The Accuracy of Statistical Models to Link Serial Crimes: A Meta-Analysis

Abstract Purpose: Crime linkage can be expensive, time-consuming, and harmful to the well-being of law enforcement staff conducting it. Part-automation of this process has been explored recently, whereby statistical models could be used to prioritise potential crime links for human attention. This meta-analysis evaluates the overall effectiveness of statistical models at predicting crime links, as measured by the Area Under the Curve (AUC). Methods: Following a systematic literature search, 29 papers were included in the meta-analysis, and quality assessed. A three-factor random effects model was used to analyse the data for each crime category. Behavioural domains [across all statistical models used] that yielded the greatest accuracy for linkage predictions were further evaluated through a subgroup analysis for behavioural domain, which examined the difference in average AUC between statistical models used. Results: Most studies focused on the crimes of sexual assault and burglary. These studies also produced the largest number of effect sizes on which to base a meta-analysis. Results indicate greater accuracy when using behavioural domains related to geographical information or that aggregated all modus operandi (MO) behaviours. Further, the most effective statistical model to link crimes is dependent on the crime type. Conclusions: Some findings support the use of classification models over data reduction tools for the purpose of developing crime linkage decision-support tools. It is important for studies to highlight the practical value of statistical models used to identify linked crimes in research. For example, this could be accomplished by reporting the hit and false alarm rates of crimes identified as linked. There is sufficient evidence to support the use of geographical data to identify linked acquisitive offences, and the use of aggregated behavioural domains to identify linked sexual offences. However, research should now focus on statistical models and behavioural domains to identify linked offences for the under-researched crimes of homicide, car theft, robbery, and arson.