Prediction of Occurrence of Discrete Events - Geomathematics: Theoretical Foundations, Applications and Future Developments - page 164

Fig. 5.18 Weighted logistic regression applied to gold deposits ( circles ) in Meguma Terrane,
Nova Scotia. ( Top ) Posterior probability map with 91 unique conditions for seven binary patterns
without missing data, unit cell size ¼ 1km 2 ;( Bottom ) t -value map for Fig. 5.18a (Source:
Agterberg et al. 1994 , Plate 5a, b)

deviations. Table 5.3 shows regression coefficients in comparison with WofE
weights. Estimated standard deviations are shown as well. Finally, Fig. 5.19 is an
evaluation of the goodness of WLR fit. The difference between expected and
observed relative frequencies is plotted against posterior probability. The absolute
value of the largest difference of this type is 0.0775. This is less than the
Kolmogorov-Smirnov statistic ( 0.1426; 95 % two-tailed test) from which it may

be concluded that the fit of the logistic model is good.
It can be seen in Table 5.3 that the seven contrasts show similarity with the seven
regression coefficients. Theoretically in discrete multivariate analysis ( cf . Andersen
1990 ; Christensen 1990 ), it can be shown that if, asymptotically, explanatory
variables are conditionally independent, then logistic regression and the procedure
used in WofE produce identical results (cf. Agterberg 1992 ).

¼

5.2.2 Comparison of Logistic Model
with General Linear Model

Applications of the general linear model to study the occurrence of large copper
deposits in the Abitibi Volcanic Belt were discussed in the previous chapter. In a
discussion of a paper by Agterberg and Robinson ( 1972 ), Tukey ( 1972 ) suggested