Causal Effect of Mandatory Elementary Education Textbooks on Literacy and Cognition of Low Socioeconomic Status Older Adults

Abstract

Background:

Previous research has suggested that early-life education can help build cognitive reserve, providing resilience against late-life cognitive decline. The research on education and late-life cognitive health, however, has focused more on developed nations and higher levels of education. This research examines the effects of an elementary school policy on literacy, education, and late-life cognitive health.

Methods:

We analyzed, through a shift-share instrumental variable approach, the causal effect on late-life cognitive health of an exogenous policy change that introduced free and mandatory textbooks in 1960 for elementary schools in Mexico. This change essentially made elementary school attendance universal. For our analysis, we used the Mexican Health and Aging Study (MHAS) and Census datasets linked to 1960 literacy rates for cohorts born between 1930 and 1965. The MHAS and Censuses include data on literacy and early-life education. The MHAS additionally includes data on late-life cognitive engagement and functioning.

Results:

We found that the textbook program increased literacy, years of education, and late-life cognitive engagement. We found improvements in cognition for those with lower early-life socioeconomic status.

Conclusions:

There is scarce literature analyzing the causal effect of literacy and early-life education policies in developing countries on late-life cognitive engagement and functioning. Our results, particularly for the most vulnerable populations, suggest that elementary educational policies in low- and middle-income countries can boost cognitive engagement and function in late life.

Introduction

Cognitive reserve can help preserve global cognitive function and memory domains in late life, thereby reducing the risk of dementia (Stern, 2002). Though cognitive reserve cannot be measured directly, education and cognitive engagement are proxies for cognitive reserve (Gómez-Soria et al., 2023; Stern, 2009). Previous research has found that education is associated with better cognitive functioning at older ages (e.g. Gonçalves et al., 2023; Prencipe et al., 1996; Stern, 2013; Stern et al., 1994). Cognitive engagement—i.e., active involvement in mentally stimulating activities that may require attention, problem-solving, memory, or reasoning—is positively associated with cognitive function (e.g. Gómez-Soria et al., 2023). Previous longitudinal studies have documented that cohorts with higher educational attainment and greater cognitive stimulation show higher baseline cognitive function (Gutierrez et al., 2024; Yang et al., 2024; Zhai et al., 2024). Moreover, socioeconomic status (SES) significantly influences early educational opportunities, which in turn affect cognitive outcomes in later life (McMaughan et al., 2020). Individuals from higher SES backgrounds have shown better cognitive function in old age relative to their lower SES counterparts, highlighting the long-term impact of early educational inequities (Gireesh et al. 2024).

Education is an endogenous variable, that is, one correlated with other, unobservable variables that may bias estimates of its effects (Behrman et al., 2014). Previous findings on the effects of education may also reflect reverse causality: i.e., higher levels of education may improve cognition, but higher levels of cognition at younger ages may lead to greater educational attainment. There is scarce research on the causal effects of education on cognition at older ages (e.g. Glymour et al., 2008; Banks and Mazzonna, 2012). Most research on this topic focuses on developed rather than developing nations. Several studies in the United States and the United Kingdom have examined the long-term effects of raising the upper age bound of compulsory schooling for youths. Glymour et al. (2008) found U.S. increases in compulsory schooling led to improved memory at older ages. Nguyen et al. (2016) found that state-level increases in compulsory schooling and school characteristics helped form a substantial protective effect of education on dementia risk. Also in the United States, Amin et al. (2025a) found that increasing education from primary to secondary levels had potentially large positive effects on cognition, though the upper bounds of such effects were smaller than what many previous studies showed for compulsory schooling laws. The effects were greater when increasing education from secondary to tertiary levels. Banks and Mazzonna (2012) found that the 1947 reform in the United Kingdom, which increased the school-leaving age from 14 to 15, had a significant positive impact on memory and executive functioning in older males, with somewhat weaker evidence for females. Barcellos et al. (2025) found that the 1972 increase in school-leaving age from 15 to 16 in England, Scotland, and Wales led to reduced incidence of Alzheimer’s Disease and Related Dementias (ADRD) and delayed its onset.

As noted, there is little research regarding the causal effects of education on cognition at older ages in developing countries. Nevertheless, Huang and Zhou (2013) found in China that completing primary school significantly improves cognition. Amin et al. (2025b), in a study of the causal effects of schooling on memory performance for middle-aged and older adults in China, Ghana, India, Mexico, Russia, and South Africa found that increasing schooling from no education to primary school had significant positive effects on memory scores in all these nations but Russia. This suggests that promoting primary education could have lasting cognitive benefits, but that the effects of additional high school or university were less consistent.

The burden of ADRD on low- and middle-income countries (LMIC) is expected to increase in the coming decades. The proportion of the population with ADRD that lives in LMIC is predicted to increase to 70.5% by 2050 (Prince et al., 2015). Given the limited research on how early-life education influences cognition among older adults with low socioeconomic status, we explore the causal effects of education on cognitive engagement and cognition among LMIC adults. In this work, we assess the effects of a policy for increasing access to education on literacy, early-life education, cognitive engagement, cognition, and likely dementia for Mexican older adults. Like other LMIC, Mexico is experiencing rapid population aging, high poverty rates amongst older adults, and high dementia prevalence (Angel et al., 2017).

We analyze the effects on cognition among older adults who benefited from a 1960 free textbook program. Our conceptual framework is based on the Cognitive Reserve Model and the Life Course Approach. We use longitudinal data from the Mexican Health and Aging Study (MHAS) between 2001 and 2021 as well as the 2000 and 2020 Mexican Censuses. These data help us quantify the effects of national childhood educational policies on late-life cognitive engagement and cognition decades later. Because the textbook program was introduced simultaneously in every state, we can leverage the pre-existing differences in literacy rates to measure how the policy helped close gaps that existed before the intervention. We do so by using a reduced form shift-share analysis in which the share corresponds to each state’s pre-existing illiteracy rate and the shift is the nationwide introduction of free and mandatory textbooks. We validate our shift-share design by demonstrating a monotonic tercile gradient across all three datasets, with states with the lowest 1960 literacy rates showing the largest post-policy gains. Our results are consistent when adding controls for linear state-specific convergence trends. We complement this binary treatment design with a continuous exposure specification, following Duflo (2001), that exploits the staggered grade-by-grade rollout of textbooks to measure 0 to 6 years of cohort-specific exposure, interacted with the state illiteracy rate. These approaches allow us to estimate the causal effects of education on cognition by level of early-life socioeconomic status, which plays a critical role in shaping educational access (Rakesh et al., 2025). By examining the effects of a policy for increasing literacy, we highlight policies that could be implemented in other LMICs.

Evolution of the Education System in Mexico

Mexican education during the Spanish Colonization was organized by the Catholic Church and focused on religious instruction and the classics, typically for the elite and with limited access for indigenous populations. Efforts to provide more inclusive and egalitarian education surged after Mexican Independence in 1821. The 1857 Constitution established that education should be secular and free, ideals reflected in the subsequent Liberal Reforms (1857–1861) and the presidency of Benito Juárez (1858–1872). The presidency of Porfirio Diaz (1876–1911) expanded public education in urban areas but not in rural areas with large indigenous populations.

The Mexican Revolution (1910–1920), which emphasized education as a means for promoting social mobility, further boosted access to education. The 1917 Constitution made education secular, compulsory, and free. José Vasconcelos, named the first Minister of Education in 1921, implemented extensive policies to reduce illiteracy, building schools, libraries, and cultural missions in rural areas, as well as a free school lunch program for students in poverty (Solana et al., 2001).

Further reforms developed as increases in young populations placed pressure on the elementary education system. The presidency of Adolfo López Mateos (1958–1964) featured an “Eleven-Year Plan” to guarantee access to elementary education. The Plan further expanded the number of elementary schools and training programs for teachers and disbursed free elementary textbooks in Spanish throughout the nation. Education for indigenous populations has historically been more marginalized (Gobierno de México et al., 2023), with bilingual Spanish/Indigenous instruction only beginning in the 1980s.

Mexican Elementary Education Policies for Cohorts Born between 1930 and 1965

The provision of free Spanish-language textbooks did much to make compulsory education de facto. The program formally launched on February 12, 1959, when the National Commission for Free Textbooks (CONALITEG) was created to produce and distribute free textbooks to all students in basic education. To estimate the number of free elementary textbooks required per grade in 1960, the Ministry of Education conducted a nationwide school census in 1959. Also in 1959, textbooks were written by prominent experts in the field and designed and printed by private printers. The distribution involved a major joint effort among institutions in Mexico; CONALITEG worked with the Ministry of National Defense and the Postal Service to distribute the free textbooks. Distribution began on January 16, 1960, and enabled students to have access to necessary learning materials. The distribution was successful nationwide, with most textbooks arriving on time before the start of the academic year (see Alvanrez de Testa, 1992; Avila and Muñoz, 1999; Barriga Villanueva, 2011; Hernández Luna, 1986; Villa Lever, 2009). Table A1 in the Online Appendix shows the grade, year, and subjects of the distributed books. As Figure 1 shows, the policy appears to have helped reduce illiteracy among persons 15 and older throughout Mexico. Illiteracy has decreased in states throughout the nation since 1960, when the textbook program began, and in subsequent decades as young students who benefited from the program reached adulthood.

Figure 1.

Mexico’s Illiteracy Rate

State-level illiteracy rates among the population aged 15 years and older in Mexico in 1960, 1970, 1980, and 1990.

Another policy that may have led to variation in the effects of the 1960s educational reforms was the standardization of the academic year. In 1966, Mexican schools that followed the “A” calendar (February to November) unified their academic year with the “B” calendar (September to June). This meant that, from 1966 onwards, classes throughout the country would meet from early September to late June. While the change was implemented to create a more uniform and consistent school calendar for all students, students who had been following calendar type “A” may have missed one to five months of instruction in 1966, the transition year (Gobierno de México et al., 2023). Table A2 and Figure A1 in the Online Appendix show the academic calendar type by state before 1966.

Conceptual Framework and Hypotheses

To assess how these academic changes may have affected the cognitive health of Mexican students decades later, we relied on the Cognitive Reserve model and the Life Course theory. According to the Cognitive Reserve Model, individuals with higher cognitive reserve exhibit better cognitive functioning. Cognitive reserve allows the brain to better cope with pathological damage, reducing the risk of dementia (Scarmeas and Stern, 2004; Stern, 2009). Cognitive skills and abilities acquired through education, occupation, and late-life cognitive engagement build cognitive reserve through the life course (Stern et al., 2019). According to the Life Course theory, gender, cultural background, childhood socioeconomic status, age, a cohort’s historical time and place, education and employment trajectories, and institutions all affect aging and health (Settersten, 2018). In particular, the life-course cumulative advantage/disadvantage hypothesis suggests that opportunities and constraints accumulate across the life course and generate disadvantage or advantage in later adulthood (e.g. Dannefer, 2020; O’Rand, 1996).

In this study, we examine a policy intervention—free textbooks—that may reduce early-life disparities in literacy and education for the most vulnerable population and its effects on late-life cognition. Higher literacy and early-life education may promote late-life cognitive engagement, which builds cognitive reserve and, ultimately, may improve late-life cognition, reducing cumulative disadvantages and health disparities at older ages. Figure 2 summarizes our conceptual framework.

Figure 2.

Conceptual Framework

The conceptual framework illustrates how improved access to educational resources (free elementary textbooks) may increase literacy and years of education, leading to greater cognitive engagement and cognitive reserve, and is also associated with better health outcomes that support better cognitive function. Gender, childhood socioeconomic status (SES), urban/rural residence, and indigenous background are considered determinants of educational opportunities.

We focus on cohorts born between 1930 and 1965, covering the last cohorts to reach school age before the free textbook program and the first cohorts to benefit from it. We accounted for gender in our conceptual framework as an ascribed characteristic that may impact access to education. Mexican women have faced multiple disadvantages and challenging contextual conditions, including labor market barriers and lack of access to healthcare, that may adversely affect health and cognitive functioning, while Mexican men have had fewer disadvantages in accessing education but better job prospects and access to healthcare, leading to better health trajectories and late-life cognition (Blinded for review; O’Rand, 1996). Other mediators that determine access to education are childhood socioeconomic status, being born in an urban or rural area, and belonging to an indigenous community.

We hypothesized that the textbook program introduced in 1960, by giving access to learning materials and standardizing content for elementary education, increased literacy and years of schooling. We also hypothesized that resulting higher levels of literacy and education improved late-life cognitive engagement and ultimately cognition at older ages. For cognition, we hypothesized that this policy would increase the likelihood of normal cognitive function while reducing the likelihood of cognitive impairment and dementia. Finally, we hypothesized that the effect of the policy would be strongest among those most disadvantaged, specifically individuals with lower childhood socioeconomic status.

Methods

Data

We use data on individuals from the 2000 and 2020 Mexican Censuses, and from the MHAS 2001–2021. These data sources capture the same cohorts at different points in time: when aged 35 to 70 in the 2000 Census, when aged 50 years old or older in MHAS, and when aged 55 to 90 in Census 2020.

For the 2000 Census (XII Censo General de Población y Vivienda), we drew on responses to the 10.6 percent household sub-sample; for the 2020 Census (XIV Censo General de Población y Vivienda), we used the extended questionnaire administered to a subsample of households. We employed Census data is distributed by INEGI through IPUMS-International (Ruggles et al., 2025). This questionnaire includes all items from the basic questionnaire and additional questions on a wide range of demographic, socioeconomic, and health characteristics, including literacy, years of education, and dwelling characteristics (INEGI, 2021).

The MHAS is a nationally representative panel study of the Mexican population 50 years or older and their spouses. Five waves of the survey are available for the years 2001, 2003, 2012, 2015, 2018, and 2021. Its questionnaire and question phrasing are comparable to those of the Health and Retirement Study (Wong et al., 2017).

Sample

We complement the MHAS analysis with microdata from the 2000 and 2020 Mexican Censuses to validate the education results at two distinct points in the life course of the cohorts of interest. The 2000 Census was a single-questionnaire enumeration of all private dwellings; we use the publicly available 10.6-percent sample of household microdata. The 2020 Census extended questionnaire was administered to roughly four million households (about 1 in 10), oversampling rural areas and population groups of policy relevance. From each source, we selected respondents born between 1930 and 1965 to ensure comparability with the MHAS sample of individuals aged 50 and older, and dropped those born between 1950 and 1953, who were only partially exposed to the free textbook policy. This yielded analytic samples of 2,308,674 respondents from the 2000 Census, observed at ages 35 to 70, and 2,115,205 respondents from the 2020 Census, observed at ages 55 to 90. Together, the two Censuses allow us to measure the impact of full exposure to the policy on educational attainment at the end of the cohorts’ working lives and again at older ages, while the MHAS provides the cognitive outcomes for the same cohorts.

MHAS waves include 71,543 person-year observations from 25,337 respondents, 50 years old or older. We analyzed cohorts born between 1930 and 1965 (N=72,552 person-year observations from 17,723 unique respondents). We excluded returned migrants from the United States who first emigrated before reaching age 12, that is, those who may have had elementary schooling in the United States (N=104 person-year observations). For our main analysis specification, we also dropped individuals who had only partial exposure to the textbook policy, that is, those born between 1950 and 1953. The final analytical sample included 50,093 person-year observations from 12,293 unique respondents. The USC Institutional Review Board approved the study protocol (Protocol approval number UP-21-00199).

Dependent Variables Censuses 2000 and 2020.

Our Census dependent variables included total years of education, having incomplete elementary education (i.e., fewer than six years of education), and literacy (i.e., ability to read and write a message in Spanish, coded ‘1’ if ‘yes’ and ‘0’ otherwise) (INEGI, 2000, 2017).

Dependent Variables MHAS 2001–2021.

Our MHAS dependent variables included total years of education, incomplete elementary education, and literacy. We also constructed a cognitive engagement index (0–4) reflecting the number of self-reported leisure activities over the previous year, including attending a training course; reading books, magazines, or newspapers; doing puzzles (e.g., crossword); or playing board games (e.g., cards, dominoes, chess). We included in our analysis an indicator of having at least two self-reported cognitive engagement activities. Proxy questionnaires for MHAS participants unable to respond did not include cognitive engagement questions. Our primary outcomes were the cognitive measures in MHAS. MHAS scales, scoring, imputation, and classification of these items have been described in detail (Michaels-Obregon et al., 2022). We examined the categorical variable that classifies respondents as (1) normal, (2) cognitive impairment, no dementia (CIND), and (3) dementia (Langa et al., 2023; Michaels-Obregon et al., 2022). We combined CIND and dementia categories into a 0/1 indicator of likely dementia (1=Dementia/CIND).

Main Covariates.

Our first main covariate was exposure to the free textbook reform in 1960. We estimated years of exposure to the reform based on the respondent’s birth year and school grade. Individuals born after 1953 had full exposure to the reform; those born earlier had partial or no exposure. We created a dichotomous variable with value 1 indicating exposure to the reform and 0 otherwise, excluding, as noted earlier, individuals born between 1950 and 1953. Our second main covariate was the 1960 adult illiteracy rate in a respondent’s state of birth, relying on restricted data where necessary for this linkage in MHAS (MHAS, 2022).

Covariates.

For our analysis of Census data, our covariates were: age, gender, state of birth, indigenous, and urban/rural location. For our MHAS analysis, our covariates were age, gender, state of birth, childhood socioeconomic status, and urban/rural location. For early-life socioeconomic status (SES), we constructed a continuous variable ranging from 0 to 6, with higher scores indicating greater childhood socioeconomic disadvantage. This score used six items reflecting life conditions before age ten: absence of a toilet in the home, frequently going to bed hungry, not regularly wearing shoes, dropping out of school (respondent or sibling) to support the family, family members sleeping in the same room used for cooking, and receiving financial assistance (Saenz et al., 2022). We classified individuals as high early-life SES if they did not report any early-life deprivations, low early-life SES if they reported one or two deprivations, and very low early-life SES if they reported three or more deprivations. Because we measured the long-run effects of change in literacy, we did not include as covariates any indicators that could be influenced by early education.

Other Relevant Covariates.

We also included indicators for whether individuals were six to 12 years of age in 1966 and, if so, their school calendar that year (1= type A; 0= type B).

Empirical Strategy

Our empirical strategy exploits the textbook program as a natural experiment to study the effects of improved access to educational resources. We sought to estimate the causal impact of increased educational attainment on cognitive engagement and cognition outcomes among older adults. Because the policy was implemented simultaneously in all Mexican states, we could not rely on standard difference-in-differences comparisons. Instead, in our main specification, we used a shift-share design, interacting an individual’s exposure to the reform with the 1960 illiteracy rate in their state of birth. This design exploited a common aggregate ‘shift’ (the timing of the national reform) interacted with a cross-sectional ‘share’ of baseline exposure to generate plausibly exogenous differences in the expected intensity of treatment between units. Identification comes from contrasts within the unit whose size is proportional to the ‘share.’ As a result, the interaction coefficient could be interpreted as an intent-to-treat effect per unit of the share. For more on this identification strategy, see Adão et al. (2019) and Goldsmith-Pinkham et al. (2020). To complement our analysis, we estimated the marginal effect of each additional year of textbook exposure following Duflo (2001) where we exploited the grade-specific rollout of textbooks in 1960 to measure the effect.

Our main design captured variation in the policy’s impact across states: while every child received free textbooks, those from states with higher baseline illiteracy rates were expected to benefit more. By providing free textbooks, the policy reduced the direct cost of education for families and likely narrowed educational disparities, especially for disadvantaged children. We hypothesized that the introduction of free textbooks had the greatest impact on children facing the most severe early-life disadvantages—those who otherwise might have remained illiterate.

Our analysis involved a reduced-form shift-share approach, using the introduction of free textbooks as an instrument, and comparing its effect relative to illiteracy rates before program introduction in states where individuals were born. Our intuition was that, although the policy was implemented nationally, it affected each state differently, with the expected impact depending on each state’s educational level prior to implementation. The reduced form model estimated in this analysis is

$$Y_{irt}={\beta }_0+{\beta }_1\left({IlliteracyRate}_{1960}\right.\mathrm{*}\left.{FreeTextbooks}_{\mathit{irt}}\right)+{\beta }_2{Age}_i+{\beta }_3{X}_{it}+{\delta }_s+{\gamma }_t+{ϵ}_{it}$$

where $Y_{irt}$ is the outcome variable, including years of education, literacy, cognitive engagement, and an indicator for normal cognition or likely dementia. ${FreeTextbooks}_{\mathit{irt}}$ indicates whether the individual had full exposure to the textbook policy, that is, throughout elementary schooling. ${IlliteracyRate}_{1960}$ is the illiteracy rate for individuals aged 15 and above in the individual’s state of birth prior to the reform. The vector $X_{it}$ is a set of sociodemographic control variables that include gender, rural residence, indigenous language use, and early-life disadvantage indicators. We also controlled for the 1966 standardization of academic schedules across Mexican states. We included birth-state fixed effects ${\delta }_s$ to control for all time-invariant differences across states that could be correlated with 1960 illiteracy and later outcomes. We included birth-year (cohort) fixed effects to control for nationwide shocks and secular trends common to all states. We estimated this model for educational outcomes using the 2000 and 2020 Mexican Censuses and MHAS to ensure the effects in both models were comparable. Because cognition outcomes were not available in the Census data, we relied exclusively on MHAS data for their analysis. We included wave-fixed effects to net out any difference in survey instruments, sampling, or survey conditions across MHAS waves. We included age-fixed effects to account for the potentially non-linear relationship between age and our outcomes (especially cognition), thereby avoiding reliance on parametric functional forms for aging.

The interaction term $\left({IlliteracyRate}_{1960}\right.\mathrm{*}\left.{FreeTextbooks}_{\mathit{irt}}\right)$ takes values between zero and one. The coefficient of interest ${\beta }_1$ therefore measures the maximum effect of the policy, e.g., for a hypothetical state with a 100 percent illiteracy rate. Because such a rate is not feasible, we scaled ${\beta }_1$ by 0.1 to analyze the effect of a higher illiteracy rate of 10 percentage points at baseline, allowing us to express our findings in more realistic terms. We additionally interpret our effects using the empirically observed interquartile range (IQR) of the 1960 state illiteracy rate across our sample (24.3 percentage points, corresponding to the difference between a state at the 25th percentile (20.0%) and a state at the 75th percentile (44.3%) of pre-policy illiteracy). The distribution of the 1960 state-of-birth illiteracy rate ranged from a minimum of 12.8% to a maximum of 60.5%, with a mean of 33.7% and a standard deviation of 14.1%. While we included only individuals with complete exposure to the policy in our main specification, analyses that included individuals with partial exposure yielded similar results, as discussed below. To understand the heterogeneous causal effects of the policy, we analyzed the data by gender and early-life SES. We corrected for multiple hypothesis testing using the Romano–Wolf correction (Romano and Wolf, 2005).

Validity of the Shift-Share Identification Strategy

Our shift-share design relies on the 1960 state-level illiteracy rates. This identifies the cross-sectional share that explains the variation in treatment intensity stemming from the implementation of the textbook program. For this design to identify causal effects, three conditions must be met. First, the 1960 illiteracy rate must be a valid measure of how much literacy in each state could improve because of the free-textbook policy. Second, post-policy cohorts in higher-illiteracy states must experience disproportionately larger literacy and education gains. Third, these differential gains must not be driven by pre-existing convergence patterns unrelated to the textbook policy. To validate these conditions, we used our three main data sources: the MHAS panel (2001–2021) and the 2000 and 2020 Mexican Censuses.

First, we examine whether states with lower baseline literacy in 1960 saw larger literacy improvements for treated cohorts. Figure 3 compares the change in state literacy rates from individuals born before and after 1950 to the 1960 literacy rate for each state. Results from all three data sources show a strong negative relationship, as states with the lowest literacy in 1960 experienced the largest increases in literacy after the universal textbook policy. This is consistent with our hypothesis that the free textbook policy had the greatest impact in states with the largest literacy gaps.

Figure 3.

Change in literacy for cohorts before and after 1950, and the 1960 literacy rate using MHAS, 2000, and 2020 Census Data

Figure 3 shows the 1960 state literacy rate and changes in literacy among cohorts born before and after 1950 using MHAS, the 2000 Mexican Census, and the 2020 Mexican Census.

Second, we divided the population from each data source by terciles based on the 1960 state literacy rate. Figure 4 shows the literacy rate by birth cohort for states with high, middle, and low literacy in 1960. We observe a similar pattern across the three sources, with cohorts born before 1950 showing a greater dispersion in literacy based on the 1960 literacy rate ranking. This gap decreases dramatically for the cohorts exposed to the free textbook policy. Although literacy rates improve across all birth cohorts born after 1950, we observe a higher rate of change in states with low and middle literacy levels in 1960 than in states with high literacy levels at that time.

Figure 4.

Average literacy rate by birth year and 1960 literacy level using MHAS, 2000, and 2020 Census Data

Figure 4 shows average literacy rates by birth cohort across states, grouped by their 1960 literacy levels (high, middle, and low), using MHAS, the 2000 and 2020 Mexican Censuses.

Next, we estimated a model that interacted policy exposure with terciles of baseline literacy in 1960 (see Online Appendix Table A3). This model allows us to identify whether the changes in literacy after the policy differ by the pre-existing heterogeneity in state literacy rates. The results show a clean monotonic rise across all three datasets, with the policy effect on illiteracy being greatest for states with low literacy in 1960 and least for those with high literacy that year.

Third, we assess whether secular convergence—the possibility that high-illiteracy states were improving before the policy—can explain our findings. We do so through multiple strategies. In the regression framework, we estimate a continuous interaction model in which the policy exposure indicator is interacted with the 1960 state-level illiteracy rate. This allows us to test whether states with higher baseline illiteracy experienced differentially larger gains for exposed cohorts. We then augment this model with a state-specific linear trend (birth year × 1960 illiteracy rate), which directly absorbs any linear convergence between high- and low-illiteracy states. If the exposure × illiteracy interaction remains significant after this adjustment, then the estimated effect cannot be attributed to secular convergence. We also test for differential pre-trends among pre-policy cohorts (born before 1950). We find neither the continuous interaction of birth year with the 1960 illiteracy rate nor tercile-specific differential trends to be statistically significant, indicating no evidence of systematic convergence before the reform. Finally, placebo tests reassigning 1960 state illiteracy rates randomly (100 draws) yield coefficients centered at zero, far from our main estimate, across all three datasets (Figure 9).

Figure 9:

Effects for Randomized Draws of 1960 Illiteracy Rate by State on Literacy using MHAS, Census 2000, and Census 2020 Data

Figure 9 shows a placebo test based on randomized assignment of 1960 state illiteracy rates. The distribution of estimated treatment effects from random draws is shown for MHAS and the 2000 and 2020 Mexican Censuses.

Marginal Effect of Textbook Exposure

Our main specification compares cohorts with full exposure to the textbook program to those with no exposure to it, excluding individuals from cohorts (born between 1950 and 1953) who were partially exposed. To complement and add robustness to our identification strategy, we exploit the grade-specific rollout of textbooks documented in Table A1. The staggered rollout by grade means that children in different birth cohorts had varying levels of exposure to free textbooks depending on the grade they were attending in 1960. We use this rollout to estimate the marginal effect on outcomes stemming from an additional year of textbook exposure.

Following Duflo (2001), who exploited similar cohort-by-region variation in exposure to Indonesia’s school construction program, we constructed a continuous exposure variable ranging from 0 to 6 years of free textbook exposure based on each cohort’s grade-specific overlap with the textbook rollout schedule. Next, we interact this variable with the 1960 state illiteracy rate to estimate the following model:

$$Y_{irt}={\beta }_0+{\beta }_1({ExposureYears}_i\mathrm{*}{IlliteracyRate}_{1960})+{\beta }_2{Age}_i+{\beta }_3{X}_{it}+{\delta }_s+{\gamma }_t+{ϵ}_{it}$$

where ${ExposureYears}_i$ ranges from 0 to 6. Consequently, the coefficient ${\beta }_1$ measures the effect of one additional year of textbook exposure in a state with a given illiteracy level. This specification includes all birth cohorts from 1930 to 1965, including the partially exposed (1950–1953), which is excluded in our main specification.

Results

Table 1 presents descriptive statistics of individuals who were and were not exposed to the free textbook policy during their elementary school years. Panels A and B show results for our analysis of Census data: individuals in both data sources who were exposed to the textbook policy had higher levels of literacy and education. Panel C presents results for analysis of MHAS data with similar results. It also shows that those who were exposed to the textbook program reported a smaller average number of early life disadvantages, higher household incomes in adulthood, higher cognitive engagement, and lower rates of likely dementia.

Table 1:

Descriptive Statistics

Indicator	No Exposure	Free Textbooks Exposure	Total Population
Panel A: Census 2000 Data
Literacy	0.77	0.90	0.86
Incomplete Elementary	0.59	0.32	0.42
Years of Schooling	4.64	7.44	6.45
Female	0.52	0.52	0.52
Observations	988,319	1,655,573	2,643,892
Panel B: Census 2020 Data
Literacy	0.78	0.91	0.87
Incomplete Elementary	0.56	0.26	0.36
Years of Schooling	4.99	8.18	7.16
Female	0.54	0.53	0.53
Observations	752,887	1,362,318	2,115,205
Panel C: MHAS Data
Literacy	0.78	0.92	0.84
Incomplete Elementary	0.65	0.30	0.50
Years of Schooling	4.07	7.33	5.52
Age	67.74	56.29	62.65
Female	0.53	0.56	0.54
Birth Year	1941	1958	1949
Rural	0.56	0.50	0.53
Household annual income (USD Power Purchase Parities 2012)	792.30	675.74	740.50
Early life disadvantages score	2.10	1.98	2.05
Cognitive engagement score (1 or more affirmative answer)	0.58	0.70	0.65
Normal cognition	0.72	0.87	0.79
Likely dementia	0.28	0.13	0.21
Proxy Respondent	0.08	0.05	0.07
Year-respondent observations	41,906	21,809	63,715

Note: Authors calculations using Mexico’s 2000 Census, 2020 Census and MHAS data

Literacy and Education

Table 2a and 2b, using MHAS and Census data, show the estimated effects of the 1960 textbook program on later adult literacy and educational attainment.

Table 2a:

Effect of the Introduction of Free and Mandatory Textbooks on Educational Outcomes

	(1)	(2)	(3)
	Literacy	Incomplete Elementary	Years of Schooling
Panel A: Total Exposure to Free Textbooks
FreeTextbooks* Illiteracy Rate 1960	0.470***	−0.439***	2.441**
Robust Standard Error	(0.061)	(0.127)	(1.001)
95% Confidence interval	(0.346, 0.594)	(−0.697, −0.181)	(0.399, 4.482)
adj. R-sq	0.152	0.307	0.322
N	60,970	60,970	60,970
Panel B: Years of Exposure to Free Textbooks
Exposure Years* Illiteracy Rate 1960	0.076***	−0.069***	0.382**
Robust Standard Error	(0.010)	(0.019)	(0.153)
95% Confidence interval	(0.056, 0.096)	(−0.106, −0.031)	(0.069, 0.694)
adj. R-sq	0.150	0.296	0.314
N	68,462	68,462	68,462
Panel C: Years of Exposure to Free Textbooks with State-Specific Linear Trends
Exposure Years* Illiteracy Rate 1960	0.052**	−0.049**	0.493*
Robust Standard Error	(0.024)	(0.023)	(0.257)
95% Confidence interval	(0.056, 0.096)	(−0.094, −0.004)	(−0.011, 0.997)
adj. R-sq	0.150	0.296	0.314
N	68,462	68,462	68,462
Age FE	Yes	Yes	Yes
State born FE	Yes	Yes	Yes
Birth Year FE	Yes	Yes	Yes
Wave FE (for MHAS)	Yes	Yes	Yes

Notes: Authors estimations using MHAS data. Robust standard errors clustered by state in parentheses. All reported p-values are adjusted using the Romano–Wolf multiple hypothesis correction.

^*p<0.10

^**p<0.05

^***p<0.01

Table 2b.

Effect of the Introduction of Free and Mandatory Textbooks on Educational Outcomes

	(1)	(2)	(3)
	Literacy	Incomplete Elementary	Years of Schooling
Panel A: Total Exposure to Free Textbooks, 2000 Census
FreeTextbooks* Illiteracy Rate 1960	0.379***	−0.169	0.524
Robust Standard Error	(0.027)	(0.115)	(0.769)
95% Confidence interval	(0.325, 0.433)	(−0.404, 0.066)	(−1.043, 2.092)
adj. R-sq	0.122	0.188	0.216
N	2,308,674	2,188,153	2,188,153
Panel B: Years of Exposure to Free Textbooks, 2000 Census
Exposure Years* Illiteracy Rate 1960	0.063***	−0.029	0.100
Robust Standard Error	(0.004)	(0.019)	(0.126)
95% Confidence interval	(0.054, 0.071)	(−0.067, 0.010)	(−0.156, 0.356)
adj. R-sq	0.117	0.177	0.206
N	2,640,822	2,502,225	2,502,225
Panel C: Total Exposure to Free Textbooks, 2020 Census
FreeTextbooks* Illiteracy Rate 1960	0.381***	−0.244**	1.289
Robust Standard Error	(0.030)	(0.117)	(0.906)
95% Confidence interval	(0.321, 0.441)	(−0.482, −0.006)	(−0.559, 3.137)
adj. R-sq	0.105	0.172	0.187
N	2,117,141	2,115,581	2,115,581
Panel D: Years of Exposure to Free Textbooks, 2020 Census
Exposure Years* Illiteracy Rate 1960	0.063***	−0.042**	0.234
Robust Standard Error	(0.005)	(0.019)	(0.147)
95% Confidence interval	(0.053, 0.073)	(−0.079, −0.005)	(−0.054, 0.522)
adj. R-sq	0.101	0.163	0.181
N	2,443,704	2,441,918	2,441,918
Age FE	Yes	Yes	Yes
State born FE	Yes	Yes	Yes
Birth Year FE	Yes	Yes	Yes

Notes: Authors estimations using Mexico’s 2000 and 2020 Census data. Robust standard errors clustered by state in parentheses. All reported p-values are adjusted using the Romano–Wolf multiple hypothesis correction. Estimates for the 1990 Census are reported in the Online Appendix.

^*p<0.10

^**p<0.05

^***p<0.01

Column (1) of Table 2a indicates that a 10-percentage-point higher baseline illiteracy rate was associated with an additional 4.7-percentage-point increase in the likelihood of being literate among MHAS respondents exposed to the full textbook program. Column (1) of Table 2b shows similar results for Census respondents, with a 3.8 percentage-point increase in 2000 and a 3.8 percentage-point increase in 2020. Using the empirically observed IQR of the 1960 state illiteracy rate (24.3 percentage points), the corresponding effect in Table 2a is an 11.4 percentage-point increase in the likelihood of being literate among MHAS respondents exposed to the full textbook program. In Table 2b, this corresponds to a 9.2 percentage-point increase for the 2000 Census, and a 9.3 percentage-point increase for the 2020 Census. The difference in the MHAS and Census indicators reflects sample composition and the ages at which effects are measured. While MHAS oversamples older and lower-SES respondents among whom literacy gains were concentrated, Census 2000 encompasses all exposed individuals aged 35 to 70, and Census 2020 encompasses them aged 55 to 90.

Panel B of Table 2a and Panels B and D of Table 2b show the estimated effects of years of exposure to free textbooks. Results indicate that an additional year of exposure, coupled with a 10-percentage-point decline in baseline illiteracy, led to a statistically significant increase in literacy of 0.76 percentage points for MHAS respondents, 0.63 percentage points for the 2000 Census respondents, and 0.63 percentage points for the 2020 Census respondents (equivalently, 1.8, 1.5, and 1.5 percentage points, respectively, at the IQR scaling). If we consider an exposure equivalent to the six years of elementary school in Mexico, the results are not statistically different from those of our total exposure to free textbooks using our shift-share identification design, which adds to the robustness of our findings.

To confirm that the effects on education outcomes due to exposure to the textbook policy are not strictly due to convergence in literacy rates before policy implementation, we present an additional specification which augments the model with state-specific linear trends that absorb any linear convergence that may have occurred. Results in Panel C of Table 2a indicate that the policy continued to have a positive effect on educational outcomes after controlling for possible secular convergence.

Column (2) of Table 2a shows that closing the baseline illiteracy gap by 10 percentage points decreased the proportion of MHAS respondents failing to complete elementary school by a statistically significant 4.4 percentage points. These effects correspond to a 6.7 percent decrease in having incomplete elementary schooling. Using the empirically observed 24.3-percentage-point interquartile range instead, this corresponds to a 10.7-percentage-point decrease in incomplete elementary schooling among MHAS respondents, or a 16.4 percent decrease. Column (3) results show that textbook policy increased years of education among respondents. We find similar results in magnitude and significance for 2020 Census respondents, but for 2000 Census respondents, the estimates have the same sign but are not statistically significant.

Overall, these results indicate that the textbook program had a substantial and lasting positive impact on literacy, and some degree of positive effect – based on year of measurement – in the probability of completing elementary schooling, and years of schooling. All literacy coefficients were in the same direction and were statistically significant, though higher for MHAS respondents.

Because the free textbook policy was universal, it raises the question of which population groups may have benefited most from it. To answer this question, we disaggregated our analysis by gender and early-life socioeconomic status in MHAS. Figure 5 and Figure 6 present our results for individuals with full exposure to the program and for those with one year of exposure to it. Results show a larger effect on literacy in both measures for women than for men. This result is consistent with estimations using the 2000 and 2020 Censuses (see Appendix Figures A2 and A3). Results for incomplete elementary education show a different trend across genders. Both identifications show a larger decrease in the likelihood of having incomplete elementary education among men than among women. We do not find significant effects by gender for the effect of the program on years of education.

Figure 5.

Education Outcomes by Population Group in MHAS

Figure 5 shows the estimated effects of full exposure to the free elementary textbook program on education outcomes in MHAS, stratified by population group. Panel A shows literacy, Panel B shows incomplete elementary education, and Panel C shows years of schooling. Estimates are presented separately by gender and childhood socioeconomic status (SES), with vertical bars indicating 95% confidence intervals.

Figure 6.

Effect of an Extra Year of Exposure on Education Outcomes by Population Group in MHAS

Figure 6 shows the estimated marginal effects of one additional year of exposure to the free elementary textbook program on education outcomes in MHAS, stratified by population group. Panel A shows literacy, Panel B shows incomplete elementary education, and Panel C shows years of schooling. Estimates are presented separately by gender and childhood socioeconomic status (SES), with vertical bars indicating 95% confidence intervals.

Because the textbook intervention targeted elementary education, we examined whether early-life socioeconomic disadvantages mediated the relationship between textbook access and educational outcomes. Using MHAS data, we differentiated the policy’s impact by early-life SES. For literacy, the effects were more pronounced for those with lower early-life SES. While results across early-life SES categories suggest that the free textbook policy improved literacy across all SES groups, the largest and most statistically robust literacy gains were among individuals from low-SES backgrounds before age 10. Early-life SES affects completion rates for elementary education. In both identification strategies, we observe an effect for those with high and low SES, but no effect for those with very low SES.

Given small sample sizes in the MHAS, we were only able to analyze effects on indigenous populations with Census 2000 and 2020 data. We found that the effects for all education outcomes were greater for non-indigenous populations. We found a positive and statistically significant effect of literacy on indigenous populations, but no effect on elementary school completion or years of education (see Appendix Figures A2 and A3).

In summary, we found that literacy effects were more pronounced for women, non-indigenous populations, and individuals with more pronounced early-life disadvantages. For years of education and incomplete elementary, we found the most progress among men, non-indigenous populations, and those with the least early-life disadvantages.

Cognition

We turn next to the effect of the textbook program on late-life cognitive outcomes. Table 3 presents our results, using MHAS data from Waves I to VI. We derived these estimates from our main reduced-form specification. Panel A, Column (1) of Table 3 shows that closing the gap in baseline illiteracy rates by 10 percentage points was associated with a 3.9-percentage-point increase in the likelihood of having higher cognitive engagement among cohorts exposed to the textbook policy. This is equivalent to a 6.8 percent increase in the likelihood of having higher late-life cognitive engagement. Using the empirically observed 24.3-percentage-point interquartile range instead, this corresponds to a 9.6-percentage-point increase in the likelihood of higher cognitive engagement, equivalent to a 16.6 percent increase. Columns (2) and (3) present the effects of the textbook policy for normal cognition and likely dementia. Although estimates were as hypothesized—positive for normal cognition, negative for likely dementia—they were not statistically significant. Still, our results for late-life cognitive engagement suggest that early-life exposure to improved educational resources helped boost mental stimulation and cognition at older ages. Panel B of Table 3 presents the results for an additional year of exposure to free textbooks. Results are similar in significance and direction to those of the total exposure. We find that an extra year of exposure to the free textbook policy for the same 10-percentage point decrease in the illiteracy gap was associated with a 0.68 percentage-point increase in cognitive engagement (a 1.7 percentage-point increase at the IQR scaling). If we extrapolate this value to the six years of elementary schooling, we obtain a point estimate very similar to that in Panel A. Results for normal cognition and likely dementia are not statistically significant.

Table 3:

Effect of the Introduction of Free and Mandatory Textbooks on Cognition Outcomes

	(1)	(2)	(3)
	Cognitive Engagement	Normal Cognition	Likely Dementia
Panel A: Total Exposure to Free Textbooks
FreeTextbooks* Illiteracy Rate 1960	0.397***	0.050	−0.050
Robust Standard Error	(0.070)	(0.044)	(0.044)
95% Confidence interval	(0.254, 0.541)	(−0.040, 0.140)	(−0.140, 0.040)
adj. R-sq	0.111	0.083	0.083
N	41,275	57,045	57,045
Panel B: Years of Exposure to Free Textbooks
Exposure Years* Illiteracy Rate 1960	0.068***	0.010	−0.010
Robust Standard Error	(0.013)	(0.007)	(0.007)
95% Confidence interval	(0.042, 0.093)	(−0.005, 0.024)	(−0.024, 0.005)
adj. R-sq	0.110	0.080	0.080
N	47,213	64,064	64,064
Age FE	Yes	Yes	Yes
State born FE	Yes	Yes	Yes
Birth Year FE	Yes	Yes	Yes

Notes: Authors estimations using MHAS data. Robust standard errors clustered by state in parentheses. All reported p-values are adjusted using the Romano–Wolf multiple hypothesis correction.

^*p<0.10

^**p<0.05

^***p<0.01

Because the textbook policy led to substantial heterogeneity in educational outcomes across population groups, we also assessed whether it had a differential impact on these population groups. Figure 7 presents the total policy’s effects disaggregated by MHAS population groups, and Figure 8 presents the results for an extra year of exposure. Panel A in both figures shows a similar pattern: cognitive engagement for those exposed to the textbook policy increased for both men and women. Persons with very low early-life SES benefited more than those with other SES levels, though those with low SES also benefited significantly. Our results for total policy exposure indicated that a 10-percentage-point increase in baseline illiteracy was associated with a 5.2 percentage-point increase in cognitive engagement among those with very low early-life SES (a 12.6 percentage-point increase at the IQR scaling). This effect differed significantly from that for those with higher SES. Panels B and C present disaggregated results for the likelihood of having normal cognition and likely dementia. There were no statistically significant results by gender.

Figure 7.

Cognition Outcomes by Population Group in MHAS Data

Figure 7 shows the estimated effects of full exposure to the free elementary textbook program on late-life cognitive outcomes in MHAS, stratified by population group. Panel A shows cognitive engagement, Panel B shows normal cognition, and Panel C shows likely dementia. Estimates are presented separately by gender and childhood socioeconomic status (SES), with vertical bars indicating 95% confidence intervals.

Figure 8.

Effect of Extra Year of Exposure on Cognition Outcomes by Population Group in MHAS

Figure 8 shows the estimated marginal effects of one additional year of exposure to the free elementary textbook program on late-life cognitive outcomes in MHAS, stratified by population group. Panel A shows cognitive engagement, Panel B shows normal cognition, and Panel C shows likely dementia. Estimates are presented separately by gender and childhood socioeconomic status (SES), with vertical bars indicating 95% confidence intervals.

To test whether there was a differential impact of the policy on cognition outcomes by early-life SES, we estimated our preferred model independently for each SES category. We found statistically significant results for those with a very low early-life SES with both identification approaches. A 10-percentage-point increase in baseline illiteracy for them was associated with a statistically significant 1.6 percentage-point decrease in likely dementia (a 3.9 percentage-point decrease at the IQR scaling).

In summary, we found that cognitive engagement effects stemming from the textbook policy were statistically significant for both men and women, and that cognition effects were more pronounced for individuals with very low early-life SES. The results suggest that the cognitive benefits of the textbook policy, particularly for maintaining normal cognition and reducing likely dementia, were statistically significant among individuals from the most disadvantaged early-life backgrounds. We also found that the textbook policy increased cognitive engagement for both men and women.

Robustness Tests

To assess the consistency of our estimates, we implemented several robustness checks that address potential sources of bias in our empirical strategy. First, given that our main specification measures the effect of the policy on individuals with full exposure to it, we assessed the effect among individuals who were only partially exposed to the policy. Second, we re-estimated all models using a balanced panel of MHAS respondents who were interviewed in every wave. Third, we re-estimated all models, including only those who never migrated from their childhood communities. These analyses sought to minimize concerns that selective attrition or migration, both of which could be correlated with early-life educational opportunities, could affect our estimates of late-life cognitive engagement or cognition. Fourth, we estimated our analysis with a narrower sample of birth cohorts to make sure that the differences between the oldest individuals before the intervention and the youngest ones after it were not driving our results. Finally, we examined whether our results could have arisen spuriously from idiosyncratic correlations between the exposure and pre-policy illiteracy rates.

Online Appendix Figure A4, Panel A shows that the effects on literacy in both the Census data and MHAS data were positive and statistically significant. This was true regardless of whether we included the partially treated individuals in our analysis (Census-PE and MHAS-PE), we used a balanced panel (MHAS-B), or we only considered individuals who were still living in their childhood communities (MHAS-S). MHAS analyses using balanced and unbalanced samples yielded similar point estimates, which is relevant for ensuring that MHAS panel attrition and mortality did not drive our estimates. Similarities in results using Census data and those using MHAS data suggest our estimates are reliable.

Panel B shows the results for incomplete elementary education. All models here, except that for individuals still living in their childhood communities, were statistically significant. Similarly, Panel C shows model results regarding years of education were statistically significant, except for those using the balanced panel and those still in their childhood communities.

Online Appendix Figure A5 shows robustness checks for our cognition outcomes. Panel A shows that the results for cognitive engagement were all statistically significant and of similar magnitude. Panels B and C, which present the effects on cognition outcomes for the total population, show results of similar magnitude but lack statistical significance for all outcomes.

In sum, we addressed concerns regarding internal migration during childhood, which could lead to misclassification of treatment intensity by replicating our main analyses but restricted to individuals who reported still living in their birth communities. The results were remarkably consistent with our baseline estimates, supporting the conclusion that migration-related misclassification did not bias our findings. The effects on literacy, elementary education completion, cognitive engagement, and the likelihood of cognitive impairment remain similar in magnitude and statistical significance when we compare MHAS (our preferred specification) to B-MHAS (balanced sample) models. This consistency suggests that attrition is not driving our findings.

We also estimated two probit models where the outcome variables were whether an individual left or remained in the sample after 2012. We used all observable characteristics and their interactions with the treatment variable. We found no significant effects for observable characteristics by treatment status on the probability of leaving the sample due to attrition or mortality. Appendix Table A4 presents our results for attrition, and Table A5 presents our results for mortality. In addition, we examined the characteristics of individuals who left the sample by treatment status. Most of those who exited were in the non-treatment group and were more likely to have proxy respondents or likely dementia. Individuals who left the MHAS panel were more likely to have not been exposed to the textbook policy and to have poorer health and lower cognitive functioning. Consequently, we believe that if attrition were to bias our estimates, the direction of that bias would indicate that our estimates are likely lower bounds of the true effects.

Next, we examined whether our findings on cognitive engagement remain robust in the face of two possible sources of selection bias: differential proxy reporting and differential mortality. Table A6, Panel A shows that policy exposure does not predict requiring a proxy interview, ruling out the most direct channel through which missing proxies could bias our estimates. Panel B compares the headline cognitive engagement estimate, which excludes proxies because their questionnaires omit the activity items, with a worst-case lower bound that assigns 0 (no engagement) to every proxy with a missing value. The coefficient falls by only about 7 percent and remains highly significant, indicating that proxy missingness does not drive the cognitive engagement effect. Panel C addresses selective mortality by collapsing to one observation per respondent and regressing a deceased indicator on the shift-share interaction, yielding a small, negative, and statistically insignificant coefficient.

As an additional robustness check, Tables A7 and A8 re-estimate the education and cognitive models using a narrower birth cohort range (1945–1960) that limits exposure to selective mortality among older respondents. Across both tables, the sign, significance, and approximate size of the policy-by-illiteracy interaction remain stable across all outcomes and under both exposure specifications, with only modest coefficient attenuation relative to the wider-cohort baseline, which aligns more with modest differences between cohorts than with a mortality-related bias. Consistent with the primary findings, cognitive engagement remains strongly significant while normal cognition and likely dementia remain non-significant, suggesting that the findings are not attributable to the inclusion of the oldest cohorts.

Finally, we examined whether our results could have arisen spuriously from idiosyncratic correlations between exposure and pre-policy illiteracy rates. We did so by reassigning the 1960 state illiteracy rates randomly across individuals’ states of birth. Figure 9 shows that when using these randomized illiteracy distributions, the estimated treatment effects are centered around zero with no systematic patterns in either the Census or MHAS samples. This confirms that the interaction of previous illiteracy in the true state-of-birth with exposure to the textbook program was essential for generating the effects we document.

Together, these robustness checks confirm that our empirical strategy credibly isolated the causal effect of the textbook policy and that the improvements in education and cognition were not artifacts of sample composition or model specification.

Discussion

Learning to read and write a message in Spanish improved late-life cognitive engagement for all population groups and for those with the steepest disadvantages that are associated with a lower likelihood of CIND and likely dementia, as the cognitive reserve model would predict for individuals with low levels of socioeconomic status (Scarmeas et al., 2001; Stern et al., 1994; Valenzuela and Sachdev, 2006). Our results on literacy and cognitive engagement are novel, as previous literature focused on years of education (e.g. Glymour et al., 2008; Banks and Mazzonna, 2012). Previous studies of Mexico have focused on years of education and found consistent evidence that, as in other countries, higher levels of education were positively associated with late-life cognitive functioning (Mejia-Arango and Gutierrez, 2011). We contribute to the literature by analyzing respondents at the lower end of educational attainment levels.

Prior studies found that cognitive engagement with leisure activities (e.g., attending a training course and playing board games) was higher for individuals with more years of education (Robertson et al., 2023; Sommerlad et al., 2023). Our work found an impact from learning to read and write on cognitive engagement. These causal effects persisted after adjusting for sociodemographic characteristics and several robustness checks related to attrition, internal migration, mortality, and whether exposure to the policy was partial or full. We also found that the signs of the coefficients of the sociodemographic variables were consistent with previous literature (e.g. Downer et al., 2016; Saenz et al., 2020; Wong and Díaz, 2007).

By exploiting a universal but differentially consequential policy, we showed that the largest and most persistent gains in literacy, education, and cognition outcomes were concentrated in some of the most disadvantaged population groups: those with the lowest early-life SES. Our results indicated that a 10-percentage-point higher pre-policy illiteracy rate in one’s state of birth increased the probability of being literate later in life by 4.7 percentage points overall (Table 2a) —equivalent to an 11.4 percentage-point increase when scaled to the empirically observed interquartile range (25th to 75th percentile) of the 1960 state illiteracy rate—, with greater gains for women (6.5 percentage points, or 15.8 percentage points at the IQR scaling) and for those from a low early-life SES (4.4 percentage points, or 10.7 percentage points at the IQR scaling) or very low early-life SES background (4.3 percentage points, or 10.4 percentage points at the IQR scaling). These literacy improvements led to higher cognitive engagement across all population groups and a higher likelihood of normal cognition alongside a lower probability of cognitive impairment for those with very low early-life SES. Such results indicate that universal policies such as the textbook program can help those with the greatest disadvantages.

Our paper also contributed an innovation in research design. Because the free-textbook policy was nationwide, we could not rely on a standard difference-in-differences design. We instead used a reduced-form shift-share strategy that interacted previous state-of-birth illiteracy rates with cohort-specific exposure to reform. Identification relied on variation in state and cohort of birth: cohorts with full exposure in states that had greater levels of illiteracy saw a greater impact than those in states with lower illiteracy rates. We employed a comprehensive set of fixed effects to ensure this variation remained credibly exogenous. These included state-of-birth fixed effects to address time-invariant differences across states, birth-year fixed effects to address nationwide changes affecting all states, survey-wave fixed effects to account for measurement differences across MHAS waves, and age-fixed effects to capture the nonlinear relationship between aging and cognition. We also controlled for the 1966 school calendar unification, which resulted in state-level differences in length of education after the introduction of the free textbook policy. Our interaction coefficient was an intent-to-treat effect per unit of baseline illiteracy, which we could interpret by scaling the coefficient by 0.1 to reflect a 10-percentage-point difference in baseline illiteracy or by scaling the coefficient at the observed IQR of 0.243, which corresponds to the difference between a state at the 25th percentile and a state at the 75th percentile of pre-policy illiteracy rate.

This design addressed the common issue of having ‘no untreated group’ in national initiatives, while also ensuring the treatment contrast was aligned with those who could benefit the most from the reform.

Our reduced-form shift-share design interacted a common national reform with cross-sectional variation in pre-policy illiteracy to generate plausibly exogenous differences in treatment intensity across cohort and state of birth for respondents. For causal interpretation, we required the national rollout of free textbooks in 1960 to be exogenous to potential long-run outcomes; the 1960 state illiteracy rate to be predetermined and conditionally exogenous after controlling for state-, cohort-, wave-, and age-fixed effects and observed covariates; and the link between intensity and effect to be approximately linear. Identification also required there to be no omitted state-specific shocks correlated with both baseline illiteracy and adult cognition that differentially move with policy exposure, and spillovers through selective migration to be limited.

The robustness checks further strengthened our conclusions. The similarity of estimates between the balanced and unbalanced MHAS panels suggests that attrition, more common among individuals with poorer health and cognition, likely attenuated our results. Similarly, placebo tests using randomly assigned 1960 illiteracy rates to the state of birth consistently yielded null results, further showing that our identification strategy captured genuine, policy-driven variation rather than spurious correlations. Using random assignment of illiteracy distributions did not reproduce the observed effects, supporting the credibility of our empirical strategy. Altogether, we found that the 1960 textbook policy generated long-term improvements in literacy and cognitive function, particularly among populations disadvantaged in early life, and that these results were robust to attrition, migration, and data-driven specifications. Our findings underscore the long-term importance of universal education policies that reduce disparities in access to learning materials.

Limitations

Our analysis has several limitations. First, the reform’s universal rollout prevented us from using an untreated counterfactual for the analysis. Our shift–share design mitigates this limitation but does not eliminate concerns that states with high illiteracy rates could have followed different post-1960 trajectories for reasons unrelated to textbooks. Second, we measured program exposure by exposure to elementary education after the reform and did not observe the intensity of treatment (e.g., use of grade-specific textbooks or classroom quality). Our analysis, therefore, likely underestimated the effects of intensity of treatment and could not address differences in instructional quality. Third, at the time of their MHAS surveys, our cohorts were still relatively young among populations at risk of dementia. Assuming that cognitive benefits accumulate with age, this suggests that our estimates may be lower bounds on long-term effects that have not materialized yet.

In addition, the MHAS did not ask cognitive engagement questions for respondents with proxies. This limits the precision of our cognitive engagement estimates. For context, we note that 74 percent of MHAS respondents with proxies are individuals who were born before and thus not exposed to the policy. This means that we are assessing cognitive engagement among healthier persons, reinforcing our suggestion that our results may be lower bounds of the true effect of early-life education.

Conclusion

Learning to read and write a message in early childhood can boost cognitive engagement at older ages and, for the most disadvantaged populations, reduce the likelihood of dementia. This suggests that investing in early-life education is crucial for LMICs seeking to improve late-life cognitive engagement that is associated with better cognition at older ages. Our findings are most relevant for developing countries where a significant proportion of the population has low basic reading and writing skills, and where basic education may have profound effects on late-life cognitive engagement and cognitive health among the most disadvantaged, among other benefits.

Highlights.

• A free school textbook program increased literacy rate and years of education

• Greater literacy helped boost late-life cognitive engagement

• Higher literacy improved cognition in those of low early-life socioeconomic status

• Improving literacy may boost cognitive health at older ages

Data Availability

Data used in this study are made available by the Mexican Health and Aging Study (MHAS). MHAS is distributed by https://www.mhasweb.org/Home/index.aspx with funding from the National Institute on Aging (grant number R01AG018016) and the Instituto Nacional de Estadística y Geografia (INEGI) in Mexico.