Bayesian Analysis for Small Subgroups
Transcript of Bayesian Analysis for Small Subgroups
![Page 1: Bayesian Analysis for Small Subgroups](https://reader030.fdocuments.in/reader030/viewer/2022012422/6176657d52428742156220a9/html5/thumbnails/1.jpg)
Bayesian Analysis for Small Subgroups
Building Strong Evidence in Challenging Contexts:
Alternatives to Traditional Randomized Controlled Trials
Intuitive Inferences with Heightened Precision
Mariel Finucane
September 22, 2016 • OPRE
![Page 2: Bayesian Analysis for Small Subgroups](https://reader030.fdocuments.in/reader030/viewer/2022012422/6176657d52428742156220a9/html5/thumbnails/2.jpg)
2
Problem:
We often have samples that are too
small to produce reliable estimates
![Page 3: Bayesian Analysis for Small Subgroups](https://reader030.fdocuments.in/reader030/viewer/2022012422/6176657d52428742156220a9/html5/thumbnails/3.jpg)
3
Problem:
We often have samples that are too
small to produce reliable estimates
Solution:
Hierarchical Bayes
![Page 4: Bayesian Analysis for Small Subgroups](https://reader030.fdocuments.in/reader030/viewer/2022012422/6176657d52428742156220a9/html5/thumbnails/4.jpg)
4
“Shrinkage”
Problem:
We often have samples that are too
small to produce reliable estimates
Solution:
Hierarchical Bayes
Borrowing strength
Reliability adjustment
Stabilization
Random effects modeling
Shrinkage
Multilevel modeling
Empirical Bayes
Small-area estimation
Variance components
Partial pooling...
![Page 5: Bayesian Analysis for Small Subgroups](https://reader030.fdocuments.in/reader030/viewer/2022012422/6176657d52428742156220a9/html5/thumbnails/5.jpg)
5
Goals for Today
• Why Hierarchical Bayes?
– Baseball
– Health policy impact evaluation
![Page 6: Bayesian Analysis for Small Subgroups](https://reader030.fdocuments.in/reader030/viewer/2022012422/6176657d52428742156220a9/html5/thumbnails/6.jpg)
6
Baseball
![Page 7: Bayesian Analysis for Small Subgroups](https://reader030.fdocuments.in/reader030/viewer/2022012422/6176657d52428742156220a9/html5/thumbnails/7.jpg)
7
Health Policy Impact Evaluation
![Page 8: Bayesian Analysis for Small Subgroups](https://reader030.fdocuments.in/reader030/viewer/2022012422/6176657d52428742156220a9/html5/thumbnails/8.jpg)
8
Health Policy Impact Evaluation
![Page 9: Bayesian Analysis for Small Subgroups](https://reader030.fdocuments.in/reader030/viewer/2022012422/6176657d52428742156220a9/html5/thumbnails/9.jpg)
9
Solution:
Bayesian models yield intuitive probabilistic inference—given the observed data, how likely is it that the true effect exceeds 0? Or 20% of
the mean? Or the cost of the intervention?
Problem:
Policymakers want more meaningful inferences than just whether they can reject the null hypothesis that a program had no effect
![Page 10: Bayesian Analysis for Small Subgroups](https://reader030.fdocuments.in/reader030/viewer/2022012422/6176657d52428742156220a9/html5/thumbnails/10.jpg)
10
Frequentist
Fixed:
parameter
Random: data
“There is a 13 percent chance of
observing such an extreme reduction
in costs if the program has no effect.”
Bayesian
Fixed: data
Random:
parameter
“There is a 97 percent chance that
the program reduced costs enough
to recover the $20 fee.”
Health Policy Impact Evaluation
![Page 11: Bayesian Analysis for Small Subgroups](https://reader030.fdocuments.in/reader030/viewer/2022012422/6176657d52428742156220a9/html5/thumbnails/11.jpg)
11
Health Policy Impact Evaluation
![Page 12: Bayesian Analysis for Small Subgroups](https://reader030.fdocuments.in/reader030/viewer/2022012422/6176657d52428742156220a9/html5/thumbnails/12.jpg)
12
Health Policy Impact Evaluation
![Page 13: Bayesian Analysis for Small Subgroups](https://reader030.fdocuments.in/reader030/viewer/2022012422/6176657d52428742156220a9/html5/thumbnails/13.jpg)
13
What’s the catch?
• Absence of conventional standards of
evidence could lead to greater
adoption of ineffective programs
• Potentially confusing to policymakers
and stakeholders
• Results depend on assumptions,
which can introduce bias
![Page 14: Bayesian Analysis for Small Subgroups](https://reader030.fdocuments.in/reader030/viewer/2022012422/6176657d52428742156220a9/html5/thumbnails/14.jpg)
14
Appendix 1: Bias/Variance Tradeoff
![Page 15: Bayesian Analysis for Small Subgroups](https://reader030.fdocuments.in/reader030/viewer/2022012422/6176657d52428742156220a9/html5/thumbnails/15.jpg)
15
Appendix 1: Bias/Variance Tradeoff
![Page 16: Bayesian Analysis for Small Subgroups](https://reader030.fdocuments.in/reader030/viewer/2022012422/6176657d52428742156220a9/html5/thumbnails/16.jpg)
16
Appendix 1: Bias/Variance Tradeoff
![Page 17: Bayesian Analysis for Small Subgroups](https://reader030.fdocuments.in/reader030/viewer/2022012422/6176657d52428742156220a9/html5/thumbnails/17.jpg)
17
Appendix 2: Hospital Performance Measurement
![Page 18: Bayesian Analysis for Small Subgroups](https://reader030.fdocuments.in/reader030/viewer/2022012422/6176657d52428742156220a9/html5/thumbnails/18.jpg)
18
“CMS seeks to report on systematic differences in patient outcome due to hospital quality…
Inherent randomness causes directly estimated hospital
effects to vary more than the systematic effects that are to be identified. Large reductions in
reported variation [via shrinkage] are appropriate for
hospital performance measures.”
Appendix 2: Hospital Performance Measurement