Communicating limitations honestly

Presenting PA results responsibly means being as clear about what your model cannot do as about what it can. Stakeholders who receive your findings — whether program directors, caseworkers, policymakers, or the public — deserve an honest account of the model’s limitations. This page outlines the most important ones to communicate.

Predictions are probabilistic, not certain

A predicted probability of 0.75 means the model estimates a 75% chance of the outcome — not that the outcome will definitely occur. A student with a 0.75 predicted probability of dropping out may graduate; a student with a 0.20 probability may not. Individual predictions carry uncertainty, and that uncertainty should be part of how results are communicated.

This matters especially when predictions are used to inform decisions about specific individuals. Stakeholders should understand that a high predicted probability is a signal worth paying attention to — not a verdict.

The model reflects historical patterns

Your model was trained on historical data, which means it has learned patterns from the past. If those historical patterns reflect inequities — for example, if certain groups have systematically received fewer resources or faced harsher consequences — the model may reproduce those patterns in its predictions. This is one reason why bias evaluation is a critical part of the PA workflow, as covered in the previous section.

It also means the model may not perform equally well for all subgroups, particularly those that were underrepresented in the training data.

The model applies to a specific context

A model trained on data from one set of schools, communities, or program sites may not generalize well to different contexts. If your model is applied to a new population or a different time period, its performance should be reassessed rather than assumed to be unchanged.

Model performance can degrade over time

Even within the same context, a model trained on older data may become less accurate over time as circumstances change — student populations shift, programs evolve, and economic conditions fluctuate. This is one reason why ongoing monitoring after deployment is important, as discussed in the next section.

The model is a tool, not a decision-maker

Perhaps most importantly: a predictive model is one input into a decision, not the decision itself. It should complement — not replace — professional judgment, knowledge of individual circumstances, and consideration of ethical implications. Stakeholders should be clear on what the model is authorized to inform and what decisions remain with people.

Communicating these limitations is not a sign of weakness in your work — it is a sign of rigor and integrity. A stakeholder who understands what the model cannot do is better positioned to use it well.