Vanisha Sharma is a Post-doctoral Scholar at the Evans Policy and Research Group at the University of Washington, Seattle. Among other projects in climate and agriculture policy, Vanisha is leading a collaboration with the Borlaug Institute of South Asia, with support from the Gates Foundation, where she is using spatial econometrics and machine learning to map out the costs and benefits of climate adaptation strategies, and heterogeneity in climate adaptation by gender, for all countries in South Asia. You can read more about her work on her website.
The problem
Does faster internet rollout benefit everyone equally? While the arrival and spread of internet has been associated with increases in economic activity, it is challenging to tease out the causal welfare effects of internet because of endogeneity of internet supply given region-level income and demand, and variation in timing of the rollout.
How can we identify causal effects of internet roll-out on income inequality?
Non-random roll out of mobile phone towers makes identification of treatment effects difficult. This is mainly because mobile companies tend to estimate demand before roll out, thereby targeting more developed areas first, which in turn have higher levels of income and consumption. Identification comes from district-level variation in quasi-random roll out of 2G mobile phone towers. This rollout is quasi-random because while mobile phone companies target certain states and districts, within-district rollout is not pre-determined. In this analysis, since internet access eventually becomes universal, there is no conventional control group. Instead, in the same vein as Callaway and Sant’anna (2021), I use the ‘not-yet’ treated group for comparison with the treated group. The key identification assumption here is that the effect of unobservables would not systematically differ in the treated districts in the post-treatment period, in the absence of treatment. If this assumption holds, for each group-time unit, one can estimate the average treatment effect on treated by comparing average change in outcomes in the treated group to those in the not-yet treated group. This assumption can be tested similar to a 2×2 DID parallel trends set up, as explained in the next paragraph. I construct a variable denoting the number of years elapsed since the introduction of the first 2G tower in each district, akin to the ‘first-treated’ variable utilized in standard staggered Difference-in-Differences (DID) models as in (Callaway & Sant’Anna, 2021; Wooldridge, 2021; Sun & Abraham, 2021). This adds a multiple-period time dimension to the survey data that only has two time periods.
I test for pre-trends using the DID estimator by Callaway and Sant’Anna (2021) for district-level night-light intensity from the Socioeconomic High-Resolution Rural Urban Geographic dataset (SHRUG). More specifically, after disaggregating the staggered rollout into group-time pairs, I find that district level night lights intensity is not statistically different between each treated and control group-time pair before the arrival of 2G towers, indicating that the parallel trends assumption holds for this analysis. The parallel trends also take into account the gradual roll-out of 2G internet. I address other potential space- and time-varying effects by using extended two-way fixed effects as in Wooldridge (2021) for each treated group and time-period pair. Some examples of some space- and time-varying factors include population growth rates by state, and changing state policies on road infrastructure to name a few.
I supplement my analysis with an instrument to further address endogeneity concerns of internet roll out. I use topography as an instrument for cell phone tower construction. Previous research has shown that tower construction is a non-linear function of topography (Klonner & Nolen, 2010). Terrain ruggedness is strongly associated with exposure to treatment through demand for construction, and it is plausible to assume that it does not affect household-level income other than through the channel of construction. I find similar, though inflated, results using the 2SLS model in combination with the staggered DID.
What does the data show?
The data for this paper is a combination of four different sources, the Indian Human Development Survey (IHDS), Collins Bartholomew’s Mobile Coverage Explorer data, Socioeconomic High Resolution Rural-Urban Geographic (SHRUG) data, as well as OpencellId, an open database for global cellphone towers coverage. The main variable of interest for this analysis is the variation in treatment exposure, measured by the duration since the arrival of 2G mobile phone towers in each district. The main outcome variable is annual household income as reported in the household survey data from IHDS.
Since there are only two waves of the IHDS data, I am unable to test the parallel trends assumption for the aggregate district-level outcomes from the sample. However, since the treatment is at the district level, I use district-level DMSP night lights data as a proxy for annual income as in Henderson, Storeygard, and Weil (2012). Figure 1 below shows compelling evidence of trends statistically different between control and treated groups with increasing exposure to treatment.
Fig. 1: Pre-trend Test on Night Light Intensity with Temporal Exposure to Treatment.
There are two main results in this paper. First, I find that the richest quintile significantly benefits from internet rollout and witnesses significant increases in annual income over 5 years, of up to 75% of the base income in 2005. The group-time ATT increases over time. Second, the lower quintiles are significantly worse off, with much lower increases of up to 30% in the staggered DID model, and significant decreases in the 2SLS model. The magnitude of these effects first increases, and then decreases with exposure to 2G internet. See Figure 2.
Fig. 2: ATT By Income Quintiles
What are some mechanisms that may explain the results?
I also explore two potential mechanisms through which 2G internet may have this impact on household income. The first channel is through labor market outcomes. I find that households earning wages from non-farm or family businesses are significantly more likely to benefit from 2G internet than those earning from agriculture or salaried jobs. The increases in shares of business income are not matched by increasing number of hours worked in the business. This is suggestive of an income effect through cost reduction in fixed costs, search costs, and greater outreach to bigger customer base via better phone connection, as opposed to transitioning from other jobs to businesses (Jensen, 2007; Aker, 2010; Klonner & Nolen, 2010).
For the second potential pathway, I investigate whether exposure to the 2G internet is associated with migration. I utilize an index that classifies a household’s locality from highly rural to highly urban, graded on a scale of 1 to 4, as the dependent variable. The analysis reveals slight yet statistically significant upticks in the likelihood of residing in a more urban setting compared to the untreated group. Irrespective of income, as exposure to towers progresses beyond the three-year mark, the inclination towards migrating to more urban locales diminishes. This observation aligns with intuition, suggesting that improved connectivity in rural regions diminishes the necessity for relocation to urban areas (Aker & Mbiti, 2010).
What are some policy implications?
With the advent of 5G and subsequent technologies enabling faster internet, it is becoming even more important to understand the impacts of faster internet along the income distribution. The uneven distribution of economic benefits of internet technology underscores the need for targeted interventions to bridge the digital divide, ensuring that the socio-economic gains from internet access are equitably shared. This may involve tailored support for sectors and communities less able to capitalize on internet connectivity, alongside broader initiatives to enhance digital literacy and infrastructure in under-served areas.
References
Aker, J. C. (2010). Information from markets near and far: Mobile phones and agricultural markets in Niger. American Economic Journal: Applied Economics, 2(3), 46-59. https://doi.org/10.1257/app.2.3.46
Aker, J. C., & Mbiti, I. M. (2010). Mobile phones and economic development in Africa. Journal of Economic Perspectives, 24(3), 207-232. https://doi.org/10.1257/jep.24.3.207
Callaway, B., & Sant’Anna, P. H. C. (2021). Difference-in-differences with multiple time periods. Journal of Econometrics, 225(2), 200-230. https://doi.org/10.1016/j.jeconom.2020.12.001
Jensen, R. (2007). The digital provide: Information (technology), market performance, and welfare in the South Indian fisheries sector. The Quarterly Journal of Economics, 122(3), 879-924. https://doi.org/10.1162/qjec.122.3.879
Henderson, J. V., Storeygard, A., & Weil, D. N. (2012). Measuring economic growth from outer space. American Economic Review, 102(2), 994-1028. https://doi.org/10.1257/aer.102.2.994
Klonner, S., & Nolen, P. J. (2010). Cell phones and rural labor markets: Evidence from South Africa. In A. Banerjee, R. Benabou, & D. Mookherjee (Eds.), Understanding poverty (pp. 23-48). Oxford University Press.
Sun, L., & Abraham, S. (2021). Estimating dynamic treatment effects in event studies with heterogeneous treatment effects. Journal of Econometrics, 225(2), 175-199. https://doi.org/10.1016/j.jeconom.2020.09.006
Wooldridge, J. M. (2021). Two-way fixed effects, the two-way Mundlak regression, and difference-in-differences estimators. Econometrics Journal, 24(1), 1-14. https://doi.org/10.1093/ectj/utaa045
Feature Image from: https://www.utwente.nl/en/centrefordigitalinclusion/Blog/02-Digitale_Kloof/
