2. AN ANTHROPOMETRIC MEASURE OF EDUCATIONAL INEQUALITY

Following Baten and Mumme (Reference Baten and Mumme2010), this paper measures educational inequality by the difference in mean years of schooling between the taller half and the shorter half of the female population. Height has been increasingly used by economic historians as a proxy for living standards in situations in which direct data (e.g. wage) are lacking (Komlos and Baten Reference Komlos and Baten2004). There is robust evidence in anthropometric research that people of upper socio-economic strata tend to be taller than those of lower strata (Steckel Reference Steckel1995). Therefore, in the absence of information on social classification (e.g. income and occupation), taller people can represent the better-off socio-economic stratum, and shorter people can represent the worse-off socio-economic stratum (López-Alonso and Vélez-Grajales Reference López-Alonso, Vélez-Grajales, Bértola and Williamson2017). In this regard, the difference in mean years of schooling between the taller half and the shorter half is a proxy for educational inequality between the better-off stratum and the worse-off stratum, namely the socio-economic inequality in education. This method falls within the «gradient» or «socio-economic» approach, which measures inequality in social indicators by differences between groups with different socio-economic characteristics (Sahn and Younger Reference Sahn and Younger2006). Height can be seen as a socio-economic characteristic, because at the population level it is mostly determined by living standards. The data in this paper show that biologically better-off, namely taller, people tend to be educationally better-off, as 281 out of the total 284 observations of taller–shorter differences in years of schooling are positive in favour of taller individualsFootnote ⁷. The three negative exceptions are all close to zero.

The height-based approach to measuring social classification is particularly relevant for Latin America, where pre-1980s data are quite lacking, especially at a sub-country level. Since adult height is mostly determined by nutrition and environment during early childhood, in particular the first 3 years following birth, it is a good proxy for childhood living standards (Baten Reference Baten2000; Cole Reference Cole2003). Thus, inequality in living standards of a particular period in history can be inferred from height information. For instance, the difference between the taller half and the shorter half of people who were born in 1960 reflects inequality between the better-off half and the worse-off half of that population between approximately 1960 and 1963. This method overcomes the unavailability of historical data on social classification. Moreover, height data cover a wider range of social groups than do income data. This advantage is particularly important for Latin America, which features high informality, high unemployment and poor rural population.

A question may be raised concerning the genetic determinant of height. At an individual level, height indeed has a genetic basis (Magnusson et al. Reference Magnusson, Rasmussen and Gyllensten2006, p. 662), while at a population level, genetic differences across individuals are mostly averaged out and mean height is mostly determined by living standards (Baten and Mumme Reference Baten and Mumme2010, p. 284). However, genetic differences account for height differences between different ethnic groups, even at a population level, unless the distinct ethnic groups have had comparable living standards for several generationsFootnote ⁸. Differences in dietary habits may also explain height differences between ethnic groups (Blum Reference Blum2013). These two issues are relevant for Latin America, where ethnic heterogeneity is great (e.g. indigenous people vs. white people). The availability of ethnic information in the database does not allow for controlling for ethnic differences in gene and dietary habits. However, this should not be a major problem, because, within a given region, ethnic composition tends to be relatively homogeneous and this paper does not compare heights across regionsFootnote ⁹. Second, according to Leunig and Voth (Reference Leunig, Voth, David and Thomas2001), some factors other than nutritional intake and the intensity of manual labour influence height, such as heating and the concentration of lead in water. However, these factors are closely correlated with household conditions. Poor families are more likely to suffer from the so-called «fuel poverty» and the replacement of lead pipe is slower in poorer areas. Thus, the influences of these factors on height are actually included in the overall influences of household conditions on height. Third, significant improvement in living standards at later ages (e.g. adolescence) basically does not reverse height stunting caused by poor living standards earlier in life, because the influence of environment on height at later ages is negligible in comparison with the influence of the environment in early childhood (Baten Reference Baten2000). Thus, height robustly reflects early childhood living standards. Finally, heights converge as people approach their biological height limits. However, this is more relevant for today's developed countries rather than Latin America in the last century. Even in England, men of the first and the second strata were still 3 cm taller than those of the fourth and the fifth strata in the 1980s (Floud et al. Reference Floud, Floud, Gregory and Wachter1990).

3. DATA AND REGIONAL CLASSIFICATION

This paper uses various rounds of the DHS on the nine countries, covering women aged between 15 and 49 years of age. Women aged 15–19 are dropped because people younger than 20 years of age may not have reached their final height (Moradi and Baten Reference Moradi and Baten2005, p. 4; Akachi and Canning Reference Akachi and Canning2015, p. 115). Height shrinking after 50 years of age is not a problem because the upper bound of age in the DHS is 49. Another potential bias is that women under 22 years of age may not have completed schooling. If so, the level of schooling of women aged 20 and 21 may be underestimated. However, this downward bias is not found due to the very low proportion of women enrolled in tertiary education. Sampling weight and de-normalisation across survey rounds are taken into account in the calculations. Table 1 shows sample size for each country, which is not necessarily proportional to country size. For example, Brazil has the smallest sample size because it was surveyed in only two rounds after 1988Footnote ¹⁰.

TABLE 1 SAMPLE SIZES AT THE SUB-COUNTRY LEVEL

Source: Computed with data from the DHS database.

The nine countries, which are chosen purely based on data availability as explained in section 1, can be dichotomised as the four large South American countries and the five small Central American and Caribbean countries. These countries feature great heterogeneity. Ethnic composition is a telling example. For instance, while Bolivia, Peru and Guatemala have significant indigenous populations, Brazil and Colombia are dominated by white people and people of mixed race. Fixed-effects regression is thus used to control for unobservable or unmeasurable country-specific factors.

The empirical analysis is carried out at a sub-country level. The four types of regions are pre-defined in the DHS database. Capital region includes the capital city and major cities; non-capital region refers to all regions other than capital region; urban region includes the capital city, major and small cities and towns; and rural region refers to regions other than urban regionFootnote ¹¹. The four partly overlapping regions consist of two mutually exclusive pairs: capital vs. non-capital regions and urban vs. rural regions. Capital regions are singled out not only because of developing countries' typical great disparity between capital and non-capital regions, the so-called «primacy effect», but also because of the Latin America-specific characteristic that both the ISI model and the neoliberal reform favoured capitals and large cities more so than other regions. The state-led industrialisation prior to the 1980s was biased towards capital-intensive industry, which clustered in capitals and large cities, and led to the geographical concentration of wealth, especially in large Latin American countries (Baer Reference Baer1972, pp. 107–108). During the neoliberal period, capital regions benefited from advantageous access to foreign information and foreign investors' preference, which enlarged interregional inequality (Baten and Fraunholz Reference Baten and Fraunholz2004, p. 73; Baltzer and Baten Reference Baltzer and Baten2008, p. 193). These asymmetries between capital and non-capital regions can be generalised to the case of urban vs. rural regions.

However, policies in favour of capital and urban regions have side effects. The ISI period witnessed great internal migration largely due to poor economic situations in rural regions and better living standards and preferential policies in urban regions (Frankema Reference Frankema2008). However, newly established industries, which tended to be capital-intensive, did not generate enough employment opportunities, especially for unskilled labours (Frankema Reference Frankema2012). In the 1950s and the 1960s, manufacturing employment grew much slower than the urban population (Baer Reference Baer1972, pp. 106–107). Consequently, unskilled labours were oversupplied, and informality dramatically increased in urban regions, which increased urban inequality. The increasing inequality was not reversed during the neoliberal period (Goldberg and Pavcnik Reference Goldberg and Pavcnik2004; Oostendorp and Quang Reference Oostendorp and Quang2010).

The great internal migration leads to a potential migration bias for regional classificationFootnote ¹². For instance, if an individual was born in rural region and moved to capital or urban region at her school age, then her place of early childhood residence is different from the place where she received schooling. That is to say, interregional migrants' heights reflect the living standards of their places of early childhood residence, but they received schooling in other regions. In order to avoid this discrepancy, this paper combines individuals' birthplace information with their mobility records in the database. An individual is classified under a given region (i.e. capital, non-capital, urban or rural) if she has always lived in this region since she was born, or if she had moved to this region by 5 years old and did not move out thereafter, except for within-region mobility. On the one hand, this treatment ensures that one's regional classification matches her place of early childhood residence, which determines her height. On the other hand, as people usually start schooling at around 6 years old, having moved to a region by 5 years old without moving out thereafter ensures that she received schooling in this region. This treatment leads to the loss of some observations, but not to a large extent.

4. YEARS OF SCHOOLING AND EDUCATIONAL INEQUALITY

Figure 1 shows mean years of schooling for the taller half and the shorter half of women in each type of region for each country from the birth cohort of 1950–1954 to the birth cohort of 1990–1994. Since people normally start schooling at around 6 years of age, data from the birth cohort of 1950–1954 to the cohort of 1990–1994 actually reflect educational inequality from approximately the period of 1956–1960 to the period of 1996–2000 or slightly later. Except for Brazil, all countries show increasing years of schooling across birth cohorts, especially prior to the 1980sFootnote ¹³. In Haiti, rural women's years of schooling even increased from less than 1 year in the 1950s to around 6 years in the 1990s. Different regions within a country share quite similar trends. The taller half and the shorter half also follow similar trends. The increase in years of schooling tends to be weaker in the 1980s and the 1990s for countries with available data. However, it is unclear if this growth stagnation was due to the debt crisis or a «ceiling effect» after mean years of schooling had reached a certain level. At any rate, the recession did not reduce years of schooling.

Source: Computed with data from the DHS database.

FIGURE 1 MEAN YEARS OF SCHOOLING OF THE TALLER HALF AND THE SHORTER HALF.

Women born in capital regions have the most years of schooling for almost all birth cohorts, regardless of the taller half or the shorter half. Rural regions always have the fewest years of schooling. The capital–non-capital and urban–rural gaps are considerable. For the birth cohort of 1950–1954, taller women born in capital regions have approximately 8–12 years of schooling across countries, except for Haiti and Nicaragua, whereas years of schooling for those born in rural regions are approximately 0–4. According to the Barro-Lee data, women born between 1951 and 1955 in the United Kingdom and France have on average 9.47 and 9.35 years of schooling (national average), respectively. Thus, the better-off Latin American women born in the early 1950s in capital regions have comparable years of schooling with women born during the same period in the most advanced countries. This phenomenon still holds true for subsequent birth cohorts. By contrast, rural regions in these countries have schooling level similar to that of other underdeveloped countries. Thus, a simple national average for years of schooling masks the great regional disparity. For instance, according to the Barro-Lee data, the national mean years of schooling for women born between 1951 and 1955 in Bolivia and Peru are 5.51 and 7.07, respectively. Figure 1, however, shows that taller (shorter) women born between 1950 and 1954 in capital regions of Bolivia and Peru have 10.33 (5.09) and 11.65 (9.09) years of schooling, respectively. By contrast, taller (shorter) women from the same birth cohort in rural regions have 2.66 (1.31) years of schooling in Bolivia and 3.04 (2.34) in Peru.

Figure 2 depicts the trend of differences in mean years of schooling between the taller half and the shorter half of women from the birth cohort of 1950–1954 to the birth cohort of 1990–1994. Taller women have more years of schooling than shorter women in 281 out of a total of 284 observations. Inequality in years of schooling tends to be persistent and fluctuant across cohorts. However, birth cohorts in the 1950s and the early 1960s in capital regions show obvious reductions in educational inequality in all countries except for the Dominican Republic, Nicaragua and Peru. The 1950s and the 1960s witnessed the peak of the ISI model and a wave of progressive reforms led by populist and nationalist movements in some Latin American countries (e.g. the Bolivian Revolutionary Nationalist Movement—MNR, the historical Brazilian Labour Party—PTB and the Guatemalan Revolutionary Action Party—PAR). If the ISI model and progressive reforms can explain this educational equalisation in capital regions, it may be argued that they only mattered in capital regions where the government had a stronger influence and the endemic social structure was susceptible to reforms and changes.

Source: Computed with data from the DHS database.

FIGURE 2 TALLER–SHORTER DIFFERENCE IN MEAN YEARS OF SCHOOLING.

In countries with available data, educational inequality tends to decrease for birth cohorts in the 1980s and the 1990s, regardless of region. This finding is in sharp contrast to the well-documented increase in economic inequality during the recession, which corroborates the observation of opposite dynamics between educational inequality and economic inequality (Cox Reference Cox, Attewell and Newman2010). Figure 1 shows that, for the birth cohorts in the 1980s and the 1990s, stagnation of the taller half's increase in years of schooling is more significant than that of the shorter half's. The shorter half in Colombia and Honduras did not even experience stagnation. Thus, educational equalisation for the «recession cohorts» may have been due to the greater (weaker) negative impact of the recession on the better-off half (worse-off half) in educational terms, provided that it was not due to a «ceiling effect». Interregional differences in educational inequality are quite small, except for in Bolivia and Guatemala. Rural regions always have the lowest inequality, regardless of country or cohort. An explanation for this phenomenon is that both better-off and worse-off people in rural regions have few years of schooling, leaving limited space for inequality.

5. DETERMINANTS OF EDUCATIONAL INEQUALITY

This paper focuses on policy and political determinants of educational inequality, because the government's growing role in the Latin American economy during the 20^th century makes it necessary to examine how the government has influenced inequality (Arroyo Abad and Lindert Reference Arroyo Abad, Lindert, Bértola and Williamson2017). The particular interest lies in trade liberalisation, democratisation and educational expansion, which represent three landmark changes in Latin America's economic, political and social spheres in the second half of the last century. Trade policy was the key issue in both the paradigm shift from the early export-led model to the ISI model and the shift from the ISI model to the neoliberal model; democratisation was a historic transition in Latin America during the Third Wave of Democracy in the 1980s; and expansion of primary education was a great social programme prevalent in Latin America in the last century. Moreover, Latin America's tax reform in the 1970s and the 1980s is another variable of interest. This section discusses these four policy and political factors in Latin America, especially their impact on inequality, and then introduces the measurement of these variables and other control variables.

5.1 Trade Liberalisation

Trade may be the most important and contentious issue in Latin America's contemporary economic history. In the 1950s and the 1960s, trade restriction was the core policy tool of the ISI regime (Taylor Reference Taylor1998). In turn, in the 1980s and the 1990s, trade liberalisation was a core component of neoliberal reform. Table 2 shows the periods of trade restriction and trade liberalisation in the nine countries from 1950 to 1994.

TABLE 2 TRADE RESTRICTION AND TRADE LIBERALISATION BETWEEN 1950 AND 1994

Source: Based on Sachs and Warner (Reference Sachs and Warner1995).

The influence of trade liberalisation on inequality, mostly economic inequality, has provoked great debates. According to the Stolper–Samuelson theorem, trade openness increases demand for and return to a country's abundant production factor. Thus, openness should reduce inequality in developing countries, which are abundant in less-skilled and unskilled labour. By contrast, the skill premium argument suggests that trade openness generates skill premium by introducing skill-biased technology and benefits skilled labour more than unskilled labour, which increases inequality (Green et al. Reference Green, Dickerson and Arbache2001). As Latin America became a labour-abundant continent in the last century, trade liberalisation should have an equalising effect. However, empirical studies on Latin America tend to reject the Stolper–Samuelson theorem and support the skill premium argument (Castilho et al. Reference Castilho, Menéndez and Sztulman2012). Wood (Reference Wood1997) argues that Latin America's trade liberalisation coincided with that of large labour-abundant Asian countries, and those Asian countries' lower labour costs offset Latin America's cost advantage in unskilled products. A second explanation for the disequalising effect of trade liberalisation is the so-called «skill-enhancing trade» (Robbins Reference Robbins1995). Particularly, the presence of large Asian countries drove Latin American countries to produce semi-skilled products, which generated skill premium (Oostendorp and Quang Reference Oostendorp and Quang2010, p. 4).

However, how trade liberalisation has influenced educational inequality is under-studied. From the supply side, the influence can be analysed through how trade liberalisation affects social spending, because educational spending is a major part of social spending. The «efficiency hypothesis» argues that trade liberalisation motivates the government to cut down social spending in exchange for competitiveness; as such, inequality increases. By contrast, the «compensation hypothesis» suggests that trade liberalisation motivates the government to raise social spending to maintain political stability in the uncertain post-liberalisation environment. Therefore, it can be argued that post-liberalisation educational provision is partly determined by how the government perceives and deals with opportunities and challenges. Kaufman and Segura-Ubiergo (Reference Kaufman and Segura-Ubiergo2001) find evidence in support of the «efficiency hypothesis» for Latin America. From the demand side, increases in economic inequality after trade liberalisation may translate into corresponding increases in educational inequality. However, skill premium generated by trade liberalisation may encourage poor people to attend school, thus reducing educational inequality, provided that education is affordable. Baten and Mumme (Reference Baten and Mumme2010) find that, in the post-war period, trade liberalisation reduced educational inequality in developing countries, including Latin America. Similarly, Sahn and Younger (Reference Sahn and Younger2006) find that, in Latin America, educational inequality decreased while income inequality increased in the late 1980s and the 1990s. These findings suggest opposite dynamics between educational inequality and economic inequality, because only a small portion of schooling translates into real output (World Bank 2005, p. 68; Cox Reference Cox, Attewell and Newman2010, p. 5). Thus, an equalising effect of trade liberalisation on education is more likely to hold. However, this equalising effect is expected to take place mainly in capital and urban regions. In rural regions, this effect may be weak, if any, due to limited access to external market and limited tradable sectors.

Therefore, how do we measure trade liberalisation? First, it is important to distinguish between «natural openness» and trade liberalisation. Studies that focus on the former usually measure trade openness by deriving from the gravity equation the share of trade in GDP that is purely explained by geography and country size (e.g. Frankel and Romer Reference Frankel and Romer1999). Studies focusing on the latter tend not to use trade share, as it reflects many contributors to trade other than trade policy. As argued by Rodríguez and Rodrik (Reference Rodríguez and Rodrik2000), trade policy indeed affects trade volume, but the outcome of trade policy should not be confused with that of trade volume. Some studies use tariff rates or revenues to measure trade liberalisation (e.g. Santos-Paulino and Thirlwall Reference Santos-Paulino and Thirlwall2004). However, the availability of data on tariff rates and duty revenues for Latin America prior to the 1970s does not allow for using this measure in this paper. Moreover, both official and effective tariff rates have been extensively criticised as a measure of trade liberalisation, primarily because prohibitive tariff rates on particular goods may highly depress their traded volume and the actual effect of tariff rates has to be analysed together with other trade policy instruments.

This paper measures trade liberalisation by two distinct variables. The first is the dummy variable based on trade liberalisation date (Sachs and Warner Reference Sachs and Warner1995). A 5-year birth cohort is classified under trade liberalisation (coded as 1), if at least three out of the 5 years are classified as liberalised period. It is important to note that the trade liberalisation date compiled by Sachs and Warner (Reference Sachs and Warner1995) is different from their well-known cross-sectional trade liberalisation dummy exclusively for the period of 1970–1989. The former is constructed based on review of the literature for each country case, whereas the latter is based on the famous five criteria on tariff, nontariff barrier, state monopoly of exports, black market premium and socialist economyFootnote ¹⁴. Although the cross-sectional trade liberalisation dummy has been extensively criticised (e.g. Rodríguez and Rodrik Reference Rodríguez and Rodrik2000), the data on trade liberalisation date correctly reflect changes in trade policies and are much less subject to the criticism (Wacziarg and Welch Reference Wacziarg and Welch2008). Moreover, following Fajnzylber and Lederman (Reference Fajnzylber and Lederman1999) and Giavazzi and Tabellini (Reference Giavazzi and Tabellini2005), this paper considers the Sachs and Warner trade liberalisation date as a signal of broader market-friendly policy and institutional changes, namely economic liberalisation, because such policy and institutional changes tend to accompany trade liberalisation, especially in Latin AmericaFootnote ¹⁵.

The second variable is constructed, following Wei (Reference Wei1999), by deriving the residual between observed trade share and fitted share (the «natural openness») based on geography, population, language and incomeFootnote ¹⁶. According to Wei (Reference Wei1999), this residual includes the effect of trade policy. A larger residual indicates a larger proportion of trade that is attributable to factors other than geography and population (e.g. trade policy), namely higher trade freedom. This paper extends the original model of Wei (Reference Wei1999) by adding GDP per capita as explanatory variable and using panel data. Clemens and Williamson (Reference Clemens and Williamson2011, p. 13) cite several statistics and show that, in history, at least half of the growth in the world trade and developed countries' trade can be explained by income growth, namely demand and supply factors, rather than reduction in trade barriers. Therefore, controlling for income makes the residual closer to the «policy component of trade». However, this residual has its drawbacks as a proxy for trade liberalisation. During the ISI period, the government restricted imports in basic manufactured goods but promoted imports in capital goods. As a result, in the 1950s and the 1960s, the share of imports in Latin America's GDP did not decline (Baer Reference Baer1972, p. 103). Moreover, even after controlling for income, the «residual» still includes many contributors to trade volume in addition to trade policy. In this regard, this residual is similar to a black box, as it is unclear to what extent it is driven by trade policy. Thus, it is used only as a secondary option. In brief, the Sachs and Warner dummy mainly reflects subjective policy attitudes, while the residual openness mainly reflects objective outcomes of policy changes and other unspecified factors.

Can the causality run from educational inequality to trade policy? Baten and Mumme (Reference Baten and Mumme2010) consider this not to be a major problem, in light of the finding of Baltzer and Baten (Reference Baltzer and Baten2008) that lower economic inequality in Latin America did not stimulate openness. However, stronger rejections of the possibility of endogeneity can be derived from history. Trade policy, either restriction or liberalisation, was the core and symbolic policy tool of both the ISI model and the neoliberal model. Because the emergence of these two models is due to factors exogenous to educational inequality, trade restriction and trade liberalisation should be considered as exogenous. Latin America's ISI was driven by «big facts», such as the turndown of international trade and the collapse of the global financial system caused by the two world wars and the Great Depression, the success of the Soviet state-led industrialisation, and the promotion by the Latin American structuralist school (Baer Reference Baer1972; Ocampo and Ros Reference Ocampo, Ros, Ocampo and Ros2011). Moreover, Latin America's trade restriction policy also had its root in the earlier balance of payments management as a response to the volatile global market and capital inflow (Ocampo and Ros Reference Ocampo, Ros, Ocampo and Ros2011, p. 5). The neoliberal reform in the 1980s was a response to the debt crisis under the influence of international agencies, to the success of the East Asian model, and to the failure of the Soviet model, which are the «big facts» as well in the words of Lindauer et al. (Reference Lindauer, Pritchett, Rodrik and Eckhaus2002) and Ocampo and Ros (Reference Ocampo, Ros, Ocampo and Ros2011). Moreover, Latin America's neoliberal reform was also part of the global wave of neoliberalisation following the «Washington Consensus». Finally, in the following regressions, trade liberalisation is treated with a lag structure, which further avoids endogeneity.

5.2 Democracy

Latin America's inequality is rooted in the exclusionary institutions established in the Spanish colonial period and has been perpetuated by elites' privilege-based institutional arrangement (De Ferranti et al. Reference De Ferranti, Perry, Ferreira and Walton2004; Prados de la Escosura Reference Prados De La Escosura, Hatton, O'Rourke and Taylor2007). Inequality in political rights between elites and the rest of the population impeded educational development and disequalised educational distribution (Engerman et al. Reference Engerman, Mariscal, Sokoloff, Eltis, Lewis and Sokoloff2009). Latin America's elites resisted the redistribution of income and wealth through tax-supported education, because of the privilege-based social structure and low benefits from investing in public education (Frankema Reference Frankema2009; Lindert Reference Lindert2010). By contrast, the evidence that the Northern United States led its Southern counterpart in primary education in the 19^th century shows how equal distribution of political rights promoted redistribution towards tax-supported education (Go and Lindert Reference Go and Lindert2010). Brown (Reference Brown1999) and Brown and Hunter (Reference Brown and Hunter2004) find a positive influence of democracy on education in Latin America. However, a condition for democracy to take equalising effect is sufficient political consciousness of the poor (Bollen and Jackman Reference Bollen and Jackman1985). For many Latin American countries in the last century, this condition might not hold, especially for rural regions. Moreover, the functioning of democratic governments in some Latin American countries was subject to the strong presence of business and military elites (Brown and Hunter Reference Brown and Hunter1999). Economic inequality is another factor that prevents democracy from operating (Cárdenas and Helfand Reference Cárdenas, Helfand, Durlauf and Blume2011). Thus, although democracy can in theory weaken the elites' resistance to redistribution, it is likely that social structure restricts the functioning of democracy in Latin America. This leads the actual effect of democracy to be an empirical question.

Another concern stems from the diverse political spectrum of the Latin American non-democratic regimes. Among the nine countries under study, non-democratic rulers range from the leftist Nicaraguan Sandinista National Liberation Front—FSLN to the rightist Haitian Duvalier family. Different non-democratic regimes may have had opposite influences on inequality. This complicates the question on the role of democracy. In this paper, democracy is measured by the dummy variable from the Dictatorship and Democracy Database (Cheibub et al. Reference Cheibub, Gandhi and Vreeland2010). If a 5-year birth cohort has at least 3 years under democracy, this cohort is classified under «democracy». Table 3 lists democratic periods in the nine countries between 1950 and 1994.

TABLE 3 PERIODS OF DEMOCRACY BETWEEN 1950 AND 1994

Source: Based on Cheibub et al. (Reference Cheibub, Gandhi and Vreeland2010).

6. EMPIRICAL STRATEGY AND RESULTS

Separate panel regressions are used to examine the determinants of educational inequality in capital regions, non-capital regions, urban regions and rural regions, respectively. The dependent variable is the difference in mean years of schooling between taller and shorter women in each region of each country for each birth cohort. All data are arranged quinquennially. Table A1 in Appendix provides summary statistics for the panel used in regressions in Table 4Footnote ¹⁸. Country-invariant and year-variant shocks (e.g. the debt crisis), namely time fixed effects, are controlled by using quinquennial dummies. Because people usually start schooling at around 6 years old, schooling data for each 5-year birth cohort (e.g. 1950–1954) actually reflect educational inequality for the 5-year period more or less 6 years later (e.g. 1956–1960). Thus, there are two sets of regressions.

TABLE 4 DETERMINANTS OF EDUCATIONAL INEQUALITY: FIXED-EFFECTS REGRESSIONS WITH LAGGED EFFECT

Note: Robust standard errors in parentheses: ***P < 0.01, **P < 0.05, *P < 0.1. Dependent variable (Edu. ineq.): taller–shorter difference in mean years of schooling. SW refers to using the dummy of liberalisation date (Sachs and Warner Reference Sachs and Warner1995) to measure trade liberalisation. Residual refers to using the residual between observed trade share in GDP and fitted share to measure trade liberalisation.

The first set of regressions takes a lag structure by using the values of all explanatory variables during each birth cohort (e.g. 1950–1954) to estimate that birth cohort's educational inequality, which realises in the 5-year period around 6 years later (e.g. 1956–1960). For instance, educational inequality of the birth cohort of 1950–1954, which realised during approximately 1956–1960, is regressed on trade liberalisation during 1950–1954. This lag of approximately 6 years allows the explanatory variables to take effect. Lagged effect is particularly important for trade liberalisation and democracy, as their influence on educational decision tends to operate via a lag (Wood and Ridao-Cano Reference Wood and Ridao-Cano1999; Galor and Mountford Reference Galor and Mountford2008; Cárdenas and Helfand Reference Cárdenas, Helfand, Durlauf and Blume2011; Blanchard and Olney Reference Blanchard and Olney2017). The second set of regressions is concerned with the possible contemporaneous effect on educational inequality of GDP per capita, educational expansion, tax reform, war and demographic composition, and thus uses the contemporaneous form of these variables (e.g. using war dummy during 1955–1959 to estimate educational inequality of the birth cohort of 1950–1954, which realised in 1956–1960). Due to data availability, I use data from the next 5-year period (e.g. 1955–1959) rather than exactly the 5-year period 6 years later (e.g. 1956–1960), but this should not have been a problem. Trade liberalisation and democracy are still arranged with a lag structure due to the clear lagged effect indicated in the literature. The Hausman test favours fixed-effects model, and Huber–White robust standard errors are used to deal with heteroskedasticity.

Table 4 presents the results from the lag regressions. Trade liberalisation is defined by Sachs and Warner's dummy in columns denoted by «SW». In columns denoted by «Residual», trade liberalisation is defined by the residual between observed trade share and fitted share. Adjusted R ² shows that the model has much higher explanatory power for capital regions than other regions. The SW dummy reduces educational inequality by 0.43 years in capital regions but is statistically insignificant for other regions. In contrast, increasing Residual by one percentage point raises educational inequality in rural regions by 0.02 years. The magnitude of impact is negligible given the limited within-country dispersion of Residual by only several percentage points. Residual is insignificant for other regions. However, due to the black-box feature of Residual, it is unclear to what extent its effect is driven by trade policy. Thus, it may not be advisable to interpret much from this result. Shift to democracy increases educational inequality in capital regions by 0.71 years and 0.67 years in the two specifications, but it is insignificant for other regions. This is quite a large impact, and it contradicts the expectation that democracy serves as an equalising force. Increasing the primary enrolment ratio by one percentage point reduces educational inequality by 0.04 or 0.03 years in capital regions. This magnitude of impact is not small considering that the within-country standard deviation of enrolment ratios is around ten percentage points. This variable is insignificant for other regions. Neither tax reform nor war is statistically significant for any regions. The two demographic variables are barely significant. GDP per capita is significant in all specifications, increasing educational inequality in capital and urban regions but reducing it in non-capital and rural regions. A one-standard-deviation increase in logged GDP per capita (0.39 or 39 per cent) increases educational inequality by around 0.64 years (0.39×1.64) in capital regions but reduces it by 0.27 or 0.32 years in rural regions. However, this interpretation has to be treated cautiously, because within-country standard deviation of GDP per capita greatly varies across countries, ranging from 0.07 in Bolivia to 0.45 in Brazil.

Table 5 presents results for the contemporaneous regressions, which are basically similar to the lag regressions. The statistical significance of SW dummy's equalising effect in capital regions increases to the 1 per cent level. Democracy's disequalising effect in capital regions increases to over 0.9 years. Educational expansion loses its significance in capital regions in column 2. Tax reform becomes significant in capital regions with quite a large disequalising effect of 0.75 years and 0.74 years in the two specifications, respectively, indicating the disequalising effect of regressive taxation. For non-capital and rural regions, GDP per capita loses its significance. The share of school-aged people (Immature) becomes significant in non-capital and rural regions, with a small disequalising effect. Finally, specifications with quadratic terms for the enrolment ratio do not show Kuznets effect. Adding mean years of schooling and the quadratic term in the model does not show statistical significance either, implying the absence of educational Kuznets effect. For the parsimony of the model, mean years of schooling are not included in the final regressions.

TABLE 5 DETERMINANTS OF EDUCATIONAL INEQUALITY: FIXED-EFFECTS REGRESSIONS WITH CONTEMPORANEOUS EFFECT

Note: Robust standard errors in parentheses: ***P < 0.01, **P < 0.05, *P < 0.1. Dependent variable (Edu. ineq.): taller–shorter difference in mean years of schooling. SW refers to using the dummy of liberalisation date (Sachs and Warner Reference Sachs and Warner1995) to measure trade liberalisation. Residual refers to using the residual between observed trade share in GDP and fitted share to measure trade liberalisation.

Finally, are the results driven by influential observations? Table 6 provides results for regressions after dropping influential observations detected by influence statisticsFootnote ¹⁹. Only regressions with the SW dummy are presented, because regressions with Residual do not show any qualitative differences. Dropping influential observations significantly increases adjusted R ² for all regions. The model's explanatory power for capital and urban regions is still higher than that for non-capital and rural regions. Regressions for capital regions do not show qualitative differences in comparison with full-sample regressions. The SW dummy becomes statistically significant with equalising effect for non-capital and urban regions in the lag regressions, but remains insignificant for rural regions. Educational expansion becomes significant with equalising effect in urban regions and remains insignificant in non-capital and rural regions. Democracy becomes significant with small disequalising effect in urban regions in the lag regression. The two demographic variables become significant, implying that demographic composition may be an important factor for educational inequality, which is beyond the scope of this paper. In summary, dropping influential observations leads to a sharper contrast between capital and urban regions on the one hand and rural regions on the other.

TABLE 6 ROBUSTNESS CHECKS: FIXED-EFFECTS REGRESSIONS WITH LAGGED AND CONTEMPORANEOUS EFFECT

Note: Robust standard errors in parentheses: ***P < 0.01, **P < 0.05, *P < 0.1. Dependent variable (Edu. ineq.): taller–shorter difference in mean years of schooling. SW refers to using the dummy of liberalisation date (Sachs and Warner Reference Sachs and Warner1995) to measure trade liberalisation.