Region Specific Geospatial Models for Prediction of Maltreatment Risk

Background and Purpose:Risk for child maltreatment is an important issue which has gained much momentum in both research and policy domains. Successful modeling of risk for maltreatment improves identification children with characteristics observable at birth by service providers. While prediction models for at risk children exist, previous work has not introduced localized geographic information of residence as a factor in predicting risk, which may contribute to differences in observed maltreatment risk. Omnibus risk models may oversimplify patterns of risk over large geographic areas, removing regional nuances in factors contributing to maltreatment risk. Generalized geographically weighted regressions are introduced as method for identifying region specific influence of risk factors in predicting maltreatment.

Methods:Administrative data from two record sources were probabilistically linked from California: (1) birth records for the 2006 cohort year and (2) child welfare records for the first 5 years of life (2006 through 2011). Customized probabilistic linkage algorithms were used to match records based on identifying information for both children and their parents. Geographic coordinates for residential address at birth were generated through a geocoding process applied to birth records. Generalized linear models were fit using birth record information to estimate risk for a report of alleged maltreatment on a training set and these models were applied to a test set to assess model stability. Similarly, generalized geographically weighted regressions were estimated for a training set and applied to a test set. Comparison of model accuracy and recommendations for inclusion of spatial data will be provided by presenters.

Results: Predictive risk models for maltreatment produced risks scores for the cohort population with an AUC = 0.809 (CI_95%= 0.801-0.817). GLM residual plots indicated model misfit with dense clusters of misfit geospatially located across several regions of California. Geographically weighted regression identified a collection of models, differentiated by coordinate locations, in which predictor weights were allowed to vary. Confidence intervals for parameter estimates were obtained from bootstrap samples and risk scores calculated. Regions with scores at the lower and upper quartiles were identified. Predictors such as mother’s citizenship, paternity, and low birthweight were interpreted as having variability across regions. Presenters will address regions with dense populations for community level variability in model parameter estimates.

Conclusions and Implications: The location of family residence at birth provides important information about the likelihood a child is reported for alleged maltreatment within the first five years of life. When spatial information is used to model risk, there are regional differences in the contribution of predictors to overall risk. These regional differences indicate specific factors may feature into risk quite differently across individuals based on the location of residence at birth. Furthermore, these regional differences in predictor weights do not universally indicate higher risk for all predictors in a given region. Results indicate the importance of locally weighted models when assessing risk to provide more targeted interventions to prevent maltreatment.