Literacy in British India

As British rule spread in India, former indigenous schools were either replaced or incorporated into the new colonial education system. This was a slow and uneven process that was largely complete by the end of the nineteenth century (Nurullah and Naik 1951). Public education funding, which was meager in the 1850s when the British Crown took control from the East India Company, increased from 1.5 percent of the colonial budget in 1882 to 5.2 percent in 1921, a sum that was still below 1 percent of national income (Chaudhary 2016). Did the transition to colonial schooling increase literacy? Unfortunately, we cannot answer that question because there are few comparable estimates of literacy before the 1880s. We know indigenous village schools were common in the early nineteenth century. They attracted boys from different backgrounds, but few girls (Rao Reference Rao2020). Yet, these accounts offer few specifics on literacy. One notable exception is a Scottish missionary named William Adam. He estimates literacy was around 4 percent (ability to read and write) to 6 percent (ability to sign) across a handful of districts in eastern India in the 1830s (Adam and Long Reference Adam and Long1868). It is hard to extrapolate, however, from these estimates because we do not know if Adam’s districts were positively or negatively selected compared to the Indian average.

Beginning with the 1881 census, we know that male literacy increased from 6 percent in 1881 to 11 percent in 1921, while female literacy increased from under 1 percent to 2 percent (Online Appendix Table A2). Literacy increased because more children went to school. Indeed, enrollment increased faster than literacy, from one in ten children attending school in 1891 to just over one in five in 1921 (Chaudhary 2016). This is not to say people did not learn to read and write outside of formal schools. Rather, schools offered a natural venue to acquire functional literacy in a society where the majority of adults were not literate.

The increase in enrollment mirrors the increase in schools. Between 1891 and 1921, schools per 100,000 people almost doubled from 44 to 70, with publicly managed and funded schools tripling from 9 to 30, while privately managed publicly funded schools almost doubled from 26 to 41 per 100,000 people (Chaudhary Reference Chaudhary2010a). Public funding was used to increase the number of schools and reduce fees. Although public primary education was not free, fees were not a significant barrier for skilled workers. For example, annual primary school fees in 1900 represented less than 0.5 percent of the annual wages of a skilled laborer (Chaudhary 2016).

On the expenditure side, public funding of education was decentralized to provinces in the 1870s, with further decentralization of primary education to districts and municipalities in the 1880s. The decentralization led to big differences in public spending across provinces, for example, between the coastal provinces of Bombay and Bengal. Bombay received more public funds and built a network of publicly funded and managed schools charging low fees. Unlike Bombay, Bengal received fewer public funds and subsidized privately managed aided schools, charging higher fees to build their network (Chaudhary Reference Chaudhary2010a). ^{Footnote 5} Other provincial systems fell in between those of Bombay and Bengal.

Yet, these differences in public spending across provinces did not translate into differences in outcomes, namely enrollment or literacy. The coastal provinces of Bengal, Bombay, and Madras had higher enrollment and literacy in each decade, with male literacy averaging 20 percent compared to 11 percent in the interior for Central Provinces and United Provinces. Apart from regional differences, castes ranked higher in the Hindu caste hierarchy were more literate and better educated. In comparison, literacy among lower castes, also known as depressed castes or Scheduled Castes in modern India, averaged 1.6 percent. Tribal groups had even lower literacy at under 1 percent. ^{Footnote 6}

Against this backdrop of low but varying literacy, few scholars have looked at the effects of demand shifters in explaining levels of schooling. Much of the scholarship argues that poor public funding led to low literacy, which is a reasonable conclusion given the national patterns (Chaudhary 2016). Yet, differences in public spending alone cannot explain the differences in outcomes across and within provinces. To that end, we study if and how railroads affected the demand for basic literacy by exploiting temporal and spatial variation within British India.

Railroads in Colonial India

Unlike schooling, the British were early promoters of railroads in India, building an extensive rail network. ^{Footnote 7} The first passenger line opened in 1853, connecting Bombay to Thane, a distance of 20 miles. Prompted by mercantile interests in Britain, the early lines connected the ports of Bombay, Calcutta, and Madras to the interior. Given the few good roads and navigable rivers, British firms hoped railroads would lower the costs of exporting raw cotton from India and of importing British manufactured goods to new Indian markets (Thorner Reference Thorner1951). Indeed, the British believed that goods traffic would be significant while passenger traffic would be insignificant. They proved to be wrong, with passenger traffic accounting for 35 percent of revenues.

Indian railroads were built by British firms with British financing, albeit subsidized by a guaranteed dividend backed by the Government of India (GOI). Such firms were the main players up until the 1870s, when the GOI began to build lines. This was followed by mixed public-private partnerships in the 1880s. Such partnerships were the norm until the 1920s (Sanyal Reference Sanyal1930). Route mileage expanded quickly in the early decades, especially from 1881 to 1901. Total route miles increased from 9,890 in 1881 to 17,308 in 1891, 25,363 in 1901, 32,839 in 1911, and then 37,266 in 1921 (Bogart and Chaudhary Reference Bogart and Chaudhary2016).

Figure 1 maps the spread of the network from 1881 to 1921. The important ports were connected to the interior before 1881. Many lines crossed the densely populated Indo-Gangetic plain, with fewer interior lines on the Deccan plateau. Early proposals, such as the Kennedy plan in 1852, called for lines parallel to the coast in order to economize on costs. Some were never built because subsequent officials opted for more expensive routes cutting through mountains (Davidson Reference Davidson1868). We return to the Kennedy plan below in order to construct an instrument for route placement.

Figure 1 RAIL NETWORK 1881–1921

Source: See text for details.

Although British firms built the railroads, the GOI dictated route placement. What guided their decisions? Military, commercial, and famine concerns were cited as the main drivers in official correspondence (Hurd Reference Hurd1983). Following the Sikh Wars in the 1840s and the Indian Mutiny in 1857, the British were cognizant of the need to transport troops and supplies across the country at low cost. Existing transport routes were of poor quality and slow, which made it necessary to station troops at multiple locations in the event of an uprising (Parliamentary Papers 1854). On the commercial side, British merchants lobbied for Indian railroads that would connect the ports to cotton-growing regions in the interior and from the eastern and western ports to Delhi in the north. Another consideration was famine. Following devastating crop failures and famines in the 1870s, the GOI built “protective lines” in famine-prone regions. Finally, a few small lines were built connecting plains to the hill stations. While not random, the railroad network across districts was not uniformly indicative of positive or negative selection in affecting the subsequent increase in literacy. Rather, a mix of factors affected where and when railroads were built. Coastal districts with important ports were connected early, as were those in the Ganga plains. Yet, a few cotton-growing interior districts were connected before 1881, as were districts closer to Afghanistan. Neither group would be considered positively selected for rail access. To address the endogeneity of railroads, we compare cohorts within districts in panel models and use an instrumental variables strategy among other cross-sectional models.

Conceptual Framework

To motivate the empirical exercise, we describe a simple framework linking railroad exposure to schooling in this sub-section. In colonial India, parents usually made the decision to send their children to school and to keep them in school. What affected this decision? Among rural households, cultivators and tradesmen sent their boys to the village primary school, as did some laborers. For cultivating families, the opportunity cost of child labor was higher because children helped their families, especially in the sowing and harvest seasons (Hartog Reference Hartog1929; Sharp Reference Sharp1914, 1919). This was less of a concern for tradesmen.

There was little legal compulsion to send your child to school. Weak compulsory schooling laws were introduced in a few towns in the late 1910s, but with rare enforcement (Nurullah and Naik 1951). Fees were common in primary schools, but as noted earlier, they were low. Scholarships were also available to defray the costs, though landless families were probably less aware of such opportunities compared to tradesmen and landed cultivators (Sharp Reference Sharp1914). If rural families wanted to send their children for more schooling beyond the local or neighboring village school, they would send them to a secondary school in a nearby town. Some urban schools had hostels, but it was more common for children attending an urban school to stay with extended family. Fewer than 10 percent of children in urban schools stayed in a hostel. ^{Footnote 8}

Learning to read, write, and count effectively was an important skill for families engaged in trade and commerce. With basic literacy, rural men could better perform their occupational tasks. They could also work as teachers and postmen in their villages, earning wages well above those of skilled labor in most provinces (Chaudhary 2016). Attending a secondary school led to further opportunities in the colonial bureaucracy, law, and professional employment. This was a draw for landed zamindar families looking to transition from rural to urban living.

Among urban households, richer families would send their children to the primary stages of middle or English high schools. Fees were higher at these schools, which nonetheless did not deter the growing demand for English education among rich and middle-class Indian families during this period (Basu Reference Basu1974). ^{Footnote 9} Learning English was a necessary skill to secure well-paying government and other service sector jobs. Some girls would also attend urban schools, albeit with fewer job prospects. Teaching was a common occupation for literate women of poor means, most of whom would have been educated in a town. Indeed, girls from rural families were less likely to attend primary school or move to a nearby town for more schooling.

How did railroads affect this household decision-making? By reducing price dispersion and increasing trade, railroads increased agricultural incomes in colonial India (Donaldson Reference Donaldson2018). This, in turn, would have generated income and substitution effects. Increasing agricultural incomes would lead families to consume more schooling if schooling is a normal good (i.e., income effect). For rural families, higher incomes would also relax credit constraints on sending children, mostly boys, to higher-quality secondary schools.

Yet, rising agricultural incomes would increase the opportunity cost of child labor leading some families, in particular cultivators and landless laborers, to send fewer children to school. More agricultural income could also lead to the expansion of rural primary schools if land taxes increased. But there is no evidence of such a supply side response because land taxes did not increase in sync with agricultural incomes (Kumar Reference Kumar1983). Apart from agriculture, railroads also had small and positive effects on city growth (Fenske, Kala, and Wei Reference Fenske, Kala and Wei2023). This would increase the returns to education due to agglomeration and more service and professional employment. While agricultural and non-agricultural channels likely affected male literacy, non-agricultural channels may have been more important for female literacy.

Unlike transportation costs that fall immediately with the opening of a railway, many of these conceptual links from railroads to literacy take time to develop. Indeed, spillovers from railroads are likely to grow as exposure to railroads increases. For income effects to operate in the agricultural sector, channels such as price discovery (Aker Reference Aker2010), the formation of new links between buyers and sellers (Jensen and Miller Reference Jensen and Miller2018), learning about alternative crops (Munshi Reference Munshi2004), and learning about the return to education in agriculture (Foster and Rosenzweig Reference Foster and Rosenzweig1995), all involve frictions that prevent them from being immediate. They require the development of new networks and relationships. Credit constraints, similarly, would only ease gradually because rural credit markets characterized by high interest rates and scarcity (Nath Reference Nath2022) may be slow to transform.

Similarly, opportunities in the non-agricultural sector would increase gradually with a railway connection. Such growth requires firms to establish, relocate, and grow, all of which takes time. The growth of the bureaucracy in any city connected to the railway would not be immediate. Urbanization would also not increase the number of workers and consumers immediately. Furthermore, if railroads contribute to agglomeration effects, they would also affect city growth rates and, hence, the growth of urban returns to education. Such conceptual links suggest that the duration of railroad exposure is perhaps better suited to capture the effect of railroads on schooling because it allows for differential effects between a place connected to a railway for one year compared to another that is connected for ten years. A simple indicator for the presence of railroads may not capture these links.

Against this background, more exposure to colonial railroads could lead to more literacy and schooling if (1) the income effect of rising agricultural incomes exceeded the substitution effect of increasing the opportunity cost of child labor, (2) if rising incomes relaxed household income constraints of sending children to school, (3) if railroads increased opportunities in the non-agriculture sector such as in industry and services, and (4) if rising urbanization increased the returns to education.

Measuring Literacy

The colonial census reports literacy by gender and age bins. Since 1901, the census also reports English literacy. Despite its richness, enumerating literacy over time is difficult because of changes in definition and measurement. In the 1881 and 1891 censuses, individuals were classified into three categories: literate, learning, and illiterate. Yet, enumerators were given no guidance on measuring literacy or accounting for learners apart from an age threshold (Gait Reference Gait1913). Age bins were also different across provinces.

Beginning with the 1901 census, the “learning” category was dropped and literacy was reported for standard age bins: those under age 10, aged 10 to 15, aged 15 to 20, and those over age 20. A uniform definition was adopted, namely “the ability to read and write.” Yet again, enumerators were not given official guidance on measuring literacy. Some provinces used a rigorous standard, while others enumerated individuals as literate if they could sign their name (Gait Reference Gait1913). It was only in 1911 that a uniform standard, the ability to read and write a short letter, was introduced. This makes literacy in the 1901 and later censuses difficult to compare. For example, many children under age 10 were counted as learners in the 1891 census, and some children under 10 were recorded as literate in the 1901 census, but not in subsequent censuses. To get around these issues, our panel analysis uses cohort literacy from the 1911 and 1921 censuses, when literacy was uniformly measured. Our cross-sectional analysis uses total literacy in each census year.

Online Appendix Table A1 shows literacy by cohort, gender, and language from 1901 to 1921. These are crude literacy rates equal to the number of literates in each group divided by the population of that group. Men were more literate than women, though this gender gap narrowed over time. English literacy was low in absolute terms but sizeable as a share of total literacy. Almost 15 percent of literate individuals in 1921 were, for example, also literate in English. Most children typically learned to read and write in a vernacular language before learning English (Sharp Reference Sharp1914). So, English literacy was, in particular, a measure of upper tail human capital. Online Appendix Table A2 shows total, male, and female literacy for each cross-section, while Online Appendix Figure 1 shows the distribution of total, male and female literacy. While the distribution of literacy was highly skewed in 1881, it became more dispersed by 1921.

Measuring School Enrollment

Unlike literacy, which measures the stock of human capital, enrollments capture the flow of human capital. As we expect railroads to affect the stock of literacy by increasing the flow of children into school, we complement our analysis of literacy with an analysis of school enrollment.

District enrollments are not reported in the colonial census. Rather, they are reported in many district gazetteers, which are less uniform. Nonetheless, we construct a series on primary and secondary enrollment between 1894 and 1911, the years with the most uniform data. ^{Footnote 12} This is an unbalanced panel, as most provinces report enrollment for a subset of years.

Primary school enrollment is recorded as the number of children enrolled in primary schools divided by the cohort under age 10. It averaged 6 percent in 1901 and 1911, compared to 3 percent for secondary enrollment, which is children in schools other than primary schools, divided by the cohort aged 10 to 15. ^{Footnote 13} A shortcoming of these data is that many secondary schools have attached primary classes, so some primary-aged children will then be included in secondary enrollment. For example, 47 percent of children in English secondary schools in 1912 were in primary classes, increasing to more than 60 percent in Assam and Eastern Bengal (Sharp Reference Sharp1914). Such primary classes were of higher quality than regular vernacular primary schools. As stated in Richey (Reference Richey1923), “The fact is that a very large percentage of the boys receiving elementary education in towns are not attending primary schools but the preparatory departments of secondary schools. It is only parents of the poorest class who send their boys to municipal primary schools” (p. 109). While this introduces some measurement error in enrollment, we are unaware of district-level enrollment data for all primary school children, regardless of school type.

Measuring Access to Railroads

To estimate the effect of railroads, we follow Fenske, Kala, and Wei (Reference Fenske, Kala and Wei2023) to code the opening dates of railway access in each district. Fenske, Kala, and Wei (Reference Fenske, Kala and Wei2023), following a procedure similar to Donaldson (Reference Donaldson2018), construct a polyline shapefile of the Indian railway network with an opening date for each segment. These dates are based on the 1934 edition of History of Indian Railroads Constructed and In Progress. For each listed railway line, they record the opening dates along with the beginning and end points of each line. We intersect this shapefile of railway lines with a map of modern sub-districts. Using a GIS mapping of colonial districts to these modern sub-districts, we compute the earliest year that each colonial district is connected to the railroad.

We use the date of opening to measure the duration of railroad exposure in a district, which, as we noted earlier, better captures the conceptual links from railroads to schooling. We refine this idea further in our panel analysis and assume the duration of railroad years affects literacy only up to the beginning of elementary school. If, for example, a railroad arrived in a district in 1893, it would not affect literacy for the cohort 20 and above in 1901 because they would be age 12 and above in 1893 and would have finished primary school. So, cohort 20 and above would have no exposure to railroads (coded as 0). In contrast, the arrival of railroads in 1893 would affect cohorts under age 10 in 1901 because many of them would not have presumably started elementary school as railroads arrived.

Since the age bins do not perfectly correspond to elementary school years, we use the youngest age in the bin to measure cumulative exposure up to elementary school. Our measure is the number of years a railroad has been operating in a district minus the number of years since the youngest member of a cohort would have regularly begun elementary school at age 6. ^{Footnote 14} Denote this number of years since schooling began as y(c). For cohorts aged 20 and above, y(c) is 14. For cohorts aged 15–20, it is 9. For cohorts aged 10–15, it is 4. For cohorts aged below 10, it is 0. For cohort c, y(c) years since schooling began, in district d, with a railroad that opened in year r, measured in census year t, we define our treatment measure RailroadYears _cdt as:

(1) $${RailroadYear{s_{cdt}} = \left\{ {\begin{array}{} {max\left\{ {t - r - y\left( c \right),\,0} \right\}\,if \,r \leq t} & \cr {\,\,\,\,\,\,\,\,\,\,\,\,\,\,\,\,\,\,\,\,\,\,\,\,\,\,\,\,\,\,\,\,\,\,\,\,\,\,\,\,\,\,\,\,\,\,\,0 \,\,if \,r > t} \end{array}}\right\}}$$

It may well be that railroads continue to affect schooling up until a child ends elementary school. We therefore construct a second measure, which is the number of years a railroad has operated in a district minus the number of years since the youngest member of a cohort would have regularly finished elementary school. ^{Footnote 15} This measure assumes railroads affect literacy up to age 12 for the index age in a cohort. In Equation (1), this is equivalent to replacing y(c) with 8 for the cohort aged 20 and above, 3 for cohorts aged 15 to 20, and zero for the cohorts aged 10–15 and 0–10.

As constructed, the two measures bound the duration of exposure to railroads and assume that railroads affect literacy only up to the beginning or end of elementary school. Our first measure assumes parents decide whether to enroll their children in school based on cumulative exposure to the railroad up to the beginning of elementary school (age 6). Our second measure assumes parents decide whether to keep their children in elementary school based on cumulative exposure to the railroad up to the end of elementary school. ^{Footnote 16}

For the cross-sectional analysis, we count the number of years a district has been connected to a railroad in each census. As seen in Online Appendix Table A2, 52 percent of districts were connected to a railroad by 1881, increasing to 96 percent by 1921, with much of the increase happening before 1901. The railroad years measure thus better illustrates the variation across districts. For example, the number of railroad years averaged 7.6 years across districts in 1881, increasing to 22 years in 1901 and 40 years by 1921.Footnote ¹⁷

Geographic and Socioeconomic Controls

We construct a rich set of geographic variables to control for the geographical selection of railroad exposure. We collect data on the latitude and longitude coordinates of the centroid of the district, which we compute ourselves. We control for ruggedness from Nunn and Puga (Reference Nunn and Puga2012). We control for altitude, precipitation, temperature, slope, and suitability for growing specific crops such as cotton, dryland rice, wetland rice, and wheat, averaged over raster cells within a district. These are taken from the Food and Agriculture Organization of the United Nations’ Global Agro-Ecological Zones (FAO-GAEZ) data portal. Since proximity to the coast and rivers likely influenced railroad access, we include indicators for rivers and coastal districts as captured in Natural Earth Data’s shapefile maps of rivers and coasts. We also control for medieval ports recorded by Jha (Reference Jha2013). We control for the seasonality of rainfall. In particular, using data on historic rainfall from Matsuura and Willmott (Reference Matsuura and Willmott2018), we compute the Feng, Porporato, and Rodriguez-Iturbe (Reference Feng, Porporato and Rodriguez-Iturbe2013) entropy-based measure of seasonality. Finally, we control for the Kiszewski et al. (Reference Kiszewski, Andrew Mellinger, Malaney, Ehrlich Sachs and Sachs2004) index of the stability of malaria transmission.

Apart from geography, we control for the scale of urbanization before the advent of railroads using the population of cities enumerated in Chandler and Fox (Reference Chandler and Fox1974), circa 1850. These cities range in population from 11,000 to 580,000 across 52 districts. This effectively controls for more urban districts that were likely to be connected with railroads before less urban districts. We also control for the religious and caste composition of a district, including the share of Brahmans, traditional Hindu elites, Muslims, Christians, and tribal groups. Such shares are intended to capture historical differences in education among groups that may be correlated with railroad access. These data are taken from the colonial censuses.

Synthetic Panel

We estimate the following model using the log of the literacy rate by year, district, and cohort as the outcome:

(2) $${\text{ln}}\left( {LiteracyRat{e_{cdt}}} \right){\text{ }} = \beta RailroadYear{s_{cdt}} + {\theta _d} + {\delta _p} \times {\eta _t} + {\delta _p} \times {\gamma _c} + {\unicode{x03F5}_{cdt}}$$

In this model, LiteracyRate _cdt is literacy for cohort c in district d and census year t. We use log literacy because it is a highly skewed variable, as shown in Online Appendix Figure 1. We estimate the model for t ∈ {1911,1921} and cohort c ∈ {0-10,10-15,15-20, 20+}. RailroadYears _cdt measures the cumulative years of railroad exposure for cohort c in district d in year t.

We control for several fixed effects. First, district fixed effects, θ _d , capture unobservable time-invariant district features that lead some districts to get railroads before others and that may correlate with literacy. Second, we include interactions of province × year and province × cohort fixed effects captured by δ _p × η _t and δ _p × γ _c to control for provincial changes in census enumeration methods by year and cohort. Such flexible controls address most measurement concerns related to literacy and account flexibly for omitted variables at the province and cohort level that may change over time. We cluster standard errors by district to account for serial correlation over time.

In this setup, we identify the effects of railroads using variation in cumulative exposure to railroad years across cohorts within districts over time. The key identifying assumption is that such exposure in railroad years is uncorrelated with the error term ε _cdt . We believe this is a reasonable assumption given the flexible fixed effects included in the model. As a robustness check, we run the same analysis using the 1901 census and controlling for district fixed effects and province × cohort fixed effects. Since we use only the 1901 census for this exercise, changes in the standards used to measure literacy in different censuses are not an issue.

Cross-Section

We complement the panel methods with two cross-sectional models as follows.

ORDINARY LEAST SQUARES

We exploit the complete data from 1881 to 1921 using repeated cross-sections. For each census year, we estimate a separate OLS regression of the following form in t ∈{1881,1891,1901,1911,1921}:

(3) $${\text{ln}}\left( {LiteracyRat{e_{dt}}} \right){\text{ }} = \beta RailroadYear{s_{dt}} + {\gamma ^\prime }{x_{dt}} + {\delta _p} + {\unicode{x03F5}_{dt}}$$

In this equation, ln(LiteracyRate _dt ) is the log literacy rate in district d in year t. Unlike cohort literacy in the synthetic panel, this measure picks up adult literacy because everyone who is literate is included in total literacy regardless of age. RailroadYears _dt is the number of years district d in year t has had a railroad. This is 0 if the district is unconnected in t. The vector x _dt includes the GIS controls, pre-rail urbanization, and social controls described in the previous section. We also include province fixed effects captured by δ _p . Finally, ε _dt is the error term.

Such a regression may generate biased estimates of the causal effect of railroads. For example, if more developed districts were the first to receive railroads, our estimate of railroad years would be biased upward because it would conflate the effects of railroads with those of prior development. On the other hand, if famine-prone areas received access early on, then our estimates would likely have a negative bias. To address such endogeneity concerns, we employ instrumental variables.

INSTRUMENTAL VARIABLES

We construct two instruments for RailroadYears _dt that exploit different sources of variation. First, we build a tree spanning the 67 British military cantonments that existed as of 1864 before major expansion of railroads. If military concerns drove the placement of railroads, we expect cantonments where army troops were stationed to get early access. Using Prim’s algorithm, we construct the shortest tree that spans these 67 military cantonments. Figure 2 shows a map of this tree superimposed on the 1881 railway network. After constructing the tree, we compute the distance of each district from the spanning tree. We then use the log of (one plus) distance to this tree as an instrument for RailroadYears _dt .

Figure 2 MAP OF MILITARY CANTONMENT SPANNING TREE

Notes: Spanning tree drawn in black. 1881 railway network drawn in grey.

Source: See text for details.

According to Kulkarni (Reference Kulkarni1979), two factors determined the location of cantonments. First, these places had to be “suitable for outdoor training round the year” (p. 214). This favored areas at moderate elevations. Second, cantonments could not be located near ravines where an enemy could hide. Both factors favor a particular aspect of climate and geography that is plausibly exogenous to other factors affecting human capital, conditional on the geographic and socio-economic controls. In addition, the northern Indian plains were more vulnerable to attack as compared to hilly areas, which in turn motivated the establishment of cantonments (Kulkarni Reference Kulkarni1979). Our analysis includes province fixed effects that capture this dimension of location.

Our second instrument uses Major J. P. Kennedy’s 1852 proposal for building railroads. Major Kennedy was the consulting engineer for the GOI and pushed for building low-cost railroads that, in his view, would confer innumerable benefits “to the growth of everything connected with the extension of British interests in India as well as with the industry, the wealth, and the comfort of its vast population” (Parliamentary Papers 1854, p. 3). Yet, Major Kennedy was aware of the costs of building railroads. So, he emphasized lower-cost routes connecting the ports with the interior. In particular, his plan called for a network in “strict harmony with the natural advantages” of the country. Unlike routes that would cut through the Eastern and Western coastal ranges of India, his plan called for routes that favored softer gradients, following the coast and natural topography.

Donaldson (Reference Donaldson2018) used portions of the Kennedy plan that were not implemented to construct placebo lines. In many cases, however, Kennedy’s routes were adopted, as seen by comparing the Kennedy plan in Figure 3 to the actual network in Figure 1. In other cases, however, more expensive routes were selected. Hence, we are assuming here that, conditional on controls, the 1852 Kennedy plan is uncorrelated with factors that would affect literacy other than access to railroads. To construct the instrument, we convert the map of Kennedy’s proposal into a polyline shapefile. We then calculate the shortest distance of each district from this route. We use the log of (one plus) distance to the lines in the Kennedy plan as an instrument for RailroadYears _dt . We report results using the two instruments in the same regression and the results of each instrument used individually.

Figure 3 MAP OF 1852 KENNEDY PLAN

Notes: 1852 Kennedy Plan drawn in black. 1881 railway network drawn in grey.

Source: See text for details.

Enrollment

Table 4 shows the results of enrollment for the panel and cross-sectional methods. As seen in the top panel, where we include district and year fixed effects, increasing exposure to railroads has a positive and significant effect only on secondary enrollment. Indeed, the coefficient on primary enrollment is negative, albeit insignificant. It would be possible for railroads to increase secondary enrollment without similarly increasing primary enrollment if, for example, they had no effect on the extensive margin, but raised the continuation rate into secondary education. Given that many of the secondary schools in the data combined secondary schooling with elite primary education, these results may also reflect greater enrollment in elite primary schools. What these results rule out is the interpretation that railroads led to an increase in children attending basic vernacular primary school.

Table 4 ENROLLMENT

Notes: Robust standard errors in parentheses. *** p<0.01, ** p<0.05, * p<0.1. These cross-sectional models include province fixed effects; social controls namely the share of Brahmans, Christians, Muslims, and Tribes; the GIS controls namely area, latitude, longitude, altitude, precipitation, slope, temperature, ruggedness, malaria transmission, seasonality, indicators for coastal districts, rivers, and medieval ports, and suitability for specific crops such as cotton, dryland rice, wetland rice, wheat, and tea; and the city population recorded in Chandler and Fox (Reference Chandler and Fox1974) circa 1850.

Source: See text for details.

In terms of magnitude, the effects of railroad exposure increase secondary enrollment by 0.55 standard deviations in Specification (6) of the panel analysis and by 0.42 standard deviations of secondary enrollment for the 1911 IV specification. In comparison to literacy, these standardized β coefficients are larger for both the panel and cross-sectional models. This is unsurprising. We would expect bigger effects of railroads on the flow of children into school compared to the stock of literates because of high dropout rates, with many children leaving primary school before completing 3 to 4 years of schooling, which educationists in this period argued was necessary to become literate (Parulelar Reference Parulelar1939).

Agricultural Income and Land Taxes

Railroads had a large effect on price convergence, trade, and agricultural income in India (Donaldson, Reference Donaldson2018). Are rising agricultural incomes then a mediator between railroads and higher literacy? In Tables 5 and 6, we conduct a mediation analysis suitable for an OLS framework (Imai, Keele, and Tingley Reference Imai, Keele and Tingley2010; Imai et al. Reference Imai, Keele, Tingley and Yamamoto2011). Similar to enrollment and literacy, the mediators are logged in these regressions. Table 5 shows the mediation results for total literacy, and Table 6 shows these results for secondary enrollment. ^{Footnote 25} As seen in Specifications (1) and (2) in the top panel, the coefficient on income is small, negative, and insignificant.

Table 5 MEDIATORS: TOTAL LITERACY, OLS

Notes: Robust standard errors in parentheses. *** p<0.01, ** p<0.05, * p<0.1. The outcome is log literacy in the respective year. These cross-sectional models include province fixed effects; social controls namely the share of Brahmans, Christians, Muslims, and Tribes; the GIS controls namely area, latitude, longitude, altitude, precipitation, slope, temperature, ruggedness, malaria transmission, seasonality, indicators for coastal districts, rivers, and medieval ports, and suitability for specific crops such as cotton, dryland rice, wetland rice, wheat, and tea; and the city population recorded in Chandler and Fox (Reference Chandler and Fox1974) circa 1850.

Source: See text for details.

Table 6 MEDIATORS: SECONDARY ENROLLMENT, OLS

Notes: Robust standard errors in parentheses. *** p<0.01, ** p<0.05, * p<0.1. The outcome is log secondary enrollment in the respective year. These cross-sectional models include province fixed effects; social controls namely the share of Brahmans, Christians, Muslims, and Tribes; the GIS controls namely area, latitude, longitude, altitude, precipitation, slope, temperature, ruggedness, malaria transmission, seasonality, indicators for coastal districts, rivers, and medieval ports, and suitability for specific crops such as cotton, dryland rice, wetland rice, wheat, and tea; and the city population recorded in Chandler and Fox (Reference Chandler and Fox1974) circa 1850.

Source: See text for details.

In Specifications (3) and (4), we also rule out a link from agricultural income to education via public funding. Surcharges on existing land taxes were a key funding source for district boards that managed rural primary education. While there could be a positive link in theory from railroads to agricultural income to land taxes, we find that land taxes per capita are not a significant mediator for literacy or secondary enrollment. These results are unsurprising because land taxes were revised infrequently in most parts of the country in the late nineteenth and early twentieth centuries (Kumar Reference Kumar1983). While railroad years are correlated with agricultural income, they are uncorrelated with land taxes.Footnote ²⁶ There is no evidence that rising agricultural incomes mediate our railroad results.

Apart from agricultural income, railroads may have increased non-agricultural income and urbanization, which in turn would have increased the returns to education thus linking railroads to education. We indirectly test whether increasing returns to skill, urbanization, and rising non-agricultural incomes play a mediating role by looking at income tax revenues, urbanization share, and the share of workers in industry and services. Income taxes were assessed on non-agricultural income using a schedule that varied by income source. Since income from agriculture was not taxed, this measure captures income from industrial and professional employment. The share of non-agricultural workers and income taxes are both proxies, then, for returns to education.

Using data on income taxes from the District Gazetteers for 1901 and 1911, Specifications (5) and (6) in the top panel of Tables 5 and 6 show that income taxes have a positive and significant coefficient for both literacy and secondary enrollment. They mediate 30 percent to 46 percent of the effects of railroads on literacy and 25 percent to 43 percent of the effects on secondary enrollment. Rising non-agricultural incomes may have led to income effects and eased liquidity constraints, leading more families to “buy” schooling for their children.

In the bottom panel, we look at urbanization and the share of workers in industry and services. To measure urbanization, we compute the share of the population in a district living in towns with at least 5,000 persons using census data on city populations from Fenske, Kala, and Wei (Reference Fenske, Kala and Wei2023). We use labor force data from Fenske, Gupta, and Yuan (Reference Fenske, Gupta and Yuan2022) to construct these measures using the decennial census. Similar to income taxes, urbanization mediates between 38 percent and 48 percent of the effects on total literacy and a smaller share of secondary enrollment at 9 percent to 16 percent. Service sector employment also appears to partially mediate the results, but less so than urbanization and income taxes. It mediates anywhere from 6 percent to 16 percent of the effect of railroads on literacy and secondary enrollment. Lawyers and public administrators, among other professionals, were part of the service sector. Such workers were more educated than the rest of the population and were paid higher wages than other skilled occupations. These measures do, however, conflate income effects with rising returns to education. We have no way of disentangling these channels and hence interpret these results as evidence of their joint importance.