Research

Access to basic infrastructure is a critical component of quality of life and an important measure of economic development. However, on-the-ground data about infrastructure access, especially in low-income countries, is often sparse and costly to collect. We train and calibrate a machine learning model to extract data on infrastructure access for each 6.72x6.72km area of Africa from satellite images. The model achieves accuracy levels of 77.1% to 84.7%. We show the value of this novel dataset with two applications. First, we use a spatial regression discontinuity design to study how much of the heterogeneity in infrastructure access across countries comes from differences in institutional quality. Second, we study the role of political favoritism in explaining within country heterogeneity.

We predict asset ownership in Africa making use of Landsat-8 satellite imagery and ground truth data from the Demographic and Health Surveys (DHS). We trained a ResNet-18 model} (He et al., 2016), a sub-class of a Convolutional Neural Network (CNN), and used it to predict asset ownership levels over 6.72 km by 6.72 km patches of land in Africa. We used transfer learning to train our model to perform a regression task, initializing our model with weights originally used to classify images. The performance of our best model is much better than the performance of our baseline models, but our model might be overfitting, and the predictions it generates are far off from our ground truth data. We obtained lower $R^2$ validation set values than Yeh et al. (2020), a paper that attempts our same task.

Using administrative data about the grades of Business and Economics undergraduate students from Chile, and exploiting the fact that students are randomly assigned to their first semester classes, I want to examine the existence of peer effects among students, and to study how different class compositions affect their outcomes. In 2012, the University of Chile implemented a college special admission program that targets high achieving vulnerable students from public schools, and I want to see how students that belong or don’t belong to this program interact within their own group and with the other group, and how students should be assigned to classes in order to maximize their academic outcomes. Besides the fact that in this context the peers of the students are determined in an exogenous way, another advantage of this setup is these college students are assigned on average to 6 classes during their first semester, and they interact with different students in each of these classes, so this allows me to use instrumental variables in a similar fashion to Bramoulle et al. (2009), computing measures of the outcomes or characteristics of the peers of the peers of a student, in contrast to other type of studies where all students belong to the same peer group, as it is the case with roommate or classroom studies. Besides that, the network structure can be exploited to compute network characteristics that could be able to explain how the interactions between these groups might determine the outcomes of the students in this context.

In this project I revisit the findings of Michalopoulos & Papaioannou (2014), exploring the role of national institutions on local development levels in Africa. The authors use 3rd and 4th degree RD polynomials, define the bandwidth of their RD design in arbitrary ways, and there are some pixels were the distance to the border is not computed properly, so to overcome those flaws and check how robust their findings are, I plan to use the optrdd package from Imbens & Wager (2019) to estimate an Optimized Regression Discontinuity model, a data-driven model that doesn’t rely on polynomial or bandwidth definitions.

Understanding the impact of the built environment and urban populations on climate change and air quality is a growing challenge as the percentage of the world’s population that resides in urban areas continues to rise. In particular, human activity levels have been shown to be directly linked with ground-level ozone, which is responsible for several health and climate effects. We explore this relationship through the use of a multimodal learning architecture that predicts ozone concentrations (parts per billion) in urban areas from satellite and street-level imagery. The model comprises two Convolutional Neural Networks (CNN), one trained on satellite images of each location to learn higher-level features such as geographical characteristics and land use, and another trained on multiple street-level images of each location to learn ground-level features such as motor vehicle activity. The feature representations learned from each sub-model are concatenated and passed through several fully connected layers to predict the ozone level of the location. This concatenated model achieves a test RMSE of 11.70 ppb. This approach can be used to inform urban planning and policy by providing an insight into the particular urban features that aggravate ozone concentrations.

Long-standing literature argues that social capital is closely implicated in labour market outcomes. However, this hypothesis has yet to be tested in Latin America, the most unequal region in the world. We focus on Chile, one of the most stratified countries in Latin America. This study examines the relationship between social capital and four measures of status attainment, including job prestige and employment income. We use data from the first wave of the Longitudinal Social Study of Chile (ELSOC), a representative survey of the Chilean urban population aged 18–75 years. We analyse a subsample of 1,351 individuals who are currently employed. A Bayesian model of over-dispersion with relational data is used to estimate the size of the network, a novel measure of social capital. We analyse the data set using linear and logistic regression models and a complementary path analysis, first estimating models for the entire sample, and then splitting the sample into three groups to evaluate differences within individuals’ socioeconomic background. Results indicate that contacts’ occupational prestige has a positive association with job prestige and employment income, while the size of the network increases individuals’ salaries and labour participation. We also observe that social capital flows through stratified networks which tend to favour individuals from high socioeconomic backgrounds. We discuss the need to conduct more in-depth evaluations of how better creation of social capital and its effects on status attainment could be closely linked to positions of privilege and advantage accumulation processes in highly unequal contexts.

Taking advantage of georeferenced data from Chilean students, we estimate the impact of commuting time over academic achievement. As the commuting time is an endogenous variable, we use instrumental variables and fixed effects at school level to overcome this problem. Also, as we don’t know which mode of transport the students use, we complement our analysis using machine learning methods to predict the transportation mode. Our findings suggest that the commuting time has a negative effect over academic performance, but this effect is not always significant.