HISTORICAL BACKGROUND

EMPIRICAL STRATEGY AND DATA

Empirical Strategy

In this article, we use price data covering M South Asian markets. Each observation is a market-pair, indexed ij. For product p, traded between markets i and j, we estimate:

(1) \[{\rho _{ij}}^p = {\beta ^p}{\text{LinguisticDistanc}}{{\text{e}}_{{\text{ij}}}} + {x_{ij}}^p\prime {\gamma ^p} + {\delta _i}^p + {\eta _j}^p + {\varepsilon _{ij}}^p.\]

In Equation (1), \[{\rho _{ij}}^p\] is the correlation coefficient for the price of p between markets i and j. LinguisticDistance _ij, described later, captures linguistic distance between the two markets. \[{x_{ij}}^p\] is a vector of controls. We use this to account for a wide set of dissimilarities between i and j that may correlate with linguistic distance and with the degree of price integration. In our baseline estimations, \[{x_{ij}}^p\] includes a constant, as well as controls for proximity (log distance in kilometers between the markets, whether both markets are coastal, and whether both markets are connected by the same river), geographic similarity (the correlations in precipitation and temperature between the markets, and their absolute differences in: altitude, latitude, longitude, rainfall, temperature, land quality, ruggedness, malaria, humidity, precipitation, and terrain slope), agricultural similarity (absolute differences in suitabilities for growing banana, chickpea, cocoa, cotton, groundnut, dryland rice, oil palm, onion, soybean, sugar, tea, wetland rice, white potato, wheat, or tomato), other measures of similarity (whether the markets are in the same province, and their religious distance), and characteristics of the data (first year, last year, and the number of years in which the price is available for both markets).

One limitation of our empirical strategy is the possibility that our control variables are measured with greater error than our principal right-hand-side variable of interest, that is, LinguisticDistance _ij. This could lead to our estimates of β ^p being overstated. We note, then, that linguistic distance may be interpreted more broadly, for example, as a measure of greater ancestral distance. \[{\delta _i}^p\] and \[{\eta _j}^p\] are fixed effects for market i and market j. The sample is all market pairs ij such that i ≠ j, i > j, and there are sufficient observations to compute \[{\rho _{ij}}^p\]. That is, we have at most \[\frac{{{M^2} - M}}{2}\] observations in any one regression. We cluster standard errors by both market i and market j in the baseline (Cameron, Gelbach, and Miller Reference Cameron, Gelbach and Miller2011). Because of the possible spatial dependence induced by forming every pairwise combination of markets, we show results in the Online Appendix in which we cluster at alternative levels and compute Conley (Reference Conley1999) standard errors.

PRICES

Our data on prices are taken from three editions (1921, 1907, and 1885) of Wages and Prices in India. These are initially in reported in sers per rupee: we invert this measure to obtain nominal prices. For 206 markets in modern-day Pakistan, India, Bangladesh, and Burma, these data provide prices for more than a dozen crops: Arhar Dal, Bajra, Barley, Gram, Jawar, Kangni, Maize, Marua, Rice, Salt, Wheat, Bulrush Millet and Similar, Great Millet and Similar, and Lesser Millets. The data covers both British India and the Princely States. These do not represent all markets in India—almost every populated place would have a market of some sort. Rather, these are markets in which the colonial government collected price data. More populous districts and districts in British India are more likely to appear in the data, and, in provinces such as Coorg that have few districts, at least one district is likely to be present.

In most of our results, we focus on the three most commonly reported prices: rice, wheat, and salt. The data do not allow us to consider differences between different varieties of wheat or salt. However, we also show that estimates of Equation (1) with several other crops produce similar results. The price data cover the period 1861 through 1921, with many markets entering our data for the first time in 1869. While the data-collection methods differed across markets in early years, from 1872 onwards uniform fortnightly returns of retail prices were used. ^{Footnote 4} So long as there are at least three years in which a price is reported in both markets i and j, we can compute a correlation coefficient for that product for the ij pair. This quantity, \[{\rho _{ij}}^p\], is our principal dependent variable.

In Figure 1, we provide intuition for our results by mapping the correlation between the price of rice in a single market, the largely Punjabi-speaking city of Ludhiana, with the price of rice in all other markets in our data. It is clear from the figure that rice prices track those in Ludhiana more closely in regions that speak more closely-related languages such as Hindi and Gujarati and less closely in regions that speak more distantly-related languages such as Burmese and Telugu. These regions are, however, also closer in physical proximity to Ludhiana, and many of the markets that most closely track prices in Ludhiana lie on the Indo-Gangetic Plain. Thus, our analysis relies on estimation of Equation (1) to demonstrate that the correlation between linguistic distance and price integration cannot be explained away by other observable differences in proximity or geography.

Figure 1

LUDHIANA: RICE PRICE CORRELATIONS

Source: Wages and Prices in India.

LINGUISTIC DISTANCE

To compute linguistic distances among the markets in our data, we use two additional data sources. These are the 1901 Census of India and version 19 of the Ethnologue Global Dataset. For each district that existed in 1901, the census data report the number of speakers of each language. For example, the three most commonly spoken languages reported for Ludhiana District are “Punjabi” (665,476), “Hindostani” (2,970), and “Kashmiri” (1,224). We assign each market to the language composition of the district that contained it in 1901. For consistency with the Ethnologue data on distances, we aggregate these to the level of ISO language codes. For Ludhiana, the three most commonly spoken languages become pan, hin, and kas. The data do not, unfortunately, mention second languages.

To compute the distances among these languages, we turn to Ethnologue. Every language in this source is categorized using a language tree with a maximum number of 15 branches. These classifications are based on several sources, the most important of which is Frawley (Reference Frawley2003). Such “cladistic” measures have become widely used in economics (Desmet, Ortuño-Ortín, and Wacziarg Reference Desmet, Ignacio and Romain2012; Gomes Reference Gomes2014). ^{Footnote 5}

Following Esteban, Mayoral, and Ray (Reference Esteban, Laura and Debraj2012), we take the distance d _mn between any two languages m and n as:

(2) \[{d_{mn}} = 1 - {\left( {\frac{{{\text{SharedBranches}}}}{{15}}} \right)^\delta }.\]

Similarly following Esteban, Mayoral, and Ray (Reference Esteban, Laura and Debraj2012), we choose δ = 0.05 as a baseline and use δ = 0.5 for robustness. To aggregate these to distances among markets, given population shares of languages m and n in each district i and j of s _mi and s _nj, we follow Spolaore and Wacziarg (Reference Spolaore and Romain2009) and compute linguistic distance among districts as:

(3) \[L{D_{ij}} = \sum\nolimits_m {\sum\nolimits_n {({S_{mi}} \times {S_{nj}} \times {d_{mn}}).} } \]

In Figure 2, we map the linguistic distances among every district in our data and Ludhiana. While it is evident that the markets at which languages more closely related to Punjabi are spoken are geographically close to Ludhiana, it is also clear that this correlation of linguistic and physical distance is not perfect. Distances change relatively rapidly over space when the linguistic composition of the population similarly changes rapidly. Further, regions that are relatively similar in physical distance can be quite dissimilar in their linguistic distance. Punjabi and Bengali, for example, both share the branches Indo-European, Indo-Iranian, and Indo-Aryan. Punjabi and Tamil, by contrast, share no branches, as Tamil is a Dravidian language. And yet the distance between the Punjab and Bangladesh is not markedly different than the distance between the Punjab and Tamil Nadu. The log distance in kilometers between Ludhiana and Dacca is 7.40, whereas it is 7.76 between Ludhiana and Madurai.

Figure 2

LUDHIANA: LINGUISTIC DISTANCES

Source: Census of India 1901.

ADDITIONAL CONTROLS

Some of our control variables are computed directly. Distance in kilometers is computed using the latitude and longitude of the market. “Both coastal” and “both connected by the same river” indicators are computed in ArcMap using a shapefile of district boundaries. “Minimum year,” “maximum year,” and “number of common observations” are computed directly from the price data.

The “same province” indicator is based on the provinces that contained each market in 1901. The “religious distance” variable is computed using the same equation as Equation (3), taking the religious composition of each district as reported in Table 8 of the 1921 Census (Literacy By Religion). We assume that the distance d _qr between any religion q and r is 1 if q ≠ r and 0 if q = r. ^{Footnote 6}

Data on land quality are taken from Ramankutty et al. (Reference Ramankutty, Foley, John and Kevin2002) and have been used in several economic studies, such as Michalopoulos (Reference Michalopoulos2012) and Ashraf and Galor (Reference Ashraf and Oded2011). ^{Footnote 7} It is an index based on soil and climate characteristics and is not particular to any one type of agriculture. “Ruggedness” is the measure of terrain ruggedness initially introduced by Nunn and Puga (Reference Nunn and Diego2012). ^{Footnote 8} Our measure of “malaria prevalence” was originally created by Kiszewski et al. (Reference Kiszewski, Andrew, Andrew, Pia, Sonia and Jeffrey2004). ^{Footnote 9} Altitude data are taken from the Consultative Group for International Agricultural Research’s Shuttle Radar Topography Mission 30 dataset. ^{Footnote 10} Means of precipitation, temperature, and suitabilities for specific crops are taken from the Food and Agriculture Organization’s Global Agro-Ecological Zones data portal. ^{Footnote 11} Similar suitability measures have been used by Alesina, Giuliano, and Nunn (Reference Alesina, Paola and Nathan2013) and Alsan (Reference Alsan2015). Correlations in rainfall are computed using the Matsuura and Willmott (Reference Matsuura and Cort2007) gridded series. ^{Footnote 12} We join each market to the nearest point in these data and compute correlations in annual rainfall over the period 1900–2000. Humidity data are taken from the Climatic Research Unit at the University of East Anglia. ^{Footnote 13}

Like many studies that control for geographic confounders with historical outcome variables, we are compelled to use present-day raster data (e.g., Alsan (Reference Alsan2015) and Nunn and Puga (Reference Nunn and Diego2012)). We expect that this will add measurement error to our right-hand-side variables, but that it is unlikely this measurement error will induce spurious correlation between linguistic distance and market integration. For the variables that require geographic data (i.e., the coastal and river indicators, as well as those using raster data), we begin with a district map for modern India. ^{Footnote 14} We compute the coastal and river indicators at this level, and compute other geographic variables by averaging over raster points within a district. If a market in our data shares the name of a modern-day district (or an updated name, as in the case of Benares and Varanasi), we have a unique match between the market and the modern district polygon. Otherwise, we match all districts that split from the erstwhile district that previously shared the name of the market to that market.

Summary Statistics

Summary statistics are presented in Table 1. Some general patterns are apparent from this table. First, relative to a maximum number of observations of \[\frac{{{{206}^2} - 206}}{2} = 21,115\], we typically have fewer pairwise correlation coefficients. This is because not all products are traded in all markets. Second, while the degree of price integration is relatively high (>0.8 for both wheat and rice), there is variation in price integration both across space and across markets. Some market pairs exhibit negative price correlations. Market integration is more limited for salt than for rice and wheat; the average price correlation for salt (<0.35) is lower, and more than a quarter of these correlations are negative. One possible explanation of this lower correlation is the limited number of inland production sites for salt; this limits arbitrage opportunities in response to shocks, causing lower average salt price correlations across markets. Linguistic distances range from close to 0 (i.e., market pairs in which both markets are dominated by the same language) to 1 (i.e., market pairs in which the dominant languages spoken are unrelated).

Table 1

SUMMARY STATISTICS

Source: See the text.

MECHANISMS

In this section, we outline the mechanisms suggested in both the economic and historical literatures that provide plausible links between linguistic distance and market integration. We then assess these empirically to the extent our data allow.

GENETIC DISTANCE

To evaluate whether linguistic distance operates as a proxy for a broader set of barriers to the transmission of information, technology, and culture, we compute a measure of the genetic distance among the markets in our data. We show that, while linguistic distance and genetic distance are correlated, neither one is a “sufficient statistic” that fully accounts for the coefficient of the other.

We obtain data on genetic distance from Pemberton, DeGiorgio, and Rosenberg (Reference Pemberton, Michael and Rosenberg2013). Similar to the data used by Spolaore and Wacziarg (Reference Spolaore and Romain2009), these data contain pairwise Weir and Cockerham (Reference Weir and Cockerham1984) F _ST coefficients based on differences in allele frequencies from microsatellites. While the raw data report coefficients based on 5,795 individuals from 267 human populations, we restrict ourselves to the data on ethnic groups indigenous to South Asia. These are the Balochi, Brahui, Burusho, Hazara, Kalash, Makrani, Pathan, Sindhi, Assamese, Bengali, Gujarati, Hindi, Kannada, Kashmiri, Konkani, Malayalam, Marathi, Marwari, Miso, Oriya, Parsi, Punjabi, Tamil, and Telugu. While these groups cover the majority of the population in our sample, there are some major missing groups, of which Urdu is the largest.

Following Spolaore and Wacziarg (Reference Spolaore and Romain2009), given population shares of groups m and n in districts i and j of s _mi and s _nj with genetic distance \[{F_{ST}}^{mn}\], we compute genetic distance among districts as:

(5) \[G{D_{ij}} = \sum\nolimits_m {\sum\nolimits_n {({S_{mi}} \times {S_{nj}} \times {F^{mn}}_{ST}).} } \]

Note that we re-scale s _1i and s _2j as fractions of the population matched to the genetic data, rather than as fractions of the full district population. We present a map of genetic distances from Ludhiana in Online Appendix Figure A1. This has many similarities to Figure 2. Other regions of South Asia that are proximate to the Punjab are more genetically similar, although it is clear that South Indian groups in Dravidian-speaking regions are more genetically dissimilar, conditional on physical distance. The apparent proximity with Burma is overstated due to the lack of coverage of major Burmese populations in the genetic data.

Our aim is to assess whether linguistic distance proxies for broader (and possibly deeper) barriers to the diffusion of information, culture, and technology. We re-estimate Equation (1), first with genetic distance as an outcome, and second with genetic distance as an additional control. We report the results in Table 3. Linguistic and genetic distance are correlated, even conditional on our baseline fixed effects and controls.^{Footnote 17} Genetic distance itself predicts less market integration and diminishes the coefficient on linguistic distance, but does not fully eliminate it in any specifications where linguistic distance was significant in Table 2. With fixed effects and controls, the change in coefficient on linguistic distance is slight when compared with Table 2. These results imply that, while linguistic distance may indeed proxy for other differences across populations, its relationship with market integration cannot be fully accounted for by the additional transaction costs imposed by barriers to the diffusion of beliefs, traditions, and practices stemming from ancestral distance.

Table 3

GENETIC DISTANCE

* = Significant at the 10 percent level.

**= Significant at the 5 percent level.

*** = Significant at the 1 percent level.

Notes: Standard errors clustered by market i and market j in parentheses. All regressions are OLS and include a constant. Controls are minimum year, maximum year, number of observations, ln(distance) in km, both coastal, connected to river, rainfall correlation, temperature correlation, and absolute differences in: altitude, latitude, longitude, rainfall, temperature, land quality, ruggedness, malaria, humidity, precipitation, slope, religion, and suitabilities for growing banana, chickpea, cocoa, cotton, groundnut, dryland rice, oil palm, onion, soybean, sugar, tea, wetland rice, white potato, wheat, and tomato. Fixed effects are for markets i and j.

Source: See the text.

COARSE AND FINE DISTINCTIONS

We show that it is the highest-level distinctions in our data, such as those between Indo-European and Dravidian languages, that drive our results. This is, however, a crude proxy, and we cannot rule out the possibility that languages here proxy for past patterns of migration and state formation that themselves shaped markets and trade routes.

Recall that, in our baseline analyses, we computed the distance between any two languages m and n as:

\[{d_{mn}} = 1 - {\left( {\frac{{{\text{SharedBranches}}}}{{15}}} \right)^\delta }.\]

While this follows the convention in the literature, it does not allow us to distinguish whether coarser distinctions (e.g., those between Indo-European and Dravidian languages) or lesser divisions (e.g, those between Bengali and Punjabi) drive our results. We replace d _mn with a dummy for having ≤N shared branches, for N = {1, …, 15}. We re-estimate Equation (1), and present our results in Figures 6, 7, and 8. These correspond to Column (4) with fixed effects and controls. In all three figures, it is clear that coarser distinctions matter more than finer ones. Indeed, we show in Online Appendix Table A5 that limiting our sample only to district pairs in which the dominant language in both districts is Indo-European leads to coefficient estimates on linguistic distance that, while still negative, are generally insignificant and less robust across specifications. That is, our results are driven by coarser language distinctions, particularly those that separate major language families.

Figure 6

RESULTS BY LEVEL: WHEAT

Source: Authors’ estimates.

Figure 7

RESULTS BY LEVEL: SALT

Source: Authors’ estimates.

Figure 8

RESULTS BY LEVEL: RICE

Source: Authors’ estimates.

Consider a language such as Gujarati (Indo-European, Indo-Iranian, Indo-Aryan, Intermediate Divisions, Gujarati, Gujarati). It has no branches in common with a Dravidian language such as Tamil. It shares one branch with languages such as Yiddish that are Indo-European but not Indo-Iranian. It shares two branches with languages such as Balochi that are Indo-Iranian but not Indo-Aryan. It shares three branches with an Indo-Aryan language such as Hindi that is classified under “Western Hindi” rather than “Intermediate Divisions.” It shares four branches with a language such as Nepali that is within these “Intermediate Divisions,” but is not within the Gujarati sub-class. It shares five branches with other Gujarati languages (such as Jandavra). In all three figures, language divisions with two common branches or fewer yield visibly greater differences than finer distinctions. These results suggest that our main results derive from divisions on the scale of Gujarati-Tamil, Gujarati-Yiddish, and Gujarati-Balochi, rather than from finer distinctions as those among Gujarati and Hindi, Nepali, or Jandavra. These coarser distinctions are those that have been shown before to correlate with conflict, redistribution, and public goods provision—suggesting they are correlated with deeper differences in preferences—as opposed to finer distinctions that inhibit coordination and integration (Desmet, Ortuño-Ortín, and Wacziarg Reference Desmet, Ignacio and Romain2012). This is suggestive evidence that our results are driven not simply by ease of communication, but also by more fundamental differences in preferences.

MISSING MARKETS

To test whether missing markets, due, for example, to differences in tastes drive the correlation between linguistic distance and market integration, we evaluate whether linguistic distance predicts whether two given markets report a certain good’s price in the same year, and whether markets that are more linguistically distant from their neighbors experience more volatile prices. When we look at the situation for major crops, we find little evidence of missing markets increasing with linguistic distance. Only limited evidence suggests that prices are more variable at markets that are more linguistically different from those around them.

We take two approaches. First, we test whether linguistic distance predicts how frequently prices are available for two markets in the same year. Taking \[{N_{ij}}^p\] as the number of common price observations at markets i and j for product p, we estimate Equation (1), except that we now take \[{N_{ij}}^p\] as the dependent variable, and no longer control for minimum year, maximum year, or the number of common observations. Results are presented in Table 4. There is only weak evidence of missing markets correlating with linguistic distance; while we find a negative correlation between linguistic distance and \[{N_{ij}}^p\] for wheat, no such correlation is available for salt or rice. We find similar failures of linguistic distance to predict\[{N_{ij}}^p\] when using lesser crops from the data such as barley and maize, although we do not report these here. One explanation of the different result for wheat is the greater variability of the outcome variable: the standard deviation of the number of common years for wheat is 22.6, versus 8.8 for salt and 9.7 for rice. That is, as wheat is reported less often in many markets, there is more variation to be explained.

Table 4

MISSING MARKETS: NUMBER OF COMMON YEARS

* = Significant at the 10 percent level.

**= Significant at the 5 percent level.

*** = Significant at the 1 percent level.

Notes: Standard errors clustered by market i and market j in parentheses. All regressions are OLS and include a constant. Controls are ln(distance) in km, both coastal, connected to river, rainfall correlation, temperature correlation, and absolute differences in: altitude, latitude, longitude, rainfall, temperature, land quality, ruggedness, malaria, humidity, precipitation, slope, religion, and suitabilities for growing banana, chickpea, cocoa, cotton, groundnut, dryland rice, oil palm, onion, soybean, sugar, tea, wetland rice, white potato, wheat, and tomato. Fixed effects are for markets i and j.

Source: See the text.

As a second approach, we evaluate whether markets that are more linguistically distant than those within a set radius experience prices that are more volatile. Our logic here is that linguistic distance from neighbors may lead to more volatile prices because of reduced trade and arbitrage. For each market i, we keep the other markets within 500 kilometers and take the average of their linguistic distance from i (denoted \[\overline {{\text{LinguisticDistanc}}{{\text{e}}_{{\text{ij}}}}} \]) as well as the average of the controls (denoted \[\overline {{x_{ij}}^p} \]). We estimate:

(6) \[C{V_i}^p = {\beta ^p}{\text{ }}\overline {{\text{LinguisticDistanc}}{{\text{e}}_{{\text{ij}}}}} + \overline {{x_{ij}}^p} \prime {\gamma ^p} + {\varepsilon _i}^p.\]

In Equation (6), \[C{V_i}^p\] is the coefficient of variation of the price of product p at market i. We estimate Equation (6) by ordinary least squares (OLS) and report robust standard errors. Results are presented in Table 5. While we find evidence that wheat prices are more volatile at markets that are more linguistically distant from others in their neighborhood, we find no similar evidence for rice or salt. The differences by crop here are somewhat puzzling, as it is rice prices that are most volatile in our data, as measured by the coefficient of variation.

Table 5

MISSING MARKETS: VOLATILITY

* = Significant at the 10 percent level.

**= Significant at the 5 percent level.

*** = Significant at the 1 percent level.

Notes: Robust standard errors in parentheses. All regressions are OLS and include a constant. Controls are minimum year, maximum year, number of observations and averages of ln(distance) in km, both coastal, connected to river, rainfall correlation, temperature correlation, and absolute differences in: altitude, latitude, longitude, rainfall, temperature, land quality, ruggedness, malaria, humidity, precipitation, slope, religion, and suitabilities for growing banana, chickpea, cocoa, cotton, groundnut, dryland rice, oil palm, onion, soybean, sugar, tea, wetland rice, white potato, wheat, and tomato.

Source: See the text.

TRADING COMMUNITIES

To evaluate whether the presence of trading networks sharing a common tongue drives our results (e.g., as might be the case if small communities of traders have lower costs of establishing themselves in regions where the dominant language resembles their own), we correlate linguistic distance with the common presence of communities such as the Marwaris or Parsis. We find little evidence that the co-presence of these communities correlates with linguistic distance.

We focus on one group that has received particular attention in the literature: the Marwaris. By 1920, between 200,000 and 400,000 Marwaris, most of them working as traders, lived outside of the Rajputana Agency (Markovits Reference Markovits2008). These traders drew on capital and personnel from throughout the subcontinent. They gained dominant positions in regional trade, importing, exporting, and moneylending. These communities held assets jointly in patrilineal extended families, sharing information and personnel (Roy Reference Roy2014).

For each pair of markets i and j, we estimate the absolute difference in Marwari share, or \[A{D_{ij}}^{{\text{Marwari}}} = {\text{ }}|{S_i}^{{\text{Marwari}}} - {S_j}^{{\text{Marwari}}}|\]. We then estimate Equation (1) with \[A{D_{ij}}^{{\text{Marwari}}}\] as both an outcome and as a control. That is, we test whether linguistic distance predicts the colocation of Marwaris across district pairs, and the degree to which the co-presence of this trading community can account for the conditional correlation between lingusitic distance and market integration. Results are presented in Table 6. There is little evidence of linguistic distance driving differences in the presence of this trading community, and little evidence that it explains price integration.

Table 6

TRADING COMMUNITIES

* = Significant at the 10 percent level.

**= Significant at the 5 percent level.

*** = Significant at the 1 percent level.

Notes: Standard errors clustered by market i and market j in parentheses. All regressions are OLS and include a constant. Controls are minimum year, maximum year, number of observations, ln(distance) in km, both coastal, connected to river, rainfall correlation, temperature correlation, and absolute differences in: altitude, latitude, longitude, rainfall, temperature, land quality, ruggedness, malaria, humidity, precipitation, slope, religion, and suitabilities for growing banana, chickpea, cocoa, cotton, groundnut, dryland rice, oil palm, onion, soybean, sugar, tea, wetland rice, white potato, wheat, and tomato. Fixed effects are for markets i and j.

Source: See the text.

Results are similar if we perform the same exercise for the other communities listed, although we do not report these for space. While we cannot observe all these communities in our data, several are recorded in the census either as linguistic or religious groups. In particular, we are able to observe the Parsis, Afghanis, Gujaratis, Khatris, Memons, Multanis, and Sindhis. We also observe the Vanis, but they are not present in the markets in our data. Since the English could also be potentially thought of as another migrant mercantile community, we also consider their presence. Results are again similar, and again not reported, using the English. Our results are particularly unlikely to be explained by the spread of the English language: less than one-tenth of 1 percent of the population in the 1901 census is recorded as “English” by language.

Alternatively, if we replace the absolute difference in the population share of a minority group with the maximum for a market pair, results are very similar. Because a group is often present in one market and not another, the maximum across a pair is highly correlated with the absolute difference in shares. Similarly, we find little correlation between linguistic distance and the minimum presence of a trading community across a market pair, and our results are not generally sensitive to controlling for this minimum. Again, we omit these results for space.

LITERACY

In a related test for the costs of information, we examine whether linguistic distance correlates with differences in literacy rates. While linguistically distant markets have more dissimilar literacy rates, this does not diminish the correlation of linguistic distance with market integration.

For data on literacy, we use the 1921 Census of India. These data report literacy at the district level, and we match each market to the district that contains it. As with the presence of trading communities, for each community, we take this difference as both an outcome and as a control. We present results in Table 7. More linguistically distant markets have more dissimilar literacy rates, but this does little to predict price correlations, or to explain away their correlation with linguistic distance.

Table 7

LITERACY RATE

* = Significant at the 10 percent level.

**= Significant at the 5 percent level.

*** = Significant at the 1 percent level.

Notes: Standard errors clustered by market i and market j in parentheses. All regressions are OLS and include a constant. Controls are minimum year, maximum year, number of observations, ln(distance) in km, both coastal, connected to river, rainfall correlation, temperature correlation, and absolute differences in: altitude, latitude, longitude, rainfall, temperature, land quality, ruggedness, malaria, humidity, precipitation, slope, religion, and suitabilities for growing banana, chickpea, cocoa, cotton, groundnut, dryland rice, oil palm, onion, soybean, sugar, tea, wetland rice, white potato, wheat, and tomato. Fixed effects are for markets i and j.

Source: See the text.

INFRASTRUCTURE

Finally, we examine whether linguistic distance proxies for shared preferences over public goods, in particular, those that facilitate trade. We show that more linguistically distant markets spend less time both connected to the railway network, but, nonetheless, this does not fully account for our main result.

Following a procedure similar to Donaldson (Reference Donaldson2018), we use the 1934 edition of History of Indian Railways Constructed and In Progress to identify the year each market became connected to the colonial railway. This source divides the Indian railway system into segments (e.g., “Karimganj to Badarpur”) with a date of opening (in this example, 4-12-96) and length in miles (in this example, 12.00). We use these data to code the first date at which the district containing each market was connected to the Indian Railway system. For each market pair ij, we can then identify the number of years up to 1921 that both markets were connected to the railway system. We then estimate Equation (1) with this variable as both an outcome and as a control. We present results in Table 8. More linguistically distant markets spend more time both connected to the railroad; however, this does little to predict price correlations or explain away their correlation with linguistic distance. One possible contributing factor to these results is the nature of the Indian railways, which were often built to track pre-existing trade routes (Andrabi and Kuehlwein Reference Andrabi and Michael2010).

Table 8

RAILWAY CONNECTIONS

* = Significant at the 10 percent level.

**= Significant at the 5 percent level.

*** = Significant at the 1 percent level.

Notes: Standard errors clustered by market i and market j in parentheses. All regressions are OLS and include a constant. Controls are minimum year, maximum year, number of observations, ln(distance) in km, both coastal, connected to river, rainfall correlation, temperature correlation, and absolute differences in: altitude, latitude, longitude, rainfall, temperature, land quality, ruggedness, malaria, humidity, precipitation, slope, religion, and suitabilities for growing banana, chickpea, cocoa, cotton, groundnut, dryland rice, oil palm, onion, soybean, sugar, tea, wetland rice, white potato, wheat, and tomato. Fixed effects are for markets i and j.

Source: See the text.

ROBUSTNESS

Selection on Unobservables

In this section, we demonstrate the robustness of our results to selection on unobservables. We present a number of additional exercises in the Online Appendix.

To demonstrate robustness to selection on unobservables, we use the approach of Altonji, Elder, and Taber (Reference Altonji, Elder and Taber2005) as implemented by Bellows and Miguel (Reference Bellows and Edward2009) and Nunn and Wantchekon (Reference Nunn and Leonard2011). We estimate Equation (1) with either a limited set of controls or with a full set of controls, and compute:

(7) \[AET = \frac{{{\beta ^{{\text{FullControls}}}}}}{{{\beta ^{{\text{RestrictedControls}}}} - {\beta ^{{\text{FullControls}}}}}}.\]

We report results where the restricted set of controls is either empty or contains only ln(Distance). Larger values of this statistic imply that the selection on unobservables would need to have a larger effect on β relative to that of observables in order to be consistent with a true β of 0. Results are presented in Table 9. The coefficient estimates for wheat are sensitive to controls regardless of what is in the base set of controls, but are not as sensitive to the addition of fixed effects. Results for salt and rice appear sensitive to adding fixed effects and controls together, but this is driven by ln(Distance). Once this is included as a baseline control, AET is negative (i.e., controls push β away from zero) or greater than one. That is, we find that the estimate of β is sensitive to controls for wheat, while for salt and rice, the estimate of β is no longer sensitive to controls once ln(Distance) has been included.

Table 9

ALTONJI-ELDER-TABER STATISTICS

* = Significant at the 10 percent level.

**= Significant at the 5 percent level.

*** = Significant at the 1 percent level.

Notes: Standard errors clustered by market i and market j in parentheses. All regressions are OLS and include a constant. Controls are minimum year, maximum year, number of observations, ln(distance) in km, both coastal, connected to river, rainfall correlation, temperature correlation, and absolute differences in: altitude, latitude, longitude, rainfall, temperature, land quality, ruggedness, malaria, humidity, precipitation, slope, religion, and suitabilities for growing banana, chickpea, cocoa, cotton, groundnut, dryland rice, oil palm, onion, soybean, sugar, tea, wetland rice, white potato, wheat, and tomato. Fixed effects are for markets i and j.

Source: See the text.