The CHAMP Center Predictive and Causal Modeling Laboratory is the CHAMP Center’s engine for developing the predictive and causal models that will eventually be transformed into prevention tools. These models are used to inform our understanding of how to identify children at risk for maltreatment exposure and its devastating effects, and to determine which factors to target with interventions to reduce such risks. This work involves close collaboration between our teams at the NYU Grossman School of Medicine and the University of Minnesota, who prepare and analyze existing datasets comprising many thousands of variables collected from many thousands of children.

Within these data sets, specific outcomes are identified, such as exposure to physical abuse or aggressive behavior. Then, the other variables in the data set are used by our modeling methods to determine the predictive and causal models that best fit the data.

Predictive modeling uses advanced Machine Learning Predictive Classification methods. We utilize Markov Boundary (MB) feature selection to identify the simplest and most accurate predictive model possible from the data. Models created this way are much easier to implement in practice than those without MB feature selection because they can generate highly accurate predictions using a small number of features, making practical assessment feasible.

Predictive models, even if they accurately determine a child’s risk level for an outcome, cannot reveal ways to reduce that risk because the true causes behind the outcome may confound the predictive factors in the model. Such factors cannot be targeted for intervention because changing a confounded factor will not alter the outcome.

Accordingly, we apply advanced Causal Data Science methods to identify intervention targets to reduce the risk of maltreatment or its consequences. These methods can safeguard findings from confounding from all measured variables and, in many cases, can identify whether unmeasured or latent causes confound a factor’s association with the outcome.

When a factor’s relationship with an outcome is shown to be unconfounded by both measured and latent variables, its causal status is assured. This allows the strength of this relationship to be used to estimate the factor's effect on the outcome without bias. When such effects are estimated without bias, they reflect the expected change in the outcome that would result from a specific change in the causal factor if intervention were applied. The CHAMP Center identifies factors that show promise for prevention by obtaining such unbiased estimates of effect.