Creating winners and losers: date of birth, relative age in school, and outcomes in childhood and adulthood Pablo A. Peña August 2015 Abstract This article presents three contributions. First, we estimate the causal effect of relative age (i.e., how old a person is relative to her school classmates) on test scores for a sample of students born between 1992 and 2004 using a policy experiment that shifted four months the cutoff date for school eligibility in Mexico. We exploit three distinct sources of identification and obtain similar estimates. One year of relative age confers and advantage in test scores that ranges between 0.2 and 0.4 standard deviations. Second, using date of birth as a proxy for relative age in school in a sample of Mexican adults born between 1960 and 1980, we estimate relative-age effects in six labor and marriage market outcomes: college attainment, employment status, earnings, having employer-provided medical insurance, college attainment of the spouse, and number of children. We find significant relative-age effects in the six outcomes. Third, we present a theoretical model to show that, starting with the same relative-age effects in academic performance in childhood, wealth could magnify, attenuate or even reverse the sign of relative-age effects in adulthood outcomes. This “anything goes” theoretical result is useful in interpreting different estimates of relative-age effects in adulthood outcomes across countries. Altogether, the findings indicate that institutional features of educational systems create “winners” and “losers” in academic performance via relative age in school, ultimately affecting labor and marriage market outcomes in adulthood, although possibly to varying degrees depending on other conditions such as wealth. Key words: relative-age effect; school-entry age. JEL Classification: I21, J01, J13 1 1 Introduction Everywhere in the world formal education has rules that make the birthdate of a child a determinant of how old she is relative to her classmates. Elementary instruction is organized in grades that are one year long. At the same time, there are cutoff dates for school eligibility: children must be of a certain age by a specific date of the year in order to be allowed to enroll in school. The combination of birthdates spread along the calendar, cutoff dates for school eligibility, and year-long grades, mechanically produces age differences of up to one year among classmates. A well-established fact in the literature is that on average older students outperform academically their younger classmates. Studies in a number of countries have documented systematic differences in test scores favoring older students within the same grade. Those differences are usually referred to as “relative-age effects.” Panel I of Table A in the appendix presents a summary of studies on relative-age effects in academic performance.1 Most estimates suggest that a difference of approximately one year of age between classmates is associated with a difference of 0.1 to 0.5 standard deviations in tests scores. There are two potential explanations for relative age effects in academic performance. The first is that gaps in test scores are a reflection of differences in maturity that would not be observed if students were tested at the exact same age. In other words, the observed relative-age effects could be an artifact of age-at-test differences without necessarily implying real gaps in skills ceteris paribus. The second explanation is that, when exposed to the same educational experiences, younger students are less effective at learning than their older classmates. In this case, relative age differences would result in real gaps in skills. The best available evidence supports the first explanation over the second. Black, Devereux and Salvanes (2011) use IQ tests given at different dates to individuals of similar age and find that relative-age effects are mostly due to age-at-test differences. Once they control for age at test, the age of a person relative to her school classmates has a small and negative effect on test scores. Even if relative age effects were solely due to age differentials at the moment of testing, they are relevant from an economic perspective. Albeit illusory, systematic gaps in test scores among classmates could affect human capital investment decisions under incomplete information. Economic theory predicts that people with greater “ability” have an incentive to invest more in their human capital (Becker 1964). However, ability is neither entirely nor directly observable. 1 Table A should be interpreted with caution because of potential “publication bias” favoring significant results. Researchers finding no relative-age effects might decide not to attempt the publication of their results. Franco, Malhotra and Simonovits (2014) provide evidence of publication bias in the social sciences. They find that authors are less likely to write up and submit null findings. 2 2 To make decisions regarding investments in human capital, people must infer their ability from their interaction with the world. How well someone does in school, particularly in standardized test scores, is commonly seen as an indicator of ability. Students, parents, teachers, and educational authorities drawing inferences on ability from signals biased in favor of older students might turn illusory ability differentials into real differences in later outcomes that affect welfare, such as educational attainment and earnings. 3 Several studies have found relative-age effects in intermediate outcomes that shape students’ academic paths (see Panel II of Table A.) According to their results, relatively older students are less likely to be classified as having a learning disability, and they get into more demanding or higher-quality tracks. They are more likely to have intentions to attend college, and to enroll in flagship post-secondary institutions.4 To some degree, the evidence for adulthood outcomes points in the same direction. Of the seven studies shown in Panel III of Table A, six find relativeage effects in educational attainment, and four find relative-age effects on wages or earnings. Overall, the evidence indicates that relative age in school could be creating gaps that persist into adulthood. Properly accounting for relative-age effects in academic performance could increase fairness and enhance efficiency. If ability signals were corrected by relative age, identical individuals with different academic performance solely due to their relative age would not be treated differently by the school system, their parents and even themselves. At the same time, more human capital investment could be channeled to truly high-ability individuals otherwise perceived as low-ability because of their lower relative age. To properly account for relative-age effects we need unbiased estimates of the causal effect of relative age on academic performance. Such task presents two challenges. The first is the endogeneity of relative age. Parents can affect their children’s relative age by postponing their school entry—due to lack of readiness or to give them an edge. School officials can also affect relative age through grade retention of underperforming students. The second challenge is the endogeneity of birthdates. To some extent, parents can choose in what season they have children, and parents with different attributes might aim for different seasons. Thus, children’s characteristics might not be independent of their season of birth. The endogeneity of relative age 2 Following Becker (1964), “‘ability’ would be measured by the average rate of return on human capital” (p. 98), and those “who produce more human capital from a given expenditure [would be said to] have more […] ‘ability’” (p. 124). 3 An example of how biased signals early in life can turn into differentials in adult outcomes is the case of professional hockey players documented by Barnsley, Thompson, and Barnsley (1985), and made widely known by Malcolm Gladwell’s bestseller Outliers: the story of success. 4 The study of Bellari and Pellizzari (2008) on grades in college is an exception. They find negative relativeage effects. The authors argue their result might be driven by positive relative-age effects in non-cognitive skills. 3 and season of birth implies that Ordinary Least Squares estimates of relative-age effects might be biased. Some studies have resorted to institutional rules to address the endogeneity challenges and identify the causal effect of relative age on test scores. Following Bedard and Dhuey (2006), the most frequent approach is to instrument students’ relative age with their “assigned” or “expected” relative age, that is, their relative age computed under the assumption that students enroll on time are not held back by parents or retained by school officials. A second approach is to use differences in school eligibility cutoff dates across countries or local jurisdictions as a source of identification (Bedard and Dhuey 2006, Elder and Lubotsky 2009, Sprietsma 2010, Crawford, Dearden and Meghir 2010). A third approach is to use the discontinuity in relative age created by the school eligibility cutoff date as the identification source (McEwan and Shapiro 2008, and Crawford, Dearden and Greaves 2014). 5 The general validity of the identification strategies mentioned above has been questioned. In the case of using assigned relative age as an instrumental variable, a potential problem is the violation of the monotonicity assumption discussed by Barua and Lang (2009). They argue that greater assigned relative age increases actual relative age for some students but it could also reduce actual relative age for others by decreasing their likelihood of being held back, creating potential biases of unknown magnitude and direction. In the case of estimates based on the discontinuity around the cutoff date, a potential threat is the precise manipulation of birthdates around the cutoff. Shigeoka (2014) provides compelling evidence of this type of manipulation in Japan, where redshirting is not permitted. There seems to be a deliberate effort to manipulate birthdates to postpone school entry—although that does not seem the case in the studies of McEwan and Shapiro (2008) and Crawford, Dearden and Greaves (2014). Even if the estimates of relative-age effects were valid from a technical perspective, there is an issue of how general their interpretation can be. Most studies of relative-age effects in test scores focus on high-income countries. According to Table A only three countries in Africa, Asia, and Latin America have been studied: Japan, Korea, and Chile. The role of countries’ wealth in magnifying or attenuating relative-age effects is theoretically unclear. In consequence, whether the results for high-income countries can be generalized to lower-income contexts remains an open question that needs to be addressed empirically. This study presents three contributions. First, it provides estimates of relative-age effects in test scores in Tlaxcala, one of the poorest states of Mexico—its income per capita is comparable to that of Jamaica, Namibia or Thailand. We use test score data in grades 3 to 9 for students born 5 Other studies claim to use the discontinuity as source of identification but instead of a regression discontinuity design they use fixed-effect models with monthly- or quarter-level data to estimate relative-age effects. Those studies were not counted as using the discontinuity as the source of identification. 4 between 1992 and 2004, and estimate relative-age effects using a policy experiment: an unanticipated reform that shifted four months the cutoff date for school eligibility. To our knowledge, this is the first study that uses a change in the cutoff date for school eligibility to identify relative age effects. 6 We estimate relative age effects by comparing pre- and post-reform cohorts across seasons of birth using a difference-in-differences approach. At the same time, since we have exact birthdates (something rare in the literature) we estimate the impact of relative age at the unanticipated post-reform cutoff date using a regression discontinuity approach. For comparability with other studies, we also estimate relative age effects using assigned relative age as instrumental variable—as most other studies. Lastly, we compare sideby-side the estimates produced by the three distinct identification sources using the same data. Relative-age effects in academic performance among school children are relevant to the extent they make a difference in adulthood—otherwise they would be nothing but a curious fact. The second contribution of this study is the estimation of relative-age effects in six outcomes in the labor and marriage markets using a sample of Mexican adults born between 1960 and 1980. The outcomes studied are: college attainment, employment status, earnings, having employerprovided medical insurance, college attainment of the spouse, and number of children. Relative age in school is proxied with the date of birth. People who presumably were older relative to their school classmates on average attained more college education, earn more, and have more educated spouses. Relative age is associated with a higher probability of having employerprovided medical insurance among men, and with a higher probability of being occupied and having fewer children among women. The diversity in the results for adulthood outcomes found in the literature (see panel III of Table A) compels us to theorize reasons why relative-age effects in adulthood outcomes could be observed in some countries but not in others, especially given that relative-age effects in test scores are observed everywhere. The third contribution of this study is a theoretical model à la Becker to analyze the role of wealth in the persistence of relative-age effects from childhood into adulthood. The model shows that it is a priori unclear whether we should expect larger relativeage effects in adulthood in poorer countries relative to rich countries, or vice versa. Wealth can magnify, attenuate or even reverse relative-age effects in adulthood. Based on this “anything goes” result, we can expect differences in relative-age effects in adulthood across countries due to wealth disparities. The remainder of the article is organized as follows. Section 2 presents the estimates of the causal effect of relative on test scores in childhood. Section 3 shows the estimates of relative-age 6 Smith (2010) uses a policy change in the starting-school date in Canada that affected some students but not others. The change was repealed soon after, students were reassigned, and the distribution of relative age was unaffected. 5 effects in adulthood outcomes. Section 4 presents the theoretical model. The findings are summarized in Section 5. 2 Relative-age effects in childhood In this section we present the estimates of the causal effects of relative-age on test scores in grades 3 to 9 using a sample of students in Tlaxcala, Mexico. The novelty of our estimates presented lies in the use of an unanticipated change in the cut-off date for school eligibility as part of the identification strategies. Additionally, this is the first study of relative-age effects in test scores in a low-income context. 2.1 Institutional background The state of Tlaxcala is located 75 miles east from Mexico City. It has a population of 1.17 million and a land area of approximately 1,550 square miles—it is similar to Rhode Island in both respects. In 2006, Tlaxcala had an income per capita of roughly half the national average. As the rest of Mexico, Tlaxcala has a comprehensive educational system—without academic tracking. Students can enroll in preschool approximately at age three, and they start school approximately at age six. They are expected to attend three years of preescolar, six years of primaria (grades 1 to 6), three years of secundaria (grades 7 to 9), and three years of bachillerato (grades 10 to 12). In order to enroll in preschool, students in Tlaxcala must be at least three years old by December 31 of the year of enrollment. To enroll in elementary school, they must be at least six years old by December 31 of the year of enrollment. Dropout rates at each level are crucial for the appropriate interpretation of the estimates of relative-age effects in test scores for different grades. Since students’ relative age might affect the decision to drop out of school, there might be self-selection into the sample of students for whom we observe test scores. In Tlaxcala the dropout rate in primaria is 0.7%, and 97.0% of those who complete that level continue in school. In secundaria the dropout rate is 15.6%, and 99.5% of those who complete it start the next level, bachillerato, where the dropout rate is 35.8%. 7 Since we use data for grades 3 to 9, the estimates for the higher grades must be cautiously interpreted since some degree of self-selection into staying in school could be taking place. 7 Sistema Nacional de Información Estadística Educativa of Mexico’s Ministry of Education, academic year 2011-12. 6 2.1.1 Test score data We use test scores from ENLACE, a high-stakes national standardized test given between 2006 and 2013 at the end of every academic year (around the start of June) to students in grades 3 to 9 and 12.8 The test is grade-specific and was designed, administered and graded by the federal Ministry of Education. It covers three subjects: Spanish, Math, and a rotating third subject (Science, Ethics, etc.) ENLACE constitutes a virtual census of the student population in the grades covered. The Ministry of Education of Tlaxcala provided the 2009 and 2012 individual test score data for every student in the state, grades 3 to 9. The data include the Unique Population Registry Key (CURP) of every student, which contains the birthdate as stated in the birth certificate. Some birthdates are incomplete, and some others are invalid or anomalous (extremely young or old) potentially due to data-entry mistakes. Of the 162,186 students tested in 2009, 4,870 had no CURP recorded, two had invalid information, and nine were anomalous (born before 1980 or implied enrolling at age four or younger.) For the 2012 test, out of the 161,295 test takers, 2,619 had no CURP recorded, 20 were invalid, and two were anomalous. Table 1 shows the composition of the sample of analysis by year of assessment (2009 or 2012), year of birth of the students, and grade attended. The observations without CURP, or with an invalid or anomalous birthdate are grouped into the “N.A.” category. Relative to previous studies, the data analyzed here have two distinctive features. First, they cover seven contiguous grades. Therefore, we are able to observe at the same point in time students of the same age enrolled in different grades. In contrast, most studies rely on singlegrade test score data. Second, our data provide exact dates of birth, allowing the study of age differences of up to one day. Most studies in the literature rely on coarser birthdate data and measure age differences in months. We define the relative age of a student as the difference between her age (in days) and the average age for her grade in the state (in days) divided by 365.25. We standardize test scores by grade, setting the mean to zero and the standard deviation to one. Figure 1 shows relative age and test scores in the 2012 data for students born between January 1997 and December 2003, by month of birth. The top panel shows average and median relative age of test-takers. The bottom panel shows average test scores in Spanish and Math, in standardized units relative to the students in the same grade. On average, students born in January are older than their classmates, whereas those born in December are younger. The relationship is starker for median relative age. The differences in median relative age between students born in January and December reach almost a full year. 8 ENLACE is scheduled to be replaced in 2015 with a test that will not be high-stakes. 7 The bottom panel shows a pattern in test scores that mimics the relative age pattern. Those born earlier in the year on average have higher test scores than their classmates, and those born later have lower test scores. For some years of birth the differences between those born in December and those born in January exceed 0.2 standardized units. It is worth emphasizing that Figure 1 shows students grouped by date of birth regardless of grade attended. For instance, December 2001 includes all students born in that month regardless of the grade they were attending at the time of the test. In other words, the patterns in relative age and test scores in Figure 1 include redshirting and grade retention. It is clear that relative age and test scores are not independent from date of birth. On average, children born earlier in the year are older relative to their classmates and perform better in standardized tests. 2.1.2 The reform The December 31 cutoff date is relatively recent in Tlaxcala. Prior to 2003 the cutoff date for school eligibility was September 1. The shift took place in the context of a federal push to allow earlier enrollment. In 2002 federal and state authorities agreed to make the September 1 cutoff flexible for ineligible students who showed “enough maturity.” 9 In 2006 the federal law was reformed and the cutoff date was moved from September 1 and December 31. 10 Figure 2 shows relative age and test scores in the 2009 data for students born between September 1993 and December 2000, by month of birth.11 Relative age and test scores show a saw-tooth pattern similar to that in Figure 1. However, the location of the highest and lowest points in Figure 2 differs across cohorts. The shaded areas mark the months of September to December. In the first four years shown, average and median relative age are higher in the shaded areas than in the non-shaded areas. In the last four years shown, the relationship is reversed: average and median relative age are lower in the shaded areas. The reform changed the distribution of relative age. Students born in September through December became relatively younger and the rest became relatively older. The change in the pattern in relative age is matched by the pattern in tests scores. For students born prior to 1997, average test scores are higher in the shaded areas. For students born in 1997 or after, average test scores are lower in the shaded areas. In sum, the shift in the cutoff date for school eligibility from September 1 to December 31 modified the relative age of students born in the same dates of different years. Thus, the reform provides a policy experiment in which, holding constant the season of birth, relative age was 9 Agreement 312 published in Diario Oficial de la Federación, April 15, 2002. Decree published in Diario Oficial de la Federación, June 20, 2006. 11 A potential factor that in principle could have affected the 2009 test score data is the influenza epidemic in Mexico, which caused the application of the test to be postponed in two states. However, Tlaxcala was not one of those states. 10 8 exogenously modified across cohorts. We can estimate the impact of relative age on test scores contrasting pre- and post-reform cohorts across seasons of birth, using a difference-in-differences approach. Additionally, the lack of anticipation of the reform provides an appropriate setting to estimate the impact of relative age on test scores using a regression discontinuity approach around the new cutoff date. 2.1.3 Birth patterns The policy change took place years after the conception and birth of the cohorts who enrolled shortly before or shortly after the reform. The reform was in effect in 2003 and there are no records indicating it was publicly discussed before that. Students whose conception or birth could have been affected by the reform would be expected to enroll approximately seven years later, in 2010, and to be in second grade by 2012—consequently they would not be in the 2012 data. To explore whether there was any response to the reform in terms of the timing of births, we analyze the number of births by date. Figure 3 shows the distribution of students in the sample by distance in days to December 31—the new cutoff date. The top panel shows the 2009 data. The cutoff date when those students were born was September 1. There are two facts worth noting. First, there is a spike in the number of students born in the days leading up to September 1. The five birthdates with the highest number of students are August 28 to September 1. Second, there are no noticeable patterns in other parts of the year. The bottom panel of Figure 3 shows the 2012 data. For those students the cutoff date for school eligibility was September 1 when they were born, but it was December 31 when they enrolled in school. The pattern is virtually the same as in the top panel. The same spike is observed in the days leading up to September 1. More importantly, there is no noticeable spike around December 31. There is no evidence of manipulation of births in the vicinity of the new cutoff date, which is consistent with the reform not being anticipated. It is interesting to note that the spike in the number of students born in the days leading up to September 1 indicates that birthdates were manipulated aiming at an earlier enrollment. In other words, parents’ were trying to enroll children earlier than they were supposed to. This contrasts with the evidence for Japan presented by Shigeoka (2014). In the Japanese case, redshirting is not permitted and parents seem to try to increase their children’s relative age in school by manipulating the timing of births, switching them from below to above the cutoff. In the case of Tlaxcala, parents seem to deliberately try to decrease their children’s relative age in school by doing the opposite, manipulating birthdates and switching them from above to below the cutoff. Besides the precise manipulation of birthdates, another concern is self-selection into gestational seasons. Parents of different characteristics might aim for their children to be born in 9 different parts of the calendar. Buckles and Hungerman (2013) found that in the US mother characteristics vary by season of birth of the child. To explore if something similar is observed in Tlaxcala we use the National Survey of Occupation and Employment, a household survey collected by Mexico’s National Institute of Statistics and Geography. The survey records exact birthdates of all household members, among other demographic and economic variables. In five non-overlapping waves of the survey collected between 2005 and 2013 we identified 9,361 children born between July 3, 1997 and July 2, 2003 residing with their mothers in Tlaxcala. Figure 4 shows the fraction of children whose mother has some college education (over 12 years of schooling) and the average age of the mother when the child was born. Date of birth is expressed as the distance to the closest December 31. The chart also shows fitted quadratic polynomials with 95% confidence intervals. The fraction of children whose mother has some college is lower for birthdates early in the calendar year. Therefore, if anything, children likely to be older relative to their classmates seem to have less educated mothers, which would lead us to expect them to have lower—not higher—test scores. Maternal age at childbirth shows no discernible pattern. Figure 4 does not suggest a clear threat to the validity of the empirical strategies described below. 2.2 Empirical strategies The reform and the distinctive features of the data allow the separate application of three different empirical strategies. First, using 2009 and 2012 data we compare relative age and test scores between cohorts that started school right before and right after the reform with a difference-in-differences approach across seasons of birth. Second, we use a regression discontinuity approach around the new unanticipated cutoff date for school eligibility on the 2012 data. Lastly, we instrument relative age with assigned relative age using the 2012 data. The three strategies estimate different counterfactuals. Consequently, their resulting estimates might differ without necessarily being inconsistent. The difference-in-differences estimates shed light on the effect of students “trading places” in the distribution of relative age: those who would have been the older become the younger and vice versa. The regression discontinuity estimates inform us of the relative-age effects for students born virtually at the same time but placed with peers that on average are either significantly younger or significantly older. Finally, the instrumental variable estimates provide information on relative-age effects attributable to “small” age differences between students in the same grade. A comparison of the three sets of estimates is valuable per se. Most studies rely on one source of identification alone. The use of three different sources of identification on the same data sheds light on the comparability of estimates with different caveats and subject to different potential biases. 10 2.2.1 Comparison of pre- and post-reform cohorts We define the “pre-reform cohort” as the students born in 1995-96 who took ENLACE in 2009, regardless of grade, and the “post-reform cohort” as all students born in 1998-99 who took ENLACE in 2012, also regardless of grade. The pre-reform cohort was expected to start school approximately in 2001 or 2002, whereas the post-reform cohort was expected to start school approximately in 2004 or 2005. The two cohorts are presumably fully covered by the test score data (see Table 1). The measure of academic performance for the pre-reform cohort is given by the 2009 test scores, and for the post-reform cohort it is given by the 2012 test scores. All tests scores are standardized by grade and year, providing a metric of relative performance comparable for students attending different grades and born in different years. 12 Students in both cohorts were roughly the same age by the time of assessment: between 12.42 and 14.42 years. The reform opened the possibility of enrolling earlier only to children born between September 1 and December 31. Children born between January 1 and August 31 were not directly affected by the reform. In order to estimate the causal effect of relative age on test scores we compute a difference-in-differences Wald estimate. Let t = 0 denote the pre-reform cohort and t = 1 the post-reform cohort. Let b = 0 denote birthdates between January 1 and August 31, and b = 1 the birthdates between September 1 and December 31. Lastly, let y be the test score and r the relative age. The difference-in-differences Wald estimate is: {E[ y | b 0, t 1] E[ y | b 0, t 0]} {E[ y | b 1, t 1] E[ y | b 1, t 0]} {E[r | b 0, t 1] E[r | b 0, t 0]} {E[r | b 1, t 1] E[r | b 1, t 0]} (1) The numerator is the gain in test scores for students born in January 1 to August 31 relative to the gain for students born in September 1 to December 31 (i.e., the difference in test-score differences.) The denominator is the gain in relative age for students born in January 1 to August 31 relative to the gain for those born in September 1 to December 31 (i.e., the difference in relative-age differences.) The ratio of the differences in differences is our Wald estimate of the causal effect of relative age on test scores. We compute the estimate and its standard errors by bootstrap. A caveat of the use of the shift in the cutoff date as the source of identification is that the reform changed not only the distribution of relative age but also the distribution of absolute age. 12 This standardization also helps addressing “grade inflation” in the terms of Koretz (2009), Chapter 10. Since was a high-stakes exam, it is not surprising that average test scores grew in time in a significant manner (close to half a standard deviation) in 2006-2013. ENLACE 11 The average age of students in any given grade decreased after the reform.13 Teachers presumably made adjustments to cope with younger pupils, and therefore the education process most likely was not held constant. This could affect learning differently for students of different ages. Our difference-in-differences estimates should be interpreted bearing that in mind. 2.2.2 Discontinuity around the post-reform cutoff The 2012 data provide an appropriate setting for a regression discontinuity approach because seven contiguous grades (3 to 9) are tested at the same time. Thus the data cover students with birthdates only a few days apart who attend different grades at the same point in time. Figure 1 shows clear differences between test scores of students born right before and right after the December 31 cutoff. It is worth remarking that average test scores are shown by date of birth, regardless of grade. The discrete jumps observed around December 31 are net of any mechanism that parents or schools employ to cope with the cutoff date (e.g. redshirting or grade retention.) We perform both parametric and non-parametric regression discontinuity analyses comparing the relative age and the test scores of students born around December 31. For our non-parametric analysis we restrict our sample to students born in neighborhoods of ±30, ±5 and ±1 days around the cutoff. We estimate differences in relative age and test scores between students born before the cutoff and students born after the cutoff. With those differences we compute Wald estimates of the impact of relative age on test scores (see Hahn, Todd and van der Klaauw 2001). Let z = 0 denote birthdates before the cutoff and let z = 1 denote birthdates after the cutoff. The non-parametric regression-discontinuity Wald estimate is: E[ y | z 1] E[ y | z 0] E[r | z 1] E[r | z 0] (2) The numerator is the difference in test scores between students born after the cutoff and students born before the cutoff. The denominator is the difference in relative age between students born after the cutoff and students born before the cutoff. The Wald estimate in (2) and its standard errors are obtained by bootstrap. Since narrow neighborhoods around the cutoff imply fewer observations, to preserve the statistical power of our non-parametric analysis we pool together all students born around December 31, regardless of year of birth. To obtain separate estimates for every year of birth we adopt a fuzzy regressiondiscontinuity parametric approach that is analog to our non-parametric approach. Based on the 13 This is not specific to the reform in question. Every change in the distribution of relative age is accompanied by a change in the distribution of absolute age. 12 patterns in Figure 1 we assume a linear relationship between relative age and test scores. We estimate the impact of relative age on test scores instrumenting relative age with a cubic polynomial in the date of birth and a discontinuous increase at the cutoff. A caveat of the regression discontinuity estimates is their local interpretation. The effect of relative age could be different for students born in other seasons. However, at least as far as the variables shown in Figure 4, there is no indication of seasonal patterns to the extent found in the US, suggesting that self-selection into gestational seasons could be less of an issue in our sample. 2.2.3 Post-reform assigned relative age as an instrumental variable Following Bedard and Dhuey (2006) and others, we also estimate the impact of relative age on test scores using assigned relative age as an instrumental variable. Assigned relative age is defined as the number of days between the date of birth of the student and December 31 of her year of birth, divided by 365.25. Thus, assigned relative age always takes values between zero and one, whereas actual relative age can be greater than one (when there is redshirting or retention) or negative (when age-ineligible students enroll.) As in the literature, we assume a linear relation between test scores and relative age, and estimate the model separately by grade, for grades 3 to 9 in the 2012 data. A caveat of this approach is the potential violation of the monotonicity assumption discussed by Barua and Lang (2009). However, it is valuable to produce estimates comparable to other in the literature with our data and also compare them to the estimates produced with the other two strategies. 2.3 Results Table 2 shows the comparison of the pre- and post-reform cohorts. The top panel shows results for the full sample. The first set of columns shows the number of observations by cohort divided into two groups: students born in January 1 to August 31, and students born in September 1 to December 31.The composition of the cohorts by period of birth is very similar: 32.6% of the 1995-96 cohort was born in September 1 to December 31, compared to 32.1% of the 1998-99 cohort. The second set of columns shows average relative age and the standard error in parentheses. Among students born in January 1 to August 31, those in the 1995-96 cohort were on average 0.079 years younger than their classmates, whereas those in the 1998-99 cohort were on average 0.096 years older than their classmates. The change in relative age between cohorts amounts to an increase of 0.172 years. For students born in September 1 to December 31, those in the 199596 cohort were on average 0.246 years older than their classmates, whereas those in the 1998-99 cohort were on average 0.268 years younger than their classmates. In this case, the change in 13 relative age between cohorts amounts to a decrease of 0.515 years. The difference in differences is 0.687 years of relative age. In other words, in comparison to the 1995-96 cohort, students born in January 1 to August 31 in the 1998-99 cohort gained 0.687 of relative age. By symmetry, the loss for students born in September 1 to December 31 has the same magnitude. The two sets of columns at the right hand side of Table 2 show average test scores in standard deviations (σ.) In the pre-reform cohort, students born in January 1 to August 31 have an average score of -0.027σ in Spanish and -0.023σ in Math. In contrast, in the post-reform cohort, students born in January 1 to August 31 have an average score of 0.047σ in Spanish and 0.038σ in Math. Thus, for those born in January 1 to August 31 average scores increased 0.074σ in Spanish and 0.061σ in Math across cohorts. In the pre-reform cohort, students born in September 1 to December 31 have an average score of 0.064σ in Spanish and 0.062σ in Math. In contrast, in the post-reform cohort, students born in September 1 to December 31 have an average score of -0.072σ in Spanish and -0.078σ in Math. Thus, for those born in September 1 to December 31 average scores decreased 0.136σ in Spanish and 0.140σ in Math across cohorts. The difference in differences in average test scores is 0.210σ in Spanish and 0.201σ in Math. In other words, when comparing the 1995-96 and 1998-99 cohorts, average test scores grew a fifth of a standard deviation among those born in January 1 to August 31 relative to those born in September 1 to December 31. The middle and bottom panels of Table 1 show separate results for boys and girls. Although there are gender differences in levels (boys perform better in Math and girls perform better in Spanish) the changes are roughly similar in magnitude. Table 3 shows the difference-in-differences Wald estimates of the impact of relative age on test scores described in equation (1). The estimates and their standard errors were obtained by bootstrap using 1000 repetitions. The estimate of the impact of one year of relative age is 0.305σ in Spanish and 0.292σ in Math. The impact is larger in Spanish for boys (0.325σ), and it is larger in Math for girls (0.315σ.) Table 4 shows the regression discontinuity non-parametric analysis. The top panel includes boys and girls together. The first set of columns shows the number of observations born before and after the cutoff date in neighborhoods of ±30, ±5, and ±1 days around the cutoff date. A narrower neighborhood provides a finer analysis at the expense of a smaller sample size. The ±1 day neighborhood provides the finest analysis with 324 students born on December 31 (right before the cutoff) and 290 students born on January 1 (right after the cutoff.) The second set of columns in Table 4 shows average relative age for those born before the cutoff and after the cutoff, and the average difference. Those born before the cutoff are relatively younger, and those born after the cutoff are relatively older. The average difference in relative age for those born at different sides of the cutoff but only one day apart is 0.641 years. The two sets of columns at the right hand-side of Table 4 show averages for test scores. Those born 14 before the cutoff have lower average test scores than those born after the cutoff. The difference for those born at different sides of the cutoff but only one day apart is 0.253σ in Spanish and 0.192σ in Math. The middle and bottom panels of Table 4 show separate estimates for boys and girls. The sample sizes are reduced by approximately one half and standard errors are larger. The qualitative results are similar. However, for the narrower neighborhoods the differences are larger for boys than for girls. Table 5 presents the regression-discontinuity Wald estimates of the impact of relative age on test scores. The estimates and their standard errors were obtained by bootstrap with 1000 repetitions. The estimates of the impact of one year of relative age are for the neighborhoods of ±30, ±5, and ±1 days are 0348σ, 0.339σ, and 0.395σ in Spanish and 0.365σ, 0.347σ, and 0.305σ in Math, respectively. The impact estimates are similar for boys and girls. Table 6 shows parametric estimates of the discontinuities in relative age and tests scores around December 31 of the years 1997 to 2002. Each coefficient was estimated in a separate regression including students born within a ±1 year neighborhood of December 31 of the year shown. The estimates for relative age show the results of regressing relative age on a cubic polynomial in the date of birth together with a dummy variable for the discontinuity at the December 31 cutoff. The estimates range between 0.387 and 0.837 years of relative age, depending on gender and age (i.e., the year of the cutoff analyzed.) Table 6 presents two different estimates for tests scores. The “sharp” estimates are the result of regressing test scores on a cubic polynomial in the date of birth and a discontinuity at the December 31 cutoff. They are interpreted as the difference in the test score averages at the discontinuity. The “fuzzy” estimates are interpreted as the causal effect of relative age on test scores at the discontinuity. They were computed instrumenting relative age with a cubic polynomial in date of birth, and a discontinuity at the cutoff. Sharp estimates range between 0.101 and 0.305σ in Spanish, and between -0.013 and 0.34σ in Math. Fuzzy estimates range between 0.222 and 0.478σ in Spanish, and between 0.067 and 0.555σ. Only one fuzzy estimate is not statistically significant (for girls around December 31, 1997). There is no clear trend in the estimates across years, suggesting relative-age effects do not dilute in the age range analyzed (from 9 to 15 years.) Table 7 shows the instrumental variable estimates of the impact of relative age on test scores, by grade and gender. Each coefficient was estimated in a separate regression including only students in the grade shown. The estimates range between 0.160 and 0.377σ in Spanish, and between 0.094 and 0.390σ in Math, depending on the grade. Although the estimates remain significant for 9th grade, they are smaller. This could suggest some dilution with age. However, the number of students in 9th grade (16,133) is much smaller than the number of students in 8th grade (20,683). An important fraction of 9th-graders might have dropped out of school before 15 taking the 2012 test. Dropouts make the estimates across grades not entirely comparable because relative age could be a factor in the decision to abandon school. In order to compare the results of the three different empirical strategies we focus on students born around the 1998 discontinuity and those who were in 7th or 8th grade in 2012. Table 8 presents a side-by-side comparison. The largest discrepancy is found for girls in math: the regression discontinuity estimate is 0.497 and the instrumental variable estimate for 8th grade is 0.185. Besides those two estimates, the rest are reasonably similar, ranging between 0.2 and 0.4σ. These results are generally consistent with what has been found in previous studies. More importantly, this unique comparison shows that different sources of identification point at relative-age effects estimates of similar magnitudes. Weather we consider variations in relative age produced by trading places in the distribution of age holding season of birth constant, by being born shortly before instead of shortly after the cutoff, or by being born in different dates but attending the same grade, we obtain similar estimates for the advantage in test scores conferred by relative age. 3 Relative-age effects in adulthood The existence of relative-age effects in test scores is well-documented (see Panel I of Table A.) However, the relevance of those estimates would be hard to justify if we did not observe relativeage effects in adulthood outcomes as well. In this section we estimate relative-age effects in labor and marriage market outcomes using a sample of Mexican adults. To the best of our knowledge, this is the first study of relative-age effects in adulthood outcomes outside of highincome countries. 3.1 Institutional background Mexico is a middle-income country with a GDP per capita (at purchasing power parity) of roughly half of Korea’s and a third of the US’. It has a comprehensive educational system, without tracking or streaming. Like virtually any other country, for many years Mexico has had cutoff dates for school eligibility. Unfortunately, there is not much information on those cutoff dates prior to the 1990s. It is common knowledge that September 1 was the cutoff date, but it is unclear since when or to what extent it was enforced—or even known. We do not know how often the rules were bent to enroll ineligible students. For the purpose of this study we take September 1 as the cutoff date for school eligibility, but we should entertain some rule-bending as probable. An issue that arises in the study of birthdates among Mexican adults is the potential difference between actual and official birthdates—the latter is the one recorded in birth certificates. It is not unusual to meet Mexican adults who claim having been born on a different 16 date to the one recorded in their birth certificate—my own mother is an example. In the past, many people were born in private homes and the birthdate officially recorded depended entirely on what parents reported to the authorities, which could have been inaccurate. In some cases civil registry officials recorded the date of registration as the date of birth. Due to this potential issue, relative age could be proxied with error: the birthdate used for enrollment decisions (from the birth certificate) could be different from the observed birthdate (as recorded in the survey.) 3.2 Survey data Our analysis is based on the National Survey on Occupation and Employment (ENOE) collected by Mexico’s National Institute for Statistics and Geography. ENOE is a nationally representative household survey with a quarterly-rotating panel structure. A distinctive feature of ENOE is that it records the birthdates of all the members of the surveyed households as stated by the selected respondent. The sample of analysis consists of eight non-overlapping waves: 2005-I, 2006-II, 2007-III, 2008-IV, 2010-I, 2011-II, 2012-III, and 2013-IV. The sample has a total of 3,280,136 observations. Of those observations, 3,083,698 have the birthdate fully specified.14 The sample of analysis is restricted to the 828,860 individuals born between March 3, 1960 and March 2, 1980. The birthdate bounds were chosen to provide a symmetric sample around September 1. Everyone in the sample was approximately 24 to 54 years old by the time of the survey. Six adult outcomes are analyzed, and the number of valid observations in the sample varies by outcome. The first outcome is whether the person has some college education. It was defined as a dummy indicating if the person attained any education at a four-year college level or beyond. The second outcome is employment status defined as a dummy that indicates being occupied versus not occupied. The third outcome is earnings, in logarithm. The fourth outcome is a dummy indicating whether the person has employer-provided medical insurance. Since employer-provided medical insurance is mandatory, it is proxy for working in the formal sector. The fifth outcome is a dummy variable indicating whether the spouse (or the cohabitating significant other) attained some college education. The variable was constructed for individuals who live with their spouse or significant other in the same household. The sixth outcome is the number of children and it is only recorded for women. Table 9 shows the distribution of the sample by year of birth, together with the average values for the six outcomes analyzed. The fraction of adults with some college ranges between 16 and 23%. The percent occupied is rather stable and slightly over 70%. The logarithm of earnings is higher for older cohorts, which is consistent with the well-known experience wage 14 Of the 196,438 without a fully specified birthdate, 32,753 only specify year and month, 97,149 only specify year and 66,536 have no information. 17 premium. The percent with employer-provided medical insurance ranges between 59 and 64%. The percent whose spouse has some college is stable around 17%. Finally, the average number of children for females ranges between 3.34 and 1.57, steadily declining from older to younger cohorts. Figure 5 shows the composition of the sample by day of birth relative to September 1, expressed as the distance in days to the closest September 1, in 73 five-day bins. There is no apparent bunching in the number around September 1. Thus, at least in terms of self-reported birthdates, there is no pattern in the density of the sample around the cutoff date that could invalidate a regression discontinuity design. Figure 6 shows averages for the outcomes of interest by date of birth, expressed as the distance in days to the closest September 1, in 73 five-day bins. The percentage of respondents with some college shows different patterns at each side of the cutoff. The same can be said for the percent occupied, the percent whose spouse has some college, and the number of children. For the logarithm of earnings and the percent with employer-provided medical insurance no pattern is obvious. The patterns in Figure 2 are not controlled for gender and time trends. The econometric analysis described in the next section lays out two models to properly isolate relative-age effects. 3.3 Empirical strategies Since we do not observe time starting-school for the adults in the sample, we proxy their relative age in school with the distance between the date of birth and September 1. We use two approaches for the estimation of relative-age effects in adult outcomes. First, we use a regression discontinuity design around September 1. The equation to estimate is: yi [ s (i ) w(i ) t 1 k bik ] 0 1 xi 2 xi2 0ci 1ci xi 2ci xi2 i 4 (3) where yi is the outcome analyzed, δs(i) and φw(i) are state and survey-wave fixed effects, respectively, bi is birthdate in days, xi is a variable denoting the shortest distance in days from the birthday of person i to September 1 (it can be positive or negative), divided by 365, and ci is a dummy variable indicating whether the birthday falls in September 1 to March 2. The third term at the right-hand side of (3) is a fourth-degree polynomial that works a control for a smooth trend across cohorts. The terms in brackets are controls, and the remaining terms at provide the regression discontinuity design. The parameter of interest is γ0 and it is interpreted as the discontinuous effect of the September 1 cutoff for school eligibility on outcome yi. The model (3) is also estimated restricting the data to a window of ±0.25 years around the cutoff date, and using a cubic (instead of quadratic) polynomial in the running variable xi. 18 In our second approach we use a month-of-birth fixed effects model. This second model provides more flexibility, recognizing at the same time that the discontinuity might not be sharp—some people were probably held back while others were likely enrolled in spite of being age-ineligible. The fixed-effects specification we estimate is: yi [ s (i ) w(i ) t 1 k bik ] 0 t 1t dt (i ) i 4 11 (4) The terms in brackets are the same controls as in model (1). dt(i) is a dummy indicating whether individual i was born in month t, and ηt is a fixed effect for the month of birth t. The omitted month is September, and t=1 indicates October and t=11 indicates August. The coefficients ηt are interpreted as differences relative to September. The model in (4) allows us comparing outcome differences around September 1 with differences around other dates in the year. We estimate model (4) separately for different five-year age-at-survey groups. 3.4 Results Table 10 shows the regression discontinuity estimates, with and without covariates (survey-wave and state-of-birth fixed effects, and a smooth trend across cohorts.) Each coefficient is the result of a separate regression. The standard error and the number of observations are presented below each estimate. The discontinuity estimates are significant for four outcomes: educational attainment, earnings, the marriage market and fertility. At the discontinuity, the fraction of adults with some college increases 1.44 percentage points (the average is 18.3%), earnings are 2.4% higher, the fraction of adults who have a spouse with some college increases 1.1 percentage points (the average is 17.1%), and the average number of children decreases in 0.04 (the average is 2.51). The middle and bottom panels show separate results for males and females. The main differences are in the fraction occupied and the fraction with employer-provided medical insurance. For females, the discontinuity implies an increase of one percentage point in the fraction occupied. For males the estimate is not significant. In terms of employer-provided medical insurance, for males the fraction increases in 0.9 percentage points but is only marginally significant, and for females it is not significant. When higher-order polynomials are used, the significance decreases. The second row within each panel shows the results using a cubic polynomial in the running variable. The estimates remain significant only for the fraction with some college, and when split by gender, the significance only remains for females. The third row within each panel shows a quadratic specification with a shortened window of ±0.25 years around September 1. The overall estimates remain significant for the fraction with some college and for earnings. However, when split by gender, the results for males are not 19 significant. For females, the results are significant for the fraction with some college, earnings, and percent with employer-provided medical insurance. The loss in magnitude and significance using a higher order polynomial or a smaller window around the cutoff date is a result of the data not showing a sharp change on September 1. That is consistent with an environment where some people were bending the eligibility rule, or where the date reported in the survey and the date used for school eligibility might be different. Table 11 shows the estimates of the fixed-effects model. Each column within a panel is a separate regression. The standard error is reported below each estimate. The coefficients for each month are reported relative to September (the omitted category.) Estimates with and without covariates (dummies for state of birth and survey wave) are presented for each of the outcomes. Figure 7 provides a graphic summary of the results for the specifications with covariates. The solid markers indicate significance at 95% confidence. There are some noteworthy results in Figure 7. First, for the first five outcomes the only significant estimates are negative—there are no positive and significant estimates. In the case of the number of children, the only significant estimates are positive. In other words, relative to people born in any other month, those born in September on average are more likely to have attained some college education, have higher earnings, are more likely to be occupied, are more likely to have employer-provided medical insurance, have more educated spouses, and have fewer children. Second, there is a downward trend from October to August in the first five outcomes, and an upward trend for number of children, although with varying degrees. The clearest case is for the percent with some college. Third, the largest difference in absolute value between estimates for adjacent months is observed between August and September—with the exception the percent occupied among males, and the percent with employer-provided medical insurance among females or overall. In other words, the greatest variation across adjacent months is observed around September 1. An issue treated in the literature when studying earnings is substitution along the lifecycle (see Black et al. 2011, and Frederickson et al. 2013). People who attain more education also postpone their earnings and forgo an experience premium. Thus, comparisons that do not take into account timing might be misleading. A similar argument can be made about fertility. Females attaining more education might postpone childbearing. To explore if the estimates in Table 11 are the result of timing, several fixed-effects models are estimated separately for five categories of age-at-survey. Since the sample includes different waves of the survey between 2005 and 2013, age at the moment of the survey and year of birth are not collinear—people of the same cohort are observed at different ages. Table 12 shows the estimates of the difference between August and September for five-year categories of age-at-survey: 25-29, 30-34, 35-39, 40-44 and 45-53. The model is the same as in 20 Table 11, but only the coefficient on the dummy for the month of August is shown. Each coefficient is estimated in a separate regression with dummies for state of birth and survey wave. Below each estimate are the standard error and the number of observations in the regression. There is variation across age groups. The estimates for the percent with college are all negative and in the range from -2.2 to -0.4%, but not all are significant. For males, the significant estimates are observed at ages 30-34 and 35-39, whereas for females they are observed at ages 35-39, 40-44, and 45-53. In the case of the percent occupied, some point estimates are positive and others are negative, and the only significant estimate is for ages 40-44 and it is positive. For females they point estimates are all negative and two are significant: for ages 40-44 and 45-53. In terms of the logarithm of earnings, all but one estimate (for males 35-39) are negative. All significant estimates are negative: ages 30-34 and 45-53 for males, and 35-39, 40-44 for females. In the case of the percent with employer-provided medical insurance, the only significant estimates are negative and observed for males at ages 35-39 and 40-44. In case of the percent with spouse with some college, all point estimates are negative, but only the estimates for females ages 30-34 and 35-39 are significant. Finally, the point estimates for number of children (females only) are all positive, and they are significant for ages 35-39 and 40-44. The main finding from the analysis across age groups is that the differences in adult outcomes observed around September 1 are not limited to young adults. For all six outcomes we observe significant and sizeable differences among adults age 35 or older. In other words, relative-age effects in adult outcomes do not seem to be just an artifact of timing. However, there is a great deal of variation across estimates which, in principle, could be due to relatively small sample sizes.15 That is what the estimates pooling males and females suggest. They have greater sample sizes and provide more similar estimates across age groups. In sum, the empirical analysis indicates that the six outcomes studied are not independent from date of birth. People born in months that presumably made them older relative to their classmates on average attain more college education, have higher earnings, and have more educated spouses. Relatively older males are more likely to have employer-provided medical insurance. Relatively older women are more likely to be occupied and on average have fewer children. Although the results of the regression discontinuity analysis are not robust to cubic or higher-order polynomials or smaller windows of analysis, the results from the month-fixed effects model indicate there is an inflection point between August and September. People look “the most different” across those two months. No other month is above September in terms of educational attainment, earnings, or spousal educational attainment. No other month is below 15 For every 10,000 people in the sample, approximately 849 are born in August and 821 in September. Those would be the observations providing the identification of the August-September differences. 21 September in terms of fertility. Additionally, the results do not seem to be driven by timing— relative-age effects are observed well pass age 35. 4 Theoretical model of wealth and persistence of relative-age effects The previous sections show significant relative-age effects in childhood and adulthood. The results for childhood are in line with prior studies. The results for adulthood add to a handful of studies with differing findings. Of the group of countries analyzed so far, Mexico has the lowest income per capita—roughly between one third and one half of the other countries’ GDP per capita. A natural question is whether ex ante we would expect relative-age effects in adulthood to be larger or smaller in Mexico, even if all countries had the same relative-age effects in academic performance in childhood. It is easy to imagine how wealth could attenuate the persistence of relative-age effects from childhood into adulthood. Human capital investment in a poorer country might be limited to the individuals showing the best academic promise, whereas a richer country could afford a more egalitarian investment. It is equally easy to imagine wealth magnifying persistence. More resources might widen the investment gap between individuals with different levels of ability, whereas in the extreme situation with nothing to be invested, there would not be a gap. The model below formalizes the idea that, under assumptions commonly used in human capital theory, the relationship between wealth and persistence of relative-age effects from childhood into adulthood cannot be signed. In other words, the model provides an example of how “anything goes.” Let us start by assuming a household where altruistic parents decide how much to invest in the human capital of two children. The household can be interpreted as an economy where a social planner allocates human capital investments across individuals. The only difference between the two siblings is their age relative to their classmates. Sibling 1 is relatively older, whereas sibling 2 is relatively younger. Because of a greater maturity, sibling 1 performs significantly better in school than sibling 2. As a consequence, sibling 1 is perceived as having greater ability than 2—even though they have identical ability. Based on children’s perceived abilities, parents maximize: u(c0 ) v(c1 ) v(c2 ) (5) subject to the following two constraints: w c0 y1 y2 22 ci r ( y i ; a i ) where u(c0) is the utility of parental consumption c0. v(ci) is the utility of consumption of child i. The parameter δ is a measure of altruism of parents towards their children. yi is the investment in the human capital of child i, and w is parental wealth. r(yi;ai) stands for the lifetime earnings of child i when investment in human capital is yi and perceived ability is ai. Utility functions u and v are assumed increasing and concave. Lifetime earnings are also assumed increasing and concave in investment, i.e. r′(y;a) > 0, and r″(y;a) ≤ 0. Given the same investment, higher-ability individuals are assumed to have greater lifetime earnings and a higher marginal return, i.e. if aj > ak then r(y;aj) > r(y;ak) and r′(y;aj) > r′(y;ak) for all y. It should be emphasized that the decision-maker does not observe true ability—the siblings in the model have the same. The decision is based solely on the signal given by academic performance. Parents or society make investments decision believing that a1 and a2 are their children’s abilities. However, the actual ability for both children is a. Thus, child i’s actual earnings are r(yi;a). The model can be loosely interpreted as a societal choice problem. Under that interpretation, streaming or tracking could be one of the ways in which a less egalitarian investment could materialize in an economy. Merit-based scholarships would be another. In those cases more investment resources would go to individuals perceived by society as having greater ability. On the opposite end, free tuition and affirmative action policies could be ways in which a more egalitarian investment could occur. Differences in wealth could cause societal choices to differ, and those different choices would be reflected in different educational systems and different admission and financial aid policies. In order to solve for optimal investment in the model, the first order conditions require equalizing the marginal utility produced by the investment in each child to the marginal utility of parental wealth, represented by the Lagrange multiplier λ: v(c1 )r( y1; a1 ) v(c2 )r( y2 ; a2 ) (6) Expression (6) makes explicit the trade-off between efficiency and equality in the investment decision. Efficiency implies investing in a way that equalizes the marginal returns on human capital—when they are equal there is nothing to be gained in terms of earnings by reallocating investment. Equality implies allocating the investment so that no gains in the sum of children’s utility can result from reallocating resources—they are equally happy at the margin. The efficiency-equality trade-off is not constant across different levels of wealth. However, 23 expression (6) does not provide an indication of whether the gap in optimal investments across siblings widens or narrows with greater parental wealth. It is unclear whether the difference y1-y2 grows or shrinks for smaller values of λ. The relationship can only be signed by making further assumptions on functional forms and parameter values. In order to formally establish that the relationship between wealth and persistence can go either way, we can assume specific functional forms and parameters, solve equation (6) for different values of λ (i.e. different levels of wealth), and then measure the gap in optimal investments y1-y2 and the ratio of earnings r(y1;a)/r(y2;a). The magnitudes of the gap in investments and the ratio of earnings are interpreted as the extent to which relative-age effects persist into adulthood. Let us assume that children’s utility function has the following form: v(ci ) ci (7) and that lifetime earnings is defined as: r ( yi ; ai ) ai yi 1 exp( ai yi ) ai (8) with 0 < γ < 1, ai ≥ 1 and yi ≥ 0. Verifying that both equations are increasing and concave is straightforward. Given those functional forms the marginal utility of investing in the human capital of child i is: 1 v(r ( yi ; ai ))r( yi ; ai ) ai yi 1 exp( ai yi ) ai 1 a exp(a y ) i i i (9) Assume the actual ability of both children is a = 1.25. However, the perceived ability of sibling 1 is a1 = 1.5, and the perceived ability of sibling 2 is a2 = 1.0. Assume also the parameter values μ = 3.2 and γ = 0.948. Based on those assumptions, it is possible to numerically find the investments that solve (9) for different levels of marginal utility of parental wealth λ. Those solutions are depicted in Figure 8. Figure 8 shows the gap in human capital investments y1-y2 that parents would choose for different levels of parental wealth. It also shows the ratio of earnings that would result from those investments, defined as r(y1;a)/r(y2;a)—notice that the ratio is computed using actual ability a for both siblings. In the horizontal axis is the marginal utility of parental wealth, λ. The order of the axis is reversed to be interpreted as a measure of parental wealth—greater parental wealth is equivalent to a lower λ. 24 Figure 8 shows a non-monotone relationship between the gap in investments and parental wealth. There are ranges where the gap in investment decreases with parental wealth, and there are also ranges where the opposite holds. The same can be said for the earnings ratio—given that siblings have the same ability it is only a reflection of the gap in investments. There are three values of λ (1.53, 1.65 and 2.34) where the investment gap is zero. For those levels of wealth the perceived difference in ability results in identical investments. Thus, relative-age effects in adulthood do not exist at those three levels. Moreover, there are two ranges of λ where the parents invest more in the sibling with lower perceived ability (when the marginal utility is below 1.53 or between 1.65 and 2.34). In those cases the relative-age effects are reversed in adulthood: the relatively older sibling has lower educational attainment and lower earnings than the relatively younger sibling. The patterns shown in Figure 8 in no way constitute a general result. They are just an illustrative example. The functional forms and the parameters used were picked precisely to exemplify that in a simple model of human capital investment “anything goes”: wealth could magnify, attenuate, or even reverse relative-age effects. Although highly stylized, the model shows that from a theoretical perspective we do not know whether we should expect relative-age effects in adult outcomes to be greater in wealthy countries or in poor countries—even when they have similar relative-age effects in childhood. Thus, the evidence on relative-age effects in adulthood from high-income countries is not necessarily informative of what happens in lower-income countries. 5 Summary of findings The results of the policy experiment that shifted four months the cutoff date for school eligibility indicate that the way formal education is structured creates academic “winners” and “losers.” Three different sources of identification (difference-in-differences comparing pre- and postreform cohorts, a regression discontinuity design around the new cutoff, and using “assigned” relative age as an instrument for relative age) provide similar estimates of the impact of relative age on academic performance. One year of relative age confers and advantage in test scores that ranges between 0.2 and 0.4 standard deviations. Such advantage does not vanish by grade 9 or age 15. In a human capital investment context, those gaps in academic performance could be costly for society on efficiency and equity grounds: investment in high-ability individuals might be discouraged solely because they were born in the “wrong” part of the calendar. This study also finds relative-age effects in labor and marriage market outcomes. People who, based on their birthdate, were presumably older relative to their school classmates on average attained more college education, earn more, and have more educated spouses. Relatively older males are more likely to have employer-provided medical insurance, and relatively older 25 women are more likely to be occupied and have fewer children. The evidence for adults is consistent with the gaps observed in childhood having lasting effects on outcomes related to welfare. The fact that some studies for other countries have found little or no relative-age effects in adulthood is not necessarily in conflict with the evidence presented here. The theoretical model shows that such differences across countries could be the result of differences in wealth—even if all countries have the same relative-age effects in childhood. Prior studies on adulthood outcomes are limited to high-income countries. Further evidence from middle- and low-income countries would help clarifying the role of wealth in the persistence of relative-age effects into adulthood. Overall, the results of this study have clear policy implications. Standardized test scores carry a lot of weight as measures of children’s potential in the opinion of parents, teachers, students, and the general public. Their relevance in the public’s view was epitomized by the November 2012 issue of the magazine Wired, which showed on its cover the top scorer in ENLACE and called her “the next Steve Jobs.” Scores from standardized tests provide parents, teachers, students, and authorities an accessible metric to draw inferences about ability. Such test scores could be adjusted to better reflect ability and improve the allocation of human capital investment. Crawford, Dearden and Greaves (2014) have suggested a policy response in the UK “to adjust nationally set and administered tests appropriately by age and to provide feedback on the basis of these adjusted scores” (p. 30). In the specific case of Mexico, some standardized tests have direct consequences on the educational outcomes and career paths of students. For instance, enrollment in public high schools in the metropolitan area of Mexico City is determined in a competitive process based on the scores obtained in standardized tests. Every year hundreds of thousands of 9th-graders take one of the two tests given by Metropolitan Commission of Public Secondary Education Institutions. As part of the process, students state their preferences over public high schools, ranking up to 20 options. Students are allocated to the school of their highest preference, provided there are slots available. Test scores define the order in which students are allocated until all the spots are filled. In this context, relatively younger students have a handicap. Some of them are missing potentially life-changing opportunities due to their relative-age. The most coveted schools are the ones that automatically grant a spot in Mexico’s National Autonomous University (UNAM) upon graduation with a mediocre GPA. UNAM is arguably the most prestigious university in the country, and it is also tuition-free. A small disadvantage in test scores could be the difference between having access to it or not. Adjusting—or at least better communicating—test scores used as signals of ability in order to avoid biases against relatively younger students is a simple policy that could enhance both 26 equity and efficiency. With appropriate adjustments equally able individuals would not be treated differently, and truly abler individuals would be easier to identify. References Barnsley, R.H., Thompson, A.H., & Barnsley, RE. (1985). Hockey success and birthdate: The relative age effect. Canadian Association for Health, Physical Education, and Recreation, 51, 23-28. Barua, R., & Lang, K. (2009). School Entry, Educational Attainment and Quarter of Birth: A Cautionary Tale of LATE. NBER Working Paper No. 15236. Becker, Gary S. (1964). Human Capital. A Theoretical and Empirical Analysis with Special Reference to Education, third edition, The University of Chicago Press. Bedard, K., & Dhuey, E. (2006). The persistence of early childhood maturity: International evidence of long-run age effects. The Quarterly Journal of Economics, 121(4), 1437–1472. Billari, F., & Pellizzari, M. (2012). The younger, the better? Age-related differences in academic performance at university. Journal of Population Economics 25(2), 697-739. Black, S. E., Devereux, P. J., & Salvanes, K. G. (2011). Too young to leave the nest? The effect of school starting age. The Review of Economics and Statistics, 93(2), 455–467. Buckles, K. S. & Hungerman, D. M. (2013). Season of Birth and Later Outcomes: Old Questions, New Answers. 95(3). 711-724. Cascio, E., & Schanzenbach, D. W. (2007). First in the Class? Age and the Education Production Function NBER Working Paper No 13663. Crawford, C., Dearden, L. & Greaves, E. (2014). The drivers of month-of-birth differences in children’s cognitive and non-cognitive skills. Journal of the Royal Statistical Society: Series A (Statistics in Society). DOI: 10.1111/rssa.12071. Crawford, C., Dearden, L. & Meghir, C. (2010). When you are born matters: the impact of date of birth on educational outcomes in England. DoQSS Working Paper No. 10-09 Datar, A. (2006). Does delaying kindergarten entrance give children a head start? Economics of Education Review, 25(1), 43–62. Dhuey, E. & Lipscomb, S. (2008). What makes a leader? Relative age and high school leadership. Economics of Education Review 27(2): 173–183. Dhuey, E. & Lipscomb, S. (2010). Disabled or young? Relative age and special education diagnoses in schools. Economics of Education Review 29(5): 857–872. Dobkin, C., & Ferreira, F. (2010). Do school entry laws affect educational attainment and labor market outcomes. Economics of Education Review, 29(1), 40–54. 27 Elder, T. E., & Lubotsky, D. H. (2009). Kindergarten entrance age and children’s achievement: Impacts of state policies, family background, and peers. The Journal of Human Resources, 44(3), 641–683. Franco A., Malhotra N., Simonovits G. (2014). Social science. Publication bias in the social sciences: unlocking the file drawer. Science 345(6203): 1502-5. Fredriksson P. & Ockert, B. (2013). Life-cycle effects of age at school start. The Economic Journal. DOI: 10.1111/ecoj.12047. Fukunaga, H., Taguri, M. & Morita. S. (2013). Relative age effect on Nobel laureates in the UK. JRSM Open, 4: 1-2. Gibbons, L., Belizán, J., Lauer, J., Betrán, A., Merialdi, M., & Althabe, F. (2010). The Global Numbers and Costs of Additionally Needed and Unnecessary Caesarean Sections Performed per Year: Overuse as a Barrier to Universal Coverage World Health Report (2010) Background Paper, No 30. Gladwell, Malcolm. (2008). Ouliers: the story of success, Little, Brown and Co. Grenet, J. (2011). Academic performance, Educational Trajectories and the Persistence of Date of Birth Effects. Evidence from France. Manuscript. Hahn, Jinyong, Petra Todd, and Wilbert van der Klaauw. 2001. “Identification and Estimation of Treatment Effects with a Regression-Discontinuity Design.” Econometrica, 69(1): 201–09. Hanushek, E. & L. Woessmann. (2006). Does Educational Tracking Affect Performance and Inequality? Differences-in-Differences Evidence across Countries. Economic Journal, 116, C63–C76. Kawaguchi, D. (2011). Actual age at school entry, educational outcomes, and earnings. Journal of the Japanese and International Economies, 25(2), 64–80. Koretz, D. (2009). Measuring Up: What Educational Testing Really Tells Us. Cambridge, MA: Harvard University Press. Lawlor, D., H. Clark, G. Ronalds, & D. Leon. (2006). Season of birth and childhood intelligence: Findings from the Aberdeen Children of the 1950s cohort study. British Journal of Educational Psychology 76(3), 481–499. McEwan, P. J., & Shapiro, J. S. (2008). The benefits of delayed primary school enrollment: Discontinuity estimates using exact birth dates. The Journal of Human Resources, 43(1), 1– 29. Mühlenweg, A., Blomeyer, D., Stichnoth, H., Laucht, M. (2012). Effects of age at school entry (ASE) on the development of non-cognitive skills: Evidence from psychometric data. Economics of Education Review 31(3): 68–76. Mühlenweg, A. M., & Puhani, P. A. (2010). The evolution of the school-entry age effect in a school tracking system. The Journal of Human Resources, 45(2), 407–438. 28 Nam, K. (2014). Until when does the effect of age on academic achievement persist? Evidence from Korean data. Economics of Education Review, 40, 106–122. Puhani, P. A., & Weber, A. M. (2007). Does the early bird catch the worm? Instrumental variable estimates of early educational effects of age of school entry in Germany. Empirical Economics, 32(2–3), 359–386. Robertson, E. (2011). The effects of quarter of birth on academic outcomes at the elementary school level. Economics of Education Review, 30(2): 300-311. Schneeweis, N., & M. Zweimüller. (2014). Early Tracking and the Misfortune of Being Young. The Scandinavian Journal of Economics 116(2), 394–428. Shigeoka, H. (2014). School Entry Cutoff Date and the Timing of Births. Available at SSRN: http://ssrn.com/abstract=2297711 or http://dx.doi.org/10.2139/ssrn.2297711 Smith, J. (2010). How valuable is the gift of time?: the factors that drive the birth date effect in education. Education Finance and Policy 5: 247–277. Sprietsma, M. (2010). Effect of relative age in the first grade of primary school on long-term scholastic results: international comparative evidence using PISA 2003. Education Economics 18(1), 1-32. Zweimüller, M. (2013). The effects of school entry laws on educational attainment and starting wages in an early tracking system. Annals of Economics and Statistics 111/112, 141-169. 29 TABLE 1—Number of students by grade and year of birth in the two tests Birth year 3 4 5 Grade 6 7 8 9 Total ENLACE 2009 1989 1990 1991 1992 1993 1994 1995 1996 1997 1998 1999 2000 2001 N.A. Total 2 24 79 361 1,986 19,861 11 582 22,906 6 29 107 441 2,311 19,949 18 484 23,345 1 5 24 155 727 3,144 19,148 27 4 22 168 829 6,333 18,084 23 506 23,737 509 25,972 1 25 133 758 7,321 14,346 24 978 23,586 1 2 22 138 673 7,009 13,189 32 2 14 120 802 6,504 12,289 21 910 21,976 912 20,664 3 16 143 970 7,337 20,254 21,546 21,569 21,772 21,843 21,962 19,879 11 4,881 162,186 ENLACE 2012 1992 1993 1994 1995 1996 1997 1998 1999 2000 2001 2002 2003 2004 N.A. Total 4 45 241 1,830 21,149 3 386 23,658 2 14 57 397 2,027 20,684 3 14 62 354 3,460 20,880 11 441 23,625 346 25,127 2 9 67 387 2,235 18,986 15 288 21,989 1 16 87 417 2,333 19,270 9 9 96 571 2,966 17,689 28 421 22,554 464 21,823 3 13 76 575 5,424 16,112 21 295 22,519 3 13 86 689 6,091 19,578 20,506 21,948 22,897 23,163 22,525 21,152 3 2,641 161,295 N.A.: the birthdate was not available, was prior to 1980 or implied enrolling in school at age four or younger. 30 TABLE 2—Difference-in-differences analysis of the impact of the reform on relative age and test scores Date of birth Observations Year of birth 1995-96 1998-99 All Jan. 1-Aug. 31 29,039 28,826 Sep. 1-Dec. 31 14,076 13,628 Relative age: mean (s.e.) Year of birth First 1995-96 1998-99 diff. Sep. 1-Dec. 31 14,461 14,477 7,055 6,862 0.093 0.172 (0.003) (0.004) -0.268 -0.515 (0.004) (0.006) 0.687 (0.007) -0.027 0.047 0.074 (0.006) (0.006) (0.008) 0.064 -0.072 -0.136 (0.008) (0.009) (0.012) 0.210 (0.015) -0.023 0.038 (0.006) (0.006) 0.062 -0.078 (0.008) (0.009) 0.061 (0.008) -0.140 (0.012) 0.201 (0.015) -0.034 (0.004) 0.279 (0.006) 0.124 0.158 (0.004) (0.006) -0.233 -0.512 (0.007) (0.009) 0.670 (0.011) -0.243 (0.008) -0.131 (0.012) -0.143 0.100 (0.008) (0.012) -0.249 -0.118 (0.012) (0.017) 0.219 (0.020) -0.057 0.016 (0.008) (0.008) 0.002 -0.105 (0.012) (0.012) 0.073 (0.012) -0.107 (0.017) 0.180 (0.021) -0.123 (0.004) 0.214 (0.005) 0.063 0.186 (0.004) (0.005) -0.304 -0.517 (0.005) (0.008) 0.703 (0.009) 0.188 0.238 0.050 (0.008) (0.008) (0.011) 0.260 0.107 -0.153 (0.012) (0.012) (0.017) 0.203 (0.020) 0.012 0.060 (0.008) (0.008) 0.124 -0.049 (0.012) (0.012) 0.049 (0.012) -0.173 (0.017) 0.222 (0.020) Second diff. Girls Jan. 1-Aug. 31 Sep. 1-Dec. 31 14,578 14,348 7,021 6,766 Second diff. Math: mean (s.e.) Year of birth First 1995-96 1998-99 diff. -0.079 (0.003) 0.246 (0.004) Second diff. Boys Jan. 1-Aug. 31 Spanish: mean (s.e.) Year of birth First 1995-96 1998-99 diff. Notes: Based on ENLACE 2009 (1995-96 cohort) and ENLACE 2012 (1998-99 cohort) in Tlaxcala. Standard errors in parentheses. Test scores were standardized at the state-level by grade and year. TABLE 3—Difference-in-differences Wald estimates of the impact of relative age on test scores All Boys Girls Spanish 0.305 (0.022) 0.325 (0.032) 0.289 (0.030) Math 0.292 (0.022) 0.268 (0.032) 0.315 (0.030) Estimates and standard errors obtained by bootstrap with 1000 repetitions. 31 TABLE4—Regression discontinuity non-parametric analysis of the impact of relative age on test scores Observations ±30 days ±5 days ±1 day All Before cutoff 10,507 1,716 324 After cutoff 10,567 1,806 290 Boys Before cutoff 5,323 856 155 After cutoff 5,241 894 147 Girls Before cutoff 5,184 860 169 After cutoff 5,326 912 143 Difference Difference Difference Relative age: mean (s.e.) ±30 days ±5 days ±1 day Spanish: mean (s.e.) ±30 days ±5 days ±1 day Math: mean (s.e.) ±30 days ±5 days ±1 day -0.245 (0.005) 0.338 (0.003) 0.583 (0.006) -0.267 (0.013) 0.373 (0.007) 0.640 (0.015) -0.273 (0.028) 0.368 (0.018) 0.641 (0.034) -0.077 (0.010) 0.126 (0.010) 0.203 (0.014) -0.095 (0.024) 0.121 (0.024) 0.217 (0.033) -0.115 (0.056) 0.137 (0.058) 0.253 (0.080) -0.086 (0.010) 0.127 (0.010) 0.213 (0.014) -0.089 -0.036 (0.024) (0.055) 0.134 0.155 (0.024) (0.061) 0.223 0.192 (0.034) (0.082) -0.220 (0.008) 0.350 (0.004) 0.57 (0.009) -0.270 (0.018) 0.384 (0.010) 0.654 (0.021) -0.290 (0.038) 0.367 (0.020) 0.657 (0.044) -0.223 (0.013) -0.024 (0.014) 0.199 (0.019) -0.230 (0.033) -0.011 (0.034) 0.219 (0.047) -0.323 (0.076) 0.055 (0.084) 0.378 (0.113) -0.135 (0.014) 0.079 (0.014) 0.214 (0.020) -0.138 -0.080 (0.034) (0.082) 0.110 0.160 (0.034) (0.092) 0.248 0.241 (0.048) (0.123) -0.271 (0.007) 0.326 (0.004) 0.597 (0.008) -0.265 (0.018) 0.362 (0.010) 0.627 (0.021) -0.257 (0.040) 0.369 (0.031) 0.626 (0.052) 0.074 (0.013) 0.274 (0.013) 0.200 (0.019) 0.038 (0.033) 0.250 (0.032) 0.212 (0.046) 0.075 (0.078) 0.222 (0.079) 0.147 (0.112) -0.034 (0.014) 0.175 (0.013) 0.209 (0.019) -0.040 0.004 (0.034) (0.074) 0.157 0.150 (0.032) (0.081) 0.198 0.146 (0.047) (0.110) Notes: Based on ENLACE 2012 in Tlaxcala, students born between 1997 and 2003. Standard errors in parentheses. Test scores were standardized at the state-level by grade and year. TABLE 5—Regression-discontinuity non-parametric Wald estimates of the impact of relative age on test scores All Boys Girls ±30 days 0.348 (0.023) 0.349 (0.024) 0.348 (0.024) Spanish ±5 days 0.339 (0.053) 0.338 (0.053) 0.341 (0.055) ±1 day 0.395 (0.131) 0.397 (0.132) 0.395 (0.129) ±30 days 0.365 (0.024) 0.365 (0.024) 0.365 (0.024) Estimates and standard errors obtained by bootstrap with 1000 repetitions. 32 Math ±5 days 0.347 (0.055) 0.348 (0.055) 0.352 (0.056) ±1 day 0.305 (0.129) 0.301 (0.134) 0.300 (0.134) TABLE 6—Regression-discontinuity parametric estimates of the impact of relative age on test scores Discontinuity on Observations Dec. 31 of: All 1997 19,979 1998 21,109 1999 22,691 2000 22,880 2001 22,927 2002 22,095 1997 10,015 1998 10,621 1999 11,405 2000 11,407 2001 11,469 2002 11,223 1997 9,964 1998 10,487 1999 11,286 2000 11,473 2001 11,457 2002 10,871 Relative age Spanish Sharp Fuzzy Math Sharp Fuzzy 0.411 (0.015) 0.719 (0.016) 0.704 (0.014) 0.500 (0.015) 0.835 (0.011) 0.598 (0.006) 0.139 (0.038) 0.267 (0.037) 0.227 (0.035) 0.195 (0.035) 0.225 (0.036) 0.246 (0.036) 0.400 (0.066) 0.353 (0.035) 0.287 (0.032) 0.343 (0.040) 0.309 (0.028) 0.440 (0.044) 0.099 (0.039) 0.284 (0.037) 0.232 (0.035) 0.192 (0.035) 0.270 (0.035) 0.312 (0.036) 0.247 (0.065) 0.359 (0.035) 0.316 (0.032) 0.330 (0.040) 0.352 (0.028) 0.526 (0.044) 0.387 (0.023) 0.715 (0.024) 0.665 (0.021) 0.510 (0.022) 0.834 (0.016) 0.577 (0.009) 0.175 (0.053) 0.223 (0.052) 0.207 (0.049) 0.226 (0.051) 0.160 (0.051) 0.247 (0.050) 0.478 (0.099) 0.376 (0.049) 0.352 (0.047) 0.423 (0.059) 0.265 (0.040) 0.462 (0.063) 0.208 (0.054) 0.240 (0.053) 0.197 (0.050) 0.182 (0.051) 0.192 (0.051) 0.324 (0.050) 0.454 (0.102) 0.223 (0.050) 0.336 (0.048) 0.351 (0.059) 0.316 (0.041) 0.555 (0.064) 0.437 (0.018) 0.724 (0.020) 0.745 (0.017) 0.488 (0.019) 0.837 (0.014) 0.623 (0.008) 0.101 (0.054) 0.305 (0.052) 0.234 (0.049) 0.169 (0.048) 0.275 (0.049) 0.230 (0.051) 0.301 (0.084) 0.326 (0.047) 0.222 (0.042) 0.268 (0.053) 0.356 (0.037) 0.407 (0.060) -0.013 (0.055) 0.326 (0.052) 0.261 (0.049) 0.203 (0.049) 0.340 (0.049) 0.292 (0.051) 0.067 (0.085) 0.497 (0.048) 0.295 (0.042) 0.309 (0.053) 0.387 (0.037) 0.486 (0.060) Boys Girls Standard errors in parentheses. Each coefficient was estimated in a separate regression including students born in the years immediately before and after December 31 of the year shown. Sharp estimates computed using birthdate as the running variable with a cubic polynomial. Fuzzy estimates were computed instrumenting relative age with a cubic polynomial in birthdate and a discrete increase at the cutoff date. 33 TABLE 7—Instrumental Variable estimates of the impact of relative age on test scores Grade All 3 4 5 6 7 8 9 0.263 (0.027) [23,269] 0.321 (0.029) [23,184] 0.297 (0.034) [24,781] 0.277 (0.034) [21,690] 0.302 (0.030) [22,029] 0.280 (0.032) [20,683] 0.177 (0.028) [16,133] Spanish Boys 0.255 (0.039) [11,874] 0.261 (0.040) [11,833] 0.284 (0.051) [12,346] 0.264 (0.049) [10,866] 0.320 (0.043) [11,127] 0.273 (0.044) [10,345] 0.193 (0.041) [7,729] Girls 0.274 (0.037) [11,395] 0.377 (0.041) [11,349] 0.311 (0.046) [12,435] 0.289 (0.045) [10,824] 0.279 (0.040) [10,901] 0.283 (0.045) [10,338] 0.160 (0.038) [8,404] All 0.316 (0.027) [23,269] 0.361 (0.029) [23,184] 0.290 (0.034) [24,781] 0.290 (0.034) [21,690] 0.340 (0.030) [22,029] 0.206 (0.032) [20,683] 0.120 (0.028) [16,133] Math Boys 0.330 (0.039) [11,874] 0.329 (0.041) [11,833] 0.260 (0.051) [12,346] 0.302 (0.050) [10,866] 0.363 (0.044) [11,127] 0.228 (0.045) [10,345] 0.148 (0.041) [7,729] Girls 0.304 (0.037) [11,395] 0.390 (0.040) [11,349] 0.318 (0.046) [12,435] 0.278 (0.045) [10,824] 0.316 (0.040) [10,901] 0.185 (0.046) [10,338] 0.094 (0.039) [8,404] Standard errors in parentheses. Number of observations in square brackets. Each coefficient was estimated in a separate regression. Relative age was instrumented with assigned relative age, computed as the number of days between birthdate and December 31 of the year of birth, divided by 365.25. TABLE 8—Comparison of estimates of the impact of relative-age on test scores Spanish Math Regression Instrumental Instrumental Regression Instrumental Instrumental Difference in discontinuity variables, 7th variables, 8th Difference in discontinuity variables, 7th variables, 8th differences (parametric) grade grade differences (parametric) grade grade All 0.305 0.353 0.302 0.280 0.292 0.359 0.340 0.206 (0.022) (0.035) (0.030) (0.032) (0.022) (0.035) (0.030) (0.032) Boys 0.325 0.376 0.320 0.273 0.268 0.223 0.363 0.228 (0.032) (0.049) (0.043) (0.044) (0.032) (0.050) (0.044) (0.045) Girls 0.289 0.326 0.279 0.283 0.315 0.497 0.316 0.185 (0.030) (0.047) (0.040) (0.045) (0.030) (0.048) (0.040) (0.046) Standard errors in parentheses. Difference-in-differences bootstrapping estimates for cohorts born in 1995-96 and 1998-99. Regression discontinuity fuzzy estimates around the December 1998 discontinuity. See Tables 3, 6 and 7. 34 TABLE 9—Composition of the sample of analysis of adulthood outcomes % with employer% with % Logarithm of % with spouse Number Year of provided medical some college occupied earnings with some college of children birth insurance 1960 17.8 [30,094] 70.4 [30,094] 8.32 [16,012] 60.1 [30,094] 16.7 [16,136] 3.34 [22,429] 1961 18.1 [33,541] 70.9 [33,541] 8.32 [17,935] 60.0 [33,541] 17.1 [17,946] 3.25 [25,085] 1962 18.1 [36,330] 72.1 [36,330] 8.33 [20,001] 59.7 [36,330] 17.5 [19,567] 3.19 [27,184] 1963 18.3 [37,462] 72.6 [37,462] 8.33 [20,696] 59.4 [37,462] 17.6 [20,330] 3.09 [28,002] 1964 17.9 [38,461] 73.2 [38,461] 8.33 [21,574] 60.0 [38,461] 17.5 [20,732] 3.01 [28,860] 1965 17.3 [39,246] 73.3 [39,246] 8.32 [22,282] 59.4 [39,246] 17.3 [21,390] 2.96 [29,319] 1966 17.0 [40,189] 73.7 [40,189] 8.31 [22,996] 59.4 [40,189] 17.1 [21,784] 2.88 [29,758] 1967 16.8 [39,608] 73.2 [39,608] 8.32 [22,748] 60.5 [39,608] 17.2 [21,643] 2.80 [29,261] 1968 16.8 [42,057] 73.8 [42,057] 8.32 [24,270] 60.1 [42,057] 17.5 [22,688] 2.71 [30,752] 1969 16.6 [42,076] 73.4 [42,076] 8.31 [24,275] 60.0 [42,076] 16.9 [22,814] 2.65 [30,657] 1970 16.6 [44,617] 73.5 [44,617] 8.30 [25,776] 59.8 [44,617] 16.8 [23,961] 2.57 [31,775] 1971 16.9 [42,450] 72.9 [42,450] 8.31 [24,245] 61.1 [42,450] 17.2 [22,854] 2.48 [30,153] 1972 16.8 [45,430] 72.9 [45,430] 8.30 [26,297] 60.3 [45,430] 17.1 [24,475] 2.39 [31,683] 1973 17.3 [45,014] 72.7 [45,014] 8.30 [25,973] 60.8 [45,014] 16.8 [24,266] 2.30 [30,774] 1974 18.2 [45,398] 72.2 [45,398] 8.29 [26,024] 61.7 [45,398] 17.3 [24,505] 2.19 [29,954] 1975 18.6 [44,697] 71.5 [44,697] 8.28 [25,423] 62.3 [44,697] 16.8 [24,214] 2.10 [28,625] 1976 19.1 [43,519] 71.3 [43,519] 8.27 [24,694] 62.6 [43,519] 17.0 [23,544] 2.00 [26,978] 1977 20.1 [43,546] 71.2 [43,546] 8.26 [24,712] 62.6 [43,546] 16.8 [23,573] 1.89 [25,817] 1978 22.3 [42,839] 71.1 [42,839] 8.28 [24,135] 63.3 [42,839] 17.4 [23,018] 1.76 [23,777] 1979 23.1 [44,224] 70.6 [44,224] 8.27 [24,758] 63.9 [44,224] 16.9 [23,642] 1.62 [22,720] 1980 23.3 [8,062] 70.3 [8,062] 8.23 [4,540] 64.2 [8,062] 16.0 [4,258] 1.57 [3,992] 1960-80 18.3 [828,860] 72.3 [828,860] 8.30 [469,366] 60.9 [828,860] 17.1 [447,340] 2.51 [567,555] Notes: Based on ENOE 2005-I, 2006-II, 2007-III, 2008-IV, 2010-I, 2011-II, 2012-III, and 2013-IV. Only includes adults born between March 3, 1960 and March 2, 1980. 35 TABLE 10—Regression discontinuity estimates of the impact of an additional year of relative-age on in adulthood outcomes Percent with some college Covariates All Quadratic Cubic Quadratic with ±0.25 Males Quadratic Cubic Quadratic with ±0.25 No Yes Percent occupied No Yes Logarithm of earnings No Yes Percent with employerprovided medical insurance No Yes Percent with spouse with some college No Yes 1.527 1.436 (0.254) (0.252) [825,837] [825,837] 0.994 0.903 (0.340) (0.337) [825,837] [825,837] 0.902 0.798 (0.363) (0.360) [413,626] [413,626] 0.450 (0.295) [825,837] 0.479 (0.395) [825,837] 0.170 (0.419) [413,626] 0.422 (0.294) [825,837] 0.413 (0.394) [825,837] 0.101 (0.418) [413,626] 0.0237 0.0239 (0.007) (0.007) [468,123] [468,123] 0.0075 0.0087 (0.010) (0.009) [468,123] [468,123] 0.0199 0.0199 (0.010) (0.010) [234,586] [234,586] 0.544 (0.322) [825,837] 0.156 (0.431) [825,837] 0.580 (0.457) [413,626] 0.488 (0.319) [825,837] 0.262 (0.427) [825,837] 0.620 (0.454) [413,626] 1.172 (0.300) [565,413] 0.499 (0.401) [565,413] 0.714 (0.428) [282,930] 1.075 (0.297) [565,413] 0.405 (0.398) [565,413] 0.611 (0.425) [282,930] 1.553 1.442 (0.392) (0.388) [379,968] [379,968] 0.836 0.736 (0.524) (0.520) [379,968] [379,968] 0.392 0.277 (0.559) (0.554) [190,997] [190,997] -0.213 (0.253) [379,968] -0.022 (0.338) [379,968] -0.293 (0.358) [190,997] -0.229 (0.252) [379,968] -0.013 (0.338) [379,968] -0.296 (0.357) [190,997] 0.0177 0.0171 (0.009) (0.008) [280,284] [280,284] -0.0003 -0.0009 (0.011) (0.011) [280,284] [280,284] 0.0031 0.0005 (0.012) (0.012) [140,979] [140,979] 0.929 (0.486) [379,968] -0.046 (0.650) [379,968] 0.104 (0.691) [190,997] 0.922 (0.480) [379,968] 0.226 (0.643) [379,968] 0.239 (0.683) [190,997] 0.944 (0.396) [274,774] 0.401 (0.530) [274,774] 0.308 (0.566) [138,021] 0.836 (0.393) [274,774] 0.297 (0.526) [274,774] 0.189 (0.562) [138,021] Females Quadratic Number of children No Yes 1.483 1.406 0.992 0.977 0.0359 0.0357 0.232 0.129 1.409 1.335 -0.046 -0.041 (0.330) (0.328) (0.446) (0.444) (0.012) (0.012) (0.410) (0.407) (0.444) (0.440) (0.015) (0.015) [445,869] [445,869] [445,869] [445,869] [187,839] [187,839] [445,869] [445,869] [290,639] [290,639] [445,850] [445,850] Cubic 1.061 0.990 0.483 0.409 0.0151 0.0178 0.561 0.583 0.643 0.546 -0.020 -0.016 (0.442) (0.439) (0.597) (0.595) (0.016) (0.016) (0.549) (0.546) (0.594) (0.589) (0.020) (0.019) [445,869] [445,869] [445,869] [445,869] [187,839] [187,839] [445,869] [445,869] [290,639] [290,639] [445,850] [445,850] Quadratic with 1.287 1.199 0.186 0.128 0.0400 0.0424 1.193 1.172 1.124 1.034 -0.030 -0.023 ±0.25 (0.472) (0.468) (0.634) (0.632) (0.017) (0.017) (0.582) (0.578) (0.635) (0.629) (0.021) (0.021) [222,629] [222,629] [222,629] [222,629] [93,607] [93,607] [222,629] [222,629] [144,909] [144,909] [222,622] [222,622] Notes: Based on ENOE 2005-I, 2006-II, 2007-III, 2008-IV, 2010-I, 2011-II, 2012-III, and 2013-IV. Only includes adults born between March 3, 1960 and March 2, 1980. Standard errors in parentheses. Number of observations in square brackets. Each coefficient is the result of a separate regression. The covariates are dummies for state of birth and survey wave. 36 TABLE 11—Fixed-effects estimates the impact of relative-age on adulthood outcomes Percent with some college Covariates All Constant October November December January February March April May June July August Males Constant October November December January February March April May June July August Females Constant October November December January Percent occupied Logarithm of earnings Percent with employerprovided medical insurance No Yes No Yes No Yes No Yes 19.502 (0.147) -0.035 (0.208) -0.670 (0.210) -1.044 (0.207) -1.264 (0.207) -1.321 (0.211) -1.735 (0.207) -1.822 (0.210) -2.181 (0.207) -2.424 (0.208) -1.859 (0.207) -1.591 (0.207) 825,837 11.763 (0.639) -0.058 (0.206) -0.667 (0.208) -1.049 (0.205) -1.265 (0.205) -1.351 (0.209) -1.487 (0.206) -1.522 (0.208) -1.841 (0.206) -2.086 (0.206) -1.631 (0.206) -1.425 (0.205) 825,837 72.628 (0.171) 0.128 (0.241) -0.001 (0.243) -0.295 (0.240) -0.206 (0.240) -0.445 (0.245) -0.248 (0.241) -0.433 (0.244) -0.616 (0.241) -0.305 (0.241) -0.434 (0.241) -0.552 (0.240) 825,837 74.679 (0.746) 0.141 (0.241) -0.005 (0.243) -0.284 (0.240) -0.200 (0.239) -0.447 (0.244) -0.193 (0.240) -0.367 (0.243) -0.559 (0.240) -0.243 (0.241) -0.434 (0.240) -0.519 (0.239) 825,837 8.327 (0.004) -0.007 (0.006) -0.011 (0.006) -0.027 (0.006) -0.030 (0.006) -0.030 (0.006) -0.037 (0.006) -0.027 (0.006) -0.050 (0.006) -0.049 (0.006) -0.031 (0.006) -0.027 (0.006) 468,123 8.285 (0.013) -0.006 (0.006) -0.009 (0.006) -0.022 (0.006) -0.027 (0.006) -0.027 (0.006) -0.036 (0.006) -0.025 (0.006) -0.044 (0.006) -0.043 (0.006) -0.029 (0.006) -0.027 (0.006) 468,123 61.469 (0.186) 0.104 (0.263) -0.037 (0.266) -0.304 (0.262) -0.199 (0.261) -0.312 (0.267) -1.007 (0.263) -0.864 (0.266) -1.054 (0.262) -1.370 (0.263) -0.723 (0.262) -0.626 (0.261) 825,837 21.723 (0.227) 0.225 (0.320) -0.618 (0.323) -1.113 (0.320) -1.160 (0.318) -1.345 (0.325) -1.585 (0.319) -1.779 (0.324) -1.939 (0.320) -2.413 (0.320) -1.947 (0.319) -1.459 (0.318) 379,968 21.381 (0.701) 0.250 (0.317) -0.568 (0.320) -1.023 (0.317) -1.102 (0.315) -1.280 (0.322) -1.385 (0.316) -1.551 (0.321) -1.706 (0.317) -2.125 (0.317) -1.781 (0.316) -1.316 (0.315) 379,968 92.649 (0.146) 0.085 (0.207) 0.002 (0.208) 0.073 (0.206) 0.066 (0.205) -0.134 (0.210) 0.056 (0.206) 0.002 (0.209) -0.122 (0.206) 0.199 (0.206) 0.266 (0.206) 0.125 (0.205) 379,968 90.047 (0.455) 0.083 (0.206) -0.004 (0.208) 0.089 (0.206) 0.108 (0.205) -0.096 (0.210) 0.027 (0.205) -0.056 (0.209) -0.194 (0.206) 0.166 (0.206) 0.227 (0.205) 0.105 (0.205) 379,968 8.467 (0.005) 0.001 (0.007) -0.007 (0.007) -0.023 (0.007) -0.018 (0.007) -0.021 (0.007) -0.032 (0.007) -0.022 (0.007) -0.041 (0.007) -0.040 (0.007) -0.027 (0.007) -0.020 (0.007) 280,284 8.449 (0.015) 0.004 (0.007) -0.003 (0.007) -0.016 (0.007) -0.015 (0.007) -0.017 (0.007) -0.031 (0.007) -0.020 (0.007) -0.036 (0.007) -0.034 (0.007) -0.027 (0.007) -0.019 (0.007) 280,284 17.595 (0.192) -0.224 (0.270) -0.718 (0.274) -0.932 (0.269) -1.338 (0.269) 8.229 (0.841) -0.280 (0.268) -0.745 (0.271) -0.988 (0.266) -1.380 (0.266) 55.432 (0.259) 0.441 (0.365) -0.042 (0.369) -0.114 (0.363) -0.300 (0.363) 57.615 (1.142) 0.444 (0.363) -0.059 (0.368) -0.113 (0.361) -0.310 (0.362) 8.118 (0.007) -0.011 (0.010) -0.018 (0.010) -0.024 (0.010) -0.046 (0.010) 8.023 (0.022) -0.011 (0.010) -0.015 (0.010) -0.020 (0.010) -0.042 (0.010) 37 Percent with spouse with some college No Yes 53.032 (0.809) 0.069 (0.261) -0.035 (0.264) -0.270 (0.260) -0.147 (0.259) -0.328 (0.265) -0.920 (0.261) -0.735 (0.264) -0.823 (0.260) -1.089 (0.261) -0.531 (0.260) -0.520 (0.259) 825,837 18.228 (0.174) -0.039 (0.245) -0.692 (0.247) -0.943 (0.244) -1.599 (0.244) -1.221 (0.249) -1.470 (0.244) -1.622 (0.247) -2.035 (0.244) -2.080 (0.245) -1.334 (0.244) -1.322 (0.244) 565,413 17.262 (0.545) -0.038 (0.243) -0.609 (0.245) -0.866 (0.242) -1.511 (0.242) -1.138 (0.247) -1.347 (0.242) -1.447 (0.245) -1.792 (0.242) -1.823 (0.243) -1.189 (0.242) -1.205 (0.242) 565,413 50.791 (0.281) 0.538 (0.397) 0.285 (0.400) -0.658 (0.396) 0.087 (0.394) -0.456 (0.403) -1.094 (0.395) -1.067 (0.401) -1.430 (0.397) -1.884 (0.397) -1.251 (0.395) -1.142 (0.394) 379,968 55.698 (0.867) 0.507 (0.393) 0.214 (0.396) -0.626 (0.392) 0.141 (0.390) -0.550 (0.399) -0.946 (0.391) -0.947 (0.397) -1.157 (0.392) -1.475 (0.392) -1.065 (0.391) -1.064 (0.390) 379,968 14.880 (0.229) 0.245 (0.325) -0.455 (0.326) -0.745 (0.324) -0.986 (0.322) -1.017 (0.329) -1.177 (0.322) -1.186 (0.327) -1.802 (0.323) -1.660 (0.323) -1.240 (0.322) -1.029 (0.322) 274,774 12.219 (0.724) 0.290 (0.322) -0.393 (0.324) -0.665 (0.321) -0.944 (0.320) -0.951 (0.327) -0.991 (0.320) -0.983 (0.324) -1.547 (0.321) -1.363 (0.321) -1.089 (0.320) -0.892 (0.319) 274,774 70.641 (0.238) -0.411 (0.336) -0.293 (0.340) -0.269 (0.334) -0.516 (0.334) 63.549 (1.047) -0.436 (0.333) -0.245 (0.337) -0.225 (0.331) -0.460 (0.331) 21.413 (0.258) -0.346 (0.364) -0.902 (0.368) -1.219 (0.361) -2.209 (0.361) 25.170 (0.810) -0.409 (0.360) -0.809 (0.364) -1.184 (0.358) -2.080 (0.358) Number of children No Yes 2.461 (0.009) -0.005 (0.013) -0.002 (0.013) 0.006 (0.012) 0.024 (0.012) 3.810 (0.037) -0.001 (0.012) 0.008 (0.012) 0.016 (0.012) 0.041 (0.012) February -1.269 -1.373 -0.426 -0.504 -0.037 -0.035 -0.340 -0.298 -1.455 -1.370 0.007 0.029 (0.274) (0.272) (0.370) (0.369) (0.010) (0.010) (0.341) (0.338) (0.369) (0.366) (0.013) (0.012) March -1.879 -1.568 -0.634 -0.537 -0.044 -0.042 -0.866 -0.796 -1.723 -1.709 0.083 0.032 (0.271) (0.268) (0.365) (0.364) (0.010) (0.010) (0.336) (0.334) (0.363) (0.360) (0.013) (0.012) April -1.831 -1.447 -0.567 -0.420 -0.032 -0.026 -0.818 -0.714 -2.067 -1.972 0.099 0.044 (0.273) (0.271) (0.369) (0.368) (0.010) (0.010) (0.339) (0.337) (0.367) (0.363) (0.013) (0.012) May -2.334 -1.901 -0.588 -0.467 -0.058 -0.049 -0.973 -0.781 -2.330 -2.136 0.113 0.059 (0.269) (0.267) (0.364) (0.362) (0.010) (0.010) (0.334) (0.332) (0.361) (0.358) (0.013) (0.012) June -2.447 -2.063 -0.856 -0.758 -0.064 -0.057 -0.865 -0.655 -2.427 -2.258 0.118 0.070 (0.272) (0.269) (0.366) (0.365) (0.010) (0.010) (0.337) (0.335) (0.365) (0.362) (0.013) (0.012) July -1.798 -1.499 -1.180 -1.160 -0.040 -0.035 -0.190 0.016 -1.430 -1.328 0.097 0.058 (0.271) (0.268) (0.365) (0.364) (0.010) (0.010) (0.336) (0.333) (0.362) (0.359) (0.013) (0.012) August -1.687 -1.499 -0.975 -0.931 -0.039 -0.038 -0.268 -0.154 -1.606 -1.544 0.072 0.046 (0.269) (0.266) (0.363) (0.361) (0.010) (0.010) (0.334) (0.331) (0.361) (0.358) (0.012) (0.012) 445,869 445,869 445,869 445,869 187,839 187,839 445,869 445,869 290,639 290,639 445,850 445,850 Notes: Based on ENOE 2005-I, 2006-II, 2007-III, 2008-IV, 2010-I, 2011-II, 2012-III, and 2013-IV. Only includes adults born between March 3, 1960 and March 2, 1980. Standard errors in parentheses. Each column within each panel represents a separate regression. The covariates are dummies for state of birth and survey wave. 38 TABLE 12—August-September difference across age groups in adulthood outcomes, fixed effects model Age group All 25-29 30-34 34-39 40-44 45-53 Males 25-29 30-34 34-39 40-44 45-53 Percent with some college Percent occupied Logarithm of earnings Percent with employer- Percent with provided spouse with medical some college insurance -1.182 (0.733) [73,280] -1.470 (0.445) [180,821] -1.766 (0.389) [217,623] -1.111 (0.407) [202,942] -1.453 (0.477) [151,171] -0.783 (0.822) [73,280] -0.442 (0.515) [180,821] -0.083 (0.461) [217,623] -0.364 (0.477) [202,942] -1.339 (0.564) [151,171] -0.0277 (0.017) [41,720] -0.0266 (0.012) [104,007] -0.0178 (0.011) [126,018] -0.0293 (0.012) [116,314] -0.0366 (0.015) [80,064] -0.284 (0.839) [73,280] 0.163 (0.549) [180,821] -1.182 (0.504) [217,623] -1.574 (0.530) [202,942] 0.879 (0.615) [151,171] -0.530 (0.914) [34,811] -1.466 (0.536) [113,869] -1.990 (0.464) [154,101] -0.734 (0.473) [150,327] -0.714 (0.546) [112,305] -1.287 (1.104) [33,671] -2.211 (0.673) [83,149] -1.615 (0.596) [99,950] -0.422 (0.632) [93,697] -1.119 (0.759) [69,501] -0.461 (0.756) [33,671] 0.010 (0.432) [83,149] -0.011 (0.379) [99,950] 1.059 (0.399) [93,697] -0.621 (0.516) [69,501] -0.0105 (0.019) [25,727] -0.0260 (0.013) [63,420] 0.0058 (0.013) [75,585] -0.0252 (0.014) [68,584] -0.0449 (0.018) [46,968] -0.935 (1.288) [33,671] 0.319 (0.831) [83,149] -1.568 (0.760) [99,950] -2.800 (0.789) [93,697] 0.300 (0.919) [69,501] -0.235 (1.291) [15,134] -1.336 (0.742) [53,489] -0.972 (0.621) [74,373] -1.054 (0.607) [74,815] -0.355 (0.689) [56,963] Females 25-29 Number of children -1.042 -0.910 -0.0544 0.190 -0.814 0.0504 (0.979) (1.212) (0.031) (1.042) (1.275) (0.031) [39,609] [39,609] [15,993] [39,609] [19,677] [39,608] 30-34 -0.836 -0.507 -0.0207 -0.148 -1.557 0.0069 (0.592) (0.773) (0.020) (0.694) (0.766) (0.022) [97,672] [97,672] [40,587] [97,672] [60,380] [97,669] 34-39 -1.891 -0.089 -0.0479 -0.877 -2.928 0.0508 (0.510) (0.698) (0.018) (0.645) (0.681) (0.022) [117,673] [117,673] [50,433] [117,673] [79,728] [117,667] 40-44 -1.692 -1.724 -0.0419 -0.531 -0.476 0.0939 (0.525) (0.734) (0.020) (0.686) (0.721) (0.026) [109,245] [109,245] [47,730] [109,245] [75,512] [109,238] 45-53 -1.674 -1.727 -0.0293 1.222 -1.213 0.0151 (0.596) (0.847) (0.024) (0.791) (0.842) (0.033) [81,670] [81,670] [33,096] [81,670] [55,342] [81,668] Notes: Based on ENOE 2005-I, 2006-II, 2007-III, 2008-IV, 2010-I, 2011-II, 2012-III, and 2013-IV. Only includes adults born between March 3, 1960 and March 2, 1980. Standard errors in parentheses. Number of observations in square brackets. Each coefficient is the result of a separate regression. The regressions include dummies for state of birth and survey wave. 39 FIGURE 1—Relative age and test scores in 2012 by date of birth Source: ENLACE 2012 test scores for Tlaxcala, grades 3 to 9. Relative age was computed as the difference in days between the date of birth of each student and the average for all students in the same grade, divided by 365.25. Test scores were standardized at the state level by grade. Observations were grouped by month of birth. 40 FIGURE 2— Relative age and test scores in 2009 by date of birth Source: ENLACE 2009 test scores for Tlaxcala, grades 3 to 9. Relative age was computed as the difference in days between the date of birth of each student and the average for all students in the same grade, divided by 365.25. Test scores were standardized at the state level by grade. Observations were grouped by month of birth. 41 FIGURE 3—Test-takers by distance between day of birth and December 31 Notes: Based ENLACE 2009 and 2012, grades 3 to 9 in Tlaxcala. The horizontal axis shows the number of days between the birthdate of the students and the closest December 31. The top panel only includes test-takers in 2009 born between July 3, 1994 and July 2, 1999. The bottom panel only includes test-takers in 2012 born between July 3, 1997 and July 2, 2003. 42 FIGURE 4—Mother characteristics by distance between day of birth and December 31 Notes: Based on ENOE 2005-I, 2006-II, 2007-III, 2008-IV, 2010-I, 2011-II, 2012-III, and 2013IV. It includes 9,361 children born between July 3, 1997 and July 2, 2003, residing in Tlaxcala with their mother. The horizontal axis shows the number of days between the birthdate of the child and the closest December 31. The solid lines show the fit of quadratic polynomials at each side of December 31, and the shaded areas show their 95% confidence intervals. 43 FIGURE 5—ENOE sample: number of respondents by birthdate in five-day bins Distance in days to the closest September 1 Notes: Based on ENOE 2005-I, 2006-II, 2007-III, 2008-IV, 2010-I, 2011-II, 2012-III, and 2013-IV. Only includes adults born between March 3, 1960 and March 2, 1980. The horizontal axis shows the number of days between the birthdate of the person and the closest September 1. The solid lines show the fit of quadratic polynomials at each side of September 1, and the shaded areas show their 99% confidence intervals. 44 FIGURE 6—Adulthood outcomes around the September cutoff, averages for five-day bins % with some college % occupied Logarithm of earnings % with employer-provided medical insurance % with spouse with some college Number of children (women only) Distance in days to the closest September 1 Notes: Based on ENOE 2005-I, 2006-II, 2007-III, 2008-IV, 2010-I, 2011-II, 2012-III, and 2013-IV. Only includes adults born between March 3, 1960 and March 2, 1980. 45 FIGURE 7—Estimates of month-of-birth fixed effects in adulthood outcomes relative to September % with some college % occupied Logarithm of earnings % with employer-provided medical insurance % with spouse with some college Number of children (women only) Month of birth (September is the omitted category) Notes: Based on ENOE 2005-I, 2006-II, 2007-III, 2008-IV, 2010-I, 2011-II, 2012-III, and 2013-IV. Only includes adults born between March 3, 1960 and March 2, 1980. Estimates produced with fixed-effects models with covariates (see Table 11). Baseline month is September. Solid markers indicate estimate is significant at 95% confidence. 46 FIGURE 8—Simulation results to illustrate that “anything goes”: wealth can magnify, attenuate or reverse relative-age effects in adulthood Notes: Based on a simulation using the functional forms and parameters mentioned in the text. The gap in investments illustrates differences in the allocation of resources to siblings perceived as having different abilities. It is computed as the investment in the high-ability sibling minus investment in the low-ability sibling. The ratio of earnings is calculated using those investments and actual ability of each sibling (which is assumed to be the same.) It is computed as the earnings of the high-ability sibling divided by the earnings of the low-ability sibling. 47 APENDIX TABLE A—Relative-age effects: literature review Author(s) Country and data source Outcome Technique Main results IV relative age with assigned relative age. Controls for season of birth effects are possible because of the cross-country variation in cutoff dates for school eligibility. The difference between the oldest and the youngest in 3rd or 4th grade ranges between 0.20-0.47σ across the 11 countries analyzed. Of the 19 countries only two (Denmark and Finland) do not have significant estimates. For the rest they range in 0.14-0.42σ. For the pooled samples they are 0.26-029. σ in 4th grade and 0.16-0.18σ in 8th grade. Panel I. Test scores Bedard and Dhuey (2006) Datar (2006) Lawlor, Clark, Ronalds and Leon (2006) Austria, Belgium, Canada, Czech Republic, Denmark, England, Finland, France, Greece, Iceland, Italy, Japan, New Zealand, Norway, Portugal, Slovak Republic, Spain, Sweden, United States. Trends in International Mathematics and Science Study (TIMSS) 1995 and 1999, Early Childhood Longitudinal Study (ECLS) and National Education Longitudinal Study (NELS). Sample sizes across grades and countries range between 2,920 and 25,062. United States. Early Childhood Longitudinal Study-Kindergarten Class. Representative sample of children who entered kindergarten in school year 19981999. Sample size in main regressions ranges between 13,039 and 13,777. Scotland. Aberdeen Children of the 1950s cohort study. Sample size is 12,150. Cascio and US. Project STAR, Tennessee. Sample Schanzenbach (2007) size is 5,719. Puhani and Weber (2007) Germany. Progress in International Reading Literacy Study (PIRLS) 2001. Sample size ranges between 1,123 and 6,591. McEwan and Shapiro Chile. Eight annual surveys of first (2008) graders (JUNAEB 1997-2004) and census of fourth graders from SIMCE 2002, TIMSS 1998 Elder and Lubotsky United States. Early Childhood (2009) Longitudinal Study-Kindergarten cohort (ECLS-K) and the National Educational Longitudinal Survey of 1988 (NELS:88) Kawaguchi (2009) Crawford, Dearden and Meghir (2010) Smith (2010) Math and science test scores in grades 4 and 8 (3 for ECLS and 8 for NELS) Math and reading test scores at the start of kindergarten and the end of first grade. IV using as instrument the number of days between child’s 5th birthday and the school’s cutoff date, and state’s kindergarten cutoff date. Different domains of OLS with age at childhood starting school and intelligence at ages season of birth fixed 7, 9, and 11. effects. Reading and math Stanford Achievement Test scores at the end of kindergarten. PIRLS tests scores in grade 4. IV using expected age. Age at starting primary school and age relative to class peers were both associated with the different measurements of childhood intelligence. Test-score differential between children who enter kindergarten with a one-year difference in age is 0.71σ. IV using assigned One year of relative age has an relative age. Month- effect of 0.40σ in PIRLS in 4th level observations. grade. Test scores in grades RDD 4 (JUNAEB) and 8 (TIMSS) Math and reading test scores, grade repetition Delaying entrance to kindergarten one year increases test scores in 0.6-0.8σ in kindergarten, and results in gains of 0.07-0.10σ in the first two years of school. The benefits are larger for at-risk children. Reading: 17-19 percentiles in kindergarten, 14 in 1, 11 in 3, 11 in 5, and 6 in 8 Math: 24-25 percentiles in kindergarten, 18 in 1, 12 in 3, 9 in 5, and 4 in 8 Japan. TIMSS 2003. Sample size ranges Test scores in grades OLS with quarter-of- 4th graders born in Jan-Mar (the between 2,453 and 4,558. 4 and 8. birth fixed effects. youngest) score 0.186-0.222σ below those born in Apr-Jun (the oldest). For 8th graders the estimates are 0.113-0.159σ. England. Census of all children attending Test scores at ages 7, Month-fixed effects, 0.35σ at 7, 0.21σ at 11 and 0.13σ at state (public) schools in England, 11, 14 and 16. comparing versus 16 national achievement (Key Stage) test Sep., Different cutoff results and some background dates by LEA, IV characteristics date of birth, home postcode and a school identifier Canada. British Columbia. Students born Repetition of 3rd 2SLS, identifying One additional year of age at test between May 1985 and December 1987. grade and test scores SSA effect using a reduces the probability of repeating Administrative records for 3rd grade and in 10th grade. temporary variation 3rd grade in 6 percentage points, Foundation Skills Assessment in 10th in policy that created and increases test scores in 0.090σ grade. Sample size ranges between differences. in numeracy and 0.107σ in literacy. 94,428 and 109,956. 48 IV: actual age is instrumented with predicted age Increase test scores by more than 0.3σ Sprietsma (2010) Black, Devereux and Salvanes (2011) Grenet (2011) Robertson (2011) Crawford, Dearden and Greaves (2014) Nam (2014) Belgium, Canada, Denmark, France, Iceland, Italy, Japan, Korea, Latvia, Norway, New Zealand, Poland, Portugal, Spain, Sweden, Yugoslavia. Program for International Student Assessment (PISA) 2003. Sample size ranges between 2,160 and 10,465. Norway. Norwegian military records from 1980 to 2005. Sample size is 652,215, all males. France. Panel Primaire de l’Éducation Nationale (PPEN) 1997, with a sample size of 9,342 for 1st grade and 7,653 for 3rd grade. Panel Secondaire de l’Éducation Nationale (PSEN) 1995, with a sample size of 16,790 for 6th grade, 10,894 for 9th grade, and 5,460 for 11th grade. Diplôme National du Brevet (DNB) 2004, with sample size of 781,391 for 9th grade. US. Suburban school district of Chicago. Administrative records and test scores of Illinois Standards Achievement Test (ISAT) 2004-2007 in grades 3, 5 and 8. Sample size in main regressions ranges between 1,604 and 1,730. UK. Avon Longitudinal Study of Parents and Children, children born in August and September. Sample size of 982. Millennium Cohort Study, cohort members who were born in England between September 2000 and August 2001. Sample size is 5,019. Test scores at age 15 OLS with controls for institutional settings (grade retention, vocational education, etc.) In 10 out of 16 countries the results are positive and significant. In those countries older students (by 11 months) outscore younger peers by 0.096-0.226σ in reading and 0.108-0.305σ in math. IQ approximately at IV with estimated age 18. school starting age. On year of additional age-at-test increases score in 0.10σ, and SSA decreases the score by 0.03σ. Test scores in grades IV relative age with Older students (born in Jan.) score 1, 3, 6, 9 and 11. assigned relative age. 0.66σ above the youngest (born in Dec.) in 1st grade, 0.40-0.53σ in 3rd, 0.25-0.30σ in 6th, and up to 0.18σ in 9. Not significant effect is found in 11th grade. ISAT scores reading OLS using quarterand math grades 3, 5 of-birth fixed effects. and 8. Probability of being retained. National achievement test scores at age 7, IQ and Wechsler objective language dimensions (comprehension and expression), locus of control, self-esteem at age 8. Korea. Korean Education and Test scores in grades Employment Panel, grades 9 (2004) and 7, 8, 9, and 12. 12 (2007). TIMSS, grade 9 (1999, 2003, 2007). Korean education Longitudinal Survey, grades 7 (2005), 8 (2006) and 9 (2007). Korea Youth Panel Survey, grades 9 (2003) and 12 (2009). Sample size ranges between 1,229 and 6,789. RDD. Day-level observations. Students born in Sep.-Nov. perform better than those born in other quarters. The effects are up to 0.24σ in 3rd grade, 0.20σ in 5th, and 0.16σ in 8th grade. They are also less likely to have been retained. Older students do better when tests are given at the same point in time but not when given at the same age. IV relative age with In grades 7-9 the effects range assigned relative age between 0.13-0.31σ comparing older (born in Mar.) versus younger (born in Feb.) The effects are not significant or negative in 12th grade. Panel II. Intermediate outcomes Bedard and Dhuey (2006) Puhani and Weber (2007) US and Canada. Administrative data Participating in prefrom BC, Ministry of Education, Canada. university program NELS restricted use, US (CA), writing the SAT and enrolling in 4-year college in the US Germany. State of Hessen, administrative records for academic year 2004-05 for students who entered school between 1997 and 1999. Sample size ranges between 32,059 and 182,676. Billari and Pellizzari Italy. Bocconi University administrative (2008) records of applicants and students who enrolled from 1995 to 1998. Sample size is 5,269 for students and 12,676 for applicants. Dhuey and Lipscomb US. Project Talent, 10th through 12th (2008) graders attending high school in 1960. Sample size ranges between 250,069 and 264,986. National Longitudinal Study of the High School Class of 1972, high school seniors. Sample size ranges Probability of attending higher track secondary track. Relatively older students in British Columbia and the US are more likely to participate in preuniversity academic programs during the final years of high school, and more likely to enter a flagship postsecondary institution in the United States. IV using assigned One year of relative age increases relative age. Month- by 12 percentage points the level observations. probability of attending the highest secondary track. Grades in college, time to completion, and admission test scores. IV using the incidence of private pre-schools in the province of birth Younger students have higher graduation marks and higher admissions test scores. Indicator of whether the student was sports team captain or club president. Self-reported leadership skill OLS using quarterly relative-age dummies. Relative age measured with respect to the statewide school Relatively oldest 25 percent of students are between 4 and 11 percent more likely to hold a leadership position than the relatively youngest. 49 IV relative age with assigned relative age. Month-level observations. between 15,960 and 15,968. High School and Beyond, 1980 senior and sophomore classes. Sample size ranges between 18,031 and 18,066. Dhuey and Lipscomb US. ECLS 1998-2004, with a sample size (2010) of 8,120. NELS 1988, with a sample size of 16,870. ELS 2002, with a sample size of 12,140. Mühlenweg and Puhani (2010) Grenet (2011) Schneeweis and Zweimüller (2014) (Project Talent). Probability of disability classification in kindergarten, and grades 1, 3, 5, 8 and 10. Germany. State of Hessen. Probability of Administrative records of students born attending, upgrading in June or July in general and vocational to, or downgrading schools in academic years 2002-03 from the highest or a through 2006-07. Sample size ranges higher secondary between 10,192 and 11,077 per academic track. year. entry cutoff date. IV using assigned An additional month of relative age relative age. Month- decreases the likelihood of level observations. receiving special education services by 2–5 percent. IV using assigned Younger students (born in July, relative age. Month- after the cutoff) are 8-19 percentage level observations. points less likely to be in the highest track. The effects are mitigated: in 10th grade younger students are more likely to upgrade to and less likely to downgrade from a higher track. 2000-2005 Scolarité data base. Sample Admission into Multinomial probit. Younger students are about 21 size of 772,561. academic track, Month-level percentage points less likely to vocational track, observations. choose a high-track school than dropping out. their older peers in grade 9 Austria. Administrative student-level Choosing, upgrading IV using assigned Strong positive relative-age effect data from the city of Linz, with a simple or downgrading relative age. Month- on track choice in grades 5–8, it size of 25,248. PISA 2003 and 2006, different tracks. level observations. persists beyond grade 8 for students with a sample size of 8,136. from less-favorable socioeconomic backgrounds and students in urban areas (younger students are about 21 percentage points less likely to choose a high-track school than their older peers in grade 9) Panel III. Adulthood outcomes Kawaguchi (2009) Dobkin and Ferreira (2010) Black, Devereux and Salvanes (2011) Grenet (2011) Japan. Employment Status Survey 2002. Years of education, Adults born in Apr. 1968-Mar. 1972, employment status who were 30-34 years old in 2002. and income. Sample size ranges between 16,310 and 27,801. OLS with quarter-of- Younger males have 0.13 fewer birth fixed effects. years of education and 3.9% lower Age in months. earnings. Younger females have 0.08 fewer years of education and are 0.019 more likely to be employed. United States. 2000 Decennial Census Educational RDD. Age in days. Younger individuals are more Long Form for the states of California attainment and, likely to complete more years of and Texas (approximately 15% of the wages, employment, schooling below the college level. population in each state.) Individuals household income, The effects are below 0.01. No over the age of 30. Sample size in main house ownership, effects found at the college level or regressions ranges between 479,500 and house value, marital on other variables. 767,302. status. Norway. Norwegian Registry Data. Earnings at age 24- IV. Age in months. SSA lowers earnings up to 11.6% Individuals born between 1932 and 1970. 35, educational for females and 9.9% for males at Sample size ranges between 220,418 and attainment, the beginning of their careers (ages 701,676. childbearing, full24-25) and the effects disappear by time employment, age 30. For full time workers the mental health, receipt figures are 4.1 and 3.8%, of social assistance. respectively. No effect on educational attainment. SSA makes boys 0.005 less likely to have poor mental health at age 18 from (average is 0.07.) And reduces in 0.018 (average is 0.08) the probability of teenage pregnancy. France. Enquête Emploi (1990-2002). Age left full-time OLS using monthly- Older women (born in Dec.) stay Individuals born between 1945 and 1965. education, highest fixed effects. Age in 0.07 fewer years in school. Older Sample size ranges between 74,129 and held qualification, months. individuals are more likely to have 102,532. probabilities of “academic qualification” employment, and less likely to hold a “vocational working part-time, qualification.” Older men are less and being a civil likely to be unemployed. Older servant, and hourly individuals are more likely to be wage. civil servants, and have 0.8-2.3% higher hourly wages (not controlling for educational 50 Fredriksson and Ockert (2013) Zweimüller (2013) Nam (2014) attainment). Fuzzy RDD with SSA increases educational two-sample IV. Age attainment in 0.16 years, only a in months. timing effect on earnings, primeage earnings unaffected, for low people with educated parents, earnings in prime age increase with SSA Austria. Austrian Social Security Starting wages, type OLS using assigned Younger students are more likely to Database. Individuals who entered the of entry job, relative age as the pursue an apprenticeship and less labor market between 1997 and 2000, educational explanatory variable. likely to have higher education. and their wages are observed up until attainment, wages 1 RDD. Age in Older males (born in Sep.) are less five years after entry. Sample size ranges to 5 years after job months. likely than to work in a blue-collar between 153,777 and 169,987. entry. job and significantly more likely to work in a white collar-job. Wage penalty of 1.4-1.9 percentage points for students born in August (the youngest) compared to students born in September (the oldest.) After five years of labor market experience, the wage penalty amounts to 0.8-1.1 percent. Korea. August Supplementary Data of Attending two- or OLS using assigned No significant effect. the 2010 Economically Active four-year college, relative age as the Population Survey. Individuals born attending four-year explanatory variable. between March 1976 and February 1986. college, employment, Age in months. Sample size ranges between 2,657 and holding a regular job, 5,274. and wages. Sweden. Administrative data from Statistics Sweden. Individuals born between July 1935 and June 1955. Sample size ranges between 307,585 and 2,037,166. Educational attainment in years of schooling, employment and wages 51