Learning English in a Globalizing World: Information. Technology and Factor Supply Response in India

Transcription

1 Learning English in a Globalizing World: Information Technology and Factor Supply Response in India (Previously circulated as "Speaking English in a Globalizing World: Information Technology and Education in India") Gauri Kartini Shastry Harvard University September 2007 Abstract A number of studies suggest that globalization and trade liberalization in developing countries have increased returns to education and inequality. A key question is whether this will lead to changes in factor supplies that mitigate this inequality. I study how Indian trade liberalization in the early 990s impacted human capital investment by exploiting variation in English learning costs across districts. This variation is driven by linguistic diversity which compels individuals to learn additional languages; most people choose between English and Hindi, the native lingua franca. Individuals whose native language is linguistically further from Hindi have lower relative opportunity costs of learning English, because they nd Hindi harder to learn and often su er psychic costs since many non-hindi speakers feel Hindi was imposed on them. I demonstrate that more bene ts of globalization go to districts with lower costs of learning English: I show that these districts experienced greater growth in information technology presence and greater growth in school enrollment, but smaller increases in the skilled wage premium. JEL codes: F6,I2,O5 Correspondence: shastry@fas.harvard.edu. I am grateful to David Cutler, Esther Du o, Caroline Hoxby, Michael Kremer for their advice and support, Petia Topalova for assistance with the National Sample Surveys, David Clingingsmith for access to additional data and all participants of the Labor Economics/Public Finance and Development Economics lunch workshops at Harvard, the IZA/World Bank Conference on Employment and Development 2007 and the Federal Reserve Board of Governors International Finance workshop for their comments. In addition, I thank Filipe Campante, Davin Chor, Quoc-Anh Do, Eyal Dvir and Michael Katz for useful conversations and assistance. Finally, I thank Daniel Tortorice for invaluable support and encouragement. All errors are mine.

2 Introduction How do households respond to globalization? Economists often debate the costs and bene ts of globalization for individual households. Much theoretical and empirical research has highlighted e ects on inequality and skilled wage premiums; a fundamental factor in this discussion is how skilled labor supply responds. In this paper, I study India s trade reforms in the early 990s and show that households increase the human capital accumulation of their children. More importantly, how globalization impacts school enrollment and wage premiums depends on the elasticity of the English-speaking skilled labor supply. I show that areas with greater English literacy received more bene ts from globalization: they had more exporting information technology (IT) rms and greater growth in school enrollment. There are two characteristics of Indian trade liberalization worth noting; they correspond to two motivating themes in this paper. First, the Indian experience with trade reforms contradicts some predictions of standard trade theory. Under a simple Heckscher-Ohlin trade model with two goods and two countries, the poor country should specialize in unskilled labor intensive industries after liberalization. Like in other countries, trade liberalization in India did not follow this pattern: much of Indian trade growth occurred in high-tech service industries that hired skilled, English speakers. The literature has explored several explanations for this phenomena in various countries, e.g. emphasizing speci c elements of the reforms or institutions, international ows of capital (Feenstra and Hanson 996, 997) and complementarity between skill levels across countries (Kremer and Maskin 2006). I focus instead on the consequences of this type of liberalization, abstracting from why India exports skilled labor intensive goods. One important consequence is the impact on skilled wage premiums: several recent empirical studies in Latin America found that trade liberalization increased skilled wage premiums 2. There is little existing evidence on East Other explanations are discussed in Goldberg and Pavcnik (2004). 2 See Goldberg and Pavcnik (2004) for a review of the literature, which includes, e.g., Hanson and Harrison (999), Feenstra and Hanson (997), Feliciano (993) and Cragg and Epelbaum (996) on Mexico, Robbins, Gonzales and Menendez (995) on Argentina, Robbins (995a) on Chile, Robbins (996a) and Attanasio, Goldberg and Pavcnik (2004) on Colombia, Robbins and Gindling (997) on Costa Rica and Robbins (995b,

3 Asia 3, but conventional wisdom indicates a much smaller e ect on wage premiums. A possible explanation for this di erence is that Latin America experienced a much smaller increase in the supply of skilled workers than East Asia. This speaks to the possible role played by factor supply in mitigating globalization s e ect on inequality. India provides a promising context to study this hypothesis because trade liberalization created skilled jobs and there is variation in the elasticity of factor supply. The second characteristic of trade liberalization in India, and the second motivating theme, is the central role played by language skills. Due to nature of India s service exports (e.g. in IT), exporting rms need to hire English-speaking skilled workers. I exploit variation in the cost of learning English generated by historical di erences in language policies. Countries have struggled with whether to encourage people to retain diverse local languages, choose a single native language or promote a global lingua franca, such as English. The choice of medium of instruction in public schools has far-reaching consequences. 4 Native language instruction can strengthen national identity, while instruction in a foreign language may be less accessible for poor households. At the same time, if the global language is used by coveted white-collar jobs, public school instruction in that language may increase economic opportunities for the poor. 5 The ability to integrate better with the world economy may bring more trade bene ts to places that promote English. Since much technological progress happens in English, speaking English may facilitate the adoption of new technologies. This paper provides evidence on the labor supply response to globalization. I show that the e ect of trade liberalization varies with labor supply elasticity, by exploiting exogenous variation in the cost of learning English. Most measures of the cost of learning English are endogenous. State governments that care more about access to global opportunities may advance the teaching of English, but also pursue other policies that promote trade. I utilize variation driven by linguistic and historical forces that is exogenous to such outward- 996b) on Uruguay. 3 See Wood (997) for a survey, Lindert and Williamson (200) and Wei and Wu (200). 4 See Human Development Report (2004) and Angrist, Chin and Godoy (2006). 5 See Lang and Siniver (2006), Angrist and Lavy (997) and Munshi and Rosenzweig (2006). 2

4 oriented policies and to reverse causality. In fact, this variation caused people to learn English for non-trade related reasons even in 96, long before IT trade in the 990s was contemplated. In India, linguistic diversity motivates individuals to learn a second language even to communicate locally. The common choices are English and Hindi, the native lingua franca. Individuals whose mother tongue is linguistically further from Hindi have a lower relative opportunity cost of learning English because they nd Hindi more di cult to learn, but also because they are more likely to su er psychic costs when speaking Hindi, a language that many feel was imposed on them as a national language. Over time, these historical tendencies led to the growth of institutions promoting English instruction in places where native languages are further from Hindi. I rst show empirically that linguistic distance from Hindi induces people to learn English and schools to teach English. Next, I con rm that the impact of globalization during the 990s has varied with pre-existing di erences in linguistic distance from Hindi. I gathered and examined data on the information technology (IT) sector in India, an industry that grew primarily due to economic liberalization and technological progress. I use IT to refer to both software rms and business-process outsourcing such as call centers and data entry rms. I show that IT rms were more likely to locate and employ more workers in districts with lower costs of learning English, as measured by linguistic distance. Turning next to human capital accumulation, I show that districts where the native languages are linguistically further from Hindi experienced greater increases in urban school enrollment from 993 to 2002 even relative to pre-existing trends. Finally, I use micro-level data to demonstrate that, in the 990s, districts with lower costs of learning English experienced a smaller increase in the skilled wage premium. A simple theoretical model provides intuition for these results. In all of these results, I show how pre-existing di erences in linguistic diversity explain changes over time in employment and wages, controlling for region-speci c trends; I estimate not just correlations but the causal impact during a period of liberalization. The implications of this paper are two-fold. The rst implication concerns the bene ts of 3

5 policies promoting a global language (and other export-relevant skills) in a developing country prior to liberalization. Indian districts with more English speakers were better positioned to gain from trade. 6 The second implication explains mixed evidence concerning globalization s impact on wage inequality and provides evidence on the long term consequences of globalization. While trade reforms may increase wage inequality in the short run, this e ect could be dampened in the long run as factor supply responds. 7 My results may also allay concerns that globalization only bene ts the wealthy: I nd an increase in school enrollment even in early grade school, which is unlikely to be driven by the wealthy. The paper is organized as follows. Section 2 provides background information on trade liberalization and information technology in India and section 3 describes the Indian linguistic context. In section 4, I discuss the empirical strategy and identi cation. Section 5 describes a simple theoretical model to provide intuition for the empirical ndings. Section 6 provides results regarding IT rm presence and school enrollment growth and a robustness check using variation in the percent of English speakers in 96. Finally, section 7 provides suggestive evidence on returns to education and section 8 concludes. 2 Background on trade liberalization and IT During much of the post-colonial period, India heavily protected its economy through strict investment licensing and import controls. In the late 970s and 980s, the government took small steps towards integrating with world markets. The Open General Licensing list (items that did not require a license) was revised to include many more capital goods and intermediate inputs. Tari rates were raised substantially, but due to the extensive licensing, 6 Munshi and Rosenzweig (2006) show that increases in returns to English dwarfed those in returns to education and that enrollment rose in English-medium schools from 980 to 2000 using a household survey from a suburb of Mumbai, India. I show that educational attainment overall rises using nationally representative data. I also explore a trade-related explanation for the rise in the returns to English. 7 Edmonds, Pavcnik and Topalova (2005) also study the relationship between economic liberalization and education in India and nd an adverse impact on schooling. While the authors isolate the e ect on school enrollment of reduced family income due to import competition, I focus on the impact of increases in job opportunities and returns to English education from exports and integration with world markets. 4

6 they did not e ectively restrict imports (Panagariya 2004). Even as late as 990, however, numerous obstacles limited trade. The average tari was 79%, and 65% of imports faced non-tari barriers (Panagariya 2003). The policies in the 980s raised India s growth rate, but heavy borrowing led to a balance-of-payments crisis in 99, which resulted in a comprehensive shift towards an open economy. Reforms ended most import licensing requirements for capital goods and reduced tari rates dramatically. Import licensing on consumer goods was not eliminated until much later in 200 and agricultural products were still subject to high tari s. The March 992 Export-Import Policy also dramatically increased the number of items that could be exported (Panagariya 2004). The government also revised the heavy regulation of service sectors, such as insurance, banking, telecommunications and infrastructure. These sectors were opened up to private participation and foreign investment. The government freed up telephone services in the 994 National Telecommunications Policy (revised in the New Telecom Policy of 999 due to technological progress). Foreign direct investment in e-commerce was free of all restrictions and foreign equity in software and electronics was granted automatic approval, particularly for IT exporters. These policy changes led to remarkable growth in Indian exports; the annual growth rate rose by 3.3 percentage points in the 990s from the 980. Service exports grew more rapidly than manufacturing exports; even within manufacturing, capital or skilled labor intensive sectors grew faster (Panagariya 2004). Along with technological progress, this policy shift led to remarkable growth in IT services outsourced to India. By 2004, India was the single largest destination for foreign rms to purchase IT services, contributing two-thirds of global software outsourcing and half of business process outsourcing. In 2005, IT outsourcing accounted for 5% of India s GDP and was forecasted to contribute 7% to projected growth to 200 (The Economist 2006). Employment growth has been strong over the past decade; from 56,000 professionals in 990-9, the sector employed 83,500 in 2003, implying an annual growth rate of more than 20% (NASSCOM 2004). In particular, the IT sector increased job opportunities for young, educated work- 5

7 ers; the median IT professional is 27.5 years old and 8% have at least a bachelor s degree (NASSCOM 2004). An entry-level job in a call center pays on average Rs. 0,000 ($230) considered very high for a rst job (The Economist 2005). IT rms are relatively free to locate based on the availability of educated, English-speaking workers, due to their young age, export focus and reliance on foreign capital. It is important to note that much of the liberalization in the early 990s was a shock; the small steps in the 980s were ad-hoc and did not re ect a systematic shift towards an outward-oriented economy. An underlying assumption for my identi cation strategy is that these reforms were not implemented di erently in areas with many English speakers in order to gain from trade in services. This is unlikely since the balance-of-payments crisis that motivated the policy shift and the advances in telecommunications and information technology were unrelated and unexpected. The return to speaking English was higher in the 990s than in the 980s both because of technological progress and trade policy. 3 Linguistic distance from Hindi 3. Language of government and media of instruction The 96 Census of India documented 652 mother tongues from ve distinct language families native to India. These families are quite diverse; while linguists assert familial relationships between languages as distinct as English and Hindi (both are Indo-European), they are unwilling to connect some languages native to India such as Hindi and Kannada, the language spoken in Karnataka. As of 99, 22 languages were native to more than one million people and 4 languages had more than 0,000 native speakers. Due to this diversity, many individuals need to learn a second language to communicate at a local level, especially in urban areas (Clingingsmith 2006). Of all multilingual individuals who are not 6

8 native speakers, 60% chose to learn Hindi and 56% chose to learn English 8. These two are clearly the dominant choices: at 6%, Kannada was the next most popular second language. The choice of whether to learn Hindi or English depends on an individual s mother tongue. Someone whose mother tongue is similar to Hindi will nd it easier to learn Hindi relative to someone whose mother tongue is more di erent. Those close to Hindi have a greater opportunity cost of learning English. Historical forces have ampli ed this tendency. During the British occupation, English was established as the language of government and of instruction. After India won independence in 947, a nationalist movement to make an indigenous language the o cial language favored Hindi, since it was spoken by the most people. This movement was opposed by non-native Hindi speakers, but after much debate, the o cial status of Hindi was written into the constitution. This led to riots in non Hindispeaking areas, the most violent of which occurred in Tamil Nadu in May 963. Speakers of other languages felt at a disadvantage speaking Hindi and nally, in 967, the government made Hindi and English joint o cial languages (Hohenthal 2003). This history explains higher levels of English literacy among people who speak languages linguistically distant to Hindi. In fact, in some states more people speak English than Hindi. This theory has ambiguous predictions for native Hindi speakers. While they have less need to learn a second language, English would allow them to interact with more additional people than any other choice. I examine their behavior empirically. Since schools often require a second language, we would expect Hindi natives to be more likely to learn English. Over time, this relationship became institutionalized through media of instruction. British colonizers drove early movements in Indian education by setting up urban schools taught in English. Their goal was fostering an elite class to help govern the country (Nurullah and Naik 947, Kamat 985). By 850, missionary societies and princely states had set up rural schools in native languages. In 854, after the British government committed to educating the entire population (Dakin, Ti en and Widdowson 968), education spread to lower classes 8 More than three hundred million people speak Hindi as a rst language, while only 80,000 are native English speakers; thus, many more people spoke Hindi than English. 7

9 and an increasing number of schools taught in native languages. University education in major cities was still primarily in English. Even today, it is the main medium of post-tertiary instruction (Hohenthal 2003). In 993, according to the Sixth All India Educational Survey, there were over 28 media of instruction in urban primary schools; Hindi was most common with 38% of schools and English was second with 9%. At the secondary level, this di erence was smaller: Hindi was the media of instruction in 29% of schools and English in 20%. 3.2 Measuring linguistic distance As there is no universally accepted measure of language distance, I calculate three independent measures. My preferred measure was developed in consultation with an expert on Indo-European languages, Jay Jasano, the Diebold Professor of Indo-European Linguistics and Philology at Harvard University. This measure is based on classifying languages into ve degrees of distance from Hindi (see gure ). A second measure is based on a language family tree from the Ethnologue database, one of the most comprehensive listings of languages. Figure 2 provides an extract from this language tree that includes all languages found in the Indian census. I de ne distance as the number of nodes between two languages. For example, Urdu is two degrees from Hindi, while Marwari is four degrees. In order to link all languages, I assume there is a node connecting the di erent families. Another measure is taken from glottochronology, a method to estimate the time of divergence between languages (Swadesh 972). This method takes a list of 20 core meanings and uses expert judgments on how many cognates exist between a pair of languages. 9 I use the percent of cognates shared between each language and Hindi from the Dyen et al. (992) dataset of 95 Indo-European languages. Table provides example core meanings in English, Bengali and Hindi as well as cognate judgments. Since the dataset contains only Indo-European languages, I assume languages in other families have only 5% of words 9 Glottochronology involves a formula to convert this measure into the time of divergence, but is currently out of favor among linguists. The percent of cognates is still an acceptable measure of similarity. 8

10 in common with Hindi. 0 Reassuringly, the correlation between these 3 measures is quite high ( between degrees and nodes, and respectively between these two measures and the percent of cognates). Using the Jay Jasano measure and the 99 Census of India, I calculate a district s linguistic distance from Hindi in two ways: ) the population-weighted average distance of all native languages and 2) the share of the population who speak languages at least 3 degrees away (hereafter, distant speakers ). Table 2 provides summary statistics on language. 4 Identi cation and empirical speci cations 4. Linguistic distance and learning English This identi cation strategy - using variation in the cost of learning English driven by linguistic distance from Hindi - depends on two assumptions. First, linguistic distance from Hindi must predict the cost of learning English and second, it must not be correlated with omitted factors that a ect schooling or exports. I could then instrument for English learning costs with distance to Hindi. However, a comprehensive district-level measure of this cost is unavailable, partly due to data limitations and partly because the cost of English is multidimensional. We would want to take into account how many schools teach English and how many adults speak English at the very least, both unavailable at the district level. Instead, I use this variation in two ways: rst, I present reduced form results using the weighted average and percent distant speakers measures of linguistic distance. Second, I proxy for the cost of learning English with the percent of schools teaching in the regional mother tongue, the only information available at the district level. Since English is not the regional mother tongue anywhere in India, this proxy is a lower bound for schools that do 0 Linguists use 5% as a signi cance level when determining whether two languages are related: languages are unrelated if less than 5% of words are cognates. The lowest percent of cognates with Hindi among Indo-European languages is 4.6%. For Indo-European languages not in Dyen et al. s list, I use the value of the closest language in the tree that also appears in the list. Close matches exist for all 2 such languages. 9

11 not teach in English. In this subsection, I provide implicit evidence for the rst assumption by demonstrating that linguistic distance from Hindi predicts English literacy using data on second languages spoken by state and mother tongue. In the next subsection I describe the empirical speci cations to test the e ects of globalization and in the following subsection, I discuss support for the second assumption. Using data from the Census of India (96 and 99), I estimate E lkt = D l + 0 X lk + t + g + lkt where E lkt is the percent of native speakers of language l in state k, region g and year t who choose to learn English (conditional on being multilingual), D l is the distance of language l from Hindi, and t is a year xed e ect. I include region xed e ects, g, for north, northeast, east, south, west and central India. The vector X lk includes the share of language l speakers in state k and an indicator for whether language l is the state s primary language. I weight observations by the number of native speakers and cluster the standard errors by state. The results con rm the relationship between learning English and linguistic distance to Hindi (table 3). In column I include only dummy variables for year, region, and each distance from Hindi; in column 2 I include all controls. Languages degree away from Hindi are the omitted group. Two points are worth making. First, individuals at a linguistic distance of zero (including native speakers of Hindi and Urdu ) are very likely to learn English, since they already know Hindi. This creates a non-monotonicity in the relationship between linguistic distance and English learning. In gure 3, I plot the coe cient on each distance from Hindi with a 95% con dence interval to convey this result graphically. The propensity to learn English dips signi cantly when we move slightly away from native Hindi speakers and then rises as native tongues get more di erent. In my main speci cation, I account for the non-monotonicity by controlling for the percent of native Hindi speakers. I also calculate weighted average linguistic distance only among non-hindi-speakers. I can Hindi and Urdu are often considered the same language. Separating them gives similar results. 0

12 also exploit this non-monotonicity since districts with more Hindi speakers, all else equal, are going to have lower costs of learning English. Second, linguistic distance from Hindi signi cantly predicts how many multilingual individuals learn English. The p-values at the bottom of the table test whether the dummy variables are jointly di erent from zero. This result is obtained whether or not I include the dummy variable for a distance of zero. However, the individual dummy variables are not linearly increasing. This is partly due to Hindi speakers but also due to languages at four degrees away. These languages - Kashmiri, Nepali, Sindhi and Dogri (mainly spoken in Jammu and Kashmir) - are likely to not follow this trend since many native speakers are living in other parts of India due to ethnic and political con icts. When I control for the share of native speakers in the state, this second non-monotonicity is resolved. In my speci cations below, I include many district controls to account for this nding. It is also important to examine both the weighted average and percent distant speakers measures, since the latter does not impose a linear structure. Columns 3-4 in table 3 test both reduced form measures of linguistic distance. One linguistic degree increases the percent of multilinguals who learn English by 8.9 percentage points. Speaking a language 3 or more degrees away from Hindi increases the percent of English speakers by 42 points relative to speakers of languages and 2 degrees away. Columns 5-8 con rm these ndings for 96 and 99. The results are robust to using the overall share of English learners or the number of English speakers as the dependent variable. The control variables matter as expected: individuals are more likely to learn English in 99 than 96 and they are more likely to learn English the more other people in their state share their mother tongue. This second result is reasonable since speakers of minority languages are likely to learn the regional language before English. I reject the hypothesis that speaking a language any distance from Hindi has a uniform e ect on English learning by testing the equality of all linguistic distance xed e ects. The results in table 4 lend further credence to the use of linguistic distance. Columns -2

13 show linguistic distance to Hindi does not in uence the decision to become multilingual. The control variables have the expected sign: more people become multilingual in 99, and fewer learn second languages when their mother tongue is spoken by more neighbors. Columns 3-4 demonstrate that linguistic distance to Hindi is measured sensibly since individuals further from Hindi are less likely to learn Hindi. 2 I next explore how distance from Hindi of languages spoken in a district predicts whether schools teach English by estimating M ij = D j + P j + 0 2Z j + i + g + ij () where M ij is a measure of language instruction at school level i (primary, upper primary, secondary or higher secondary) in area j (either a state or district), D j captures the linguistic distance from Hindi of languages spoken in area j, and P j is child population (aged 5-9, in millions). The vector Z j includes average household wage income, average income for individuals with secondary schooling, the percent of adults with regular jobs, the percent of households with electricity, and the percent of people who: have graduated from college, have completed high school, are literate, are Muslim, or regularly use a train. All of these are measured for urban areas in 987. I include the distance to the closest of the 0 biggest cities and the percent of native English speakers. I control for the percent of people who are native Hindi speakers to account for the non-monotonicity described above and an indicator variable for whether the district is on the coast to account for potential trade routes 3. To ensure that these results are not driven by large Hindi-speaking populations in the "Hindi Belt" states who are particularly poor due in large part to corruption and government inef- ciency, I include an indicator variable for the following states: Bihar, Uttar Pradesh (and Uttaranchal), Madhya Pradesh (and Chhattisgarh), Haryana, Punjab, Rajasthan, Himachal Pradesh, Jharkhand, Chandigarh and Delhi. In addition, I include region and school level 2 All results in tables 3 and 4 are robust to including state xed e ects and to clustering by native language, to account for correlated error terms between speakers of a language across states. 3 All results are also robust to including a dummy variable for bordering a foreign country. 2

14 xed e ects. The data appendix describes the sources. 4 The results show that linguistic distance from Hindi predicts the percent of schools that teach in English (columns -2 of table 5) or teach English as a second language at the state level (columns 3-4). An increase in one degree in the average distance from Hindi increases the percent of schools teaching in English by about 20 percentage points and the percent of schools teaching English by 33 points. More Hindi speakers also increases the teaching of English, but only when linguistic distance is measured by the percent distant speakers. Finally, I examine the percent of schools in a district that teach in the regional mother tongue, a noisy lower bound on how many schools do not teach in English (columns 5-7). The coe cient is negative as expected, but not signi cant; allowing a more exible relationship con rms the in uence of linguistic distance (column 7). 4.2 Empirical speci cations Using this variation, I rst test empirically whether globalization has had a larger impact in districts with lower costs of learning English. I show this by studying the IT sector. Growth in IT is one potential bene t of promoting a global language, since IT rms hire mostly educated English speakers. I estimate T jt = D j + 0 Z j + 0 2W j + t + g + jt (2) where T jt (for technology) is a measure of IT presence in district j in year t and D j measures linguistic distance. Z j is as in equation () and W j includes log population, the number of elite engineering colleges, the distance to the closest airport, and the percent of non-migrant adults with an engineering degree in 987. I cluster the standard errors by district. The measures of IT presence include the existence of any headquarters or branches, the age of the oldest rm, the number of headquarters and branches and the number of 4 The percent of schools teaching in the mother tongue can be greater than due to noise in the data. 3

15 employees. 5 The data, described in the appendix, contains only rm-level, not branch-level, employment; I assign employees evenly across branches to estimate employment by district. Summary statistics can be found in table 6. To study whether school enrollment responds to globalization di erently in districts with di erent costs of learning English, I estimate log (S ijt ) log(s ijt ) = D j I (t = 2002) + log (S ijt ) + 0 2P jt (3) + 3 I (t = 2002) + 0 4Z j I (t = 2002) + 5 B j + j + t + gt + it + ijt where S ijt is urban enrollment at grade level i in district j, region g and time t 2 f987; 993; 2002g, I () is an indicator function, D j and Z j are as above 6 and P jt is a vector of log child population at time t and t. The district e ect, j, allows independent enrollment trends and it is a cohort e ect. Region xed e ects are interacted with time, allowing for region-speci c changes in trend. I use only data from urban areas and cluster by district. I also include B j, a proxy for labor demand growth for educated workers, to control for other factors that may in uence demand for education. B j measures predictable changes in skilled labor demand. This variable, calculated along the lines of Bartik (99) and Gelber (2007), is a weighted average of national industry employment growth rates where the weights are pre-liberalization industrial composition of district employment. This control should not be correlated with local labor supply shocks but it may pick up some of the e ect of liberalization through national employment growth. Details are in the data appendix. Since the data consist of the number of students enrolled, not enrollment rates, I control for population among 5-9 year-olds from the 99 and 200 Census - I cannot use a more precise population for each grade since children start school at di erent ages and often fail 5 I take the natural log of headquarters and branches plus one and employment plus one to avoid having to drop districts with no IT presence. 6 Since I already include district xed e ects, Z j is interacted with I (t = 2002), allowing pre-liberalization di erences to have a di erent e ect post reform. The results are robust to not including these controls. 4

16 to be promoted. Enrollment increased dramatically, by 32%, between 993 and 2002 (table 7), but the increases were greater in the highest grades. 4.3 Linguistic distance and omitted factors It is important to pinpoint the source of identi cation in the empirical tests. Most measures of the cost of learning English, e.g. the number of English speakers, are likely to be correlated with unobserved characteristics that in uence schooling or export-oriented policies of local governments. If a government cares about access to global opportunities, it may both promote English and provide incentives for foreign direct investment. We may also worry about reverse causality, since outsourcing rms often set up English training centers. The variation from linguistic distance to Hindi su ers less from these problems. Distance to Hindi impacts the cost of learning English in a manner that is orthogonal to preferences of local governments and export-oriented motives before 990. The results in columns 5-6 of table 3 help discredit this concern. The tendency for ethnic groups speaking languages further from Hindi to learn English was evident in 96, before anyone could anticipate the 990s trade liberalization and the large returns to speaking English. This supports the assumption that these groups were not more forward-looking or outward-oriented in the 980s; it is unlikely they anticipated the trade bene ts to learning English. Similarly, government policies or English-language opportunities will not a ect the linguistic distance of a language from Hindi. Large movements of people across districts may alter the distance from Hindi of languages spoken in an area, but migration in India is still quite low. According to the 987 National Sample Survey, only 2.3% of individuals in urban areas had migrated in the past ve years, only 6.8% had moved outside their current district and only 2.4% had moved outside their current state. Nevertheless, linguistic distance to Hindi is not randomly distributed across India. Figure 4 maps the distribution of percent distant speakers. Much of the variation is across region (indicated by thick black lines). Since the geographic variation may be correlated with other 5

17 factors that in uence schooling or exports, such as weather, agricultural productivity or culture, it is very important to include region xed e ects and region-speci c trends in all empirical speci cations. I also include district control variables, even allowing them to have di erent e ects before and after liberalization. In speci cation (3), the xed e ects and interactions ensure that I am identifying o deviations from a district s own trend and the region s change in trend in schooling. This strategy - using preexisting di erences to predict changes in school enrollment, allowing for district and region e ects and controlling for other preexisting di erences - rules out other explanations for these results. To demonstrate the residual variation that I use, in gure 5, I map the residuals from a regression of linguistic distance on region xed e ects and vector Z j. This variation is much cleaner since it is less likely to be correlated with omitted variables. It is also important to ensure that linguistic distance is not correlated with the supply of schools. If it were, we would worry that districts further from Hindi simply had greater preferences for education. In row () of table 8, I present estimates from a regression of the log number of schools on linguistic distance to Hindi, log child population and controls in vector Z j. Neither measure of linguistic distance predicts the number of schools in 993 or the number of schools o ering courses in science or commerce (results not shown). Finally, districts that are linguistically distant from Hindi do not, by de nition, speak the language spoken by the most people. One worry would be that these districts were less integrated with the rest of the country and this may have impacted the evolution of industries or interstate trade. In table 8 rows (2)-(8), I present estimates from a regression of the percent of the labor force employed in speci c industries on vector Z j to demonstrate that linguistic distance has had almost no impact on district industrial evolution. 7 7 These results are robust to not including Z j, running the regression at the individual level, or de ning the outcome variable as the share of the adult population or the total sum of workers in those industries. 6

18 5 Theoretical model Before discussing the empirical results, it is informative to examine theoretical predictions. I describe a simple model of how the e ects of globalization vary with di erent English learning costs. Consider an economy with two districts that di er only in this cost. It is not necessarily true that English-skilled labor in the low cost district (LC) is more elastic than in the high cost district (HC), but it is a reasonable interpretation when districts do not di er greatly and most individuals do not speak English, as is the case in India. I specify production processes for two goods, one of which is tradeable, and a schooling decision between English- or Hindi-medium instruction. I solve the model before and after trade liberalization to compare the changes in education and the skilled wage premium. After trade reforms, the cost of transporting the traded good falls su ciently to facilitate its production and export from this economy. 8 English and Hindi skilled workers are equally productive in the domestic sector, but only English speakers can produce the globally traded good. 9 Goods travel freely between districts, but workers do not. Trade reforms cause a positive demand shock for skilled labor, but the impact depends on the elasticity of supply. Identical shocks would cause a greater increase in employment, but a smaller increase in the wage premium, for educated workers where the supply is more elastic ( gure 6). However, it is unlikely the districts would experience the same shock; more exporting rms should locate in LC where English speakers are cheaper to hire. In this case, there would still be a larger increase in skilled employment in LC, but the relative change in the wage premium depends on the size of the demand shocks ( gure 7). To show this, I assume the global industry needs a second factor that is xed in the short run; the importance of this factor constrains the di erence between the labor demand shocks. 8 I abstract from why India would export a skilled labor intensive good. 9 We can think of this good as IT or any good or service requiring English speaking workers. To simplify the model, I assume that the traded good is not consumed domestically. 7

19 5. Production processes for traded and non-traded goods The non-traded good, Y, is consumed in both districts and produced using LY Y = min ; H Y + E Y L H where L > H where L Y, H Y and E Y are the quantities of unskilled, Hindi skilled and English skilled labor used. Since Hindi and English skilled workers are perfect substitutes, rms hire the cheaper workers. Prior to trade liberalization, English speakers earn q E = q H working in the Y industry. The amount of Y produced is determined by the availability of labor. After trade liberalization, rms who start producing the traded good, X, set up in either district, taking the price of X, p X, as given. The production function is X = F E X where 0 < < where E X is the amount of English skilled labor used and F is the exogenous endowment of the xed factor that earns a return r F. We can think of F as infrastructure (such as telecommunication networks) that is slow to change or immobile entrepreneurs. 5.2 Schooling decisions Individuals live for one period and work as unskilled labor or get instantaneous education in English or Hindi and work as skilled labor. P people are born each period and they di er in a parameter c i which is distributed uniformly over [0; ]. This parameter a ects the relative cost of studying in English versus Hindi; one could think of it as partly corresponding (inversely) to the linguistic distance to Hindi of the individual s mother tongue. Individuals with a low c i nd English easier to learn than individuals with a higher c i. Studying in Hindi costs (t H + c i ) w U where t H is xed (0 < t H < ) and w U is the unskilled wage. Studying in English costs t E + j c i wu where t E is xed (0 < t E < ) 8

20 and j is a district-speci c parameter. This parameter di ers in the low cost and high cost districts: LC < HC. This does not correspond exactly to the Indian context since c i is distributed identically in both districts, but its relationship with English learning costs di ers. This di erence allows for a simpler model, but does not drive the results. Given this simpli ed cost structure, I assume that t E < t H, allowing education in English to be cheaper than in Hindi for some people to ensure that we have some English speakers in autarky. The results are robust to assuming t E = t H but the calibrated model would be less realistic. To simplify the algebra, I let t E = 0 and t H = t. Individuals maximize lifetime income with respect to h,e 2 f0; g, denoting Hindi or English schooling: max h; e 8 >< >: j c i w U + max fw E ; w H g if e=, h=0 tw U c i w U + w H if e=0, h= w U if e=0, h=0 9 >= >; where w H and w E are the wages for Hindi and English skilled workers, respectively. Since all skilled workers are equally productive in the Y sector, w E w H. When solving the individual s problem, I ignore the unrealistic case in which there are no Hindi-skilled workers. Thus, people with low values of c i get English schooling, those in the middle study in Hindi and those with higher c i remain unskilled. Note that even when w E = w H, as in autarky, some individuals will study English. Letting H and E be the number of Hindi and English skilled workers, respectively, I de ne two additional terms: total education, ED = H + E, and the weighted average return to skill, bq = w HH+w E E w U (H+E). 5.3 Characterizing the equilibrium In equilibrium, all labor markets must clear. For unskilled workers this condition is L Y = P w H w U tw U w U (4) 9

21 Skilled-labor market clearing depends on whether the demand for English speakers exceeds their initial supply. Recall that even when w E = w H, some individuals choose to study in English. If, in equilibrium, the demand for English speakers is less than this initial supply (case A), then w E = w H. The market clearing condition is H Y + F w E (p X ( )) wh w U tw U = P w U (5) If the demand for English workers exceeds the natural supply, then w E > w H and no English speakers work in the Y industry (case B). The labor market clearing conditions are H Y = P w H w U tw U w U w U t + w E w H w U j! (6) F w E (p X ( )) = P w U t + w E w H w U j (7) These labor market clearing conditions, zero pro t conditions for each sector and the production function for X close the model. Good Y is the numeraire. The equilibrium without any trade is a special case of A, described in Proposition below, when F = 0. Since the demand for English skilled workers rises after trade liberalization, the wage for English workers has to rise. Now that fewer English speakers are working in the Y industry, the wage for Hindi skilled workers rises as well to keep the ratio of skilled to unskilled workers in Y production constant. To compare how these changes di er in districts with di erent levels of j, we have to rst solve the equilibrium in both case A and B. Proposition Case A. If, in equilibrium, we = w H, i.e. the demand for English skilled workers is less than or equal to the natural supply, then E is falling and H is rising in the cost of learning English, j. Total education, ED, the amount of X produced and the average return to education, bq, are independent of j. Proof. All proofs are found in the appendix. 20

22 If the two districts LC and HC are both in this case, they will have identical wages, production of X and returns to education. They will also have identical total education, but the low cost district will have a higher proportion of English speakers. I next solve case B. Proposition 2 Case B. If, in equilibrium, we > w H, i.e. the demand for English skilled workers is greater than the natural supply, then E is falling and H is rising in the cost of learning English, j. Total education, ED, and the amount of X produced are both falling in j, but the e ect of an increase in j on the average return to education, bq, is ambiguous. A district is no longer in case A when the demand for English workers is greater than the natural supply; the high cost district will leave case A at a lower value of F since it will exhaust its smaller natural amount of English speakers sooner. Proposition 3 (in the appendix) provides the necessary and su cient condition for case A. The intuition behind the e ect of j on returns to education is simple. English-skilled workers are less elastic in HC; therefore the wage must rise more. Similarly, the greater elasticity of Hindi-skilled workers in HC results in a smaller increase in the Hindi wage. These changes are constrained, however, by the constant skilled-unskilled labor ratio in Y production and the relative size of the demand shocks for English speakers. When X production is intensive in F (a high ), the amount of X that can be produced is more constrained by the amount of F available: the demand for English speakers in the two districts cannot be too di erent. Thus, the return to education increases by more in HC ( gure 7a). If X production is less intensive in F, the demand shock for skilled labor in LC is much bigger than in HC, increasing the return to education by more in LC ( gure 7b). I test two main predictions of this model, assuming districts in India are in case B (since there is evidence of a return to speaking English 20 ). First, the district with a lower English learning cost should produce more X and second, the low cost district will have more educational attainment. Finally, I provide suggestive evidence that the average return to education rises by more in higher cost districts. 20 See Munshi and Rosenzweig (2006). 2

23 6 Impact of linguistic distance from Hindi 6. Information technology Estimating equation (2) indicates a strong positive e ect of linguistic distance from Hindi on IT presence (table 9). In Panel A, I drop the ten most populous cities (as of 99) since IT rms are likely to locate there regardless of English speaking manpower; in Panel B, I include these cities and an interaction with linguistic distance. The cost of learning English, when measured as the weighted average or percent distant speakers, predicts whether any IT rm establishes a headquarters or branch in a district (columns -2). An increase in one degree from Hindi of the average speaker s mother tongue (three-fourths of a standard deviation) results in a 4% increase in the probability of any IT presence (the dependent variable mean is 2%); a 20% increase in the percent distant speakers (half a standard deviation) increases the probability by 6%. The magnitude of these e ects is economically signi cant: about a fourth of the e ect of housing one of 26 elite engineering colleges. Columns 3-4 show that IT headquarters were established earlier in areas linguistically further from Hindi, by approximately one year per 20% distant speakers. 2 Linguistic distance from Hindi also explains geographic variation in the log number of headquarters and branches and log employment when divided evenly by branch (columns 5-8), though these coe cients are more likely signi cant when using the percent distant speakers. Twenty percent more distant speakers increases the number of establishments by 0% and employment by almost 50%. A back-of-the-envelope calculation suggests that having 20% more distant speakers would increase the percent of English speakers by about 6% (from % to.7%). Thus, having 3000 more English speakers would attract 0.05 more IT establishments (at the mean of 0.5) and 36 more IT employees (at the mean of 72). Given 2 Many districts may be very unlikely to receive IT rms for a number of reasons. While these reasons should be orthogonal to linguistic distance, I con rm this result by using rm-level data to focus on districts with any IT rms between 995 and 2003 (results available, not shown). I also use the time variation to study whether rms locate in cities linguistically further from Hindi earlier or whether they branch out to smaller cities that are further from Hindi later as the large cities get more crowded. The results suggest that the second is true, but there is insu cient data and variation in time. 22