Infrastructure Complementarities and Local
Economic Growth: Evidence from Electriﬁcation
and Highway Construction in Brazil∗
Harris Selod1 , Jevgenijs Steinbuks1 , Ian Trotter2 , and Brian
Blankespoor1
1
The World Bank
2
cosa, Brazil
Universidade Federal de Vi¸

March 14, 2024

Abstract
This paper uses panel data over four decades in Brazilian municipal-
ities to study the separate and joint impacts of highway and electricity
infrastructure access on local economic outcomes. The identiﬁcation strat-
egy employs diﬀerence-in-diﬀerence estimators with staggered adoption
design and several treatments. Results show strong contemporaneous
eﬀects of electrifying municipalities that already have access to a highway,
whereas electriﬁcation or highway provision alone may, at best, have no
eﬀect. Infrastructure investments also facilitated long-lasting structural
transformation eﬀects, with both types of infrastructure access spurring
the growth of the industrial output share.

JEL: O13, O18, O47, O54
Keywords: Brazil, electricity, highways, infrastructure complementarities,
local economic development

∗ The authors thank Cl´ement de Chaisemartin, Niclas Moneke and St´ ephane Straub as
well as attendees of the NCID workshop on Energy and Environmental Issues in Developing
Countries, the European Meeting of the Urban Economics Association, the annual workshop
on the Empirical Methods in Energy Economics conference, and the Urbanization and Poverty
Reduction Research conference for useful advice, comments and suggestions.

1
1 Introduction
It is well established in the economic literature that infrastructure investment
is a major driver of economic development, particularly through its impact on
structural transformation (Berg et al., 2017). Infrastructure investments can
have sizeable multiplier eﬀects on local economies by boosting productivity
through better factor allocation (Vagliasindi and Gorgulu, 2021), facilitating
trade through improved access to markets (Jedwab and Storeygard, 2021), or
removing barriers to entry that prevent the development of local economic activ-
ities (Perez-Sebastian and Steinbuks, 2017). As various types of infrastructure
may diﬀerentially contribute to growth, it is important to disentangle their eﬀects
and how these eﬀects may vary when provided in isolation or in combination.
In particular, are diﬀerent infrastructure types complements or substitutes in
driving local economic eﬀects? The answer to this question bears much policy
relevance as policymakers need to optimize costly infrastructure investments
and their distribution over localities. It is also important for policymakers to
understand how these eﬀects may be nuanced and vary according to location
and sector.
In this paper, we study the joint impact of access to transportation (high-
ways)1 and electricity infrastructure on local gross value added (GVA) using
panel data over four decades and more than 3,000 Brazilian municipalities. The
study ﬁlls an important knowledge gap by focusing on infrastructure complemen-
tarities. It estimates the causal eﬀect of infrastructure complementarities using
the diﬀerence-in-diﬀerences model with multiple and potentially heterogeneous
treatments (De Chaisemartin and D’Haultfoeuille, 2022b; De Chaisemartin and
d’Haultfoeuille, 2020), and an identiﬁcation strategy that relies on the plausible
exogeneity of the timing and common trends across municipalities in the absence
of treatment. It then establishes a causal mechanism of infrastructure investment
impacts, distinguishing between productivity and relative input use channels.
We ﬁnd that neither the provision of highways nor electricity alone impacts
GVA per capita (our main variable of interest) or, separately, GVA or population.
At the same time, we ﬁnd a strong complementarity associated with the joint
provision of highways and electricity, which increases GVA per capita at the time
the infrastructure investment is made. The eﬀect, however, disappears after one
1 We focus on highways because they are Brazil’s main transportation mode for both people

and freight/cargo. Brazil has the fourth-largest highway system in the world. In 2014, road
cargo transport accounted for over 60% of Brazil’s transportation matrix (Sandoval, 2014).

2
or two periods. It percolates through increased output thanks to productivity
gains, whereas reallocation of labor and capital has a small, if any, impact.
We also ﬁnd evidence of structural transformation associated with providing
single and joint infrastructure types. Speciﬁcally, the provision of highways and
electricity leads to a higher GVA share of industry in the local economy, and
this eﬀect persists over time. These results withstand robustness and sensitivity
checks and are reinforced by an alternative identiﬁcation strategy using two-way
ﬁxed eﬀects with instrumental variables.
Our paper is closely related to several recent papers in the literature ex-
ploring the interactions between diﬀerent types of infrastructure and economic
development in low- and lower-middle-income countries. Moneke (2020) studies
how big push infrastructure investments aﬀect welfare through a positive TFP
erez-Sebasti´
shock in Ethiopia in a spatial general equilibrium model. P´ an, Ser-
rano Quintero, and Steinbuks (2023) endogenize the government’s decisions to
invest in the transport and electricity networks in a multi-sector quantitative
spatial equilibrium model that incorporates the quality of the infrastructure
networks, which determine sectoral productivities and trade costs. Using a
reduced-form econometric approach Vanden Eynde and Wren-Lewis (2021) and
Abbasi et al. (2022) ﬁnd evidence of strong economic complementarities for
roads and electricity in the context of rural India and the Sub-Saharan Africa
region, respectively. Our paper extends the approaches of the latter two studies
using a novel econometric approach and focusing on the long-term eﬀects in a
middle-income country.

2 Background and Data
2.1 Policy Context
Brazil sporadically started building roads and hydropower plants in the late
nineteenth century, with the ﬁrst highways constructed in the 1920s. This trend
slowly continued in the ﬁrst half of the twentieth century and signiﬁcantly accel-
erated in the 1960s and 1970s. The impetus came from President Kubitschek’s
prioritization of public infrastructure to support industrialization and fulﬁll
his electoral platform of “ﬁfty years of progress in ﬁve.” During his presidency
(1956–1961), Kubitschek launched a major national development plan (Piano de
Metas or “Targets Plan”) focusing on transportation and energy sectors. This
led to a major expansion of both electricity and transportation systems along

3
ılia (Ayres et al., 2019). This
with the creation of the new capital city, Bras´
impetus was continued under subsequent governments with a series of national
development plans in 1967, 1973, and 1976 emphasizing the construction of
hydroelectric dams and roads improvement and highway construction (Leturcq,
2018).2
The plans required a large public investment eﬀort toward infrastructure.
In the initial Targets Plan, 71 percent of public investments targeted energy,
¯
transportation, and communications (Obhara, 1974). These investments were
sustained in the national development plans that followed3 . They were under-
taken at a time when electricity demand increased due to population growth
and urbanization. New investments then slowed down – starting with transport
followed by electricity – during the “lost decade” of the 1980s and 1990s as
Brazil’s governments focused on addressing the macroeconomic crisis (Raiser
et al., 2017). This led to a progressive deterioration of the transport network as
investments did not cover maintenance needs (Correa and Ramos, 2010).
All in all, infrastructure investments resulted in a signiﬁcant expansion of
the electricity and highway networks. Lipscomb, Mobarak, and Barham (2013)
report that the Brazilian transmission network in Brazil grew at an average rate
of 8.9 percent per year between 1950 and 2000, resulting in a 72-fold increase
from 2,359 kilometers to 167,443 kilometers over the period. As a consequence,
household access to electricity rose from 48% in 1970 to 93% in 2000. The total
length of highways increased from 41,189 km in 1970 (of which 3,488 km of
ılia) to 90,788 km in 2000 (of which 7,440
radial highways originating from Bras´
ılia–calculation by authors based
km were radial highways originating from Bras´
on DNIT data).

2.2 Data
We start with creating a longitudinal dataset with four points in time (1970, 1980,
1996, and 2000) across 3,102 Brazilian municipalities for which economic data are
available.4 It covers more than 93 percent of the Brazilian population throughout
the study period.5 Our data comprise the municipal Gross Valued Added (GVA),
2 The 1973 plan aimed to create 8 radial roads, 17 longitudinal highways, 24 transverse

roads, 27 diagonal roads, and 62 connecting highways (Pereira and Lessa, 2011).
3 e.g., 47 percent in 1967, ibid.
4 Note that the data is unequally spaced in time, which will not aﬀect the scope and

magnitude of our results as shown in Section 4.4.
5 We work with administrative boundaries at the municipality level from the Brazilian

ıstica, IBGE )
Institute of Geography and Statistics (Instituto Brasileiro de Geograﬁa e Estat´

4
population, and labor and capital by sector of activity from Brazil’s national
accounts data disseminated by Brazil’s Institute for Applied Economic Research
omica Aplicada, IPEA).6 Unfortunately, the data for
(Instituto de Pesquisa Econˆ
labor and capital is available only for a subset of municipalities and years.
We use georeferenced data on roads for each decade from 1970 to 2000
obtained from Brazil’s National Department of Transport Infrastructure (De-
partamento Nacional de Infraestrutura de Transportes, DNIT) at a scale of
approximately 1:1,000,000. From the road data, we subset the road segments
that are paved, duplicated, or in the progress of duplication.7 For the purpose
of identiﬁcation (see next section), we separate out radial roads extending from
Brazilia towards the 25 state capitals8 along with two coastal highways based
on DNIT classiﬁcation.9 For each time period, we construct the start and end
points of the DNIT radial and coastal highway road segments that existed at the
time. Our road access measure is a dummy variable equal to 1 for municipalities
that have a radial road within 50 km of its centroid and 0 otherwise.
Electricity access data over the period 1970 to 2000 come from various
sources including the Geographic Information System of the Electric Sector
c˜
(Sistema de Informa¸ aﬁcas do Setor El´
oes Geogr´ etrico, SIGEL) of the Brazil
encia Nacional de Energia El´
National Electricity Regulatory Authority (Agˆ etrica,
ANEEL), the Business Information System for the Electricity Sector (Sistema
c˜
de Informa¸ etrica, SIESE ) of a major
oes Empresariais do Setor de Energia El´
Brazilian electric utility, Eletrobras, and Brazil’s and Cadastral Geographic
Information System of the National Interconnected System (Sistema De In-
c˜
forma¸ aﬁcas Cadastrais Do Sistema Interligado Nacional, SINDAT )
oes Geogr´
of the Brazil’s Operator of the National Electricity System (Operador Nacional
etrico, ONS ). As in Lipscomb, Mobarak, and Barham (2013),
do Sistema El´
we deﬁne electricity access in a given year as the share of municipality area
located within 50 kilometers of a transmission line, distribution substation, or
autonomous power plant.
Table 1 shows variable means for population, gross-valued added, and capital-
using the 1970 data as a baseline reference. The 568 missing municipalities for which economic
data are not available only represent 6 percent of the country’s population in 1970 and 7
percent in 2000.
6 See www.ipeadata.gov.br.
7 More precisely, the road segment is deﬁned as one of the following: ‘Pavimen-
tada’,‘Duplicada’, and ‘Em obra de Duplica¸ ˜o’.
ca
8 The state of Tocantins and its capital, Palmas, was not yet created in 1970.
9 Radial highways include road segments with labels ‘010’, ‘020’, ‘030’, ‘040’, ‘050’, ‘060’,

‘070’, ‘080’, ‘090’. Coastal highways include segments with labels ‘101’ and ‘116’.

5
labor ratio across municipalities, overall and for the industry sector, in the
study’s four periods.

Table 1: Variable Means (3,102 municipalities)

1970 1980 1996 2000
Population 28.16 35.92 47.17 50.93
Gross Value Added (GVA) 93.59 230.59 287.65 307.75
GVA per capita 1.56 3.42 3.08 3.77
GVA, Industry 32.48 92.35 93.33 84.66
Capital-Labor Ratio 11,023 34,279 N/A N/A
Capital-Labor Ratio, Industry 33,272 77,434 N/A N/A
Note: Population and labor are in million people; GVA and capital are
in million R$ (constant 2010 prices). The capital-labor ratio is available
for 1231 municipalities.

2.3 Infrastructure Investments in Our Data
Figure 1 shows the spatial evolution of radial and coastal highways and electricity
access in Brazil by comparing infrastructure access for 1970 and 2000, which are
the starting and end dates for our study. The maps show that electriﬁcation and
radial and coastal highway expansion mostly occurred towards the Northeast and
Center-West regions, whereas investments in other regions were modest. Figure
2 further details the number of municipalities by their degree of electriﬁcation
in 1970 and 2000. We see that the bulk of electriﬁcation investment consisted
in full electriﬁcation of formerly non-electriﬁed municipalities. Only a small
share of municipalities received partial electriﬁcation. As we construct a binary
measure of electriﬁcation for the empirical analysis, we focus on the case of full
electriﬁcation (i.e., municipalities with 100 percent of their population having
electricity access).
Figure 3 shows the roll-out of both radial and coastal highway investment
and full electriﬁcation. Consistent with the roll-out of the above investment
plans, infrastructure access increased signiﬁcantly. In 1970, 1,901 municipalities
(i.e., 61 percent of all municipalities) did not have access to power or highway
infrastructure, 975 municipalities only had electricity access, 117 municipalities
only had highway access, and 109 municipalities had both infrastructure types.

6
1970

2000

Figure 1: Changes in Infrastructure Access (1970-2000)

7
Figure 2: Electriﬁcation Rate across Municipalities in 1970 and 2000

Figure 3: Sankey diagram of infrastructure investment ﬂows, 1970-2000

Note. N: No infrastructure; H: Highway access; E: Electricity access; B: Both highway
and electricity access.

The bulk of the investment between 1970 and 2000 was the electriﬁcation of

8
424 of the municipalities that lacked any infrastructure and of 32 municipalities
that already had highway access. Highways were provided to 93 municipalities
that had no infrastructure and to 6 municipalities that were already electriﬁed.10
Only 9 municipalities were simultaneously granted highway and electricity access.

3 Identiﬁcation strategy and estimation sample
We are interested in estimating the average causal eﬀect of infrastructure pro-
vision on local economic outcomes at the municipality level (i.e., output per
capita, output, input use, and productivity measures). As infrastructure is
not necessarily placed independently of local economic outcomes, a potential
reverse causality problem arises. Also, given that some variables correlated with
both infrastructure and economic outcomes might not be observed, an omitted
variable bias might also be present. We address both of these issues, leading to
endogeneity problems, using two diﬀerent approaches.
Our main approach estimates a diﬀerence-in-diﬀerences model with multiple
and potentially heterogeneous treatments. Similar to Vanden Eynde and Wren-
Lewis (2021), our identiﬁcation strategy relies on the plausible exogeneity of the
timing and common trends across municipalities in the absence of treatment.
Following Bird and Straub (2020) and Morten and Oliveira (2023), we speciﬁcally
ılia, which was planned in the
focus on the radial network originating from Bras´
ılia
1950s to become the new capital of Brazil. Because the creation of Bras´
motivated the development of the current radial highway system linking it to
other state capitals (Morten and Oliveira, 2023), we can obtain a sample of
exogenously treated municipalities by removing these state capitals along with
ılia. To address the potential contamination eﬀects of road placement, we
Bras´
restrict the analysis to “inconsequential areas”, excluding urban nodes that
might have motivated highway construction. We also exclude municipalities
that have received non-radial highways. Finally, to ensure the comparability of
context, we only keep municipalities within 150 km of treated municipalities
in the control group. This leaves us with 391 municipalities, either without
any infrastructure or endowed with a radial highway in 1970. We observe 193
instances of infrastructure investments, either by electriﬁcation or radial highway
provision, between 1970 and 2000. In section 4.4, we explore the sensitivity of
this assumption by expanding the control group to municipalities within 200 km
10 This small number appears reasonable because electriﬁcation requires some means of

transport. Moneke (2020) makes a similar observation for Ethiopia.

9
of treated municipalities. To address the possibility of spatial spillovers (that
could lead to potential contamination eﬀects), we also perform a robustness
check that excludes municipalities within 50 km of treated municipalities from
the control group.
The empirical speciﬁcation that we estimate is as follows:

Yit = ζN,E I t = ti,t E I t = ti,t−1 H + ζN,H I t = ti,t H I t = ti,t−1 E
+ ζH,B I t = ti,t E I t = ti,t−1 H + ζE,B I t = ti,t H I t = ti,t−1 E (1)
+ ζN,B I t = ti,t E ∩H I t = ti,t−1 H I t = ti,t−1 E + αi + γt + εit

where the term tP
i,t is the time at which municipality i ﬁrst got access to
infrastructure P = {E, H }, αi and γt are municipality and time ﬁxed eﬀects, and
εit is the i.i.d. error term. The speciﬁcation above is estimated by the diﬀerence-
in-diﬀerences estimator (DIDM ) robust to heterogeneous eﬀects proposed by
De Chaisemartin and d’Haultfoeuille (2020) and extended for multiple treatments
by De Chaisemartin and D’Haultfoeuille (2022b). As noted by De Chaisemartin
and d’Haultfoeuille (2020), the DIDM estimator could fail if municipalities that
received one or several types of infrastructure experienced diﬀerent trends than
municipalities that did not receive any infrastructure. To test the parallel trends
assumption, we use the placebo estimator proposed by De Chaisemartin and
d’Haultfoeuille (2020) that compares the outcome’s evolution from t − 2 to t − 1,
in municipalities that became treated between t − 1 and t. If the parallel trends
assumption is violated, the placebo estimator will be signiﬁcantly diﬀerent from
0.

4 Results, Robustness Checks, and Discussion
4.1 Aggregate Output and Population
In this section, we present the DIDM estimator (De Chaisemartin and D’Haultfoeuille,
2022b; De Chaisemartin and d’Haultfoeuille, 2020) of the eﬀect of the separate
and joint provision of the electricity and highway access on various municipality
level outcomes. Our main focus is on gross value added per capita, which is a
normalized measure of output (allowing comparisons of impact across municipal-
ities of diﬀerent sizes). It also provides a measure of the apparent productivity
of municipalities. The DIDM estimator estimates the average treatment eﬀect at
the time when a municipality receives the infrastructure across all municipalities

10
that become treated at some point. All outcomes are expressed in natural
logarithms and have semi-elasticity interpretations.
We use binary measures of infrastructure access and consider whether a
municipality has access to a highway or is fully electriﬁed. We distinguish
across three switching scenarios depending on whether a municipality without
any infrastructure type receives electricity access only (denoted N2E), highway
access only (denoted N2H), or whether a municipality with highway access
becomes electriﬁed (H2B). As we have seen in Figure 3, there are only a few
transitions where an electriﬁed municipality is provided with a highway (E2B),
or a municipality is provided with both highway and electricity access (N2B).
We exclude those rare switching cases from the analysis as they would not have
meaningful interpretations.11
Tables 2-5 present the estimates of the impact of infrastructure provision
on municipal gross value added (GVA) per capita, GVA, population, and GVA
shares. The ﬁrst row presents the estimates for the entire period of the study
(1970-2000), whereas the next three rows show estimates for the sub-periods
1970-1980, 1980-1996, and 1996-2000 respectively. The penultimate row presents
the dynamic estimator of De Chaisemartin and D’Haultfoeuille (2022a), which
measures eﬀects one period after the infrastructure provision. The last row
shows the placebo estimator (see section 3) of the eﬀect of the separate and joint
provision of electricity and highway access. To preserve space, we only report
the number of observations and treated municipalities (switchers) in Table 2 as
these numbers are the same for all tables.
Table 2 focuses on municipal gross value added per capita, our main variable
of interest. We see from the ﬁrst row that neither the provision of highways
nor electricity alone has a statistically signiﬁcant impact on value-added per
capita. This contrasts previous research on the impact of electricity and highway
provision on output per capita and productivity (Berg et al., 2017; Bird and
Straub, 2020; Fernald, 1999; Perez-Sebastian, Steinbuks, et al., 2020). These
previous studies, however, did not consider whether municipalities receiving a
type of infrastructure investment already had received the other type. At the
same time, the electriﬁcation of municipalities already endowed with a highway
increases the value added per capita by 29 percent.12 This result highlights
11 The standard errors cannot be accurately estimated when the number of switching munici-
palities is too small.
12 The percentage change in the outcome is given by the formula exp(β )-1, where β is the

reported regression coeﬃcient.

11
the strong impact of providing the two types of infrastructure. We do not ﬁnd
statistically signiﬁcant dynamic estimates (row 5), which suggests that the eﬀect
of infrastructure placement is temporary. Placebo estimates of the provision
of electricity, whether alone or following highway provision, are much smaller
and statistically insigniﬁcant. This indicates that the parallel trends assumption
is unlikely to be violated for electricity provision. For highway provision, the
placebo eﬀect is positive and signiﬁcant, preventing us from concluding whether
the provision of highways alone impacted GVA per capita.

Table 2: DIDM estimates of infrastructure provision (GVA per Capita)

Dependent variable: Gross Value Added per Capita (log)
N2E N2H H2B
1970-2000 (E) coef. 0.036 0.021 0.252∗∗
s.e. (0.046) (0.088) (0.101)
Obs./Switchers 390 / 102 338 / 55 274 / 36
1970-1980 coef. -0.030 0.273 0.081
s.e. (0.057) (0.174) (0.175)
Obs./Switchers 168 / 35 134 / 21 89 / 4
1980-1996 coef. 0.139∗∗ -0.035 0.276∗∗
s.e. (0.069) (0.121) (0.132)
Obs./Switchers 133 / 44 113 / 22 104 / 23
1996-2000 coef. -0.063 -0.319∗∗∗ 0.269∗∗
s.e. (0.105) (0.081) (0.129)
Obs./Switchers 89 / 23 91 / 12 81 / 9
1970-2000 (D) coef. 0.074 0.170∗ 0.268
s.e. (0.061) (0.100) (0.170)
Obs./Switchers 234 / 79 213 / 43 165 / 27
Placebo coef. -0.110 0.135 0.089
s.e. (0.067) (0.080) (0.087)
Obs. 222 204 166
Notes: Bootstrapped standard errors (200 replications) in parentheses.
∗
p<0.1, ∗∗ p<0.05, ∗∗∗ p<0.01. (E): estimated eﬀect of the treatment
at the time period of the infrastructure provision. (D): estimated eﬀect
one period after the infrastructure provision.

The results by period (rows 2-4) show a more contrasted picture. There were
positive, albeit marginally signiﬁcant, impacts of highways on output per capita
in the 1970-1980. The provision of electricity alone positively aﬀected output per
capita in 1980-1996. By contrast, the period 1996-2000 shows negative impacts
on gross value added per capita of the separate provision of highways. This

12
could be due to the sequencing of investments targeting lagging regions in the
later period, as discussed in Section 2.1. We see suggestive complementarity
eﬀects for the last period, as the point estimate of the electriﬁcation eﬀect for
municipalities already endowed with a highway (column H2B) is greater than
that for electriﬁcation only (column N2E).

Table 3: DIDM estimates of infrastructure provision (Gross Value Added)

Dependent variable: Gross Value Added (log)
N2E N2H H2B
1970-2000 (E) coef. 0.035 -0.004 0.325∗∗∗
s.e. (0.047) (0.081) (0.125)
1970-1980 coef. 0.067 0.190 0.469
s.e. (0.071) (0.184) (0.309)
1980-1996 coef. 0.063 -0.018 0.284∗
s.e. (0.071) (0.112) (0.167)
1996-2000 coef. -0.069 -0.317∗∗∗ 0.367∗∗
s.e. (0.103) (0.103) (0.156)
1970-2000 (D) coef. 0.101 0.112 0.390∗
s.e. (0.075) (0.121) (0.220)
Placebo coef. -0.183∗∗∗ 0.138 0.046
s.e. (0.070) (0.094) (0.109)
Notes: Bootstrapped standard errors (200 replications)
in parentheses. ∗ p<0.1, ∗∗ p<0.05, ∗∗∗ p<0.01.
(E): estimated eﬀect of the treatment at the time period of
the infrastructure provision.
(D): estimated eﬀect one period after the infrastructure
provision.

Tables 3 and 4 decompose the previous ﬁndings by showing how municipality
gross value added and population responded to infrastructure investment. The
ﬁrst row of Table 3 shows a strong complementarity eﬀect as the gross value
added in municipalities endowed with a highway that was subsequently electriﬁed
increases by 38 percent compared to untreated municipalities. The provision of
electricity alone or highways alone did not have a statistically signiﬁcant impact
on GVA. The ﬁrst row of Table 4 does not show any statistically signiﬁcant
impact on the population in municipalities due to any type of infrastructure
provision. Taken together, these results suggest that the increase in gross
value added per capita (Table 2) is solely driven by the eﬀects of gross value
added resulting from the electriﬁcation of municipalities already endowed with a

13
highway. We do not ﬁnd any signiﬁcant dynamic eﬀects except for a positive
and marginally signiﬁcant eﬀect on GVA for the electriﬁcation of municipalities
already endowed with a highway (row 5). The placebo estimates of highway
provision and electriﬁcation in municipalities already endowed with highways are
not statistically signiﬁcant, lending credibility to the parallel trend assumption.
However, this is not the case for the provision of electricity alone, preventing
us from concluding whether the provision of electricity alone had an impact on
either GVA or population.
A closer examination by the period of the electriﬁcation of municipalities
already endowed with a highway (rows 2, 3, and 4 of Tables 3 and 4) indicates
that the complementarity eﬀect mainly occurred in the 1996-2000 period. In that
period, providing electricity in places already supplied with a highway positively
increased gross value added by 44 percent and population by 10 percent increase.

Table 4: DIDM estimates of infrastructure provision (Population)

Dependent variable: Population (log)
N2E N2H H2B
1970-2000 (E) coef. -0.001 -0.025 0.073
s.e. (0.025) (0.033) (0.066)
1970-1980 coef. 0.097∗∗ -0.083 0.388
s.e. (0.041) (0.063) (0.342)
1980-1996 coef. -0.076∗ 0.016 0.008
s.e. (0.041) (0.044) (0.073)
1996-2000 coef. -0.006 0.002 0.098∗∗∗
s.e. (0.023) (0.031) (0.022)
1970-2000 (D) coef. 0.027 -0.058 0.121
s.e. (0.055) (0.062) (0.116)
Placebo coef. -0.073∗ 0.003 -0.044
s.e. (0.041) (0.057) (0.067)
Notes: Bootstrapped standard errors (200 replications)
in parentheses. ∗ p<0.1, ∗∗ p<0.05, ∗∗∗ p<0.01.
(E): estimated eﬀect of the treatment at the time period of
the infrastructure provision.
(D): estimated eﬀect one period after the infrastructure
provision.

14
4.2 Structural Transformation
We now turn to the impact of infrastructure provision on the industry share in
Gross Value Added. The ﬁndings (summarized in Table 5) point towards evidence
of structural transformation. There are also noticeable heterogeneous eﬀects
associated with the separate and joint provision of electricity and highways.
Providing electricity (columns N2E) increases, on average, the share of
industry by 5 percent as compared to municipalities that were not electriﬁed.
The impact is consistently positive and signiﬁcant only in the 1970-1980 period.
The placebo test validates the impact of electriﬁcation. For electricity access,
the size of dynamic eﬀects is comparable to the average eﬀect. This indicates
that the impact of electriﬁcation on the industry GVA share persists over time,
indicating a long-lasting eﬀect on the economy’s structure.
We ﬁnd little average eﬀect of highway connection (columns N2H) on the
industry GVA share but a positive dynamic impact (6 percentage points), which is
three times as large as the average eﬀect and is statistically signiﬁcant. Contrary
to electricity provision, this indicates that highway provision has a long-lasting
impact on the GVA share of the industry sector. The placebo test for highway
provision does not reject this ﬁnding.
Both electricity and highway eﬀects indicate structural transformation through
an increase in the industry share. This is consistent with the previous literature
on the impact of electriﬁcation and structural transformation (Gaggl et al., 2021;
Perez-Sebastian, Steinbuks, et al., 2020) and previous studies that found that
highways facilitate input use for industry and spur the development of industries
trading heavier goods that can be transported by roads (Duranton, Morrow, and
Turner, 2014; Jaimovich, 2019).
Finally, we look at the impact on the the industry GVA share following
the electriﬁcation of municipalities that were already connected to the highway
network (columns H2B). We ﬁnd no statistically signiﬁcant eﬀect. This suggests
no complementarity in the sense that the share of industry does not change
when both types of infrastructure are present. The placebo test does not reject
the validity of these results.

15
Table 5: DIDM estimates of infrastructure provision (Industry GVA share)

Dependent variable: Industry GVA share
N2E N2H H2B
∗∗∗
1970-2000 (E) coef. 0.049 0.020 0.027
s.e. (0.016) (0.020) (0.029)
1970-1980 coef. 0.088∗∗∗ 0.030 0.042
s.e. (0.025) (0.042) (0.069)
1980-1996 coef. 0.027 0.026 -0.000
s.e. (0.023) (0.028) (0.040)
1996-2000 coef. 0.034 -0.009 0.090
s.e. (0.033) (0.020) (0.059)
1970-2000 (D) coef. 0.051∗∗ 0.058∗∗ 0.025
s.e. (0.020) (0.027) (0.053)
Placebo coef. 0.023 0.022 -0.009
s.e. (0.022) (0.015) (0.023)
Notes: Bootstrapped standard errors (200 replications)
in parentheses. ∗ p<0.1, ∗∗ p<0.05, ∗∗∗ p<0.01.
(E): estimated eﬀect of the treatment at the time period of
the infrastructure provision.
(D): estimated eﬀect one period after the infrastructure
provision.

4.3 Causal Mechanisms
We now investigate in greater detail the possible mechanisms that can ex-
plain how infrastructure provision aﬀects output and productivity. Assuming a
Cobb-Douglas aggregate production technology at the municipal level, we focus
speciﬁcally on the capital/labor ratio (capital intensity) and the total factor
productivity (TFP), which captures technological improvements (see derivation
and estimation procedure of TFP in Appendix 1). Capital and labor data are
available only for the period 1970-1980, which, fortunately, is a period when
signiﬁcant investment was made. However, as we explain in section 2.2, the data
are only available for some municipalities, leaving us with a smaller sample of
municipalities during that period. We, therefore, ﬁrst apply the DIDM estimator
to this restricted subsample to make sure that the main results in the second
row of Table 2 are unchanged. Appendix Table A1 shows that this is indeed
the case. Because the subsample does not include municipalities with highway
access that were subsequently electriﬁed during the period, we cannot estimate

16
complementarity eﬀects. This said, our results for the provision of single infras-
tructure types indicate that the subsample does not seem to suﬀer from sample
selection bias.

Table 6: DIDM estimates of infrastructure provision
on TFP and capital-labor ratio: 1970-1980

All Sectors Industry
N2E N2H N2E N2H

K/L coef. -0.128 -0.103 -0.361 -0.292
s.e. (0.137) (0.180) (0.255) (0.492)
TFP coef. 0.166∗ 0.367∗∗ 0.618∗∗∗ 0.100
s.e. (0.094) (0.173) (0.199) (0.241)
Obs./Switchers 87 / 25 89 / 19 87 / 25 89 / 19
Notes: Bootstrapped standard errors (200 replications) in parentheses.
∗
p<0.1, ∗∗ p<0.05, ∗∗∗ p<0.01. Dependent variables are in natural logs.

Table 6 shows the impact of electriﬁcation and highway provision on capital
intensity and TFP, pooling all sectors together, and separately for the industry
sector. We do not observe any statistically signiﬁcant eﬀects of the provision of
electricity or highways on the capital/labor ratio at the aggregate level. On the
contrary, the provision of either infrastructure type has strong positive eﬀects
on TFP. Electriﬁcation increases TFP by almost 18 percent, although this eﬀect
is only marginally signiﬁcant.
When focusing on the industry sector, we see that electriﬁcation signiﬁcantly
increases TFP, which is 86 percent larger in electriﬁed municipalities. This
is consistent with the previous literature documenting positive total factor
productivity eﬀects in the industry sector following electriﬁcation (Kassem,
2021). These ﬁndings are also consistent with our estimated changes in industry
GVA share following electriﬁcation (see Table 5), with TFP being the driver of
this change (for a detailed explanation of this mechanism, see Perez-Sebastian,
Steinbuks, et al., 2020).
Similar to electriﬁcation, we ﬁnd a positive and signiﬁcant eﬀect (a 44 percent
increase) of highway provision on aggregate TFP. The eﬀect of highway provision
on the industry sector TFP is positive but not statistically signiﬁcant. This is
consistent with the absence of signiﬁcant eﬀects of highway provision on the
industry GVA share for the 1970-1980 period in Table 5).

17
Overall, the above results indicate that infrastructure-induced growth per-
colated through productivity and, thus, more eﬃcient use of production inputs
facilitated by better access to electricity, predominantly for the industry sector
and highways for the aggregated economy.

4.4 Robustness Tests
We ﬁnd that there is no signiﬁcant diﬀerence in the distribution of gross value-
added and population across treatments and controls before treatment occurs,
as shown in Appendix Figures A2 and A3. Consistent with Kahn-Lang and
Lang (2020), this gives us further conﬁdence that our diﬀerence-in-diﬀerences
estimates are plausible.
Next, we reestimate our speciﬁcations from Tables 2, 3 and 4 using the
doubly robust diﬀerence-in-diﬀerences estimator proposed by Callaway and
Sant’Anna (2021). Figures 4 and 5 show the corresponding event study graphs
for our key results (the impact of electriﬁcation of municipalities previously
endowed with highways on GVA per capita and the impact of electriﬁcation
of municipalities without highways on the industry GVA share). Appendix
Figure A4 and Table A2 show event study graphs and the estimated coeﬃcients
for all results. The event study graphs and estimation results conﬁrm our
earlier ﬁndings. First, only joint infrastructure placement has a signiﬁcant
though contemporaneous eﬀect on GVA per capita (due to a temporary increase
in GVA without eﬀect on population, see Appendix Table A2). Second, the
electriﬁcation of municipalities without highways has a longer-lasting eﬀect on
structural transformation, increasing the industrial GVA share. The estimated
coeﬃcients also carry the same signiﬁcance as in our main speciﬁcation (i.e.,
using De Chaisemartin and D’Haultfoeuille (2022b) and De Chaisemartin and
d’Haultfoeuille (2020)) for all types of infrastructure investments and outcome
variables. If anything, all coeﬃcients measuring complementarity eﬀects (i.e.,
electriﬁcation in the presence of highways) are larger, indicating that our main
estimates are likely to be conservative.
We also replicate our main results from Table 2 using a less stringent measure
of electricity access, i.e., considering municipalities that are 80 percent electriﬁed
instead of fully electriﬁed. The results presented in Appendix Table A3 are
broadly of the same magnitude and signiﬁcance as those presented in Table 2.
One diﬀerence is that we lose the signiﬁcance of the eﬀect of electriﬁcation of
municipalities already endowed with a highway for the 1996-2000 period (which

18
Figure 4: Event study graphs: GVA per Capita, H2B

Note. The event study graph is based on the estimator of Callaway and Sant’Anna
(2021). H2B: highway access followed by electriﬁcation.

was only marginally signiﬁcant in Table 2). The only other diﬀerence is that
we ﬁnd a positive dynamic eﬀect for the provision of highways (N2H). However,
this result is not robust to the placebo test. Importantly, the placebo test for
complementarity (H2B) remains valid.
Additionally, in Table A4, we consider a larger sample of control municipalities
located up to 200 km from a treated municipality. The results closely match those
of Table 2 except that the positive complementarity eﬀect for the period 1980-
1996 becomes statistically signiﬁcant, and the positive and signiﬁcant dynamic
eﬀect emerges for the provision of highways. As for the previous robustness
check, the placebo test for complementarity (H2B) also remains valid.
In Table A5, we exclude from the control group municipalities within 50 km of
a treated municipality to reduce the possibility of local spillovers. Results are very
similar to those of Table 2. We only lose the signiﬁcance of the complementary
eﬀect for the period 1980-1996 (which was marginally signiﬁcant in Table 2).
Again, the placebo test remains valid for the electriﬁcation of municipalities
already endowed with a highway.
Finally, we explore the implications of data unequally spaced in time. We

19
Figure 5: Event study graphs: Industry GVA share: N2E

Note. The event study graph is based on the estimator of Callaway and Sant’Anna
(2021). N2E: no access to infrastructure followed by electriﬁcation.

follow recommendations from the Stata command did_multiplegt provided
by De Chaisemartin and d’Haultfoeuille (2020) and rerun the regression in two
diﬀerent ways: First, we remove the 1996 observations and append observations
identical to 1980 but with the year equal to 1990, and consider 1990 treatments
instead of 1996 treatments. This is a conservative approach as it assumes that the
outcomes are delayed. Second, we keep the 1996 outcomes (which the command
treats as if they were observed in 1990) but consider 1990 treatments. Appendix
tables A6 and A7 show that the scope and magnitude of results for our main
variable of interest, GVA per capita, are little changed. The impact on GVA per
capita of electriﬁcation of municipalities endowed with highways (H2B) is only
slightly diminished while signiﬁcance remains. The same ﬁndings hold for other
variables of interest.13
All in all, these results conﬁrm the robustness of our main ﬁndings.14
13 These results are available upon demand.
14 We carried out the same robustness checks for Tables 3, 4 and 5 and also found comparable
results.

20
4.5 Alternative Identiﬁcation Strategy
Our quantitative placebo, balance, and sample sensitivity tests and the use of
alternative DID estimators give us reasonable conﬁdence that our identiﬁcation
approach using diﬀerence-in-diﬀerences estimation is justiﬁed. To strengthen
our inference, we show in this section that our main results are also robust to an
alternative identiﬁcation strategy.
Speciﬁcally, we estimate an instrumental variable model for data with two-
way ﬁxed eﬀects. As with the main speciﬁcation, we exclude the state capitals,
the areas covered by non-radial highways, and keep municipalities within 150
km of treated municipalities in the control group. We regress our outcomes
of interest (log GVA, log population, and log GVA per capita) on a dummy
variable indicating the presence of a radial highway in the municipality, a dummy
variable indicating whether the municipality is electriﬁed, and on the interaction
between the two treatments (to assess complementarity eﬀect). To address the
endogeneity of electriﬁcation, we use the time-variant instrument developed by
Lipscomb, Mobarak, and Barham (2013) and further reﬁned in Szerman et al.
(2022). To address road placement endogeneity, we use a time-variant instrument
based on the natural least cost path between the urban nodes (see, e.g., Faber
(2014), Donaldson (2018) or Bird and Straub (2020) who proposed time-invariant
versions of the least-cost path instrument). Appendix A2 details the estimation
and construction of our instruments.
Table 7 shows the estimation results of the instrumental variable model.15
Consistent with our ﬁndings using our main identiﬁcation strategy reported
in section 4.1, we see no eﬀect of highway provision alone on GVA per capita,
GVA, or population. However, the estimated impact of electricity provision
alone diﬀers from the ﬁndings in section 4.1. Instead of ﬁnding no eﬀect, we now
see a negative and statistically signiﬁcant eﬀect of electriﬁcation on GVA and a
positive and statistically signiﬁcant eﬀect on population. This translates into a
negative and statistically signiﬁcant eﬀect of electricity provision on GVA per
capita. We are not concerned about this diﬀerence since the placebo tests had
failed for electricity provision in our main identiﬁcation strategy. The new results
imply that GVA is smaller in electriﬁed places that lack highway access. This is
consistent with the literature that documents how workers and activities leave
unproductive peripheral areas when infrastructure is improved (Faber, 2014).
15 See Appendix Table A10 for the ﬁrst-stage estimation results.

21
Table 7: IV estimates of infrastructure provision (second-stage)

Dependent variable:
GVA per GVA (log) Population (log)
capita (log)
(1) (2) (3)
∗∗∗ ∗∗∗
Electricity Access −2.341 −1.366 0.975∗∗∗
(0.386) (0.462) (0.214)

Highway Access −0.554 −0.406 0.148
(0.496) (0.576) (0.276)

Electricity Access X 3.372∗∗∗ 2.898∗∗∗ −0.475
Highway Access (0.951) (1.110) (0.532)

Wald Test χ2 3844.2∗∗∗ 2577.3∗∗∗ 2501.3∗∗∗
Observations 1,536 1,536 1,536
Note: Robust standard errors clustered at the state level in parentheses.
∗
p<0.1,∗∗ p<0.05,∗∗∗ p<0.01.

However, these negative eﬀects associated with the provision of electricity
alone are fully compensated when both types of infrastructure are present,
conﬁrming our main ﬁndings regarding infrastructure complementarity in section
4.1. The overall eﬀect of joint infrastructure provision (given by the sum of the
three coeﬃcients on electricity access, highway access, and their interaction) is a
twofold increase in GVA and an increase of GVA per capita by 61 percent.

5 Conclusions
This paper is among the few to investigate the eﬀects of diﬀerent types of
infrastructure provision on local economic development and the ﬁrst to investigate
the issue over a long period (3 decades). Focusing on several decades is necessary
not only because investments are staggered in time but also because eﬀects may
take time to materialize. In our case, the study period allows us to capture the
bulk of infrastructure investment in Brazil, which started in the 1970s, and to
assess how these large investment programs aﬀected targeted municipalities.
In contrast with previous results from the literature, we do not ﬁnd strong

22
evidence of single-type infrastructure provision on output and apparent labor
productivity (which we measure with municipal GVA and GVA per capita).
We do ﬁnd, however, evidence of an impact of the joint provision of electricity
and highway infrastructure, which we estimate to increase GVA by 38 to 100
percent and GVA per capita by 29 to 61 percent, depending on the econometric
approach. Importantly, the impacts diﬀer across decades, being highly signiﬁcant
in some and not signiﬁcant in others. This arguably reconciles ﬁndings from
the literature where short-term and long-term eﬀects can signiﬁcantly vary (Lee,
Miguel, and Wolfram, 2020).
In addition, we ﬁnd a positive eﬀect of infrastructure provision (both highways
and electricity separately) on TFP, though our data restricts the analysis of
TFP impacts to a single decade.
Finally, we ﬁnd evidence of infrastructure investment playing a role in fa-
cilitating structural transformation. The separate provision of highways and
electricity increases the local GVA share of the industry by about 5 percent.
These eﬀects are found to persist over time.
These results have an important policy implication: the government should
consider infrastructure provision as a package. If the objective is to promote
local economic development, then failing to improve transport access along with
electriﬁcation may not be eﬀective.

References
Abbasi, Mansoureh et al. (2022). Roads, Electricity, and Jobs: Evidence of Infras-
tructure Complementarity in Sub-Saharan Africa. Policy Research Working
Paper 9976. The World Bank.
Ali, Rubaba et al. (2015). Highways to Success or Byways to Waste: Estimating
the Economic Beneﬁts of Roads in Africa. The World Bank, Washington
D.C.
en, and Patrick J Kehoe (2019). The
Ayres, Joao, Marcio Garcia, Diogo A Guill´
Monetary and Fiscal History of Brazil, 1960-2016. Working Paper 25421.
National Bureau of Economic Research.
Balestra, Pietro and Jayalakshmi Varadharajan-Krishnakumar (1987). “Full
Information Estimations of a System of Simultaneous Equations with Error
Component Structure”. Econometric Theory 3.2, pp. 223–246.

23
org Peters, and Colin Vance (2021). Development Eﬀects
Bensch, Gunther, J¨
of Electriﬁcation in Brazil–A Comment on Lipscomb et al.(2013). USAEE
Working Paper.
Berg, Claudia N, Uwe Deichmann, Yishen Liu, and Harris Selod (2017). “Trans-
port Policies and Development”. The Journal of Development Studies 53.4,
pp. 465–480.
ephane Straub (2020). “The Brasilia Experiment: the Hetero-
Bird, Julia and St´
geneous Impact of Road Access on Spatial Development in Brazil”. World
Development 127, p. 104739.
Bun, Maurice JG and Teresa D Harrison (2019). “OLS and IV Estimation of
Regression Models Including Endogenous Interaction Terms”. Econometric
Reviews 38.7, pp. 814–827.
Callaway, Brantly and Pedro HC Sant’Anna (2021). “Diﬀerence-in-Diﬀerences
with Multiple Time Periods”. Journal of Econometrics 225.2, pp. 200–230.
Correa, Vivian Helena Capacle and Pedro Ramos (2010). “A Precariedade do
ario Brasileiro para o Escoamento da Produ¸
Transporte Rodovi´ ˜o de Soja do
ca
ca
Centro-Oeste: situa¸˜o e perspectivas”. Brazilian Journal of Rural Economy
and Sociology (Revista de Economia e Sociologia Rural-RESR) 48.2, pp. 1–26.
Croissant, Yves and Giovanni Millo (2008). “Panel Data Econometrics in R: the
plm Package”. Journal of Statistical Software 27.2.
ement and Xavier D’Haultfoeuille (2022a). Diﬀerence-in-
De Chaisemartin, Cl´
diﬀerences Estimators of Intertemporal Treatment Eﬀects. Working Paper
29873. National Bureau of Economic Research.
— (2022b). Two-way Fixed Eﬀects and Diﬀerences-in-Diﬀerences Estimators
with Several Treatments. Working Paper 30564. National Bureau of Economic
Research.
ement and Xavier d’Haultfoeuille (2020). “Two-way Fixed Ef-
De Chaisemartin, Cl´
fects Estimators with Heterogeneous Treatment Eﬀects”. American Economic
Review 110, pp. 2964–96.
Donaldson, Dave (2018). “Railroads of the Raj: Estimating the impact of trans-
portation infrastructure”. American Economic Review 108.4-5, pp. 899–934.
Duranton, Gilles, Peter M Morrow, and Matthew A Turner (2014). “Roads and
Trade: Evidence from the US”. Review of Economic Studies 81.2, pp. 681–724.
Faber, Benjamin (2014). “Trade Integration, Market Size, and Industrialization:
Evidence from China’s National Trunk Highway System”. Review of Economic
Studies 81.3, pp. 1046–1070.

24
Fernald, John G (1999). “Roads to Prosperity? Assessing the Link between Public
Capital and Productivity”. American Economic Review 89.3, pp. 619–638.
Gaggl, Paul, Rowena Gray, Ioana Marinescu, and Miguel Morin (2021). “Does
Electricity Drive Structural Transformation? Evidence from the United
States”. Labour Economics 68, p. 101944.
Jaimovich, Esteban (2019). “Roadways, Input Sourcing, and Patterns of Special-
ization”. European Economic Review 120, p. 103319.
Jedwab, Remi and Adam Storeygard (2021). “The Average and Heterogeneous
Eﬀects of Transportation Investments: Evidence from sub-Saharan Africa
1960–2010”. Journal of the European Economic Association jvab027.
Kahn-Lang, Ariella and Kevin Lang (2020). “The Promise and Pitfalls of
Diﬀerences-in-Diﬀerences: Reﬂections on 16 and Pregnant and Other Appli-
cations”. Journal of Business & Economic Statistics 38.3, pp. 613–620.
Kassem, Dana (2021). Does Electriﬁcation Cause Industrial Development. Work-
ing Paper.
Klein Goldewijk, Kees, Arthur Beusen, Gerard Van Drecht, and Martine De
Vos (2011). “The HYDE 3.1 Spatially Explicit Database of Human-Induced
Global Land-Use Change over the Past 12,000 Years”. Global Ecology and
Biogeography 20.1, pp. 73–86.
Lee, Kenneth, Edward Miguel, and Catherine Wolfram (2020). “Does Household
Electriﬁcation Supercharge Economic Development?” Journal of Economic
Perspectives 34.1, pp. 122–44.
Leturcq, Guillaume (2018). Dams in Brazil: Social and Demographical Impacts.
Springer.
Lipscomb, Molly, A Mushﬁq Mobarak, and Tania Barham (2013). “Development
Eﬀects of Electriﬁcation: Evidence from the Topographic Placement of Hy-
dropower Plants in Brazil”. American Economic Journal: Applied Economics
5.2, pp. 200–231.
Moneke, Niclas (2020). “Infrastructure and Structural Transformation: Evidence
from Ethiopia”. PhD thesis. The London School of Economics and Political
Science (LSE).
Morten, Melanie and Jaqueline Oliveira (2023). “The Eﬀects of Roads on Trade
and Migration: Evidence from a Planned Capital City”. American Economic
Journal: Applied Economics forthcoming.
¯
Obhara, Yoshinori (1974). “Brazilian Economic Development Since 1956: A Study
on Policies Reﬂecting the Stages of Growth”. The Developing Economies
12.2, pp. 133–154.

25
Pereira, Luiz AG and Simone Narciso Lessa (2011). “O Processo de Planeja-
ario no Brasil”. Caminhos
mento e Desenvolvimento do Transporte Rodovi´
de Geograﬁa 12.40, pp. 26–46.
Perez-Sebastian, Fidel and Jevgenijs Steinbuks (2017). Public Infrastructure
and Structural Transformation. Policy Research Working Paper 8285. World
Bank.
e F´
Perez-Sebastian, Fidel, Jevgenijs Steinbuks, Jos´ eres, and Ian Trotter (2020).
Electricity Access and Structural Transformation: Evidence from Brazil’s
Electriﬁcation. Policy Research Working Paper 9182. World Bank.
erez-Sebasti´
P´ an, Fidel, Rafael Serrano Quintero, and Jevgenijs Steinbuks (2023).
Spatial Misallocation of Complementary Infrastructure Investment: Evidence
from Brazil. Policy Research Working Paper 10650. The World Bank.
Raiser, Martin et al. (2017). Back to Planning: How to Close Brazil’s Infras-
tructure Gap in Times of Austerity. Technical Report 117392-BR. World
Bank.
alise das Implica¸
Sandoval, Daylyne Maerla Gomes Lima (2014). “An´ ˜es Log´
co ısticas,
abeis e Financeiras da Escolha do Servi¸
Con´ co de Transporte”. PhD thesis.
ılia, Faculdade de Tecnologia.
Universidade de Bras´
Stock, James H and Jonathan H Wright (2000). “GMM with Weak Identiﬁcation”.
Econometrica 68.5, pp. 1055–1096.
cao, Molly Lipscomb, and Ahmed Mushﬁq
Szerman, Dimitri, Juliano Assun¸
Mobarak (2022). Agricultural Productivity and Deforestation: Evidence from
Brazil. Working Paper.
Tobler, Waldo (1993). Three Presentations on Geographical Analysis and Model-
ing. Technical Report.
Vagliasindi, Maria and Nisan Gorgulu (2021). What Have We Learned about the
Eﬀectiveness of Infrastructure Investment as a Fiscal Stimulus? A Literature
Review. World Bank Policy Research Paper 9796. The World Bank.
Van Beveren, Ilke (2012). “Total Factor Productivity Estimation: A Practical
Review”. Journal of Economic Surveys 26.1, pp. 98–128.
Vanden Eynde, Oliver and Liam Wren-Lewis (2021). Complementarities in
Infrastructure: Evidence from Rural India. CEPR Discussion Paper 16139.
Verdin, Kristine L et al. (2007). Development of a Global Slope Dataset for
Estimation of Landslide Occurrence Resulting from Earthquakes. Technical
Report.
Wooldridge, Jeﬀrey M (2010). Econometric Analysis of Cross Section and Panel
Data. MIT press.

26
Appendix
A1 Estimation of Total factor Productivity
We assume that municipal economies follow a Cobb-Douglas production of the
form

α β
Yit = Ait Kit Lit (A1)

where Yit , Ait , Kit and Lit are, respectively, the GVA, TFP, capital stock
and total labor in municipality i at time t, α and β are the Cobb-Douglas
elasticities. We estimate the production function by log-regressing municipal-
level GVA on local capital and labor. We estimate both a pooled model,
which assumes common production technology, and a ﬁxed-eﬀects model (our
preferred speciﬁcation), allowing for heterogeneous production technologies
across municipalities (see Van Beveren, 2012). Tables A8 and A9 show the
estimation results. Recovering the estimates for α and β , we then calculate TFP
α β
as Yit /Kit Lit . As a robustness check, we also estimated the production function
parameters imposing homogeneity of degree 1 (i.e., assuming a constant return
to scale), leading to similar eﬀects of infrastructure access on TFP as in Table 6.

A2 Estimation Using Instrumental Variables
Empirical Speciﬁcation

The ﬁrst-stage regressions that we estimate are as follows:

Xit = αi + Bit Zit + γt + εit (A2)

where i is the municipality index, t is the time index, Xit is the vector of variables
of interest, i.e., the access to road indicator, the access to electricity indicator
and their interaction term, Zit is the vector of instrumental variables, Bit is a
matrix of coeﬃcients to be estimated, αi and γt are vectors of municipality and
year ﬁxed eﬀects, and εit is a vector of i.i.d. error terms.
Note that the functional form of the interaction model necessitates additional
instrumental variables to satisfy the order condition for identiﬁcation (Bun and
Harrison, 2019; Wooldridge, 2010). That is, vector Zit should include instruments
for electricity access, road access, and the interaction of electricity access and
road access. We instrument for the interaction term by the interaction of the

27
instruments for each term separately. Doing so satisﬁes the rank identiﬁcation
condition as long as instrumental variables for electricity and road access terms
are not perfectly colinear. However, owing to the nonlinear nature of the
interaction term, weak identiﬁcation problems may arise (Stock and Wright,
2000). We test for the possibility of weak identiﬁcation using the Wald test for
panel data (Wooldridge, 2010).

The second-stage regression that we estimate is as follows:

Yit = δi + Xit λit + µt + ηit (A3)

where Yit is an outcome variable (such as municipality gross value added or
total local employment), Xit is the vector of estimated variables of interest in
the ﬁrst stage, λit are our coeﬃcients of interest, δi and µt are municipality and
year ﬁxed eﬀects, and ηit is the error term.

The model is estimated using the instrumental variable transformation
method of Balestra and Varadharajan-Krishnakumar (1987) with the plm package
in R software (see Croissant and Millo, 2008).

Construction of the instruments

As all the time-invariant exogenous spatial variation in infrastructure placement
will be accounted for by municipality ﬁxed eﬀect, we must construct time-varying
instruments for both electricity and road placements, as described below.
We use the access to electricity instrumental variable developed by Lipscomb,
Mobarak, and Barham (2013) and further reﬁned in Szerman et al. (2022).16
The instrument takes advantage of the fact that hydropower accounts for the
majority of electricity generation in Brazil, accounting for 88 to 92 percent of
electricity generation in the country between 1980 and 2000, according to the US
International Energy Statistics database.17 The power potential of a hydropower
plant depends on local topological and hydrological characteristics, such as e.g.,
slope, elevation, and the amount of water available. These characteristics are
plausibly exogenous to the unobserved local conditions that may be correlated
with economic growth. The construction of an instrument involves three key
steps. The ﬁrst step calculates the total spending budget for hydropower
16 The original instrument of Lipscomb, Mobarak, and Barham (2013) was criticized for

inconsistent demarcation of the Amazon region throughout the diﬀerent steps of their instrument
construction, see Bensch, Peters, and Vance (2021).
17 See https://www.eia.gov/international/overview/world.

28
plants based on the actual construction of major dams across the entire country
each decade. The second step calculates a cost factor that sorts potential
locations by their geographic suitability. The ﬁnal step applies the suitability
predictions to the areas where hydropower plants were actually built. Using
the predictions of the estimated construction site for each hydropower plant,
a hypothetical transmission network that depends solely on topological and
hydrological characteristics is constructed and used as an instrumental variable.
Detailed description of constructing the instrument can be found in Lipscomb,
Mobarak, and Barham (2013) and Szerman et al. (2022).
We construct the “natural path” instrumental variable for highways following
Ali et al. (2015) based on the least amount of time between the 25 existing state
capitals in the base period18 and Bras´
ılia. It incorporates local terrain using
slope derived from a Digital Elevation Model (NASA’s Shuttle Radar Topography
Mission) from Verdin et al. (2007) and a historical land cover dataset (circa
1900) from HYDE (Klein Goldewijk et al., 2011). We use the same walking
velocity from Tobler (1993) :

W = 6 ∗ exp(−3.5|S +0.05|) (A4)

where W is the hiking velocity (kph) and S is the slope or gradient. We
modify the walking speed by land cover classes from HYDE using the weights
in Ali et al. (2015). We use the cost connectivity algorithm in ESRI ArcGIS to
solve the least cost path.
We construct the natural path for each time period by taking the result of the
intersection of the endpoints of a road segment for each radial corridor (in a time
period) with the nearest point on the natural path line (See Figure A1).19 That
is, each time period, we only consider the least cost segments closest to existing
roads in a given time period. The portion of the least cost path increases with
time, along with the construction of the existing roads. This makes it possible
to use in a panel regression.
We deﬁne the instrumental variable for road access as the binary variable
indicating the presence of a natural path road within 50 kilometers of the
geographic center of a locality. We also consider other buﬀer sizes, leading to
comparable results.
18 Even though the state of Tocantins was only established in 1988, we conservatively treat

it as an urban node motivating highway construction.
19 We have a total of 8 radial corridors.

29
Table A10 describes the results of the ﬁrst-stage regression (equation A2).

Appendix Tables and Figures

Table A1: DIDM estimates of infrastructure provision
on Gross Value Added per capita in 1970-1980
using a sample with available capital stock

Dependent variable:
Gross Value Added per Capita (log)
N2E N2H
coef. -0.030 0.288∗∗
s.e. (0.064) (0.144)
Obs./Switchers 87 / 25 89 / 19
Notes: Bootstrapped standard errors (200 replications)
in parentheses. ∗ p<0.1, ∗∗ p<0.05, ∗∗∗ p<0.01. All

Table A2: Doubly robust diﬀerence-in-diﬀerences estimator of
infrastructure provision (Callaway and Sant’Anna, 2021)

Period: 1970-2000
N2E N2H H2B
log(GVA/POP) coef. 0.037 0.130 0.314∗∗
s.e. (0.050) (0.094) (0.133)
Obs./Switchers 672 / 168 492 / 123 413 / 89
log(GVA) coef. 0.067 0.047 0.457∗∗∗
s.e. (0.054) (0.112) (0.165)
Obs./Switchers 672 / 168 492 / 123 413 / 89
log(POP) coef. 0.030 -0.084 0.143
s.e. (0.045) (0.057) (0.123)
Obs./Switchers 672 / 168 492 / 123 413 / 89
Notes: Bootstrapped standard errors (1000 replications) in parentheses.
∗
p<0.1, ∗∗ p<0.05, ∗∗∗ p<0.01.

30
Table A3: DIDM estimates of infrastructure provision
on Gross Value Added per capita
using 80% electriﬁcation rate proxy for electricity access

Dependent variable: Gross Value Added per Capita (log)
N2E N2H H2B
1970-2000 (E) coef. 0.016 0.073 0.220∗∗
s.e. (0.050) (0.091) (0.092)
Obs./Switchers 376 / 101 286 / 47 250 / 45
1970-1980 coef. 0.017 0.404∗∗ 0.131
s.e. (0.071) (0.179) (0.111)
Obs./Switchers 160 / 31 113 / 17 82 / 4
1980-1996 coef. 0.069 0.014 0.274∗
s.e. (0.089) (0.141) (0.149)
Obs./Switchers 129 / 42 96 / 19 97 / 26
1996-2000 coef. -0.063 -0.336∗∗∗ 0.150
s.e. (0.101) (0.089) (0.108)
Obs./Switchers 87 / 28 77 / 11 71 / 15
1970-2000 (D) coef. -0.053 0.190∗ 0.292∗
s.e. (0.084) (0.099) (0.166)
Obs./Switchers 219 / 73 179 / 36 144 / 30
Placebo coef. -0.168∗∗ 0.155 -0.008
s.e. (0.078) (0.082) (0.092)
Obs./Switchers 216 173 149
Notes: Bootstrapped standard errors (200 replications) in parentheses.
∗
p<0.1, ∗∗ p<0.05, ∗∗∗ p<0.01. (E): estimated eﬀect of the treatment
at the time period of the infrastructure provision. (D): estimated eﬀect
one period after the infrastructure provision.

31
Table A4: DIDM estimates of infrastructure provision
on Gross Value Added per capita
keeping controls within 200 km of treated municipalities

Dependent variable: Gross Value Added per Capita (log)
N2E N2H H2B
1970-2000 (E) coef. 0.023 0.043 0.228∗∗
s.e. (0.044) (0.083) (0.097)
Obs./Switchers 450 / 102 401 / 55 310 / 36
1970-1980 coef. -0.035 0.289 0.087
s.e. (0.062) (0.175) (0.184)
Obs./Switchers 188 / 35 155 / 21 101 / 4
1980-1996 coef. 0.130∗∗ 0.013 0.241∗
s.e. (0.055) (0.114) (0.133)
Obs./Switchers 153 / 44 134 / 22 116 / 23
1996-2000 coef. -0.094 -0.332∗∗∗ 0.259∗
s.e. (0.103) (0.075) (0.149)
Obs./Switchers 109 / 23 112 / 12 93 / 9
1970-2000 (D) coef. 0.047 0.218∗∗ 0.219
s.e. (0.059) (0.100) (0.169)
Obs./Switchers 274 / 79 255 / 43 189 / 27
Placebo coef. -0.112∗ 0.161∗∗ 0.079
s.e. (0.065) (0.073) (0.085)
Obs./Switchers 262 246 190
Notes: Bootstrapped standard errors (200 replications) in parentheses.
∗
p<0.1, ∗∗ p<0.05, ∗∗∗ p<0.01. (E): estimated eﬀect of the treatment
at the time period of the infrastructure provision. (D): estimated eﬀect
one period after the infrastructure provision.

32
Table A5: DIDM estimates of infrastructure provision
on Gross Value Added per capita
excluding controls within 50 km of treated municipalities

Dependent variable: Gross Value Added per Capita (log)
N2E N2H H2B
1970-2000 (E) coef. 0.037 0.023 0.230∗∗
s.e. (0.044) (0.086) (0.102)
Obs./Switchers 381 / 102 335 / 55 258 / 36
1970-1980 coef. -0.033 0.283∗ 0.074
s.e. (0.070) (0.168) (0.171)
Obs./Switchers 165 / 35 133 / 21 85 / 4
1980-1996 coef. 0.144∗ -0.036 0.234∗
s.e. (0.077) (0.115) (0.132)
Obs./Switchers 130 / 44 112 / 22 98 / 23
1996-2000 coef. -0.060 -0.325∗∗∗ 0.289∗
s.e. (0.101) (0.077) (0.148)
Obs./Switchers 86 / 23 90 / 12 75 / 9
1970-2000 (D) coef. 0.079 0.172∗ 0.237
s.e. (0.060) (0.098) (0.185)
Obs./Switchers 228 / 79 211 / 43 155 / 27
Placebo coef. -0.112 0.143 0.070
s.e. (0.069) (0.078) (0.099)
Obs./Switchers 216 202 156
Notes: Bootstrapped standard errors (200 replications) in parentheses.
∗
p<0.1, ∗∗ p<0.05, ∗∗∗ p<0.01. (E): estimated eﬀect of the treatment
at the time period of the infrastructure provision. (D): estimated eﬀect
one period after the infrastructure provision.

33
Table A6: DIDM estimates of infrastructure provision
on Gross Value Added per capita
using treatment in 1990 and outcome copied from 1980 onto 1990

Dependent variable: Gross Value Added per Capita (log)
N2E N2H H2B
1970-2000 (E) coef. 0.016 0.058 0.209∗
s.e. (0.043) (0.070) (0.116)
Obs./Switchers 407 / 102 338 / 55 286 / 36
1970-1980 coef. -0.030 0.273 0.081
s.e. (0.065) (0.175) (0.169)
Obs./Switchers 168 / 35 134 / 21 89 / 4
1980-1990 coef. 0.000 0.000 0.000
s.e. (0.000) (0.000) (0.000)
Obs./Switchers 133 / 27 113 / 22 104 / 11
1990-2000 coef. 0.068 -0.211∗ 0.343∗
s.e. (0.090) (0.118) (0.175)
Obs./Switchers 106 / 40 91 / 12 93 / 21
1970-2000 (D) coef. 0.023 0.131 0.160
s.e. (0.055) (0.118) (0.192)
Obs./Switchers 234 / 62 213 / 43 162 / 15
Placebo coef. -0.092∗∗∗ 0.097 0.028
s.e. (0.032) (0.071) (0.039)
Obs./Switchers 239 204 178
Notes: Bootstrapped standard errors (200 replications) in parentheses.
∗
p<0.1, ∗∗ p<0.05, ∗∗∗ p<0.01. (E): estimated eﬀect of the treatment
at the time period of the infrastructure provision. (D): estimated eﬀect
one period after the infrastructure provision.

34
Table A7: DIDM estimates of infrastructure provision
on Gross Value Added per capita
using treatment in 1990 and outcome copied from 1996 onto 1990

Dependent variable: Gross Value Added per Capita (log)
N2E N2H H2B
1970-2000 (E) coef. 0.029 0.021 0.174∗∗
s.e. (0.038) (0.082) (0.084)
Obs./Switchers 407 / 102 338 / 55 286 / 36
1970-1980 coef. -0.030 0.273∗ 0.081
s.e. (0.068) (0.164) (0.172)
Obs./Switchers 168 / 35 134 / 21 89 / 4
1980-1990 coef. 0.139∗∗ -0.035 0.206
s.e. (0.070) (0.104) (0.195)
Obs./Switchers 133 / 27 113 / 22 104 / 11
1990-2000 coef. 0.006 -0.319∗∗∗ 0.175
s.e. (0.085) (0.085) (0.108)
Obs./Switchers 106 / 40 91 / 12 93 / 21
1970-2000 (D) coef. 0.032 0.170∗ 0.140
s.e. (0.057) (0.102) (0.194)
Obs./Switchers 234 / 62 213 / 43 162 / 15
Placebo coef. -0.055 0.135∗ 0.142
s.e. (0.059) (0.082) (0.102)
Obs./Switchers 239 204 178
Notes: Bootstrapped standard errors (200 replications) in parentheses.
∗
p<0.1, ∗∗ p<0.05, ∗∗∗ p<0.01. (E): estimated eﬀect of the treatment
at the time period of the infrastructure provision. (D): estimated eﬀect
one period after the infrastructure provision.

35
Table A8: Production function estimates

Dependent variable:
Gross Value Added (log+1)

Pooled FE
∗∗∗
Capital (log) 0.70 0.23∗∗∗
(0.01) (0.03)

Labor (log) 0.28∗∗∗ 0.34∗∗∗
(0.02) (0.03)

Constant −2.70∗∗∗
(0.13)

Municipality FE No Yes
Time FE No Yes
F Statistic 3,777.56∗∗∗ 114.35∗∗∗
Observations 2,462 2,462
Note: Robust standard errors in parentheses.
∗
p<0.1, ∗∗ p<0.05, ∗∗∗ p<0.01

36
Table A9: Production function estimates for the industry sector

Dependent variable:
Gross Value Added (log+1)

Pooled FE
Capital (log) 0.47∗∗∗ 0.27∗∗∗
(0.02) (0.02)

Labor (log) 0.26∗∗∗ 0.18∗∗∗
(0.02) (0.02)

Constant −0.64∗∗∗
(0.05)

Municipality FE No Yes
Time FE No Yes
F Statistic 8,239.91∗∗∗ 351.93∗∗∗
Observations 2,462 2,462
Note: Robust standard errors in parentheses.
∗
p<0.1, ∗∗ p<0.05, ∗∗∗ p<0.01.

37
Table A10: IV estimates of infrastructure provision (ﬁrst-stage)

Dependent variable:
Electricity Highway Electricity X
Access Access Highway Access

(1) (2) (3)
Electricity Access IV −0.323∗∗∗ 0.474∗∗∗ −0.084∗∗∗
(0.026) (0.105) (0.009)

Roads Access IV 0.052 0.830∗∗∗ 0.247∗
(0.129) (0.065) (0.141)

Electricity Access IV X 0.058 −0.615∗∗∗ 0.061
Roads Access IV (0.161) (0.091) (0.178)

Municipality FE Yes Yes Yes
Time FE Yes Yes Yes
Wald test χ2 156.8∗∗∗ 193.9∗∗∗ 100.6∗∗∗
Observations 1,536 1,536 1,536
Note: Robust standard errors in parentheses. ∗ p<0.1, ∗∗
p<0.05, ∗∗∗
p<0.01.

38
Figure A1: This map shows the time-variant natural path.

39
Figure A2: Box Plots of Municipal GVA by Investment Sequence

Note. The vertical axis measures gross value added or population at the beginning of
the period. Box plots group municipalities by initial status and investment occurring
during the period. H2B: highway access followed by electriﬁcation; H2H: highway
access not followed by any investment; N2E: no access to infrastructure followed by
electriﬁcation; N2H: no access to infrastructure followed by highway provision; N2N:
no access to infrastructure throughout the period. Treatment groups are H2B, N2E
and N2H. Control groups are H2H and N2N.

40
Figure A3: Box Plots of Municipal Population by Investment Sequence

41
Figure A4: Event study graphs

Note. The event study graphs are based on the estimator of Callaway and Sant’Anna
(2021). H2B: highway access followed by electriﬁcation; N2E: no access to infrastructure
followed by electriﬁcation; N2H: no access to infrastructure followed by highway
provision.