Machine Learning (ML) and Artificial Intelligence (AI) are well underway to becoming a staple in quantitative economics. Regularly, new fascinating applications hit national or international media. From AI outperforming doctors diagnosing breast cancer, over identifying dialects in naked mole rats to monitoring hiring discrimination, it seems no field is immune to the breakthroughs promised by AI. Similarly, a wide and expanding set of machine learning methods has been put forward. This poses challenges to researchers outside computer science wanting to incorporate such methods into their research.

Almost a year ago a blog post appeared on Towards Data Science suggesting an alternative to the common correlation coefficient called the Predictive Power Score (PPS). The PPS is promoted for being able to measure nonlinear relationships (asymmetrically) for both numeric and categorical variables. Without going into the technical details behind the PPS algorithm, I was wondering whether an extension to the classic correlation setup couldn’t produce similar (or even better) results.

Updated version: January 5th, 2021

Although Belgium has largely been spared, reported measles infections in the rest of Europe reached record numbers in 2018, with 82,596 children and adults in 47 of 53 countries affected; sadly enough, 72 of them did not survive the disease [1]. The outbreak of measles coincides with a growing distrust of vaccines in general. While precise numbers are difficult to obtain, it is striking that Europeans under 65 have less confidence in the safety and importance of vaccines than people over 65 [2].

Recently, a group of students and academicians raised their concerns about the current state of economics, especially the way it is taught in academia (1–3). Many of the criticisms raised are not new and go back to Keynes’ and Mises’ well-known objection against a quantitative approach to economics (see Figure 1 for an overview of some of these critiques and their counter-arguments). The authors of Thinking like an Economist? find it particularly troubling that economics students in the Netherlands are primarily being taught quantitative and mathematical research skills. This parallels a growing tendency towards technical complexity in current economic practices. Yet, it contrasts sharply with Keynes’ scepticism towards quantitative economics as a viable field of interest:

No one could be more frank, more painstaking, more free from subjective bias or parti pris than Professor Tinbergen. There is no one, therefore, so far as human qualities go, whom it would be safer to trust with black magic. That there is anyone I would trust with it at the present stage or that this brand of statistical alchemy is ripe to become a branch of science, I am not yet persuaded. But Newton, Boyle and Locke all played with alchemy. So let him continue. (4 p. 156)

In this blog I will discuss different approaches to adjust standard errors for panel data. As panel data often contains both a time and spatial dimension, considerations of serial and spatial correlation often require more than the standard heteroskedasticity-robust standard errors. A popular choice is clustering on the time, group or both levels; clustering on a supra-group level is also possible in case correlation between the groups might be problematic (see future blogs). Here I will restrict myself to clustering on the group level(s) and compare those results to the (less common) Newey–West and Driscoll–Kraay adjusted standard errors. Code (Stata) and data (Egger and Nelson, 2011) to reproduce the results can be found here.