EVOLUTION OF THE GEOGRAPHIO OONCENTRATION OF INDUSTRY^

The Geographic Concentration of Industry: Does NaturalAdvantage Explain Agglomeration?

By GLENN ELLISON AND EDWARD L . GLAESER*

Scholars in many fields of economics havebecome very interested in Silicon Valley-styleagglomerations of individual industries (J.Vemon Henderson, 1988; Michael E. Porter,1990; Paul Kmgman, 1991). These agglom-erations are striking features of the economiclandscape and may provide insights into thenature of the increasing-retums technologiesand spillovers that are thought by many to bebehind endogenous growth and businesscycles.

In our previous work (Ellison and Glaeser,1997), we noted that agglomerations mayarise in two ways. In addition to explanationsbased on localized industry-specific spillovers,there is a simpler altemative: an industry willbe agglomerated if firms locate in areas thathave natural cost advantages. For example, thewine industry (the second most agglomeratedindustry in our study) is siu-ely affected by thesuitability of states' climates for growinggrapes. If firms' location decisions are highlysensitive to cost differences (as found byDennis Carlton [1983], Timothy J. Bartik[1985], and Henderson [1997], among oth-ers), then natural advantages may account fora substantial portion of observed geographicconcentration.

* Discussants: Yannis Ioannides, Tufts University;Thomas Nechyba, Stanford University; Sukkoo Kim,Washington University-St. Louis.

* Ellison: Department of Economics, MassachusettsInstitute of Technology, Cambridge, MA 02139, andNBER; Glaeser: Department of Economics, Harvard Uni-versity, Cambridge MA 02138, and NBER. We thank theNational Science Foundation for support through grantsSBR-9515076 and SES-9601764. Ellison was also sup-ported by a Sloan Research Fellowship. We thank YannisIoannides for valuable comments and David Hwang, SteveJens, Alan Sorensen, and Marcus Stanley for researchassistance.

In this paper we use the term ' 'nattu'al ad-vantage" fairly broadly. Some possible ex-amples would be that none of the more than100,000 shipbuilding workers in the U.S. in1987 worked in Colorado, Montana, or NorthDakota and that the highest concentration ofaluminum production (which uses electricityintensively) is in Washington (which has thelowest electricity prices). We will also speakof the concentration of the mbber and plasticfootwear industry in North Carolina, Florida,and Maine and its absence from Alaska andMichigan as possibly refiecting natural advan-tages in the labor market. The industry is anintensive user of unskilled labor and faces tre-mendous competition from imports; hence wewould expect to see it locate in low-wagestates.

The simplest way to find effects of naturaladvantages on industry locations is to regresseach industry's state-level employment onstates' resource endowments as in SukkooKim (1999). A problem with this approach,however, is that one can easily think of morepotential advantages than there are states in theUnited States and fit each industry's employ-ment distribution perfectly. We identify ef-fects of natural advantages without overfittingby imposing cross-industry restrictions re-quiring the sensitivity of location decisionsto the cost of a particular input to be related tothe intensity with which the industry uses theinput.

Using such an approach to estimate the ef-fects of natural advantage on the 1987 loca-tions of four-digit manufacturing industries,we find that industry locations are related toresource and labor-market natural advantages.Our primary goal is to see how much of thegeographic concentration of industries re-ported in Ellison and Glaeser (1997) can be

311

312 AEA PAPERS AND PROCEEDINGS MAY 1999

attributed to natural advantages. In our pre-ferred specification, we attribute about one-fifth of the concentration to observable naturaladvantages. Given that we are using only asmall number of variables that capture advan-tages very imperfectly, we would guess that atleast half of the concentration reported inEllison and Glaeser (1997) is due to naturaladvantages. Nonetheless, there remain a num-ber of highly geographically concentrated in-dustries in which interfirm spillovers seemimportant.

I. Measuring Geographic Concentration

Following Ellison and Glaeser (1997), as-sume that industry /' consists of K plants thatsequentially choose locations from the set ofstates S in order to maximize profits. Supposethat the profits received by plant k when lo-cated in state s, 7r,vt,, are given by

log 7r,vt, = log TTi,

where TT,, is the expected profitability of locat-ing in state s given observed state-specificcosts, the function g^ refiects the effect on theprofitability of state s of spillovers given thatplants \, ... , k — \ have previously chosenlocations Uy,..., i;^-,, TJ,, is an unobserved com-mon random component of the profitability oflocating in state s, and e,fcs is an unobservedfirm-specific shock.

In our model, plants in an industry may clus-ter relative to aggregate activity for three rea-sons: (i) more plants will locate in states withobserved cost advantages; (ii) more plants willlocate in states with unobserved cost advan-tages; and (iii) plants will cluster if spilloversare geographically localized.

Equilibrium geographic concentration iseasy to compute for a particular specificationin which the importance of unobserved naturaladvantages and spillovers are captured by pa-rameters y"" and y ' € [0, 1]. The importanceof unobserved natural advantages is reflectedin the variance of 77,5. Specifically, we assumethat 7?,, is such that 2(1 - y'"')-Ki,e''"ly''^ hasa x^ distribution with £(7r,^e''") = 7r,, andVar(7r,,e''") = ^ " ^ ^ / ( l - -y""). Spillovers are

of an all-or-nothing variety: with probabilityy^ a "crucial spillover" exists between eachpair of plants. If such a spillover exists be-tween plant k and plant €, then plant k receivesnegative infinity profits if it does not locate inthe same state as plant €; otherwise its profitsare independent of €'s location.

When the plants choose locations to maxi-mize profits in this model, the expected shareof employment located in state s is £(5,^) =^is = TTij/S^TT/̂ . An index of geographicconcentration beyond that accounted for byobserved natural advantage is

where H is the Herfindahl index of the plants'shares of industry employment. The propertythat makes this an appealing index (which fol-lows from the same argument as in Ellison andGlaeser [1997]) is that £ ( 7 ) = y"" + 7 ' -

The concentration index will reflect bothunobserved natural advantages and localizedspillovers. The index in Ellison and Glaeser(1997) is simply this index with the crudemodel where 5,̂ is assumed to equal state5's share of overall manufacturing employ-ment. As we better account for the effects ofnatural advantages on state-industry shares,we would expect the index of concentration toget smaller.

n. Does Natural Advantage AffectIndustry Location?

In our empirical work we assume that av-erage state-industry profits are

log TTfa = ao log (pop,,) + a, log (mfg,)

where pop^ and mfg^ are the shares of totalU.S. population and manufacturing employ-ment in state s, i indexes inputs to the pro-duction process, ye^ is the cost of input € instate s, and za is the intensity with which in-dustry i uses input (. The specification in-cludes multiplicative industry dummies, (5,, to

VOL 89 NO. 2 GEOGRAPHIC CONCENTRATION OE INDUSTRY 313

account for the fact that observed cost differ-ences will affect location decisions more insome industries than in others, both because ofdifferences in the magnitude of the plant-specific shocks and because of transportationcosts. For example, while one can easily imag-ine the fur industry concentrating in responseto moderate cost differences, it seems hard toimagine that the concrete industry would con-centrate geographically even if there wereenormous differences in the costs of rocks andcoal. In the estimation, the multiplicative in-dustry effects are constrained to be normega-tive. Note that we have economized on thenumber of parameters by assuming that the ef-fect on industry profitability of the differencein the cost of a particular input is proportionalto the intensity with which the industry usesthe input, rather than estimating a separate co-efficient for each input for each industry.

Given this specification ofthe effects of nat-ural advantages, expected state-industry em-ployment shares are

r exp(- (2 pop?» mfg?' e x p ( - 6; 2 /3es' e

a)

We estimate this relationship for the 1987state employment shares of four-digit man-ufacturing industries by nonlinear leastsquares using 16 interactions designed to re-flect advantages in natural resource, labor,and transportation costs. We normalize all ofthe interaction variables to have a standarddeviation of 1 and choose their signs so thatthe /8 coefficients are expected to be positive.A larger estimated /? value indicates thatvariation in the cost/intensity of use of thatinput has a larger effect in aggregate on thedistribution of industries.

Results from estimating a base model with-out the multiplicative industry dummies arepresented in Table IA. The first column givesthe name of the variable. Each is described intwo lines: the first being the state-level inputprice variable (usually a proxy rather than aprice) and the second being the industry-levelvariable reflecting intensity of use or the sen-sitivity of location decisions to input costs.

The second column gives the coefficient,which should be interpreted as the percentageincrease in the state's share of total industryemployment caused by a one-standard-deviation increase in the explanatory variable;f statistics for the coefficient estimates aregiven in parentheses. Thus the coefficient of0.170 for the variable electricity price X elec-tricity usage means that the state's share ofindustry employment increases by about 17percent (e.g., from 10 percent to 11.7 percent)with a one-standard-deviation increase in thisvariable. Table IB shows the two industriesfor which the industry component of each vari-able is largest. For example, electricity is mostimportant (as a fraction of value added) forprimary aluminum and alkalies and chlorine.Table IB also shows the states where the inputcost is lowest (Washington, Idaho, and Mon-tana for electricity) and the state where it ishighest (Rhode Island).

The first six variables in the table (a-f ) aredesigned to reflect the costs of six commoninputs: electricity, natural gas, coal, agricul-tural products, livestock products, and lumber.All are highly significant and of the expectedsign. The coefficients on several of these vari-ables are among the largest we find, indicatingthat these variables reflect a substantial com-ponent of natural advantage.

The next six variables (g-1) relate to laborinputs. The first three are the average manu-facturing wage in the state interacted with (g)wages as a share of value added, (h) the frac-tion of industry output that is exported, and (i)the fraction of U.S. consumption of the outputgood that is imported. Interactions (h) and (i)examine whether industries that are morecompetitive intemationally are more wage-sensitive. All of these variables, however, willnot matter if average wage differences are at-tributable to differences in labor productivity.Interactions (g) and (i) are significant andhave the expected positive sign, but the coef-ficients are fairly small. We find no evidenceof exporting industries concentrating in low-wage states.

The other labor input variables are designedto capture differences in the relative prices ofdifferent types of labor. Variable (j) is the in-teraction of the share of the adult populationin the state without a high-school degree with

314 AEA PAPERS AND PROCEEDINGS MAY 1999

TABLE 1—EFFECT OF "NATURAL ADVANTAGES"ON STATE-INDUSTRY EMPLOYMENT

A. CoefficientState variable X industry variable (f statistic)

(a) Electricity price X 0.170electricity use (17.62)

(b) Natural gas price X 0.117natural gas use (6.91)

(c) Coal price X 0.119cojil use (4.55)

(d) Percentage faimland X 0.026agricultural inputs (2.58)

(e) Per capita cattle X 0.053livestock inputs (5.08)

(0 Percentage timberland X 0.152lumber inputs (11.98)

(g) Average mfg wage X 0.059wages/value added (4.11)

(h) Average mfg wage x -0.014exports/output (-1.28)

(i) Average mfg wage X 0.036import competition (3.10)

(j) Percentage without HS degree X 0.157percentage unskilled (7.38)

(k) Unionization percentage X 0.100percentage precision products (12.17)

(1) Percentage with B.A. or more X 0.170percentage executive/professional (12.70)

(m) Coast dummy X -0.031heavy exports (-2.20)

(n) Coast dummy X 0.017heavy imports (0.92)

(o) Population density X 0.043percentage to consumers (3.68)

(p) (Income share - mfg share) X 0.025percentage to consumers (4.49)

B. Industries where most Best statesVariables" important (SIC) [worst state]

(a) Primary aluminum (3334) WA, ID, MTAlkalies and chlorine (2812) [RI]

(b) Brick and clay tile (3251) AK, LA, TXFertilizer (2873-4) [HI]

(c) Cement (3241) MT, NV, WYLime (3274) [VT]

(d) Soybean oil (2075) NE, ND, SDVegetable oil (2076) [DC]

TABLE 1—Continued.

Variables'Industries where most

important (SIC)Best states

[worst state]

(e) Milk (2026) SD, NE, MTCheese (2022) [MD]

(0 Sawmills (2421) AK, MT, IDWood preserving (2491) [DC]

(g) Industrial patterns (3543) MS, NC, ARAuto stampings (3465) [MI]

(h) Oil and gas machinery (3533) MS, NC, ARRice milling (2044) [MI]

(i) Dolls (3942) MS, NC, ARTableware (3263) [MI]

(j) Apparel (23) MS, KY, WVTextiles (22) [AK]

(k) Machine tools (354) MI, NY, HIJewelry (391) [SD]

(1) Computers (357) DC, MA, CT(Periodicals) (2721) [WV]

(m) Rice milling (2044)Industrial gases (2813)

(n) Nonferrous metals (3339)Petroleum refining (2911)

(o) Potato chips (2036) DC, NJ, RIJewelry (3411) [AK]

(p) Potato chips (2036) FL, CA, NYJewehy(3411) [NC]

'Letters in this column refer to state and industry vari-ables in part A of the table.

the share of workers in the industry who areunskilled. Next, (k) is the interaction ofunionization in the state (as a proxy for thepresence of skilled workers) widi the fractionof employees in the industry who are precisionproduction workers. Variable (1) is the inter-action of the fraction of the adult populationin the state with bachelors' degrees or moreeducation with the fraction of industry workerswho are executives or professionals. All ofthese variables have a powerful positive effect.

The final four variables (m-p) relate totransportation costs. The first two (m and n)are designed to examine whether industriesthat are intensive importers or exporters ofheavy goods tend to locate on the coast. Nei-ther ofthe estimates is positive and significant.

VOL 89 NO. 2 GEOGRAPHIC CONCENTRATION OF INDUSTRY 315

The next two variables (o and p) are meant tocapture the idea that firms will reduce trans-portation costs or improve their marketing bylocating closer to their customers. They are in-teractions of the share of the industry's outputthat is sold to consumers with population den-sity and with the difference between a state'sshare of income and its share of manufacturingemployment. Both are significantly positivelyrelated to employment.

The coefficients on the natural advantagesin specifications that include multiplicativedummies for two-digit and three-digit indus-tries are similar. The tendency of labor-intensive industries to locate in low-wagestates appears more pronounced in these re-gressions, while estimates of the effects due tounskilled labor, import competition, and in-come share minus manufacturing share be-come insignificant or negative.

m . Does Natural Advantage ExplainAgglomeration?

Our greatest motivation for studying naturaladvantage is a desire to know whether it canaccount for a substantial portion of observedgeographic concentration. Table 2 illustratesthe effect on measured geographic concentra-tion of accounting for observed natural advan-tages. Each row reports on the distribution ofindustry agglomeration indexes y obtainedfrom a particular model of natural advantage.The first row describes the concentration indexof Ellison and Glaeser (1997), which corre-sponds to the trivial model £(5 , J = mfg^. Themean value of y in this model is 0.051. Onlya few industries have negative y's. (This isnoteworthy because the model has no trans-portation costs leading firms to spread outwhen serving local markets.) We regard the 28percent of industries with y > 0.05 as showingsubstantial agglomeration. For comparison,the y of the automobile industry (SIC 3711)is 0.127. Such extreme agglomeration is un-common but far from unique: 12.8 percent ofmanufacturing industries have a y greater than0.1.

The second row shows the results when weintroduce the 16 cost/intensity of use interac-tions but do not allow industries to differ inthe sensitivity of location decisions to ob-

TABLE 2—ESTIMATES OF RESIDUAL GEOGRAPHICCONCENTRATION AFTER ACCOUNTING FOR OBSERVED

NATURAL ADVANTAGES

Model

ABCD

y

0.0510.0480.0450.041

Percentage of industries with y in range

<0.0

2.83.93.14.4

0.00-0.02

39.939.942.942.9

0.02-0.05

29.230.129.429.8

0.05-0.10

15.313.713.513.3

>0.1

12.812.411.19.6

Notes: Models A - D are different models of natural advantage: (A)no cost variables; (B) cost interactions introduced; (C) cost inter-actions plus dummies for two-digit industries; (D) cost interactionsplus dummies for three-digit industries.

served cost differences. The mean y declinesslightly to 0.048, and the overall distributionlooks quite similar. The third and fourth rowsdescribe the concentration indexes foundwhen we allow for multiplicative dummies foreach two- and three-digit industry, respec-tively. In these models, natural advantageshave greater explanatory power, reducing themean values of y to 0.045 and 0.041, respec-tively. We conclude that 20 percent of mea-sured geographic concentration can beattributed to a few observable naturaladvantages.

The fraction of industries that are extremelyagglomerated in this measure declines, butonly moderately: 9.6 percent of industries stillhave y greater than 0.1 in the latter specifica-tion. Another notable feature of the distribu-tion of y is that the index is negative for onlya very few industries. The finding that virtuallyall industries are at least slightly agglomeratedis apparently fairly robust to the introductionof measures of cost advantages.

rV. Conclusion

Industries' locations are affected by a widerange of natural advantages. About 20 percentof observed geographic concentration can beexplained by a small set of advantages. Wethink that this result is particularly notablegiven the limits on our explanatory variables.For example, nothing in our model can explainwhy there is no shipbuilding in Colorado, norcan it predict that soybean-oil production isconcentrated in soybean-producing states, asopposed to being spread among all agricultural

316 AEA PAPERS AND PROCEEDINGS MAY 1999

States. We hope that, in the future, others willprovide better estimates than we have beenable to give here. We conjecture that at leasthalf of observed geographic concentration isdue to natural advantages.

At the same time, there remain a large num-ber of highly concentrated industries where itseems that agglomeration must be explainedby localized intraindustry spillovers. Simplecost differences can not explain why the furindustry, the most agglomerated industry inour sample, is centered in New York. We seethe attempt to provide a clearer understandingof the sources of these spillovers as an impor-tant topic for future research. Some resultsalong these lines are described in Guy Dumaisetal. (1997).

REFERENCES

Bartik, Timothy J. "Business Location Deci-sions in the United States: Estimates of theEffects of Unionization, Taxes, and OtherCharacteristics of States." Journal of Busi-ness and Economic Statistics, January 1985,J ( l ) , pp. 14-22.

Carlton, Dennis. "The Location and Employ-ment Decisions of New Firms: An Econo-metric Model with Discrete and Continuous

Exogenous Variables." Review of Econom-ics and Statistics, August 1983, 65(3), pp.440-49.

Dumais, Guy; Ellison, Glenn and Glaeser,Edward L. "Geographic Concentration asa Dynamic Process." National Bureau ofEconomic Research (Cambridge, MA)Working Paper No. 6270, November1997.

Ellison, Glenn and Glaeser, Edward L. "Geo-graphic Concentration in U.S. Manufactur-ing Industries: A Dartboard Approach."Journal of Political Economy, October1997, 705(5), pp. 889-927.

Henderson, J. Vernon. Urban development:Theory, fact, and illusion. New York: Ox-ford University Press, 1988.

"The Impact of Air Quality Regula-tion on Industrial Location." Annalesd'Economie et de Statistique, January-March 1997, (45), pp. 123-37.

Kim, Sukkoo. "Regions, Resources, and Eco-nomic Geography: Sources of U.S. RegionalComparative Advantage, 1880-1987 . "Regional Science and Urban Economics,January 1999, 29(1), pp. 1-32.

Knigman, Paul. Geography and trade. Cam-bridge, MA: MIT Press, 1991.

Porter, Michael E. The competitive advantageof nations. New York: Free Press, 1990.