Do economic effects justify the use of fiscal incentives?

Murray, Matthew N.

1. Introduction

Recruitment of large industrial and commercial facilities is a key aspect of most local development strategies. Policy makers presumably act on the assumption that new sitings will create positive shocks and significant net benefits for the local community. It is this premise that justifies the significant inducements and incentives that appear very large relative to the magnitude of the locating firm. (1) Economists have dabbled around the margin of this debate. Oates and Schwab (1991), for example, argue that interjurisdictional fiscal competition is efficiency enhancing by bringing business tax payments in closer alignment with services received from the public sector. Inman and Rubinfeld (1996), on the other hand, argue that competition for mobile capital may lead to a local public sector that is too small. This paper explores the question of whether large, mobile companies, the typical targets of aggressive industrial recruitment activities and recipients of lucrative incentives, have positive net impacts on regional economics. In the spirit of Jacobson, LaLonde, and Sullivan (1993), we apply the tools of the program evaluation literature to address this question.

There are several reasons why the net economic effect of a new location could in fact be well short of the direct (or gross) employment and investment effects that are the basis for granting concessions. First, concessions erode the fiscal gains of location. Tax incentives limit revenue gains, and other concessions (such as site acquisition and development) require expenditure of public funds. Foregone tax revenues and higher public expenditures mean that state and local governments must either provide fewer public services or impose higher taxes on existing industry and residents to maintain balanced budgets.

Second, a significant new location can easily crowd out other economic activity. For example, Porter (1999) found that megasports events, specifically Super Bowls, have no real effects on spending in host communities, a result he attributes in part to the crowding-out phenomenon. The 1996 Atlanta Olympics provides another example. Spending by Olympic patrons created significant short-term jobs and income. At the same time, the Olympics crowded out other economic activity, as fewer other special events were held during the summer of 1996 and many local businesses not tied to the Olympics lost sales. In fact, anecdotal evidence suggests more than complete crowding out since the total number of visitors to Atlanta during 1996 was down 8.8% from 1995. (2) Also, wages were driven up by labor shortages during both the construction and operational phases. (3)

More generally, the location of a large company can crowd out other economic activity by shifting sales from existing firms, congesting local infrastructure, and raising prices in factor markets; Porter (1999) noted the possibility of diminishing returns in production. The factor market effects can be pronounced, especially for existing firms that produce in highly competitive national or international markets. Further, large companies (like Mercedes Benz and BMW) may prefer locations without other significant, visible firms sited in the same jurisdiction so that they can dominate the business and civic community. (4) Thus, the incentive for other firms to locate may be diminished by the preexisting presence of large firms in an area.

The business location literature has focused on the ways in which market forces and public policies influence firm location, investment, and job creation. (5) This paper reverses the focus of previous research by modeling the discrete location behavior of large firms with an eye on estimating the net impact of location on traditional measures of regional economic growth. Although our reduced form model does not explicitly take into account the concessions that are granted to firms nor the specific channels through which firm location positively or negatively influences regional economic performance, we are nonetheless able to identify the net effect of large firm location on regional income and employment. Panel data techniques and extensive statistical tests are applied to a primary database of large companies that made location decisions in the 1980s. We control for nonrandom site selection on the part of firms and nonrandom company selection on the part of communities in order to estimate the time path and magnitude of net economic effects on host communities. The primary finding is that the location of a large firm has no measurable net economic effect on local economies when the entire dynamic of location effects is taken into account. Thus, the siting of large firms that are the target of aggressive recruitment efforts fails to create positive private sector gains and likely does not generate significant public revenue gains either.

The next section of the paper details the model of firm and community selection leading to an empirical model of regional economic growth. An important aspect of this model is the discrete and noncompetitive process that typifies the siting of large firms. This is followed by discussion of the data used in the empirical model. Empirical findings are then reported and discussed.

2. Conceptual Framework

Economic growth and decline occur as firms locate, expand, contract, and exit the economy in response to market forces and public policies. The focus of this analysis is on the role that large, newly locating firms play in the regional growth process and the extent to which the large firms influence the propensity of other firms to locate, expand, contract, or exit. The large firms can be viewed as demanders of sites, and the communities can be seen as suppliers, either directly (through industrial parks) or indirectly (through accommodating land use controls). (6) The negotiations that characterize the siting process of large new companies suggest the presence of market power on both the demand and supply sides of the market. Other firms are modeled separately (below) as competitive agents in the regional economy.

Large firms determine their demand for sites by examining the expected profitability associated with each site based on a region's demand, cost, policy structure, and amenities. Note that firms may not choose locations with the best profit potential, possibly accepting lower market returns in exchange for amenities (see Fox and Murray 1990). During the negotiating process with communities, firms reveal their general intentions including planned production, capital investment, employment, and nonpayroll spending. Communities, as discussed below, may choose to induce location through provision of incentives that improve site-specific profits or may seek to limit access to sites through zoning or other restrictions. Together, market forces and information on incentives enable firms to formulate expectations on intersite profitability. Generally this information is not available to a researcher.

Just as firms evaluate site-specific profits, states and localities determine their willingness to supply sites based on evaluations of potential returns from the location of large companies within their jurisdiction. (7) In practice states and substate jurisdictions have separate indices of expected returns from large company locations. The state would have greater interest in state-wide benefits and costs, whereas localities would be expected to hold a more parochial view of benefits and costs. It is difficult to know exactly how specific states and communities will evaluate the returns to location since the weights applied to benefit and cost factors will vary dramatically. Narrowly, the calculus might reflect the perceived surplus (deficit) of revenues over public sector costs. Other factors that might be included are economic effects such as jobs, income, and changes in industry structure as well as the prestige effects that accrue to political leaders when a large location occurs.

In cases where net economic, fiscal, and/or political surpluses are anticipated, states and communities may seek to increase the probability of a large firm locating through grants, abated taxes, provision of training services, site development assistance, and other concessions. There is presumably some maximum amount that a community would be willing to give away to attract a large new company, some or all of which may be provided to a locating firm through concessions and other forms of support. A stick, such as zoning or other regulatory controls, might be used instead of a carrot in cases where policy makers perceive net costs instead of net benefits from a specific location. Information on pieces of incentive packages actually accepted by firms is often available, whereas unaccepted incentive offers are generally unobserved.

The noncompetitive location process begins as firms reveal their production and investment plans and scrutinize alternative sites. Communities then make their best offer based on expected returns. Firms respond by accepting or rejecting community bids or by making a counter offer. (8) A new round of negotiations could begin if the offer is not accepted. Generally, it is impossible to observe the sites examined by a large firm and the firm's assessment of intersite profitability as well as incentives offered and incentives accepted. But actual company locations are observed to take place in specific regions, in which case it can be assumed that the location satisfies both the company's required minimum or reservation level of profit and the community's required return on available sites.

An important question is the ex post effect of location on regional economic performance. One approach to estimating the economic effect of a large company's intervention is to specify a detailed structural model capturing the optimization behavior of firms, factors of production, and policy makers. Unfortunately, we cannot directly estimate nor indirectly identify the component structural equations corresponding to firms' expected profitability and communities' expected returns from large company locations. To do so would require detailed information on each company and each community including the range of factors, especially incentives or restrictive policies, considered in evaluating rates of return across sites. However, a structural model is not necessary to measure the net economic effect that large firms have on the siting area. Net effects can be assessed through a reduced form model that differentiates between the competitive behavior of most locating, expanding, contracting, and exiting firms and the noncompetitive behavior and location of many large, newly locating firms and their potential host communities.

The behavior of competitive industry is reflected in a stylized reduced form model of regional economic growth,

(1) [E.sub.it] = [X.sub.it]B,

where [E.sub.t] is some aggregate measure of region i's economic activity in period t (such as the level of employment or the growth in personal income), [X.sub.it] includes the attributes of the economy that encourage/retard growth (including taxes and labor costs), and B are the agent's responses to market and public policy variables. As a reduced form, this specification encompasses the behavior of firms, factors of production, and policy makers.

For empirical implementation, Equation 1 is amended to accommodate the location of large firms as follows:

(1) [E.sub.it] = [X.sub.it]B + [D.sub.i][alpha] + v,

(2) [MATHEMATICAL EXPRESSION NOT REPRODUCIBLE IN ASCII.]

(3) v = [v.sub.i] + [v.sub.t] + [v.sub.it].

The variable D reflects the presence or absence of a company location in a given community, and a is the estimated mean impact of location on the specific measure of substate growth. (9) The term a, which measures only employment or income effects in the analysis that follows, is a subset of the returns that communities consider when evaluating the returns to large firm location. The term v is expanded to accommodate a two-way error structure for panel data applications including time-specific and placespecific effects on economic activity. A random-effects model assumes each term is independently and identically distributed as (O, [[isgma].sup.2.sub.v]); alternatively, a fixed-effects model assumes that [v.sub.i] and [v.sub.t] are fixed parameters to be estimated.

3. Data

An exhaustive process was followed to generate a viable database and an appropriate econometric model, and each step is discussed in detail below. The setup for the empirical analysis relies on three data-intensive issues and three econometric issues based on concerns prior to development of the econometric findings. This section addresses the three data issues: (i) identifying locations and the general issue of control regions, (ii) selecting the measure of economic performance, and (iii) choosing an appropriate unit of regional analysis.

A set of treatment counties and two sets of control counties were developed for the study. The treatment group is a 100% sample of all counties where locations of large firms have occurred. Large locations were defined as all new sitings between 1980 and 1989 that involved an expected employment of at least 1000 people. The sitings range in size from 1000 employees to 7594 employees with a median of 1200. The number of locations by year is given in Table 1 and by state of location in Table 2. The requirement that locations take place in the 1980s permits the time path of economic performance before and after the location to be examined. A very comprehensive process was used to identify the sitings of new firms versus firms that simply relocated or changed ownership. A primary data collection effort was undertaken as every state's economic development agency or other appropriate agency was contacted to determine where and when such locations occurred. (10) In addition, secondary data were collected by reviewing a number of popular publications such as Site Selection Magazine, which commonly provides listings of large firms and their sites of location. Locating firms and local government agencies were also contacted when necessary to ensure the data on place, time of location, and announced employment were accurate. Locations that proved to be expansions at existing sites were omitted. No limitations were placed on the industries in which firms could operate. The industry could not always be identified, but the data include firms in manufacturing; services; construction; finance, insurance, and real estate; retail and wholesale trade; and transportation, communications, and public utilities industries. Because of the size of these firms and the nature of their products or services, it can be assumed that they are producing their goods or services for national or international markets.

Since the location of large firms is not a random event, control groups composed of untreated (nonlocation) counties were necessary to identify the impact of location on treated (location) counties. (11) One control group was formed by taking a 10% random sample of all nonlocation counties. (As described below, this control group indicated evidence of selectivity bias, and pooling was rejected.) A second control group was developed that reflected a larger regional economy and was composed of all Metropolitan Statistical Areas (MSAs) as of 1996. (12) The possibility of contamination bias (i.e., the inclusion of a control county or MSA where a large location has in fact occurred) exists to the extent that the primary and secondary data collection efforts failed to identify all large firm locations.

The 93 locations that occurred in MSAs took place in 46 different MSAs located in 26 states. The control sample includes 188 MSAs located in 33 states (reliable data were unavailable for 17 states). The location MSAs were on average larger with a median population of 852,372 in 1990 compared with a median population of 193,889 in the control MSAs. As can be seen in Figure 1, the location MSAs grew faster. For example, the median location MSA had a compound annual employment increase of 2.36% during the 1980s compared with 1.84% for the control MSAs.

[FIGURE 1 OMITTED]

The empirical work is undertaken using both personal income and employment growth as separate indicators of economic effects from large locations. However, the basic findings are the same so that the results are only provided for employment effects, which are measured for both the county and the broader area in which the location occurs. The county is the area of primary economic impact, and this well-defined local government unit is often responsible for at least part of the local government location incentives as well as policies that may discourage firm location. The area is defined as the entire MSA for metropolitan locations or the location county and all contiguous counties for nonmetropolitan locations. The area is expected to encompass a wider set of the economic effects that result from locations.

A time series for both the location and nonlocation communities is necessary to study the time path of economic performance. (13) Employment and income data for all counties and areas in the samples were obtained for 1972-1995. (14) For purposes of this study, the relevant time dimension is the number of years before or after the location, rather than the specific calendar year in which the location occurred. Thus, data for each treatment place (either county or area) were centered on the location year. Time is then reflected as the number of years before or after the location. Centering was not directly possible for control counties since by definition no location has occurred. Control places were randomly centered (without replacement) across the years 1980-1989, and time is measured relative to the number of years before or after the randomly selected centering point.

A balanced and an unbalanced data panel are used alternately in the empirical analysis. (15) Both panels include the entire cross section and a constant, although different, length for the time series. The difference is that the balanced panel includes a fixed number of years before and after the location. Specifically, the balanced panel is composed of annual data for five years before and five years after the location plus the year of the location. Thus, the time series for every place has exactly the same number of years before and after the location. The unbalanced panel has a time series using all years from 1972 to 1995. However, the number of years before and after the location varies based on the calendar year in which the data are centered. Thus, for a 1980 location, eight years of data are available before it and 15 years after it; however, for a 1989 location, 17 years of data are available before and six years of data after.

Unweighted average employment growth in treatment areas was generally higher than for the control MSAs, both before and after the location (see Figure 1). (16) Of course, the growth in some prelocation years potentially can be explained at least partially by construction activity, whereas growth in the postlocation years can be explained by the new firm's operations. But the location areas had higher growth in 23 of the 32 years shown in the figure. The higher growth rates suggest that there is something systematically different about the two samples that could go beyond the direct locational effects. (17) One explanation is that there is a different growth-generating process arising because the geographic site of location areas is consistently preferred over nonlocation MSAs, a potentiality that is addressed in the empirical analysis. This same phenomenon heightens concerns over the potential for selection bias, since locations seem to occur predominantly in high-growth areas. On the other hand, the reverse is true for location counties (see Figure 2). The location counties experienced consistently slower growth than the average control county, suggesting that firms locate in slow-growing counties in high-growth metropolitan areas.

[FIGURE 2 OMITTED]

4. Empirical Analysis

The key issue for this paper is estimation of [alpha] in Equation 2. Proper econometric analysis of this question requires a number of steps. First, the reduced form equation necessary to estimate Equation 2 must be developed. Second, a statistical analysis must be undertaken to determine whether the treatment and control panels described in the data section can be pooled for estimation purposes. Third, it is necessary to test whether the error term and the right-hand side variables are correlated (i.e., that selection bias is present), which would lead to biased parameter estimates. Finally, the econometric estimates can be generated once the previous econometric issues have been settled. The remainder of this section describes the detailed reasoning and findings for the first four steps, whereas the econometric estimates are provided in the following section.

Proxy variables are used in the reduced form equation to control for market and public policy influences on growth. First, GDP growth is used to control for macroeconomic trends, national product demand, and the national policy environment. Weak national growth may depress product demand, in turn reducing the absolute attractiveness of all potential sites and vice versa. Second, measures of state growth are used to control for cross-sectional variation in the business climate including factor market, public policy, and regulatory conditions across state and substate regions. The assumption is that measures of statewide growth embody, as a composite, the diverse structural elements like tax burdens and labor costs that influence growth within counties and MSAs. Third, having accounted for national and state market conditions, substate regions may still display differential growth due to variations in industry mix and other unique place-specific factors. For example, during recovery from recession an area characterized by a dominant manufacturing sector may enjoy stronger growth than regions with a large agricultural or service sector. Thus, the percentages of employment in manufacturing and in agriculture are included as separate regressors. (18) Finally, we utilize random and fixed effects to control for remaining features of the regional economy.

Location of a large business is evaluated as a noncompetitive activity occurring through a negotiated process that takes place simultaneously with the normal growth process of MSAs and counties. The effect of these locations is examined through two indicators. The first, called Location, is a dummy variable with a value of one in the location and all subsequent years and a zero otherwise. Location always takes on a zero value for control places. The second is a set of six dummy variables with a value equal to zero in prelocation years and one in the location year (Location 1), the year after location (Location 2), and so forth. The first approach requires that the economic effects from a location be the same for all years after the location; the second allows for the impacts to vary by year after the location.

The empirical analyses are performed separately for counties (control and location counties) and MSA areas (control and location areas). Estimation of the model depends on the relationship between v in Equation 4 and the right-hand-side variables of Equation 2. Of particular interest is the variable D, since the presence of any correlation would lead to biased parameter estimates. In accounting for these problems, we followed Bassi (1984) who suggested a step-by-step approach to model specification and identification of intervention impacts in her analysis of job training programs. (19) The approach explores alternative estimators--ordinary least squares (OLS), random effects, and fixed effects--that sequentially accommodate a more sophisticated structure between v and D based on model specification tests. For a given estimator, we first conducted F-tests to determine whether the prelocation economic growth process captured by Equation 2 is the same for the treatment and control groups. This is essential to ensuring that a picks up any regime shift from location as opposed to other causes of different patterns of economic growth across the two groups. Alternative control groups or estimators are required if there are underlying differences in the growth patterns. The F-tests are estimated using prelocation data, but no simple determination of what is meant by a prelocatinn time period can be made because some growth effects from construction or pre-announcement expectations may exist prior to the actual location. Thus, two time periods were examined: one for the time up to three years before the location and one for the time up to one year before the location. The shorter time period was used to minimize any bias in the estimates from pre-announcement expectations or construction activity prior to the location.

The F-tests were conducted using each control group and its counterpart of location areas, that is, all MSA areas and the 10% random sample and both the MSA area and location county treatment groups. State employment growth, the GDP growth rate, and various subsets of the control variables that account for the economic structure (the percentages of employment in manufacturing and in agriculture) were used in the equations. The F-tests, estimated using a wide variety of model specifications, were always significant using the county control group, indicating that there were systematic differences in the growth processes of the location and control counties during the prelocation period. Although the remaining analysis emphasizes the results for the MSA control group, we also present estimates for counties, but offer the caution that these findings may be biased if fixed effects do not adequately control for different regional patterns of growth. Little evidence was found of systematic differences between the MSA control group and location counties.

The second step is to look for correlations between the error term v and the regressors in the growth equations, in particular D, to determine if selection bias is present. Correlations between the two could provide evidence that firms systematically selected high (or low) growth places for location and that differences in growth between location and control places are due to some preexisting differences, rather than to the effects of a large facility. A series of growth equations was estimated using data from the two prelocation time periods, OLS, and random effects estimators. (20) Separate equations were run using the county database and the area database. The key issue is whether dummy variables D for the location counties or MSA areas were significant in the prelocation time period, providing evidence of selection bias as firms chose places that were already growing at a different pace. In addition to the dummy variable, every combination of the regressors listed above was tried.

Evidence is found that selection based on fixed effects is taking place in the location of employment in the area specifications with the dummy variable being positive and statistically significant in the OLS equations. The dummy variable is significant in most of the random effects equations, but no estimates could be obtained with the industry structure variables present. The fixed effects model is an acceptable control for nonrandom selection when selection takes place on the basis of fixed factors and therefore is the preferred estimator for the area equations. (21) The t-values on the fixed effects are always significant at a higher degree of confidence for the panel data up to one year before location than for the panel up to three years before location, suggesting the occurrence of a growth spurt immediately before location that is perhaps linked to construction of the large facility or expectations regarding future supplier linkages. Overall, these findings indicate that businesses are likely to select high-growth places for locations, but the higher growth can be explained with fixed effects or with characteristics of the local economic structure. The dummy variable was negative but only statistically significant in one of the county equations, suggesting more limited evidence of selectivity bias.

Growth Equation Results

A listing of empirical results for employment growth equations is presented in Table 3 for counties and in Table 4 for MSAs. (22) In each case, the dependent variable is annual growth in employment for the place. The preferred fixed effects results are presented when the estimated variance is positive; OLS and random effects estimates are also shown for comparison. Equations based on a balanced time series, with the same number of years of data before and after each location, and an unbalanced time series, with the total available number of years, are provided for each equation.

The state growth variable, which is included to account for cross-section effects of the full range of policy variables and growth conditions, is highly significant in every equation. In general, the coefficient estimates are very close to 1, indicating a strong relationship between statewide factors and regional economic growth. On average, substate regions grow at a rate roughly commensurate with the growth of the state itself. GDP, intended to pick up time series effects of the national policy environment and national market conditions, is entered into every equation and is significant in some cases. (23) The percentages of employment in manufacturing and in agriculture, which operate as fixed effects across areas, are generally significant. The fixed effects estimator could not be used with these variables since they include no intertemporal variation. (24)

The Location variable, taking the value one in the location and all subsequent years, is never significant, suggesting the large plant locations had no effect on growth. There is some evidence that location has a net positive effect using the annual location variables (Location 1-Location 6), although the results are mixed. With the unbalanced equations and the preferred fixed effects model, significant effects are found contemporaneously with the location (Location 1) and one (Location 2) and three (Location 4) years after location. An F-test of the combined significance of the annual location variables was insignificant. Further, the only effect with the balanced equation is negative five years after the location. At least one location variable is significant in each of the OLS equations and the unbalanced equations with the random effects model. An F-test of the combined effect of including the six annual location variables was significant only in the OLS equation without fixed effects for industry structure.

5. Discussion

Together these findings yield little support for the case that large company locations appreciably influence the path of regional economic growth. The descriptive data indicate that large companies are choosing low-growth counties in rapidly growing MSAs. Low-growth counties may offer less congested infrastructure, relatively lower priced land, more aggressive recruitment efforts, and so on. At the same time, the locating firm will be able to exploit agglomeration economies, infrastructure, amenities, the availability of business services, and so forth that are generally available in the broader metropolitan area, as well as to take advantage of the visibility and prestige associated with a rapidly growing regional economy.

Once we control for national and state economic conditions as well as time and place fixed effects, there is little evidence of positive or negative growth impacts associated with the location of large firms. (25) These findings are consistent with the longer term development of this research project where a wide variety of estimators, data, and model specifications failed to produce a consistent pattern of positive or negative results. (26) The strongest evidence for positive growth impacts comes from the preferred fixed effects model applied to MSAs, but even here there is no consistent finding of any economic stimulus. When the analysis is conducted at the county level, where there is less confidence that control and location counties can be pooled, in only one instance does the coefficient of the location intervention variable show statistical significance. More generally, across all estimators (including the less preferred OLS and random effects models), there are only three instances of positive growth effects and four instances of negative growth effects.

What might explain the findings that large firms have little or no net impact on regional growth? Perhaps the political leaders' and/or communities' goals in business and industrial recruitment were to enhance prestige and community visibility or to promote changes in industry structure (for example, by replacing old, cyclical jobs with new jobs). If these were in fact the communities' objectives, they may or may not have been realized. One thing seems clear: recruitment did not lead to more rapid regional growth. In all likelihood, the absence of significant growth impacts means that large companies simply displace other sources of job and income growth in the regional economy. As discussed above, the underlying dynamics of this displacement may reflect the fiscal consequences of granting incentives, the crowding of regional infrastructure, and/or higher prices in local factor markets. Since growth would have taken place absent the large firm's location, typical business recruitment strategies focused on larger enterprises along with the liberal granting of tax and other incentives to the same firms is simply not a cost-effective economic development strategy.

Table 1. Large Firm Locations by Year

Year Number of Locations

1980 14
1981 10
1982 6
1983 5
1984 8
1985 9
1986 10
1987 17
1988 17
1989 13
Total 109

Table 2. Large Firm Locations by State

Alabama 6
Arkansas 3
Colorado 3
Connecticut 4
Florida 28
Georgia 1
Illinois 1
Indiana 12
Kansas 7
Kentucky 1
Louisiana 1
Maryland 2
Massachusetts 3
Michigan 1
Mississippi 3
Missouri 3
Nebraska 2
North Carolina 7
Ohio 1
Oklahoma 1
Rhode Island 1
South Carolina 2

Table 3. Regression Estimates for County Location

 State Employment Percent
 Growth GDP Growth Agriculture

Ordinary least squares
 Balanced 0.99 *** -0.00006
 Unbalanced 0.98 *** 0.00003
 Balanced 0.99 *** -0.00005
 Unbalanced 0.97 *** 0.00004
 Balanced 0.98 *** -0.00001 0.06 *
 Unbalanced 0.96 *** 0.00009 0.07 ***
 Balanced 0.98 *** -0.00006 0.06 ***
 Unbalanced 0.96 *** 0.0001 0.07 ***
Fixed effects
 Balanced 0.94 *** -0.0002
 Unbalanced 0.93 *** 0.0003 **
 Balanced 0.94 *** 0.002
 Unbalanced 0.93 *** 0.0003 **
Random effects
 Balanced 0.95 *** 0.0002
 Unbalanced 0.95 *** -0.0002 *
 Balanced 0.95 *** 0.0002
 Unbalanced 0.94 *** 0.0002 **
 Balanced 0.95 *** 0.0002 0.07
 Unbalanced 0.94 *** 0.0002 ** 0.06
 Balanced 0.95 *** 0.0002 0.07
 Unbalanced 0.94 *** 0.0002 * 0.06

 Percent
 Manufacturing Location Location 1

Ordinary least squares
 Balanced -0.0009
 Unbalanced -0.001
 Balanced -0.0009
 Unbalanced -0.001
 Balanced -0.03 *** -0.001
 Unbalanced -0.02 *** -0.001
 Balanced -0.03 *** -0.001
 Unbalanced -0.02 *** -0.002
Fixed effects
 Balanced 0.01
 Unbalanced 0.002
 Balanced 0.002
 Unbalanced 0.003
Random effects
 Balanced 0.0006
 Unbalanced 0.001
 Balanced 0.001
 Unbalanced 0.001
 Balanced -0.03 *** 0.0005
 Unbalanced -0.02 * 0.002
 Balanced -0.03 *** 0.001
 Unbalanced -0.02 *** 0.001

 Location 2 Location 3 Location 4

Ordinary least squares
 Balanced
 Unbalanced
 Balanced 0.0003 0.003 0.0005
 Unbalanced 0.0001 0.002 0.0003
 Balanced
 Unbalanced
 Balanced -0.0002 0.002 -0.00005
 Unbalanced -0.0003 0.002 -0.0002
Fixed effects
 Balanced
 Unbalanced
 Balanced 0.003 0.006 0.002
 Unbalanced 0.004 0.007 ** 0.003
Random effects
 Balanced
 Unbalanced
 Balanced 0.002 0.005 0.001
 Unbalanced 0.003 0.005 * 0.003
 Balanced
 Unbalanced
 Balanced 0.002 0.005 0.001
 Unbalanced 0.003 0.005 * 0.003

 Location 5 Location 6

Ordinary least squares
 Balanced
 Unbalanced
 Balanced -0.003 -0.006 *
 Unbalanced -0.003 -0.006 **
 Balanced
 Unbalanced
 Balanced -0.004 -0.006 **
 Unbalanced -0.004 -0.007 ***
Fixed effects
 Balanced
 Unbalanced
 Balanced -0.002 -0.005
 Unbalanced -0.0003 -0.003
Random effects
 Balanced
 Unbalanced
 Balanced -0.002 -0.005
 Unbalanced -0.001 0.004
 Balanced
 Unbalanced
 Balanced -0.002 -0.005
 Unbalanced -0.001 -0.004

* t-test significant at the 0.10 level.

** t-test significant at the 0.05 level.

*** t-test significant at the 0.01 level.

Table 4. Regression Estimates for Metropolitan Statistical
Area (MSA) Location

 State
 Employment Percent
 Growth GDP Growth Agriculture

Ordinary least squares
 Balanced 0.99 *** -0.00001
 Unbalanced 0.99 *** 0.00007
 Balanced 0.99 *** 0.00003
 Unbalanced 0.99 *** 0.00007
 Balanced 0.95 *** 0.0002 0.27 ***
 Unbalanced 0.95 *** 0.0003 ** 0.34 ***
 Balanced 0.95 *** 0.0002 0.27 ***
 Unbalanced 0.95 *** 0.0003 *** 0.33 ***
Fixed effects
 Balanced 0.97 *** 0.0001
 Unbalanced 0.97 *** 0.0002
 Balanced 0.97 *** 0.0001
 Unbalanced 0.97 *** 0.0002 *
Random effects
 Balanced 0.97 *** 0.00009
 Unbalanced 0.98 *** 0.0001
 Balanced 0.97 *** 0.0001
 Unbalanced 0.97 *** 0.0002
 Balanced 0.96 *** 0.0001 0.27 ***
 Unbalanced 0.97 *** 0.0002 * 0.33 ***
 Balanced 0.96 *** 0.0001 0.27 ***
 Unbalanced 0.97 *** 0.0002 * 0.33 ***

 Percent
 Manufacturing Location Location 1

Ordinary least squares
 Balanced 0.003
 Unbalanced 0.002
 Balanced 0.003 *
 Unbalanced 0.002
 Balanced -0.02 *** 0.001
 Unbalanced -0.01 *** 0.002
 Balanced -0.01 *** 0.003
 Unbalanced -0.01 *** 0.002
Fixed effects
 Balanced 0.0008
 Unbalanced 0.003
 Balanced 0.0001
 Unbalanced 0.003 *
Random effects
 Balanced 0.001
 Unbalanced 0.003
 Balanced 0.001
 Unbalanced 0.003
 Balanced -0.02 ** 0.001
 Unbalanced -0.01 ** 0.002
 Balanced -0.02 ** 0.001
 Unbalanced -0.01 ** 0.003

 Location 2 Location 3 Location 4

Ordinary least squares
 Balanced
 Unbalanced
 Balanced 0.004 ** 0.004 ** 0.004 **
 Unbalanced 0.004 ** 0.003 * 0.003
 Balanced
 Unbalanced
 Balanced 0.004 ** 0.003 * 0.003
 Unbalanced 0.003 * 0.003 0.002
Fixed effects
 Balanced
 Unbalanced
 Balanced 0.001 -0.0005 0.0002
 Unbalanced 0.005 ** 0.003 0.004 *
Random effects
 Balanced
 Unbalanced
 Balanced 0.003 0.002 0.002
 Unbalanced 0.004 ** 0.003 * 0.003 *
 Balanced
 Unbalanced
 Balanced 0.002 0.001 0.001
 Unbalanced 0.004 ** 0.003 0.003 *

 Location 5 Location 6

Ordinary least squares
 Balanced
 Unbalanced
 Balanced 0.001 0.0001
 Unbalanced 0.0006 0.0003
 Balanced
 Unbalanced
 Balanced 0.0006 0.0002
 Unbalanced -0.00009 0.0005
Fixed effects
 Balanced
 Unbalanced
 Balanced -0.003 -0.004 *
 Unbalanced 0.0002 -0.0004
Random effects
 Balanced
 Unbalanced
 Balanced -0.001 -0.002
 Unbalanced 0.001 -0.0003
 Balanced
 Unbalanced
 Balanced -0.002 -0.002
 Unbalanced -0.000006 -0.0005

* t-test significant at the 0.10 level.

** t-test significant at the 0.05 level.

*** t-test significant at the 0.01 level.

(1) Political explanations for incentive packages also exist (see Fox and Mayes 1994).

(2) The number of conventions held in Atlanta fell 10.9%, and the number of convention attendees dipped 10.4% in 1996. Net of Olympic visitors and travel to Atlanta was down 36% in 1996. (Atlanta Convention and Visitors Bureau 2001).

(3) For example, see Business Bulletin, Wall Street Journal, September 28, 1995.

(4) On the other hand, some firms may choose locations where they benefit from agglomeration benefits and direct supplier linkages derived from the location of other large firms.

(5) Recent reviews of the literature are in Fisher (1997) and Wasylenko (1997).

(6) For a discussion of the supply and demand for industrial sites in competitive markets, see Fox (1978).

(7) Economic and fiscal impact studies are common ways of evaluating some of the private and public sector benefits and costs associated with a newly locating company. In practice, such studies typically examine gross rather than net economic and fiscal impacts, often ignoring important factor market and consumer substitution effects.

(8) This is the process that took place between the Boeing Corporation and well over a dozen states in 2003.

(9) It is not necessary to estimate explicitly the firm and community selection equations to identify [alpha]. See Bassi (1984) and Heckman and Robb (1985).

(10) Some states, including Alaska, California, Hawaii, and New York, were omitted from the sample because they either did not cooperate in the data collection or the data were not of acceptable quality.

(11) Greenstone and Moretti (2003) use an alternative experimental design that includes counties that became hosts to large new companies and runner-up counties. They argue that the use of framer-up counties is superior to randomly chosen control groups, since the former can be viewed as close substitutes to sites actually chosen for location, thus reducing any unobserved heterogeneity in the empirical analysis. However, this may not be the case in practice since no information is available on why potential sites were revealed or not revealed by firms and communities. For example, firms may behave strategically and reveal a runner-up county only because it made a lucrative incentive offer that served the company's interest of eliciting a higher bid from the county of ultimate location. In such a case, unobserved heterogeneity may still exist across location and runner-up counties.

(12) Ninety-three of the locations were in MSAs. The control group only includes MSAs from states where data on locations were used.

(13) Depending on the specific application, cross-section, and/or panel data, including pre- and postlocation information, may be adequate to identify the impact of location (Heckman and Robb 1985).

(14) In what follows we discuss only the employment data as a similar pattern emerged for income data.

(15) An unbalanced panel does not lead to econometric problems when the lack of balance is random and not the result of agent choice. See Wooldridge (2002).

(16) The number of observations drops for years above six and below eight because of the truncation at the ends of the data set.

(17) The control MSAs grew faster in seven of the prelocation years and only two postlocation years.

(18) The economic structure variables are the shares of employment five years before the location so that there is no simultaneity between the share and the location. This approach precludes time series variation in the economic structure data.

(19) More generally, see Heckman and Robb (1985) and Friedlarlder, Greenberg, and Robins (1997).

(20) This test cannot be run using the fixed effects estimator.

(21) No evidence of selection bias is found using income as the dependent variable.

(22) Results for the income equations are available from the authors upon request. The location variables are never statistically significant in the county income equations and are significant in only two cases in the MSA equations.

(23) The effects of national growth may also be picked up in the state growth variable.

(24) Random effects estimates are based on a one-way random en-or, including only the error across areas.

(25) By contrast, Greenstone and Moretti (2003) found evidence that wages in the one-digit industry and county of location were stimulated by the siting of a large firm. These results do not directly conflict with ours since Greenstone and Moretti's (2203) findings are for a single industry and a single county. This more narrowly construed result fails to account for the possibility of other activity being crowded out in surrounding counties or in other industries. They separately estimated effects on other industries in the location county and on contiguous counties and generally found a positive, although statistically insignificant, effect. However, a direct comparison with our results would require that their analysis be conducted for the overall economy in the broad economic area in which the location occurs. They also concluded that property values in the location county were increased, but the property value analysis was based on a significantly restricted sample. Further, they speculated that the effects were caused by the state providing part of the location subsidies, but payment of the subsidies increased the chance of property value losses in other parts of the state (including other parts of the MSA) that could result in no net effect, as was concluded in the current study.

(26) Fox and Murray (1998) provide a preliminary version of the analysis that is confined to only locating areas and only 24 company locations.

References

Atlanta Convention and Visitors Bureau. 2001. Unpublished data.

Bassi, Laurie. 1984. Estimating the effect of training programs with non-random selection. Review of Economics and Statistics 66:36-42.

Burstein, Melvin L., and Arthur J. Rolnick. 1996. Congress should end the economic war for sports and other businesses. The Region 10:35-6.

Business Bulletin. 1995. Wall Street Journal, 28 September.

Fisher, Ronald C. 1997. The effects of state and local public services on economic development. New England Economic Review March/April:53-56.

Fox, William. 1978. Local taxes and industrial location. Public Finance Quarterly 6:93-114.

Fox, William, and David Mayes. 1994. "Are Tax Incentives Too Large?" Paper presented at the Proceedings of the Eighty-Seventh Annual Conference of the National Tax Association, Charleston, SC.

Fox, William, and Matthew Murray. 1990. Local public policies and interregional business development. Southern Economic Journal 57:413-27.

Fox, William, and Matthew Murray. 1998. Incentives, firm location decisions and regional economic development. In Local government tar and land use policies in the United States, edited by Helen F. Ladd. Northhampton, MA: Edward Elgar, pp. 168-81.

Friedlander, Daniel, David H. Greenberg, and Philip K. Robins. 1997. Evaluating government training programs for the economically disadvantaged. Journal of Economic Literature 35:1809-55.

Greenstone, Michael, and Enrico Moretti. 2003. Bidding for industrial plants: Does winning a "million dollar plant" increase welfare? NBER Working Paper No. 9844.

Heckman, James J., and Richard Robb, Jr. 1985. Alternative methods for evaluating the impact of interventions. In Longitudinal analysis of labor market data, edited by James J. Heckman and Burton Singer. Cambridge, UK: Cambridge University Press, pp. 156-245.

Inman, Robert P., and Daniel L. Rubinfeld. 1996. Designing tax policy in federalist economics: An overview. Journal of Public Economic 60:307-34.

Jacobson, Louis S., Robert J. LaLonde, and Daniel G. Sullivan. 1993. Earnings losses of displaced workers. American Economic Review 83:685-709.

Oates, Wallace E., and Robert M. Schwab. 1991. The allocative and distributive implications of local fiscal competition. In Competition among states and local governments, edited by Daphne Kenyon and John Kincaid. Washington, DC: Urban Institute Press, pp. 127-45.

Porter, Philip K. 1999. Mega-sports events as municipal investments: A critique of impact analysis. In Sports economics: Current research, edited by John Fizel, Elizabeth Gustafson, and Lawrence Hadley. Westport, CT: Praeger, pp. 61-74.

Wasylenko, Michael. 1997. Taxation and economic development: The state of the economic literature. New England Economic Review March/April:37-52.

Wooldridge, Jeffrey M. 2002. Econometric analysis of cross section and panel data. Cambridge, MA: MIT Press.

William F. Fox * and Matthew N. Murray ([dagger])

* Department of Economics and Center for Business and Economic Research, 1000 Volunteer Boulevard, 100 Glocker Building, The University of Tennessee. Knoxville, TN 37996-4170, USA; E-mail billfox@utk.edu; corresponding author.

([dagger]) Department of Economics and Center for Business and Economic Research, 1000 Volunteer Boulevard, 100 Glocker Building, The University of Tennessee, Knoxville, TN 37996-4170, USA; E-mail mmurray1@utk.edu.

Received July 2001; accepted August 2003.