JMIR Publications

ABSTRACT

Background:

Child maltreatment is associated with multiple negative outcomes at the individual and societal level. Children sufferers of maltreatment are at greater risk of a host of negative outcomes (e.g., psychological disorders, substance use, violent delinquency, suicidality, educational outcomes).

Objective:

In order to prevent and to ameliorate child maltreatment a combination of geospatial smoothing via a risk terrain modeling framework and counterfactual modeling are proffered here to identify risky areas and to determine optimal (re-)allocation of services to maximally improve maltreatment outcomes.

Methods:

A three stage process is proposed which can iteratively be applied within a collaborating jurisdiction to enable responsive and sustained achievement of identified child welfare outcomes. This process makes use of two analytic approaches: geospatial smooth- ing via a risk terrain framework and counterfactual modeling. Risk terrain modeling (RTM) is a spatial analytic approach that uses spatial machine learning methods to estimate the risk of maltreatment based on prior cases of maltreatment and risk factors of the built environment. Using prior validated cases of maltreatment, violent crime data and built environment data we estimate a series of machine learning models to geospatially smooth the historically identitied places at increased risk of child maltreatment. Areas identified as higher risk receive exten- sive services associated with preventing or limiting child maltreatment such as pre/postnatal care, subsidized daycare and parental counseling. We make use of counterfactual explanation modeling to optimally align service allocation to maximally improve maltreatment outcomes for future service allocations within a collaborating jurisdiction. This technique leverages a statistical model associating household-level information with maltreatment outcomes in order to explore combinations of services which would be predicted to achieve optimal and practical recommendations for future service allocation efforts.

Results:

The household level counterfactual recommendations described in the previous section will be aggregated into actionable recommendations. The primary actionable recommendation will directly inform future iterations of Stage 1 by providing a model-informed approach for determining eligible regions. There are several methods and strategies for aggregating the resulting household-level counterfactual service arrays. The first step is to select, for each household observation, which of the 2n service array permutations to select. The second step is to aggregate the household-level counterfactual service arrays into spatial subregions. These subregions could be the fishnet grids specified in Stage 1, or any other chosen subregion definition such as ZIP Code, census tract, or school district. The simplest method is to, for each service type, sum the number of households in the subregion that were determined to benefit from the service. Constraints can be introduced to this logic, such as service availability and cost. Algorithmic fairness is also a potential consideration during aggregation, with possibilities for both measuring and balancing metrics such as "recourse fairness".

Conclusions:

This protocol sets forth a novel approach for the allocation of supportive services for families at risk of child maltreatment through geospatial smoothing via a RTM framework and the maximization of service impact through CEM. Child maltreatment is an unfortunate, and ubiquitous, issue in the United States. This proposal builds on jurisdiction-wide public health strategies in order to allocate services in a data-informed fashion and further align future iterations of the allocation strategy using outcomes-based counterfactual modeling at the household level. The flexibility of the proposed methodology enables its application regardless of the collaborating jurisdiction’s preferences and constraints.