Can harm be predicted? On the development and validation of a statistical model for predicting harm in missing person incidents

A small but significant proportion of missing episodes result in serious harm or death. In this study, we developed and validated a statistical model for predicting missing incidents where harm occurs. Data were provided by two police forces in England and Wales for the period. January 2015 to December 2021. Of the 44,294 missing incidents we analysed, 4% were recorded by the police as resulting in harm (n =1902). Ten variables were found to significantly increase the risk of harm, including increased age, female sex, suicide ideation, mental health concerns and being harmed in a previous missing episode. What predicted harm was also shown to vary by age group. Using a standard train/test framework, our statistical model yielded an acceptable level of predictive performance – an area under the receiver operating characteristic curve score of 0.75 – but was not superior to the current police risk assessment method both in terms of recall (the proportion of harm cases that were successfully identified) and precision (the proportion of identified cases which actually resulted in harm). If generalisable, our findings (1) call for a re-examination of the questions currently used in police missing person risk assessments and (2) suggest that a validated risk prediction model can complement police decision making in missing person investigations.