If you're interested in how the authors did the analysis, the fun details – the data sources and analytic/econometric methods – are in the supplementary materia

If you're interested in how the authors did the analysis, the fun details – the data sources and analytic/econometric methods – are in the supplementary materials (link at end of article). Ironically enough, I assume that if they hadn't managed to publish a higher-level article in Science, the actual paper (hypothetically published elsewhere) would have been the 116 (+/- epsilon) pages of detailed measures and methods, result displays, and sensitivity analyses/robustness checks. If I read the article correctly, they systematically detect when emissions policies are adopted by looking at sudden increases/"breaks" in "stringency indices" of a country's emissions policies (within a particular industrial sector.) The machine learning comes in their automated procedure for identifying these breaks (which is well out of my expertise – my longitudinal analysis training is in good old fashioned multilevel modelling and structural equation modelling.) They then do the standard trick of using a binary before/after "treament" indicator variable. Once that is done, though, it looks like you can choose your method of choice for detecting significant changes in emissions pre-post intervention.