Résumé
The ability to link serial crimes accurately by relying on crime scene behaviours may be affected by the measure used to assess linking accuracy and by the coefficient used to assess across-crime similarity. To examine these issues, the present study first manipulated two factors - the variability of across-crime similarity scores used to make linking decisions and variability in the base rate of linked crimes - to determine the robustness of several measures of linking accuracy - the correlation coefficient (r), the area under the curve (AUC), and the odds ratio (OR) - to these manipulations. Both r and the AUC were sensitive to restriction of range in the predictor variable, however the AUC was more affected. The AUC was robust to base rate variability while r was not. Both r and the AUC could easily be interpreted, whereas problems were encountered with the OR. Based on these findings, both r and the AUC were then used to assess the level of linking accuracy that could be achieved when using one of three similarity coefficients -Jaccard’s coefficient (J), the simple matching index (S), and the Sorensen-Dice index (SD) - to measure across-crime similarity. Based on analyses using both serial burglary and serial rape data, the results demonstrated that, regardless of what measure of linking accuracy was used, no significant differences were found between the different similarity coefficients except in the case of S, in a single instance when it performed significantly worse than J and SD.