Understanding and communicating PA results

At this stage in the PA workflow, you have built a model, evaluated its predictive performance, and assessed potential bias — all using held-out test data. You now have what the workflow was designed to produce: a predicted probability of the outcome for each individual in your test data.

But having predictions isn’t the same as understanding them — or being able to explain them to others. This section focuses on two closely related goals:

• Understanding: What patterns do your predictions reveal? What is driving them?
• Communicating: How do you explain those patterns clearly to stakeholders — program staff, leadership, or policymakers — who need to act on the results?

These goals are practical, not just technical. Stakeholders will have questions. They’ll want to know who the model identifies as high-risk, what factors are driving those predictions, and whether the patterns make sense given what they know about the population. Being able to answer those questions clearly is just as important as building a model that performs well.

This section explores three questions that are especially useful for both understanding and communicating PA results:

1. How are predicted probabilities distributed within and across groups?
2. Who has low or high predicted probabilities of the outcome?
3. What are the most important predictors in my model?

Together, these questions help you describe what the model is showing, who it flags as higher- or lower-risk, and what factors are driving those predictions — all of which are essential for making informed decisions about how to improve services, and for explaining your model to the people who will use it.