B andily, P.; Rauch, F.
A icle — Published Ve sion
Wi hin‐ci y oads and u ban g ow h
Jou nal o Regional Science
P o ided in Coope a ion wi h:
John Wiley & Sons
Sugges ed Ci a ion: B andily, P.; Rauch, F. (2024) : Wi hin‐ci y oads and u ban g ow h, Jou nal o
Regional Science, ISSN 1467-9787, Wiley, Hoboken, NJ, Vol. 64, Iss. 4, pp. 1236-1264,
h ps://doi.o g/10.1111/jo s.12699
This Ve sion is a ailable a :
h ps://hdl.handle.ne /10419/306120
S anda d-Nu zungsbedingungen:
Die Dokumen e au EconS o dü en zu eigenen wissenscha lichen
Zwecken und zum P i a geb auch gespeiche und kopie we den.
Sie dü en die Dokumen e nich ü ö en liche ode komme zielle
Zwecke e iel äl igen, ö en lich auss ellen, ö en lich zugänglich
machen, e eiben ode ande wei ig nu zen.
So e n die Ve asse die Dokumen e un e Open-Con en -Lizenzen
(insbesonde e CC-Lizenzen) zu Ve ügung ges ell haben soll en,
gel en abweichend on diesen Nu zungsbedingungen die in de do
genann en Lizenz gewäh en Nu zungs ech e.
Te ms o use:
Documen s in EconS o may be sa ed and copied o you pe sonal
and schola ly pu poses.
You a e no o copy documen s o public o comme cial pu poses, o
exhibi he documen s publicly, o make hem publicly a ailable on he
in e ne , o o dis ibu e o o he wise use he documen s in public.
I he documen s ha e been made a ailable unde an Open Con en
Licence (especially C ea i e Commons Licences), you may exe cise
u he usage igh s as speci ied in he indica ed licence.
h p://c ea i ecommons.o g/licenses/by/4.0/
Recei ed: 28 No embe 2022
|
Re ised: 27 Feb ua y 2024
|
Accep ed: 10 Ma ch 2024
DOI: 10.1111/jo s.12699
RESEARCH ARTICLE
Wi hin‐ci y oads and u ban g ow h
P. B andily
1
|F. Rauch
2
1
London School o Economics, London, UK
2
Heidelbe g Uni e si y, Heidelbe g, Ge many
Co espondence
F. Rauch, Heidelbe g Uni e si y, Heidelbe g,
Ge many.
Email: [email p o ec ed]
Funding in o ma ion
Depa men o In e na ional De elopmen ,
UK Go e nmen ; Wo ld Bank G oup
Abs ac
In his pape we s udy he ole o wi hin‐ci y oads layou in
os e ing ci y g ow h. Wi hin‐ci y oads ne wo ks ha e no
been s udied ex ensi ely in economics al hough hey a e
essen ial o acili a e human in e ac ions, which a e a he
co e o agglome a ion economies. We build and compu e
se e al simple measu es o oads ne wo k and cons uc a
sample o o e 1800 ci ies and owns om Sub‐Saha an
A ica. Using a simple econome ic model and wo
ins umen al a iable s a egies based on he his o y o
A ican ci ies, we hen es ima e he causal impac o wi hin‐
ci y oads layou on u ban g ow h. We ind ha o e he
ecen decades, ci ies wi h g ea e oad densi y and oad
e enness in he cen e g ew as e .
KEYWORDS
oad layou , Sub‐Saha an A ica, u ban planning, u banisa ion
1|INTRODUCTION
Ci ies enhance p oduc i i y and consump ion bene i s h ough a ious mechanisms desc ibed in a la ge li e a u e in
economics. Wi hin his adi ion, all mechanisms p oposed o agglome a ion e ec s ela e o one essen ial ea u e
o he u ban en i onmen : i acili a es sho ‐dis ance in e ac ions be ween economic agen s. This ea u e is a he
co e o he main heo ies on agglome a ion o ces, ega dless o whe he hey desc ibe p oduc ion o consump ion.
In his seminal ypology Ma shall (1920) sugges ed knowledge spillo e s, labou ma ke ex e nali ies and p oduc ion
linkages as d i e s o u ban p oduc i i y; mo e ecen ly, Du an on and Puga (2004) highligh he impo ance o
J Regional Sci. 2024;64:1236–1264.1236
|
wileyonlinelib a y.com/jou nal/jo s
This is an open access a icle unde he e ms o he C ea i e Commons A ibu ion License, which pe mi s use, dis ibu ion and
ep oduc ion in any medium, p o ided he o iginal wo k is p ope ly ci ed.
© 2024 The Au ho s. Jou nal o Regional Science published by Wiley Pe iodicals LLC.
We hank Julia Bi d, S e en B akman, Paul Collie , Vic o Cou u e, Ha y Ga e sen, Ve non Hende son, Richa d Ho nbeck, Guy Michaels, Ma Tu ne ,
Tony Venables, a numbe o anonymous e e ees and a ious semina pa icipan s o help ul commen s. We g a e ully acknowledge he gene ous suppo
o a Global Resea ch P og am on Spa ial De elopmen o Ci ies, unded by he Mul i Dono T us Fund on Sus ainable U baniza ion o he Wo ld Bank and
suppo ed by he UK Depa men o In e na ional De elopmen .
sha ing, ma ching, and lea ning. In his pape we in es iga e he ole o oads and s ee s ne wo k wi hin ci ies. We
hus ocus on a cen al componen o he u ban se ing ha could di ec ly p omo e o discou age in e ac ions
be ween agen s. Mo e speci ically, we s udy he impac o he densi y and shape o he oad g id wi hin A ican
ci ies on popula ion g ow h in ecen decades.
To s udy hese speci ic u ban ea u es, we measu e a ious s a is ics on he oad ne wo ks wi hin u ban
cen es o a la ge numbe o A ican ci ies and show he associa ion be ween some o he layou 's ea u es and
u ban g ow h. Combining sa elli e image y and Open S ee Map, an open‐sou ce geog aphic da a se mapping
oads and s ee s all o e he wo ld, we build simple s a is ics o he oad layou s a a p ecise geog aphic scale. We
belie e ou ocus on wi hin‐ci y layou s o be a concep ual inno a ion o he economic li e a u e whe e he opic
has emained la gely uns udied so a . We show ha measu es o he cen al oad g id, in pa icula i s densi y and
e enness, a e associa ed wi h ecen popula ion g ow h.
We ocus on he link be ween hese measu es o he wi hin‐ci ies oads layou and u banisa ion
speci ically in he con ex o Sub‐Saha an A ica. We chose his ocus o di e en easons. Fi s , his is he
egion o he wo ld ha is cu en ly expe iencing he as es a e o u banisa ion (B yan e al., 2020).
Unde s anding he ole o his u ban in as uc u e in he capaci y o ci ies o gene a e g ow h be e could
help designing policies ha bene i he billions o people ha li e o a e expec ed o li e in A ican ci ies in
he nex decades. Second, he oad layou is mo e meaning ul in an en i onmen ha elies hea ily on oad
anspo a ion, and cu en ly does no ha e an ex ensi e u ban ain o unde g ound ne wo k. Thi d, i is
cos ly o change a oad layou once i is in place. Planning ahead is he e o e c i ical in A ica's u banisa ion,
which is cha ac e ised by low‐income le els and whe e ‘ wo‐ hi ds o he ci ies a e o be buil ’(B yan e al.,
2020). Fou h, a ious s udies epo high le el o a ic conges ion in A ican ci ies and he need o be e
oad in es men s. Hidden ex e nal cos s o conges ion we e es ima ed a up o 5% o ci ies’GDPs in Daka o
Abidjan, which is mo e han wice he es ima es o Eu opean ci ies (Ce e o, 2013)while‘in a sample o 30
ci ies a ound he wo ld, he 8 A ican ci ies ank in bo om 12 spo s o oad densi y’(Lall e al., 2017). Finally,
some a gue in a ou o a ‘g id s uc u e, which (…) enhanced a el e iciency’(as s a ed in he las majo
epo om he Wo ld Bank on A ica's Ci ies ‐idem), al hough o he bes o ou knowledge he e is li le
empi ical esea ch in economics so a o assess hishypo hesis.Ou ocusonwi hin‐ci y e ec s
complemen s a wide li e a u e ha inds impo an posi i e e ec s o anspo a ion in as uc u e
connec ing ci ies ( ecen ly e iewed e.g. by Cui e al., 2023 and Cao e al., 2023).
To es ima e he impac o oad layou we ely on wo obse a ions: he oad layou is bo h di icul o
change and dependen o he con ex and a ailable anspo a ion echnologies a he ime o i s
cons uc ion. The pe sis ence o oad layou o e ime is exempli ied in ci ies o he ancien wo ld (like
Je usalem o Pa is, among many o he s, whe e he Roman's ca do a e s ill impo an s ee s oday) o he new
(as New Yo k whe e he ‘Commissione s’Plan’o 1811 s ill s uc u es Manha an oday). Mo e ecen ly, his
was also one conclusion om a la ge‐scale analysis o ci ies a ound he wo ld by Ba ing on‐Leigh and
Milla d‐Ball (2019). In he case o A ica in pa icula , ci ies a e o en much newe . Ba uah e al. (2021)
p o ide empi ical e idence o layou pe sis ence in A ican ci ies. Toge he , all hese obse a ions imply ha
due o his o ic easons some ci ies may be acciden ally s uck wi h be e o wo se oad g ids, which would
bene i o hu hei g ow h po en ial. I a ci y by his o ic acciden de elops a good oad layou i will be
ewa ded by long las ing popula ion g ow h, as people mo e in o bene i om he agglome a ion bene i s
gene a ed in he ci y cen e. I on he o he hand a ci y is s uck wi h a subop imal g id, i may each a peak
popula ion le el beyond which i inds i di icul o g ow any u he .
Ou main measu e o he quali y o he oad ne wo k is a simple measu e o oad densi y in he cen e o
owns and ci ies, he eby es ing whe he a lack o oad quan i y cons ains u ban g ow h. We also compu e a
second and seconda y simple measu e o he dis ance o he nea es oad o a andom poin in he ci y
cen e, as well as a numbe o measu es o egula i y and o ien a ion o he oad g id. To inc ease
compa abili y ac oss di e en ci ies we keep he a ea o ci y cen e o which we compu e hese measu es
BRANDILY and RAUCH
|
1237
cons an ac oss ci ies. All measu es we use in his pape ha e he ad an age o being in ui i e,
s aigh o wa d o compu e and simple o in e p e . Ou conclusion ocuses on he linea e ec o ou
measu es o oad densi y and spa ial dis ibu ion, which a e he mos obus and quan i a i ely mos
impo an o all he ne wo k ea u es we conside . To add ess measu emen e o and po en ial endogenei y,
we ely on wo ins umen al a iable s a egies based on he age o a ci y. We accep ha he da a a ailable
o us a e limi ed. In pa icula , we canno obse e his o ic oad g ids o a la ge panel o ci ies in a way ha is
compa able. Thus, we ely on p oxy da a. Despi e he una oidable limi a ions in his se ing, we hink ha ou
p oxima e measu es a e s ill good enough o his analysis o unco e a eal e ec . Ou IV app oaches also
su e om weak i s s ages in some speci ica ions. Ye we hink ha he consis ency o esul s ac oss OLS
and a ious IV speci ica ions, despi e hei occasional weaknesses (which we show o comple eness), o e all
poin s o sugges i e e idence conce ning he impo ance o oads in he ci y cen e.
While he e is a subs an ial li e a u e on he e ec o oads o o he o ms o anspo a ion on u ban shape
( ecen examples include Baum‐Snow, 2007; Du an on & Tu ne , 2012; Fabe , 2014; Jedwab & Mo adi, 2016;
Jedwab & S o eyga d, 2016; Michaels, 2008 and S o eyga d, 2016), his li e a u e ypically has concen a ed on
oads connec ing ci ies wi h one ano he . Less is known on oads and oad layou wi hin ci ies (excep ions include
Baum‐Snow e al., 2017, who conside he e ec o u ban highways in China, Akba e al., 2023 who s udy mobili y
in Indian ci ies and Cou u e e al., 2018 who sugges ha es ic ing a ic would b ing wel a e gains). Ou
app oach is mo e ocused on wi hin‐ci y a ia ion and includes smalle oads han hese exis ing s udies. As we
men ioned, he impo ance o oad g id o u ban de elopmen has long been ecognised. Fulle and Rome (2014)
w i e: ‘ge ing he g id igh will likely be mo e impo an han en o cing building codes on s uc u es o imposing
limi s on densi y’. Ye , o ou knowledge, he e is a gap in esea ch quan i a i ely assessing he impo ance o ci ies’
oad layou in economics. The opic has o cou se been s udied in u ban planning o en i onmen al s udies. Layou
shapes ha e been shown o be associa ed wi h anspo decisions, a el ime and local CO2 emissions. No ably, in
a ecen se ies o a icles, Ba ing on‐Leigh and Milla d‐Ball (2019 and, 2020) also use Open S ee Map o p o ide
global‐le el measu es o wi hin‐ci y oads ne wo ks and con i m esul s abou anspo decisions. Thei ocus is no
on agglome a ion economies no do hey speci ically s udy he Sub‐Saha an A ican con ex . While his egion has
known a apid u banisa ion in ecen decades and will mos likely con inue o do so, esea ch is s ill needed on his
con ex . In a li e a u e e iew on de eloping‐wo ld ci ies, B yan e al. (2020) insis on he need o mo e esea ch on
u ban mobili y in A ican ci ies.
O e all, ou esul s sugges ha dense oad ne wo ks a e associa ed wi h mo e popula ion g ow h. Ou
in e p e a ion o his empi ical ela ionship is ha be e connec ed ci y‐cen e os e in e ac ions,
agglome a ion economies and e en ually he g ow h o ci ies. We also ind e idence ha , gi en a speci ic
densi y o oad, he e en dis ibu ion o oads ac oss space in he cen e is also a signi ican de e minan o
g ow h, and mo e so han he g id‐o ien a ion o he numbe o nodes in he ne wo k. The ype o ea u es
we analyse he e is associa ed wi h an inc eased numbe o des ina ions ha can be eached wi hin a gi en
ime ame. We ake away om ou esul s ha such an inc ease in connec i i y is impo an o he g ow h
o ci ies. Al hough we acknowledge limi s o ou iden i ica ion s a egies, we hope ha he consis ency o ou
esul s is indeed indica i e o a causal ela ionship be ween oad ne wo k and ci y pe o mance (an
iden i ica ion mos ly absen om he li e a u e ou side o economics). By hese indings, we also p o ide
e idence o a hypo hesis ad anced by Collie and Venables (2016) ha he oad ne wo k is an impo an
p edic o o ecen popula ion g ow h, and ha many ci ies and owns in Sub‐Saha an A ica a e cons ained
in hei ela i e g ow h due o lack o oad densi y.
The pape is o ganised as ollows: Sec ion 2de elops ou hypo hesis in g ea e de ail, discusses how i
connec s wi h he li e a u e and p esen s ou main measu es o oads ne wo k. Sec ion 3desc ibes he cons uc ion
o he da ase s used and p esen s desc ip i e s a is ics. Sec ion 4de elops he empi ical s a egy and discusses he
ins umen al a iable s a egies we use. Sec ion 5p esen s he main esul s. We p esen a se ies o obus ness
checks in Sec ion 6. Sec ion 7concludes.
1238
|
BRANDILY and RAUCH
2|HYPOTHESIS AND MAIN MEASURES
Ou esea ch ques ion, ou hypo hesis and ou choice o ins umen al a iable a e shaped by connec ing se e al
indings om he u ban economics li e a u e. Fi s , we assume ha he e u ns om agglome a ion a e la gely
gene a ed in he ci y cen e. This is one o he s anda d assump ions used o example in monocen ic ci y models
( ollowing Mills, 1981), cen al place heo y (see e.g. Ch is alle , 1933) as well as o he heo ies on he economics o
ci ies. This sugges s ha people li e a ound a ci y cen e o cen al business dis ic , o which hey a el o wo k
and enjoy leisu e ime. The agglome a ion e u ns a e o e whelmingly gene a ed in ha cen e. Consis en wi h his
li e a u e, we pay special a en ion o he oad g id in he cen e o ci ies a he han he whole oad g id s uc u e
o e he en i e u ban agglome a ion. Al hough we also s udy he ci y as a whole o compa ison in ou obus ness
sec ion, ou p e e ed measu e cap u es only he cen e. Empi ical e idence o he monocen ic model has been
ound o a a ie y o ci ies, by demons a ing g adien s o wages and house p ices, as well as om commu e lows
( o ecen e idence o A ica, see An os e al., 2016 o La com e al., 2017). I any hing, monocen ici y has been
ound o be mo e p onounced in A ica han in he de eloped wo ld (Ce e o, 2013; G o e Goswami & Lall, 2016).
Fu he down in his d a we also p o ide empi ical e idence ha indeed an addi ional squa e me e o oad buil in
he ci y cen e has a s onge e ec on popula ion g ow h han an addi ional squa e me e o oad buil anywhe e in
he wide ci y. To de ine a ci y cen e we pick he cen oid o he b igh es pixel o nigh ligh o a ci y and d aw a
ci cle a ound i . In ou main speci ica ion, his ci cle has a adius o 1 km. This app oach has he addi ional ad an age
ha we do no need o de ine he edge o a ci y and how i changes o e ime, and ha ha ing simila sized a eas
makes he measu e mo e compa able ac oss ci ies. This ci cle also ensu es ha we don' compa e cen es wi h he
edges o ci ies. We do howe e adjus he cen al a ea o a ci y in he case whe e we ind i e s, coas s, o a bo de
(see Sec ion 3.1 o mo e de ails).
Second, by de ining he ci y cen e as he cen oid o he b igh es spo o nigh ligh , we only ha e one single
cen e o each own o ci y. This again is a s anda d assump ion made in monocen ic ci y models. Bu i is also a
es ic ion we impose o ou da a despi e po en ial cases o mul ipola ci ies. In he con ex o ou s udy his
es ic ion is likely o ha e a limi ed impo ance, especially gi en ha we also include many small and medium sized
owns, o which monocen ici y is e en mo e plausible (gi en hey may no be la ge enough o suppo wo
cen es). An os e al. (2016) p o ide e idence ha is consis en wi h monocen ic cen es o a ew la ge A ican
ci ies.
Thi d, we ake om he li e a u e ha he loca ion o he ci y cen e does no change much o e ime, and ha
ci ies ypically expand a ound his o ic ci y cen es. This is ue, o ins ance, i he ounda ion o a ci y ollowed
loca ional undamen als ha do no change o e ime. E idence can be seen in he many ci ies in he wo ld ha a e
s ill s uc u ed a ound Medie al o Ancien si es. A ecen empi ical pape making his assump ion is Ha a i (2020),
e idence o he su i al o Roman owns in Eu ope o wo cen u ies can be ound in Michaels and Rauch (2018).
The e is also ample e idence o pe sis ence o Roman oads (e.g. Bo asso e al., 2022 o De Benedic s e al., 2023).
Following his iew, a ci y wi h a be e wo king ci y cen e will de elop a la ge commu e zone a ound i o e
ime. This obse a ion is impo an o in o m ou iden i ica ion s a egy since i allows us o assume ha he his o ic
cen e and he mode n cen e coincide.
Fou h, we belie e ha he oad ne wo k is o c ucial impo ance o acili a e he agglome a ion e u ns,
wha e e hese e u ns may be. This could be iola ed i mos o he agglome a ion e u ns happen
o e whelmingly wi hin buildings and a e less dependen on connec ion in he cen e. E en in his case people
would ha e o ge o hese buildings, bu quan i a i ely he ue agglome a ion e ec s may be poo ly
app oxima ed by ou measu es. I could be iola ed i some s ic e sion o he ‘ undamen al law o oad
conges ion’(Du an on and Tu ne , 2011) applies, whe eby oad cons uc ion does no elie e conges ion and
hence does no lowe he anspo ime o human in e ac ions. This ou h poin is he main hypo hesis o
ou pape , and we p o ide empi ical e idence ha indeed he oad measu es we use in his pape ha e
p edic i e powe o popula ion g ow h.
BRANDILY and RAUCH
|
1239
Fi h, we ake i as gi en ha he e is pa h dependence o he oad layou . Once a layou has been es ablished,
i is cos ly o change i , and he g id om 1 yea o e laps s ongly wi h he g id om ano he yea . While he e a e
examples o adical changes in he layou o ci y maps (Haussmann's eno a ion o Pa is in he 19 h cen u y is a
amous example), he g ea majo i y o cases ha come o mind sugges indeed a endency o lock‐in.
1
Using a la ge
da a ex ac ion om Open S ee Map and co e ing a la ge sha e o wo ld's ci ies, Ba ing on‐Leigh and Milla d‐Ball
(2019) s udy wi hin ci ies oads layou and no ably conclude ha ‘s ee ‐ne wo k sp awl is a pa h‐dependen
p ocess. Because s ee s a e one o he mos pe manen ly de ining ea u es o ci ies’. In economics, Ba uah e al.
(2021) compa e colonial and mode n s ee maps o some A ican ci ies. They ind ha ‘ he spa ial s uc u es o
ci ies in Sub‐Saha an A ica a e s ongly in luenced by he ype o colonial ule expe ienced’and hey poin o ‘high
cos s o acqui ing new igh s o way in an al eady buil ‐up ci y’. Thei examples ha con as colonial and mode n
maps show s ong pe sis ence o oad layou . Compa ing 1960s maps o A ican ci ies o he same s ee da a we
use, hey no ably conclude ha ‘ oads ha we e in place 55 yea s ago gene ally emain in place oday’. She ze
e al. (2016) a gue ha Chicago zoning o dinances om 1923 ha e a la ge e ec on he spa ial dis ibu ion o
economic ac i i y in he ci y oday han geog aphy o anspo ne wo ks. I his poin was no ue, ci ies would be
able o sel ‐co ec , ou i s s ages would no be s a is ically signi ican , and ou measu es o he oad g id would be
poo p edic o s o u u e popula ion g ow h.
Six h, we conside i likely ha he op imal oad g id a a gi en ime depends on he a ailable anspo echnology.
The ideal g id in he age o walking and he age o he ho se di e s om he ideal g id in he age o he ca . Fo example,
he la e migh equi e b oade oads and migh a oid c ossings o oads ha de . I so, he age o a ci y in luences i s
ini ial g id, which in u n in luences i s g id la e on. This obse a ion helps us o de elop ou ins umen al a iable, which
uses a co ela e o he age o a ci y as an ins umen o he ini ial oad g id speci ica ion.
Taken oge he , hese six obse a ions imply ha ini ial di e ences o oad layou , e en i om a e y dis an
pas , can lead o di e en long un popula ion de elopmen s. Due o small di e ences in geog aphy and local
his o y, some owns may acciden ally s umble upon a mo e success ul layou han o he s. These ini ial di e ences
depend o some deg ee on he age o a ci y, which is obse able.
These poin s do no ell us wha a success ul layou should look like. We use he Open S ee Map da a se o
measu e se e al ea u es o wi hin‐ci y oads like o al leng h, numbe o nodes and in e sec ions, o ien a ion o each
segmen o oad (bea ing), e c. We hen use his measu emen ei he di ec ly o in o de o compu e ele an
cha ac e is ics o he layou . In pa icula , we build a simple measu e o densi y o oads in he ci y, measu ed by oad
kilome es o e a ea conside ed (which usually is he ci y cen e only). When compu ing his densi y a iable, we coun a
oad eco ded in he open s ee map da a wi h nlanes n imes. Such a simple coun o he densi y o oads is he mos
di ec way o measu e he densi y o he anspo ne wo k. On a e age, mo e oads imply ewe di e sions om he
sho es dis ance be ween wo poin s, and mo e oads help ease conges ion. The e could also be oo many oads, which
we don' belie e o be likely in ou sample. This s ikes us as a simple and s aigh o wa d measu e o oad layou in
ci ies. This measu e also has he ad an age o no aking any s ong s and on he co ec ela ionship be ween he
obse able ea u es and he ac ual a el speed o esiden s. The la e no being ob ious as ha ing oads wi h mul iple
lanes may be help ul in bigge ci ies, bu was e ul in smalle ones; c ossings may slow a ic down, bu enable use ul
a el combina ions, e c. Since pa o his ques ion is also empi ical, we will s udy he impac o di ec ly‐measu ed
ea u es like he numbe o in e sec ions on ci y g ow h (ei he con olling o o al oad leng h o no ).
Wha we y o cap u e is how well‐connec ed people and i ms in he ci y cen e a e. Road densi y migh be
misleading i all he oads a e concen a ed in a small subse o he ci y. In an ex eme example, a ci y migh ha e a
oad‐hea y pa king lo in he cen e, bu ew oads o he wise. Such a ci y would appea as ha ing a high oad
densi y by ou densi y measu e, bu would no enable as and s aigh o wa d connec ions. To add ess he
concen a ion o oads addi ionally, we compu e a a iable measu ing he concen a ion o oads. We build he
1
Examples o la ge and success ul ci ies whose cen al oad layou s s ill e lec hei Ancien o Medie al pas abound in Eu ope. Fo e idence o pa h
dependence o ci ies see Michaels and Rauch (2017). Fo e idence o pa h dependence o in as uc u e in A ica see Koco nik‐Mina e al. (2020).
1240
|
BRANDILY and RAUCH
e enness app oach as a measu e o how easy i is o access oad ne wo k om a andom poin in he ci y cen e.
This a iable cap u es how egula ly oads a e sp ead o e he a ea o measu emen . Facing a simila p oblem,
Donaldson and Ho nbeck (2016) c ea e 200 andom poin s wi hin US coun ies, calcula e he dis ance om each
poin o he nea es ail oad, and ake he a e age o hese nea es dis ances. Following his algo i hm, we c ea e
12 poin s in each ci cle, measu e he dis ance o he nea es oad o each o hese poin s, and compu e he a e age
nea es dis ance o he oad om hese 12 poin s.
2
Because his measu e dec eases as he oads a e less e enly
dis ibu ed ac oss space, we call i an ‘une enness index’. To illus a e his a iable, conside Figu e 1. This igu e
p esen s wo ci ies wi h s ylised oad layou s. Bo h ci ies ha e he same oad densi y in he cen e by cons uc ion.
The ci y on he igh has a sho e dis ance om a andom poin o he nea es oad.
In addi ion o hese wo main a iables o densi y and une enness, we compu e a hi d a iable measu ing he
o ien a ion o oads. Using he dis ibu ion o he (segmen s o ) s ee s’bea ing we compu e an He indahl index o
concen a ion. This simple and anspa en measu e akes i s highes alues when all s ee s un in ei he one o
wo di ec ions, and wo only (like a pe ec g id does). We call his measu e he ‘o ien a ion He indahl index’.We
desc ibe u he he cons uc ion o all measu es in 3.3.
3|DATA
We build a comple e and consis en sample o ci ies, combining sa elli e image y and geospa ial popula ion da a in
addi ion o o he da a sou ces. Fo each o hese ci ies, we obse e cen e's oad layou and measu e a se o
desc ip i e s a is ics. This sec ion b ie ly desc ibes da ase s and measu emen , all sou ces and p ocesses a e
de ailed u he in he da a appendices Aand B.
FIGURE 1 This igu e p esen s wo ci ies wi h s ylised oad layou s. Bo h ci ies ha e he same oad densi y in
he cen e by cons uc ion. The ci y on he igh has a sho e dis ance om a andom poin o he nea es oad, a
p ope y we call ‘e enness’.
2
Using 12 poin s is a low numbe in compa ison o 200. We do so o compu a ional con enience. As a jus i ica ion, we also eco d he mean dis ance
om he cen e o he ci cle o he nea es oad, which is e ec i ely he same exe cise wi h ha pa ame e se o one. The co ela ion be ween he
exe cise o 1 and 12 poin s is a ound 0.9. Gi en his la ge co ela ion, we belie e ha all such measu es would gi e ai ly simila nume ic es ima es.
BRANDILY and RAUCH
|
1241
3.1 |De ining ci ies: bounda ies and cen es
Ou main sou ce o he measu emen o ci ies loca ion and bounda ies is sa elli e image y o luminosi y a nigh .
This sou ce has he double ad an age o being comple e (i.e. i co e s he en i e con inen ) and consis en (i.e. all
he ci ies a e de ined in he same way). Fi s , we iden i y li a eas om NOAA's DMSP‐OLS sa elli e eco d (c .
Suppo ing in o ma ion: Appendix A) by keeping pixels ha emi ligh a leas wice o e he 5 yea ly obse a ions
om 2008 o 2012. Con iguous li a eas a e hen agg ega ed using a GIS so wa e o c ea e unions. Each union
ep esen s a ci y, and we conside i s oo p in as he ci y bounda ies. Figu e 2illus a es ou sample. As he Figu e
shows, we conside ci ies in almos all coun ies o Sub‐Saha an A ica, wi h heal hy spa ial a ia ion. As a
FIGURE 2 We iden i y 2,779 ci ies om 40 coun ies.
1242
|
BRANDILY and RAUCH
obus ness check, we also ely on he ‘U ban Cen e Da abase’(UCDB) om he EU Commission JRC o delimi a
second sample o A ican ci ies. We discuss he de ini ion o his sample in Suppo ing in o ma ion: (Appendix D).
Ou ocusisonci ycen es,whe eweassumeagglome a ion e ec s o ake place. Because, o he bes o ou
knowledge, he e is no da a se p o iding p ecise coo dina es o ci ies’cen e we a he ollow a sys ema ic da a d i en
de ini ion. Fo each ci y, we ake as ‘cen e’ he cen oid o he b igh es a ea (i.e., whe e pixels ha e he highes ligh alue
obse ed in he ci y). In p ac ice, his b igh es a ea some imes is a single ligh pixel, bu some imes is a la ge a ea o
con iguous pixels, in which case we compu e i s cen oid. In cases o mul iple b igh es a eas o equal b igh ness in a ci y
we choose he la ges one. This me hod iden i ies a unique ci y cen e o each ci y.
Two easons a leas suppo his de ini ion o a ci y cen e: i s , because nigh ligh s glow, i is o en he
case ha b igh es a ea is close o he geog aphic cen oid o he ci y. Second, a he same ime, nigh ligh s a e
also known o be linked wi h economic ac i i y, e en a he local le el. Figu e 3illus a es he de i ed ci y
cen es o wo small owns. The igu e also shows he ci y ex en , compu ed om ligh s da a, and he oad
ne wo k o hese owns. Impo an ly, we chose o use ea ly images o nigh ligh s (1994‐1996) o de ine ci y
cen e.Thisis o bo hempi icaland heo e ical easons. Empi ically, A ican ci ies emi ed less ligh in he
ea ly pe iod which acili a es he iden i ica ion o he main his o ical cen e. Theo e ically we a e in e es ed in
he lock‐in o his o ical cen e and he e o e wan o iden i y whe e his cen e was as ea ly as possible. This
also educes he sample o ci ies ha we e impo an enough in he mid‐1990 o emi some ligh . This selec ion
mainly excludes e y small owns and is cohe en wi h ou ins umen al a iable s a egies (c . Sec ion 4.2).
In Suppo ing in o ma ion: (Appendix B.1) we discuss al e na i e measu es o he ci y cen e and show ha
hey yield e y simila ou comes.
FIGURE 3 Example o ci y ex en de ini ion. Ci y ex en and cen e a e de ined using nigh ligh luminosi y (in
la e—2008–2012—and ea ly—1994–1996—pe iod, espec i ely).
BRANDILY and RAUCH
|
1243
TABLE 2 Desc ip i e s a is ics o he main a iables o 1850 ci ies o he main sample.
NMean sd Min Median Max
Whole Ci y
To al Road Leng h (km) 1850 243 878 0 73 28621
Me es o Road pe Sq. Km 1850 2491 2177 0 1850 17,569
A ea (Sq. Km.) 1850 94.6 279 2 40 7408
Ci y Cen e (1 km adius)
To al Road Leng h (km) 1850 19.6 16.1 0 15 87
Me es o Road pe Sq. Km 1850 6403 5160 2 5205 27,703
A ea (Sq. Km.) 1850 3.10 0.21 1 3 3
Dis . Cen e o closes Road (m) 1850 133 167 0 62 975
Road O ien a ion He indahl 1850 0.19 0.19 0 0 1
A e age Dis . Poin s o Road (m) 1850 189 155 10 150 1078
Dis ances om Ci y Cen e
Dis ance o Daka (km) 1850 4066 2071 78 3487 8272
Dis ance o Djibou i (km) 1850 3583 1451 34 3773 6507
Dis ance o Cai o (km) 1850 4007 1046 918 3875 6738
Dis ance o Cape Town (km) 1850 4338 1311 628 4838 7168
Dis ance o he closes (km) 1850 2065 830 34 2135 3678
Dis ance o he sea (km) 1850 451 350 0 389 1711
TABLE 3 Baseline model and OLS es ima ions. d cen e measu es oad densi y in he ci y cen e. The se o
con ols include a iables measu ing uggedness, ele a ion, minimum, maximum and a e age empe a u e,
p ecipi a ion, dis ance o he lakes and he coas and mala ia isk. Robus s anda d e o in pa en heses. S a s
deno e signi icance a 10 (*), 5 (**) and 1 (***) pe cen .
(1) (2) (3) (4) (5) (6)
Δpop Δpop Δpop Δpop Δpop Δpop
2000–2015 2000–2015 2000–2015 2000–2015 2000–2015 2000–2015
d cen e 0.0014*** 0.0012*** 0.0012*** 0.0012*** 0.0011*** 0.0011***
(0.0003) (0.0003) (0.0003) (0.0003) (0.0002) (0.0002)
ln pop 2000 −0.0060*** −0.0069*** −0.0086*** −0.0081*** −0.0068*** −0.0068***
(0.0015) (0.0016) (0.0017) (0.0018) (0.0014) (0.0014)
ln dis sea 0.0006 −0.0005 0.0000 −0.0004 −0.0004 −0.0007
(0.0005) (0.0007) (0.0013) (0.0017) (0.0015) (0.0014)
Obse a ions 1,450 1,450 1,450 1,450 1,856 1,890
Coun y FE Yes Yes
P o ince FE Yes Yes Yes
Con ols Yes Yes Yes Yes
Nige ia included Yes Yes
Madagasca included Yes
1250
|
BRANDILY and RAUCH
4.2 |Ins umen al a iable s a egies
One conce n wi h he empi ical s a egy based on Equa ion 1is he po en ial ole o e e se causali y. I ci ies ha
g ow as e also build mo e oad o imp o e he OSM da a se as e , hen ou OLS model would no iden i y he
causali y o he oad ne wo k on popula ion g ow h. To ci cum en his conce n, we implemen wo ela ed
ins umen al a iable (IV) s a egies using a ci y's age as an ins umen o i s ci y‐cen e oad ne wo k. As discussed
in Sec ion 2, he cu en layou o a ci y‐cen e is likely o be a di ec p oduc o he ini ial, his o ic layou . In he IV
eg essions we con ol o ini ial popula ion densi y i sel in hese eg essions. We expec o see some oad ne wo k
ea u es changing ac oss ci ies o di e en ages. Fo example, a ci y ounded be o e he in en ion and adop ion o
ca s would ha e a highe densi y sui able o olde echnologies. We expec mo e mode n ci ies, buil in he e a o
inc easing adop ion o ca s, ha ing bigge oads and also smalle oad densi ies.
To measu e he ounda ion age o a ci y, we ely on he A icapolis (2019) da a se , ha eco ds his o ic
popula ion o a la ge numbe o ci ies in Sub‐Saha an A ica. The oldes yea in ha da a se is 1950, and ou
ins umen consis s o a simple dummy a iable, indica ing hose ci ies ha exis ed in ha yea acco ding o his
da a se (i.e. ha had 1 o mo e inhabi an s in 1950). One downside o his ‘age da a se ’is ha i is only a ailable
o a subse o ou ci ies co e ed by A icapolis and hus educes ou sample (see Suppo ing in o ma ion: Appendix
B.3 o mo e de ails). Howe e , we don' see any eason ha his selec ion is ela ed o ou main a iables o
in e es , and hence likely o in oduce selec ion bias.
The e could be conce ns ha he age o a ci y migh in luence ecen popula ion g ow h h ough channels
o he han he oad layou in he ci y cen e. To mi iga e his pa ially, we measu e he age c udely as an indica o
a iable. To add ess his conce n mo e ho oughly, we also cons uc a second, ela ed ins umen , measu ing he
dis ances om he ci ies o Cape Town, Cai o, Daka and Djibou i. These ou ci ies ep esen connec ion nodes o
he wo main colonial powe s o he con inen . I a signi ican pa o colonial expansion was conduc ed ia hese
po s, dis ance o hese ou ci ies should in luence he yea o c ea ion o mode n ci ies on he con inen . In
hese speci ica ions we con ol o he dis ance o he sea and again o ini ial popula ion densi y. The con ol o
he dis ance o he sea ensu es ha i is no he dis ance o po s in gene al ha d i es his ins umen , bu only
dis ance o speci ic his o ic po s. As a gued o he i s ins umen : i he age o a ci y co ela es wi h he ini ial
oad layou , hen he dis ance o hese ou ci ies will also co ela e wi h he ini ial oad layou . Using he dis ance
om ci ies o each o hese ou poin s is a simplis ic ision o his o y ha cap u es enough in o ma ion ha i
migh unc ion as an ins umen .
U banisa ion in A ica was acili a ed o a la ge deg ee by he colonial powe s, especially he B i ish and he
F ench, he wo main pa ies in he sc amble o A ica. As la e as 1880, abou 80 pe cen o A ica we e uled
by A ica's own kings and queens. Colonial powe s we e ound nea he coas , wi h a B i ish s onghold a ound
he po o CapeTown, and F ench holdings a ound Daka (Boahen, 1985). The B i ish in asion o Cai o o 1882
es ablished a second s onghold on he con inen o he B i ish ha las ed well in o he 20 h cen u y (Hou ani
e al., 2004). Soon i became he ision o he B i ish o connec hese wo cen es, and build he ‘Cape o Cai o’
Railway line. His o ically, hese wo ci ies played an impo an pa as connec ion poin s o B i ish colonial
A ica, and many expedi ions, a els and ade o igina ed in one o he o he . F ance's in luence on he eas e n
coas o A ica was ini ia ed by ea ies wi h he ule s o wha is oday Djibou i om 1883, and he c ea ion o
ou pos s in mode n day Daka . Soon he F ench ambi ion became o es ablish an Eas ‐Wes link be ween i s
po s in Daka and Djibou i. F ench and B i ish expansions clashed nea he own o Fashoda, in he Fashoda
inciden o 1898 (Ba es 1984). In schoolbooks, his episode in his o y is some imes illus a ed wi h maps ha
show he ci ies o Cai o, Cape Town, Daka and Djibou i, wi h ou a ows mee ing nea Fashoda. This
simpli ied schoolbook iew o colonial expansion in A ica is he model we ollow in he cons uc ion o ou
ins umen s, aking he dis ance o hese colonial o igin po s as ins umen o he age o a own, and
consequen ly i s ini ial oad g id.
BRANDILY and RAUCH
|
1251
4.2.1 |Fi s ‐s age eg essions
Table 4shows he co ela ion be ween ou main igh hand side a iables, which a e he measu es o oad
densi y and une enness, and he dis ance o he ou ins umen ci ies. This is e ec i ely a i s s age
eg ession. As expec ed, we obse e a signi ican nega i e o no ela ionship be ween he dis ance o he
po s and oad densi y in he ci y cen e in Columns (1) ‐(4), which could e lec ha newe ci ies ha
we e designed wi h mode n anspo a ion in mind ely on a less dense oad g id. I is nega i e in he case o
h ee o ou ou ci ies and insigni ican ly di e en om ze o a i e pe cen le el o s a is ical signi icance in
hecaseo CapeTown.
6
When we join ly include all ou ci ies in one eg ession and add a ious con ol
a iablesasincolumn(5)wecon inue oobse enega i e coe icien s, ypically a a high le el o s a is ical
signi icance. Coe icien s become weake when we include coun y ixed e ec s, as in column (6). This seems
plausible o us since we expec he dis ance o hese colonial po s o ma e less wi hin coun ies han o
he longe dis ances ac oss he whole con inen . Howe e , i c ea es a weak ins umen p oblem o esul s
wi h coun y ixed e ec s. Hence, we show he main esul s la e wi h and wi hou coun y ixed e ec s. In
ou IV s a egy we p e e o use he ou dis ances as sepa a e a iables o simply using he minimum
dis ance gi en ha B i ish and F ench colonial legacies could ha e in luenced ci ies in di e en ways (Ba uah
e al., 2021). In columns (7) and (8) we epea he es ima ion om columns (5) and (6) bu using une enness
on he le ‐hand side, a he han densi y in jus he ci y cen e. Again, we ind a s ong ela ionship wi hou
coun y ixed e ec s and a much weake one when we include coun y ixed e ec s.
Incolumns(9)and(10)wep o idemo edi ec e idence ha dis ance o hese his o ic po s is indeed
co ela ed wi h he age o a ci y. Fo his analysis we use he da a om he ‘A icapolis’p ojec . Keeping only
ci ies in ou da a se ha a e also in he A icapolis educes he sample o 1,107 obse a ions. We hen
compu e an indica o o ci ies ha , acco ding o A icapolis, did no ha e any popula ion in 1950. As
columns (9) and (10) show, he e is a signi ican ela ionship be ween hese dis ances and he age o a ci y.
Once we include ou con ol a iables and coun y ixed e ec s, he ela ionship is posi i e, which con i ms
ha ci ies ha a e u he away om hese po s a e mo e likely o be new. Columns (11) and (12) a e hen
i s s age eg essions when using he age indica o o ci ies as an ins umen . Finally, in Columns (11) and
(12) we show ha his simple indica o a iable o ci ies ha a e olde han 1950 co ela es s ongly
signi ican ly wi h he densi y o oads in he ci y cen e and une enness in he expec ed way.
4.2.2 |Exclusion es ic ion
These IV s a egies could unco e he causal ela ionship be ween he oad ne wo k and popula ion g ow h
only unde he exclusion es ic ion ha he age o a ci y in luences i s ecen popula ion
g ow h only h ough i s impac on he oad layou . In ou p e e ed 2SLS es ima ion we always include
ini ial popula ion size and dis ance o he sea as well as coun y (o p o ince) ixed e ec s. This ensu es ha
only he e ec o a ci y's age h ough hese wo channels ha is ime‐in a ian bu coun y‐speci ic is aken
in o accoun .
I ,aswea gued,ageandloca iono ci iesa e ela ed, hen geog aphic ac o s could po en ially in luence
ci y g ow h h ough o he channels han jus he ci y‐cen e oad layou . The inclusion o he dis ance o he
coas does no only add ess he conce n ha he ou po s (Cai o, CapeTown, Daka and Djibou i) s ill play a
la ge oleas anspo a ionhubs oday,bu i alsopicksupo he in luences hecoas mayha eon helocal
economy. We also no e ha he e a e la ge po s a ailable along he A ican coas , apa om hese ou
6
On a e age, he coe icien in columns (1)‐(4) is −0.87, pe haps close o he dis ance coe icien o −1 some imes es ima ed o dis ance coe icien s
(Rauch, 2016).
1252
|
BRANDILY and RAUCH
TABLE 4 Fi s s age eg essions. new is an indica o o ci ies ounded a e 1950. Va iables ln dis e e o log dis ance o ei he Cai o, Cape Town, Daka , Djibou i, o
he sea. d cen e measu es oad densi y in he ci y cen e; une en is he une enness index. The se o con ols consis s o a iables measu ing uggedness, ele a ion,
minimum empe a u e, maximum empe a u e, a e age empe a u e, p ecipi a ion, dis ance o he big lakes, dis ance o he coas and clima ic condi ions o mala ia. S a s
deno e signi icance a 10 (*), 5 (**) and 1 (***) pe cen .
(1) (2) (3) (4) (5) (6) (7) (8) (9) (10) (11) (12)
d cen e d cen e d cen e d cen e d cen e d cen e une en une en new new d cen e une en
new −2.284*** 54.42***
(0.356) (9.682)
ln dis sea −0.0318 −0.121* −0.198*** −0.0783 −0.0833 −0.0894 5.369 10.79** 0.0279*** 0.0188 0.0999 8.439*
(0.0734) (0.0726) (0.0760) (0.0715) (0.133) (0.143) (4.029) (4.677) (0.00829) (0.0175) (0.166) (4.947)
ln pop 2000 1.901*** 1.886*** 1.821*** 1.892*** 1.979*** 1.955*** −44.48*** −43.95*** −0.135*** −0.146*** 1.534*** −31.36***
(0.0937) (0.0925) (0.0928) (0.0947) (0.0975) (0.107) (2.862) (2.856) (0.00991) (0.0104) (0.149) (3.848)
ln dis Cap 0.365 −5.494*** −3.271* 115.5*** 32.13 0.0800 0.563***
(0.303) (0.893) (1.824) (33.76) (59.52) (0.0986) (0.212)
ln dis Dhak −0.399** −2.651*** −1.431** 56.9*** 17.08 0.113** 0.0226
(0.164) (0.419) (0.620) (17.41) (20.25) (0.0447) (0.0701)
ln dis Dji −0.738*** −0.545 0.383 −9.919 −43.01 −0.0847** 0.0552
(0.225) (0.545) (1.637) (10.85) (27.23) (0.0360) (0.0516)
ln dis Cai −2.826*** −8.564*** −7.877*** 128.9*** 141.6** 0.203* 0.673**
(0.548) (1.164) (2.658) (31.51) (69.81) (0.114) (0.264)
Obse a ions 1450 1450 1450 1450 1450 1450 1450 1450 1107 1107 1107 1107
Con ols Yes Yes Yes Yes Yes Yes Yes
Coun y FE Yes Yes Yes Yes Yes
BRANDILY and RAUCH
|
1253
ci ies.
7
In addi ion o seapo s, ai po s now play a subs an ial oad in connec ing A ica, which u he
educes he impo ance o hese ou his o ic po s in con empo a y ade. Also, o add ess conce n ela ing
o he loca ion o ci ies wi hin he con inen we also p o ide obus ness speci ica ions in which we con ol
o ma ke access measu es o each ci y in ou da a se . Including such measu es ba ely change he
magni ude and signi icance o ou coe icien s o in e es .
Ano he conce n o ou ins umen al a iable s a egies is ha o he colonial legacies could equally
depend on age. Fo example, he e may be physical in as uc u e o he han oads whose age
depends on hese dis ances in a simila way, and ha ha e a simila e ec on u u e de elopmen . We
can' dismiss his conce n en i ely bu no e some limi s o his conce n. Fi s , as we a gued p e iously, we
hink ha he oad layou is pa icula ly p one o lock‐in, compa ed o o he ypes o in as uc u e.
Addi ionally, oad cons uc ion and ci y planning we e o cen al impo ance o all po en ial colonial
in es men s. Second, building econs uc ion a es in A ica a e e y high. Michaels e al. (2021)es ima ean
annual eplacemen a e o i e pe cen o housing in Tanzania, Hende son e al. (2016) ind a eplacemen
a e o 3.6 pe cen in Kenya. Gi en hese high eplacemen a es, mos houses and mos o he physical
in as uc u e has been ebuil se e al imes since he ea ly 20 h cen u y, and ha e much ime o con e ge
om an ini ial s eady s a e o a new one. Pa h dependen beha iou o in as uc u e quali y is a less likely
in such a apidly changing en i onmen . O cou se, also he physical cha ac e is ics ( he ‘quali y’and size) o
he oads we e changed and upda ed mul iple imes in he g ea majo i y o ou owns and ci ies. I is,
howe e , he layou and loca ion o oads, he plan, ha we hink is ha de o change and ha pe sis s e en i
oads a e econs uc ed many imes. Thi d, a la ge sha e o public in as uc u es in Sub‐Saha an A ica
happens o be concen a ed in capi al ci ies (Bekke & The bo n, 2012), which a e a small sha e o he sample,
andwhicha eexcluded om heanalysisinoneo ou obus ness checks. I public policy p og ammes a e
less common ou side hese ci ies, hen he e a e ewe channels h ough which adminis a i e legacies could
mani es . No e also ha , in ou baseline speci ica ions we always include a coun y (o p o ince) ixed e ec .
This also add esses o he conce ns ela ed o ins i u ional legacies. Any ins i u ional se ing ha is se (and
in a ian ) a he coun y (p o ince) le el would be abso bed by he ixed e ec s. Finally, we a e eassu ed by
he ac ha he key esul s om he IV speci ica ions a e obus o many al e na i e speci ica ions o sample
changes.
5|RESULTS
5.1 |Resul s om OLS es ima ion
To es ima e he impac o a ious ea u es o he ci y‐cen e oad ne wo k on ci ies’g ow h and p ospe i y, we i s
es ima e Equa ion 1. In his speci ica ion, we always include a coun y ixed e ec as well as (log) ini ial popula ion
densi y and minimum dis ance o he sea (c . Sec ion 5.1 o mo e de ails). We in es iga e he ela ionship be ween
ou simple a iables o oad ne wo k and popula ion g ow h.
Table 5 epo s he esul s. Column (1) shows ha a g ea e oad densi y in he ci y cen e co ela es posi i ely
and s ongly s a is ically signi ican ly wi h popula ion g ow h in he pe iod 2000–2015. The measu e o oad
densi y in he ci y cen e ‘d cen e’is measu ed in km/km
2
and has a mean o a ound 6. Inc easing he oad densi y
in he ci y cen e by 1 km pe km
2
, is associa ed wi h mo e han one‐ en h o a pe cen o highe popula ion g ow h
annually. Accumula ed o e 15 yea s, his adds up o a o al popula ion ha is mo e han 1.55 pe cen la ge
7
The busies po s o A ica a e (in o de descending in size acco ding o o al ca go olume): Richa d's Bay, Saldanha Bay, Alexand ia, Damie a, Eas Po ,
Apapa, Casablanca, Skikda, Poin e Noi e, Mombasa, Abidjan, Tange , Bejaia, Jo Las a and Tin Can Island (AAPA ‐Ame ican Associa ion o Po
Au ho i ies, 2015).
1254
|
BRANDILY and RAUCH
TABLE 5 Resul s om he baseline OLS eg ession model. d cen e measu es oad densi y in he ci y cen e; n cen e is he numbe o oads’nodes in he ci y cen e;
une en is he une enness index; o ien a ion H is he o ien a ion He indahl index. Robus s anda d e o . S a s deno e signi icance a 10 (*), 5 (**) and 1 (***) pe cen .
(1) (2) (3) (4) (5) (6) (7) (8) (9) (10)
Δpop Δpop Δpop Δpop Δpop Δpop Δpop Δpop Δpop Δpop
2000–2015 2000–2015 2000–2015 2000–2015 2000–2015 2000–2015 2000–2015 2000–2015 2000–2015 2000–2015
d cen e 0.0012*** 0.0013** 0.0009*** 0.0012*** 0.0007 0.0012 0.0011*
(0.0003) (0.0006) (0.0003) (0.0003) (0.0007) (0.0008) (0.0006)
n cen e 0.00003*** −0.00000 0.0000 0.0000 0.0000
(0.0000) (0.00002) (0.0000) (0.0000) (0.0000)
une en −0.00003*** −0.00002** −0.00002** −0.00001 −0.00001*
(0.0000) (0.0000) (0.0000) (0.0000) (0.0000)
o ien a ion H −0.0083 −0.0002 0.0037 −0.0004 0.0037
(0.0052) (0.0048) (0.0052) (0.0051) (0.0052)
ln pop 2000 −0.0069*** −0.0064*** −0.0069*** −0.0061*** −0.0072*** −0.0050*** −0.00691*** −0.00712*** −0.00691*** −0.00616***
(0.00163) (0.00162) (0.00160) (0.00152) (0.00167) (0.0014) (0.0017) (0.0016) (0.0016) (0.0015)
ln dis sea −0.0005 −0.0005 −0.0005 −0.0007 −0.0006 −0.0006 −0.0005 −0.0006 0.0008 0.0006
(0.0007) (0.0007) (0.0007) (0.0007) (0.0007) (0.0007) (0.0007) (0.0007) (0.0014) (0.0005)
Obse a ions 1450 1450 1450 1450 1450 1450 1450 1450 1450 1450
Coun y FE Yes Yes Yes Yes Yes Yes Yes Yes
P o ince FE Yes
BRANDILY and RAUCH
|
1255
because o his addi ional oad leng h. Below we mul iply his e ec wi h ou p e e ed es ima e. In column (2) we
show he es ima ed impac o he numbe o nodes (i.e. any in e sec ion wi h one oad o mo e) wi hin he ci y‐
cen e also co ela es posi i ely wi h ci y g ow h bu column (3) shows ha his e ec disappea s when he densi y
o oads is conside ed. In o he wo ds, he numbe o nodes co ela es wi h ci y g ow h inso a as i is a di ec
p oduc o he o al amoun o oad. In columns (4) o (7) we u n o ou wo indices o spa ial o ganisa ion, he
une enness and o ien a ion He indahl indices, espec i ely. We desc ibe hese indices in Sec ion 3.3. The main
esul in column (4) indica es ha as he e enness o he oad g id ( espec i ely, he index) inc eases (dec eases) so
does he popula ion g ow h o e 2000–2015. In column (5) we also add ci y‐cen e oad densi y as an independen
a iable. The conce n he e is ha wo ci ies wi h iden ical oad densi y as measu ed by ou d cen e measu e may
ne e heless ha e qui e di e en lows in he cen e. Simila o he esul in column (4), he nega i e sign associa ed
o he une enness index indica es ha la ge a e age dis ances o he nea es oad ha m popula ion g ow h,
holding oad densi y cons an , as expec ed.
8
In column (6) we es ima e he co ela ion be ween he o ien a ion
He indahl index and popula ion g ow h: a nega i e co ela ion appea s ha is no s a is ically signi ican a
con en ional le els. This coe icien sh inks and emains insigni ican in column (7), when oad densi y is con olled
o . As explained in Sec ion 3.3, bo h he une enness and he o ien a ion He indahl indices e en ually measu e
di e en hings, al hough hey co ela e qui e s ongly (0.42). One could o ins ance hink o a ci y‐cen e whe e all
oads un no h o sou h (highes He indahl index) ye a e all concen a ed in one speci ic neighbou hood (highes
une enness); o hese wo o ces, he une enness seems a dominan ac o o ci y g ow h.
In column (8) we es ima e a model including all he oad ne wo k ea u es a once (i.e. s ee densi y, numbe o
nodes, he une enness index and he o ien a ion He indahl index). We d aw wo main conclusions om his
exe cise: i s , he only ea u e whose impac s emain signi ican is he une enness index. Such esul s indica e ha
he spa ial dis ibu ion o he ne wo k ma e s e en a e he o al amoun o oads and nodes, as well as hei
o e all bea ing concen a ion a e conside ed. Second, al hough he signi icance o he oad densi y collapses, he
poin es ima e emains qui e simila in mos speci ica ions. We ake his esul as an indica ion o he obus ness o
ou esul and o he impo ance o oad densi y. No e also ha ci y‐cen e oad densi y di ec ly co ela es wi h all
o he independen a iables, and i is he e o e no su p ising ha he s anda d e o s inc ease in column (8)
compa ed o column (1). In column (9) and (10) we subs i u e ei he (wi hin‐coun y) p o ince ixed e ec s o no
ixed e ec , espec i ely, o he coun y ixed e ec . Column (9) is a s a egy o educe he po en ial measu emen
bias (Seidel, 2023; c . Sec ion 3.3) while column (10) mi o s subsequen speci ica ion in he emainde o his pape .
In bo h cases, he e ec o he une enness index is educed and emains signi ican only a he 10% le el in column
(10). The coe icien o oad densi y emains close o o he es ima es in e ms o magni ude, bu is weake in e ms
o s a is ical signi icance. Wha Table 5shows is ha o he a iables es ed, only d cen e and une en show a
signi ican and somewha obus co ela ion wi h popula ion g ow h, while he o he a iables don' .
9
We con inue
he mo e de ailed in es iga ion wi h hese wo cen al a iables.
5.2 |Resul s om he 2SLS es ima ion
We nex u n o he IV es ima es co esponding o he OLS esul s de i ed so a . Fi s , we use an indica o a iable
o new ci ies as an ins umen o he oad layou in he ci y cen e in Table 6. This a iable indica es ci ies ha
we e ounded a e 1950 acco ding o he A icapolis da a se (c . Sec ion 4.2 o mo e de ails). Gi en ha no all
8
The une enness index is exp essed in me es so he coe icien in column (4) can be ead as: wi hin a gi en coun y, ci ies whe e he e enness is 1 m
smalle also had a popula ion g ow h ha is 3 × 10
−5
highe (gi en ini ial ci y popula ion and i s dis ance o he sea coas ). Accumula ed o e 15 yea s his
implies a popula ion ha is mo e han 0.045 pe cen la ge .
9
Conside he epo ed coe icien s in Column (10) o Table 5. In combina ion wi h he s anda d de ia ions epo ed in Table 2we can see ha an inc ease
by one s anda d de ia ion has an e ec o 0.018 o d cen e, o 0.0017 o une en, bu only 0.00007 o o ien a ion on he ou come.
1256
|
BRANDILY and RAUCH
TABLE 6 IV using age as ins umen . d cen e measu es oad densi y in he ci y cen e; une en is he une enness index. The i s s age F s a is ic epo s he F es o he
excluded ins umen , which is an indica o o ci ies ha we e ounded a e 1950. The se o con ols consis s o a iables measu ing uggedness, ele a ion, minimum
empe a u e, maximum empe a u e, a e age empe a u e, p ecipi a ion, dis ance o he big lakes, dis ance o he coas and clima ic condi ions o mala ia. S a s deno e
signi icance a 10 (*), 5 (**) and 1 (***) pe cen .
(1) (2) (3) (4) (5) (6) (7) (8)
Δpop Δpop Δpop Δpop Δpop Δpop Δpop Δpop
2000–2015 2000–2015 2000–2015 2000–2015 2000–2015 2000–2015 2000–2015 2000–2015
d cen e 0.00574*** 0.00653*** 0.00920** 0.00621***
(0.00217) (0.00176) (0.00391) (0.00159)
une en −0.0002*** −0.0003*** −0.0002*** −0.0003***
(0.0000) (0.0000) (0.0000) (0.0000)
ln dis sea 0.00110 0.00030 −0.00017 −0.00075 0.0019 0.0022 0.00016 0.00298
(0.00071) (0.00075) (0.00098) (0.0011) (0.0017) (0.0015) (0.0017) (0.00195)
ln pop 2000 −0.0153*** −0.0135*** −0.0169*** −0.0159*** −0.0245*** −0.0160*** −0.0187*** −0.0174***
(0.00483) (0.00410) (0.00433) (0.00392) (0.0084) (0.0040) (0.00402) (0.00361)
Obse a ions 1107 1107 1107 1107 1107 1107 1107 1107
Coun y FE Yes Yes Yes Yes
Con ols Yes Yes Yes Yes
Fi s s age F 18.7 15.8 45.2 26.8 8.3 14.6 42.1 30.5
BRANDILY and RAUCH
|
1257
ou owns and ci ies ea u e in he A icapolis da a se , his educes he size o ou sample o 1,107. This ins umen
ollows he same logic as he po dis ances, which is ha olde ci ies a e mo e likely o be s uck wi h subop imal
g ids; bu i measu es he age o a ci y di ec ly. The ad an age o his s a egy is ha we ge a s ong i s s age,
wi h he excep ion o he speci ica ion in column (5). All coe icien s on oad densi y a e posi i e and s a is ically
signi ican , while magni udes on une enness a e nega i e and s a is ically signi ican . Magni udes a e la ge han in
he OLS es ima e, which migh be explained by a e e se causali y p oblem in he OLS eg ession, and some scope
o measu emen e o . Quali a i ely, his able suppo s he OLS indings on densi y and une enness and shows
ha bo h a e s ongly associa ed wi h popula ion g ow h in ecen yea s. As expec ed, a ci y wi h a dense oad g id
has a highe popula ion g ow h, while a mo e une en oad g id leads o less popula ion g ow h. This con i ms ha a
ci y wi h a egula s ee g id consis ing o oads ha a e e enly dis ibu ed has a ou able condi ions o g ow.
Table 7p esen s esul s co esponding o he OLS esul s in Table 5, bu his ime es ima ed as 2SLS using
dis ances o colonial po s as ins umen s. Coe icien s e ain he expec ed signs, posi i e o oad densi y and
nega i e o une enness. This able su e s om weak i s s ages in columns wi h coun y ixed e ec s. We
s ill epo he coe icien o compa ison. We no e ha coe icien s a e simila in magni ude o he ones
es ima ed using he o he ins umen . O e all, bo h IV s a egies seem o con i m he quali a i e esul s om
ou baseline OLS es ima ion. The impac o ci y‐cen e oad leng h densi y emains posi i e and signi ican wi h
he 2SLS es ima ions. Simila ly, he coe icien associa ed o he une enness index indica es ha be e sp ead
oads ha e a posi i e impac on ci y g ow h o e he 2000‐2015 pe iod and 2SLS only con i ms his conclusion.
Fo bo h measu es and bo h ins umen s, he poin es ima e o he coe icien is g ea e in 2SLS han in OLS,
sugges ing ha he po en ial bias in OLS a enua es he ue e ec o oads. We acknowledge han hese wo
IVs ha e limi a ions, bu we ind i wo h epo ing ha hey yield compa able conclusions ega ding bo h he
di ec ion o he e ec o oad ne wo k and he di ec ion o he po en ial bias in OLS. Coe icien s again ha e
he expec ed signs as in p e ious ables.
In e ms o magni ude, a coe icien o 0.005 on oad densi y implies ha i a ci y inc eases i s cen al oad
densi yby1kmpe uni ci clea ea,i s o alpopula ion g ows by an addi ional 0.5 pe cen pe yea . We can
compa e his wi h he mean o 6.4 km in ou sample, and a s anda d de ia ion o 5.2. Accumula ed o e 15
yea s, ou p e e ed es ima e hus leads o a popula ion ha is7pe cen la ge inaci y ha hasanaddi ional
kilome e o oad in he cen e. On une enness, a coe icien o −0.0002 sugges s ha inc easing he
une enness measu e by 100 educes popula ion g ow h by 0.02 pe cen pe yea . Gi en ha he mean
une enness is a ound 200, his is a less quan i a i ely s ong e ec han he one we epo conce ning oad
densi y.
6|ROBUSTNESS CHECKS
In his subsec ion, we e i y ha ou esul s emain obus when we a y some o he choices guiding ou main
speci ica ion and sample selec ion. These esul s a e shown in Table 8. The 6 columns epo s ou 6 main
es ima ions: he main independen a iable is oad densi y in columns (1) o (3) and une enness in columns (4) o (6);
o bo h independen a iables we p esen he esul s associa ed o ou basic OLS (columns (1) and (4),
espec i ely), ou dis ance IV (columns (3) and (6)) and ou age IV (columns (2) and (5)). Panel A displays he baseline
esul s (i.e., a speci ica ion ha uses he ull se o con ol a iables including log ini ial popula ion and dis ance o
he sea as well as a coun y ixed e ec ). Gi en he coun y ixed e ec s, he dis ance IV esul s ha e low s a is ical
powe h oughou his able, as in p e ious esul s.
Panel B epea s he main exe cise bu excludes capi al ci ies. I could be ha capi al ci ies ollow unique
de elopmen pa hs wi hin hei coun ies, gi en hei be e ins i u ions, public goods p o isions, and in e na ional
connec ions (Bekke & The bo n, 2012). As Panel B shows, excluding capi al ci ies does no change he coe icien s
in ei he magni ude o s a is ical signi icance in any meaning ul way.
1258
|
BRANDILY and RAUCH
TABLE 7 IV using dis ances as ins umen . d cen e measu es oad densi y in he ci y cen e, une en is he une enness index. The Fi s s age F es s a is ic epo s he
mul i a ia e F es o he excluded ins umen s. The se o con ols consis s o a iables measu ing uggedness, ele a ion, minimum empe a u e, maximum empe a u e,
a e age empe a u e, p ecipi a ion, dis ance o he big lakes, dis ance o he coas and clima ic condi ions o mala ia. S a s deno e signi icance a 10 (*), 5 (**) and 1 (***)
pe cen .
(1) (2) (3) (4) (5) (6) (7) (8)
Δpop Δpop Δpop Δpop Δpop Δpop Δpop Δpop
2000–2015 2000–2015 2000–2015 2000–2015 2000–2015 2000–2015 2000–2015 2000–2015
d cen e 0.00242*** 0.00513** 0.00333*** 0.00358
(0.000828) (0.00253) (0.00104) (0.00237)
une en −0.0000 −0.00028** −0.0000* −0.000163*
(0.0000) (0.0001) (0.0000) (0.0000)
ln dis sea 0.000694 0.000326 −0.000256 −0.00120 0.000962 0.00103 0.000109 0.00154
(0.000486) (0.000530) (0.000671) (0.000997) (0.00122) (0.00127) (0.00131) (0.00168)
ln pop 2000 −0.00802*** −0.00630** −0.0143*** −0.0167*** −0.0116*** −0.00850*** −0.0132** −0.0134***
(0.00239) (0.00259) (0.00547) (0.00603) (0.00288) (0.00289) (0.00520) (0.00463)
Obse a ions 1450 1450 1450 1450 1450 1450 1450 1450
Coun y FE Yes Yes Yes Yes
Con ols Yes Yes Yes Yes
Fi s s age F 58.6 12.2 4.8 7.1 32.6 6.4 9.2 7.2
BRANDILY and RAUCH
|
1259
