Bayesian perspectives for epidemiological research: 1...
Transcript of Bayesian perspectives for epidemiological research: 1...
![Page 1: Bayesian perspectives for epidemiological research: 1 ...people.musc.edu/~elg26/talks/CPCjournalclub.sept2007.pdf · Bayesian perspectives for epidemiological research: 1. Foundations](https://reader036.fdocuments.net/reader036/viewer/2022063019/5fe0c95775e7eb76581da6fe/html5/thumbnails/1.jpg)
Bayesian perspectives for epidemiological research: 1. Foundations and basic methods.Sander Greenland, International Journal of Epidemiology, 2006. 35:765-775 (with discussion)
Cancer Prevention and Control Journal ClubSeptember 13, 2007Presenter: Liz Garrett-Mayer
![Page 2: Bayesian perspectives for epidemiological research: 1 ...people.musc.edu/~elg26/talks/CPCjournalclub.sept2007.pdf · Bayesian perspectives for epidemiological research: 1. Foundations](https://reader036.fdocuments.net/reader036/viewer/2022063019/5fe0c95775e7eb76581da6fe/html5/thumbnails/2.jpg)
Introduction to Bayesian Statistics
Who knows what Bayesian statistics is?Who remembers what Bayes’ theorem is? (We’ll get back to Bayes’ theorem later)“Bayesian inference is statistical inference in which evidence or observations are used to update or to newly infer the probability that a hypothesis may be true” (wikipedia)Critical idea: UPDATE
![Page 3: Bayesian perspectives for epidemiological research: 1 ...people.musc.edu/~elg26/talks/CPCjournalclub.sept2007.pdf · Bayesian perspectives for epidemiological research: 1. Foundations](https://reader036.fdocuments.net/reader036/viewer/2022063019/5fe0c95775e7eb76581da6fe/html5/thumbnails/3.jpg)
Key Ideas of Bayesian Approach
Prior: What we believe about unknown parameters BEFORE observing data
Posterior: What we believe about unknown parameters AFTER observing data
![Page 4: Bayesian perspectives for epidemiological research: 1 ...people.musc.edu/~elg26/talks/CPCjournalclub.sept2007.pdf · Bayesian perspectives for epidemiological research: 1. Foundations](https://reader036.fdocuments.net/reader036/viewer/2022063019/5fe0c95775e7eb76581da6fe/html5/thumbnails/4.jpg)
Update?
The data UPDATES our prior beliefsHow is the data represented?
The likelihood functionA model-based representation of the data
Prior x likelihood = posterior(technically: Prior x likelihood posterior)
What the heck does that mean?Let’s use the first example…
∝
![Page 5: Bayesian perspectives for epidemiological research: 1 ...people.musc.edu/~elg26/talks/CPCjournalclub.sept2007.pdf · Bayesian perspectives for epidemiological research: 1. Foundations](https://reader036.fdocuments.net/reader036/viewer/2022063019/5fe0c95775e7eb76581da6fe/html5/thumbnails/5.jpg)
Example from Greenland
(8.3%)
(2.5%)
![Page 6: Bayesian perspectives for epidemiological research: 1 ...people.musc.edu/~elg26/talks/CPCjournalclub.sept2007.pdf · Bayesian perspectives for epidemiological research: 1. Foundations](https://reader036.fdocuments.net/reader036/viewer/2022063019/5fe0c95775e7eb76581da6fe/html5/thumbnails/6.jpg)
Let’s start with what we knowLikelihoodThis is the probability model that we assumeSetting:
Large NOR (or RR)
Common approach:Use a normal approximation
Assume log(RR) ~ NormalMean = log(ad/bc)Variance = 1/a + 1/b + 1/c + 1/d
Familiar? I HOPE SO!
![Page 7: Bayesian perspectives for epidemiological research: 1 ...people.musc.edu/~elg26/talks/CPCjournalclub.sept2007.pdf · Bayesian perspectives for epidemiological research: 1. Foundations](https://reader036.fdocuments.net/reader036/viewer/2022063019/5fe0c95775e7eb76581da6fe/html5/thumbnails/7.jpg)
Greenland’s calculations
Mean = log(3*193/(5*33)) = 3.51Variance = 1/3 + 1/33 + 1/5 + 1/193 = 0.57Then what?95% CI:
Exponentiate:
)74.2,22.0(57.096.1)51.3log(
var96.1)ˆlog(
−=±
± ψψ
)4.15,80.0()74.2,22.0exp( =−
![Page 8: Bayesian perspectives for epidemiological research: 1 ...people.musc.edu/~elg26/talks/CPCjournalclub.sept2007.pdf · Bayesian perspectives for epidemiological research: 1. Foundations](https://reader036.fdocuments.net/reader036/viewer/2022063019/5fe0c95775e7eb76581da6fe/html5/thumbnails/8.jpg)
“What the data says”
OR = 3.5195% CI: (0.80, 15.4)
FREQUENTIST SO FAR……
![Page 9: Bayesian perspectives for epidemiological research: 1 ...people.musc.edu/~elg26/talks/CPCjournalclub.sept2007.pdf · Bayesian perspectives for epidemiological research: 1. Foundations](https://reader036.fdocuments.net/reader036/viewer/2022063019/5fe0c95775e7eb76581da6fe/html5/thumbnails/9.jpg)
What about this “Prior” business?
The prior QUANTIFIES what we know or expect before doing the studyFrequentist: “I know nothing”
Regardless of prior researchRegardless of simple intuition
Bayesian: “I know something”Might be quite a lotMight be not much, but still quantifiable
![Page 10: Bayesian perspectives for epidemiological research: 1 ...people.musc.edu/~elg26/talks/CPCjournalclub.sept2007.pdf · Bayesian perspectives for epidemiological research: 1. Foundations](https://reader036.fdocuments.net/reader036/viewer/2022063019/5fe0c95775e7eb76581da6fe/html5/thumbnails/10.jpg)
Priors0.
00.
40.
8
0.02
5
0.05 0.1
0.25 0.5 1 2 4 10 20 40
Greenland Prior
RR
Den
sity
![Page 11: Bayesian perspectives for epidemiological research: 1 ...people.musc.edu/~elg26/talks/CPCjournalclub.sept2007.pdf · Bayesian perspectives for epidemiological research: 1. Foundations](https://reader036.fdocuments.net/reader036/viewer/2022063019/5fe0c95775e7eb76581da6fe/html5/thumbnails/11.jpg)
Priors0.
00.
40.
8
0.02
5
0.05 0.1
0.25 0.5 1 2 4 10 20 40
Greenland Prior
RR
Den
sity
![Page 12: Bayesian perspectives for epidemiological research: 1 ...people.musc.edu/~elg26/talks/CPCjournalclub.sept2007.pdf · Bayesian perspectives for epidemiological research: 1. Foundations](https://reader036.fdocuments.net/reader036/viewer/2022063019/5fe0c95775e7eb76581da6fe/html5/thumbnails/12.jpg)
Priors0.
00.
40.
8
0.02
5
0.05 0.1
0.25 0.5 1 2 4 10 20 40
Greenland Prior
RR
Den
sity
![Page 13: Bayesian perspectives for epidemiological research: 1 ...people.musc.edu/~elg26/talks/CPCjournalclub.sept2007.pdf · Bayesian perspectives for epidemiological research: 1. Foundations](https://reader036.fdocuments.net/reader036/viewer/2022063019/5fe0c95775e7eb76581da6fe/html5/thumbnails/13.jpg)
Priors0.
00.
40.
8
0.02
5
0.05 0.1
0.25 0.5 1 2 4 10 20 40
Greenland Prior
0.0
0.4
0.8
1e-0
4
0.00
1
0.01 0.1 1 10 100
1000
1000
0
FrequentistRR
RR
Den
sity
Den
sity
![Page 14: Bayesian perspectives for epidemiological research: 1 ...people.musc.edu/~elg26/talks/CPCjournalclub.sept2007.pdf · Bayesian perspectives for epidemiological research: 1. Foundations](https://reader036.fdocuments.net/reader036/viewer/2022063019/5fe0c95775e7eb76581da6fe/html5/thumbnails/14.jpg)
What do you think?
Is it reasonable to exclude RR’s>10?RR’s>100?RR’s>1000?
Frequentists exclude nothing
![Page 15: Bayesian perspectives for epidemiological research: 1 ...people.musc.edu/~elg26/talks/CPCjournalclub.sept2007.pdf · Bayesian perspectives for epidemiological research: 1. Foundations](https://reader036.fdocuments.net/reader036/viewer/2022063019/5fe0c95775e7eb76581da6fe/html5/thumbnails/15.jpg)
Getting the posterior
Combine the prior and the data (likelihood)How?
Multiply the prior x likelihood ORAdd the log-prior + log-likelihood
Resulting distributionposterior ORlog-posterior
Sound hard?For some prior-likelihood combinations, very simple“Conjugate” prior
Example: normal is easy
![Page 16: Bayesian perspectives for epidemiological research: 1 ...people.musc.edu/~elg26/talks/CPCjournalclub.sept2007.pdf · Bayesian perspectives for epidemiological research: 1. Foundations](https://reader036.fdocuments.net/reader036/viewer/2022063019/5fe0c95775e7eb76581da6fe/html5/thumbnails/16.jpg)
Posterior Distribution
Define the following:
Then, the distribution of the posterior is:
),(~
),(~)ˆlog(
2
2
τμ
σμψ
mNand
N
⎟⎟⎟⎟
⎠
⎞
⎜⎜⎜⎜
⎝
⎛
⎟⎠⎞
⎜⎝⎛ +
+
+ −1
22
22
22 11,11
)ˆlog(
στστ
σψ
τm
NOUCH!
Note thatσ is thestandarderror!
![Page 17: Bayesian perspectives for epidemiological research: 1 ...people.musc.edu/~elg26/talks/CPCjournalclub.sept2007.pdf · Bayesian perspectives for epidemiological research: 1. Foundations](https://reader036.fdocuments.net/reader036/viewer/2022063019/5fe0c95775e7eb76581da6fe/html5/thumbnails/17.jpg)
Not as bad as it seems….
m = 0, τ2 = 0.5Posterior becomes:
⎟⎟⎠
⎞⎜⎜⎝
⎛++
=⎟⎟⎟⎟
⎠
⎞
⎜⎜⎜⎜
⎝
⎛
⎟⎠⎞
⎜⎝⎛ +
+
−
12,
12)ˆlog(12,12
)ˆlog(
2
2
2
1
2
2
2
σσ
σψ
σσ
σψ
NN
For large n, what happens?
![Page 18: Bayesian perspectives for epidemiological research: 1 ...people.musc.edu/~elg26/talks/CPCjournalclub.sept2007.pdf · Bayesian perspectives for epidemiological research: 1. Foundations](https://reader036.fdocuments.net/reader036/viewer/2022063019/5fe0c95775e7eb76581da6fe/html5/thumbnails/18.jpg)
Can you take just one more mathy slide?
Note our estimates:ψ = log(3.51)σ2 = 0.57
( )27.0,59.012,12
)ˆlog(1
2
2
2NN =
⎟⎟⎟⎟
⎠
⎞
⎜⎜⎜⎜
⎝
⎛
⎟⎠⎞
⎜⎝⎛ +
+
−
σσ
σψ
![Page 19: Bayesian perspectives for epidemiological research: 1 ...people.musc.edu/~elg26/talks/CPCjournalclub.sept2007.pdf · Bayesian perspectives for epidemiological research: 1. Foundations](https://reader036.fdocuments.net/reader036/viewer/2022063019/5fe0c95775e7eb76581da6fe/html5/thumbnails/19.jpg)
0.0
0.2
0.4
0.6
0.8
1.0
0.02
5
0.05 0.1
0.25 0.5 1 2 4 10 20 40
PriorLikelihoodPosterior
Maybe a picture would bring us back to earth…..
![Page 20: Bayesian perspectives for epidemiological research: 1 ...people.musc.edu/~elg26/talks/CPCjournalclub.sept2007.pdf · Bayesian perspectives for epidemiological research: 1. Foundations](https://reader036.fdocuments.net/reader036/viewer/2022063019/5fe0c95775e7eb76581da6fe/html5/thumbnails/20.jpg)
0.0
0.2
0.4
0.6
0.8
1.0
0.02
5
0.05 0.1
0.25 0.5 1 2 4 10 20 40
PriorLikelihoodPosterior
Maybe a picture would bring us back to earth…..
![Page 21: Bayesian perspectives for epidemiological research: 1 ...people.musc.edu/~elg26/talks/CPCjournalclub.sept2007.pdf · Bayesian perspectives for epidemiological research: 1. Foundations](https://reader036.fdocuments.net/reader036/viewer/2022063019/5fe0c95775e7eb76581da6fe/html5/thumbnails/21.jpg)
Maybe a picture would bring us back to earth…..
0.0
0.2
0.4
0.6
0.8
1.0
0.02
5
0.05 0.1
0.25 0.5 1 2 4 10 20 40
PriorLikelihoodPosterior
![Page 22: Bayesian perspectives for epidemiological research: 1 ...people.musc.edu/~elg26/talks/CPCjournalclub.sept2007.pdf · Bayesian perspectives for epidemiological research: 1. Foundations](https://reader036.fdocuments.net/reader036/viewer/2022063019/5fe0c95775e7eb76581da6fe/html5/thumbnails/22.jpg)
What about other priors: flatter0.
00.
20.
40.
60.
81.
0
0.02
5
0.05 0.1
0.25 0.5 1 2 4 10 20 40
PriorLikelihoodPosterior
![Page 23: Bayesian perspectives for epidemiological research: 1 ...people.musc.edu/~elg26/talks/CPCjournalclub.sept2007.pdf · Bayesian perspectives for epidemiological research: 1. Foundations](https://reader036.fdocuments.net/reader036/viewer/2022063019/5fe0c95775e7eb76581da6fe/html5/thumbnails/23.jpg)
What about other priors: steeper0.
00.
20.
40.
60.
81.
0
0.02
5
0.05 0.1
0.25 0.5 1 2 4 10 20 40
PriorLikelihoodPosterior
![Page 24: Bayesian perspectives for epidemiological research: 1 ...people.musc.edu/~elg26/talks/CPCjournalclub.sept2007.pdf · Bayesian perspectives for epidemiological research: 1. Foundations](https://reader036.fdocuments.net/reader036/viewer/2022063019/5fe0c95775e7eb76581da6fe/html5/thumbnails/24.jpg)
What if we had more data?What if we had twice as much data?
0.0
0.2
0.4
0.6
0.8
1.0
0.02
5
0.05 0.1
0.25 0.5 1 2 4 10 20 40
PriorLikelihoodPosterior
![Page 25: Bayesian perspectives for epidemiological research: 1 ...people.musc.edu/~elg26/talks/CPCjournalclub.sept2007.pdf · Bayesian perspectives for epidemiological research: 1. Foundations](https://reader036.fdocuments.net/reader036/viewer/2022063019/5fe0c95775e7eb76581da6fe/html5/thumbnails/25.jpg)
What if we had more data?What if we had 10 times as much data?
0.0
0.2
0.4
0.6
0.8
1.0
0.02
5
0.05 0.1
0.25 0.5 1 2 4 10 20 40
PriorLikelihoodPosterior
![Page 26: Bayesian perspectives for epidemiological research: 1 ...people.musc.edu/~elg26/talks/CPCjournalclub.sept2007.pdf · Bayesian perspectives for epidemiological research: 1. Foundations](https://reader036.fdocuments.net/reader036/viewer/2022063019/5fe0c95775e7eb76581da6fe/html5/thumbnails/26.jpg)
So….easy, right?
There are even simpler approximationsCombining the prior and likelihood:
Computationally easyStatisticians have already done all the hard part!Approximations work pretty well!
![Page 27: Bayesian perspectives for epidemiological research: 1 ...people.musc.edu/~elg26/talks/CPCjournalclub.sept2007.pdf · Bayesian perspectives for epidemiological research: 1. Foundations](https://reader036.fdocuments.net/reader036/viewer/2022063019/5fe0c95775e7eb76581da6fe/html5/thumbnails/27.jpg)
OK…so what does Sander Greenland have to say about all of this?
A LOT!Bayesian statistical methods have been around for a long timeThey were “pushed out” awhile back in favor of frequentist methods
‘randomized’ trials and surveyscomputation
However, epidemiologic studies are generally not randomizedBayesian approaches allow you to ‘adjust’ for some of the subjectivity involved
control for biasadjust for knowledge outside the trial
![Page 28: Bayesian perspectives for epidemiological research: 1 ...people.musc.edu/~elg26/talks/CPCjournalclub.sept2007.pdf · Bayesian perspectives for epidemiological research: 1. Foundations](https://reader036.fdocuments.net/reader036/viewer/2022063019/5fe0c95775e7eb76581da6fe/html5/thumbnails/28.jpg)
Education in Bayesian approaches is lacking
Some think that the Bayesian approach is too complicatedNot really….needs to be put into contextBalance of objective versus subjectiveConsideration of implications of the prior is a useful exercise in many situationsFocus has been on the computational differencesBut, the focus should be on the philosophical differences
![Page 29: Bayesian perspectives for epidemiological research: 1 ...people.musc.edu/~elg26/talks/CPCjournalclub.sept2007.pdf · Bayesian perspectives for epidemiological research: 1. Foundations](https://reader036.fdocuments.net/reader036/viewer/2022063019/5fe0c95775e7eb76581da6fe/html5/thumbnails/29.jpg)
Why Epi?
‘Huge uncertainties’measurement errorprocess generating the datalack of information about confounderslack of knowledge of what potential confounders may be
Precise computation ‘not necessary’Informative priors can help!
![Page 30: Bayesian perspectives for epidemiological research: 1 ...people.musc.edu/~elg26/talks/CPCjournalclub.sept2007.pdf · Bayesian perspectives for epidemiological research: 1. Foundations](https://reader036.fdocuments.net/reader036/viewer/2022063019/5fe0c95775e7eb76581da6fe/html5/thumbnails/30.jpg)
Frequentism vs. Subjective Bayes
Arguments against Bayes:assumptions and models are subjectivepropaganda (via the prior)
Actually: these can also be considered arguments against Frequentism
frequentists pick models toorecall the ‘frequentist’ prior (or lack thereof
Bayesians are often just more explicit which ends up encouraging criticism.
![Page 31: Bayesian perspectives for epidemiological research: 1 ...people.musc.edu/~elg26/talks/CPCjournalclub.sept2007.pdf · Bayesian perspectives for epidemiological research: 1. Foundations](https://reader036.fdocuments.net/reader036/viewer/2022063019/5fe0c95775e7eb76581da6fe/html5/thumbnails/31.jpg)
Priors
Not arbitraryCan arise from different sources
data from other studiesmine versus yoursshould allow for possible biasesshould allow for lack of generalizability across studies
![Page 32: Bayesian perspectives for epidemiological research: 1 ...people.musc.edu/~elg26/talks/CPCjournalclub.sept2007.pdf · Bayesian perspectives for epidemiological research: 1. Foundations](https://reader036.fdocuments.net/reader036/viewer/2022063019/5fe0c95775e7eb76581da6fe/html5/thumbnails/32.jpg)
The other part: the likelihood
another assumption!both Bayes and Frequentists requires a modelTHIS IS NOT NECESSARILY AN OBVIOUS CHOICECan try more than one model(and can try more than one prior)Generally: results are more sensitive to likelihood choice than prior choice
![Page 33: Bayesian perspectives for epidemiological research: 1 ...people.musc.edu/~elg26/talks/CPCjournalclub.sept2007.pdf · Bayesian perspectives for epidemiological research: 1. Foundations](https://reader036.fdocuments.net/reader036/viewer/2022063019/5fe0c95775e7eb76581da6fe/html5/thumbnails/33.jpg)
“Objective Bayes”
Use vague priorsTo Greenland, probably even more terriblemake the presumption of complete ignorance very obvious!
“They make these unrealistic ‘non-informative’priors explicit, and so produce posterior intervals that represent the inference of no one (except perhaps someone who understood nothing of the subject under study or even the meaning of the variable names).”
![Page 34: Bayesian perspectives for epidemiological research: 1 ...people.musc.edu/~elg26/talks/CPCjournalclub.sept2007.pdf · Bayesian perspectives for epidemiological research: 1. Foundations](https://reader036.fdocuments.net/reader036/viewer/2022063019/5fe0c95775e7eb76581da6fe/html5/thumbnails/34.jpg)
Frequentist vs Bayesian Divergence
“Frequentist methods pretend that the models are laws of chance in the real world”
“The subjective Bayesian interpretation is much less ambitious (and less confident)…insofar as it treats the models and the analysis results as systems of personal judgements, possibly poor ones, rather than as some sort of objective reality”
![Page 35: Bayesian perspectives for epidemiological research: 1 ...people.musc.edu/~elg26/talks/CPCjournalclub.sept2007.pdf · Bayesian perspectives for epidemiological research: 1. Foundations](https://reader036.fdocuments.net/reader036/viewer/2022063019/5fe0c95775e7eb76581da6fe/html5/thumbnails/35.jpg)
Frequentist ‘fantasy’
95% CIs are presented as if they account for random error” (e.g., measurement error)No mention of random error is generally mentioned Frequentist approaches do not account for biasesThe approach treats observational data as if it arose from a randomized experiment or random sample“Letting the data speak for themselves” assumes lack of biases, which is often false assumption
![Page 36: Bayesian perspectives for epidemiological research: 1 ...people.musc.edu/~elg26/talks/CPCjournalclub.sept2007.pdf · Bayesian perspectives for epidemiological research: 1. Foundations](https://reader036.fdocuments.net/reader036/viewer/2022063019/5fe0c95775e7eb76581da6fe/html5/thumbnails/36.jpg)
Some “EASY” Bayesian approaches:Information Weighting
Recall the exampleObserved data:
log(ψ) = log(3.51)var(log(ψ)) = 0.57
Prior assumption:log(ψ) = 0var(log(ψ)) = 0.5
![Page 37: Bayesian perspectives for epidemiological research: 1 ...people.musc.edu/~elg26/talks/CPCjournalclub.sept2007.pdf · Bayesian perspectives for epidemiological research: 1. Foundations](https://reader036.fdocuments.net/reader036/viewer/2022063019/5fe0c95775e7eb76581da6fe/html5/thumbnails/37.jpg)
Information Weighting
Generate a weighted average of the means, weighting by the variances
Note: exp(0.587) = 1.80
587.0
))51.3(log()0()log(posterior
57.01
5.01
57.01
5.01
=
++
=ψ
![Page 38: Bayesian perspectives for epidemiological research: 1 ...people.musc.edu/~elg26/talks/CPCjournalclub.sept2007.pdf · Bayesian perspectives for epidemiological research: 1. Foundations](https://reader036.fdocuments.net/reader036/viewer/2022063019/5fe0c95775e7eb76581da6fe/html5/thumbnails/38.jpg)
Remember this? Same result!0.
00.
20.
40.
60.
81.
0
0.02
5
0.05 0.1
0.25 0.5 1 2 4 10 20 40
PriorLikelihoodPosterior
![Page 39: Bayesian perspectives for epidemiological research: 1 ...people.musc.edu/~elg26/talks/CPCjournalclub.sept2007.pdf · Bayesian perspectives for epidemiological research: 1. Foundations](https://reader036.fdocuments.net/reader036/viewer/2022063019/5fe0c95775e7eb76581da6fe/html5/thumbnails/39.jpg)
How can we interpret this?
Both the DATA and the PRIOR are weakBoth have similar varianceBut, posterior mean is quite different than ‘simple’ mean
frequentist RR = 3.51Bayesian RR = 1.80 (=exp(0.266))(prior RR = 1)
Interestingly, average RR in other studies is about 1.70
![Page 40: Bayesian perspectives for epidemiological research: 1 ...people.musc.edu/~elg26/talks/CPCjournalclub.sept2007.pdf · Bayesian perspectives for epidemiological research: 1. Foundations](https://reader036.fdocuments.net/reader036/viewer/2022063019/5fe0c95775e7eb76581da6fe/html5/thumbnails/40.jpg)
Seems like a big difference!
Yes and noDepends on the sample sizeHere, the data were quite weakN = 234, but very few cases
![Page 41: Bayesian perspectives for epidemiological research: 1 ...people.musc.edu/~elg26/talks/CPCjournalclub.sept2007.pdf · Bayesian perspectives for epidemiological research: 1. Foundations](https://reader036.fdocuments.net/reader036/viewer/2022063019/5fe0c95775e7eb76581da6fe/html5/thumbnails/41.jpg)
Extendable?
So, sort of easyBut, not practical….when is the last time we didn’t adjust for age, race and gender!Can still do this with adjustment All you need is
the adjusted estimates of RR and its varianceyour prior mean and variance
![Page 42: Bayesian perspectives for epidemiological research: 1 ...people.musc.edu/~elg26/talks/CPCjournalclub.sept2007.pdf · Bayesian perspectives for epidemiological research: 1. Foundations](https://reader036.fdocuments.net/reader036/viewer/2022063019/5fe0c95775e7eb76581da6fe/html5/thumbnails/42.jpg)
Another approach:Including “prior data”
Frequentist approach to Bayes!Include “pseudo-data” that is consistent with your prior beliefsCan do some simple calculations to figure out how muchProcess:
construct data equivalent to the prioradd those data to datasetresulting estimate and its variance are actually posterior estimates
Aside: This is a popular approach in Bayesian phase I trials!
![Page 43: Bayesian perspectives for epidemiological research: 1 ...people.musc.edu/~elg26/talks/CPCjournalclub.sept2007.pdf · Bayesian perspectives for epidemiological research: 1. Foundations](https://reader036.fdocuments.net/reader036/viewer/2022063019/5fe0c95775e7eb76581da6fe/html5/thumbnails/43.jpg)
Discussion
Data alone say nothingneed at least a model!both frequentists and Bayesian must choose a model
If frequentist results arise from ‘perfect randomized trial’
prior not very necessaryvagueness appropriate
observational studies not perfect!Bayesian statistics become worthwhile exactly when frequentist approaches break down