Figure 4

Multivariable regression of prediction errors (MAE). Between city MAEs were regressed against an indicator for whether the predicted and training cities were in the same country, differences in population density and distances to city center. Separate regressions were carried out for the best-off and worst-off deciles. Coefficients of ‘differences in distances to CBD’ and ‘differences in population density’ show change in error for each unit of difference in t-values. (See Materials and Methods for further details of variables included and how the difference in t-values were computed)