7
The “Shrinkage” Debate In traditional statistics, the observed rate is thought to best represent the truth (n outcomes/k trials) Bayesian statistics considers observed data in the context of prior information Empirical Bayes derives prior information from the data e.g., the “best” guess is a rate somewhere between the observed rate and the overall rate Accomplished using hierarchical modeling

The “ Shrinkage ” Debate

  • Upload
    tana

  • View
    35

  • Download
    0

Embed Size (px)

DESCRIPTION

The “ Shrinkage ” Debate. In traditional statistics, the observed rate is thought to best represent the truth (n outcomes/k trials) Bayesian statistics considers observed data in the context of prior information Empirical Bayes derives prior information from the data - PowerPoint PPT Presentation

Citation preview

Page 1: The  “ Shrinkage ”  Debate

The “Shrinkage” Debate

• In traditional statistics, the observed rate is thought to best represent the truth (n outcomes/k trials)

• Bayesian statistics considers observed data in the context of prior information

• Empirical Bayes derives prior information from the data

– e.g., the “best” guess is a rate somewhere between the observed rate and the overall rate

• Accomplished using hierarchical modeling

Page 2: The  “ Shrinkage ”  Debate

Empirical Bayes Approach: Shrink to the average mortality

0%

20%

Mo

rtal

ity

rate

(%

)

Mortality rates forhigh-risk surgery 10%

Mo

rtal

ity

rate

(%

)

15%

5%

Overall mean mortality rate

Adjusted for reliability

Observed mortality rates

Box size is proportional to the hospital caseload

Page 3: The  “ Shrinkage ”  Debate

Traditional approach

• Widely used in performance measurement• CMS Hospital Compare website• Massachusetts Cardiac Surgery Report Card

• Advantages: • Innocent until proven guilty

• Disadvantages:• Assumes small hospitals are average• Ignores the volume-outcome relationship

Page 4: The  “ Shrinkage ”  Debate

“The Hospital Compare model [standard shrinkage] underestimates the typically poorer performance of low-volume hospitals”

Page 5: The  “ Shrinkage ”  Debate

Described methods for shrinking towards mortality for a hospital’s volume group

Page 6: The  “ Shrinkage ”  Debate

Composite Measure Approach: Shrink to the mortality for volume group

0%

20%M

ort

alit

y ra

te (

%)

Mortality rates forhigh-risk surgery 10%

Mo

rtal

ity

rate

(%

)

15%

5%

Observed mortality rates

Low volume

Medium volume

High volume

Mortality rates

Composite mortality

Page 7: The  “ Shrinkage ”  Debate

Which approach is better?

• When trying to identify the “best” hospitals, incorporating hospital volume is better • Center of excellence model• e.g., Leapfrog Group’s “Survival Predictor”

• But from a quality improvement perspective, volume is not actionable, so it may make sense to shrink to the mean• MBSC risk- and reliability-adjusted outcomes