A. Basic Specification
Results of our basic specification using the two analysis samples are presented in Table 5. Our basic specification estimates the linear probability model including all variables displayed in Table 2, except the father’s age because father’s age is highly correlated with mother’s age. Because estimates from the teen and young adult samples are similar, we first discuss the similarities in the two results, and then discuss the differences.
As is typical for these types of regressions, our coefficients imply strong links between a child’s education and parental schooling. The better the parental schooling, the more likely it is that youths will seek higher education. As discussed earlier, the findings could reflect parental preferences over a child’s education or a child’s generic abilities inherited from higher-educated parents or both. In addition, higher educational achievements are also positively associated with an increase in the mother’s age and the father’s employment status, but negatively related to the mother’s employment status. There is no observed difference in the education of children raised in female- and male-headed households.
Two variables of a child’s characteristics deserve special attention. First, our results indicate that the sibship size produces a small and negative marginally significant effect after controlling for all other factors. Our estimates suggest that adding one sibling reduces the chance of going to general high school by 0.3% and general college by 0.4%. Second, our estimates indicate that being the first born largely increases the chance of enrollment in general high school and college, by 6.6% and 3.2%, respectively.
Our results demonstrate a strong link between a child’s education and housing variables. A positive correlation is observed between a child’s education and the floor space. An increase of 100 square meters, for instance, is associated with an increase in the chance of enrollment in general high school and college by 1.5% and 1.2%, respectively. Likewise, children living in owner-occupied houses have a higher chance of getting into general high school or college, as are children living in newer houses.
However, interpreting these results requires caution. It is possible that the results reflect the fact that
parents that are more willing to invest in a house are likely to create positive benefits for their children’s learning.15 It is also possible that these coefficients may reflect our inability to control for household income. Perhaps new, larger, self-owned houses produce a positive effect simply because they are associated with a child’s family’s well-being. We will discuss this issue later.
Youths who have recently moved from other locations (migrated 3–5 years ago) are less likely to be enrolled in general high school. The greater the distance they moved, the larger the negative effect on a child’s education. Because the housing effect usually takes time to materialize, this effect should be attributed to residential stability in an earlier period, i.e., at the time of junior high school.
Residential stability is valuable to teens probably because they do not have to learn to adapt to a new social network (junior high schools are usually very close to where teens live).16 Notice that the effect of residential stability is less evident among teens. This could be because general high school admissions are based on every young adult’s test score. As a result, many young adults cannot benefit from the existing social network as they did at junior high school because they must attend distant general high schools.
One parameter of particular interest is the household crowdedness. Similar to Goux and Maurin (2005), our results also confirm the importance of private space on a child’s education.17 Nevertheless, its effect is more complex and possibly nonlinear. For instance, teens growing up in families of medium crowdedness are more likely to enroll in high school than those in high- or low-crowdedness houses. For young adults, those raised in medium crowdedness perform equally well as those in low-crowdedness houses. Notice that our estimation also controls for a house’s floor space. Changing from
15 Green and White (1997) explained why home ownership might positively influence children’s cognitive and behavioral outcomes. First, there is a stronger investment incentive for owners compared with renters. Better physical home environments increase the probability of success of the children of owners. Second, compared with renters, there is higher self-esteem among owners, resulting in greater emotional support for the children.
Finally, there is greater geographic stability creating a neighborhood network that is likely to promote a child’s outcome.
16 A longer tenure (or less mobility) often implies a more stable home and school (peer) environment, which helps to invest in building social capital that enhances a child’s outcomes. Therefore, a longer tenure tends to lead to better outcomes for children. For details, see Coleman (1988).
17 Because our estimation setting is different from that of Goux and Maurin (2005), a comparison may be inappropriate. However, we estimated an additional model using the average number of rooms per person as the
high- to low-crowdedness houses does not refer to an increase in floor space and number of rooms at the same time. Instead, the effect should be interpreted as increasing the private space, but reducing the shared space, in a household (e.g., smaller living room). From our estimates, it appears that there is an optimal mix of private and shared space that helps a child’s schooling.
As stated earlier, there is a risk of bias generated from our sample selection. Most notable is the restriction on the eldest sibling’s age and on coresiding with a mid-aged adult. If such a restriction induces a new bias into the estimation, we should observe differences in results that use only 17-year-old youths from the teen sample and only 19-year-17-year-old youths from the young adult sample. This can be seen from Table 5, where we list estimated results of youths from the age cohorts of 17 and 19, respectively. As demonstrated in the table, we observe only modest differences between results using the full sample and half of the sample. None of the estimated coefficients, however, changes its sign after restricting the sample, and the vast majority remain statistically significant. These results imply that our sample selection, at most, results in small biases in the estimation.
B. Effects of Area Dummies
One key concern regarding our findings is whether our results demonstrate the importance of housing variables or just the inability to control for unobserved family heterogeneity. For instance, strong associations between a child’s education and housing variables found in the estimation could possibly be caused by failure to control for the household’s income, one kind of unobserved family heterogeneity. To address this question properly, it is important to show some evidence that adding area dummies indeed mitigates the concern of unobserved family factors. Table 6 lists the estimated results using area dummies at the town, village, and lin levels, respectively. For the purpose of comparison, we also include results without controlling for neighborhood fixed effects. As indicated from this table, the total number of area dummies at each level is 0, 364, 7508, and 91,929, respectively, and a slightly smaller number for young adults. Given that the number of area dummies varies so much, it is not surprising that regressions controlling for different levels of neighborhood effects yield dramatically different estimates. For instance, the coefficient of sibship size in the teen sample changes from –0.016 when there are no area controls to –0.006 and –0.003 when controlling at
the village and lin level, respectively. In fact, the Hausman test suggests that any two sets of estimates are statistically different.18If positive relationships between housing variables, especially floor space, ownership status, housing age, and child’s education, are posited as another channel to display the income effect, we should anticipate the effect becoming smaller when looking across children residing in the same neighborhood. Families residing in the same neighborhood should have similar family assets or potential earnings. Throughout the table, however, estimates of housing variables continue to show significant effects on the youth’s educational outcome, some of which become even larger after controlling for many more area dummies. While it is still possible that our results are biased because large variations exist within the same neighborhood, the results do not seem to suggest that our findings are driven by unmeasured household income.
Another way to examine the effect of area dummies is to compare our results with findings in previous studies accounting for unobserved endogeneity through IV methods or family dummies.
Generally, these studies found the coefficient on sibship size changes from statistically significant in OLS estimation to insignificant in IV estimation [e.g., Angrist, Lavy, and Schlosser (2005), Black, Devereux, and Salvanes (2005)]. It is therefore interesting to see whether adding more area dummies generates a similar result. From the table, it is clear that the coefficient of sibship size diminishes when a finer level of area controls is included. At the level of the lin, the coefficient of sibship size for teens is only marginally significant at the 10% level Obviously, a finer area control reduces the effect of sibship size, a sign supporting the reduction of unobserved family heterogeneity.
C. IV Estimation
So far we have shown that estimates with area fixed effects exhibit a pattern similar to recent studies employing IV strategy. Nevertheless, it is still uncertain whether neighborhood dummies are good controls for unobserved family heterogeneity. A more convincing method is to compare area fixed effect results with IV results so that the extent of endogenous bias can be determined.
18 The smallest chi-square value occurs when comparing results of village fixed effects with those of lin fixed
Nevertheless, this is not easy because our regression includes, in addition to the number of siblings, a variety of variables characterizing a child’s housing environment. Unless we are able to find an instrument for every housing variable, implementing a full-scale IV estimation is extremely difficult.
In light of this difficulty, we have decided to conduct IV estimation in a different way. We first use multiple births and preferences toward a mixed sibling-sex composition to construct the instrument of sibship size.19 Through exogenous variations because of multiple births at the third-born and sibling-sex composition of the oldest three children, we can look at the effect of three or more births on the educational outcome of the first- and second-born child in families with at least three children.
Neighborhood dummies are also included to aid family controls.
The estimates in the first two columns of Table 7 report the first-stage and IV results for teens and young adults in families of at least three children. Because our sample is reduced to less than half of its original size because of the restriction on the number of children, we control for the village instead of the lin fixed effects. All instruments are significant in the first stage. Family size goes up by 0.83–0.87 in response to multiple births at the third born. Likewise, the family size increases by 0.38–0.42 for families whose first three siblings are girls; this reflects Taiwanese parents’ preference for boys over girls.
Controlling for the village fixed effects, IV results again show that the number of siblings has little effect on the child’s education. Moreover, we do not observe clear differences in the coefficients of housing variables between regression results and IV results. The vast majority of housing variables still hold their original signs and magnitudes. To formally examine whether IV results differ from area fixed effects results, we reestimate the area fixed effects model using village dummies based on this new sample.20 The Hausman Test shows the chi-square value for these two sets of estimates is 3.5 for the teen sample and 0.29 for the young adult sample; both fail to reject the null hypothesis that the IV results and regression results are statistically indifferent.
19 Taiwan Census data only record the age of each family member. Therefore, multiple births are identified by checking whether two consecutive children share the same age. It is possible that our method overstates the number of multiple births for families whose age gap between two consecutive children is less than one year.
Nevertheless, we believe the likelihood of a mother having two children in one year is limited.
20 The Hausman test is conducted based on the 21 explanatory variables in the regressions. Coefficients of fixed dummies are not considered.
Results reported above account for the potential endogeneity between sibship size and child’s education. However, our estimates can still be subject to biases if housing variables are endogenously determined based on the number of children (e.g., parents may decide to move to a bigger house once they have more children). Given we cannot find an instrument for every housing variable, we restrict our sample to those who have moved into their current residence one year before the second child was born. For these families, the chance that their housing variables are correlated with the exogenous variations in sibship size (e.g., multiple births or sex composition) should be considerably lower, and therefore should shed some light on the effect of housing variables. The remaining columns of Table 7 present the first-stage and IV estimates for this particular sample. Although the first-stage results continue to confirm the validity of our instruments, coefficients on many housing variables become insignificant after imposing the restriction, likely because of a much smaller sample imposed by the move-in year constraint. Nevertheless, the majority of housing variables still hold their sign, showing at least some evidence of their importance.
D. Gender Differences
When discussing housing variables, one often-raised question is whether gender differences exist.
Do boys need a bigger house? Do girls have special needs for private space? To explore this possibility, Table 8 expands the estimation by allowing for gender interactions on three variables:
first-born, floor space, and household crowdedness. As expected, first-born boys have a higher school enrollment than first-born girls; this is likely because boys in Taiwan’s society are subject to more social pressure than girls.
Both household crowdedness and floor space exhibit some gender differences. In addition, those gender differences seem to change for different ages. The chance of school enrollment is higher for boys raised in households with larger floor space, but there is no observable gender difference in the teen sample. On the contrary, girls raised in medium-crowdedness households have a higher chance of getting into high school. However, the gender difference disappears in the young adult sample. It appears that different housing needs exist for boys and girls at different times in their lives.
5. Conclusions
Understanding factors that determine the children’s educational attainments is an important research question in the social sciences. The answer is not only crucial for human capital formulation, a key driver of economic growth, but also essential for income distribution purposes because education is considered a driver for income mobility.21 Among those components, housing environment provided by the parents is often considered of great relevance [Haveman and Wolfe (1995)]. While it is widely believed that a better housing environment stimulates a child’s learning, there is limited evidence as to the causal link between housing environment and a child’s schooling.
In this study, we seek to uncover the effect of housing environment on children’s educational attainments. Differing from Goux and Maurin (2005), who use exogenous variations in the child’s private space as instruments, we control for unobserved family heterogeneity through their residential choices. In general, families living within a close distance share similar parental preferences, household assets, and earning potential. In addition, children in the same neighborhood typically go to the same school. Using the Taiwan census files that include the unique address information of every household in the records, we compare the chance of general high school or general college enrollment for youths of the same age and in the same neighborhood. After controlling for area fixed effects using tens of thousands of area dummies, our results indicate the importance of housing variables in determining a child’s schooling. The educational attainment of children is positively associated with increases in floor space, increases in residence stability, and the ownership status, but negatively related to increases in building age. The results are robust even accounting for the endogeneity between sibship size and child’s education using IV estimation.
Several findings deserve special attention. First, a first-born child, particularly a boy, is more likely to perform well in school. While the finding may reflect the fact that parents, particularly those in Taiwan, tend to put more pressure on first-borns, our finding is consistent with Black, Devereux, and Salvanes (2005) who argue that birth order, not family size, matters for a child’s outcome. A more
21 According to Haveman and Wolfe (1995), the government’s spending on children in terms of primary and secondary education in 1992 is estimated to be 235 billion, or roughly 4% of GDP in U. S. In Taiwan, the spending on compulsory education is a little less than 3 percent of GDP.
careful analysis that explores a full range of effects of birth order and possibly its interactions with housing variables may be necessary.
Second, our results are different from the findings of Goux and Maurin (2005) regarding the effect of a child’s private space. Although our results also confirm the importance of household crowdedness, its effect appears to be nonlinear because the chance of school enrollment is higher for children raised in medium-crowdedness houses than those in low-crowdedness houses. Further investigations on the effect of household crowdedness may also be necessary to uncover the exact impact.
Finally, and most importantly, our identification uses the area fixed effects to control for unobserved family heterogeneity. While we have shown evidence supporting this approach, we caution readers that there might still be uncontrolled family factors, such as genetic differences or interactions between parents and children, in the estimation.
The main contribution of this paper is to provide causal evidence regarding the effect of housing environment on a child’s education. Although many studies have attempted to establish the link between housing environment and children’s educational achievement (e.g., ownership, residential stability), our paper appears to be the first that offers a complete picture of the effect of a wide range of housing variables. Our paper has demonstrated the importance of a few housing variables (e.g., tenure status and house floor space). Future studies could use our findings as the basis to consider more effective policy instruments to enhance children’s educational attainments in designing housing policy.
References
Aaronson, D. (2000), “A Note on the Benefits of Homeownership,” Journal of Urban Economics, 47, 356–369.
Angrist, J. D., Lavy, V and J. D. Schlosser (2005), “New Evidence on the Causal Link between the Quantity and Quality of Children,” Mimeo.
Becker, G., and H. G. Lewis (1973), “On the Interaction between the Quantity and Quality of Children,” Journal of Political Economy, 81(2), 279–288.
Becker, G., and N. Tomes (1976), “Child Endowments and the Quantity and Quality of Children,”
Journal of Political Economy, 84(4), 143–162.
Black, S., P. J. Devereux, and K. G. Salvanes (2005), “The More the Merrier? The Effect of Family
Black, S., P. J. Devereux, and K. G. Salvanes (2005), “The More the Merrier? The Effect of Family