A. Data Source
The data for this study are derived from the 2000 Taiwan census, conducted every 10 years by the Directorate of General Budgeting, Accounting, and Statistics. The Taiwan census files collect information using a detailed questionnaire similar to that used to create the PUMS files for the US censuses (long-form), except that income-related variables are excluded. At each household, the interviewer records each individual’s basic demographics (race, sex, age, and marital status), educational attainment, relationship with the head of household, working and employment status within the past two weeks, as well as the industry in which he or she works. In addition, the interviewer records the residence’s structure (number of living rooms, bedrooms, kitchens, and bathrooms), tenure status (rent or own), years lived in the residence, and the location from which the family last moved. The residence information is further linked with the housing registry from the Ministry of Interior to ascertain the floor space of the house, the building year, and the major construction material used for the residence. More importantly, the Taiwan census includes a scrambled, but unique, address for every household’s residence. As seen below, this unique address information plays an essential role in the analysis.
The advantage of using the Taiwan census is that the files contain the full sample of Taiwan residents, around 22 million in total or 300,000 individuals in most age cohorts. The large sample size, together with the detailed address information, provides a good opportunity to analyze the effect on educational attainment of the housing environment. Ideally, we would examine the link using the final education levels of all family adult respondents and their current housing information. In practice, however, this is not possible because the census files do not record family information of those who no longer reside with their parents and siblings. Obtaining the complete family background is therefore difficult, especially for adult respondents because a large portion of them do not coreside with parents
and siblings. Moreover, the census files report only the respondent’s relationship with the head of the household, but not with other members. Although we could match their relationships according to each member’s age and gender, the identification becomes quite complicated when there are more than three adults in a household (e.g., coresiding with a brother or sister-in-law).
B. Sample Selection
For the purposes of this study, we restrict the sample in several ways. We select households with at least one unmarried child aged between 15 and 20 at the time of the census, of which the eldest sibling is no older than 22. We focus on the younger sample to reduce the bias resulting from incomplete family information. We restrict the sample to ages over 15 because compulsory schooling in Taiwan ends at junior high school (9th grade). To avoid mistakes arising from matching parents, we keep only nuclear families in the sample, eliminating households that live with grandparents, relatives, or other friends. Furthermore, we drop households in which children are raised by a single parent to reduce complications because different family structures may also affect a child’s education. Finally, we include only samples that have stayed in the residence for at least three years because the housing effect usually takes a longer time to materialize.
To demonstrate the impact of exclusion criteria, Table 1 lists the observed number of youths aged from 15 to 20 for each selection criteria. The first column lists the total number of youths in the census by age cohort. As indicated by these numbers, the number of respondents peaks at the age of 19 and then gradually declines as their age rises; this pattern is consistent with the number of births between 1980 and 1985 (ages 15 to 20 in 2000) in Taiwan.7 The vast majority of youths, particularly younger ones, coreside with their families. This can be seen from the difference between the first and the second columns, which shows the number of youths who live with at least one adult aged 35 or older.
Nevertheless, more and more youths, especially those older than 20 years, choose to live alone for either marriage or work reasons. That youths live alone for other reasons may increase the risks of matching complete family information, a point we will return to later.
7 The number of respondents obtained from the census data is very close to the birth numbers between 1970 and 1975; the difference is less than 3 percent in every age cohort.
The largest reduction in sample size occurs when restricting the sample to nuclear families. This is not surprising because about 67% of the elderly in Taiwan coreside with their children.8 Among these nuclear families, around 20% of the youth do not have valid parental information: either they are growing up in single-parent families (around 60% are single mothers) or are no longer coresiding with both parents. Another 10–20% are removed because of the age restriction of the eldest sibling; the older the respondent, the more likely they are to be removed by this age constraint. Finally, around 7%
are eliminated because they have stayed in the current residence for less than three years. The final sample size consists of a little over one third of the original sample. Still, we have around 100,000 respondents in each age cohort.
C. Measure of Educational Attainment
Before describing our analysis sample, it is important to first discuss our measures of educational attainment. Previously used measures include the highest completed level of education [Boehm and Schlottmann (1999), Angrist, Lavy and Schlosser (2005), Black, Devereux and Salvanes (2005)], private school attendance [Conley and Glauber (2005)], held back in school grade [Conley and Glauber (2005), Goux and Maurin (2005)], test scores [Guo and VanWey (1999), Haurin, Parcel, and Haurin (2002)], dropping out [Green and White (1997)], and graduating from school by a certain age [Aaronson (2000)]. Because our data are derived from the census files, we cannot make distinctions between the quality of the youth’s school (e.g., school ranking), or the youth’s academic performance within the school. Therefore, we adopt a measure similar to the one used in Conley and Glauber (2005) that compares the respondent’s age with the highest schooling that he or she is currently enrolled in or has completed so far. The education system in Taiwan is similar to that of the United States, except that compulsory schooling is nine instead of 12 years. Therefore, from the age of six, children are required to take six years of elementary school and three years of junior high school. After finishing junior high school, those seeking additional education can go to senior high school (three years) and even higher after graduating from high school. Suppose a child of age 16 reports his or her highest
schooling is junior high school. Then he or she either did not proceed to higher education or had been held back a grade in previous school years. By examining one’s age and highest schooling, we can compare a child’s educational attainments with those of peers in the same age cohort.
There are, however, two complications with this measure. First is that the cut-off birthday for school admission may result in some children starting school late.9 For instance, a September-born child may be almost one year older than a child born the following August but they are in the same school grade. Because the census data only record an individual’s age (in years) at the time of the census interview, we are unable to determine whether a child meets the full age requirement at the time of school enrollment. Thus, some 15-year-old children may already be in senior high school, while others are still in junior high school.10 Second, there are two types of senior high schools (general versus vocational) and colleges (general versus junior) in Taiwan. Although the quality difference between various types of schools is small in some countries, the gap is large here because students are enrolled into schools based on their test scores on school entrance exams. Generally, general high schools are more difficult to enter, as are general colleges.11To resolve these difficulties, we first restrict the sample to youths of ages 16 and 17, and ages 19 and 20. Youths aged 15 and 18 are removed because their educational measures are harder to define. Next, we check if the respondent’s reported schooling matches the highest schooling of his or her age. More specifically, we examine if youths of ages 16 or 17 attended general high schools (nonvocational), and whether youths of ages 19 or 20 attend general colleges (nonjunior). In the discussion that follows, we refer to the younger sample as the “teen” sample and the older sample as the “young adult” sample.
D. Description of Analysis Sample
We work with two analysis samples, both described in Table 2. To demonstrate the effect of our sample selection criteria, we continue to present sample statistics by age cohorts. In total, there are
9 The cut-off birthday in Taiwan is similar to that of the United States: children must be six years old (full) by September 1st to be enrolled in the school.
10 The 2000 Taiwan census is conducted at the end of that year. Therefore, roughly half of all 15-year-old children are in junior high school, while the rest are in senior high school.
11 For instance, the minimum score for entering a public high school in Taipei in 2004 was 220 points, about 30 points higher than that of public vocational schools. Likewise, the minimum score for entering general college is considerably higher than that of junior college in Taiwan.
283,959 teens and 188,937 young adults because more young adults are removed during the selection process. In both samples, except for youths aged 20, we have more males than females, reflecting the special gender preference in Taiwan.12 Because of the sibling’s age restriction, a higher proportion of first-borns are observed in young adults than teens. No significant difference, however, is observed in the average number of siblings among different age cohorts.
The educational attainment of youths is listed in the first set of rows of Table 2. A little over half of teens were enrolled in general high schools at the time of the census; 35% were in vocational high schools, while the remaining teens were out of school. The variation in schooling among young adults is larger. About 40–50% of young adults continued schooling after high school (e.g., general or junior colleges), while another 40–50% chose to stop after general or vocational high schools. Only 5–10%
of young adults stopped their education after compulsory schooling.
One concern with our educational measure is whether the cut-off birthday affects schooling. If that is the case, we should observe a large discrepancy in schooling between two consecutive ages. Table 2 provides some evidence regarding this concern. For teens, there are only limited schooling differences between ages. In fact, the proportion of those attending general high school for 17-year-old youths is actually lower than that of 16 year olds, showing that the cut-off is not a concern for teens.
The schooling comparison among young adults is a little bit complicated. Our data for a child’s education show a rising trend of schooling between the two age cohorts. For instance, the proportion of youths attending general college increased from 17% to 25%, and attending junior college increased from 22% to 30%. Nevertheless, this observation seems unlikely to be because of the cut-off birthday because the number of young adults in each age cohort enrolled in general and junior colleges remains almost the same.13 Instead, the increase in schooling reflects the fact that those who did not seek higher education left home for work. Because our sample removes youths that live alone, young adults
12 The observation that there are more 20-year-old females than males is likely to reflect the fact that males are more likely to work away from home. As a result, the category of youths aged 20 that coreside with parents is dominated by females.
that live with their family at the age of 20 tended to enroll in higher education. In other words, the rising schooling trend is primarily because of our selection criteria, a point that we return to later.
Table 2 also reports variables describing the parental background of the youths, including age, education, and work status. The average parental age of young adults is two years older than that of teens, reflecting the age difference between teens and young adults. In both the teen and the young adult samples, mothers are less likely to have acquired higher levels of education than fathers, especially for colleges or above. Likewise, the difference in working status between fathers and mothers is quite large. Over 90% of fathers in both samples hold a full-time job, while only around 60% of mothers do. Nevertheless, in some families mothers shoulder more economic burden than fathers, with about 10% of the sample being female-headed households.
The Taiwan census data include a wide range of descriptions of housing environment, including floor space, number of rooms, age of building, tenure status, and the location from which the family last migrated. The floor space of the house is measured by square meter. On average, the typical respondent lives in a building 10 to 20 years old, with 3.5 rooms, and 130 square meters. To better account for overcrowdedness, we construct three dummies that compare the number of bedrooms in a house with the number of children in a family. Typically, parents share a bedroom, so the comparison is based on the remaining bedrooms (minus the parents’ bedroom) and the number of children. A household is considered as having high crowdedness if some children share a room, medium crowdedness if every child has his or her own room, and low crowdedness if every child has more than one room. By this standard, more than 60% of respondents live in a house with medium crowdedness; the rest reside in households with limited private space. These rates remain almost unchanged with respect to the teen or young adult sample.
More than 90% of youths live in self-owned households, reflecting the high rate of owner-occupied houses in Taiwan. In most cases, the youths in the sample have been at the same residence for more than 10 years; less than 14% of youths moved into the current residence within the last five years, of which around 3% moved within the local vicinity (within the same village); the rest migrated from other villages.
E. Area Dummies and Family Heterogeneity
Before showing the estimation results, it is useful to first describe the area dummies, which aim to control for unobserved family heterogeneity. Because the census data record detailed address information, area dummies can be constructed from the highest level (county) to the lowest (lin). For instance, Taipei, the capital of Taiwan, consists of 12 towns, 435 villages, and 9741 lins. The average number of square kilometers of a town, village, and lin in Taipei are 22.6, 0.624, and 0.028, respectively.14 Not surprisingly, as seen in Table 3, the sample number in an area drops sharply as the level of government jurisdiction moves from towns to lins. While there are, on average, 780 teens and 520 young adults in a town, each lin accommodates only 3.1 teens and 2.4 young adults. From the percentile distribution based on lin, at least half of lins have only one teen and one young adult at the time of the census. Despite this, there is still a great deal of variation in many other lins in the sample.
This can be seen from the numbers in parentheses, showing that the number of teens or young adults at the first quartile, based on the whole sample, is 3 and 2, respectively.
If area dummies are good controls for family heterogeneity, we should observe that the extent of variation within a neighborhood declines when a smaller neighborhood is used. To demonstrate the relationship between family heterogeneity and area dummies, Table 4 shows “within” and “between”
standard deviations (SD) of housing environment variables. Because these SDs may exhibit different patterns in cities and rural regions, we further separate our sample into two groups based on the number of residents in the town: large towns (more than 100,000 residents) and small towns (less than 100,000 residents). For the purpose of exposition, we only list these numbers at the village and lin level. Consistent with our expectation, “between” SD rises and becomes larger than “within” SD for the vast majority of housing variables as the neighborhood level moves from village to lin.
Nevertheless, we do not find a clear difference in SD between large and small towns in the sample.