Effect of climate and geography on worldwide fine resolution economic activity

country or a relatively large area, are: Gabon (central West Africa), South-West USA (parts of California and Arizona), and a few Balkan countries (Albania, Montenegro, and Bosnia and Herzegovina). Specifically, the latter area displays a negative bias, with the geography-climate prediction underestimating GCP-PC, and vice-versa in the former two cases. According to Eq (1) these discrepancies may indicate that the combined contribution of institutions and national resources (term δ) is not negligible: this term would be positive in the case of the Bal- kan countries above, with institutions and/or national resources therefore boosting economic productivity compared to the geography and climate baseline, and negative in the other two areas. However, as we will show later, these discrepancies are not present when smaller sec- tions of the GCP-PC data (i.e. split into terciles) are modelled, which may imply that the results for the all-grid case exhibit a minor shortcoming in the RF statistical modelling with a very large amount of data being fit (nearly 13,000 grid points) and/or in the GCP-PC data, particu- larly for countries where the reporting of economic productivity figures is less established. Top tercile of GCP per capita Here we subsample the GCP dataset by retaining only the top tercile of GCP-PC. The RF error again levels off after six or seven variables. From an initial nMAE of 0.68 with one predictor, this reaches a value of 0.28 after six variables and only improves by about 0.02 with the addi-

tion of the following four predictors–the nMAE with the top ten predictors is 0.26 (Fig 3 and Table 3). Compared to the all-grid-point case, latitude is not the most important predictor for the top tercile; rather it now ranks third, after the top quartile of air pressure (MSLP top Q) and the distance from major rivers. The top predictor, MSLP top Q, indicates that high air pressure is a key factor in determining the GCP-PC for countries in the top tercile. Also, and as in the case of all-grid-point, variations in climatic variables, rather than their mean, are still Fig 6. Difference between predicted and observed GCP-PC. Difference between predicted (using the top 10 predictors) and observed GCP-PC, namely (f–GCP-PC) in Eq (1). Top left: all grid points. Top right: top tercile of GCP-PC. Bottom left: middle tercile of GCP-PC. Bottom right: bottom tercile of GCP-PC. https://doi.org/10.1371/journal.pone.0229243.g006 PLOS ONE Effect of geography on fine resolution economy PLOS ONE | https://doi.org/10.1371/journal.pone.0229243 March 2, 2020 12 / 20