week 7: bayesian inference · week 7: bayesian inference – p.8/21. two programs demonstrating...
TRANSCRIPT
![Page 1: Week 7: Bayesian inference · Week 7: Bayesian inference – p.8/21. Two programs demonstrating MCMC sampling Two excellent programs exist to demonstrate how MCMC finds peaks in](https://reader035.vdocuments.site/reader035/viewer/2022081522/5fc3f9c39cd8872ee91d3188/html5/thumbnails/1.jpg)
Week 7: Bayesian inference
Genome 570
February, 2016
Week 7: Bayesian inference – p.1/21
![Page 2: Week 7: Bayesian inference · Week 7: Bayesian inference – p.8/21. Two programs demonstrating MCMC sampling Two excellent programs exist to demonstrate how MCMC finds peaks in](https://reader035.vdocuments.site/reader035/viewer/2022081522/5fc3f9c39cd8872ee91d3188/html5/thumbnails/2.jpg)
Bayes’ Theorem
Conditional probability of hypothesis given data is:
Prob (H | D) =Prob (H & D)
Prob (D)
SinceProb (H & D) = Prob (H) Prob (D | H),
substituting this in:
Prob (H | D) =Prob (H) Prob (D | H)
Prob (D)
The denominator Prob (D) is the sum of the numerators over all possiblehypotheses H, so it is the sum of the numerators over those, giving
Prob (H | D) =Prob (H) Prob (D | H)
∑
H Prob (H) Prob (D | H)
Week 7: Bayesian inference – p.2/21
![Page 3: Week 7: Bayesian inference · Week 7: Bayesian inference – p.8/21. Two programs demonstrating MCMC sampling Two excellent programs exist to demonstrate how MCMC finds peaks in](https://reader035.vdocuments.site/reader035/viewer/2022081522/5fc3f9c39cd8872ee91d3188/html5/thumbnails/3.jpg)
A visual example of Bayes’ Theorem
P(H 1) P(H ) P(H )2 3
2P(D | H )2
P(H |D ) = 3 2+ +
12P(D | H )
prior probabilities
likel
ihoo
ds
posterior probability
2P(D | H 3 )
Week 7: Bayesian inference – p.3/21
![Page 4: Week 7: Bayesian inference · Week 7: Bayesian inference – p.8/21. Two programs demonstrating MCMC sampling Two excellent programs exist to demonstrate how MCMC finds peaks in](https://reader035.vdocuments.site/reader035/viewer/2022081522/5fc3f9c39cd8872ee91d3188/html5/thumbnails/4.jpg)
A dramatic, if not real, example
Example: “Space probe photos show no Little Green Men on Mars!”
priors
posteriors
no
yes no
yes
noyes
no
yes
likelihoods
0
no
yes
1
Week 7: Bayesian inference – p.4/21
![Page 5: Week 7: Bayesian inference · Week 7: Bayesian inference – p.8/21. Two programs demonstrating MCMC sampling Two excellent programs exist to demonstrate how MCMC finds peaks in](https://reader035.vdocuments.site/reader035/viewer/2022081522/5fc3f9c39cd8872ee91d3188/html5/thumbnails/5.jpg)
Calculations for that example
Using the Odds Ratio form of Bayes’ Theorem:
Prob (H1|D)
Prob (H2|D)=
Prob (D|H1)
Prob (D|H2)
Prob (H1)
Prob (H2)
︸ ︷︷ ︸ ︸ ︷︷ ︸ ︸︷︷︸
posterior likelihood priorodds ratio ratio odds ratio
For the odds favoring their existence, the calculation is, for the optimistabout Little Green Men:
4
1×
1/3
1=
4/3
1= 4 : 3
While for the pessimist it is
1
4×
1/3
1=
1/12
1= 1 : 12
Week 7: Bayesian inference – p.5/21
![Page 6: Week 7: Bayesian inference · Week 7: Bayesian inference – p.8/21. Two programs demonstrating MCMC sampling Two excellent programs exist to demonstrate how MCMC finds peaks in](https://reader035.vdocuments.site/reader035/viewer/2022081522/5fc3f9c39cd8872ee91d3188/html5/thumbnails/6.jpg)
With repeated observation the prior matters less
If we send 5 space probes, and all fail to see LGMs, since the probabilityof this observation is (1/3)5 if there are LGMs, and 1 if there aren’t,we get for the optimist about Little Green Men:
4
1×
(1/3)5
1= =
4/243
1= 4 : 243
while for the pessimist about Little Green Men:
1
4×
(1/3)5
1= =
1/972
1= 1 : 972
Week 7: Bayesian inference – p.6/21
![Page 7: Week 7: Bayesian inference · Week 7: Bayesian inference – p.8/21. Two programs demonstrating MCMC sampling Two excellent programs exist to demonstrate how MCMC finds peaks in](https://reader035.vdocuments.site/reader035/viewer/2022081522/5fc3f9c39cd8872ee91d3188/html5/thumbnails/7.jpg)
A coin tossing example
Theprior
0.0 0.2 0.4 0.6 0.8 1.0
p0.0 0.2 0.4 0.6 0.8 1.0
p
Thelikelihoodfunction
0.0 0.2 0.4 0.6 0.8 1.00
0.0 0.2 0.4 0.6 0.8 1.0
Theposterior
0.0 0.2 0.4 0.6 0.8 1.0
p
0.0 0.2 0.4 0.6 0.8 1.0
p
11 tosses with 5 heads 44 tosses with 20 headsWeek 7: Bayesian inference – p.7/21
![Page 8: Week 7: Bayesian inference · Week 7: Bayesian inference – p.8/21. Two programs demonstrating MCMC sampling Two excellent programs exist to demonstrate how MCMC finds peaks in](https://reader035.vdocuments.site/reader035/viewer/2022081522/5fc3f9c39cd8872ee91d3188/html5/thumbnails/8.jpg)
Markov Chain Monte Carlo sampling
To draw trees from a distribution whose probabilities are proportional tof(t), we can use the Metropolis algorithm:
1. Start at some tree. Call this Ti.2. Pick a tree that is a neighbor of this tree in the graph of trees. Call
this the proposal Tj.
3. Compute the ratio of the probabilities (or probability densityfunctions) of the proposed new tree and the old tree:
R =f(Tj)
f(Ti)
4. If R ≥ 1, accept the new tree as the current tree.5. If R < 1, draw a uniform random number (a random fraction between
0 and 1). If it is less than R, accept the new tree as the current tree.6. Otherwise reject the new tree and continue with tree Ti as the
current tree.7. Return to step 2.
Week 7: Bayesian inference – p.8/21
![Page 9: Week 7: Bayesian inference · Week 7: Bayesian inference – p.8/21. Two programs demonstrating MCMC sampling Two excellent programs exist to demonstrate how MCMC finds peaks in](https://reader035.vdocuments.site/reader035/viewer/2022081522/5fc3f9c39cd8872ee91d3188/html5/thumbnails/9.jpg)
Two programs demonstrating MCMC sampling
Two excellent programs exist to demonstrate how MCMC finds peaks in atwo-dimensional space. The user can place the peaks and can run bothregular and “heated” chains to explore the space.
Paul Lewis’s Windows program MCRobot which is available athttp://www.eeb.uconn.edu/people/plewis/software.php
John Huelsenbeck’s similar Mac OS X program McmcApp (alsocalled iMCMC) which is available athttp://cteg.berkeley.edu/software.html(hint: to place a peak on the run window, before running move thecursor on the window while holding down the mouse button. Thepeak is much wider than the initial ellipse, so keep that smallish).
We will demonstrate one of these in class.
Week 7: Bayesian inference – p.9/21
![Page 10: Week 7: Bayesian inference · Week 7: Bayesian inference – p.8/21. Two programs demonstrating MCMC sampling Two excellent programs exist to demonstrate how MCMC finds peaks in](https://reader035.vdocuments.site/reader035/viewer/2022081522/5fc3f9c39cd8872ee91d3188/html5/thumbnails/10.jpg)
Bayesian MCMC
We try to achieve the posterior
Prob (T) Prob (D | T) / (denominator)
and this turns out to need the acceptance ratio
R =Prob (Tnew) Prob (D | Tnew)
Prob (Told) Prob (D | Told)
(the denominators are the same and cancel out. This is a greatconvenience, as we often cannot evaluate the deonominator, but we canusually evaluate the numerators).
Note that we could also have a prior on model parameters too, and as wemove through tree space we could also be moving through parameterspace.
Week 7: Bayesian inference – p.10/21
![Page 11: Week 7: Bayesian inference · Week 7: Bayesian inference – p.8/21. Two programs demonstrating MCMC sampling Two excellent programs exist to demonstrate how MCMC finds peaks in](https://reader035.vdocuments.site/reader035/viewer/2022081522/5fc3f9c39cd8872ee91d3188/html5/thumbnails/11.jpg)
Using MrBayes on the primates data
BovineLemurTarsierCrab E.MacRhesus MacJpn MacaqBarbMacaqOrangChimpHumanGorillaGibbonSquir MonkMouse
0.416
0.899
1.00
0.999
1.00
0.997
0.986
0.905
0.949
0.938
0.856
Frequencies of partitions (posterior clade probabilities)Week 7: Bayesian inference – p.11/21
![Page 12: Week 7: Bayesian inference · Week 7: Bayesian inference – p.8/21. Two programs demonstrating MCMC sampling Two excellent programs exist to demonstrate how MCMC finds peaks in](https://reader035.vdocuments.site/reader035/viewer/2022081522/5fc3f9c39cd8872ee91d3188/html5/thumbnails/12.jpg)
Issues to think about with Bayesian inference
Where do you get your prior from?
Week 7: Bayesian inference – p.12/21
![Page 13: Week 7: Bayesian inference · Week 7: Bayesian inference – p.8/21. Two programs demonstrating MCMC sampling Two excellent programs exist to demonstrate how MCMC finds peaks in](https://reader035.vdocuments.site/reader035/viewer/2022081522/5fc3f9c39cd8872ee91d3188/html5/thumbnails/13.jpg)
Issues to think about with Bayesian inference
Where do you get your prior from?Are you assuming each branch has a length drawnindependently from a distribution? How wide a distribution?
Week 7: Bayesian inference – p.12/21
![Page 14: Week 7: Bayesian inference · Week 7: Bayesian inference – p.8/21. Two programs demonstrating MCMC sampling Two excellent programs exist to demonstrate how MCMC finds peaks in](https://reader035.vdocuments.site/reader035/viewer/2022081522/5fc3f9c39cd8872ee91d3188/html5/thumbnails/14.jpg)
Issues to think about with Bayesian inference
Where do you get your prior from?Are you assuming each branch has a length drawnindependently from a distribution? How wide a distribution?Or is the tree drawn from a birth-death process? If so, what arethe rates of birth and death?
Week 7: Bayesian inference – p.12/21
![Page 15: Week 7: Bayesian inference · Week 7: Bayesian inference – p.8/21. Two programs demonstrating MCMC sampling Two excellent programs exist to demonstrate how MCMC finds peaks in](https://reader035.vdocuments.site/reader035/viewer/2022081522/5fc3f9c39cd8872ee91d3188/html5/thumbnails/15.jpg)
Issues to think about with Bayesian inference
Where do you get your prior from?Are you assuming each branch has a length drawnindependently from a distribution? How wide a distribution?Or is the tree drawn from a birth-death process? If so, what arethe rates of birth and death?Or is there also a stage where the species studied are chosenare selected from all extant species of the group? How do youmodel that? Are you modelling biologists’ decision-makingprocesses?
Week 7: Bayesian inference – p.12/21
![Page 16: Week 7: Bayesian inference · Week 7: Bayesian inference – p.8/21. Two programs demonstrating MCMC sampling Two excellent programs exist to demonstrate how MCMC finds peaks in](https://reader035.vdocuments.site/reader035/viewer/2022081522/5fc3f9c39cd8872ee91d3188/html5/thumbnails/16.jpg)
Issues to think about with Bayesian inference
Where do you get your prior from?Are you assuming each branch has a length drawnindependently from a distribution? How wide a distribution?Or is the tree drawn from a birth-death process? If so, what arethe rates of birth and death?Or is there also a stage where the species studied are chosenare selected from all extant species of the group? How do youmodel that? Are you modelling biologists’ decision-makingprocesses?
Is your prior the same as your reader’s prior?
Week 7: Bayesian inference – p.12/21
![Page 17: Week 7: Bayesian inference · Week 7: Bayesian inference – p.8/21. Two programs demonstrating MCMC sampling Two excellent programs exist to demonstrate how MCMC finds peaks in](https://reader035.vdocuments.site/reader035/viewer/2022081522/5fc3f9c39cd8872ee91d3188/html5/thumbnails/17.jpg)
Issues to think about with Bayesian inference
Where do you get your prior from?Are you assuming each branch has a length drawnindependently from a distribution? How wide a distribution?Or is the tree drawn from a birth-death process? If so, what arethe rates of birth and death?Or is there also a stage where the species studied are chosenare selected from all extant species of the group? How do youmodel that? Are you modelling biologists’ decision-makingprocesses?
Is your prior the same as your reader’s prior?If not, what do you do?
Week 7: Bayesian inference – p.12/21
![Page 18: Week 7: Bayesian inference · Week 7: Bayesian inference – p.8/21. Two programs demonstrating MCMC sampling Two excellent programs exist to demonstrate how MCMC finds peaks in](https://reader035.vdocuments.site/reader035/viewer/2022081522/5fc3f9c39cd8872ee91d3188/html5/thumbnails/18.jpg)
Issues to think about with Bayesian inference
Where do you get your prior from?Are you assuming each branch has a length drawnindependently from a distribution? How wide a distribution?Or is the tree drawn from a birth-death process? If so, what arethe rates of birth and death?Or is there also a stage where the species studied are chosenare selected from all extant species of the group? How do youmodel that? Are you modelling biologists’ decision-makingprocesses?
Is your prior the same as your reader’s prior?If not, what do you do?Use several priors?
Week 7: Bayesian inference – p.12/21
![Page 19: Week 7: Bayesian inference · Week 7: Bayesian inference – p.8/21. Two programs demonstrating MCMC sampling Two excellent programs exist to demonstrate how MCMC finds peaks in](https://reader035.vdocuments.site/reader035/viewer/2022081522/5fc3f9c39cd8872ee91d3188/html5/thumbnails/19.jpg)
Issues to think about with Bayesian inference
Where do you get your prior from?Are you assuming each branch has a length drawnindependently from a distribution? How wide a distribution?Or is the tree drawn from a birth-death process? If so, what arethe rates of birth and death?Or is there also a stage where the species studied are chosenare selected from all extant species of the group? How do youmodel that? Are you modelling biologists’ decision-makingprocesses?
Is your prior the same as your reader’s prior?If not, what do you do?Use several priors?Just give the reader the likelihood surface and let them providetheir own prior?
Week 7: Bayesian inference – p.12/21
![Page 20: Week 7: Bayesian inference · Week 7: Bayesian inference – p.8/21. Two programs demonstrating MCMC sampling Two excellent programs exist to demonstrate how MCMC finds peaks in](https://reader035.vdocuments.site/reader035/viewer/2022081522/5fc3f9c39cd8872ee91d3188/html5/thumbnails/20.jpg)
An example
Suppose we have two species with a Jukes-Cantor model, so that theestimation of the (unrooted) tree is simply the estimation of the branchlength between the two species.
We can express the result either as branch length t, or as the netprobability of base change
p =3
4
(
1 − exp(−4
3t)
)
Week 7: Bayesian inference – p.13/21
![Page 21: Week 7: Bayesian inference · Week 7: Bayesian inference – p.8/21. Two programs demonstrating MCMC sampling Two excellent programs exist to demonstrate how MCMC finds peaks in](https://reader035.vdocuments.site/reader035/viewer/2022081522/5fc3f9c39cd8872ee91d3188/html5/thumbnails/21.jpg)
A flat prior on p
0.0 0.5 0.750.25
p
Week 7: Bayesian inference – p.14/21
![Page 22: Week 7: Bayesian inference · Week 7: Bayesian inference – p.8/21. Two programs demonstrating MCMC sampling Two excellent programs exist to demonstrate how MCMC finds peaks in](https://reader035.vdocuments.site/reader035/viewer/2022081522/5fc3f9c39cd8872ee91d3188/html5/thumbnails/22.jpg)
The corresponding prior on t
0 1 2 3 4 5
t
So which is the “flat prior”?Week 7: Bayesian inference – p.15/21
![Page 23: Week 7: Bayesian inference · Week 7: Bayesian inference – p.8/21. Two programs demonstrating MCMC sampling Two excellent programs exist to demonstrate how MCMC finds peaks in](https://reader035.vdocuments.site/reader035/viewer/2022081522/5fc3f9c39cd8872ee91d3188/html5/thumbnails/23.jpg)
Flat prior for t between 0 and 5
0 1 2 3 4 5
t
Week 7: Bayesian inference – p.16/21
![Page 24: Week 7: Bayesian inference · Week 7: Bayesian inference – p.8/21. Two programs demonstrating MCMC sampling Two excellent programs exist to demonstrate how MCMC finds peaks in](https://reader035.vdocuments.site/reader035/viewer/2022081522/5fc3f9c39cd8872ee91d3188/html5/thumbnails/24.jpg)
The corresponding prior on p
0.0 0.750.50.25
p
Week 7: Bayesian inference – p.17/21
![Page 25: Week 7: Bayesian inference · Week 7: Bayesian inference – p.8/21. Two programs demonstrating MCMC sampling Two excellent programs exist to demonstrate how MCMC finds peaks in](https://reader035.vdocuments.site/reader035/viewer/2022081522/5fc3f9c39cd8872ee91d3188/html5/thumbnails/25.jpg)
The invariance of the ML estimator to scale change
^t̂ = 0.383112
( p = 0.3 )
0 1 2 3 4 5−14
−12
−10
−8
−6
t
ln L
Likelihood curve for twhen 3 sites differ out of 10
Week 7: Bayesian inference – p.18/21
![Page 26: Week 7: Bayesian inference · Week 7: Bayesian inference – p.8/21. Two programs demonstrating MCMC sampling Two excellent programs exist to demonstrate how MCMC finds peaks in](https://reader035.vdocuments.site/reader035/viewer/2022081522/5fc3f9c39cd8872ee91d3188/html5/thumbnails/26.jpg)
The invariance of the ML estimator to scale change
0.0 0.2 0.4 0.6 0.8−14
−12
−10
−8
−6
ln L
p
p = 0.3^
^( t = 0.383112 )Likelihood curve for p
when 3 sites differ out of 10
Week 7: Bayesian inference – p.19/21
![Page 27: Week 7: Bayesian inference · Week 7: Bayesian inference – p.8/21. Two programs demonstrating MCMC sampling Two excellent programs exist to demonstrate how MCMC finds peaks in](https://reader035.vdocuments.site/reader035/viewer/2022081522/5fc3f9c39cd8872ee91d3188/html5/thumbnails/27.jpg)
When t has a wide flat prior
0
1
2
3
4
5
1 10 100 1000
T
t 95% interval
MLE
The 95% two-tailed credible interval for t with varioustruncation points on a flat prior for t
Week 7: Bayesian inference – p.20/21
![Page 28: Week 7: Bayesian inference · Week 7: Bayesian inference – p.8/21. Two programs demonstrating MCMC sampling Two excellent programs exist to demonstrate how MCMC finds peaks in](https://reader035.vdocuments.site/reader035/viewer/2022081522/5fc3f9c39cd8872ee91d3188/html5/thumbnails/28.jpg)
What is going on in that case is ...
0
T is so large this is < 2.5% of the area
T
Week 7: Bayesian inference – p.21/21