thy neighbor’s jobs: geography and labor market dynamics

Regional Science and Urban Economics 33 (2003) 663–693www.elsevier.com/ locate/econbase

T hy neighbor’s jobs: geography and labor marketdynamics

a b ,*Spencer P. Glendon , Jacob L. VigdoraWellington Management Inc., Boston MA, USA

bDuke University, Durham NC, USA

Received 27 October 1999; received in revised form 23 December 2002; accepted 2 January 2003

Abstract

This paper develops three explanations for the extent of correlation between neighboringgeographic areas’ economic outcomes. Export-oriented firms in neighboring counties mightindependently produce similar goods, or might be linked directly through the production ofintermediate inputs. In either case, counties are exposed to similar demand shocks. Finally,regions share markets for goods and services that are both produced and consumed locally.Empirical results suggest that much of the ‘risk’ associated with economic decline inneighboring regions can be attributed to industrial similarity rather than direct dependenceof jobs in one area on jobs in another. 2003 Elsevier B.V. All rights reserved.

Keywords: Employment spillovers; Local labor markets; Regional development

JEL classification: E32; R11; R23

1 . Introduction

It is deeply intuitive that adjoining geographic areas, whether bounded bystreets, natural topography, or political delineations, are closely linked econ-omically. Not only are such areas likely to share common attributes such as

*Corresponding author. Terry Sanford Institute of Public Policy, Duke University, Durham, NC27708, USA. Tel.:11-919-613-7354; fax:11-919-681-8288.

E-mail address: [email protected](J.L. Vigdor).

0166-0462/03/$ – see front matter 2003 Elsevier B.V. All rights reserved.doi:10.1016/S0166-0462(03)00006-1

mailto:[email protected]

664 S.P. Glendon, J.L. Vigdor / Regional Science and Urban Economics 33 (2003) 663–693

climate and accessibility, they often feature similar industrial composition andproduction technologies. Additionally, neighboring areas share markets for manylocally produced and consumed goods. While a small but growing body ofliterature has considered these issues (Krugman, 1991; Ghosh and Wolf, 1997;Hanson, 1998), much economic analysis of geographic regions overlooks thisrichness of interaction between neighbors.

This paper examines the links in economic fluctuations that exist between UScounties and tests three hypotheses regarding causal mechanisms underlying thelinks. In doing so, we model counties as having two basic categories ofemployment: export-sector employment, which produces goods for a globalmarket, and local-sector employment, which provides goods and services toconsumers in the immediate area. Labor demand in the export sector is determinedexogenously from the perspective of the locality. Local sector labor demand isdetermined endogenously.

The potential exposure of neighboring counties to similar exogenous export-sector demand shocks motivates two of the three hypotheses. The first might betermed an ‘indirect’ link: counties might produce similar goods for consumption inthe global marketplace; therefore when demand for a certain product declinesemployment reductions will tend to occur in several neighboring countiessimultaneously. This hypothesis implies that a shock to one county should beaccompanied by a shock of similar relative magnitude in a neighboring county,even if the first county is small relative to the second. We test this hypothesis byexamining spatial correlation inrelative employment shocks across neighboringcounties. Evidence suggests that this causal channel is significant; correlation inrelative employment shocks can be detected in regions up to 300 miles indiameter. Indirect export-sector links are most pronounced in industries heavilyreliant on fixed-location inputs: agriculture and mining.

The second ‘direct’ export-sector link involves firms directly associated with theproduction process in the originating county. A shock to a producer in one countyaffects suppliers of intermediate goods in the surrounding area. Unlike theindirect-link hypothesis, this theory implies that theabsolute size of employmentshocks matters. A 1% employment loss in a relatively small county should havefewer cross-border repercussions than an equivalent percentage point loss in alarge county. Our test of this hypothesis focuses on spatial correlation in absoluteemployment shocks across counties. Evidence supports this link, especially in themanufacturing sector, where upstream–downstream linkages are presumably mostimportant.

The final hypothesis concerns the effects of exogenous labor demand shocks onendogenously determined local sector employment. Decreases in export-sectoremployment in one county may in turn lead to decreased demand for local-sectorgoods both in the original county and in neighboring areas. As with the directexport-sector hypothesis, this explanation implies that employment in one countydepends to some extent on continued employment in another county. The data

S.P. Glendon, J.L. Vigdor / Regional Science and Urban Economics 33 (2003) 663–693 665

show surprisingly little support for this hypothesis, suggesting that cross-sectoreffects are very localized in nature.

In addition to examining the development of employment shocks acrossgeographic space, we will also consider development over time. Our motivationfor this component of analysis can be seen in Fig. 1. This graph, more fullyexplained in Section 2 below, shows the predicted impulse response to an initialemployment shock measured at two levels of geographic aggregation within theUS: county and state. The significant difference in impulse responses implies thatthe measured after-effects of an initial shock are greater when larger geographicareas are used as the basis for analysis. Simple contemporaneous correlation ofshocks across counties within states is not sufficient to explain this pattern. Ourevidence suggests that this pattern arises because employment shocks, both withinthe export sector and across sectors, take time to propagate across space. Theafter-effects of an initial employment shock are proportionally larger in countiessurrounding the origin of the shock.

These results have many implications for both practitioners and scholars ofregional development. Regions that share risk from employment fluctuations couldbenefit from region-specific automatic stabilization policies. A rational fiscalfederalist system would explicitly consider this risk-sharing in assigning re-sponsibilities to varying levels of government. Our findings do suggest thatcounties share risk, in the sense that employment patterns in neighboring countiesmove together. Losses in one county tend to be associated with losses in nearby

Fig. 1. Impulse response to an employment shock dashed lines represent boundaries of 95% confidenceintervals.


counties in the long run. Much of the risk that individual counties are exposed to,however, is ‘diversifiable’, since cross-county correlations are rooted to a largeextent in similarity of export industries rather than direct upstream–downstream orexport-local sector linkages. That is to say, counties with an export employmentportfolio relatively uncorrelated with that of their neighbors can expect the impactof downturns in those neighboring areas to be relatively limited. Spillovers fromone county’s export sector to its neighbor’s local sector are quite small inmagnitude.

Section 2 describes our data source and motivates the temporal focus of ourstudy. Section 3 describes our three hypotheses and the relevant empirical tests foreach. Details on our methodology and the main empirical findings are presented inSection 4. Section 5 concludes.

2 . Employment dynamics in counties and states

The data used for the estimation of employment dynamics in this paper arederived from the Regional Economic Information System (REIS). The REISprovides excellent data on employment and earnings by one-digit sector at the

1county and state levels. Our sample uses data covering the period from 1969 to1996. Unlike similar datasets, the REIS is not a sample but rather a full census ofemployers based on Census Bureau ES-202 data, which in turn comes from

2employers’ Social Security and tax records. According to the Census Bureau, theES-202 data cover approximately 98% of all private sector employment. In thissection we use REIS data to investigate employment fluctuation patterns at varyinglevels of geographic aggregation.

2 .1. Autocorrelation patterns

Previous research has sought to determine the importance and long runimplications of economic shocks to geographic regions. Bartik (1991) andBlanchard and Katz (1992) estimate the impact of employment fluctuations onlocal outcomes such as unemployment rates and wage levels for metropolitan areasand states, respectively. Clark (1998) shows that idiosyncratic shocks to groups ofstates are a very large component of the overall variation in employment. One

1The REIS data also provides information on population, transfer payments, and local economicconditions including unemployment.

2While other datasets including County Business Patterns also have the variables we care about, theyare based on samples and are therefore prone to substantial statistical noise. In an autoregressiveframework, such noise can lead to exaggerated estimates of mean reversion, since movements thatappear to be shocks are frequently only measurement error. Further research in this area might makeuse of non-public use data at the Census Bureau, such as the Longitudinal Research Database or theStandard Statistical Establishment List, both of which provide greater levels of industry detail.


result that appears consistently in this research is that regional shocks toemployment are permanent, in the sense that employment levels tend not to revertto their previous value following a fluctuation. Blanchard and Katz show that themain equilibrating force operating in labor markets following an employmentshock is household migration. A second common result is that regional employ-ment shocks exhibit autocorrelation. That is, a one-time shock tends to be followedby further movements in the same direction.

Both of these empirical patterns—the permanence and autocorrelation ofemployment shocks—can be observed in Fig. 1, which plots impulse responsefunctions derived from the following county-level autoregression:

Dx 5r(L) Dx 1a 1 u , (1)jt jt21 j jt

and the corresponding state-level autoregression:

Dx 5b(L) Dx 1a 1 u , (2)jt jt21 j jt

wherej indexes jurisdictions (either states or counties),t indexes time,x representsthe log of aggregate employment,a is a jurisdiction-specific fixed effect, andu is

3an idiosyncratic shock. The vector of autocorrelation coefficientsr(L), whereL isthe lag operator, indicates the degree to which initial employment shocks areamplified or dampened in subsequent periods. The autoregressions used to produceFig. 1 both use a fourth-order polynomial forr. The inclusion of the fixed effectterm, a , reflects our belief that there are persistent long-term differences inj

4employment growth across regions. There are two important notes to this analysis.First, employment here and throughout the paper is expressed relative to nationalemployment, so that fluctuations are measured relative to national trends. Thisnormalization eliminates national macroeconomic trends as a source of variation inemployment. Second, the county-level autoregression in Eq. (1) is weights eachobservation by its share of state employment, so that states may be properlythought of as aggregations of counties.

The impulse response functions generated from county- and state-level data aresignificantly different from one another. The confidence intervals, generated by aMonte Carlo procedure for each autoregression, do not overlap. This divergenceimplies that the after-effects of an employment shock are more pronounced in data

3The functional form in Eq. (1) reflects the assumption that employment is an integrated AR(p)process, wherep is the order of the polynomialr(L) in Eq. (1) orb(L) in Eq. (2). In Eq. (1), forexample, we assume that

x 5 x 1njt jt21 jt

andn 5r(L)n 1a 1e .jt jt21 j jt

On the assumption of a unit root in relative employment, see Blanchard and Katz (1992).4The detrending of the data implied by the inclusion of fixed effects may bias our results towards

finding less persistence in employment shocks (Campbell and Mankiw, 1987).


with greater geographic aggregation. An intuitive explanation for this pattern isthat the after-effects of an isolated employment shock spill over across counties.Aggregating employment data to the state level therefore permit the observation ofa greater proportion of these after-effects. As discussed in the following subsec-tion, this intuition does not necessarily hold theoretically—it is possible, undercertain circumstances, to observe impulse response functions like these even whencounties are autarkic, with completely independent trends in employment. Weperform a simulation that suggests this possibility is highly unlikely, confirmingthe intuitive interpretation.

The coefficient estimates underlying Fig. 1 are displayed in Table 1. Thedivergence in county and state responses begins the year following a shock: thestate-level first-order autocorrelation parameter exceeds its county-level counter-part by roughly one-third. Second- and third-order autocorrelations are alsogreater when measured at the state level. The fourth-order parameter breaksthis mold—there is some dampening of employment shocks in the fourth year

5in both states and counties. As Fig. 1 illustrates, the net effect of these coefficient

Table 1County and state employment autoregressions

Independent variable Dependent variable:D ln(Employment),relative to national trend

County level data State level data(estimates ofr) (estimates ofb )

L*D ln(Employment) 0.290 0.391(0.004) (0.030)

2*L D ln(Employment) 20.008 0.019(0.004) (0.032)

3*L D ln(Employment) 0.022 0.083(0.004) (0.033)

4*L D ln(Employment) 20.031 20.080(0.004) (0.031)

2R 0.330 0.398N 70748 1127

Note: Standard errors in parentheses.L represents the lag operator. Regressions include county andstate fixed effects. Observations from Alaska and Hawaii are excluded from analysis. All independentvariables are measured relative to national trend. Observations in the county regression are weighted bythe county’s average share of state employment over the years 1969–1996.

5The significantly negative fourth-order parameter estimates here are not inconsistent with themonotonically decreasing impulse response functions depicted in Fig. 1. The complete lagged responseto a shock in the fourth year is not simplyr , the fourth-order autocorrelation parameter, but rather:4

2 2 2 4r 1 2r r 1r 1r r 1 2r r 1r .4 3 1 2 1 2 1 2 1


6differences is a significant disparity in the amplification of an initial shock.

2 .2. The case for autarkic counties

Are the employment dynamics shown in Fig. 1 and Table 1 consistent with amodel where counties are autarkic, having no systematic relationship withsurrounding areas? Put differently, is it possible to justify the observed relationship

7betweenb(L) andr(L) without hypothesizing systematic links between counties?Theoretical analysis of this question, which appears in Appendix A, shows that

it is indeed technically possible forb(L) to exceedr(L) when counties areautarkic. Intuitively, this occurs when counties within a state experience sys-

8tematically different growth rates. Theory proves to be a rather unhelpful guide,however, in assessing whether the observed relationship betweenb(L) andr(L) islikely under conditions of autarky. To better gauge this possibility, we performed asimulation.

We began by simulating the process of employment growth in counties.Employment dynamics are assumed to follow the process in Eq. (1), using thepoint estimates forr(L) listed in Table 1. Counties were assigned an initialemployment level from a lognormal distribution: initial log employment in eachsimulated county is a draw from theN(ln(40 000),1) distribution. Counties werealso assigned fixed effect (or alternatively, trend) parameters randomly from a

6Input–Output models of regional employment, such as the Bureau of Economic Analysis’ RIMS IIand the proprietary IMPLAN and REMI models, produce multiplier estimates that increase with thesize of the geographic region under consideration. This regularity is imbedded in the assumptionsunderlying each model—new firms are modeled as purchasing a certain quantity of inputs from localproducers. As the geographic size of a region increases, more of these intermediate producers areincluded in the multiplier estimate. Note that an Input–Output interpretation of Fig. 1 is only tenable ifsome cross-county multiplier effects occur with a time lag. This interpretation coincides withHypothesis 2: correlated shocks to export industries, discussed in Section 3.3 below.

7The observed difference in county and state impulse responses might result from a highererror-to-truth ratio in county employment data, which would bias the coefficients in the countyemployment autoregression towards zero relative to the state coefficient. Our state data is simply a sumof our county observations, so such a scenario could only occur if county employment data covarynegatively. There is some evidence that this may occur, because some employment within a state isassigned to various counties in a shifting manner. Running autoregressions with sectors prone to thisactivity (i.e. transportation, communication, and utilities) omitted leads to very similar impulseresponse functions.

8Suppose, for example, that one county in a state doubles in size while 99 other counties of equalinitial size remain stagnant. Normalizing the initial employment level in each county to 1, the 100%shock at the county level, growth from 1 to 2, represents a 1% shock at the state level, or employmentgrowth from 100 to 101. Suppose further that the growing county increases employment from 2 to 3 theyear following the initial shock, a 50% increase. The state therefore grows from 101 to 102. Themagnitude of this after-effect relative to the initial shock is 50% measured at the county level, butequals 99% at the state level.


normal distribution. In each period, log employment levels were subjected to arandom, independent, normally distributed shock. The simulation was allowed torun for 35 time periods.

To simulate states, we simply aggregated the simulated employment levels for60 simulated counties over the 35 year period. Once 50 simulated states wereaccumulated, we ran the regression model indicated in Eq. (2), allowing state fixedeffects and four lagged terms on the right hand side. This entire procedure wasrepeated 1000 times to generate a simulated distribution ofb(L) coefficients underthe assumption of autarky. Summary statistics for this distribution are printed inTable 2.

The results indicate that it is essentially impossible to replicate the aggregatedresults of Table 1 while maintaining the assumption of county autarky. The actualcoefficient estimates forb , b and b all lie above the 95th percentile of their1 2 3

9respective distributions. The actual estimate forb falls in the lower part of the4

simulated distribution, at approximately the 10th percentile. The joint probabilityof observing the vector of coefficientsb in Table 1 when counties are autarkic isthus on the order of 1 in 100 million. According to this simulation, if countieswere truly autarkic, we would actually expect state-level autocorrelation measuresto be less than county-level measures. The results in Fig. 1 and Table 1 arepatently inconsistent with the assumption of autarky.

2 .3. Beyond autarky

The rejection of the autarkic county hypothesis corresponds to our basicintuition. Indeed, the notion that neighboring areas share economic fates might beconsidered patently obvious. A natural question to ask at this point is whether

Table 2Simulated distribution of state autoregression coefficients

Independent variable Dependent variable:D ln(Employment),relative to national trend

5th Percentile Median 95th Percentile

L*D ln(Employment):b 0.211 0.254 0.30012*L D ln(Employment):b 20.084 20.039 0.01023*L D ln(Employment):b 20.050 20.003 0.04134*L D ln(Employment):b 20.092 20.049 20.0094

Note: The coefficient distributions reported here are derived from a simulation described in detail inAppendix A. The simulation assumes that counties are autarkic and employment dynamics behaveaccording to the point estimates in Table 1. The parameter distributions reflect the expected distributionof state autoregression coefficients when counties are autarkic.

9The estimate ofb reported in Table 1 actually lies above all 1000 of the coefficients generated in1

the simulation. The estimate ofb lies above 997 of them.3


simple contemporaneous correlation in the shocks affecting local areas couldcreate the discrepancy noted in Fig. 1. Simple logic shows that the answer is ‘no’.

Consider the case where employment fluctuations in all counties within a stateare perfectly correlated: a 1% shock to one county implies a 1% shock to allcounties within the state. It is relatively straightforward to show that if employ-ment in all counties follows the process in Eq. (1), and all counties are exposed tothe same employment shocks in each period, then the estimated coefficients in Eq.(2) should exactly match those in Eq. (1). Intuitively, if all counties in a stateexperience a 1% shock, which is augmented by an additional 0.29% in thesubsequent period, then the state in aggregate experiences a 1% shock augmentedby an additional 0.29% in the subsequent period.

To explain Fig. 1, we must presume both that employment shocks are correlatedacross counties, and that some portion of this correlation occurs with a time lag.The following section introduces three different models of regional employmentdynamics consistent with these two presumptions.

3 . Justifying cross-county autocorrelation

3 .1. Distinctions between export and regional industries

In this section, we will formally introduce the distinction between export and10local industries discussed in the introduction. Export industries produce for a

market extending beyond the region and are subject to demand shocks that can beconsidered exogenous to the region. The level of demand for export goodsproduced in jurisdictionj will be considered a scalar here, implying that the exportsector inj is best thought of as producing a unique composite commodity. Thedemand for local output from jurisdictionj is endogenously determined, andpresumed to be directly related to the overall level of production both inj and inareas neighboringj. The local sector here can also be thought of as producing acomposite commodity.

Within this framework, we identify three possible reasons for correlation influctuations across localities. First, export-oriented firms located in neighboringareas may produce similar goods, owing to similarity of natural resourceendowments or pooling in the market for skilled labor or other inputs. Similarityof the composite export goods creates the potential for spatial correlation indemand shocks. Second, export-oriented firms may rely on other firms within theregion that supply intermediate goods. These upstream–downstream linkages in

10This distinction is introduced to the regional economic literature by North (1955), who uses theterms ‘export’ and ‘residentiary’ industries.


production create ‘direct’ correlation between demand shocks, as distinguished11from the ‘indirect’ mechanism described above. Third, the multiplier effect of

export demand shocks on employment in local industries may spill over acrosslocalities.

The models developed in this section assume that innovations in employmentresult from shocks to labor demand, rather than changes in labor supply. In thisview, labor supply in a locality is elastic and reacts quickly to changes in demandconditions. Wage responses to employment demand shocks are temporary innature, as combinations of migration and changed commuting patterns quicklyarbitrage any differentials arising from demand innovations. Previous authors(Bound and Holzer, 1996; Holzer, 1989; Blanchard and Katz, 1992) have testedthe validity of these assumptions by various methods, including instrumentalvariables techniques and by examining the relationship between employment andwage innovations. These tests generally support the assumption that year-to-yearfluctuations in employment in geographic areas, especially relative to local trends,result from shocks to labor demand.

3 .2. Hypothesis 1: indirect correlation in export-sector demand shocks

Consider an economic region consisting ofJ localities, indexedj[[1, J].Export-sector employment in each locality,E, is exposed to idiosyncratic shocksthat may amplify or diminish over time. In addition to this serial autocorrelation,shocks may be spatially autocorrelated. The log change in export-sector employ-ment in any locality j at time t can thus be expressed as:

D ln(E ); e 5r (L)e 1O f (L)e 1a 1 u (3)jt jt j jt21 j it21 j jti±j

To incorporate the possibility of contemporaneous correlation of demand shocksacross localities, we will, without loss of generality, assume that:

u 5c u 1v , j [ [2, J], (4)jt j 1t jt

12wherev is distributed independently ofu . Here,c measures the correlation ofjt 1t j

export sector employment shocks between localities 1 andj. The system ofJequations indicated by (3), withJ21 modified to reflect the contemporaneouscorrelation in (4), can be estimated as a vector autoregression using the method ofseemingly unrelated regressions. While it should be possible to estimate the

11Upstream–downstream linkages may show up in various temporal order. For example, in somesituations suppliers of intermediate inputs may lay off workers before the final producer does, if ordersfor the intermediate inputs cease before production stops downstream.

12Generality is maintained here because it is possible to reparamaterize Eq. (4) to express shocks inall localities exceptj as a function of current shocks inj plus an independent error term.


parametersr andf using data on only one region, in practice we will use panelj j

data for estimation purposes.An important assumption to note in this model is that the correlation between

employment shocks in one locality and another does not depend on the relativesize of the two localities. Effectively, the model implies that a 1% shock in CookCounty, IL, should have the same proportional effect on export-sector employmentin neighboring DuPage County as a 1% shock in Wapello County, IA, has onneighboring Mahaska County, even though the ratio of county populations is 6:1in the former case and 1.6:1 in the latter. This is a plausible assumption ifcross-county correlations reflect only similarity in export industries, rather thandirect connections between firms in one county and firms in its neighbor. Thisemphasis on the relative, rather than absolute size of shocks distinguishes theempirical test of the indirect correlation hypothesis from the direct correlationhypothesis discussed in the following section.

3 .3. Hypothesis 2: direct correlation in export-sector demand shocks

If shocks to export sector employment propagate geographically throughupstream–downstream linkages between firms, then the absolute size of the initialshock should matter at least as much as the relative size of a shock. Using theexample from the preceding paragraph, a 1% shock to Cook County, which inabsolute magnitude represents about 6% of DuPage county’s employment, mightreasonably be expected to have a greater cross-county impact than a 1% shock toWapello county, which would represent about 1.6% of Mahaska county’semployment.

To test the second hypothesis, we will focus on the absolute implications forshocks of a given absolute magnitude. In essence, we will test whether theabsolute impact of the loss of 1000 jobs in Cook County for the surrounding areaequals the absolute impact of the loss of 1000 jobs in Wapello County. Such amodel would be appropriate, for example, if a typical firm with 1000 employeesrelied on a firms using a relatively constant amount of labor in other counties toprovide intermediate inputs. The estimated equations will be similar to Eq. (3)above, with the substitution of employment levels for logarithms:

13DE 5r (L)DE 1O f (L)DE 1a 1 u . (5)jt j jt21 j it21 j jt

i±j

Eq. (4), which introduced the possibility of contemporaneous correlation inemployment fluctuations in neighboring areas, will apply in this case as well.

13As in all empirical specifications, employment here will be normalized by national employment toeliminate national macroeconomic trends.


3 .4. Hypothesis 3: spillovers in the local sector

Previous research (e.g. Hildebrand and Mace, 1950; Bartik, 1991; Guthrie,1995) has sought to determine the effect of shocks to export-sector employment onother sectors of the local economy. We posit that the demand for labor in the localsector in any one localityj is a function of the level of employment not only inthat locality, but in its surroundings as well. Presuming that employment in thelocal sector,Y , is integrated and autoregressive, our model is:jt

J14

DY 5n (L) DY 1O m (L) DE 1O g (L) DY 1a 1 u . (6)jt j jt21 j jt j it j jtj51 i±j

This model resembles its immediate predecessor, in that the absolute change inemployment in other sectors and/or localities is of greatest importance indetermining the absolute change in local sector employment. The presence oflagged employment change terms on the right-hand side indicates that local-sectoremployment does not fully respond to export-sector shocks instantaneously.Moreover, there may be inertia in local sector employment trends, as indicated bythe lagged dependent variable. The presence of other localities’DY terms on thejt

right-hand side reflects the presumption that employees in all sectors in neigh-boring jurisdictions demand local-sector output from localityj, and thereforecurrent realizations of local sector employment in one jurisdiction influence localsector employment in neighboring jurisdictions. Unfortunately, with this presump-tion it is impossible to identify the parametersn , m and g in the J-equationj j j

15system implied by (6). Rather than attempt to recover these parameters throughconstraints or functional form assumptions, we will simply estimate a reduced-form version of Eq. (6).

J

DY 5 v (L) DY 1O m (L) DE 1a 1 u (7)jt j jt21 j jt j jtj51

14This formulation can be attributed to Hildebrand and Mace (1950) and North (1955). Tiebout(1956) notes that causality may sometimes run in reverse in this scenario: the existence of a regionalsector can enable the development of an export sector. We believe that Tiebout’s caution is of limitedrelevance in late twentieth century America, where most regions have developed some level ofproduction for export.

15To illustrate this problem, consider the simple case with two regions, where local sectoremployment in either region is determined by the process

Y 5 aE 1 bE 1 cY1 1 2 2

Y 5 dE 1 eE 1 fY .2 1 2 1

The reduced form of this system is given by the following equations:

dc 1 a ec 1 b]] ]]Y 5 E 1 E1 1 212 cf 12 cf

af 1 d bf 1 e]] ]]Y 5 E 1 E2 1 212 cf 12 cf

from which it is not possible to recover the structural parameters.


Here, them coefficients represent reduced-form cross-industry multiplier effects.j

The regional sector spillover explanation for cross-county correlation carriesdifferent implications than the export sector hypotheses outlined above. If theeconomic links between counties derive purely from similarity or interdependenceof export-oriented firms, then it should be possible to sever them, by simplydiversifying the set of industries in one locality away from the set in another. In

16other words, a social planner could essentially fabricate autarky. If neighboringareas are primarily linked by their common market for local goods, however, thenthe management of industrial composition cannot produce autarky. To ensureemployment growth at home, local planners should take coordinated steps toensure the health of the export sector regionwide, since localities within the regionshare significant risk (Persson and Tabellini, 1996).

4 . Methodology and results

4 .1. Data and methods

The framework described above requires that we divide employment into tosectors corresponding to export and regional production. We define export sectorindustries as those related to mining, manufacturing and agriculture. Regionalsector industries include wholesale and retail trade, services, finance, insuranceand real estate, transportation, communication and utilities. This classificationscheme clearly misidentifies some finer industries; some manufacturing activityclearly serves regional markets, for example, and some service sector industriescater to a national or global market. With our data restriction to one-digit

17industries, however, we consider this classification to be the best one available.To check our assumptions, we will perform additional analysis using onlymanufacturing and retail trade as representatives of the export and local sectors,respectively.

Counties will serve as the basic geographic unit of analysis here. We begin byidentifying counties that can be considered central to a multi-county region. These‘employment center’ counties are identified on the basis of two criteria. First, the

16Glaeser et al. (1992) examine the implications of metropolitan industrial diversity, and show thatdiversification predicts greater rates of growth over time.

17The use of one-digit industries as proxies for the export and local sector creates the possibility ofbiased estimates of unknown sign. There are two basic sources of bias. First, since one-digit industriesare noisy measures of export industries, we can expect attenuation bias to push our estimates towardszero. Second, to the extent that the ‘noise’ is correlated positively or negatively across counties, ourresults may exhibit a bias in the same direction. The final direction and magnitude of this bias isuncertain. The sensitivity checks described in the text, however, tend to corroborate the results reportedhere. Future research might expand on these results by making use of non-public use datasets such asthe SSEL.


population centroid of at least one other county must lie within 25 miles of its owncentroid. Second, the total average employment in that county over the years

181969–1996 must exceed that of every other county within a 50 mile radius. Ourestimates of the influence of central counties on surrounding counties musttherefore be qualified. They identify the impact of a ‘typical’ central county on its‘typical’ surroundings.

Radiating out from each central county we then construct concentric rings, 25miles in width, at progressively greater distances. We take all counties withpopulation centroids falling within a given ring and aggregate their data into oneobservation. The observation representing the innermost ring thus groups togethercounties that are less than 25 miles from the relevant employment center. Theoutermost ring in this analysis groups together counties that are between 275 and

19300 miles from the employment center. From each employment center wetherefore have a wide geographical region—with area slightly greater than the landarea of Texas—over which we can track employment activity.

4 .2. Evaluating hypothesis 1: indirect correlations in shocks to export industry

Eq. (3), modified to allow for contemporaneous correlation as in Eq. (4), wasestimated using two lags of export-sector employment in the central county plus12 concentric rings. Selected coefficients from the resulting 13-equation au-

20toregressions are presented in Table 3. These coefficients measure the correlationin employment shocks between employment center counties and areas successivelyfurther away. Additionally, they illustrate the propagation of shocks over time.Each column in the table contains results from a single regression, with logemployment changes in the indicated area, either the central county or one of the12 concentric rings around it, serving as the dependent variable. Rows of the tablecompare the values of relevant parameters for different geographic areas.

The first row of Table 3 compares estimates forc, the coefficient measuring the

18Using this definition, there are a total of 205 central counties in the US, slightly fewer than thenumber of Metropolitan Statistical Areas (MSAs) and Consolidated MSAs. As of June, 1999, therewere 274 MSAs and CMSAs in the continental US. Most central counties are the centers of MSAs. Acomplete list of central counties appears as Table A1.

19By extending the analysis of surrounding areas to a distance of 300 miles, we inevitably includesome central counties in the ‘rings’ surrounding other central counties. Under ordinary circumstances,such methodology would create concerns regarding correlation of the error terms across observations inour regression analysis. However, since the motivating hypothesis of this exercise is that correlationacross counties in the same geographic region does exist, we are equally concerned that autocorrelationwould exist in a more restrictive analysis.

20The following parameters are estimated in the regression models but not reported in Table 3. In thejemployment center county regression, estimates off , for 1# j #10, i.e. the impact of lagged

employment changes in concentric rings on the center. Additionally, the regressions for each concentricjring i omit estimates off for j ± i, the impact of lagged employment changes in other rings on current

changes in ringi employment.

S.P.G

lendon,J.L.V

igdor/

Regional

Scienceand

Urban

Econom

ics33 (2003) 663–693

677

Table 3The impact of relative employment shocks within the export sector

Parameter Dep. Var.: Dependent variable:e in 25-mile ring around employment center with outer radius equal to:t

e in Emp. 25 50 75 100 125 150 175 200 225 250 275 300t

Center county miles miles miles miles miles miles miles miles miles miles miles miles

c – 0.023 0.019 0.037 0.033 0.041 0.029 0.034 0.040 0.035 0.037 0.043 0.034(0.018) (0.014) (0.013) (0.013) (0.013) (0.012) (0.012) (0.012) (0.012) (0.012) (0.012) (0.012)

r 0.217 0.186 0.134 0.228 0.322 0.253 0.285 0.299 0.337 0.260 0.214 0.289 0.2201

(0.014) (0.018) (0.024) (0.027) (0.028) (0.030) (0.032) (0.033) (0.034) (0.035) (0.035) (0.034) (0.033)

r 20.087 20.046 20.006 20.067 20.149 20.007 20.149 20.067 20.098 20.150 20.143 20.073 20.2032

(0.014) (0.018) (0.024) (0.028) (0.029) (0.031) (0.032) (0.034) (0.036) (0.036) (0.036) (0.034) (0.033)

f – 0.018 0.046 0.036 0.024 0.030 0.028 0.032 0.027 0.034 0.030 0.025 0.0281

(0.017) (0.014) (0.013) (0.013) (0.012) (0.012) (0.012) (0.012) (0.012) (0.012) (0.012) (0.012)

f – 0.033 0.029 0.024 0.038 0.022 0.014 0.016 0.026 0.023 0.019 0.024 0.0142

(0.017) (0.013) (0.013) (0.012) (0.012) (0.012) (0.012) (0.012) (0.012) (0.012) (0.012) (0.011)2R 0.059 0.080 0.109 0.141 0.159 0.153 0.156 0.170 0.174 0.164 0.163 0.169 0.164

N 5100 5100 5100 5100 5100 5100 5100 5100 5100 5100 5100 5100 5100

Note: Standard errors in parentheses. Parameter estimates listed in boldface are significant at the 5% level.


contemporaneous correlation in employment shocks between employment centercounties and the indicated concentric ring. There is significant evidence of a small,but significant and persistent, degree of spatial correlation across a wide geog-raphic area. When an employment center county experiences a 1% shock toemployment in the export sector, surrounding counties to a distance of 300 milestend to experience shocks on the order of 0.03–0.04%. Somewhat surprisingly, thecontemporaneous correlation appears to be smallest within 50 miles of the centralcounty, though the coefficients in the first two columns are statistically indis-tiguishable from those in most of the other columns.

This pattern confirms earlier research by Ghosh and Wolf (1997), whodocument strong correlations in employment movements between states thatdiminish with distance. While certainly of substantial general interest, thisevidence does not constitute an explanation for the pattern shown in Fig. 1.Contemporaneous correlation in employment shocks within states does not implythat measures of autocorrelation should increase when aggregated to highergeographic levels. There are two coefficient patterns that could produce anexplanation for Fig. 1. First, the autocorrelation coefficients in concentric ringscould be of greater magnitude than those in the central county. Second,employment in the concentric rings could display a delayed response to the initialcentral county shock itself. The remaining rows in Table 3 provide evidence onboth patterns.

The second and third rows of parameter estimates in Table 3 estimateautocorrelation coefficientsr in the employment center county and surroundingrings. The coefficients show a high degree of similarity across rings, indicatingthat the effect of an area’s own past on current innovations is roughly similar in allparts of a region. In all areas, the estimates ofr indicate that there is a significant1

positive relationship between the current change in export sector employment andthe change in the previous period. The value ofr is negative and significant in2

most specifications, indicating a small degree of reversion two years after a shockto the export sector.

There is not much clear evidence to suggest that autocorrelation coefficients arelarger in concentric rings than they are in central counties. While many estimatesof r significantly exceed the central county estimate of 0.217, many estimates of1

r are significantly less than the central county estimate of20.087.2

The fourth and fifth rows of parameter estimates in Table 3 examine laggedcross-county correlation coefficients, the components of the vectorf from Eq. (3)above. The estimates reported here record the effect of lagged central countyemployment innovations on contemporary changes in concentric rings. Othercoefficients measuring inter-ring intertemporal correlation are omitted from thetable. Here, a significant and telling pattern emerges. Up to a radius of 300 miles,an innovation in central county employment predicts same-signed movements inring employment with a 1- and 2-year lag. Estimates off , the 1-year lagged1

effect of central county employment shocks, are statistically significant at the 5%


level in 10 out of 12 cases, and are generally of similar magnitude to corre-sponding estimates ofc. Estimates off , the 2-year lagged effect, are statistically2

significant in four out of 12 cases and overall somewhat smaller than estimates ofeitherc or f . The cross-area effects of a central county employment shock are1

21thus considerably amplified over time.The implications of Table 3 for regional employment dynamics can be seen in

Fig. 2, which uses the vector autoregression point estimates derived from Eq. (3)to simulate employment dynamics in a region with typical initial employment

22levels in the center and all concentric rings. The figure plots impulse responsefunctions for regions of systematically increasing size, starting with only thecentral county and then proceeding to add successive rings. For each function

Fig. 2. Export sector impulse response functions 1% shock to export sector in central county.

21Table 3 omits a large number of VAR coefficients. The central county equation controls for twolagged employment changes in each ring. Each ring equation controls for two lagged employmentchanges in all other rings. Among other findings, current employment changes in the central county aresignificantly affected by lagged employment changes in other counties within 50 miles. The feedbackeffects implied by these coefficients are incorporated into the impulse response functions in Fig. 2. Thefull set of coefficients is available from the author upon request.

22The average region had 32 840 export sector jobs in the central county; 24 101 in the ring withouter radius equal to 25 miles, 47 954 in the 50-mile ring; 95 695 in the 75-mile ring; 158 108 in the100-mile ring; 176 328 in the 125-mile ring; 236 116 in the 150-mile ring; 258 745 in the 175-milering; 281 222 in the 200-mile ring; 316 049 in the 225-mile ring; 345 138 in the 250-mile ring; 362 503in the 275-mile ring; and 374 965 in the 300-mile ring.


plotted on the graph, the initial shock to export sector employment is normalized23to a value of one.

The uppermost series plotted in Fig. 2 shows the impulse response foremployment center counties. This function reflects the reaction to an initial shockin that county, along with feedback effects that occur when central countyemployment affects surrounding counties which in turn affect the central county.The within-county long-run impact of a 1% negative shock to export-sectoremployment is a roughly 1.16% reduction in employment. The magnification ofthe employment shock in this case is smaller than that recorded for counties in Fig.1, which used employment in all sectors rather than the export sector alone. As theregion under consideration increases in size, the after-effects of an employment

24shock become larger relative to the initial size of the shock. Within a 50-mileradius, representing an area roughly equivalent to Massachusetts in size, the initialcentral county shock is magnified by 30% in the long run. Within a 125-mileradius, representing an area equivalent to Ohio in size, the initial shock ismagnified by 70% in the long run.

The median state in the US has a land area equivalent to that of a circle with aradius between 125 and 150 miles. Interestingly, the 125- and 150-mile impulseresponse functions closely approximate the state-level impulse response functionfor overall employment depicted in Fig. 1. Contemporaneous and lagged correla-tions in shocks to export sector employment thus appear to be a viable explanationfor the patterns observed in Fig. 1.

Aggregation of counties into regions larger than 300 miles in diameter continuesto increase the relative magnitude of a shock’s after effects. In a region with aradius of 225 miles, representing an area roughly equivalent to California in size,the long-run impact of an export-sector shock is roughly twice the initialmagnitude. Further enlargements of the region have little impact on impulseresponse functions–employment dynamics in a region with a 300-mile radius look

25quite similar to those in a region with a radius of 225 miles.A general inspection of the impulse response functions reveals some points of

consistency. Employment levels in the export sector settle into a new long-run

23This normalization implies that impulse response functions for larger regions have been magnifiedconsiderably. The vector autoregression implies, for example, that a 1% shock to a central county ofaverage size is associated with a 0.05% shock in a region with a 300-mile radius. The 300-mile radiusimpulse response is therefore multiplied by a factor of 20.

24For each region, the initial shock consists of the 1% impulse to the central county plus anycontemporaneous effects indicated by thec coefficients in Table 2.

25Using the average region sizes listed in footnote 22, a 328 job (1%) shock to employment in acentral county predicts a contemporaneous loss of 963 other export-sector jobs in the surrounding300-mile region. In the long run, the central county export sector losses are magnified to 380 jobs; theoverall loss for the 300-mile region is roughly 2500 jobs. It is important to note that the hypothesizedrelationship between central county and regional employment estimated here is purely correlational; inno sense do central county employment changes ‘cause’ changes in the surrounding area.


steady state 2–4 years after the initial shock. Adjustment time increases with thearea of the region considered. In larger regions, there is evidence of ‘overshooting’the new long-run employment level, a pattern also visible in Blanchard and Katz

26(1992). Overall, Table 3 and Fig. 2 provide significant evidence of contempora-neous and lagged correlation in shocks to export-sector employment acrosscounties, and show that this pattern can explain the difference between county andstate impulse responses observed in Fig. 1. The most likely mechanism underlyingthese correlations, which apply to relative rather than absolute shocks, is similarityin the composite export good across neighboring areas.

4 .3. Evaluating hypothesis 2: direct correlations in shocks to export industry

The direct correlation hypothesis implies that an export-sector shock of givenabsolute magnitude should have consistent implications for absolute changes inemployment in neighboring counties. The hypothesis motivated Eq. (5), whichrelated absolute changes in employment in one county to current and past absolutechanges in surrounding areas. Table 4 presents selected results from a 13-equationvector autoregression of absolute export-sector employment changes in centralcounties and corresponding concentric rings on lagged employment changes in allareas, as well as current changes in the central county for ring areas. As in theprevious section, two lags of all employment variables are used. Employmentlevels in each geographic area are measured as a fraction of total nationalemployment, to eliminate broad macroeconomic trends.

The first row of Table 4 presents estimates ofc, which measures thecontemporaneous correlation in employment changes between areas. In contrastwith the previous table, there is essentially no evidence of significant correlation inabsolute employment shocks here, except in the innermost ring surrounding theinnermost ring. The loss of one export-sector job in the innermost ring predicts aloss of 0.04 export-sector jobs in the innermost ring. Coefficients in other rings aregenerally smaller than their standard errors in absolute value—four of these 11coefficients are negative.

The amplification of absolute employment shocks in central counties andsurrounding rings, as measured by the parametersr and r , follows a similar1 2

pattern in the year following a shock. In most areas, for every job lost in the initialyear, an additional 0.25–0.4 jobs disappear in the subsequent year. In all areas,losses are partially offset in the second year following a shock, although theestimated magnitude of this effect varies substantially between center and rings. Inmost rings, the second autocorrelation coefficient exceeds the first autocorrelation

26Blanchard and Katz (1992) present some evidence that might explain this overshooting. In theirdata, states that experience a negative employment shock simultaneously experience wage decreases—consistent with the hypothesis that labor demand shocks drive the process. Blanchard and Katz surmisethat wage declines prompt outmigration of workers.

682S.P.

Glendon,

J.L.Vigdor

/R

egionalScience

andU

rbanE

conomics

33 (2003) 663–693

Table 4The impact of absolute employment shocks within the export sector

Parameter Dep. Var.: Dependent variable:DE in 25-mile ring around employment center with outer radius equal to:t

DE in Emp. 25 50 75 100 125 150 175 200 225 250 275 300t


c – 0.044 20.008 0.028 0.053 0.004 0.044 0.073 0.052 0.072 20.036 20.084 20.005

(0.016) (0.018) (0.034) (0.057) (0.059) (0.083) (0.082) (0.077) (0.093) (0.107) (0.103) (0.105)

r 0.382 0.344 0.239 0.240 0.377 0.328 0.315 0.322 0.380 0.333 0.288 0.309 0.2851

(0.014) (0.016) (0.021) (0.021) (0.021) (0.024) (0.020) (0.023) (0.028) (0.024) (0.024) (0.025) (0.024)

r 20.089 20.508 20.247 20.375 20.403 20.379 20.508 20.439 20.333 20.425 20.441 20.405 20.4462

(0.014) (0.016) (0.022) (0.022) (0.022) (0.024) (0.021) (0.024) (0.029) (0.024) (0.025) (0.025) (0.025)

f – 0.005 20.003 20.001 0.067 0.017 0.017 20.004 0.008 0.089 0.136 0.051 0.0341

(0.016) (0.017) (0.033) (0.056) (0.058) (0.082) (0.081) (0.077) (0.092) (0.106) (0.102) (0.104)

f – 0.078 0.050 0.098 0.153 0.083 0.098 0.108 0.113 0.179 0.149 0.067 0.0472

(0.015) (0.017) (0.033) (0.055) (0.057) (0.081) (0.080) (0.076) (0.091) (0.104) (0.101) (0.102)2R 0.147 0.288 0.191 0.205 0.233 0.230 0.258 0.243 0.232 0.230 0.236 0.233 0.236

N 5100 5100 5100 5100 5100 5100 5100 5100 5100 5100 5100 5100 5100

Note: Standard errors in parentheses. Parameter estimates listed in boldface are significant at the 5% level.


coefficient in absolute value, while being of opposite sign. The long run impact ofan isolated export-sector shock to any particular ring would therefore beapproximately equal to the initial impact.

As was the case in Table 3, there is some evidence of lagged responses tocentral county employment shocks in concentric ring areas, as measured by theparametersf andf . Unlike the previous table, here the primary response to an1 2

employment innovation in the central county occurs after a lag of two years. Theestimates off , in Table 4’s fourth row, are uniformly insignificant and fall on1

either side of zero. Estimates off , by contrast, are uniformly positive, and2

statistically significant within 100 miles of the central county. Put together, theestimates suggest that when 10 jobs are lost in the export sector in a central countyin a given year, another four jobs will be lost in the central county in the followingyear, and another four jobs will be lost in surrounding counties to a distance of 100miles the year afterward. Point estimates suggest further losses in more distantrings with a 2-year lag, but these effects are estimated relatively imprecisely and

27are therefore statistically insignificant.The implications of the dynamics described in Table 4 are plotted in Fig. 3. To

construct this figure, a hypothetical central county, having initial employmentlevels equal to the national average for such counties, was presumed to lose 1000export-sector jobs. The implications of these losses were then traced out for thecentral county and 12 concentric rings, each of which is presumed to contain initialemployment equal to the national average for similar rings. The resulting predictedemployment levels were then summed into regions of varying size and trans-formed into logarithms. Fig. 3 plots the resulting impulse responses for theseregions, normalizing the magnitude of the initial shock to 1 in all cases.

Comparison of Figs. 2 and 3 reveals both similarities and differences. In bothgraphs, the ratio of long-term to initial employment change increases as the size ofthe area under consideration increases. The impulse response functions for regionshaving similar land area to the median state (radii of 125 or 150 miles) looksimilar to the state-level impulse response in Fig. 1. On the basis of these graphs, itappears that either the indirect correlation mechanism or the direct correlationmechanism could be justified on the basis of the data.

Some discrepancies between Figs. 2 and 3 bear mentioning here. The impulseresponse functions in Fig. 3 are relatively tightly spaced 1 year after the initialshock, reflecting the insignificance ofc and f parameters in the underlying1

27Table 4 omits a large number of VAR coefficients. The central county equation controls for twolagged employment changes in each ring. Each ring equation controls for two lagged employmentchanges in all other rings. Among other findings, current employment changes in the central county aresignificantly affected by lagged employment changes in other counties within 25 miles, and twice-lagged employment changes within 50 miles. The feedback effects implied by these coefficients areincorporated into the impulse response functions in Fig. 3. The full set of coefficients is available fromthe author upon request.


Fig. 3. Export sector impulse response functions 1000 job shock to export sector in central county.

vector autoregression. In the second year following the shock, impulse responsesdiverge dramatically in Fig. 3. This pattern contrasts with Fig. 2, where significantdivergence occurs in the year immediately following the shock. A second contrastconcerns the amplitude of impulse responses in the largest regions. In Fig. 2, theimpulse response functions converge when the radius of the area consideredexceeds 225 miles. In Fig. 3, the impulse responses for regions greater than 400miles in diameter expand considerably. It is important to note that many of theregression parameters producing the impulse response functions in Fig. 3 aremeasured imprecisely, suggesting that the differences among responses shown inthat graph may not be statistically significant.

To this point, this analysis suggests both that the relative size of employmentinnovations in one county influence relative innovations in neighboring counties,and that the absolute size of employment changes in one county influence absolutechanges in neighboring counties. Moreover, both estimated relationships appearstrong enough to explain the discrepancy between state and county impulseresponses on their own. To sort out these alternative explanations, and test thesensitivity of the analysis to the assumption that export-sector industries areadequately captured at the one-digit SIC level, the vector autoregressionsmotivated by Eqs. (3) and (5) were re-estimated, restricting the export sector toinclude only manufacturing industries.

28The results of this exercise differ from those reported in Table 3 and 4. Within

28The coefficient estimates derived from this estimation are available from the author upon request.


the manufacturing sector, there is no significant evidence of contemporaneous orlagged correlation in relative employment shocks across counties—point estimatesof c are uniformly negative and largely insignificant, point estimates off andf1 2

are less than 0.01 in absolute value in 11 and nine out of 12 equations,respectively. By contrast, absolute shocks to manufacturing employment in centralcounties have both immediate and lasting implications for absolute levels ofmanufacturing employment in surrounding counties. Point estimates suggest thatthe loss of 1000 manufacturing jobs in a central county is associated with thepermanent loss of an additional 1600 jobs in the surrounding region to a radius of300 miles.

These results suggest that the two mechanisms evaluated to this point apply todifferent segments of the overall export sector. Manufacturing industries inneighboring counties are connected primarily through upstream–downstreamlinkages—a shock to central county industry has few implications unless it is largein absolute terms. Resource extraction and agricultural industries follow the othermodel, where the primary link between neighboring counties is similarity of thecomposite export good.

4 .4. Evaluating hypothesis 3: spillovers in the local sector

Section 3.4 above presented a model in which local-sector employment in anyone area responds to the level of employment in the export sector within that areaplus overall employment in surrounding areas. Table 5 presents selected resultsfrom a vector autoregression designed to test this model in reduced form,regressing absolute changes in local sector employment on lagged changes, pluscontemporary and lagged employment changes in export sector employment forcentral counties and all concentric rings. This vector autoregression uses one lag ofall relevant independent variables.

The first row of Table 5 shows the cross-sectoral impact of contemporarychanges in export-sector employment in the central county of each region. The lossof ten jobs in the central county export sector predicts the immediate loss of

29between two and three additional jobs in the local sector in that same county. Inthe long run, as Fig. 4 illustrates, the long-run impact of a 10-job loss in the export

29In percentage terms, this relationship is substantially less than one-for-one, especially consideringthat employment in the local sector tends to exceed that in the export sector. One interpretation of theseresults is that the impact of a shock to the export sector on the local sector is dampened to the extentthat export producers in neighboring counties continue their preceding level of activity. In essence, thisdampening is a form of ‘risk-sharing’ for local-sector producers in different counties. Asdrubali et al.(1996) reach similar conclusions in their study of US states between 1963 and 1990. They show thatthere is a substantial degree of risk sharing between states, enough to smooth about 75% of the typicalshock to gross state product. They identify shared credit and capital markets, as well as federalgovernment action, as mechanisms by which risk sharing occurs.

686S.P.

Glendon,

J.L.Vigdor

/R

egionalScience

andU

rbanE

conomics

33 (2003) 663–693

Table 5The impact of export sector shocks on the local sector

Independent Dep. Var: Dependent Variable:DY in 25-mile ring around employment center with outer radius equal to:t

variable DY in Emp. 25 50 75 100 125 150 175 200 225 250 275 300t


DE 0.227 20.031 20.048 20.064 20.101 20.084 20.125 20.107 20.126 20.155 20.221 20.159 20.105t

(center co.) (0.007) (0.015) (0.012) (0.017) (0.037) (0.031) (0.047) (0.047) (0.044) (0.049) (0.051) (0.052) (0.055)

DE 20.005 20.019 0.039 0.030 0.062 0.070 0.103 0.080 0.040 0.083 0.153 0.197 0.150t21

(center co.) (0.008) (0.015) (0.012) (0.017) (0.037) (0.031) (0.046) (0.046) (0.043) (0.049) (0.051) (0.052) (0.054)

DY 0.479 0.447 0.559 0.587 0.584 0.602 0.589 0.599 0.594 0.594 0.638 0.618 0.619t21

(0.010) (0.010) (0.008) (0.007) (0.008) (0.007) (0.008) (0.007) (0.007) (0.007) (0.007) (0.007) (0.007)2R 0.472 0.392 0.482 0.540 0.572 0.595 0.597 0.602 0.616 0.616 0.658 0.645 0.649

N 5100 5100 5100 5100 5100 5100 5100 5100 5100 5100 5100 5100 5100

Note: Standard errors in parentheses. Parameter estimates listed in boldface are significant at the 5% level. Regressions are weighted by the averagelevel ofemployment in a region’s central county over the years 1969–1996.


Fig. 4. Local sector impulse response functions 1000 job shock to export sector in central county.

sector is between four and five lost jobs in the local sector. In the county of origin,there is significant evidence of cross-industry multiplier effects.

The remaining columns of Table 5 reveal that negative shocks to export-sectoremployment in central counties actually have a statistically significant, thougheconomically small,positive impact on local sector employment in surroundingareas, extending to a radius of 300 miles. Summed across rings, the point estimatessuggest that the job losses created by a negative export sector shock in the centralcounty are almost exactly offset by local sector employment increases in thesurrounding area.

Before discussing this surprising finding, it is important to note the remainingcoefficients in Table 5. Coefficient estimates in the second row indicate that insome concentric rings, the employment gains experienced in the year of a shock tothe central county are partially or completely offset by losses in the following year.Estimates in the third row suggest that employment changes in the local sectortend to be amplified significantly in years following the initial change, with aninitial gain or loss of one job predicting 0.5–0.6 additional jobs gained or lost,

30respectively, in the subsequent year. Any local sector gains following an exportsector loss are to a large extent temporary.

30Table 5 omits a large number of VAR coefficients. The central county equation controls for currentand once-lagged export-sector employment changes in each ring. Each ring equation controls forcurrent and once-lagged employment changes in all other rings. Among other findings, thesecoefficients indicate that local-sector employment changes in the center and each ring correlatesignificantly with contemporaneous changes in one or two adjacent rings. The full set of coefficients isavailable from the author upon request.


A second important caveat is that these coefficients report the response to acentral county export-sector employment shock, and neglect the fact that central-county shocks are frequently accompanied by export-sector shocks in surroundingrings. Taking these effects into consideration significantly reduces the magnitudeof the local sector effects implied in Table 5.

How ought we interpret these findings? At the outset, it is important to note thatinterpretation of the coefficients in this table must be undertaken with caution, aseach represents some combination of the parameters in the structural Eq. (6)above. With that in mind, it appears that the local sector in surrounding countiesplays a role in absorbing shocks to the center in the short run. Local sector firmsmay take advantage of lower wages immediately following an export-sector shock,while automatic stabilization policies such as unemployment insurance buoydemand for local goods and services. After a time period as short as 1 year, thisabsorptive capacity dissipates. In the long run, the effect of a one-time centralcounty export-sector shock on local sector employment is small, especially in the

31context of larger regions.Long-run local sector effects, though small in absolute size, are large relative to

initial effects. This contrast can be seen in Fig. 4, which plots local-sector impulseresponse functions to a one-time loss of 1000 jobs in the central county exportsector within regions of varying size. Initial responses are normalized to one for all

32regions shown. In larger regions, the long-run response to a shock of given sizeis larger relative to the instantaneous response. Thus, although the the evidencepresented in this section challenges the hypothesis that shocks to the export sectorsignificantly affect local sector employment in surrounding counties, the evidencethat exists does provide a partial explanation for the state-county discrepancy inFig. 1. Shocks may appear to have more severe long-run effect in largergeographic areas because short-term employment shifts across county boundariesdisguise the initial size of the shock.

5 . Implications and conclusions

Neighboring geographic regions tend to share economic fates. This paper has

31To check whether the patterns in Table 5 might be an artifact of the assignment of one-digitindustries to local and export sectors, we re-estimated the underlying vector autoregression using onlymanufacturing employment as the export sector and only retail trade employment as the local sector.The results of that VAR are qualitatively quite similar: the model predicts significant local sectorresponse to export sector shocks within the central county, but not in the concentric rings surroundingit. The coefficients from this exercise are available from the author upon request.

32This normalization prevents us from showing the impulse responses for regions with a radius of100 miles and above, since the initial local-sector response to a negative export sector shock is actuallypositive within regions of at least this size. Generally, though, the patterns in these larger regionsconforms to that displayed in the graph: as the size of the region increases, the long-run response looksworse relative to the short-run response.


sought to expose the mechanisms underlying these ties, focusing on threehypotheses. Counties may experience correlated employment fluctuations becausethey produce similar products for the national market, because firms in neigh-boring counties may directly cooperate in the production of some goods, orbecause employment shifts in one county change the demand for locally producedand consumed goods within a multi-county area.

In addition to seeking direct evidence to evaluate these hypotheses, we havetested whether any of them can explain the empirical regularity shown in Fig. 1:the after-effects of an employment shock increase in relative magnitude at higherlevels of geographic aggregation. Such a pattern would emerge if labor demandshocks propagate across county lines, but take time to do so. Thus we soughtevidence of serial correlation across two dimensions: space and time.

Our results provide significant evidence of cross-county correlation in shocks toexport sector employment, whether measured in relative or absolute terms.Relative export-sector employment fluctuations spread widely, affecting regions upto 600 miles in diameter. Our analysis suggests that the correlation of relativefluctuations is greatest in resource-oriented industries such as mining andagriculture. In manufacturing industries, employment innovations of large absolutemagnitude carry the greatest implications for surrounding areas. This distinctionsuggests, intuitively, that direct links in production involving intermediate inputsare most important in the manufacturing sector. The intercounty, intertemporaltransmission of both relative and absolute shocks to the export sector can explainthe discrepancy observed in Fig. 1.

We find less evidence to support the hypothesis that intersectoral ‘multipliereffects’ spill over across county boundaries. While local-sector firms respondnegatively to an adverse export-sector shock both in the short and long run, thereis evidence of a temporary positive local-sector response in surrounding areas, andlittle if any response in the long run. We note that in spite of the relativeunimportance of the local sector in propagating employment shocks, the evidencethat exists supports a third explanation for the pattern in Fig. 1: across counties,the local-sector response to an export sector shock becomes more negative overtime.

Previous authors have discussed the risks that any geographic region faces whena neighboring region falls on economic difficulty. Our analysis implies that a largeportion of this ‘risk’ is diversifiable2that is, it arises because of industrialsimilarity rather than direct dependence of employment in one region on continuedemployment in a neighboring area. A county that produces a composite exportgood that differs markedly from that of its neighbors will retain a large degree ofindependence in employment fluctuations.

Much remains to be understood in the analysis of geography and labor marketdynamics. The role of interlocality migration and wage differentials in mitigatingor propogating local demand shocks has not been considered here. It is reasonableto expect that counties of varying sizes have varying impacts on their surround-ings, and this paper has only focused on relatively large counties. The role of


government redistributive policy in propagating shocks across neighboring areasalso merits examination. Although this paper does not fully consider theseadditional lines of inquiry, it serves to underscore the importance of devotingfuture resources to them.

A cknowledgements

We are grateful to Gerald Carlino, David Cutler, Edward Glaeser, ClaudiaGoldin, Lawrence Katz, John Leahy, Andrei Shleifer, Elizabeth RichardsonVigdor, two anonymous referees, and participants in the Harvard Urban, Regionaland Transportation Economics Workshop for helpful comments on previous drafts.Glendon thanks Harvard’s Research Training Group in Positive Political Economyand the Alfred P. Sloan Foundation for financial support. Vigdor thanks theNational Science Foundation Graduate Fellowship program, the John M. OlinFoundation, and the Earle A. Chiles Foundation for financial support. Remainingerrors are the authors’ responsibility.

A ppendix A. Aggregation under the assumption of autarky

Consider the case of two counties. Label the level of employment in the firstcounty X , and in the second countyY . Suppose that employment innovations int t

the two counties obey the following processes:

X Xt t21]] ]]ln 5a 1r ln 1e (A.1)S D S DX XtX Xt21 t22

and

Y Yt t21]] ]]ln 5a 1r ln 1e (A.2)S D S DY YtY Yt21 t22

where the parameterr , 1 is known. Note that these processes are identical tothose identified in Eq. (1), with a simple version of the lag polynomial. Aregression pooling these two observations would provide an unbiased estimate ofthe parameterr so long as the standard assumptions regarding ordinary least

33squares regression are met.

33If counties are of unequal size, and therefore have heteroskedastic error terms, it may be desirableto weight observations to achieve efficient estimation. Even when unweighted, though, the estimate ofr

remains unbiased. Weighting observations may also be justified when counties of different sizes followdifferent data-generating processes, and the objective is to capture the process experienced by the‘mean’ worker.


Suppose further that the sum of employment in the two counties follows asimilar process:

X 1Y X 1 Yt t t21 t21]]]] ]]]]ln 5g 1b ln 1h (A.3)S D S D tX 1 Y X 1 Yt21 t21 t22 t22

How will the aggregated autocorrelation coefficientb compare to the disaggre-gated coefficientr?

It is useful to rewrite Eq. (A.3) as

X Y X Yt t t21 t21]] ]] ]] ]]ln a 1 (12 a) 5g 1b ln b 1 (12 b) 1h (A.4)S D S D tX Y X Yt21 t21 t22 t22

where a5X /(X 1Y ) and b5X /(X 1Y ). Substituting exponen-t21 t21 t21 t22 t22 t22

tiated versions of (A1) and (A2) into the left hand side of (A4), then exponentiat-ing both sides of the equation, we arrive at:

r rX Y Xt21 t21 t21(a 1e ) (a 1e ) (g1h )X Xt Y Yt t]] ]] ]]a e 1 (12 a) e 5e b 1 (1S D S D SX Y Xt22 t22 t22

bYt21]]2 b) (A.5)DYt22

This equation becomes much simpler if we omit error and fixed effect terms, butthe resulting expression leads to the same conclusion. It is not possible todetermine the relationship betweenb and r. Ignoring the fixed effects and error

rterms, in the case wherea5b, r must exceedb, since the functionx is convexby assumption. Under certain conditions whenb . ( , )a and X /X ,t21 t22

( . )Y /Y , it is possible forb to exceedr. It is therefore not possible tot21 t22

determine theoretically whether the estimates ofb shown in Table 1 are consistentwith the autarkic model given the estimates ofr shown.

Table A.1. Employment center counties

Calhoun, AL Nez Perce, ID Rapides, LA New York, NY Madison, TNHouston, AL Twin Falls, ID Middlesex, MA Onondaga, NY Shelby, TNJefferson, AL Adams, IL Emmet, MI Buncombe, NC Sullivan, TNMadison, AL Champaign, IL Grand Traverse, Cumberland, NC Angelina, TX

MIMobile, AL Cook, IL Kent, MI Guilford, NC Bowie, TXMontgomery, AL Marion, IL Wayne, MI Mecklenburg, NC Brazos, TXBoone, AR Peoria, IL Blue Earth, MN New Hanover, NC Dallas, TXCraighead, AR Sangamon, IL Hennepin, MN Pitt, NC Erath, TXPulaski, AR Allen, IN Lyon, MN Wake, NC Gray, TXSebastian, AR Marion, IN Olmstead, MN Bowman, ND Harris, TXWashington, AR St. Joseph, IN Stearns, MN Burleigh, ND Jefferson, TXFresno, CA Tippecanoe, IN Adams, MS Cass, ND Kerr, TXSacramento, CA Vanderburgh, IN Forrest, MS Ramsey, ND Lamar, TXSanta Clara, CA Vigo, IN Harrison, MS Allen, OH McLennan, TXAlamosa, CO Cerro Gordo, IA Hinds, MS Cuyahoga, OH Midland, TX


Denver, CO Dubuque, IA Lauderdale, MS Franklin, OH Nueces, TXEl Paso, CO Linn, IA Lee, MS Hamilton, OH Potter, TXMontrose, CO Polk, IA Lowndes, MS Lucas, OH Sutton, TXOtero, CO Scott, IA Washington, MS Carter, OK Taylor, TXHartford, CT Wapello, IA Adair, MO Custer, OK Travis, TXSussex, DE Woodbury, IA Boone, MO Jackson, OK Victoria, TXDist. of Columbia Finney, KS Cape Girardeau, Oklahoma, OK Wichita, TX

MOAlachua, FL Ford, KS Greene, MO Tulsa, OK Salt Lake, UTDuval, FL Riley, KS Jackson, MO Multnomah, OR Chittenden, VTEscambia, FL Saline, KS Jasper, MO Wasco, OR Norfolk City, VAHillsborough, FL Sedgwick, KS Livingston, MO Allegheny, PA Richmond City,

VALeon, FL Seward, KS St. Louis City, Lycoming, PA Roanoke City, VA

MOOrange, FL Shawnee, KS Silver Bow, MT Philadelphia, PA Benton, WASt. Lucie, FL Thomas, KS Brown, NE Charleston, SC Chelan, WABibb, GA Fayette, KY Douglas, NE Florence, SC King, WAChatham, GA Jefferson, KY Hall, NE Greenville, SC Whatcom, WADougherty, GA McCracken, KY Madison, NE Richland, SC Harrison, WVFulton, GA Pike, KY Red Willow, NE Codington, SD Kanawha, WVGlynn, GA Pulaski, KY Scotts Bluff, NE Hughes, SD Wood, WVLowndes, GA Warren, KY Washoe, NV Minnehaha, SD Ashland, WIMuscogee, GA Caddo, LA Bernalillo, NM Pennington, SD Brown, WIRichmond, GA E. Baton Rouge, Curry, NM Walworth, SD Eau Claire, WI

LAWare, GA Lafayette, LA Santa Fe, NM Yankton, SD La Crosse, WIAda, ID Natchitoches, LA Albany, NY Davidson, TN Marinette, WIBonneville, ID Orleans, LA Broome, NY Hamilton, TN Milwaukee, WIIdaho, ID Ouachita, LA Erie, NY Knox, TN Oneida, WI

R eferences

Asdrubali, P., Sorenson, B., Yosha, O., 1996. Channels of interstate risk sharing: United States1963–1990. Quarterly Journal of Economics 111, 1081–1110.

Bartik, T.J., 1991. Who Benefits from State and Local Economic Development Policies. W.E. UpjohnInstitute for Employment Research, Kalamazoo, MI.

Bound, J., Holzer, H., 1996. Demand Shifts, Population Adjustments, and Labor Market OutcomesDuring the 1980s, NBER Working Paper[5685.

Blanchard, O.J., Katz, L.F., 1992. Regional evolutions. Brookings Papers on Economic Activity 1,1–73.

Campbell, J.Y., Mankiw, N.G., 1987. Are output fluctuations transitory? Quarterly Journal ofEconomics 102, 857–880.

Clark, T.E., 1998. Employment fluctuations in US regions and industries: the roles of national,region-specific, and industry-specific shocks. Journal of Labor Economics 16, 202–229.

Ghosh, A.R., Wolf, H.C., 1997. Geographical and Sector Shocks in the US Business Cycle, NBERWorking Paper[6180.

Glaeser, E.L., Kallal, H.D., Scheinkman, J.A., Shleifer, A., 1992. Growth in cities. Journal of PoliticalEconomy 100, 1126–1150.


Guthrie, S.J., 1995. Government Policy and Economic Behavior: The Case Study of National Defense.Doctoral Dissertation, Harvard University.

Hanson, G.H., 1998. Market Potential, Increasing Returns, and Geographic Concentration, NBERWorking Paper[6429.

Hildebrand, G.H., Mace, A., 1950. The employment multiplier in an expanding industrial market: LosAngeles County, 1940–1947. Review of Economics and Statistics 32, 241–249.

Holzer, H., 1989. Employment, unemployment and demand shifts in local labor markets. Review ofEconomics and Statistics 73, 25–32.

Krugman, P., 1991. Increasing returns and economic geography. Journal of Political Economy 99,483–499.

North, D.C., 1955. Location theory and regional economic growth. Journal of Political Economy 63,243–258.

Persson, T., Tabellini, G., 1996. Federal fiscal constitutions: risk sharing and moral hazard.Econometrica 64, 623–646.

Tiebout, C.M., 1956. Exports and regional economic growth. Journal of Political Economy 64,160–164.

thy neighbor’s jobs: geography and labor market dynamics

Documents