Adrian TORCHIANA soutiendra sa thèse "Essays in Applied Econometrics with Missing Data" lundi 9 octobre 2017, 10:00, Salle MF 323.
Directeur de thèse : Pierre DUBOIS, Directeur scientifique de TSE & professeur, UT1 Capitole.
Le jury sera composé de :
Monsieur Xavier d’HAULTFOEUILLE, Professeur d’Economie, CREST
Monsieur Laurent GOBILLON, Professeur Paris School of Economics
Monsieur Thierry MAGNAC, Professeur d’Economie, UT1 Capitole
Résumé (en anglais):
In the first chapter, I consider the bias that might arise in production function estimation when sales are used as a stand-in for production. In practice, researchers typically observe sales and not production; the two are distinct because firms manage inventory through time, and items sold during an arbitrary accounting period were not necessarily produced contemporaneously. I show using simulations that using sales as a stand-in for production can bias production function regressions when firms manage inventory dynamically. I then go to the data: I study a French administrative dataset called FICUS, which covers the universe of French firms from 1994 to 2007, and allows me to observe sales, production, labor, and capital, at the firm-year level. I perform my analysis at the four-digit industry level, and show that the bias from using sales as a stand-in for production is small in most industries, suggesting that researchers who observe only sales generally need not worry that results derived from production function estimation are invalid. However, in certain industries where changes in inventory are common, the bias is non-negligible.
In the second chapter, which is joint work with Paul T. Scott, Ted Rosenbaum, and Eduardo Souza-Rodrigues, we show that misclassification in remotely sensed land cover data leads to biased estimates of both land areas and land cover transition rates, and propose a correction based on a hidden Markov model. Using simulations and a high-quality validation dataset, we show that our method produces consistent estimates of land use transition probabilities, whereas naive estimates of transition rates are erroneously high. A broad implication is that applied researchers should carefully consider and control for the impact of errors in remote sensing when studying the determinants of land use change. Importantly, our method produces consistent estimates of land cover transition probabilities without requiring ground-truth validation data, which are typically difficult to obtain. This is relevant for policy: for example, monitoring land cover is a central point of climate negotiations, and NGOs and other organizations evaluating changes in countries' deforestation rates may want to apply our method.
In the third chapter (which is also joint work, with the same coauthors as the second), we apply our HMM method to study deforestation in Brazil. We develop a model of Amazonian deforestation and regrowth that allows us to predict how levels of Amazonian biomass and agricultural land respond to transportation costs and agricultural commodity prices in both the short- and long-run. In our model, land managers balance forest clearing costs against discounted future returns to agricultural production when deciding whether to clear a parcel of forest. Our empirical strategy relies on transportation costs computed using detailed spatial data describing Brazil’s paved and unpaved road network, as well as on estimates of deforestation rates derived from satellite sensor data, using the methodology in the second chapter. We plan to extend the model in future work.
