The purpose of the paper is to expand upon the RAAM framework by identifying the specific conditions that “predict” the presence of positive supportive family relationships through CART analyses (a Machine Learning approach). We propose and test a theory-driven data mining approach.
Methods. Data (n = 1,047) was obtained from one of the largest studies of homeless youth (ages 13-24) in the United States recruited from three drop-in centers between October of 2011 and June of 2013.
Results. In regard to family support, the pruned CART analyses suggests that youth who indicate an informal living situation, saw a deceased person, and did not move to LA to pursue the entertainment business have one or more supportive family members. On the contrary, youth who report hard drug use, witnessed the physical attack of a family member, and whose living arrangements are not informal do not identify a family member as a support person. For the training sample, the sensitivity was 83 % and the specificity was 61 %. For the validation sample that examined the remaining sample, the sensitivity was 74 % and the specificity was 42 % . The area under the curve (AUC) =0.77 for the training sample and AUC = 0.63 for the validation sample. The misclassification rates for developmental and validation CART models were 26 % and 39 %, respectively.
Conclusions and Implications. These findings that youth who have informal living situations may have more supportive adults in their networks and thus, future interventions for this particular subgroup may want to focus on strengthening and supporting familial relationships. However, the study also indicates that family reunification interventions may not be appropriate for youth who are literally homeless and witnessed family violence and use hard drugs. Moreover, ML models when informed by theory can uncover new insights, particularly moving beyond simple direct linear effects. Our results show that complex interactions among several variables differentiate those YEH with supportive family from those who do not. Future work may benefit from the classification of subgroups with unique sources of risk and resilience not only among homeless youth but other marginalized populations in need of intervention refinement.