The dissertation consists of three essays that analyze production efficiency, size effects and human mobility using a panel of 591 Chinese agricultural households in the later 1990s.;The first essay estimates technical efficiency through the framework of a translog stochastic production frontier with a behavioral inefficiency component. The results reveal a decreasing trend of output elasticities with respect to labor and fertilizer. The marginal effects of inefficiency terms show significant output gains by eliminating land fragmentation, improving access to education in rural area, and promoting specialization and mechanization.;The second essay examines land heterogeneities and the empirical inverse relationship in developing countries between farm size and productivity. When we utilize the egalitarian principle during land allocation in China and use imputed quality constant land rather than actual land area in the regression, the inverse relationship between farm size and productivity disappears. Hence, the strong inverse relationship that some studies have found is undoubtly due to a number of methodological problems, i.e., the failure to account properly for land quality differences.;The third essay analyzes the decision-making of rural Chinese households on whether to allocate efforts exclusively to the family farm, taking off-farm work locally or migrating to another region. We observe statistically significant state dependence between the current period response and past behavior. Simulated probability changes support our hypothesis that the average schooling of household labor and household size play important roles in Chinese farm household job-location decisions.;The dissertation advances several econometric methodologies. We obtain the close form of marginal effects of inefficiency terms in Battese & Coelli (1995) model, as well as their variance estimator. We propose two methods to summarize curvature conditions of flexible functional forms. In the second essay, we apply Hahn-Hausman (2002) test to examine the validity of the instrumental estimation. We derive the Murphy-Topel two-step estimation variance estimator in the linear case. The third essay extends the dynamic discrete choice model of Wooldridge (2002a,b) to trichotomous setting. We use the maximum simulated likelihood estimation to reduce the computation burden in estimating the random effects multinomial logit model.