UC Davis

This dissertation covers various topics ranging from migration to wealth measures in developing countries and to policy analysis in general. In Chapter 1, I study the internal migration in China and aim to understand its effect on the human capital accumulation of the later generations. In Chapter 2, I investigate novel wealth measures and look for ways to understand the low-cost wealth metrics. In Chapter 3, I discuss a doubly-robust method to estimate causal effects for panel data in the presence of effect heterogeneity.In Chapter 1, I investigate the mechanisms through which parental migration affects the schooling outcomes of children left behind in rural China. This issue affects 61 million children. Previous literature on this topic focuses on estimating the net effect of migration, whereas this paper disentangles the net effect into different mechanisms of policy interests. I establish a theoretical framework to incorporate three essential and widely-studied mechanisms that migration could affect left-behind children’s school performance: parental accompaniment, child’s study time, and investment in children. Motivated by the theoretical model’s solution, I apply the structural equation model to estimate the influence through different mechanisms. I propose an identification strategy based on instrumental variables and the Heckman selection model. Using the model on rural household survey data from nine provinces, I find that the effects through parental absence and investment are both significantly negative with large sizes. In contrast, the impact through the child’s study time is insignificant with a negligible size. The surprising negative effect through investment is mainly driven by reduced nutrition investment by the de facto custodians, who may not have compatible incentives to allocate the remittances to the child. Through a refined subgroup analysis, I find that girls are suffering ten times more from the underinvestment than boys, revealing a shocking gender inequality in rural China. In Chapter 2 (co-authored with Professor Ashish Shenoy), I aim to understand novel poverty measures. For under-developed countries, wealth measures are essential for measuring economic growth, policy design, and setting development goals. In particular, I focus on the use of nighttime light data. Unlike standard wealth measures based on national accounts or expenditure surveys, nighttime light data has the advantages of high frequency, low cost, and precision over small geographic units. These advantages make it an ideal substitute for in-person surveys. Nighttime light data has been a popular wealth measure in recent years, and previous papers mainly argue that nighttime light intensity and gross domestic production levels are highly correlated globally. However, we analyze the relationship between night light growth and economic growth for 179 countries or regions and find heterogeneity in the correlation. To deal with the heterogeneity, we propose a weighted least squares estimator for the average correlation coefficient by properly re-weighting each country. We find a significant and positive average correlation among middle-income countries. Moving beyond the average association, we apply the LASSO regression to identify and estimate non-zero individual correlation coefficients. This is inspired by the sparsity of country-level associations observed in the preliminary analysis. We further apply the ”knockoff” method to control the false discovery rate among the selected countries. In Chapter 3 (co-authored with Professor Dmitry Arkhangelsky, Professor Guido W. Imbens, and Lihua Lei), we develop a novel method for causal inference with observational panel data, which overcomes the limitations of existing methods. Cross-sectional models account for treatment assignment using methods such as inverse probability weighting. We extend this approach to panel data. Taking the case of staggered adoption as an example, we model the adoption time with duration models such as the Cox hazards model. As long as the information about the assignment mechanism is accurate, our method works under substantially weaker assumptions than the traditional methods. As a byproduct, we characterize the class of experimental designs under which the conventional methods are guaranteed to produce consistent estimates of the causal effects. The method from our paper can be widely applied to empirical analysis, such as program evaluation.