Effect of climate and geography on worldwide fine resolution economic activity

very important: the month-to-month variations for the dew-point temperature (despite having temperature in its name, dew point temperature, rather confusingly, is not a temperature; while it also depends on temperature, it is closely associated with relative humidity) and the overall variations of air pressure (MSLP SD) rank in the top five most important variables. The nMAE is reduced to 0.31 after the fifth variable, MSLP SD. Next, in sixth position, distance to rivers is again an important predictor, with a further reduction of nMAE by ca. 0.04. Table 4 directly compares the top ten predictors in both the all-grid-point and top tercile cases. It is evident the large overlap amongst them, with eight variables in common (see col- oured boxes), and another variable related to air pressure in both cases (MSLP top Q and MSLP mean). The RF model yields again a remarkable geographical prediction/reconstruction of the top tercile of GCP-PC (S3 Fig). While the top quartile of air pressure (top predictor) sufficiently reproduces the overall geographical pattern of GCP-PC, subsequent predictors clearly enhance the fit (S4 Fig). For instance, evident improvements over a few areas in Russia, Libya, the Ara- bian peninsula and northern/western Australia are seen with the inclusion of the distance from major rivers. Latitude and the month-to-month variations of dew-point temperature fur- ther noticeably improve the fit to the GCP-PC, particularly over Libya, southern USA and northern Mexico in the case of latitude, and Arabian peninsula and Spain in the case of dew-

point temperature. Subsequent predictors only modify marginally the fit. When compared with the measured GCP-PC, there do not seem to be systematic biases in the predicted (using the top 10 predictors) GCP-PC (Fig 6 top right). Mid and bottom terciles of GCP per capita To complement the analysis of the top tercile, we concisely discuss the results for the middle and bottom terciles of GCP-PC. In terms of overall statistics (Table 3), the middle terciles yields slightly worse results than the top tercile across most of the metrics, notably for the RF performance, which is however still considerably superior to GB, and ML. The bottom tercile, instead, yields very marginally better results than the top tercile. Further, in terms of actual predictors, Table 4 shows that there is a good level of agreement for all four cases, with four predictors out of ten in common when all four cases are taken together, with further four-five in common amongst three cases. When cases are taken in pairs the number of common pre- dictors range between six and eight predictors. Further, even in the case of the middle and bot- tom terciles the RF model is capable of reproducing well the observed respective GCP-PC using the top six predictors (S5–S8 Figs). This assessment demonstrates that the statistical, and physical, link between GCP-PC and geography and climate variables is robust. As with the top tercile, there do not seem to be systematic biases in the predicted (using the top 10 predictors) GCP-PC (Fig 6 bottom panels), except perhaps for a couple of areas in the case of the bottom

tercile: north-west Myanmar and around the border of Zimbabwe (Fig 6 bottom right), though these are likely due to GCP-PC data issues. Behaviour of top variables and their potential physical link with GCP-PC We now take a closer look at the behaviour of the top predictors, particularly for the all-grid- point case, as a way to gain a better understanding of how such predictors affect GCP-PC, focussing especially on the climatic variables. Table 4 shows the linear correlation coefficient between the top ten predictors and the corresponding GCP-PC, for all-grid-point and the three terciles, respectively. It is interesting to note that individually none of the predictors has a very high linear correlation with GCP-PC, in any of the four cases. The highest correlation coefficients are with Solar Radiation median (0.45), MSLP positive step-one change (0.42), PLOS ONE Effect of geography on fine resolution economy PLOS ONE | https://doi.org/10.1371/journal.pone.0229243 March 2, 2020 13 / 20