Essays on family economics

• Main Author: Wang, Lu
• Format: Text
• Language: unknown
• Published: No Publisher Supplied 2021
• Subjects: DML
• Online Access: https://dx.doi.org/10.7282/t3-cbvn-ye53; https://rucore.libraries.rutgers.edu/rutgers-lib/66758

This dissertation discusses inter-generational behaviors on family economics and emphasizes the power of advanced econometric tools as semiparametric double-index model and Machine Learning approaches. The first chapter is jointly written with Yixiao Jiang, which focuses on inter-generational transfer motives from elderly parents to children above 18 years old. There are two types of motives documented in the literature: the altruistic motive, wherein larger transfers are given to poorer children, and the exchange motive, wherein larger transfers are given to the children that better serve their parents. The micro-economic theory of inter-generational transfer argues that altruism and exchange motives imply different transfer responses to children's income. Yet the existing empirical literature is mixed due to the difficulty in finding creditable exclusion restrictions. We uniquely draw a distinction between the transfer motives by estimating the income-transfer derivative in a semiparametric, two-part model. This econometric model consists of a double-index, semiparametric binary response first-stage for whether a transfer occurs, and a linear second stage for the transferring amount. Compared to models employed by extant studies, this approach preserves identification when an exclusive restriction is difficult to find, makes fewer parametric assumptions, and accounts for heteroscedasticity explicitly. Using the Health and Retirement Study dataset, this more flexible approach documents a positive relationship between children’s income and the transfer amount, which supports the exchange motive and overturns results from traditional parametric models. However, the models we use in the first chapter suffer from four potential drawbacks. First, there’re still assumptions for the functional form of the model. For example, in our Transfer Incidence step we assume that the dependent variable is a function of two linear indices. This will result in a bias if the assumption fails to fit the empirical data. Second, there might not be sufficient controls. Since our data is observational data, the categories where children’s household income falls into might not be randomly assigned, which calls for including the right confounders/controls to avoid a bias. In the first chapter, we only controlled for parents’ and children’s characteristics. But other sets of variables like parent-child interactions, parents’ wealth structures and health conditions might affect the distribution systematically and therefore need to be controlled for. We are faced with a dilemma that failing to include the right controls will result in a biased estimation of the parameter of interest yet adding too many controls will incur a huge computational burden and we may even lose identification. Third, even if we incorporate the correct controllers, models in the first chapter failed to adequately account for endogeneity from children's income. Fourth, the marginal effects of children's income within $10k-$70k remain unclear since the results are not statistically significant. We can only draw a conclusion that compared with base category < $10k, those who earn more more than $70k tends to get more transfer amount from parents. But the pattern for children whose income falls within $10k-$70k is still not clear. In Chapter 2, I try to solve the aforementioned problems and re-estimate the effects of children's income to downstream transfer amounts from parents. I employ two novel identification strategies, namely Double Selection Model (DS henceforth) and Double Machine Learning (DML henceforth). Both methods can retain the consistency of parameters of interest, handle many potential high-dimensional data, be immune to noises with respect to nuisance parts and no longer rely on traditional consistent model selection. All these properties enable me to argue that sample selection error can be reduced or even eliminated if many raw confounders are included in the potential control pool. I test against this assumption using two different robustness checks and conclude that the results with and without correcting sample selection error in a left censored model are similar, providing evidence that sample selection corrections greatly diminished in importance. Furthermore, these two novel approaches represent a statistically clear nonlinear pattern with respect to the effects of children's income on transfer amount, which the approach in Chapter 1 failed to achieve. In general, the results are consistent with those found in the previous chapter. When children are in a lower income category, the altruistic motive dominates when parents make downstream transfer decision. As children's income increase, the exchange motive becomes more important. The only difference is the magnitude of the dominance with regards to the exchange motive for higher-income children. The DS and DML models both find that though exchange motive dominates as children's income increase, the effect is weaker as compared to that was found using identification strategy in previous chapter. Chapters 1 & 2 discuss the downstream transfers from parents to children. However, the inter-generational transfers can be upstream from children to parents. In Chapter 3, I study the opposite direction, the effects of children’s educational advantages on parents' income. Adult children’s educational attainment could influence parents’ income in many different ways. First, well-educated adult children may help their parents develop a healthy lifestyle and improve their parents’ general health, which in turn, would improve the parents’ productivity and work performance. Second, well-educated adult children may have better knowledge of financial arrangements and investment strategies for helping their parents score higher other income. Third, well-educated adult children could give direct suggestions to their parents’ career development and use their network and resources to help boost their parents’ career. I start with the OLS approach to study the effects of children’s educational attainment on parents’ income, net the effects of parents’ education, age, gender and race and find out that as children’s average years of education increases by 1, the return of children’s education on parent’s total (wage) income increases by 5.9% (3.3%). These results suffer from endogeneity derived from omitted variables and simultaneity bias. The literature reported two sets of omitted variables that can cause endogeneity. Specifically, variables measuring children's ability tend to bias down the OLS estimates and variables describing children's family backgrounds have a two-way effect. On the one hand, omitted variables that describe children's family background bias OLS estimates just the same way as how ability bias the estimates, which rises up the schooling effect of children. On the other hand, as shown in the literature, researchers in 10 papers show that OLS estimates tend to bias down the true effects of schooling. Therefore, I re-estimate the education effects of children via the DML approach which can incorporate a large amount of controllers like family backgrounds, family structure as well as interaction between family members. Not only are DML estimates powerful for handling concerns about endogeneity from unobserved heterogeneity and aforementioned omitted variables, they are also immune to noises with respect to controllers. Results from this novel approach demonstrate that as children’s average years of education increases by 1, the return of children’s education on parent’s total income increases by 5.6%, which is statistically significant at 99% confidence level. However, we failed to conclude that DML results are different from OLS results from the same sample. A possible reason is the two-way effects of family backgrounds, which will be explained in detail in Chapter3.