AI Dynamic Models And Revalidation

AI dynamic models & revalidation (Concern 7)

ML-derived selection procedures enable analysis that wasn't feasible before, and many vendors refresh the algorithm frequently — sometimes after every test administration. Such "dynamic" procedures update validation evidence and normative data in near real time. That creates a dilemma and several operational problems.

The core dilemma

When a dynamic model changes, the change could reflect either:

• a real change in the nature of the applicant sample and/or the job (which should drive an update), or
• sample idiosyncrasies or instabilities over time that should not be capitalized on.

Distinguishing the two is hard — and capitalizing on instability degrades the model.

Why "dynamic" raises problems

1. Documentation must keep up. Validity evidence must be documented in a technical report. So every update to the underlying selection process requires updating the technical report (validity and normative-data characteristics). Out-of-date reports are problematic for legal defensibility and for HR records maintenance over time. (See technical-validation-report.)
2. Score adjustments & grandparenting. Each new selection process may require adjusting scores already in the database, or a policy on how to treat scores produced by different processes at different times. Continual adjustment creates administration problems: a candidate qualified today may be unqualified tomorrow (or vice versa). Changes can include altering predictor weights, adding/removing predictors, and changing interpretations (e.g., adjusting cutoff scores). Grandparenting policies are administratively hard to manage in high-volume programs. (See selection-decisions-and-scoring.)
3. Disparate-treatment / disparate-impact risk. A result of frequent change is that different candidates are evaluated on different variables depending on when they applied. If that variation relates to a protected characteristic, there's a disparate-impact specter; even without group-level disparate impact, disparate-treatment concerns can arise — and the appearance of such treatment can trigger applicant dissatisfaction and complaints.
4. Applicant-pool shifts. Large shifts in the applicant pool can materially affect reliability, validity, range restriction, or range enhancement on big data — so monitor key applicant-pool characteristics (demographic, educational). Future shifts in available technologies/data/algorithms may change the applicant database and indicate the ML model should be updated.
5. Defining group comparisons. Frequent algorithm changes make it harder to define relevant applicant pools for adverse-impact analysis or appropriate normative groups for comparison.

Revalidation & norms updating

Employers have always revalidated and updated norms; with AI the difference is the frequency. Traditionally, revalidation was triggered when the job changed, the test was compromised, the applicant pool shifted substantially, or enough time elapsed to question validity in a legal challenge — and it was undertaken at well-spaced intervals because it was laborious. Today's computing power makes continuous updating far less laborious, which raises a genuinely open question: how often should validation be refreshed to accommodate the nature of new applicant data?

Questions to ask (from the article)

• Are dynamic algorithms and norms useful?
• How should results from dynamic algorithms be documented to comply with existing and future legal and administrative requirements?
• How frequently should tests be revalidated and norms updated?
• What are the indicators that revalidation and updates to norms are needed?

Pitfalls

• Capitalizing on sample instability and mistaking it for real change.
• Letting technical reports/normative documentation fall out of date as the model drifts.
• Unmanaged score adjustments → candidates flipping qualified/unqualified; messy grandparenting.
• Evaluating different applicants on different variables → disparate-treatment exposure (and optics).
• Not monitoring applicant-pool shifts that invalidate prior validity/range assumptions.
• Unstable pool/normative definitions that make adverse-impact analysis incoherent.

Checklist

• Each model/norm update justified as real change, not instability
• Technical report and normative documentation updated with every change
• Score-adjustment / grandparenting policy defined and administrable
• Risk that candidates are scored on different variables (disparate treatment) assessed
• Applicant-pool characteristics monitored for shifts affecting validity/range
• Revalidation/norms-update cadence and trigger indicators defined
• Adverse-impact pools and normative groups remain definable under the update regime

See also

ai-validity-evidence · ai-selection-legal-landscape · generalizing-validity-evidence · technical-validation-report · administration-documentation (review/updating, records) · selection-decisions-and-scoring (cutoffs, norms)

Source: Tippins, Oswald & McPhail (2021), Concern: "Changes to Technologically Enhanced Systems" (Dynamic models and norms; Revalidation and norms updating).