Firm location in a polycentric city: the effects of taxes and agglomeration economies on location decisions

Transcription

1 Environment nd Plnning C: Government nd Policy 2007, volume 25, pges 671 ^ 691 DOI: /c0649 Firm loction in polycentric city: the effects of txes nd gglomertion economies on loction decisions Jnet E Kohlhse Deprtment of Economics, University of Houston, Houston, TX , USA; e-mil: jkohlhse@uh.edu Xihong Ju Houston ^ Glveston Are Council, 3555 Timmons Lne, Houston, TX 77227, USA Received 6 April 2006; in revised form 31 July 2006 Abstrct. The uthors explore the determinnts of firm loction in polycentric city with the id of dt for the Houston region. Firm loction is modeled in discrete-choice frmework, with eight employment centers nd outlying res used s possible choices. Agglomertive nd dispersive forces re explicitly treted, s re txes nd other chrcteristics tht vry over spce. The findings suggest tht property txes hve lrge deterrent effects on firm loctions for the four industril groups nlyzed here: oil nd gs; mnufcturing; finnce, insurnce nd rel estte (FIRE); nd services. When gglomertion economies re present, they re weker thn the tx effects nd re positive for only the FIRE nd services industril groups. 1 Introduction Tody's cities present n interesting lndscpe tht differs drmticlly from the lyout presented by trditionl monocentric city models. Monocentric models ssume tht employment is concentrted in the centrl business district (CBD), with decentrlized employment found t low densities t other loctions. However, for mny lrge urbn res, like the Greter Houston re, the sptil lndscpe of firms often shows severl other loctions or employment subcenters (in ddition to the CBD) where reltively high employment densities occur. Such cities re termed `polycentric' nd offer chllenge to reserchers not only to understnd how, why, nd where these subcenters pper, but lso to chrcterize ny potentil ttrctive forces generted by the subcenters for firm loction. Recent reserch on employment subcenters in urbn res hs focused on the definition nd identifiction of employment subcenters (Crig nd Ng, 2001; McMillen, 2001), the effects of subcenters on lnd vlues or rel estte vlues, nd the effects of subcenters on the sptil distribution of employment densities nd popultion densities in urbn res (Crig nd Kohlhse, 2003; McMillen nd McDonld, 1998; Smll nd Song, 1994). In contrst, in the present study we model intrurbn firm loction in discrete choice frmework. In prticulr, we empiriclly test the determinnts of firm loction mong eight employment centers in Hrris County, Texsöthe most populted county in the Houston metropolitn re. Houston is n interesting city in which to study the loction decisions of firms for mny resons. Perhps most importntly, mrket forces dominte the decision-mking process for firm loction. The decisions re mde in n tmosphere of miniml government regultion on lnd use: Houston is the only mjor US city without centrlized zoning (Seign, 1972). Fiscl policies do, however, impct firms in tht txes nd services cn vry over the mny jurisdictions within the re. The presence of severl concentrtions of employment my be indictive of gglomertive forces operting t the microgeogrphic level. And relting to trnsporttion, Houston lies on flt, homogeneous plne, hs well-developed highwy network for trnsporting goods nd people, nd hs well-defined export node t Port of Houston (the Houston Ship Chnnel).

2 672 J E Kohlhse, X Ju Our study focuses on two importnt questions bout firm loction within polycentric city: whether or not there is empiricl evidence of the presence of gglomertive forces in nd ner to the subcenters; nd whether or not fiscl vribles re importnt determinnts of firm loction. To exmine the role of fiscl vribles nd potentil gglomertive forces, the loction preferences of firms locting within subcenters s well s firms locting outside subcenters re exmined. Both logit nd mixed-multinomil logit models re estimted to provide empiricl evidence for the rguments. Previous empiricl reserch on gglomertion economies hs been conducted using vriety of pproches for vrious levels of sptil ggregtion nd industril ggregtion (for excellent survey rticles see Durnton nd Pug, 2004; Quigley, 1998; Rosenthl nd Strnge, 2004). Severl studies hve been conducted t the intermetropolitn-re level by exmining ggregte urbn re production functions. For exmple, Moomw (1986; 1998), Henderson (1986), nd Segl (1976) present evidence tht productivity differences mong urbn res exist nd re function of the size of the urbn re. Agglomertion vribles hve been specified either s the totl popultion of the city, or s the totl employment in the city. Rich detil t the locl geogrphic level is lost in such ggregte pproches. Other recent reserch hs focused on smller sptil units, such s firms ggregted to the zipcode level (Rosenthl nd Strnge, 2003, 2005; Shukl nd Wddell, 1991) or t the censustrct level (Rosenthl nd Strnge, 2005). A few studies hve been ble to exploit detiled individul plnt dt in the UK (Durnton nd Overmn, 2002) nd the US (Henderson, 2003). Our study looks t loction decisions mde by individul firms within single urbn re. The present study is one of few to use discrete choice model which ttempts to explicitly ccount for gglomertion economies nd fiscl vribles. Extensive use is mde of geogrphic informtion systems (GIS) softwre to code ddresses nd to crete sptilly disggregted vribles. Another importnt feture is comprtive nlysis of four mjor industril groups, rther thn the trditionl focus on only the mnufcturing sector. The pper is orgnized s follows. In prt 2 we present two lterntive qulittive choice models in which to model firm's loction decision. In prt 3 we discuss the dt nd the unique gglomertion proxy vribles creted for the study. Empiricl results from estimting logit nd mixed-multinomil logit model re presented in prt 4, nd results compred by the four industril groups: oil nd gs; mnufcturing; finnce, insurnce nd rel estte (FIRE); nd the services sector. Conclusions nd policy implictions re offered in prt 5. 2 Discrete choice pproch to firm loction decisions We ssume firms hve lredy decided to locte in given metropolitn re nd re fced with the current decision of where to locte within the region. To model the firm's loction choice, qulittive choice frmework is developed in which the firm is ble to choose mong number of discrete sites on the bsis of expected future profit levels t the lterntive sites. Following the development in Greene (2003, chpter 21) nd Wooldridge (2002, chpter 15) we cn specify model of the probbility of choosing mongst set of L sites. Let firm m's expected profit eqution t site i be: V mi ˆ f S ki, F mj e mi, k ˆ 1, :::, K, i ˆ 1, :::,L, m ˆ 1, :::,M, j ˆ 1, :::,J, (1) where V mi is the present discounted vlue of firm m's expected future profits t site i over its lifetime. The vector S represents site-specific vribles, indexed by k, which impct firm m's expected present vlue t loction i. It is through the site-specific

3 Firm loction in polycentric city 673 vribles tht potentil gglomertion economies operte, s do the fiscl vribles. The vector F represents firm-specific vribles, indexed by j, which influence firm m's present vlue expecttion t loction i. The function f(s ki, F mj ) mkes up the deterministic portion of firm m's present vlue expecttion t site i. The error term, e mi, is tht prt of firms m's present vlue expecttion t site i not explined by the function. The firm is ssumed to hve unbised expecttions of the profitbility of ech site: the firm simply chooses the site with the highest expected present vlue. The probbilities generted in the discrete choice model rise becuse the resercher does not hve ll the informtion the firm uses to ssess the profitbility of lterntive sites. Eqution (1) represents the expected present vlue tht firm m will generte if it loctes t site i. However, the only V mi observed for ech locting firm is the vlue of the relized choice. Firm m will follow the decision rule: N mi ˆ 1, if V mi > V mt, t 2 T m, 8 t 6ˆ i, (2) 0, otherwise, where N mi is n index of L site choices open to firm m nd T m is the set of L lterntives, of which t is subset. The firm is ssumed to know the prt of the profit eqution (1) determined by the resercher, f(.), s well s the error term. In contrst, to the resercher, the error term in eqution (1) represents the missing firm or site chrcteristics. Substituting eqution (1) into the decision rule, eqution (2), gives: N mi ˆ 1, iff (S ki, F mj e mi > f S kt, F mj e mt, t 2 T m 8 t 6ˆ i, (3) which cn be rewritten s N mi ˆ 1, iff S ki, F mj f S kt, F mj > e mt e mi, t 2 T m 8 t 6ˆ i, (3b) 0, otherwise, (3c) Assuming the difference (e mt e mi ) follows probbility distribution, the probbility tht firm m will choose site i (N mi ˆ 1) for ech lterntive i cn be estimted using the logit function. We express the function s being composed of chrcteristics of the choices (sites) s well s the chooser (firms)ö vrint often termed the `mixed-multinomil logit' (see the discussion in Long nd Freese, 2003, section 6.7.5): P N mi ˆ 1jS ki, F mj ˆ exp f S ki, F mj Š, (4) X L i ˆ 1 exp f S ki, F mj Š where P(N mi ˆ 1) is the probbility tht firm m will choose site i. A relted discrete choice model is the simple logit, where we model the choice between `concentrted' or dispersed loction. In such specifiction, the choice of loction cn be viewed s zero ^ one choice, where N ˆ 1 if firm m loctes inside ny one of the eight employment centers, nd N ˆ 0 if firm m loctes in the rest of Hrris County. Then the probbility tht firm m loctes in concentrted loction becomes: P N m ˆ 1jz m ˆ exp f z m Š 1 exp f z m Š, (5) where z is vector of vribles ssocited with ech firm m.

4 674 J E Kohlhse, X Ju 3 Dtset nd creted vribles The focus of our study is the Houston metropolitn reöin prticulr, firm loctions within its centrl county, Hrris. (1) Figure 1 shows cumultive popultion nd employment distributions for the Houston MSA (Metropolitn Sttisticl Are) in 1990 s function of distnce from the CBD. The figure shows tht employment is more centrlized thn popultionömore thn 50% of employment lies within 10 miles of the CBD wheres only bout 30% of the popultion does. This reltive sptil distribution is intriguing in itself nd motivtes one of our reserch questions: wht re the forces tht bind firms together, even in the presence of decentrlizing popultion? To explore the issues, we nlyze firm loction in the context of set of discrete concentrtions of employment, tht is, the CBD nd other employment `subcenters'. We use the seven employment subcenters in the Houston region, identified by the method of quntile smoothing splines described in pper by Crig nd Ng (2001): Bytown, Psden, LPorte, Cler Lke, the Glleri, Crrilon, nd Greenspoint. (2) Figure 2 shows the eight employment centers defined by their respective 1990 census trct. (Herefter we use the term `employment centers' s n inclusive term to indicte the CBD nd the other employment subcenters.) Cumultive employment nd popultion shre Distnce from CBD (miles) Figure 1. Houston metropolitn sptil distributions, Popultion Employment Aside from the CBD, Houston's strongest employment subcenter is the Glleri re, clled `Uptown' by some rel estte professionls. This is retil nd office re bordering on the innermost circumferentil highwy (I-610) nd the min southwest freewy (US 59). Cler Lke is the re south nd est of the CBD tht contins NASA. Crillon is n re bout 5 miles west of the Glleri. Greenspoint is ner (1) The lnd re of Hrris County is bout 1730 squre miles. About 30% of the county's lnd re is covered by the 540 squre mile city of Houston. (2) Even though the Houston MSA is mde up of five counties, the CBD nd ll employment subcenters identified by Crig nd Ng (2001) lie within its centrl county, Hrris County. To determine the loction of the employment subcenters, Crig nd Ng use nonprmetric specifiction (quntile smoothing splines) to evlute the upper til of employment densities. Their seven subcenters lie on three concentric rings (defined t 6, 13, nd 21 miles) round the CBD. In their method, res with employment densities t the 95th percentile or bove, conditionl on the distnce from the CBD, tht lso pper to influence neighboring res, re identified s employment subcenters. The geogrphy of the eight employment centers re smll res, defined by the boundry of n pproprite 1990 census trct.

5 Firm loction in polycentric city 675 Employment centers 1990 census trct boundry Figure 2. Houston re employment centers nd 1990 census trct boundries. the mjor irport, on n rteril freewy north of the CBD. Bytown, Psden, nd LPorte re ll industril res in the neighborhood of the Houston Ship Chnnel. Individul estblishment-level dt used for this study were tken from commissioned Dun & Brdstreet MrketPlce file for Hrris County, Texs The full dtset contins 72 vribles nd observtions. The dt re t the estblishment level or plnt level, but we choose to use the term `firm' for ese of exposition in the rest of the pper (in our dt, most of the estblishments re stnd-lone firms). Firm-specific chrcteristics in the dtset include nme nd ddress of firm, number of employees, nnul sles, Duns number, 1990 census trct number, four-digit SIC codes, nd yer estblished. Sixteen industries t the two-digit SIC level were selected for nlysis nd then brodly grouped into four industril groups, s reported in tble 1. The groupings were constructed to cpture industries tht re likely to produce complementry products or services. The underlying ssumption is tht dvntges of gglomertion could be detected more esily in these brodly interrelted sectors. For exmple, industries with complementry products my tend to locte ner ech other in order to sve on trnsporttion costs of inputs nd outputs, or to enhnce informtion trnsmission Tble 1. Mjor industril groups by stndrd industry clssifiction (SIC) (source: Stndrd Industril Clssifiction Mnul, 1987). Group Nme SIC code 1 Oil nd gs 13, 28, 29, 30 2 Mnufcturing 33, 35, 37, 38 3 FIRE 60, 63, 65, 67 4 Services 73, 75, 80, 87 Finnce, insurnce, nd rel estte.

6 676 J E Kohlhse, X Ju through frequent fce-to-fce contct. Potentil sources of the externl economies my include lbor-mrket pooling, input shring, knowledge spillovers, vilbility of consumption externlities, nd others (see the survey by Durnton nd Pug, 2004). The first brod sector contins oil-relted nd gs-relted industries. They re oil nd gs extrction (SIC 13), chemicl nd llied products (SIC 28); petroleum refining (SIC 29); nd rubber nd miscellneous plstics products (SIC 30). The second group, mnufcturing, includes: primry metl industries (SIC 33); commercil mchinery nd computer equipment (SIC 35); trnsporttion equipment (SIC 37); nd mesuring, nlyzing, nd controlling instruments (SIC 38). The third groupöfireöcontins: depository institutions (SIC 60); insurnce crriers (SIC 63); rel estte (SIC 65); nd holding nd other investment compnies (SIC 67). The fourth group ws designed for service industries; it contins: business services (SIC 73); utomotive repir, services, nd prking (SIC 75); helth services (SIC 80); nd engineering, ccounting, reserch, mngement, nd relted services (SIC 87). In order to crete sptilly detiled vribles, the ddresses of the individul firms were geocoded, but not ll firms' ddresses were ble to be geocoded, for myrid of resons. The overll mtching rte ws bout 67%, rnging from 56% (SIC 63, insurnce crriers) to 77% (SIC 37, trnsporttion equipment), bringing the originl smple size of down to for the sixteen industries. The smple size ws further reduced to fter firms with zero employees were deleted from the smple. (3) For the empiricl work, two smples of firms were exmined in two different discrete choice frmeworks. First, the entire smple of bout firms (herefter referred to s the `full' smple) ws used to estimte the choice between concentrted loction (ny of the eight employment centers) or dispersed loction (nywhere else in the rest of Hrris County, Texs). Second, smller smple of firms, bout 4600, ws used to investigte further those firms choosing concentrted loctions in one of Houston's eight employment centersöherefter these re referred to s the `centers' smple. Severl vribles, both dependent nd independent, were constructed for the dt nlysis. Bsed on the geocoded ddress of ech firm, firms were ssigned prticulr loction ctegory: one of the eight employment centers or the `rest of Hrris County'. Any firm lying within the census trct contining n employment center ws ssigned tht center s loction. Two types of discrete dependent vribles were creted. The first dependent vrible ws binry vrible, where zero is ssigned to firms locting in the rest of Hrris County nd one ws ssigned to firms locting in ny one of the eight employment centers. The second dependent vrible ssigned firms loction ctegory numbered from 1 to 8, with the CBD ssigned loction 1. We then used the CBD s the bse (omitted) ctegory for estimting the mixed-multinomil logit on the centers smple. Tble 2 shows the distribution of the number of firms by loction ctegory. For the mixed-multinomil logit, two types of independent vribles were constructed: those tht re chrcteristic of the eight choices, nd those tht re chrcteristic of the individul firm. Tble 3 reports mens nd stndrd devitions for the two dtsets used in the estimtionöthe full dtset nd the centers dtset. Seven choice-specific independent vribles were constructed whereby the vlues of the vribles differ for ech of the discrete choices. The vribles include three mesures of gglomertion, three tx vribles, popultion density, n ccessibility vrible, nd the distnce to the centroid of ech center from the nerest firm of prticulr industry group. (3) For the sixteen industries, the percentge of observtions with zero reported employees rnged from low of 4% for SIC 30örubber nd plsticsöto high of 41% for SIC 67öholding compnies.

7 Firm loction in polycentric city 677 Tble 2. Loction ptterns of firms in the `full' nd `centers' smples. Numbers of firms mnufcturing oil nd gs FIRE services ll 4 groups Full smple Rest of Hrris County Subcenters Totl Centers smple CBD b Bytown Psden LPorte Cler Lke Glleri Crrilon Greenspoint Totl FIREÐfinnce, insurnce, nd rel estte. b CBDÐcentrl business district. We focus on the geogrphy of gglomertive forces within the city, nd proxy potentil gglomertive forces with three different mesures. One vrible focuses on the number of nerby firms, nd the other two on nerby employment. The first vrible, FIRMS, is defined s the number of firms in the sme industril group within 1-mile rdius of the census trct contining n employment center. This ws constructed to cpture the ttrctiveness of firm to other firms in the sme brod industril ctegory. This mesure is intended to cpture the microgeogrphy of locliztion economies nd is most closely relted to some of the work of Rosenthl nd Strnge (2005) for New York City, who found tht the impct gglomertion economies ttenute with distnce nd re strongest for 1-mile rdius. We hypothesize tht links to other like firms nd other complementry firms re wht chrcterizes the bsis for gglomertion economies within cities. We lso constructed two other proxy mesures of gglomertive forces. As n lterntive to the number of nerby firms to proxy for locliztion economies, we define the vrible IND EMP to be the totl number of employees in the sme industry. Counting the number of employees rther thn the number of firms should control for industril orgniztion chrcteristics. For exmple, few lrge firms my hve mny employees, nd the source of the gglomertion economy my be from potentil lbormrket pooling s proxied by the entire employee bse. The finl mesure is our ttempt to cpture the ide of urbniztion economies on microgeogrphicl scle. Trditionlly, urbniztion economies re thought of s the benefits tht given firm receives from the expnsion of the entire industril bse of n urbn re. In our cse, however, we define sptilly limited mesure: the benefits firm gets from the expnsion of the entire industril bse within given subregion of the metropolitn re. To do so, we define the third gglomertion vrible, TOT EMP, to be the totl number of employees in ll industril sectors (not just the four studied in detil here) within 1-mile rdius of the census trct center. The dt for this vrible were creted by the Houston ^ Glveston Are Council, nd re lso bsed on Dun nd Brdstreet dt.

8 678 J E Kohlhse, X Ju Tble 3. Men chrcteristics (with stndrd devitions shown in prentheses) of `full' nd `centers' smples. Mnufcturing Oil nd gs FIRE Services Full smple Firm size b (153.39) (126.62) (119.69) (153.30) Number of firms Subcenters smple Firm chrcteristics Firm size b (309.92) (164.57) (101.97) (82.63) Number of firms Choice chrcteristics c vrying by industry c Agglomertion number of firms (14.93) (78.063) ( ) (165.72) in sme industry number of (0.602) (3.766) (2.252) (4.361) employees in sme industry (thousnds) Totl number of (49.109) (49.109) (49.109) (49.109) employees in ll industries (thousnds) Distnce to firm, (0.253) (0.253) (0.200) (0.162) sme industry (miles) common to ll industries c Tx ($ per $100 vlue) (0.223) (0.223) (0.223) (0.223) Popultion density (3.033) (3.033) (3.033) (3.033) (thousnds) Distnce to highwy (2.970) (2.970) (2.970) (2.970) (miles) FIREÐfinnce, insurnce, nd rel estte. b Averged over ll firms. The fiscl vrible we use, TAX, mesures the locl tx burden imposed by vrious government nd qusi-government bodies. Property-tx vribles were collected for twenty-three Independent School Districts (ISDs), twenty-five cities, nd over 300 Municipl Utility Districts (MUDs) within Hrris County, Texs, for the yer County property txes re not included since ll observtions lie within Hrris County. The tx rtes re effective property tx rtes mesured in dollrs per $100 property vlution. The ISD nd city-tx dt were gthered from the Hrris County Apprisl District. The MUD tx dt were retrieved from Municipl Service's dt file. Ech tx vrible ws geogrphiclly coded by census trct nd ssigned to the pproprite site-specific vrible. The Houston re is unusul in its method of providing severl locl public services, such s wter nd gs. Ares outside the city limits often re provided with locl services by specil districts clled Municipl Utility Districts, or MUDs. MUDs re creted for specific purposes within limited geogrphicl re. In 1990, 313 MUDs existed within Hrris County outside city boundries. One unique feture of the empiricl study described here is tht it is the first (to the best of the uthors' knowledge) to

9 Firm loction in polycentric city 679 incorporte the MUD tx rtes into firm loction decision study. MUD tx rtes were typiclly excluded in previous studies of Houston, due to lck of geocoded dt source. (4) We creted three vribles to describe the economic environment of ech employment center. These vribles serve to cpture elements of ny comprtive dvntge given center my hve within the sptil structure of the metropolitn re. The first vrible, DENSITY, mesures the popultion density of ech of the employment centers nd is clculted by dividing the 1990 popultion by lnd re for ech employment center. The vrible is ssumed to cpture the generl economic environment of the center. The men popultion density is 3983 persons per squre mile. The highest vlue of popultion density is persons per squre mile, locted ner the Glleri. The second vrible, DHIGHWAY, is used to ccount for the reltive ccessibility of ech employment center within the generl highwy trnsporttion network of the metropolitn re. The mesure of the highwy ccess of n employment center is defined s the distnce from the centroid of ech center to the nerest freewy entrnce, nd mesured using ArcView GIS. The selected freewys in this study re Interstte Highwy 10 (I-10), Interstte Highwy 45 (I-45), Stte Highwy 59 (US-59), Interstte Highwy Loop 610 (I-610), nd Beltwy 8. The third vrible is intended to cpture n indiction of within-center lnd prices or rents. A typicl lnd-rent function would hve peks in the centroids of ech employment center nd decline with distnce from ech center. As proxy for the expected lnd-price premium ner employment centers, we define the vrible, DNEAR, to be the distnce from the centroid of ech subcenter to the nerest firm of given industril group. This mesure is dmittedly very crude but, in the bsence of lnd rent dt, it seems resonble lterntive. There is one firm-specific independent vrible included in this dtset, EMPLOYEES, which is obtined from Dun nd Brdstreet's dt file directly. The vrible reflects the size of ech firm nd is mesured by the number of employees t ech plnt loction. This mesure is used to control for differences mong firms tht might be due to internl scle economies. The Dun nd Brdstreet dtset reports the qulity of the employee dt, so in constructing our dtset we used only firms whose employment dt were reported s n ctul (not estimted) vlue tht ws not zero. 4 Empiricl results Two clsses of discrete choice models re estimted: for the choice of dispersed versus concentrted loction; nd for the choice mong the eight employment centers. Coefficient estimtes re reported in tbles 4 nd 5 ^ 8 nd elsticities re reported in tbles 4, 9, nd 10. The elsticities were mesured t the observtion level nd then verged. Tble 11 reports comprison of the predicted nd ctul number of firms for model 1 (described below). Results re presented for the four industril groupsöoil nd gs, mnufcturing, FIRE, nd services. Initilly, our intention ws to estimte discrete choice model with nine choices: the first eight choices being mong the eight employment subcenters nd the ninth choice being the `rest of Hrris County'. However, the definition of the choice-specific vribles such s tx rte nd popultion density were problemtic for this ninth choice s it hd such lrge lnd re nd vriety of vlues. Therefore we present two sets of estimtes: one set isoltes the dispersed ^ concentrted choice while the second set looks in prticulr t concentrted loction choices. (4) We would like to thnk Dr Ronld Welch of Welch Assocites for providing us with geocoded boundry file for the Hrris County MUDs. In recent pper, Plmon nd Smith (1998) use MUD tx rtes in study of housing-price cpitliztion.

10 680 J E Kohlhse, X Ju Tble 4. Concentrted versus dispersed loction: logit-estimtion results (with z-sttistics shown in prentheses) for the `full smple'. Estimted model Mnufcturing Oil nd gs FIRE Services vrible Employees (2.49)* (2.37)* (0.54) (0.61) Constnt ( 19.00)** ( 10.69)** ( 27.17)** ( 61.41) Log likelihood Likelihood rtio w 2 (1) b Probbility > w n Correctly clssified (%) Clculted elsticity Employees *significnt t the 5% level, two-tiled test; **significnt t the 1% level, two-tiled test. Note: Dependent vrible is dichotomous: 1 ˆ locte in ny of the 8 employment centers, 0 ˆ locte in rest of Hrris County. FIREÐfinncil, insurnce, nd rel estte. b Likelihood rtio test, clculted test sttistic with 1 degree of freedom. 4.1 Dispersed versus concentrted loction To estimte firm's choice between concentrted versus dispersed loctions, we use simple logit with firm-specific vribles s regressors. By necessity our model is modest: becuse of dtset limittions, the only firm-specific vrible vilble for the nlysis is the number of employees t the estblishment. Given these limittions, the function from eqution (5) cn be expressed s follows: f z m ˆ 0 2 EMPLOYEES, (6) where EMPLOYEES is the totl number of employees t single plnt loction. The coefficients nd elsticities reported in tble 4 show tht loction choices by firms differ by industril sector. Both for mnufcturing nd for oil nd gs, lrger firms re more likely to choose concentrted loctions. In contrst, firm size is not source of loction preference for firms in the FIRE nd services sectors. The finding of the heterogeneity of loction choices to firm size results crries over into the more detiled model estimted below. 4.2 Choice of employment center Loction preferences of firms tht choose loction tht cn be chrcterized by concentrted employment densities were exmined in mixed-multinomil logit frmework. Firms locting outside employment centers re excluded from the following estimtions, nd becuse no MUDs exist in the employment centers, the tx vrible is composed only of ISD nd city tx rtes. Results re reported for the four brod industril groups in tbles 5 ^ 10: tbles 5 ^ 8 present coefficient estimtes for three models differentited by gglomertion mesure; tbles 9 nd 10 present elsticities, clculted t the observtion level nd then verged.

11 Firm loction in polycentric city 681 Tble 5. Choice of employment center for mnufcturing firms (with z-sttistics shown in prentheses. Vrible Model 1 Model 2 Model 3 Choice-specific vrible Number of firms in ( 1.57) sme industry Number of employees ( 3.22)** in sme industry (thousnds) Totl number of ( 1.88) + employees in ll industries (thousnds) Popultion density (5.52)** (6.08)** (5.80)** (thousnds) Tx ( 6.12)** ( 6.82)** ( 6.33)** Distnce to highwy ( 2.52)* ( 1.81) ( 2.36)** (miles) Distnce to firm ( 4.13)** ( 5.56)** ( 4.69)** (miles) Firm-specific vrible Employees Psden ( 0.71) ( 0.63) ( 0.71) LPorte ( 0.45) ( 0.77) ( 0.48) Cler Lke (0.19) (0.38) (0.26) Glleri ( 0.50) ( 0.82) ( 0.56) Crrilon ( 0.55) ( 0.65) ( 0.58) Greenspoint ( 1.68) ( 2.12)* ( 1.79) + Number of firms b Log likelihood Likelihood rtio w 2 12 _ + significnt t 10% level, two-tiled test; *significnt t 5% level, two-tiled test; **significnt t 1% level, two-tiled test. Note: Bse ctegory is CBD. Bytown omitted due to lck of dt. b Number of observtions equls 8 (number of firms). c Likelihood rtio test, clculted test sttistic with 12 degrees of freedom. The function f(.) of the mixed-multinomil logit eqution, eqution (4), cn be expressed in the following three models, differentited by mesure of gglomertion: Model 1 f S ki, F mj ˆ 1 DENSITY 2 TAX 3 FIRMS 4 DHIGHWAY 5 DNEAR X8 i ˆ 1 i 5 EMPLOYEES CENTER i, (7) Model 2 f S ki, F mj ˆ 1 DENSITY 2 TAX 3 IND EMP 4 DHIGHWAY 5 DNEAR X8 i ˆ 1 i 5 EMPLOYEES CENTER i, (8)

12 682 J E Kohlhse, X Ju Tble 6. Choice of employment center for oil nd gs firms (with z-sttistics shown in prentheses. Vrible Model 1 Model 2 Model 3 Choice-specific vrible Number of firms in ( 3.43)** sme industry Number of employees ( 0.71) in sme industry (thousnds) Totl number of ( 1.59) employees in ll industries (thousnds) Popultion density (10.50)** (9.37)** (9.73)** (thousnds) Tx ( 12.33)** ( 12.16)** ( 12.64)** Distnce to highwy ( 3.89)** ( 4.23)** ( 4.19)** (miles) Distnce to firm ( 8.22)** ( 6.96)** ( 6.95)** (miles) Firm-specific vrible Employees Bytown (1.78) (1.51) (1.57) Psden (0.93) (0.87) (0.89) LPorte (1.33) (1.86) (1.71) + Cler Lke (0.16) ( 0.22) ( 0.07) Glleri (3.17)** (3.09)** (3.12)** Crrilon (3.71)** (3.80)** (3.78)** Greenspoint ( 0.13) ( 0.30) (0.13) Number of firms Log likelihood Likelihood rtio w 2 12 _ + significnt t 10% level, two-tiled test; **significnt t 1% level, two-tiled test. Note: Bse ctegory is CBD. Number of observtions equls 8 (number of firms). b Likelihood rtio test, clculted test sttistic with 12 degrees of freedom. Model 3 f S ki, F mj ˆ 1 DENSITY 2 TAX 3 TOT EMP 4 DHIGHWAY 5 DNEAR X8 i ˆ 1 i 5 EMPLOYEES CENTER i, (9) where the index is over the eight employment centers. The CBD (i ˆ 1) is used s the bse ctegory, so tht 6 ˆ 0. The gglomertion vribles (FIRMS, IND EMP, nd TOT EMP) mesure potentilly ttrctive forces within 1-mile bnd of the census trct contining ech center. The vrible FIRMS is the number of firms from the sme brod industril group, IND EMP is the number of employees in the sme industry, nd TOT EMP is the totl number of employees in ll industries. The vrible TAX is the sum of city nd school district tx rtes, DNEAR is the distnce from the centroid of ech subcenter to the nerest firm in the industril group, nd DHIGHWAY is the stright-line distnce from the centroid of ech subcenter to the nerest freewy. In the estimtion procedure, ech site-specific vrible genertes one prmeter estimte nd ech chooser-specific vrible genertes seven prmeter estimtes becuse the CBD choice

13 Firm loction in polycentric city 683 Tble 7. Choice of employment center for FIRE (finnce, insurnce, nd rel estte) firms (with z-sttistics shown in prentheses). Vrible Model 1 Model 2 Model 3 Choice-specific vrible Number of firms in (15.51)** sme industry Number of employees (10.02)** in sme industry (thousnds) Totl number of (10.67)** employees in ll industries (thousnds) Popultion density (10.50)** (16.11)** (16.21)** (thousnds) Tx ( 3.66)** ( 5.69)** ( 7.88)** Distnce to highwy ( 6.94)** ( 13.13)** ( 12.23)** (miles) Distnce to firm (0.90) (7.10)** (7.89)** (miles) Firm-specific vrible Employees Bytown ( 0.4) ( 0.92) ( 1.03) Psden ( 2.68)** ( 1.98)* ( 1.87) LPorte (0.01) (0.42) (0.38) Cler Lke ( 0.82) ( 0.61) ( 0.56) Glleri )* ( 0.40) ( 0.42) Crrilon ( 1.23) ( 0.81) ( 0.80) Greenspoint ( 0.70) ( 1.71) ( 1.60) Number of firms Log likelihood Likelihood rtio w 2 12 _ + significnt t 10% level, two-tiled test; *significnt t 5% level, two-tiled test; **significnt t 1% level, two-tiled test. Note: Bse ctegory is CBD. Number of observtions equls 8 (number of firms). b Likelihood rtio test, clculted test sttistic with 12 degrees of freedom. prmeter is normlized to zero. For the chooser-specific vrible (firm size), we specified the vrible s n interction term, EMPLOYEES CENTER, siscommon prctice in mixed-multinomil models (Long nd Freese, 2003) Agglomertion economies We found evidence tht both gglomertion economies nd gglomertion diseconomies exist in Houston nd tht the results differ by brod industril group (tble 11). Furthermore, we found evidence consistent with the opertion of different types of gglomertion economiesöour prticulr version of locliztion nd urbniztion economies. We interpret the signs of the significnt coefficients on the gglomertion vribles s being indictive of the presence or bsence of gglomertive forces. This is becuse the signs of the coefficients of the choice-specific vribles re the sme s the signs of the mrginl effects in mixed-multinomil logit (Greene, 2003). The coefficients of the three gglomertion proxy vribles, FIRMS, IND EMP, ndtot EMP re significnt t the 12% level or better, with the exception of the coefficient on IND EMP for the oil nd

14 684 J E Kohlhse, X Ju Tble 8. Choice of employment center for services firms (with z-sttistics shown in prentheses). Vrible Model 1 Model 2 Model 3 Choice-specific vrible Number of firms in (16.10)** sme industry Number of employees (15.77)** in sme industry (thousnds) Totl number of (13.26)** employees in ll industries (thousnds) Popultion density (15.15)** (16.64)** (19.10)** (thousnds) Tx ( 12.01)** ( 10.58)** ( 10.48)** Distnce to highwy ( 10.91)* ( 12.66)** ( 15.24)** (miles) Distnce to firm ( 6.40)** (2.03)** (4.94)** (miles) Firm-specific vrible Employees Bytown (0.79) (0.21) (0.02) Psden ( 1.55) ( 1.36) ( 1.34) LPorte (0.04) (0.29) (0.59) Cler Lke ( 0.53) ( 0.21) ( 0.19) Glleri ( 1.05) ( 0.64) ( 0.05) Crrilon (0.05) (0.29) (0.39) Greenspoint ( 1.54) ( 1.66)* ( 2.64)** Number of firms Log likelihood Likelihood rtio _ w 2 *significnt t 5% level, two-tiled test; **significnt t 1% level, two-tiled test. Note: Bse ctegory is CBD. Number of observtions equls 8 (number of firms). b Likelihood rtio test, clculted test sttistic with 12 degrees of freedom. gs sector. Moreover, the results re remrkbly consistent cross the three mesures of gglomertive forces. Locliztion economies pper to operte in two of the four sectors. Nerby firms nd employees in the sme industril groups pper to be n ttrctive force for only the FIRE nd services industril groups. In contrst, nerby firms nd employees of the sme industril group pper to be repelling forces for firms in the oil nd gs group s well s for the mnufcturing group. The results re similr when we exmine our mesure for urbniztion economiesöthe totl number of employees cross ll industries. There ppers to be evidence of urbniztion economies for firms in FIRE nd services, but diseconomies for firms in mnufcturing, nd oil nd gs. The mgnitude of the ttrctive nd dispersive effects cn be exmined in the context of the elsticities reported in tble 9. Quntittively, most responses re smllös evidenced by most elsticities being < 1. Exceptions re for firms in the FIRE sector which re reltively sensitive to the mkeup of the Glleri nd CBD (severl elsticities > 1). For exmple, 10% increse in the totl number of employees in ll industries increses the probbility of FIRE firm locting in the CBD by bout 17%, nd locting in Glleri by bout 12%. In contrst, repelling nd ttrcting forces cn be found by compring the elsticities of the oil nd gs group to the service group.

15 Firm loction in polycentric city 685 Tble 9. Elsticities for gglomertion vribles by center. Agglomertion Mnufcturing Oil nd gs FIRE Services vrible Number of firms in sme industry CBD Bytown Psden LPorte Cler Lke Glleri Crrilon Greenspoint Number of employees in sme industry CBD Bytown Psden LPorte Cler Lke Glleri Crrilon Greenspoint Totl number of employees in ll industries CBD Bytown Psden LPorte Cler Lke Glleri Crrilon Greenspoint Note: Elsticities were evluted t the observtion level nd then verged by choice. FIREÐfinnce, insurnce, nd rel estte. A 10% increse in the number of sme-group firms in the CBD decreses the probbility of n oil nd gs firm locting there by 6%; but, 10% increse in the totl number of sme-group firms in the CBD increses the probbility tht service firm loctes in CBD by over 6%, nd firm in the FIRE sector by bout 10% Property txes Results in tble 10 (nd summrized in tble 12) indicte tht, lthough proximity vribles do ply role in determining firm loction, firms re more significntly impcted by public policy in their loction decisions. The significnce t the 1% level nd the negtive signs of the property-tx coefficients imply tht firms tret the property txes s loction deterrents. The mgnitudes of tx elsticities differ by industril group, but re elstic in most cses. Previous reserch by others hs found evidence tht txing jurisdictions cn influence the mount of ctivities tking plce within them (see the summry in Brtik, 1991; Chrney, 1983; Finney, 1994; Fox, 1981; McGuire, 1985; McHone, 1986; Wsylenko, 1980; 1980b). The question is tht of how, nd to wht degree, do txes ffect firm loction.

16 Tble 10. Elsticities for other vribles by center. Vrible Mnufcturing Oil nd gs FIRE Services Choice-specific vrible Tx CBD Bytown Psden LPorte Cler Lke Glleri Crrilon Greenspoint Popultion density CBD Bytown Psden LPorte Cler Lke Glleri Crrilon Greenspoint Distnce to highwy CBD Bytown Psden LPorte Cler Lke Glleri Crrilon Greenspoint J E Kohlhse, X Ju

17 Tble 10 (continued). Mnufcturing Oil nd gs FIRE Services Distnce to firm CBD Bytown Psden LPorte Cler Lke Glleri Crrilon Greenspoint Firm-specific vrible Employees CBD Bytown Psden LPorte Cler Lke Glleri Crrilon Greenspoint Notes: Models differ by mesure of gglomertion: model 1 uses number of firms in the sme industry; model 2, the number of employees in the sme industry; nd model 3, the totl number of employees in ll industries. Elsticities re evluted t the observtion level nd then verged by choice. FIREÐfinnce, insurnce, nd rel estte. Firm loction in polycentric city 687

18 688 J E Kohlhse, X Ju Tble 11. Evidence of gglomertion economies in the centers model. Industry Sign of gglomertion effect Type gglomertive force FIRMS IND EMP TOT EMP Mnufcturing neg neg*** neg* dispersive Oil nd gs neg*** neg neg wek dispersive FIRE pos*** pos*** pos*** ttrctive Services pos*** pos*** pos*** ttrctive *significnt t 10% level; ***significnt t 1% level. Fire, insurnce, nd rel estte. Tble 12. Evidence of property-tx effects in the centers model. Industry Tx-effect sign elsticities rnge Loction impct (bsolute vlues) Mnufcturing neg*** 3.7 ± 7.6 deterrent Oil nd gs neg*** 4.4 ± 15.3 deterrent FIRE neg*** 0.7 ± 3.3 deterrent Services neg*** 1.4 ± 2.8 deterrent *** significnt t 1% level. Fire, insurnce, nd rel estte Oil nd gs firms show the most elstic responses to txes, with elsticities rnging from pproximtely 4 to 15, followed by mnufcturing, with elsticities rnging from bout 4to 8. Firms in FIRE nd services lso view property txes s loction deterrent but re somewht less sensitive to chnges, with most elsticities flling between 1 nd 3. The single inelstic response is for the Glleri choice for firms in the FIRE sector, where the elsticity is 0.7. The inelstic response mkes sense in tht this loction hs the highest predicted probbility. If the choice probbilities re unbised estimtes of the proportion of firms in ech center, the chnge in choice probbilities due to the tx chnges my be thought of s prediction of the chnge in the proportion of new firms locting in the centers. Therefore, the elsticities could be interpreted s the percentge chnges in the expected reltive frequency of smple of new firms opening in the centers resulting from percentge chnge in the tx rte of the respective center. For exmple, compring the CBD nd Glleri choices for model 1, 1% increse in property txes in those centers would led to bout 9% decrese in the reltive frequency of oil nd gs firms locting in Glleri nd 5% decrese for them in the CBD. In contrst, for FIRE firms, the tx effects re much smller. A 1% increse in property txes would led to 0.7% decrese in the reltive frequency of locting in Glleri nd 1% for the CBD. Compring our findings to other studies bsed on differing methodologies, we find further support for elstic responses to txes. Chrney's (1983) elsticity of 2.52, or Fox's (1981) elsticity of 4.43, re close to mny of the tx elsticities in this study in mgnitudes. Since Chrney's (1983) dependent vrible is firm density, her clculted tx elsticity is perhps more comprble to tht of the present study thn is Fox's elsticity; Fox's dependent vrible is the lnd re by jurisdiction devoted to industril production Other independent vribles As determinnt of firm loction, popultion density is significnt t the 1% level for ll industril groups. The signs of the coefficients on DENSITY re ll positive nd

19 Firm loction in polycentric city 689 imply tht firms choosing employment center loctions re ttrcted to loctions with higher popultion densities. To the extent tht popultion densities reflect generl economic outcomes, it cn be concluded tht firms choose to locte in economiclly vibrnt res. The coefficients of ll the highwy-distnce vribles, DHIGHWAY, hve negtive signs, nd re significnt t the 5% or 1% level for ll models. The exception is for the mnufcturing sector, where the coefficient is significnt t the 10% level in model 2. From the elsticity results reported in tble 10, we re ble to predict how responses differ by sector. For exmple, bsed on model 1, 10% increse in distnce to the nerest highwy would reduce the probbility of firm locting in Psden by bout 8% for mnufcturing firms, by bout 11% for the oil nd gs firms, by bout 14% for FIRE, nd 11% for the service firms. The vrible DNEAR mesures the reltive ttrctiveness of proximity to the centroid of ech employment center. The signs of the distnce vribles vry by model nd sector. Expected negtive signs occur for ll three models estimted for mnufcturing nd oil nd gs. But for FIRE nd services severl positive coefficients re found, especilly when the employee-count mesures of gglomertive forces re used. Responses by FIRE nd services firms re in generl inelstic nd re the lest responsive to chnges in distnce. The mny positive inelstic results could indicte tht these types of firms prefer to locte wy from the centroid of employment centers, in order to tke dvntge of potentilly lower rents. An lterntive explntion my be tht mny of the FIRE nd services firms prefer to locte fr from the centroid of the employment centers in order to cpture lrger locl mrket re for their products. (5) For exmple, bsed on model 3, 10% increse in distnce from the centroid of center to firm in the sme industry would reduce the probbility of firms locting in Glleri by 10% for oil nd gs nd 12% for mnufcturing, but increse the probbility by 4% for FIRE nd 2% for services. The one firm-specific vrible, EMPLOYEES CENTER, is our proxy of firm size, nd is included to mesure internl economies of scle. If n increse in the number of employees per firm significntly reduces the probbility of the mrginl firm locting in tht subcenter, then lrger firms do not fvor tht subcenter. The results show mixed results for the firm-size effect nd, s indicted erlier, the results for the mixedmultinomil logit mirror erlier results found in the dispersed model. Firm size is more likely to be n importnt determinnt of loction for firms in the mnufcturing, nd oil nd gs sectors thn for firms in the FIRE nd services sectors. Mny of the signs of the coefficients re indeterminte [nd the mrginl effects nd elsticities need not be of the sme sign s the coefficient for the firm-specific vribles in mixed-multinomil setting (Greene, 2003)] nd only few of them re significnt t the 5% or 10% level. The sttisticl significnce of the firm-size vrible is industry-specific nd choice-specific nd elsticities re quite smll. The one exception to the inelstic pttern occurs for the mnufcturing sector, for which responses re elstic: for exmple, if mnufcturing firm's own employment incresed by 10%, the probbility tht the firm would locte in the Greenspoint subcenter would decline by between 16% nd 22%. 5 Summry nd conclusion Our reserch is one of few ttempts to model the workings of within-city gglomertive forces using firm-level dt in discrete-choice model. Moreover, fiscl vribles operte t fine level of sptil detil, so tht our results re likely to be n improvement on previous work tht hs been done t higher levels of sptil ggregtion. (5) We would like to thnk referee for this insight.