.u.‘ iii»? .. a.
hagqumﬁuw , v.5...

. : .
aunﬁmaunﬁmgn... . , . , . imme . . int...
5:. , . .
haw... .. x
s: :ﬁw.”
.2. . . , .. _
Benn... . . . . . .

2.. I‘ll).
1mﬂﬂs..§1
x... 12h»...
.. Wu!) .4019... 3....
1t..17t’mv. .13.! a9“. a.

.z .
V . ........J.w.....
3.95. g
:2. 3.5%:
‘3
h. 11"; :I
. . a . c
.9: .xi... , 3
_ ,. ..

Kauai: x...)
u. Jab ._
’7‘: g II»! H!!! t
. .u- L
I £9.00”
u.
I;

 

 

 

’ n l
.1§¢l¢...:‘

51.5.3.3}
3"). Yai
€35.59 :6.

. . . . . ‘ .. lkuﬂnv.
.23....”«1 . . . , .mmmvmmsﬂ p

7.593...) .~

51.1., 1
yaw..." .141 H...
I. {Emu 1:... In.
.1 . . I.
. h r 9.... Z:
3.: 1x3

 

l .J‘..

. . ._. £55913: 5.. up
5......32 .. 1.....3mniﬂz 3mg... 1%
. .

. hurt-3.7). .z. .

c

. .u‘ilé, .ll... HI. .1. q .5 .t .
. 2 I. . L: 83.7..atlkts .

“inﬂhﬂbi... 2.. 1 5.. V l

I

I. .63 BM. . vtnmdba» .. .vwi. 1.2.. hi 3.:
anﬁ%§§%§ ‘_ . a... >3

 

[may UBRAHY
2 Michigan State
20% University

 

 

 

This is to certify that the
dissertation entitled

THREE ESSAYS ON REGIONAL ECONOMIC
DEVELOPMENT AND LOCAL DETERMINANTS OF
POVERTY

presented by

Denys Nizalov

has been accepted towards fulﬁllment
of the requirements for the

Doctoral degree in Agricultural Economics

 

 

 

 

 

ajor Professo s Ignature

Egg/:95

Date

 

MSU is an Afﬁrmative ActiorVEqual Opportunity Institution

 

- _._u--I-c-o---------n-t-uﬁv-~--~----n-n-.-—

 

PLACE IN RETURN BOX to remove this checkout from your record.
TO AVOID FINES return on or before date due.
MAY BE RECALLED with earlier due date if requested.

 

DATE DUE

DATE DUE

DATE DUE

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

2/05 p:/ClRC/DaleDue.indd-p.1

THREE ESSAYS ON REGIONAL ECONOMIC DEVELOPMENT AND LOCAL
DETERMINANTS OF POVERTY

By

Denys Nizalov

A DISSERTATION
Submitted to
Michigan State University
in partial fulﬁllment of the requirements
for the Degree of
DOCTOR OF PHILOSOPHY

Department of Agricultural Economics

2006

ABSTRACT

THREE ESSAYS ON REGIONAL ECONOMIC DEVELOPMENT AND LOCAL
DETERMINANTS OF POVERTY

By

Denys Nizalov

This work focuses on several important issues of regional development, poverty, and
public policy. It consists of three essays. The ﬁrst essay examines the relationship
between the characteristics of small communities in Michigan and variation of poverty
rates across the State. The poverty model is based on the production behavior of
community residents. It is estimated using the Census 2000 data. The difference in
regional poverty rates is explained primarily by variation in the quality and quantity of
community labor supply. Signiﬁcant differences are detected among rural, metropolitan
and metropolitan adjacent communities. Also, higher poverty rates in rural areas tend to
persist over time, implying that urban and rural poverty should be treated with separate

policies.

The second essay analyzes the impact of economic development policies and highway
infrastructure improvements on grth of per capita income and jobs in Michigan
counties. The policies considered for analysis have signiﬁcant impact on grth
outcomes. However this effect is non-linear. Signiﬁcant heterogeneity in policy effects is
also detected. The impacts are different with respect to average income level in a county
as well as between metro and non-metropolitan areas. In addition, cross-policy effects are
found. The study uses improved measurement of policy treatment while accounting for

possible spillover effects.

The third essay analyzes the relationship between local economic growth and the
distribution of businesses across size categories. The distribution is measured by the
employment share in businesses of various sizes and by a business distribution index. The
index provides a measure of the extent to which the local economy deviates from an
equal employment share in each of nine business size categories. Strong links between a
county’s business size distribution and its economic growth are found. There is also a
difference between the optimal income and job growth enhancing distributions. The
results also indicate that the optimal growth-enhancing distribution of employment has a
higher share of the smallest businesses (with 1 to 4 employees) than it is currently the
average. In summary, the results support increased policy emphasis on encouraging small
business start-ups and development. However, the optimal development strategy depends

on the initial distribution of businesses within a local economy.

ACKNOWLEDGEMENTS

I would like to express my gratitude and the deepest appreciation to my thesis
advisor, Scott Loveridge for his help, guidance, and support, which greatly accelerated
my movement towards the degree completion. I am deeply indebted to Allan Schmid for
his advice and encouragement. I also thank my thesis committee members, John Giles
and James Oehmke, for their help during my studies and for their careful reading of the
manuscript.

I thank everybody in the Department of Agricultural Economics. I greatly
appreciate the friendship of many graduate students. It was very pleasant to be part of
such a unique international community of graduate students in Cook Hall.

Last but not least I thank my American friends, Karen and David Glossops, my
mother and my mother-in-law for their continuous love and support. I dedicate this work
to my wife, Olena, and my children, Anastasiya and Nikolai, for their love, understanding

and patience.

iv

TABLE OF CONTENTS

LIST OF TABLES ............................................................................................................ vii
LIST OF FIGURES ......................................................................................................... viii
ESSAY]. POVERTY IN MICHIGAN SMALL COMMUNITIES: DEMAND VERSUS
SUPPLY OF LABOR ......................................................................................................... l
1 . l . INTRODUCTION .................................................................................................... 1
1.2. MOTIVATION, BACKGROUND AND PREVIOUS STUDIES ........................... 4
1.3. REGIONAL POVERTY MODEL: DESIGN AND ESTIMATION ....................... 7
4. DESCRIPTIVE STATISTICS AND PRELIMINARY EVIDENCE ....................... 19
1.5. RESULTS AND IMPLICATIONS ........................................................................ 20
1.6. CONCLUSIONS .................................................................................................... 28
ESSAY 2 ........................................................................................................................... 30
2.1. INTRODUCTION AND OBJECTIVES ................................................................ 31
2.2. MICHIGAN DEVELOPMENT POLICIES ........................................................... 34
2.3. EMPIRICAL MODEL ........................................................................................... 40
2.4. DATA DESCRIPTION .......................................................................................... 45
2.5. RESULTS ............................................................................................................... 47
2.5.1. Base Model ....................................................................................................... 48
2.5.2. Policy Effect Heterogeneity ............................................................................. 49
2.6. CONCLUSIONS .................................................................................................... 55
ESSAY 3. DOES FIRM SIZE DISTRIBUTION MATTER? AN EMPIRICAL TEST OF
THE ENTREPRENEURIAL PIPELINE THEORY ........................................................ 58
3.1. INTRODUCTION .................................................................................................. 58

3.2. BUSINESS SIZE DISTRIBUTION AND IMPLICATIONS FOR ECONOMIC
GROWTH ...................................................................................................................... 6O

3.3. EMPIRICAL MODEL AND DATA ...................................................................... 64

3.4. RESULTS AND IMPLICATIONS ........................................................................ 71

3.5. SUMMARY AND CONCLUSIONS ..................................................................... 77
APPENDIX 1.1. HOUSEHOLD PRODUCTION MODEL ............................................ 80
APPENDIX 2.1: MODEL SPECIFICATION .................................................................. 84
APPENDIX 3.1. JUSTIFICATION OF THE ESTIMATION TECHNIQUE ................. 85
APPENDIX 3.2: MODEL SPECIFICATION .................................................................. 87
TABLES ........................................................................................................................... 89
FIGURES ........................................................................................................................ 106
REFERENCES ............................................................................................................... 118

vi

LIST OF TABLES

Table 1.1.
Table 1.2.
Table 1.3.
Table 1.4.
Table 1.5.
Table 2.1.
Table 2.2.
Table 2.3.
Table 2.4.
Table 2.5.
Table 3.1.
Table 3.2.
Table 3.3.
Table 3.4.
Table 3.5.
Table 3.6.

Distribution of Community Poverty Rates in Michigan .................................. 89
Descriptive Statistics of Michigan Communities ............................................ 90
Test for Difference, F-statistics ....................................................................... 9]
Estimates of regional poverty model ............................................................... 92
First Stage Estimation for All Communities .................................................... 93
Summary Statistics: Policy Treatment, Michigan, 1993-2000 ........................ 95
Summary Statistics: Michigan Counties, 1993-2000 ...................................... 96
Base Model ...................................................................................................... 97
Michigan Development Policy Effects by Income Level ................................ 98
Intensity of Policy Application ........................................................................ 99
Business Size Distribution: Descriptive Statistics ......................................... 100
Summary Statistics: Michigan counties, 1988-2000 .................. 101
Estimation Results ......................................................................................... 102
Distribution Index and County Economic Growth ........................................ 103
Alternative Estimation Methods .................................................................... 104
Correlation of Business Size Distribution Measures with Control Variables 105

vii

LIST OF FIGURES

Figure 1.1. Poverty Rate, 1999 ....................................................................................... 106
Figure 1.2. Michigan Urban/ Rural Areas by Census Block Groups, 1999 .................... 107
Figure 1.3. Distribution of the Community .................................................................... 108

Figure 1.4. Joint Signiﬁcance of the Labor Demand and Labor Supply Characteristics 108
Figure 2.1. Michigan Economic Growth Patterns .......................................................... 109

Figure 3.1. Share of Business Establishments by Size Category for 83 Michigan
Counties .................................................................................................................... 110

Figure 3.2. Distribution of Employment across the Business Size Categories for 83

Michigan Counties. .................................................................................................. 111
Figure 3.3. Distribution of the Business Size Index across Michigan Counties ............. 112
Figure 3.4. Distribution Index and Income Growth ........................................................ 113
Figure 3.5. Distribution Index and Job Growth .............................................................. 114
Figure 3.6. Index for Business Size Distribution, Michigan Counties. .......................... 115
Figure 3.7. Distribution of Employment and Job Growth .............................................. 116
Figure 3.8. Distribution of Employment and Income Growth ........................................ 117

viii

ESSAY]. POVERTY IN MICHIGAN SMALL COMMUNITIES: DEMAND

VERSUS SUPPLY OF LABOR

1.1. INTRODUCTION

Poverty attracts considerable attention from researchers and policy makers
(Glennerster, 2002; Lipton & Ravallion, 1995). As a result, a large number of policies
and programs have been developed to alleviate poverty. Among them are subsidies, job
training, minimum wage and job creation incentives, etc. Nevertheless, there is a debate
on what would be the most efﬁcient from the government standpoint - improving labor
demand or increasing the quality of the labor supply in areas with high poverty. These
and many other issues related to poverty still need to be investigated. Among them are
whether the determinants of poverty are different between urbanized and rural
communities, whether the sources of persistent regional poverty are beyond the
conditions of the labor market, and, whether the high poverty rates in rural areas are just a
reﬂection of imperfect poverty deﬁnition. This study provides answers to some of these
questions.

Unlike most of the previous literature, this study focuses on explaining
neighborhood poverty variation rather than focusing on the poverty status of individuals
or the poverty variation across large geographic areas (e. g., county or state). This focus
on the neighborhood (small community) is important for several reasons. Unlike the
literature on individual poverty, the neighborhood model would reveal the net effect of
the common resident’s behavior on a community’s poverty rate. It also eliminates the

effect of individual unobservable determinants of poverty and endogenous location

effects that dominate the individual’s behavior. Neighborhood behavior would
incorporate interdependencies among individual residents, which previous literature
addresses only partially.

The focus on the neighborhood also has beneﬁts in comparison to the other
literature on regional poverty. Even though a neighborhood is rarely an independent
administrative unit, many decisions of the state, federal or local government have a
neighborhood scale. Opening or closing a school or military base, expanding a road or
public transportation network, placing a job training or a poverty alleviation program
affect directly an area that is much smaller than a county or a state. Moreover, when the
effect of such a decision is studied on the county or state level, it is often found to be
small and is anyway subject to aggregation error bias.

The geographic scale of this study is also important because the variation of
poverty rates across smaller communities is much larger than the variation across larger
geographic areas. For example, in the year 1999, when the State of Michigan had about
10 percent of its population in poverty, poverty rates varied from three to about nineteen
percent across the counties. At the same time, the variation of poverty across smaller
spatial units such as Census Tracts and Block Groups ranged from zero to one hundred
percent. Thus, focusing on the smaller spatial units with wider variation of poverty rates
would allow studying common determinants of poverty with higher precision. Also,
smaller areas represent a more homogeneous population, which allows more consistent
results to be obtained. Finally, the smaller spatial units allow for the testing of whether
the determinants of rural and urban poverty are different. This difference is hard to

measure at a larger geographic scale.

There is a broad literature on poverty that shows that the sources of poverty
variation across geographic areas are related to differences in historical, natural and
human-made regional characteristics, as well as demographic characteristics of local
population. However, modeling regional poverty as well as the empirical estimation of
such a model is viewed in the literature as an unﬁnished task (Weber and Jensen, 2004).
This paper confronts the modeling task and derives a relationship between community
characteristics and poverty rates based on production behavior of households. The
aggregation procedure is used for this purpose. This approach is similar to the micro-
foundation tradition in the macroeconomics literature (Forni and Lippi, 1997, 1999), and
is new to the ﬁeld of regional science. The resulting model explains the role of the labor
market in determining community poverty rates, and provides a solid base for the
empirical analysis. The model is estimated using Census 2000 data on the smallest
publicly available spatial units — partitions of the Census Block Groups'— for the State of
Michigan. This estimation procedure provides for careful treatment of the aggregation
error, non-randomness of regional characteristics and possible bias from omitted
variables and simultaneity. The focus on the State of Michigan in this empirical analysis
also has a number of beneﬁts. The State is somewhat isolated from spillover effects from
the neighboring states by the natural barriers of the Great Lakes. By considering just one
state, the estimation avoids many unobservable disturbances in macroeconomic,
institutional and legal environments, as all the communities share common political and
historical background. The use of a single data source (US Census Bureau) insures

consistency in deﬁnition and measurement approach.

 

1 Partition of a Census Block Group combines the Census Blocks within the same administrative and
statistic boundaries. See www.census. gov for deﬁnitions. Average population of this unit in Michigan is

The rest of this paper is laid out as follows. Section 1.2 provides further
motivation behind the research questions and reviews brieﬂy the previous literature.
Section 1.3 discusses the regional poverty model and the estimation procedure. Section
1.4 describes characteristics of Michigan communities. Estimation results and policy

implications are presented in section 1.5. Section 1.6 concludes the paper.

1.2. MOTIVATION, BACKGROUND AND PREVIOUS STUDIES

As mentioned earlier, variation in poverty across the smaller communities is
wider than across larger geographic units. The spatial distribution of community poverty
rates also has some interesting patterns that cannot be observed among larger spatial
units. The map of Michigan’s community poverty rates (Figure 1.1) shows that most of
the communities with above state average poverty rates are located in the northern,
relatively rural part of the state. However, there are 826 communities (6.9% of total
number of communities — Table 1.1) with extremely high poverty (30% and above),
located in primarily urbanized areas. As can be seen on the map, the communities with
the high poverty are adjacent to each other, and the boundaries of these contiguous high
poverty areas do not coincide with the county boundaries. Thus, the county is not the best
unit for regional poverty analysis. This spatial concentration of poor and non-poor areas
has also been found in other states (Crandall and Weber, 2004). However, the questions
regarding the drivers behind this spatial concentration, the conditions of labor demand or

the characteristics of labor supply remain unanswered.

 

about 1,300 residents.

In order to ﬁnd an explanation to the observed distribution of poverty, we ﬁrst
refer to the previous literature. This literature can be put into two broad categories2 (see
Weber and Jensen, 2004 for a recent review). The ﬁrst group explains the probability of
an individual with given characteristics being poor. The explanatory factors include:
education, health, working experience, household composition, asset endowment and
other individual and household characteristics (e.g., Shimeles et al., 2000; Strauss and
Thomas, 1995, 1998; Schultz, 1988). Some studies in this group also include regional and
neighborhood characteristics in their explanations of individual poverty status. Galster
(2003), in his review, points to neighborhood speciﬁc poverty effects. These effects work
through behavioral norms, expectations, information sharing standards, etc. However,
Ihlanfeldt (1999) suggests that the effect of these characteristics on individuals is often
subject to the endogenous location bias.

The second group of poverty studies explores the distribution of poverty across
geographic areas. Factors such as climate, industrial structure of the regional economy,
conditions of the labor and other markets, history of settlement and migration affect this
distribution (e.g., Levernier et al., 2000; Partridge and Rickman, 2003; Blank, et al.,
1993; Powers and Dupuy, 1994; Triest, 1997). In particular, Blank et a1. (1993) ﬁnd that
wage disparities and changes in family composition are the major determinants of cross-
state differences in poverty rates. Levernier et a1. (2000) ﬁnd that education is another
key factor in poverty reduction and that the effect is stronger in non-metropolitan areas. It

is also found that higher labor force participation by women is associated with a lower

 

2 There is also a large literature on macroeconomic and institutional conditions such as variations in laws,
terms of trade, foreign exchange and interest rates in explaining poverty variation across larger geographic
areas (e.g., World Bank, 2001). However, these conditions are common for the studied communities and,
thus, are not considered further.

poverty rates. However, for the smaller regions migration plays the key role explaining
variations in regional poverty (Madden, 2003; Goetz, 1999). It is also found that the
impact of the labor market on poverty differs over time and space. Changes in the
industrial structure of local economies (such as plant or mine closings) increase regional
poverty in the short run but they have smaller long run effects (Levernier et al., 2000).
The latter can be explained by responses in migration, unemployment, wages and prices
to employment shocks (Blanchard et al., 1992). It is also found that the job growth effect
on poverty depends on the composition of households in a region, industrial structure,
employment rate and initial rate of poverty (Partridge and Rickman, 2003, 2005; Crandall
and Weber, 2004). Bartik (2001) develops that idea further arguing for a large-scale labor
demand program targeting poor regions.

The literature on regional poverty, however, has some gaps. First, there is a need
for a structural regional model to justify the deterministic role of various factors. Most of
the previous studies include into regional models characteristics of natural environment,
economic structure, public and community institutions, social norms, cultural
environment, and demographic characteristics of the population (see Blank, 2004 for a
comprehensive review of this approach). However, this earlier model remains open
ended. It does not suggest which factors in each group should be included and why, and
what would be a source of endogeneity. Thus, many results of the previous studies are
subject to omitted variable bias.

Second, the systematically higher rural poverty rates found in the literature
(Weber and Jensen, 2004) may only reﬂect the problem with the poverty deﬁnition,

where the lower cost of living (prices) in some areas is ignored (Fisher, 1997;

Glennerster, 2002). Moreover, in rural and urban areas labor demand and labor supply
conditions may play a functionally different role.

Another issue related to the regional poverty is that it persists over time (e.g.,
Partridge and Rickman, 2003; 2005). This persistency reﬂects the effect of some
unobservable factors that do not change over time. However, whether the sources of
persistence are beyond the conditions of the labor market and whether the persistence is
the same for rural and urban areas is not known. This difference (if exists) would have
important implications both for the empirical analysis and for the policy design.

The issues mentioned earlier are formalized in the following hypotheses:

0 Characteristics of the labor supply have stronger impact on community poverty rates
than the conditions of labor demand.

0 There is a signiﬁcant difference between the rural and urban poverty responses to the
conditions of the local labor demand and supply (heterogeneity).

0 Poverty in rural communities tends to persist over time even after controlling for the
labor market conditions and characteristics of the labor force.

The next section presents the theoretical and empirical models for testing these

hypotheses.

1.3. REGIONAL POVERTY MODEL: DESIGN AND ESTIMATION

In order to bridge the gap in the literature and to gain a better understanding of the
variation in poverty rates across the small communities, we derive a regional poverty
model by taking the following steps. First, a household income function is derived and
plugged into the Census Bureau poverty deﬁnition. Second, the resulting household

poverty model is aggregated over the population of a small geographical area

(community)3. This aggregated model relates the community poverty rates to the income
generating patterns of the local population.

Poverty is a multidimensional phenomenon and can be determined and measured
differently4. The Census Bureau deﬁnition of poverty is used for the purposes of this
studys. According to this deﬁnition, a household is considered as poor (p=1) if its total
monetary income 02,-) is below the poverty line (2). The line (known also as a poverty
threshold6) depends on the total size of the household (h,). The same thresholds are used
throughout the United States (do not vary geographically). Thus, the poverty deﬁnition

can be presented by the following ﬁmction:

(1.1) Pi:f(Yis Z(hi))+ 8.

where a, is an individual household error. This error is due primarily to misreporting a

household income and/or a household structure. It is assumed that the error is i.i.d. and
follows standard normal distribution. The income (yi), however, is an outcome of the
household production activity and is a composite function. The household production
model (Singh et al., 1986) was used to derive the relationship between the income and the
characteristics of a household, conditions of the labor market, and other factors (see

Appendix 1.1 for derivation). The household income function takes the following form:

(1.2) yi = l‘jarl'thi‘i'ﬂ' ij.(Wj, rj, k', C ij, hi)+Vij

 

3 One may argue that justifying the community model there is no need to start at the household level.
However, current approach reveals several theoretical and estimation issues. Among them are endogeneity,
and possible inconsistency in earlier regional models due to an aggregation error. These problems cannot
be revealed otherwise.

4 For a discussion of poverty deﬁnitions see Duclos, J-Y (2002).

S More details on the deﬁnition at http://www.census.gov/hhes/poverty/povdef.html

where y,- is real income; r is a rate of return (interest rate) to the physical assets (a), w is a
return to the stock of labor (I) of a given quality for the household i residing in the area j.
The last component (v) is a vector of subsidies, transfers and taxes. Function It I denotes
a real economic proﬁt of a household production, which depends on the interest and wage
rates, costs (c) and a vector of prices (k) and other conditions of a region that determine
production opportunities of a household. The proﬁt is a result of household utility
optimization and is conditioned on a set of household characteristics (h). This component
absorbs the majority of the variation in the household incomes within a region,
conditional on the stock of labor and capital. Under equal conditions external to a
household, and the independent distribution of household characteristics, a real proﬁt for
one household becomes a loss for another. For example, a decision of one household to
take a particular job closes this opportunity for the other households in the neighborhood,
holding the household characteristics equal7. The individual characteristics responsible
for such choices include entrepreneurial ability, preferences, access to information, etc.
As a next step in the modeling, the household poverty status function (1 .1) and
the income function (1.2) are aggregated over the population of a small geographic area
(d)8. In other words, all residents of a small community d (i e d ) are represented by a
single observation9. The aggregated (or community) poverty model takes the following

form:

 

6 For history of development of the poverty threshold see Fisher (1997).

7 The literature on neighborhood effects (e.g., Ihlanfeldt, 1999) suggests that this assumption is not realistic
in cases when endogenous sorting is present or if there is any other sort of interdependencies within a
community. We consider relaxing this assumption later in this section.

8 These communities are much smaller than the local labor market j (e.g., neighborhood, Census Block).

9 Which is often a case with the public use datasets (e.g., Census Bureau provides data on Tracts, Counties,
States, etc.)

(1.3) Pd=Fd(Yd, 2(Hd))+Ed

(1.4) Yd‘= rjAd+wjrd+7r d‘(w, r, k, c, H)+vd

where P, Y, A, T, 72' , V, H and E are the distributions of corresponding household
characteristics within a community d. The functional relationships between the
community characteristics and the poverty rate may be different for different
communities (e. g., rural and urban). The source of this difference is related to the
differences in the unobservable preferences (utility functions) common to a group of
communities (e. g., determined by the culture or historical background, or conditions of
local markets). So far, the aggregated poverty function (Fd) is allowed to be a community
speciﬁc. The relationship between the poverty rate and the community characteristics is
obtained by substituting the equation (1.4) into the equation (1.3).

The poverty deﬁnition (1.1), however, relates the poverty outcomes to the
observed monetary income, while the equations (1.2) and (1 .4) explain the real economic
income. Substituting one for the other introduces an additional source of measurement
error due to the unobservable real proﬁt and partially unobservable transfers. Fortunately,
the aggregation procedure provides a solution to this and some other problems.

An advantage of the aggregation is that it allows separating the variation in the
poverty outcomes caused by common factors from the impact of individual speciﬁc
factors. Common factors include the characteristics of a region such as the interest and
wage rates, while the examples of the individual factors could be an unobserved ability,
preferences or measurement error. The impact of the independent individual factors
(observable or not) on the community outcomes vanishes with the aggregation (Forni and

Lippi, 1997).

10

Estimation Model

The community poverty model, derived earlier, provides a solid base for the
testing the hypotheses of this study. Using the terms of the model, the study evaluates
whether the conditions of the labor demand (the return to labor - w) or the characteristics
of the community labor supply (the stock of labor — T) have more power in explaining the
variation in the community poverty rates. The regional return to labor (w) 10 is controlled
by the county level average wage, unemployment rate and occupational structure1 1. The
characteristics of the labor force within a community (T) are controlled by the following
variables. A quantity (stock) of the labor force is proxied with an average number of
adults per household. The quality of labor force is described with a distribution of
educational attainments, experience (proxied with the average age of a working cohort”),
the number of dependent children, and persons with disabilities per household”.

Another hypothesis is that some community characteristics that affect the income
generating behavior may be unobservable to the analyst and stay unchanged for years.
Such characteristics may affect the accumulation of the labor force and return to it. This
effect is called a persistent poverty (e.g., Partridge & Rickman, 2003; 2005). To control

for persistent poverty and to test for the difference in it between rural and urban

 

10 Goetz (1999) explains poverty and migration outcomes with an expected wage on a local labor market,
which equals a product of an average wage and a probability of being employed: wage *(I-unemployment
rate). .

11 County characteristics are assumed to be exogenous to the characteristics of a small community d. A
possible critique may be that the ﬁrms’ location decisions are inﬂuenced by the characteristics of the local
labor force. However, ﬁrms consider conditions of the local labor market as a whole, but not a very small
portion of it.

12 Adults of an age 16 to 65 are considered as the working age cohort.

13 A more direct measure of health status is not available.

11

communities, a measure of poverty rate from the Census 1990 is included into the model

(P90)M. Thus, the estimation model takes the following form:

(1.5) Pd: F d (Wj, Td, P9Od) +Sd+md+ed+€j

where s and m are the vectors of controls and ed+ej is a composite error term. The
discussion of their use follows. This model becomes a regional analog to the earning
capacity model known in the microeconomic literature on poverty (e.g., Triest, 1997).

As follows from the theoretical model, the analysis of the reduced form the
poverty model (1.5) rests on several assumptions. First is that the characteristics of the
labor market (wT) should not correlate with the conditions of the capital market (rA).
Economic theory treats capital and labor as substitutes in the production function.
However, the omission of this term is not harmful to the analysis of the lower part of the
income distribution. The poor do not hold any signiﬁcant production assets, so the return
to the assets is expected to have little impact on the poverty outcomes. As a result, this
assumption is reasonable.

Second, it is necessary that the household proﬁt does not correlate with the stock
of labor. However, Bertrand, et al. (2004) ﬁnd that poor persons (with lower physical and
human capital) pay higher relative price for the individual mistakes (wrong choices) in
the production activity. For that reason, the poor may be more risk averse and have
relatively lower proﬁt, holding other factors ﬁxed. It follows that the proﬁt and the stock
of labor (7) should correlate positively. Thus, the omission of the former may cause the

upward bias in the parameter estimates for the labor characteristics (T). However, if the

 

14 However, this effect should be distinguished from the lagged response of regional poverty to exogenous

12

households’ unobserved proﬁts are independent within a community, the aggregated
community proﬁt ( 7t ) tends toward zero conditional on the other factors, and the bias
disappears”.
Third, if the system of taxes and subsidies is in general progressive, the
correlation between the transfers and the stock of labor is negative. The omission of the
transfer term (v) brings some downward bias in the estimated parameters. However, we
tend to think about a reverse causation with respect to the poverty status (poverty causes

subsidy, but not vice versa). As a result, the omission of the transfer term does not bias

the estimates of the reduced form model.

Aggregation

Since one of the key assumptions of this analysis relies on the properties of
aggregation procedure, these properties require a careful treatment“. First, the properties
of the error term change with aggregation. Second, an independence of the individual
errors within the communities is a necessary condition.

The ﬁrst issue is that the impact of the idiosyncratic error (8,) on the community

outcome is diminishing with the number of households over which the aggregation is
made17 (Forni and Lippi, 1997). The speed of convergence of the error term to its mean
value depends on two factors. The ﬁrst is the degree of correlation among the shocks
common to the households. The second is the explanatory power of the aggregated
model. The residual impact of the household idiosyncratic error on the community

poverty rate depends on the size of the population over which the aggregation is

 

shocks. The latter effect remains in the error term.
15 The independence assumption is explained and relaxed later in this section.

13

performed. For that reason, the size is controlled econometrically and is denoted by the
vector (s). This vector includes the size of the community population and its area.

However, if the error term a, co-varies among the community residents, the aggregated

error converges to some non-zero constant as the size of population increases. The
covariation may be due to the common unobservable factors such as conditions of the
labor market, local infrastructure, common historical or cultural background, or other
neighborhood effects. If the aggregated error is positive, than the community has a
poverty generating neighborhood effect, while the communities with a negative value of
the error have a poverty reducing unobservable neighborhood effect. This aggregated
error term is denoted by ed+ej in the equation (1.5). The ﬁrst term (ed) represents
independent unobservable neighborhood effects. The second term ((2,) is for unobservable

covariates in the neighborhood effect common to the local labor markets.

Nonrandom Observations

Selective migration creates a serious difﬁculty for poverty analysis. Groups of
households with common characteristics tend to select common areas of residence
causing correlation of the individual observable and unobservable factors over space
(Goetz, 1999; Madden, 2003). The effect of these common unobservables does not
disappear with the aggregation constituting the neighborhood effect. The problem arises
when those unobservable individual factors or the unobservable regional factors
impacting migration and residential location correlate with the poverty outcomes. This

correlation may cause bias in the estimation results. On the positive side, selective

 

16 For theoretical treatment of aggregation see Fomi and Lippi (1997, 1999).
17 Under the assumption that the individual error has a ﬁnite variance.

l4

migration makes the communities relatively more homogeneous, making them an
attractive unit for analysis.

To correct for the selective migration effect, Strauss and Thomas (1995) suggest
including the factors determining the migration patterns into the analytical model. In this
study, those factors are the share of retirees in the local population (to control for the
destination of the retirees’ migration) and a binary variable for minor civil divisions with
a college or university (to control for the migration of students). In the Census data set,
migration is partially observed for the Block Groups, thus the share of in-migrants in
local population is included into the model as well. The vector of controls for migration is
denoted by (m) in the model. As found by F itchen (1994), most people that change the
residential location, tend to stay within the same basic community. This ﬁnding is
consistent with the endogenous location arguments, but also shows that community
characteristics persist over time. Thus, the controls for migration (m), poverty rate in
1990 (P90) and conditions of the labor markets (w) would provide a sufﬁcient control for

the selective migration.

Omitted Variable Bias

Even though several controls are used for the unobserved covariates, the results
may suffer from omitted variable bias due to the unobserved conditions of the local labor
market denoted by e. To deal with this problem a spatial ﬁxed effect estimation
procedure is used. Following this procedure, a set of binary variables for the local labor
markets is included into the model. The markets are proxied by the commuting zones
(Tolbert and Sizer, 1996). The zones include several counties connected by the

commuting patterns of their residents. About 80% of the residents live and work within

15

the same commuting zone. These binary variables absorb the impact of all observable and
unobservable factors that are constant within the zones, including the common error term
(e1). These controls absorb also the systematic differences in the price levels (and
associated errors in the poverty measurements) across the labor markets. Also, they

provide an additional control for the migration.

Endogeneigy Bias

The endogeneity of some explanatory variables may bias the estimation results. In
particular, the number of adults, dependent children and the share of in-migrants may
reﬂect the outcome of the poverty-coping strategies. For example, in the better-off areas
people may have better opportunities to support larger families. On the other hand, an
area with a lower poverty may attract wealthier migrants, while poor communities may
attract more poor families (e. g., F itchen, 1995). Thus, the two-stage estimation procedure
is applied (Foster & McLanahan, 1996; Wooldridge, 2001, 2003).

Three groups of instrumental variables are used. The ﬁrst group includes the

8: an average age of heads of family

following characteristics of the neighboring areasl
households, an average household size and a proportion of females. The second group
includes ten-year lag for county-level wage rate, average number of adults per household,
and their interaction. The third group consists of a binary variable for the coastal

counties. These variables are statistically signiﬁcant in predicting the instrumented

variables and do not belong to the poverty model (see Table 1.5 for more details.)

 

18 County average characteristics (excluding one township to which the observation belongs) are taken as
characteristics of neighboring area.

16

Error Structure

The estimation procedure described earlier is designed to obtain the unbiased
estimates of the community poverty model. However, the remaining error term (ed) is not
independent across the observations. There are two sources of this correlation. First, a
spatial correlation of the poverty rates has been found in the literature (e.g., Partridge and
Rickman, 2005; Crandall and Weber, 2004). Second source is related to the construction
of the explanatory variables in our empirical model. Some of them are more aggregated
than the dependent variable (e. g., the binary variable for the college towns), which may
cause a correlation in the error (Moulton, 1986). To account for this autocorrelation and

other forms of hetroskedasticity, fully robust standard errors are estimated”.

Rural/Urban Difference

Another hypothesis is tested by estimating the difference in the poverty function
(Fd) among rural and urban communities. This sort of heterogeneity in the model is
another possible source of bias. When the response of some communities to the
explanatory factors is functionally different from other communities, bundling these
responses into a single estimation model produces a function, which is a weighted sum of
the responses in these different groups of communities. As F omi and Lippi argue (1997)
the simplest solution to this problem is to disaggregate the population to a meaningful
number of relatively homogeneous sub-populations and to estimate the model separately

for those subgroups. Another approach is to carefully model the heterogeneity of

 

19 A less conservative approach is the estimation of the cluster robust errors. However, it also brings a risk
of misspeciﬁcation of the error structure.

17

parameters within the same econometric model by including interactions of the
explanatory variables with the sub-population dummies.

It is found in the literature that the heterogeneity of a poverty model may relate to
urban/rural or metropolitan/non-metropolitan differences (Weber and Jensen, 2004).
Thus, we disaggregate Michigan communities into those four categories (Figure 1.2) and
test for the signiﬁcance of the functional difference.

The southern part of the State includes metropolitan and metropolitan adjacent
areas (Beale Codes 0, 1, 2, 3, 4, 6 and 8). The northern, relatively rural, part includes
Michigan non-metropolitan counties. Those two parts have different migration and
commuting patterns as well as the structure of the regional economies”. Urbanized
communities are put in a separate category for each part of the State. A community is
considered as urbanized if it is located inside the Urbanized Areas or Urbanized
Clusters21 deﬁned by the Census Bureau. Thus, the ﬁrst sub-region is the southern
urbanized area, which is referred to as Metropolitan. This category includes urban
communities in all seven Michigan metropolitan areas. The non-urbanized part of
southern Michigan is called Metropolitan Adjacent and contains farms and sparsely
populated residential communities. Next sub-region includes urbanized communities in
the northern half of the State. It is referred to as Rural Towns. It includes northern cities
(none are more than 20,000 population) plus many smaller towns and villages. The rest of

the non-urbanized northern part of the State is called Rural.

 

20 Southern Michigan is heavily industrialized and also contains some of the Michigan’s best farmlands.
The northern half of the State has poorer soils and a shorter growing season. Forests are the dominant land
use. Mining has been a major employer in some northern counties, but has been declining. Tourism plays
an important role in northern Michigan. This area is a site for second homes and retirement. The northern
part of the State is “very dependent upon unearned income and government earnings [such as public
administration, education, social services]” (Erickcek & Watts, 2003), which all together contributed more
than 30 percent of the total personal income in 2000.

18

1.4. DESCRIPTIVE STATISTICS AND PRELIMINARY EVIDENCE

The empirical model derived earlier is estimated using the data from the Census
2000 Summary File 3 for the State of Michigan. Partitions of the Census Block Groups
are used as the units of analysis. These units are the smallest geographic areas available
for the public use from the Census Bureau, and the only ones that have a clear distinction
between the urban and rural areas22 (U .S. Census Bureau, 2003). The partitions are much
more homogeneous by the boundary design than the larger units such as Block Groups,
Tracts, counties or states used in the previous literature. The downside of the small
geographic units is that the data imputation problem may be more sever. However, this
issue is not investigated further in this paper.

The total number of the Census Block Group partitions in Michigan is 13,707.
The study sample, however, excludes 530 observations with no land area, 885 units with
no population and 274 units for other reasons”. The resulting sample includes 12,018
units with the average population of 1,339 in on average 509 households. Out of the total
sample, 1871 communities (15.6% of total) are the rural; 517 (4.3%) are in the rural
towns; 6775 (56.3%) are the metropolitan; and 2855 (23.8%) are the metropolitan
adjacent (Table 1.2).

The size of communities is different across the sub-regions of the state. The rural

communities have on average 678 residents hosting altogether about 7% of Michigan

 

21 See http://www.census.gov/geo/www/ua/ua_2k.html for ofﬁcial Census Bureau deﬁnition.

22 Urban/Rural ﬁeld (UR) of the identiﬁcation section of Summary File 3 data is used.

23 Among them 26 with no housing units, two partitions with no reported income and one with no
household. Also excluded are units with unusual demographic characteristics. Among them, 67 had no
adult of working age, 88 areas had no male or female adult of age above 25 years, 34 had average number
of working age male or female adults of more than 10 per household and six units had average number of
dependents per household more than ﬁve. In addition, 50 observations with no worker over 16 years are
excluded.

l9

population (Table 1.2). The metropolitan communities are the largest on average — 1,485
residents — hosting about 72% of Michigan population. Spatial distribution of the poverty
is different between Michigan sub-regions. Metropolitan adjacent areas have the lowest
average poverty rate24 (6.2%), while the highest rate is in the nrral towns (14%). Many
rural and central city communities have a long history of poverty.

The characteristics, used to describe the poverty distribution are also different
among the sub-regions. The most signiﬁcant difference is observed in the distribution of
the labor demand conditions. The metropolitan communities face the highest average
wage rate (about $36.5 thousand), while the lowest average wage (about $23.8 thousand)
is in the rural communities. The residents of the metropolitan adjacent communities face
the lowest unemployment rate (5.3%), while the rural communities have the highest rate
(7.8%). The distribution of the labor force characteristics is less different among the sub-
regions, with the education level being slightly higher in the urbanized communities. At
the same time, the variables that designed to control for the migration effect display the
differences in the distributions. Northern communities (rural towns and rural) have a
higher share of retirees and in-migrants, while the urbanized communities (metropolitan

and rural towns) have a higher probability of being a part of a college town.

1.5. RESULTS AND IMPLICATIONS
The regional poverty model is estimated using the procedure described earlier”.

The dependent variable — poverty rate — is measured at the level of the Block Group

 

24 Percentage of the total households with income below the poverty threshold is used as the community
poverty rate.

25 The ZSLS procedure was compared with OLS and GMM with and without the proposed control
variables. The 2SLS results are signiﬁcantly different from the OLS and the estimates with no controls.
However 2SLS and GMM estimates are similar.

20

partitions. The distribution of the poverty rates at this geographic level is skewed. There
are more communities with lower poverty rates than the high poverty communities
(Figure 1.3). Thus, the dependent variable is transformed with the natural logarithm”,
and the estimates represent the semi-elasticity (a percentage change) of the poverty rate
to the changes in the regional characteristics. The results are presented in Tables 1.3 and
1.4.

The major ﬁnding of this study is that the responses of the poverty rates are
jointly different in each sub-region from the responses in other parts of the state. This
difference is tested by computing a joint signiﬁcance of the interactions of the model
parameters with binary variables for the sub-regions (a variation of the Chow test). The
results of this test (F -statistics) are reported on Table 1.3. The biggest difference (F-
statistics equals 11.63) is found between the rural and metropolitan communities, while
the smallest is between the rural and rural town communities (F-statistics is 0.86). This
difference is driven by speciﬁcs of preferences and unobservable constraints reﬂected in
production decisions. This ﬁnding implies that the estimates of the poverty model for the
entire state do not represent, as a rule, the behavior of any of the sub-regions. The state-
wide estimates would change from a state to state simply because the proportions of the
sub-regions would change as well. Another important implication is that poverty
alleviation programs and policies have to be, as a rule, sub-region speciﬁc and target
those factors that are the most signiﬁcant in a speciﬁc area (accounting for the cost

efﬁciency of the program). The question of what effects the variation more — the labor

 

26 The estimates of the models with the transformed and non-transformed measure of the poverty rate are
similar in sign and signiﬁcance level. However, the transformed model is more efﬁcient as should be
expected.

21

demand or the labor supply conditions - in different sub-regions, becomes even more
important.

The characteristics of the labor supply have a higher impact on the variation of
poverty than the characteristics of labor demand for all sub-regions of the state (as
measured by joint signiﬁcance of these two groups of variables - see Figure 1.4). This
result conﬁrms the conclusions reached by Weber and Jensen (2004, p.20) in their review
of the literature. The effect of the labor supply characteristics is particularly strong in the
metropolitan communities (F-statistics is 76.62). Relative to the total explanatory power
of the model, the supply characteristics have the lowest effect on poverty in the
metropolitan adjacent communities.

In contrast, the labor demand conditions have an insigniﬁcant effect27 on poverty
distribution in all parts of the state when the quality of labor is controlled. This
counterintuitively implies that better demand conditions (e. g. higher wages, lower
unemployment) have limited poverty reduction effects. Most likely, the new jobs are
taken by those with better qualiﬁcations, leaving the poor behind. It also implies that
communities have to improve the quality of the labor supply before they can beneﬁt from
stronger labor demand conditions. With that, we turn to the detailed description of the
results for each of the explanatory factors in the model. Whether an estimate for a
particular sub-region differs individually from the estimate for rural communities28 is
indicated by the boldface font in Table 1.4. For example, more people with disabilities

correlates with increased poverty to a greater extent in the metropolitan communities than

in the rural counterparts, and the difference is statistically signiﬁcant.

 

27 The estimates of the labor demand effect are similar whether the commuting zone ﬁxed effects are
controlled or not.

22

In Table 1.4, Column 1 presents the estimation results where all communities in
the state are used. Columns 2 through 5 present the estimates for the rural, rural town,
metropolitan and metropolitan adjacent communities, respectively. Comparing the results
among the columns we can see that in most cases the state-wide estimates reﬂect an
average effect in different sub-regions weighted by the number of observations and the

level of signiﬁcance.

Labor Market Conditions

Three variables control for the characteristics of the labor demand in the model.
Among them, only the wage rate in rural and metropolitan communities exhibits a
signiﬁcant relationship with the poverty rate (Table 1.4). An increase in the county level
average wage rate by 1% (which is equivalent of $238 for rural communities and $365
for the metropolitan communities) decreases the poverty rate by 0.7 percent, conditional
on other factors. The estimates of the effect of labor demand conditions (except the
occupational structure) are not statistically different among the sub-regions, which

probably reﬂects their lower overall signiﬁcance.

Labor Force Characteristics

The number of working age adults in an average household is the most important
determinant of the community poverty rare in all regions of the state. An additional
working age adult is associated with a two to three times decrease in the regional poverty
rate (Table 1.4). It implies that an increase in the average income associated with an

additional adult is larger than the corresponding increase in the poverty threshold. The

 

28 Rural sub-region is taken for control group as an example.

23

highest effect is observed in metropolitan areas where the conditions of the labor demand
are better. This result indicates that policies reducing the number of single parent
households may have a large poverty reduction effect. Larger families can enjoy
economies of scale in the household production not available to single-adult families

(e. g., sharing childcare and other duties, improving access to information, employing
specialization and distributing risks).

A correlation between the average number of children under 16, and the poverty
rate is signiﬁcant only for rural communities. The positive sign on this effect reﬂects the
quality-quantity trade off in children found in the literature. Families with higher income
tend to have fewer children, but to invest more in the quality of their education and
health. However, in our study the two stage estimation procedure reduces this circular
effect, and what is observed is a reﬂection of the poverty deﬁnition. One additional
person in a household with no income lifts the poverty threshold and, thus, the
community poverty rate. This result is similar to what is found in the literature (e. g.
Levernier et al., 2000).

The educational level of the local population is one measure of labor quality. The
effect of it is quite uniform across the sub-regions of the state. A one percentage point
increase in the share of people with incomplete high school education29 increases regional
poverty by one to two percent, reﬂecting relatively lower quality of labor. An increase of
share of population with higher levels of education than the high school is associated with
some decrease in the poverty rate. However, this effect is statistically and practically
small reﬂecting the fact that the higher level of education affects the distribution of

income at the levels higher than poverty line.

24

The average age of the local working age cohort serves as a proxy for experience.
It has different impact on poverty in different sub-regions. Every additional year in the
average age of the local working cohort decreases poverty in urban areas at a decreasing
rate (the relation has a U-shaped form). The interpretation is that people with more
experience can earn more. The relationship between the age and poverty outcomes in
non-urbanized areas, however, has a positive sign implying that physical health of youth
has a higher return than experience (the relation has an inverse U-shaped form). It implies
that people with more experience would be better off moving to urbanized areas.

Average number of people with disabilities per household is another characteristic
of the labor force. Every additional person with disabilities corresponds to an increase in
the community poverty rate. This effect is particularly strong in the metropolitan areas,
where an additional person with disabilities is associated with 0.85 times higher poverty
rate. This estimate is much higher than a similar effect in rural communities. This result
implies that society does not provide full insurance to the families with disabled

members, and people with disabilities cannot fully cover their living costs.

Persistent Poveﬂy

Location in a county with a higher poverty rate in 1990 is associated with higher
poverty in 1999, which points to persistence of regional poverty. In rural communities
this effect is particularly strong. One percent higher poverty rate in a county 10 years ago
is associated with 0.45 percent higher poverty rate in 1999. The results for other regions
are not statistically signiﬁcant. Partridge and Rickman (2005), and Weber and Jensen

(2004) ﬁnd similar evidence for rural areas. However, our results highlight differences in

 

29 The share of population with complete high school degree is the omitted category.

25

the persistent poverty across different types of communities. Persistence of regional
poverty implies that its sources do not disappear over time constituting a regional poverty
trap. Regional poverty traps may happen due to the conditions of local infrastructure,
topography, weather or other natural or human made factors. Escaping from the trap

often requires collective action and outside help. This result calls for a more active role of

government in poverty reduction, particularly in rural areas.

Control Variables

The use of the controls is necessary to obtain consistent results. Even though not
all of them have an interpretation besides controlling for the unobservables, excluding
these variables would bias the results.

The share of in-migrants is found to be not statistically signiﬁcant. The result is
smaller in magnitude to what has been found in the literature (e. g., Partridge and
Rickman, 2005). The share of retirees captures speciﬁcs of their migration in Michigan.
A higher share of people over 65 is associated with lower poverty, controlling for other
factors. It reﬂects the fact that older people bring non-labor income (pensions, insurance,
dividends, etc.) into communities and generate local demand for goods and services.
Increasing the share of retirees in a local population by a one-percent point is associated
with a reduction in the poverty rate by about seven percent in the urbanized areas”. This
result is also similar to the ﬁnding by Partridge and Rickman (2005) when they control

for conditions of labor demand.

 

30 Note that adding wealthier people to a region can reduce the percentage of households in poverty
without necessarily doing anything for those in poverty. On the other hand, adding an external income
stimulates local demand for goods and services, which in turn may have poverty reduction effect. The
estimates represent the net effect.

26

College students represent a special case of temporary chosen poverty. A binary
variable for college towns controls for communities with a high share of students. The
metropolitan communities with colleges have poverty rates that are 0.34 times higher
than in the communities with similar characteristics, but no college. The presence of
voluntary low-income students should not be confused with poverty of the working
population.

Controls for the aggregation error (population and area size) are found to be

signiﬁcant, indicating that the residual impact of the individual idiosyncratic error (8,) is

a potential problem. The estimate for population size in metropolitan areas is the smallest
(0.12 vs. 0.25 in rural areas). It indicates that the individual unobservables have a lower
effect on the community outcomes. The units in metropolitan areas are more populated
with relatively more homogeneous residents, which supports our argument. Thus, adding
a person to a larger community would have smaller effect on the aggregate characteristics
of this community. This result is comparable to previous ﬁndings in the literature (e.g.,
Partridge and Rickman, 2005; Levernier et al., 2000), where the population size effect is
much smaller for larger spatial units (county). The positive sign on the population size
reﬂects most likely the effect of the boundary design.

The results of this study show a lot of similarities with previous ﬁndings in the
literature. However, current analysis shows important differences in poverty responses
between urbanized and non-urbanized communities, and articulates that quality and

quantity of local labor supply are the most signiﬁcant determinants of poverty.

27

1.6. CONCLUSIONS

This paper takes a new approach to the analysis of poverty distribution across
small communities that is based on production behavior of community residents. For that
purpose, the study relies on the Census 2000 data on partitions of the Census Block
Group. The results reveal signiﬁcant differences in the poverty responses of rural,
metropolitan and metropolitan adjacent communities to the characteristics of regional
labor demand and supply. It follows that a policy response to community poverty should
also be selective and region-speciﬁc.

It is also found that the characteristics of the labor supply play a more important
role in determining community poverty than the conditions of the labor demand. Thus,
the improvements of the labor demand conditions are less powerful poverty reduction
instruments than the improvements in the quality of labor.

Among the characteristics of labor supply, the number of adults per household
(quantity of labor force) is the most important factor. However, adults have different
opportunities to earn income in different sub-regions. Those opportunities are beyond the
conditions of labor demand and may reﬂect varying barriers and costs of labor force
participation. Among these barriers could be poor transportation and communication
infrastructure, lower availability of childcare facilities, etc. Quality of labor is also
important. Communities with higher proportions of people with disabilities have higher
poverty. It implies that improvements in the system for support and assistance for the
disabled are needed.

A high persistency of poverty is found in nrral areas. As the number of people in

agriculture, mining and forestry has declined, some small rural communities have lost

28

their reason for being. Some communities have experienced high poverty rates for
decades, which calls for a policy intervention.

While many theoretical and estimation issues were considered, the results are not
without points for critique. However, the proposed model provides solid empirical and
theoretical ground for further regional poverty research and policy analysis. In particular,
it allows incorporating factors such as the cost of labor force participation, access to
information, etc. into the poverty analysis. Also, further research is needed to determine

if the results are valid for institutional environments in other states.

29

ESSAY 2

Nizalov, D., and S. Loveridge. (2005). “Regional Policies and Economic Growth: One
Size Does Not Fit All,” Review of Regional Studies, 35(3), pp. 266-290.

30

2.1. INTRODUCTION AND OBJECTIVES

The growth rate and the distribution of personal income are among the major
characteristics of well being of a local population. They are also issues of concern for all
levels of government. As a result, a large number of federal and local economic
development policies and programs are aimed at personal income growth. Such programs
include improvement of public infrastructure, job training, grants, loans, tax abatements
for business expansion and retention, etc. For example, in the year 2000 Michigan had
more than 40 programs and policy tools speciﬁcally aimed at economic development
(CRC 2001). However, with many states currently facing budget difﬁculties, the role of
state government in economic development is likely to be challenged with increasing
frequency. The effectiveness of development policies and programs in raising income is
therefore an important question.

There is a large literature on economic development program evaluation, but there
is no consistency in the results. Some authors found no signiﬁcant relationship between
policies and growth, while ﬁndings of others vary greatly (for reviews, see Bartik 1991;
Fisher 1997; Wasylenko 1997; G055 and Phillips 1999; Buss 2001). The reasons for the
variation in ﬁndings are numerous. Among them are the quality of data (Buss 2001);
deviation of true political goals from optimal economic development path of a region
(Wolcoff 1992; Dewar 1998; Thomburgh 1998); imperfections in policy evaluation
methods (Bartik 1991; Buss 2001) and shortcomings in policy design.

This paper addresses some of the above-mentioned concerns. However, there are
some less studied issues that contribute to the lack of consistency in empirical policy

assessments. The ﬁrst of them is that the same policies may have signiﬁcantly different

31

effects in different geographic areas due to the heterogeneity of local economies. This
heterogeneity leads to an aggregation error and non-systematic empirical results (Forni
and Lippi 1997). Separate estimation of policy effects in counties with relatively high and
low income, different industry structure, and level of urbanization is undertaken to assess
the extent of the aggregation error issue.

Some support for the heterogeneity concern can be found in the literature. Loh
(1993) ﬁnds that effect of development policies on job growth differs across industries.
He ﬁnds also that the effects of different policies on the same outcome are different, as
well as the effect of the same policy on different measures of growth. Similar evidence is
found by Dalenberg and Partridge (1995), Luce (1994), and Papke (1991). Fisher (1997)
suggests that this heterogeneity can be explained by different marginal return to factors
targeted by the policies across industries and sectors. The marginal return may also differ
with respect to average income level, and between metropolitan and non-metropolitan
areas. Another source of heterogeneity is related to possible differences in the size of the
marginal effects of policy under different scales of policy application. If under these
circumstances, the intensity of policy application is systematically different among sub-
regions of a state, the marginal policy effect should also be different.

Another rarely addressed problem in policy analysis is that a large number of
policies and programs are implemented simultaneously. Several of them are either
mutually exclusive or additive by design. On the one hand, this fact may lead to mixing
the impacts of different policies. On the other hand, programs may have mutually
increasing or decreasing effect. To address these issues, we consider the effect of several

policies at the same time.

32

The above-mentioned issues can be restated in the following hypotheses:

o The impacts of development policies vary with respect to an average income level,
industrial structure, and degree of urbanization of local economy.

0 The policy effects are not linear.

o The effect of any development policy is conditional on presence of other policies in
a region.

To test these hypotheses, this article analyzes impacts of Michigan economic
development policies and improvements in highway infrastructure on county per capita
income and job growth in the 1990s. Focusing on the economy of just one state helps
control for unobservable factors such as the legal system, culture, other statewide
institutions, and natural amenities. Loh (1993) mentions that such an approach would
greatly improve the quality of data on development policies since comparability is no
longer an issue. This approach allows excluding variation in macroeconomic factors as a
potential source of bias.

Several other issues raised in the literature are also addressed. First, an attempt is
made to improve the measurement of policy treatments. Previous studies use the presence
or absence of policy treatment in an area (e. g., state) to identify the policy effect. Other
studies look at the number of development programs available in a region (for review see
Fisher and Peters 1997). In contrast, to quantify the policy shock to a local economy, we
use more precise measures. The intensity of policy use is measured by number of
businesses receiving a beneﬁt (e. g., tax break) or by area of a zone receiving a policy

beneﬁt. This approach allows assessing the effect of marginal change in the treatment on

33

growth outcomes (e. g., providing policy beneﬁts to one additional business) -— an issue
that has important policy implications.

Second, a policy treatment may have a signiﬁcant spillover effect on neighboring
areas (Bartik 1991; Papke 1993, 1994; Fisher 1997). This effect arises due to regional
production and consumption relationships (multiplier effect), migration, and commuting
patterns. While policies are applied to small territories of a neighborhood or a township
size, a county is used as a unit of analysis“. The area of a county captures additional
beneﬁts of a policy applied to a small territory within a county.

The rest of the paper is structured in the following way. Part 2.2 provides a brief
introduction to the Michigan economy and explains the nature of the economic
development policies under consideration. Next, we turn to the justiﬁcation of the
empirical model used for policy analysis. Part 2.4 describes the data. Discussion of the

results is presented in part 2.5. Part 2.6 concludes the paper.

2.2. MICHIGAN DEVELOPMENT POLICIES

Michigan is a heavily industrialized state. While its population accounts for 3.5
percent of the US total (Census Bureau 2000), the state contribution to the national
manufacturing output was 5.4 percent in 2000 (BEA 2005). Motor Vehicles
Manufacturing is an important industry for the Michigan economy. While major facilities
are in metropolitan areas, the auto companies have outsourced many components to small

manufacturers across the state, providing employment outside the metropolitan areas.

 

31 One could argue that a county is a too large area to capture the effect of treatment applied to just one
business establishment or the effect of a small tax-free zone. However, if anything we would underestimate
the policy effect since local commuting zones normally include several counties (Tolbert and Sizer 1996).
Supply and demand chains are even more widely distributed than commuting.

34

During 19903 Michigan real personal income per capita32 grew at an average
annual rate of 1.9 percent, slightly outpacing the national average of 1.6 percent. The job
grth was on average 0.9 percent per year, compare to 0.6 percent of the national
average (BEA 2005).

Even though the state economy went through structural changes contributing to
the growth, this period is also known for an aggressive role of the state and local
government in economic development (Bartik et a1. 2003; CRC 2001). Those efforts were
mirrored in the sharp increase of the state’s “Governor’s Cup” ranking published by Site
Selection, an economic development trade publication. The Governor’s Cup is an annual
ranking of states based on the reported number of major new investments in business re-
location and expansion”. After placing 15th in 1993, and 22nd in 1994, Michigan
recovered to 7th place in 1995 and 6h in 1996. In 1997, Michigan started a string of ﬁrst
place rankings that continued for four consecutive years before it dropped to 4th place in
2001, 2nd in 2002, and 5th in 2003.

However, the contribution of the government to this improvement may be quite
limited; the various types of business expansion (jobs created, new buildings or ﬁnancial
investments) included in the ranking may not. depend exclusively on the existence of a
particular policy. For example, with respect to job creation, Faulk (2002) ﬁnds that only
around 25 percent of new jobs claimed by policy beneﬁciaries can be attributed to the
policy effect. The rest would have been created anyway. A similar number is presented

by Papke (1994). Moreover, Fisher and Peters (1997) argue that a policy effect of the job

 

32 Real income is considered throughout this paper.

33 Based on number of new and expanded facilities in terms of jobs, square footage, or dollars invested.
The ranking does not control for state population. According to the US Census Bureau, Michigan is the 8th
most populous state.

35

growth can be negative due to factor substitution. Thus, a careful evaluation of policy
effects on growth is necessary.

This article focuses on a group of policies initiated during 19905 that have been
implemented in a substantial proportion of Michigan counties. The effect of those new
policies is compared to a more traditional development tool: road infrastructure. The
policies considered for the analysis are the following:

o The Michigan Economic Growth Authority (MEGA) program grants eligible
businesses with Single Business and Income tax credits for 8 to 20 years. The program
targets “large-scale investment and job creation [and retention], as well as attraction of
technology-intensive business” (CRC 2001 , p. 20). The purpose of the program is to
support the projects that would not occur without the policy support”. There is no
targeting of speciﬁc geographic area. However, businesses located in federally
designated distressed zones have a lower eligibility threshold. Also some industry
restrictions apply.

The program started in 1995, and by the end of 2000 there were 107 authorities
located in 30 counties which directly and indirectly created 145,542 jobs”. The number
of MEGAs per 100,000 of county residents measures the treatment. The intensity of this
normalized treatment ranges from 0 to 14.3 and averages 0.3 MEGAs per county (see
Table 2.1 for more details).

Bradshaw (2002) reports that roughly 70 percent of new jobs generated by
expanded or new businesses are created during the ﬁrst year of the expansion. For that

reason we would expect much of the effect of MEGAs on job growth to be short run.

 

34 “... applicants mast certify that the project would not occur absent the MEGA grant.” (CRC 2001, p.
20).

36

New employment leads to additional labor income, which in turn is partially diverted to
neighboring communities by commuters. Thus, it is reasonable to expect MEGA’s effect
on income growth to be lower than the effect on job growth. However, the consumption
linkages of new workers and supply/demand linkages of ﬁrms beneﬁting from MEGAs
would amplify the policy effect in the longer run.

Even though county economic conditions do not enter the formal eligibility
criteria, endogeneity of MEGA’s effects may occur if more of the eligible business
establishments are attracted to counties with higher growth potential. If this were true, a
simple correlation between number of MEGA and growth would include upward bias.
The issue of endogeneity of policy treatment is addressed in more detail in section three
of this paper.

0 The Renaissance Zone (RZ) program provides ﬁrms and individual residents of the
state’s most economically distressed areas with a waiver of virtually all state and local
taxes for the purpose of revitalization of those areas (CRC 1998, 2001). This program
uses Michigan State Housing Development Authority deﬁnition of an “economically
distressed area” for targeting (CRC 1998). This deﬁnition is based on an area’s average
income level and change in real property values.

The program was established in 1996. In the ﬁrst round 11 zones were created in
14 counties for terms ranging from 10 to 15 years. The zone areas were from 110 to 2202
acres. The second round was in 2000 and nine new zones were created increasing the
total number of counties with the zones to 29. The creation of new zones ended in
December of 2002. By that time 34 Renaissance Zones were functioning. The RZ policy

treatment is measured in acres of the zone per 1,000 county residents. Intensity of the

 

35 Based on estimates by Michigan Economic Development Corporation.

37

treatment averages 0.7 acres per 1,000 residents with a range from 0 to 104 (see Table 2.1
for details).

The design of this program beneﬁts primarily small businesses and individual
residents (CRC 1998). For that reason, we would expect the policy effect to come
through higher income of self-employed residents and property owners. Papke (1994)
analyzed a similar program in Indiana, and found that the zone residents take only 4% of
all jobs and 15-20% of new jobs in the zone”. Thus, the spillover effect of the program is
expected to be large.

The estimated effect of R2 on growth may be biased downward. The policy
targets areas with lower potential for growth, and counties with larger per capita area of
the zone would also have more distressed areas, controlling for population density.
However, Wooldridge (2002) considering the example by Papke (1994) argues that the
presence of a policy treatment in a county is not conditional on county performance once
a zone is designated. Thus, there is no signiﬁcant endogeneity in policy treatment after
controlling for county speciﬁc conditions. The same argument can be applied to other
policies under consideration.

0 The Brownﬁeld Development Authority (BDA) program targets re-development of
“blighted”, “functionally obsolete” and contaminated sites (CRC 2001, p. 31). BDAs are
allowed to use tax increment for redevelopment of the brownﬁeld sites. In addition there
are state bonds “dedicated to brownﬁeld remediation”. Also, the BDAs projects are
eligible for Single Business Tax credits. Michigan adopted the program in 1996 and by

the end of 2000 there were 170 Authorities located in 53 counties. We use the number of

 

36 For review of literature on Enterprise Zone programs in US, see Papke (1993).

38

BDAs in a county per 100,000 residents to quantify the policy treatment. Its intensity
averages 0.4 authorities per county (see Table 2.1 for details).

The immediate job growth effect generated by the BDA37 program is expected to
be small because it is limited to temporary jobs associated with the environmental
remediation phase of the redevelopment effort. However, BDAs bring previously
underutilized land into active economic use increasing property value. For that reason the
effect of BDAs on income growth is expected to be positive and signiﬁcant. Two factors
operate in opposite directions in terms of bias. More distressed areas would potentially
have higher number of brownﬁelds available for redevelopment, but on the other hand
higher demand for sites is expected in faster growing areas.

0 Improvement in highway infrastructure is another development tool. It may lead to
a reduction in commuting time and transportation cost, contributing to long-term growth
of both income and number of jobs. The effect of highways is measured in terms of miles
of federal, state and major county highways and roads per 1,000 residents. Michigan has
on average 3.8 miles per 1,000 county residents (see Table 2.1 for details). Construction
of major roads follows a long-term plan and is funded primarily through state and federal
budgets, reducing the bias in the policy treatment.

The data on the policies come from the Michigan Economic Development
Corporation, Citizens Research Council of Michigan and the State of Michigan
government. Policy treatment is measured as of January 1 of a given year.

The set of policies under consideration does not include some important programs
such as direct incentives, job training, loan guarantees, subsidies, etc. This exclusion is

driven by the data availability and spatial scale of policy implementation. However, the

39

policies considered in this study are good representatives of four groups of development
incentives widely used in other states. MEGA represents a class of ﬁrm-level tax
abatements. R2 are akin to a wide range of programs targeting revitalization of
economically distressed areas (e. g., Enterprise Zones). BDA represents programs that do
not have job growth or business creation/expansion as a primary objective. And ﬁnally,
highway system expansion is a typical example of infrastructure improvement activity.
Moreover, “[of] all the public services examined for an inﬂuence on economic
development, transportation services, and highway facilities especially, show the most

substantial evidence of a relationship” (Fisher 1997, p.54).

2.3. EMPIRICAL MODEL

There are several approaches to the development policy analysis. Among other
case studies, simulations and econometric analysis are used most frequently. Each of
them has its own beneﬁts and weaknesses (Bartik 1991; Sadoulet and deJanry 1995;
Fisher and Peters 1997; G053 and Phillips 1999). Econometric analysis is the most
appropriate to use for testing heterogeneity of policy effects with the policy data
described above.

Simple correlation analysis of policy treatments and growth will likely to produce
biased estimates of policy effects since, as we argue above, some policies target
speciﬁcally areas with lower growth rate, while some others are likely to be applied more
in the areas with higher potential for growth. However, the fact that policy treatments are
assigned by state or federal agencies partially reduces endogeneity of policy shock to the

current county growth conditions. Other sources of endogeneity are also possible. One is

 

37 For discussion see Mayer and Lyons (2000), DeSousa (2000)

40

due to endogenous migration or sorting”. That is that areas affected by policies attract
population with some common unobservable characteristic correlated with growth.
Another source is due to normalization of policy treatment in our study by population.
Since the size of population responds to the growth conditions through migration (Goetz
1999), faster growing regions would experience endogenous reductions in per capita
policy treatment, biasing estimates downward.

There are two ways to circumvent the endogeneity problem. First is to estimate
the policy effect using instrumental variables (e.g., G033 and Phillips 1999). Second is to
control for policy selection criteria within the estimation model. Unfortunately, a
plausible set of instruments for the policies under consideration is not available. That
leaves us with the second approach, which, however, does not guarantee control for all
sources of bias (Besley and Case 2000). Nevertheless, we take several steps to minimize
the uncontrolled endogeneity.

First, we eliminate bias in normalized policy treatment created by endogenous
population change in the following way. The treatment is normalized by population in a
year of implementation of that treatment. This normalized treatment is carried over time
with no accounting for change in population in later periods. Any changes in the policy
treatment itself (e. g., a new authority or additions to a Renaissance Zone) are normalized
by population in a year of that change and added to the amount of treatment from the
previous period.

Second, we include into the estimation model controls for policy targeting.

The ﬁrst candidates for the controls would be income and job levels per capita since

 

38 We do not consider an endogenous sorting as a serious problem. If that sorting (business relocation and
population migration) is a response to the treatment, we consider it as a part of policy outcome for a county,

41

policies target areas with low income and employment. However, those variables
are endogenous to the growth by construction of the variables (future level is a
function of the current level and the growth rate - which is ruled out for the ﬁxed
effect estimation procedure used in our analysis). Instead, we include the following
variables that may impact the targeting and inﬂuence the growth. The share of
people with a bachelor’s degree, denoted by educ controls for the level of education
(or human capital) of a county population. The number of jobs per capita in
manufacturing (manuf), government (gov) and farming (farm) serve as controls for
the industry structure of each county’s economy. Density of business establishments
(dens) controls for the degree of concentration and urbanization”. These variables
explain also a large portion of variation in income and job levels.

Design and implementation of the policies takes time. Thus, county economic
conditions should be controlled in a year when the policy designation decisions are made.
Therefore, the control variables are used with a one-year lag. This also allows avoiding
simultaneity bias in the control variables.

The resulting estimation model takes the following form (2.1)“):

(2.1) GR, '4 = ﬂo + ,6, policy, + ,6, educ” + ,63 manuﬁ-1 + ,6 4 gov ,-1

+ ﬂ5farm ,./+ ,B6dens,-/+ ,6 7year, +e+u,,

 

even though no effect may be observed at a macro-level.

39 In particular, the density variable helps control for possible systematic differences in policy application
between rural and urban areas.

40 The indexes for counties are suppressed.

42

where GR,--, is the growth rate between years t and t’ (t’>t). Vector e denotes
unobserved county conditions that are ﬁxed over time, and u is an idiosyncratic
error.

Policies may have different primary objectives (e. g., growth of large vs.
small businesses, growth vs. redistribution of income, etc). Two aspects of policy
objectives, income and job growth in a county, are considered in turns. Thus, GR,--,
represents in turn each of the two growth outcome measures.

The timing of the growth response to the policy treatment is an important issue.
Some of the responses, such as change in production and employment levels, are
relatively fast, while others, such as relocation and changes in business birth rates, may
take a long time to occur. Also, some policies may have a short-lived effect, while others
impact long-run growth patterns. This timing differs also for different development
outcomes (e. g., job vs. income growth) (Papke 1994).

One may argue that short-run ﬂuctuations in a growth rate may be driven
primarily by ﬂuctuations in demand and industry specialization. It would be problematic
to separate a policy effect if designation of the policy treatment is also driven by these
factors. These issues represent a major challenge for our empirical analysis. To address
this issue, we ﬁrst consider two measures of the growth outcomes: over one-year period
(annual growth) and three-year average growth rates. The annual measure is designed to
capture the most immediate policy effect (if any). The second measure averages out some
of the annual ﬂuctuations, and is designed to capture more persistent policy effects.
Secondly, to control for demand and industry ﬂuctuations (which are most likely to be

common for most of the state) we include a set of year binary variables (year) together

43

with measures of industry structure (as described above) into the model. The year
dummies control also for common shocks such as inﬂation, national business activity
conditions, etc.

A vector of normalized policy treatments (Policy) is of primary interest. To test
the hypothesis about the cross-policy effects, the interaction terms among the policy
treatments are also included in the base model. Nonlinearity of the growth response is
tested by inclusion of square terms for policy treatment. To test for heterogeneity of
policy effects, the model is estimated separately for counties with high and low level of
per capita income. The difference between counties with high, and low per capita number
of jobs in manufacturing, and between metropolitan and non-metropolitan counties are
also considered and discussed brieﬂy.

The model omits a number of factors relevant to growth. Among them are local
taxes, natural amenities, quality of local government, spatial interdependence of the
regions, etc. Most of them can be a source of omitted variable bias since they may
determine the application of policy. However, these factors have little variability over
time. We assume also, that other conditions that describe county eligibility for the
treatment or attractiveness of a county to MEGA eligible businesses stay stable over time,
conditional on other controls. Using eight periods of annual data is consistent with this

assumption“. A county speciﬁc effect (e) denotes all time invariant conditions.

 

41 In using ﬁxed effect transformation of the data, we have to consider a tradeoff between the number of
factors that stay stable and the time horizon. The longer the time span, the fewer the number of factors that
stay constant and, thus, are controlled by this estimation technique. The eight-year time period, considered
in our analysis, is thought to be a good balance. It is reasonable to assume that unobservable county
conditions, which impact an amount of policy treatment are stable over this time period conditional on
other variables in the model. On the other hand, Figure 2.1 shows that we have a large variation of, for
example, annual income growth around the state average of 1.8 percent per year over 1993-2000. The
policy treatment has also a signiﬁcant variation.

44

A ﬁxed effect Feasible Generalized Least Squares (FGLS) procedure is used for
estimation of the model. The main beneﬁt of the selected procedure is that it helps control
for unobserved ﬁxed factors in the growth model (e) and provides efﬁcient estimates of
standard errors. Fixed effect estimation reveals how the change in the treatment for the
same entity associates with change in the growth rate controlling for other factors. The
FGLS procedure is particularly helpful in estimating three-year average grth where a
signiﬁcant serial correlation in error is present. Alternative procedures are mentioned in

section ﬁve of this paper.

2.4. DATA DESCRIPTION

Local Area Annual Estimates (BEA 2005) is the primary source of the data for the
dependent and control variables for 83 of Michigan counties over the years 1993 — 2000,
making a total of 664 annual county-level observations. This time frame is determined by
the data availability. To get consistent FGLS estimates, the year of 1993 is not used for
estimation after ﬁxed effect transformation is performed on the full set of data. Thus,
only 581 observations are used by the FGLS procedure.

Four dependent variables are used in our analysis, making a set of four equations.
They are annual and three year average growth in personal income and number of jobs.
As an example, consider the year 2000. The annual growth for this year is change in
income or job levels between years 2001 and 2000. Three-year average grth is
computed as the change between the years 2003 and 2000 divided by three. Figure 2.1
depicts the patterns of both growth outcomes in Michigan over 1993-2000 period. There
is a 58 percent correlation between annual and three-year average income growth and 71

percent correlation for the job grth measures. There is a 45 percent positive correlation

45

between the annual income and job growth for Michigan counties. Thus, these variables
reﬂect some different aspects of economic growth.

The average annual income growth for Michigan counties was 1.7 percent,
varying from minus 11.6 to positive 14.7 percent. Jobs grew at an average rate of 1.1
percent, ranging from minus 15.8 to plus 24.0 percent (see Table 2.2 for descriptive
statistics). In our sample, the average share of adults with a bachelor’s degree (used to
control for education level) is 8.1 percent. There are on average 101.1; 68.1 and 5.6 jobs
per thousand county residents in manufacturing, government and farming respectively.
The average concentration of business is 25.1 establishments per square mile.

We pay particular attention to potential heterogeneity of the policy effects across
different types of local economies. To compare the growth patterns, we have divided the
state into counties with above and below the median per capita personal income levels,
called “high” and “low” income counties respectively. To avoid artiﬁcial truncation, we
consider the average income level over eight years for each county for the assignment
into the appropriate income group.

The other dimensions of potential heterogeneity are industrial structure of the
local economy and degree of urbanization. Counties with more and less than median per
capita number of jobs in manufacturing are treated as separate sub-regions for that
purpose. The average number of jobs over eight years for each county is considered for
this purpose. Counties are also distinguished as metropolitan and non-metropolitan. The
metropolitan part includes 41 Michigan metropolitan and metropolitan adjacent counties
(Beale Codes 0, 1, 2, 3, 4, 6 and 8). The non-metropolitan part includes other 42

Michigan non-metropolitan, relatively rural, counties.

46

There is a strong correlation among the different classiﬁcations of county
economies. For example, 39 percent of metropolitan counties also have a high number of
manufacturing jobs, and 45 percent of metropolitan counties are in the high income
category. About 55 percent of the high income counties have also a high number of
manufacturing jobs. However this correlation is not perfect, which provides an
opportunity to use those characteristics as potentially different dimensions of policy

effects.

2.5. RESULTS

The estimation results are presented in Tables 2.3 and 2.4. Each table contains
ﬁxed effect F GLS estimates of policy treatments and control variables on one and three-
year average growth rates”. The selected estimation procedure was compared with
random effect and ﬁrst difference estimation procedures, and was more efﬁcient”. The
results of the ﬁrst difference and ﬁxed effect estimations are similar, implying that our
model provides sufﬁcient control for endogeneity (Wooldridge 2002).

We tested if the inclusion of the control variables is helpful by regressing the
treatment by each policy on the controls and on other policies. The results indicate that
the controls are statistically signiﬁcant in explaining the variations in policy treatment.
We found a systematic cross-policy relationship (see Appendix 2.1 for details). For that
reason, the policy variables are interacted with each other and de-meaned’“. Thus, the
estimates represent marginal policy effects on grth in a county with average treatment

by other policies.

 

42 Stata‘m 9.0 was used for estimation.
43 As expected when the standard assumptions hold.

47

2. 5. 1. Base Model

The estimates of the policy effects for all Michigan counties (the base model) are
presented in Table 2.3. The results for the income growth are in Columns 1 and 2, and for
the job growth in Columns 3 and 4. Very few individual policy coefﬁcients are
statistically signiﬁcant. The directions of policy effects on income and job growth are
somewhat similar.

The effect of highway improvements is positive, but decreasing (the quadratic
term is negative) and has a relatively large t-statistic. The results on three-year average
job grth (Column 4) imply that one additional mile of highways per 1,000 residents
increases the average growth rate by 1.2 percent (for a county with average values of
other policy treatments). The effect of every additional mile is decreasing by 0.1 percent.
The turning point of 12.6 miles per 1,000 of residents is within the range of values in the
sample. However, it is far above the state average of 3.8 miles. It implies that highway
improvements have the potential to facilitate job growth in most Michigan counties. The
effect of the highways is higher in counties with larger areas in Renaissance Zones (the
interaction term between the highway’s and RZ’s treatments is positive and signiﬁcant).

Other estimates of the policy effects (which are occasionally signiﬁcant) do not
provide a strong basis for judgment. Policy variables are jointly signiﬁcant in explaining
variations in county growth rate (with exception for one-year income growth — Column 1
- see Chi-square statistics for all policy variables at the bottom of the table 2.3). The joint
policy effect comes primarily through a direct policy impact (as opposed to the

interactions with other policies). Interaction terms are jointly signiﬁcant in affecting the

 

44 Without de-meaning the estimates would represent the marginal policy effect in a county with no

48

three-year average growth (Columns 2 and 4 - see Chi-square statistics for policy
interactions).

Most of the estimates for the control variables are statistically signiﬁcant. Since
these variables have signiﬁcant correlation with policy treatments (Appendix 2.1), they
provide a strong control for unobservable policy targeting. The model has a sufﬁcient
power in explaining variation in growth rates. Both log—likelihood and Wald statistics are
high (see the bottom of the Table 2.3).

Even though we found that the policies are jointly signiﬁcant in explaining
income and job growth, and that the interactions of their effects are signiﬁcant, there is
not much to say about the effects of individual policies. This is the point where many
researchers would stop further investigation of the issue, blaming downward selection
bias, lack of reliable instrumental variables, and multicollinearity. However, we
hypothesize that policy effects are heterogeneous and change with the type of local
economy. This unaccounted heterogeneity hides statistical and practical signiﬁcance of

the policy effects.

2. 5. 2. Policy Effect Heterogeneity

What are the implications of the heterogeneity problem? If this heterogeneity
exists, but is ignored, the estimation results would present a weighted average of the
effects in the separate (relatively homogeneous) sub-regions. The standard error would be
also inﬂated. Moreover, the results will be inconsistent when a larger population is
considered and the proportion of the sub-regions changes. To show whether

heterogeneity is present, we estimate the model separately for counties with high and low

 

treatment by other policies (interaction term equal to zero in this case).

49

income. A summary of the results is presented in Table 2.4. Columns 1-4 contain the
estimates of the income growth. The estimates of the job growth equation are presented
in columns 5-8. Similar separate equations are estimated for metropolitan/non-
metropolitan counties and for counties with above and below the median number of jobs
in manufacturing (results are summarized below).

Testing the signiﬁcance of the difference in the estimation results between the
sub-regions involves the following steps. First, we interact the explanatory variables with
binary variables for the sub-regions. Second, we insert these new interactions into the
model and estimate this augmented model on the full set of observations. Statistical
signiﬁcance of the interactions of the policy variables with the sub-region dummy
indicates the signiﬁcance of the difference in policy effect between the parts of the state
under consideration. In the results tables we have denoted with boldface font the
estimates that are individually different at 10% signiﬁcance level between the sub-
regions (estimates of the difference are not presented in the paper)“. The joint
signiﬁcance of the differences is assessed with Chi-square statistics presented at the

bottom of each panel.

Heterogeneity with respect to the income level

Our exploration of heterogeneity in policy effects focuses on the differences with
respect to income level. The estimation results in Table 4 indicate that the difference in
policy effects is statistically signiﬁcant jointly for the three-year average growth for both
income and job growth outcomes (see Chi-square statistics for the difference). The joint

signiﬁcance of the policy effects on job growth is higher in lower income counties (see

50

Chi-square statistics for all policy variables — Columns 7 and 8). It implies that difference
in income levels is an important source of heterogeneity.

The individual differences in policy effects can be easily seen when the model is
estimated separately for high and low-income counties. The individual difference is
observed for most of the variables. Some of these differences are statistically signiﬁcant.
In particular, the BDA effect on the annual growth becomes statistically different
between the sub-regions of the state. The effect on annual income growth (Column 1)
becomes individually signiﬁcant in counties with high income. The difference implies
that the linear effect is more negative in these areas. The effect becomes positive more
quickly than in the low-income counterparts (the quadratic term is positive and larger).
The effect on the annual income growth becomes positive in the high-income counties
after the point of about nine BDA per 100,000 residents. The effect of this policy on the
three-year average job growth (Column 8) is positive and signiﬁcant in low-income
counties. Comparing the results for BDAs with the previous table (where no effect was
detected) we conclude that uncontrolled heterogeneity of a county economy with respect
to the income level hides the signiﬁcance of this policy effect.

The effect of highways is also found to be different with larger effect in the high-
income counties. Another conﬁrmation to the heterogeneity hypothesis is the fact that the
linear effect of the highways on the three-year average income growth differs in sign and
magnitude (Columns 2 and 4), while the results from the previous table are insigniﬁcant.

Some signiﬁcant differences in the effects between the high and low-income

counties are found for MEGA and some of the policy interactions. However, none of the

 

45 This and other results not reported in the paper are available from authors upon request.

51

RZ effects becomes statistically signiﬁcant. The direction of policy effects on income and
job growth are consistent with the previous table.

The gain from estimation of individual policy effects by separate sub-regions is
not always large. The weakness of heterogeneity evidence (for RZ in particular) is due to
several possible reasons. First, the primary source of heterogeneity in the policy effects
may be different than the income level. Second, the deﬁnition of the sub-regions may be
imperfect. Alternatively, there may be no heterogeneity or no signiﬁcant policy effect.

We explore some of these issues in turn.

Metro/ Non-Metro Difference

We continue the discussion of heterogeneity by considering the difference in the
policy effects between metropolitan and non-metropolitan areas. The difference in the
joint effect on income growth is signiﬁcant (results are not presented). It implies that
joint policy impact is different in metropolitan and non-metropolitan counties. It serves as
additional evidence in favor of the heterogeneity hypothesis. The job growth model is not
different statistically.

The individual difference is most obvious (and signiﬁcant) for the effect of R2 on
income growth, and highways on job growth. The effect of RZ on three-year average
income growth in metropolitan counties becomes positive and signiﬁcant, but with
decreasing rate. The turning point is around 26 acres of zone per 1,000 metropolitan
residents. This point is within the range of the treatment values, but is far above the
average of 0.7 acres. It implies that metropolitan counties can beneﬁt more from
application of this policy. Also, an increase in application of this policy in counties with

more than 26 acres per 1,000 residents is counterproductive.

52

The effect of the highways on job grth is higher in non-metropolitan counties.
It is positive but with decreasing rate. The turning point is around 13-16 miles per 1,000
non-metropolitan residents for both the annual and three-year average growth. This point
stays within the range of the policy treatment values in the sample, but is above the
average. This result implies that the highway improvements are an effective tool in
stimulating job growth in non-metropolitan areas. Considering this along with the RZ
results we can also conclude that the source of heterogeneity is different for different
policies, but heterogeneity stays as a signiﬁcant source of error in estimation results that

hides the true policy effect.

Heterogeneity with respect to industry structure

Heterogeneity of policy effects with respect to industrial structure of the local
economy adds one more dimension to the analysis. The results are similar to the case of
metropolitan/non-metropolitan heterogeneity. However, the policy effects are not
signiﬁcantly different among the sub-regions (with few exceptions).

Even though statistical and practical signiﬁcance is observed, results should be
interpreted with caution. The measures of heterogeneity are far from perfect. For
example, Papke (1991) shows that the effect of infrastructure varies signiﬁcantly across
industries within the manufacturing sector. The measure of urbanization used in this
study might not capture all the differences between urban, fringe and rural areas. Also,
other sources of heterogeneity such as unemployment rate are possible. For example,
while Bartik (1991) and other authors (e.g., G055 and Phillips 1999) argue that the policy
effect should be higher in regions with higher unemployment rate due to higher social

beneﬁts, others argue that when policies stimulate labor demand under low

53

unemployment, more people would be attracted to the labor force and thus the effect
should be higher under low unemployment.

Measures of the policy treatment could also be better. Highway miles do not
measure all of the changes in quality of the infrastructure. Moreover, selection bias might
not be completely excluded. In addition, there are costs associated with the use of the
policy tools, which may offset the beneﬁts. Those costs are not controlled for in our
model.

The measures of growth also have weaknesses. Averaging reduces all of the
short-terrn effects. If a policy has an effect on, for example, one-year growth but with
some time lead, this effect would be underestimated. On the other hand, averaging
reveals longer-term effects even if they are relatively small comparative to the short-run
noise. Nevertheless, the results provide some insight on how the policy effects change

over time and space.

54

2.6. CONCLUSIONS

In this paper we implemented a quantitative assessment of the impact of Michigan
economic development policies and highway infrastructure improvements on personal
income and job growth. We ﬁnd that the policies have signiﬁcant impact on the growth
outcomes. However, the effects vary across different types of local economies. Ignoring
this heterogeneity leads to underestimation of the policy effects. Signiﬁcantly different
policy effects are detected between areas with high and low levels of income and
different urban structure. The lines along which the individual policy effect may change
are different for different policies. For example, Brownﬁeld Development Authority
treatment is more beneﬁcial for low-income communities, which may lack sufﬁcient
market pressure for redevelopment of underutilized land resources. The highways beneﬁt
more non-metropolitan counties where the reduction of transportation cost may be more
signiﬁcant. Renaissance Zones beneﬁt more metropolitan counties. The difference in the
Renaissance Zone’s effect may be due to the relatively higher beneﬁts provided by the
policy in metropolitan areas where businesses and residents may experience higher tax
loads, and where agglomeration effects may be higher. While determining the precise
sources of heterogeneity requires additional study, it is clear from our results that
economic development resources can be used more efﬁciently if policies target the areas
most likely to produce higher returns.

The marginal effects of policies change with the scale of policy treatment,
implying that there is a minimum threshold necessary for a positive impact. Also there is

a point after which policies become counterproductive. Signiﬁcant interactions between

55

 

the policy effects are detected. The results imply that the policies have mutually
reinforcing effect in most of the cases.

The policies have little correlation with ﬂuctuations in annual growth. However,
they change more persistent growth patterns as measured by the three-year average
growth rate.

Some of the results have broader implications. The policy treatment such as a
Renaissance Zone can be considered as a change in tax regime external to local residents
and businesses. Thus, implementation of the policy can be treated as a “natural

”46. The results imply that lower taxes stimulate local growth holding ﬁxed

experiment
other factors (such as quality of infrastructure and public services).

Comparison of policy effects on alternative growth outcomes opens a wide range
of opportunities for analysis. Job growth should correspond to income growth trends
since wage is a part of personal income. Keeping this in mind, we can look at the
difference in policy effects between income and job growth as changes in income
distribution patterns. For example, if we observe a positive effect on job grth but not
on income growth we can suspect that either part of the jobs go to commuters or that a
substitution effect in the sources of income takes place (e. g., unemployment subsidies are
substituted with some wage income without raising well-being). The other extreme point
would be when we observe a positive effect on income growth without growth or even
with decline in the number of jobs. This situation implies that the source of income

growth is either an increase in property value, or increase in proprietary income, or

substitution of labor for other factors. The ideal policy would have positive impact on

 

46 Bartik (1991, 1997) argues that taxes and public spending are highly endogenous to growth and without
an external shock it is impossible to determine their impact with any precision.

56

both growth outcomes. Our results show that in general Michigan policies have similar
impact on both income and job growth. It implies that Michigan economic development
efforts are beneﬁcial to the well being of state residents. However, more rigorous analysis
of this issue is required.

In the light of the current trends in economic development when communities,
especially in rural areas, are increasingly vulnerable to globalization processes, it is
extremely important to tailor the development policies to the speciﬁcs of the targeted
areas. Our results show that one-size-ﬁts-all approach is not appropriate for regional
economic development. The results can be used to set priorities among different policy
tools and contribute to better understanding of the development policy effects. They also

allow a more accurate expectation of the policy impact.

57

ESSAY 3. DOES FIRM SIZE DISTRIBUTION MATTER? AN EMPIRICAL

TEST OF THE ENTREPRENEURIAL PIPELINE THEORY

3.1. INTRODUCTION

Encouraging entrepreneurship has been a subject of concern in the economic
development literature for some time (Birch 1987; Hanham et al., 1999; Johanisson,
1991; Walstad, 1994), but has generally failed to capture the attention of mainstream
local economic development practitioners (Loveridge, 1996). The practitioners continue
to focus on industrial recruitment and business retention. These traditional approaches
must be reassessed as increased competition from lower wage countries in both
manufacturing and services makes successful recruitment and even retention ever more
difficult. These new challenges are shifting the focus of regional development to the role
of amenities in driving job creation (Green, 2001; Florida, 2002), and fostering a renewed
attention to the role of entrepreneurship in local development (Von Bargen et. al., 2003;
Walzer, 2004).

Much has been written on how to make it easier and less risky to start businesses
(Watson et al. , 1998; Sohl and Rosenberg, 2003), and on identifying start-ups with high
growth potential (Birch and Medoff, 1994). However, whether a community has enough
of entrepreneurs remains an unanswered question. The expression “too much of a good
thing” might well apply to the supply of small businesses in some communities, while
other communities rely heavily on employment by a small number of large businesses.

Neither of these extremes is likely to be good for sustainable growth in a long run.

58

Lichtenstein and Lyons (2006) raise this distribution issue. They argue that small,
medium-size and large businesses are interdependent within a regional economy, and that
there are entrepreneurs at all sizes of businesses. Entrepreneurial development cannot be
limited to a small business. Moreover, some small businesses lack entrepreneurial
characteristics (Muske and Woods 2004; Lichtenstein and Lyons op. cit. ). On the other
hand, mid-sized and large ﬁrms can be very entrepreneurial (Lichtenstein and Lyons, op.
cit). Entrepreneurial development requires a balanced distribution of businesses within
local economy. The interdependence of businesses comes from the local production and
consumption linkages as well as from various externalities (e. g., small businesses create
local amenities, large businesses create a critical mass for many local business services,
etc.). Businesses of different sizes have mutual beneﬁts in developing entrepreneurial
skills, having successﬁil role models, and training the region’s labor force. Illustrations of
such interdependences include, for example, the health care industry, where large
regional hospitals are often accompanied by a number of small clinics, individual
practitioners and medical equipment suppliers. Another example is the automobile
industry, where large companies are supported by small subcontractors. Lichtenstein and
Lyons suggest that regional entrepreneurial growth depends on how easily local
entrepreneurs can improve their skills and expend their businesses. It follows also that an
appropriate economic development strategy may differ in response to a business mix in a
particular region. Reﬂecting this fact, Lichtenstein and Lyons (2006) deﬁne three groups
of strategies: small business incubating, improvement of overall business performance,

and selective attraction of large businesses.

59

This paper explores whether that balanced distribution or, in Lichtenstein and
Lyons’ (2006) term, a community’s “pipeline of entrepreneurs” has an empirical
relationship with measures of overall economic growth. An index is introduced that
summarizes the distribution of employment across businesses size categories, and then
the relationship between the index and county economic grth is estimated.

To make the analysis robust to unobserved interstate variation in growth
conditions such as natural amenities, the legal system, culture, or other statewide
institutions, the study focuses on only one state, Michigan. This approach also excludes
variance in macroeconomic factors as potential source of bias. However, the ﬁndings are
applicable to other regions of the United States and other countries.

The paper is structured in the following way. Part 3.2 reviews literature on the
relationship between business size and economic growth, and provides further
motivation. Part 3.3 discusses the issues of measuring business distribution, the empirical
model, and describes the data. Part 3.4 presents the results and implications, while Part

3.5 concludes.

3.2. BUSINESS SIZE DISTRIBUTION AND IMPLICATIONS FOR ECONOMIC
GROWTH

There is a debate in the literature and among practitioners whether regional
economies beneﬁt more from incubating small business start-ups, facilitating expansion
of existing businesses, or from attracting large business establishments from other
regions. On one hand, there is a common perception that small business is the major

contributor to job growth and economic development (e. g., Birch, 1987). Small business

60

is associated with entrepreneurship and innovation — widely accepted engines of growth
(Malecki, 1994; McQuaid, 2002). Empirical studies reviewed by Martinez (2005) found
positive correlation between output growth and share of small business. Also, a larger
share of small and medium-size businesses can make a local economy more ﬂexible and
less vulnerable to industry-speciﬁc shocks.

On the other hand, Davis et al. (1993, 1994) sharply criticize this perception.
Looking at business establishment level data for manufacturing, they argue that the
association of small business with job growth is driven by endogenous transition of the
establishments among the size categories and by disregarding the job contraction, which
is also higher among small businesses. They conclude that there is “no strong or
systematic relationship between net job growth rates and either ﬁrm or plant size [in
manufacturing]” (Davis et al., 1993, p.25). This result is similar to earlier ﬁndings in the
literature (see for review Lucas, 1978). In addition, not every small business is
entrepreneurial (Muske and Woods, 2004). Many new small businesses, especially
branch plants, franchises, retail shops etc., are replicas of existing business approaches
(Malecki, 1994). Moreover, the link between the growth of small business and income
grth is even more questionable. Jobs in the small business sector are on average less
persistent (Davis, 1993), have fewer fringe beneﬁts, and lower wages (Ettlinger, 1997).
These arguments provide a rationale for practitioners to devote substantial resources to
attraction of large businesses to their regions.

However, the conclusions by Davis et al. are also subject to critique. Changes in
employment by individual business establishments are not independent of one another.

Outsourcing of jobs by large businesses to smaller counterparts may support the job

61

creation role of small business, while on the economy-wide scale there may be no growth
at all. Moreover, business establishments are linked by production process and ownership
ties, which make the growth of one establishment conditional on the growth of others.
Also, investments into large business attraction may never pay-off since businesses tend
to act strategically in requesting the development incentives (Gabe and Kraybill, 2002)
and they may move to another state or abroad.

An interesting conceptual reconsideration of the relationship between the business
distribution and regional economic development is provided by Lichtenstein and Lyons
(2006). They argue that effectiveness of business incubating, improvements of overall
business performance, or selective business attraction strategies depends on speciﬁc
conditions of a business mix in the local economy. To describe this mix, Lichtenstein and
Lyons introduce the notion of an “entrepreneurial pipeline”: a two-dimensional “map” of
business distribution within a community. These dimensions are the life stages and
entrepreneurial skill levels of businesses. The pipeline helps describe the
interdependences that exist among the businesses within a region that are beyond
production and consumption. These interdependences include peer pressure of the
businesses at the same level of development, and the role models of businesses at a more
advanced stage. It follows that it is not any single part of the pipeline (e. g., large scale
plants) that are more important than the other, but rather, the important thing is to have
sufﬁcient numbers of businesses in all categories. In other words, communities should
have a balanced distribution of businesses. Lichtenstein and Lyons argue that
development policies should be community speciﬁc and target development of that part

of the business distribution that is underrepresented. If, for example, a community suffers

62

a low business start-up and survival rates economic developers should rely more on the
incubating strategy, while communities that do not have a sufﬁcient number of more
mature medium and large size businesses should put more efforts into selective attraction.
Communities that seem to have sufﬁcient proportion of businesses at all life stages and
all levels of entrepreneurial skills would beneﬁt from employing development strategies
that target overall improvement in business performance such as infrastructure
improvements, upgrading of education, communication and public services.

This paper aims at exploring the validity of Lichtenstein and Lyons’ “pipeline”
conceptual framework through an empirical test. However, to do this test, the notion of
the “pipeline” needs to be operationalized. Even though not all the small businesses are
new, the life stages of business development are closely linked with the size of
businesses. Before getting big, all businesses started as small (Lichtenstein and Lyons,
2006). Even though there are entrepreneurs at business of all sizes, it takes an
entrepreneur with a higher level of skills to successfully operate large business (Lucas,
1978). Thus, the size distribution of businesses may reﬂect both dimensions of
Lichtenstein and Lyons’ “pipeline”, so, the size distribution of businesses within a local
economy is used to test the concept. In particular, we test whether there is a signiﬁcant
link between the distribution and two common measures of local economic development:

income and job growth rates.

63

3.3. EMPIRICAL MODEL AND DATA
An empirical test of the pipeline hypothesis faces two challenges — developing a
tractable measure of the ﬁrm size distribution, and estimating that measure’s relationship

to local growth. We address each of them in turn.

Measures of the Size Distribution

County Business Patterns (Census Bureau, 2005) provides data on the number of
business establishments in nine employment size groups by county. The smallest group
includes establishments with one to four employees, while establishments in the largest
group employ more than 1,000 workers. Figure 3.] depicts the shares of business
establishments by the size groups for Michigan for the years 1988 — 2000.

As one might expect, the average share of business establishments in each size
group decreases with increasing size of the business (Figure 3.1). The range of the
county-level values for each group also decreases from 40 percent (44 to 84) for the
smallest size group (1 to 4 employees) to a range of 0.5 percent for the establishments
with more than 1,000 employees (Table 3.1, Panel A). In each category above 19
employees, there is at least one county with no ﬁrms in that size range (minimum share
of establishments is zero). This distribution is typical for the Great Lakes region of the
United States.

However, the establishment distribution does not reﬂect the fact that larger
businesses have more impact on the local economy. Thus, the number of establishments

in each category is weighted by a measure of employment size. This way, the distribution

64

of employment shares across business size categories (groups)47 is constructed. The
resulting employment distribution is surprisingly close to uniform48 (Figure 3.2). If the
employment shares were equal across the nine categories, each group would have 1/9 or
11 percent of total employment. This equal share point is within one standard deviation
from the mean for ﬁve out of nine groups and within two standard deviations across all
groups.

This distribution of employment among the business size categories can serve as
the basis for exploring the pipeline theory. A starting hypothesis is that the deviation of
the employment shares from a uniform distribution has negative impact on the growth of
the regional (county) economy.

To test the entrepreneurial pipeline concept the following steps are taken. First,
we regress county growth rates on the shares of employment in each size group. If the
pipeline theory is valid, then the measures of the distribution should have signiﬁcant
correlation with the growth controlling for other covariates”. If, in addition, the optimal
distribution is uniform, we would expect the shares of the underrepresented groups (those
with average below the equal share: 1-4, 250-499, 500-999, 1000<) to have a positive
association with growth. The shares of the over-represented groups would have a

negative association. An alternative way to test the same hypothesis is to look at the

 

47 The ideal way to construct the weights is to use the average size of a business establishment within each
group. However these data are not readily available, so we use the lower size bound in each group as a
weight. This makes our employment distribution over-weight the smaller businesses in each size category.
The error is larger for the size groups with wider employment ranges. Nevertheless, the reconstructed
distribution of employment serves the purposes of the current line of inquiry. Tests involving application of
alternative weights did not change our empirical results (alternative results are not reported).

48 The share of the size categories encompassing a wider range of values is underestimated.

49 The study omits the effect of non-employer businesses (establishments with zero employed workers) on
regional growth. This group includes self-employed entrepreneurs and non-active businesses. The effects of
these two types of non-employer businesses go most likely in opposite directions. If the net effect is
positive and the share of this business category correlates positively with share of small business, than the

65

 

absolute deviation of each group from the equal share. In this setting a larger deviation is
hypothesized to be associated with lower growth.

Each approach, however, would suffer from multicolinearity among the size
groups, and would provide only limited insights into the optimal growth enhancing
distribution. To enhance the analysis, the distribution needs to be represented with a
single statistic. As an example, we have constructed a business distribution index (3.1),
which is the sum of the absolute deviations of the employment shares in each group from

the uniform distribution (11%).

(3.1) Index = gill xi —1 1%|
where x is the share of employment in each of the nine size groups respectively. By
construction, a positive deviation from 11% in one category is offset by an equal negative
deviation in other categories. Thus, by summing absolute values, the size of the deviation
from a perfectly even distribution is counted twice. We therefore divide the total Stun of
absolute deviations by two.

The interpretation of the index is straightforward. It represents the share of total
employment that would have to be reallocated among the groups to achieve a perfectly
uniform distribution of the employment across the business size categories. The larger is
the index, the more unequal is the current distribution of employment across the size
groups. If the uniform distribution is optimal for growth, the index should have a negative

association with measures of growth. However, if the equal distribution is not optimal,

 

effect of the latter on grth would be overestimated. However, this omission does not cause any selection
bias since the grth outcomes are observed for all counties for each year.

66

ugh——

there is a positive value of the index (optimal deviation from the uniform distribution)
that maximizes growth.

For Michigan counties the average value of this index is 20.35%. It ranges from
12.88% to 61.47% (see Table 3.1, Panel C). The distribution of the index across
Michigan counties (averaged over the study period) is interesting as well. The southern
part of the state, which includes all Michigan metropolitan and metropolitan adjacent
counties, has a lower value of the index than the northern (primarily rural) counties
(Figure 3.3). This map may support the initial hypothesis that more equal distribution of
employment (captured by lower values of the index) may be attributable to more dynamic
local economies. Whether lower values of the index are associated with a higher grth
is tested by plotting the values of the index against two measures of growth: the annual
growth of per capita personal income and growth in the number of jobs per 1,000 county
residents.

Figures 3.4 and 3.5 show the relationship between the county’s distribution index
(averaged over the study period) and the two measures of growth. In contrast to the initial
expectations these ﬁgures demonstrate a positive correlation between the growth and the
index. However, the results are biased since the distribution may correlate with industry
structure (Davis, 1993), education and other determinants of growth (Barro 1998;
Démurger et al. 2001; Rupasingha et al. 2002). Thus, these covariates should be

controlled econometrically.

67

Estimation procedure

The relationship between the two measures of annual growth rate and the
measures of the business size distribution (size_distribution) is estimated using the
model that includes several controls for the growth conditions of a county economy.
These conditions may also affect the distribution of businesses. The two measures
of the distribution discussed earlier - the shares of employment and the index — are

considered in turn. The model takes the following form50 (3.2):

(3. 2) GRu+l= ,60 + ,8, size_distribution, + ,62 educ, ‘+ ,63 manuﬁ + ,6 4farm,

+ ,6 5gov,+ ,6 6denst, + ,6 year, +e+u,,

where GR“. 1 is the growth rate between years t and t+l of both per capita income
and jobs per 1,000 county residents. The share of population with a bachelor’s
degree (educ) controls for the level of education. Year binary variables (year)
control for time speciﬁc shocks common to all counties (e. g. inﬂation, national
business activity conditions, etc.) and common trends, such as declining number of
jobs in manufacturing. The number of jobs per capita in manufacturing (manuf),
farming (farm) and government (gov) controls for the industry structure of local
economy. The number of business establishments per square mile (denst) controls
for the degree of concentration and urbanization. This last variable correlates with
the measures of business size distribution by construction (they both use the total
number of business establishments). However, this correlation is small; for example

the correlation between business density and the distribution index is about 0.01.

 

50 The indexes for counties are suppressed.

68

 

The model omits a number of potential covariates. Among them are local taxes,

natural amenities, quality of local government, spatial interdependence of the regions, as
well as other geographic and institutional conditions. However, most of these factors
have little variability over time. Vector e denotes this unobserved county speciﬁc
conditions that are ﬁxed over time. Finally, it represents an idiosyncratic error.

One may argue that conditional grth model employed in our analysis should
include income and job levels to control for the state dependency. However, those
variables are endogenous to the grth by construction of the variables (the future level
is a function of the current level and the growth rate —- which is ruled out for the ﬁxed
effect estimation procedure used in our analysis). Nevertheless, the control variables used
in the model explain a large portion of variation in the income and job levels and, thus,
control for their effects.

A ﬁxed effect Feasible Generalized Least Squares (FE FGLS) procedure is used
for the estimation of the model. The main beneﬁt of the selected procedure is that it helps
control for unobserved ﬁxed factors in the growth model (e) and provides efﬁcient
estimates of standard errors. Fixed effect estimation reveals how the change in the
business distribution for the same entity associates with change in the growth rate
controlling for other factors. Unrestricted F GLS procedure helps to mitigate the effects of
spatial and serial autocorrelation in errors. Even though the estimation approach does not
control for all sources of bias (Besley and Case 2000) it is the most efﬁcient among
feasible alternatives (see Appendix 3.1 for comparison of alternative procedures). To get
consistent FGLS estimates, one year (1988) is not used for the estimation after the ﬁxed

effect transformation is performed on the full set of data.

69

Whether the inclusion of above-mentioned controls improves the estimation can
be veriﬁed by regression of the measures of business distribution on the control variables.
Appendix 3.2 presents the results for the business distribution index and the shares of
employment in 1-4, 20-49 and 500-99 employee ﬁrms (Columns 1 to 4 respectively).
There is a signiﬁcant correlation between the controls and the measures of the pipeline,

which supports the inclusion of the former into the empirical model.

I_)a_tg

The relationship between the grth and the business distribution is estimated
using data on 83 Michigan counties over the thirteen-year period of 1988 - 2000. The
total sample includes 1079 observations. However, only 996 observations (years 1989-
2000) are used by the FGLS procedure. In addition to the County Business Patterns
(Census Bureau, 2005) this study uses Local Area Annual Estimates (BEA, 2005), which
provide data on income, jobs, population and most other variables.

The mean value of real per capita personal income growth is 1.29 percent per year
(Table 3.2). The average job growth is 1.67 percent per year. There is 0.21 positive
correlation between income growth and job growth. Thus, they reﬂect slightly different
dimensions of economic development. This relatively low correlation between the two
measures of growth reﬂects the fact that new jobs, particularly in metropolitan areas,
beneﬁt a large pool of commuters from other counties. This fact also explains why the
correlation between the index and income growth (Figure 3.4) is lower than the
correlation with job growth (Figure 3.5). Also the growth rates of jobs and wages are

different for most of industries.

70

Over the thirteen years under consideration, Michigan’s economy went through
structural changes contributing to growth. The share of smaller businesses on average
decreased while the share of businesses with 50 to 249 and 500 to 999 employees
increased by a few percentage points. Cross sectional variation of the shares decreased
for groups of 1 to 499 employees. The average value of the business distribution index
did not vary signiﬁcantly, while the range of values trended downward after peaking in
1992 (Figure 3.6). So the employment distribution becomes more similar across
Michigan counties over the period of analysis.

Regarding the control variables, the average share of population with a bachelor’s
degree in our sample is 6.00 percent (Table 3.2). There are on average 76.55 jobs per
1,000 of residents in manufacturing, 13.07 in farming and 73.49 employed by
government. An average un-weighted concentration of business is 4.15 establishments
per square mile. Additional descriptive statistics and data sources can be found in Table

3.2.

3.4. RESULTS AND IMPLICATIONS

The main estimation results are presented in Tables 3.3 and 3.4. The tables
contain FE F GLS estimates of the conditional correlation between the measures of
business size distribution and annual growth rate of jobs and per capita income.

The characteristics of the business distribution are signiﬁcant determinants of
income and job growth. For all speciﬁcations, the measures of distributions are jointly
signiﬁcant. Table 3.3 presents estimates of the relationship between the growth and the
shares of employment by business size categories. Columns 1, 2 and 3 display the effects

on the job growth, while Columns 4, 5 and 6 display the effect on the income growth.

71

Columns 1 and 4 present estimates for all Michigan counties. In addition, the model is
estimated separately for metropolitan and non-metropolitan counties (Columns 2, 5 and
3, 6 respectively). The metropolitan part includes 41 Michigan metropolitan and
metropolitan adjacent counties (Beale Codes 0, 1, 2, 3, 4, 6 and 8). The non-metropolitan
part includes other 42 Michigan non-metropolitan, relatively rural, counties. The reason
for this separate estimation is that the growth models are functionally different between
these two sub-regions of the state (N izalov and Loveridge, 2005).

The estimates on the employment shares represent the effect of marginal change
in the share in comparison to the omitted category — businesses with more than 1,000
employees. A common feature to both grth outcomes (Columns 1 and 4) is that the
effect of the smallest size category (with 1 to 4 employees) is positive, meaning that the
effect of micro-enterprises (individual practitioners and family businesses) on grth is
higher than the effect of the largest enterprises. The relative effect of each group is
decreasing with the size of the business: estimates become less positive (see also Figures
3.6 and 3.7).

An important difference between the job growth and income growth results
should be highlighted. The positive effect of smaller business is more distinctive for the
job growth, while the regional income growth beneﬁts more from only the smallest
businesses as well as the omitted group of the largest enterprises (relative effect of other
categories is negative indicating that their effect is inferior). This difference between the
two models is consistent with the literature, which questioned the effect of small business
on income growth. It follows that there is a conﬂict between job and income grth

enhancing distributions of the business across the size categories. The job growth

72

enhancing distribution would have a larger share of employment in small businesses with
size from 1 to 99 employees and smaller share of larger businesses than it is on average.
On the other hand, an income grth enhancing distribution would have a larger share of
both the smallest (1-4 employees) and the largest (more than 1,000 employees)
businesses than we observe in an average Michigan county.

The direction of this relationship is somewhat different from what we
hypothesized earlier. We expected that an increase in share of groups that are on average
below the equal share (1-4, 250-499, 500-999) would have positive association with
growth. That suggestion is true only for the smallest group (1-4).

Many estimates in Columns 1 and 4 of Table 3.3 are imprecise (statistically
insigniﬁcant) which makes it hard to interpret the results. There are two explanations for
the large standard errors. First, there is a large correlation among the parts of business
distribution (Appendix 3.1), which leads to inﬂated errors. A solution to this, as we
argued earlier, would be an alternative measure of the distribution (the index considered
later). Second, the growth models are functionally different between the metropolitan and
non-metropolitan areas (N izalov and Loveridge, 2005). Thus, a separate estimation of the
model for both sub-regions is necessary.

When we consider the job grth model for metropolitan and non-metropolitan
areas (Columns 2 and 3), this functional difference becomes obvious. For metropolitan
areas, a larger share of the largest businesses (omitted category) is inferior comparing to
other categories the later have positive association with growth. The positive effect is

largest for the smallest businesses, but is gradually decreasing with increasing business

73

size. It implies that a job growth enhancing distribution would have more employment in
smaller businesses than in the larger ones.

In non-metropolitan areas this relationship looks quite different. All business
categories, except the smallest one, have a smaller effect than the largest businesses
(coefﬁcients are negative). Thus, to facilitate the job growth in non-metro areas the
emphasis should be on attracting or growing large businesses and facilitating start-ups of
the smallest. This metro/non-metro difference in the growth model may explain the
difference in policy effects that target small business development found in the literature
(e.g., Nizalov and Loveridge, 2005).

On the other hand, the income growth model does not reveal any important
difference between metropolitan and non-metropolitan areas (estimates in Columns 5 and
6 are similar). For both sub-regions it is beneﬁcial to facilitate the start-ups of the
smallest businesses and increase the share of the largest.

Regarding the results where the distribution is presented with the index (Table
3.4), the positive coefﬁcient implies that the growth-enhancing distribution of businesses
is different from the uniform (11% employment share per size group). The quadratic term
for the index is negative, which means that the positive effect of the deviation from
uniformity is decreasing, and that there is a point after which a further deviation from the
even distribution has a negative impact on growth. The turning point for the job grth is
34.1%, while it is 25.7% for the income growth. These points are above the average value
of the index in our sample (20.35%). The results allow splitting Michigan counties into
three groups. The ﬁrst group includes counties with lower than optimal values of the

index (below 20-22%). Counties in this group would beneﬁt from having more

74

employment in smaller businesses. The second group includes counties with the index in
a range of 22-3 5%. The distribution in this group is close to being the optimal,
controlling for other growth conditions. Finally, the third group includes counties with
too high value of the index (above 35%). The distribution in this group is overrepresented
with smaller businesses. These three groups can be clearly seen on the map of the index

distribution (Figure 3.3).

 

The groups match well the three development strategies, presented by

 

Lichtenstein and Lyons (2006). In the light of our ﬁndings, it seems to be reasonable that

 

the ﬁrst group would beneﬁt more from a business incubating strategy, while the second
and third groups would beneﬁt more from the performance enhancing and the selective
attraction strategies respectively. Thus, the results indicate that the majority of counties
can do better by focusing effort on managing the “pipeline of entrepreneurs and
enterprises” and adjusting the business distribution so that it is closer to the optimal.

Regarding the estimates on control variables, a higher level of education is
negatively related to growth. This variable reﬂects a negative relationship between the
initial level of income (with which the education correlates positively) and growth. Other
controls are also signiﬁcant. The estimates are robust to changes in the speciﬁcation of
business distribution variables.

The paper, however, has some limitations. By considering annual grth we are
focusing our attention on the short run relationship between the size distribution and
growth. The long-run effects may be different and should be addressed separately. Also,
the data that we use in this study does not provide us with information on non-employer

entities (self-employed entrepreneurs), which make a large portion of all business

75

 

establishments. This exclusion, however, does not compromise our results. The number
of non-employer entities is correlated with the share of the smallest businesses.

The deﬁnitions of size groups are taken as provided by the County Business
Patterns, without considering alternatives and possible implications. Average size of
establishments in each group can be estimated more precisely, and the distribution index
can be constructed in several other ways (e. g. sum of square deviations from national
averages). Moreover, considering larger areas, such as commuting zones (Tolbert and
Sizer, 1996) or states may better capture the interdependences that exist among the
businesses of different sizes. Future work may explore these questions, and document

whether our ﬁndings can be replicated in other states.

76

3.5. SUMMARY AND CONCLUSIONS

This paper analyzes the relationship between economic growth of Michigan
counties and measures of the distribution of employment across businesses of different
size. This study uses data on Michigan counties from the County Business Patterns
(Census Bureau, 2005) and Local Area Annual Estimates (BEA, 2005), for the years
1988-2000. The distribution is measured by the share of employment in establishments in
various size categories and by a distribution index. The index provides a measure of the
extent to which the local economy deviates from an equal percentage of employment in
each business size category.

The major ﬁnding is that there are strong links between the business size
distribution and regional economic growth. The results indicate that the growth-
optimizing distribution of employment by business size is not uniform. We found also
that optimal job and income grth enhancing distributions would have a higher share of
the smallest businesses than is currently the average. However, a larger share of
businesses with more than 1,000 employees would also facilitate income growth in some
counties.

There are several direct policy implications. The appropriate policy actions (e. g.,
incubating small business or recruiting and growing large businesses) depend on the
current distribution of businesses in a particular region, supporting ideas by Lichtenstein
and Lyons (2006). In Michigan, most counties (particularly in the southern part of the
state) seem to need pay more attention to smaller businesses, given our results. A regional
economy needs to maintain sufﬁciently large share of self-employed entrepreneurs,

family businesses and other smallest (micro) enterprises to enhance growth of other parts

77

 

 

of local economy. Unlike industrial recruitment or infrastructure-based economic

development strategies, policies to stimulate small business formation need not be costly,

but may require thoughtful effort to achieve cultural and attitudinal shifts. Some
examples of low cost small business-friendly cultural shifts follow.

0 Local civic groups could feature the small business of the month, year, etc. that could
inspire others to try opening a business. The Edward Lowe Foundation recently began
doing this in Michigan with its “Fifty Companies to Watch” awards. Businesses
receiving the award report easier access to credit and better networks as a result of the
award.

0 Communities can work to seed formation of peer-to-peer networks of small business
owners (Lichtenstein et al., 2004; Muske and Woods, 2004).

0 Small businesses could be featured in K-12 learning experiences as mentors. The Junior
Achievement program offers some opportunities but mentoring could be implemented
more widely and in a broader set of contexts (Emery et al., 2004).

0 Small business examples and role playing could be featured in basic math and social
studies in K-12 curricula (Hanham et al., 1999).

o Institutions of higher learning could be encouraged to offer entrepreneurship
specializations within a variety of programs of study.

In conclusion, it is evident that in Michigan, the notion of an entrepreneurial
pipeline is a concept that is supported by the modeling results, at least in terms of the
numbers of businesses in various size categories. The entrepreneurial pipeline should
help inform policy as decision makers choose to pursue various competing strategies for

local economic development.

78

The analysis presented in this paper is not without points for critique. One may
argue, for example, that counties, used as units of analysis, are too small to capture all the
complexity of business interdependencies. Also, the ﬁndings presented in this paper may
be sensitive to the deﬁnition of the index or to the speciﬁcs of Michigan economy. These

and some other issues are the subjects to the further research.

79

 

 

APPENDIX 1.1. HOUSEHOLD PRODUCTION MODEL

The speciﬁcation of a model explaining poverty (choice of variables) is informed
by the theory of household production. A household is involved in a set of activities using
a stock of physical and human assets and investing in them. The purposes of those
activities are to reproduce the assets (ﬁrst of all human), to meet social standards, to
enjoy consumption and to increase wealth (the stock of assets). Those purposes are
represented by a utility function U(.). Real income is considered as a constraint on utility
maximization and it serves as an intermediary of that process. Observed monetary
income is often considered as a proxy for real income (though we know that income and
happiness are not well correlated after basic needs are met).

A household production model (Singh et a1, 1986) describes the interdependence
of production and consumption activities of a household. According to the model, a
household maximizes utility (equation A1) subject to a set of constraints (equations A2-
A4).
(A1) Max U(Xh, Xm, l, H)
where X}, is a vector of home-produced and consumed at home goods like agricultural
products, housing, childcare etc. A vector X," is a set of market-purchased consumption
goods, I is a leisure time; H, is a set of household characteristics (often referred as taste
shifters) such as age, gender, cultural or ethnic background etc.

The constraints are of the following nature. The ﬁrst one (A2) is a budget
constraint — expenditures must equal income including borrowing.

(A2) Budget: kam+knn=kh(q-Xh)+wLo+v

80

 

where k,,. is a vector of prices for market purchased goods, k}, is a vector of prices for
home produced goods, w is a vector of wage rates for a stock of labor, [.0 is number of
hours working out of home (in the labor market). Vector v includes non-labor income and
will be described later (A5). Vector q is a set of home-production outputs. It is described
by a production function (A3). And ﬁnally, kn is a vector of prices for market purchased
production inputs n.

(A3) Production: q=q(Lh, n, Ah)

where q(.) is a household production function, L}, is a vector of a household labor inputs
into a home production, It is a vector of market purchased production inputs, and A}, is a
vector of physical assets used in household production.

The nature of the next constraint (A4) is that amount of labor (human) assets have
physical (and biological) limits T. Vector T describes those limits, which in turn are an
outcome of household characteristics H (e. g., number of adults in a household). The
endowment of time can be spent on leisure, home production or on work outside the
home.

(A4) Time: Lh+ Lo+l=T

The last component of a budget constraint — non-labor income v — should be
described separately (A5). It includes return on investments (A -Ah). Where r is a rate of
return, assuming that all the assets not used in a household production are invested. The
other components is a vector of subsidies, transfers and taxes, which could depend on
household characteristics and income.

(A5) Non-labor income: v=r(A-Ah)+s

Combining equations (A2) to (AS) the budget constraint takes the form:

81

 

(A6) kam + thh+wl=rA+wT+khq(Lh, n, Ah)-th-rAh-knn +s

This expression equates a household’s real expenditures to a real income. Expression
(A7) 7: =khq(Lh, n, Ah)-th-rAh-knn

is an equivalent of economic proﬁt of a household production.

Solving a household utility maximization problem we can get the optimal
household income y‘ as:

(A8) y’ = rA+wT+7r ‘(w, r, kn, kh, H)+s

The level of household income depends on the initial stock of assets (human and
physical), exogenous to the household return to assets (prices and costs), transfers and a
household ability to extract income out of its assets (proﬁt).

Our derivation is based on the assumption that a household does not hire out-of-
household labor, but rather buys necessary inputs (including labor) on a market. Second,
we assume perfect asset markets (both physical and human). In other words there is no
cost to participate in labor or capital or other markets, and perfect information about the
markets is available. Often it is not true, especially when we are considering regional
poverty. This assumption can be relaxed in further projects to evaluate an impact of
market imperfections such as costs of labor force participation on regional poverty.

Another issue, not addressed in the model, is factor mobility. Capital does move
to utilize low cost labor as witnessed by the textile industry moving first from New
England to the south of the US, and then to Asia. But, the process has not been sufﬁcient
to eliminate rural poverty either in the south or in Asia. Likewise, labor of a given quality
does migrate from low wage areas to higher wage areas as witnessed by the migration of

Mexicans and Asians (and others) into the United States and the migration of people from

82

 

Appalachia to industrial cities in the north. But, similarly to the case of capital, labor
migration has not eliminated poverty at the source of the migration. This suggests there
are formal and cultural barriers to the equalization of factor returns. However, it is
beyond the scope of this modest research project to investigate these variables.

The model developed here takes the stock of human capital as a given. It does not
further inquire into how that stock was produced — how individuals and the broader

community reached investment decisions.

83

 

APPENDIX 2.1: MODEL SPECIFICATION

Whether the control variables explain non-random variations in policy treatment
caused by targeting can be veriﬁed. We estimate the correlation between the policy
treatment and the controls. Table 2.5 presents the results of this estimation. It shows a
signiﬁcant correlation between the treatment and other variables included in the models,
which supports a concern about non-random policy placement. It shows also a systematic
relationship among the policy treatments. For example, areas with higher per capita
highway mileage have more of BDA per capita (coefﬁcient on highways in the BDA
equation is positive — Column 3). On the other hand, there are fewer MEGAs in areas
with more RZ (coefﬁcient on R2 is negative -— Column 4), holding other factors ﬁxed.

Regarding the controls, a higher level of education associates with fewer
applications of MEGA, but more BDA. Characteristics of county industry structure are
also signiﬁcant in explaining the variations in the amount of policy treatment. A negative
coefﬁcient on business concentration indicates that the policies systematically target
areas with fewer business establishments, holding other factors ﬁxed. Joint signiﬁcance
of the control variables is also high (see Chi-square statistics at the bottom of the table).

Thus, the use of the above-mentioned controls improves results of the policy analysis.

84

APPENDIX 3.1. JUSTIFICATION OF THE ESTIMATION TECHNIQUE

To compare alternative estimation procedures, the relationship between the
income growth and the distribution index is estimated using pooled ordinarily least
squares (POLS) (Columns 1 and 2, Table 3.5), ﬁxed effect (FE) (Columns 3 and 4), ﬁrst
difference (FD) (Column 5), Arellano-Bond (ABond) (Column 6), and ﬁxed effect
feasible generalized least squares (FE F GLS) (Column 7) methods. Results are presented
in Table 1.5. Since different estimation methods have different use of the lagged values,
the comparison is limited to the period 1990-2000, so that the same number of
observation is used. For that reason the FE F GLS estimates in Table 1.5 are somewhat
different from the results presented in Table 1.4 of the main part of this paper.

When we compare POLS estimates (Column 1, 2) with FE and FD results
(Columns 3-7) we can observe a signiﬁcant difference both in terms of estimated
parameters and a goodness-of-ﬁt. This difference suggests the presence of unobserved
ﬁxed effect and makes FE and FD estimates superior relative to POLS. Addition of time
varying control variables (education, industry structure, business concentration and year
dummies) reduces omitted variable bias further in a signiﬁcant way. The difference is
observed for both POLS (Columns 1 and 2) and FE (Columns 3 and 4) estimates.

Consistency of FE and FD estimates rely heavily on a strict exogeneity
assumption. The assumption implies that the explanatory variables have to be exogenous
for all time periods (not just contemporaneously). If this assumption holds the results of
FD and FE estimation procedures (Columns 4 and 5) should be similar. However, some
difference is observed for estimates on education. This difference can be treated as a sign

of endogeneity problem (Wooldridge, 2002, Ch.11) or that this variable is predetermined.

85

A potential remedy for this problem is provided by Arellano-Bond estimation
technique (Arellano and Bond, 1991). Result is presented in Column 6. These estimates
are somewhat in between of the FE and FD results. However, ABond estimates may be
inconsistent due to the presence of serial correlation in errors. Thus, the FE procedure is
used as a base for the analysis. Comparing FE and FD, the former is more efﬁcient for
two reasons. First bias in FD parameters does not decrease with the number of time
periods. Second, with a small serial correlation in error term (which is true in our case)
FE procedure is more efﬁcient (Wooldridge, 2002, Ch. 10).

Finally, to improve the efﬁciency of the FE estimates the standard errors are
corrected for spatial and serial autocorrelation by applying FGLS procedure on FE
transformed data (Column 7). This step increases the goodness of ﬁt of the model
comparing to the original FE procedure.

The results, obtained by these different procedures display signiﬁcant similarity,
which allows to conclude that they are robust to the alternative technique. However, FE

FGLS is used for the rest of the paper as the most efﬁcient.

86

APPENDIX 3.2: MODEL SPECIFICATION

Table 3.6 presents estimates of the association of the control variables with the
business distribution index and employment share in selected business size categories.
The results imply that a higher level of education reduces inequality in the business
distribution (parameter on education is negative in Column 1). A higher share of people
with bachelor’s degree is also associated with reduction in employment share of the
smallest (1-4 employees) and medium (500-999 employees) businesses holding other
factors ﬁxed. A higher share of manufacturing and a larger number of business
establishments per square mile makes the deviation of business distribution from
uniformity larger (coefﬁcients in Column 1 are positive). Employment in the government
sector and farming is also a signiﬁcant determinant of business size distribution. In
summary, signiﬁcance of the control variables in explaining the variations in the
measures of business distribution justiﬁes their presence in the estimation model.

Tabulation of business distribution introduces substantial multicolinearity into the
model, as was argued above. The share of employment in both the small (1-4 and 20-49
employees) and medium (500-999 employees) businesses strongly correlate with the
shares of employment in other size groups. However, some interesting results can be
observed. By construction, the estimates on the shares are biased toward minus one. An
increase of the share in one group by one percent tends to decrease the share of
employment in other group by one percent holding other factors ﬁxed. Results in Column
2 reveal a positive and signiﬁcant correlation among the shares of employment in the
smaller size groups (1-4, 5-9, 10-19 and 20-49 employees). The results imply that growth

of smaller businesses is mutually dependent, lending support to Lichtenstein and Lyons’

87

theory. This effect can come through a larger supply and demand base among the
businesses, role modeling, experience or other unobserved common conditions. There is

no such relationship for the larger business groups.

88

TABLES

Table 1.1. Distribution of Community Poverty Rates in Michigan

 

 

 

 

 

 

 

Poverty Rate Number of Communities % of Communities
0% -— 10% 7,665 63.8%
10.1%-30% 3,527 29.3%
30.1% and above 826 6.9%

 

Data: Based on Census 2000, US Census Bureau

89

 

Table 1.2. Descriptive Statistics of Michigan Communities

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

Variable’ All State Rural Rural Metropolitan Metropolitan
Towns Adjacent
Mean Mean Mean Mean Mean
(Std. (Std. (Std. (Std. Dev.) (Std. Dev.)
Dev.) Dev.) Dev.)
Number of observations 12,018 1,871 517 6,775 2,855
[% of total] i] 00.0] H 5.6] (4.3] (56.3] [23.8]
Population, 100 person 13.4 6.8 10.1 14.9 10.8
(9.1) (4.4) (5.9) (9.8) (5.3
Poverty rate‘, % 10.2 9.7 14.2 11.1 6.2
(10.8) (6.1) (11.0) (12.0) (5.0)
County Poverty Rate in 1990", 12.9 15.4 14.2 12.8 12.0
% (5.6) (4.5) (4.2) (6.0) (4.2)
County Wetge"",_$,0003 34.0 23.3 24.9 36.5 29.7
(6.2) (2.5) (2.2) (4.8) (4.5)
County Unemployment, % 5.9 7.8 7.3 5.8 5.3
(2.0) (2.4) (2.2) (2.0) (1.3)
County Occupation Structures, % 19.1 19.6 14.9 18.1 23.2
(9.9) (8.7) (9.0) (10.0) 48.8)
Working Age Adults, persons per 1.8 1.6 1.6 1.7 1.8
household (0.6L (0.5) (0.7) (0.7) (0.2)
Dependants (children under 16), 0.8 0.7 0.6 0.8 0.8
persons per household (0.3) (0.3) (0.2) (0.3) (0.2)
People with Disabilities, persons 0.8 0.9 0.8 0.8 0.8
er household (0.4) (0.3) (0.4) (0.4) (0.3)
Incomplete High School‘, % 16.9 17.6 15.0 17.3 15.1
(11.2) (8.3) (8.2) (12.3) (7.3)
Completed High School, % 31.3 39.1 34.2 28.8 37.4
(1 1.0) (9.0) (10.7) (10.7) (8.9)
Less than College, % 30.3 28.6 30.4 30.1 31.7
(7.8) (7.3) (7.3) (7.9) (7.1)
Bachelor Degree, % 13.6 9.6 13.2 14.8 10.5
(9.7) (6.4) (8.0) (10.4) (6.7)
Graduate Degree, % 7.0 5.1 7.3 9.0 5.3
(8.9) (4.6) (6.5) (9.8) (5.1)
Average Age of working cohort 38.6 41.1 38.4 38.1 39.9
(16 to 65), Lears (3.3) (2.9) (4.0) (3.3) (2.2)
In-Migrationz, % 17.3 20.2 23.2 17.0 16.4
(12.5) (9.7) (14.0) (13.4) (8.5)
Share of retirees (65 and Older), 12.3 16.5 17.7 12.1 10.7
% (7.6) (8.1) (9.5) (7.7) (5.2)
College Town (dummy) 0.3 0.01 0.4 0.4 0.0<
(0.5) (0.1) (0.5) (0.5) (0.1)

 

Data: ' Census 2000, SF3 is a data source unless indicated otherwise; " Census 1990; "'Bureau of

Economic Analysis

‘Percentage of households in poverty.
2A3 percentage of people 5 years and older leaving outside of county of current residence in 1995.
3 Average wage per job in a county in 1998 and 1999
4Education for population 25 years and over
5 Share of production, transportation and material moving occupations in total employment in a county

90

 

Table 1.3. Test for Difference, F-statistics

 

 

 

 

 

 

 

Rural Rural Towns Metropolitan
Rural 0.86 1163'"
Rural Towns 0.86 331'"
Metropolitan 11.63‘" 3.31m
Metro. Adjacent 3.59'" 1.66" 5.52'

 

 

 

 

* signiﬁcant at 10%; “ signiﬁcant at 5%; “‘* signiﬁcant at 1%

91

Table 1.4. Estimates of regional poverty model

 

+

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

Pove
All State Rural Rural Metropolitan Metro.
Towns Adjacent
3 County Wage+ -l.25 m -0.72 .. -0.41 -0.75 " -0.26
g (0.48) (0.33) (1.66) (0.37) (0.28)
g \ County 0.00< -002 -0.01 0.01 0.00<
Q (a; Unemployment (0.02) (0.02) (0.06) (0.07) (0.05)
g County -0.00< -0.00< -0.01 -0.00< 0.00<
.3 Occupational (0.00<) (0.00<) (0.01) (0.00<) (0.00<)
Structure °
Adults -5.05 -l.82 -2.58 -3.36 -193
(1.18)‘" (0.84)" (0.86)"° (0.95)” (1.22)
Dependents 0.66 1.16 -131 1.02 -1.38
(Children <16y_) (0.79) (0.75) (1.37) (0.53)‘ (1.41)
Incomplete School 0.02 0.01 0.01 0.02 0.02
(%) (0.01)” (0.00<)“’ (0.01) (0.01)” (0.01)"
E Incomplete College -0.00< -0.01 . -0.01 -0.00< 0.00<
e» (%) (0.00<) (0.00<) (0.01) (0.00<) (0.01)
§: Bachelor Degree -0.00< -0.01 -0.00< -0.01 -0.00<
if (%) (0.01) (0.01) (0.01) (0.01) (0.01)
3 Graduate Degree 0.01 -0.01 -0.00< 0.01 -0.00<
3 (m (0.01)’ (0.01) (0.01) (0.00<)" (0.01)
People with 0.93 0.17 0.64 0.85 0.36
Disabilities (0.20)” (0.1 1) (0.38)‘ (0.21)“ (0.19)‘
Age -1.26 0.27 -0. 70 -134 -0.02
(0.37)” (0.1 1)” (0.38)‘ (0.36)“ (0.19)
Age Squared 0.01 -0.00< 0. 01 0. 02 -0.00<
(0.00<)” (0.00<)” (0.00<) (0.00<)“ (0.00<)
County Poverty 0.03 0.45 0.67 -0.01 0.31
Rate, 1990 (0.13) (0.21)" (0.45) (0.25) (0.19)
In-Migration -0.03 0.01 -0.02 -0.02 0.02
(0.02)‘ (0.01) (0.02) (0.01) (0.02)
Share of Retirees -0.12 -0.02 —0.07 -0.07 -0.07
g (0.03)” (0.02) (0.02)” (0.02)” (0.04)
g College Town 0.27 0.09 0.18 0.34 0.23
g (dummy) (0.12)“ (0.22) (0.26) (0.06)‘“ (0.31)
9 Population 0.14 0.25 0.25 0.12 0.14
(0.01)” (0.01)” (0.03)” (0.01)“ (0.01)“
Area 0.02 0.01 -0.04 -0.07 0.02
(0.00<)” (0.00<)” (0.16) (0.06) (0.00<)“
Constant 29.58 -6.61 17.21 30.26 4.41
(8.12)”‘ (2.96)“ (8.31)“ (7. 74)“ (6.32)
Observations 12018 1871 517 6775 2855
F-statistics 88.63W 58.21'" 22.75'" 124.85'" 87.14'"

 

Robust standard errors in parentheses; + Variable is in a natural logarithm form; *, **, *" signiﬁcant at
10%, 5%, 1% respectively; Commuting zone dummies are suppressed for the display purposes; Boldface
font for the parameters that are different from estimates for rural communities at 10%.

92

 

Table 1.5. First Stage Estimation for All Communities

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

Number of Number of Share of In-
Adults Children Migrants
County Wage‘“ -014 0.17 -12.02
(0.09) (0.09)* (2.90)***
County Unemployment 0.01 0.02 -0.46
(0.005)* (0.00<)*** (0.16)***
County Occupational Structure -0.00< -0.00< -0.01
(0.00<)* (0.00<) (0.02)
Incomplete School (%) -0.00< 0.00< 0.10
(0.001)" (0.001)*** (0.02)***
Incomplete College (%) -0.00< 0.00< 0.02
(0.001) (0.001) (0.02)
Bachelor Degree (%) 0.00< 0.00< 0.26
(0.001) (0.001) (0.03)***
Graduate Degree (%) 0.00< 0.00< 0.20
(0.001)*** (0.001) (0.03)**"‘
People with Disabilities 0.17 0.07 -0.10
(0.02)*** (0.02)*** (0.41)
Age -0.21 0.04 -7.27
(0.04)*** (0.02)“ (0.47)***
Age Squared 0.00< -0.00< 0.08
(0.00<)*** (0.00< *** (0.01)***
Share of Retirees -0.03 -0.02 -0.05
(0.001)*** (0.00<)*** (0.02)"
College Town (binary) -0.07 -0.02 -2.55
(0.01)*** (0.01) (0.27)***
County Poverty Rate, 1990 -0.10 -0.06 -0.06
(0.02)*** (0.02)*** (0.65)
Population 0.00< -0.00< 0.20
(0.00<) (0.00<)*** (0.02)***
Area 0.00< 0.00< 0.05
(0.00<)*** (0.00<) (0.01)***
Average Size of Household in -0.08 -0.16 20.13
Neighboring Area (0.07) (0.06)*** (2.29)***
Coastal County (binary) 0.05 0.03 -2.45
(0.01)*** (0.01)*** (O.31)***
Adults per Household in 1990 0.83 1.11 -80.48
(0.23)*** (0.22)*** (8.19)***
County Wage in 1990 0.00< 0.00< -0.01
(0.00<)*** (0.00<)*"‘* (0.001)***
Adults 90 * Wage 90 -0.00< -0.00< 0.01
(0.00<)*** (0.00<)*** (0.00<)*"‘*

 

93

 

Table 1.5. (Cont)

 

 

 

 

 

 

 

 

 

Number of Number of Share of In-
Adults Children Migrants

Proportion of Females in 0.01 -0.02 -0.06
Neighboring area (0.01) (0.01)*** (0.18)
Average Age of Family -0.20 -0.02 7.93
Households’ Heads in (0.09)" (0.07) (3.23)”
Nﬂboring Area
Average Age of Heads Squared 0.00< -0.00< -0.07

(0.00<)" (0.00<) (0.03)"
Constant 8.70 1.23 29.36

@15)‘ * * (1.77) (77.40)
Observations 12018 12018 12018
Adjusted R-squared 0.40 0.32 0.31
F-statistics for Instruments 7.77*** 1222*" 2576*"

 

 

 

 

Robust standard errors in parentheses; "' signiﬁcant at 10%; " signiﬁcant at 5%; *** signiﬁcant at 1%

94

 

Table 2.1. Summary Statistics: Policy Treatment, Michigan, 1993-2000

 

 

 

 

 

 

Variable Mean Std. Min Max
Dev.

Number of Michigan Economic Growth Authorities 0.272 0.634 0 14.271

(MEGA) in a county8 per 100,000 residents

Area of the Renaissance Zone, acresa per 1,000 0.664 5.031 0 104.083

residents

Number of Brownﬁeld Redevelopment Authorities in 0.396 1.043 0 22.986

a countyb per 100,000 residents

Federal, State and major county highways, milesc per 3.807 5.218 0.494 85.995

 

1,000 residents

 

 

 

 

Data: a) Michigan Economic Development Corporation
b) Victor Land Use Institute, Michigan State University Extension
c) Michigan Department of Transportation

Note: Policy treatment as of January 01 for a given year. Statistics for 664 population weighted county-

level observations.

95

 

Table 2.2. Summary Statistics: Michigan Counties, 1993-2000

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

Variable Mean Std. Min Max
Dev.
Annual Income Growtha, % 1.731 2.602 -11.565 14.725
Three-Year Average Income Growth“, 1.317 1.562 -6.841 5.308
°/o
Annual Job Growtha, % 1.073 2.325 -15.780 23.985
Three-Year Average Job Growth”, % 0.587 1.644 -8.610 12.234
Percentage of Population with 8.140 3.217 2.977 15.650
Bachelor Degree , %
Number of Jobs in Manufacturing per 101.086 34.197 9.462 196.835
1000 Residentsa
Number of Government Jobs per 1000 68.062 37.082 33.502 230.062
Residentsa
Number of Jobs in Farming per 1000 5.607 7.351 0.000 43.976
Residents8
Business Establishment Densityc 25.107 22.710 0.091 58.876

 

Data: a) Based on Local Area Annual Estimates (BEA 2005)
b) Based on Census of population 1990, 2000. Data for 1993-99 is imputed by linear extrapolation
c) County Business Patterns (Census Bureau 2005)
Note: Statistics for 664 population weighted county-level observations.

96

 

Table 2.3. Base Model

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

Income Growth Job Growth
Annual 3-Year Average Annual 3-Year Average_
(1) (2) (3) (4)
Highway 0.003 -0.053 0.443 1.220
(0.795) (0.376) (1.190) (0.560)"
Highway Squared -0.055 -0.031 -0.084 -0.097
(0.031)‘ (0.015)” (0.047)‘ (0.022)”
RZ 0.056 0.018 0.030 -0022
(0.036) (0.017) (0.054) (0.026)
RZ Squared -0.001 -0.000 -0.001 0.000
(0.000) (0.000) (0.001) (0.000)
BDA 0.034 -0.046 -0.106 -0.035
(0.092) (0.043) (0.137) (0.065)
BDA Squared 0.003 0.005 0.000 0.003
(0.006) (0.003) (0.009) (0.004)
MEGA 0.006 -0.263 0.497 0.096
(0.324) (0.153)‘ (0.485) (0.228)
MEGA Squared 0.016 0.058 -0042 0.006
(0.067) (0.032)’ (0.100) (0.047)
Highway *Rz 0.002 0.001 0.003 0.002
(0.001 f (0.000)” (0.001 )‘" (0.001)”
Highway *BDA 0.002 0.003 0.009 0.007
(0.005) (0.002) (0.008) (0.004)’
Highway *MEGA 0.010 -0.008 0.017 0.004
(0.027) (0.013) (0.040) (0.019)
RZ*BDA -0.001 -0.001 -0.000 -0001
(0.001) (0.001 ) (0.002) (0.001)
RZ‘MEGA 0.045 0.020 -0044 -0.010
(0.049) (0.023) (0.073) (0.034)
BDA*MEGA 0.002 0.042 0.038 0.006
(0.051) (0.024)‘ (0.077) (0.036)
Educationmy -0.662 -0.704 m -0.092 -0.080
(0.527) (0.249) (0.788) (0.371)
Manufacturing,” 0034 .. -0.036 -0040 -0.045
(0.017) (0.008) (0.026) (0.012)
Covemment(,., -0.023 -0.027 0042 -0.057
0.015) (0.007)” (0.023)‘ (0.01 1 )”‘
FarmingM) 0.040 0.020 -0.151 -0.087
(0.137) (0.065) (0.204) (0.096)
Business ConcentrationM) 0.316 -0.178 0.259 -0.413
(0.364) (0.172) (0.544) (0.256)
Constant -0.970 -0.249 . 1.326 0.741
(0.310) (0.146) (0.463) (0.218)
Direct Policy: cm 8.44 12.23' 12.28' 21.87
Policy int: Ch12 9.81 17.84" 10.09 23.44'"
All Policy: Chiz 19.46 40.63'" 29.14‘" 69.1 1'"
Log likelihood -1185.34 -750.11 -1419.16 -981.62
Wald chi2 232.45'" 928.67'" 188.17‘" 419.27'"

 

 

Standard errors in parentheses. *, **, and ""‘ signiﬁcant at 10%, 5%; and 1% respectively. Fixed Effect
FGLS estimates presented. Year dummies are suppressed for display purposes. Observations: 581. Number
of counties: 83

97

Table 2.4. Michigan Development Policy Effects by Income Level

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

Income Growth Job Growth
High Income Low Income Highlncome Low Income
Annual 3-Year Annual 3-Year Annual 3-Year Annual 3-Year
Avergage Average AveLage Average
(1) (2) (3) (4) (5) (6) (7) (8)
Highway -0779 2.752 0.369 -0.822 -3.219 1.994 1.056 2.048
(2.012) (1.046)” (1.081) (0.461)‘ (2.794) (1.346) (1.776) (0816)”
Highway -0. 100 -0.023 -0057 -0.007 -0.093 0.075 -0.1 1 1 -0137
Squared (0.1 14) (0.059) (0.042) (0.018) (0.158) (0.076) (0.069) (0.032)“
RZ 1.1 12 0.293 0.037 -0.002 0.627 0.515 0.047 -0.038
(0.943) (0.491) (0.041) (0.018) (1.310) (0.631) (0.068) (0.031)
RZ 0.053 0.025 -0.000 0.000 0.024 0.016 -0.001 0.000
Squared (0.049) (0.026) (0.001) (0.000) (0.069) (0.033) (0.00 1L (0.000)
BDA -1.074 -0.352 0.160 0.01 1 -0.655 -0.446 0.062 0.178
(0.603)‘ (0.313) (0.137) (0.059) (0.837) (0.403) (0.226) (0.104)‘
BDA 0.120 0.01 1 -0. 005 0.002 0.078 0.029 -0. 006 -0.005
Squared (0.043)“ (0.022) (0.007) (0.003) (0.060) (0.029) (0.012) (0.005)
MEGA 4.605 1.799 -0.271 -0.348 0.429 1.168 0.556 0.036
(3.082) (1.603) (0.364) (0.155)“ (4.279) (2.062) (0.599) (0.275)
MEGA -0415 -0.206 0.054 0.079 0.076 -0.063 -0.072 0.001
Squared (0.217)‘ (0.113)‘ (0.081) (0.034)“ (0.301) (0.145) (0-133L (0.061)
Highway 0.072 0.008 0.002 0.001 0.094 0.051 0.004 0.003
*RZ (0.070) (0.036) (0.001)“ (0.000)‘ (0.097) (0.047) (0.002)" (0.001)”
Highway 0030 0.007 -0.000 0.001 -0.007 0.001 0.007 0.003
*BDA (0.020) (0.011) (0.006) (0.003) (0.028) (0.014) (0.010) (0.004)
Highway 0.095 0.014 -0.008 -0021 -0015 0.002 0.037 0.007
*MEGA (0.054)‘ (0.028) (0.035) (0.015) (0.075) (0.036) (0.057) (0.026)
RZ’BDA -0.261 -O.108 -0.002 -0.001 -0024 -0.055 -0.001 -0.001
(0.232) (0.120) (0.001 (0.001) (0.322) (0.155) (0.002) (0.001)
RZ*MEGA 1.157 0.515 0.023 0.018 0.328 0.368 -0.040 -0004
(1.211) (0.630) (0.051) (0.022) (1.681) (0.810) (0.084) (0.039)
BDA -0.036 0.089 -0.062 -0004 0.224 0.109 -0040 -0.067
*MEGA (0.090) (0.047)‘ (0.065) (0.028 (0.124)‘ (0.060)‘ (0.107.)_ (0.049
All Policy: 19.31 25.86" 13.98 34.42 9.94 17.62 25.30 62.99
cm
Difference: 19.33 29.66'" 10.58 25.34"
ChiZ
Log -554.08 -361.81 -594.35 -349.96 -650.54 -435.85 -736.99 -513.79
hkehhood
Wald ch12 239.30'" 704.74'" 109.36‘" 433.34“ 112.18'" 297.68'" 111.71'" 211.09'"
Observations 294 294 287 287 294 294 287 287
Number of 42 42 41 41 42 42 41 41
Counties

 

 

 

 

 

 

 

 

 

 

Standard errors in parentheses. *, **, and "" signiﬁcant at 10%, 5%; and 1% respectively. Fixed Effect
FGLS estimation procedure is used. Control variables are suppressed for display purposes. Estimates in
boldface font indicate that difference between areas with higher and lower per capita income is individually
signiﬁcant at 10% level.

98

Table 2.5. Intensity of Policy Application

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

(1) (2) (3) (4)
Highway RZ BDA MEGA
Highway 3.856 0.748 -0214
(2.354) @409)’ (0.209)
RZ 0.001 0.038 -0.010
(0.001) (0.007)” (0.004)”
BDA 0.008 1.240 0.042
(0.004)‘ (0.233)” (0.021)"
MEGA -0008 -1252 0.162
(0.008) (0.465)‘" (0.081)"
Educationm) 0.047 -1.338 1.918 m -0.458 ..
(0.041) (2.361) (0.403) (0.208)
Manufacturing,” 0.006 -0.269 0.029 0.001
(0.001)” (0.076)"° (0.013)“ (0.007)
Govemmen(,.., 0.005 -0047 0.047 -0032
(0.001) (0.065) (0.01 1) (0.006)
Farming...) 0.015 1.373 .. 0.187 . 0.128 ..
(0.01 1) (0.611) (0.106) (0.054)
Business Concentration“) -0.083 m -3.748 .. -0.778 m -0.076
(0.029) (1.630) (0.283) (0.145)
Constant -0034 -4.785 0.246 -0.629
(0.024) (1.356) (0.238) (0.1 19)
Controls: Chi2 50.44 31.85 60.77 50.50
Log likelihood 279.45 -2068. 17 -1052.03 -660. 16
Wald ch12 180.19‘W 119.53'" 366.70'" 165.56"

 

 

Standard errors in parentheses. *, ", and "”" signiﬁcant at 10%, 5%; and 1% respectively. Fixed Effect
FGLS estimation procedure is used. Year dummies are suppressed for display purposes. Observations: 581.
Number of counties: 83.

99

Table 3.1. Business Size Distribution: Descriptive Statistics

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

A. Establishment share, %
Size Group Mean Std. Dev. Min Max
(employees)
1-4 57.65 6.86 44.55 83.93
5-9 20.51 2.60 7.55 26.76
10-19 11.59 2.22 3.57 16.01
20-49 6.58 1.91 0 11.62
50-99 2.07 .87 0 4.59
100-249 1.17 .55 0 2.82
250-499 .29 .21 0 1.12
500-999 .10 .12 0 .59
1,000< .04 .07 0 .49
B. Employment share, %
Size Group Mean Std. Dev. Min Max
(employees)
1-4 5.94 3.08 2.03 25.14
5-9 9.84 2.97 4.54 31.40
10-19 10.87 3.00 5.56 38.96
20-49 11.89 2.59 0 25.69
50-99 9.02 3.07 0 24.27
100-249 10.02 3.94 0 43.29
250-499 5.81 4.14 0 22.08
500-999 3.69 4.40 0 20.68
1,000< 2.68 4.60 0 34.55
| C. Size Distribution Index, % |
I Mean Std. Dev. Min Max I
| index 20.35 5.28 12.88 61.47 J

 

 

 

 

 

Note: Statistics not weighted, based on 1079 observations for 83 Michigan counties.

Data: County Business Patterns, US. Census Bureau, 1988-2000.

100

Table 3.2. Summary Statistics: Michigan counties, 1988-2000

 

 

 

 

 

 

 

 

 

Variable Mean Std. Dev. Min Max
Income growth“, % 1.29 2.70 -11.56 14.72
Job growth“, % 1.67 6.84 -39.28 65.29
Population with BA/BS degree", % 6.00 2.38 2.28 15.65
Jobs in manufacturing", number per 1000 76.55 40.93 7.96 196.83
residents
Jobs 1n government , number per 1000 73.49 36.12 35.50 230.06
res1dents
. . a
Jobs in farming , number per 1000 13.07 9.12 .10 45.70
res1dents
. . c C .
Busmess establishment den81ty , number 415 8.96 .09 58.88

per per square mile

 

 

 

 

 

Data: 21) Based on Local Area Annual Estimates, Bureau of Economic Analysis. Farm employment is the

number of workers engaged in the direct production of agricultural commodities, either livestock
or cr0ps; whether as a sole proprietor, partner, or hired laborer. The manufacturing as the SIC
division. Government includes Federal civilian, military, and state and local.
b) Based on Census of population 1990, 2000. Data for 1988-89, 1991-99 is imputed by linear

extrapolation

c) Based on County Business Patterns, Census Bureau

Note: Statistics for 1079 un-weighted observations

101

 

Table 3.3. Estimation Results

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

Job Growth Income Growth
(1) (2) (3) (4) (5) (6)
A11 Metro Non- All Metro Non-
Metro Metro
Employment Share 1.83 3.1 1 1.69 0.12 0.05 0.18
in 14 Employee (0.31)” (1.06)” (0.37)” (0.12) (0.43) (0.13)
Businesses
Employment 0.02 1.13 -0.08 -0.09 -0.10 -0.08
in 5-9 (0.16) (0.41 )"‘ (0.19) (0.06) (0.17) (0.07)
Employment 0.20 1.68 -0. 13 -0.07 -0.06 -0.06
in 10-19 (0.16) (0.34)“ (0.20) (0.06) (0.14) (0.07)
Employment 0.39 0.53 0.27 -0.03 -0.15 0.02
in 20-49 (0.14)” (0.28)‘ (0.18) (0.06) (0.11) (0.06)
Employment 0.15 0.26 0.06 -0.04 -0.32 0.03
in 50-99 (0.12) (0.23) (0.15) (0.05) (0.10)” (0.05)
Employment -023 0.25 -0.40 -004 -009 -0.00<
in 100249 (1.10)“ (0.19) (0.13)” (0.04) (0.08) (@4)
Employment -0.08 0.53 -033 -0.06 -013 -002
in 250499 (0.09) (0.14)”' (0.12)” (0.03)‘ (0.06)” (0.04)
Employment -0.16 0.20 -025 -0.06 -0.08 -005
in 500999 (0.08)‘ (0.12) (0.11)“ (0.03)‘ (0.05) (0.04)
Education -0.69 -044 -0.48 -005 -020 -0.01
(0.69) (0.95) (1.03) (0.26) (0.39) (0.36)
Manufacturing 0.00< 0.09 -0.03 -0.04 -0.02 ~0.08
(0.03) (0.03)” (0.05) (0.01)” (0.01) (002)”
Government -0.01 -0.09 -0.03 -0.02 -0.01 -0.03
(0.02) (0.06) (0.02) (0.01)" (0.02) (0.01)“‘
Farming 0.02 0.59 -0.22 -0.15 -004 -032
(0.22) (0.29)" @38) (0.08)‘ (0.12) (0.14)“
Business Density -0.09 -O.62 -l.53 0.26 0.33 -0.60
(0.54) (0.60) (2.58) (0.21) (0.24) (0.91)
Constant -0.94 -2.10 0.14 -2.76 -3.36 -250
(0.86) (1.02)” (1.38) (0.33)‘” (0.42)” (0.48)“
Observations 996 492 504 996 492 504
Number of counties 83 41 42 83 41 42
Log likelihood -3123.85 -1417.69 -1644.57 -2161.81 -975.28 -1117.16
Wald ch12 299.60 209.83 174.47 519.26 661.69 155.33

 

 

 

 

 

Standard errors in parentheses; "‘ signiﬁcant at 10%; " signiﬁcant at 5%; **"‘ signiﬁcant at 1%; Time
dummies are suppressed for display purpose

102

 

Table 3.4. Distribution Index and County Economic Growth

 

 

 

 

 

 

 

 

 

 

 

 

 

 

Job Growth Income Growth
Index 0.68 0.26
(0.26)” (0.09)”
Index Squared -0.01 -0.005
(0.005)” (0.002)”
Education -0.29 -0.03
(0.73) (0.26)
Manufacturing 003 -0.04
(0.03) (0.01)”
Government -0.00< -0.02
(0.02) (0.01)”
Farming 0.03 -0. 16
(0.23) (0.08)‘
Business Density -0.05 0.27
(0.57) (0.20)
Constant -0.49 -2.71
(0.90) (0.32)”
Log likelihood -3188.70 -2161.35
Wald chi2 141.41 520.66

 

Standard errors in parentheses; 996 annual observations in 83 counties; * signiﬁcant at 10%; " signiﬁcant

at 5%; "* signiﬁcant at 1%; Time dummies are suppressed for display purpose

103

 

Table 3.5. Alternative Estimation Methods

 

Income Growth

 

 

 

 

 

 

 

 

 

 

 

 

 

 

POLS POLS FE FE FD ABond FEFGLS
Index 0.002 0.059 0.110 0.205 0.079 0.234 0.215
(0.068) (0.060) (0.118) (0.102)” (0.160) (0.098)“ (0096)“
Index Squared -0000 -0001 -0002 -0004 -0001 -0005 -0004
(0.001) (0.001) (0.002) (0.002)” (0.003) (0.002)” (0002)“
Education 0.082 0.001 -1.853 -0902 -0051
(0.034)“ (0.326) (0.599)” (0.559) (0.292)
Manufacturing 0005 -0.043 -0. l 3 1 -0.105 -0.042
(0.002)” (0.012)” (0.019)” (0.015)” (0.011)”
Government -0.007 -0.018 -0.092 -0.057 -0.018
(0.003)” (0.008)“ (0.022)” (0.013)” (0.008)”
Farming 0.003 -0059 -1314 -0.689 -0.076
(0.007) (0.095) (0.325)” (0.226)” (0.088)
Business -0.004 0.321 0.244 0.103 0.323
Density (0.005) (0.245) (0.448) (0.493) (0.222)
Income -0.106
Growth, one (0.038)”
year lag
Constant 1.507 -0.650 0.181 1.189 1.534 0.515 -1.753
(0.897)’ (0.973) (1.572) (2.692) (0.472)“ (0.407) (0.299)”
R-squared 0.00 0.27 0.00 0.29 0.24
Total F- 1.34 22.77 0.69 19.27 24.38 26.94 21.36
statistics
Log likelihood -1985.89

 

 

 

 

 

 

 

 

Standard errors in parentheses; 913 annual observations in 83 counties; "‘ signiﬁcant at 10%; ** signiﬁcant

at 5%; *" signiﬁcant at 1%; Time dummies are suppressed for display purpose

104

 

 

 

Table 3.6. Correlation of Business Size Distribution Measures with Control

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

Variables
(1) (2) (3) (4)
Index Employment Employment Employment
14 20.49 500.999
Education -0.47 -0.05 0.44 -0.38
- (0.26)‘ (0.07) (0.15)” (0.27)
Manufacturing 0.05 -0.00< -0.04 -0.02
(0.01)“ (0.00<) (1.01)“ (0.01)”
Government 001 -0.00< 0.00< -0.02
(0.01) (0.00<) (0.00<) (0.01)"
Farming 0.00< 0.08 0.01 0.14
(0.08) (0.02)“ (0.05) (0.09)
Business Density 0.50 0.21 -0.30 0.03
(0.20)" (0.06)“ (0.12)” (0.21)
Employment Share in 1-4 0.43 -0.26
Employee Businesses (0.07)”. (012)”
Employment Share in 5-9 0.15 0.00< -0.28
(0.02)” (0.03) (0.06)”
Employment Share in 10-19 0.13 -0.40 -0.43
(0.02)” (0.03)” (0.06)”
Employment Share in 20-49 0.10 -0.48
(0.02)” (005)”
Employment Share in 50-99 0.01 -0.40 -0.48
(0.01) (0.02)” (004)”
Employment Share in 100-249 -0.06 -0.30 -0.50
(0.01)” (0.02)“ (0.04)”
Employment Share in 250-499 -0.03 -0.23 -0.51
(0.01)” (0.02)“ (0.03)”
Employment Share in 500-999 -0.02 -0.15
(0.01)” (0.02)”
Constant -0.46 0.21 -0.07 -0.49
(0.32) (0.09)” (0.19) (0.34)
Observations 996 996 996 996
Number of ﬁps 83 83 83 83
Log likelihood -215305 -869.17 -1618.09 -2189.23
Wald chi2 87.21 768.40 695.83 429.99

 

 

 

 

 

 

Standard errors in parentheses. "' signiﬁcant at 10%; ** signiﬁcant at 5%; “”" signiﬁcant at 1%
Fixed Effect FGLS estimation procedure is used. Year dummies are suppressed for display purposes. 913
annual observations for 83 Michigan counties over 1989-2000 time period are used for estimations.

105

 

 

FIGURES

Community Poverty Rate
Michigan, 1999

Census Block Groups
Poverty Rate, °/o
0.0 - 11.3
gig 11.4-28.5
- 28.6 - 78.0
E] County Boundary

 

Q§t_a: Based on Census 2000, US Census Bureau
Figure 1.1. Poverty Rate, 1999

106

 

Michigan Communities

 

  
 
  
   
 

Ru rallU rban
Communities
Rural
- Metropolitan / Rural Towns
Metropolitan Adjacent
Road
lnfrastructu re
- - a Limited Access Highway
------ Principal Highway

I:| County Boundary

Detroit

 

 

M: Based on Census 2000, US Census Bureau
Figure 1.2. Michigan Urban/ Rural Areas by Census Block Groups, 1999

107

_

 

 

Michigan Community Poverty Rates, 2000

   
 

 

 

UF?_
0)
C
.9
‘89“
81
.o
o
“6
‘E
a)
2m~
a)
O.
O- . = = =-
0 20 4 60 80 100

0
Poverty Rates
Source: US Cersus Bureau

Figure 1.3. Distribution of the Community

 

F-Statistics __ _,
0| \1 O N
0 0| 0 01

N
01
sssssss -

All State Rural Rural Metro. Metro.

 

Source: Based on the estimation results

Figure 1.4. Joint Significance of the Labor Demand and Labor Supply
Characteristics

108

 

Michigan Economic Growth

   

Growth, %

3%

Your

-3%

 

 

 

Data: Bureau of Economic Analysis; Inﬂation Adjusted.
Figure 2.1. Michigan Economic Growth Patterns

109

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

Share
80%
70%
50% _ .
' Max
50%4-"7 +30
Mean
40%”l-1 'SD
2U%--
10%-- HT :l:
0% i I l I 1 1.7—‘1 i ‘
zaesseaagsee
c1 1:) c1 ‘7‘ V 0‘ C,
‘—' N to C) d .
e a § '-

 

 

 

Data: County Business Patterns, US. Census Bureau, 1988-2000.
Figure 3.1. Share of Business Establishments by Size Category for 83 Michigan
Counties.

110

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

Share Max
40% +80
Mean
30% -SD
Min
20%
H. r7 ' 7
10% ' 71 r: ' In 11%
"‘ .1 -1 ~ ii
0% ’ ’I I I’ I’ I I I I " .
1r 0) as m m m m m V 3128
:— tn 7 ‘1 m R: Q 8 ‘3
E E: 8 9 § § §

 

 

 

Data: County Business Patterns, US. Census Bureau, 1988-2000.
Figure 3.2. Distribution of Employment across the Business Size Categories for 83

Michigan Counties.

111

 

Average Index of Busines Size Distribution,
Michigan, 1988-2000

     

Michigan
Average Distribution Index '
‘7 1153-183

[L] 18.4 -22.5
- 22.6 - 32.7
- 32.8-45.3

 

Data: Based on County Business Patterns, Census Bureau

 

 

Figure 3.3. Distribution of the Business Size Index across Michigan Counties

112

Entrepreneurial Pipeline Theory

Business Size Distribution and Income Growth, Michigan, 1988-2000

 

 

 

CO -v
. Corr-0.03
" e

° N —1 e ‘ '
9\ $5.52
1:. e ".‘e. A. . I
*" l:-'.. =.' ' "
3 P - a '- _
E n". g. I
0 .
E 1:)
Q
0
ﬂ
—

‘T -t

 

 

 

Y r

1 0 '20 30 40 50
Distribution Index

Note: For average values by county
Data: Based on County Business Patterns (Census Bureau, 2005)
and Local Area Annual Estimates (BEA, 2005)

Figure 3.4. Distribution Index and Income Growth

113

Entrepreneurial Pipeline Theory

Business Size Distribution and Job Growth, Michigan, 1988-2000

 

 

 

 

8.0-1 .
e Com028

v.

0

e\ mg

.15

H

E

‘- C)

U .

D

g a
(:14

 

 

10 20 30 40 50
Distribution Index

Note: For average values by county
Data: Based on County Business Patterns (Census Bureau, 2005)
and Local Area Annual Estimates (BEA, 2005)

Figure 3.5. Distribution Index and Job Growth

114

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

Index,%6
62
—-—-+Std. Dev.
———--Std.Dev.
Min
—Max
‘Q—‘-’~. .'. -.'¢—.-¢'
12 1 1 1 1 1 1 1 1 1 1 1 1
co OO O\ 05 ON 3 ON ON ON 03 O\ O\ O
eeeee_aeeeeem

 

 

 

Data: Based on County Business Patterns, US. Census Bureau, 1988-2000.
Figure 3.6. Index for Business Size Distribution, Michigan Counties.

115

 

Job Growth

 

 

I All I Metro 121 Non-Metro I

 

 

Figure 3.7. Distribution of Employment and Job Growth

116

 

Income Growth

0.15

0.05

-0.05

-0.15

-0.25

 

-0.35

 

I All I Metro 121 Non-Metro

 

 

Figure 3.8. Distribution of Employment and Income Growth

117

REFERENCES

Arellano, M. and S. R. Bond. (1991). “Some Tests of Speciﬁcation for Panel Data: Monte
Carlo Evidence and an Application to Employment Equations.” Review of Economic
Studies, 58(2), pp. 277-97.

Barro, R.J. (1998). Determinants of Economic Growth: A Cross-Country Empirical
Study. MIT Press: Cambridge, MA.

Bartik, T.J. (1991). Who Beneﬁts from State and Local Economic Development Policies?
Upjohn Institute for Employment Research: Kalamazoo, MI.

Barik, T.J. (1997). “Discussion,” New England Economic Review, March/April, pp.67-
71.

Bartik, T. J. (2001). Jobs for the Poor: Can Labor Demand Policies Help? New York:
Russell Sage Foundation.

Bartik, T.J., P.Eisinger and G.Erickcek. (2003). “Economic Development Policy in
Michigan,” In C. Ballard, P. Courant, D. Drake, R. Fisher and E. Gerber (eds.), Michigan
at the Millennium (pp. 279-297). Michigan State University Press: East Lansing, MI.

BEA (Bureau of Economic Analysis). (2005). Local Area Annual Estimates. Available at:
www.bea.gov , January.

Bertrand, M.; Mullainathan, Sendhil & Shaﬁr, Eldar. (2004). “A Behavioral-Economics
View of Poverty,” The American Economic Review. 94(2): pp. 419-423

Besley, T. and A. Case. (2000). "Unnatural Experiments? Estimating the Incidence of
Endogenous Policies," Economic Journal, Royal Economic Society, 110(127), F672-94.

Birch, D. (1987). Job Creation in America. The Free Press. New York.

Birch, D., and J. Medoff. (1994). “Gazelles,” in Solomon, LC, and AR. Levenson
(Eds.), Labor Markets, Employment Policy, and Job Creation (pp. 159-168). Westview.
Boulder, Co.

Blanchard, O. J .; L. F. Katz; R.E Hall and B. Eichengreen. (1992). “Regional
Evolutions,” Brookings Papers on Economic Activity, 1992(1): pp. 1-75

Blank, R., D.Card, F.Levy and J .Medoff. (1993). “Poverty, Income Distribution, and
Growth: Are They Still Connected?” Brookings Papers on Economic Activity, 2: p. 285

Blank, R. (2004). Poverty, Policy and Place: How Poverty and Policies to Alleviate
Poverty Are Shaped by Local Characteristics. Rural Poverty Research Center, Working
Paper 04-02

118

 

Bradshaw, T. (2002). “The Contribution of Small Business Loan Guarantees to Economic
Development,” Economic Development Quarterly, 16(4), 360-69.

Buss, T.F. (2001). “The Effect of State Tax Incentives on Economic Growth and Firm
Location Decisions: An Overview of the Literature,” Economic Development Quarterly,
15(1), 90-105.

Census Bureau. (2000). Decennial Census of Population and Housing. US. Department
of Commerce. Available at: www.census. gov , January.

_. (2003). Technical Documentation. Summary File 3. 2000 Census of Population and
Housing.

. (2005). County Business Patterns. US. Department of Commerce. Available at:
www.census.go_v_ , January.

Crandall, M. and B. Weber. (2004). Local Social and Economic Conditions, Spatial
Concentration of Poverty, and Poverty Dynamics. Rural Poverty Research Center,
Working Paper 04-04

CRC (Citizens Research Council of Michigan). (1998). Michigan Renaissance Zones in
the Economic Development Context. Report No. 324, November.

. (2001). Survey of Economic Development Programs in Michigan. Report No. 334,
May.

Dalenberg, D. R. and M. D. Partridge. (1995). “The Effect of Taxes, Expenditures, and
Public Infrastructure on Metropolitan Employment,” Journal of Regional Economics,
35(4), pp. 617-40.

Davis, S., J. Haltiwanger and S. Schuh. (1993). Small Business and Job Creation:
Dissecting the Myth and Reassessing the Facts, NBER Working Paper No. 4492.

Davis, S., J. Haltiwanger and S. Schuh. (1994). “Small Business and Job Creation:
Dissecting the Myth and Reassessing the Facts,” Business Economics, 29(3), p.13.

De’murger, S., Sachs, J. D., Woo, W.T., Bao, S., Chang, G. and Mellinger, A. (2001).
Geography, Economic Policy and Regional Development in China. CID Working Paper
No. 77. Available at: www.cid.harvard.edu/cidwp/077.htm , January, 2002.

DeSousa, C. (2000). “Brownﬁeld Redevelopment versus Greenﬁeld Development: A
Private Sector Perspective on the Costs and Risks Associated with Brownﬁeld
Redevelopment in the Greater Toronto Area,” Journal of Environmental Planning and
Management, 43(6), pp. 831-53.

Dewar, ME. (1998). “Why State and Local Economic Development Programs Cause So
Little Economic Development,” Economic Development Quarterly, 12(1), pp. 68-87.

119

Duclos, J -Y (2002). Poverty and Equity: Theory and Estimation

Emery, M., M. Wall, and D. Macke (2004). “From Theory to Action: Energizing
Entrepreneurship Strategies (E2) Strategies to Aid Distressed Communities Grow Their
Own,” Journal of the Community Development Society, 35(1), pp. 82-96.

Erickcek, G. and B. Watts. (2003). An Economic Development Benchmarking System for
Rural Michigan. W.E.Upj ohn Institute for Employment Research.

Ettinger, N. (1997). “An Assessment of the Small-Firm Debate in the United States,”
Environment and Planning A, vol. 29, pp. 419-442.

Faulk, D. (2002). “Do State Economic Development Incentives Create Jobs? An Analysis
of State Employment Tax Credits,” National Tax Journal, 55, pp.263-80.

Fisher, G. M. (1997) “The Development of the Orshansky Thresholds and Their
Subsequent History as the Ofﬁcial US. Poverty Measure,” Poverty Measurement
Working Papers, US. Census Bureau. available at
http://www.census.gov/hhes/poverty/povmeas/papers/orshansky.htrnl

Fisher, RC. (1997). “The Effect of State and Local Public Services on Economic
Development,” New England Economic Review, March/April, pp. 53-67.

Fisher, P. and A. Peters. (1997). “Tax and Spending Incentives and Enterprise Zones,”
New England Economic Review, March/April, pp. 109-30.

Fitchen, J .M. (1995). “Spatial Redistribution of Poverty through Migration of Poor
People to Depressed Rural Communities,” Rural Sociology, 60(2): pp. 181-201.

Fitchen, J .M. (1994) “Residential Mobility among the Rural Poor,” Rural Sociology,
59(3): pp. 416-36.

Florida, R. (2002). The Rise of the Creative Class. Basic Books. New York.

F omi, M. and M. Lippi. (1997). Aggregation and the Microfoundations of Dynamic
Macroeconomics. Clarendon Press: Oxford.

Fomi, M. and M. Lippi. (1999). “Aggregation of Linear Dynamic Microeconomic
Models,” Journal of Mathematical Economics, 31 : pp. 131-158.

Foster, E. M. and McLanahan, S. (1996). “An Illustration of the Use of Instrumental
Variables: Do Neighborhood Conditions Affect a Young Person’s Chance of Finishing
High School?” Psychological Methods, 3(1): pp. 249-260.

Gabe, T.M. and D.S. Kraybill. (2002). “The effect of State Development Incentives on
Employment Growth of Establishments,” Journal of Regional Science, 42(4), pp. 703-30.

120

Galster, G. (2003). “Investigating Behavioural Impacts of Poor Neighborhoods: Towards
New Data and Analytic Strategies,” Housing Studies. 18(6): pp. 893-914

Glennerster, H. (2002). “United States Poverty Studies and Poverty Measurement: The
Past Twenty-Five Years,” Social Service Review. 76(1): pp.83-107

Goetz, S. (1999). “Migration and Local Labor Markets,” in S. Loveridge (ed.), The Web
Book of Regional Science, Regional Research Institute, West Virginia University.

Goss, E. and J. Phillips. (1999). “Do Business Tax Incentives Contribute to a Divergence
in Economic Growth?” Economic Development Quarterly, 13(3), 217-28.

Green, G. (2001). “Amenities and Community Economic Development: Strategies for
Sustainability,” The Journal of Regional Analysis and Policy, 31(2), pp. 61 —75.

Hanham, A., S. Loveridge, and B. Richardson. (1999). “A National School-Based
Entrepreneurship Program Offers Promise,” Journal of the Community Development
Society, 30(2), pp] 15-129.

Ihlanfeldt, KR. (1999). “The Geography of Economic and Social Opportunity in
Metropolitan Areas” in Governance and Opportunity in Metropolitan America (edited by
Altshuler, A. A), Washington, DC. National Academy Press.

J ohanisson, B. (1991). “University Training for Entrepreneurship: Swedish Approaches,”
Entrepreneurship and Development, 3, pp. 67-82.

Levernier, W., M. Partridge and D. Rickman. (2000). “The Causes of Regional Variation
in US. Poverty: A Cross-County Analysis,” Journal of Regional Science, 40(3): pp.473-
497.

Lichtenstein, G. and T. Lyons. (2006, Forthcoming). “Managing the Community’s
Pipeline of Entrepreneurs and Enterprises: A New Way of Thinking about Business
Assets,” Economic Development Quarterly.

Lichtenstein, G. A., T. S. Lyons, and N. Kutzhanova. (2004). “Building Entrepreneurial
Communities: The Appropriate Role of Enterprise Development Activities,” Journal of
the Community Development Society 35(1), pp. 5-24.

Lipton, M. and M. Ravallion. (1995). "Poverty and Policy," in Handbook of Development
Economics Vol. 3 (edited by Jere Behrman and TN. Srinivasan), North Holland,
Amsterdam.

Loh, E. (1993). “The Effects of Job-Targeted Development Incentive Programs,” Growth
and Change, 24, 365-83.

Loveridge, S. (1996). “On the Continuing Popularity of Industrial Recruitment,"
Economic Development Quarterly. 10(2), pp. 151-158

121

Lucas, E.L., Jr. (1978). “On the Size Distribution of Business F irms,” The Bell Journal of
Economics, 9(2), pp. 508-23.

Luce, T.F., Jr. (1994). “Local Taxes, Public Services and Firm Location,” Public Finance
Quarterly, 22(2), pp. 139-76.

Madden, J. F. (2003). “The Changing Spatial Concentration of Income and Poverty
among Suburbs of Large US Metropolitan Areas,” Urban Studies. 40(3): pp. 481-503.

Malecki, E.J. (1994). “Entrepreneurship in Regional and Local Development,”
International Regional Science Review, 16(1&2), pp. 119-53

Martinez, J. AB. (2005). “Equilibrium Entrepreneurship Rate, Economic Development
and Growth. Evidence from Spanish Regions,” Entrepreneurship & Regional
Development, 17 (March), pp. 145-61.

Mayer, P. and T. Lyons. (2000). “Lessons from Private Sector Brownﬁeld Redevelopers:
Planning Public Support for Urban Regeneration,” Journal of the American Planning
Association, 66(1), pp. 46-57

McQuid, R. W. (2002). Entrepreneurship and Regional Development Policies, Working
Paper, Employment Research Institute, Napier University, Edinburgh, UK.

Moulton, BR. (1986). “Random Group Effects and the Precision of Regression
Estimates,” Journal of Econometrics. 32: pp. 385-97.

Muske, G., and M. Woods. (2004). “Micro Businesses as an Economic Development
Tool: What they Bring and What they Need,” Journal of the Community Development
Society, 35(1), pp. 97-116.

Nizalov, D., and S. Loveridge. (2005). “Regional Policies and Economic Growth: One
Size Does Not Fit All,” Review of Regional Studies, 35(3), pp. 266-290.

Papke, L. (1991). “Interstate Business Tax Differentials and New Firm Location:
Evidence from Panel Data,” Journal of Public Economics, 45, 47-68.

. (1993). “What Do We Know about Enterprise Zones?” in J .Poterba (ed.), Tax
Policy and the Economy, vol. 7, pp. 37-72.

. (1994). “Tax Policy and Urban Development: Evidence from the Indiana
Enterprise Zone Program,” Journal of Public Economics, 54, 37-49.

Partridge, M. and D. Rickman. (2003). The Evolution of US. County Poverty Rates
During the 1990s: Is there a Spatial Dimension? Presented on the 50th annual Regional
Science Association International meeting, Philadelphia, November.

122

. (2005). High-Poverty Nonmertopolitan Counties in America: Can Economic
Development Help? Presented on the 51St annual Regional Science Association
International meeting, Seattle, WA.

Powers, ET. and M. Dupuy. (1994). “Understanding Differences in Regional Poverty
Rates,” Economic Commentary, Nov 15,

Rupasingha, A., S. Goetz, and D. Freshwater. (2002). “Social and Institutional Factors as
Determinants of Economic Growth: Evidence from The United States Counties,” Papers
in Regional Science, 81(2), pp. 139-55.

Sadoulet, E. and A. de J anvry, 1995. Quantitative Development Policy Analysis. The
Johns Hopkins University Press: Baltimore.

Schultz, TR (1988). “Education Investments and Returns,” in H.Chenery and
T.N.Srinivasan (eds.), Handbook of Development Economics, V.3, Amsterdam: North
Holland Press, sections 3 and 4.

Shimeles, A., B.Kebede and A. Bigsten. (2000). Poverty, Income Distribution and Labor
Markets in Ethiopia.

Singh, 1., Squire, L., Strauss, J. (1986). Agricultural Household Models: Extensions,
Applications and Policy. World Bank.

Site Selection, various issues. Conway Data Inc.

Sohl, J ., and W. Rosenberg. (2003). “The US Angel and Venture Capital Market: Recent
Trends and Developments,” Journal of Private Equity, 6(2), pp. 7-17.

Strauss, J. and Thomas, D. (1995) Human Resources: Empirical Modeling of Household
and Family Decisions, Section 5.

Strauss, J. and Thomas, D. (1998). “Health, Nutrition and Economic Development,”
Journal of Economic Literature, 36(2), pp. 766-817.

Thomburgh, D., 1998. “A Path of Smarter Federal Leadership in Economic
Development: Learning, Leveraging, and Linking,” Economic Development Quarterly,
12(4), pp. 291-98.

Tolbert, C. M. and M. Sizer, 1996. US. Commuting Zones and Labor Market Areas: A
1990 Update. Economic Research Service, Staff Paper No. AGES-9614, September, 39

p.

Triest, R. (1997). “Regional Differences in Family Poverty,” New England Economic
Review. Jan/Feb, p.3

Von Bargen, P., D. Freedman, and E. Pages. (2003). “The Rise of the Entrepreneurial
Society,” Economic Development Quarterly, 17(4), pp. 315-324.

123

Walstad, W. (1994). Entrepreneurship and Small Business in the United States. Institute
for Entrepreneurship Education, Center for Entrepreneurial Leadership, Inc. Ewing
Marion Kauffman Foundation. Kansas City, Mo.

Walzer, N. (2004). “Introduction: Entrepreneurship in Community Development,”
Journal of the Community Development Society, 35(1), pp. 1-4.

Wasylenko, M., 1997. “Taxation and Economic Development,” New England Economic
Review, March/April, pp. 37-52.

Watson, K., S. Hogarth-Scott, and N. Wilson. (1998). “Small Business Start-Ups:
Success Factors and Support Implications,” International Journal of Entrepreneurial
Behavior and Research, 4(3), pp. 217-238.

Weber, B. and L. Jensen. (2004). Poverty and Place: A Critical Review of Rural Poverty
Literature. Rural Poverty Research Center, Working Paper 04-03

Wolcoff, M. (1992). “Is Economic Development Decision Making Rational?” Urban
Affairs Quarterly, 27(3), 340-55.

Wooldridge J .M. (2002). Econometric Analysis if Cross Section and Panel Data, MIT
Press.

Wooldridge J .M. (2003). Introductory Econometrics: A Modern Approach, (2d ed.)
Thomson South-Western.

124

1111111111111111111