Convert error from absolute to relative

proportion_ci outputs absolute values, but plt.errorbar requires
values relative to the original data.

Update results images

Author: endolith

examples/merrill_1984_fig_2c_2d.py

@@ -142,10 +142,14 @@
                    'Plurality'):
         x, y = zip(*sorted(count[method].items()))
         CE = np.array(y)/y_cw
+
+        # Add 95% confidence interval error bars (Clopper-Pearson exact method)
         ci = np.empty((2, len(y)))
         for i in range(len(y)):
             ci[:, i] = binomtest(y[i], y_cw[i]).proportion_ci()
-        plt.errorbar(x, CE*100, ci*100, fmt='-', label=method)
+        yerr = ci - CE
+        yerr[0] = -yerr[0]
+        plt.errorbar(x, CE*100, yerr*100, fmt='-', label=method)
         table.append([method, *CE*100])
 
     print(tabulate(table, ["Method", *x], tablefmt="pipe", floatfmt='.1f'))