34
Bayesian Hospital Mortality Rate Estimation and Standardization for Public Reporting Edward I. George Veronika Roˇ ckov´ a, Paul R. Rosenbaum, Ville A. Satop¨ a and Jeffrey H. Silber The Wharton School, University of Pennsylvania [email protected] Workshop on New and Evolving Roles of Shrinkage in Large-Scale Prediction and Inference Banff International Research Station, Canada April 11, 2019 1 / 34

Bayesian Hospital Mortality Rate Estimation and Standardization … 1... · 2019. 4. 12. · Reported indirectly and directly standardized mortality rates Major di erences between

  • Upload
    others

  • View
    0

  • Download
    0

Embed Size (px)

Citation preview

Page 1: Bayesian Hospital Mortality Rate Estimation and Standardization … 1... · 2019. 4. 12. · Reported indirectly and directly standardized mortality rates Major di erences between

Bayesian Hospital Mortality Rate Estimationand Standardization for Public Reporting

Edward I. George

Veronika Rockova, Paul R. Rosenbaum, Ville A. Satopaa and Jeffrey H. Silber

The Wharton School, University of Pennsylvania

[email protected]

Workshop on New and Evolving Roles ofShrinkage in Large-Scale Prediction and Inference

Banff International Research Station, CanadaApril 11, 2019

1 / 34

Page 2: Bayesian Hospital Mortality Rate Estimation and Standardization … 1... · 2019. 4. 12. · Reported indirectly and directly standardized mortality rates Major di erences between

Prologue - Is shrinkage estimation really safe?

Observe xi | µi ∼ N(µi , vi ), i = 1, ..., p independently.

Hierarchical model: µ1, . . . , µpiid∼ N(µ, v).

E [µj | data] = vvi+v xi + vi

vi+v µ

eB shrinkage estimator: µj =vµ

vi+v xi + vivi+v x

µj strongly shrinks xi towards x when vi >> v

Justifications for this estimation rely critically on the priorassumption that

µ1, . . . , µp all have the same mean µ and variance v !

2 / 34

Page 3: Bayesian Hospital Mortality Rate Estimation and Standardization … 1... · 2019. 4. 12. · Reported indirectly and directly standardized mortality rates Major di erences between

Some History - The National Halothane Study (1969)

856,500 surgeries under anesthesia at 34 Hospitals from1959-1962

16,840 deaths within 6 weeks of surgery (about 2%)

Compared the effect of general anesthetic agents,especially halothane, on postoperative mortality

Reported indirectly and directly standardized mortality rates

Major differences between hospitals were observed,even after standardization

3 / 34

Page 4: Bayesian Hospital Mortality Rate Estimation and Standardization … 1... · 2019. 4. 12. · Reported indirectly and directly standardized mortality rates Major di erences between

Copyright © National Academy of Sciences. All rights reserved.

National Halothane Study: a Study of the Possible Association Between Halothane Anesthesia and Postoperative Hepatic Necrosis; Report. Edited by John P. Bunker [and Others]http://www.nap.edu/catalog.php?record_id=19006

4 / 34

Page 5: Bayesian Hospital Mortality Rate Estimation and Standardization … 1... · 2019. 4. 12. · Reported indirectly and directly standardized mortality rates Major di erences between

Copyright © National Academy of Sciences. All rights reserved.

National Halothane Study: a Study of the Possible Association Between Halothane Anesthesia and Postoperative Hepatic Necrosis; Report. Edited by John P. Bunker [and Others]http://www.nap.edu/catalog.php?record_id=19006

5 / 34

Page 6: Bayesian Hospital Mortality Rate Estimation and Standardization … 1... · 2019. 4. 12. · Reported indirectly and directly standardized mortality rates Major di erences between

Copyright © National Academy of Sciences. All rights reserved.

National Halothane Study: a Study of the Possible Association Between Halothane Anesthesia and Postoperative Hepatic Necrosis; Report. Edited by John P. Bunker [and Others]http://www.nap.edu/catalog.php?record_id=19006

6 / 34

Page 7: Bayesian Hospital Mortality Rate Estimation and Standardization … 1... · 2019. 4. 12. · Reported indirectly and directly standardized mortality rates Major di erences between

Public Reporting of Hospital Mortality Rates Today

Medicare’s web based “Hospital Compare”:

To provide the public “with information on how well thehospitals in your area care for all their adult patients withcertain medical conditions” such as heart attacks.(U.S. Department of Health and Human Services, 2007)

Available at http://www.hospitalcompare.hhs.gov

7 / 34

Page 8: Bayesian Hospital Mortality Rate Estimation and Standardization … 1... · 2019. 4. 12. · Reported indirectly and directly standardized mortality rates Major di erences between

A Typical Medicare Mortality Rate Report

Figure: Comparing heart attack mortality rates with Hospital Compare8 / 34

Page 9: Bayesian Hospital Mortality Rate Estimation and Standardization … 1... · 2019. 4. 12. · Reported indirectly and directly standardized mortality rates Major di erences between

Medicare Public Reporting

Hospital Compare 2008: Out of 4311 hospitals, 4302 of them(99.8%) are “no different than U.S. National rate” and zerohospitals are ”worse than U.S. National rate”.

How did Hospital Compare reach these conclusions?

- The smaller the hospital volume, the more its mortality rateestimate is “shrunk” to the overall mean.

- Medicare’s justification: Estimates for small volume hospitalsrely on the pooled data of all hospitals: “this pooling affordsborrowing of statistical strength that provides more confidencein the results.”

Hospital Compare’s approach is being copied as the“Gold Standard” for general performance comparisons.

UH-OH! Hospital Compare’s estimates contradict the conventionalwisdom that mortality rates are higher at low volume hospitals!!!

9 / 34

Page 10: Bayesian Hospital Mortality Rate Estimation and Standardization … 1... · 2019. 4. 12. · Reported indirectly and directly standardized mortality rates Major di erences between

Hospital Compare’s Reported Mortality Rates by Volume

●●

● ●

●●

●●

● ●

●●

● ●

● ●

●●

●●

●●

●●

● ●

●●

● ●●

●●

●●

● ●

●●

●●

● ●

●●

●●

●●

● ●●

● ●

●●● ●

●●

●●

●●

● ● ●

●●

●●

●●

●●

●●

●●

●●

●●

● ●

● ●

●●

●●

●●●

●●

●●

●●●

●●

●●

●●

●●

● ●●

● ●

●●

●●

●●

●●

● ●● ●●

● ●

●●

● ●● ●

●●●

● ●

●●

● ●

●●

●●

●●

●● ●

●●

● ●●

●●

●●

●●

●●

● ●●

●●

●● ●

●● ●

● ●●

●●● ●

●●

● ●

●●

●●

●●

● ●

●●

●●

●●

● ●

●●

●●

●●

●●●

● ●

●●

● ●

●●

●●

● ● ●●

● ●●

●●

● ●

●●●●

●● ●

●●

●●

● ●

●●●●

●●

●●●

●●

●●

●●

● ●●

●●

●●

●●

● ●

●●

●●

●●

●●

●●

●●

● ●

●●

●●

●●

● ●

● ●

●●

●●●

●●

●●

●● ●

●●

●●

● ●

●●

●●

●●● ●

●●

●●

●●

●●

● ●

●●

●●

●●

●●●

●●

●● ●

●●

● ●

● ●

●●

●● ●

●●

●●●

● ●

●●

●●●

● ●●

●●●● ●

●●

● ●

●●

●●

● ●

●●

● ●

●●●●●

●●

●●

● ●

●●

●●

●●

● ●●

●●

●●

●●

●●

●●

●●

● ●

● ●

●●

● ●

●●●●

● ●

●●

● ●

●●

●●

● ●

● ●●

●●

● ●

●●

●●

● ●●

● ●

●●●

●●

●●

●●

●●

● ●●●

●●

●●

● ●

●●

● ●

●●

●●

●●

● ●

●●

●●

●●

● ●

●●

●●

●●

●●

●●

●●●

●●

●● ●

●●

●●

●●

●●

● ● ●

● ●

●● ●

●●

●●

●●

●●

●●

●●

● ● ●

●●

● ●

● ●

●●

●●

●●

●●●

●●

●●

●● ●

●●

●●

●●

●●

●●

●●

●●●

●●

●●

●●

●●

●●

●●

●●

● ●

●●

●●

●●●

●●

●●

●●

●●

●●

●●

● ●

● ●

● ●

● ●

● ●

●●

●●

● ●

●●

● ●

●●

●●

● ●

● ●

● ●

●●

●●

●● ●

● ●

● ●●

● ●

●●

●●

● ●

●●

● ●●

●● ●

●●

●●

●●

●●

●●

● ●

● ●

●●

●●

●●

●●

●●

●●

●●

●●●

●●

●●

●● ●

● ●

●●

● ●

● ●

●●

●●

● ●

●●

● ●

● ●

●●

●● ●

●●

●●

●●

● ●

●●

●●

● ●

●●●

●●

●●

●●

●●

●●

●●

●●

●●

●●

●●

●●

●●

●●

● ●

● ●●

●●

●●●

●●

●●

●●

●●

●●

●●

●●

●●

●●

● ●

●●

●●

● ●

●●

●●

●●

● ●

●●

●●

●●

● ●●

●●●

●●

● ●

●●

●●

● ●

●●

●●

●●

●●

●●

●●

●●

●●

●●

●●

●●

● ●

● ●

● ●

● ●●●

●● ●●

●●

●●

● ●

●●

●●

● ●

●●

●●

●●

●●

●●

● ●

●●

●●

●●

●●

● ●

●●

●●●

● ●

●●

●●

● ● ●●

●●

●●

●●

● ●

● ●

● ●

●●

●●

●●

●●

●●

●●

●●

●● ●

●●●

●●

● ●

●●

● ●

●●

● ●

● ●

● ●

●●

●●

1 5 10 50 100 500 1000

0.0

0.2

0.4

0.6

0.8

1.0

Volume

O

Figure: Observed Hospital Rates

●● ●

●●

●●

●●

● ●

●●

●●

●●

●● ●

●●●

● ●

●●

●●

●●

● ●

●●

●● ●

●●

●●

● ●

●●

●●

●●

●●

●●

●●

●●

●●

● ●●

●●

●●

● ●

●● ●

●●

●●

●●

●●

● ●●

● ● ●

●●

●●

● ●

●●

●● ●

●●

● ●●

●●

● ●

●●

● ●

●●

●●

●●

●●

●●

●●

●●

●●

● ●●

● ●●

●● ●

●●

● ●

● ● ●●

●●

●●

●●

● ●●

●●

●●

●●

●●

●●

●●

●●

●●

●●

●●●

●●

●●

●●

●● ●

●●

●●●

●●

●●

●●

●●●●

● ●

●●

●●

●● ●

●●

●●

● ●

●● ●●

● ●

●●

●●

●●

●●

●●

●●

● ●

●●

●●

●●

●●

●●

●●

●●

●●

●●

●● ● ●

● ●●

● ●

● ●●

●●

●●

● ●

●●

● ●

● ●

●● ●

●●

●●

●●

●●

●●

● ●

●●

●●

● ● ●

●●

●●

●●

●●

●●

●●

● ●

●●

●●

●●

●●

●●

●●

●●●

●●

● ●

●●

●●

● ●

●●●

● ●

●●

●●

●●

●●

●●

●●

●●

●●

●●

●●

●● ●●

●●●

●●

●●

● ●●

●●

●●●

●●

● ●

●●

●●

●●

●●

●●

● ●

●●

●●

●●

●●

●●

●●

● ●

●●●

●●

● ●

●●

●● ●

●●

●●

●●

●●

●●●

●● ●

●●

●●

●●

●●

●● ●● ●● ●

●●

● ●

●●●

●●

●●

●●

●●

●●●

●●

●●

●●

●●●

●●

●●

●●

●●● ●

●●●

●●

●●

●●

●●●● ●● ●

● ●●●●●

●●

●●

●●●

●●

●●

●●

●●

●●

●●

● ●

●●

●●

●●

●●

●●●

●●

●●

● ●●

●●

● ●

●●

●●

●●

●●

●●

● ●●

●●

●●

●●

●●

● ●●

●●

●●

●●

●●

●●

●●

● ●● ●●

● ●●

●●

● ●

●●

●●

●●

●●●

●●●

●●

●●

●●

●●

●●

●●

●●

●● ●

● ●

●●

●●

●●

●●

●●

●●

●●

● ●

●●

●●

●●●

●●

●●

●● ●●

●●

●●

●●

●●

●●

●●

●●

●● ●

●●

●●

●●

●●

●●

● ●

●●

●●

●●

● ●●

● ●

●●

●●●

● ●

●●

●●

●●●

●● ●●

● ●●● ●

●●

●●

●●

●●

●●

●●

●●●

●●

●●●

●●

●●●

●●

●●

●●

●● ●

● ●

●●

●●

●●●

●●

● ●

●●

●●

●●

●●

● ●

●●

●●

● ●

● ●

●●

●●

●●

●●

●●

●●

●●●

●●

●●

●●

●● ●

●●●

●● ●

●●

●●

●●

● ●

● ●

●●

●●

● ●

●●●

●●

●●

●●

●●

●●

●●

●●

●● ●

●●

●●

●●

● ●●

●●

●●●

●●

●●

●●●●

● ●●●

●●

● ●

●●

● ●

●●

●● ●

●● ●

●●

●●

● ●●

●●

●●

● ●

● ● ●

●●

●●

●● ●

●●

●●

●●●

●●●

●●

●●

●●

●●

●●

●●

●●

●●

●●

●●

●●

●●

●●

● ●

● ●

●●

● ●

●●

● ●

●●

●●

●●

●●

● ●

●●

● ●

● ●

● ●

●●

●●

●●

●●●

●●

●●

●●

●●

●●

●●

●●

● ●●●

●●

●●

●●●

●●

●●●●●

●●

●● ● ●

●●

●●

●●

● ●

●●

● ●

●●

●●●

●●

●●

●●●

●●

●●

●●

●●

●●

●●

●●

● ●

●●

●●

●●

● ●

●●

●●

●●

● ●●

●●

●●

●●

● ●●

●●

● ●

● ●

●●

●●

●●●

●●

●●

●●

● ●●

●●

●●

●●

●●

●●●

●●

● ●

●●

●● ●

● ●

●●

●●

●●

●●

●●●●

●●

●●

●●

●● ●

●●

● ●

●●

●●

●●

● ●

●●

● ●

●● ●

●●

●● ●

●●

●●

● ●

●●

●●

●●

●●

●●

●●

●●

●●

● ●

●●

● ●

●●

●●

●● ●

●●

●●

●●

● ●●●

●●

● ●

●●

●●

●●

●●

● ●

● ●

●●

●●●

●●

●●●●

●●

●●

●●●

●●

●●

●●

●● ●

●●

●●

●●

●●

●●

●●

●●

● ●

●●

●●

●●

●●

●●●

● ●

●●

●●

●●

●●

●●

●●

● ●

●●

●●

●●

●●

●● ●

● ●

●●

●●

●●

●●

● ●

●●

● ●

●●

●●

●●

●●

●●

● ●

●●

●●

●●

●●

● ●

● ●

●●

●●

●●

●●

●●

●●●●

●●

●●

●●

●●

●●

●●

●●

●● ●

●●

●●

●●

●●

●●

●●●

● ●

●●●●

●●

● ●

●●

●●

●● ● ●

●●

● ●●

● ●● ●

●●●

●●

●●

●●

● ●

●●

●●

●●

●●

●●

●●

●●

●● ●

● ●

● ● ●●

●●

●●

● ●●

●●●

●●

●●

●●

●●

●●

●●●

●●

●●

● ●●

●●

●● ●

●●

●●

●●

●●

●●

●●

●●

●●

●●

●●

● ●●

●●

●●●

● ●

●●

●●

●●

●●

●●

●●

●●

●●

●●

●●

●● ●

●●

●●

●●

● ●

●●

●●●

● ●

●●

●●

●●

●●●

●●

1 5 10 50 100 500 1000

0.10

0.15

0.20

0.25

0.30

Volume

PIS

Figure: Reported Hospital Rates

Two Major Steps:

1 Hospital Compare begins with a log linear random effects model topredict hospital mortality rates.

2 These rates are then standardized (indirectly) to adjust for patientcase-mix differences. 10 / 34

Page 11: Bayesian Hospital Mortality Rate Estimation and Standardization … 1... · 2019. 4. 12. · Reported indirectly and directly standardized mortality rates Major di erences between

Administrative Data from Medicare Billing Records

Medicare data on AMI (Acute Myocardial Infarction) cases fromJuly 1, 2009 to December 31, 2011.

377, 615 AMI patients admitted to 4, 289 hospitals.

56, 567 deaths within 30 days of admission (about 15%)

27 patient characteristics (e.g. age, heart failure, hypertension etc)

4 hospital characteristics (volume, resident-to-bed ratio,nurse-to-bed ratio, PCI)

Training: first 2 years. Validation: remaining 6 months.

11 / 34

Page 12: Bayesian Hospital Mortality Rate Estimation and Standardization … 1... · 2019. 4. 12. · Reported indirectly and directly standardized mortality rates Major di erences between

Hospital Compare’s Random Effects Model

log

(phj

1− phj

)= αh + x ′hjβ,

where

phj = P(Yhj = 1): 30-day mortality rate at hospital h for patient j ,(h = 1, 2, . . . ,H and j = 1, 2, . . . , nh).

αh ∼ N(µ, σ2): hospital random effects.

x ′hjβ: patient fixed effects (based on patient characteristics xhj).

Fit using PROC GLIMMIX in SAS, Krumholz et al. (2006ab).

12 / 34

Page 13: Bayesian Hospital Mortality Rate Estimation and Standardization … 1... · 2019. 4. 12. · Reported indirectly and directly standardized mortality rates Major di erences between

Hospital Compare Model Estimates

●●●

● ●

● ●

●●

●●

●●

● ●

●●

●●

●●

● ●

●●

●●

●●

●●●

●●

●●

●●

●●●

●●

●●

● ●

● ●

● ●

●●

●●

●●

●●

●●

●●

●● ● ●

●●

● ●

●●

● ●

●●● ●

●●

●●

● ●

●● ●

● ●●●●

● ●

●●

● ●

●●

●●

● ●

● ●

●●

●●

●● ●

● ●

● ●

● ●

●●

●● ●

●●

● ●

●●●

●●

●●

● ●

●●

●●

● ●

●●

●●

●●

●●

●●●

● ●●

●●

● ●

●●

● ●

●●

●●●

● ● ●●

● ●

●●

●●

●●

●●

●●

●●

● ●

● ●

●●

●●

●●

●●

●●

●●

●●

●●

● ●

●●

● ●

●●

● ●●

●●

●●

●●

●●●

●●●

●●

●●

●●

●●

● ●●●

●● ●

●●

●●

●●

● ●

● ●

●●

●●

●●

● ●

● ●

●●

●●●

●●

●●

●●

●●

●●

●●

●●

●●

●●

●●

●●

●●

●●

●●

●●

●●

●●

●●

●●

●●

●●

●●

●●

●●

● ●

●●

●●

●●

●●

●●

● ●

●●

●●

●●

● ●

●●

●●

●●

●●

● ●●

●●

●●

● ●

●●

●●

●●

● ●●

●●

●●

●●

●●

● ●

●●

●●

●●

●●

●●

●●

●●

●●

●●

●●

●●

●●

●●

●●

● ●

●●

●●

●●

● ●

●●

●●

●●

● ●

●●

●●

●●

●●

●●

●●

●●

●●

●●

●●

● ●●

●● ●

●●

●●

●●

● ●

●●

●● ●

●●

●●

●●

●●

●●

● ●

●●

●●●

●●

● ●●

●●

●●

●●●

● ●

●●

● ●

●●●

●●

●●

●●

● ●

●●

●●

●●●

●●

● ●

●●

●●

●●

●●

●●

●●

●●●

●●

● ●

●●

●●●

●●

●●

●●

●●

●●

●●

●●

●●

● ●

●●

●●

●●

●●

●●

●●

●●

●●

● ●

●●

● ●

●●

●●

●●

●●

●●●

● ●●

●●

●●●

● ●

●●

●●

●●

● ●

●●

●●

●●

●●

●●

●●

●●

●●

●●

●●● ●

●● ●

●●

●●

●●

●●

●●

●●

●●●

●●

●●

● ●

● ●●

● ●

●●●

●●

●●●

●●

●●

● ●

● ●

●●

●●

●●

●●

●●

●● ●

●●

●●

●●

●● ●

●●

●●

●●●

●●

●●

●●●

●●

●●

●●

●●

●●●

●●

●●

●●

● ●

● ●

●●

●●

●●

●●

●● ●

●●

●●

●●

●●

● ●

●●●

●●

● ●

●●

●●

●●

●●

● ●

●●

● ●

● ●

●●

●●

●●

●●

● ●

●●

●●

●●

●●

●●

● ●

●●

●●

●●

● ●

●● ●

●●

●●

●●

●●

●●

●●

● ●

● ●

●●

●●

●●

●●

●●

●●●

●●

● ●

●●

●●●

●●

●●

●●

●●

●●

●● ●

●●

●●

●●

●● ●

● ●

●●

●●

●●

●●

●●

●●

●●

● ●●

●●●

●●

●●

● ●

●●

● ●

●●

●●

●●

●●

●●

●●

●●

● ●

●●

●●

●●

●●

●●

●●

●●

● ●

●●

●●

●●

●●

●●

●●

● ●

● ●

●●

●●

●●

●●

●●

● ●

●● ●

●●

●●

●●

●●

●●

● ●

●●

●●

●●

●●

●●

●●

●●

●●

●●

●●

●●

●●

●●

● ●

● ●

●●

●●●

●●

●●

●●

●●

●●

●●

● ●

●●

●●

● ●●

●●

●●

●●

●●

● ●

● ●

●●

●●

●●

●●

● ●

●●

●●

● ●

●●

●●

●●

● ●

● ●

●●●

●●

●● ●

●●

●●

●●

●●

● ●

●●

● ●

●●

●●●

●●

●●

●●

●●

1 5 10 50 100 500 1000

0.0

0.2

0.4

0.6

0.8

Volume

P

Figure: Ph Hospital Mortality Rates

● ●

●●

● ●

●●

●●

●● ●

●●

● ●

●●

●●

● ●

●●

● ●

●●

● ●

●●

●●

●●

●●

●●

● ●●

●●

●●

●●

● ●

●●

●●

●●

●●

● ●●

● ●●

●●

● ●

●●

●● ●

●●

●●

●●

●●

●●

●●

●●

●●

●●

●●

● ●●

●●●

●●

●● ●

● ●

● ●

● ● ●●

●●

●●

● ●

●●

●●

●●

●●

●●

●●

●●

●●

● ●

● ● ●●

●●

●●

● ●

●●●●

● ●

●●

●●

● ●

●●

● ●

●● ●

● ●

●●

●●

●●

●●

●●

●●

●●

●●

●●

● ● ●

● ●●

●●

●●

●●

● ●

●●

●●

● ●

●●

●●

●●

●●

●●

● ● ●

●●

●●

●●

● ●

●●

●●

●●

●●●

● ●

● ●

●●

●●

●●

● ●

●●

●●

●●

●●

●●

●●

●●●●

●●●

● ●

●●

●●

●●

●●

●●

●●

●●

●●

●●

●●

●●

●●

●●

●●

●●

●●

●●

●●

●●

●●

●●

●●

● ●

●●

●●

●●

●●

●● ●●●

● ●●

●●

● ●

●●

●●●

●●

●●

●●

●●

●●

●●

●●

●●● ●

●●●

●●

●●

●●

●●●●●

● ●

●●

●●●●

●●

●●

●●

●●

●●

●●

●●

● ●

●●

●●

●●

● ●

● ●

● ●

●●

●●

● ●●

●●

● ●

●●

● ●●

●●

●●

● ●

●●

●●

●●

● ●●

● ●●

● ●

● ●

●●

●●●

●●●

●●

●●

●●

●●

●●

●● ●

●●

●●

●●

●●

●●

●●

●●

● ●

●●

●●●

●● ●●

●●

●●

●●

●●

●●

●●

●●

●● ●

●●

●●

●●

●●

● ●

●●

●●

● ●●

● ●

●●

●●

●●

●●

●●

●●

●● ●●

● ●●● ●

●●

●●

●●

●●●

●●

●●●

●●

●●

●●

●●

●●

●●●

●●

●●

●●●

●●

● ●

●●

●●

●●

●●

● ●

●●

●●

● ●

● ●

●●

●●

●●

●●

●●

●●●

●●

●●

●● ●

●●●

●●●

●●

● ●

●●

●●

●●

● ●

●●●

●●

●●●

●●

●●

●●

●●

●●

●●

● ●

●●

●●

●●

●●●

●●

●●

●●

●●

●●

●●

●●

●●

●●

● ●

● ●●

● ●

●●●

●●

●●

●●

●● ●

●●

●●●

●●●

●●

● ●

●●

●●

● ●●

●●

●●

● ●

●●

●●

●●

●●

● ●

● ●

●●

●●

● ●

●●

●●

●●

●●

● ●

●●

●●

●●

●●

●●●

●●

●●

●●

●●

● ●●●

●●

●●

●●

●●●●

●●●

●●

●●

●●

●●

●●

●●

●●

●●

●●●

●●

●●●

●●

●●

● ●

●●

●●

●●

●●

● ●

●●

●●

● ●●

●●

●●

●●

● ●

●●

●●

● ●

● ●

●●

●●

●●

●●

●●●

●●

●●

●●

●●●

● ●

●●

●●

●●

●●

●●

●●

●●

●●

●●

● ●

●●

●●

●●

●●

● ●

●●

● ●

●●

●●

●●

●●

●●

●●

●●

●●

●●

● ●

●●

● ●

●●

●●

● ●

●●

●●

●●

● ●

●●

● ●

● ●

●●

●●

● ●

●●

●●

●●●

●●

●●

●●

●●●

●●

●●

●●

●●

●●

●●

● ●

●●

● ●

●●

●●

●●

● ●

●●

●●

●●

●●

● ●

●●

● ●

●●

●●

●●●

● ●

●●

●●

●●

●●

● ●

●●

●●

●●

●●

●●

● ●

● ●

●●

●●

● ●

●●

●●

●●

●●

●●

●●

●●

●●

●●

● ●

●●

●●

●●

●●

● ●

●●

●●

● ●

●●

●● ● ●

●●

● ●●

●●

● ●

●●

●●

●●

● ●

●●

●●

●●

●●

●●

●●

● ●●●

●●

●●

● ●

●●●●

●●

●●

●●

●●

●●

●●

●●

●●

●●

●●

● ●

●●

● ●

●●

●●

● ●

●●

●●

● ●●

●●

●●

● ●

●●

●●

●●

●●

●●

●●

●●

●●

●●

●●

●●

●●

●●

●●

●●

●●

●●

1 5 10 50 100 500 1000

−6.5

−6.0

−5.5

−5.0

Volume

αFigure: αh Hospital Effects

Ph = 1nh

∑nhj=1 phj have shrunk the raw observed mortality rates.

Strong shrinkage of the αh’s for small volume hospitals.

Under this model, mortality rate Ph variation at small volumehospitals driven primarily by patient case-mix differences.

13 / 34

Page 14: Bayesian Hospital Mortality Rate Estimation and Standardization … 1... · 2019. 4. 12. · Reported indirectly and directly standardized mortality rates Major di erences between

Shouldn’t Hospital Characteristics be in the Model?

Hospital effects have been modeled as αh ∼ N(µ, σ2),completely random with the same mean and variance!

This assumption is leading to the strong shrinkage of the αh’sfor small volume hospitals!

Available hospital characteristics have been left out!

Proponents of the Hospital Compare approach argue thatincluding hospital characteristics in the model would be“unfair” to the hospitals.

But isn’t excluding hospital characteristics “unfair” to thepeople seeking accurate information?

14 / 34

Page 15: Bayesian Hospital Mortality Rate Estimation and Standardization … 1... · 2019. 4. 12. · Reported indirectly and directly standardized mortality rates Major di erences between

Will Adding Hospital CharacteristicsMake a Difference?

15 / 34

Page 16: Bayesian Hospital Mortality Rate Estimation and Standardization … 1... · 2019. 4. 12. · Reported indirectly and directly standardized mortality rates Major di erences between

Suppose we elaborate αh ∼ N(µ, σ2) to αh ∼ N(µh, σ2h)

µh and σ2h can now be functions of hospital characteristics!

Model µh σ2h

(C,C) Constant Constant(L,C) Linear(log volh) Constant(S,L) Spline(log volh) Log-Linear(volh)(SL,L) Spline(log volh)+Linear(ptcah, ntbrh, rtbrh) Log-Linear(volh)

(SLI,L) (SL,L) + (log volh × agehj) Interaction

(C,C) is equivalent to Hospital Compare.

Each subsequent model nests the previous one.

Fully Bayesian implementations with non-influential, vague priors.

MCMC calculations via posterior augmentation with Polya-Gammalatent variables (Polson, Scott, and Windle 2013).

16 / 34

Page 17: Bayesian Hospital Mortality Rate Estimation and Standardization … 1... · 2019. 4. 12. · Reported indirectly and directly standardized mortality rates Major di erences between

Emancipating the Means and Variances with Volume

● ●

●●

● ●

●●

●●

●● ●

●●

● ●

●●

●●

● ●

●●

● ●

●●

● ●

●●

●●

●●

●●

●●

● ●●

●●

●●

●●

● ●

●●

●●

●●

●●

● ●●

● ●●

●●

● ●

●●

●● ●

●●

●●

●●

●●

●●

●●

●●

●●

●●

●●

● ●●

●●●

●●

●● ●

● ●

● ●

● ● ●●

●●

●●

● ●

●●

●●

●●

●●

●●

●●

●●

●●

● ●

● ● ●●

●●

●●

● ●

●●●●

● ●

●●

●●

● ●

●●

● ●

●● ●

● ●

●●

●●

●●

●●

●●

●●

●●

●●

●●

● ● ●

● ●●

●●

●●

●●

● ●

●●

●●

● ●

●●

●●

●●

●●

●●

● ● ●

●●

●●

●●

● ●

●●

●●

●●

●●●

● ●

● ●

●●

●●

●●

● ●

●●

●●

●●

●●

●●

●●

●●●●

●●●

● ●

●●

●●

●●

●●

●●

●●

●●

●●

●●

●●

●●

●●

●●

●●

●●

●●

●●

●●

●●

●●

●●

●●

● ●

●●

●●

●●

●●

●● ●●●

● ●●

●●

● ●

●●

●●●

●●

●●

●●

●●

●●

●●

●●

●●● ●

●●●

●●

●●

●●

●●●●●

● ●

●●

●●●●

●●

●●

●●

●●

●●

●●

●●

● ●

●●

●●

●●

● ●

● ●

● ●

●●

●●

● ●●

●●

● ●

●●

● ●●

●●

●●

● ●

●●

●●

●●

● ●●

● ●●

● ●

● ●

●●

●●●

●●●

●●

●●

●●

●●

●●

●● ●

●●

●●

●●

●●

●●

●●

●●

● ●

●●

●●●

●● ●●

●●

●●

●●

●●

●●

●●

●●

●● ●

●●

●●

●●

●●

● ●

●●

●●

● ●●

● ●

●●

●●

●●

●●

●●

●●

●● ●●

● ●●● ●

●●

●●

●●

●●●

●●

●●●

●●

●●

●●

●●

●●

●●●

●●

●●

●●●

●●

● ●

●●

●●

●●

●●

● ●

●●

●●

● ●

● ●

●●

●●

●●

●●

●●

●●●

●●

●●

●● ●

●●●

●●●

●●

● ●

●●

●●

●●

● ●

●●●

●●

●●●

●●

●●

●●

●●

●●

●●

● ●

●●

●●

●●

●●●

●●

●●

●●

●●

●●

●●

●●

●●

●●

● ●

● ●●

● ●

●●●

●●

●●

●●

●● ●

●●

●●●

●●●

●●

● ●

●●

●●

● ●●

●●

●●

● ●

●●

●●

●●

●●

● ●

● ●

●●

●●

● ●

●●

●●

●●

●●

● ●

●●

●●

●●

●●

●●●

●●

●●

●●

●●

● ●●●

●●

●●

●●

●●●●

●●●

●●

●●

●●

●●

●●

●●

●●

●●

●●●

●●

●●●

●●

●●

● ●

●●

●●

●●

●●

● ●

●●

●●

● ●●

●●

●●

●●

● ●

●●

●●

● ●

● ●

●●

●●

●●

●●

●●●

●●

●●

●●

●●●

● ●

●●

●●

●●

●●

●●

●●

●●

●●

●●

● ●

●●

●●

●●

●●

● ●

●●

● ●

●●

●●

●●

●●

●●

●●

●●

●●

●●

● ●

●●

● ●

●●

●●

● ●

●●

●●

●●

● ●

●●

● ●

● ●

●●

●●

● ●

●●

●●

●●●

●●

●●

●●

●●●

●●

●●

●●

●●

●●

●●

● ●

●●

● ●

●●

●●

●●

● ●

●●

●●

●●

●●

● ●

●●

● ●

●●

●●

●●●

● ●

●●

●●

●●

●●

● ●

●●

●●

●●

●●

●●

● ●

● ●

●●

●●

● ●

●●

●●

●●

●●

●●

●●

●●

●●

●●

● ●

●●

●●

●●

●●

● ●

●●

●●

● ●

●●

●● ● ●

●●

● ●●

●●

● ●

●●

●●

●●

● ●

●●

●●

●●

●●

●●

●●

● ●●●

●●

●●

● ●

●●●●

●●

●●

●●

●●

●●

●●

●●

●●

●●

●●

● ●

●●

● ●

●●

●●

● ●

●●

●●

● ●●

●●

●●

● ●

●●

●●

●●

●●

●●

●●

●●

●●

●●

●●

●●

●●

●●

●●

●●

●●

●●

1 5 10 50 100 500 1000

−6.

5−

6.0

−5.

5−

5.0

Volume

α

: (C,C)

●●

●●

●●

●●

●●

●●

●●

●●

●●

●●

● ●

● ●

●●

●●

●●

●●

●●

●●

●●

●●

●●

● ●

●●

●●

●●

● ●

● ●

●●

●●

●●

●●

●●

●●

●●

●●

●●

●●

●●

●●

●●

● ●

●●

●●

●●

●●

●●

●●

●●●

●●

●●

● ●

●●

●●

●●

●●

●●

●●

●●

●●

●●

●●

●●●

●●

●●

●●

●●

●●

●●

●●

●●

● ●

●●

● ●

●●

●●

●●●

● ●

●●

●●

●●

●●

●● ●

● ●

●●

●●

●●

● ●

●●

● ●

●●

●●

●●

● ●

● ●

●●

●●

●●

●●

●●

●●

●●

●●

●●

●●

●●

●●

● ● ●

●●

●●

●●

●●

●●

● ●

●●

●●

●●

●●

●● ●●

●●

●● ●

●●

●●

●●

●●

●●

●●

●●

●●

● ●

●●

●●

●●

●●

●●

●●

●●

● ●

●●

●●

●●

●●

●●

●●

●●

●●

●● ●

●●

● ●

●●

●●

●●

●●

●●

●●

● ●

●●

●●

●●

●●

●●

●●

●●

● ●

● ●●

●●

●●

●●

●●

●●

●●

●●

●●

● ●

●●

●●

●●

●●

● ●●●

● ●

●●

● ●

●●

●●

● ●

● ●

●●

●●

●●

●●●

●●

● ●

●●

●●

●●

●● ●

●●●

●●

● ●

●●

●●

● ●

●●

● ●

●●

●●

●●

● ●

● ●

●●

●●

●●

●●

●●

● ●●

● ●

●●

●●

● ●

●●

●●

●●

●●

●●

● ●

●●

●●

●●

●●

●● ●

●●

●●

●●

●●

●●

●●

●●

●●

● ●●

●●

●●

●●

●●

●●

●●

●●

●●●

● ●

● ●

● ●

●●

●●

●●

●●

●●

● ●

● ●●

●●

●●

● ●

●●

●●

●●

● ●

●●

● ●

●●

●●

●●

●●

●●

●●

●●

●●●

●●

●●

●●

● ●

● ●

●●

●●

● ●

●●

●●

●●

● ●

●●

●●

●●

●●

●●

●●

●●

●●

●●

●●

●●

●●

●●

●●

●●

●●

●●

●●

●●

●●

●●

● ●

●●

●●

●● ●

●●

● ●

●●

● ●

●●

●●

●●

●●

●● ●

● ●

●●

●●

●●

●●

●●

●●

●●

●●

●●

●●

●●

●●

●●

●●●

●●

● ●

● ●

●●

●●

● ●

●●

●●

●●

●●

1 5 10 50 100 500 1000−

6.5

−6.

0−

5.5

−5.

0

Volume

α

: (L,C)

●●

●●

●●

●●

●●

●●

●●

● ●

●●

●●

●●

●●

●●

●●

●●

●●

● ●

●●

●●

●●

●●

● ●

●●

● ●

● ●

●●

●●

●●

● ●

●●

●●

●●

●●

●●

●●

●●

●●

●●

●●

● ●

●●

●●●

● ●

●●

● ●

●●

● ●

●●

● ●

●●

●●

● ●

●●

●●

● ●

●●

●●

●●

●●

●●

● ●

● ●

●●

●●

●●

●●

●●

●●

● ●

●●

●● ●

●●●

●●●

●●

●●

●●

●●

●●

● ●

●●

●●

●●

●●

●●

●●

●●

● ●

●●

●●

● ●

●●

●●

●●

●●

●●

●●

●●

●●

● ●

●●

●●

●●

●●

● ●

●●

● ●

● ●

●●

●●

● ●●

●●

●●

● ●

●●

●●

●●

●●

●●

● ●

●●

●●

●●

●●

●●

●●

● ●

●●

●●

●●

●●

●●

●●

●●

●●

●●

●●

● ●

●●

●●

●●

●●

●●

●●

●●

● ●

●●

●●

● ●

●●

●●

●●

●●

●●

● ●

●●

●●

● ●

●●

●●

●●

●●

●●

●●

●●

●●

● ●

● ●

●●

● ●

●●

● ●

●●

●●

● ●

●●

●●

●●●

● ●

●●

●●

●●

●● ●

●●

●●

●●

●●

●●●

●●

●●

●●

●●

●●

● ●

●●

●●

●●

●●

● ●

●●

●●

●●

●●

● ●

●●

●●

●●

●●

●●

● ●

●●

●●

●●

●●

●●

●●

●●

●● ●

●●

●●

●●

●●

●●

●●

● ●

● ●

● ●●

● ●

●●

●●

●●

●●

●●

●●

● ●

●●

● ●

●●

●●

● ●

●●

●●

●●

●●●

●●

●●●

●●

● ●●

●●

●●

●●

●●

●●

●●

●●

●●

●●

●●

●●

●●

●●

●●

● ●

● ●

● ●

●●

● ●

●●

●●

●●

●●

● ●

●●

●●

●●

●●

●●

●●

● ●

● ●

●●

●●

●●

● ●

● ●

●●

●●

● ●

● ●

●●

●●

●●

●●

●●

●●

●●

●●

●● ●

●●

● ●

●●

● ●

●●

●●

●●

●●

●●

● ●

●●

●●

●●

●●

●●

●●

●●

● ●

●●

●●

● ●

● ●

● ●

●●

●●

●●●

1 5 10 50 100 500 1000

−6.

5−

6.0

−5.

5−

5.0

Volume

α

: (S,L)

Figure: Posterior means of αh vs volh

Dramatic improvements over the (C,C) model.

Data speak clearly because simpler models are nested.

Higher mortality rates at low volume hospitals.

17 / 34

Page 18: Bayesian Hospital Mortality Rate Estimation and Standardization … 1... · 2019. 4. 12. · Reported indirectly and directly standardized mortality rates Major di erences between

Adding More Hospital Characteristics and an Interaction

● ●

●●

●●

●●

●●

●●

●●●

● ●

●●

●●

●●

● ●

●●

●●

●●

●●

●●

●●

●●

●●

●●

●●

● ●

●●

● ●

● ●

●●

●●

●●

● ●

●●

●●

●●

● ●

●●

● ●

● ●

● ●●

●●

● ●

●●

● ●

●●

●●

●●

● ●

●●

●●

●●

●●

●●

●●

●●

●●

●●

●●

●●

●●

●●

●●

● ● ●●

●●

●●

● ●

●●

●●

●●

●●

●●

●●

● ●

●●

● ●

●●

●●

●●

●●●

●●

●●

●●

●●

●●

●●

●●

●●

● ●

●●

●●

● ●

●●

●●

●●

● ●

●●

●●

●●

●●

● ●

● ●

●●

●●

●●

●●

●●

● ●

●●

●●

●●

●●

●●

●●

●●

●●

●●

●●

●●

●●

● ●

●●

●●

●●

●●

●●

●●

●●

●●

● ●

●●

●●

●●

●●

● ●

●●

●● ●

●●

●●

● ●

● ●

●●

●●

●●

●●

●●

●●

●●

●●

●●

●●●

● ●

● ●●

●●

●●

●●

●●

●●

●●

●●

● ●

● ●

●●

●●

●●

●●

●●

●●

●●

●●

●●

● ●

●●

● ●

●●

●●

●●

●●

●●

●●

●● ●

●●

●●

●●

●●

●●

●●

● ●

●●

●●

●●

● ●

● ●

●●

●●

●●

●●

●●

● ●

●●

● ●

●● ●

●●

●●

● ●

●●

● ●

●●

●●

●●

●●

●●

●● ●

●●

●●

●●

●●

●●

●●

●●

● ●

●●

●●●

●●

●●

●●

●●

●● ●

●● ●

● ●

● ●

●●

●●

●●

●●

● ●

● ●●

●●

●●

●●

●●

●●

●●

● ●

●●

●●

●●

●●

●●

●●

●●

●●

●●

●●

●●

●●

●●

●●

● ●

●●

●●

●●

● ●

●●

●●

● ●

●●

●●

●●

●●

●●

●●

●●

●●

●●

●●

●●

●●

●●

●●

●●

●●

● ●

●●

●●

●●

●●

●●

●●

● ● ● ●

●●

●●

●●

●●

●●

●●

●●

●●

●●

●●

●●

●●

●●

●●

●●

● ●

●●

●●

●●

●●

●●

● ●

1 5 10 50 100 500 1000

−6.

5−

6.0

−5.

5−

5.0

Volume

α

: (SL,L)

●●

● ●

●●

●●

● ●

●●

●●

●●

●●

●●

●●

●●

●●

●●●

●●

● ●

●●

●●●

●●

●●

●●

●●

●●

●●

●●

●●

●●

● ●

●●

●●

●●

● ●

●●

●●

● ●

●●

●●

● ●

●●

●●

●●

●●

●●

●●

●●

●●

●●

●●

●●

● ●

●●

●●

●●

● ●

● ●

●●

● ●

● ●

● ●

●●

●●

● ●

●●

●●

●●

●●

●●

●●

● ●

●●●

●●

●●

● ●

●●

●●

● ●

●●

●●

●●

●●

●●

●●

● ●

● ●

●●

●●

●●●

●●

● ●

●●●

●●

●●

●●

●●

●●

●●

● ●●

●●

●●

●●

●●

●●

●●

●●

●●

●●

●●

●●

●●

●●

●●

●●

●●

●●

●●

●●●

●●

●●

●●

●●

●●

●●

●●

●●

●●

●●

●●

●●

●●

●●

●●

● ●●

●●

●●

●●

● ●

●●

●● ●●

●●

●●

●●

● ● ●

●●

●●

●●

●●

●●

●●

● ●

●●

●●

●●

●●

●●●

●●

●●

●●

● ●

●●

●●

●●

●●

●●

●●

●●

● ●

●●

●●

●●

●●

● ●

● ●

● ●

●●

●●

●●

●●

●●

●●

●●

●●

●●

●●

●●

●●

●●

●●

●●

●●

●●

●●

●●

●●

●●

●●

●●

●●

●●

●●

●●

● ●

●●

●●

●●

●●

●●

●●

●●

● ●●

●●

●●

●●

● ●

●●

●●

●●

●●

●●

●●

● ●●

●●

●●

●●

●●

●●

●●

●●

●●

●●

●●

●●

●●

●●

●●

●●

●●

●●

●●

●●

●●

●●

●●

●●

●●

●●

●●

●●

●●

●●

●●

● ●

● ●

●●

●●

●●

●●

●●

●●

●●

●●

●●

●●

●●

●●

1 5 10 50 100 500 1000

−7.

0−

6.5

−6.

0−

5.5

−5.

0−

4.5

−4.

0−

3.5

Volume

α

: (SLI,L)

Figure: Posterior means of αh vs volh

Refinements continue to support higher mortality rates at lowvolume hospitals

18 / 34

Page 19: Bayesian Hospital Mortality Rate Estimation and Standardization … 1... · 2019. 4. 12. · Reported indirectly and directly standardized mortality rates Major di erences between

Have Predictions Really Improved?

19 / 34

Page 20: Bayesian Hospital Mortality Rate Estimation and Standardization … 1... · 2019. 4. 12. · Reported indirectly and directly standardized mortality rates Major di erences between

Model Comparisons via Predictive Bayes Factors

In-sample Bayes Factors are unreliable with vague priors.

Instead, use out-of-sample Predictive Bayes factors versus (C,C)

BFMi/MCC=

P(yout | yin,Mi )

P(yout | yin,MCC )

Model (L,C) (S,L) (SL,L) (SLI,L)

log BF 27.54 32.13 35.46 37.96

Vast successive improvements. (SLI,L) clearly best.

20 / 34

Page 21: Bayesian Hospital Mortality Rate Estimation and Standardization … 1... · 2019. 4. 12. · Reported indirectly and directly standardized mortality rates Major di erences between

Matched Sample Comparisons of Model Predictions

Predictive BF’s gauge overall model fit.

But what about the calibration of model predictions withfuture mortality rates on particular segments of patients?

For this purpose, we compared each model’s predictions toout-of-sample mortality rates on two sets of patients:

LV: Patients at low-volume hospitals (bottom 20%)HV: Matched patients at high-volume hospitals (top 20%)

Controlling patient risk characteristics through matchingprovides a clearer comparison of predicted mortality ratesbetween low- and high-volume hospitals.

21 / 34

Page 22: Bayesian Hospital Mortality Rate Estimation and Standardization … 1... · 2019. 4. 12. · Reported indirectly and directly standardized mortality rates Major di erences between

Matching Strategy

Five HV patients are matched to each LV patient.

Matching is based on minimizing weighted distance betweenpatient characteristics, propensity scores and expectedmortalities.

An example of patient distancesHV Patients

LV Patients 1 2 3 4 5a 1176.30 1371.56 482.97 399.51 380.02b 1190.85 1389.88 498.10 427.68 394.97c 816.24 1017.29 122.94 63.94 25.97d 1120.22 1330.56 437.57 359.39 328.57

Note that patient c is likely to be matched to patient 5.

22 / 34

Page 23: Bayesian Hospital Mortality Rate Estimation and Standardization … 1... · 2019. 4. 12. · Reported indirectly and directly standardized mortality rates Major di erences between

Out-of-Sample Comparisons

Low VolumeHigh Volume

MatchedHigh Volume

All

Observed Mortality 28.3 19.8 12.4

(C,C) 23.1 21.6 12.7(SLI,L) 29.6 21.0 12.4

Table: Out-of-sample predicted mortality compared against observedmortality in the matched study of low and high volume hospitals.

At low-volume hospitals, (C,C) is poorly calibrated.It predicted 23.1% when the actual observed was 28.3%.

23 / 34

Page 24: Bayesian Hospital Mortality Rate Estimation and Standardization … 1... · 2019. 4. 12. · Reported indirectly and directly standardized mortality rates Major di erences between

Indirect and Direct Standardization

24 / 34

Page 25: Bayesian Hospital Mortality Rate Estimation and Standardization … 1... · 2019. 4. 12. · Reported indirectly and directly standardized mortality rates Major di erences between

Standardizing Mortality Rates for Public Reporting

Mortality rate Ph at hospital h influenced by patient case-mix.

Standardize Ph to eliminate this effect of case-mix variation.

Two approaches:

Indirect standardization (used by Hospital Compare)Direct standardization

Both approaches make use of the fact that the mortality ratefor patient xhj at any hospital h∗ can be obtained via

ph∗(xhj) = logit−1(αh∗ + x ′hjβ).

Note that ph∗(xhj) is a counterfactual unless h∗ = h.

25 / 34

Page 26: Bayesian Hospital Mortality Rate Estimation and Standardization … 1... · 2019. 4. 12. · Reported indirectly and directly standardized mortality rates Major di erences between

Indirect Standardization

P ISh = (Ph/Eh)× y ,

where

Eh = 1nh

∑nhj=1

[1H

∑Hh∗=1 ph∗(xhj)

]Eh: Average mortality rate of hospital h patients had theybeen treated at all H hospitals.

y : national average mortally rate (≈ 15%).

Some drawbacks:

- lacks probabilistic justification.

- fails to eliminate case-mix variation effects, except for (C,C).

- systematically underestimates actual hospital mortality rates.

26 / 34

Page 27: Bayesian Hospital Mortality Rate Estimation and Standardization … 1... · 2019. 4. 12. · Reported indirectly and directly standardized mortality rates Major di erences between

●● ●

●●

●●

●●

● ●

●●

●●

●●

●● ●

●●●

● ●

●●

●●

●●

● ●

●●

●● ●

●●

●●

● ●

●●

●●

●●

●●

●●

●●

●●

●●

● ●●

●●

●●

● ●

●● ●

●●

●●

●●

●●

● ●●

● ● ●

●●

●●

● ●

●●

●● ●

●●

● ●●

●●

● ●

●●

● ●

●●

●●

●●

●●

●●

●●

●●

●●

● ●●

● ●●

●● ●

●●

● ●

● ● ●●

●●

●●

●●

● ●●

●●

●●

●●

●●

●●

●●

●●

●●

●●

●●●

●●

●●

●●

●● ●

●●

●●●

●●

●●

●●

●●●●

● ●

●●

●●

●● ●

●●

●●

● ●

●● ●●

● ●

●●

●●

●●

●●

●●

●●

● ●

●●

●●

●●

●●

●●

●●

●●

●●

●●

●● ● ●

● ●●

● ●

● ●●

●●

●●

● ●

●●

● ●

● ●

●● ●

●●

●●

●●

●●

●●

● ●

●●

●●

● ● ●

●●

●●

●●

●●

●●

●●

● ●

●●

●●

●●

●●

●●

●●

●●●

●●

● ●

●●

●●

● ●

●●●

● ●

●●

●●

●●

●●

●●

●●

●●

●●

●●

●●

●● ●●

●●●

●●

●●

● ●●

●●

●●●

●●

● ●

●●

●●

●●

●●

●●

● ●

●●

●●

●●

●●

●●

●●

● ●

●●●

●●

● ●

●●

●● ●

●●

●●

●●

●●

●●●

●● ●

●●

●●

●●

●●

●● ●● ●● ●

●●

● ●

●●●

●●

●●

●●

●●

●●●

●●

●●

●●

●●●

●●

●●

●●

●●● ●

●●●

●●

●●

●●

●●●● ●● ●

● ●●●●●

●●

●●

●●●

●●

●●

●●

●●

●●

●●

● ●

●●

●●

●●

●●

●●●

●●

●●

● ●●

●●

● ●

●●

●●

●●

●●

●●

● ●●

●●

●●

●●

●●

● ●●

●●

●●

●●

●●

●●

●●

● ●● ●●

● ●●

●●

● ●

●●

●●

●●

●●●

●●●

●●

●●

●●

●●

●●

●●

●●

●● ●

● ●

●●

●●

●●

●●

●●

●●

●●

● ●

●●

●●

●●●

●●

●●

●● ●●

●●

●●

●●

●●

●●

●●

●●

●● ●

●●

●●

●●

●●

●●

● ●

●●

●●

●●

● ●●

● ●

●●

●●●

● ●

●●

●●

●●●

●● ●●

● ●●● ●

●●

●●

●●

●●

●●

●●

●●●

●●

●●●

●●

●●●

●●

●●

●●

●● ●

● ●

●●

●●

●●●

●●

● ●

●●

●●

●●

●●

● ●

●●

●●

● ●

● ●

●●

●●

●●

●●

●●

●●

●●●

●●

●●

●●

●● ●

●●●

●● ●

●●

●●

●●

● ●

● ●

●●

●●

● ●

●●●

●●

●●

●●

●●

●●

●●

●●

●● ●

●●

●●

●●

● ●●

●●

●●●

●●

●●

●●●●

● ●●●

●●

● ●

●●

● ●

●●

●● ●

●● ●

●●

●●

● ●●

●●

●●

● ●

● ● ●

●●

●●

●● ●

●●

●●

●●●

●●●

●●

●●

●●

●●

●●

●●

●●

●●

●●

●●

●●

●●

●●

● ●

● ●

●●

● ●

●●

● ●

●●

●●

●●

●●

● ●

●●

● ●

● ●

● ●

●●

●●

●●

●●●

●●

●●

●●

●●

●●

●●

●●

● ●●●

●●

●●

●●●

●●

●●●●●

●●

●● ● ●

●●

●●

●●

● ●

●●

● ●

●●

●●●

●●

●●

●●●

●●

●●

●●

●●

●●

●●

●●

● ●

●●

●●

●●

● ●

●●

●●

●●

● ●●

●●

●●

●●

● ●●

●●

● ●

● ●

●●

●●

●●●

●●

●●

●●

● ●●

●●

●●

●●

●●

●●●

●●

● ●

●●

●● ●

● ●

●●

●●

●●

●●

●●●●

●●

●●

●●

●● ●

●●

● ●

●●

●●

●●

● ●

●●

● ●

●● ●

●●

●● ●

●●

●●

● ●

●●

●●

●●

●●

●●

●●

●●

●●

● ●

●●

● ●

●●

●●

●● ●

●●

●●

●●

● ●●●

●●

● ●

●●

●●

●●

●●

● ●

● ●

●●

●●●

●●

●●●●

●●

●●

●●●

●●

●●

●●

●● ●

●●

●●

●●

●●

●●

●●

●●

● ●

●●

●●

●●

●●

●●●

● ●

●●

●●

●●

●●

●●

●●

● ●

●●

●●

●●

●●

●● ●

● ●

●●

●●

●●

●●

● ●

●●

● ●

●●

●●

●●

●●

●●

● ●

●●

●●

●●

●●

● ●

● ●

●●

●●

●●

●●

●●

●●●●

●●

●●

●●

●●

●●

●●

●●

●● ●

●●

●●

●●

●●

●●

●●●

● ●

●●●●

●●

● ●

●●

●●

●● ● ●

●●

● ●●

● ●● ●

●●●

●●

●●

●●

● ●

●●

●●

●●

●●

●●

●●

●●

●● ●

● ●

● ● ●●

●●

●●

● ●●

●●●

●●

●●

●●

●●

●●

●●●

●●

●●

● ●●

●●

●● ●

●●

●●

●●

●●

●●

●●

●●

●●

●●

●●

● ●●

●●

●●●

● ●

●●

●●

●●

●●

●●

●●

●●

●●

●●

●●

●● ●

●●

●●

●●

● ●

●●

●●●

● ●

●●

●●

●●

●●●

●●

1 5 10 50 100 500 1000

0.10

0.15

0.20

0.25

0.30

Volume

PIS

: (C,C)

●●

●●

●●

●●

●●

●●●

●●

●●

●●

● ●

●●

●●

●●

●●

●●

● ●

●●

● ●

●●

●●

●●

● ●

●●

● ●

●●

●● ●

● ●

● ●

●●

● ●

●●

●●

●●

●●

●●

●●

●●

●●

●●

●●

●● ●

● ●

●●

●●

●●

●●

●●

●●

●●

● ●

● ●

●●

●●

● ●●

●●

●●

● ●

●●

●●

● ●

●●

●●●

●●

●●

● ●

●●

●●

●●●

●●

●●

●●

●●

●●

● ●

●●

●●

●●

●●●

●●

●●●

● ●

●●

●●

●●

●●

●●

●●

●●

●●

●●

●●

●●

●●

●●

● ●●

●●

●●

●●

●●

●●

●●

●●

●●

●●

●●

●●

●●

●●

● ●●

● ●

●●

●●

●●

●●

●●

●●

●●

●●

●●

● ●

●●

●●

●●

●●

● ●

●●

●●

●●

●●

●●

●● ●

●●

●●

●●

●●

●●

●●

●●

●●

●●

● ●

●●

●●

●●

●●

●●

●●

●●

●●

●●

● ●

●●

●●

● ●

●● ●

● ●

●●

● ●

●●

●●

●●

●●

●●

●●

●●

● ●

● ●

●●

●●

●●

●●

●●

●●

●●

●●

● ●

●●

●●

●●

●●

●●

●●

●●

●●

● ●

●●

●●

●●●

●●

●●

● ● ●

●●

●●

●●

●●

●●

●●

●●

● ●

●●

●●

●●

●● ●

●●

●●

● ●

●●●

●●

●●

●●

●●

●●

●●

●●

●●

● ●●

● ●

●●

●●

● ●

●●

●●

●●

●●

●●

●●

●● ●●

●●

●●●

●●

●●

●●

●●

●● ●

● ●●

●●

● ●

●●

●●●

●●

●●

● ●

● ●

●●

●●

●●

●●

●●

● ●

●●

●●

●●●

●●

● ●

●● ●

●●

●●

●●

●●

●●

●●

●●

● ●

●●

●●

●●●

● ●●

●●

●●

●●

●●

●●

●●

●●

●●

●●

●●

● ● ●

●●

●●

●●

●●●

●●

●●

●●

● ●

● ●

●●

●●

● ●

●●

●●

●●

● ●

●●●

●●

●●

●●

●●

●●

●●

●●

●●

●●

●●

●●

●●

● ●

●● ●

●●

●●

●●

●●

●●

●●

●●●

● ●●

●●

●●

●●

●●

●●

●●

●●

●●

●●

●●

●●

● ●

● ●

●●

● ●

●●●

●●

●●●

● ●

●●

●●

●●

●●

●●

● ●●●

●●

●●

●●

●●

●●

●●

●●

●●

● ●

● ●

●●

●●

●●

●●

●●

●●

●●●

●●

●●●

●●

● ●

● ●

● ●

●●

●●

●●●

●●

●●●

●●

● ●

●●

● ●

●●●

●●

●●

●●

● ●

●●

●●

●●

●●

●●

●●

●●

●●

●●

● ●

●●

●●

●●

● ●

●●

●●

●●

●●

●●

●●

●●

●●

●●

●●●

●●

●●

● ●

●●

●●

●●

●●

●● ●

●●●

●●

● ●●

●●

●●

●●

●●

●●

●●

●●● ● ●

● ●

●●

●●

●●

●●

●●

●●

●●

● ●

●●

●●

●●

●●

●●

●●

●●

●●

●●

●●

●● ●

●●

● ●

● ●

●●

●●

●●

● ●

●●

●●

●●

●●

● ●

●●

1 5 10 50 100 500 1000

0.10

0.15

0.20

0.25

0.30

Volume

PIS

: (L,C)

● ●

●●

●●

●●

● ●

●●

●●

●●

●●

●●

● ●

●●

●●

●●

● ●

●●

●●

●●

● ●

●●

●●●●

●●

● ●

●●

●●

● ●

●●

●●

●●

● ●

●●

●●

● ●

●●

●●

●●

●●

● ●

●●

●●

●●

● ●

●●●

●●

● ●

●●

●●

● ●

● ●

●●

● ●

●●

●●

● ●

●●

●●

●●

●●●

●●

●●

●●

●●

●●

●●●

●●

●●●

●●

●●

●●

●●

●●

●●

●●

●●●

●●

●●

●● ●

●●

●●

● ●

●●

●●

●●

●●

●●

●●

●●

●●

●●

●●

●●●

●●

●●

●●

●●

●●

●●

●●

●●

●●

●●

●●

●●

●●

●●

●●

●●

●●

●●

● ●

● ●●

●●

●●

● ●

●●

●●

● ●

●●

●●

● ●

●●

●●

●●

●●

●●

●●

●● ●

●●

●● ●

●●

● ●

●●

●●

●●

●●

●●

● ●

● ●

●●

● ●

●●●

●●

●●

●●

●●

●●

● ●

●●

●●●

●●

●●

●●

●●

●●

●●

●●

●●

●●

●●

●●

●●

●●

●●

●●

● ●●

●●

●●

●●

●●●●

●●

●●

●● ●

● ●

● ●

●●

●●

●●

●● ●

●●

●●

●●

●●

●●

●●

●●

●●

●●

●●

● ●●●

●●

●●

●●

●● ●

●●

●●

●●

● ●

●●

●●

●●

●●

●●

●●

●●

●●●

●●

● ●

●●

●●

●●

●●

●●

●●

●●

●● ●

●●

● ●

●●

● ●

●●

●●

●●

● ●

●●

●●

●●

● ●

●●

●●

●●●

●●

●●

● ●

●●

● ●

●●

●●

●●●

●●

● ●

●●

●●

●●

●● ●

●●

●●●

●●

● ●

●●

●●

● ●

●● ●

●●

●●

● ●

●●

●●

● ●

●●

●●

●●

●●

●●

●●

●●

● ●

● ●●

●●

●●

●●

●●

●●

● ●●

●●

●● ● ●

●●

●●

● ●

●●

●●

●●

● ●

● ●

●●

●●

●●

●●

●●

●●

●●

●●

●●

●●

● ●●

●●

●●

●●

●●

● ●

●●

● ●

●●●

●●

●●

●●

●●

●●

●●

● ●

●●

● ●

●●

●●

●●

●●

●●

●●

●●

● ● ●

●●

●●

●●

●● ●

●●

●●

● ●

●●

● ●

●●

●●

●●

●●

●●

●●

●●

●●

●●

●●

●●

●●

●●

●●

●●

●●

●●

● ●

● ●

●●●

●●

●●

●●

●●

●●

●●

● ●

● ●

●●

●●

● ●

● ●

●●

●●

●●

●●

● ●

●●

●●

●●

●●

●●

●●

●● ●

●●

●●

●●●

1 5 10 50 100 500 1000

0.10

0.15

0.20

0.25

0.30

Volume

PIS

: (SLI,L)

Figure: Indirectly Standardized Mortality Rates P ISh vs. volh.

27 / 34

Page 28: Bayesian Hospital Mortality Rate Estimation and Standardization … 1... · 2019. 4. 12. · Reported indirectly and directly standardized mortality rates Major di erences between

Direct Standardization

PDSh =

1

N

H∑h∗=1

nh∗∑j=1

ph(xh∗j), N =H∑

h∗=1

nh∗

PDSh : Average mortality rate of all N patients had they been

treated at hospital h.

Benefits of this approach

+ easier to understand.

+ an interpretable, almost linear scaling of αh.

+ eliminates the effect of case-mix variation.

+ is correctly calibrated to actual mortality rates.

28 / 34

Page 29: Bayesian Hospital Mortality Rate Estimation and Standardization … 1... · 2019. 4. 12. · Reported indirectly and directly standardized mortality rates Major di erences between

● ●

●●

●●

●●

● ●

●●

●● ●●

●●

● ●

●●

● ●

● ●

●●

● ●

●●

●●

●●

●●●

●●

●●

●●

●●

●●

● ●●

●●

●●

●●

●●

●●

●●

●●

●●

● ●● ●

●●

●●

●●

●●

●●

●●

●●●

●●●

● ●●

●●

● ●

●●

●●

●●

●●

●●

● ●

●●

●●

●●

● ●●

● ●●

●●

●● ●

●●

● ●

● ●

● ● ●●

●●

●●

●●

●●

●●

●●

●●

●●

●●

●●

●●

●●

●●

● ●●

●●

● ●

● ●

●● ●

●●

●●

●●

●●

●●

●●

●●●●

● ●

●●

●●

● ●

●●

●●

● ●

●● ●●

● ●

●●

●●

●●

●●

●●

●●

●●

●●

●●

●●

●●

●●

●●

●●

●●

● ●

● ● ●●

●●

●●

●●●

●●

●●

● ●

●●

●●

●●

●●

●●

●● ●

●●

●●

●●

● ●

● ●●●

●●

●●

●●

●●

●●

● ●

●●

●●

●●

●●

●●

●●

●●

●●●

●●

●●

● ●

●●

●●●

●●

●●

●●

●●

●●

●●

●●

●●

●●

●●

●●●●

●●●

●●

●●

●●● ●

●●

●●

●●

●●

●●

●●

●●

● ●

●●

●●

●●

●●

● ●

●●

●● ●

●●●

●●

●●

●● ●

●●

●●

●●

●●

●●●

● ●

●●

● ●

●●

●● ●● ●●

● ●

●●

●●

●●

●●

●●

●●

●●

●●

●●

●●

●●●

●●

● ●

●●

●●● ●

●●●

●●

●●●

●●

●●●●

●● ●

●●●●●●

●●

●●

●●

●●

●●

●●

●●

●● ●

●●

●●

●●

●●

●●

●●

●●

●●

●●

● ●

● ●●

●●

●●

●●

●●

● ●

●●

● ●●

●●

● ●

●●

●●

● ● ●

●●

●●

● ●

●●

●●

●●

●●

● ●

● ●●

●● ●

●●

●●

●●

●●

●●●

●●

● ●

●●

●●

●●

● ●

●●

●●

●●●

●●

●●

● ●

●●●

●●

●●

●●

●●

● ●

●●

●●

●●

●●

●● ● ●

●●

●●

●●

●●

●●

● ●

●●

●● ●

● ● ●

●●

●●

●●

● ●

●●

●●

● ● ●●

● ●

●●●

● ●

●●

●●

●●●

●● ●

● ●

●● ●

●●

●●

●●

●●

● ●

●●

●●●

●●

● ●●

●●

●●

● ●●

●●

●●

●●●

● ●

●●

●●

●●

● ●

●●

●●

●●

●●

●●

● ●●

●●

●●

●●●

● ●

● ●

●● ●

●●

●●

●●

●●●

●●

●●

●●

●●

●●●

●● ●

●●

●●

● ●

● ●

●●

●●

●●

● ●

●●●

●●

●●●

●●

●●

●●

●● ●●●

●●

● ●●

●●

●●

●●

●●

●●

●●

●●

● ●●●

●●

● ●

●●

●●

●●

●●

●●

●●

●●

● ●●

●●●

●●

●●

●●●

●●

●●

●●

● ●●

●●

●● ●

●●

●●

●●●

●●

● ●

● ●

●●

●●

● ●●

●●

●●

● ●

●●

●●

●●

●●

●●

●●

●●

●●

●●

●●

●●

●●

●●

● ●

● ●

●●

● ●

●●

●●

●●

● ●

●●

●●

●●

●●

●●

●●

●●

●●

●●

● ●

●●

●●

●●●

●●●

●●

● ●●●●

●●

●● ● ●

●●

●●

●●

● ●

●●

●●

●●

●●●

●●

●●

●●

●●

●●

●●

●●

● ●

●●

●●

●●

●●

● ●

●●

●●

●●

● ●●●

●●

●● ●

● ●●

●●

● ●●

● ●

●●●

●●

●●

●●

●●●

●●

●●

●●

●●

●●

●●

●●

●●

●● ●●

●●

●●

●●

●●●

●●

●●

● ●

●●

●●

●●●

●●

●●

● ●

●●

●●

●●●

●●

● ●

●●

●●

●●

●●

●●

●●

●●

●●

●●

● ●

●●

●●

●●

●●

●●

●●

●●

● ●

●●

●●

● ●

●●

●●

●●

●●

● ●

● ●

●●

●●

●●●

●●

●●

●●

● ●

●●

●●

●●

●●

●●

●●

●●

● ●

●●

● ●

●●

●●

●●

●●

●●

●●

●●

●●

● ●

●●

●●

●●

●●

●●

●●

● ●

●●

●●

●●

●●

● ●

●●

●●●

●●

●●

●●

●●

● ●

●●

●●

●●

●●

●●

●●

●●

● ●

● ●●

●●

●●

●●

● ●

●●

●●

●●

●●

●●

●●●

●●

●●

●●

●●

● ●

●●

●●

●●

●●

●●

●●

●●●

●●

● ●

●●

●●

●● ●●

●●

●●

● ●●

● ●● ●

●●

●●

●●

●●

● ●

●●

●●

●●

●●

● ●

●●

●●

●●

●●

●●

●●

●●

●●

●● ●

●●●

●●

●●

●●

●●

●●

●●●

●●

●●●

● ●

●●

●●

●●

● ●

●●

● ●

●●

●●

● ●

●●

●●

● ●●●

●●

●●

● ●

●●

●●●

●●

●●

●●

● ●

●●

●●

●●

●● ●

●●

●●

●●

●●

●●

● ●

●●

●●

●●●

1 5 10 50 100 500 1000

0.10

0.15

0.20

0.25

0.30

Volume

PD

S

: (C,C)

●●

●●

●●

●●●

●●

●●

● ●

●●●

● ●

●●

● ●

● ●

●●

●●

● ●

●●

●●

●●

●●

●●

●●

●●

●●●

● ●

●●

●●

● ●

●●

● ●

●●

●●

● ●

●●

●●

●●

●●

●●●

●●

●●

●●

●●

●●

●●

●●

● ●

●●

●●

● ●

●●

●●

●●

●●

● ●

●●●

●●

● ●

●●

●●

●●

●●

●●

●●

●●

●●

●●

●●

●●

●●

● ●

●●

●●

●●

●●

●●

●●

●●

●●

●●

●●

●●

●●

●●

●●

●●●

●●●

● ●

●●

●●

●●

●●

●●

●●

●● ●

● ●●

●●

●●

●●

●●

●●

●●

●●

●●

●●

●●

●●

●●

● ●

●●

●●

●●

● ●

● ●

●●

●●

●●

●●

●●

●●

●●

● ●

●●

●● ●

● ●

● ● ●

●●

●●

●●

●●

●●

●●

●●

●●

●●

●●

●●

● ●●

●●

●●

●●

●●

●●

●●

●●

●●

●●

●●

● ●

●●

●●

●●

●●

● ●

●●

●●

● ●●

●●

● ●

●●

●●●

●●

●●

●● ●

●●

● ●

●●

●●

●●

●●

●●

●●

●●

●●

●●

● ●

●●

●●

●●

●●

●●

●●

●●

●●

●●

●●

●●

● ●●

●●

●●

●●

●●

●●

●●

●●

●●

● ●

●●

●●

●●

●●

● ●

●●

●●

●●

●●

●●

●●

●●

● ●

●●

●●

●●

●●

●●

●●●

●●

● ●

●●

●●

●●

●●

●● ●

● ●●

●●

● ●

●●

●●●

● ●

●●

● ●

●●

●●

●●

●●

● ●

● ●

●●

●●

●●

●●

●●

● ●●

●●

● ●

●●●

●●

●●

● ●

● ●

●●

●●

●●

●●

●●

●●●

●●

● ●

●●

●●

●●

●●

●●

●●

●● ●

●●

●●

●●

●●

●●

●●

●●

●●

● ●●

●●

●●

●●

●●

●●

●●

●●

●●

●●●

● ●

● ●

● ●

●●

●●

●●

●●

●●

●●

●●

● ●

●●

●●

● ●●

●●

●●

●●

●●

●●

●●

●●

●●

●●

● ●

●●

● ●

●●

●●

●●

●●

●●

●●

●●●

●●

●●

●●

● ●

●●

● ●

● ●

●●

●●

●●

●●

●●

●●

● ●

●●

●●

●●

●●

●●

●●

●●

● ●

●●

●●

●●

●●●

●●

●●

● ●

●●

●●

●●

● ●

●●

●●

●●

●●

●●

●●

●●

●●

● ●

●●

●●

●●

●●

●●

●●

●●

●●

●●●

●●

●●

●●

●● ●

● ●

●●

●●

●●

●●

●●

●●

●●

●●

●●

●●

● ●

●●

●●

● ●

● ●

●●●

●●

● ●

●●

●●

● ●

●●

● ●

●●

●●

●●●

1 5 10 50 100 500 10000.

100.

150.

200.

250.

30

Volume

PD

S

: (L,C)

●●

●●

●●

●●

●●

●●

●●

●●

● ●

●●

●●

●●

●●

●●

● ●

●●

●●

●●

●●

●●

●●

●●

●●

● ●

● ●

● ●

● ●

●●

●●

●●

●●

●●

●●

●●

● ●

●●

●●

●●

●●

●●

● ●

●●

●●

●●

● ●

●●

●●●

● ●

● ●

●●●

●●

●●

●●

●●

●●

● ●

●●

●●

●●

●●

●●

●●

●● ●

● ●

● ●

●●

●●

●●

● ●

●●

●●

●●

●●

●●

●●

●●

●●

● ●●

●●

●●

●●

●●

●●

●●

●●

●●

● ●

●●

●●

●●

● ●

●●

●●

● ●

●●

●●

● ●

● ●

● ●

●●

●●

●●

● ●

●●

●●

●●

●●

● ●

●●

●●

●●

●●

●●

●●

●●

●● ●

●●

● ●

● ●

●●

● ●

●●

●●

●●

●●

● ●

●●

●●

●●

●●

●●

●●

●●

●●

●●

●●

● ●

●●

●●

●●

●●

●●

●●

●●

●●

● ●

●●

●●

●●

●●

●●

●●

●●

●●

●●

● ●

●●

●●

● ●

● ●

●●

●●

●●

●●

● ●

●●

●●

●●

●●

●●

●●

●●●

● ●

● ●

●●●

●●

●●

●●

●●

● ●

●●

●● ●

●●●

●●

●●

●●

●●

●●

●●

●●

● ●

●●

●●

●●

●●

●●

●●

●● ●

● ● ●

●●●

●●

●●

●●

●●

● ●

●●

●●

●●

●●

●●

●●

● ●

●●

●●

●●

●●●

●●

●●

●●

●●

●●

●●

●● ●

●●

●●

●●

●●

●●

●●

●●

● ●

●●

●●

● ●●

●●

●●

● ●

●●

●●

●●

●●

● ●

●●

●●

●●

●●

●●

●●

●●

●●

●●

● ●

●●

●●

●●

●●●

●●

● ●

●●

●●

●●

● ●

●●

●●

● ●

●●

● ● ●

●●

●●

●●● ●●

●●

●●

●● ●

●●

●●

●●

● ●

●●●

●●

●●

● ●●

●●

●●

●●

●●

●●

●●

●●

●●●

●●

●●

● ●

●●

●●

●●

● ●

● ●

●●

●●

●●

●●

●●

●●

●●

●●

●●

●●

●●

●●

● ●●

1 5 10 50 100 500 1000

0.10

0.15

0.20

0.25

0.30

Volume

PD

S

: (SLI,L)

Figure: Directly Standardized Mortality Rates PDSh vs. volh.

red horizontal line - average mortality rate over patients

blue horizontal line - average mortality rate over hospitals

29 / 34

Page 30: Bayesian Hospital Mortality Rate Estimation and Standardization … 1... · 2019. 4. 12. · Reported indirectly and directly standardized mortality rates Major di erences between

Mortality Rate Uncertainty QuantificationThe variation of the PDS

h posterior mean estimates is smallerat the low volume hospitals.However, the posterior mean variation should not be confusedwith posterior uncertainty of the estimates which is conveyedby the full posterior distribution of the PDS

h values.

●●

●●

●●

●●

●●

●●

●●

●●

●●

●●

●●

●●

●●

●●

●●

●●

●●

●●

●●

● ●

●●●

●●●

●●●

●●●

●●●

●●●

●●●

●●●

●●●

●●●

●●

●●

●●

●●

●●

●●

●●

●●

●●

●●●●

●●

●●●●●

●●

●●●●●

●●

●●●●●

●●

●●●●●

●●

●●●●●

●●

●●●●●

●●

●●●●●

●●

●●●●●

●●

●●●●●

●●

●●●

●●

●●

●●●●

●●

●●

●●

●●●●

●●

●●

●●

●●●●

●●

●●

●●

●●●●

●●

●●

●●

●●●●

●●

●●

●●

●●●●

●●

●●

●●

●●●●

●●

●●

●●

●●●●

●●

●●

●●

●●●●

●●

●●

●●

●●●●

●●

●●

●●●●●

●●

●●●●●

●●

●●●●●

●●

●●●●●

●●

●●●●●

●●

●●●●●

●●

●●●●●

●●

●●●●●

●●

●●●●●

●●

●●●

●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●

●●

●●●●

●●

●●●●

●●

●●●●

●●

●●●●

●●

●●●●

●●

●●●●

●●

●●●●

●●

●●●●

●●

●●●●

●●

●●●●

●●●●

●●●●

●●●●

●●●●

●●●●

●●●●

●●●●

●●●●

●●●●

●●●●

●●●●

●●●●

●●●●

●●●●

●●●●

●●●●

●●●●

●●●●

●●●●

●●●

●●●●

●●

●●●●

●●

●●●●

●●

●●●●

●●

●●●●

●●

●●●●

●●

●●●●

●●

●●●●

●●

●●●●

●●

●●●●

1 2 5 10 25 50 100 200 400 800

0.1

0.2

0.3

0.4

Volume

PD

S

: (C,C) model

●●●

●●●

●●●

●●●

●●●

●●●

●●●

●●●

●●●

●●●

●●

●●●●●

●●

●●

●●●●●

●●

●●

●●●●●

●●

●●

●●●●●

●●

●●

●●●●●

●●

●●

●●●●●

●●

●●

●●●●●

●●

●●

●●●●●

●●

●●

●●●●●

●●

●●

●●●●●

●●

●● ●

●●

●●

●●

●●

●●

●●

●●

●●

●●

●●

●●

●●

●●

●●

●●

●●

●●

●●

●●

●●

●●

●●

●●

●●

●●

●●

●●

●●

●●

●●

●●

●●●

●●

●●●●●●●

●●

●●●●●●●

●●

●●●●●●●

●●

●●●●●●●

●●

●●●●●●●

●●

●●●●●●●

●●

●●●●●●●

●●

●●●●●●●

●●

●●●●●●●

●●

●●●●

●●●

●●

●●●

●●

●●●

●●

●●●

●●

●●●

●●

●●●

●●

●●●

●●

●●●

●●

●●●

●●

●●●

●●

●●

●●●●●

●●

●●●●●

●●

●●●●●

●●

●●●●●

●●

●●●●●

●●

●●●●●

●●

●●●●●

●●

●●●●●

●●

●●●●●

●●

●●●

●●●●●

●●

●●●●●

●●

●●●●●

●●

●●●●●

●●

●●●●●

●●

●●●●●

●●

●●●●●

●●

●●●●●

●●

●●●●●

●●

●●●●●

●●

●●

●●

●●●

●●

●●●

●●

●●●

●●

●●●

●●

●●●

●●

●●●

●●

●●●

●●

●●●

●●

●●●

●●

●●●

1 2 5 10 25 50 100 200 400 800

0.1

0.2

0.3

0.4

Volume

PD

S

: (L,C) model

●●

●●●

●●●

●●●

●●●

●●●

●●●

●●●

●●●

●●●

● ●

●●

●●

●●

●●

●●

●●

●●

●●

●●

●●

●●

●●

●●

●●

●●

●●

●●

●●

●●

●●

●●●●●●●

●●

●●●●●●●●

●●

●●●●●●●●

●●

●●●●●●●●

●●

●●●●●●●●

●●

●●●●●●●●

●●

●●●●●●●●

●●

●●●●●●●●

●●

●●●●●●●●

●●

●●●●●●●●

●●

●●

●●●

●●

●●●

●●

●●●

●●

●●●

●●

●●●

●●

●●●

●●

●●●

●●

●●●

●●

●●●

●●

●●●

●●

●●

●●

●●

●●

●●

●●

●●

●●

●●

●●●●

●●

●●●

●●

●●●●

●●

●●●

●●

●●●●

●●

●●●

●●

●●●●

●●

●●●

●●

●●●●

●●

●●●

●●

●●●●

●●

●●●

●●

●●●●

●●

●●●

●●

●●●●

●●

●●●

●●

●●●●

●●

●●●

●●

●●●●

●●

●●●

●●

●●

●●

●●

●●

●●

●●

●●

●●

●●

●●

●●

●●

●●

●●

●●

●●

●●

●●

●●

●●

●●

●●

●●

●●

●●

●●

●●

●●

●●●●

●●●●

●●●●

●●●●

●●●●

●●●●

●●●●

●●●●

●●●●

●●●●

1 2 5 10 25 50 100 200 400 800

0.1

0.2

0.3

0.4

Volume

PD

S

: (SLI,L) model

Figure: PDSh posterior uncertainty at 10 hospitals of varying volume.

30 / 34

Page 31: Bayesian Hospital Mortality Rate Estimation and Standardization … 1... · 2019. 4. 12. · Reported indirectly and directly standardized mortality rates Major di erences between

Hospital Classification by Mortality Rates

The credibility intervals for PDSh can be used to classify

hospitals into Low, Average and High mortality according towhether its 95% interval is entirely below, intersects or isentirely above the overall average morality rate of 15%.

All Hospitals Lower Volume Quartile Upper Volume QuartileCounts

(%)Low Average High Low Average High Low Average High

(C,C) 33 4333 30 0 1116 0 32 1047 20(0.752) (98.57) (0.68) (0.00) (100.00) (0.00) (2.91) (95.27) (1.82)

(SLI,L) 58 3310 1028 0 210 906 57 1038 4(1.32) (75.30) (23.38) (0.00) (18.82) (81.18) (5.19) (94.45) (0.36)

Table: Hospital Classifications by Low, Average and High Mortality Rates.

31 / 34

Page 32: Bayesian Hospital Mortality Rate Estimation and Standardization … 1... · 2019. 4. 12. · Reported indirectly and directly standardized mortality rates Major di erences between

Attribute Effects

With PDSh values, meaningful insights into relationships

between hospital mortality rates and PTCA, NTBR are RTBRare readily obtained.

●●●

●●

●●

●●

●●

PTCA = 0 PTCA = 1

0.10

0.15

0.20

0.25

0.30

PD

S

: PDSh vs PCI

●●

●●

●●

● ●●

●●

●●

●●

●●

● ●

● ●

● ●

●●

●●

●●

● ●

●●

●●

●●

●●

●●

● ●

●●

●●

●●

●●

●●

●●

●●

●●

●●

●●

●●

●●

●●

●●

● ●●

●●

●● ●

●●

●●

●●

●●

●●

●●

●●

●●

●●

●●

●●

●●

●●

●●

●●

● ●

●●

●●

●●

●●

●●

●●

● ●

● ●

●●

●●●

●●

● ●

●●

● ●

● ●

●●

●●

●●

●●

●●

●●

● ●

●●●

● ●

●●

● ●

●●

●●

● ●●

● ●

●●

●●

●●

●●

● ●

●●

●●

●●

●●

●●

●●

●●

●●

● ●

● ●

● ●

●●

●●

●●

●●

●●

●●

●●

● ●

●●

● ●

● ●

●●

● ●

●●

●●

●●

● ●

●●

●●

●●

●●

●●

●●

●●

●●

●●

● ●

●●

●●

●●●

● ●

●●

●●

●●●

●●

●●

●● ●

●●

●●

●●

●●

●●

●●

● ●

●●

●●

●●

●●

●●

●●

● ●

● ●

●●

● ●●

●●

●●

●●

●●

●●

●●

●●

●●

●●

●●

●●

●●

●●

●●

●●

●●

●●

●●

●●

●●

●●

●●

●●

●●

●●

●●

●● ●

●●

●●

●●

●●

●●

●●

●●

●●●

●●

● ●

●●

●●

●●

●●

●●

●●

●●

●●

●● ●

●●

● ●

● ●

●●

●●

●●

●●

●●

●●

● ●

●●

●●

●●

●●

●●

● ●

●●

● ●

●●●

●●

●●

●●

● ●

●●

●●

●●

●●

●● ●

●●

●●

●●●

●●

●●

●●

●●

● ●

●●●

●●

● ●

●●

●●

●●●

●●

● ●

●●

●●

● ●

●●●

●●

●●

● ●

●●

●●

● ●

●●

●●

●●

●●

● ●

●●

●●

● ●

●●

●●

● ●

●●

●●

●●

●●●

●●

● ●

●●

●●

●●

● ●

●●

● ●

● ●

●●

●●

●●

●●

●●

● ●

●●

●●

●●

●●

●●

●●

●●

●●

●●

● ●

● ●

●●

● ● ●●

●●

●●

●●

●●

●●

●●

● ●

●●

●●

●●

●●

●●

● ●

●●

●●

●●

●●

●●

●●

● ●

●●●

0.2 0.5 1.0 2.0 5.0

0.10

0.15

0.20

0.25

0.30

Nurse−to−Bed−Ratio

PD

S

: PDSh vs NTBR

●●

●●

●●

●●●

●●

●●

●●

●●

●●

●●

●●

●●

●●

●●

●●

●●

●●

●●

●●

●●

●●

●●

●●

●●

● ●

●●

●●

●●

● ●

●●

●●

●●

●●

●●

●●

● ●●

●●

●●●

●●

●●

●●

●●●

●●

●●

●●

●●

●●

●●

●●

●●

●●

● ●

●●

●●

●●

●●

●●

● ●

●●

●●

●●

●●

●●

●●●

●●

● ●

●●

●●

●●

●●

●●

●●

●●

●●

●●

●●

●●●

●●

●●

●●

●●

●●

●●●

● ●

●●●

●●

●●

●●

●●

●●

●●

● ●

●●

●●

●●

●●

●●

●●

●●

●●

●●

●●

●●

●●

●●

●●

●●

●●

●●

●●

●●

●●

●●

●●

●●

●●

●●

●●

●●

●●

●●

●●

●●

●●

●●

●●

●●

●●

●●

●●●

●●

●●

●●

●●●

●●

●●

●● ●

●●

●●

●●

●●

●●

●●

●●

●●

●●

●●

●●

●●

● ●

●●

●●

●●

●● ●

●●

●●

●●

●●

●●

●●

●●

●●

●●

●●

●●

●●

●●

●●

●●

●●

●●

●●

●●

●●

●●

●●

●●

●●

●●

●●

●●●

●●

●●

●●

●●

●●

●●

●●

●● ●

●●

●●

●●

●●

●●

●●

●●

●●

●●

●●

●● ●

●●

●●

●●

●●

●●

●●

●●

●●

●●

● ●

●●

●●

●●

●●

●●

●●

●●

●●

●●●

●●

●●

●●

●●

●●

●●

●●

●●

●● ●

●●

●●

●●●

●●

●●

●●

●●

●●

● ●●

●●●

●●

●●

●●

●● ●

●●

● ●

●●

●●

● ●

●●●

●●

●●

●●

●●

●●

●●

●●

●●

●●

●●

●●

●●

●●

● ●

●●

●●

●●

●●

●●

●●

●●●

●●

●●

●●

●●

●●

●●

●●

●●

●●

●●

●●

●●

●●

●●

●●

●●

●●

●●

●●

●●

●●

●●

●●

●●

● ●

●●

●●

●●●●

●●

●●

●●

●●

●●

●●

●●

●●

●●

●●

●●

●●

●●

●●

●●

●●

●●

●●

●●

●●

●●●

Resident−to−Bed−Ratio

PD

S

0.10

0.15

0.20

0.25

0.30

0 0.5 1 1.5 2 2.5

: PDSh vs RTBR

Figure: PDSh under the (SLI,L) model.

32 / 34

Page 33: Bayesian Hospital Mortality Rate Estimation and Standardization … 1... · 2019. 4. 12. · Reported indirectly and directly standardized mortality rates Major di erences between

Conclusions

Strong evidence that hospital characteristics and interactions shouldbe included in the model.

Indirect standardization fails to eliminate the effect of case-mixvariation and underestimates actual mortality rates.

Directly standardized rates should be the new gold standard forpublic reporting and for further analyses of what influences mortality.

Dilemma

Should Medicare publicly report the alarmingly high mortality ratesat the low volume hospitals?

33 / 34

Page 34: Bayesian Hospital Mortality Rate Estimation and Standardization … 1... · 2019. 4. 12. · Reported indirectly and directly standardized mortality rates Major di erences between

Thank you!

References:

1 Mortality Rate Estimation and Standardization for Public Reporting:Medicare’s Hospital Compare.E. George, V. Rockova, P. Rosenbaum, V. Satopaa, J. Silber.Journal of the American Statistical Association 2017.

2 Improving Medicare’s Hospital Compare Mortality Model.J. Silber, V. Satopaa, N. Mukherjee, V. Rockova, W. Wong, A. Hill,O. Even-Shoshan, P. Rosenbaum, E. George.Health Services Research 2016.

34 / 34