Machine Learning (ML) algorithms are powerful data-driven tools for approximating high-dimensional or non-linear nuisance functions and, among other things, handling sparsity. In this paper, we bring the power of ML algorithms (i.e., Lasso, classification and regression trees, and random forests) into statistical methods to estimate the impact of policy interventions using panel data when the effect of confounding regressors on the outcome are potentially highly non-linear. We use Double Machine Learning (DML) (Chernozhukov et al., 2018) for the estimation of causal effects of homogeneous treatments with unobserved individual heterogeneity (fixed effects) and no unobserved confounding by extending Robinson (1988)’s partially linear regression model. We investigate three alternative approaches to handle the unobserved individual heterogeneity: the within-group (WG, mean-centred) estimator, first-difference (FD) estimator, and correlated random effect (CRE) estimator (Mundlak, 1978). Monte Carlo simulations show that there are gains from the use of DML with flexible learners with non-linear non-smooth functions of the covariates, but least squares estimates are more robust with linear and non-linear smooth problems. In addition, we observe that the sequence of estimated causal effects from tree-based learners is not normally distributed. The DML method has many potential applications in social sciences regarding the effect of a policy intervention on outcomes of interest. We provide an illustrative example with the introduction of the national minimum wage in the UK.
Presented by:
Dr Annalivia Polselli (University of Essex)
Date & time:
18 Oct 2023 12:30 pm - 18 Oct 2023 13:30 pm
Venue:
The seminar will be held in person at 2N2.4.16 and online. Please contact the seminar organisers for the zoom call details at iserseminars@essex.ac.uk.
External seminars home