Data and Country Level Measure of Distance

For a cross-country estimation of the relationship of linguistic distance to economic outcomes, we need an algorithm to determine distance between any two languages and a measurement strategy to calculate average distance for any population of its language to that of the official language. In order to conceptualize the notion of distances between languages, the measure based on Ethnologue’s linguistic tree diagrams is used. The distance between any two languages i and j based on Fearon (Reference Fearon2003) is defined as

From Equation (1) we see that if two languages belong to different language families, i.e., the number of common nodes between them is 0, their distance is equal to 1, which by construction is the maximum distance between any two languages. The value of λ determines the relative distance between two languages which belong to the same family compared to two languages that belong to different families. For instance, consider Spanish and Catalan belonging to the Indo-European language family and having seven branches in common.Footnote ⁶ Choosing a value of λ equal to 0.5 would imply the distance between Spanish and Catalan is equal to 0.116. Choosing a lower λ, such as 0.05, would give greater weight to the similarity in the earlier nodes, and the distance between Spanish and Catalan would fall to 0.012. Of course, if two languages differ at the first node, as would be the case for Spanish and Tamil, whatever the value of λ the distance score would remain at 1. As no theoretical basis has been established for choosing the correct value of λ, following Fearon (2003), we fix the value of λ equal to 0.5 in our analysis.Footnote ⁷

We can now calculate a weighted measure of average distance of a country’s population from the official language. The official language(s) of the countries included in the regression on Tables 3 and 4 are shown in the Excel file accompanying the Online Appendix. The data on the number and size of linguistic groups in the country comes from the data of Fearon (Reference Fearon2003), which take into account all linguistic groups that form at least 1% of the population share.Footnote ⁸

The average distance from the official language (ADOL) for any country i is calculated as

where n are the number of linguistic groups in the country, P_ij refers to the population share of group j in country i, and d_jo refers to the distance of group j from the official language. The coding rules when there is more than one official language depend on whether there is a group associated with an official language in which social and political mobility is possible for monolinguals of that language (e.g., Germans in Switzerland, Afrikaners in South Africa) or whether the group associated with that official language must have proficiency in another official language for full mobility prospects (e.g., Urdu speakers in Pakistan). For the former, language distance equals zero. In the case of the latter, language distance equals one-half the distance between their indigenous language and the less prestigious official language plus one-half the distance between their language and the more prestigious official language.Footnote ⁹

The constructed measure of ADOL is distinct from indices of linguistic diversity used in the literature (Alesina et al. Reference Alesina, Devleeschauwer, Easterly, Kurlat and Wacziarg2003; Desmet, Weber, and Ortuño-Ortín Reference Desmet, Weber and Ortuño-Ortín2009; Greenberg Reference Greenberg1956); while measures of linguistic diversity are concerned with the level of linguistic heterogeneity within a country, our index measures how distant the official language of a country is from the languages spoken within a country. As the choice of official language is not restricted to a set of indigenous languages, countries that are classified as having low levels of linguistic diversity nonetheless maybe linguistically distant from the official language. To see this, consider countries such as Angola, Burundi, Lesotho, Rwanda, Swaziland, and Zambia; all have a value of linguistic diversity as measured by the Greenberg index of less than 0.005, however their average distance from the official language is at least 0.50, as all use a nonindigenous imperial language as their official one.Footnote ¹⁰

The measure of ADOL is closest in spirit to the peripheral index proposed by Desmet, Ortuño Ortín, and Weber (Reference Desmet, Ortuño Ortín and Weber2005). Their index measures the distance of all peripheral groups to the dominant central group, which in their application referred to the largest linguistic group in the country. Here we extend their application where the official language is the language of the largest linguistic group in the country. In our application, ADOL differs from the original application of the peripheral index when the official language is not the language of the largest ethnic group, such as Amharic in Ethiopia, or when a country has adopted a nonindigenous language to act as their official language, as is the case in most postcolonial states in sub-Saharan Africa and South Asia.Footnote ¹¹

Figure 1 shows a color-coded map of the world depicting the average distance from the official language for the sample of countries included in our study. For illustrative purposes, Table 1 also provides the average language distance scores for a selected set of ethnic groups and countries.Footnote ¹²

FIGURE 1.

World Distribution of Average Distance from Official Language

Note: The white colored areas refer to countries on which information on language distance is not available.

TABLE 1.

Distance from Official Language for the Three Largest Ethnic Groups for Selected Countries

Note: According to the coding rules, Afrikaans speakers are treated as if Afrikaans were the only official language; hence their value for distance is zero.

Table 2 in turn shows descriptive statistics for a range for interesting socioeconomic variables for the entire sample, as well as by quartiles of language distance. Strikingly, all variables considered are seen to be monotonic with respect to ADOL.

TABLE 2.

Socioeconomic Outcomes by Quartiles of Language Distance

Note: In the parentheses are provided the number of observations followed by the standard deviation.

Why does the Distance from the Official Language Matter?

Outlining a clear theoretical mechanism is essential in order to understand through which channels choice of official language affects socioeconomic development. The framework will subsequently guide us in our empirical exercise. It will also enable a theoretically founded interpretation of the results, as sketched below with a formal exposition provided in the Online Appendix. The two main facets of socioeconomic development that our theory links to official language choice are human capital formation and health.

Individuals in our framework are assumed to be utility maximizers and choose the level of human capital and preventive health behavior to maximize their wellbeing. The cost of human capital formation for any individual i is assumed to be a function of their ability, the distance of individual i from the official language of the country, and to the amount of exposure of individual i to the official language.

In our theory, the first assumption is the greater the distance of individual i to the official language, the higher the cost of obtaining human capital and participating in the economy. This first assumption implies that all else equal, a native French speaker would face a lower cost of learning Italian than a native German speaker, as Italian is structurally closer to French than German, and hence obtaining higher human capital. The second assumption states that the greater the exposure to the official language, the lower the costs of obtaining human capital and participation in the economy. The second assumption in turn implies, all else equal, Akan speakers from Ghana would face lower learning and participation costs and obtain higher human capital due to the use of English as the official language in the United States as compared to in Ghana, as their level of exposure to English would be much higher in the United States.Footnote ¹³

The health behavior of individuals is assumed to be affected through two distinct channels. The first one, directly linked to official language choice, is through language acting as a barrier to health care access or to the comprehension of pertinent health information (Bowen Reference Bowen2001; Djité Reference Djité2008, Chapter 3; Higgins and Norton Reference Higgins and Norton2009; Underwood, Serlemitsos, and Macwangi Reference Underwood, Serlemitsos and Macwangi2007).Footnote ¹⁴ Evidence supplied by Translators Without Borders (TWB) provides pertinent information. In a study conducted in Kenya after the outbreak of the Ebola crisis, they were able to gauge the importance of language in comprehending health related information. Randomly sampled rural and urban respondents were first tested on their knowledge regarding channels of Ebola transmission. Before treatment, the average share of correct answers was as low as 8 percent. Subsequently, a random half of the study population was provided information regarding Ebola transmission using posters in Swahili, and the other half using posters in the official language of English. Post-treatment the Swahili-treatment group had on average 92 percent of the answers correct whereas the English-treatment group had only 16 percent of the answers correct.Footnote ¹⁵

To be sure, governments can use multilingual messaging to transmit health education to all their citizens. However, it is costly to transmit technical information in languages that have not been standardized, as would happen if those languages had official status. Indeed, much of the primary health care information, especially in Africa, is only available in the official language.Footnote ¹⁶ In any event, our claim about the favoring of official language in the health sector is not that it is an absolute impediment to available health care, but that on average, the probability of successful access and treatment is reduced.Footnote ¹⁷

The second channel through which language policy affects health behavior is indirect and works through the conduit of human capital. The reasoning being that education matters for the ability of individuals to be able to process and use information regarding best health practices (refer to Dupas (Reference Dupas2011, 435–6) and the citations contained therein for an overview on the complementarities between education and health behavior; also refer to De Walque (Reference De Walque2007 Reference De Walque2009) on the relationship between education, HIV, and preventive sexual behavior in sub-Saharan Africa, and De Walque (Reference De Walque2010) on the relationship between education and smoking behavior).

It is important to note that our measure of ADOL subsumes both the theoretical concepts of distance and exposure to the official language. Referring to the formal exposition in the Online Appendix, the notion of distance from official language is self-evident from Equation (2); for the case of exposure, in the cross-country analysis we attribute the distance of other ethnic groups (i ≠ j) in the country to be a measure of exposure of the ethnic group i to the official language. The intuition here is that the greater the average linguistic distance of all other groups to the official language, the less likely the other groups j ≠ i will be relying on the official language in their everyday lives, and therefore the lower the exposure of members of group i to the official language. As the measure takes into account the distance of all ethnic groups, the concepts of group distance and exposure are both captured by the same measure.

Recent research strongly suggests the independent and joint influence of both measures for health outcomes. A series of recent papers (Aoki and Santiago Reference Aoki and Santiago2015; Bleakley and Chin Reference Bleakley and Chin2004 Reference Bleakley and Chin2008 Reference Bleakley and Chin2010; Clarke and Isphording Reference Clarke and Isphording2015) examines the link between language proficiency of immigrants in the destination country language, i.e., English and education, labor market, and of particular interest to us here, health outcomes in the United States, United Kingdom, and Australia. They all follow the methodology suggested by Bleakley and Chin (Reference Bleakley and Chin2004), which uses the age-at-arrival (a measure of exposure to English) interacted with a non-English-speaking origin country dummy (a measure of distance) to predict English proficiency. The reasoning is based on the “critical period hypothesis” of language acquisition (Lenneberg, Chomsky, and Marx Reference Lenneberg, Chomsky and Marx1967). In other words, young children find it easier than older children or adults to learn languages.

Bleakley and Chin (Reference Bleakley and Chin2010) data allow us to show that the joint effect of exposure and distance to the official language, English, leads to higher rates of disabilities later in life. Our reanalysis of this data (available from authors on request) specifies a dummy outcome variable based on whether an individual reports that he or she suffers from a disability that hampers work. The analysis shows that the coefficient on English proficiency (jointly determined due to exposure and distance to English) reduces the probability that an individual reports such a disability by 6.5 percentage points.Footnote ¹⁸ Clarke and Isphording (Reference Clarke and Isphording2015), relying on the same identification strategy (but here using a continuous variable for language distance) with data on immigrants to Australia, also find large effects of our two channels (high exposure and low distance) on physical health.

The main result from these studies is that immigrants from non-English-speaking countries (i.e., with high language distance) who arrive after the age of nine (signifying late exposure to English) have significantly worse health outcomes in later life. These data provide strong empirical support for our theoretical claim on both the individual and the joint effects on health outcomes of linguistic exposure and distance.Footnote ¹⁹

The Choice of Proxies for our Dependent Variable

The discussion in the previous section assumes that the choice of official language influences the level of human capital in society by affecting the cost of acquisition. A measure of human capital is thus a natural outcome variable to explore. Since this cannot be measured directly, we need reasonable proxies, and for this we need to address two issues. First, available measures of human capital, such as years of schooling or enrollment rates, mostly capture quantity and not quality, which obscures the variation in the levels of learning that students at the same grade level exhibit across countries. The problem becomes especially pronounced as enrollment levels and years of schooling have sharply risen in developing countries over the past decades, but learning outcomes have either stagnated or even worsened. For instance, in some countries in sub-Saharan Africa, up to 40 percent of young people who have attended primary school for five years have neither the essential skills to avoid lapsing into illiteracy, nor the minimal qualifications to secure a job (UNECOSOC 2011). Similarly the latest available round of Demographic and Health Survey (DHS) data from 35 sub-Saharan African countries shows that 33 percent of the males recorded as having between four and seven years of schooling are still unable to read a complete sentence. This implies that available quantitative measures of human capital might be a poor indicator of actual stock of knowledge, especially for developing countries.

A second issue relates to the time it takes to translate values on average distance to observable changes in levels of human capital. If language choices for postcolonial states were made post–World War II, it might take two generations for the effects of this choice to affect standard outcome variables in a significant way such as output per worker.

Our proposed solution to these problems is to use four distinct measures, each with different advantages, and to show that our results are robust across these different measures, allowing us to combine them for general analysis. For our most direct measure, one that captures the actual level of knowledge (or human capital), we rely on test scores from comparable student achievement tests across countries.Footnote ²⁰ Using such a measure however comes at a potential cost. These internationally comparable test scores are available only for 70 countries, and these include only six from sub-Saharan Africa.

As an indirect measure of human capital, here working through the channel of health, we measure life expectancy. As just discussed, we assume that populations with high rates of human capital, controlling for country wealth, are better able to take advantage of modern health resources and communicate successfully with medical staff, thereby improving diagnoses and implementation of remedies. These differences in knowledge (based on test scores) and life expectancy ultimately (albeit slowly) translate into differences in levels of wealth and productivity, as captured by GDP per capita and output per worker, our third and fourth proxies, and both (rather noisy) economic variables that should also be affected by average language distance. Indeed, the down side of using a purely income based measure such as GDP per capita is that it fails to account for the fact that certain countries that are rich in natural resources concentrate income in the hands of a few individuals. Consequently, for such countries, GDP per capita is a poor indicator of the true state of development for the majority of the population. Figure 2 shows a strong negative relation between ADOL and the four dependent variable of interests.

FIGURE 2.

Scatter plot of ADOL and the Four Socio-Economic Variables of Interest

As noted, none of these four proxies is perfect. Given this lack of a perfect composite measure of socioeconomic development we undertake the approach of first presenting our basic regressions with each of the above four dependent variables: a measure of cognitive skills, life expectancy, log GDP per capita and log output per worker. After presenting our initial results in support of our thematic framework, we then adopt the strategy of using the standardized score on the Human Development Index (zHDI) as our preferred dependent variable. This index includes health, education, and wealth measures, and is strongly correlated with the four component measures.Footnote ²¹ The rationale of using zHDI as the dependent variable, for robustness exercises and further empirical analysis, is based on the fact that not only does it capture all four dimensions outlined by our theory, albeit imperfectly, but also avoids losing valuable observations.

Cross Country Regressions

In order to explore the correlation between the dependent variables of interest and ADOL, we estimate a reduced form regression that takes the form

where in all specifications we estimate robust standard errors. The results are shown in Table 3, where the DV_i in columns (1) and (2) is a measure of cognitive skills taken from the work of Hanushek and Woessmann (Reference Hanushek and Woessmann2012). Columns (3) and (4) in turn use life expectancy in 2010 as the dependent variable to explore the effect of ADOL on health. Columns (5) and (6) consider log GDP per capita in 2005 in 2005 constant dollars and finally columns (7) and (8) use log output per worker from the work of Hall and Jones (Reference Hall and Jones1999) as a measure of productivity.

TABLE 3.

Regressions of Distance on Cognitive Scores, Life Expectancy, Log GDP per Capita, and Log Output per Worker

Notes: **p < 0.10; **p < 0.05; ***p < 0.01. Robust SE’s in parentheses and standardized coefficients in square brackets.

X_i refers to a vector of controls and in all eight specifications shown in Table 3; besides our measure of ADOL, we control for three additional confounding factors. First, we control for ethnolinguistic fractionalization (ELF), a measure that takes into account linguistic distance between all ethnic group dyads, and based on Fearon (Reference Fearon2003). The concepts of ELF and ADOL, as explained in the section on data and country level measures of distance, are distinct; however, empirically the correlation between the two measures is 0.57 and thus it is important to account for it in a multivariate framework. The choice of the measure of ELF is inspired by the work of Desmet, Weber, and Ortuño-Ortín (Reference Desmet, Weber and Ortuño-Ortín2009) who show that accounting for distance between groups in diversity measures is important, though once distance is accounted for the choice between the exact nature of the index used—diversity, peripheral heterogeneity, or polarization—is empirically irrelevant.Footnote ²²

Second, we include a measure of institutional quality from the Polity-IV data set, quantifying the extent of institutionalized constraints on the decision-making power of chief executives averaged over the years 1960 to 2000.Footnote ²³ As we are interested in understanding the effects of language policy choices on socioeconomic development, the third control we include is the level of log GDP per capita in the year of independence, i.e., before official language choices were instituted and hence account for the previous level of development which were largely unrelated to contemporary language policy choices.Footnote ²⁴

Columns (2), (4), (6), and (8) additionally include continent dummies. The inclusion of continent dummies implies that the coefficient on ADOL is being estimated based on the difference in language distances between countries within a continent, and the dependent variable of interest. On the one hand, the inclusion of continent dummies ensures that the effect we are capturing is not being driven by the black box of across continental differences. On the other hand, if our objective is to explain what makes countries in any continent distinct, the inclusion of continent fixed effects by definition will imply that these differences, if they are correlated with the independent variable of interest, are relegated to the black box of fixed effects. As we later contend (in the section on the instrumental variable approach) that geography is a key factor affecting language policy choices, they are consequently correlated with continents. For this reason, the inclusion of continent dummies absorbs a large part of the effect of language distance, though there remains much variance to be explained.

For the dependent variable life expectancy we additionally control for the percentage of people ages 15–49 who are infected with HIV, to ensure that our estimates are not only capturing differences in HIV prevalence rates. For log GDP per capita we control for the availability of natural resources, namely percent of world oil, gold, iron, and zinc reserves, and number of minerals present in a country.Footnote ²⁵

In all eight specifications ADOL is seen to be both substantively and statistically an important correlate of the four dependent variables. To have an intuitive understanding of the magnitude of the effect imagine a country such as Ghana switching from using English to Akan, the language of the largest ethnic cluster, as their official language. This reduces the ADOL from 1 to 0.18, and moves Ghana up 13 and 11 spots in terms of their ranking on cognitive tests scores and life expectancy, respectively, and 21 ranks up in the case of log output per worker.

Table 4 in turn considers the standardized value of the HDI in 2010, a composite measure of the facets of socioeconomic development outlined by our theory, as the dependent variable. ADOL by itself explains around 55% of the cross-country variation in the HDI, and together with all controls 76% of the cross-country variation in the levels of HDI are accounted for in the regression. The largest drop in the coefficient occurs between columns (4) and (5) when we include continent dummies.

TABLE 4.

Regressions of Distance on zHDI

Notes: *p < 0.10; **p < 0.05; ***p < 0.01. Robust SE’s in parentheses and standardized coefficients in square brackets.

Finally, Table 5 shows that the correlation documented between ADOL and HDI in Table 4 cannot be attributed to any particular region of the world. Columns (2)–(6) in Table 5 drop Africa, the Americas, Asia, Europe, and Oceania, respectively, and the coefficient on average distance remains, both substantively and statistically, an important correlate of HDI.

TABLE 5.

Regressions of Distance on zHDI Excluding One Continent at a Time

Notes: Column (1) considers the entire sample; columns (2)–(6) drop Africa, Americas, Asia, Europe, and Oceania, respectively. *p < 0.10; **p < 0.05; ***p < 0.01. Robust SE’s in parentheses and standardized coefficients in square brackets.

Theoretically Inspired Controls and Some Robustness Checks

We now explore other potentially important factors that have been highlighted in the literature as important in explaining cross-country income differences to evaluate the robustness of our results.

Taking into account new insights on deep historical sources of economic performance (Ashraf and Galor Reference Ashraf and Galor2013; Bockstette, Chanda, and Putterman Reference Bockstette, Chanda and Putterman2002; Michalopoulos and Papaioannou Reference Michalopoulos and Papaioannou2013; Nunn Reference Nunn2009), we add a measure of genetic diversity, genetic diversity squared, and the index of state antiquity to the specification given by column (5) of Table 4. The results are shown in column (2) of Table 6. The addition of these controls does not affect the precision or magnitude of the coefficient on average distance.

TABLE 6.

Robustness Tests of Regressions of Distance on Standardized Value of HDI

Notes: *p < 0.10; **p < 0.05; ***p < 0.01. Robust SE’s in parentheses and standardized coefficients in square brackets.

The historical origin of a country’s laws has been shown to be correlated to a broad range of economic outcomes (Shleifer, Lopez-de Silanes, and La Porta Reference Shleifer, Lopez-de Silanes and La Porta2008). In column (3) of Table 6 we additionally control for the legal origin of the countries. As can be seen this control does not affect the precision or magnitude of our estimates.

The data on GDP at independence is measured in a common denominator, i.e., at constant 1990 prices, for all countries in our sample. However, given the date of independence between countries varies widely, the same income levels in different eras might imply a different stage of development. Alternatively, the timing of independence itself may contain information on a country’s wealth. In order to address this concern of comparability across eras, we consider only the sample of countries that gained independence after 1945 and re-estimate Equation (3) for all five dependent variables of interest. The results in Table 7 show that ADOL is still statistically significant and an economically meaningful predictor of the socioeconomic variables considered.Footnote ²⁶

TABLE 7.

Regressions of Distance on Cognitive Scores, Life Expectancy, Log GDP per Capita, Log Output per Worker, and zHDI in 2010—Sample of Countries Independent Post-1945

Notes: The dependent variables in columns (1)–(5) are cognitive scores, life expectancy in 2010, log GDP per capita in 2005, log output per worker from the work, and zHDI in 2010, respectively. *p < 0.10; **p < 0.05; ***p < 0.01. Robust SE’s in parentheses and standardized coefficients in square brackets.

We need also to ask how robust our findings are to contemporary changes in the international political economy, from an era of import substitution growth models (where there may have been an advantage to the promotion of indigenous languages) to an era of globalization (where the premium on English would be revealed) (Rodrik Reference Rodrik1990). Perhaps our results supporting the role of languages that are proximate to that of the local populations were appropriate for the first era, but not for the second? We examine this possibility in column (3) of Table 8, by replacing GDP at independence with zHDI in 1990, and find that the effect remains significant both statistically and substantively in the 1990–2010 period. Globalization, in other words, has not lessened the importance of average distance for human development.

TABLE 8.

Regressions of Distance on Standardized Value of HDI in 1990 and 2010

Notes: In column (1) and (3) the dependent variable is zHDI in 2010; in column (2) it is zHDI in 1990. *p < 0.10; **p < 0.05; ***p < 0.01. Robust SE’s in parentheses and standardized coefficients in square brackets.

In the Online Appendix we conduct a series of robustness tests and show the correlation is robust to additional controls for geography, climate, and alternative measures of ELF and institutions.

An Instrumental Variable Approach

To provide evidence that the documented relationship between ADOL and socioeconomic development is indeed causal, we now undertake a strategy of using an instrument that is correlated with ADOL but uncorrelated with other country characteristics.

We identify the availability of a written tradition as one of the important factors affecting language policy choices. The rationale is that in the absence of a written language states first need to invest in creating a standardized orthography, vocabulary, and modern scientific terminology before a language can be utilized to fulfill the functions of an official language. Thus many states in the face of uncertainty associated with the cost and returns involved in the creation of written language might resort to using the colonial language. The proposed relationship finds strong support when we observationally examine availability of written traditions and choice of official language. Looking across the globe, nearly every country that had a writing script for an indigenous language has adopted at least one indigenous language as at least co-official. This factor can explain the language policy choices observed in sub-Saharan African. Most sub-Saharan African countries (with Ethiopia, Tanzania, and Liberia as exceptions) did not possess a writing tradition and are characterized by the usage of only the colonial language as the official language. To empirically test whether availability of a writing tradition has any explanatory power, we regress our measure of distance from official language on a dummy for having a writing tradition.Footnote ³⁶ The results are shown in Table 9.

TABLE 9.

Factors Affecting Average Distance from Official Language

Notes: *p < 0.10; **p < 0.05; ***p < 0.01. Robust SE’s in parentheses and standardized coefficients in square brackets.

The availability of a writing tradition is seen to be a statistically significant predictor of ADOL. In columns (2) and (3) we control for log GDP per capita at independence and log population in 1500 (as a proxy for levels of development in the Middle Ages), respectively. The two wealth related factors are not only seen to be statistically insignificant but also their explanatory power is seen to be less by a factor of 40–60 as compared to the hypothesized factor.

The regressions shown in Table 9 thus provide support to the assertion that possessing a writing tradition is an important determinant of ADOL. However, the indicator variable cannot be used as an instrument, as states which had a writing tradition, as compared to those which did not, arguably also differ on other important unobservable characteristics which might affect socioeconomic development.

Drawing from the work of Diamond (Reference Diamond1997), we hence propose using distance from the sites at which writing was independently invented as an instrument for ADOL.Footnote ³⁷ He argues that geography was a crucial factor as to why a set of polities—Tonga’s maritime protoempire, the Hawaiian state emerging in the late 18th century, all of the states and chiefdoms of subequatorial Africa and sub-Saharan West Africa, and the largest native North American societies, those of the Mississippi Valley and its tributaries—did not acquire writing before the expansion of Islam and the arrival of the Europeans.

Writing was independently invented in Mesopotamia (Sumer) around 3200 BCE, in China around 1200 BCE, and in Mesoamerica around 600 BCE, and then diffused through trade and exchange to the rest of the world. The rationale for using the distance from the site of invention as an instrument is that the further the distance from the site of invention, the less likely is a country to have obtained the writing tradition through the process of diffusion, and consequently based on the evidence in Table 9 will have a higher ADOL. Observe that using the distance from the site of invention as an instrument exploits the exogenous component of the probability of having a writing tradition, i.e., geography. The key underlying assumption for it to be a valid instrument is that the distance from these sites of invention should have no independent impact on socioeconomic development today, except through the channel of affecting the probability of possessing a writing tradition.

To operationalize the measure we calculate the great-circle distance, using the Haversine formula, from each of the sites of invention to every other country in our sample. We then take the minimum of the distance from the three sites as the measure of distance from the place of invention of writing. Figure 3 shows the relationship between the shortest distance from the sites where writing was invented and the ADOL; as hypothesized the distance from official language is seen to be increasing in the distance from where writing was invented. The IV estimates for the five dependent variables of interest are shown in Table 10.

FIGURE 3.

Reduced Form Relationship Between ADOL and Distance from Site of Invention of Writing

TABLE 10.

IV Regressions of Distance on Cognitive Scores, Life Expectancy, Log GDP per Capita, Log Output per Worker, and zHDI in 2010

Notes: *p < 0.10; **p < 0.05; ***p < 0.01. Robust SE’s in parentheses and standardized coefficients in square brackets.

Columns (1), (3), (5), (7), and (9) regress cognitive test scores, life expectancy, log GDP per capita, log output per worker, and zHDI, respectively, on ADOL instrumented for by the minimum distance from the sites of invention of writing. In panel (B) the first stage regressions of distance from the sites of invention of writing on ADOL are shown. Inspecting the F statistics shows that all, except in column (1), meet or exceed the value of 10, and in most cases are greater than 30, suggesting that distance from the site of invention is a strong instrument for ADOL.Footnote ³⁸ In panel A are the results of the second stage; we see that the predicted value of ADOL is a statistically significant and economically important predictor of all the socioeconomic variables. The point estimates slightly exceed the OLS estimates in Tables 3 and 4.

In columns (2), (4), (6), (8), and (10) we additionally add the three controls outlined before in the section on cross country regressions—linguistic diversity accounting for distance, constraints on the executive, and log GDP per capita at independence. We additionally control for an America dummy and the proportion of population of European descent in 1975. The reason is that the majority of the population on the American continent can be classified as either settlers or individuals of mixed race heritage (also known as “mestizos”), whose mother tongue is a language which the settlers brought along with them. Thus for these countries distance from the site of invention of writing is not an important determinant of ADOL. Again the ADOL is seen to be a statistically significant predictor of the levels of socioeconomic development. Results remain stable adding genetic diversity, genetic diversity squared, and latitude to these estimations, as shown in Table A.12 of the online Appendix.

A potential concern with the estimates in Table 10 is that the distance from the sites of invention of writing could be correlated with other factors affecting socioeconomic development. If for instance we were to assume that distance from these earliest sites of invention of writing was responsible not only for acquiring the writing tradition but also a determinant of quality of state institutions and/or governance, then we would be violating the exclusion restriction for our instrument to be valid. In order to assess whether this is a cause for concern we run reduced form regressions of the minimum of the distance from the sites of invention of writing on the three most widely used measures of state institutional capacity and governance: (i) average protection against expropriation risk from the Political Risk Services (PRS) group averaged over the years 1995–05; (ii) social infrastructure combining government anti-diversion policies and openness to international trade from the work of Hall and Jones; and (iii) constraints on the executive from Polity-IV and averaged over the years 1960–2000. The results are shown in Table 11.

TABLE 11.

IV Falsification Test—Regressions of Distance from Sites of Invention of Writing on Development Outcomes

Notes: *p < 0.10; **p < 0.05; ***p < 0.01. Robust SE’s in parentheses and standardized coefficients in square brackets.

The distance from the sites of inventions of writing is not a significant correlate of any of the three measures of state institutions or governance, with the F statistic taking a value of less than 1 in all three regressions. Thus the IV results confirm the negative relationship between ADOL and socioeconomic development estimated by the OLS, and suggest that the OLS estimates may be a lower bound of the true effect of ADOL.

Finally, to gauge the economic magnitude of the IV estimates, again consider Ghana adopting Akan, the language of its largest ethnic cluster, as its official language instead of English. Such a change would move Ghana 23, 24, and 31 positions up in the ranking of countries on cognitive test scores, life expectancy, and log output per worker. Alternatively it would move Ghana from the $7{\text{th}}$ , $22{\text{nd}}$ and $21{\text{st}}$ percentile of the distribution of cognitive test scores, life expectancy, and log output per worker to the $38{\text{th}}$ , $40{\text{th,}}$ and $47{\text{th}}$ percentile, respectively.

Data

The data come from the Indian National Family Health survey (NFHS 3) of the year 2005–06. We consider the sample of males and females aged 15–54 years to estimate the effect of individual language distance on various socioeconomic outcomes of interest. The data provide information on the native language of the respondent, typically a proxy for the language of one’s ethnic group even if the respondent has only limited facility in it, and state of residence, which allows us to calculate the language distance for individuals from the official state language. The data set also provides information on relevant individual characteristics such as age, religion, caste, educational attainment, a wealth index, employment status, nature of occupation, as well as knowledge and adoption of health practices.

Identification Strategy

We estimate the effect of the distance of an individual’s native language from the official state language on six variables. The first two are proxies for human capital: (i) years of education and (ii) a dummy variable for whether the individual is literate. The next two measure health knowledge and practices: (iii) an indicator variable for whether the individual has ever heard of AIDS; (iv) whether the household uses a mosquito bed net for sleeping.Footnote ⁴⁰ The final two measure occupation and wealth outcomes: (v) whether the individual holds a white-collar jobFootnote ⁴¹ ; (vi) an indicator for whether the individual falls in the top quintile of the income distribution.

Comparing across Indian states indicates large variations in their levels of socioeconomic development, which are important to account for in any empirical exercise. Accordingly in all our specifications we account for state fixed effects.Footnote ⁴²

A naive comparison of language distance and socioeconomic outcomes based on native speakers (nonmigrants) vs non-native speakers (migrants) resident in the same state fails to account for the fact that natives and migrants might differ along unobservable dimensions which are not accounted for and which might be correlated with language distance.Footnote ⁴³ In order to address this concern we restrict ourselves to the sample of individuals who report as having always lived in the same state, or in other words we exclude any first generation migrants. For nonmajority language speakers, our data include both members of rooted minority groups (who by rights in the Indian Constitution can receive primary education in their mother tongues) and individuals whose families were migrants in recent generations (who do not have concentrations of their population that would make them eligible for indigenous language instruction in public education). To the extent that rooted minorities are getting the indigenous language instruction that the Constitution affords them, our results seeking to estimate the effect of not receiving mother-tongue education would be an underestimate. On the Constitutional formula for minority language instruction, see Sridhar (Reference Sridhar1996).

As we observe individuals belonging to the same linguistic groups in states having different official languages, we are able to account for any linguistic-specific group differences through the inclusion of language group fixed effects. In sum, our identification strategy ensures that the estimated effect of language distance is not due to any time invariant state or linguistic group characteristics.

Data

The SACMEQ II round collected data on around 40,000 students, 5,300 teachers, and 2,000 school heads from 2000 primary schools.Footnote ⁴⁴ The dataset provides information on standardized student achievement tests in reading and mathematics across the 13 countries for pupils currently in the sixth grade.Footnote ⁴⁵ The scores are standardized with a mean of 500 and standard deviation of 100. Moreover the standardized scores are provided for essential reading and math tests as well as for a comprehensive math and reading test. The data also provide a categorical indicator which captures whether students meet the minimum and desirable reading levels of SACMEQ. These are the main pupil related outcomes which form the dependent variables of interest. The dataset also provides extensive information on the students’ socioeconomic background such as parents’ education, possessions, housing quality, availability of extra lessons outside the classroom (often referred to as tuitions), support at home for homework, and school absences. It also asks a question regarding usage of the medium of instruction, English, at home, which is divided into the category of never, sometimes, and often. The dataset also collects information regarding teachers, headmasters, schooling infrastructure, and quality. It also allows us to identify the classroom to which each student belongs. Control variables and descriptive statistics are provided in Table 13.

TABLE 13.

Educational Data for East and Southern Africa: Selected Descriptive Statistics

The descriptive statistics convey the gravity of the problem facing the educational sector in Africa. About 60% of the students do not reach the minimum reading level. When the bar is fixed at the desirable reading level, about 86% of the students are classified as not reaching that level, and this in spite of vast foreign aid expenditures over the previous decade directed at the educational sector (Devarajan and Fengler Reference Devarajan and Fengler2013). Obviously, fundamental factors affecting student achievement have not yet been addressed, which directs attention to the exposure channel.

Identification Strategy

To test for exposure, the key independent variable of interest is the frequency with which pupils use English at home. Regarding the usage of English at home, 23% report as never using English at home, 55% report using English sometimes at home, and 21% report using English often at home. We construct a binary indicator which takes the value 0 in case the student never uses English at home and the value 1 if the students use English often or sometimes at home.Footnote ⁴⁶ As all students are Africans, in the data their distance to the official language, English, is equidistant and equal to 1.Footnote ⁴⁷ This means there is no need to control for the effect of individual level distance from the official language. The choice of our independent variable, use of English at home, is inspired by the work of Dustmann, Frattini, and Lanzara (Reference Dustmann, Frattini and Lanzara2012) who show that the single most important factor in explaining differences between immigrant and native children PISA tests scores in OECD countries is the language spoken at home.

Recall that more than 70–80% of the population in most African countries do not speak the official language and this is especially true for the older generations. It is therefore not surprising that the variable “using English at home” captures a rather small increment in academic success. In the data around 70% of the pupils who do not reach the minimum reading level still claim to use English at home. Given the low level of skills the pupils themselves possess it can be inferred that the exposure to English that takes place even at home is not comparable in quantity or quality in any way to the exposure that language minority students in advanced industrial countries, for instance as immigrants, experience while learning in a majority language. Thus the reported levels of high usage might still be very low in quality and quantity when compared to conventional exposure to the medium of instruction in countries where it is spoken by the local population as a native language. That said, given that our measure of exposure captures low quantity and quality of exposure, if it still turns out to be a significant explanatory factor of student performance, this would imply that the estimate should be considered to be the very lower bound of the effects of exposure.

The data identify the classroom to which each student belongs. We have information on 33,141 students in 4,686 classes across the 11 countries.Footnote ⁴⁸ We are hence able to account for classroom fixed effects in our analysis. Taking classroom fixed effects implies that common factors—such as teachers, school infrastructure, and other unobservables—that affect student performance at the classroom level are accounted for. We can now estimate the effect of using English at home, which is our proxy for exposure to the medium of instruction, on test scores with class fixed effects and controls at the level of the student’s home.