The Geography of Intergenerational Mobility in Latin America and the Caribbean

,


I Introduction
Intergenerational mobility (IGM) has gained interest in the economic literature among other things thanks to its importance for equity, social cohesion, and economic growth. Its observed correlation with income inequality, commonly named "the Great Gatsby Curve", has contributed to the desire for understanding IGM given the documented rise of inequality over the last decades in rich countries (see Corak, 2013).
In the case of the developing economies, the Latin America and the Caribbean (LAC) region is of particular interest because of its historically documented high levels of income inequality relative to other regions of the world. 1 However, the scarcity of high-quality data (e.g. long panel data sets or tax records with linked generations) has limited the study of IGM in income. These constraints are also common in other regions (e.g., Africa), so efforts to document IGM on a global scale have taken an alternative path given by the measurement of mobility in education. These measures are of interest in and of themselves, but they are also a proxy for economic status given the close relationship between education and income.
Recent studies have used household and public opinion surveys with retrospective information about parents' education to document the levels of IGM in education in LAC at the country level (for an example, see Hertz et al., 2007;Narayan et al., 2018;Neidhöfer, Serrano, & Gasparini, 2018). However, analyzing IGM at a more geographically disaggregated level, as argued in Narayan et al. (2018), is valuable because it can help researchers understand the importance of localized patterns and drivers of IGM, as shown for the case of developed countries. Along these lines, Chetty, Hendren, Kline, and Saez (2014) state that the United States can be better described as a collection of societies, some of which are "lands of opportunity" with high rates of mobility across generations, and others in which only a few children escape poverty. Thus far, this type of analysis has not been conducted in LAC countries as a whole due to the inadequacy of most survey data for this purpose. This paper fills that gap in the literature by generating estimates of IGM in education at smaller geographical levels.
In this paper, I estimate intergenerational mobility in education for LAC countries at a disaggregated regional level using data from 91 censuses. The analysis covers 24 countries spanning more than half a century (between 1960 and 2012). I rely on samples of co-residents (i.e., children living with their parents or older relatives). To minimize the impact of coresidence, I investigate mobility in education at the bottom of the educational attainment distribution by focusing on primary education, which can be measured with a high degree of confidence between ages 14 and 18. Furthermore, an important share of the population does not attain more than primary education in the period analyzed and this focus allows me to create indicators that are directly comparable to the estimates recently generated for 27 countries in Africa (see Alesina, Hohmann, Michalopoulos, & Papaioannou, 2021), a continent that shares the feature of having high levels of income inequality despite its lower levels of income and higher poverty rates.
The estimates of upward (and downward) mobility measured as the likelihood of finishing (or failing to finish) primary education, conditional on having parents who failed to finish (or who were able to finish) primary school, show wide cross-country and within-country heterogeneity. In LAC, the distance between the most and least upwardly mobile countries is similar to what has been recently documented in Africa, although the least mobile countries in Africa are less mobile than the least mobile country in LAC. I find only small differences by gender, but I do document a declining trend in the mobility gap between urban and rural populations. At the sub-national level, there is heterogeneity in mobility across districts/provinces, and some countries show lower levels of mobility in the northern regions (e.g., Brazil), whereas the opposite is true for Mexico. However, the variability is much lower in countries with lower number of regions and less population. The level of mobility at the sub-national level is highly positively correlated to the share of primary completion of the previous generation, which suggests a high level of inertia. In addition, geographical correlates do not appear to be highly correlated to mobility except for distance to the capital.
Similarly, some proxies of economic development like the share of employment in industry and agriculture at the beginning of the sample period seem to be associated to the levels of mobility at the district level.
This paper contributes to several strands of the economic literature. First, it adds to the literature on intergenerational mobility in general (see Black & Devereux, 2011, for a survey) but specifically to the literature focusing on the geography of socioeconomic mobility that recently received more attention in part because of the work of Chetty et al. (2014), which shows important variation across commuting zones in the United States. Second, it adds to the recent wave of research that looks at intergenerational mobility in education (see Emran & Shilpi, 2021;Torche, 2019, for recent surveys focused on developing countries).
This set of papers includes on the one hand those that use household survey data or opinion surveys. For example, Hertz et al. (2007), Narayan et al. (2018), and Van der Weide, Lakner, Gerszon Mahler, Narayan, and Ramasubbaiah (2021) that document IGM for a very large set of countries across the world, 2 and Neidhöfer et al. (2018) that focus on 18 countries from Latin America. This paper expands over them in terms of country coverage of the region and cross-country comparability. In addition, these estimates use the same type of data of recent estimates available for Africa, allowing a cross-regional comparison that was not available. 3 On the other hand and more closely related to this paper, it contributes to the pool of studies using administrative data or census data. For instance, Asher, Novosad, and Rafkin (2021) study mobility among different marginalized groups and analyze geographic differences in India; Card, Domnisoru, and Taylor (2022) use 1940 census data to study the role of school quality in mediating upward mobility in the US; Van der Weide, Ferreira de Souza, and Barbosa (2020) study mobility at the sub-national level in Brazil; and most closely related to this paper, Alesina et al. (2021) document patterns of IGM in Africa using census data and estimate regional childhood exposure effects using migrants. To the best of my knowledge, this is the first paper to document IGM at a very disaggregated regional level for almost the entire population in LAC.
The paper is organized as follows. Section II describes data and methodology. Section III reports the main descriptive results at country level and the geography of mobility. Section IV looks at correlates of intergenerational mobility. Section V concludes with final remarks.

II Data and Methodology
Three sources of data have been typically used to estimate intergenerational mobility: (1) cross-sectional samples of adult populations with retrospective questions about parental education, for example, Narayan et al. (2018) use household survey data that covers the 96% of the world population; (2) panel data long enough in its time dimension to include the socioeconomic or educational attainment of two generations, for example, Celhay, Sanhueza, and Zubizarreta (2010) use the Chilean CASEN to estimate mobility in schooling and income; and (3) administrative/registry data with linked information for parents and adult children, for example, Chetty et al. (2014) use tax records in the U.S. to estimate income mobility.
In the case of Latin America, most of the literature has used household survey data or public opinion surveys (see for example, Hertz et al., 2007;Narayan et al., 2018;Neidhöfer et al., 2018) given that long panel data as well as administrative/registry data that allow the researcher to link generations are rare. In contrast, in this paper, I use census data obtained from IPUMS International (Integrated Public Use Microdata Series, IPUMS, 2019), hosted at the University of Minnesota Population Center, which reports harmonized representative samples (typically 10%) of full census micro data sets for a large number of countries. In particular, I use 91 samples of population and housing censuses from 24 countries, which are run to compute the total population and contain an educational attainment question in their questionnaire. 4 The key advantage of this data set is that it contains the entire population (or at least a large share of it publicly available) at a point in time, allowing me to analyze mobility at a very disaggregated geographical level. However, the main disadvantage of this data set is that it does not link all the individuals with their parents because both (individuals and parents) need to be part of the same household. Below, I explain how this is addressed and I refer to recent evidence showing that the coresidence bias is likely to be very small for the indicators used in this paper.  Table A1 in the Appendix for the details about the fraction of the data available by census), and they represent 91 samples of these 24 countries drawn at various points from 1960 to 2012.

II.1 Countries and smaller administrative units
In terms of geography, IPUMS reports residence at the time of the interview for at most two levels of administrative units in which the households were enumerated. These variables contain the geographies for every country harmonized spatio-temporally to provide spatially consistent boundaries across samples in each country. This allows me to assign individuals to "coarse" (roughly similar to states in the U.S.) and "fine" administrative units (roughly similar to counties in the U.S.). The sample spans 400 provinces (admin-1 units) and 6,684 districts (admin-2 units). The baseline estimates will make use of the former to avoid issues derived from having a reduced number of observations per administrative unit but estimates using the latter are also reported in the Appendix.

II.2 Linking generations and coresidence
The data collection is organized at the household level, so it is possible to link only those individuals who live in the same household at the time of the interview. The data set includes a variable that by means of 62 different values details the relationship between the individual and the head of the household. Based on this variable, I classify individuals into five different generations where the head corresponds to generation zero (see Table 1), and based on the generation number I use individuals who live with at least one member of the immediately previous generation, where these old generation members are considered as "pseudo-parents". 5 Table 1 provides the details of the assignment.  Figure 1 shows the unweighted average rate of coresidence by age in the sample pooling all 5 A similar approach is followed by Alesina et al. (2021) with Census data from Africa.
the countries and years. There are rates above 90% for individuals before reaching 18 years old that then start decreasing more rapidly getting close to 40% for people who are 25 years old. When the coresidence rate is computed with samples that distinguish urban/rural or gender, I find negligible differences in the former and a steeper fall in the rate of coresidence by age for women relative to men (see Figure A1 in the Appendix). Coresidence rate by age Notes: Coresidence is defined as living with at least one relative of the immediately previous generation. The data in the graph is unweighted. Figure 2 disaggregates the coresidence rate by country, displaying some variability in the magnitude of the fall in coresidence with age. This figure also suggests that the magnitude of the fall in coresidence around age 25 is driven by Brazil, which is the most populated country in Latin America. Table A3 in the Appendix provides coresidence rates by country for different age groups.
A concern associated with the use of coresidents is that it may generate bias in the es- timates of IGM as individuals who reside with their parents may systematically differ from those not residing with them (see for example, Emran, Greene, & Shilpi, 2018;Emran & Shilpi, 2021;Francesconi & Nicoletti, 2006). However, Munoz and Siravegna (2021) show that the average coresidence bias when computing upward mobility (measured as the likelihood of completing primary for those whose parents did not complete primary) for individuals aged 21-25 years (with coresidence rates of less than 50% on average) is approximately 2%. In addition, the ranking obtained using these coresident samples closely follows the one obtained with a sample that includes all children (the Spearman rank correlation between the estimates with the full sample and those with coresident samples is 0.91). Given these findings, the potential for coresidence bias in my estimates is small as they are computed using individuals aged 14-18 (or 14-25) years, a group with much higher rate of coresidence. 6

II.3 Education
Why is education a suitable variable by which to measure IGM? Education as a measure of socioeconomic status relative to income in the context of developing countries has at least three advantages: (1) it contains less measurement error, reducing potential attenuation bias (see Solon, 1992); (2) it is fixed early in the life cycle, which avoids the life-cycle bias found in studies that use income (see Haider & Solon, 2006), in addition, education is closely linked to income and it is important by itself in terms of human development; and (3) it can be completely attributed to a specific individual, while income sometimes is hard to assign within a household (e.g., a household with multiple persons and home production, which may be specially relevant in the case of rural populations in poor countries).
There are two questions about educational attainment in the data set. The first one reports the total years of schooling completed by each individual (formal schooling regardless of the track or kind of study), and the second one is re-coded by IPUMS to capture educational attainment in terms of the level of schooling completed 7 and contains four categories: (1) less than primary completed, (2) primary completed, (3) secondary completed, and (4) university completed. In the main analysis of the paper, I use the latter variable, which has a lower number of missing values and it is available for more countries than the former. 8 This variable applies, to the extent possible, the United Nations standard of six years of primary schooling, three years of lower secondary schooling, and three years of higher secondary schooling.
In the sample, a majority of individuals report levels of education that correspond to less than completed secondary and near 50% less than primary (see Figure 3a), which supports the focus on primary completion that I will detail later as most of the action happens at lower levels of completion. 9 In addition, although the level of education in Latin America 7 It does not necessarily reflect any particular country's definition of the various levels of schooling in terms of terminology or number of years of schooling.
8 Years of schooling is not available for Brazil 2010, Cuba 2002, Saint Lucia 1991, Suriname 2012, Trinidad and Tobago 1970, and Uruguay 2011 An additional reason to focus on primary completion is that these estimates are directly comparable to and the Caribbean has been increasing across cohorts (see Figure 3b), the continent still shows a share of around 60 percent with at most nine years of schooling in the most recent cohort (those born in the 1980s), which roughly corresponds to the completion of lower secondary education. Nonetheless, I also provide an Appendix using estimates that focus on the completion of secondary level. Figure 4 shows the transition matrix for individuals older than 25 to get a rough idea of the patterns of intergenerational education mobility present in the data set. 10 This plot highlights that the action is terms of mobility happens in the lower two levels of educational attainment, qualitatively similar to what can be seen in Alesina et al. (2021) for the African continent. The same mosaic plot can be found by country in the Appendix. Two countries that stand out in terms of low and high levels of parental attainment of primary education the ones recently documented in Alesina et al. (2021) for Africa. 10 I use individuals older than 25 years as younger ones are unlikely to have completed their education. The main analysis uses younger individuals as the focus is on primary completion. Figure A3 in the Appendix reproduces this mosaic with individuals aged 14-25 years.

II.4 Methodology
For each individual in the sample, I analyze the relationship between their own educational attainment against the average attainment of individuals one generation older living in the same household, rounded to the nearest integer. For this I consider a measure of absolute 11 Saint Lucia shows similar patterns but contains a much smaller population. 12 The same plot using individuals aged 14-25 years can be found in the Appendix (see Figure A4).   intergenerational mobility that reflects the likelihood that a child completes a strictly higher or lower education level than the members of the immediately previous generation in the household (parents and/or extended family members, such as aunts and uncles).
Upward mobility at the country level. To estimate upward IGM, I estimate the following econometric specification, pooling observations from all the censuses and countries: where y up icoyt is a dummy variable that takes a value equal to one when individual i completes at least primary education and zero otherwise. The parameters γ b o , γ b y , θ t refer respectively to fixed effects by decade-cohort of the individual i, decade-cohort of the generation above that co-resides with individual i, and census year. This regression uses a sample of individuals with ages between 14 and 18 (or 14 to 25), for whom the generation above (parents or older relatives) have on average less than primary education. Hence, α up c is the parameter of interest and measures the likelihood of completing primary for children whose "parents" did not complete primary net of cohort and census year effects.
This empirical approach has been used in Alesina et al. (2021) with data from Africa and delivers a measure of mobility comparable between countries that captures some long-term patterns over half a decade by netting out common (across countries) birth cohorts and census year effects. Downward mobility at the country level. To estimate downward IGM, I use a similar econometric specification, pooling observations from all the censuses and countries: where y down icoyt is a dummy variable that takes a value equal to one when individual i does not complete primary education and zero otherwise. The parameters γ b o , γ b y , θ t again refer respectively to fixed effects by decade-cohort of the generation above that co-resides with individual i, decade-cohort of the individual i, and census year. This regression uses a sample of individuals with ages between 14 and 18 (or 14 to 25), for whom the generation above (parents or older relatives) have on average completed at least primary education. 13 Hence, α down c is the parameter of interest and measures the likelihood of failing to complete primary for children whose "parents" completed at least primary school net of cohort and census year effects.
Upward and downward mobility at finer geographical level. To estimate IGM at a more disaggregated level (i.e., provinces or districts), I use the following econometric specifications run country by country: 13 I use the average attainment of the generation above instead of the maximum to make these estimates directly comparable to those in Alesina et al. (2021). However, this decision makes little difference as I explain in the robustness section later.
where the variables and subscripts in common have similar interpretation as in Equations 1 and 2, and the additional subscript r refers to the district or province according to the level of geographical dis-aggregation used in the analysis (provinces as the baseline estimates and districts as an additional exercise reported in the Appendix).
Why is primary education a suitable variable by which to measure IGM?
The focus on primary education is based on the fact that a non-negligible share of the population in Latin America and the Caribbean has an educational attainment of less than primary as shown in the previous subsection. Furthermore, this focus makes the analysis directly comparable to the recent work of Alesina et al. (2021) in Africa and allows me to minimize the potential bias that comes from using samples of co-residents. Nonetheless, the focus on the lowest level of education can also be justified from a conceptual point of view.
Development policy discussions often claim that the poorest should not be left behind and this focus is related to the school of moral philosophy exemplified by the principle of justice proposed by Rawls (1971). 14 Robustness. As a robustness check, I compute upward and downward mobility using some alternative options in terms of data construction. First, I use the maximum attainment of the generation above instead of average. This change produces estimates with differences that are negligible (for example, Pearson correlation coefficient between the measures using average versus maximum at the country, province, and district level are approximately 1).
Second, I estimate mobility using a sample of individuals linked to (probable) parents as done by IPUMS (2019). This change produces estimates that are also highly correlated (for example, Pearson correlation coefficient between the measures using average versus maximum at the country, province, and district level are 0.98, 0.97 and 0.93, respectively).
Alternative measures of IGM. I estimate a set of additional measures of intergenerational mobility, which are less focused on the bottom of the educational attainment distribution. In contrast to the estimates that focus on primary education, these measures are computed using individuals between ages 19 and 25. First, I estimate upward and downward mobility considering secondary education instead of primary. Second, I estimate upward mobility as the likelihood of finishing at least secondary education for those whose generation above were not able to complete primary school. These indicators are more prone to suffer from coresidence bias but they still provide valuable information. For example, Munoz and Siravegna (2021) show that the rank correlation between indicators of upward mobility using secondary level computed with all children versus coresidents is approximately 0.86. 15 Table 2 summarizes the estimates of mobility at the country level. On average, close to 50 percent of children with parents that did not finish primary education (from now on, illiterate parents) are able to complete primary. On the other hand, downward mobility is close to 10 percent, as one out of 10 children with parents that finished primary education (from now on, literate parents) do not complete primary.

III.1 Country-level estimates
There is substantial heterogeneity within LAC countries. The probability of completing primary for children of illiterate parents ranges from 18% in Guatemala to 87% in Jamaica.
In the case of downward mobility, the estimated probability of not completing primary for children of literate parents ranges from being null in Jamaica to 23% in Haiti. The heterogeneity found in upward mobility in Latin America (e.g., the 69 percentage points gap between Jamaica and Guatemala) is relatively similar to the one documented for African countries (e.g., the 75 percentage points gap between South Africa and South Sudan) by Alesina et al. (2021), although with higher minimum and maximum values. Furthermore, the level of upward mobility among countries in LAC shows substantial overlap with that of Africa. Countries such as Haiti, Guatemala, and Nicaragua with the lowest levels of upward 15 Using 72 country and 5-year birth cohorts that span 18 countries in Latin America. mobility in LAC are more upwardly mobile than the five lowest (Malawi, Ethiopia, Sudan, Mozambique, and South Sudan) out of the 27 countries for which Alesina et al. (2021) provide estimates.  1970,1980,1990,2000,2011 .  1963,1975,1985,1996,2006,2011 .668 .  1960,1970,1980,1990,2000,2010 .  1962,1972,1982,1992,2002.432 .463 .116 .096 118,082 207,766 Colombia 1973,1985,1993,2005 .  1960,1970,1980,1991,2000,2010.367 .422 .171 .128 10,755,296 18,713,402 El Salvador 1992,2007.342 .374 .164 .138 85,402 150,582 Haiti 1971,1982,2003.212 .266 .226 .178 104,465 183,588 Nicaragua 1971,1995,2005.194 .238 .223 .18 93,635 167,740 Guatemala 1964,1973,1981,1994,2002 . Notes: Columns (1) and (2) give upward-IGM estimates. They reflect the likelihood that children, aged 14-18 and 14-25, whose parents have not completed primary schooling will manage to complete at least primary education. Columns (3) and (4) give downward-IGM estimates. They reflect the likelihood that children, aged 14-18 and 14-25, whose parents have completed primary schooling or higher will not manage to complete primary education. Columns (5) and (6) give the number of observations used to estimate the country-specific IGM statistics (children whose parental education is reported in the censuses). Countries are sorted from the highest to the lowest level of upward IGM in the 14-18 sample (column (1)). "mean" gives the unweighted average of the 24 country-estimates.  Notes: Upward mobility reflects the likelihood that children, aged 14-18, whose parents have not completed primary schooling will manage to complete at least primary education. Downward mobility reflects the likelihood that children, aged 14-18, whose parents have completed primary schooling or higher will not manage to complete primary education. Both estimates are net of cohort and census year effects. significantly negatively correlated (see Figure A13 in the Appendix).
Country-level estimates of intergenerational mobility focused on secondary education can be found in Table A7 of the Appendix. The level of upward (downward) mobility is considerably lower (higher) and the samples smaller. Similar to the estimates using primary education, there is significant variation across countries. In the case of upward mobility measured as the likelihood that children complete at least secondary education when their parents were not able to complete primary, we see lower levels of mobility at the country level as one may have expected (see Table A10 in the Appendix).

III.1.a Urban-rural
Given that an important feature of most developing countries is the gap in living standards between rural and urban residents (see Lagakos, 2020), I explore the heterogeneity  1950s, 1960s, 1970s, 1980s other countries mean among countries with data 1950s-1980s (b) Downward mobility Notes: These estimates correspond to the probability of completing at least primary education for those whose parents did not finish primary school in the case of upward mobility and probability of not completing primary education for those whose parents completed primary school in the case of downward mobility. They are estimated for individuals aged 14-18 years by country, birth decade of the "children" and urban/rural status of the household residence.
in IGM between these populations and document how they have evolved across birth cohorts. I do so by estimating upward and downward mobility 16 by country, birth decade of the "children" and urban/rural status of their residence. Figure 7 reports the gap between the upward/downward mobility in urban-rural areas over birth cohort. I find a positive gap that has been declining from 36 percentage points (i.e., upward mobility in urban areas is on average 36 percentage points higher than in rural areas for the cohort born in years [1950][1951][1952][1953][1954][1955][1956][1957][1958][1959] to 20 percentage points as one moves towards older birth cohorts. Similarly, the gap in downward mobility is closing from below moving from 29 percentage points for 1950 birth decade to 15 percentage points for 1980 birth decade. Figure A14 and Figure A15 in the Appendix show estimates by sub-population rather than the gap between them for countries with data for at least 4 decades, suggesting that the gap has been decreasing because of an increase (decrease) in upward (downward) mobility. 16 The probability of completing at least primary education for those whose parents did not and the probability of not completing primary for those whose parents complete primary school, respectively.

III.1.b Gender
As discussed in a recent survey on IGM in developing countries (see Torche, 2019), gender gaps in education have been disappearing or even moving in favor of women. I examine whether these patterns hold in this novel data set by estimating IGM for males and females separately and documenting how the gap between these populations has evolved across birth cohorts. I estimate upward and downward mobility 17 by country, birth decade of the "children" and gender. I do not find systematic differences by gender for older birth cohorts but it appears that there is a trend towards higher upward mobility for women as they have 3 percentage points higher upward mobility in the 1980s birth cohort (see Figure 8) while the gap in downward mobility move around similar values (the gap in favor of women is approximately 3 percentage points for 1980s birth cohort) with a flatter trend. Figure A16 and Figure A17 in the Appendix show estimates by sub-population rather than the gap between them for countries with data for at least 4 decades, suggesting that the gap has been increasing in favor of women because of an increase (decrease) more than proportional for them in upward (downward) mobility.

III.1.c Evolution over time
As mentioned in the data section, the coverage over time is unbalanced with some countries spanning more years than others. This limits the analysis of trends over time and the conclusions that can be derived from comparisons between them at given points in time or for a given cohort. Nevertheless, I document estimates of mobility by country for the different birth cohorts that are available. Figure 9 reports these estimates. It is clear how the level of upward mobility has been increasing at the same time that downward mobility has been falling. This is not surprising given the fact that educational attainment has increased in   1950s, 1960s, 1970s, 1980s other countries mean among countries with data 1950s-1980s (b) Downward mobility Notes: These estimates correspond to the probability of completing at least primary education for those whose parents did not finish primary school in the case of upward mobility and probability of not completing primary education for those whose parents completed primary school in the case of downward mobility. They are estimated by country, birth decade of the "children" and gender.  1940 1950 1960 1970 1980 1990 (b) Downward mobility Notes: The estimates are done by birth decade cohort of the children.
the region over the last decades. Table 3 summarizes the estimates of mobility at the province level. These results show that there are countries with substantial variance in mobility levels across provinces. This is for example the case of Paraguay, Mexico, Guatemala, Bolivia, and Peru, where the difference in upward mobility between the most upwardly mobile to the least upwardly mobile is more than half the range found in the case of countries in Latin America. However, there are also particular cases with either high or low upward mobility at the country level and a very small variation within country, such as Jamaica and Haiti, although somewhat expected as they correspond to countries with small number of administrative units and population.

III.2 Spatial variation of intergenerational mobility in LAC
In the case of downward mobility, the variability is much smaller. However, Paraguay stands out as a case where the range between the provinces with the minimum and the maximum level of downward mobility is relatively wide. Figure 10 and 11 maps the same estimates that are summarized in Table 3. We can see some interesting patterns in some countries. For example, Mexico shows a somewhat lower level of upward mobility in the south and you can identify a lighter spot in the middle of the country that corresponds to the region of the capital. In contrast, Brazil shows much lower level of mobility in the northern regions and higher mobility in the East coast near the states of São Paulo and Rio de Janeiro. Overall, the continent shows higher levels in the south, especially in the Pacific coast and some heterogeneous level in the case of Islands of the Caribbean region with important contrasts between Cuba and Haiti.
In the Appendix, I report similar estimates (see Table A6) and maps (see Figure A5 and A6) at the district level, which corresponds to the finest administrative unit available in the data set. The patterns are qualitatively similar, however, given the level of dis-aggregation there are some districts with just few observations used for the estimation that produces 17 The probability of completing at least primary for those whose parents did not and the probability of not completing primary for those whose parents complete primary school, respectively. estimates that end up outside the [0,1] range.
Similarly, summary statistics of alternative estimates of intergenerational mobility that consider secondary education at the province and district levels can be found in the Appendix (see Table A8, A9, A11, and A12). They are consistent with the country-level estimates, in the sense that relative to my baseline estimates using primary education, they show lower levels of upward mobility, higher levels of downward mobility, smaller samples, and significant within-country variation. Notes: This table shows summary statistics for province level estimates of IGM. Upward reflects the likelihood that children, aged 14-18, whose parents have not completed primary schooling will manage to complete at least primary education. Downward reflects the likelihood that children, aged 14-18, whose parents have completed primary schooling or higher will not manage to complete primary education. "Total" shows the unweighted summary statistics across all provinces. The columns "Nmin" and "Nmean" report respectively the smallest and average sample size across provinces. Countries are sorted from the highest to the lowest average level of upward IGM across provinces (column "mean"). Provinces with less than 50 observations are omitted.

Figure 10: Upward Mobility in LAC
Notes: Upward mobility reflects the likelihood that children, aged 14-18, whose parents have not completed primary schooling will manage to complete at least primary education. Notes: Downward mobility reflects the likelihood that children, aged 14-18, whose parents completed primary schooling will not manage to complete at least primary education.

IV Correlates of Intergenerational Mobility
In this section, I explore a set of correlates of regional IGM with the aim of uncovering a set of stylized facts that help characterize its geography. A necessary caveat is that the set is relatively small given the difficulty of collecting data that is comparable for all the administrative units. 18 An additional and perhaps more important caveat is that the analysis does not provide any causal interpretation and is solely descriptive.
I run univariate regressions pooling all the countries linking IGM to geographical and initial conditions that have been discussed in previous studies on intergenerational mobility outside the continent (for example, see Alesina et al., 2021). This is done by estimating the following econometric specification: where d = [up, down], the dependent variable corresponds to the measure of upward or downward intergenerational mobility previously estimated for province/district r in country c, η d c denote country fixed effects, Z cr and β d are respectively the covariate and the coefficient of interest. The latter summarizing the linear association between intergenerational mobility and the covariate.

IV.1 Education of the old generation
First, I analyze the share of the old generation that was able to complete primary education. Alesina et al. (2021) finds this measure to be strongly associated with mobility in Africa.
This correlate in part reflects the initial outcomes at the province/district-level for parents. I compute this variable using an econometric specification similar to the one used to compute   1940 1950 1960 1970 1980 1990 (b) Downward mobility Notes: These estimates are computed by birth decade cohort of the children. mobility at regional level (see Equation 3) run country by country: where e icroyt is a dummy variable equal to 1 if the completed educational level of the old generation observed for individual i from country c region r is at least primary. Similar to before, γ b o and γ b y are birth-decade fixed effects for parents and children, and θ t a census year fixed effect. In other words, δ cr estimates the share of "parents" who complete primary by region netting out cohort and census year effects.
I find a strong positive (negative) correlation between upward (downward) mobility and literacy of the old generation (see Figure A18). This suggests the existence of a high level of inertia, confirming the findings of Alesina et al. (2021). Similar patterns are found at the country-birth cohort (see Figure 12).

IV.2 Other covariates
Given the high level of inertia, the correlation analysis of the remaining correlates is performed one by one and also partialing out the effect of the educational attainment of the old generation. The idea is to test whether any potential relationship with the covariate of interest remains after removing the effect of the covariate on "initial conditions". This is done estimating the following specification (in addition to equation 4): α up/down cr = η c + βZ cr + γW cr + cr (6) where m up/down cr corresponds to the measure of upward or downward IGM for province/district r in country c, η c denote country fixed effects, W cr is the share of literacy of the parents in region r, and Z cr and β are respectively the covariate and the coefficient of interest. Alesina et al. (2021) has shown that some geographical characteristics are also correlated to the level of intergenerational mobility. In this paper I consider distance to the capital, distance to the border, and distance to the coast. I also consider other characteristics of the districts that are proxies of the level of development at the beginning of the period of study.
These are the urban share of the population, the share of employment in agriculture, the share of employment in industry, and the share of employment in service. These last four covariates are computed restricting the sample to only individuals born before 1960.
The results are reported in Figure 13 for upward and downward mobility respectively.
Although upward mobility seems to be correlated with most of the proxies of development, the correlations become insignificant at the 5% when controlling by education of the old generation in all the cases. Only the share of employment in industry, which is positively associated with upward mobility, is statistically significant at the 10%. In the case of downward mobility, I find a significant correlation at the standard level, even conditioning on education of the old generation, with the share of employment in industry and agriculture, although with opposite signs. Higher share of employment in agriculture is associated with higher downward mobility while higher share of employment in industry is associated with lower downward mobility.
In the case of geographical correlates, distance to the border and coast are not significantly correlated to either measure of mobility. This is in line with Alesina et al. (2021) in the case of the border but differ relative to their results for the coast. However, distance to the capital is negatively (positively) correlated to upward (downward) mobility although weakly (statistically significant at the 5% for upward and at 10% for downward mobility).

V Final Remarks
This paper examines intergenerational educational mobility for countries in Latin America and the Caribbean at a disaggregated regional level using census data spanning more than half a century. I investigate mobility in education at the bottom of the educational attainment distribution by focusing on the likelihood of completion of primary education for those whose parents did not complete the level, which can be measured with a high degree of con-fidence between ages 14 and 18. Similarly, I measure downward mobility as the likelihood of not completing primary for those whose parents were able to complete at least primary school. I find wide cross-country and within-country heterogeneity. In LAC, the distance between the most and least upwardly mobile countries is relatively close to what has been recently documented in Africa, although the least mobile countries in Africa are less mobile than any country in LAC. Similarly, the median country in LAC shows higher upward mobility than the median country in Africa. I do not find significant differences by gender, but I do document a declining trend in the mobility gap between urban and rural populations.
Within-country mobility shows a variety of patterns. For example, there are countries with higher mobility in the northern regions (e.g., Mexico), whereas others show higher mobility in the southern regions (e.g., Brazil). The level of heterogeneity within country also varies country by country with the lowest levels found in the smallest and less populated ones.
In terms of correlates within countries, the level of mobility is highly correlated to the share of primary completion of the previous generation, which suggests a high level of inertia.
In addition, upward mobility appears weakly positively correlated to the share of employment in industry and distance to the capital, whereas downward mobility is significantly correlated to the shares of employment in industry and agriculture, and only weakly correlated to distance to the capital.
Given the unbalanced nature of the data set in terms of coverage over time and across countries, further research could shed more light on potential determinants of mobility in Latin America by focusing on the analysis of particular countries with relatively high data coverage, such as Chile, Mexico, or Brazil, which makes the collection of correlates by administrative unit easier. This paper contributes to this goal by creating the estimates of mobility at a disaggregated geographical level and making them available in an online data appendix for future research.

Appendices
In this Appendix I provide details on the sample construction and some additional tables and graphs.  Table A2 reports sample size from raw data to samples restricted by age and by availability of information on education. Table A3 reports the rates of co-residency by country for different ages. Table A4 reports the rates of co-residency by country-sample for different ages.      Table A6 reports district-level estimates of intergenerational mobility. Figure A5 and A6 displays maps of mobility at the district-level for LAC. Figure A7 and A8 displays maps of mobility at the district-level for LAC using secondary education. Table A7, A8, and A9 report estimates of IGM that consider secondary education. Table A10, A11, and A12 report estimates of IGM that consider the likelihood of completing secondary education when parents completed less than primary. Figure A13 displays the negative relationship between upward and downward mobility. Figure A14 displays estimates of upward mobility by urban/rural status for selected countries. Figure A15 displays the estimates of downward mobility by urban/rural status for selected countries. Figure A16 displays estimates of upward mobility by gender for selected countries. Figure A17 displays the estimates of downward mobility by gender for selected countries. Figure A18 shows scatter plots between IGM and share of the old generation that completes at least primary education by district.  Notes: This table reports the total sample size by country-year Census, and for restricted population by age and keeping only observations with information of education for children and parents.

B Rates of co-residence
This table shows the co-residency rate by country for different ages. The co-residence rate is the total number of individuals that co-reside with at least one member of an immediately older generation in the household divided by the total number of individuals in the age group. The sample considers only individuals for whom their own educational attainment and the relationship to household head are observed.

(b) Coresidence rate by gender
Notes: Coresidence is defined as living with at least one individual of the immediately previous generation. The data in the graph is unweighted. Figure A2: Comparison of IGM with all versus coresident children Notes: The source of these estimates is Munoz and Siravegna (2021). It shows the relationship between estimates of the conditional probability of completing at least primary school for individuals whose parents did not complete primary using two data sources. One set of estimates. computed with census data, use individuals aged 21-25 that coreside with at least one parent. The second set of estimates use the equivalent five birth-cohorts of each census sample with data from Latinobarometro where individuals are asked about the educational attainment of their parents. These 72 estimates span 18 countries in Latin America.

C Schooling by cohort
In this section, I summarize the education level by country and cohort using data on individuals at least 25 years old.    Notes: This table shows summary statistics for district-level estimates of IGM. Upward reflects the likelihood that children, aged 14-18, whose parents have not completed primary schooling will manage to complete at least primary education. Downward reflects the likelihood that children, aged 14-18, whose parents have completed primary schooling or higher will not manage to complete primary education. "Total" shows the unweighted summary statistics across all districts. The columns "Nmin" and "Nmean" report respectively the smallest and average sample size across districts. Countries are sorted from the highest to the lowest average level of upward IGM across districts (column "mean"). Districts with less than 50 observations are omitted.
E District-level maps of mobility Figure A5: Upward Mobility in LAC Notes: Upward mobility reflects the likelihood that children, aged 14-18, whose parents have not completed primary schooling will manage to complete at least primary education. This graph uses provinces for St. Lucia, Jamaica, Trinidad and Tobago and Suriname that do not have a finer administrative units in the data set.
Notes: Downward mobility reflects the likelihood that children, aged 14-18, whose parents completed at least primary schooling will not manage to complete primary education. This graph uses provinces for St. Lucia, Jamaica, Trinidad and Tobago and Suriname that do not have a finer administrative units in the data set.