As I have done throughout much of my career, I continue to work on new methodology for analysing panel data structures – where we observe the same units, say, companies, over multiple time periods. My recent research on panel data has focused on a couple of areas that are important in empirical research. First, how to best handle unbalanced panels – where not all units are observed in all time periods -- in nonlinear models. I show that extensions of the correlated random effects approach are natural and straightforward to implement. Second, I have been looking at how to combine instrumental variables methods and panel data methods to allow two sources of endogeneity in nonlinear panel data models. The two sources are unobserved factors that remain roughly constant over time – such as company organisation, managerial talent, culture – and shocks that can vary over time (such as hidden costs unobserved to the data analyst). I find that combining correlated random effects and control function approaches leads to convincing estimation while also being easy to implement in standard econometrics packages.
The course I taught in the PhD programme had an emphasis on panel data method, but I started with quick reviews of ordinary least squares and instrumental variables methods, emphasising the modern perspective that all models are approximations and then interpreting our estimates in that light. I then used this background to cover the standard estimators for linear panel data models – particularly random effects, fixed effects, and first differencing. In addition, we covered instrumental variables versions of all these methods. An important feature of the course is I showed how the correlated random effects (CRE) approach can unify random and fixed effects approaches in the linear case. I then covered nonlinear models for limited dependent variables, including the CRE approach for panel data. Several students were very interested in nonlinear dynamic panel data models, and I was able to fit in material on that, too. Some staples from causal inference, including difference-in-differences estimation, where one exploits a before-after, control-treatment group design, were also covered.
A solid grounding in general econometrics is needed to learn and apply machine learning. One must learn about the sources of endogeneity in models – including those that are used in the finance industry – before tackling more sophisticated estimation methods. I think of machine learning methods as a tool that takes some of the arbitrariness and tedium out of choosing predictors and control variables. It does not usually help in formulating interesting problems about company behaviour and the effects of interventions on financial assets. One must still think very carefully about issues of endogeneity and design before using machine learning to simplify the task of selecting controls.
I found the students in my EDHEC course to be exceptionally motivated and talented. This is not surprising: the same traits that make one successful in industry, an eagerness to know more and the ability to absorb complicated ideas, are the same traits that lead to successful scholars. Frankly, I thought the students would want more direct instruction on how the methods I was teaching could be immediately helpful to them. I was naïve. The students have the experience to know that new knowledge can be helpful both immediately but also down the road. A typical PhD student may be a bit too concerned with getting a grade and just moving on to the next task.
My experience is mainly in publishing research in econometric methodology. But through students and colleagues, and in serving on editorial boards, I find that empirical work that is transparent and relatively simple has the best chance of having impact and being published. In other words, choose econometric methods that are up to the task, but do not use fancy methods just for the sake of making them fancy. Try different methods that are accepted has having good features to show robustness of your findings. Most importantly, do not become dogmatic about methods. There are lots of good ideas out there and trying different approaches can provide important insights.