active in their local factor markets, and their industry
a¢ liation, which determines their exposure to aggregate productivity shocks and real estate
intensity. Wages and land prices clear the local markets and are determined endogenously,
conditional on the industry composition of the local market. Firms’ investment and hiring
decisions are also endogenously determined in equilibrium.
The calibration of the model generates the main empirical patterns observed in the data:
High beta areas have more pro-cyclical wages and real estate prices (i.e., higher covariance
between aggregate productivity and local factor prices) than low beta areas. In addition, more
procyclical wages act as a natural hedge against shocks, making the returns of …rms in high beta
areas less sensitive to aggregate shocks. Therefore, their expected returns are lower relative to
…rms from the same industry but located in low beta areas. This is especially true for …rms
that belong to industries with low real estate intensity.
In order to simplify our empirical and theoretical analysis, we make several assumptions
that merit some discussion. First, we assume that there is no local factor mobility (labor
and land) between di¤erent local markets. Though land is truly immobile, it is possible for
labor to move across markets in response to shocks. Nevertheless, at annual frequency, jobrelated mobility is very low. Between 2012 and 2013, for instance, only 3.8% of households
moved across county lines, and job-related moves make up roughly one third of these moves.2
Several factors contribute to low labor mobility, especially at higher (business cycle) frequency.
Moretti (2011) argues that in the short run, frictions in labor mobility and in the housing supply
constrain the ability of workers and housing stock to fully adjust to shocks. Frictions in land
and housing supply play an important role in the basic spatial equilibrium models (Rosen, 1979;
2
Estimates are from the Census Bureau report, Reason for Moving: 2012 to 2013. The other two major reasons
for moves, family-related (e.g., change in marital status) and housing-related (e.g., wanted better neighborhood)
reasons, each account for one-third of inter-county moves. Chen and Rosenthal (2008) investigate individual
migration decisions using IPUMS data. Consistent with the Census statistics, they …nd that among movers, an
important reason for moving is the availability of local amenities (e.g., climate, temperature, non-metropolitan
areas, etc.), which is not directly related to wages. In addition, Kennan and Walker (2011) develop an econometric
model of optimal migration, and estimate the moving costs to be $312,146 for the average hypothetical mover,
and $80,768 for the average actual mover. These …gures exceed expected income gains for most people, and
Kennan and Walker (2011) argue that many of the moves that do occur are motivated by reasons other than
other than income gains.
3
Roback, 1982). These models suggest that a shock to a local labor market is both re‡ected in
worker wages and capitalized into rents/housing and land prices. In this case, the movement
in house prices discourages labor mobility.3 Consistent with this view, we …nd evidence that
house prices, like wages, are also more sensitive to economy wide shocks in high beta areas.4
Therefore, inter-market labor mobility cannot fully absorb the di¤erential e¤ects of aggregate
shocks, leaving relative factor prices unchanged.
We also assume that the location choice of the …rms is exogenous. Starting with Marshall
(1920), there is a large urban economics literature that studies the causes and e¤ects of agglomeration. Likewise, the issue of industrial clustering is well documented and studied in the
literature.5 Most of the work in this area is geared toward understanding the di¤erences in clustering across industries rather than the individual …rm’s location decision within its industry.6
Our focus is the e¤ect of location on the risk of the …rm compared to its peers in its industry.7
While we focus on the geographic segmentation of the factors of production, we abstract from
similar segmentation in the …rms’ product markets. This condition is satis…ed for industries
that produce tradable goods where …rms’ products are not con…ned to be sold in the local
markets where they are produced. However, the sales of certain industries-the retail sector
in particular-are predominantly local (i.e., nontradable) and are naturally a¤ected by local
economic conditions. Mian and Su… (2012) document that job losses in the nontradable sector
during the Great Recession are signi…cantly higher in high leverage counties, implying that
worsening household balance sheets in those areas led to sharp demand declines for nontradable
goods. Therefore, local area characteristics such as local betas can impact both the input and
the output prices for these nontradable industries, making inference about …rm risk di¢ cult.
3
Recently, sharp declines in house prices and a breakdown of the Beveridge Curve (historically strong negative
relationship between job vacancies and unemployment rate) during the Great Recession sparked a debate about
whether the two observations are related. Sterk (2015) develops a general equilibrium model where declines in
house prices lower households’ home equity levels, reducing the geographic mobility of the unemployed homeowners.
4
We don’t model the household side and therefore don’t consider the feedback e¤ects between house prices
and wages. Nevertheless, the di¤erential e¤ect of aggregate shocks on local good prices such as housing would
work as an additional mechanism to amplify the relationship between local betas and the cyclicality of wages.
5
See Ellison, Glaeser, and Kerr (2010) for a recent contribution to this area.
6
Almazan, Motta, and Titman (2007) present a model of a …rm’s location choice in this category.
7
While we assume that the …rms’location choice is exogenous, in reality, they might be using this local beta
mechanism to manage their risk. Speci…cally, inherently risky …rms may locate in high beta areas to mitigate
their risk. However, this would result in higher risk for the …rms located in high beta areas, which goes against
our …ndings. Since we …nd the opposite, that means either the hedging e¤ect of cyclical wages in high beta areas
is actually stronger than what we measure in the data, so that we …nd these e¤ects despite the …rms’mitigating
endogenous location choices, or the …rm’s location choice is exogenous to our mechanism.
4
Consistent with our hypothesis, we con…rm that the relationship between local betas and …rm
risk is indeed stronger for the …rms that produce tradable goods.
We examine other predictions for how local betas are likely to impact labor and asset
markets. In particular, we …nd that the positive relationship between local betas and the
procyclicality of wages is stronger for industries with lower unionization rates, consistent with
the view that unionization increases the frictions in wage adjustments. We also …nd that the
negative relationship between local betas and expected equity returns are mainly found in …rms
with geographically focused operations for which …rm headquarter location is a better proxy
for where the …rm actually operates.
We perform several robustness checks for our main …ndings. First, we test the robustness
of our results to two alternative measures of local beta. The …rst measure aggregates industry
betas, which are computed as the beta of industry total factor productivity (TFP) on the
aggregate TFP. The second measure is estimated directly using MSA GDPs. Next, we test the
asset pricing implications using an expanded sample period, starting in 1970, to ensure that our
results hold beyond the main sample period of 1986-2011. Third, we change the asset pricing
test assets from individual …rm returns to industry-MSA portfolio returns to ensure that our
results are not driven by few outlier …rms. Finally, we check the robustness of the regression
results to various assumptions for the correlation structure of the residuals. We con…rm that
the results are robust to these variations.
Our paper is related to a growing strand of literature that studies the connection between
…rms’location, economic activity, and …nancial performance. On the real side, Dougal, Parsons,
and Titman (2015) document that …rms’investments are sensitive to the investments of other
…rms headquartered in the same area. They argue that the shared local vibrancy facilitate the
positive externality among …rms in the same area. Engelberg, Ozoguz, and Wang (2010) …nd
that fundamentals of …rms in industry clusters have stronger comovement. They interpret this
…nding as a possible outcome of …rms’exposure to the same local labor markets. Chaney, Sraer,
and Thesmar (2012) study the e¤ect of changes in the value of real estate portfolios of …rms
on those …rms’ investments. Speci…cally, they calculate the change in the real estate values
based on the changes in property prices in …rms’headquarter locations. Our analysis proposes
a source for heterogeneity in real estate prices by exploring the cyclicality of the local economy.
5
On the …nancial side, several papers investigate the relationship between …rms’ location
and stock returns. Pirinsky and Wang (2006) study the correlations between stock returns
of …rms headquartered in the same area and …nd that their returns move together. Garcia
and Norli (2012) show that the returns of geographically focused …rms exceed the returns of
geographically dispersed …rms. Korniotis and Kumar (2013) document that local economic
conditions are useful in predicting the returns of …rms in that area. The common theme in
these studies is that …rms in the same area share the same group of investors. Our paper
complements this body of work by exploring a channel in which …rms in the same area share
the same factors of production.
Our paper is also related to the growing body of work in asset pricing in which asset returns
are tied to the labor market frictions. Danthine and Donaldson (2002); Gourio (2007); Berk and
Walden (2013); and Favilukis and Lin (2012a, b) among others study the e¤ects of wage rigidity.
In an equilibrium setting, wage rigidity leads to labor-induced operating leverage, resulting in
volatile and cyclical pro…ts accompanied by high and countercyclical risk premia. While there
is no exogenously imposed wage rigidity in our model, there are endogenous variations in wage
‡uctuations across local labor markets, which lead to a more subtle form of operating leverage.
Implications of labor adjustment costs are investigated in Merz and Yashiv (2007); Belo, Lin,
and Bazdresch (2014); Belo and Lin (2012); among others. In the presence of labor adjustment
costs, the …rms’market values and expected returns are related to their hiring behavior. Chen
and Zhang (2011) and Kuehn, Petrosky-Nadeau, and Zhang (2013) study asset pricing with
labor market search. Search frictions in the labor market endogenously generate volatility in
unemployment rates and wage rigidity and match the dynamics of equity returns. Donangelo
(2014) studies the implications of labor mobility on asset pricing, …nding that industries that
rely on a more ‡exible labor force face greater risk. Our work adds to this literature by
incorporating a previously unexplored friction, the local nature of labor markets. We address
the heterogeneity in local labor markets and investigate its implications on the …rms operating
in these markets.
Finally, our paper contributes to the literature on capital heterogeneity and asset pricing.
Eisfeldt and Papanikolaou (2011) consider organization capital; Hansen, Heaton, and Li (2005)
and Li and Liu (2012) investigate intangible capital; and Belo and Lin (2012) and Jones and
Tuzel (2013) study the implications of having inventories as part of …rms’ productive assets.
6
Tuzel (2010) studies the asset pricing implications of …rms’ real estate holdings, …nding that
…rms that own more structures (real estate) are less ‡exible, hence riskier, and earn higher risk
premia. In this paper, we study the implications of local beta on the propagation of aggregate
shocks to local real estate prices and examine how this mechanism a¤ects …rms’returns. We
show that in high beta areas, real estate holdings of …rms magnify the e¤ects of aggregate
shocks, hence adding to the “risky real estate” argument in Tuzel (2010).
The paper is organized as follows. Section 1 presents a simpli…ed version of the model with
closed-form solutions with the goal of building intuition. Section 2 describes the data used in
our empirical analysis and introduces our local beta measure. Section 3 presents our empirical
results relating GDP shocks and local betas to wages, real estate returns, and …rm returns.
Section 4 presents our full model and quantitative results. Section 5 concludes.
1
Simpli…ed Model
In this section we present a simpli…ed version of the model where labor is the only factor of
production and discuss the basic intuition. Section 4 develops the full model.
Each local market, m, is populated by a measure one of workers, and a continuum of
measure one of in…nitely lived …rms producing a homogeneous good using labor. There are two
industries in the economy where one is more cyclical than the other, j 2 f1; 2g. Local markets
di¤er in their compositions of the two types of industries. Let sm 2 (0; 1) denote the fraction
of …rms that belong to the high cyclicality industry in market m.
The output of …rm i in industry j of market m during time t is given by:
I
l
Yijm;t = At j Zit Lijm;t
:
(1)
Lijm;t denotes the labor used in production by …rm i. Labor share in the …rm’s production
function is given by
l,
where
l
2 (0; 1): Aggregate productivity is denoted by at = log (At ) :
at follows the AR(1) process:
at+1 =
where "at+1
i.i.d. N 0;
2
a
a at
+ "at+1
(2)
. Ij is a scaler that represents the industry cyclicality and scales
7
the e¤ect of the aggregate productivity on the …rm’s production, where I1 = Ihigh > 1 and
I2 = Ilow < 1: The idiosyncratic …rm productivity, zit = log(Zit ), follows the AR(1) process:
zi;t+1 =
where "zi;t+1
i.i.d. N 0;
2
z
z zit
+ "zi;t+1
(3)
. "zi;t+1 are also uncorrelated across …rms.
Firms have access to the local labor markets, which are competitive. Labor is free to move
between …rms in the same area, but not across di¤erent labor markets.8 Therefore, the marginal
product of labor is equalized among …rms in the same area. Hiring decisions are made after
…rms observe the productivity shocks and labor is adjusted freely. Each period the …rm chooses
the amount of labor, Lijm;t ; to maximize operating pro…ts of the …rms, which is output minus
labor expenses:9
ijm;t
= Yijm;t
Wm;t Lijm;t
(4)
where Wm;t denotes the wage rate in local labor market m at time t.
The …rst order condition for the …rm’s optimization problem leads to the …rms’labor choice:
Wm;t =
Ij
l 1
l At Zit Lijm;t
1
1
l
Lijm;t = Zit
Ij
l At
Wm;t
(5)
!
1
1
l
The aggregate labor supply in each market m is normalized to be 1. In equilibrium, the
aggregate labor demand in each market equals the supply,
XZ
Lijm;t di = 1;
(6)
j
8
We use the terms “local market” and “local area” interchangeably throughout the paper.
In this setting, the …rm’s problem is a static pro…t maximization problem. Maximizing period pro…ts is
equivalent to maximizing …rm value.
9
8
where j 2 f1; 2g: Using lognormality of Zit , we can derive the equilibrium wage rate as:
Ihigh
1
l
Wm;t =
where
=
le
2(1
2
z
2 )(1
z
l)
s m At
+ (1
Ilow
1
l
s m ) At
!1
l
(7)
is a constant.
Wage rate increases in aggregate productivity, At , therefore it is procyclical. More importantly, wage cyclicality increases with sm ; the fraction of …rms that belong to the high risk
industry. Replacing Wm;t and Lijm;t in equation 4 leads to maximized …rm pro…ts:
Ij
1
ijm;t
1
where
= (1
l) e
l
= At
2(1
2)
z (1
Ihigh
1
l
1
1
l
Zit
s m At
+ (1
Ilow
1
l
sm ) At
!
l
(8)
2
z l
2
l)
is a constant.
Let’s de…ne the elasticity of …rm pro…ts to aggregate productivity,
ijm;t ;
as a measure of
…rm risk:10
@
ijm;t
=
ijm;t
@At
=
"
sm
l Ihigh
1
l
Ihigh
1
l
At
Ihigh
1
l
ijm;t
At
Equation 9 shows that
risk industries), and
Ij
ijm;t
sm At
ijm;t
+ (1
+ (1
sm )
Ij
l Ilow
1
Ilow
1
l
s m ) At
l
!
Ilow
1
l
At
#
(9)
> Ihigh > 1 for …rms with Ij = Ihigh (i.e., …rms from the high
< Ilow < 1 for …rms with Ij = Ilow , regardless of the local markets.
The wages in the local market, which is a mix of high risk and low risk industries, are less
cyclical than the output of high risk industries, and more cyclical than the output of low risk
industries. This leads to even higher (lower) cyclical variation in the pro…ts of high (low) risk
industries.
Now let’s de…ne the sensitivity of …rm risk,
ijm;t ,
to the share of high risk industries in a
local market, sm . Note that higher sm implies a riskier local market.
10
Donangelo (2014) de…nes a similar risk measure and calls it labor-induced operating leverage.
9
l
@
ijm;t
sm
1
=
l
Ilow +Ihigh
1
l
(Ilow
Ihigh ) At
Ihigh
1
l
s m At
Since Ilow
+ (1
Ilow
1
l
s m ) At
(10)
!2
Ihigh < 0 and all other terms are positive, we …nd
@
ijm;t
sm
< 0, implying that
in areas with higher share of high risk industries, the risk of the individual …rms is lower than
their industry peers located in areas with lower share of high risk industries. This geographical
comparison holds for …rms within both high risk and low risk industries. Wages are more
cyclical in local markets where the majority of …rms belong to the high risk industry (higher
sm ) and serve as a hedge against ‡uctuations in aggregate productivity. On the other hand,
local markets where the majority of …rms belong to the lower risk industry, i.e., lower sm
areas, experience smaller ‡uctuations in wages, which leads to the ampli…cation of the e¤ect of
aggregate shocks on …rm pro…ts relative to industry peers, resulting in higher risk. Thus, the
industrial composition of the local market renders a novel operating leverage channel.
It is useful to emphasize that while the risk of individual …rms, conditional on their industry,
is lower in markets with a higher share of high risk industry, unconditionally, …rms in these
areas are still riskier than the …rms elsewhere.11 Therefore, throughout the paper, all …rm level
analysis is performed conditional on the …rm’s industry.
The static nature of the simpli…ed model with only labor allows us to make predictions,
based on closed form expressions, on how the riskiness of the local economy a¤ects the wage
cyclicality and the risk of the …rms operating there. In the dynamic full model presented in
Section 4 …rms own and use land, another local factor, and equipment, in addition to labor.
The full model generates additional predictions based on the input composition of …rms, with
a focus on land holdings, and quanti…es the qualitative insights of the present section.
11
To show this, we write total pro…ts of local markets as the sum of pro…ts of individual …rms located in that
area,
m;
and de…ne
m
=
@ m
@At
m
At
:
@ m
@sm
Ihigh +Ilow
1
l
=
0
At
Ihigh
1
B
l
@sm At
(Ihigh
+[1
are riskier in areas with a higher share of high risk industries.
10
Ilow )
12
Ilow
1
l
sm ]At
C
A
> 0, implies that, unconditionally, …rms
2
Data and Measurement
In this section we describe the data used in the paper and discuss the measurement of critical
variables. To conduct the empirical analysis we combine a number of data sets. Appendix
Table A1 summarizes all datasets used in the paper. Internet Appendix provides more details
about data sources and variable construction.
Local Beta The key variable in this paper is the beta of local economies,
local
m .
We compute
local beta as the average of the GDP betas of the industries operating in that area, weighted
by the employment share of the industries. Speci…cally,
local
m;t
=
X
wi;m;t
ind
i;t ;
(11)
i
where wi;m;t represents the employment share of industry i in market m in year t, and
ind
i;t
represents the beta of industry i in year t.
We classify the local markets by Metropolitan Statistical Areas (MSA). MSAs are geographic
entities de…ned by the O¢ ce of Management and Budget that contain a core urban area of 50,000
or more population. Each MSA consists of one or more counties that contain the core urban
area, as well as adjacent counties that have a high degree of social and economic integration
(as measured by commuting to work) with the urban core.12 There are 373 unique MSAs in
our sample.
To calculate the weight of industries in each MSA, we collect industry employment data at
the MSA level from the County Business Patterns (CBP) dataset published by the U.S. Census
Bureau.13 We compute industry share (weight) wi;m;t as the ratio of each industry’s employment
in an MSA to the total reported MSA employment in year t. While most MSAs have diverse
economic base featuring many industries, there is heterogeneity in industrial diversity of MSAs.
Figure 1 plots a histogram of industrial dispersion of employment within each MSA, computed
12
The term “Core Based Statistical Area” (CBSA) refers to both metro and micro areas. Currently,
the Census Bureau uses the MSA and metro CBSA terms interchangeably. For more information, see
http://www.census.gov/population/metro/.
13
Disaggregated data is at times suppressed for con…dentiality reasons. However, in these situations, the
Census Bureau provides a “‡ag” that tells us of the range within which the employment number lies. Like Mian
and Su… (2012), we take the mean of this range as a proxy for the missing employment number in such scenarios.
11
as a Her…ndahl index of industry employment shares.
To calculate industry betas, we obtain annual data on the industry value added, as a measure
of industry output, from the Bureau of Economic Analysis (BEA). Industry shock is the growth
in the real industry value added where nominal data are de‡ated by GDP de‡ators to calculate
real value added. Industry betas
ind
i;t
are calculated as the slope coe¢ cients from the regressions
of industry shock (real industry value added growth) on aggregate shock (real GDP growth),
using data up to year t. Table 1 reports the industries with the highest and lowest betas in 2011.
The industries with the lowest betas operate broadly in the food manufacturing, health care, and
oil sectors. These industries have negative or near zero betas in our sample. The industries with
the highest betas operate in heavy manufacturing (primary metal, transportation equipment,
nonmetallic mineral and wood) or the …nancial sector, with betas around 3. Replacing industry
employment weights, wi;m;t ; and industry betas,
ind
i;t ,
in equation 11, we obtain MSA betas
over 1986-2011.
Figure 2 plots the histogram of MSA betas as of 2011, last year of observation in our sample.
Most MSA betas are between 0.8 and 1.2, and betas are positively skewed. Table 2 summarizes
the MSAs with the lowest and the highest betas as of 2011 to gain more perspective into local
betas. MSAs with the highest local betas are typically dominated by the heavy manufacturing
industries and accomodation sector, while those with the lowest local betas have economies
based on food manufacturing, health care, or education service industries. There is no particular
relationship between the local beta and size of an MSA (as measured by employment). The
correlation coe¢ cient between local beta and employment, computed using the sample of all
MSAs in 2011, is less than 0.1.
Figure 3 compares the economic performance of the MSAs with the highest and lowest
local betas over the period of 2001-2011. We obtain data of each MSA’s real GDP from the
BEA.14 The top panel plots the average real GDP of the two groups of the MSAs along with the
national GDP. The bottom panel plots annual real GDP growth. The …gures show that high
beta areas experienced steady growth during the expansion years until 2007, but experienced
a bigger reduction in both GDP levels and growth during the great recession (2008-2009). The
lowest beta areas, on the other hand, experienced neither a big increase nor a signi…cant drop
14
To the best of our knowledge, there is no publicly available data before 2001. This is also one of our main
motivations for constructing our benchmark measure of local beta from industry betas as in equation 11.
12
in output over the same time period. These …ndings support the validity of our local betas,
constructed from local industry shares and industry betas, to capture the economic risk of local
areas.
Factor Prices We measure local factor prices, wages and real estate prices, using data from
several di¤erent sources. We obtain wage data at MSA by industry level from the Quarterly Workforce Indicators (QWI) dataset of the Longitudinal Employer-Household Dynamics
(LEHD) program at the U.S. Census Bureau.15 We also study hourly occupational wages for
metropolitan areas from the Occupational Employment Statistics (OES) program of the Bureau
of Labor Statistics. We use both broad occupation de…nitions with 22 major occupation groups
and detailed occupation de…nitions with 854 detailed occupations.
We measure MSA housing prices with the house price indexes (HPI) from the Federal
Housing Finance Agency. Commercial real estate returns take into account the income and
appreciation of all commercial property types (o¢ ce, retail, industrial, apartment, and hotel)
and are provided by the National Council of Real Estate Investment Fiduciaries (NCREIF
NPI). Commercial real estate rent data are from CoStar.16
Firm Location For our …rm level analysis, we identify a …rm’s location with its headquarter location from Compustat.17 We supplement the data with headquarter location change
information from Compact Disclosure, compiled by Engelberg, Ozoguz, and Wang (2010).18
To further assess the validity of this identi…cation, we link our Compustat-CRSP sample
to ReferenceUSA U.S. Businesses Database and collect employment data for all headquarter,
branch, and subsidiary locations of …rms in our sample.19 This allows us to create an employ15
The main advantage of using QWI data over other sources, such as the CBP or QCEW, is that QWI reports
average wages for virtually all industries in all areas, whereas CBP and similar programs do not disclose wages
for many industry-area combinations for con…dentiality reasons.
16
HPI and CoStar rent data are available at the MSA level. NPI is available at the MSA level for most areas,
and at the metropolitan division level for 11 MSAs, which are subgroups of MSAs. For those areas, we take the
averages of HPI returns for metropolitan divisions and use that as a measure for the MSA return.
17
Chaney, Sraer, and Thesmar (2012) argue that headquarters and production facilities tend to be clustered
in the same state and MSA and headquarters represent an important fraction of corporate real estate assets.
They hand-collect information on …rm headquarter ownership using their 10K …les, and …nd that …rms that
report headquarter ownership also have positive real estate ownership based on Compustat data. Therefore,
they conclude that headquarter location is a reasonable proxy for …rm location.
18
Compustat reports only the most recent headquarter location of …rms. Compact Disclosure discs provide
current headquarter location of …rms and covers the years 1990-2005. There are roughly 300 headquarter location
changes over this time period.
19
We collect the most recent employment numbers from ReferenceUSA U.S. Businesses Database in November
13
ment map for each of roughly 2000 …rms in the linked sample.20 We …nd that 63% of the …rms
in our linked sample have at least 50% of their employment in their headquarter MSA. For the
median …rm in our sample headquarter location accounts for 72% of total employment. While
60% of the …rms have more than half of their employees in their headquarter MSA, there is
signi…cant heterogeneity among …rms of di¤erent size (market capitalization). Appendix Table
A2 shows that headquarter MSA is a much better location proxy for smaller …rms. Sorting
…rms into size quintiles, we …nd that almost 80% of …rms in the smallest size quintile have more
than half of their employment in their headquarter MSA, and about 55% of these …rms have
virtually all their employment in the same MSA. For the …rms in the biggest size quintile, those
statistics are roughly 50% and 10%. To address the concern that headquarter location may be
a noisy measure of where the …rm operates and owns assets, we will also examine our …rm level
results with a subsample of smaller …rms.
In …rm level regressions, we make all comparisons within the industry. To check whether
…rms from an industry have their headquarters all clustered in one or few MSAs, we compute a
measure of the geographic dispersion of …rms in each industry.21 Figure 4 plots the histogram
of this geographic dispersion metric. The …gure shows that most industries have large variation
in …rm locations, while few industries are more geographically focused, yet still include …rms
from several di¤erent MSAs.
Financial and Accounting Data Data on real estate holdings and …rm employees are
from Compustat. We apply standard …lters to the Compustat data and exclude …rms without
positive sales (SALE) and assets (AT). Following Fama and French (1993), in order to avoid the
survival bias in the data, we include …rms in our sample after they have appeared in Compustat
for two years. Following Tuzel (2010), we measure the real estate holdings of the …rms as the
sum of buildings (FATB) and capitalized leases (FATL). We replace missing values with zero.
To calculate the real estate ratio (RER), we scale the real estate holdings with the number of
employees (EMP).
2014 for businesses that are active at that time.
20
Note that this is the employment map created from ReferenceUSA database. If ReferenceUSA misses any
of the …rms’ establishments, employees of those establishments don’t show up in our employment map. Since
ReferenceUSA only reports domestic establishments, international employees of the …rms are not included in the
employment map.
21
Speci…cally, the geographic dispersion measure is constructed as the Her…ndahl index of how the number of
…rms in an industry (from Compustat) are divided among the MSAs.
14
Monthly stock returns are from the Center for Research in Security Prices (CRSP). Similar
to Fama and French (1993), our sample includes …rms with ordinary common equity as classi…ed
by CRSP, excluding ADRs, REITs, and units bene…cial interest. We match CRSP stock return
data from July of year t to June of year t + 1 with accounting information (Compustat) for
…scal year ending in year t
1 as in Fama and French (1992, 1993), allowing for a minimum of
a six month gap between …scal year-end and return tests.
3
Empirical Analysis
In the …rst part of our empirical analysis, we study the e¤ect of aggregate GDP shocks on local
factor prices, in particular on wages and real estate prices, conditional on the beta of the local
market. In the second part, we study the relationship between local beta and …rms’risk and
returns.
3.1
Local Factor Prices
Our …rst hypothesis is that business cycles (shocks to aggregate GDP) will a¤ect the wages in
an area more if the local aggregate of industries that operate in that market is prone to business
cycle shocks; i.e., if the area has a high local beta. This prediction is theoretically derived in
equation 7. We test this hypothesis and report the results in Table 3. Speci…cally, we run the
following panel regressions with …xed e¤ects:
wageind;m;t = b0 + b1 shockt
+Time
where
local
m;t 1
+ b2
local
m;t 1
(12)
Industry Dummies + MSA Dummies +
ind;m;t ;
wageind;m;t is the percent change in wage per employee in each industry, market (MSA),
and year triplet; shockt is the aggregate real GDP growth in that year;
local
m;t 1
is the local beta
constructed from the GDP betas of the industries operating in that area, computed as in
Section 2.22 Due to time-industry …xed e¤ects, coe¢ cient estimates re‡ect the variation within
the same industry and year. We expect to …nd a positive estimate for the interaction term,
b1 , implying that wage growth in high beta areas covaries more with GDP shocks, relative to
22
The complete speci…cation also includes shockt as an additional regressor, which drops due to time …xed
e¤ect in the regression.
15
wage growth in the same industry but lower beta areas. We present results for several di¤erent
speci…cations for the entire sample, subsamples, and di¤erent controls. Panel A uses data from
LEHD, our main data source for wages at the MSA-industry level, and Panel B presents the
results using MSA-occupation level hourly wage data from OES.23 We cluster standard errors
at the MSA-industry level.
Consistent with our hypothesis, we …nd that the interaction term is uniformly positive
and signi…cant. In our main speci…cation using LEHD data (columns 1 and 2 in Panel A), the
estimates imply a roughly 15 basis points di¤erence in wage growth for a one standard deviation
increase in real GDP (2.5%) between MSAs in the highest and lowest beta quintiles (0.25 beta
spread). We have similar …ndings based on occupational wage data. Columns 1 and 2 of Panel
B present results for 22 major occupation groups, whereas columns 3 and 4 present results
for 854 detailed occupation de…nitions, where estimates are both economically and statistically
more signi…cant.
An implicit assumption in our hypothesis is that labor markets are competitive and there
are no major frictions to the adjustment of employment or wages. A violation of this condition
may arise due to the prevalence of labor unions in certain industries. In the context of wages,
Kimbell and Mitchell (1982) report that labor contracts in unionized industries are characterized
by multi-year contracts with built-in in‡ation adjustments. Chen, Kacperczyk, and OrtizMolina (2011) argue that the presence of powerful unions substantially reduces …rms’operating
‡exibility. In order to mitigate these potential concerns due to union involvement, we also
consider subsamples of non-unionized industries and occupations.24 We de…ne non-unionized
industries (occupations) as industries (occupations) with unionization rates lower than the
median unionization rate of all industries (occupations) in that year.25 We expect our main
…ndings to hold more strongly for non-unionized industries and occupations.26 We report the
23
Both samples have their advantages. LEHD data is disaggregated to industry level, whereas OES data
is disaggregated to occupation level, and hence includes occupation, rather than industry controls. Another
di¤erence between the two data sets is LEHD includes total wages for the period, whereas OES includes hourly
wages.
24
We obtain data on the unionization rate for industries and occupations from www.unionstats.com, compiled
by Barry Hirsch and David Macpherson from the Current Population Survey and updated annually. The database
is described in Hirsch and Macpherson (2003). The industry and occupations are based on Census codes. We
use crosswalks between the Census industry and occupation codes used in the unionization dataset and NAICS
industry classi…cation codes in LEHD and Standard Occupational Classi…cation (SOC) codes used in OES wage
datasets.
25
Our results are qualitatively similar when we use di¤erent cuto¤s for unionized industry de…nitions.
26
OES and unionization data use di¤erent occupation classi…cations. We can match roughly 2/3 of the 854
detailed occupation de…nitions in OES to the unionization data. The unmatched occupations are included in the
16
regression results for non-unionized industries in columns 3 and 4 of Panel A and the nonunionized occupations in columns 5 and 6 of Panel B. We …nd that excluding highly unionized
industries and occupations from our main sample strengthens the results and slightly increases
the magnitudes of the coe¢ cients for the cross terms.
In our benchmark sample, we do not distinguish industries based on the geographic segmentation in their product markets. Our implicit assumption is that local beta does not have a
substantial e¤ect on the output demand/prices of the …rms, which may not hold for industries
that produce nontradable goods that are sold to locals. Therefore, we consider a subsample that
excludes the nontradable industries. Following Mian and Su… (2012), we de…ne nontradable industries as the retail sector and restaurants (SIC 52-59, NAICS 44-45, and 722) and create a
subsample of tradable industries by excluding these.27 The results, presented in columns 5 and
6 of Panel A, show that our main results hold, and even get slightly more signi…cant for the
sample of tradable industries.
Overall, Table 3 demonstrates that local wages are more sensitive to systematic shocks in
local markets with higher betas. The signi…cant weight of risky industries in the high local
beta areas ampli…es the e¤ects of aggregate shocks to the wages of all employees in the area,
regardless of their industry and occupation a¢ liation. This implies that employees in high beta
areas are more exposed to aggregate shocks than their counterparts in low beta areas. Even
though we take the location choice of the employees as exogenous, in equilibrium, employees
should be indi¤erent between locating to di¤erent areas, at least in the long run. To the extent
that employees care about their labor income risk, they should require to earn higher wages in
high beta areas. The last columns of Table 3 (columns 7 and 8) investigate this hypothesis. In
Panel A, we regress the level of annual wages on local betas controlling for year-industry …xed
e¤ects; in Panel B, we regress hourly wages and control for year-occupation …xed e¤ects. We
…nd that wages rise with local beta and most estimates are highly signi…cant. Wages in the
MSAs in highest beta quintile are approximately $1000 to $2700 higher (annually) than their
counterparts in the lowest beta quintile MSAs, in 1990 dollars.
Beside wages, aggregate shocks should have a bigger impact on real estate prices and rents
in high beta areas. Commercial real estate is a major local input to the …rms, so good (bad)
low unionization subsample for a conservative estimate.
27
Identifying nontradable industries is not a trivial task. Even the retail sector, which is the prime example of
nontradable sector, may not be completely nontradable due to non-local sales through the internet, catalog, ...
17
shocks would lead to increased (decreased) demand for this type of assets. Since the supply of
commercial real estate is inelastic in the short run, change in demand should have an impact on
the prices and market rents, and variations in demand will be bigger in high beta areas. Beside
commercial real estate, systematic shocks could also have a bigger e¤ect on house prices in
high beta areas due to two separate channels. The …rst channel is due to increased (decreased)
demand for housing from households as a result of increasing (decreasing) wages in the area.
The second channel is due to spillovers from the increasing (decreasing) commercial real estate
prices, since both types of real estate share a common input, land.
In order to test the e¤ect of aggregate shocks on real estate returns, we run the following
panel regressions with …xed e¤ects:
re
rm;t
= b0 + b1 shockt
local
m;t 1
+ b2
local
m;t 1
+Time Dummies + MSA Dummies +
(13)
ind;m;t :
re represents housing returns in columns 1 and 2,
The results are presented in Table 4. rm;t
commercial real estate returns in columns 3 and 4, and commercial real estate (o¢ ce buildings)
rent growth in columns 5 and 6 of Table 4. Like wage regressions, we cluster the standard errors
at the MSA level. We expect to …nd a positive estimate for the interaction term, b1 , implying
that real estate returns in high beta areas are more sensitive to GDP shocks relative to their
counterparts in low beta areas. Table 4 presents results with and without MSA …xed e¤ects.
Consistent with our hypothesis, we …nd that the interaction terms are positive and signi…cant in
all speci…cations. The coe¢ cient estimates imply roughly a 0.6% di¤erence in housing returns,
2.5% di¤erence in commercial real estate returns, and 1.5% di¤erence in commercial real estate
rent growth, respectively, between MSAs in the highest and lowest beta quintiles, in response
to a one standard deviation change in real GDP.
Our panel regression results in Table 3 and 4 show that prices of local factors of production, wages and real estate, are more sensitive to aggregate shocks in areas with more cyclical
economies. We also adopt an alternative methodology where we test the same hypotheses using
two-stage cross sectional regressions and …nd consistent results. The details of this speci…cation
and the results are available in the Internet Appendix.
18
3.2
Firm Level Results
Section 3.1 demonstrates that GDP shocks have a bigger impact on local factor prices such as
the wages and real estate prices in high beta areas compared to low beta areas. Since labor
and commercial real estate are major inputs to the …rms, the di¤erential e¤ect of the aggregate
shocks on the local input prices should be an additional channel for how these shocks a¤ect
…rms. We next study the e¤ect of this mechanism on the returns of the …rms located in areas
with di¤erent local betas.
The greater sensitivity of wages to aggregate shocks in high beta areas provide a natural
hedge for the …rms, mitigating the e¤ect of the shocks on the …rms, and lead to lower …rm
risk. At the same time, real estate values are more sensitive to shocks in those areas. Since
the …rm value is partly derived from the value of its capital, including corporate real estate,
this mechanism would imply higher …rm risk in high beta areas. So, for the …rms that own real
estate, the two channels have opposite e¤ects on the relationship between local beta and …rm
risk.28
We begin to investigate the implications of local betas for …rm risk by estimating conditional equity betas for …rms as in Lewellen and Nagel (2006). Conditional equity betas are
estimated using short-window regressions (one year) and monthly returns, and do not require
the speci…cation of conditioning information. We correct for non-synchronous trading following
the methodology described in Lewellen and Nagel (2006).29 We use conditional equity beta as
a proxy for the …rm’s risk, and examine the e¤ect of local betas on the conditional equity betas
by running panel regressions with …xed e¤ects30 :
cond
f irm;t
= b0 + b1
+Time
local
m;t 1
(14)
Industry Dummies + controlsf irm;t +
28
f irm;t
For …rms that do not own but lease real estate, leasing will create an additional hedging mechanism in the
presence of procyclical rents (assuming that market rent changes will be re‡ected to their leases, which would
be true if they are signing a new lease aggreement, or renegotiating their existing lease). The e¤ect would be
similar to the labor e¤ect.
29
Speci…cally, we regress monthly excess stock returns on the excess returns of the market, and one lag of the
market portfolio. Conditional beta is the sum of the coe¢ cients for the contemporaneous and lagged market
returns.
30
In panel regressions of equity betas (equation 14) and …rm returns (equation 15) we equally weight the
observations. We show that larger …rms tend to operate in dispersed geographic areas, for which headquarter
location is a weaker proxy for where the …rm operates. Local betas should have a larger e¤ect among small …rms
since the betas are more precisely estimated for these …rms.
19
where
cond
f irm;t
represents the conditional equity beta . We expect the regression coe¢ cient
b1 < 0, re‡ecting lower risk of …rms in high beta areas.31 In order to tease out the e¤ects of
the labor and real estate channels, we create subsamples based on …rm real estate exposure.
The idea is that …rms with low exposure to real estate should not be a¤ected by the real estate
channel, so, …rms in high beta areas should have lower industry-adjusted risk due to the labor
channel. As the real estate exposure of the …rms increase, we expect this mechanism to get
weaker, and maybe switch sign.
Table 5 reports the results for the entire sample of …rms, and subsamples based on the real
estate ratio (RER), de…ned as the real estate holdings scaled by the number of employees of the
…rm. This ratio attempts to quantify the relative importance of the real estate versus the labor
channel for the …rm. Panel A report the results for the entire sample, Panel B sort the …rms
into low (below median) and high (above median) RER subsamples. In columns 1 to 4 of Panel
B, subsamples are formed based on …rm-level RER.32 In columns 5 to 8, subsamples are formed
by calculating the average RER for each industry and sorting industries based on this ratio
(below or above median).33 We present results with and without controlling for well-known
…rm level predictors of returns, such as size, book-to-market ratio, leverage, pro…tability, and
investment ratio. Standard errors are clustered at the …rm level.
We …nd that in the entire sample, and especially in the low RER subsamples (columns 1, 2,
5, and 6 of Panel B), b1 is negative and signi…cant, implying that …rms in high beta areas have
lower risk, as measured by their conditional equity betas, than their industry peers in lower
beta areas.34 Using the entire sample of …rms, our b1 estimates imply that conditional equity
betas for the …rms located in the highest MSA beta quintile are roughly 0.1 lower compared to
the …rms in the lowest MSA beta quintile. The implied spread is slightly higher in the sample of
…rms with low real estate exposure. As the real estate holdings of the …rms increase (in columns
3, 4, 7, and 8 of Panel B), the di¤erence in the equity betas declines in magnitude, and loses
its statistical signi…cance, implying that the real estate channel starts to dominate and cancels
31
This is true if the labor channel dominates the real estate channel for the entire sample of …rms.
In Tables 5 to 8, all …rms in our Compustat sample with valid real estate ratios are sorted into subsamples
before the Compustat sample is merged with returns from CRSP. This leads to variations in the sizes of the
subsamples after the return data is merged.
33
It is well known that some industries need and hold more real estate than others (Tuzel, 2010). Sorting …rms
based on industry RER helps reduce the measurement errors individual …rms face.
34
Many …rms in the low RER subsample have zero real estate holdings. Since virtually all …rms need some
real estate to operate, these …rms are most likely leasing substantial amounts of real estate, hence essentially
have negative real estate exposure.
32
20
out the e¤ect of the labor channel for this subsample. The results are both qualitatively and
quantitatively similar whether we sort on the basis of …rm level RER or industry level RER, so
they are robust to measuring real estate exposure in di¤erent ways.
Next, we investigate the relationship between local betas and …rm returns. Since low risk
implies low expected returns, we expect to …nd a negative relationship between local betas
and future stock returns, which serve as a proxy for expected returns. Similar to our tests for
conditional equity betas, we examine the e¤ect of local betas on future …rm returns by running
panel regressions with …xed e¤ects:
rfe irm;t+1 = b0 + b1
+Time
local
m;t
(15)
Industry Dummies + controlsf irm;t +
f irm;t
where rfe irm;t+1 is the excess …rm return from July of year t + 1 to June of year t + 2. We expect
the regression coe¢ cient b1 < 0, re‡ecting lower expected returns of …rms in high beta areas.
This should especially be the case for the …rms with low real estate exposure, since the labor
channel is likely to dominate in those …rms. As the real estate exposure of the …rms increase,
we expect this mechanism to get weaker due to the e¤ects of the real estate channel.
Table 6 follows the same format as Table 5, and reports the results for …rm returns using the
entire sample of …rms and also subsamples based on the real estate ratio (RER). We …nd that
in the entire sample, and especially in the low RER subsamples, b1 is negative and signi…cant,
implying that …rms in high beta areas have lower expected returns than their industry peers
in lower beta areas. Using the entire sample of …rms, our b1 estimates imply roughly 1% lower
expected returns for the …rms located in the highest MSA beta quintile compared to the lowest
beta quintile. The implied spread doubles to approximately 2% and is highly signi…cant in the
sample of …rms with low real estate exposure. Like our …ndings in Table 5, the di¤erence in
the expected returns declines and loses its statistical signi…cance as the real estate holdings of
the …rms increase, and the results are robust to measuring real estate exposure at the …rm or
the industry level.
In Table 7, we consider various re…nements to our baseline sample of low real estate …rms.
In the …rst 4 columns of Panel A, we study …rms from the tradable sectors, which are the
sectors whose output are traded in the entire market, and not limited to the local market. For
21
nontradable industries local betas can impact both the input and the output prices, making
inference about …rm risk di¢ cult.35 Consistent with this prediction, we con…rm that the di¤erence in the expected returns of …rms located in low and high beta areas is slightly bigger for the
sample of tradable …rms. In columns 5 through 8 of Panel A, we consider …rms from tradable
industries with low unionization rates. It is widely accepted that unionization increases the frictions in wage adjustments and reduces …rms’ operating ‡exibility. Therefore, we expect that
the strength of the labor channel, and the resulting lower risk and expected returns of …rms in
high beta areas would be more prevalent for …rms from industries with lower unionization rates.
Also consistent with this prediction, our b1 estimates from non-unionized industries are higher
in absolute value, implying a bigger spread between the expected returns of …rms located in
high and low beta areas.
Panel B of Table 7 considers subsamples of …rms with geographically focused operations,
for which …rm headquarter location is a better proxy for where the …rm operates. We use two
instruments for geographical focus. Our …rst instrument is based on …rm size: Based on the
employment map of …rms created from ReferenceUSA database, we …nd that smaller …rms are
much more likely to focus their operations and employment in the headquarter MSA (Table
A2). The …rst subsample consists of …rms with market values below the median market value of
all …rms in that year. As our second instrument, following Garcia and Norli (2012), we classify
…rms as geographically focused if few state names are mentioned in the …rms’annual reports.36
Garcia and Norli (2012) report that the median …rm mentions …ve states in its 10-K, and the
average state count for the …rms in the highest geographical focus quintile is two. We classify
…rms that mention two or fewer states as geographically focused …rms.37 Our prediction is that
…rm headquarter location will be a less noisy measure of …rm location in these subsamples;
hence, we expect a stronger relationship between local beta and expected returns. We con…rm
that the b1 estimates go up signi…cantly (in absolute value), more than doubling the size of
the original coe¢ cients in some speci…cations, though their statistical signi…cance is somewhat
35
Following Mian and Su… (2012), we de…ne nontradable industries as the retail sector and restaurants. The
retail sector poses additional challenges for our empirical analysis. Most public retail …rms operate in disperse
geographic areas. For example, Garcia and Norli (2012) report that retailers such as Sears, Darden Restaurants,
Barnes & Noble, O¢ ce Max, and many others operate in more than 45 states each. For such …rms, headquarter
location would be a poor proxy for where the …rm operates. By excluding …rms from the retail sector, we improve
the match between a …rm’s headquarter location and the market where it mainly operates.
36
This is a noisy measure of geographic focus as state names can be mentioned for reasons besides being
physical locations of the …rms. The main advantage of this measure is that it is available for a big cross section
of …rms.
37
This subsample has many fewer observations than (about 1/10th) our baseline sample.
22
lower due to smaller sample sizes.
In the last part of our …rm level analysis, we form portfolios of …rms based on the beta of their
local markets, and investigate their future returns. We report the average industry-adjusted
returns of the portfolios in Table 8. We construct equal-weighted, rather than value-weighted
portfolios. As we discuss in Section 2, larger …rms tend to operate in dispersed geographic
areas, for which headquarter location is a weaker proxy for where the …rm operates.38 Since
local betas are more precisely measured for smaller …rms, we expect them to have a stronger
relationship between their expected returns and local betas. In order to form the portfolios,
we simultaneously sort the …rms based on their local betas and real estate ratios. Sorting on
local betas is performed within each industry to account for within-industry variation in local
betas. Industry-adjusted returns are computed by subtracting mean returns of each industry
from individual …rm returns at each month. In addition to industry-adjusted returns, we
present their alphas estimated as intercepts from the Fama-French 3-factor model. Results are
presented for portfolios constructed from the entire sample, tradable sectors, non-unionized
tradable sectors, and small …rms, that we …nd to be more geographically focused. Panel A
measures real estate exposure with …rm-level real estate ratio, Panel B uses industry-level real
estate ratio.
We …nd, within the …rms with low real estate exposure, that industry-adjusted portfolio
returns decline monotonically as the beta of the local market increases, while there is no signi…cant relationship between the returns and local betas of the …rms with high real estate
exposure. For the baseline sample, the spread between the returns of low and high beta portfolios is 2.3%.39 The spreads in portfolio returns get somewhat larger for the subsamples we
consider.40 Though future returns decline almost monotonically as local betas rise, most of
the spread in returns comes from the low industry-adjusted expected returns of the highest
local beta portfolio. We see similar patterns in the 3-factor alphas. These results are both
qualitatively and quantitatively consistent with the panel regression results presented in Tables
38
Almost 80% of …rms in the smallest size quintile have more than half of their employment in their headquarter
MSA, and about 55% of these …rms have virtually all their employment in the same MSA. For the …rms in the
biggest size quintile, those statistics are roughly 50% and 10%. (See Appendix Table A2)
39
We present equal weighted portfolio returns. As expected, value weighted return spreads are smaller and
not statistically signi…cant.
40
Due to much smaller sample size, there is considerably more noise, and larger standard errors, in the portfolios
formed from the geographically focused …rms, measured as …rms that mention 5 or fewer states in their annual
reports. We do not consider the subsample of …rms that mention one or two states as this restriction makes the
sample too small to construct meaningful within-industry portfolios.
23
6 and 7.
Overall, our results provide supporting evidence for our hypothesis that aggregate shocks
have di¤erential e¤ects on the …rms based on the local beta of the area and the real estate
exposure of the …rms. Among the …rms with low real estate exposure, …rms located in high
local beta areas are less a¤ected by the systematic shocks, and have lower expected returns
relative to their industry peers located in low local beta areas, due to hedging e¤ects in labor
costs. For …rms with high real estate exposure, changes in labor costs and real estate prices
cancel each other; hence, we do not …nd di¤erential location e¤ect on risk and returns.
We conduct a number of additional robustness exercises and present the results in the
Internet Appendix. We test the robustness of our results to two alternative measures of local
beta. The …rst measure aggregates industry betas, which are computed as the beta of industry
total factor productivity (TFP) on the aggregate TFP. The second measure is estimated directly
using MSA GDPs. Next, we test the asset pricing implications using an expanded sample period,
starting in 1970, to ensure that our results hold beyond the main sample period of 1986-2011.
Third, we change the asset pricing test assets from individual …rm returns to industry-MSA
portfolio returns to ensure that our results are not driven by few outlier …rms. Finally, we check
the robustness of the regression results to various assumptions for the correlation structure of
the residuals. We con…rm that the results are robust to these variations.
4
Full Model
In this section, we generalize the simple model presented in Section 1 where labor is the only
factor of production, and consider asset pricing in a production economy with three types of
inputs. Two of the inputs (land, representing real estate, and labor) are local inputs in limited
supply. We build heterogeneity into the industry composition (hence, risk) of the local markets.
Our main questions are: (i) How does local risk a¤ect local factor prices? (ii) How does local
risk a¤ect the risk of the …rms operating in those markets? For …rms that have high exposure
to land, which channel (labor or real estate) dominates in …rm returns?
24
4.1
Firms
Similar to the simpli…ed model, there exists a continuum of measure one of in…nitely lived
…rms in each local market m that are subject to aggregate and …rm level productivity shocks,
and produce a homogeneous good. However, now …rms own and use equipment and land, in
addition to labor. Moreover, …rms belong to one of the four industries categorized by both risk
(i.e., high or low risk) and real estate intensity (i.e., high or low real estate intensity).
The production function for …rm i is given by:
Yijm;t = F (At ; Zit ; Lijm;t ; Sijm;t ; Kijm;t )
j
I
(16)
j
l
s
k
= At j Zit Lijm;t
Sijm;t
Kijm;t
:
Lijm;t denotes the labor used in production by …rm i of industry j in local market m during period t. Sijm;t denotes the beginning of period t land holdings (real estate) of …rm
i. Kijm;t denotes the beginning of period t equipment holdings of …rm i. Labor, land, and equipment shares in the …rm’s production function are given by
2 (0; 1): All …rms have the same labor share
l:
l,
j
s;
and
j
k
where
high
,
s
low )
k
j
s+
j
k
=
To capture the heterogeneity in real estate
intensity of industries, we assume that …rms belong to either a low real estate (
or a high real estate (
l+
low ,
s
high
),
k
industry. We assume that half of the …rms belong to high
real estate intensity industries, and the other half belong to the low real estate industries.41
The processes for aggregate and …rm productivity, At and Zit ; and industry risk scaler, Ij , are
described in Section 1.
Local labor markets are competitive and labor is free to move between …rms in the same
area. Therefore, the marginal product of labor equals the wage rate:42
FLijm;t
= FL (At ; Zit ; Lijm;t ; Sijm;t ; Kijm;t )
(17)
= Wm;t
where Wm;t is the wage that clears the local labor market m at time t.
41
In total we have four industries, j 2 f1; 2; 3; 4g; categorized by the two orthogonal characteristics: Industry
high
risk scaler, Ij 2 fIhigh ; Ilow g, and real estate intensity, ( js , jk ) 2 f( low
), ( high
; low
s ; k
s
k )g:
42
This is an assumption we make for convenience. Frictions could be introduced through labor adjustment
costs, a la Belo, Lin, and Bazdresch (2014), wage rigidity, as in Favilukus and Lin (2012b), or a labor market
search mechanism similar to Kuehn, Petrosky-Nadeau, and Zhang (2013).
25
The capital accumulation rule for equipment is:
Kijm;t+1 = (1
)Kijm;t + Iijm;t
where Iijm;t denotes investment in equipment and
(18)
denotes the depreciation rate of installed
equipment.
Purchases and sales of land, and equipment investment are subject to quadratic adjustment
s
k
costs given by gijm;t
and gijm;t
:
with
k; s
g s (Sijm;t+1 ; Sijm;t ) =
1
2
g k (Iijm;t ; Kijm;t ) =
1
2
s
k
(Sijm;t+1 Sijm;t )2
Sijm;t
Iijm;t
Kijm;t
(19)
2
Kijm;t
(20)
> 0: In this speci…cation, investors incur no adjustment cost when net investment
is zero, i.e., when the …rm replaces its depleted equipment stock and maintains its equipment
level, and when the …rm does not change its land holdings.
Firms are equity …nanced. Dividends to shareholders are equal to:
Dijm;t = Yijm;t
Wm;t Lijm;t
Pm;t (Sijm;t+1
Sijm;t )
Iijm;t
s
gijm;t
k
gijm;t
(21)
where Pm;t is the land price that clears the local land market m at time t. At each date t,
…rms choose fSijm;t+1 ; Iijm;t ; Lijm;t g to maximize the equity value of the …rm, which can then
be computed as the solution to the dynamic program:
Vijm;t =
max
fIijm;t ;Sijm;t+1 ;Lijm;t g
fDijm;t + Et (Mt;t+1 Vijm;t+1 )g ;
(22)
subject to equations 2-19, where Mt;t+1 is the stochastic discount factor between time t and t+1.
The …rst order conditions for the …rm’s optimization problem leads to two pricing equations,
which are derived in the Internet Appendix. The returns to the …rm are de…ned as:43
43
We do not assume constant returns to scale in the production function; i.e., l + s + k 2 (0; 1). In the
presence of constant returns to scale, …rm return would be equivalent to the weighted average of returns to land
S
K
and equipment investment, Ri;t+1
and Ri;t+1
; where weights are the shares of land and equipment in the …rm’s
total capital stock. With slightly decreasing returns to scale, …rm returns slightly diverge from the weighted
average of investment returns.
26
F
Rijm;t+1
=
4.2
Vijm;t+1
:
Vijm;t Dijm;t
(23)
Local Markets
Firms have access to the local labor and land markets. All land is owned and utilized by the
local …rms, and all labor is employed by these …rms. Each local market m is populated by a
measure one of …rms, a measure one of workers, and is endowed with a measure one of land.
We assume that labor is not mobile across local labor markets.
There is heterogeneity in the industry composition (high or low risk) of local markets. The
fraction of …rms from the high risk industries in each market is denoted by sm 2 (0; 1): Beside
their industry composition, local markets are ex-ante identical. In equilibrium, local wages and
land prices clear the local labor and land markets.
4.3
The Stochastic Discount Factor
Since the purpose of our model is to examine the cross sectional variation across …rms, we use
a framework where time series properties of returns are matched by using an exogenous pricing
kernel. Following Berk, Green, and Naik (1999) and Zhang (2005), we directly parameterize
the pricing kernel without explicitly modeling the consumer’s problem. As in Jones and Tuzel
(2013) and Imrohoroglu and Tuzel (2014), the pricing kernel is given by:
a
t t+1
log Mt+1 = log
log
where ,
0
> 0, and
driven by the
t
1
t
=
0
+
1
2
2 2
t a
(24)
1 at
< 0 are constant parameters. The volatility of Mt+1 is time-varying,
process. This volatility takes higher values following business cycle contractions
and lower values following expansions, implying a countercyclical price of risk as the result.
This pricing kernel shares a number of similarities with Zhang (2005). Mt+1 , the stochastic
discount factor from time t to t + 1, is driven by
a ,
t+1
the shock to the aggregate productivity
process in period t + 1. The volatility of Mt+1 is time-varying, driven by the
t
process. This
volatility takes higher values following business cycle contractions and lower values following
27
expansions, implying a countercyclical price of risk as the result.44 In the absence of countercyclical price of risk, the risk premia generated in the economy does not change with economic
conditions. Empirically, existence of time varying risk premia is well documented (Fama and
Schwert (1977); Fama and Bliss (1987); Fama and French (1989); Campbell and Shiller (1991);
Cochrane and Piazzesi (2005); Jones and Tuzel (2013); among many others).
4.4
Equilibrium and Calibration
Solving our model generates the pricing functions for local land prices Pm;t and local wages
Wm;t as well as …rms’investment and hiring decisions as functions of the state variables, …rms’
industry and local industry shares, sm . Since the stochastic discount factor is speci…ed exogenously, the solution does not require economy-wide aggregation. However, local land prices and
wages are determined endogenously; so the solution requires aggregation at the local market
level, m. The aggregate local state is (
m ; A),
where
m
is the current distribution of local
…rms over holdings of capital (equipment and land), and …rm level productivity. For the individual …rm, the relevant state variables are its capital holdings (Kijm;t ; Sijm;t ), its …rm level
productivity Zit , and the aggregate local state (
m;t ; At ).
The role of the aggregate local state
is to allow the …rms to predict future land prices and wages.
(Kijm;t ; Sijm;t ; Zit ;
m;t ; At )
ijm;t
faced by …rm i of industry j in local market m at time t.
A recursive competitive equilibrium is a law of motion Hm , where
market and industry speci…c individual policy functions
0
and Sijm
= 'jm (
(i) (
jm ; 'jm )
ijm );
represents the state
and pricing functions Pm (
m ; A)
jm
0
m
= Hm (
m ; A);
0
and 'jm , where Kijm
=
and Wm (
m ; A);
local
jm ( ijm )
such that:
solve the …rms’investment problem,
(ii) Pm clears the local land market
XZ
0
Sijm
di = 1;
j
44
A countercyclical price of risk is endogenously derived in Campbell and Cochrane (1999) from time varying
risk aversion; in Barberis, Huang, and Santos (2001) from loss aversion; in Constantinides and Du¢ e (1996)
from time varying cross sectional distribution of labor income; in Guvenen (2009) from limited participation; in
Bansal and Yaron (2004) from time varying economic uncertainty; and in Piazzesi, Schneider, and Tüzel (2007)
from time varying consumption composition risk.
28
and Wm clears the local labor market
XZ
Lijm di = 1;
j
where the aggregate labor and land supply in each market m is normalized to be 1, and j 2
f1; 2; 3; 4g:
(iii) Hm is generated by
jm
and 'jm :
We solve for the equilibrium prices and allocations recursively using the approximate aggregation idea of Krusell and Smith (1998). Internet Appendix provides details of the model
solution.
We calibrate the model at annual frequency. Table 9 presents the parameters used in the
calibration. We adapt the parameters of the …rm level productivity process from the production
function estimations in Imrohoroglu and Tuzel (2014). The persistence of the …rm productivity
process,
z;
is 0.7, and the conditional volatility of …rm productivity,
z,
is 0.27. The parameters
of the production function are typical values used in the literature. The share of labor
following Cooley and Prescott (1995). The shares of equipment and land,
j
k
and
j
s;
l
is 0.6
are set
to 0.21 and 0.09, respectively, in the low real estate intensity industries, and 0.11 and 0.19 in
the high real estate industries. We calibrate these parameters to match the interquartile range
for the structures/(structures+equipment) ratio of NIPA industries.45 We model technology as
slightly decreasing returns to scale, with
l
+
j
s
+
j
k
=
= 0:9:
We take the parameters of the aggregate productivity from King and Rebelo (1999) and
annualize them. Their point estimates for
a
and
a
are 0.979 and 0.0072, respectively, using
quarterly data, implying annual parameters of 0.922 and 0.014. The depreciation rate for …xed
capital,
is set to 8% annually, which is roughly the midpoint of values used in other studies.
Cooley and Prescott (1995) use 1.6%; Boldrin, Christiano, and Fisher (2001) use 2.1%; and
Kydland and Prescott (1982) use a 2.5% quarterly depreciation rate.
We set the industry risk parameters Ilow to exp( 0:9) and Ihigh to exp(0:9).46 We solve and
45
Data is taken from NIPA Private Fixed Assets by Industry tables. We construct the ratio of current cost
structures/(structures+equipment) for industries. The ratios are quite stable over time. The time series average
for the 25th percentile is 0.42, and 75th percentile is 0.76 over the 1955-2013 period.
46
Even though industry risk scalers vary signi…cantly between low and high beta industries, their output betas
(computed as the slope coe¢ cient from the regression of industry output growth on aggregate output growth)
vary much less. The main reason for this is procyclical ‡uctuations in capital stock due to time varying price of
29
simulate the model for two local markets representing the low and high beta markets. sm ; the
fraction of …rms from the high risk industry, is set to 0.1 and 0.9, respectively, for the low and
high beta areas. We compute local betas
local
m
as the average of industry output growth betas,
weighted by the employment share of industries, similar to its data counterpart described in
equation 11. Our parameters lead to
local
m
of 0.9 for the low beta area, and 1.1 for the high
beta area. We can interpret the high (low) beta area in the model as an area with local beta
roughly one standard deviation above (below) the average local beta in the data. In each area,
half of the …rms belong to high real estate intensity industries, and the other half belong to the
low real estate industries, where real estate intensity and industry risk scalers are orthogonal
industry attributes.47
We choose the pricing kernel parameters ,
0;
and
1
to match the average riskless rate and
the …rst two moments of aggregate value-weighted excess stock returns reported in Imrohoroglu
and Tuzel (2014). The discount factor
roughly 1%.
0
and
1
are 3:2 and
is 0:99, which implies an annual risk free rate of
12, respectively, and generate annual excess mean returns
and standard deviation of 6:2% and 17%, respectively. The adjustment cost parameters,
k,
s
and
are both set to 1 to replicate the value-weighted average (annual) volatility of investment
to capital ratio of 16% reported in Imrohoroglu and Tuzel (2014).48
In order to compute the model statistics, we perform 100 simulations of the model economy
with 2,000 …rms over 50 periods (years).
4.5
Quantitative Results
In this economy, …rms optimally make their investment and hiring decisions to maximize …rm
value. The optimality conditions dictate that …rms invest (hire) until the marginal cost of
investing (hiring) equals the marginal bene…t. The marginal bene…t of investing and hiring
increases in productivity. Therefore, everything else equal, the demand for labor and land
risk, which dominate the procyclical ‡uctuations in industry and aggregate output. Therefore, modest di¤erences
in industry output betas require relatively stark di¤erences in industry scalers.
47
In a robustness analysis, we added another area with a more diversi…ed economy to our analysis and solved the
model for three areas. We con…rmed that our quantitative results are almost the same when we use observations
from all three areas, or from any two of the areas. Therefore, adding more areas with di¤erent industry shares
is unlikely to have a material e¤ect on our results.
48
The investment to capital ratio in data is not separately calculated for equipment and land as investment
data is not available in disaggregated form. s = k = 1 leads to 15% volatility in I/K for equipment, and 16%
volatility for land.
30
will be increasing in aggregate productivity. Since both land and labor are in limited supply,
in equilibrium, the market clearing condition can only be satis…ed if the marginal cost of
hiring (wage rate) and investing in land (a¢ ne in land prices) are also increasing in aggregate
productivity. So, a good (bad) aggregate productivity shock leads to increases (decreases) in
wage rate and land prices. This e¤ect is more pronounced in areas where a larger fraction of
…rms belong to the high risk industry (Ij = Ihigh ) since the aggregate productivity shocks have
bigger e¤ects on the marginal bene…t of investing and hiring of …rms that belong to the high
risk industry.
Table 10 demonstrates this result by running regressions similar to the ones we run in section
3.1, using simulated data from the model economy. We run regressions of the form:
log(Wm;t ) = b0 + b1 outputt
local
m
+ b2
local
m
+ Time dummy +
m;t
(25)
log(Pm;t ) = b0 + b1 outputt
local
m
+ b2
local
m
+ Time dummy +
m;t
(26)
where Wm;t and Pm;t are wage rate and land prices in each area, and outputt is log aggregate
output of all the …rms in the economy.
The …rst column of Table 10 reports the wage regression, and the second column reports the
land price regression results. Both regressions produce positive and highly signi…cant estimates
for b1 ; implying that wage growth and land price growth in high beta areas covary more with
aggregate output shocks, relative to their counterparts in lower beta areas.
We next investigate the e¤ect of local risk (local beta) on the risk and expected returns of
the …rms operating in those areas. The greater sensitivity of wages to aggregate shocks in high
beta areas provides a natural hedge for the …rms operating there, mitigating the e¤ect of the
shocks on the …rms. This mechanism leads to lower sensitivity of returns to aggregate shocks for
…rms in high beta areas, lower overall risk, and lower expected returns for the …rm. While the
full model can not be solved analytically, the results of the simple model presented in Section 1
allows us to derive closed-form expressions for local wage rate and …rm pro…ts, and are useful
in clarifying this channel and gaining intuition into this result. Equation 7 shows that, while
local wage rate is cyclical in all areas, the sensitivity of wage rate to aggregate productivity
At increases with sm ; the fraction of …rms from the high beta industry. This, in turn, reduces
the sensitivity of …rm pro…ts to aggregate shocks in high beta areas, as shown in Equations
31
9 and 10. Since risk is de…ned as (the inverse of the) covariation of returns with the pricing
kernel (equation 24), and the pricing kernel is monotonically decreasing in shocks to aggregate
productivity, lower sensitivity of pro…ts to productivity shocks implies lower risk.
While the labor mechanism lowers the risk of the …rms in high beta areas relative to their
industry peers located elsewhere, the land mechanism leads to the opposite result. Given that
land values are more sensitive to aggregate shocks in high beta areas, and that the …rm value is
partly derived from the value of its hand holdings, greater variation in land prices would imply
higher sensitivity of …rm returns to aggregate shocks in high beta areas.
The model does not have a rental market. All capital (equipment and land) is owned by the
…rms. All …rms optimally own some land since the marginal product of land goes to in…nity
as land ownership approaches zero, however, …rms that belong to high real estate intensity
industries endogenously maintain a higher ratio of land to equipment, compared to others.
Therefore, there is cross sectional heterogeneity in the …rms’exposures to the labor versus real
estate channels.
In order to test these predictions in the model, we replicate the …rm-level analysis presented
in Section 3.2 using simulated data from the model economy. The results are presented in Table
11. In Panels A and B we run regressions of the form:
= b0 + b1
local
m
+ Industry
Time dummy +
f irm;t
(27)
rfe irm;t+1 = b0 + b1
local
m
+ Industry
Time dummy +
f irm;t
(28)
cond
f irm;t
separately for …rms that belong to industries with low and high real estate intensity. rfe irm;t
is the excess …rm return, where raw …rm returns are de…ned by equation 23.
cond
f irm;t
is the
conditional beta of the …rm, estimated by running regressions of the excess …rm returns on
the excess market return. Both regressions include industry-time …xed e¤ects; therefore, all
comparisons are within the …rms in the same industry and same year.
Panels A and B of Table 11 present the results of the …rm conditional beta and return
regressions described in equations 27 and 28. The …rst column runs the regressions using
the entire sample of …rms, whereas columns 2 and 3 use low and high real estate intensity
subsamples, respectively. Labor channel implies lower risk, and lower expected returns for
…rms in high beta areas, which would lead to negative b1 estimates. Land price channel implies
32
positive b1 estimates. It is not clear ex-ante which channel would dominate the regression
coe¢ cient. However, b1 is expected to be lower (more negative, or less positive) in the low real
estate subsample compared to the estimate from the high real estate subsample. We …nd that b1
is negative in all speci…cations and samples, implying that, overall, the labor channel dominates
the land price channel. However, the absolute value of the coe¢ cients are much larger for the
…rms from low real estate industries: -0.10 versus -0.06 for conditional beta regressions (both
statistically signi…cant); -0.83 versus -0.05 for the …rm return regressions (only the former is
statistically signi…cant). These results are consistent with our empirical results described in
section 3.2, where we …nd that the labor channel overall dominates the real estate price channel
in the data.
5
Conclusion
We show that the industrial composition of local markets and, in particular, the cyclicality of
the major industries in the areas matter for how the systematic shocks a¤ect the …rms located
in those areas. We calculate the “local beta”for the metropolitan areas in the U.S., which is the
average of the industry GDP betas, weighted by the industry shares in each area. We …nd that
aggregate GDP shocks have more pronounced e¤ects on local factor prices, such as wages and
real estate prices, in high beta areas compared to the lower beta areas. These local factors, on
average, account for more than 75% of the economic output, so ‡uctuations in their prices are
relevant for the …rms in the area. Larger e¤ect of shocks on wages we found in high beta areas
provides a natural hedge for the …rms operating there, compared to their industry peers in low
beta areas, mitigating the e¤ect of these shocks on …rm’s returns. In addition to wages, high
beta areas experience bigger ‡uctuations in local real estate prices due to aggregate shocks, as
the demand for these assets changes more drastically. Since …rms have di¤erent exposures to
real estate, the implication of this real estate channel for …rms varies. Speci…cally, for …rms
with high real estate holdings (long position in real estate), being located in high beta areas
also means that part of the …rms’value is more exposed to aggregate shocks due to the more
procyclical real estate prices. Therefore, the real estate channel would o¤set the e¤ect of the
labor channel, making the e¤ect of local beta on …rm risk less signi…cant for …rms with high
real estate holdings.
33
We develop a theoretical model where …rms belong to either a high risk (more cyclical) or a
low risk (less cyclical) industry, and local areas vary in their composition of industry makeup.
Each area features a continuum of …rms that use labor, land (real estate), and equipment in their
production. Land and labor markets clear within each market. The model generates patterns
similar to our main empirical results. Speci…cally, we con…rm that land and labor prices are
more procyclical in high beta areas. Larger ‡uctuations in wages reduce the sensitivity of …rm
returns to systematic shocks in high beta areas, leading to lower risk for these …rms. These
results are stronger for …rms with low real estate exposure.
References
[1] Almazan, Andres., Motta, Adolfo D., and Sheridan Titman. 2007. Firm Location and the
Creation and Utilization of Human Capital. Review of Economic Studies 74, 1305-1327.
[2] Bansal, Ravi W. and Amir Yaron. 2004. Risks For the Long Run: A Potential Resolution
of Asset Pricing Puzzles. Journal of Finance 59:1481-509.
[3] Barberis, Nicholas, Huang, Ming and Tano Santos. 2001. Prospect Theory and Asset Prices.
Quarterly Journal of Economics 116:1-53.
[4] Belo, Frederico, and Xiaoji Lin, and Santiago Bazdresch. 2014. Labor Hiring, Investment and Stock Return Predictability in the Cross Section. Journal of Political Economy
122:129-177.
[5] Belo, Frederico and Xiaoji Lin. 2012. Labor Heterogeneity and Asset Prices: The Importance of Skilled Labor. Working Paper.
[6] Berk, Jonathan B., Richard C. Green and Vasant Naik. 1999. Optimal Investment, Growth
Options, and Security Returns. Journal of Finance 54:1553-607.
[7] Berk, Jonathan, and Johan Walden. 2013. Limited Capital Market Participation and Human Capital Risk. Review of Asset Pricing Studies 3:1-37.
[8] Boldrin, Michele, Lawrence Christiano and Jonah Fisher. 2001. Habit Persistence, Asset
Returns, and the Business Cycle. American Economic Review 91:149-66.
34
[9] Campbell, John Y. 1996. Understanding risk and return. Journal of Political Economy
104:298–345.
[10] Campbell, John Y. and John H. Cochrane. 1999. By Force of Habit: A ConsumptionBased Explanation of Aggregate Stock Market Behavior. Journal of Political Economy
107:205–51.
[11] Campbell, John, and Robert Shiller. 1991. Yield Spreads and Interest Rate Movements: A
Bird’s Eye View. Review of Economic Studies 58:495-514.
[12] Chaney, Thomas, Sraer, David, and David Thesmar. 2012. The Collateral Channel: How
Real Estate Shocks A¤ect Corporate Investment. American Economic Review 102:6, 23812409.
[13] Chen, Jason, Kacperczyk, Marcin and Hernan Ortiz-Molina. 2011. Labor Unions, Operating Flexibility, and the Cost of Equity, Journal of Financial and Quantitative Analysis
46:1, 25-58.
[14] Chen, Long, and Lu Zhang. 2011. The Stock Market and Aggregate Employment. Journal
of Financial Economics 99: 385–399.
[15] Chen, Yong and Stuart S. Rosenthal. 2008. Local Amenities and Life-Cycle Migration: Do
People Move for Jobs or Fun? Journal of Urban Economics 64:519-537.
[16] Cochrane, John H., and Monika Piazzesi. 2005. Bond Risk Premia. American Economic
Review 95:138-60.
[17] Constantinides, George and Darrell Du¢ e. 1996. Asset Pricing with Heterogeneous Consumers. Journal of Political Economy 104:219–240.
[18] Cooley, Thomas F., and Edward C. Prescott. 1995. Economic Growth and Business Cycles.
In T. F. Cooley (ed.), Frontiers of Business Cycle Research, Princeton University Press.
[19] Danthine, Jean-Pierre, and John B. Donaldson. 2002. Labor Relations and Asset Returns.
Review of Economic Studies 69, 41–64.
[20] Donangelo, Andres. 2014. Labor Mobility: Implications for Asset Pricing. Journal of Finance 68, 1321-46.
35
[21] Dougal, Casey, Parsons, Christopher A., and Sheridan Titman. 2015. Urban Vibrancy and
Corporate Growth. Journal of Finance 70, 163–210.
[22] Eisfeldt, Andrea L. and Dimitris Papanikolaou. 2013. Organization Capital and the CrossSection of Expected Returns. Journal of Finance 68, 1365–1406.
[23] Ellison, Glenn, Edward L. Glaeser, and William R. Kerr. 2010. What Causes Industry
Agglomeration? Evidence from Coagglomeration Patterns. American Economic Review
100, 1195–1213.
[24] Engelberg, Joseph, Arzu Ozoguz, and Sean Wang. 2010. Know Thy Neighbor: Industry
Clusters, Information Spillovers, and Market E¢ ciency. Working Paper.
[25] Fama, Eugene F., and Robert R. Bliss. 1987. The Information in Long-Maturity Forward
Rates. American Economic Review 77:680-92.
[26] Fama, Eugene F., and Kenneth R. French. 1989. Business Conditions and Expected Returns on Stocks and Bonds. Journal of Financial Economics 25:23-49.
[27] Fama, Eugene F. and Kenneth R. French. 1992. The Cross Section of Expected Stock
Returns. The Journal of Finance 47:427-65.
[28] Fama, Eugene F. and Kenneth R. French. 1993. Common Risk Factors in the Returns on
Stocks and Bonds. Journal of Financial Economics 33, 3-56.
[29] Fama, Eugene F. and James MacBeth. 1973. Risk, Return and Equilibrium: Empirical
Tests. Journal of Political Economy 81, 607–636.
[30] Fama, Eugene F., and William Schwert. 1977. Asset Returns and In‡ation. Journal of
Financial Economics 5:115-46.
[31] Fan, Joseph J. H., Sheridan Titman and Garry Twite. 2012. An International Comparison
of Capital Structure and Debt Maturity Choices. Journal of Financial and Quantitative
Analysis 47:23-56.
[32] Favilukis, Jack and Xiaoji Lin. 2012a. Does Wage Rigidity Make Firms Riskier? Evidence
from Long-Horizon Return Predictability. Working Paper.
36
[33] Favilukis, Jack and Xiaoji Lin. 2012b. Wage Rigidity: A Solution to Several Asset Pricing
Puzzles. Working Paper.
[34] Garcia, Diego and Oyvind Norli. 2012. Geographic Dispersion and Stock Returns. Journal
of Financial Economics 106:3, 547-565.
[35] Gourio, Francois. 2007. Labor leverage, …rms heterogeneous sensitivities to the business
cycle, and the cross-section of returns. Working Paper.
[36] Guvenen, Fatih. 2009. A Parsimonious Macroeconomic Model for Asset Pricing. Econometrica 77:1711-50.
[37] Hansen, Lars P., Heaton, John C., and Nan Li. 2005. Intangible Risk? In C. Corrado, J.
Haltiwanger, and D. Sichel (ed.), Measuring Capital in the New Economy, University of
Chicago Press, Chicago:111-152.
[38] Hirsch, Barry T., and David A. Macpherson. 2003. Union Membership and Coverage Database from the Current Population Survey: Note. Industrial and Labor Relations Review
56:349-54.
[39] Imrohoroglu, Ayse, and Selale Tuzel. 2014. Firm Level Productivity, Risk, and Return.
Management Science 60:2073-2090.
[40] Jones, Christopher S., and Selale Tuzel. 2013. Inventory Investment and the Cost of Capital. Journal of Financial Economics 107, 557-579.
[41] Jones, Christopher S., and Selale Tüzel. 2013. New Orders and Asset Prices. Review of
Financial Studies 26:115-157.
[42] Kennan John and James R. Walker. 2011. The E¤ect of Expected Income on Individual
Migration Decisions. Econometrica 79:211-251.
[43] Kimbell, Larry J.and Daniel J.B. Mitchell. 1982. Labor Market Contracts and In‡ation.
In: Bailey, M.N. (ed.), Workers, Jobs and In‡ation, Brookings, 199-238.
[44] King, R. G. and S. T. Rebelo. 1999. Resuscitating Real Business Cycles. In J. B. Taylor and
M. Woodford (ed.), Handbook of Macroeconomics, edition 1, chapter 14,Elsevier:927-1007.
37
[45] Korniotis, George M., and Alok Kumar. 2013. State-Level Business Cycles and Local Return Predictability. Journal of Finance 68:1037-1096.
[46] Krusell, Per. and Anthony A. Smith, Jr. 1998. Income and Wealth Heterogeneity in the
Macroeconomy. Journal of Political Economy 106:867–96.
[47] Kuehn, Lars-Alexander, Petrosky-Nadeau, Nicolas, and Lu Zhang. 2013. An Equilibrium
Asset Pricing Model with Labor Market Search. Working Paper.
[48] Kydland, F. and E. Prescott. 1982. Time to Build and Aggregate Fluctuations. Econometrica 50:1345-70.
[49] Lewellen, Jonathan, and Stefan Nagel. 2006. The conditional CAPM does not explain
asset-pricing anomalies. Journal of Financial Economics 82:289–314.
[50] Marshall, Alfred. 1920. Principles of Economics. London: Macmillan.
[51] Merz, Monika, and Eran Yashiv. 2007. Labor and the Market Value of the Firm. American
Economic Review 97, 1419–1431.
[52] Mian, Atif, and Amir Su…. 2012. What Explains High Unemployment? The Aggregate
Demand Channel. Working Paper.
[53] Moretti, Enrico. 2011. Local Labor Markets. In: O Ashenfelter and D Card (eds.), Handbook of Labor Economics, 4B Amsterdam: Elsevier-NorthHolland, 1237-1314.
[54] Novy-Marx, Robert. 2013. The Other Side of Value: The Gross Pro…tability Premium.
Journal of Financial Economics 108:1-28.
[55] Petersen, Mitchell. 2009. Estimating Standard Errors in Finance Panel Data Sets: Comparing Approaches. Review of Financial Studies 22, 435-480.
[56] Piazzesi, Monika, Schneider, Martin and Selale Tüzel. 2007. Housing, Consumption, and
Asset Pricing. Journal of Financial Economics 83:531-69.
[57] Pirinsky, Christo A., and Qinghai Wang. 2006. Does Corporate Headquarters Location
Matter for Stock Returns? Journal of Finance 61: 1991-2015.
[58] Roback, Jennifer. 1982. Wages, Rents and the Quality of Life. Journal of Political Economy
90:6, 1257–1278.
38
[59] Rosen, Sherwin. 1979. Wage Based Indexes of Urban Quality of Life. In: Miezkowski, Peter
N., Straszheim, Mahlon R. (eds.), Current Issues in Urban Economics. Johns Hopkins
University Press, Baltimore, MD, 74–104.
[60] Sterk, Vincent. 2015. Home Equity, Mobility, and Macroeconomic Fluctuations. Journal
of Monetary Economics 74:16-32.
[61] Tuzel, Selale. 2010. Corporate Real Estate Holdings and the Cross Section of Stock Returns.
Review of Financial Studies 23:2268-302.
[62] Zhang, L. 2005. Value premium. Journal of Finance 60:67-103.
39
Figure 1. Histogram of Industrial Dispersion of Employment within MSAs. The figure illustrates the distribution of industrial dispersion of employment within each MSA, calculated as the Herfindahl-Hirschman Index (HHI) of industry
employment shares in 2011.
Figure 2. Histogram of MSA Betas. The figure illustrates the distribution of MSA Betas as of 2011.
40
Figure 3. Local Beta and Fluctuations in Local GDP. The figure illustrates the average real GDP and GDP growth
of the top and bottom 15 MSAs based on their local betas over the 2001-2011 period. MSAs are classified based on local betas
in 2011. Top panel plots average real GDP, normalized to 1 in 2001, bottom panel plots the average real GDP growth.
41
Figure 4. Histogram of MSA Dispersion of Firms within Industries. The figure illustrates the distribution of the
dispersion of Compustat firm locations within each industry in 2011. For each industry we count the number of firms located in
each MSA. The dispersion of firm locations is calculated as the HHI of MSA firm-count shares in each industry.
42
Table 1
Highest and Lowest Beta Industries
The table presents the 3-digit NAICS industries with lowest (Panel A) and highest (Panel B) betas as of 2011. The
industry betas, β ind , are calculated as the slope coefficients from the regressions of industry shock (real industry value
added growth) on aggregate shock (real GDP growth) from 1978 to 2010.
NAICS
Industry Title
β ind
Panel A. Lowest β ind Industries
211
311
312
213
622
Oil and Gas Extraction
Food Manufacturing
Beverage and Tobacco Product Manufacturing
Support Activities for Mining
Hospitals
−0.757
−0.712
−0.712
−0.245
−0.067
Panel B. Highest β ind Industries
331
525
321
336
327
Primary Metal Manufacturing
Funds, Trusts, and Other Financial Vehicles
Wood Product Manufacturing
Transportation Equipment Manufacturing
Nonmetallic Mineral Product Manufacturing
43
3.623
3.479
3.259
3.104
2.863
Table 2
Highest and Lowest Beta MSAs
The table presents the summary statistics for the MSAs with lowest (Panel A) and highest (Panel B) betas as of
local
2011 and the transition probability matrix of local beta quintiles. Local betas, βm
are calculated as the average of the industry betas operating in that area, weighted by the employment share of industries. Representative
industry is the industry that has the highest employment share in that MSA among all industries with location quotient above 3.5. Location quotient is the ratio of an industry’s share of regional employment to its share of the
entire economy. % of Employment reports the fraction of jobs from the representative industry in that MSA. #
Employment reports the number of employees in 2011 for the MSA and the employment rank among all MSAs.
CBSA
MSA Title
Local
βm
Representative Ind.
Panel A. Lowest
41140
32900
34900
27060
25260
23580
40340
33260
43580
34060
24140
26980
38220
47580
40660
St. Joseph, MO-KS
Merced, CA
Napa, CA
Ithaca, NY
Hanford-Corcoran, CA
Gainesville, GA
Rochester, MN
Midland, TX
Sioux City, IA-NE-SD
Morgantown, WV
Goldsboro, NC
Iowa City, IA
Pine Bluff, AR
Warner Robins, GA
Rome, GA
0.71
0.75
0.76
0.78
0.79
0.81
0.82
0.83
0.83
0.84
0.86
0.86
0.87
0.87
0.87
local
βm
% of Emp.
# Emp.
Emp. Rank
14.9%
15.8%
11.9%
37.6%
13.8%
14.5%
21.6%
13.1%
13.2%
7.9%
7.2%
5.7%
6.6%
10.6%
5.2%
48762
39914
56022
45545
22896
59760
86211
65689
65462
45865
34069
65159
25139
34278
32780
261
302
230
281
359
223
178
215
216
277
327
217
355
326
333
24.9%
34.2%
22.6%
24.7%
6.9%
6.8%
11.8%
5.1%
10.6%
6.6%
23.3%
7.5%
16.6%
4.4%
4.4%
102109
49793
31770
54824
40314
110969
123517
36093
242354
49204
730747
65748
105276
42132
90369
160
253
337
236
299
149
136
320
76
259
34
214
152
295
176
MSAs
Food Manuf.
Food Manuf.
Bevg. & Tobac. Manuf.
Educ. Service
Food Manuf.
Food Manuf.
Hospitals
Supp. Mining
Food Manuf.
Chemical Manuf.
Food Manuf.
Truck Transport.
Food Manuf.
Food Manuf.
Food Manuf.
local
Panel B. Highest βm
MSAs
21140
37700
29020
19140
18020
43900
25860
11500
48620
34740
29820
29140
35980
31900
26100
Elkhart-Goshen, IN
Pascagoula, MS
Kokomo, IN
Dalton, GA
Columbus, IN
Spartanburg, SC
Hickory-Lenoir-Morganton, NC
Anniston-Oxford, AL
Wichita, KS
Muskegon-Norton Shores, MI
Las Vegas-Paradise, NV
Lafayette, IN
Norwich-New London, CT
Mansfield, OH
Holland-Grand Haven, MI
1.73
1.48
1.33
1.29
1.28
1.26
1.24
1.24
1.24
1.23
1.23
1.23
1.23
1.20
1.20
Transp. Equip. Manuf.
Transp. Equip. Manuf.
Transp. Equip. Manuf.
Textile Product Mills
Transp. Equip. Manuf.
Transp. Equip. Manuf.
Furniture Manuf.
Fabric. Metal Manuf.
Transp. Equip. Manuf.
Prim. Metal Manuf.
Accommodation
Transp. Equip. Manuf.
Accommodation
Fabric. Metal Manuf.
Fabric. Metal Manuf.
44
Table 3
Local Wages and Local Beta
Panel A reports the effect of aggregate shocks on the industry wage growth in an MSA, conditional on the local beta,
local
local
βm
. Panel B reports the effect of aggregate shocks on the occupational wage growth in an MSA, conditional on βm
.
local
Calculation of βm is described in Table 2. Wage growth is annual in Panel A, hourly in Panel B, all in real terms.
Aggregate shock (Shock ) is the aggregate real GDP growth in that year, in %. Regression sample period is 1990-2011
in Panel A (LEHD Data), 1999-2011 in Panel B (OES Data). Non-unionized industries (occupations) are industries
(occupations) with unionization rates lower than the median unionization rate of all industries (occupations) in that
year. Tradable industries are all industries excluding the retail sector and restaurants. All standard errors are clustered
at the MSA level, presented in parantheses. ∗ , ∗∗ , and ∗∗∗ represent significance level of 10%, 5%, and 1%, respectively.
Panel A. Annual Wage for Industries
Wage Growth (%)
All Industries
(1)
local
βm
local
Shock × βm
Ind.×Year FE
MSA FE
Observations
R2
Non-Union Industries
(2)
∗∗∗
−1.70
(0.29)
0.24∗∗
(0.12)
X
409294
0.05
(3)
(4)
∗∗∗
−0.45
(0.76)
−1.82
(0.35)
0.24∗
(0.13)
0.34∗∗
(0.16)
X
X
409294
0.05
−0.23
(0.90)
0.34∗∗
(0.17)
X
X
X
222549
0.06
222549
0.06
Wage Level
Tradable Industries
(5)
∗∗∗
−1.83
(0.31)
0.30∗∗
(0.13)
(1990 $)
(6)
(7)
(8)
−0.40
(0.81)
1276.13
(996.47)
2734.57∗∗∗
(480.78)
X
X
X
442591
0.64
0.30∗∗
(0.14)
X
X
X
343477
0.04
343477
0.04
442591
0.57
Panel B. Hourly Wage for Occupations
Wage Growth (%)
Broad Occupations
(1)
local
βm
local
Shock × βm
Occ.×Year FE
MSA FE
Observations
R2
Detailed Occupations
(2)
∗∗∗
−1.30
(0.33)
0.44∗∗∗
(0.14)
X
76986
0.08
(3)
∗
−1.12
(0.67)
0.38∗∗∗
(0.13)
X
X
76986
0.09
(4)
∗∗∗
−1.22
(0.20)
0.12
(0.46)
0.29∗∗∗
(0.07)
X
1028541
0.04
0.24∗∗∗
(0.06)
X
X
1028541
0.05
45
Wage Level
Detailed Non-Union Occ.
(5)
(6)
∗∗∗
−1.25
(0.21)
0.31∗∗∗
(0.08)
X
758607
0.04
0.53
(0.50)
(1990 $)
(7)
(8)
∗∗
0.77
(0.35)
0.47∗∗∗
(0.11)
0.27∗∗∗
(0.08)
X
X
758607
0.04
X
1349174
0.83
X
X
1349174
0.86
Table 4
Real Estate Returns and Local Beta
The table reports the effect of aggregate shocks on the real estate returns in the MSA, conditional on the local beta,
local
local
is described in Table 2. Housing returns are the annualized changes in the FHFA house
. Calculation of βm
βm
price indexes in each MSA. Commercial real estate returns are the total annualized returns to all property types in
each MSA, from NCREIF. Rent growth is the annualized growth in office building rents in each MSA, from CoStar.
Aggregate shock (Shock ) is the aggregate real GDP growth in that year, in %. Regression sample period is 1986-2011.
Regressions are at the quarterly basis. The commercial real estate regression includes property type fixed effect (Office,
Industrial, Retail, Apartment, Hotel). All standard errors are clustered at the MSA level, presented in parantheses. ∗ ,
∗∗
, and ∗∗∗ represent significance level of 10%, 5%, and 1%, respectively.
Housing Returns
Commercial Real Estate Returns
Rent Growth
(1)
local
βm
local
Shock × βm
Time FE
MSA FE
Observations
R2
(2)
∗∗∗
−1.87
(0.56)
1.16∗∗∗
(0.25)
(3)
(4)
∗∗∗
−2.68
(4.38)
9.25
(6.17)
1.09∗∗∗
(0.26)
4.16∗
(2.14)
3.52∗
(2.08)
3.82
(1.18)
(5)
−4.36
(2.77)
2.72∗∗
(1.13)
(6)
0.18
(7.54)
1.98∗
(1.14)
X
X
X
X
X
X
X
X
X
36268
0.45
36268
0.46
10267
0.52
10267
0.53
5411
0.18
5411
0.21
46
Table 5
Panel Regression of Conditional Equity Betas and Local Beta
The table reports the relationship between the conditional equity betas (βfcond
irm ) of the firms located in an MSA and the
local
local
local beta, βm
. Calculation of βm
is described in Table 2. Conditional equity betas for firms are computed each
calendar year from short window regressions using monthly data, correcting for non-syncronous trading, as in Lewellen
and Nagel(2006). All independent variables are lagged for one year. In Panel A, we regress conditional equity beta on
local beta, and other firm level control variables. Log BM and Log Size are the log of the firm’s book to market ratio
and market equity constructed following Fama and French (1992). Leverage is firm’s market leverage as in Fan, Titman
and Twite (2012). P rof itability is gross profit measure as in Novy-Marx (2013). Investment is the investment ratio
as in Dougal, Parsons, and Titman (2015). In Panel B, the Subsamples are sorted based on RER, defined as (buildings
+ capital leases) / Employees. Columns 3-6 use firm level RER, columns 7-10 use industry level RER, computed as
the average RER of firms in each industry. Regression sample period is 1986-2011. Standard errors are clustered by
firms and are presented in parentheses. ∗ , ∗∗ , and ∗∗∗ represent significance level of 10%, 5%, and 1%, respectively.
Panel A: Controlling for Firm Characteristics
(1)
local
βm
−0.35∗∗∗
(0.12)
(2)
(3)
−0.35∗∗∗
(0.12)
(4)
−0.33∗∗∗
(0.12)
(5)
−0.34∗∗∗
(0.12)
(6)
−0.33∗∗∗
(0.12)
−0.03∗
(0.01)
Log BM
−0.04∗∗∗
(0.01)
−0.05∗∗∗
(0.01)
0.30∗∗∗
(0.05)
Leverage
0.34∗∗∗
(0.05)
−0.37∗∗∗
(0.05)
Profitability
−0.36∗∗∗
(0.05)
0.46∗∗
(0.18)
Investment
X
105334
0.09
X
105334
0.09
−0.30∗∗
(0.12)
−0.10∗∗∗
(0.01)
Log Size
Ind.×Time FE
Observations
R2
(7)
−0.36∗∗∗
(0.12)
X
105334
0.09
X
105334
0.09
X
105334
0.09
X
105334
0.09
0.58∗∗∗
(0.18)
X
105334
0.09
Panel B: Subsample by Real Estate Holdings
Low RER Firms
(1)
local
βm
−0.45∗∗∗
(0.16)
(2)
−0.40∗∗
(0.16)
High RER Firms
(3)
(4)
−0.18
(0.19)
−0.12
(0.18)
Low RER Industries
(5)
−0.48∗∗∗
(0.14)
(6)
−0.44∗∗∗
(0.14)
High RER Industries
(7)
(8)
−0.19
(0.20)
−0.11
(0.19)
Log BM
−0.08∗∗∗
(0.02)
−0.10∗∗∗
(0.02)
−0.09∗∗∗
(0.02)
−0.10∗∗∗
(0.02)
Log Size
−0.03∗∗∗
(0.01)
−0.07∗∗∗
(0.01)
−0.06∗∗∗
(0.01)
−0.05∗∗∗
(0.01)
Leverage
0.41∗∗∗
(0.08)
0.25∗∗∗
(0.08)
0.43∗∗∗
(0.07)
0.23∗∗∗
(0.08)
Profitability
−0.15∗
(0.08)
−0.59∗∗∗
(0.06)
−0.07
(0.07)
−0.64∗∗∗
(0.07)
Investment
0.31
(0.29)
0.84∗∗∗
(0.25)
0.39
(0.24)
0.77∗∗∗
(0.28)
Ind.×Time FE
Observations
R2
X
44851
0.11
X
44851
0.11
X
52760
0.09
X
52760
0.10
47
X
61162
0.10
X
61162
0.10
X
44167
0.08
X
44167
0.08
Table 6
Panel Regression of Equity Returns and Local Beta
The table reports the relationship between the future returns of the firms located in an MSA and the local beta,
local
local
βm
. Calculation of βm
is described in Table 2. In Panel A, we regress future monthly return on local beta, and
other firm level control variables. Log BM and Log Size are the log of the firm’s book to market ratio and market
equity constructed following Fama and French (1992). Leverage is firm’s market leverage as in Fan, Titman and
Twite (2012). P rof itability is gross profit measure as in Novy-Marx (2013). Investment is the investment ratio as
in CDougal, Parsons, and Titman (2015). Future returns are measured in the year following the portfolio formation,
from July of year t+1 to June of year t+2, and annualized (%). In Panel B, the Subsamples are sorted based on RER,
defined as (buildings + capital leases) / Employees. Columns 3-6 use firm level RER, columns 7-10 use industry level
RER, computed as the average RER of firms in each industry. Regression sample period is 1986-2011. Standard errors
are clustered by firms and are presented in parentheses. ∗ , ∗∗ , and ∗∗∗ represent significance level of 10%, 5%, and 1%,
respectively.
Panel A: Controlling for Firm Characteristics
(1)
local
βm
(2)
(3)
(4)
(5)
(6)
−5.24∗∗
−4.09∗
−4.77∗∗
−5.64∗∗
−4.89∗∗
(2.23)
(2.33)
(2.31)
(2.25)
(2.22)
(2.25)
6.52∗∗∗
(0.28)
Log BM
−1.22∗∗∗
(0.12)
6.68∗∗∗
(0.99)
Leverage
−1.85∗
(1.09)
7.84∗∗∗
(0.91)
Profitability
9.76∗∗∗
(0.97)
−19.63∗∗∗
(4.12)
Investment
X
1138028
0.15
X
1138028
0.15
−5.19∗∗
(2.37)
5.92∗∗∗
(0.33)
−1.97∗∗∗
(0.11)
Log Size
Ind.×Time FE
Observations
R2
(7)
−5.13∗∗
X
1138028
0.15
X
1138028
0.15
X
1138028
0.15
X
1138028
0.15
−9.73∗∗
(4.17)
X
1138028
0.15
Panel B: Subsample by Real Estate Holdings
Low RER Firms
(1)
local
βm
(2)
−10.29∗∗∗
−10.38∗∗∗
(3.45)
(3.55)
High RER Firms
(3)
−1.13
(3.39)
(4)
−1.32
(3.68)
Low RER Industries
(5)
(6)
−7.90∗∗
−8.06∗∗
(3.08)
(3.15)
High RER Industries
(7)
(8)
−1.70
(3.36)
−1.58
(3.65)
Log BM
6.88∗∗∗
(0.52)
4.99∗∗∗
(0.45)
6.84∗∗∗
(0.45)
4.92∗∗∗
(0.48)
Log Size
−1.34∗∗∗
(0.19)
−1.35∗∗∗
(0.18)
−1.30∗∗∗
(0.15)
−1.02∗∗∗
(0.19)
Leverage
−3.93∗∗
(1.72)
−1.75
(1.63)
−2.30
(1.41)
−0.40
(1.75)
Profitability
10.21∗∗∗
(1.52)
Investment
−4.81
(6.82)
Ind.×Time FE
Observations
R2
X
484464
0.16
X
484464
0.16
9.43∗∗∗
(1.32)
15.27∗∗∗
(1.41)
−18.52∗∗∗
(5.59)
X
576199
0.17
X
576199
0.17
48
4.39∗∗∗
(1.32)
−10.21∗
(5.68)
X
658523
0.15
X
658523
0.15
−10.11∗
(6.04)
X
479505
0.15
X
479505
0.15
Table 7
Panel Regression of Equity Returns and Local Beta for Subsamples
The table reports the relationship between the future returns of the firms located in an MSA and the local beta,
local
local
βm
, for various subsamples. Calculation of βm
is described in Table 2. RER subsamples are sorted based on
RER, defined as (buildings + capital leases) / Employees. Industry level RER is computed as the average RER of
firms in each industry. Tradable industries are all industries excluding the retail sector and restaurants. Non-unionized
industries are industries with unionization rates lower than the median unionization rate of all industries in that year.
Two subsamples of geographically focused firms are firms with below-median market capitalization, and firms that
mention two or fewer states in their 10-Ks. Log BM and Log Size are the log of the firm’s book to market ratio and
market equity constructed following Fama and French (1992). Leverage is firm’s market leverage as in Fan, Titman
and Twite (2012). P rof itability is gross profit measure as in Novy-Marx (2013). Investment is the investment ratio
as in Dougal, Parsons, and Titman (2015). Future returns are measured in the year following the portfolio formation,
from July of year t+1 to June of year t+2, and annualized (%). Regression sample period is 1986-2011. Standard errors
are clustered by firms and are presented in parentheses. ∗ , ∗∗ , and ∗∗∗ represent significance level of 10%, 5%, and 1%,
respectively.
Panel A. Subsamples for Tradable Industries and Non-Union Industries
Tradable Industries
Low RER Firms
(1)
local
βm
(2)
−11.57∗∗∗
(3.45)
Tradable, Non-Union Industries
Low RER Industries
(3)
−11.91∗∗∗
(3.55)
Low RER Firms
(4)
−8.58∗∗∗
(3.08)
(5)
−8.91∗∗∗
(3.16)
Low RER Industries
(6)
−15.09∗∗∗
(4.14)
−16.01∗∗∗
(4.23)
(7)
−10.54∗∗∗
(3.56)
(8)
−10.95∗∗∗
(3.65)
Log BM
6.97∗∗∗
(0.52)
6.82∗∗∗
(0.45)
7.07∗∗∗
(0.63)
7.03∗∗∗
(0.51)
Log Size
−1.41∗∗∗
(0.19)
−1.35∗∗∗
(0.15)
−1.50∗∗∗
(0.23)
−1.45∗∗∗
(0.18)
Leverage
−4.06∗∗
(1.74)
−2.26
(1.42)
−5.18∗∗∗
(2.00)
−3.30∗∗
(1.55)
10.20∗∗∗
(1.55)
Profitability
X
470862
0.16
10.89∗∗∗
(1.86)
−9.77∗
(5.71)
−5.88
(6.87)
Investment
Ind.×Time FE
Observations
R2
15.56∗∗∗
(1.43)
X
470862
0.16
X
646084
0.15
X
646084
0.15
16.04∗∗∗
(1.59)
−9.10
(8.55)
X
345218
0.15
X
345218
0.15
−9.77
(7.06)
X
513762
0.15
X
513762
0.15
Panel B. Subsamples for Geographically Focused Firms
Tradable, Market Size Below Median
Low RER Firms
(1)
local
βm
−16.76∗∗∗
(5.38)
(2)
−17.38∗∗∗
(5.55)
Tradable, Two or Fewer States in 10-K
Low RER Industries
(3)
−13.14∗∗∗
(5.00)
Low RER Firms
Low RER Industries
(4)
(5)
(6)
(7)
(8)
−12.44∗∗
(5.15)
−32.94∗
(17.92)
−31.07∗
(18.13)
−27.29∗∗
(13.65)
−25.68∗
(14.13)
Log BM
7.02∗∗∗
(0.73)
6.87∗∗∗
(0.66)
6.45∗∗∗
(2.31)
8.65∗∗∗
(1.70)
Log Size
−6.05∗∗∗
(0.60)
−5.25∗∗∗
(0.52)
−3.39∗∗∗
(0.93)
−2.71∗∗∗
(0.63)
Leverage
−7.37∗∗∗
(2.58)
−4.41∗∗
(2.10)
−5.58
(6.37)
−2.26
(4.81)
6.94∗∗∗
(2.06)
Profitability
−0.28
(10.33)
Investment
Ind.×Time FE
Observations
R2
11.40∗∗∗
(1.98)
X
258646
0.16
X
258646
0.16
−1.93
(8.74)
X
325705
0.14
X
325705
0.14
49
11.18∗∗
(5.21)
5.71
(6.67)
−5.99
(24.55)
47.68
(33.33)
X
41460
0.26
X
41460
0.27
X
58436
0.21
X
58436
0.22
Table 8
Industry-Adjusted Returns for Local Beta Sorted Portfolios
Tradable
Non-union firms
Tradable firms
All firms
Tradable
Small Firms
0.36
(1.21)
−0.26
(1.48)
0.20
(1.45)
Ind-adjusted return
FF 3-factor alpha
1.25
(1.10)
0.71
(0.92)
0.57
(0.96)
0.62
(0.86)
0.33
(0.92)
0.59
(0.86)
0.32
(0.91)
2
0.95
(1.04)
FF 3-factor alpha
0.87
(1.05)
0.57
(0.91)
FF 3-factor alpha
Ind-adjusted return
0.47
(0.91)
0.49
(0.92)
FF 3-factor alpha
Ind-adjusted return
0.36
(0.93)
Ind-adjusted return
Low
local
βm
1.04
(0.94)
0.86
(1.02)
0.73
(0.79)
0.32
(0.91)
0.59
(0.65)
0.34
(0.76)
0.40
(0.64)
0.14
(0.73)
3
2.45∗
(1.26)
2.57∗∗
(1.26)
2.75∗∗
(1.26)
3.46∗∗
(1.50)
−1.96∗∗∗
(0.73)
−2.10∗∗∗
(0.76)
−2.18∗∗∗
(0.73)
−2.60∗∗∗
(0.89)
−0.50
(0.63)
−0.28
(0.67)
−0.39
(0.66)
−0.66
(0.78)
3.71∗∗
(1.49)
3.27∗
(1.86)
3.79∗∗
(1.81)
−2.76∗∗∗
(0.87)
−3.53∗∗∗
(1.06)
−3.59∗∗∗
(1.04)
−0.88
(1.10)
−1.24
(1.07)
−0.75
(0.76)
2.28∗
(1.26)
Low-High
−1.91∗∗
(0.75)
High
local
βm
−0.39
(0.64)
4
Low RER Firms
0.39
(1.48)
0.04
(1.49)
0.19
(1.23)
0.35
(1.25)
−0.18
(0.86)
−0.19
(0.87)
−0.02
(0.80)
−0.01
(0.81)
Low
local
βm
Panel A. Portfolios Sorted by Local Beta and Firm Level RER
1.41
(1.28)
1.54
(1.30)
0.92
(1.02)
1.06
(1.02)
0.33
(0.70)
0.18
(0.71)
0.08
(0.65)
−0.01
(0.66)
2
0.47
(1.14)
0.96
(1.17)
1.10
(0.85)
1.19
(0.88)
0.98
(0.61)
1.08∗
(0.64)
0.81
(0.58)
0.91
(0.61)
3
2.09∗
(1.08)
2.48∗∗
(1.08)
0.84
(0.92)
0.90
(0.93)
0.58
(0.71)
0.77
(0.73)
0.86
(0.65)
1.06
(0.67)
4
High RER Firms
−1.49
(1.02)
−1.28
(1.08)
−0.86
(0.99)
−0.63
(1.09)
−0.86
(0.70)
−0.47
(0.78)
−0.53
(0.63)
−0.25
(0.69)
local
High βm
1.88
(1.88)
1.32
(1.91)
1.05
(1.77)
0.98
(1.80)
0.67
(1.23)
0.27
(1.25)
0.51
(1.13)
0.24
(1.15)
Low-High
The table presents the average industry-adjusted returns and their alphas from the Fama-French 3-factor model for portfolios sorted on the local beta,
local
. At the end of each June, firms are simultaneously sorted based on their local betas and real estate ratios (RER). Sorting on local betas is performed
βm
within each industry. Industry-adjusted returns are computed by subtracting mean returns of each industry from individual firm returns. Calculation of
local
is described in Table 2, RER is defined in Table 5. Panel A uses firm level RER, Panel B uses Industry RER. Results are presented for the entire
βm
sample, tradable industries, non-unionized industries and geographically focused firms (market cap below median), which are explained in Table 7. Sample
period is 1986-2011. Newey-West standard errors with 1 lag are presented in parentheses. ∗ , ∗∗ , and ∗∗∗ represent significance level of 10%, 5%, and 1%, respectively.
50
Tradable
Small Firms
Tradable
Non-union firms
Tradable firms
All firms
51
1.48
(1.16)
1.76
(1.14)
FF 3-factor alpha
0.96
(0.81)
FF 3-factor alpha
Ind-adjusted return
0.99
(0.82)
0.96
(0.74)
FF 3-factor alpha
Ind-adjusted return
1.04
(0.74)
1.08
(0.72)
FF 3-factor alpha
Ind-adjusted return
1.14
(0.73)
Ind-adjusted return
local
Low βm
0.91
(0.87)
0.62
(0.90)
0.37
(0.66)
0.32
(0.66)
0.16
(0.61)
0.08
(0.62)
0.08
(0.60)
0.03
(0.61)
2
1.14
(0.85)
1.07
(0.86)
0.89
(0.64)
0.63
(0.69)
0.83
(0.53)
0.61
(0.58)
0.64
(0.52)
0.43
(0.57)
3
2.31∗∗
(1.08)
2.44∗∗
(1.06)
2.40∗∗
(1.10)
2.55∗∗
(1.08)
2.44∗∗
(1.21)
2.63∗∗
(1.20)
4.02∗∗
(1.68)
4.63∗∗∗
(1.61)
−1.16∗
(0.62)
−1.36∗∗
(0.57)
−1.36∗∗
(0.64)
−1.59∗∗∗
(0.59)
−1.45∗∗
(0.73)
−1.66∗∗
(0.68)
−2.55∗∗
(1.00)
−2.87∗∗∗
(0.91)
−0.07
(0.55)
−0.12
(0.54)
−0.03
(0.56)
−0.06
(0.56)
−0.12
(0.61)
−0.22
(0.61)
0.17
(0.89)
−0.03
(0.88)
Low-High
local
High βm
4
Low RER Industry
−1.86
(1.79)
−2.52
(1.84)
−0.96
(1.96)
−0.82
(1.98)
−0.92
(0.87)
−1.15
(0.89)
−0.97
(0.83)
−1.14
(0.84)
local
Low βm
Panel B. Portfolios Sorted by Local Beta and Industry Level RER
2.76∗
(1.45)
2.40
(1.52)
2.33
(1.51)
2.29
(1.58)
0.93
(0.87)
0.53
(0.96)
0.54
(0.79)
0.22
(0.87)
2
Table 8 – Continued
Industry-Adjusted Returns for Local Beta Sorted Portfolios
0.17
(1.16)
0.69
(1.18)
1.50
(1.19)
1.77
(1.20)
0.70
(0.62)
0.79
(0.63)
0.49
(0.59)
0.58
(0.59)
3
1.16
(1.14)
1.68
(1.15)
0.85
(1.34)
0.72
(1.35)
0.63
(0.66)
0.90
(0.67)
0.87
(0.60)
1.15∗
(0.62)
4
High RER Industry
−1.38
(1.17)
−1.64
(1.18)
−2.12
(1.35)
−2.20
(1.38)
−0.80
(0.77)
−0.69
(0.78)
−0.53
(0.70)
−0.52
(0.71)
local
High βm
−0.48
(2.16)
−0.88
(2.19)
1.16
(2.59)
1.38
(2.59)
−0.11
(1.26)
−0.47
(1.26)
−0.44
(1.16)
−0.62
(1.16)
Low-High
Table 9
Model Parameter Values
Parameter
Description
Value
αl
αshigh , αslow
αklow , αkhigh
Ilow , Ihigh
β
γ0
γ1
ηk
ηs
δ
ρa
σa
ρz
σz
Labor share
Land share
Equipment share
Industry risk scalers
Discount factor
Constant price of risk parameter
Time varying price of risk parameter
Adjustment cost parameter for equipment
Adjustment cost parameter for land
Equipment depreciation rate
Persistence of aggregate productivity
Conditional volatility of aggregate productivity
Persistence of firm productivity
Conditional volatility of firm productivity
0.60
0.19, 0.09
0.11, 0.21
e−0.9 , e0.9
0.99
3.2
-12
1
1
0.08
0.922
0.014
0.7
0.27
Table 10
Model-Implied Factor Price Regressions
The table reports the effect of aggregate output shocks, ∆outputt , on the wage (Wm,t ) and land price (Pm,t ) growth in
local
local
is computed as average of the industry output betas operating in that
. βm
an area, conditional on the local beta, βm
area, weighted by the employment share of industries. All values are based on regressions run on data generated from
100 simulations of 2,000 firms for 50 periods (years). Point estimates are simulation medians of regression coefficients,
and confidence intervals (in parentheses) for the estimates are constructed from the 5th and 95th percentiles of the
simulated distributions of those estimates. Like the specifications presented in Tables 3 and 4, regressions include time
fixed effects.
Dependent variable:
∆ log(Wm,t )
∆ log(Pm,t )
local
∆outputt × βm
0.56
(0.25,0.91)
0.39
(0.16,0.62)
local
βm
0.00
(-0.88,0.88)
-0.01
(-0.43,0.43)
52
Table 11
Model-Implied Equity Betas and Firm Returns
The table reports the relationship between the conditional betas (βfcond
irm ) and expected returns of the firms located in
local
an area and the local beta, βm . Panel A presents the panel regression results for equity betas, Panel B presents
the panel regression results for expected equity returns. βfcond
irm is estimated running regressions of excess firm returns
local
is described in Table 10. Results are based on
on the market returns using 50-period windows. Calculation of βm
regressions run on data generated from 100 simulations of 2,000 firms for 50 periods. Point estimates are simulation
medians of regression coefficients (portfolio averages), and confidence intervals (in parentheses) for the estimates are
constructed from the 5th and 95th percentiles of the simulated distributions of those estimates. Like the specifications
presented in Tables 5, 6, and 7, panel regressions include industry × time fixed effects.
Panel A: Conditional Beta Regressions
Dependent Variable: βfcond
irm,t
local
βm
All
Low RER Industries
High RER Industries
-0.08
(-0.38,-0.02)
-0.10
(-0.45,-0.02)
-0.06
(-0.31,-0.02)
Panel B: Firm Return Regressions
Dependent Variable: rfe irm,t+1
local
βm
All
Low RER Industries
High RER Industries
-0.44
(-0.88,0.15)
-0.83
(-1.44,-0.08)
-0.05
(-0.30,0.38)
53
Appendix Tables
Table A1
List of Datasets
Data Source
Time Period
Purpose/Target
Data used in MSA β construction
BEA Industry GDP (SIC)
BEA Industry GDP (NAICS)
Census CBP
BEA MSA GDP
BEA Ind. Fixed Assets (SIC)
BEA Ind. Fixed Assets (NAICS)
BEA Ind. FTPT Emp. (SIC)
BEA Ind. FTPT Emp. (NAICS)
1947-1997
1977-2011
1986-2011
2001-2011
1947-1997
1977-2011
1948-1997
1977-2011
Calculate industry β in 1986-1997
Calculate industry β in 1998-2011
Obtain industry employment weights in each MSA
Output
Calculate an alternative measure of local β: βm
TFP
Calculate an alternative measure of local β: βm
in
TFP
Calculate an alternative measure of local β: βm
in
TFP
Calculate an alternative measure of local β: βm
in
TFP
Calculate an alternative measure of local β: βm
in
Labor data
Census LEHD
BLS OES
www.unionstats.com
1990-2011
1999-2011
1983-2011
Obtain the annual wage for each industry in each MSA
Obtain the hourly wage for each occupation in each MSA
Obtain the union coverage at industry and occupation level
Real estate data
FHFA HPI
NCREIF NPI
CoStar Rent
1975-2011
1978-2011
1982-2011
Obtain the housing price index in each MSA
Obtain the commercial real estate returns for each MSA
Obtain the average rent of office buildings in each MSA
1990-2005
Track the change of firms’ headquarter locations
1992-2008
Obtain a proxy for firms’ geographic concentration
2013-2014
Obtain firms’ employment distribution across MSAs
Firm location data
Compact Disclosure
State counts from Garcia and
Norli (2012)
www.referenceusa.com
1986-1997
1998-2011
1986-1997
1998-2011
Financial and accounting data
Obtain firms’ location, industry, and other characteristics
Calculate monthly stock returns and equity βs
Compustat (annual)
CRSP Stock (monthly)
Table A2
Employment Share of Headquarter MSA
The table reports the percentage of firms that have at least 50%, 75%, 90%, or 100% of their employment in their
headquarter MSA. We create an employment map of firms by linking our Compustat-CRSP sample to ReferenceUSA
U.S. Businesses Database and collecting employment data for all headquarter, branch, and subsidiary locations of firms
in our sample. ReferenceUSA data are collected in November 2014 for businesses that are active at that time. We
aggregate the employment counts of firm establishments at the MSA level. Employment share of the headquarter
MSA is computed by dividing total employment in that MSA by total number of employees in all MSAs. Firm size
is measured with equity market capitalization in December 2013. First panel divides the sample into two size groups
(above-below median), second panel divides it into five groups (quintile portfolios).
# Firms
≥ 50%
Percentage of Employees Working in the Headquarter MSA
≥ 75%
≥ 90%
=1
All Firms
Small Firms
Large Firms
2008
1004
1004
62.75%
71.81%
53.69%
47.46%
60.26%
34.66%
37.60%
51.89%
23.31%
27.09%
42.13%
12.05%
All Firms
Small Firms
Q2 Firms
Q3 Firms
Q4 Firms
Large Firms
2008
402
402
401
402
401
62.75%
78.61%
69.90%
59.35%
57.46%
48.38%
47.46%
69.90%
58.96%
41.90%
39.05%
27.43%
37.60%
63.68%
48.51%
31.42%
25.87%
18.45%
27.09%
55.22%
36.57%
20.95%
13.43%
9.23%
54