Diagnostics, Robustness and Model Validation
EVIDENCE LAB · 09
Diagnostics,
Robustness & Validation.
A model is not credible because it produces coefficients. Learn to interrogate assumptions, influential observations, alternative specifications and performance beyond the estimation sample.
TESTSTRESSVALIDATE
01 · READ THE RESIDUALS
What the model misses
leaves a pattern.
Random scatter
No visible structure around zero supports the chosen functional form.
Nonlinearity
A curved pattern signals that a straight-line specification misses systematic structure.
Heteroskedasticity
Residual variance changes with fitted values, invalidating conventional standard errors.
Influence
An unusual high-leverage observation may materially determine the fitted coefficients.
02 · ROBUSTNESS SEQUENCE
One estimate is a result.
A stable estimate is evidence.
Re-specify
Compare defensible controls, functional forms and transformations.
Re-estimate
Use robust or clustered uncertainty when the error structure requires it.
Stress-test
Examine influential observations, subgroups and alternative samples.
Validate
Evaluate prediction and calibration on data not used for estimation.
MODEL VALIDATION CONSOLE
Audit the model.
Earn the conclusion.
Adjust diagnostic conditions and activate corrective practices. ECONORIA updates the audit status and evidence-readiness score.
VALIDATION VERDICT
03 · VALIDATION PRINCIPLE
A model should fail
before policy does.
Internal validity
Are coefficients credible for the observed sample and design?
External validity
Will conclusions travel to other populations, places or periods?
Predictive validity
Does the model perform on genuinely unseen observations?
