Research

My research focuses on Econometrics. In particular, I am interested in improving estimators when regressors of interest are endogenous. In linear models, it means that the variables we are interested in are correlated with the error terms. As a result, the Ordinary Least Squares (OLS) estimator will be inconsistent, i.e. as the sample size increases, the value of the estimate will not get closer to the true value of the coefficient associated to the variables of interest.

More generally, I love Econometrics. I also love making econometric procedures available to all by coding them too. So you can find some scripts written in R and Julia.

 

Academic work:

This paper proposes to solve a canonical stochastic growth model, simulate its solutions and estimate the discount factor and risk aversion parameters through the corresponding Euler equations. In particular, we use the generalized method of moments (GMM) along with two recent refinements from Antoine, Bonnal & Renault (2007) and Guay & Pelgrin (2016) and the generalized empirical likelihood estimators (GEL). The model is calibrated to reflect features encountered in empirical data by matching the variance and autocovariance of interest rates. It reveals the presence of weak identification, so inference procedures robust to weak identification are also considered. We find that the continuously updated-GMM estimator behaves well in small samples, as well as the three-step GMM improvement from Antoine, Bonnal & Renault (2007). In addition, the GEL estimators are highly unstable in the presence of weak identification, and require significantly more time to numerically converge than its competitors.

 

I consider the Two Stage Least Estimator (2SLS) in an environment where the first stage is nonparametrically specified. Given a conditional moment restriction holds for the error term conditional on some valid instrument, an infinite amount of moment conditions can be used to estimate the parameter of interest. However, the number of instruments and their strength can affect the properties of the estimator. Following Bun & Windmeijer (2011), I derive a similar finite sample bias approximation of the 2SLS estimator, and use it to build an easy-to-implement criterion function for the optimal choice of instruments that is asymptotically optimal. The procedure that estimates the parameter of interest after selection of the instruments yields a lower Mean Squared Error than the 2SLS estimator for a broad range of designs, and its use over it is recommended in empirical applications.

 

In a linear model with an endogenous variable, I propose to use a kernel-based regression in the first stage to build predictions of the endogenous variable. The second-stage estimator using these predictions is consistent, and asymptotically normal. It is also efficient and includes the Two-stage Least Squares estimator as a special case. I study the effect of the bandwidth on the mean squared error of my estimator in a simulation exercise, and compare it to some existing estimators.