Conditional Independence Test for Weights-of-Evidence Modeling

Weights-of-evidence modeling is a GIS-based technique for relating a point pattern for locations of discrete events with several map layers. In general, the map layers are binary or ternary. Weights for presence, absence or missing data are added to a prior logit. Updating with two or more map layers is allowed only if the map layers are approximately conditionally independent of the point pattern. The final product is a map of posterior probabilities of occurrence of the discrete event within a small unit cell. This paper contains formal proof that conditional independence of map layers implies that T, the sum of the posterior probabilities weighted according to unit cell area, is equal to n, being the total number of discrete events. This result is used in the overall or “omnibus test” for conditional independence. In practical applications, T generally exceeds n, indicating a possible lack of conditional independence. Estimation of the standard deviation of T allows performance of a one-tailed test to check whether or not T-n is significantly greater than zero. This new test is exact and simpler to use than other tests including the Kolmogorov-Smirnov test and various chi-squared tests adapted from discrete multivariate statistics.