Flag Predictions Outside the Observed Predictor Support

Define the support of the fitted data

A prediction can fall within each predictor’s separate range while representing a combination scarcely observed in the data. Check both marginal ranges and joint coverage. A smooth model output does not reveal whether the requested scenario is supported by observations or driven mainly by functional-form assumptions.

NIST’s model-validation guidance emphasizes checking model adequacy beyond a fit statistic. Extend that discipline to the prediction inputs: a model fitted well in one region may behave poorly when asked to describe another.

Inventory the requested scenarios

List predictor values, units, categorical levels and missing-input handling for every reported prediction. Compare them with the fitted sample. Flag new categories, values outside training ranges and unusual combinations of otherwise familiar values.

For longitudinal or clinical prediction, also check whether the deployment setting, measurement method and follow-up horizon match the development data. A numerical range check cannot detect all population shifts, but it provides an auditable first screen.

Inspect a hypothetical combination

Suppose the fitting data contain adults aged 20–80 and doses from 0–100 units. All doses above 60 were observed only among participants younger than 50. A prediction for an 80-year-old at dose 100 lies within both marginal ranges yet lacks direct support for that joint combination.

The model will often return a number anyway. Label the scenario as extrapolative with respect to the joint support, inspect the assumptions driving it, and avoid presenting its interval as if it captures every uncertainty about transport to that combination. Parameter uncertainty is not the same as uncertainty about an unsupported functional form.

Decide how to report unsupported regions

Limit primary displays to a scientifically justified range where possible. If extrapolation is central to the question, state it explicitly and examine alternative plausible model forms or obtain additional data. Do not conceal unsupported regions by drawing an uninterrupted curve across them.

Use a data-density display, rug plot or clear range annotation when it helps readers see coverage. Preserve the exact scenario grid with the analysis. The final prediction report should separate what the observed data directly inform from what the model extends through assumptions, allowing users to judge whether those assumptions are acceptable for the intended decision.

Source verification: 6 November 2026. Examples are hypothetical.