3.1 Measures of distance from the GQ and NSEW highways

We have merged geo-coded data on the location of the GQ and NSEW networks into the district database.Footnote ¹ This information is used to calculate the distances of district centroids (the geographic center of the district area) from the nearest points on the GQ and NSEW networks. Figure 2 plots the highway networks (excluding parts of NSEW that were not built by 2010) and the distance of every district centroid from its nearest point on each highway. We also categorize districts into four distance bands from each highway: nodal (major metropolitan area at which the highways start and end),Footnote ² 0–40 km from the highway, 40–100 km from the highway, and more than 100 km from the highway.Footnote ³ Thus, there are eight distance bands in total.

Figure 2.

GQ and NSEW highway networks, with distances from district centroids. (a) Golden Quadrilateral (GQ). (b) North-South-East-West Highway.

Notes: GQ = Golden Quadrilateral; NSEW = North-South-East-West Highway. The green lines depict the shortest distance from district centroids (blue dots) to the actual highway placement (red lines).

Table A4 in the online appendix shows the distribution of districts across these distance bands. As explained later in the methodology section, we use these bands to assign districts to the ‘treatment’ and ‘control’ groups in the DiD estimation. Specifically, the 0–40 km distance band from the GQ (or the NSEW) identifies the GQ (or NSEW) treatment districts, while the control districts are those more than 100 km away from both highways. Table A4 therefore shows that our sample contains 72 GQ treatment districts and 40 NSEW treatment districts. There are 194 districts in the common control group. There is little overlap between districts proximate to GQ and those proximate to NSEW; it would be difficult to distinguish between the impacts of these highways were this not the case.

Figure A1 in the online appendix depicts hypothetical straight-line versions of the GQ and NSEW highways used in robustness checks. The nodal cities used to construct the straight-line version of the GQ are Delhi, Kolkata, Chennai, Bangalore, and Mumbai. The nodes used to construct the NSEW straight-line versions are cities at its northern, western, and southern extremities (Jalandhar, Porbandar, and Kanniakumari, respectively), and Jhansi in central India (where the East-West and North-South arms of the NSEW cross). We ignore the arm of NSEW going east from Jhansi because it was largely unbuilt in 2010.Footnote ⁴

Table A5 (online appendix) shows the distribution of districts according to their distance bands from the actual versus straight line versions of GQ and NSEW. In general, there is a strong but less than a perfect overlap in these distributions.

4.1 Estimating the average impacts of the highways

Formally, the underlying regression specification can be described as follows:

(1)\begin{equation}{Y_{i,t}} = \beta \times \textrm{Highwa}{\textrm{y}_i} \times \textrm{Post}_t^{\textrm{Highway}} + {\emptyset _i} + {\varphi _t} + {\varepsilon _{i,t}} \end{equation}

This regression is estimated on district-level panel data. Here, Y_i,t is an outcome of interest in district i and year t. The dummy variable $\textrm{Post}_t^{\textrm{Highway}}$ is equal to one in years after the highway completion, and zero in years prior to that. The dummy variable $\textrm{Highwa}{\textrm{y}_i}$ is equal to one in districts close to the new highways (the treatment districts) and zero otherwise. ${\emptyset _i}$ is a set of district fixed effects that controls for time-invariant district-level factors, and ${\varphi _t}$ is a set of year dummies that control for unobserved time-varying factors common to all districts. The impact of the highways is estimated by $\beta$, the coefficient on the treatment variable (the interaction $\textrm{Highwa}{\textrm{y}_i} \times \textrm{Post}_t^{\textrm{Highway}}$), which measures how the change in the outcome after the highway was built differed across control and treatment districts.

We adjust this basic specification to account for the fact that we are simultaneously estimating the impacts of two highway networks, the GQ and the NSEW. There are two factors to consider in this regard. First, in estimating the impact of either highway network, it is important to control for the presence of the other one. Second, the two networks could have had different impacts. In other words, there were two sets of treatment districts: those near GQ, and those near NSEW.

We assign districts into distance bands based on the proximity of the district centroid to the GQ. The bands are: more than 100 km from the nearest GQ point, 40–100 km from the GQ, 0–40 km from the GQ, and nodal districts.Footnote ⁵ We then interact the indicators for each GQ distance band with a variable indicating the years after the GQ was built. We repeat this process for the NSEW and include both sets of interactions on the RHS. Thus, the specification we estimate is as follows:

(2)\begin{align}{Y_{i,t}} = {\beta ^{\textrm{GQ}}} \times \textrm{G}{\textrm{Q}_i} \times \textrm{Post}_t^{\textrm{GQ}} + {\beta ^{\textrm{NSEW}}} \times \textrm{NSE}{\textrm{W}_i} \times \textrm{Post}_t^{\textrm{NSEW}} + {\emptyset _i} + {\varphi _t} + {\varepsilon _{i,t.}}\end{align}

here, $\textrm{G}{\textrm{Q}_i}$ (respectively, $\textrm{NSE}{\textrm{W}_i}$) is a vector of dummies indicating the distance band from the GQ (respectively, NSEW) to which district i belongs, while $\textrm{Post}_t^{\textrm{GQ}}$ (respectively, $\textrm{Post}_t^{\textrm{NSEW}}$) is a dummy equal to one in the years after GQ (respectively, NSEW) completion. The omitted distance band dummy corresponds to districts more than 100 km from the highway (GQ or NSEW). ${\emptyset _i}$ is a set of district fixed effects, and ${\varphi _t}$ is a set of year dummies (or state-year dummies).

Because GQ and NSEW construction commenced after 2001, the indicators $\textrm{Post}_t^{\textrm{GQ}}$ and $\textrm{Post}_t^{\textrm{NSEW}}$ are set equal to zero in the baseline year of our two-period panel (2000–01) and one in the endline year (2010–11). To account for the fact that, unlike GQ, NSEW was not fully completed by 2010–11, only those segments of NSEW that were complete as of 2010 are considered when assigning districts to distance bands around NSEW.

The impact of the GQ is measured by the ${\beta ^{\textrm{GQ}}}$ corresponding to the 0–40 km distance band from GQ, to be denoted by ${\beta ^{\textrm{GQ}, 0 - 40 }}$ hereafter. Because we are controlling for $\textrm{NSE}{\textrm{W}_i} \times \textrm{Post}_t^{\textrm{NSEW}},{\beta ^{\textrm{GQ,}0 - 40}}$ in effect measures how the post-GQ change in the outcome differed between districts 0–40 km from GQ (the GQ treatment group) and districts more than 100 km from both highways (the control group). Similarly, the impact of NSEW is measured by the ${\beta ^{\textrm{NSEW}}}$, corresponding to the 0–40 distance band from the NSEW, denoted by ${\beta ^{\textrm{NSEW}, 0 - 40 }}.$ Our main results tables thus report these two $\beta$ s.

In our specification, we use flexible state-year fixed effects instead of a common year fixed effect. This preferred specification controls for unobserved state-level differences in growth – to account for the documented divergence in economic growth across Indian states (GOI, 2017).

4.2 Estimating conditional impacts: could highway impacts have depended on local market conditions such as initial education levels?

We also test the hypotheses that the impact of the highways depended on initial conditions in districts, focusing on the initial rates of educational attainment in the district. Roberts et al. (Reference Roberts, Melecky, Bougna and Xu2020) and Alam et al. (Reference Alam, Herrera Dappe, Melecky and Goldblatt2022) use a simple policy model to argue that gaining a greater understanding of the heterogenous (or ‘conditional’) impacts of transport corridor investments – including highways – is needed to better inform policy decisions on the design of corridor investment programs. However, they also document that attempts to estimate heterogenous impacts of connectivity are scarce in the literature.

Our regression estimation thus exploits the information on varying initial conditions across districts and adopts a difference-in-difference-in-difference approach by interacting the treatment variable $(\textrm{Highwa}{\textrm{y}_i} \times \textrm{Post}_t^{\textrm{Highway}})$ with variable(s) capturing initial conditions in districts:

(3)\begin{align} {Y_{i,t}} & ={\beta ^{{\rm GQ}}} \times {\rm G}{{\rm Q}_i} \times {\rm Post}_t^{{\rm GQ}}\textrm{ + }{\beta ^{{\rm NSEW}}} \times {{\rm NSEW}_i} \times {\rm Post}_t^{{\rm NSEW}}\nonumber\\ & \quad + {\delta ^{{\rm GQ}}} \times {\rm G}{{\rm Q}_i} \times {\rm Post}_t^{{\rm GQ}} \times {Z_i} + {\delta ^{{\rm NSEW}}} \times {{\rm NSEW}_i} \times {\rm Post}_t^{{\rm NSEW}} \times {Z_i}\nonumber\\ & \quad + {\gamma _1}{\rm Post}_t^{{\rm GQ}} \times {Z_i} + {\gamma _2}{\rm Post}_t^{{\rm NSEW}} \times {Z_i} + {\emptyset _i} + {\varphi _t} + {\varepsilon _{i,t}}. \end{align}

here, ${Z_{i }}$ is a vector of initial conditions of interest in district i. The effect of initial conditions on the impact of the highways is estimated by the $\delta$ s, the coefficients on the triple interaction term between $\textrm{Highwa}{\textrm{y}_i} \times \textrm{Post}_t^{\textrm{Highway}}$ and the ${Z_{i }}$ s.

For illustration, suppose that the ${Z_{i }}$ in question is a variable measuring initial rates of educational attainment. The corresponding ${\delta ^{\textrm{GQ}}}$ (respectively, ${\delta ^{\textrm{NSEW}}}$) coefficient measures how the impact of the GQ (respectively, the NSEW) depends on the local level of educational attainment. A positive estimate of this ${\delta ^{\textrm{GQ}}}$ would imply that the impact of the GQ on the outcomes of interest was more positive in districts with higher initial rates of educational attainment.

4.3 Examining robustness to endogenous highway route: two-stage least squares estimation using instrumental variables

Ordinary least squares (OLS) estimates of the highway impacts could be biased if the placement of the highways was correlated with unobserved factors affecting local developmental outcomes. For example, it could be that the path of the highways was deliberately tilted toward locations with good growth prospects. Datta (Reference Datta2012) argues that such endogenous placement is not a major concern with the GQ and NSEW projects because they were largely highway upgrade projects and, as such, their routes were pre-determined by existing highway segments. While we agree with this reasoning, there remains a concern that in parts of the highway networks, there was room to choose between alternative pre-existing segments.

As in Ghani et al. (Reference Ghani, Goswami and Kerr2016), we address this concern by employing a two-stage least squares (2SLS) estimation strategy using IVs as a robustness check. To qualify as an instrument, a variable must be correlated with the treatment variable but uncorrelated with unobserved factors affecting the outcomes being considered. Because the highways were intended to connect certain pre-specified nodal cities, the straight-line paths connecting those nodes are close to the actual paths of the highways while arguably being exogenous to unobserved drivers of developmental outcomes in non-nodal districts. Hence, our chosen IVs measure district proximity to hypothetical straight-line versions of the GQ and NSEW highways.

For illustration, consider the hypothetical straight-line version of the GQ, constructed by replacing each highway segment connecting a pair of GQ nodal cities with the straight line joining those nodes. We generate a dummy variable indicating districts whose centroids are within 40 km of this hypothetical highway network. The interaction of this variable with $\textrm{Post}_t^{\textrm{GQ}}$ instruments for the GQ treatment variable (that is, it instruments for the interaction of $\textrm{Post}_t^{\textrm{GQ}}$ with a dummy indicating that the district is within 0–40 km of the actual GQ). A similar procedure generates the IV for the NSEW treatment variable.

5.1 Average wider economic impacts of the highways

Table 1 reports the estimation results for equation (2), our baseline DiD specification measuring the average impact of the highways. In line with prior research on the impact of the GQ on industrial activity (Ghani et al., Reference Ghani, Goswami and Kerr2016), the estimation results suggest that the GQ highway had a statistically significant positive impact on district output per capita. Looking at column (1) of table 1 with the results for GDP per capita (in logs), the point estimate of ${\beta ^{\textrm{GQ}, 0 - 40 }}$ implies that the highway increased GDP per capita by about 4 per cent.

Table 1.

Impacts of highways on income and environmental quality

Notes: The OLS estimation is based on four distance bands on the proximity of district centroid to highways (0–40 km, 40–100 km, more than 100 km, and nodal districts) interacted with post-treatments. The omitted (comparator) distance bands are those exceeding 100 km. Only the coefficients on the 0–40 km distance bands are shown for conciseness. Standard errors clustered by district in parentheses.

As for the highways' impacts on environmental measures (table 1, columns (3) and (4)), it appears that the GQ highway led to an increase in air pollution related to PM. Specifically, the GQ is estimated to have increased AOT by approximately 0.03 points. When compared to the mean increase of AOT (0.07 points) across all districts during this period, this impact is highly significant in magnitude. We do not detect a significant impact on NO₂ levels in the air in the districts in the vicinity to the GQ highway (0–40 km).

Because wind disperses particulates in the air, air pollution could have also increased downwind of the highway. To examine this, we first determine the annual average downwind direction for each GQ (0–40) district by analyzing the hourly wind dataset obtained from the European Center for Medium-Range Weather Forecasts' Climate Data Store. Utilizing the u- and v-components of the wind at various locations within a district, we find the median wind angle over the course of a year and average it for the years 2000 and 2001. The wind angle at the district level remains highly consistent from year to year, exhibiting a correlation exceeding 0.9 between both years.

We then group the GQ (40–100) districts into two categories: (i) ‘downwind districts’ are those whose centroid direction from a nearby GQ (0–40) district is within 45 degrees of the wind direction of that GQ district; and (ii) ‘near GQ districts’ are the remaining districts in the GQ (40–100) group. Online appendix table A6 shows that, compared to the control districts (those far away from GQ and NSEW, as in the rest of our estimations), there was a statistically significant increase in particulate matter pollution in districts near and downwind of GQ districts (the coefficient on Post GQ* Downwind in column (2)), but not in other districts near GQ districts (the coefficient on Post GQ* Near in column (2)). We do not observe a similar downwind impact on GDP. Hence, this result indicates that, because PM disperses in the air, ignoring downwind districts may be missing the environmental spillover of the GQ on nearby districts. It suggests that the GQ's positive impact on income was more geographically concentrated than its negative impact on environmental quality. Moreover, this pattern of spillovers also serves as a robustness check on our main result: if our result is spuriously driven by some unobserved factor that happens to be spatially correlated with GQ, then we would not expect its spatial spillovers to depend on wind direction specifically from the GQ district.

Overall, these average findings raise a question of whether some measures or local market conditions could be identified to mitigate the tradeoff between income growth and worsening air pollution. We focus on this important question next.

5.2 Heterogenous impacts: can human capital help mitigate the income–environment tradeoff?

To examine heterogeneity in the highways' impact, we estimate the triple-difference specification in equation (3) by OLS. Recall that we are interested in estimating the coefficients on the interactions between the treatment $(\textrm{Highwa}{\textrm{y}_i} \times \textrm{Post}_t^{\textrm{Highway}})$ and variables ${Z_i}$ that capture the initial conditions. Our broad hypothesis is that, as predicted by a Heckscher–Ohlin type model of comparative advantage and trade, the impact of better market access may depend on certain factor endowments such as educational attainment and cropland (which are immobile in the short to medium term). If this is correct, then low average levels of factor endowments could explain why the average district may not experience wider economic benefits and incur wider economic costs from the highway construction. From the policymaker's perspective, identifying such complementary factors can reveal how the highway investment programs could be enriched with complementary public interventions to maximize the wider economic benefits and mitigate the wider economic cost generated by these investments.

The estimation results in table 2 first consider the interactions of highway with educational attainment, as measured by the secondary school completion rate (the variable Educ). No statistically significant heterogeneity impact of connectivity on GDP per capita is found depending on the secondary school completion rate: the coefficient on the triple interaction terms with Educ is statistically not significant (table 2, column (1)).Footnote ⁶ By contrast, for the PM air pollution (measured by AOT), the coefficient on the interaction of PostGQ*GQ (0–40) with Educ is negative and statistically significant at the 5 per cent level (table 2, column (2)). This finding suggests that higher educational attainment mitigated the increase in PM air pollution caused by the GQ. The magnitude of this heterogeneity is sizable, with the estimate implying that moving from a district at the 25^th percentile of the secondary school completion rate (17 per cent) to one at its 75^th percentile (29 per cent) would reduce the impact of the GQ on AOT by 0.02 percentage points (compared with the average AOT impact of 0.06 points shown in table 1, column (2)).

Table 2.

Heterogenous Impacts of highways

Notes: The OLS estimation is based on four distance bands on the proximity of district centroid to highways (0–40 km, 40–100 km, more than 100 km, and nodal districts) interacted with post-treatments. Only the coefficients on the 0–40 km distance bands are shown for conciseness. The omitted (comparator) distance bands are those exceeding 100 km. Standard errors in parentheses. Standard errors are clustered at the district-level.

Next, we include an interaction of the highway treatment variables with a measure of the share of cultivated land in the total district area (the variable Cropland). We are interested in this interaction term because the impact of the highways could also depend on the comparative advantage in agriculture (proxied by Cropland).

Interestingly, for the AOT outcome variable (table 2, column (5)), the coefficients on the interactions of PostNSEW*NSEW(0–40) and PostGQ*GQ (0–40) with Cropland are positive and statistically significant at the 1 and 10 per cent levels, respectively. These estimates indicate that the income–air pollution tradeoff was worse in more agricultural districts.