Full text of Regional Economic Development : Selected Papers From the Third Annual Conference of the Business & Economics Research Group (BERG) : Volume 3, Number 2, 2007

View original document

The full text on this page is automatically extracted from the file linked above and may contain errors and inconsistencies.

Federal Reserve Bank of St. Louis

REGIONAL ECONOMIC
DEVELOPMENT
VO LU M E 3 , N U M B E R 2

2007

Selected Papers from the Third Annual Conference of the
Business & Economics Research Group (BERG)
Forecasting Real Housing Price Growth
in the Eighth District States
David E. Rapach and Jack K. Strauss
Educational Attainment and Recovery from Recessions
Across Metropolitan Areas
Bryan Bezold
Transferable Tax Credits in Missouri:
An Analytical Review
Paul Rothstein and Nathan Wineinger
How Well Are the States of the Eighth
Federal Reserve District Prepared for the Next Recession?
Gary A. Wagner and Erick M. Elder
The Economic Impact of
Broadband Deployment in Kentucky
David Shideler, Narine Badasyan, and Laura Taylor

REGIONAL ECONOMIC
DEVELOPMENT

Selected Papers from the
Third Annual Conference of the
Business & Economics
Research Group (BERG)

Director of Research

Editor’s Introduction

Robert H. Rasche

Thomas A. Garrett

Deputy Director of Research

Cletus C. Coughlin
Editor

Thomas A. Garrett

33
Forecasting Real Housing Price Growth
in the Eighth District States
David E. Rapach and Jack K. Strauss

Center for Regional Economics—8th District (CRE8)
Director

Howard J. Wall

Educational Attainment and
Recovery from Recessions
Across Metropolitan Areas

Subhayu Bandyopadhyay
Cletus C. Coughlin
Thomas A. Garrett
Rubén Hernández-Murillo
Natalia A. Kolesnikova
Michael R. Pakko
Christopher H. Wheeler

Bryan Bezold

53
Transferable Tax Credits in Missouri:
An Analytical Review

Managing Editor

Paul Rothstein and Nathan Wineinger

George E. Fortier
Editor

Lydia H. Johnson
Graphic Designer

Donna M. Stiller

The views expressed are those of the individual authors and
do not necessarily reflect official positions of the Federal
Reserve Bank of St. Louis, the Federal Reserve System, or the
Board of Governors.

75
How Well Are the States of the
Eighth Federal Reserve District
Prepared for the Next Recession?
Gary A. Wagner and Erick M. Elder

88
The Economic Impact of
Broadband Deployment in Kentucky
David Shideler, Narine Badasyan, and Laura Taylor

F E D E R A L R E S E R V E B A N K O F S T. LO U I S R E G I O N A L E C O N O M I C D E V E LO P M E N T

V O LU M E 3 , N U M B E R 2

2007

Regional Economic Development is published occasionally by the Research Division of
the Federal Reserve Bank of St. Louis and may be accessed through our web site:
research.stlouisfed.org/regecon/publications/. All nonproprietary and nonconfidential
data and programs for the articles written by Federal Reserve Bank of St. Louis staff and
published in Regional Economic Development also are available to our readers on this
web site.
General data can be obtained through FRED (Federal Reserve Economic Data), a database
providing U.S. economic and financial data and regional data for the Eighth Federal
Reserve District. You may access FRED through our web site: research.stlouisfed.org/fred.
Articles may be reprinted, reproduced, published, distributed, displayed, and transmitted in their entirety if copyright notice, author name(s), and full citation are included.
Please send a copy of any reprinted, published, or displayed materials to George Fortier,
Research Division, Federal Reserve Bank of St. Louis, P.O. Box 442, St. Louis, MO 631660442; george.e.fortier@stls.frb.org. Please note: Abstracts, synopses, and other derivative
works may be made only with prior written permission of the Federal Reserve Bank of St.
Louis. Please contact the Research Division at the above address to request permission.
© 2007, Federal Reserve Bank of St. Louis.
ISSN 1930-1979

V O LU M E 3 , N U M B E R 2

2007

F E D E R A L R E S E R V E B A N K O F S T. LO U I S R E G I O N A L E C O N O M I C D E V E LO P M E N T

Contributing Authors
Narine Badasyan
Murray State University
narine.badasyan@murraystate.edu

David Shideler
Murray State University
david.shideler@murraystate.edu

Bryan Bezold
St. Louis Regional Chamber and Growth
Association
bbezold@stlrcga.org

Jack K. Strauss
Saint Louis University
strausjk@slu.edu

Erick M. Elder
University of Arkansas at Little Rock
emelder@ualr.edu
Thomas A. Garrett
Federal Reserve Bank of St. Louis
tom.a.garrett@stls.frb.org
David E. Rapach
Saint Louis University
rapachde@slu.edu

Laura Taylor
Connected Nation, Inc.
ltaylor@connectky.org
Gary A. Wagner
University of Arkansas at Little Rock
gawagner@ualr.edu
Nathan Wineinger
Washington University in St. Louis
ndwinein@wustl.edu

Paul Rothstein
Washington University in St. Louis
rstein@wustl.edu

F E D E R A L R E S E R V E B A N K O F S T. LO U I S R E G I O N A L E C O N O M I C D E V E LO P M E N T

V O LU M E 3 , N U M B E R 2

2007

iii

Editor’s Introduction
Thomas A. Garrett

he Center for Regional Economics–8th
District (CRE8) at the Federal Reserve
Bank of St. Louis sponsored the third
annual conference of the Business and
Economics Research Group (BERG) in St. Louis in
June 2007. Researchers from university economics
departments and centers for business and economic research located within the Eighth Federal
Reserve District presented a wide variety of papers
on economic issues relevant to District states.1
This issue of Regional Economic Development
contains five papers from the third annual BERG
conference. David Rapach and Jack Strauss of
Saint Louis University forecast real housing price
growth for Eighth Federal Reserve District states
using an autoregressive distributive lag model.
Bryan Bezold of the St. Louis Regional Chamber
and Growth Association explores the relationship
between educational attainment in metropolitan

areas and the speed at which metropolitan areas
recover from recessions. Paul Rothstein and Nathan
Wineinger from Washington University in St. Louis
provide an analysis of Missouri’s tax credit programs and whether the design of each program is
consistent with its intended economic purpose.
Gary Wagner and Erick Elder from the University
of Arkansas at Little Rock use a Markov-switching
regression model to describe economic expansions
and contractions in Eighth Federal Reserve District
states and to forecast the probability that each
state’s revenue balance will not be adequate to
weather a recession. David Shideler and Narine
Badasyan of Murray State University and Laura
Taylor from Connected Nation, Inc., use countylevel data for the state of Kentucky to assess whether
broadband infrastructure influences the industrial
competitive of Kentucky’s counties.

The third annual BERG conference agenda, along with BERG members
and past events, can be found at http://research.stlouisfed.org/berg/.

Thomas A. Garrett is an assistant vice president and economist at the Federal Reserve Bank of St. Louis.
Federal Reserve Bank of St. Louis Regional Economic Development, 2007, 3(2), p. 32.
© 2007, The Federal Reserve Bank of St. Louis. Articles may be reprinted, reproduced, published, distributed, displayed, and transmitted in
their entirety if copyright notice, author name(s), and full citation are included. Abstracts, synopses, and other derivative works may be made
only with prior written permission of the Federal Reserve Bank of St. Louis.

V O LU M E 3 , N U M B E R 2

2007

F E D E R A L R E S E R V E B A N K O F S T. LO U I S R E G I O N A L E C O N O M I C D E V E LO P M E N T

Forecasting Real Housing Price Growth
in the Eighth District States
David E. Rapach and Jack K. Strauss
The authors consider forecasting real housing price growth for the individual states of the Federal
Reserve’s Eighth District. They first analyze the forecasting ability of a large number of potential
predictors of state real housing price growth using an autoregressive distributed lag (ARDL) model
framework. A number of variables, including the state housing price-to-income ratio, state unemployment rate, and national inflation rate, appear to provide information that is useful for forecasting real housing price growth in many Eighth District states. Given that it is typically difficult to
determine a priori the particular variable or small set of variables that are the most relevant for
forecasting real housing price growth for a given state and time period, the authors also consider
various methods for combining the individual ARDL model forecasts. They find that combination
forecasts are quite helpful in generating accurate forecasts of real housing price growth in the
individual Eighth District states. (JEL C22, C53, E37)
Federal Reserve Bank of St. Louis Regional Economic Development, 2007, 3(2), pp. 33-42.

he rollercoaster ride of the housing
market continues to receive considerable
attention in the popular and financial
press. There is currently speculation of
a precipitous drop in housing prices in certain
regions of the country after the sharp rise in
housing prices (“bubble”?) over the past decade.
Policymakers are keenly interested in housing
price fluctuations and their potential impact on
household consumption spending, as evinced by
numerous comments by former Federal Reserve
Chairman Alan Greenspan and current Chairman
Ben Bernanke. This interest appears warranted:
The median household now holds more of its
wealth in housing than in stocks and has greater
access to cash through refinancing backed by
housing wealth (Greenspan and Kennedy, 2005).
Given the substantial interest in housing price
fluctuations, the present paper investigates fore-

casts of real housing price growth in the individual
states of the Federal Reserve’s Eight District
(Arkansas, Illinois, Indiana, Kentucky, Missouri,
Mississippi, and Tennessee). We focus on forecast
horizons of four and eight quarters because these
horizons are relevant to forecasting over the business cycle, and most recent discussions of housing
price fluctuations focus on possible swings in housing prices over business-cycle horizons.1 We consider a large number of potential predictors (25)
of real housing price growth for each state. This is
motivated by a sizable literature that examines the
determinants of housing prices using in-sample
1

The literature on forecasting housing prices in the United States at
the aggregate or state level is relatively sparse, especially compared
with the massive literature on forecasting economic variables such
as U.S. gross domestic product (GDP) and inflation. The extant literature on forecasting housing prices in the United States tends to
focus on long-run trends (Hendershott and Weicher, 2002).

David E. Rapach is an associate professor of economics and Jack K. Strauss is the Simon Professor of Economics at Saint Louis University. The
authors acknowledge financial support from the Simon Center for Regional Forecasting at Saint Louis University. The authors thank participants
at the 2007 Eighth District Business and Economics Research Group (BERG) conference and a referee for very helpful comments. The results
reported in this paper were generated using GAUSS 6.0; the GAUSS programs are available at http://pages.slu.edu/faculty/rapachde/Research.htm.

© 2007, The Federal Reserve Bank of St. Louis. Articles may be reprinted, reproduced, published, distributed, displayed, and transmitted in
their entirety if copyright notice, author name(s), and full citation are included. Abstracts, synopses, and other derivative works may be made
only with prior written permission of the Federal Reserve Bank of St. Louis.

F E D E R A L R E S E R V E B A N K O F S T. LO U I S R E G I O N A L E C O N O M I C D E V E LO P M E N T

V O LU M E 3 , N U M B E R 2

2007

Rapach and Strauss

tests; see, for example, Cho (1996), Abraham and
Hendershott (1996), and Johnes and Hyclak (1999).
Potential determinants of housing prices in this
literature include income, interest rates, construction costs, and labor market variables such as the
unemployment rate and size of the labor force.2
Following Stock and Watson (1999, 2003, and
2004), we generate simulated out-of-sample forecasts of real housing price growth using an autoregressive distributed lag (ARDL) model framework.
More specifically, when forecasting real housing
price growth for a given Eighth District state, we
estimate 25 individual ARDL forecasting models
of real housing price growth, where each ARDL
model includes one of the potential predictors.
This provides a convenient framework for analyzing the forecasting ability of each of the individual
potential predictors of real housing price growth.
The plethora of potential predictors of real
housing price growth also leads us to consider
combination forecasts. Typically, it is difficult to
identify a priori the particular variable (or small
set of variables) that is most relevant for forecasting
a variable such as real housing price growth, especially because the predictive ability of individual
variables can vary markedly over time.3 Combination forecasts provide a way of incorporating
information that may be useful for forecasting in
environments with a large number of potential
predictors, and they have been shown to work well
in a number of recent forecasting applications
involving GDP growth, inflation, and employment
growth; see, for example, Stock and Watson (1999,
2003, and 2004) and Rapach and Strauss (2005 and
2007). We consider a number of different methods
for combining the individual ARDL model forecasts from the extant literature and investigate
their ability to help generate reliable forecasts of

real housing price growth in the Eighth District
states.4
Previewing our results, we find that a number
of the individual predictors are able to improve on
forecasts of real housing price growth relative to
an autoregressive (AR) benchmark model, sometimes substantially. These variables include the
housing price-to-income ratio, state unemployment
rate, and national inflation rate. However, there is
no single variable that is able to improve on the
AR model forecasts across all of the Eighth District
states at all of the forecast horizons considered,
and there are instances where a variable that performs very well for one particular state performs
poorly for another. Fortunately, we also find that
some of the forecast combining methods perform
quite well and almost always provide sizable
improvements in forecast accuracy relative to the
AR benchmark model.
The rest of the paper is organized as follows:
The next section outlines the econometric methodology, and the third section presents the empirical
results.

ECONOMETRIC METHODOLOGY
Let ∆yt = yt – yt –1, where yt is the log level of
real housing prices at time t. Furthermore, let
h

y th+ h = (1 h ) ∑ j =1 ∆y t + j ,

so that y th+h is the (approximate) growth rate of real
housing prices from time t to t + h; h is the forecast
horizon. Let xi,t (i = 1,…, n) represent one of n
potential predictors of real housing price growth.
An individual ARDL model based on the predictor
xi,t is given by
(1)

We focus on real housing price growth at the state level in this paper
primarily because it allows us to examine regional differences in
housing price fluctuations while still having a fairly large number
of potential predictors available at the state level. Although statelevel housing prices are able to capture some important geographic
differences in housing price fluctuations, as we mention in the conclusion, we are also planning to investigate forecasts of real housing
price growth for individual metropolitan areas in the Eighth District
in future research.
See Stock and Watson (2003) for evidence of this in the context of
forecasting U.S. GDP growth and inflation.

V O LU M E 3 , N U M B E R 2

2007

y th+ h = α +

q1 −1

q2 −1

j =0

∑ β j ∆y t − j + ∑ γ j x i, t − j + εth+ h ,

where ε th+h is an error term. Equation (1) can be
used to construct a set of recursive (expanding
estimation window) simulated out-of-sample forecasts of y th+h using information available at time t,
4

See Timmermann (2006) for a recent survey of forecast combining
methods.

F E D E R A L R E S E R V E B A N K O F S T. LO U I S R E G I O N A L E C O N O M I C D E V E LO P M E N T

Rapach and Strauss

and we denote the forecast of y th+h formed at time t
h
for a given predictor xi,t by ŷ i,t+h|
t . More specifically,
h
ŷi,t+h|t is calculated by plugging ∆yt –j 共j = 0,…,q1 – 1兲
and xi,t –j 共j = 0,…,q2 – 1兲 into (1), with the α, β j ,
and γ j parameters set equal to their ordinary least
squares (OLS) estimates based on data available
from the start of the sample through period t and
ε th+h set to its expected value of zero. We select the
lag lengths (q1 and q2) in (1) using the Schwarz
information criterion (SIC) and a minimum value
of zero for q1 and one for q2 (to ensure that the
potential predictor xi,t appears in (2)) and a maximum value of four for q1 and q2.5 Dividing the total
sample into in-sample and out-of-sample portions
of size R and P, respectively, we use this procedure
to generate a series of P – 共h – 1兲 recursive simulated out-of-sample forecasts for the ARDL model
that includes xi,t , which we denote as

{ ŷ

h
i, t + h t

T −h

}

t =R

Note that the lag lengths q1 and q2 are selected
anew when forming each out-of-sample forecast,
so that the lag lengths for the ARDL forecasting
model are allowed to vary through time. In our
applications in the next section, we consider 25
potential predictors, and so we will have 25 series
of h-step-ahead individual ARDL model forecasts
of real housing price growth for each of the seven
states in the Eighth District.
We also compute recursive simulated out-ofsample forecasts for an AR model, which is given
by (1) with the restriction γ j = 0 共j = 0,…,q2 – 1兲
imposed. The series of out-of-sample forecasts are
generated using a procedure analogous to that for
the ARDL forecasting model described above.7
Following much of the forecasting literature, the
5

The SIC and the Akaike information criterion (AIC) are two popular
model selection procedures. Note that we obtain similar results when
we select the lag lengths in (1) using the AIC.

Note that the first forecast uses all data available at time R to form a
h
h
forecast of y R+h
; this forecast is denoted by ŷ i,R+h|
R . The information
set is then updated by one period, and we use all data available at
h
time R + 1 to form a forecast of y共R+1兲+h
; this second forecast is denoted
h
by ŷ i,共R+1兲+h|
.
We
continue
in
this
manner through the end of the
R+1
out-of-sample period, leaving us with P – 共h – 1兲 recursive simulated
h
T–h
out-of-sample forecasts, {ŷ i,t+h|
t } t =R .

We select the lag length (q1) for the AR model using the SIC and a
minimum (maximum) value of zero (four) for q1.

AR model serves as a benchmark forecasting model.
We consider three types of methods for combining the individual ARDL model forecasts.
Some of the combining methods require a holdout
period to calculate the weights ({wi,t }in= 1) used to
combine the individual ARDL model forecasts,
and we use the first P0 observations from the outof-sample period as the initial holdout period.
This leaves us with a total of P – 共h – 1兲 – P0 outof-sample forecasts available for evaluation.8 In
our applications in the next section, we evaluate
the benchmark AR model, individual ARDL model,
and combination forecasts over the 1995:Q1–
2006:Q4 out-of-sample period. Importantly, this
period includes the bull housing market that has
prevailed in many parts of the country over the
past decade.
The first type of combining method uses simple
schemes: mean, median, and trimmed mean. The
mean (median) combination forecast is simply the
average (median) of the individual ARDL model
forecasts, while the trimmed mean combination
forecast takes the average of the individual ARDL
model forecasts after dropping the highest and
lowest individual ARDL model forecasts. Stock
and Watson (1999 and 2003) find that simple combinations of individual ARDL model forecasts
consistently outperform an AR benchmark model
(although by a fairly limited margin) with respect
to forecasting U.S. real GDP growth and inflation.9
The second type of combining procedure we
employ uses a discount mean square forecast error
(DMSFE) criterion over the holdout out-of-sample
period to determine the weights used to combine
the individual ARDL model forecasts formed at
time t; see Stock and Watson (2004). More specifically, the DMSFE combining method uses the
weights
8

Note that we use the first P0 observations from the out-of-sample
period to estimate the combining weights used to generate the first
combination forecast available for evaluation. We then use the first
P0 + 1 observations from the out-of-sample period to estimate the
combining weights used to generate the second combination forecast available for evaluation. We continue in this manner through
the end of the available out-of-sample period, leaving us with a
series of P – 共h – 1兲 – P0 out-of-sample combination forecasts available for evaluation.

The simple combining methods obviously do not require a holdout
period, as the combining weights are not estimated.

F E D E R A L R E S E R V E B A N K O F S T. LO U I S R E G I O N A L E C O N O M I C D E V E LO P M E N T

V O LU M E 3 , N U M B E R 2

2007

Rapach and Strauss

w i, t = mi−,1t

∑ j =1 m−j ,1t ( i = 1,…, n),

EMPIRICAL RESULTS
Data

where

mi, t = ∑

t − h t − h−s
θ
s =R

h
s+h

− yˆ ih, s + h s

)

and the parameter θ is a discount factor. When
θ = 1, there is no discounting, whereas θ < 1 means
that greater importance is attached to the recent
forecasting performance of the individual ARDL
models in determining the combining weights. In
the next section, we consider θ values of 1.0 and
0.9 in our applications.
The final type of combining method we use is
the “cluster” approach recently developed by Aiolfi
and Timmermann (2006) based on their C共K,PB兲
algorithm. The initial cluster combination forecast
is generated by first grouping the individual ARDL
model forecasts over the holdout out-of-sample
period,

{ŷ

h
i, s + h h

R + ( P0 −1)− ( h −1)

}

( i = 1,…,n),

s=R

into K equal-sized clusters based on the MSFE,
with the first cluster containing the individual
ARDL model forecasts with the lowest MSFE values, the second cluster containing the individual
ARDL model forecasts with the next lowest MSFE
values, and so on. The initial combination forecast
is the average of the individual ARDL model forecasts contained in the first cluster. To form the
second combination forecast, the MSFE is computed for the individual ARDL model forecasts,

{

ŷ ih, s + h h

R + ( P0 −1)− ( h −1)+1

}

Nominal housing price indices for individual
U.S. states starting in 1975:Q1 are available from
Freddie Mac. The Conventional Mortgage Home
Price Index provides a means for measuring the
typical price inflation for houses within the United
States using matched transactions on the same
property over time to account for quality changes.
Freddie Mac uses data from both purchase and
refinance-appraisal transactions, and its database
consists of over 33 million homes. The available
sample for the housing price indices ends in
2006:Q4. We convert the nominal housing price
index into real terms using the personal consumption expenditure (PCE) deflator from the Bureau
of Economic Analysis (BEA). We then compute
annualized growth rates as 400 times the differences in the log levels of real housing prices. The
annualized real housing price growth rates are
plotted in Figure 1. Note that real housing price
growth is predominantly positive over much of
our 1995:Q1–2006:Q4 out-of-sample forecast evaluation period, indicating that the individual states
of the Eighth District typically experienced fairly
strong housing markets over the past decade.10
As discussed above, we consider 25 potential
predictors of real housing price growth for each
state. Six of these are state-level variables:
• Ratio of housing price to per capita personal
income
• Real per capita personal income
• Population
• Employment

( i = 1,…,n),

• Labor force

s = R +1

• Unemployment rate
and the individual ARDL model forecasts are
again grouped into K clusters based on the MSFE.
The second combination forecast is again the average of the individual forecasts in the first cluster.
We can proceed in this manner through the end of
the available out-of-sample period to construct the
complete set of combination forecasts. Following
Aiolfi and Timmermann (2006), we consider K
values of two and three in our applications in the
next section.
36

V O LU M E 3 , N U M B E R 2

2007

Nominal personal income data are from the
BEA and are converted into per capita terms using
10

The housing price indices exhibit exaggerated saw-tooth patterns in
the first part of the sample for a number of the states. This appears
to be an artifact of the development and construction of the housing
price indices. To minimize the influence of these patterns when
estimating the forecasting models, we smooth the real housing price
growth observations up to 1984:Q4 by taking a moving average of
the current and three previous real housing price observations.
Smoothing of the early observations has been applied to the real
housing price growth rate series depicted in Figure 1.

F E D E R A L R E S E R V E B A N K O F S T. LO U I S R E G I O N A L E C O N O M I C D E V E LO P M E N T

Rapach and Strauss

Figure 1
Annualized Real Housing Price Growth, 1976:Q1–2006:Q4
20

Arkansas

Illinois

−10

−20

1980 1985 1990 1995 2000 2005

Kentucky

−20

1980 1985 1990 1995 2000 2005

Missouri

−20

−10

1980 1985 1990 1995 2000 2005

−20

1980 1985 1990 1995 2000 2005

Mississippi

−20

Indiana

1980 1985 1990 1995 2000 2005

−20

1980 1985 1990 1995 2000 2005

Tennessee

10
0
−10
−20

1980 1985 1990 1995 2000 2005

population data from the U.S. Census Bureau and
then into real terms using the PCE deflator. The
labor market variables are from the Bureau of Labor
Statistics (BLS). The housing price-to-income ratio
is a popular “valuation ratio” for housing prices
that may help to signal whether housing is overor under-valued. The income and employment
variables provide measures of the ability of households to purchase housing and are thus potentially
important determinants of housing demand. Significant changes in population can also lead to sizable
shifts in housing demand.
We also consider five regional variables as
predictors:
• Housing starts
• Building permits

• Homes for sale
• Homes sold
• Housing vacancy rate
These variables, all from the U.S. Census Bureau,
are available for each of the four U.S. Census
regions.11 These housing market variables may
provide signals of trends and supply conditions in
housing markets that affect housing prices.
Finally, 14 national variables also serve as
predictors:
• Average weekly hours in manufacturing
11

Reflecting their U.S. Census Bureau classification, we use variables
from the South region for Arkansas, Kentucky, Mississippi, and
Tennessee and variables from the Midwest region for Illinois,
Indiana, and Missouri. Note that these variables are not available
at the state level for the entire sample period we consider.

F E D E R A L R E S E R V E B A N K O F S T. LO U I S R E G I O N A L E C O N O M I C D E V E LO P M E N T

V O LU M E 3 , N U M B E R 2

2007

Rapach and Strauss

• Average weekly initial claims for unemployment insurance
• Manufacturers’ new orders for consumer
goods and materials (in chained 1982 dollars)
• Vendor performance
• Manufacturers’ new orders of nondefense
capital goods (in chained 1982 dollars)
• S&P 500 stock price index
• Real M2 money supply (in chained 2000
dollars)
• Term spread (10-year Treasury bond yield
minus the federal funds rate)
• Consumer confidence index
• PCE deflator
• Industrial production
• Commercial and industrial loans outstanding
(in chained 2000 dollars)
• Consumer installment credit outstanding
• Real effective mortgage rate
The first nine national predictors comprise
nine of the ten leading economic indicators from
the Conference Board12: These indicators potentially measure broad economic trends that can
affect the demand for housing. Data on industrial
production, commercial and industrial loans outstanding, and consumer installment credit outstanding are all from the Conference Board. These
variables include credit measures that also may
significantly influence housing prices. The nominal
effective mortgage rate is from Freddie Mac, and
we subtract the inflation rate based on the PCE
deflator to approximate a real effective mortgage
rate. The mortgage rate is an important component
of the “user cost” of housing and thus a potentially
important determinant of housing demand.
All of the predictors are transformed in an
effort to render them stationary. This involves
taking the first differences of log levels, with the
following exceptions: We use levels for the unemployment rate, housing vacancy rate, unemployment claims, vendor performance, term spread,
and consumer confidence; we use log levels for
12

The leading indicator we omit is national building permits, as this
is included as a regional predictor.

V O LU M E 3 , N U M B E R 2

2007

the housing price-to-income ratio; and we use first
differences for average weekly hours.

AR Benchmark and Individual ARDL
Model Forecasting Results
Table 1 reports forecasting results for the AR
benchmark and individual ARDL forecasting
models for each state. The table reports the MSFE
for the AR benchmark model and the ratio of the
individual ARDL model MSFE to the AR benchmark model MSFE. A ratio below unity thus indicates that the individual ARDL model has a lower
MSFE than the AR benchmark. Results are reported
for forecast horizons of four (h = 4) and eight (h = 8)
quarters.
An important result in Table 1 is that no single
predictor has an MSFE ratio that is below unity in
all states for both forecast horizons; that is, there
is no single predictor that delivers consistently
more accurate forecasts than the benchmark AR
model across all of the Eighth District states and
both forecast horizons. The PCE deflator (inflation
rate) produces an MSFE ratio below unity for all
seven states at both horizons, with one exception
(Indiana at h = 4), and many of the MSFE ratios for
the inflation rate are well below unity (for example,
0.26 for Kentucky at h = 8), indicating substantial
reductions in forecast accuracy relative to the AR
model. The state housing price-to-income ratio—
as mentioned above, a popular valuation ratio for
housing—also performs quite well for Arkansas,
Indiana, Kentucky, Mississippi, and Tennessee,
with MSFE ratios all below unity (often substantially so) at both forecast horizons. However, the
MSFE ratios for the state housing price-to-income
ratio are well above unity for both horizons for
Illinois and above unity for Missouri at the eightquarter horizon. Other predictors that perform well
for a number of Eighth District states are the state
unemployment rate and consumer confidence, but
there again are situations where the MSFE ratios
for these variables are considerably above unity.
Looking at the results in Table 1 on a state-bystate basis, the state housing price-to-income ratio
and state unemployment rate stand out for
Arkansas. These predictors generate reductions in
MSFE relative to the AR benchmark model of up
to 31 percent and 54 percent at the four- and eight-

F E D E R A L R E S E R V E B A N K O F S T. LO U I S R E G I O N A L E C O N O M I C D E V E LO P M E N T

0.46

0.70

State housing
price-to-income ratio

1.36

F E D E R A L R E S E R V E B A N K O F S T. LO U I S R E G I O N A L E C O N O M I C D E V E LO P M E N T

V O LU M E 3 , N U M B E R 2

2007

1.01

1.08
1.15

1.06

1.07

1.21

0.83

1.09

1.05

0.98

1.08

1.01

0.93

1.31

1.03

1.58

0.82

1.00

1.07

1.18

1.39

1.05

1.07

0.88

1.02

1.89

4.66

h=4

1.37

1.01

0.92

1.11

0.58

1.16

1.10

0.75

1.10

1.00

1.01

1.13

1.02

1.23

0.57

1.00

0.97

1.02

1.06

1.89

0.71

1.09

0.84

1.03

2.74

5.66

h=8

Illinois

1.03

0.93

1.21

0.99

1.08

1.03

1.25

1.12

1.01

1.13

0.96

1.06

0.90

1.20

1.17

1.06

1.29

1.11

1.29

0.69

1.13

1.17

0.81

0.95

0.84

2.15

h=4

1.07

0.99

1.26

0.93

0.96

1.20

1.76

1.04

1.02

0.94

1.10

0.67

1.18

1.23

1.00

1.09

1.01

1.13

0.47

1.28

1.31

0.50

1.13

0.69

1.93

h=8

Indiana

1.03

1.17

1.08

1.05

0.68

0.75

2.00

0.95

1.05

1.01

1.16

1.17

0.97

1.40

0.85

1.97

1.33

1.10

1.15

0.67

1.07

1.98

0.78

0.99

0.67

1.18

h=4

1.05

1.12

1.02

1.08

0.26

0.54

1.96

0.89

1.03

1.01

1.07

1.29

0.91

1.20

0.62

0.65

1.17

0.99

1.01

0.69

0.98

1.47

0.70

1.01

0.38

2.25

h=8

Kentucky

0.97

1.33

1.02

1.07

0.53

0.67

1.16

0.77

1.03

1.01

1.03

0.97

1.08

1.11

1.00

1.05

1.12

1.24

1.21

1.17

1.38

0.90

1.14

0.92

2.93

h=4

1.29

1.38

0.99

1.13

0.32

0.44

1.65

0.80

1.03

1.02

1.09

1.12

1.04

1.16

0.83

1.00

1.05

1.14

1.22

2.05

1.17

1.60

0.96

1.43

1.13

4.69

h=8

Missouri

1.11

1.37

1.04

1.03

0.63

0.88

1.30

0.92

1.06

1.03

1.11

1.14

0.98

1.16

1.08

1.15

1.18

1.03

0.99

1.08

1.23

1.02

0.58

9.70

h=4

1.13

1.22

1.01

0.98

0.43

0.56

1.01

0.90

1.04

1.00

0.91

1.05

0.96

1.08

1.10

1.02

1.01

0.95

0.94

0.89

1.01

1.18

0.75

1.01

0.35

9.44

h=8

Mississippi

1.12

1.30

1.16

1.08

0.89

0.94

1.23

1.02

1.01

1.05

1.08

0.96

1.17

0.95

1.08

1.22

1.04

1.11

1.15

1.29

1.33

1.02

1.05

0.76

3.43

h=4

1.19

1.12

1.02

1.00

0.62

0.72

1.29

0.94

1.02

1.00

0.97

1.06

0.97

1.11

0.83

0.86

1.08

1.01

0.95

1.34

1.16

1.04

1.06

0.52

4.56

h=8

Tennessee

NOTE: Entries in the AR MSFE row report the MSFE for the AR benchmark model. Entries in the other rows report the ratio of the MSFE for the individual ARDL model
that includes the predictor indicated in the first column to the MSFE for the AR benchmark model.

Real effective mortgage rate 0.95

1.23

Consumer credit
outstanding

0.98

1.04

1.05
1.28

Industrial production

Commercial/industry loans

0.82
0.64

0.87

Consumer confidence

PCE deflator (inflation rate) 0.81

0.95
0.82

0.94
0.81

Real M2

1.04

1.00

0.93

1.04

1.02

0.99

Term spread

1.00
1.05

New orders–capital goods

0.90

Vendor performance

S&P 500 index

1.14

New orders–consumer
goods

1.10

1.33
0.95

Average weekly hours

1.02
1.18

1.02
1.12

Regional homes sold

Regional vacancy rate

Unemployment claims

1.21

1.06
1.36

Regional building permits

Regional homes for sale

0.49
0.97

0.69
1.08

State unemployment rate

1.08

Regional housing starts

1.41
1.14

State employment

1.82

State labor force

1.02
1.45

State real personal income 1.01

State population

5.49

4.90

AR MSFE

h=8

h=4

Predictor

Arkansas

Individual ARDL Model Forecast Results: 1995:Q1–2006:Q4 Out-of-Sample Forecast Evaluation Period

Table 1

Rapach and Strauss

V O LU M E 3 , N U M B E R 2

2007

0.87

C共3,PB兲
0.76

0.85

0.82

0.88

0.98

0.99

0.91

h=8

0.85

0.92

0.91

0.93

1.02

0.97

0.93

h=4

0.72

0.81

0.76

0.83

0.94

0.97

0.82

h=8

Illinois

0.84

0.88

0.86

0.89

0.93

0.87

h=4

0.63

0.69

0.68

0.69

0.78

0.89

0.75

h=8

Indiana

0.76

0.81

0.84

0.85

0.98

0.90

0.86

h=4

0.52

0.61

0.69

0.71

0.84

0.91

0.76

h=8

Kentucky

0.80

0.86

0.90

0.97

0.89

h=4

0.70

0.84

0.76

0.88

1.02

1.03

0.95

h=8

Missouri

0.86

0.92

0.89

0.93

1.01

1.00

0.97

h=4

0.74

0.81

0.80

0.84

0.89

0.97

0.88

h=8

Mississippi

NOTE: Entries report the ratio of the MSFE for the combining method indicated in the first column to the MSFE for the AR benchmark model.

0.90
0.90

0.92

Trimmed mean

DMSFE, θ = 1.0

C共2,PB兲

1.03

Median

DMSFE, θ = 0.9

0.96
0.99

Mean

h=4

Combining
method

Arkansas

Combination Forecast Results: 1995:Q1–2006:Q4 Out-of-Sample Forecast Evaluation Period

Table 2

1.00

0.98

0.99

1.18

1.00

0.98

h=4

0.79

0.86

0.93

0.95

0.99

0.92

h=8

Tennessee

Rapach and Strauss

F E D E R A L R E S E R V E B A N K O F S T. LO U I S R E G I O N A L E C O N O M I C D E V E LO P M E N T

Rapach and Strauss

quarter horizons, respectively. For Illinois, the state
population, regional vacancy rate, and inflation
rate display the best performance, with reductions
in MSFE relative to the AR benchmark of up to 18
percent and 43 percent at the reported horizons.
The state housing price-to-income ratio, state population, state unemployment rate, and unemployment claims produce large reductions in MSFE
for Indiana, with reductions up to 31 percent and
53 percent at the two horizons. For Kentucky, six
predictors are able to generate sizable reductions
in MSFE relative to the AR benchmark at both
horizons: the state housing price-to-income ratio,
state population, state unemployment rate, regional
vacancy rate, inflation rate, and consumer confidence. The largest reductions in MSFE are 33 percent and 74 percent at the four- and eight-quarter
horizons, respectively. Real M2, consumer confidence, and the inflation rate lead to sizable reductions in MSFE relative to the AR benchmark for
Missouri, with reductions of up to 47 percent and
68 percent at the two horizons. Three variables
stand out for Mississippi: the state housing priceto-income ratio, consumer confidence, and inflation rate. The state housing price-to-income ratio
leads to the largest reductions in MSFE (42 percent
and 65 percent) at the two reported horizons. For
Tennessee, the state housing price-to-income ratio
and inflation rate lead to the largest reductions in
MSFE relative to the AR benchmark at both of the
reported horizons (up to 24 percent and 48 percent).

Combining Method Forecasting Results
Table 2 reports the combination forecast results
in the form of the ratio of the combining method
MSFE to the AR benchmark MSFE, so that (as in
Table 1) a ratio below unity indicates that the combining method forecast is more accurate than the
AR benchmark forecast in terms of MSFE. The
results in Table 2 show that the simple combining
methods often produce reductions in MSFE relative
to the AR benchmark of around 10 percent, and
this is in line with the findings of Stock and Watson
(1999 and 2003) in the context of U.S. GDP growth
and inflation forecasts. The DMSFE combining
method forecasts appear to perform somewhat
better than the simple combining method forecasts
in most cases, with the DMSFE combining method

based on θ = 0.9 leading to reductions in MSFE
relative to the AR benchmark of approximately
10 to 15 percent at the four-quarter horizon and
approximately 20 to 30 percent at the eight-quarter
horizon in most cases. The cluster combining
methods exhibit the best overall performance,
especially the C共3,PB兲 method. With one exception
(Tennessee at h = 4), the MSFE ratios are all well
below unity for the C共3,PB兲 method, with reductions in MSFE of up to 24 percent and 48 percent
relative to the benchmark AR model at horizons
of four and eight quarters, respectively (both for
Kentucky). The C共3,PB兲 cluster combining method
leads to average reductions in MSFE relative to
the AR benchmark model across the seven states
of approximately 15 percent and 30 percent at
horizons of four and eight quarters, respectively.
Given that it will be difficult to identify a priori
the particular predictors that are most relevant for
a given out-of-sample period, the performance of
the combining methods—especially the C共3,PB兲
method—indicates that they provide a useful way
of producing relatively accurate forecasts of real
housing price growth in the Eighth District states
in the presence of many potentially relevant
predictors.

CONCLUSION
We examine the ability of a host of economic
variables to forecast real housing price growth for
the seven individual states in the Federal Reserve’s
Eighth District. A number of variables, such as the
state housing price-to-income ratio, state unemployment rate, consumer confidence, and inflation
rate, produce forecasts that often substantially
outperform a benchmark AR model in terms of
MSFE in individual Eighth District states, but no
single variable is able to improve on the AR benchmark for all states at all reported horizons. Given
that it will be difficult to identify a priori the particular variable or small set of variables that are
best suited for forecasting real housing price growth
for a given state and time period, we also analyze
the performance of forecast combining methods.
We find that combining methods generally offer
useful means of incorporating and culling information from a large number of potential predictors

F E D E R A L R E S E R V E B A N K O F S T. LO U I S R E G I O N A L E C O N O M I C D E V E LO P M E N T

V O LU M E 3 , N U M B E R 2

2007

Rapach and Strauss

when forecasting real housing price growth in the
Eighth District states.
Finally, we briefly discuss two ways that we
are extending the research presented in this paper.
First, we are currently applying the approaches
employed in the present paper to a greater number
of individual U.S. states, including larger U.S.
states (in terms of population) that have experienced substantial increases in real housing prices
over the past decade—states for which there are
serious concerns of a housing price “bubble.” We
are also preparing to apply the approach used in
the present paper to forecasting real housing price
growth for individual metropolitan areas in the
Eighth District, as households are often interested
in forecasts of housing price growth in their more
immediate vicinity.

REFERENCES
Aiolfi, Marco and Timmermann, Allan. “Persistence
in Forecasting Performance and Conditional
Combination Strategies.” Journal of Econometrics,
November-December 2006, 135(1-2), pp. 31-53.
Abraham, Jesse M. and Hendershott, Patric H.
“Bubbles in Metropolitan Housing Markets.” Journal
of Housing Research, 1996, 7(2), pp. 191-207.
Cho, Man. “House Price Dynamics: A Survey of
Theoretical and Empirical Issues.” Journal of
Housing Research, 1996, 7(2), pp. 145-72.
Greenspan, Alan and Kennedy, James. “Estimates of
Home Mortgage Originations, Repayments, and Debt
on One-to-Four Family Residences.” Finance and
Economic Discussion Series Paper 2005-41, Federal
Reserve Board, September 2005.

V O LU M E 3 , N U M B E R 2

2007

Hendershott, Patric H. and Weicher, John C.
“Forecasting Housing Markets: Lessons Learned.”
Real Estate Economics, Spring 2002, 30(1), pp. 1-11.
Johnes, Geraint and Hyclak, Thomas. “House Prices
and Regional Labor Markets.” Annals of Regional
Science, February 1999, 33(1), pp. 33-49.
Rapach, David E. and Strauss, Jack K. “Forecasting
Employment Growth in Missouri with Many
Potentially Relevant Predictors: An Analysis of
Forecast Combining Methods.” Federal Reserve
Bank of St. Louis Regional Economic Development,
2005, 1(1), pp. 97-112.
Rapach, David E. and Strauss, Jack K. “Forecasting
U.S. Employment Growth Using Forecast Combining
Methods.” Journal of Forecasting, 2007 (forthcoming).
Stock, James H. and Watson, Mark W. “Forecasting
Inflation.” Journal of Monetary Economics, October
1999, 44(2), pp. 293-335.
Stock, James. H. and Watson, Mark W. “Forecasting
Output Growth and Inflation: The Role of Asset
Prices.” Journal of Economic Literature, September
2003, 41(3), pp. 788-829.
Stock, James H. and Watson, Mark W. “Combination
Forecasts of Output Growth in a Seven-Country Data
Set.” Journal of Forecasting, September 2004, 23(6),
pp. 405-30.
Timmermann, Allan. “Forecast Combinations,” in
Graham Elliott, Clive W.J. Granger, and Allan
Timmermann, eds., Handbook of Economic
Forecasting. Amsterdam: Elsevier, 2006, pp. 135-96.

F E D E R A L R E S E R V E B A N K O F S T. LO U I S R E G I O N A L E C O N O M I C D E V E LO P M E N T

Educational Attainment and Recovery from
Recessions Across Metropolitan Areas
Bryan Bezold
Metropolitan area business cycles vary considerably in both magnitude and duration. Some metro
areas recover rapidly from downtowns, some take longer, and some never recover. Because recent
recessions have involved employment changes at the North American Industrial Classification
System (NAICS) supersector level that were both cyclical and structural, part of a metro area’s
recovery from a recession may include the process of workers adapting to new jobs in other
industries. If worker adaptation is part of the recovery process, then metro areas with higher educational levels might be able to recover more quickly from recessions. This hypothesis is tested
with multiple regression models of the 1990 and 2001 recessions. There is a significant and negative link between college attainment and the time it took for metropolitan areas to recover from
the 2001 recession, but not from the 1990 recession. (JEL A10, J24)
Federal Reserve Bank of St. Louis Regional Economic Development, 2007, 3(2), pp. 43-52.

etropolitan areas experience economic contractions in much the
same way that the broader U.S.
economy does. However, because
economies of metropolitan statistical areas (MSAs)
differ in their demographics and other economic
characteristics, they are natural subjects to use to
test for relationships between structural economic
characteristics and economic performance during
business cycles. In this paper I use a panel of the
50 largest MSAs in the United States to test for
a relationship between the share of each MSA’s
population over 25 years of age with a college
degree and the time in months it takes for that
MSA to recover from a recession. In each case,
employment is used to define the beginning and
end of a recession and the time interval for recovery. The relationship between college attainment
and recovery length is being tested because, as
the structure of the U.S. economy changes, the
process of recovery from recession may depend
on the ability of workers to adapt to new jobs at
different firms or in different industries. If more-

educated workers are more adaptable, and moreadaptable workers are better at learning new skills
at new jobs, then there may be a relationship
between the level of education in an MSA and
the time it takes for that MSA to recover from a
recession.
During the 2001 recession, as with most recessions, the United States and most major MSAs
experienced an overall decline in nonfarm employment, a recovery to the previous peak employment
level, and then continued expansion. Nonfarm
employment declined from a peak of 132.5 million
in February 2001 to a trough of 129.8 million in
August 2003. Employment expanded at an average
monthly rate of 151,000 between September 2003
and February 2005, when the United States first
exceeded its prior peak level of employment.
Between March 2005 and January 2007, average
monthly job gains accelerated to slightly over
200,000 per month.
At the same time, however, some industries
(as defined by North American Industrial

Bryan Bezold is chief economist at the St. Louis Regional Chamber and Growth Association.

F E D E R A L R E S E R V E B A N K O F S T. LO U I S R E G I O N A L E C O N O M I C D E V E LO P M E N T

V O LU M E 3 , N U M B E R 2

2007

Bezold

Classification System [NAICS] supersectors)
declined and did not recover. U.S. manufacturing
employment, for example, declined from slightly
more than 17 million in March 2001 to 14 million
in January 2007. Employment in the information
sector also appeared to have permanently declined
from a pre-recession peak of 3.7 million to 3.0
million in January 2007. This phenomenon was
also observable in most MSAs. Approximately
3.7 million workers, or 2.8 percent of total peak
employment, appear to have been permanently
dislocated from their previous industries.
Compare this with the 1990 recession. Overall,
the U.S. experienced an employment peak, trough,
and recovery. Nonfarm payroll employment peaked
at 109.8 million in June 1990, fell to a trough of
180.2 million in May 1991, and recovered to the
previous peak level in February 1993. As with the
2001 recession, manufacturing employment began
to decline before overall nonfarm employment
began to decline. Manufacturing employment
declined by 7.3 percent, or 1.3 million jobs, from
a peak of 18.0 million in March 1989 to a trough
of 16.7 million in August 1993. Manufacturingsector employment did not fully recover to its
pre-1990 recession peak, but did rise to roughly
98 percent of its peak before beginning to decline
again in March 1998.
The structural adjustment that the United
States and most major MSAs experienced during
the 1990 recession and recovery process was thus
smaller than in the 2001 recession. The total number of manufacturing workers permanently displaced—the difference between manufacturing
employment at its pre-1990 recession peak and its
relative peak in March 1998—was 423,000. That
was much smaller than the 3 million workers who
were permanently displaced from the manufacturing sector after the 2001 recession.
It is thus the case that the U.S. economy experienced a full recovery from the 2001 recession in
terms of employment, but some industries, specifically information and manufacturing, did not.
The 1990 recession was also characterized by this
phenomenon, but to a much smaller extent, in the
manufacturing sector. Because overall nonfarm
employment in both cases recovered to and grew
past its previous peak, workers in those perma44

V O LU M E 3 , N U M B E R 2

2007

nently affected sectors either went to work in new
sectors or left the labor force and were replaced by
new entrants. This phenomenon likely took place
at the firm level as well, with workers dislocated
by the closing of firms finding work at different
firms either within or outside of that industry.
That would suggest that economies, whether
at the national or subnational level, recover from
downturns in employment because workers are
able to adapt whatever skills they may have to new
firms or new industries. Once an economy makes
this adjustment, its firms become more productive
and it resumes a pre-recession pace of expansion.
Note that in the cases of both the 1990 and 2001
recessions, average monthly nonfarm employment
gains accelerated after the level of employment
reached its pre-recession peak. The implication
here is that the business cycle has three phases:
contraction, during which many firms either go
out of business or shed workers; recovery, when
workers adjusted their skills to new firms or even
industries; and expansion, when the economy has
reorganized and can grow at a slightly higher rate.
If the sorting of workers from old firms to new
ones and old industries to new ones is a factor in
the recovery of an economy to its prior peak
employment level, then it is logical to wonder if
the adaptability of workers, their ability to learn
new skills so that they can work in different firms
or industries, would affect the time it takes for an
economy to recover from a contraction.
At this point it is important to note that the
National Bureau of Economic Research (NBER), the
official chronicler of business cycles in the United
States, characterizes recession only in terms of
contraction and expansion. There is no distinction,
in the NBER view, between recovery and expansion;
any period of growth, prior to or after exceeding a
pre-recession peak, is considered expansion. This
paper’s characterization of the business cycle as
one with three phases, contraction, recovery, and
expansion, represents a view different from the
one commonly used by economists.
Many researchers, notably Owyang, Piger, and
Wall (2005), have described business cycles at the
state level with a two-phase model, expansion
and contraction, rather than a three-phase model.
Owyang et al. (2007) also have observed that a dif-

F E D E R A L R E S E R V E B A N K O F S T. LO U I S R E G I O N A L E C O N O M I C D E V E LO P M E N T

Bezold

ferent set of variables is associated with economic
performance during the expansion and contraction
phases and that human capital measured by college
attainment is correlated with MSA and state performance during growth phases.
Other research is broadly supportive of a correlation between human capital, measured by educational attainment, and measures of MSA growth.
Glaeser and Saiz (2003) found a connection between
growth and skills measured by college attainment
for MSAs that were declining, but not for MSAs
that were growing: Having a skilled workforce was
more important to MSAs that were not experiencing in-migration than to those that did have higher
rates of in-migration. Further, Glaeser and Saiz
(2003) also concluded that “rustbelt” MSAs with
higher skill levels would be more likely to adjust
to structural economic changes due to declining
manufacturing employment than MSAs with lower
skill levels.
Silvia (2006) observed that the transition of
the U.S. economy from a purely domestic labor
market to one influenced by both national and
international factors would have a number of
results, including reduced demand for manufacturing workers but increased demand for collegeeducated workers. Silvia (2006) also noted that
skill accumulation for workers will be a lifelong
process in the face of structural economic change.
Because other research confirms correlation
between college attainment and economic growth,
and also suggests a link between skilled workers
and the ability of MSAs to adapt, then one would
expect that MSAs with higher skill levels should
recover more quickly from recessions that involve
structural and not just cyclical economic changes.
This paper tests the hypothesis that increased
worker adaptability, measured by college attainment levels, is correlated with faster recoveries
from recessions. After a contraction in employment,
workers that are permanently displaced from their
firms or industries will need to be able to adapt to
new firms or industries; also, prospective employers must believe that the workers will be able to
adapt to new environments for overall MSA
employment to recover from a contraction. Bettereducated workers should be more able to adjust to
new jobs in new industries, so MSAs with a greater

share of educated workers should need less time
to re-sort workers into jobs in new firms and
industries.

METHODOLOGY
The models in this paper use multiple regression analysis to test for a relationship between (i)
the length of time it takes for an MSA economy to
recover from its employment trough following a
recession and reach a level at or above the prerecession peak and (ii) the level of college attainment for both the 2001 and 1990 recessions. College
attainment from the 1990 and 2000 U.S. Census
surveys were used as a proxy for the general level of
workers’ adaptability. The length of time in months
for an MSA to recover from a post-recession
employment trough and reach a level at or above
the pre-recession peak was used as the dependent
variable. This number was calculated through the
inspection of seasonally adjusted nonfarm employment data for each MSA. Once the initial peak,
trough, and next peak were identified, then the
length of recovery was the number of months
between the trough and the next month that nonfarm employment rose permanently above the level
of the pre-recession peak. In many cases, the
employment data did not have a smooth curve,
but a volatile one. In these cases, the employment
level occasionally spiked to a level above the prerecession peak during the recovery period, but then
dropped back below it. In those instances, the
recovery was not considered complete until seasonally adjusted employment permanently exceeded
the pre-recession peak. This is one of several
sources of uncertainty in the model.
The independent variables for each MSA in
the model were the share of adults over the age of
25 with a college degree, the share of employment
in manufacturing and information at that prerecession peak, and the magnitude (percentage)
of each employment decline.
The share of adults with a college degree was
the variable being tested; the other independent
variables were included as controls. The magnitude
of the decline in employment was included because
the greater the number of dislocated workers, the

F E D E R A L R E S E R V E B A N K O F S T. LO U I S R E G I O N A L E C O N O M I C D E V E LO P M E N T

V O LU M E 3 , N U M B E R 2

2007

Bezold

Table 1
Descriptive Statistics for the Sample of 34 MSAs During the 2001 Recession
Variable

Median

Range

Standard deviation

Percent decline in employment, peak to trough

–0.025

0.053

0.014

0.101

0.135

0.035

Share of employment in manufacturing at pre-recession peak
Share of employment in information at pre-recession peak

0.028

0.044

0.012

Share of adults over 25 with a BA or BS

0.264

0.229

0.047

Recovery length, trough to prior peak level, in months

23.5
Minimum

Date of pre-recession employment peak

March 2000

Date of employment trough
Date of recovery to pre-recession peak employment level

35
Mode

10.2
Maximum

December 2000 September 2001

October 2001

March 2003

June 2003

August 2002

April 2005

February 2007

Table 2
Descriptive Statistics for the Sample of 44 MSAs During the 1990 Recession
Variable

Median

Range

Standard deviation

Percent decline in employment, peak to trough

–0.021

0.139

0.027

Share of employment in manufacturing at pre-recession peak

0.135

0.277

0.052

Share of adults over 25 with a BA or BS

0.209

0.263

0.045

Recovery length, trough to prior peak level, in months

Date of pre-recession employment peak
Date of employment trough
Date of recovery to pre-recession peak employment level

longer it would take to completely sort all of the
dislocated workers into new jobs. Because the
2001 recession was concentrated in the manufacturing and information NAICS supersectors, the
share of an MSA’s employment in those two sectors
would likely affect the length of time it took to
recover from a recession; more workers in affected
sectors would presumably increase the number of
workers who have to be sorted into other firms or
industries. The inclusion of these two variables
on the right-hand side of the equation reflected an
assumption that the 2001 recession was an exogenous shock to the manufacturing and information
sectors in each MSA.
46

V O LU M E 3 , N U M B E R 2

2007

16.5

17.5

Minimum

Mode

Maximum

January 1990

August 1990

September 1991

December 1990

April 1991

January 1996

May 1991

January 1992

July 2000

In the case of the 1990 recession, the share of
employment in the information sector was not
included because the information sector fully
recovered from the 1990 recession and therefore
did not produce permanently displaced workers;
it was assumed that the decline in manufacturing
was the exogenous shock to MSA economies.
Table 1 contains descriptive statistics for the model
of the 2001 recession variables; Table 2 contains
descriptive statistics for the model of the 1990
recession.
More simply put, the control variables are
selected to see whether two MSAs that had different
college attainment levels but the same magnitude

F E D E R A L R E S E R V E B A N K O F S T. LO U I S R E G I O N A L E C O N O M I C D E V E LO P M E N T

Bezold

of contraction, and an equal share of employment
in sectors most subject to the exogenous shock,
would experience recoveries of different lengths.
The analysis initially began with the 50 largest
MSAs in the United States. The eventual model of
the 2001 recession, however, is based on just 34
of them: 12 of the 50 largest MSA had not fully
recovered from the 2001 recession by January of
2007 and 4 MSAs did not experience any employment contraction during the 2001 recession. Those
16 were excluded from the model.
For the 1990 recession, there were similar limitations. The Los Angeles MSA never recovered
from the 1990 recession, and five other MSAs did
not experience a significant decline in overall
nonfarm employment. A total of six MSAs were
thus excluded from the 1990 recession model.
Of the 12 MSAs that had not yet recovered
from the 2001 recession, seven had college attainment above the U.S. average (24.4 percent) and five
had college attainment below the U.S. average.
Similarly, of the four MSAs that did not experience
an employment contraction, two had college attainment above the U.S. average and two below. The
MSAs that experienced neither recovery nor recession were distributed across the range of college
attainment, which raises an initial concern about
the proposed link between college attainment and
recovery lengths. In the case of the 1990 recession,
the only MSA not to fully recover was Los Angeles,
which had college attainment below the U.S.
average in 1990. Three MSAs with above-average
attainment, however—San Francisco, San Jose,
and San Diego—had the longest recovery times.
So, in both the 1990 and 2001 recession models,
there were some clear outliers that could not be
included in the model and that did not exhibit the
proposed relationship between college attainment
and faster recoveries.

RESULTS FOR THE MODELS OF
THE 2001 RECESSION
Results of the regression models for the 2001
recession are listed in Tables 3 through 5. Each of
the independent variables is significant at the 10
percent but not the 5 percent levels, and the model

explains approximately 30 percent of the variation
of the recovery length. The coefficient for college
attainment, –60.5, implies that if the magnitude
of the decline in employment and the shares of
employment in manufacturing and information
are held constant, a one-standard-deviation increase
in college attainment, which is 4.7 percentage
points, would decrease the length of a recovery by
2.8 months. That is roughly one-quarter of the
standard deviation of the recovery length.
Based on the results of the initial model, there
is an observed correlation between recovery length
and college attainment, but it is not of optimal significance. It is also not a large correlation and is
part of a model that leaves most of the variation in
recovery length unexplained.
These results leave two possibilities: The first
is that there is not a significant relationship between
college attainment and recovery time, either
because there isn’t really a correlation between
adaptability and recovery or because college attainment isn’t a good measure of adaptability. The
second is that such a relationship does exist and
would be better observed with some modifications
to the original model. A number of modifications
were thus attempted.
As mentioned here previously, the monthly
employment data used to identify peaks and
troughs for each MSA were volatile and the volatility contributed to the possibility that turning points
might be misidentified. To correct for this potential
error, the monthly data were compressed to quarterly frequency. The same analysis was repeated
with the same variables, transformed from months
to quarters. The results, however, did not change:
After that transformation, the coefficients for three
of the four independent variables—college attainment, share of employment in the information
sector, and share of employment in the manufacturing sector—were no longer significant at even
the 10 percent level. With quarterly rather than
monthly data, the difference between each MSA’s
contraction and recovery became smaller, whereas
the difference between each MSA’s control variables remained constant. The adjusted R 2 of the
model fell from 0.30 with the monthly data to 0.18
with the quarterly data.
Another obvious idea to further test the hypoth-

F E D E R A L R E S E R V E B A N K O F S T. LO U I S R E G I O N A L E C O N O M I C D E V E LO P M E N T

V O LU M E 3 , N U M B E R 2

2007

Bezold

Table 3
Results of Initial Regression Model of 2001 Recession (n = 34)
t-statistic

Coefficient

Variable
Constant

15.3

Percent decline in employment (peak to trough)
Share of employment in manufacturing at peak

1.5

–217.8

–1.8*

87.1

1.9*

Share of employment in information at peak

276.3

1.8*

College attainment

–60.5

–1.7*

Model statistics
F-statistic

4.6**

Adjusted R 2

0.30

NOTE: *p < 0.10, **p < 0.05.

Table 4
Results of Model of 2001 Recession with Dichotomized College Attainment Based on U.S. Average
(n = 34)
Variable
Constant

Coefficient

t-statistic

3.8

0.6

Share of employment in manufacturing at peak

126.2

Share of employment in information at peak

365.6

3.0*

–7.9

–2.5**

College attainment exceeds U.S. average (1 or 0)

2.9**

Model statistics
F-statistic
Adjusted R

6.1**
2

0.32

NOTE: *p < 0.10, **p < 0.05.

Table 5
Results of Model of 2001 Recession with Dichotomized College Attainment Based on Sample Median
(n = 34)
Variable
Constant

Coefficient

t-statistic

1.0

0.2

Share of employment in manufacturing at peak

120.2

Share of employment in information at peak

419.4

3.2**

–7.3

–2.3**

College attainment exceeds sample median (1 or 0)

2.8**

Model statistics
F-statistic
Adjusted R

5.6**
2

0.29

NOTE: *p < 0.10, **p < 0.05.

V O LU M E 3 , N U M B E R 2

2007

F E D E R A L R E S E R V E B A N K O F S T. LO U I S R E G I O N A L E C O N O M I C D E V E LO P M E N T

Bezold

esis is to expand the sample size. Five additional
MSAs were added to the sample: Raleigh-Cary,
North Carolina; Tucson, Arizona; Tulsa, Oklahoma;
Albany-Schenectady-Troy, New York; and Honolulu,
Hawaii. These MSAs were included because they
were the next-largest U.S. MSAs that experienced
both a contraction and recovery. When these five
MSAs were added to the model, however, the level
of college attainment was not significantly correlated with recovery length. As described here previously, the fact that this model correction did not
change the results could mean that the tested relationship between college attainment and recovery
length does not exist. It is also possible that the five
MSAs added to the sample were not large enough
for the reorganizing of workers to be a part of the
recovery process. The average peak employment
of the five MSAs added to the sample was 413,000,
compared with an average of 1.4 million for the
other sample of 34 MSAs. Thus, the five MSAs
added to the sample may not be large enough to
have the perceived job opportunities necessary to
convince dislocated workers to remain in the area
and look for a new job.
Another possibility is (i) that there is some
threshold of adaptability necessary for workers in
an MSA to reorganize and (ii) that being much
above that threshold is not particularly helpful to
the process of economic recovery. To test for that
result, the college attainment variable was dichotomized into a dummy variable that reflected college
attainment above the U.S. average. The results of
this model were slightly better than the original
model. With college attainment coded as a 1 if it
was above the 2000 U.S. average (24.4 percent),
the coefficient was –7.9, significant at the 5 percent
level. In this model, however, the variable reflecting the percent decline in peak-to-trough employment was no longer significant at the 10 percent
level and was removed from the model. Given that
fact, the results of this model suggest that, after
controlling for shares of employment in manufacturing and information, an MSA that had college
attainment in the year 2000 that exceeded the U.S.
average would recover to its pre-recession peak
employment level almost 8 months faster than an
MSA with below-average college attainment. Alter-

natively, having college attainment above the U.S.
average shortens recovery length in this sample
by about three-quarters of one standard deviation.
In our sample, 23 of the 34 MSAs (about twothirds) had college attainment that exceeded the
U.S. average. It is also generally true that MSAs
tend to have higher college attainment than rural
areas do, so comparing MSAs with the U.S. average
may not be appropriate. The model was run a final
time with the dummy recoded to reflect 1 if that
MSA had college attainment greater than the sample median. In the case of the 2001 recession model
with 34 MSAs, the variable for college attainment
above the sample median was significant at the 5
percent level, as were the manufacturing and information sector employment controls.

RESULTS FOR THE MODELS OF
THE 1990 RECESSION
For the 1990 recession, the same methodology
was applied, except that the share of employment
in the information sector was not used as a control
variable. The results of these regressions are in
Tables 6 through 8. For the initial model, the level
of college attainment was not significantly correlated with shorter recovery times. The model was
repeated with a dichotomized variable reflecting
college attainment above the U.S. average in 1990
or above the sample mean (both of which were
coded as 1). Neither of the dichotomized variables
was significantly correlated. The share of employment in manufacturing was also not significantly
correlated with recovery from the 1990 recession.

CONCLUSIONS
These results suggest that, after controlling for
the magnitude of peak-to-trough decline in employment in manufacturing and information, there is a
negative relationship between some measures of
college attainment and the length of time measured
in months that it takes for a large MSA to recover
from a post-recession employment trough to its
previous peak employment level for the 2001 recession. This is not the case for the 1990 recession. The
level of significance of the relationship between

F E D E R A L R E S E R V E B A N K O F S T. LO U I S R E G I O N A L E C O N O M I C D E V E LO P M E N T

V O LU M E 3 , N U M B E R 2

2007

Bezold

Table 6
Results of Initial Model of 1990 Recession (n = 44)
t-statistic

Coefficient

Variable
Constant

20.5

1.89*

Percent decline in employment (peak to trough)

–508.28

–6.98**

Share of employment in manufacturing at peak

–48.23

–1.29

College attainment

–34.32

–0.817

Model statistics
F-statistic
Adjusted R

17.21**
2

0.53

NOTE: *p < 0.10, **p < 0.05.

Table 7
Results of Model of 1990 Recession with Dichotomized College Attainment Based on U.S. Average
(n = 44)
Variable

t-statistic

Coefficient

Constant

8.82

Percent decline in employment (peak to trough)

–470.17

College attainment exceeds U.S. average (1 or 0)

–3.01

2.65**
–7.05**
–0.814

Model statistics
F-statistic
Adjusted R

25.85**
2

0.53

NOTE: *p < 0.10, **p < 0.05.

Table 8
Results of Model of 1990 Recession with Dichotomized College Attainment Based on Sample Mean
(n = 44)
Variable

t-statistic

Coefficient

Constant
Percent decline in employment (peak to trough)
College attainment exceeds sample median (1 or 0)

8.26

2.56**

–467.22

–6.99**

–2.10

–0.57

Model statistics
F-statistic
Adjusted R

24.49**
2

0.52

NOTE: *p < 0.10, **p < 0.05.

V O LU M E 3 , N U M B E R 2

2007

F E D E R A L R E S E R V E B A N K O F S T. LO U I S R E G I O N A L E C O N O M I C D E V E LO P M E N T

Bezold

the level of college attainment and recovery length
is not optimal, but the relationship reaches the 5
percent level of significance when college attainment is dichotomized into a dummy variable
reflecting above-average college attainment. The
model results do not change when smaller MSAs
are added to the sample or when the data are
compressed from monthly frequency to quarterly
frequency.
There are a number of reasons to view these
results with some skepticism. The first is that the
relationship is significant when observing the
2001 recession, but not the 1990 recession. There
are a few things worth noting about the model for
the 1990 recession in spite of the fact that there is
not a significant relationship between recovery
times and college attainments. One is that the sizes
of the coefficients for the college attainment variables are all smaller in the 1990 model than in the
2001 model. So, even if the relationship between
college attainment and recovery time was significant in the model of the 1990 recession, it would
still be weaker than that observed in the model of
the 2001 recession. Another interesting characteristic of the 1990 model, again setting aside the
question of significance, is that even with a smaller
number of right-hand-side variables, the R 2 is much
higher for the models of the 1990 recession that
the models for the 2001 recession. In fact, a simple
equation in which recovery length after the 1990
recession is associated with only a constant term
and an MSA’s percent decline in employment
explains more than half of the variation in recovery time. These differences suggest that the 1990
and 2001 recessions were distinct from one another.
Another reason for skepticism is that that 12
of the 50 largest MSAs experienced a contraction
but have not yet recovered from the 2001 recession; and some of these, most notably Boston, San
Francisco, and San Diego, have relatively high
college attainment. It is also true that some areas
with high college attainment—such as San Diego,
San Francisco, and San Jose—were among the
slowest to recover from the 1990 recession. On the
other hand, some areas with low college attainment—such as Detroit and Cleveland—still had
not recovered from the 2001 recession by the beginning of 2007. The inclusion of those MSAs might

have improved the relationship between college
attainment and recovery length.
It is also true that the independent variable in
this model, college attainment, may not be the best
proxy for worker adaptability. The fact that adding
more observations to the sample did not improve
the results is another source of uncertainty. Further
research, such as an examination of these same
MSAs during previous contractions and recoveries,
or perhaps with a more refined measure of worker
adaptability than simple college attainment, is
thus necessary.
The fact that the relationship between college
attainment and recovery length was significant for
the 2001 recession but not for the 1990 recession
could alternatively be seen to support, rather than
contradict, the proposed relationship between
college attainment as a measure of worker adaptability and faster recovery from recessions. The
period of weak employment growth following the
1990 recession has commonly been referred to as
a “jobless recovery” that differed from past recessions. The recovery from the 2001 recession was
similar in that there was also a weak employment
recovery, but different in that the apparently permanent decline in manufacturing suggests that a
structural, and not just cyclical, change took place
in the U.S. economy after the 2001 recession. After
the 2001 recession, a larger number of workers
needed to adapt themselves to new industries and
firms, so it was more important for MSAs to have
more adaptable and skilled workers to recover from
the 2001 recession as compared with the 1990
recession. The fact that the relationship between
educational attainment and recovery was stronger
for the 2001 recession than for the 1990 recession
provides some empirical support for the proposition that the 1990 and 2001 recessions were different from one another.
This finding is also consistent with previous
research that has established a correlation between
skills and growth and the ability of MSAs to adjust
to structural change (Glaeser and Saiz, 2003). That,
in turn, suggests that education and adaptability
are more important to economic recovery when
recessions have a larger component of structural
change as opposed to a larger component of cyclical
change. If observers such as Silvia (2006) are correct,

F E D E R A L R E S E R V E B A N K O F S T. LO U I S R E G I O N A L E C O N O M I C D E V E LO P M E N T

V O LU M E 3 , N U M B E R 2

2007

Bezold

and globalization will drive both continued structural changes in the U.S. economy and increased
demand for educated workers, then future economic recoveries may be more influenced by educational attainment than those in the past.

REFERENCES
Glaeser, Edward L. and Saiz, Albert. “The Rise of the
Skilled City.” Discussion Paper No. 2025, Harvard
Institute of Economic Research, 2003.
Owyang, Michael T.; Piger, Jeremy M. and Wall,
Howard J. “The 2001 Recession and the States of the
Eighth Federal Reserve District.” Federal Reserve
Bank of St. Louis Regional Economic Development,
2005, 1(1), pp. 3-16.
Owyang, Michael T.; Piger, Jeremy M.; Wheeler,
Christopher H. and Wall, Howard J. “The Economic
Performance of Cities: A Markov-Switching
Approach.” Working Paper No. 2006-056C, Federal
Reserve Bank of St. Louis, 2007.
Silvia, John E. “Domestic Implications of a Global
Labor Market.” Business Economics, July 2006, 41(3),
pp. 23-29.

V O LU M E 3 , N U M B E R 2

2007

F E D E R A L R E S E R V E B A N K O F S T. LO U I S R E G I O N A L E C O N O M I C D E V E LO P M E N T

Transferable Tax Credits in Missouri:
An Analytical Review
Paul Rothstein and Nathan Wineinger
In 2005, Missouri had 53 legally authorized tax credit programs. In this paper, the authors assemble
basic information on all of these programs and further analyze the six largest (by tax credits issued)
that include freely transferable credits. Their analysis focuses on the institutional features of these
programs, the kinds of market failures or disparities they may address, and whether the design of
each program is consistent with its economic rationale. The authors also consider whether the
evaluation of each program by the state is consistent with its economic rationale. They conclude
with a brief discussion of the transactions prices for the credits on which they have data and
whether making the tax credits refundable as well as transferable could reduce the transactions
costs associated with these programs. (JEL D61, H71, R58)
Federal Reserve Bank of St. Louis Regional Economic Development, 2007, 3(2), pp. 53-74.

n 2005, the state of Missouri had 53 legally
authorized tax credit programs. At least 32
(and at most 38) actually issued credits in
that year. The value of the credits issued was
at least $357 million dollars. This is a large number of programs, and the amount of tax revenue
forgone is significant. The revenue is only a modest share of total state general revenue, however,
about 5 percent.
An interesting feature of many Missouri tax
credits is that they are to some degree transferable.
Transferability allows an entity that has more tax
credits than tax liability to sell what he or she
cannot use. It therefore makes the credit useful to
entities that have little or no state tax liability, so
tax policy becomes a closer substitute for expenditure policy. Transferability also allows entities
that receive multiyear credits for capital projects
to sell the credits and obtain the financing they

need. On the other hand, a $1 tax credit does not
sell for $1, but the credit costs the taxpayers of
Missouri that amount when it is redeemed. Transferability therefore adds an extra dimension of
costs and benefits to a tax credit.
Of the 53 tax credit programs mentioned above,
30 have credits that were officially designated as
transferable (57 percent); and of the 32 programs we
know actually issued credits in 2005, 17 issued
transferable credits (53 percent). Transferability
does not necessarily mean that one can sell the
credits on e-Bay to the highest bidder, however. The
state classifies tax credits as transferable or not, but
the degree of transferability can be determined only
by reading the authorizing legislation. Because it
would not be practical to examine all of these programs, we focus mostly on the six largest programs
(by value of credits issued in 2005) that issued what
we consider to be “freely” transferable credits.

Paul Rothstein is an associate professor of economics in the department of economics and associate director of the Murray Weidenbaum Center
on the Economy, Government and Public Policy, at Washington University in St. Louis. Nathan Wineinger is a researcher at the Murray
Weidenbaum Center on the Economy, Government and Public Policy, and an undergraduate at Washington University in St. Louis. The authors
thank the William T. Kemper Foundation and the Weidenbaum Center for financial support, as well as the many people who gave generously of
their time and answered their questions, especially Matt Nordmann (St. Louis Equity Fund, Inc.), Brian Schmidt (Joint Committee on Tax Policy),
Jim Brentlinger (Department of Revenue), Brenda Horstman and Mike Downing (Department of Economic Development), Steve Stogel (DSC Group),
and Marty Romitti and Alan Spell (Missouri Economic Research and Information Center).

F E D E R A L R E S E R V E B A N K O F S T. LO U I S R E G I O N A L E C O N O M I C D E V E LO P M E N T

V O LU M E 3 , N U M B E R 2

2007

Rothstein and Wineinger

The central question in which we are interested
is whether each program is good public policy. We
take this question to mean, in the broadest possible
terms, is there something inefficient or unfair about
the outcomes that would result if the program did
not exist? We provide a general discussion of this
question for each of the six largest tax credit programs (by number of credits issued) that have freely
transferable credits. In doing this, we draw on
current research in local public-sector economics
and basic economic principles. Given the difficulty
of the question and the limited amount of information available about these programs, however,
we decided not to overly simplify matters just to
provide “yes” or “no” answers.
We begin our analysis by identifying and
describing the six largest programs. We then examine the way these programs are evaluated by the
Missouri Department of Economic Development
(DED), which has the formal responsibility for
reviewing them. A necessary condition for good
public policy is strong and regular administrative
review. It is therefore appropriate to consider how
these reviews help analyze the policies themselves.
We also consider three recent reviews of the programs that were not done by DED.
We then turn to our basic question. If there is
something inefficient about the outcomes that
would occur or if growth would be slower if a program did not exist, then we should be able to identify a market failure that the program addresses.
If there is something unfair about the outcomes
that would occur if a program did not exist, then
we should be able to identify the disparities or
inequities that the program addresses. Thus, our
basic question leads us to ask whether the program
in question addresses some kind of market failure
or corrects disparities or inequities. Furthermore,
the answers to these questions provide some insight
into how the program should be structured, how
its key parameters should be chosen, and what
variables should be measured to evaluate it. One
can quantify the impact of a program in any number
of ways (jobs created, services delivered, output
produced, etc.). One cannot draw meaningful conclusions from these numbers unless the rationale
for the program is clear.
54

V O LU M E 3 , N U M B E R 2

2007

In our conversations with various officials, we
found general dissatisfaction with the fact that DED
evaluates every program as if its purpose were
economic development. This is DED’s job, however, so the criticism is not really of DED but its
mandate. In doing what is expected of them, they
ignore the distinctions between correcting market
failures, reducing disparities, and promoting economic development. These are different goals, they
imply different structural features for the programs,
and they require the measurement of somewhat
different variables for proper program evaluation.
In many cases, DED’s approach forces a stark mismatch between their analysis and the analysis
suggested by economic theory.
An additional matter for concern is simply
how little analysis is actually being done. DED
tends to produce lists of raw information about
the impact of each program. In particular, they
make no attempt to produce a single, bottom-line
number of all the tangible benefits and costs for
each program. By tangible program benefits and
costs, we mean simply those impacts that can be
plausibly converted into dollars and cents. Failing
to quantify these impacts (to the greatest extent
possible) is an enormous problem. There are always
important intangible benefits and costs, as well—
the highly subjective aspects of programs that preserve state history, reduce poverty, and promote
opportunity, and which may also interfere with
property rights or reduce incentives to work. Concrete information about tangible benefits and costs
is essential for having a rational debate over these
intangible benefits and costs.
The next sections provide preliminary information about the programs, discuss the program
evaluations done by DED and others, present our
analysis of the programs, and then consider briefly
whether making the tax credits refundable in
addition to transferable would improve their cost
effectiveness.

AN OVERVIEW OF MISSOURI TAX
CREDIT PROGRAMS
We begin our analysis of Missouri transferable
tax credit programs by placing them in the larger

F E D E R A L R E S E R V E B A N K O F S T. LO U I S R E G I O N A L E C O N O M I C D E V E LO P M E N T

Rothstein and Wineinger

context of all Missouri tax credit programs and
the size of Missouri’s government. First, tax credit
programs are a popular policy tool in Missouri and
most tax credits are in fact transferable. The credits
cause the state to forgo a significant amount of
revenue in absolute terms. They cause just a modest
loss as a share of all general revenue, though.
Lists of the tax credit programs in any year are
available from two sources. The better-known is
form MO-TC, which is issued by the Missouri
Department of Revenue (DOR). This form lists the
tax credits that DOR has some role in administering
(44 in all).
A second list of tax credits appears as part of
the budget instructions issued to Missouri’s governmental departments by the Office of Administration.
This list includes tax credits that DOR has no role
in administering, but it also includes other kinds
of tax preferences.1
By using these lists and making inquiries at
the relevant government agencies, we developed a
spreadsheet with 53 tax credit programs for 2005.
(See the appendix.) We have complete information
for almost every program, but there are a few key
omissions.
We regard the value of tax credits issued under
a particular program in a given year as a good measure of the “importance” of the program in that year.
This number reflects economic demand for the
credits, which is the quantity (in dollars) that
people want to acquire on existing terms.2 We also
say that a program is “active” in a given year if it
issued any credits at all.
A brief glance at the appendix shows that, in
2005, many programs were inactive.3 Thus, the
total number of programs is not really a good indicator of the aggregate importance of tax credits in
Missouri. A more careful review of the data gives
the following basic results:
1

The list is available at www.oa.mo.gov/bp/budget.htm
(“Attachment 8”).

Strictly speaking, this is true only if the amount actually issued is
strictly less than the amount that could have been issued in that
year. This was generally, but not always, the case.

Zeros in the spreadsheet are the actual values. Missing data is indicated by “N/A.”

• Between 32 and 38 tax credit programs were
active in 2005. We cannot state the exact
number because of the lack of data for six
programs.
• The value of credits issued by the 32 active
programs for which we have data was about
$357 million dollars.
• Of these 32 programs, 17 issued credits
designated as “transferable.”4
• The value of the credits issued by these 17
programs was about $266 million dollars.
This is 74 percent of the total value of the
credits issued by the 32 active programs for
which we have data.5
On the one hand, $357 million is a significant
sum. On the other, in 2005, the Missouri state government reported about $6.933 billion in general
revenue.6 If the revenues forgone from the 32 programs could have been collected without changing
any other figures, general revenue would have been
about 5 percent higher. Gathering this revenue
would not have increased Missouri’s ranking of
42nd among U.S. states in terms of tax revenue
per capita.7

MISSOURI TRANSFERABLE TAX
CREDIT PROGRAMS
In this section, we explain our selection of the
six transferable tax credit programs, which we
examine in detail. We find that all of these programs
are primarily administered by the DED and appear
on form MO-TC.
4

This does not mean the same thing in all cases, however. Two
important programs, the Low Income Housing Tax Credit Program
and the Certified Capital Companies Program, issue transferable
credits that are not “freely” transferable. We will discuss this further
later in the paper.

If we exclude credits issued under the Low Income Housing Tax
Credits Program and the Certified Capital Companies Program, we
still find that 47 percent of all credits issued are transferable.

This is revenue apart from the earnings of utilities and certain other
operations.

The state ranking is for 2004, the most recent year for which the
data is available; see The Tax Foundation, www.taxfoundation.org/.
These figures do not include credits issued by the six programs for
which we do not have data.

F E D E R A L R E S E R V E B A N K O F S T. LO U I S R E G I O N A L E C O N O M I C D E V E LO P M E N T

V O LU M E 3 , N U M B E R 2

2007

Rothstein and Wineinger

Table 1
Tax Credit Programs Issuing Over $5 Million in Credits (FY 2005)
Tax credit program

Transferable

Amount issued

Missouri Low Income Housing Tax Credit

Yes

$83,477,412

Historic Preservation Tax Credit Program

Yes

80,213,374

Enterprise Zone Tax Benefit Program

39,066,023

Infrastructure Tax Credit Program

Yes

28,964,274

Brownfield Redevelopment Program (remediation)

Yes

14,808,297

Certified Capital Companies Program (CAPCO)

Yes

14,000,000

Neighborhood Assistance Tax Credit Program

11,263,385

Missouri Health Insurance Pool

10,015,203

Affordable Housing Assistance Tax Credit Program

Yes

9,133,829

New and Expanded Business Facility Credit

Yes

8,779,797

Missouri Business Use Incentives for Large-Scale Development (BUILD)

8,419,707

Examination Fee Tax Credits

7,576,530

Missouri Property and Casualty Guaranty Association

7,227,710

Neighborhood Preservation Tax Credit

Yes

6,784,310

NOTE: Bold typeface indicates the six programs with freely transferable credits.

Table 1 comes from data in the appendix. We
sorted the data by the amount of tax credits issued
and extracted those programs issuing more than
$5 million in credits. This gives us 14 programs.
Of the 14, eight are designated as “transferable.”
This does not mean the same thing in all cases,
however. In particular, two of the eight programs
are subject to special restrictions. For example, the
Low Income Housing Tax Credit Program follows
federal rules. These require an entity using the
credit to have an ownership interest in a housing
project. Thus, it is not enough just to have Missouri
tax liability. One cannot simply auction the credits
to the highest bidder.
Similarly, the Certified Capital Companies
Program follows rules specified by the DED: Rule
4 CSR 80-7.040(G) states that tax credits may be
sold only to insurance companies. Again, one cannot simply sell the credits to the highest bidder.
In contrast, the statutes for the remaining six
tax credits define relatively free markets for those
credits. These six transferable tax credit programs
are indicated in bold in Table 1, and they are the
focus of our analysis.
56

V O LU M E 3 , N U M B E R 2

2007

BRIEF DESCRIPTIONS OF THE
8
SIX PROGRAMS
In this section we briefly describe the six programs: We give a short statement of (i) what each
credit is for, (ii) whether the credit is awarded at
the discretion of the DED or is an entitlement to
any entity that meets the statutory criteria, (iii) the
taxes against which it can be applied, (iv) whether
there are carry-back and carry-forward provisions,
(v) and any special statutory language about transferability.

Historic Preservation Tax Credit Program
• The credit is given for 25 percent of the qualifying expenses incurred in the rehabilitation
of an approved historic structure. The total
costs of the rehabilitation must be more than
half of the acquisition cost of the property.
• The credit is an entitlement.
8

This information comes from DED (2005 and 2006) and the relevant
statutes.

F E D E R A L R E S E R V E B A N K O F S T. LO U I S R E G I O N A L E C O N O M I C D E V E LO P M E N T

Rothstein and Wineinger

• The credits can be applied to the following
taxes: income (excluding withholding), bank,
insurance premium, and other financial
institution.
• There is a carry-back of 3 years and a carryforward of 10 years.9
• By statute the credits cannot be issued to
nonprofit entities, but they can buy and sell
them.10

Infrastructure Tax Credit Program
• The credit is equal to 50 percent of a contribution made to local governments or state
agencies to finance the development of publicly owned essential public purpose infrastructure such as water, sewer, gas, electrical
systems, streets, bridges, rail spurs, storm
water drainage, and other projects.
• The credit is discretionary.
• The credits can be applied to the following
taxes: income (excluding withholding),
corporate franchise, bank, and insurance
premium.
• There is a carry-forward of 5 years.
• By statute, the credits must transact at
between 75 and 100 percent per dollar of
credit. There is also an explicit stipulation
that the seller of the credits must report the
payment received from the buyer as taxable
income and that the buyer must recognize
the difference between the face value and
his payment to the seller as taxable income.

Brownfield Redevelopment Tax Credit
Program (Remediation)
• The tax credit is worth up to 100 percent of
the costs of cleaning contaminated commercial or industrial sites that have been underutilized for at least 3 years. The project must
9

A carry-back of three years allows a taxpayer whose tax credits
exceed his current tax liability to use the excess to offset tax liability
from the three previous years. This presumably generates a refund
from the state if the taxpayer is not in arrears.

We confirmed this last point with the DED.

retain 25 jobs or create 10 new jobs; if the
property is privately owned, a city or county
government must endorse the project; and
all projects must be accepted into the state’s
Voluntary Cleanup Program.
• The credit is discretionary.
• The credits can be applied to the following
taxes: income (excluding withholding), corporate franchise, bank, insurance premium,
and other financial institution.
• There is a carry-forward of 20 years.

Affordable Housing Assistance Tax
Credit Program
• The credit is equal to 55 percent of a contribution made to a nonprofit housing organization. The contribution must be used for the
building, procurement, rehabilitation, and
in some cases basic operating expenses of a
housing organization that provides certain
types of housing, either affordable (targeted
toward persons below 50 percent of median
income) or market rate (targeted toward
“rebuilding communities” as defined by
statute).
• The credit is discretionary.
• The credits can be applied to the following
taxes: income, corporate franchise, bank,
insurance premium, other financial institution, and express company.
• There is a carry-forward of 10 years.

New and Expanded Business Facility
Tax Credit Program
• Tax credits are awarded during a 10-year
window based on capital invested in new
facilities and the creation of new jobs.
Facilities must belong to certain industries,
including manufacturing, research and
development, and computer-related services.
The credit amount varies depending on
whether the facility is owned by an existing
Missouri company or a new company and
whether the facility is in a distressed area or
not. The program is being phased out, but

F E D E R A L R E S E R V E B A N K O F S T. LO U I S R E G I O N A L E C O N O M I C D E V E LO P M E N T

V O LU M E 3 , N U M B E R 2

2007

Rothstein and Wineinger

credits could potentially be issued through
2014.11
• The credit is an entitlement.
• The credits can be applied to the following
taxes: income (excluding withholding),
insurance premium, and insurance company
retaliatory.
• There is no carry-forward for the recipient
(but see the next bullet point).
• Regarding transferability, the statutory language is similar to that for the infrastructure
tax credit program. The credits must transact
at between 75 percent and 100 percent per
dollar of credit. There is an explicit stipulation that the seller of the credits must report
the payment received from the buyer as taxable income, and the buyer must recognize
the difference between the face value and
his payment as taxable income.

Neighborhood Preservation Tax Credit
Program
• This tax credit is worth a minimum of 15
percent and maximum of 35 percent of eligible expenses for new construction or rehabilitation of owner-occupied homes incurred
in communities with median household
incomes that are low for their metropolitan
statistical area.
• The credit is an entitlement.
• The credits can be applied to the following
taxes: income (excluding withholding), corporate franchise, bank, insurance premium,
and other financial institution.
• There is a carry-back of 3 years and a carryforward of 5 years.
• There is an explicit statement that the credit
cannot be claimed in addition to any other
state tax credits.
11

Only facilities that applied for the credit on or before December 16,
2004, and began operations on or before that date are eligible. In
each year in the 10-year window in which at least $100,000 in new
capital is invested (or $1,000,000 in replacement facilities) and two
new jobs are created (25 for office jobs), an existing Missouri company receives a credit of $100 ($150 in distressed areas) per new job
and a new Missouri company a credit of $75 ($125 in distressed
areas) per new job.

V O LU M E 3 , N U M B E R 2

2007

THE SCOPE OF THE DED’S
PROGRAM EVALUATIONS
The tax credit analysis form issued by the
Missouri Office of Administration states the
following:
Per 33.282.2 RSMo, each department authorized to offer deductions, exemptions, credits or
other tax preferences shall submit the estimated
amount of such tax expenditures for the fiscal
year beginning July 1st of the following year and
a cost/benefit analysis of such tax expenditures
for the preceding fiscal year.

Pursuant to this, the DED performs an annual
analysis of the tax credits that it administers.
Table 2 compiles DED data on fiscal costs and
benefits for our six programs for 3 consecutive
years. In other words, all of the reported costs and
benefits, whether direct, indirect, current-year, or
long-run, are simply state tax revenues that are
lost or gained as a result of the programs.
Four features of the data are worth noting:
• The data on fiscal costs and benefits in the
2005 report is more complete than that in
the previous two reports.
• The direct cost, total cost, and tax credits
redeemed for each program are generally all
equal. In other words, the DED measures
program costs by the tax credits redeemed.
This may be generally reasonable, but over
the long run economic development credits
(such as the infrastructure and business
facility credits) create additional fiscal costs,
for example, through demand for additional
government services.
• Using FY 2005 data, DED forecasts fiscal
net losses in the long run (from the credits
granted in 2005) for the tax credit programs
for historic preservation, affordable housing,
and neighborhood preservation. These programs also generate net losses in the short
run.
• Two additional programs, for infrastructure
and brownfield redevelopment, generate
net losses in the short run (but net benefits

F E D E R A L R E S E R V E B A N K O F S T. LO U I S R E G I O N A L E C O N O M I C D E V E LO P M E N T

80,213,374
28,964,274
14,808,297
9,133,829
8,779,797
6,784,310

Issued

75,692,322
39,401,435
4,250,346
10,378,534
8,702,349
4,440,206

Issued

89,214,177
11,227,302
15,481,014
6,223,933
9,168,145
5,085,754

Issued

74,532,355
25,953,799
10,627,870
7,702,860
4,546,330
8,461,503

Redeemed
(also direct cost) Total cost

25,324,223
395,877
18,972,677
1,904,063
38,187,629
1,365,838

–52,116,976
31,897,315
–14,171,136
–6,240,221
10,400,889
–3,256,094

Total net
benefit (2004)

–17,655,146
–9,855,072
12,155,132
–5,907,650
30,296,647
–2,610,110

Total net
benefit (2003)

NA
NA
NA
NA
NA
NA

Benefit-cost
ratio (5-year)

1.25 to 1
0.04 to 1
5.65 to 1
0.49 to 1
7.95 to 1
0.68 to 1

Benefit-cost
ratio (5-year)

5,267,125
10,286,639
2,639,434
1,073,787
7,914,995
698,739

7,524,464
14,695,198
3,770,620
1,533,982
11,307,136
998,198

–67,007,891
–11,258,601
–6,857,250
–6,168,878
6,760,806
–7,463,305

0.66 to 1
5.21 to 1
2.89 to 1
0.32 to 1***
24.05 to 1
0.19 to 1***

Total net
Benefit-cost
Direct benefit Total benefit benefit (2005) ratio (12-year)

15,291,920
37,739,094
1,849,955
1,553,279
16,213,525
960,679

Direct benefit Total benefit

NA
NA
NA
NA
NA
NA

Direct benefit Total benefit

67,408,896
8,822,146
5,841,779** 22,250,124
16,021,091**
1,181,802
7,793,500
793,546
5,812,636** 10,214,799
4,216,773
491,053

Total cost

Redeemed
(also direct cost)

66,089,980*
10,020,578
16,101,975*
7,554,503
7,826,417*
4,001,293*

42,979,369
10,250,949
6,817,545
7,811,713
7,890,982
3,975,948

Total cost

42,979,369
10,250,949
6,817,545
7,811,713
7,890,982
3,975,948

Redeemed
(also direct cost)

NOTE: *Redeemed is slightly less than direct cost; **reports negative indirect costs; ***5-year.

Historic Preservation
Infrastructure
Brownfield Redevelopment (remediation)
Affordable Housing Assistance
New and Expanded Business Facility
Neighborhood Preservation Tax

Tax credit program FY 2005

Historic Preservation
Infrastructure
Brownfield Redevelopment (remediation)
Affordable Housing Assistance
New and Expanded Business Facility
Neighborhood Preservation

Tax credit program FY 2004

Historic Preservation
Infrastructure
Brownfield Redevelopment (remediation)
Affordable Housing Assistance
New and Expanded Business Facility
Neighborhood Preservation

Tax credit program FY 2003

Fiscal (Tax Revenue) Costs and Benefits from Each Program

Table 2

Rothstein and Wineinger

V O LU M E 3 , N U M B E R 2

2007

Rothstein and Wineinger

Table 3
Reported Impact of $1 of Redeemed FY 2005
Tax Credits on Personal Income (over 5 or
12 years)
New
personal income

Tax credit program
Historic preservation**

$11.68

Infrastructure**

137.01

Brownfield (remediation)**

64.18

Affordable housing*

5.59

Business facility**

499.70

Neighborhood preservation*

3.32

NOTE: *Indicates 5 years; **indicates 12 years.

in the long run). Only the business facility
credit generates net benefits in both the
short and long run.
Regarding tangible nonfiscal costs and benefits,
DED produces a range of information on the impact
of each program in the short and long run. For
illustration, Table 3 presents the long-run increase
in personal income for every dollar of redeemed
credits in 2005.12
From an economic perspective, there are two
basic problems with the way DED evaluates the tax
credit programs. First, DED regards each program
as if its purpose were economic development (i.e.,
to create faster growth or more jobs than would
occur otherwise). This is not true, however, because
some of these programs are more likely to correct
market failures or reduce economic disparities
than promote development. These different goals
imply different structural features for the programs
and require the measurement of somewhat different
variables for proper program evaluation. We discuss
these points in greater detail in the section on program design and evaluation. In many cases, DED’s
focus on development forces a stark mismatch
between its analysis and the analysis suggested by
economic theory.
12

This is clearly the total effect over the given time period and not the
average annual effect over the time period. We do not know whether
DED uses a discount rate in computing the total effect.

V O LU M E 3 , N U M B E R 2

2007

A second problem comes from DED’s tendency
to produce lists of raw information, like that in
Table 3. Although this information should certainly
be reported, lists of facts about impact are not a
substitute for a single, bottom-line number of all
the tangible benefits and costs for each program.
By tangible program benefits and costs, we mean
simply those impacts that can be plausibly converted into dollars and cents. Public policies often
have important intangible benefits and costs. These
highly subjective items are key components of
policies that preserve state history, reduce poverty,
and promote opportunity, but may also interfere
with property rights or reduce the incentive to
work. Concrete information about tangible benefits
and costs is essential for having an informed debate
over intangible benefits and costs. Economic
analysis cannot remove the subjective element in
program evaluation. It can, however, narrow it
substantially. Lists of raw data do not do this.13

OTHER EVALUATIONS OF THE
PROGRAMS AND OF DED
We found three recent evaluations of the
Missouri tax credit programs besides those done
by DED. We present a brief discussion of each.

The State Auditor’s Report
In February 2001, the Missouri state auditor
issued a report entitled, “Review of the State Tax
Credits Administered by the Department of
Economic Development” (Office of the State
Auditor of Missouri, Claire McCaskill, 2001). For
our purposes, the report makes three points worth
noting.
First, the report is very critical of data collection at DED. The accuracy of the collected data is
one concern, but the report focuses more sharply
on omissions in data collection:
Key data such as number of actual jobs created
per project, average wages, total investment,
industry sectors affected, street addresses for the
projects and other relevant information are not
13

For an excellent example of one way to produce a single measure of
all tangible costs and benefits, see Bartik (2005).

F E D E R A L R E S E R V E B A N K O F S T. LO U I S R E G I O N A L E C O N O M I C D E V E LO P M E N T

Rothstein and Wineinger

captured in a centralized management information system and in some cases are not captured at all. (p. 2)
Data is not maintained or monitored for 16 of
33 tax credit programs. These 16 programs are
formula-based tax credits, which are granted if
the project meets the eligibility requirements
set out in the authorizing statute. Department
of Economic Development management stated
that it is not the responsibility of the agency to
monitor the economic impact that formula-based
tax credits have on the state because they have
no discretion over whether projects qualifying
for the tax credit receive the tax credit. (p. 7)

Second, the report is very critical of “costbenefit analysis.” It is not entirely consistent in its
use of this term, but the intent is clear:
A former economist for the Department of
Economic Development, who was the former
Manager and Senior Economist for Office of
Research and Policy Analysis, said that he has
read a lot of the economic literature and attended
numerous national conferences and has come
to the conclusion that no one knows how to perform a useful cost-benefit analysis of tax credit
programs. (p. 4-5)

This comment, while not entirely without
foundation, is presented as a definitive statement
about a fundamental limit of economic analysis.
We think it more accurately reflects something else.
One cannot perform a meaningful cost-benefit
analysis of tax credit programs by treating all of
them as if they were economic development programs. Some programs address market failures
and others address disparities and inequities. Only
some are supposed to promote economic development. One can use cost-benefit analysis in all cases,
but one must measure different variables depending on the purpose of the program. Ignoring these
differences and reviewing non-development credits as if they were development credits is an incoherent exercise. It should not be surprising that no
one knows how to do it.
The auditor concludes by calling for an
“impact analysis” of each program. She had her
staff analyze four of the smaller tax credit programs.
In each case she used DED’s regional forecasting

model to consider the impact of the tax credits
issued on state revenues, employment, wages, and
gross state product. To make the simulation internally consistent, she assumed that government
spending would fall by the amount of the tax credits
redeemed. She also sent surveys to recipients of
the tax credits, asking if the credits were essential
to their decision to make their investments.14
The impact analysis conducted by the auditor
was a step in the right direction in 2001. It is now
conducted by DED for all its tax credit programs,
and the results are presented in its annual report.15

Don Phares’s Report
Don Phares (2003) of the University of Missouri
at St. Louis produced a report for the Department
of Revenue entitled, “Examining Missouri’s Tax
System: Tax Expenditures—A First Step.”16
The term “tax expenditures” in the title is very
general. It encompasses all exceptions to general
tax rules that result in less tax liability for some
entity. Tax credits are one kind of tax expenditure.
Professor Phares echoes the auditor’s call for
more data collection, noting the progress DED made
in evaluating the programs under its discretion;
he adds, however, that “proper evaluation would
allow the extent to which [the credits] meet their
intended purpose to be addressed” (p. 13). This
refrain is likely to be repeated into the indefinite
future, until such time as (i) it is recognized that
the intended purpose may be to address some
kind of market failure or to correct disparities and
inequities and not merely to promote economic
development, (ii) key program parameters are
14

Asking the recipients of the tax credits if the credits were essential
to their decisions is a highly problematic exercise. The recipients
have their own agendas, and they may answer these questions in
pursuit of those agendas. This may not entail giving a thoughtful or
candid answer to a direct question. One of the purposes of social
science and nonexperimental statistics is to provide methods for
learning about the causes of choices without relying on answers to
these kinds of questions. Furthermore, while the auditor is certainly
right that the only way to answer the question with “absolute certainty” in each case is to know the subjective thoughts of the project
managers (p. 19), a “statistical certainty” is often enough for policy
purposes.

We were told that DED does try to determine for the discretionary
programs whether the credits are essential to the investment decisions.

The report contains references to other evaluations of Missouri tax
credits that we do not discuss here.

F E D E R A L R E S E R V E B A N K O F S T. LO U I S R E G I O N A L E C O N O M I C D E V E LO P M E N T

V O LU M E 3 , N U M B E R 2

2007

Rothstein and Wineinger

Table 4
Recommendations of the Incentives Review
Committee
Tax credit program

Recommendation

Historic preservation

Maintain

Infrastructure17

Maintain

Brownfield (remediation)

Maintain18

Affordable housing

Maintain

Business facility

(Not evaluated)19

Neighborhood preservation

Improve

chosen in ways that are consistent with each program’s rationale, and (iii) each program is evaluated
in ways that are consistent with its rationale.

Incentives Review Committee Report
The most recent report on tax credit programs
in Missouri was undertaken by the Incentives
Review Committee of the Missouri Department
of Economic Development (2005), pursuant to a
request by Governor Matt Blunt. It has the efficient
title, “Report on Missouri Incentives Programs.”
This report is excellent in two respects. First,
it pays careful attention to the different purposes
of each program. Second, it recommends evaluation criteria that bear some connection to these
purposes. This makes an interesting contrast with
the auditor’s report on the issue of (broad) costbenefit analysis. Indeed, its two pages of criteria
for program evaluation really go beyond costbenefit analysis and include objectives such as
transparency and low transactions costs (although
it uses different language). The committee then
provides a brief evaluation of each program and
a recommendation to “improve,” “combine,”
“delete,” or “maintain” the program.
Unfortunately, it is not clear how rigorously
the committee used these criteria in developing
17

This is listed in the report as the “MDFB Contribution Credit.”

The summary table at the front of the report recommends “combine,”
but in the discussion of the program the recommendation is “maintain.” The former appears to be an editing error.

This program is being phased out, as noted in the section that provides brief descriptions of these programs.

V O LU M E 3 , N U M B E R 2

2007

their evaluations and recommendations. The
report gives only a brief discussion of each program
and provides little explanation for its conclusions.
Furthermore, implementing these criteria would
have required extraordinary amounts of money,
time, data, and expertise. Even with those, the
analysts would have needed to conjecture about
missing information and poorly understood behavioral responses. This would have necessitated a
sensitivity analysis to determine how the conclusions depended on the assumptions. If the authors
did these things, they do not mention them in the
report.
We report in Table 4 the recommendations in
the report for our six programs.
The only recommendation for change is for
the neighborhood preservation credit. The authors
note that
[t]he program is based on a first-come allocation
and the demand is significantly more than the
annual allocation. As such, the distribution of
funding has been on a lottery basis, which does
not provide for a concentrated redevelopment
impact. Also, some areas in the city-wide distressed areas are not lower income. (p. 40)

Their recommendation for improvement is to
[e]nact legislation to award funds on a competitive basis, requiring a comprehensive neighborhood redevelopment plan. (p. 40)

PROGRAM DESIGN AND
PROGRAM EVALUATION
Are Missouri’s tax credit programs good public
policy? We take this question to mean, in the broadest possible terms, is there something inefficient
or unfair about the outcomes that would result if
the program did not exist? We provide a general
discussion of this question for each of the six largest
tax credit programs (by number of credits issued)
that have freely transferable credits. In doing this,
we draw on current research in local public-sector
economics and basic economic principles. Given
the difficulty of the question and the limited amount
of information available about these programs,

F E D E R A L R E S E R V E B A N K O F S T. LO U I S R E G I O N A L E C O N O M I C D E V E LO P M E N T

Rothstein and Wineinger

however, we decided not to overly simplify matters
just to provide “yes” or “no” answers.
Although our analysis is qualitative and not
quantitative, we must still keep a specific thought
experiment in mind to evaluate the programs consistently. Following standard practice in the evaluation of tax policy, the initial assumption is that
the tax credit program does not exist and the state
budget is balanced. The tax credit is then introduced. This creates a shortfall in revenue that must
be met either by a reduction in spending or an
increase in other taxes: We generally consider both
cases. These changes taken together have equity
and efficiency implications.

Historic Preservation Tax Credit Program
Recall our basic question in the context of this
program. This tax credit causes Missouri to have
more historic structures than market forces alone
would provide. Is there something inefficient or
unfair about this?
The main reason that the market may fail here
is that historic structures provide benefits to more
people than just the owners of the structures. For
example, the structures may enhance local property
values or generate local tourism. These benefit
many non-owners. In the formal language of economics, the market for historic structures does not
capture the “willingness to pay” of all the beneficiaries, who receive a “positive externality.” The
result is that too few structures are preserved. This
simple point has useful implications for program
design and evaluation.
First, an efficient program must have a carefully chosen subsidy rate. This rate should induce
additional preservation activity up to the point
where the total additional benefits from this activity
(higher property values, tourism, etc.) exactly balance the total additional cost. These costs would
include the “true” cost of the subsidy itself. This
could well be more than the dollar value of the
subsidy, both because raising revenue is costly and
because any state services forgone to fund the subsidy may be worth more than their simple dollar
value. We have no way to determine whether the
current 25 percent subsidy rate is proper according
to these economic criteria.20

Second, even if the policy were well designed,
it would never pass a cost-benefit test that sums up
only fiscal costs and benefits (i.e., state tax revenues
that are lost or gained). This is true at the current
subsidy rate and should also be true at the proper
subsidy rate, given the highly indirect links between
historic preservation activities and state tax revenues. The policy should, however, pass a costbenefit test that sums up all tangible costs and
benefits. As noted above, these benefits would
include the willingness to pay of all the beneficiaries of the program. Also as noted above, a bottomline estimate of all the tangible net benefits would
allow citizens and legislators to debate, in a structured (albeit subjective) way, the intangible aspects
of the program.
Finally, we note that the program is an entitlement. Entitlement tax credits, in contrast to discretionary credits, are often criticized because they
probably give more money for activities that would
have been undertaken anyway.21 To the extent this
occurs, the program is redistributing income from
one group of taxpayers (those who fund the subsidy
or lose government benefits) to another (those who
engage in preservation activities). This may not be
entirely bad, however. If the subsidy is funded by
higher taxes and some of those taxpayers receive
positive externalities from historic structures, then
they are now simply paying for something that
they previously enjoyed for free.
More generally, one should not overreact to the
incidental income transfers that will always be part
of simple market solutions to market failures. The
alternative is to make the tax credit discretionary,
but this would certainly lead to higher administrative costs. There is a trade-off between incidental
redistribution and administrative costs that should
not be forgotten and which may not be small.

Infrastructure Tax Credit Program
This program allows local communities to
have more and better public infrastructure than
they would otherwise have. Again we ask, Is there
20

Note that these criteria have nothing to do with economic
development.

Entitlement credits also tend to be reviewed less carefully. This is a
separate issue that is emphasized in Bartik (2005).

F E D E R A L R E S E R V E B A N K O F S T. LO U I S R E G I O N A L E C O N O M I C D E V E LO P M E N T

V O LU M E 3 , N U M B E R 2

2007

Rothstein and Wineinger

something inefficient or unfair about the amount
of infrastructure in different communities that
market forces alone would provide?
Regarding efficiency, there is a large literature
in economics on the question of whether there is a
“race to the bottom” in the funding of public infrastructure. The idea is that jurisdictions (communities, states, and even countries) compete with
each other for investment and workers by cutting
taxes. The question is whether taxes can be cut so
far that even basic infrastructure that all residents
desire will not be provided. If so, then an infrastructure tax credit program may enhance efficiency
by allowing communities to finance the proper
amounts of infrastructure despite the race to the
bottom.
Whether or not a race to the bottom occurs
depends on a number of factors. These include
whether communities have the appropriate tax
instruments at their disposal, whether there is a
large or small number of communities, who owns
the land and other spatially fixed inputs to production, the objectives of the local government, and
the time frame of the analysis (short run or long
run). We think it is fair to say that, except for the
very long run, the consensus in the literature is
that in the real world there is quite likely to be a
race to the bottom in tax cuts and therefore in the
funding of public infrastructure.22
Given that a race to the bottom tends to exist,
the recommendations of economic theory for program design and evaluation are similar to the recommendations for the historic preservation program.
The program should subsidize local spending with
a carefully chosen subsidy rate. This rate should
be selected so that each jurisdiction, acting in a
decentralized way, chooses the proper quantity of
infrastructure. It can be shown that this essentially
requires a subsidy rate equal to the tax revenue
each region fears it would lose if it increased its
own tax rate (from the rate it would choose without the tax credit program). With this subsidy, the
quantity of infrastructure would be such that any
further increase in infrastructure would generate
22

The literature is large, but the key papers on which this conclusion
is based are by Zodrow and Mieszkowski (1986), Wildasin (1989),
and Myers (1990).

V O LU M E 3 , N U M B E R 2

2007

only as much consumption as would be foregone
to pay for that increase.
Recall that the actual program does not subsidize the local tax rate; rather, it funds $1 of infrastructure through a 50 cent donation from private
citizens and (presumably) a 50 cent reduction in
state services. Because the structure of this program
is so far removed from the basic economic incentives that an efficient program would take into
account, it is difficult to evaluate. Given that a market failure exists, the program may well be better
than nothing, but that is a low bar to surmount.
Finally, we note that, as with the historic
preservation program, even a well-designed infrastructure credit program would never pass a costbenefit test that sums up only fiscal costs and
benefits. There would be little or no connection
between the fiscal cost of the proper subsidy and
any extra revenues the state would receive from
any extra infrastructure attributable to the program.
The policy should, however, pass a cost-benefit
test that sums up all tangible costs and benefits.

Brownfield Redevelopment Tax Credit
Program (Remediation)
This program causes Missouri to have fewer
commercial/industrial sites with hazardous substances than market forces alone would allow. Is
there something inefficient or unfair about the
latter? After all, if land in a certain area becomes
scarce enough, then eventually even brownfields
will find buyers.
The problem with brownfields, of course, is
the problem of negative externalities. Brownfields
are ugly, they present various hazards (for example,
to ground water), and they deter development
somewhat beyond their immediate boundaries.
The externalities presumably differ wildly depending on the type of hazardous waste involved and
its proximity to different communities. This line
of thought leads to the same kind of conclusions
as have been drawn for the previous two programs,
with one major exception. The huge variation in
external damages at different locations provides
an economic rationale for making the program
discretionary, which it is.
Unfortunately, if this is the right perspective
on this program, then another feature of its design

F E D E R A L R E S E R V E B A N K O F S T. LO U I S R E G I O N A L E C O N O M I C D E V E LO P M E N T

Rothstein and Wineinger

presents a paradox. The program charges taxpayers
100 percent of the costs of cleanup, but only when
a third party, a commercial developer, is interested
in the site. This is very odd. If a brownfield is creating large enough negative externalities that citizens
are willing to pay 100 percent of the costs to clean
it up, then one should expect them to raise the
funds and do so. The broad-based taxes they would
generally use are not ideal for this purpose, but
they are practical and they work. We therefore
expect most brownfields to exist where citizens’
willingness to pay is less than the cost of cleanup.
This makes it efficient to wait until a third party—
the developer—has a willingness to pay to make
up the difference. However, because the developer
does not pay anything for the cleanup, it is unlikely
that his willingness to pay does make up the difference. The developer’s interest is simply an
expressed desire to have something when other
people pay for it. If citizens’ and the developer’s
total willingness to pay is less than the actual cost
of the cleanup, then the cleanup is inefficient. It is
also arguably unfair to have the taxpayers pay the
full cost.

Affordable Housing Assistance Tax Credit
Program
This program is obviously intended to be
redistributive. One must recognize this fact to
properly evaluate the program. An efficiency issue
is also present, however. We discuss this first.
Affordable housing policy in one state always
raises the question of whether it may lead people
from other states to move to the more generous
state. This is a variation on the race to the bottom
discussed under the infrastructure program. Unfortunately, although state action can help overcome
the problems of fiscal competition among local
communities, federal action would be required to
help overcome the problems of fiscal competition
among the states. The movement of low-income
residents in response to differences in benefits
across states has been studied for many years by
all kinds of social scientists. For the purposes of
the analysis here, we assume that these movements
are small. The proper evaluation of this program
by DED would make use of this literature.

In contrast with the historic preservation program, this program’s focus on redistribution makes
more central the issue of whether the program
creates housing that would not have been created
otherwise. In other words, it is important to know
if the program merely “crowds out” private spending on the construction of affordable housing. If
the program is ineffective in this way and funded
by higher taxes, then there is an income transfer
from taxpayers to developers and the poor are most
likely neither harmed nor helped. If the program is
ineffective and funded by reduced state spending in
other areas, then the poor may well be worse off.
It is surely unlikely that an affordable housing
program crowds out for-profit private-sector spending. It is a fair question, however, whether it crowds
out not-for-profit or philanthropic spending on
affordable housing. If it does, then it is difficult to
see how the poor could be helped, and any reductions in other programs to fund this program could
be harmful. Again, there is a literature in economics
on this issue, and the proper evaluation of this
program by DED would make use of it.
No purely redistributive program would pass
a cost-benefit test of either fiscal costs and benefits
or tangible costs and benefits. Tangible costs and
benefits would exactly balance if funds could be
raised costlessly and services delivered without
waste; but neither is true. Thus, tangible costs will
always exceed benefits. Again, this makes it important to have a single, bottom-line measure of the
tangible net costs of the program. Presenting this
information on, say, a per-beneficiary basis might
allow citizens and legislators to debate, in a structured way, whether the intangible benefits of housing assistance are sufficient to offset the costs.

New and Expanded Business Facility
Tax Credit Program
At long last, we come to an economic development program. In other words, this program does
not correct a market failure or attempt to redistribute income to assist the poor. It is an attempt to
create growth that market forces alone would not
provide.23
23

“Missed development opportunities” may result from missing capital
markets or missing insurance markets. These market failures are
qualitatively different from the externalities discussed above, however.

F E D E R A L R E S E R V E B A N K O F S T. LO U I S R E G I O N A L E C O N O M I C D E V E LO P M E N T

V O LU M E 3 , N U M B E R 2

2007

Rothstein and Wineinger

DED’s methods of analysis are directed toward
addressing this question. Even here, however, it is
possible to take the analysis much further than DED
does. A proper evaluation of economic development
programs would have the following steps, of which
the DED does only the first.
1. For all programs, determine the long-run
fiscal benefits and costs.
2. For all programs, determine the other tangible
long-run benefits and costs. Roughly, these
benefits would include the extra earnings of
residents due to improvements in local labor
markets and increases in property values,
while costs would include any extra public
services or infrastructure required.
3. Combine steps 1 and 2 to compute a single,
bottom-line measure of the tangible net
benefits or net costs of each program.
4. Perform sensitivity analysis to remedy the
problems of missing information and poorly
understood behavioral responses, which
affect steps 1 and 2. For step 1, there is uncertainty over how much activity would take
place if not for the credits and uncertainty
over the multiplier effects used to compute
long-run fiscal benefits. For step 2, the benefits and costs of the labor market effects, for
example, are sensitive to whether the new
jobs are going to people who were formerly
unemployed or not.24
The business facility credit raises tax revenue
in the short and long run. The only issue, then, is
whether the costs of additional public services that
come with development could offset the large gains
reported by the state. This seems highly unlikely.

Neighborhood Preservation Tax Credit
Program
The efficiency rationale for this program is
very similar to the rationale for the historic preservation program. Preserving neighborhoods has
benefits to people besides those who buy, sell, and
improve homes in those neighborhoods. We refer
24

This is emphasized in Bartik (2005).

V O LU M E 3 , N U M B E R 2

2007

the reader to the discussion of the historic preservation program for the basic analysis.
Finally, recall the recommendation of the
Incentives Review Committee for this program
(see the section on other evaluations of the programs and of the DED). They argue that redevelopment should be concentrated and the subsidy given
to activities that are part of a redevelopment plan.
This is correct if the external benefits of neighborhood preservation are highly dependent on concentrated activity. This seems likely. No similar
recommendation was made for the historic preservation program, but this is a sound position with
regard to efficiency if even somewhat isolated historic structures have external value.

SHOULD THE TRANSFERABLE
CREDITS ALSO BE REFUNDABLE?
We conclude by considering a reform that
would be applicable to all but the business facility
tax credit. None of the other five credits is refundable. That is to say, once an entity has offset all of
its tax liability it cannot use the remainder to
receive a refund from the state. The only way an
entity with little or no tax liability (like a nonprofit
organization) can benefit from the tax credits is by
selling them. In contrast, the business facility credit
is refundable and transferable.25
A potential problem with credits that are just
transferable is straightforward: A $1 tax credit does
not sell for $1, but the credit will cost the taxpayers
of Missouri that amount when it is redeemed.
Money that was supposed to support public programs ends up as profit to the buyer of the credits.
In contrast, if the tax credits were also refundable,
then every tax dollar spent on the tax credit would
go toward the intended activity.
25

The refundability and transferability of the business facility credit
are somewhat constrained, however. Missouri Revised Statutes,
Chapter 135, Section 135.110 states, “[T]o the extent such credits
exceed the taxpayer’s Missouri tax on taxable business income,
[they] shall constitute an overpayment of taxes and in such case, be
refunded to the taxpayer provided such refunds are used by the taxpayer to purchase specified facility items.” If the credits are sold,
the selling price must be at least 75 percent of the face value. This
seemed like a minor constraint, however, which is why we consider
the credit to be freely transferable.

F E D E R A L R E S E R V E B A N K O F S T. LO U I S R E G I O N A L E C O N O M I C D E V E LO P M E N T

Rothstein and Wineinger

Table 5
Transactions Prices for Tax Credits
Program
Historic preservation
Brownfield (remediation)
Neighborhood preservation

Observations

Average
sale price per dollar

Average transaction

3,624

90 cents

5 cents

$178,622

816

91 cents

6 cents

137, 442

1,322

87 cents

9 cents

28,658

To illustrate, we recently obtained data on all
transactions from 2002-06 for the historic preservation, neighborhood preservation, and brownfield
(remediation) credits.26 As Table 5 shows, they all
sell for about 10 cents (per dollar) below face value.
Many factors contribute to the discount on the
tax credits. For arm’s length transactions, the purchaser of the credit would take into account the
competitive return on other uses of her capital. This
in turn would depend on how long she expects
her capital to be tied up in the investment, the
riskiness of the investment, and the true net-of-tax
return. As many people with whom we spoke
emphasized, the tax consequences of transacting
and using the credits is particularly important in
evaluating the discount. For example, the profit
from purchasing tax credits at a discount and
then using them is itself taxable, and the use of
the credits also causes a taxpayer who itemizes and
pays ordinary income tax (as opposed to alternative minimum tax) to lose some of her federal
deduction for state taxes paid. Whether or not
these factors could explain the discount, especially
for short-term investments, is the subject of ongoing research.
Refundability, especially when coupled with
transferability, is a reform that merits further
exploration. It is not without its critics, however.
Refundability is of no value to agents who need
immediate liquidity. For them, transferability is
essential. This is most likely for people using the
credits as part of the financing for large capital
projects. Also, refundability requires an extra

degree of cooperation from the government. The
state is relatively passive in allowing a taxpayer to
sell a credit and take a deduction. The state must
act in sending a refund. At least during a recession,
the holder of a tax credit who has no tax liability
may still want to sell it to someone who has tax
liability, even with a discount, if he believes the
state may delay paying the refund.27 At the very
least, these points make it clear that refundability
should be considered in conjunction with transferability and not as a substitute for it.

CONCLUSIONS
One theme we have emphasized throughout
the analysis is that, before one considers the specific goals of a program, one must understand why
a program is needed at all. Is there something inefficient or unfair about the outcomes that would
result if the program did not exist? Specific goals
that exist without a careful analysis of this basic
question are likely to be arbitrary and inconsistent.
We pose this basic question for each of the six
largest programs (by value of credits issued in 2005)
that issued “freely” transferable credits and see
where it leads.
More precisely, we ask whether the program
addresses a market failure, corrects disparities or
inequities, or promotes economic development.
We then see how these questions relate to program
design and evaluation.
27

Standard deviation

Transactions prices for the affordable housing credit are not available; there seems to have been only one transaction of the business
facility credit (which is also refundable); and data on the infrastructure tax credit is still being prepared for us.

It has also been suggested to us that refundable tax credit payments
count toward the annual constitutional spending limits in Missouri
(the Hancock Amendment) and that recent rulings by the IRS may
eliminate income recognition for transferable credits. Evaluating
these claims is outside the scope of our analysis here.

F E D E R A L R E S E R V E B A N K O F S T. LO U I S R E G I O N A L E C O N O M I C D E V E LO P M E N T

V O LU M E 3 , N U M B E R 2

2007

Rothstein and Wineinger

This brings us to our second theme. It is a
thankless (and flawed) task to evaluate every program as if its purpose were economic development.
Different goals imply different structural features
for the programs. They also require the measurement of somewhat different variables for proper
evaluation. To evaluate a program that corrects an
externality, one must gather some information
about that externality. To evaluate a program that
corrects disparities or inequities, one must talk
explicitly about who gains from the program and
who loses from the state spending that will be
reduced because of the tax revenue forgone as a
result of the program.
Third, we emphasize the need for analysis.
Lists of raw data about the impact of a policy are
not the same thing as evaluation. A proper evaluation develops a single, bottom-line number of all
tangible net benefits or costs. This is the only way
to set up a rational and informed debate over the
subjective or intangible benefits and costs of a
program.
In closing, it seems to us that proper program
evaluation simply cannot be done right now in
Missouri. The Department of Economic Development has great expertise but fundamentally works
for whoever occupies the governor’s office. This
may explain their narrow focus, uniform approach,
and tendency to report data instead of conclusions.
The state auditor has the necessary independence
but lacks the expertise. Outside academics have
the expertise but lack the specialized knowledge
that accrues to people who evaluate public programs for a living. The state will have many, perhaps
even more, tax credit programs into the foreseeable
future. It ought to consider creating an organization,
perhaps akin to the Congressional Budget Office,
with the independence, expertise, and accumulated
knowledge that leads to the very best program
evaluation.

V O LU M E 3 , N U M B E R 2

2007

REFERENCES
Bartik, Tim. “Solving the Problems of Economic
Development Incentives.” Growth and Change,
Spring 2005, 36(2), pp. 139-66.
Missouri Department of Economic Development,
Incentives Review Committee. “Report on Missouri
Incentive Programs.” November 22, 2005;
www.ded.mo.gov/upload/
incentivesreviewreportnov22finala.pdf.
Missouri Department of Economic Development.
“Report on Missouri Tax Credits Administered by
the Department of Economic Development.” 2006.
Myers, Gordon. “Optimality, Free Mobility, and the
Regional Authority in a Federation.” Journal of
Public Economics, October 1990, 43(1), pp. 107-21.
Office of the State Auditor of Missouri, Claire McCaskill.
“Review of State Tax Credits Administered by the
Department of Economic Development.” Report No.
2001-13, February 23, 2001;
http://www.auditor.mo.gov/press/2001-13.pdf.
Phares, Don. “Examining Missouri’s Tax System: Tax
Expenditures—A First Step.” April 2003;
http://dor.mo.gov/tax/phares/.
Wildasin, David. “Interjurisdictional Capital Mobility:
Fiscal Externality and a Corrective Subsidy.” Journal
of Urban Economics, March 1989, 25(2), pp. 193-212.
Zodrow, George and Mieszkowski, Peter. “Pigou,
Tiebout, Property Taxation, and the Underprovision
of Local Public Goods.” Journal of Urban Economics,
May 1986, 19(3), pp. 356-70.

F E D E R A L R E S E R V E B A N K O F S T. LO U I S R E G I O N A L E C O N O M I C D E V E LO P M E N T

Rothstein and Wineinger

APPENDIX
All Tax Programs, Sorted by Amount Issued, FY 2005
Alpha
code
LHC
HPC
EZC
IDC
RTC
NAC
AHC
BFC
BUC

RCN
YOC
NEC
NGC
TDC
DPC
APU
RCC
BJI
MHC
BEC
DVC
SBI
WGC
CPC
SBG

FDA
DTC
CBC
DRC
DFH
EFC
FPC
MQJ
SCC
NEZ
REC
SDT
ISB
SMC
SCT
BFT
BTC
DAC
WEC
ATC

Name
Missouri Low Income Housing
Historic Preservation TCP
Enterprise Zone
Infrastructure TCP
Brownfield Redevelopment Program (remediation)
Certified Capital Companies (CAPCO) Program
Neighborhood Assistance
Missouri Health Insurance Pool
Affordable Housing Assistance TCP
New and Expanded Business Facility Credit
BUILD (Missouri Business Use Incentives
for Large-Scale Development)
Examination Fee Tax Credits (exam)
Missouri Property and Casualty Guaranty Association
Neighborhood Preservation Tax Credit
Youth Opportunities
New Enterprise Creation
New Generation Cooperative Incentive
Transportation Development
Development Tax Credit
Agricultural Product Utilization Contributor
Rebuilding Communities
Brownfield “Jobs and Investment”
Maternity Home
Bond Enhancement/Bond Guarantee
Shelter for Victims of Domestic Violence
Small Business Incubator
Wine and Grape Production
Charcoal Producers
Small Business Guaranty Fees/Loan Guarantee Fee
Examination Fee Tax Credits (valuation)
Examination Fee Tax Credits (registration)
Family Development Account
Brownfield (demolition)
Community Bank Investment/Community
Development Corporation
Development Reserve
Dry Fire Hydrant
Export Finance
Film Production
Missouri Quality Jobs
Missouri Business Modernization and Technology
Missouri Life and Health Guaranty Association
New Enhanced Enterprise Zone
Qualified Research Expense/Research
Skills Development Credit
Small Business Investment/Capital
Sponsorship and Mentoring Program
Shared Care
Bank Franchise Tax
Bank Tax Credit for S Corporation Shareholders
Cellulose Casings
Disabled Access
Processed Wood Energy
Special Needs Adoption

Authorization, RSMo.

Transferable

Refundable

135.350 to 135.363
253.545 to 253.561
135.200 to 135.270, 135.429
100.286(6)
447.700 to 447.718
135.500 to 135.529
32.100 to 32.125
376.975
32.105 to 32.125
135.100 to 135.150, 135.258
100.700 to 100.850

Yes
Yes
No
Yes
Yes
Yes
No
No
Yes
Yes
No

No
No
Yes
No
No
No
No
No
No
Yes
Yes

148.400
375.774
135.475 to 135.487
135.460 and 620.1100 to 620.1103
620.635 to 620.653
348.430
135.545
32.110 to 32.125
348.430
135.535
447.700 to 447.718

No
No
Yes
No
Yes
Yes
Yes
Yes
Yes
Yes
No

135.600
100.297
135.550
620.495
135.700
135.313
135.766
148.400
148.400
208.750 to 208.775
447.700 to 447.718
135.400 to 135.430

No
Yes
No
Yes
No
Yes
No
No
No
No
No
Yes

No
No
No
No
No
No
No
No
No
No
At DED
discretion
No
No
No
No
No
No
No
No
No
No
No
No

100.25
320.093
100.25
135.750
620.1875 to 620.1890
348.300 to 348.318
376.745
135.1050 to 135.1075
620.1039
N/A
135.400 to 135.429
135.348
660.055
148.064
143.471
260.285
135.490
135.300
135.325

Yes
Yes
Yes
Yes
Yes
Yes
No
Yes
No
Yes
Yes
No
No
No
Yes
No
No
Yes
Yes

No
No
No
No
Yes
No
No
Yes
No
No
No
No
No
N/A
No
No
No
No
No

F E D E R A L R E S E R V E B A N K O F S T. LO U I S R E G I O N A L E C O N O M I C D E V E LO P M E N T

V O LU M E 3 , N U M B E R 2

2007

Rothstein and Wineinger

APPENDIX, cont’d
All Tax Programs, Sorted by Amount Issued, FY 2005
Name
Missouri Low Income Housing
Historic Preservation TCP
Enterprise Zone
Infrastructure TCP
Brownfield Redevelopment Program (remediation)
Certified Capital Companies (CAPCO) Program
Neighborhood Assistance
Missouri Health Insurance Pool
Affordable Housing Assistance TCP
New and Expanded Business Facility Credit
BUILD (Missouri Business Use Incentives
for Large-Scale Development)
Examination Fee Tax Credits (exam)
Missouri Property and Casualty Guaranty Association
Neighborhood Preservation Tax Credit
Youth Opportunities
New Enterprise Creation
New Generation Cooperative Incentive
Transportation Development
Development Tax Credit
Agricultural Product Utilization Contributor
Rebuilding Communities
Brownfield “Jobs and Investment”
Maternity Home
Bond Enhancement/Bond Guarantee
Shelter for Victims of Domestic Violence
Small Business Incubator
Wine and Grape Production
Charcoal Producers
Small Business Guaranty Fees/Loan Guarantee Fee
Examination Fee Tax Credits (valuation)
Examination Fee Tax Credits (registration)
Family Development Account
Brownfield (demolition)
Community Bank Investment/Community
Development Corporation
Development Reserve
Dry Fire Hydrant
Export Finance
Film Production
Missouri Quality Jobs
Missouri Business Modernization and Technology
Missouri Life and Health Guaranty Association
New Enhanced Enterprise Zone
Qualified Research Expense/Research
Skills Development Credit
Small Business Investment/Capital
Sponsorship and Mentoring Program
Shared Care
Bank Franchise Tax
Bank Tax Credit for S Corporation Shareholders
Cellulose Casings
Disabled Access
Processed Wood Energy
Special Needs Adoption

V O LU M E 3 , N U M B E R 2

2007

Carry forward

Carry back

Multi-year

5
10
0
5
20
Until used
5
Until used
(excess over tax liability)
10
10
0

3
3
0
0
0
0
0
No

5
No
5
5
10
5
10
5
5
5
0
4
10
4
5
0
7
0
0
0
0
20
10

0
No
3
0
0
3
3
0
3
3
0
0
0
0
0
0
0
0
0
0
0
0
0

5
7
5
5
0
10
No
0
5
5
10
4
0
0
4
0
Unlimited
4
4

0
0
0
0
0
0
No
0
0
0
3 (Distressed)
0
0
0
0
0
0
0
0

0
0
0

N/A

F E D E R A L R E S E R V E B A N K O F S T. LO U I S R E G I O N A L E C O N O M I C D E V E LO P M E N T

Rothstein and Wineinger

Cap type
Annual
(100% of federal LIHTC)
None
None
Annual, flexible
None
10-year cumulative
Annual
Total of pool
Annual
None
Annual
Total of pool
Total of pool
Annual
Annual
Cumulative
Annual
Annual
Annual
Annual
Annual
None
Annual
Cumulative
Annual
Annual
None
None
None
Total of pool
Total of pool
Annual
None
Cumulative
None
Annual
None
Annual
Annual
Cumulative
Total of pool
Annual
Annual
Annual
Cumulative
Appropriation
None
None
None
Appropriation
None
None
Annual

F E D E R A L R E S E R V E B A N K O F S T. LO U I S R E G I O N A L E C O N O M I C D E V E LO P M E N T

Cap amount

Federal
deductions
Yes

10,000,000

18,000,000
11,000,000
15,000,000

16,000,000
6,000,000
20,000,000
6,000,000
10,000,000
6,000,000
6,000,000
8,000,000
2,000,000
50,000,000
2,000,000
500,000
0

4,000,000
6,000,000
None
500,000
None
1,500,000
12,000,000
9,000,000
4,000,000
10,000,000
6,000,000
13,000,000
0

V O LU M E 3 , N U M B E R 2

Yes
No
No
No
None
No
None
Yes
No
No
None
None
No
No
No
No
No
No
No
No
No
N/A
No
None
No
No
N/A
No
None
None
No
No
No
No
No
No
No
No
No
None
No
No
N/A
No
None
N/A
No
No
No
Yes
No
Yes

2007

Rothstein and Wineinger

APPENDIX, cont’d
All Tax Programs, Sorted by Amount Issued, FY 2005
Fiscal year 2005
Name

Authorized

Issued

Redeemed

Total
outstanding

Missouri Low Income Housing
Historic Preservation TCP
Enterprise Zone
Infrastructure TCP
Brownfield Redevelopment Program (remediation)
Certified Capital Companies (CAPCO) Program
Neighborhood Assistance
Missouri Health Insurance Pool
Affordable Housing Assistance TCP
New and Expanded Business Facility Credit
BUILD (Missouri Business Use Incentives
for Large-Scale Development)
Examination Fee Tax Credits (exam)
Missouri Property and Casualty Guaranty Association
Neighborhood Preservation Tax Credit
Youth Opportunities
New Enterprise Creation
New Generation Cooperative Incentive
Transportation Development
Development Tax Credit
Agricultural Product Utilization Contributor
Rebuilding Communities
Brownfield “Jobs and Investment”
Maternity Home
Bond Enhancement/Bond Guarantee
Shelter for Victims of Domestic Violence
Small Business Incubator
Wine and Grape Production
Charcoal Producers
Small Business Guaranty Fees/Loan Guarantee Fee
Examination Fee Tax Credits (valuation)
Examination Fee Tax Credits (registration)
Family Development Account
Brownfield (demolition)
Community Bank Investment/Community
Development Corporation
Development Reserve
Dry Fire Hydrant
Export Finance
Film Production
Missouri Quality Jobs
Missouri Business Modernization and Technology
Missouri Life and Health Guaranty Association
New Enhanced Enterprise Zone
Qualified Research Expense/Research
Skills Development Credit
Small Business Investment/Capital
Sponsorship and Mentoring Program
Shared Care
Bank Franchise Tax
Bank Tax Credit for S Corporation Shareholders
Cellulose Casings
Disabled Access
Processed Wood Energy
Special Needs Adoption

183,106,160
116,566,180
39,066,023
10,398,000
15,515,319
14,000,000
16,000,000
10,015,203
11,000,000
8,779,797
7,842,167

83,477,412
80,213,374
39,066,023
28,964,274
14,808,297
14,000,000
11,263,385
10,015,203
9,133,829
8,779,797
8,419,707

65,392,601
74,532,355
25,294,754
25,953,799
10,627,870
13,429,309
9,286,880
6,121,053
7,702,860
4,546,330
3,770,557

156,095,405
61,343,552
0
37,308,093
4,180,427
45,440,744
14,000,000
9,066,945
15,000,000
3,085,774
4,649,150

7,576,530
7,227,710
13,609,190
6,438,159
0
6,000,000
6,682,249
5,591,000
6,000,000
6,970,463
406,273
953,987
870,275
2,000,000
500,000
313,683
146,606
103,591
35,000
24,430
780,000
0
0

7,576,530
7,227,710
6,784,310
4,476,005
4,212,752
3,915,000
3,226,568
2,866,000
2,081,343
1,736,701
1,646,927
953,987
870,275
648,618
361,913
313,683
146,606
103,591
35,000
24,430
7,625
0
0

2,650,135
5,965,556
8,641,503
3,211,185
2,504,561
3,334,935
3,545,219
2,487,152
1,639,540
1,694,006
1,726,687
743,635
594,034
515,034
246,807
179,323
70,151
11,224
12,000
24,430
12,875
0
2,021,628

5,703,974
1,998,707
6,824,880
5,000,000
7,155,490
6,749,210
7,000,000
6,762,264
4,593,008
1,955,245
0
N/A
276,241
1,251,032
384,983
0
575,597
0
23,000
0
4,000
0
1,250,000

0
0
0
1,500,000
0
0
0
0
0
0
0
854,443
N/A
N/A
N/A
N/A
N/A
3,348,890
N/A

0
0
0
0
0
0
0
0
0
0
0

0
17,228
0
322,079
0
164,894
0
0
1,626,864
0
109,050
0

0
65,000
0
752,705
0
337,341
0
0
5,823,761
0
969,475
0
0
0
N/A
0
N/A
N/A
N/A

V O LU M E 3 , N U M B E R 2

2007

N/A
N/A
N/A
N/A
N/A
N/A
N/A

2,543,523
941,460
382,540
56,761
3,700,285
2,578,354

F E D E R A L R E S E R V E B A N K O F S T. LO U I S R E G I O N A L E C O N O M I C D E V E LO P M E N T

Rothstein and Wineinger

Special status

Phasing out

Cap reached

Phased out

Cap reached
Cap is sum of both DOA tax credits
Expired
6,000,000 cap only for FY 2005-07
Cap is sum of both DOA tax credits

Cap remaining 48,812,870
Moved to DSS

Expired

Cap reached
Issued from account, no accounts opened
Expired
Issued from account, no accounts opened
Starts FY 2006, and cap set at 7,000,000
Cap reached
None since 1998
Starts FY 2006
Expired
Repealed after no one used it
Cap reached
Not funded

Outstanding N/A, DNR authorizes, Department of Revenue
redeems no cross check

F E D E R A L R E S E R V E B A N K O F S T. LO U I S R E G I O N A L E C O N O M I C D E V E LO P M E N T

V O LU M E 3 , N U M B E R 2

2007

Rothstein and Wineinger

APPENDIX, cont’d
All Tax Programs, Sorted by Amount Issued, FY 2005
Name

Missouri agency

Housing Development Commission
Department of Economic Development
Department of Economic Development
Development Finance Board
Department of Economic Development
Department of Economic Development
Department of Economic Development
Department of Insurance
Missouri Housing Development Commission
Department of Economic Development
Development Finance Board

V O LU M E 3 , N U M B E R 2

2007

Department of Insurance
Department of Insurance
Department of Economic Development
Department of Economic Development
Department of Economic Development
Agricultural and Small Business Development Authority
Department of Economic Development
Department of Economic Development
Agricultural and Small Business Development Authority
Department of Economic Development
Department of Economic Development
Department of Social Services
Department of Economic Development
Department of Public Safety
Department of Economic Development
Department of Economic Development
Department of Natural Resources
Department of Economic Development
Department of Insurance
Department of Insurance
Department of Economic Development
Department of Economic Development
Department of Economic Development
Development Finance Board
Department of Economic Development
Development Finance Board
Department of Economic Development
Department of Economic Development
Department of Economic Development
Department of Insurance
Department of Economic Development
Department of Economic Development
Department of Economic Development
Department of Economic Development
Department of Elementary and Secondary Education
Department of Health, Division of Senior Services
Department of Revenue
Department of Revenue
Department of Revenue
Department of Revenue
Department of Natural Resources
Department of Revenue

F E D E R A L R E S E R V E B A N K O F S T. LO U I S R E G I O N A L E C O N O M I C D E V E LO P M E N T

How Well Are the States
of the Eighth Federal Reserve District
Prepared for the Next Recession?
Gary A. Wagner and Erick M. Elder

Economic downturns often force state policymakers to enact sizable tax increases or spending cuts
to close budget shortfalls. In this paper the authors make use of a Markov-switching regression
model to empirically describe the expansions and contractions in the states of the Eighth Federal
Reserve District. They use the estimated parameters from the switching regressions to form probability distributions of the revenue shortfalls states are likely to encounter in future slowdowns.
This allows them to estimate the probability that each state’s projected fiscal-year-end balances
will be sufficient to offset the fiscal stress from a recession. (JEL E6, H7)
Federal Reserve Bank of St. Louis Regional Economic Development, 2007, 3(2), pp. 75-87.

conomic downturns often force state
policymakers into difficult financial positions because of the procyclical nature
of tax bases and the countercyclical
nature of government spending. Because of the
fiscal institutions that exist in many states (such as
balanced budget rules and borrowing restrictions),
policymakers’ options for mitigating periods of
fiscal stress are effectively limited to the use of
reserve balances, spending reductions, and tax
increases.
In this paper we follow Wagner and Elder
(2006) and use a Markov-switching regression
model to empirically describe the expansions and
contractions in states of the Eighth Federal Reserve
District.1 Using the estimated parameters, we form
probability distributions of the revenue shortfalls
states are likely to encounter during the next downturn. Based on fiscal-year-end projections for 2007,
1

The Eighth District states are Arkansas, Illinois, Indiana, Kentucky,
Mississippi, Missouri, and Tennessee.

we then estimate the probability that each state’s
reserve balances will be sufficient to offset the fiscal
stress from a recession. In other words, we estimate
the probability that each state in the Eighth Federal
Reserve District will be able to avoid spending
reductions or tax increases if a recession were to
begin in 2007.
In the following sections of the paper, we
review previous research, outline our empirical
methodology and findings, and offer concluding
remarks.

STATE FISCAL CRISES AND
“OPTIMAL” RAINY DAY FUNDS
An Overview of State Fiscal Crises and
Policy Options
Although many factors contribute to periods
of state fiscal pressure, the cyclical variability of
revenue streams is generally considered to be the

Gary A. Wagner is an associate professor of economics and Erick M. Elder is a professor of economics at the University of Arkansas at Little Rock.

F E D E R A L R E S E R V E B A N K O F S T. LO U I S R E G I O N A L E C O N O M I C D E V E LO P M E N T

V O LU M E 3 , N U M B E R 2

2007

Wagner and Elder

primary cause of fiscal crises (Holcombe and Sobel,
1997; Crain, 2003). Most state tax bases, including
the primary bases of retail sales and wage and
salary income, tend to be strongly procyclical.
This means that revenue growth from these sources
expands more rapidly than a state’s economy during expansions and contracts more severely during
downturns. Given the difficulty of forecasting
recessions, state policymakers often find themselves in situations in which revenue is insufficient
to match expenditure demands.
Policymakers normally rely on spending reductions, tax increases, and reserve balances accumulated during periods of revenue growth (or savings)
to help mitigate unexpected budget shortfalls.
Although the use of savings has expanded in recent
decades, with nearly all states having a formal
“rainy day” or budget stabilization fund to institutionalize the process, spending reductions and
tax increases constitute the majority of state fiscal
adjustments made during recessions. Following the
2001 national recession, for instance, the National
Association of State Budget Officers (NASBO)
reported that states increased taxes by nearly $18
billion during fiscal years 2002 though 2004 and
also reduced budgeted spending by nearly $30 billion over the same period.2 Moreover, states used
more than $9 billion in reserve balances during
this period to help close budget gaps, which is
considerably more than the $1 billion in savings
tapped during the 1990-91 downturn (Holcombe
and Sobel, 1997).
States in the Eighth Federal Reserve District
were not immune to the difficult times associated
with the 2001 recession. Table 1 reports the tax
and spending adjustments for each state in the
Eighth District for fiscal years 2002 through 2004.
Eighth District states increased taxes by nearly
$3.3 billion from 2002 through 2004, with all of the
increases occurring in fiscal years 2003 and 2004.
Three states—Illinois, Indiana, and Tennessee—
accounted for more than 95 percent of the tax

increases in the District. Arkansas, Kentucky, and
Missouri made modest tax adjustments, while
Mississippi was the only state in the District to
avoid tax increases during the slowdown.
On the expenditure side, Illinois, Indiana, and
Missouri were responsible for more than 80 percent
of the District’s after-budget adjustments. Most of
these spending cuts occurred in 2002 and 2003,
with only Illinois and Indiana making changes in
2004. Tennessee had the smallest expenditure
adjustments during the recession, with $64 million
in cuts in 2003.
Across the nation, states relied more heavily
on spending adjustments than tax increases
(roughly 60-40) during the 2001 downturn. In the
Eighth District, however, only Arkansas, Illinois,
and Indiana had a mix of tax increases and spending reductions that approximated the adjustments
in the rest of the country. Kentucky, Mississippi,
and Missouri relied almost exclusively on spending cuts, whereas Tennessee relied almost entirely
on tax increases.

“Optimal” Rainy Day Funds
To assess how “prepared” states may be in
future recessions, it is necessary to quantify the
fiscal stress that states are likely to experience during a downturn. Previous studies have addressed
this issue from the point of view of an “optimal”
rainy day fund: If a state typically experiences fiscal stress equal to, say, 12 percent of the budget
during a downturn, then savings equal to this
amount would be sufficient to eliminate the need
for spending cuts and tax increases throughout
the slowdown.3
Early attempts to estimate the fiscal stress that
states experience, such as Pollock and Suyderhoud
(1986), Sobel and Holcombe (1996), and Navin
and Navin (1997), did so by examining the cumulative deviation in the series of interest (either
revenues or revenues plus expenditures) from a
3

Given that the spending figures reported by NASBO do not account
for spending reductions that occurred at the time state budgets were
initially adopted, the aggregate reduction in state spending due to
the recession is potentially much larger than $30 billion. For a more
detailed analysis of the 2001 recession on state fiscal health, see
Garrett and Wagner (2004).

V O LU M E 3 , N U M B E R 2

2007

The use of savings to insure against unexpected budget shortfalls
has several advantages over the use of spending cuts and tax
increases. First, expenditure reductions and tax increases are unpopular among voters and therefore may be politically costly for policymakers. Second, the use of savings is an expansionary policy, whereas
expenditure reductions and tax increases are contractionary policies.
See Wagner and Elder (2005) for an overview on the effectiveness of
state rainy day funds and reserve balances.

F E D E R A L R E S E R V E B A N K O F S T. LO U I S R E G I O N A L E C O N O M I C D E V E LO P M E N T

Wagner and Elder

Table 1
Fiscal Adjustments of Eighth District States Due to 2001 Recession (in millions of current dollars)
2002

2003

2004

Total

–2.3

109.3

107

Tax adjustments
Arkansas
Illinois

Indiana

–5.9

370

828

1,002

1,198
998.1

Kentucky

Mississippi

Missouri
Tennessee

–24.5

72.5

933.2

Eighth District total

–32.7

2,377.7

Total all states

303.8

8,018

9,550

947.3

3,292.3
17,871.8

Spending adjustments (made after budget enactment)
Arkansas

171

244

Illinois

500

202

1,320

2,022

Indiana

468.7

345.7

874.4

Kentucky

231.5

90.1

321.6

Mississippi

150.6

47.8

198.4

Missouri

750

304.7

1,054.7

Tennessee
Eighth District total
Total all states

0
2,271.8
13,668

11,27.3
11,752

1,380
3,488

4,779.1
28,908

SOURCE: Various issues of Fiscal Survey of the States, National Association of State Budget Officers.

linear trend. For example, Sobel and Holcombe
(1996) summed the cumulative shortfalls in expenditures and revenues from their respective trends
from 1989 to 1992 and found that the average state
would have needed reserves equal to 30 percent of
expenditures to maintain trend expenditures and
revenues during the 1990-91 recession. Examining
individual states over a longer period, Pollock and
Suyderhoud (1986) and Navin and Navin (1997)
find that savings equal to 11 percent and 13 percent
of the budgets in Indiana and Ohio, respectively,
would be sufficient to offset a normal downturn.
Although Pollock and Suyderhoud (1986),
Sobel and Holcombe (1996), and Navin and Navin
(1997) provide only point estimates of state fiscal

stress, it is possible to form probability distributions of state expansions and contractions using a
linear-trend approach. This would allow one not
only to estimate a distribution of shortfalls that
states are likely to experience, conditional on past
recessions, but also to calculate how much states
would need to save during expansions to insure
against those possible shortfalls that a state may
experience. However, using a linear-trend approach
to form these distributions has a serious shortcoming because the parameters from a linear-trend
model are chosen to minimize the deviation from
trend rather than to best describe the distribution
of expansions and contractions in the data. In other
words, if expansions are defined as periods above

F E D E R A L R E S E R V E B A N K O F S T. LO U I S R E G I O N A L E C O N O M I C D E V E LO P M E N T

V O LU M E 3 , N U M B E R 2

2007

Wagner and Elder

trend and contractions are defined as periods below
trend, then there is no reason to believe that the
expansions and contractions identified from a
linear-trend model will correspond to actual business cycles.4
A recent paper by Wagner and Elder (2006)
attempts to overcome this limitation by identifying
state economic cycles using a Markov-switching
regression model. The model explicitly assumes
that a data series can be characterized by a series
of distinct regimes (such as expansions and contractions), and recent work by Li, Lin, and Hsiu-hua
(2005) suggests that the model performs very well
in accurately identifying business cycle turning
points. Estimation involves jointly determining
the parameter values describing each regime that
best fit the observed data. These parameters include
an estimate of the mean growth rate of each regime,
as well as the probabilities that a given observation
came from either an expansion or contraction
regime, which are known as transition probabilities.
Wagner and Elder (2006) demonstrate how the
estimated parameters from a Markov-switching
regression model may be used to construct empirical probability distributions of state expansions
and contractions. Forming these distributions for
each state, Wagner and Elder (2006) find that the
typical state’s expected revenue shortfall is between
2.9 and 3.5 percent of revenue when the shortfall
is measured relative to zero revenue growth, and
between 13 and 16 percent of revenue when the
shortfall is measured relative to the average rate of
revenue growth during expansions. In addition,
forming distributions of complete cycles that are
based on the uncertainty in expansions and contractions, Wagner and Elder find that the average
state should save between 0.5 and 2.5 percent of
revenue during each expansion period to accumulate reserve balances sufficient to weather the next
downturn without the need for tax increases or
spending cuts.
4

A related paper by Cornia and Nelson (2003) uses value-at-risk (VaR)
to model the maximum budget shortfall that a state can expect to
experience over a fixed time period. However, because VaR cannot
model a data series as being generated by two (or more) probability
distributions, such as expansions and contractions, it is impossible
to form distributions of savings rates using VaR.

V O LU M E 3 , N U M B E R 2

2007

MARKOV-SWITCHING MODEL
AND EIGHTH DISTRICT STATE
BUSINESS CYCLES
Markov-Switching Model
The basic idea underlying regime-switching
models is that many data series appear to be generated from multiple, distinct data-generating
processes. As Hamilton (1994) notes, structural
breaks or regime changes in a data series may be
triggered by a variety of factors, including economic
downturns, policy changes, and financial crises.
If the regime changes are assumed to be predictable and known a priori, then they may simply be
modeled using dummy variables. A more practical
assumption is that the occurrence of such regime
changes is unknown.
Econometric models featuring regime changes
were first studied by Quandt (1958). Goldfeld and
Quandt (1973) extended Quandt’s simple switching
model by allowing the data series to be generated
by multiple regime switches that were governed by
a Markov process so that the timing of the switches
became dependent on the current regime in effect.
Although regime-switching regressions have a
long history, they were not widely employed as a
means of modeling business cycle movements until
Hamilton (1989) extended Goldfeld and Quandt’s
model to include serially dependent data.
Hamilton’s model was a two-regime autoregression
applied to the growth rate in real U.S. gross national
product under the assumption that the regimes
exogenously switched according to an unobserved
Markov process. He found that not only did the
model best fit the data when it identified distinct
expansion and contraction regimes, but also that
the changes between regimes closely matched the
NBER recession turning points.
Although Markov-switching models have been
widely used to examine aggregate data, they have
only recently been applied to U.S. states. Owyang,
Piger, and Wall (2005a) were the first to make use
of the model and explored the extent to which
state-level business cycles track the aggregate economy. Using Crone’s (2002) monthly coincident
index of state-level labor-market activity, they find

F E D E R A L R E S E R V E B A N K O F S T. LO U I S R E G I O N A L E C O N O M I C D E V E LO P M E N T

Wagner and Elder

that, while state cycles generally follow the aggregate economy, individual states may shift into an
expansion or contraction before the national economy shifts, continue in an expansion as the national
economy contracts, and experience a downturn
that is not associated with an aggregate downturn.
Moreover, Owyang, Piger, and Wall (2005a) find
that expansion growth rates depend positively on
the state’s education and age composition, whereas
recession growth rates depend on the state’s industry mix.
Because state revenue data are published only
annually and also include discretionary changes to
both tax rates and tax bases, they are poorly suited
for isolating state-level business cycle movements.
As a result, we follow both Owyang, Piger, and
Wall (2005a) and Wagner and Elder (2006) by using
Crone’s (2002) monthly coincident index as our
measure of state-level economic activity over the
period 1979:09–2007:01. As Owyang, Piger, and
Wall (2005a) note, an advantage of Crone’s (2002)
index is that, unlike many state-level data series,
it exhibits distinct business cycle movements at a
high frequency. On the other hand, because the
index is constructed from only labor market variables, it is not as broad a measure of economic
activity as gross domestic or gross state product.
Denoting a state’s monthly growth rate in the
.
coincident index at time t as yt , the two-regime
Markov-switching model may be expressed as

y& t = µSt + εt ,
(1)

(

)

εt ~ N 0, σ ε2 ,
µSt = µ0 + µ1 St , µ1 > 0,

where µ denotes the mean growth rate and εt is the
error term at time t assumed to be normally distributed with variance σ ε2. The growth rate in (1)
is assumed to switch exogenously between two
regimes, and the switches are governed by an unobserved regime variable, St = {0,1}. When St = 0,
.
which we refer to as the low-growth regime, yt
follows a stationary AR(0) process and is assumed
to be generated by a normal distribution with a
mean of µ 0. When St switches from 0 to 1, which
.
we call the high-growth regime, yt is presumed to

have been generated from a normal distribution
with a mean equal to µ 0 + µ 1. In short, the data.
generating process for yt is a mixture of two normal
distributions having the same variance but potentially different means.
Although St is unobserved, its behavior is
restricted to evolve according to a first-order, twostate Markov chain with the following transition
matrix:
 P (St = 0|St −1 = 0 ) P (St = 1|St −1 = 0)
P=

 P (St = 0|St −1 = 1) P (St = 1|St −1 = 1) 
(2)
1 − PLL 
 PLL
=
,
PHH 
1 − PHH

where Pij is the transition probability of St = i,
given that St –1 = j. Hence, PHH is the probability
that economic activity is in the high-growth regime
in period t, conditional on having been in the highgrowth regime in period t –1. Placing restrictions
on the behavior of St allows one to estimate the
probability that economic activity is in an expansion (or contraction) regime in each time period,
despite the fact that the underlying regime is
assumed to be latent and unobservable.5

Markov-Switching Parameter Estimates
for Eighth District States
The parameter estimates from each Eighth
District state’s Markov-switching regression are
presented in Table 2. The expected duration of
each regime, which is discussed in more detail
below, is also presented.
Given that we are updating the specification
used by Owyang, Piger, and Wall (2005a) and
Wagner and Elder (2006) with more available data,
5

We also follow Owyang, Piger, and Wall (2005a) and estimate the
models using the Bayesian Gibbs-sampling approach developed by
Kim and Nelson (1998). Our prior distributions were set equal to the
priors of Owyang, Piger, and Wall (2005a), and the joint posterior
distributions were simulated using 10,000 replications with an additional 2,000 burn-in replications. The mean parameters (µ0 and µ1)
are assumed to be normally distributed with means of 1 and –1,
respectively, and a covariance matrix that is equal to the identity
matrix. The transition probabilities, PHH and PLL , have prior beta
distributions given by β (9,1) and β (8,2), implying means of 0.9 and
0.8, respectively. For a more detailed description of the estimation
procedure, see Kim and Nelson (1998) and Owyang, Piger, and Wall
(2005a). We acknowledge use of the computer routines described in
Kim and Nelson (1999).

F E D E R A L R E S E R V E B A N K O F S T. LO U I S R E G I O N A L E C O N O M I C D E V E LO P M E N T

V O LU M E 3 , N U M B E R 2

2007

Wagner and Elder

Table 2
Markov-Switching Parameter Estimates
µ̂ 0 + µ̂ 1

µ̂ 0

P̂HH

P̂LL

E[tH]

E[tL]

E[tH] + E[tL]

Arkansas

0.334*

–0.082*

0.982

0.940

55.55

16.71

72.27

Illinois

0.362*

–0.239*

0.978

0.950

45.72

20.08

65.80

Indiana

0.337*

–0.386*

0.981

0.910

54.90

11.22

66.12

Kentucky

0.350*

–0.194*

0.980

0.936

51.82

15.72

67.54

Mississippi

0.341*

–0.075

0.975

0.930

41.19

14.45

55.65

Missouri

0.362*

–0.216*

0.980

0.941

50.48

17.12

67.61

Tennessee

0.389*

–0.057

0.979

0.936

49.40

15.79

65.20

Median

0.351

–0.194

0.980

0.937

50.48

15.79

66.28

NOTE: The reported parameters are the means of the posterior distributions; *denotes that the 90 percent highest posterior density
interval does not contain zero; and E [tH] and E [tL] denote the expected duration of expansions and contractions, respectively.

our parameter estimates are very similar to the
parameter estimates of these studies. The median
expansion and contraction (monthly) growth
rates across the Eighth District states are 0.351
and –0.194, respectively. However, as shown in
the first column of Table 2, expansion growth
rates range from a high of 0.389 in Tennessee to a
low of 0.334 in Arkansas; recession growth rates,
shown in column 2, vary from a low of –0.386 in
Indiana to a high of –0.057 in Tennessee. There is
considerably more variation in the average recession growth rates across the District than in the
average expansion growth rates. In fact, three states
in the District, Arkansas, Mississippi, and
Tennessee, have average recession growth rates
less than –0.10, whereas three states have recession
growth rates in excess (more negative) of –0.20.6
The estimated transition probabilities for each
state, PHH and PLL, demonstrate the persistence in
each regime. Given an expansion in period t –1,
our estimates indicate that the median state in the
District has a 0.980 probability of expanding in
period t. Similarly, the probability is 0.937 that a
contraction in period t –1 will be followed by a
6

See Owyang, Piger, and Wall (2005b) for a much more detailed
analysis of the expansion growth rates, contraction growth rates,
and regime turning points in Eighth District states.

V O LU M E 3 , N U M B E R 2

2007

contraction in period t. As Hamilton (1994) shows,
the expected duration of regime j may be computed
as E[tj ] = 共1 – Pjj 兲–1 for j = H,L. (These figures are
reported for each state in Table 2.) Although cycle
durations vary noticeably, the median Eighth
District state can expect to experience 50 months
of expansion (denoted E[tH ] in Table 2), followed
by nearly 16 months of contraction (denoted E[tL ]
in Table 2), resulting in an expected business cycle
of 66 months. The state with the longest expected
expansion in the District, Arkansas, will on average
experience continuous growth for one year longer
than the state with the shortest expected expansion,
Mississippi. Similarly, Illinois has the longest
expected contraction in the District, at just over 20
months, which is 9 months longer than Indiana’s
shortest expected contraction.7 Overall, the Markovswitching model identifies nearly 77 percent of
the observations as expansions, suggesting that
the normal regime for all states in the District is
one of growth.
7

Although we do not report the estimated probability that a state is
in an expansion at a given point in time (St ), it is generally very clear
whether a state is currently in an expansion or contraction regime.
Owyang , Piger, and Wall (2005b) examine these probabilities in
detail for Eighth District states, and the interested reader is referred
to their paper.

F E D E R A L R E S E R V E B A N K O F S T. LO U I S R E G I O N A L E C O N O M I C D E V E LO P M E N T

Wagner and Elder

EMPIRICAL DISTRIBUTIONS OF
STATE SHORTFALLS AND SAVINGS
RULES
Probability Distributions of State
Revenue Shortfalls
The estimated parameters of the Markovswitching regressions provide measures of the
duration and amplitude of state economic cycles.
In this section of the paper we follow Wagner and
Elder (2006) and use these parameters to form
probability distributions of state revenue shortfalls.
Although states may experience fiscal pressure
on both the expenditure and revenue sides of the
budget, Kusko and Rubin (1993) note that revenue
streams are far more sensitive to business cycle
movements than expenditures. We therefore focus
only on state revenue, implying that our estimates
should be interpreted as the minimum fiscal
stress that states are likely to experience during a
downturn.
In addition, given that state tax bases tend to be
more volatile than economic activity, we define ϕ
as the elasticity of a state’s total revenue growth with
respect to the growth in the state’s economy. Each
state’s high- and low-regime growth rates in revenue may therefore be expressed as gH = ϕ 共 µ̂ 0 + µ̂ 1兲
and gL = ϕµ̂ 0, respectively. The use of an elasticity
allows us to alter the degree of revenue variability
that is due to a state’s particular tax mix. In short,
we assume that changes in state-level revenue
growth mimic changes in the state’s economic
activity (i.e., revenue growth has the same transition probabilities as overall economic activity), but
permit revenue to be more volatile than overall
economic activity. To explore the sensitivity of
our estimates, we use three reasonable revenue
elasticities (1.2, 1.5, and 1.8) for each state.
Given that PLL is the probability that a contraction in period t–1 will be followed by a contraction
in period t, the probability that a downturn will
persist exactly tL periods is given by

exactly that length will occur, the values may then
be used to form a cumulative probability distribution. The distributions may then be used to determine how much states need to save to achieve a
given level of certainty. In addition, for a given
level of savings, the distributions can also be used
to estimate the probability that this level of savings
will be sufficient in a future downturn.
To calculate the shortfalls, we set revenue
growth at a monthly rate of gH during each expansion period and at a rate of gL during each contraction period. Shortfalls are measured relative to an
amplitude parameter, γ , that is nothing more than
the target monthly growth rate in revenue during
each recession period. Although γ may take on
any value, we calculate shortfalls using both γ = 0
and γ = gH . Setting γ = gH measures the shortfall
relative to the average expansion growth rate in
revenue, whereas setting γ = 0 measures the shortfall relative to zero revenue growth. In other words,
γ = gH will generate the level of savings required
to maintain the average expansion growth rate in
revenue throughout a downturn, while γ = 0 yields
the level of savings needed to sustain a constant
level of revenue (or zero growth rate) throughout a
slowdown.
If an expansion lasts tH periods, then the state’s
level of revenue will be equal to R0共1 + gH 兲tH, where
gH is the (per-period) expansion growth rate in
revenue and R0 is the initial level of revenue.
Assuming a contraction begins and the (per-period)
growth rate in revenue switches to gL, the level of
revenue will then be equal to R0共1 + gH 兲tH 共1 + gL 兲
after the first low-growth period and the total revenue shortfall will be equal to R0共1 + gH 兲tH [共1 + γ 兲
– 共1 + gL 兲]. Relative to revenue in the previous
expansion period, which effectively measures the
shortfall as a percentage of revenue, the shortfall
may be written as 共1 + γ 兲 – 共1 + gL 兲. Hence, for a
recession lasting tL periods, the total revenue shortfall expressed as a share of revenue will be
tL

PL (tL ) =

tL −1
PLL

−

tL
PLL
.

Therefore, if one computes (i) each state’s revenue
shortfall for a contraction lasting tL = 1, 2,…, ⬁
periods and (ii) the probability that a recession of

(3)

ς ( t L ) = ∑ (1 + γ ) − ∑ (1 + g L ) .
i =1

i =1

For each state in the Eighth District, we assume
that a recession may persist for a maximum of 30
years and construct shortfall distributions using

F E D E R A L R E S E R V E B A N K O F S T. LO U I S R E G I O N A L E C O N O M I C D E V E LO P M E N T

V O LU M E 3 , N U M B E R 2

2007

Wagner and Elder

Figure 1
Cumulative Density Functions of Revenue Shortfalls
(Elasticity = 1.2, gH = 0.351, gL = –0.194, PHH = 0.980, and PLL = 0.937)
Cumulative Probability
1.00

0.75

Constant-Revenue Shortfall
Expansion-Revenue Shortfall

0.50
Expected value = 4.72%

Expected value = 14.15%

0.25

0.00
0

Shortfall (percent of revenue)

both γ = 0 and γ = gH , which we refer to as a “constant-revenue shortfall” and “expansion-revenue
shortfall,” respectively. Figure 1 illustrates sample
cumulative density functions for both shortfalls
using the median parameter estimates in Table 2
and a revenue elasticity of 1.2.
The expected value of the constant-revenue
shortfall is 4.72 percent of revenue, whereas the
mean of the expansion-revenue shortfall is 14.15
percent. The expected shortfall values are reported
for each District state in Table 3 using revenue
elasticities of 1.2, 1.5, and 1.8.8
Depending on the revenue elasticity, the
median state in the Eighth District can expect to
encounter a constant-revenue shortfall of between
4.6 and 6.8 percent of revenue during a given
recession. Because the magnitude of a constantrevenue shortfall is a function of both the state’s
average recession growth and the expected duration of a recession, the expected constant-revenue
8

The shortfall distributions were constructed by varying the number
of periods (length of recessions) and restricting revenue to grow at
its estimated average rate each period, gL . Wagner and Elder (2006)
explored an alternative approach by varying both the length of
recessions and per-period growth rate in revenue. They find that
both approaches produced similar results.

V O LU M E 3 , N U M B E R 2

2007

shortfall in Illinois is much larger than the shortfall
for any other state in the Eighth District. In contrast,
the moderate recession growth rates in Arkansas,
Mississippi, and Tennessee imply that these states
could maintain a constant level of revenue during
a typical downturn with reserve balances of
approximately 3 percent of the budget.
The expansion-revenue shortfall estimates,
which depend on the duration of a recession plus
the difference between the average expansion and
average recession growth rates, are noticeably larger
than this 3 percent for every state in the District.
In fact, the expansion-revenue shortfalls exceed
10 percent of the state’s revenue for every elasticity
value in five of the District’s seven states and exceed
this threshold in all seven states if the elasticity is
1.5 or larger. With an elasticity of 1.5, for example,
the typical state in the Eighth District can expect a
budget shortfall between 15 and 17 percent of revenue during a “normal” downturn, but the estimates range from a high of nearly 32 percent in
Illinois to a low of just over 11 percent in Indiana
and Mississippi.
Although the expansion-revenue shortfall
estimates are quite large, it is important to recall

F E D E R A L R E S E R V E B A N K O F S T. LO U I S R E G I O N A L E C O N O M I C D E V E LO P M E N T

Wagner and Elder

Table 3
Expected Revenue Shortfalls for Eighth District States (percentage of revenue)
Expansion-revenue shortfall (γ = gH )

Constant-revenue shortfall (γ = 0)
1.2

Revenue elasticity
1.5

1.8

1.2

Revenue elasticity
1.5

1.8

2.26

2.81

3.36

12.24

15.50

18.86

Illinois

9.14

11.28

13.37

25.09

31.68

38.43

Indiana

4.64

5.74

6.81

9.07

11.33

13.60

Arkansas

Kentucky

4.64

5.75

6.84

13.89

17.51

21.19

Mississippi

1.56

1.95

2.33

9.12

11.54

14.02

Missouri

6.08

7.52

8.94

17.50

22.08

26.76

Tennessee

1.43

1.78

2.13

11.86

15.07

18.38

Mean

4.25

5.26

6.25

14.11

17.82

21.61

Median

4.64

5.74

6.81

12.24

15.50

18.86

Maximum

9.14

11.28

13.37

25.09

31.68

38.43

Minimum

1.43

1.78

2.13

9.07

11.33

13.60

Table 4
Projected Fiscal Standing of Eighth District States

Arkansas

Probability FY 2007 balance is sufficient

Projected FY 2007-end balance
(share of revenue)

Constant-revenue shortfall

Expansion-revenue shortfall

5.33

0.853

0.552

Illinois

3.34

0.512

0.336

Indiana

6.10

0.756

0.645

Kentucky

7.85

0.808

0.603

Mississippi

4.62

0.894

0.581

Missouri

8.55

0.767

0.573

Tennessee

4.77

0.958

0.652

Mean

5.79

0.793

0.563

Median

5.33

0.808

0.581

Maximum

8.55

0.958

0.652

Minimum

3.34

0.512

0.336

NOTE: FY is fiscal year. Projected FY 2007 balances were obtained from individual state budgets.

F E D E R A L R E S E R V E B A N K O F S T. LO U I S R E G I O N A L E C O N O M I C D E V E LO P M E N T

V O LU M E 3 , N U M B E R 2

2007

Wagner and Elder

that measuring shortfalls relative to expansions
provides a useful upper bound for analyzing state
fiscal stress. If policymakers were to save according
to their state’s expansion-revenue shortfalls, then,
on average, the state would be able to avoid all
expenditure reductions and tax increases for the
duration of a slowdown.
An alternative perspective is to use each state’s
shortfall distributions to determine the probability
that the state’s current level of savings is sufficient
to offset a given shortfall. We do this for each state
using the state’s fiscal-year-end projections for
2007. The results are provided below in Table 4.
We looked at Eighth District states’ fiscal year
2007 projections and found that the average state
will have reserve balances in excess of 5 percent.
Moreover, the probability that any state’s savings
is sufficient to fully offset a downturn is, for the
most part, quite high. For example, our estimates
indicate that the probability is roughly 0.80 that
District states will be able to maintain a constant
level of revenue for the duration of a recession
without relying on tax increases or spending reductions. The estimates range from a high of over 95
percent in Tennessee to a low of just over 50 percent in Illinois. In terms of hedging an expansionrevenue shortfall, we find that there is nearly a 60
percent chance that District states will be able to
continue the average expansion growth rate in revenue during a slowdown without the use of other
fiscal adjustments. In fact, all of the District states
except Illinois have at least a 55 percent chance of
escaping major fiscal adjustments relative to baseline conditions should a slowdown begin in 2007.

Probability Distributions of State
Savings Rates
Because the parameters from the Markovswitching regressions describe the distribution of
both expansions and contractions, Wagner and
Elder (2006) estimate savings rates that are based
on all of the possible expansion-contraction combinations that may occur in a given state. In this
section of the paper we show how these savings
rates are obtained, which essentially provide a
benchmark for policymakers interested in insuring
against fiscal shocks.

V O LU M E 3 , N U M B E R 2

2007

Assuming that policymakers save a fraction of
revenue (s) during each period of an expansion,
revenue will be equal to R0共1 + gH 兲 and savings
will be R0s 共1 + gH 兲 and at the end of one period.
Following tH periods of expansion, the state’s accumulated savings, compounding at a rate r, will be
given by
tH

t − j +1

R 0 s ∑ (1 + r ) H

(4)

(1 + g H ) j .

j =1

If revenue growth switches from an expansion
to a contraction, then the revenue shortfall in the
first low-growth period will be the difference
between actual revenue, R0共1 + gH 兲tH 共1 + gL 兲, and
the target level of revenue, R0共1 + gH 兲tH 共1 + γ 兲共1 – s兲,
where γ denotes the amplitude parameter specifying
the target revenue growth rate during contractions.
Prohibiting states from saving during contraction periods, the revenue shortfall in just the first
contraction period will be equal to
t
t
R0 (1 + g H ) H (1 + γ ) (1 − s ) − (1 + g H ) H (1 + g L ) .



For a recession lasting tL periods, the state’s
cumulative revenue shortfall may be written as
tL
tL
t 
i
i
(5) R0 (1 + g H ) H (1 − s ) ∑ (1 + γ ) − ∑ (1 + g L )  .
i =1
i =1



Because equation (5) is the state’s revenue
shortfall from a downturn lasting tL periods and
equation (4) is the state’s savings from an expansion
lasting tH periods, setting equations (4) and (5)
equal to one another and solving for s yields the
faction of current revenue the state must save during each of the tH expansion periods to accumulate
savings equal to the revenue shortfall. This savings
rate is given by

s (t H ,t L ) =
 tL



(1 + g H )t  ∑ (1 + γ )i − (1 + g L )i  
H

(6)

 i =1

t − j +1

∑ (1 + r ) H
j =1



(1 + g H ) + (1 + g H ) H

∑ (1 + γ )

.
i

i =1

The savings rate given by equation (6) applies
to a specific expansion length and a specific con-

F E D E R A L R E S E R V E B A N K O F S T. LO U I S R E G I O N A L E C O N O M I C D E V E LO P M E N T

Wagner and Elder

Table 5
Expected Savings Rates for Eighth District States (percentage of revenue)
Expansion-revenue shortfall (γ = gH )

Constant-revenue shortfall (γ = 0)

Arkansas

1.2

Revenue elasticity
1.5

1.8

1.2

Revenue elasticity
1.5

1.8

0.39

0.49

0.59

2.12

2.71

3.33

Illinois

1.52

1.89

2.27

4.28

5.51

6.80

Indiana

1.03

1.29

1.55

2.08

2.65

3.23

Kentucky

0.86

1.08

1.30

2.61

3.35

4.11

Mississippi

0.33

0.41

0.50

1.92

2.45

3.00

Missouri

1.08

1.35

1.63

3.17

4.07

5.01

Tennessee

0.27

0.34

0.41

2.22

2.84

3.49

Mean

0.78

0.98

1.18

2.63

3.37

4.14

Median

0.86

1.08

1.30

2.22

2.84

3.49

Maximum

1.52

1.89

2.27

4.28

5.51

6.80

Minimum

0.27

0.34

0.41

1.92

2.45

3.00

traction length. If we assume that the length of
expansions (tH ) and the length of recessions (tL )
are independent, then the probability that an
expansion persisting tH periods will be followed
by a recession lasting tL periods can be computed
as PH 共tH 兲 × PL共tL 兲, where

( )

t −1

Pj t j = Pjj j

− Pjj j for j = H , L .

Assuming that both expansions and recessions
last for a maximum of 30 years (or 360 months),
we form savings rate distributions from the
129,600 possible expansion-contraction combinations. The expected savings rate for constantrevenue shortfalls (γ = 0) and expansion-revenue
shortfalls (γ = gH ) are presented in Table 5 using
revenue elasticities of 1.2, 1.5, and 1.8.
The savings rates show the fraction of total
revenue that a state must save during each expansion period in order to accumulate sufficient savings
to offset a “normal” fiscal cycle. Assuming a revenue elasticity of 1.5 for instance, if the median
state in the Eighth District saved 1.08 percent of
revenue each expansion period, then the state
would be able to maintain a constant level of revenue during a typical downturn and would have

zero reserve balances when the downturn ends. If
the objective is to maintain the average expansion
rate of revenue growth throughout a recession, then
the median state would need savings equal to 2.84
percent during each expansion period.9 For a given
target growth rate in revenue (γ ), the closer policymakers are to achieving the state’s expected savings
rate, the more likely it is that the state will be able
to avoid expenditure reductions and tax increases
in the next recession.

CONCLUSION
Slowdowns in economic activity place tremendous strain on state budgets and frequently force
policymakers to enact sizable spending cuts and
tax increases to close budget shortfalls. Following
Wagner and Elder (2006), this paper uses a basic
Markov-switching regression model to form empirical distributions of the monthly revenue cycles in
9

The Markov-switching model assumes that the underlying regime
is not observable. However, a reasonable approach to implementing
a savings rule in practice would be to rely on the estimated probability that a state is in an expansion at a given point in time, P共St = 1兲.
The simplest possible rule for the state to follow would be to assume
that the economy is expanding (and therefore save) if P共St = 1兲 > 0.5
and save nothing otherwise.

F E D E R A L R E S E R V E B A N K O F S T. LO U I S R E G I O N A L E C O N O M I C D E V E LO P M E N T

V O LU M E 3 , N U M B E R 2

2007

Wagner and Elder

Eighth District states. We then estimate the revenue
shortfalls that District states are likely to experience
in a future recession using multiple fiscal objectives.
We find, for instance, that the typical state in
the District needs reserve balances equal to 4 to 6
percent of revenue to maintain a constant level of
revenue during the next recession without relying
on spending cuts and tax increases. If the goal is
to preserve the average expansion growth rate in
revenue during a slowdown, while simultaneously
avoiding spending or tax adjustments, then the
average state would need to have savings equal to
15 to 17 percent of revenue before the start of the
next recession. Based on fiscal-year-end projections
for 2007, we find that all of the Eighth District states
except Illinois have a greater than 50 percent
chance of avoiding major fiscal adjustments should
a slowdown begin in 2007.

REFERENCES
Cornia, Gary C. and Nelson, Ray D. “Rainy Day Funds
and Value at Risk.” State Tax Notes, 2003, 29(3),
pp. 563-67.
Crain, W. Mark. Volatile States: Institutions, Policy,
and the Performance of American State Economies.
Ann Arbor: University of Michigan Press, 2003.
Crone, Theodore. “Consistent Economic Indexes for
the 50 States.” Federal Reserve Bank of Philadelphia
Working Paper No. 02-7, 2002.
Garrett, Thomas A. and Wagner, Gary A. “State
Government Finances: World War II to the Current
Crisis.” Federal Reserve Bank of St. Louis Review,
March/April 2004, 86(2), pp. 9-25.
Goldfeld, Stephen M. and Quandt, Richard E. “A
Markov Model for Switching Regressions.” Journal
of Econometrics, 1973, 1(1), pp. 3-15.
Hamilton, James D. “A New Approach to the
Economic Analysis of Nonstationary Time Series
and the Business Cycle.” Econometrica, March 1989,
57(2), pp. 357-84.
Hamilton, James D. Time Series Analysis. Princeton:
Princeton University Press, 1994.

V O LU M E 3 , N U M B E R 2

2007

Holcombe, Randall G. and Sobel, Russell S. Growth
and Variability in State Tax Revenue: An Anatomy
of State Fiscal Crises. Westport, CT: Greenwood
Press, 1997.
Kim, Chang-Jin and Nelson, Charles R. “Business
Cycle Turning Points, A New Coincident Index, and
Tests of Duration Dependence Based on a Dynamic
Factor Model with Regime-Switching.” Review of
Economics and Statistics, May 1998, 80(2),
pp. 188-201.
Kim, Chang-Jin and Nelson, Charles R. State-Space
Models with Regime Switching: Classical and Gibbs
Sampling Approaches with Applications.
Cambridge, MA: MIT Press, 1999.
Kusko, Andrea L. and Rubin, Laura S. “Measuring the
Aggregate High-Employment Budget for State and
Local Governments.” National Tax Journal,
December 1993, 46(4), pp. 411-23.
Li, Ming-Yuan Leon; Lin, Hsiou-Wei William and
Hsiu-hua, Rau. “The Performance of the MarkovSwitching Model on Business Cycle Identification
Revisited.” Applied Economics Letters, June 2005,
12(8), pp. 513-20.
Navin, John and Navin, Leo. “The Optimal Size of
Countercyclical Budget Stabilization Funds: A Case
Study of Ohio.” Public Budgeting and Finance,
Summer 1997, 17(2), pp. 114-27.
Owyang, Michael; Piger, Jeremy and Wall, Howard.
“Business Cycle Phases in U.S. States.” Review of
Economics and Statistics, November 2005a, 87(4),
pp. 604-16.
Owyang, Michael T.; Piger, Jeremy M. and Wall,
Howard J. “The 2001 Recession and the States of the
Eighth Federal Reserve District.” Federal Reserve
Bank of St. Louis Regional Economic Development,
2005b, 1(1), pp. 3-16.
Pollock, Richard and Suyderhoud, Jack P. “The Role of
Rainy Day Funds in Achieving Fiscal Stability.”
National Tax Journal, December 1986, 39(4),
pp. 485-97.

F E D E R A L R E S E R V E B A N K O F S T. LO U I S R E G I O N A L E C O N O M I C D E V E LO P M E N T

Wagner and Elder

Quandt, Richard E. “The Estimation of the Parameters
of a Linear Regression System Obeying Two
Separate Regimes.” Journal of the American
Statistical Association, December 1958, 53(284),
pp. 873-80.
Sobel, Russell S. and Holcombe, Randall G. “The
Impact of State Rainy Day Funds in Easing State
Fiscal Crises During the 1990-1991 Recession.”
Public Budgeting and Finance, Fall 1996, 16(3),
pp. 28-48.
Wagner, Gary A. and Elder, Erick M. “The Role of
Budget Stabilization Funds in Smoothing Government
Expenditures Over the Business Cycle.” Public
Finance Review, July 2005, 33(4), pp. 439-65.
Wagner, Gary A. and Elder, Erick M. “Revenue Cycles
and the Distribution of Shortfalls in U.S. States:
Implications for an ‘Optimal’ Rainy Day Fund.”
Working paper, University of Arkansas at Little
Rock, 2006 (forthcoming in National Tax Journal).

F E D E R A L R E S E R V E B A N K O F S T. LO U I S R E G I O N A L E C O N O M I C D E V E LO P M E N T

V O LU M E 3 , N U M B E R 2

2007

The Economic Impact of Broadband Deployment
in Kentucky
David Shideler, Narine Badasyan, and Laura Taylor
Significant resources are being invested by government and the private sector in broadband
infrastructure to increase broadband deployment and use. With a unique dataset of broadband
availability (sorted by county), the authors assess whether broadband infrastructure has affected
the industrial competitiveness of Kentucky counties. Their results suggest that broadband availability increases employment growth in some industries but not others. (JEL H54, R11)
Federal Reserve Bank of St. Louis Regional Economic Development, 2007, 3(2), pp. 88-118.

s the “knowledge economy” continues
to transform our society, broadband
internet access is an essential component of infrastructure for economic
development. Broadband deployment and use is
expected to offer benefits to businesses and consumers as well as the public sector. Potential benefits of broadband usage to businesses include
productivity gains through e-commerce, integrated
supply chains, improved management (Williamson
et al., 2006), and increased productivity through
telecommuting. There are substantial foreseeable
benefits of residential broadband use, including
improved efficiency of retailing, reductions in
commuting, increased variety of home entertainment, greater availability of health care, and
improved access to educational opportunities.
In addition, broadband facilitates the delivery of
e-government services and applications, bringing
the potential to significantly enhance government
communication with its constituents. Similarly,
broadband enables online community applications, which provide additional opportunities for
individuals to contribute to society, especially the
disabled. Crandall and Jackson (2001) projected

that these benefits would lead to a $500 billion
increase in U.S. gross domestic product by 2006.
In response to these perceived benefits of
broadband, Kentucky embarked on “Prescription
for Innovation”—a unique broadband deployment
and adoption plan that leverages state, federal, and
private investment to ensure statewide broadband
availability and significantly improve technology
adoption. ConnectKentucky, a public-private partnership, has the charge of realizing the four strategic goals of Prescription for Innovation:
• full broadband deployment by the end of
2007;
• increased use of computers and the Internet;
• the creation of a meaningful online presence
for every local community;
• the development of e-community leadership
teams to form business plans and identify
applications for business, local government,
education, health care, libraries, agriculture,
tourism, and local nongovernmental
organizations.
To date, ConnectKentucky has achieved and

David Shideler and Narine Badasyan are assistant professors of economics at Murray State University. Laura Taylor is vice president of research
at Connected Nation, Inc. The authors are grateful for the generous financial and technical support of ConnectKentucky, a nonprofit public-private
partnership charged with ensuring the deployment of broadband throughout the commonwealth of Kentucky and increasing its use. They are
specifically appreciative of Wes Kerr and Leslie Lyons for their assistance in accessing the data and for their insights, which have guided this
research. The authors also acknowledge the help of their graduate research assistant, David Jennings, in data collection and formatting.

V O LU M E 3 , N U M B E R 2

2007

F E D E R A L R E S E R V E B A N K O F S T. LO U I S R E G I O N A L E C O N O M I C D E V E LO P M E N T

Shideler, Badasyan, Taylor

exceeded expectations by realizing a deployment
rate of 94 percent of Kentucky households, with
deployment planned to cover the remaining 6 percent by the end of 2007; a 73 percent increase in
the use of broadband in the home since 2004; and
a 20 percent increase in household computer ownership since 2004.
ConnectKentucky has become a model for
broadband deployment and adoption for other
states throughout the nation because of these successes; moreover, a national nonprofit organization,
Connected Nation, Inc., has been created to replicate its model in other states.
Empirical evidence ex post of broadband investments by governments and the private sector is
sparse. Quantifying many of the benefits described
above requires extensive data collection, which is
costly and time consuming. Therefore, the existing
evidence has focused on economic impacts, measured in terms of employment growth or efficiency
gains from broadband adoption. For example, Lehr
et al. (2005) estimate the impacts of broadband
availability on a number of economic indicators
such as employment growth, wages, proportion of
establishments in information technology (IT), and
rental rates between 1998 and 2002. The study
concludes that the communities in which broadband became available by 1999 experienced more
rapid growth in employment, the number of businesses overall, and the number of businesses in ITintensive sectors. Lehr et al. also observed higher
market rates for rental housing in the communities
with broadband availability. In another study,
Crandall, Lehr, and Litan (2007), finds similar
results, though the scope of analysis is limited to
only employment and output.
The present study focuses exclusively on the
economic impact that broadband deployment has
had in Kentucky’s local communities. Although
most of the early studies relied upon projections of
forward linkages, this study will look at observed
changes in economic activity related to broadband
deployment, as did Lehr et al. (2005) and Crandall,
Lehr, and Litan (2007). A major difference between
the previous studies and the present one is in our
measure of broadband availability. The previous
studies use data from Federal Communication
Commission (FCC) Form 477 to measure broadband

data availability. In the case of Lehr et al. (2005),
Form 477 data identifies the number of broadband
providers with at least one subscriber in each zip
code. Crandall, Lehr, and Litan (2007) use the statelevel Form 477 data, which provides the number
of lines available (i.e., the number of subscribers)
in each state. Our measure utilizes county-level
data aggregated from ConnectKentucky’s GIS database of broadband service as measured at the point
of service. That is, ConnectKentucky uses proprietary infrastructure data from broadband providers
to determine in which geographic areas broadband
service is offered. Measuring broadband availability
this way is superior to the previous measures
because it provides a more accurate assessment
of where broadband is available; the zip code data
exaggerates broadband availability,1 while the
state-level data is too geographically aggregated to
identify variances in broadband coverage.
Following in the spirit of Lehr et al. (2005),
this study uses an economic growth framework to
determine the impact of broadband deployment
on economic activity in Kentucky’s counties. The
next section describes our data and methodology
in more detail and is followed by our estimation
results. We conclude with a discussion of extensions and policy implications of this work.

DATA AND METHODOLOGY
Identifying the impact of infrastructure poses
several challenges that make the analysis different
from that for other economic impacts. First, a typical economic impact analysis identifies the employment creation and related economic benefits
associated with the expansion in the local economy.
Infrastructure itself does not create sustained
employment, only temporary employment associated with construction or maintenance. Second,
standard economic impact analyses are based on
backward linkages. In a traditional impact analysis,
accounting for the backward linkages among firms
1

The U.S. Government Accountability Office reported that the FCC
Form 477 zip code data overstates broadband availability because
an entire zip code is reported as having broadband if at least one
subscriber is located there; this is a poor measure of availability,
particularly in rural areas where zip codes tend to be large geographic
areas.

F E D E R A L R E S E R V E B A N K O F S T. LO U I S R E G I O N A L E C O N O M I C D E V E LO P M E N T

V O LU M E 3 , N U M B E R 2

2007

Shideler, Badasyan, Taylor

is what allows researchers to identify how growth
in one industry will lead to growth in others. However, it is often difficult to predict how specific
types of infrastructure will be utilized by industries,
particularly previously unavailable infrastructure
like broadband. Since there is no prior history of
the use of broadband in a locale, there is no way
to predict which firms will or will not utilize the
infrastructure and how use of the infrastructure
will affect the firms’ production processes. Additionally, the presence of the new infrastructure may
also make the region attractive to new firms that
will relocate to take advantage of it. Given these
difficulties, economists often estimate the economic impacts of infrastructure using a modified
growth model. (See Rupasinga, Goetz, and
Freshwater, 2000, and Lehr et al., 2005, for additional applications of this model.)
The growth model is a methodology to predict
a region’s growth over time. Simply stated, this
model predicts the economic growth of a region
during one period based upon the level of economic
activity of some previous period plus any compounded growth that would be expected to occur
between the two periods. Mathematically, this
process can be expressed as

Yt = AYtα− i e r i,
where Y represents the economic level at time t,
A is a constant, α is a scaling parameter, and e ri is
the formula for compounded growth at rate r for i
periods. The critical element of this approach is
determining the right expected growth rate, r,
between the two periods. Because of the importance of this step, the growth rate, r, is often determined statistically using multivariate regression
analysis. By transforming this growth equation
using natural logarithms, assuming that A and α
equal 1 (which are standard assumptions when
empirically testing growth models), and defining
time periods in such a way as to make i = 1, equation (1) is derived:
(1)

ln (Yt Yt −1 ) = r = r ∗ + X β + ε .

This equation simply states that the economic
growth rate is a function of the optimal growth
rate, r* (which is constant), some explanatory
90

V O LU M E 3 , N U M B E R 2

2007

variables (X ), and an error term, ε (which has a
log-normal distribution). It is reasonable to assume
that the observed rate of economic growth will
differ from the optimum growth rate in any given
period simply because of unanticipated shocks to
the economy. This is the same theoretical model
used in Lehr et al. (2005). If one takes Y to represent industrial output, instead of aggregate economic activity, this framework can also be used to
analyze the effect of broadband infrastructure on
specific industries, where the change in industrial
output is estimated by various input factors and a
random error term.
Empirically, measuring growth and identifying
explanatory variables poses some challenges.
Because output is not measured at the local level
(like gross domestic or state product), researchers
often use employment, wages, or number of establishments data as a proxy for the size of the local
economy. Given our desire to look at total economic
impacts and industrial impacts from broadband,
we use the U.S. Census Bureau’s county business
patterns data series for 2003, 2004, and 2005 as
our economic data because it provides both total
and sectoral employment at the two-digit North
American Industrial Classification System (NAICS)
level. This dataset contains private, non-agriculture
production employment data measured as of the
week of March 12 annually. Using this data, we
compute the employment growth rates of the
periods 2003-04 and 2004-05 for each of the twodigit NAICS codes. A combination of zero employment levels in rural counties and suppressed data
due to Census disclosure rules led to missing values
in the data and reduced the number of observations
available for analysis in some industries. Table 1
provides summary statistics and the number of
observations for the employment growth rates.
Additional data concerns stemmed from the
very diverse nature of counties in Kentucky. For
example, in 2004, total employment across counties ranged from 131 employees to over 400,000
employees; the average county employment was
12,681, while the median county employment
was only 3,554. This wide distribution of values
becomes even more of a concern because our
analysis uses growth rates, such that a small
increase in employment, say 25 employees, could

F E D E R A L R E S E R V E B A N K O F S T. LO U I S R E G I O N A L E C O N O M I C D E V E LO P M E N T

Employment growth in educational services

emp61_45

Employment growth in accommodation and food services

Employment growth in other services

Employment growth in unclassified

emp72_45

emp81_45

emp99_45

Employment growth in health care and social assistance

Employment growth in administrative and support
and waste management and remediation services

emp56_45

Employment growth in arts, entertainment, and recreation

Employment growth in management of companies
and enterprises

emp55_45

emp62_45

Employment growth in professional, scientific,
and technical services

emp54_45

emp71_45

Employment growth in finance and insurance

Employment growth in real estate and rental and leasing

emp52_45

emp53_45

Employment growth in transportation and warehousing

Employment growth in information

emp48_49_45

emp51_45

116

111

118

111

112

119

100

Employment growth in wholesale trade

Employment growth in retail trade

emp42_45

emp44_45_45

112
103

Employment growth in construction

Employment growth in manufacturing

emp23_45

Employment growth in mining

Employment growth in utilities

emp21_45

emp22_45

120

emp31_33_45

Total employment growth

Employment growth in forestry, fishing, and hunting

emp00_45

emp11_45

Employment growth (2004-05)

Dependent
variable

Summary Statistics

Table 1A

–1.00

–0.22

–1.00

–0.27

Minimum

2.00

0.58

1.28

1.08

0.85

0.66

2.68

1.57

0.73

1.38

1.22

0.92

3.90

0.27

2.67

1.59

0.72

0.33

1.48

1.33

0.17

Maximum

–0.64

–0.09

–0.03

–0.17

0.01

–0.29

0.01

0.04

–0.24

–0.10

0.02

–0.44

–0.06

0.01

–0.08

–0.02

–0.03

–0.15

0.00

–0.16

0.01

Mean

0.82

0.25

0.34

0.44

0.15

0.53

0.73

0.54

0.47

0.45

0.32

0.52

0.65

0.08

0.46

0.26

0.37

0.57

0.69

0.05

Standard
deviation

Shideler, Badasyan, Taylor

V O LU M E 3 , N U M B E R 2

2007

V O LU M E 3 , N U M B E R 2

2007

Percent of persons 25+ with at least a bachelors degree

Unemployment rate, 2003

Rural dummy variable (1 = rural county)

bached00

unemp03

rural

120

sat104 squared

Miles of limited-access roads

sat1042

120

Percent of county area with broadband service
as of January 1, 2004

hwyaccess

sat104

Employment growth in unclassified

emp99_34

Other

Employment growth in accommodation and food services

Employment growth in other services

emp72_34

emp81_34

Employment growth in health care and social sssistance

Employment growth in educational services

emp61_34

Employment growth in arts, entertainment, and recreation

Employment growth in administrative and support
and waste management and remediation services

emp56_34

emp62_34

Employment growth in management of companies
and enterprises

emp55_34

emp71_34

Employment growth in real estate and rental and leasing

Employment growth in professional, scientific,
and technical services

emp53_34

emp54_34

Employment growth in information

Employment growth in finance and insurance

emp51_34

emp52_34

Employment growth in retail trade

Employment growth in transportation and warehousing

emp44_45_34

emp48_49_34

Employment growth in manufacturing

Employment growth in wholesale trade

emp31_33_34

emp42_34

Employment growth in utilities

Employment growth in construction

emp22_34

emp23_34

Employment growth in forestry, fishing, and hunting

Employment growth in mining

emp11_34

Total employment growth

emp00_34

emp21_34

Employment growth (2003-04)

Independent
variable

Table 1B

0.00

0.04

0.05

0.00

0.01

–1.00

–0.51

–1.00

–0.35

–1.00

–0.23

Minimum

1.00

0.12

0.36

303.50

1.00

3.67

0.46

0.58

0.67

0.56

0.58

4.25

0.94

5.46

2.90

1.05

2.15

1.00

1.06

1.00

1.42

1.03

0.03

1.14

1.50

0.40

Maximum

0.58

0.07

0.12

26.06

0.34

0.50

–0.28

–0.02

0.01

–0.08

0.00

–0.03

–0.04

0.00

0.04

0.01

0.00

–0.08

–0.03

0.01

–0.03

0.01

0.03

–0.15

–0.04

–0.19

0.01

Mean

0.50

0.02

0.06

39.97

0.35

0.30

0.45

0.21

0.31

0.12

0.18

0.50

0.16

0.57

0.45

0.20

0.31

0.37

0.13

0.27

0.21

0.29

0.12

0.25

0.21

0.07

Standard
deviation

Shideler, Badasyan, Taylor

F E D E R A L R E S E R V E B A N K O F S T. LO U I S R E G I O N A L E C O N O M I C D E V E LO P M E N T

Shideler, Badasyan, Taylor

lead to a 19 percent increase in employment in the
county with the least employment or a 0.00006 percent increase in the county with the highest employment. To ensure that no one observation heavily
influenced the values in our results, we identified
influential observations using “studentized”
residuals in the analysis of each industry. Influential observations were those for which the residual
value exceeded 2 in absolute value; these counties
were excluded from the analysis of that industry
to ensure that the results were representative of
counties across Kentucky.2 Several industries
showed signs of heteroskedasticity, according to
White’s general test. Generalized least-squares procedures were used to correct for the heteroskedastic
error term for the following industries: wholesale
trade; transportation and warehousing; information;
real estate and rental and leasing; and professional,
scientific, and technical services.3
Data for our independent variables came from
several sources. To measure broadband availability,
we computed the area of a county for which broadband service is available using ConnectKentucky’s
GIS inventory of Kentucky’s broadband deployment and service availability. The GIS inventory
provides a comprehensive view of broadband technologies, representing digital subscriber lines (DSL),
cable modem service, and fixed wireless networks,
measured at the point of service availability (i.e.,
at the location of infrastructure placement). Coverage areas were aggregated to the county level by
Census block groups, and then the ratio of the
coverage area to total area of the county was computed. This saturation rate was our measure of
broadband infrastructure as of January 2004.
Early studies of the economic impacts of broadband, based on forward-looking models, suggest a
range of potential benefits of broadband to businesses. This includes reduced costs and increased
productivity of the workforce as well as prospects
of expansion and growth, as businesses will no
longer be constrained in their local market. Because
2

Excluding outliers explains why the number of observations reported
in Table 1 might differ from the number of observations (n) reported
in the results tables.

Because ordinary least-squares estimators are still unbiased under
heteroskedasticity, generalized least-squares techniques correct the
standard errors of the parameter estimates, but they do not affect
the value of the estimates.

there is no history of the use of broadband in a
locale, the overall expected impact of broadband
on employment growth can be twofold. On one
hand, broadband can lead to job losses, but higher
wages, through increased labor productivity. On
the other hand, it can lead to job creation as a result
of longer-term productivity increases and/or as
businesses expand their markets and venture into
regional and international markets. The overall
effect will depend on the type of industry as well
as the length and scope of broadband adoption by
a particular business.
The saturation rate squared is also included
to study the returns to scale of broadband deployment. For instance, diminishing returns, captured
by a negative coefficient of saturation squared,
would indicate that, as broadband deployment
nears its maximum (100 percent coverage of area),
its marginal effect on employment growth diminishes. In other words, if the county is nearing 100
percent served, adding an additional unit of broadband infrastructure to unserved portions of that
county would provide smaller additional benefit
in terms of job growth. This phenomenon could
be related to the increasing necessity of broadband
within the economy; as broadband service within
a county becomes ubiquitous, it becomes expected
infrastructure, and thus other economic variables
become those factors that influence variations in
job growth. It is worth noting, however, that this
study does not account for multiple layers of broadband service; that is, we have focused solely on job
growth as it relates to broadband deployment at a
given point in time. This paper does not consider
the possibility that competition among broadband
service providers may also affect job growth, and
consideration of this idea is left for future research.
Our control variables were generated from
standard, secondary sources. Educational attainment, measured as the percent of the population
25 years and older with at least a college degree
in 2000, was provided by the Kentucky State Data
Center. Educational attainment is a proxy for human
capital stock; one expects that higher educational
attainment within a county will lead to more economic activity due to the availability of more productive human capital. The level of
nontechnological infrastructure was measured as
the number of miles of limited-access highway

F E D E R A L R E S E R V E B A N K O F S T. LO U I S R E G I O N A L E C O N O M I C D E V E LO P M E N T

V O LU M E 3 , N U M B E R 2

2007

Shideler, Badasyan, Taylor

miles; these data were computed using the majorroads shapefile from the Kentucky Transportation
Cabinet, available through the Kentucky
Geography Network. From the attribute table, we
summed the length of interstates and parkways in
miles, reported as DMI_LEN_MI, for DRAWCODE
value 5 (interstates and parkways) in each county,
denoted by CO_NUMBER. One would also expect
a positive relationship between highway access and
economic activity, since more accessibility should
reduce transportation and distribution costs to
firms. Other control variables include a rural
region dummy variable to control for differences
between urban and rural places (such as population density); ConnectKentucky has a four- category
classification system (1 = rural, 2 = small metropolitan, 3 = suburban, and 4 = metropolitan)4 that
was adopted for this project. This variable takes a
value of 1 for counties designated rural, and 0 otherwise. The county unemployment rate in 2003,
as reported in the Local Area Unemployment
Statistics series from the U.S. Bureau of Labor
Statistics, was included as a proxy for the amount
of available labor in the county. If this variable
were positively correlated with employment
growth, it suggests that the industry is labor intensive so that more available labor (typically less
skilled labor) leads to more employment. If this
variable is negatively correlated with employment
growth, it suggests that the industry is not labor
intensive and/or it is able to recruit its labor from
outside of the county. Lastly, a lagged version of
the dependent variable was included to capture
any other unique characteristics about the county
and/or the industry within the county.
To determine the impact that broadband infrastructure has had on Kentucky’s local economies,
we used multivariate regression. In our analysis,
we regress the employment growth for 2004-05 for
each of the 21 two-digit industrial codes as a function of broadband saturation, saturation squared,
highway access, percentage of the population over
25 with a college degree, employment growth
between 2003 and 2004, the unemployment rate
in 2003, and the rural dummy variable.
4

ConnectKentucky developed this system to more accurately reflect
the regional differences across Kentucky that are not evident when
one uses the U.S. Department of Agriculture’s rural-urban continuum.

V O LU M E 3 , N U M B E R 2

2007

RESULTS
This section presents regression analysis of the
impact of broadband deployment on employment
growth in 20 industrial sectors (using two-digit
NAICS codes and excluding public administration)
and total employment growth in Kentucky. The
results are presented in Tables 2 through 22.
To thoroughly understand the role of broadband in economic development, we conducted
our analysis using a series of models, similar to
the structure used by Lehr et al. (2005). That is to
say, we present the results of four models for each
industry to see how the influence of saturation and
its square changes as additional controls are introduced into the model. Our first model, then, contains only saturation and saturation squared as
explanatory variables. Our second model adds the
lagged employment-growth variable. The third
model contains all control variables except saturation and saturation squared. The final model, the
most complex, contains all of our independent
variables. We also report F-statistics to determine
the overall significance of the models.
The results of greatest interest relate to the
significance and magnitude of the broadband variables. The broadband deployment variable has a
positive and significant impact on total employment
(Table 2) as well as employment growth in the following industries: mining (Table 4), construction
(Table 6), information (Table 11), and administrative, support, and waste management and remediation services (Table 16). The square of broadband
deployment is negative and significant in all of the
above industries, suggesting diminishing returns,
as explained above. Broadband’s contribution to
total employment growth ranges from 0.14 to 5.32
percent.5
5

The ranges presented are calculated across the various models for
each industry and across the range of saturation values. The lower
bound of the range corresponds to the parameter estimates of the
simplest model, for which the broadband parameter(s) are significant,
multiplied by the minimum amount of broadband saturation of the
observations used for that industry’s regression. The upper bound of
the range corresponds to the parameter estimates of the most complex
model, for which the broadband parameter(s) are significant, multiplied by the mean amount of broadband saturation of the observations
used for that industry’s regression. Because of the diminishing returns
to scale, the maximum saturation rate yields lower employment
growth than the mean.

F E D E R A L R E S E R V E B A N K O F S T. LO U I S R E G I O N A L E C O N O M I C D E V E LO P M E N T

Shideler, Badasyan, Taylor

Table 2
Regression Results for Total Employment

Intercept

Model 1

Model 2

Model 3

Model 4

–0.0163
(–1.339)

–0.0161
(–1.322)

0.0202
(0.787)

–0.0091
(–0.335)

sat104

0.1503***
(2.893)

0.1499***
(2.878)

0.1727***
(3.042)

sat1042

–0.1162***
(–2.574)

–0.1158***
(–2.559)

–0.1340***
(–2.738)

emp00_34

–0.0329
(–0.603)

hwyaccess

–0.0410
(–0.702)
0.000071
(0.662)

–0.0373
(–0.662)
0.000055
(0.529)

bached00

–0.0047
(–0.053)

–0.0694
(–0.781)

unemp03

–0.0463
(–0.163)

–0.1353
(–0.484)

0.0024
(0.256)

0.0069
(0.740)

rural

0.079

0.082

0.008

0.096

Adjusted R 2

0.062

0.057

–0.038

0.036

F-statistic
n

4.76***
114

3.28**
114

0.17
114

1.61
114

NOTE: Numbers in parentheses are t-statistics; */**/*** indicates significance at the 10/5/1 percent confidence levels.

F E D E R A L R E S E R V E B A N K O F S T. LO U I S R E G I O N A L E C O N O M I C D E V E LO P M E N T

V O LU M E 3 , N U M B E R 2

2007

Shideler, Badasyan, Taylor

Table 3
Regression Results for Forestry, Fishing, and Hunting
Model 1

Model 2

Model 3

Model 4

–0.2613
(–0.574)

–0.2414
(–0.517)

0.9713
(1.004)

0.6373
(0.543)

sat104

1.2113
(0.589)

1.1012
(0.520)

0.2237
(0.097)

sat1042

–1.4430
(–0.792)

–1.3875
(–0.745)

–0.4898
(–0.241)

Intercept

emp11_34

–0.2846
(–0.681)

–0.3323
(–0.749)

hwyaccess

–0.003825
(–1.572)

–0.003697
(–1.419)

bached00

–0.0666
(–0.023)

1.0385
(0.295)

unemp03

–12.1320
(–1.252)

–9.5106
(–0.832)

rural

–0.2611
(–0.699)

–0.1630
(–0.373)

R2
Adjusted R 2
F-statistic
n

–0.1636
(–0.387)

0.057

0.063

0.252

0.267

–0.025

–0.065

0.064

–0.018

0.69
26

0.49
26

1.34
26

0.94
26

NOTE: Numbers in parentheses are t-statistics; */**/*** indicates significance at the 10/5/1 percent confidence levels.

V O LU M E 3 , N U M B E R 2

2007

F E D E R A L R E S E R V E B A N K O F S T. LO U I S R E G I O N A L E C O N O M I C D E V E LO P M E N T

Shideler, Badasyan, Taylor

Table 4
Regression Results for Mining

Intercept
sat104
sat1042

Model 1

Model 2

Model 3

–1.1524***
(–2.704)

–1.1251**
(–2.563)

1.4845**
(2.434)

4.2753***
(2.744)

4.1584**
(2.579)

4.0498**
(2.603)

–3.1826**
(–2.451)

–3.2633**
(–2.599)

–3.2672**
(–2.595)

emp21_34

0.0885
(0.376)

hwyaccess

Model 4
0.2549
(0.348)

0.4037
(1.689)

0.3089
(1.378)

0.001596
(1.063)

0.002309
(1.614)

bached00

–1.8336
(–1.187)

–1.8579
(–1.264)

unemp03

–18.6839***
(–2.785)

–17.1996***
(–2.787)

rural

0.1482
(0.733)

0.2808
(1.430)

0.217

0.221

0.313

0.471

Adjusted R 2

0.161

0.134

0.176

0.310

F-statistic
n

3.88**
31

2.55*
31

2.28*
31

2.92**
31

NOTE: Numbers in parentheses are t-statistics; */**/*** indicates significance at the 10/5/1 percent confidence levels.

F E D E R A L R E S E R V E B A N K O F S T. LO U I S R E G I O N A L E C O N O M I C D E V E LO P M E N T

V O LU M E 3 , N U M B E R 2

2007

Shideler, Badasyan, Taylor

Table 5
Regression Results for Utilities
Model 1

Model 2

Model 3

Model 4

0.2007
(0.513)

0.1284
(0.334)

0.0478
(0.068)

1.4081*
(1.832)

sat104

–1.5917
(–1.066)

–1.0432
(–0.692)

–4.3522***
(–3.083)

sat1042

1.4710
(1.115)

1.1594
(0.889)

4.1867***
(3.113)

Intercept

emp22_34

1.7583
(1.389)

2.2656
(1.574)

1.2979
(0.972)

hwyaccess

–0.000802
(–0.569)

bached00

1.4976
(0.711)

–0.0028
(–0.002)

unemp03

–2.3573
(–0.277)

–2.3171
(–0.312)

rural

–0.2484
(–0.988)

–0.7213**
(–2.770)

R2
Adjusted R 2
F-statistic
n

–0.003194*
(–1.901)

0.068

0.168

0.408

0.675

–0.041

0.013

0.197

0.485

0.62
20

1.08
20

1.93
20

3.56**
20

NOTE: Numbers in parentheses are t-statistics; */**/*** indicates significance at the 10/5/1 percent confidence levels.

V O LU M E 3 , N U M B E R 2

2007

F E D E R A L R E S E R V E B A N K O F S T. LO U I S R E G I O N A L E C O N O M I C D E V E LO P M E N T

Shideler, Badasyan, Taylor

Table 6
Regression Results for Construction

Intercept

Model 1

Model 2

Model 3

Model 4

–0.0449
(–0.621)

–0.0114
(–0.148)

0.4658***
(3.310)

0.3001**
(1.989)

sat104

0.6704**
(2.198)

0.5592*
(1.766)

0.7974**
(2.447)

sat1042

–0.6827**
(–2.575)

–0.5877**
(–2.139)

–0.7267**
(–2.606)

emp23_34

–0.1343
(–1.266)

–0.1758*
(–1.711)

–0.1089
(–1.049)

hwyaccess

0.000146
(0.274)

0.000167
(0.319)

bached00

–1.2414***
(–2.730)

–1.3336***
(–2.820)

unemp03

–3.4686**
(–2.190)

–3.3877**
(–2.121)

rural

–0.0144
(–0.302)

0.0061
(0.130)

0.081

0.096

0.118

0.180

Adjusted R 2

0.062

0.068

0.071

0.118

F-statistic
n

4.28**
100

3.40**
100

2.51**
100

2.88***
100

NOTE: Numbers in parentheses are t-statistics; */**/*** indicates significance at the 10/5/1 percent confidence levels.

F E D E R A L R E S E R V E B A N K O F S T. LO U I S R E G I O N A L E C O N O M I C D E V E LO P M E N T

V O LU M E 3 , N U M B E R 2

2007

Shideler, Badasyan, Taylor

Table 7
Regression Results for Manufacturing
Model 1

Model 2

Model 3

Model 4

0.0098
(0.238)

0.0181
(0.430)

–0.0100
(–0.118)

–0.0047
(–0.052)

sat104

–0.0380
(–0.214)

–0.0622
(–0.347)

–0.0397
(–0.207)

sat1042

0.0226
(0.144)

0.0375
(0.238)

0.0269
(0.162)

Intercept

emp31_33_34

–0.0605
(–1.022)

–0.0546
(–0.926)

–0.0577
(–0.949)

hwyaccess

0.000203
(–0.639)

0.000191
(–0.584)

bached00

–0.0821
(–0.305)

–0.0611
(–0.216)

unemp03

0.4527
(0.475)

0.4918
(0.505)

–0.0094
(–0.329)

–0.0097
(–0.331)

rural

R2
Adjusted R 2
F-statistic
n

0.001

0.012

0.023

–0.020

–0.019

–0.030

–0.052

0.06
99

0.39
99

0.43
99

0.31
99

NOTE: Numbers in parentheses are t-statistics; */**/*** indicates significance at the 10/5/1 percent confidence levels.

100

V O LU M E 3 , N U M B E R 2

2007

F E D E R A L R E S E R V E B A N K O F S T. LO U I S R E G I O N A L E C O N O M I C D E V E LO P M E N T

Shideler, Badasyan, Taylor

Table 8
Regression Results for Wholesale Trade
Model 1

Model 2

Model 3

Model 4

–0.1101
(–0.925)

–0.0762
(–0.641)

0.2908
(1.261)

0.3092
(1.260)

sat104

0.3479
(0.683)

0.2222
(0.438)

–0.3899
(–0.710)

sat1042

–0.4463
(–0.993)

–0.3627
(–0.814)

0.0898
(0.180)

Intercept

emp42_34

0.3139*
(1.885)

0.3328**
(2.149)

0.4010*
(1.960)

hwyaccess

0.001039
(1.242)

0.001308
(1.140)

bached00

0.2426
(0.325)

0.7408
(1.240)

unemp03

–6.2595**
(–2.530)

–5.0516*
(–1.820)

rural

–0.0094
(–0.116)

–0.0219
(–0.290)

0.027

0.063

0.190

0.244

Adjusted R 2

0.006

0.032

0.144

0.183

F-statistic
n

1.27
95

2.05
95

4.17***
95

4.02***
95

NOTE: Numbers in parentheses are t-statistics; */**/*** indicates significance at the 10/5/1 percent confidence levels.

F E D E R A L R E S E R V E B A N K O F S T. LO U I S R E G I O N A L E C O N O M I C D E V E LO P M E N T

V O LU M E 3 , N U M B E R 2

2007

101

Shideler, Badasyan, Taylor

Table 9
Regression Results for Retail Trade
Model 1

Model 2

Model 3

Model 4

–0.0219
(–1.142)

–0.0224
(–1.175)

–0.0837**
(–2.054)

–0.0830*
(–1.841)

sat104

0.0422
(0.508)

0.0532
(0.642)

0.0555
(0.626)

sat1042

0.0110
(0.151)

–0.0004
(–0.006)

–0.0039
(–0.050)

Intercept

emp44_45_34

–0.1004
(–1.467)

–0.1120
(–1.546)

–0.0857
(–1.198)

hwyaccess

–0.000229
(–1.434)

–0.000274*
(–1.740)

bached00

0.2481*
(1.815)

0.1455
(1.042)

unemp03

1.0414**
(2.278)

0.8562*
(1.879)

rural

–0.0132
(–0.953)

–0.0159
(–1.138)

0.072

0.090

0.089

0.144

Adjusted R 2

0.055

0.064

0.046

0.087

F-statistic
n

4.20**
112

3.55**
112

2.06*
112

2.50**
112

NOTE: Numbers in parentheses are t-statistics; */**/*** indicates significance at the 10/5/1 percent confidence levels.

102

V O LU M E 3 , N U M B E R 2

2007

F E D E R A L R E S E R V E B A N K O F S T. LO U I S R E G I O N A L E C O N O M I C D E V E LO P M E N T

Shideler, Badasyan, Taylor

Table 10
Regression Results for Transportation and Warehousing
Model 1

Model 2

Model 3

Model 4

–0.3130*
(–1.926)

–0.3259**
(–2.029)

0.0308
(0.092)

–0.0835
(–0.210)

sat104

0.8014
(1.141)

0.9295
(1.334)

0.6747
(0.590)

sat1042

–0.6466
(–1.022)

–0.7395
(–1.180)

–0.6035
(–0.640)

Intercept

emp48_49_34

–0.3339*
(–1.812)

hwyaccess

–0.3050*
(–1.675)
0.001524
(1.211)

–0.3205
(–1.410)
0.001621
(1.240)

bached00

–0.0999
(–0.088)

–0.2570
(–0.250)

unemp03

–1.8209
(–0.501)

–2.0990
(–0.500)

rural

–0.0893
(–0.772)

–0.0673
(–0.460)

R2
Adjusted R 2
F-statistic
n

0.016

0.051

0.083

0.093

–0.006

0.019

0.030

0.017

0.71
92

1.58
92

1.56
92

1.22
92

NOTE: Numbers in parentheses are t-statistics; */**/*** indicates significance at the 10/5/1 percent confidence levels.

F E D E R A L R E S E R V E B A N K O F S T. LO U I S R E G I O N A L E C O N O M I C D E V E LO P M E N T

V O LU M E 3 , N U M B E R 2

2007

103

Shideler, Badasyan, Taylor

Table 11
Regression Results for Information
Model 1
Intercept

Model 2

Model 3

–1.3904***
(–7.608)

–1.3975***
(–7.705)

sat104

3.5404***
(4.710)

3.6109***
(4.832)

2.7246***
(4.340)

sat1042

–2.6044***
(–3.994)

–2.6591***
(–4.103)

–2.0416***
(–3.570)

emp51_34

0.2389
(1.510)

–1.0817***
(–3.303)

Model 4
–1.5642***
(–4.150)

0.2120
(1.291)

0.2453**
(2.110)

hwyaccess

0.000673
(0.538)

0.000492
(0.630)

bached00

3.6844***
(3.355)

2.6651**
(2.250)

unemp03

3.6022
(1.017)

1.9360
(0.510)

–0.1838
(–1.480)

–0.1321
(–1.180)

rural

0.266

0.286

0.291

0.420

Adjusted R 2

0.248

0.260

0.246

0.367

F-statistic

14.67***

10.70***

6.42***
84

7.87***
84

NOTE: Numbers in parentheses are t-statistics; */**/*** indicates significance at the 10/5/1 percent confidence levels.

104

V O LU M E 3 , N U M B E R 2

2007

F E D E R A L R E S E R V E B A N K O F S T. LO U I S R E G I O N A L E C O N O M I C D E V E LO P M E N T

Shideler, Badasyan, Taylor

Table 12
Regression Results for Finance and Insurance

Intercept

Model 1

Model 2

Model 3

Model 4

0.0882*
(1.916)

0.0884*
(1.936)

0.1452
(1.505)

0.1593
(1.553)

sat104

–0.1579
(–0.790)

–0.1555
(–0.784)

–0.1089
(–0.510)

sat1042

0.0672
(0.377)

0.0694
(0.392)

0.0359
(0.192)

emp52_34

–0.1429
(–1.645)

hwyaccess

–0.1546*
(–1.711)
0.000072
(0.191)

–0.1455
(–1.604)
0.000142
(0.372)

bached00

–0.3995
(–1.252)

–0.2823
(–0.849)

unemp03

–0.9763
(–0.912)

–0.8134
(–0.754)

0.0055
(0.158)

0.0059
(0.166)

rural

0.032

0.058

0.050

0.068

Adjusted R 2

0.013

0.030

0.001

0.000

F-statistic
n

1.68
103

2.04
103

1.01
103

0.99
103

NOTE: Numbers in parentheses are t-statistics; */**/*** indicates significance at the 10/5/1 percent confidence levels.

F E D E R A L R E S E R V E B A N K O F S T. LO U I S R E G I O N A L E C O N O M I C D E V E LO P M E N T

V O LU M E 3 , N U M B E R 2

2007

105

Shideler, Badasyan, Taylor

Table 13
Regression Results for Real Estate and Rental and Leasing

Intercept
sat104
sat1042

Model 1

Model 2

Model 3

Model 4

–0.3452**
(–2.361)

–0.2765*
(–1.878)

–0.9924***
(–3.622)

–1.1343***
(–2.790)

1.1267*
(1.805)

0.9297
(1.500)

0.8904
(1.220)

–0.6935
(–1.260)

–0.7373
(–1.220)

–0.8476
(–1.525)

emp53_34

–0.2013**
(–2.084)

–0.2511***
(–2.679)

–0.2359**
(–2.620)

hwyaccess

0.000231
(0.233)

0.000314
(0.330)

bached00

2.4956***
(2.869)

2.2130**
(2.370)

unemp03

8.4204***
(2.716)

7.7797*
(1.980)

rural

0.1264
(1.375)

0.1536
(1.340)

0.046

0.091

0.190

0.210

Adjusted R 2

0.024

0.060

0.142

0.143

F-statistic
n

2.11
91

2.91**
91

3.99***
91

3.14***
91

NOTE: Numbers in parentheses are t-statistics; */**/*** indicates significance at the 10/5/1 percent confidence levels.

106

V O LU M E 3 , N U M B E R 2

2007

F E D E R A L R E S E R V E B A N K O F S T. LO U I S R E G I O N A L E C O N O M I C D E V E LO P M E N T

Shideler, Badasyan, Taylor

Table 14
Regression Results for Professional, Scientific, and Technical Services
Model 1
Intercept

–0.4827***
(–3.043)

Model 2

Model 3

–0.5022***
(–3.095)

–0.9608***
(–3.444)

Model 4
–0.9995***
(–2.940)

sat104

0.9555
(1.423)

1.0098
(1.487)

0.0960
(0.120)

sat1042

–0.7463
(–1.283)

–0.7798
(–1.331)

–0.1670
(–0.240)

emp54_34

0.0487
(0.610)

0.0158
(0.215)

0.0107
(0.050)

hwyaccess

0.002047*
(1.814)

0.002136
(1.070)

bached00

2.5142***
(2.660)

2.6492***
(2.820)

unemp03

5.7840*
(1.865)

6.1294*
(1.830)

rural

–0.1020
(–1.011)

–0.0926
(–0.790)

0.021

0.024

0.180

0.184

Adjusted R 2

0.002

–0.004

0.140

0.127

F-statistic
n

1.12
109

0.87
109

4.51***
109

3.24***
109

NOTE: Numbers in parentheses are t-statistics; */**/*** indicates significance at the 10/5/1 percent confidence levels.

F E D E R A L R E S E R V E B A N K O F S T. LO U I S R E G I O N A L E C O N O M I C D E V E LO P M E N T

V O LU M E 3 , N U M B E R 2

2007

107

Shideler, Badasyan, Taylor

Table 15
Regression Results for Management of Companies and Enterprises

Intercept

Model 1

Model 2

Model 3

1.0839**
(2.426)

0.9459**
(2.135)

2.5492***
(3.864)

Model 4
2.6532***
(3.922)

sat104

–3.2493*
(–2.043)

–2.7434*
(–1.737)

–1.1233
(–0.756)

sat1042

2.2258*
(1.735)

1.8210
(1.430)

0.4818
(0.403)

emp55_34

–0.5545
(–1.553)

–0.2124
(–0.645)

–0.1776
(–0.555)

hwyaccess

0.000936
(0.766)

bached00

–3.4965**
(–2.503)

–2.5110*
(–1.725)

unemp03

–32.8718***
(–4.087)

–29.6529***
(–3.568)

rural

0.6480***
(2.855)

0.001203
(0.945)

0.5375**
(2.177)

0.204

0.279

0.517

0.590

Adjusted R 2

0.138

0.186

0.402

0.439

F-statistic
n

3.07*
27

2.97*
27

4.49***
27

3.91***
27

NOTE: Numbers in parentheses are t-statistics; */**/*** indicates significance at the 10/5/1 percent confidence levels.

108

V O LU M E 3 , N U M B E R 2

2007

F E D E R A L R E S E R V E B A N K O F S T. LO U I S R E G I O N A L E C O N O M I C D E V E LO P M E N T

Shideler, Badasyan, Taylor

Table 16
Regression Results for Administrative, Support, and Waste Management and Remediation Services
Model 1

Model 2

Model 3

Model 4

–0.8048***
(–3.749)

–0.8225***
(–3.879)

–0.9155*
(–1.938)

–1.4892***
(–3.088)

sat104

3.3569***
(3.648)

3.5039***
(3.843)

3.0858***
(3.095)

sat1042

–2.8518***
(–3.492)

–2.9873***
(–3.690)

–2.7935***
(–3.220)

Intercept

emp56_34

–0.2026*
(–1.771)

–0.1189
(–0.956)

–0.1433
(–1.208)

hwyaccess

0.001013
(0.655)

0.001463
(0.987)

bached00

2.6551*
(1.912)

2.3009
(1.642)

unemp03

8.0518
(1.449)

6.8384
(1.272)

–0.1362
(–0.883)

0.0212
(0.137)

rural

0.145

0.178

0.124

0.232

Adjusted R 2

0.123

0.146

0.066

0.159

F-statistic
n

6.69***
82

5.62***
82

2.15*
82

3.19***
82

NOTE: Numbers in parentheses are t-statistics; */**/*** indicates significance at the 10/5/1 percent confidence levels.

F E D E R A L R E S E R V E B A N K O F S T. LO U I S R E G I O N A L E C O N O M I C D E V E LO P M E N T

V O LU M E 3 , N U M B E R 2

2007

109

Shideler, Badasyan, Taylor

Table 17
Regression Results for Educational Services
Model 1

Model 2

Model 3

Model 4

–1.2298**
(–2.429)

–1.2218**
(–2.375)

–0.3593
(–0.457)

–1.7970
(–1.397)

sat104

3.4602*
(1.895)

3.4229*
(1.843)

3.4027
(1.452)

sat1042

–2.6005*
(–1.822)

–2.5801*
(–1.780)

–2.7168
(–1.479)

Intercept

emp61_34

0.1311
(0.236)

0.1183
(0.193)

0.3076
(0.484)

hwyaccess

0.001346
(0.732)

0.001518
(0.791)

bached00

1.7215
(0.885)

2.4842
(1.236)

unemp03

–4.1990
(–0.457)

2.1853
(0.211)

0.1120
(0.410)

0.2258
(0.745)

rural

0.095

0.096

0.126

0.186

Adjusted R 2

0.043

0.017

–0.011

–0.005

F-statistic
n

1.84
38

1.21
38

0.92
38

0.98
38

NOTE: Numbers in parentheses are t-statistics; */**/*** indicates significance at the 10/5/1 percent confidence levels.

110

V O LU M E 3 , N U M B E R 2

2007

F E D E R A L R E S E R V E B A N K O F S T. LO U I S R E G I O N A L E C O N O M I C D E V E LO P M E N T

Shideler, Badasyan, Taylor

Table 18
Regression Results for Health Care and Social Assistance
Model 1

Model 2

Model 3

Model 4

–0.0031
(–0.158)

–0.0058
(–0.299)

–0.0375
(–0.981)

–0.0447
(–1.066)

sat104

0.0021
(0.025)

0.0107
(0.128)

0.0987
(1.103)

sat1042

0.0364
(0.496)

0.0298
(0.408)

–0.0476
(–0.614)

Intercept

emp62_34

0.0775
(1.517)

0.0830
(1.619)

0.0963*
(1.889)

hwyaccess

0.000108
(0.673)

0.000081
(0.509)

bached00

0.1040
(0.776)

0.0091
(0.066)

unemp03

0.1593
(0.379)

–0.0564
(–0.133)

rural

0.0366**
(2.594)

0.0379***
(2.635)

0.044

0.064

0.087

0.131

Adjusted R 2

0.027

0.039

0.045

0.075

F-statistic
n

2.60*
115

2.52*
115

2.07*
115

2.31**
115

NOTE: Numbers in parentheses are t-statistics; */**/*** indicates significance at the 10/5/1 percent confidence levels.

F E D E R A L R E S E R V E B A N K O F S T. LO U I S R E G I O N A L E C O N O M I C D E V E LO P M E N T

V O LU M E 3 , N U M B E R 2

2007

111

Shideler, Badasyan, Taylor

Table 19
Regression Results for Arts, Entertainment, and Recreation

Intercept
sat104
sat1042

Model 1

Model 2

Model 3

Model 4

–0.5365***
(–2.994)

–0.5341***
(–3.042)

–0.7935**
(–2.538)

–0.9087***
(–2.713)

1.4161*
(1.985)

1.6032**
(2.268)

–0.9953
(–1.609)

emp71_34

0.7610
(1.030)

–1.1892*
(–1.931)
–0.4322*
(–1.775)

–0.6316
(–0.990)
–0.6026**
(–2.612)

–0.6093**
(–2.576)

hwyaccess

0.000458
(0.490)

0.000553
(0.545)

bached00

2.2911***
(2.692)

2.1123**
(2.385)

unemp03

6.4519*
(1.716)

5.5176
(1.414)

rural

–0.1144
(–1.156)

–0.078
(–0.732)

0.102

0.155

0.300

0.316

Adjusted R 2

0.068

0.105

0.229

0.214

F-statistic
n

2.97*
55

3.11**
55

4.20***
55

3.09***
55

NOTE: Numbers in parentheses are t-statistics; */**/*** indicates significance at the 10/5/1 percent confidence levels.

112

V O LU M E 3 , N U M B E R 2

2007

F E D E R A L R E S E R V E B A N K O F S T. LO U I S R E G I O N A L E C O N O M I C D E V E LO P M E N T

Shideler, Badasyan, Taylor

Table 20
Regression Results for Accomodations and Food Services
Model 1
Intercept

0.1260**
(2.577)

Model 2
0.1399***
(2.828)

Model 3

Model 4

0.1066
(1.061)

0.1660
(1.573)

sat104

–0.3120
(–1.486)

–0.3599*
(–1.706)

–0.3966*
(–1.734)

sat1042

0.1969
(1.052)

0.2488
(1.315)

0.2826
(1.401)

emp72_34

–0.1724
(–1.505)

–0.1674
(–1.442)

–0.1663
(–1.421)

hwyaccess

–0.000028
(–0.071)

0.000018
(0.045)

bached00

–0.3259
(–0.986)

–0.1271
(–0.373)

unemp03

–0.1845
(–0.163)

0.1745
(0.155)

rural

–0.0198
(–0.554)

–0.0308
(–0.850)

0.048

0.070

0.034

0.078

Adjusted R 2

0.028

0.041

–0.017

0.008

F-statistic
n

2.45*
100

2.41*
100

0.66
100

1.11
100

NOTE: Numbers in parentheses are t-statistics; */**/*** indicates significance at the 10/5/1 percent confidence levels.

F E D E R A L R E S E R V E B A N K O F S T. LO U I S R E G I O N A L E C O N O M I C D E V E LO P M E N T

V O LU M E 3 , N U M B E R 2

2007

113

Shideler, Badasyan, Taylor

Table 21
Regression Results for Other Services (Except Public Administration)
Model 1

Model 2

Model 3

–0.0988***
(–2.706)

–0.0893**
(–2.439)

0.0012
(0.017)

sat104

0.3207**
(2.061)

0.2834*
(1.820)

0.1758
(1.106)

sat1042

–0.2999**
(–2.200)

–0.2694*
(–1.977)

–0.1954
(–1.422)

Intercept

emp81_34

–0.1242*
(–1.712)

–0.1337*
(–1.930)

Model 4
–0.0367
(–0.493)

–0.1251*
(–1.798)

hwyaccess

0.000564**
(2.018)

0.000614**
(2.200)

bached00

0.2168
(0.918)

0.2482
(1.022)

unemp03

–1.2369
(–1.565)

–1.1539
(–1.448)

0.0059
(0.234)

0.0153
(0.590)

rural

0.045

0.071

0.162

0.191

Adjusted R 2

0.027

0.044

0.121

0.134

F-statistic
n

2.48*
108

2.66**
108

3.94***
108

3.37***
108

NOTE: Numbers in parentheses are t-statistics; */**/*** indicates significance at the 10/5/1 percent confidence levels.

114

V O LU M E 3 , N U M B E R 2

2007

F E D E R A L R E S E R V E B A N K O F S T. LO U I S R E G I O N A L E C O N O M I C D E V E LO P M E N T

Shideler, Badasyan, Taylor

Table 22
Regression Results for Unclassified
Model 1

Model 2

Model 3

Model 4

–1.5715
(–2.105)

–1.4059
(–1.759)

–0.9754
(–1.762)

–1.6043
(–1.154)

sat104

1.9049
(0.791)

1.4702
(0.580)

1.6130
(0.469)

sat1042

–1.1321
(–0.640)

–0.8312
(–0.448)

–1.0778
(–0.428)

Intercept

emp99_34

0.1637
(0.661)

0.2043
(0.823)

0.1596
(0.577)

hwyaccess

0.000975
(0.875)

0.001018
(0.716)

bached00

0.6187
(0.504)

0.5194
(0.394)

unemp03

–0.3898
(–0.061)

1.0663
(0.147)

rural

–0.0185
(–0.107)

–0.0042
(–0.022)

0.109

0.133

0.182

0.204

Adjusted R 2

0.004

–0.030

–0.110

–0.260

F-statistic
n

1.04
20

0.82

0.62

0.44
20

NOTE: Numbers in parentheses are t-statistics; */**/*** indicates significance at the 10/5/1 percent confidence levels.

In the information sector, the analysis shows a
substantial positive impact of broadband availability on employment growth, ranging from 25.27 to
87.07 percent. This growth is not surprising because
this sector contains primarily information technology jobs housed by broadband providers, computer
hardware and software related industries, and other
technology companies that are the most likely to
adopt and use broadband extensively. Additionally,
jobs within the information sector are likely to allow
or even promote working from home. At a residential level, increased broadband availability improves
the ability of these employees to telecommute,
which reduces a firm’s administrative (including
real estate) costs. This allows businesses to expand
and hire more telecommuters without incurring
the administrative costs of keeping an office.
In administrative, support, waste management
and remediation services, broadband’s contribution

to employment growth ranges between 23.74 and
84.56 percent. This is another industry that provides likely opportunities for working from home,
enabling reduced costs and potentially increased
investment in labor. Additionally, this industry
sector contains service industries such as call centers, which are highly dependent upon broadband.
In recent years, several call centers have located
or expanded in rural areas of Kentucky, but needed
broadband service to do so.
Given that construction is a secondary industry,
growth in construction depends on employment
growth in other sectors. As other sectors grow, they
demand additional facilities that create jobs in
construction. In addition, economic growth often
attracts new residents, which in turn increases the
demand for residential construction. To the extent
that we have already realized positive employment
growth due to broadband in other industries, it is

F E D E R A L R E S E R V E B A N K O F S T. LO U I S R E G I O N A L E C O N O M I C D E V E LO P M E N T

V O LU M E 3 , N U M B E R 2

2007

115

Shideler, Badasyan, Taylor

not surprising that broadband contributes to
employment growth at rates between 0.62 and
21.76 percent in the construction industry in the
2004-05 period.
Broadband deployment also had a positive and
significant impact on the mining industry. This
result is also not surprising, because the industry
relies heavily on broadband technology for many
of its production and communication processes,
including the transmission of market prices on
which production decisions are made. However,
given the small sample size in this sector we cannot generalize the result.
For some industries, namely, real estate, rental
and leasing (Table 13), arts, entertainment and
recreation (Table 19), and other services (Table 21),
broadband is positive and significant in Model 1
but becomes insignificant when we add control
variables. F-statistics indicate the overall significance of Model 4 in those industries, implying that
the variables jointly explain employment growth.
This suggests that broadband does contribute to
employment growth, though the other variables
are more influential to employment growth than
broadband. Adding control variables in Model 4
overshadows the impact of broadband deployment
on employment growth. Similarly for educational
services (Table 17), broadband deployment is
positive and significant in Models 1 and 2, but it
becomes insignificant in Model 4. However, given
the small sample size and Model 4 being insignificant, the results for educational services are
inconclusive.
For an additional set of industries, there is
weak evidence that broadband deployment affects
employment growth. Broadband deployment has
a positive sign but is not significant in any of the
four models for the retail trade (Table 9), professional, scientific, and technical services (Table 14),
and health care and social assistance (Table 18)
sectors. Although the broadband parameters are
statistically zero, the positive value does suggest
nominal correlation between broadband deployment and employment growth in these industries.
Additional evidence is found in the F-test for overall model significance: All of the models containing
the broadband parameters are statistically significant. These results seem consistent with these
116

V O LU M E 3 , N U M B E R 2

2007

industries, as they are sectors that consist chiefly
of secondary industry jobs that are dependent on
primary industries. Additionally, as has been
documented elsewhere (see Varian et al., 2002),
health care has been one of the slowest industries
to adopt broadband and still has the lowest adoption rates of any sector. Without at least a propensity to adopt, it is understandable why broadband
availability alone may not immediately affect job
growth in this sector.
For the above mentioned industries, then,
broadband infrastructure appears to lower costs
and/or make markets more accessible, leading to
employment growth. It should be noted that these
industries are made up of primarily higher wage
jobs, suggesting that broadband deployment encourages the growth of higher wage jobs.
The broadband impact is negative and significant for only one industry sector, accommodations
and food services (Table 20). The results suggest
that broadband deployment will decrease employment by 0.34 to 39.68 percent in this sector. One
explanation for this finding could be that individuals are relying more on the Internet for information
about travel destinations and hotel arrangements
rather than working through related service
providers, which may decrease employment
within the travel agency industry. An additional
and broader explanation is that broadband access
increases worker productivity such that employment declines when firms adopt broadband technologies. Given the typically low wages of this
industry (which could be indicative of low productivity), it is possible that broadband availability
enables firms to substitute technology for labor.
Although broadband shows a negative impact
for utilities (Table 5) and management of companies
and enterprises (Table 15), sample sizes are very
small for both of these industries, so the results
are not representative across Kentucky counties
and will not be considered further.
The industries where no variables are significant in any models, and none of the models is significant, are forestry, fishing, and hunting (Table 3),
manufacturing (Table 7), finance and insurance
(Table 12), and unclassified (Table 22). However,
the sample sizes in forestry, fishing, and hunting
and unclassified are too small to draw any conclu-

F E D E R A L R E S E R V E B A N K O F S T. LO U I S R E G I O N A L E C O N O M I C D E V E LO P M E N T

Shideler, Badasyan, Taylor

sions. For manufacturing and finance and insurance, the insignificance of our control variables and
models suggest that we have poor models. That is
to say, our independent variables are not explaining the variance in employment growth between
2004 and 2005. Regression results for transportation and warehousing show only previous employment growth being significant with a negative sign
(see Table 10). However, none of the models is significant according to the F-test, suggesting that
broadband infrastructure has no statistical impact
on employment growth for transportation and
warehousing.

CONCLUSIONS
Based on the results above, we conclude that
broadband deployment has a significant positive
impact on a region’s overall employment growth.
Broadband infrastructure appears to reduce costs
and/or increase market access, and thus lead to
job creation and growth in total employment. At
the sectoral level, broadband deployment positively
impacts mining; construction; information; and
administration, support, and waste management
and remediation services. Broadband deployment
does contribute to employment growth within real
estate, rental, and leasing; arts, entertainment, and
recreation; and other services; however, for these
three sectors, other economic variables appear to
be more influential to job growth than the availability of broadband. Weak evidence suggests that
broadband availability may positively impact retail
trade; professional, scientific, and technical services; and health care and social assistance, though
the impact is likely to be indirect given the supporting nature of these industries to the economy.
Broadband deployment appears to negatively
impact accommodation and food services. These
job losses, however, may be the result of substituting broadband technologies for less productive
workers, which should lead to higher wages in the
long run.
The results also suggest that broadband infrastructure contributes most to employment growth
when counties are neither sparse nor saturated in
their deployment. That is to say, employment
growth seems to be highest around the mean level

of saturation, and this result stems from the diminishing returns to scale of broadband infrastructure,
manifested by the significant but negative saturation squared term. From a productivity perspective,
this result captures the notion that a critical amount
of broadband infrastructure may be needed to sizably increase employment, but once a community
is completely built out (i.e., saturation rate equals 1),
additional broadband infrastructure will not
(indeed, cannot) further affect employment growth.
The policy implication is that investment in broadband infrastructure achieves its greatest return,
measured by employment growth, in communities
that have average saturation levels. Additionally,
policymakers may want to encourage investment
in broadband in poor counties, which also tend to
be rural and/or characterized by low-income households, so that they can benefit from the higher
levels of employment generation.
Extensions to this research are threefold. First,
we have assumed that all broadband infrastructures
are equal; however, they are not. In the United
States, broadband is typically characterized as
having an upload or download speed greater than
200 kilobits per second. Many service providers
greatly exceed this standard, though some do not.
Ideally, one would want to differentiate the broadband infrastructure to identify the speed and/or
platform that are most conducive to employment
growth. Policymakers need such information to
make wise choices about the kind of broadband
infrastructure to deploy. Second, broadband availability and broadband adoption are two very different concepts; we would like to revise this study
using measures of broadband adoption. Third, we
want to use the broadband deployment and adoption data to examine their impacts on different
demographics, such as the poor and the elderly.

REFERENCES
Crandall, Robert C. and Jackson, Charles L. “The $500
Billion Opportunity: The Potential Economic Benefit
of Widespread Diffusion of Broadband Internet
Access.” Unpublished manuscript, Criterion
Economics; www.criterioneconomics.com/docs/
Crandall_Jackson_500_Billion_Opportunity_July_
2001.pdf.

F E D E R A L R E S E R V E B A N K O F S T. LO U I S R E G I O N A L E C O N O M I C D E V E LO P M E N T

V O LU M E 3 , N U M B E R 2

2007

117

Shideler, Badasyan, Taylor

Crandall, Robert C.; Lehr, William and Litan, Robert E.
“The Effects of Broadband Deployment on Output
and Employment: A Cross-sectional Analysis of U.S.
Data.” Issues in Economic Policy, July 2007;
www3.brookings.edu/views/papers/crandall/
200706litan.pdf.
Kentucky State Data Center. Educational Attainment,
Persons 25 Years and Over. Last accessed August 9,
2007; http://ksdc.louisville.edu/sdc/census2000/
education.xls.
Kentucky Transportation Cabinet. KYTC Major Roads,
shapefile; http://kygeonet.ky.gov/metadataexplorer/.
Lehr, William H.; Osorio, Carlos A.; Gillett, Sharon E.
and Sirbu, Marvin A. “Measuring Broadband’s
Economic Impact.” Presented at the Thirty-third
Telecommunications Policy (TPRC) Research
Conference, Arlington, VA, September 2005.
Rupasingha, A.; Goetz, S.J. and Freshwater, D. “Social
Capital and Economic Growth: A County-Level
Analysis.” Journal of Agricultural and Applied
Economics, 2000, 32(3), pp. 565-72.

U.S. Census Bureau. Census 2000 Summary File 3,
Table P37: Sex by Educational Attainment for the
Population 25 Years and Over. Last accessed August
9, 2007; http://factfinder.census.gov/.
U.S. Government Accountability Office. “Broadband
Deployment Is Extensive Throughout the United
States, But It Is Difficult to Assess the Extent of
Deployment Gaps in Rural Areas.” May 2006;
www.gao.gov/new.items/d06426.pdf.
Varian, Hal; Litan, Robert E.; Elder, Andrew and
Shutter, Jay. “The Net Impact Study: The Projected
Economic Benefits of the Internet in the United States,
United Kingdom, France and Germany.” Version 2.0.
January 2002; www.netimpactstudy.com/
NetImpact_Study_Report.pdf.
Williamson, Brian; Marks, Phillipa; Lewin, David;
Bond, Justine and Lay, Helen. “Restoring European
Economic and Social Progress: Unleashing the
Potential of ICT.” Report by the Brussels Round
Table by Indepen, 2006; www.indepen.co.uk/panda/
docs/brt-main-report.pdf.

U.S. Bureau of Labor Statistics. Local Area Unemployment Statistics, Annual Average Unemployment Rate,
2003. Last accessed August 9, 2007;
http://data.bls.gov/map/servlet/map.servlet.MapTool
Servlet?state=21&datatype=unemployment&year=20
03&period=M13&survey=la&map=county&seasonal=u.

118

V O LU M E 3 , N U M B E R 2

2007

F E D E R A L R E S E R V E B A N K O F S T. LO U I S R E G I O N A L E C O N O M I C D E V E LO P M E N T

Full text of Regional Economic Development : Selected Papers From the Third Annual Conference of the Business & Economics Research Group (BERG) : Volume 3, Number 2, 2007

FRASER