METRIS™ Discuss an AI decision →
← All use cases

Use case 5 · Sensitivity analysis on national data

The denials nobody appeals

The pattern

Most denials are never appealed, so a plan never learns whether they would hold up. National data shows how large the untested group is.

Question
How many denials would fail if challenged?
Lane
All lanes
Basis
KFF national data

What METRIS does

  • Track overturn rates by service, vendor and AI tool version
  • Model which unappealed denials resemble overturned ones
  • Flag services where the hidden error is likely largest

Medicare Advantage prior authorization, 2024

Published data
Medicare Advantage prior authorization funnel 2024Decisions52.8 millionDenied4.1 millionNever appealed3.63 millionAppealed0.47 millionOverturned0.38 millionKFF. Appealed, overturned and never-appealed counts are our calculation from KFF's rates.

Of 4.1 million denials, only 0.47 million were appealed and 0.38 million partially or fully overturned. 3.63 million were never tested.

Method

hidden(r) = D · (1 − a) · r D = 4.1M denials, a = 11.5% appealed, r ∈ [0, 0.807]
Decisions
52.8 million (KFF)
Denied
4.1 million, 7.7%
Appealed
11.5% of denials
Overturned
80.7% of appeals, partially or fully

Selection into appeal is not random, so the true rate among unappealed denials is unknown. A sensitivity analysis shows the stakes across every plausible value instead of assuming one.

How many unappealed denials could be wrong

Published data · our sensitivity analysis
Wrongful denials among unappealed, by assumed rate0M1M2M3M0%20%40%60%80%Assumed share of unappealed denials that would be overturnedDenials that would not hold upOverturned on appeal: 0.38M0.36M at 10%0.91M at 25%

Appealed denials are not a random sample, so 80.7% is probably higher than the rate among unappealed denials, but no one knows by how much. Even at 10%, 0.36 million unappealed denials would not hold up, about as many as were overturned.

Reproducibility

Arithmetic on KFF's published rates; no random draws. Python with NumPy 2.5.3 and SciPy.

Sources

  1. KFF, Medicare Advantage Insurers Made Nearly 53 Million Prior Authorization Determinations in 2024, January 28, 2026

Study design and analysis: Suneeta Modekurty, SANJEEVANI AI

A statistical study of a pattern, not of any event. It identifies no organization or person. Simulated figures use synthetic data; published figures are industry-wide.