Using Experimental Data to Validate a Dynamic Behavioral Model of Human Capital Investments in Children: Evaluating the Impact of Mexico's PROGRESA Program

该项目提供了关于儿童入学和生育率的父母决策行为模型的估计，以评估在墨西哥农村进行的一项旨在提高儿童完成学业水平的对照社会实验的影响。该计划以子女出勤率为条件向父母提供补贴。社会实验在评估实验项目特征的变化以及超出实验生命周期的长期影响方面的能力往往是有限的，而且它们根本不能用于评估根本不同的项目。相反，在某种程度上，估计的行为模型是有效的，它可以用来评估反事实政策的影响。执行以下分析。首先，评估该计划对未明确纳入该计划的补贴水平的影响。该计划本身并没有实验性地改变补贴的数额。其次，评估将该计划扩展到部分或全部不符合条件的家庭的影响。第三，由于该计划只实施了两年，因此不可能通过简单地比较接受治疗的家庭和未接受治疗的家庭来评估该计划的长期影响。例如，即使该计划被实验组视为永久性的，该计划对现有家庭的影响取决于他们当前的情况，例如，他们有多少孩子。然而，该方案的长期影响可能是影响这些情况，例如，新家庭将拥有的儿童数量。第四，我们比较了教育补贴计划与对使用童工进行法律限制的截然不同的政策的影响。该项目还在方法论上做出了贡献。行为模型的结构估计需要对功能形式的辅助假设，即偏好、技术和其他约束，以及不可观察的随机元素的分布。依靠对估算中使用的数据样本元素的模型拟合检验来评估这类模型的有效性，提供了关于模型有效性的有用但远非令人信服的证据。这类模型通常受制于某种形式的“预测试”估计，因为模型的最终公式是基于先前公式对数据某些方面的拟合。在这项研究中，我们使用了一种不同的方法来验证模型。模型的有效性将根据基于随机抽取的对照组数据的模型的结构估计如何预测程序的实验影响来确定。

The project provides estimates of a behavioral model of parental decision-making about child schooling and fertility in order to evaluate the impact of a controlled social experiment in rural Mexico designed to augment completed schooling levels of children. The program provides subsidies to parents conditional on the school attendance of their children. Social experiments are often limited in their ability to evaluate variations in the characteristics of the experimental program as well as longer run effects that extend beyond the life of the experiment, and they cannot be used at all to evaluate radically different programs. In contrast, to the extent that the estimated behavioral model is valid it can be used to evaluate the impact of counterfactual policies. The following analyses are performed. First, the impact of the program at subsidy levels not explicitly part of the program is assessed. The program itself did not experimentally vary the amount of the subsidy. Second, the impact of extending the program to some or all ineligible families is assessed. Third, because the program has been in effect for only two years, it is not possible by simply comparing the treated to the untreated households to evaluate the longer run impact of the program. For example, even if the program is viewed by the experimental group as permanent, the impact of the program on the existing families is conditioned on their current circumstances, e.g., the number of children they have. However, the longer-run impact of the program may be to affect those circumstances, e.g., the number of children new households will have. Fourth, we compare the impact of the subsidy program on schooling to the radically different policy of placing legal restrictions on the use of child labor. The project also makes a methodological contribution. Structural estimation of behavioral models requires auxiliary assumptions about functional forms, i.e., of preferences, technology and other constraints, and the distributions of unobservable random elements. Assessing the validity of such models by relying on tests of model fit to sample elements of the data used in estimation provides useful, but far from compelling, evidence on the validity of the model. Such models are often subject to a form of "pre-test" estimation in that the final formulation of the model is based on the fit of prior formulations to certain aspects of the data. In this research, we use a different approach to model validation. The validity of the model will be ascertained according to how well structural estimates of the model based on data from the randomized-out control group predict the experimental impacts of the program.