1 Introduction

While the older empirical literature on trade focused on the impact of trade liberalization on country-level outcomes, such as GDP, wages, inequality, unemployment, etc., a lot of recent work has started looking at the local effects of trade liberalization using more granular data.Footnote 1 This is motivated by the fact that factors of production are highly immobile across space and sectors (at least in the short run) and hence globalization and technology shocks have heterogeneous effects across space within a country. This paper contributes to this emerging literature by studying the impact of a large episode of trade liberalization in India on district-level development where development is proxied by nighttime lights.

It is a standard practice to use per capita GDP as a measure of overall welfare or development in empirical analysis; however, district-level per capita income data are not available for all districts regularly in India. As a result, several recent papers have started using satellite nighttime lights as a proxy for district-level GDP in India.Footnote 2 The nighttime light data have the key advantage of availability at higher levels of spatial disaggregation. As a result, they have been widely accepted as a proxy for overall economic development at the national and particularly sub-national levels where data are either not available or of questionable quality.Footnote 3

We use district-level tariff data before and after the episode of large trade liberalization in India which started in 1991. The sudden and unanticipated nature of this liberalization, which is crucial for identification, has been discussed extensively in the literature.Footnote 4 The district-level tariffs are constructed by weighting industry tariffs by the initial period share of the district’s workforce in that industry. Using nighttime lights per capita as a proxy for overall development at the district level in India, we find that the districts experiencing larger decreases in tariffs had larger increases in nighttime lights per capita. In terms of magnitude, the estimates from our benchmark instrumental variable specification suggest that a decline in tariffs of 6.9 percentage points (the average district-level tariffs declined by this amount) was associated with a 13% increase in nighttime lights per capita which translates into a 2.86% increase in GDP per capita.Footnote 5 We also find heterogeneity in the effect of trade liberalization on development. In particular, districts in states with flexible labor laws benefited more from trade liberalization than districts in other states. Additionally, districts with better road infrastructure and proximity to the coast benefitted more. Somewhat surprisingly, districts with better rail networks didn’t benefit more from trade liberalization. Finally, we also looked at the female labor force participation rate which is a strong correlate of development and found that districts with higher female labor participation rates benefited more from trade liberalization.

The remainder of our paper is structured as follows. In the next section, we provide a brief review of the related literature. Section III provides a discussion of our nighttime lights and tariffs data. Section IV provides the main empirical results and section V provides concluding remarks.

2 Related literature

Our work is related to the growing literature studying the local effects of globalization, such as Chiquiar (2008) for Mexico, Kovak (2013) for Brazil, and Autor et al. (2013) for the US. The study most closely related to our work is the influential paper by Topalova (2010) which finds that rural districts more exposed to trade liberalization experienced a slower decline in poverty and a lower increase in per capita consumption. Since per capita consumption data for urban areas are not available at the district level, the analysis for urban areas is conducted at the NSS regional level which is a higher level of aggregation and the results there are less clear cut. Compared to Topalova (2010), our proxy for development, nightlights per capita, does not distinguish between rural and urban areas within a district and finds a positive effect of trade liberalization on overall development. While the NSS data, on which Topalova (2010) relies, are based on a sample of 75,000 rural and 45,000 urban households spread over 466 districts (according to 1991 district boundaries) with considerable scope for sampling bias, our satellite data provide a comprehensive and uniform coverage of all districts.

Another closely related study is Hasan et al. (2007) which studies the impact of trade liberalization on poverty in India using state-level data. Unlike Topalova (2010), they do not find evidence of a slower decline in poverty in states experiencing larger trade liberalization. Both Topalova (2010) and Hasan et al. (2007) find the gains from trade liberalization to be positively associated with labor market flexibility, a finding that we also confirm with our nightlights data. Hasan et al. (2007) also find that trade liberalization has a larger negative effect on poverty in states with a greater density of roads. This is broadly consistent with our finding of trade liberalization having a larger positive effect on nightlights per capita in districts with greater road length per capita. Neither of these papers studies how the length of railroads or the female labor force participation rates affect the impact of trade liberalization.

In a recent paper, Chor and Li (2021) study the impact of the US-China tariff war using high-frequency grid-level nighttime lights data. Using within-grid variation over time, similar to our identification strategy of within-district variation over time, they find that a 1 percentage point increase in exposure to the US tariffs was associated with a 0.59% reduction in nighttime lights.

Our paper is also related to the broader literature that has established a strong positive relationship between nighttime lights and GDP at both national and sub-national levels. Donaldson and Storeygard (2016) provide a survey of this literature. In the Indian context, Prakash et al. (2019) use nightlights as a proxy for economic activity to show that electing criminally accused politicians reduces the growth of economic activities significantly. Chodorow-Reich et al. (2020) use it to show the adverse effect of the 2016 Indian demonetization episode on economic activity. Chanda and Kabiraj (2020) use district-level nightlights per capita data to show the convergence in development over the period 1996–2010. Jha and Talathi (2021) show the adverse effects of direct British rule (compared to districts under indirect rule) on development and growth since the early 1990’s using district-level data on nightlights per capita.

3 Data

We conduct our analysis at the district level (1991 Census) within India. As per the 1991 Census of India, India was divided into 466 districts that constitute 32 states.Footnote 6 We follow Topalova (2010) and focus our analysis on 394 districts spread across 17 major states of India.Footnote 7 Since we obtain much of our data from Topalova (2010), who conducts her analysis as per the 1987–88 NSS districts, we match the districts in her analysis to the 1991 Census districts following the methodology of Castelló‐Climent et al. (2018) and Jha and Talathi (2021).Footnote 8 Summary statistics of the variables used in our analysis are presented in Table 1.

Table 1 Summary statistics
Full size table

3.1 Nighttime lights per capita

The Earth Observation Group (EOG) of the Payne Institute for Public Policy at the Colorado School of Mines maintains images of lights generated from the earth’s surface at nighttime that are captured by satellites of the National Oceanic and Atmospheric Administration (NOAA). These images are available at an annual frequency for years spanning 1992–2021. Of the many nighttime light data products available, for our analysis, we are mainly interested in nighttime light data from 1992 to 2001. The ‘stable’ nighttime lights product of the DMSP-OLS (Defense Meteorology Satellite Program—Operational Linescan System) data series is widely used in studies whose analysis spans 1992–2013.Footnote 9 For each year, the EOG makes available an annual composite image of lights generated from the earth’s surface at nighttime. For the years 1994 and 1997–2007, there are two images available for each of these years—one captured by the newer satellite and the other captured by the older satellite. Each image housed by the EOG consists of billions of pixels with each pixel emitting varying intensities of nighttime lights. These pixels are scaled to a 30-arc-second grid between 65° south and 75° north latitudes. A digital number (DN) is used to measure the luminosity of nighttime lights emitted from each pixel. The DN in the ‘stable’ nighttime lights data product ranges from 0 to 63 with a DN of 0 implying almost no light emitted and 63 representing the highest possible intensity of nighttime lights that can be emitted from a pixel.

One issue with the ‘stable’ nightlights data product is that it is top-coded at a DN of 63 and bottom-censored at a DN of 0. The top-coding and bottom-censoring prevalent in the ‘stable’ nightlights data product limits variation in intensities of nighttime lights across the brightest and dimmest areas.Footnote 10 In light of this, we obtain nighttime lights data for the years 1993 and 2001 from Bluhm and Krause (2022) (BK hereon) who correct the ‘stable’ nighttime lights images for top-coding and bottom-censoring.Footnote 11 Ideally, nighttime lights data from the pre-liberalization period, such as 1987 or 1990, would be better suited for our analysis, however, they are only available starting in 1992. Following Jha and Talathi (2021) who cite issues related to the quality of the 1992 nighttime lights data, we use data from 1993 instead as a proxy for pre-liberalization development levels.

We superimpose the 1991 district boundary shapefile of India with the respective nighttime lights images in ArcGIS and obtain the sum of lights statistic for 1993 and 2001 for each district.Footnote 12,Footnote 13 This statistic represents the total intensity of nighttime lights emitted from each district. We divide this statistic by the district population to obtain the Nightlights Per Capita measure which serves as our proxy for per capita income at the district level.Footnote 14 Figure 1A shows the 1993 raw image of nighttime lights in India and Fig. 1B shows the same for 2001. It is apparent from a visual inspection of these figures that nighttime lights increased significantly over this period. Figure 2A and B shows the distribution of nightlights per capita in 1993 and 2001 across districts in India, respectively. It turns out that on average nightlights per capita increased by 47% during this period. Some studies, such as Henderson et al. (2012), Castelló-Climent et al. (2018), and Chor and Li (2021), use Nightlights Density (Nightlights Per Area) instead as a proxy for economic activity. In light of this, we replicate our baseline results with Nightlights Density as the dependent variable and validate the robustness of our main results. Figure 3A and B shows the variation in Nightlights Density across districts in India.Footnote 15 There is an increase of 77% in the average Nightlights Density over this period which is much larger than the increase in nightlights per capita because the area remained fixed while the population grew rapidly over this period.

Fig. 1
figure 1figure 1

A Raw image of nighttime lights in 1993 for India. B Raw image of nighttime lights in 2001 for India

Full size image
Fig. 2
figure 2figure 2

A Distribution of log lights per capita in 1993 across districts in India. B Distribution of log lights per capita in 2001 across districts in India

Full size image
Fig. 3
figure 3figure 3

A Distribution of log lights density in 1993 across districts in India. B Distribution of log lights density in 2001 across districts in India

Full size image

3.2 Trade liberalization data

We obtain district-level data on tariffs and non-tariff barriers from Topalova (2007, 2010). From publications of the Ministry of Finance, India, Topalova (2007, 2010) has manually digitized detailed tariffs at the 6-digit level of the Indian Trade Classification Harmonized System (HS) for approximately 5000 product lines. She obtains a measure of average sector-level tariffs by matching the 5000 product lines to the NIC codes using the concordance of Debroy and Santhanam (1993). Topalova (2010) compiles data on Non-Tariff Barriers (NTBs) for 1997 (post-reform period) from the publication EXIM Policy, Directorate General of Foreign Trade, India. The NTBs measure is constructed as the share of products within a production sector that can be imported without any license. At the district level, the NTBs measure is an average weighted by the industry-specific employment share in that district.

4 Empirical specification and results

4.1 Estimation strategy

We use a simple two-way fixed effects model (district and time fixed effects) to identify the effects of trade liberalization on local development. Our key estimating equation is below.

$$y_{d,t} = \alpha + \beta Tariff_{d,t} + \delta_{d} + POST_{t} + \varepsilon_{d,t}$$
(1)

where \({y}_{d,t}\) is the natural log of nightlights per capita for district d in year t, \({Tariff}_{d,t}\) is the district level tariff in year t, \({\delta }_{d}\) is the district fixed effect and \({POST}_{t}\) is the time fixed effect. A positive estimate of \(\beta\) will indicate that districts experiencing greater trade liberalization experienced smaller increases in nightlights per capita while a negative estimate of \(\beta\) will indicate the opposite.

We use district-level tariff data from Topalova (2007, 2010) which is constructed as the average of industry-level tariffs weighted by the industry-specific employment shares of that district in 1991 as shown in Eq. 2 below.Footnote 16

$$Tariff_{d,t} = \mathop \sum \limits_{i} \gamma_{i,d,1991} *Tariff_{i,t}$$
(2)

In Eq. (2), \({\gamma }_{i,d,1991}\) represents the industry (i) specific employment shares in district d in 1991. This is computed as \({\gamma }_{i,d,1991}=\frac{{Workers}_{i,d,1991}}{{Workers}_{d,1991}}\) where \({Workers}_{d,1991}\) is all workers in district d.

Since the industry-specific employment shares are available separately for the rural and urban parts of a district, Topalova (2007) computes rural and urban tariffs for each district separately.Footnote 17 Our tariff for a district is the weighted average of the rural and urban tariffs for the district where weights are the shares of the rural and urban population, respectively, in the district in 1991.

While \({Tariff}_{d,t}\) correctly measures the inverse exposure of a district to trade in the sense that higher \({Tariff}_{d,t}\) implies less exposure, since non-traded goods are assigned a tariff of zero in this calculation, a causal interpretation of the effect of trade liberalization on development using this measure may be problematic for reasons discussed in detail in Topalova (2010). For example, a poor district may have a large share of workers in the non-traded sector resulting in a lower tariff in the initial period and a smaller trade liberalization (since tariffs were reduced in such a way to bring them to a uniform level). Now, if poorer districts tend to grow faster leading to convergence (a result that we verify in our context),Footnote 18 even in the absence of a causal effect of trade liberalization our OLS estimates will show that districts with lower trade liberalization grew faster suggesting a negative effect of trade liberalization on development (a positive \(\beta\) coefficient).Footnote 19 Therefore, following Topalova (2010), we use \({TrTariff}_{d,t}\) which is calculated by ignoring the workers in the non-traded sector:

$$TrTariff_{d,t} = \mathop \sum \limits_{i} \gamma_{i,d,1991} *Tariff_{i,t}$$
(3)

where \({\gamma }_{i,d,1991}=\frac{{Workers}_{i,d,1991}}{\sum_{i}{Workers}_{i,d,1991}}\), that is the denominator includes only workers in the traded sectors rather than all workers.Footnote 20 Since this measure does not depend on the share of workers in the non-traded sector and is highly correlated with \({Tariff}_{d,t}\), we use it as an instrument for \({Tariff}_{d,t}\). The first-stage results are shown in Appendix Table 4 where \({Tariff}_{d,t}\) is regressed on \({TrTariff}_{d,t}\) in column 3 and \({TrTariff}_{d,t}\) and its interaction with the dummy for the post-trade liberalization year in column 4. The coefficients of \({TrTariff}_{d,t}\) and its interaction with the post-liberalization time dummy are large and highly significant in both columns.

4.2 Empirical results

Our baseline results are presented in Table 2. In column 1, the OLS coefficient of \({Tariff}_{d,t}\) is positive but insignificant. As mentioned earlier, given that we find convergence in nightlights per capita across districts and the fact that poorer districts experienced smaller declines in tariffs, this by itself would suggest a positive \(\beta\). In column 2 we regress nightlights per capita on \({TrTariff}_{d,t}\) which can be considered a reduced form regression as we are going to use \({TrTariff}_{d,t}\) as an instrument for \({Tariff}_{d,t}\). We find a strong negative relationship between \({TrTariff}_{d,t}\) and nightlights per capita suggesting that districts experiencing greater trade liberalization had a larger increase in nightlights per capita. Column 3 presents the first IV results and here we find the estimate of \(\beta\) to be negative and significant suggesting a causal relationship between trade liberalization and nightlights per capita. That is, districts with greater trade liberalization experienced more rapid increases in nightlights per capita. Since the average decline in tariffs was 6.9 percentage points,Footnote 21 the coefficient of \({Tariff}_{d,t}\) in column 3 of − 1.78 implies a 0.069*1.78 increase in log nightlights per capita or 13% increase in nightlights per capita. Using an estimate of the elasticity of GDP per capita with respect to nightlights per capita of 0.22, this translates into 2.86% more GDP per capita. That is, districts that experienced a 6.9% decline in tariffs experienced a relative increase in GDP of 2.86%.

Table 2 Effect of trade liberalization on local development
Full size table

To alleviate the concern that changes in tariffs may be correlated with time-varying shocks that affect the level of development, in column 4, we include some initial district conditions interacted with the post-liberalization time dummy.Footnote 22,Footnote 23 The coefficient of \({Tariff}_{d,t}\) goes down but remains negative and highly significant. In column 5, we add non-tariff barriers as an additional regressor. The coefficient of \({Tariff}_{d,t}\) goes down in magnitude but remains large and negatively significant. As argued by Topalova (2010), non-tariff barriers were removed more slowly and varied a lot across goods. Therefore, their inclusion should be regarded as a robustness exercise and the coefficient of NTBs should not be given a causal interpretation.

Since trade liberalization was not the only reform, our estimates may be picking up the effects of some other reforms, such as delicensing, liberalization of FDI, and banking reforms. In column 6, we include these time-varying measures of reforms.Footnote 24 Compared to column 4, the coefficient of \({Tariff}_{d,t}\) does not change much. Finally, in column 7 we instrument \({Tariff}_{d,t}\) using both \({TrTariff}_{d,t}\) and its interaction with the post-liberalization time dummy. The coefficient of \({Tariff}_{d,t}\) drops in magnitude but remains negative and significant.

While we have used nightlights per capita as our proxy for development as is more common in the literature, some papers use nightlight density measured by the sum of nightlights divided by the area of the region as a proxy for development. In Table 5 in the appendix, we replicate all the regressions reported in Table 2 by replacing the dependent variable with nightlight density. The results are very similar to those reported in Table 2.

While we included \({TrTariff}_{d,t}\) as the measure of tariff in column 2 of Table 2 and called it our reduced form regression, Hasan et al. (2007) use this as their measure of tariff in their estimation using state-level data. For comparison with Hasan et al. (2007) as well as to show the robustness of our results to this alternative measure of tariff, in appendix Table 6, we report the results when we include initial district characteristics as well as controls for other reforms. The coefficient of \({TrTariff}_{d,t}\) is negative and significant in all cases as shown in the first 4 columns of Table 6 in the appendix.

Next, we look at some channels through which trade liberalization may affect development. This will also allow us to study if the local effects of trade liberalization are heterogeneous.

4.3 Heterogeneous effects of trade liberalization

Both Topalova (2010) and Hasan et al. (2007) found that the impact of trade liberalization on poverty varied depending on whether a state had flexible or rigid labor laws. Using Besley and Burgess’s (2004), classification of states into flexible and non-flexible with regard to labor laws where the dummy Flexible Labor Laws takes the value 1 for flexible law states and 0, otherwise we add an interaction of this dummy to \({Tariff}_{d,t}\) and report the results in columns 1 and 2 of Table 3.Footnote 25 Column 1 adds this interaction to the regression reported earlier in column 6 of Table 1 where \({Tariff}_{d,t}\) was instrumented by \({TrTariff}_{d,t}\). The interaction coefficient in column 1 of Table 3 is negative. This along with the coefficient of \({Tariff}_{d,t}\) itself implies that the impact of trade liberalization on nightlights per capita is much larger in districts in states with flexible labor laws compared to districts in other states. A similar result is obtained in column 2 when \({Tariff}_{d,t}\) is instrumented by \({TrTariff}_{d,t}\) and its interaction with the post-liberalization dummy.

Table 3 Heterogeneous effects of trade liberalization
Full size table

Next, we test if the impact of trade liberalization varies according to the distance from the coast.Footnote 26 The motivation behind using this variable is that proximity to the coast will reduce the cost of trading goods internationally and, therefore, the impact of trade liberalization on domestic prices could differ across districts depending on their proximity to the coast. As a consequence, trade liberalization can have differential effects on districts depending on their proximity to the coast. The results reported in columns 3 and 4 of Table 3 suggest that the positive effect of trade liberalization on nightlights per capita is larger the shorter the distance from the coast. Based on the estimated coefficients in column 3, we can conclude that a 6.9 percentage point reduction in tariffs at a district on the coast leads to a relative increase of 24.6% increase in nightlights per capita.Footnote 27

Next, we look at the implications of road length per capita and railroad length per capita and report the results in columns 5 to 8 of Table 3.Footnote 28 For road length per capita, the coefficients in columns 5 and 6 imply that the larger the road length per capita the greater the positive effect of trade liberalization on nightlights per capita. The coefficients in column 5 imply that the impact of trade liberalization on nightlights per capita is positive as long as the road length per capita exceeds -0.46 standard deviation below the mean.Footnote 29 Hasan et al. (2007) find that trade liberalization has a larger negative effect on poverty in states with a greater density of roads. This is broadly consistent with our finding of trade liberalization having a larger positive effect on nightlights per capita in districts with greater road length per capita. The coefficients of the interaction between \({Tariff}_{d,t}\) and railroad length per capita reported in columns 7 and 8 are positive. The positive sign implies that a better railroad network reduces the benefit of trade liberalization, however, the impact of trade liberalization on nightlights per capita remains positive.Footnote 30

Finally, we use the female labor force participation rate which is a strong correlate of development and remains very low in India compared to other countries at similar levels of development. While the average rate of 38% is quite low, it varies from an abysmally low rate of 3% to a high rate of 84%.Footnote 31 We use the interaction of female labor participation rate with \({Tariff}_{d,t}\) as an additional regressor and report the results in columns 9 and 10 of Table 3. In both cases, the interaction coefficient is negative and significant suggesting that trade liberalization is beneficial for districts with high rates of female labor force participation. A similar result is obtained when we use the share of females in the workforce.

Note that all the regressions in Table 3 include initial district characteristics interacted with post-liberalization dummy and other reform controls. The results when we do not include other reforms are presented in Table 7 in the appendix and are qualitatively similar to those presented in Table 3. Also, when we use nightlight density as an alternative measure of development and replicate the regressions in Table 3 and appendix Table 7, we get very similar results. That is, all our results presented in the paper are robust to the use of nightlight density as the dependent variable.

5 Concluding remarks

We have studied the local effects of a large episode of trade liberalization in India starting in 1991 using nightlights per capita as a proxy for overall development at the district level. Our results suggest that on average trade liberalization increased nightlights per capita in districts across India but there was considerable heterogeneity in the effect. In particular, districts in states with flexible labor laws or districts with better road networks or proximity to the coast, or higher female labor force participation rates seem to have benefited more compared to other districts.