in rainfall and temperature. Note that indeed the effects change locally. Furthermore, the choice of measure consequently assumes jumps between the knots and the magnitude of jumps are significantly different. Figure 6. Fitted values of regression without (left chart) and with spline dummy (right chart), Change in Rainfall, CBMS 2007-2010 Note: F-test for equality of dummies significant Figure 7. Fitted values of regression without (left) and with spline dummy (right), Change in TMIN, CBMS 2007-2010 Note: F-test for equality of dummies significant Another challenge on interpretability is collinearity between climate variables. For example, the three measures of rainfall change are inherently correlated since they are first and second order measures of rainfall (average and dispersion). This leads us to have four models for each of the dataset, which will be discussed in the results. Standard assumptions on the error terms are to be validated and tests for weak instrument, exogeneity and overidentification are to be undertaken through the instrumental variable regression module in Stata (see appendix). It is inevitable in some cases to encounter violation of assumptions but are taken as it is either due to data issues or due to interpretability. For instance, overidentification may be a problem in FIES dataset (Sargan’s statistic) which might be due to the number of instruments. Outliers can be menacing in CBMS-FAO | 28