Handwriting classification in a forensic intelligence context using binary logistic regression (BLR) and classification & regression tree (CRT) models

Despite the digital transformation of society, handwritten documents continue to be collected in various investigations, such as fraud investigations and drug trafficking. Recent research highlighted that handwriting offers not only comparative value (i.e. helping address a source question), but it also has the potential to infer a writer's background profile (i.e. helping address other questions than source). This study compared binary logistic regression (BLR) and classification & regression tree (CRT) models to infer a writer's cultural background based on handwriting features. An experimental two-step modelling approach was employed to distinguish Australian, Korean, and Vietnamese writers (N = 196) using categorical handwriting features coded from scanned handwritten texts. The first step was classifying Australian from non-Australian writers, and the second step was further specifically classifying non-Australians into Korean and Vietnamese. The results demonstrated how a two-step modelling framework could be operationalised for early-stage writer classification and highlighted its practical strengths and limitations. The BLR model provided statistical depths for detailed interpretation, and it achieved higher classification accuracy, 93.4% and 97.8% in each step. The CRT model also achieved a high accuracy rate, but lower than BLR with 86.7% and 94.2%. Furthermore, blind test results reflected the practical challenges and strengths of each model. The CRT model correctly classified six out of seven blind specimens while the BLR correctly classified three out of seven blind specimens. Each model presented distinct strengths as the BLR model provided rich detailed statistical outputs, such as odds ratios and significance levels, while the CRT model offered greater accessibility and usability for non-statistical experts. These findings suggest that model selection should balance interpretability, robustness and accuracy. Although more work is required until such models can be applied in practice, this study highlights the potential to extract operational insights from handwriting beyond traditional comparison methods, supporting intelligence-led workflows even when no comparison material is available.