Bayes - Medreonet
Bayes - Medreonet
Bayes - Medreonet
Create successful ePaper yourself
Turn your PDF publications into a flip-book with our unique Google optimized e-Paper software.
Background
Deductive logical schemes<br />
Result 1<br />
Result 2<br />
cause<br />
Result 3<br />
Result 4
Inductive logical schemes<br />
cause<br />
cause<br />
cause<br />
Result 1<br />
Result 2<br />
Result 3<br />
Result 4<br />
Given a series of outcomes, it is possible to hypothesize the<br />
most likely cause(s). The more the prior information we have,<br />
the better the reliability of our estimates
Two rules are important to apply <strong>Bayes</strong>'<br />
theorem:<br />
RULE OF SUM:<br />
P<br />
( A) + P( A)<br />
= 1<br />
where: P ( A) = P( A not occurring)
Two rules are important to apply <strong>Bayes</strong>'<br />
theorem:<br />
RULE OF PRODUCT:<br />
P<br />
P<br />
( A∩<br />
B) = P( B) ⋅ P( A B)<br />
( B ∩ A) = P( A)<br />
⋅ P(<br />
B / A)<br />
P<br />
( B A<br />
)<br />
=<br />
P<br />
( B<br />
) ⋅<br />
P<br />
( A B<br />
)<br />
P( A)
<strong>Bayes</strong>'<br />
theorem
P<br />
( B A)<br />
=<br />
P<br />
( B) ⋅ P( A B)<br />
P( A)<br />
If we consider B the hypothesis and A the set of collected data:<br />
P<br />
where:<br />
( hypothesis | data)<br />
=<br />
P<br />
( hypothesis)<br />
⋅ P(<br />
data / hypothesis)<br />
P<br />
( data)<br />
P(hypothesis) = Prior probability of the hypothesis is the probability of the<br />
hypothesis being true before we collect and analyze the set of data, that is<br />
the state of knowledge before the analysis of data<br />
P(data | hypothesis) = Likelihood function obtained after the experiment<br />
P(hypothesis | data) = Posterior probability of the hypothesis after the<br />
analysis of data<br />
P(data) = Probability of all possible cases giving those data
Example:<br />
Predictive value of a positive result to a diagnostic<br />
test.<br />
We take an animal belonging to a population with<br />
prevalence of infection PR to a certain disease<br />
The hypothesis is "the animal is infected"
P(hypothesis) = probability of the hypothesis being true before<br />
we collect and analyze the set of data = PR prevalence of<br />
infection<br />
We test the animal and get a positive result. This is the data<br />
collection.<br />
P(data | hypothesis) = Likelihood function obtained after the<br />
experiment = probability of a positive test result (the data we<br />
got) given that the animal is infected (that's our hypothesis) = it<br />
is the Sensitivity of our diagnostic test!<br />
P(data) = Probability of all possible cases giving those data. A<br />
positive animal can be a either True positive or a False positive.<br />
Probability of being a True positive is PR times Sensitivity of the test<br />
Probability of being a False positive is (1-PR) times (1-Specificity of the<br />
test)<br />
so, the Predictive value of a positive result to a diagnostic<br />
test, i.e. the probability of our positive animal being infected is:<br />
P<br />
( )<br />
( hypothesis)<br />
⋅ P(<br />
data / hypothesis)<br />
PR ⋅ Sensitivity<br />
P hypothesis | data =<br />
=<br />
P( data) PR ⋅ Sensitivity + ( 1−<br />
PR) ⋅( 1−<br />
Specificity)
<strong>Bayes</strong>ian inference
<strong>Bayes</strong>ian inference<br />
<strong>Bayes</strong>ian inference is a useful, powerful technique whereby<br />
newly acquired empirical data can be combined with<br />
existing information,<br />
whether that information is itself based on pre-existing<br />
empirical data or on expert opinion,<br />
to improve an estimate of the parameter(s) used to<br />
characterise a distribution.<br />
Note: <strong>Bayes</strong>ian inference is based on subjective probability (q.v.)
<strong>Bayes</strong>ian inference<br />
<strong>Bayes</strong>ian inference is a natural extension of <strong>Bayes</strong>’<br />
theorem<br />
It provides a powerful and flexible means of learning<br />
from experience<br />
As new information becomes available it enables our<br />
existing knowledge to be easily and logically updated<br />
It explicitly acknowledges subjectivity and describes<br />
the learning process mathematically<br />
We begin with an opinion, however vague, and modify<br />
it as new information becomes available
<strong>Bayes</strong>ian inference<br />
<br />
<strong>Bayes</strong>ian inference involves three steps:<br />
1. Determining a prior estimate of a parameter in the form of<br />
a probability distribution that expresses our state of<br />
knowledge (or ignorance) before any observations are<br />
made.<br />
2. Finding an appropriate likelihood function for the<br />
2. Finding an appropriate likelihood function for the<br />
observed data. The likelihood function calculates the<br />
probability of observing the data for a given value of the<br />
prior estimate of the parameter<br />
3. Calculating the posterior (i.e. revised) estimate of the<br />
parameter in form of a probability distribution of all<br />
possible values of the parameter
Step 1. Prior distributions<br />
<br />
<br />
<br />
A prior distribution expresses our state of knowledge<br />
before any new observations are made<br />
The prior distribution is not necessarily dependent on<br />
data and may be purely subjective<br />
Depending on the circumstances there are several<br />
options available:<br />
1. Uninformed priors<br />
2. Informed priors
Uninformed priors<br />
An uninformed prior does not provide any<br />
additional information to a <strong>Bayes</strong>ian inference<br />
other than establishing a possible range.<br />
For example, in some circumstances we may not have any<br />
information about the likely prevalence of infection within a<br />
herd.<br />
We might assume that, for a particular disease, the prevalence<br />
is likely to range from 0% to 30% and that any value within<br />
this range is equally as likely as any other value.<br />
This constitutes a uniform prior, Uniform(0,0.3), and has no<br />
influence on the <strong>Bayes</strong>ian inference calculation, apart from<br />
establishing a range
Other examples of uninformed priors<br />
We might want to estimate the average number of<br />
disease outbreaks per year (λ)<br />
If we assume that each outbreak is independent of<br />
every other outbreak and that there is a constant and<br />
continuous probability of a disease outbreak occurring<br />
throughout the year, then the outbreaks follow a<br />
Poisson process.
Other examples of uninformed priors<br />
The average number of outbreaks per year can also be<br />
expressed as 1/β, where β is the mean interval between<br />
events<br />
We might think it is reasonable to assign an<br />
We might think it is reasonable to assign an<br />
uninformed prior in the form of a uniform<br />
distribution, Uniform(0,x), to λ.
Other examples of uninformed priors<br />
However, we could have just as easily parameterised the<br />
problem in terms of β.<br />
Since β=1/λ our prior distribution would be 1/Uniform(0,x)<br />
which is clearly not uninformed with respect to β.<br />
probability<br />
0.35<br />
0.30<br />
0.25<br />
0.20<br />
0.15<br />
0.10<br />
prior distribution for β<br />
expressed as<br />
1/λ = 1/Uniform(0,x)<br />
0.05<br />
0.00<br />
0 0.5 1 1.5 2 2.5 3 3.5 4 4.5 5
Other examples of uninformed priors<br />
A useful technique in these circumstances to minimise the<br />
effects of re-parameterisation is<br />
to set up the prior distribution for λ as 1/λ and for β as 1/β,<br />
that is we are using β as a prior for λ and vice versa.<br />
As a result, the prior distribution is transformation<br />
invariant.
Other examples of uninformed priors<br />
While such a distribution still does not appear to be<br />
uninformed, it is the best that can be achieved in the<br />
circumstance<br />
and gives the same answer whether we undertake an<br />
analysis from the point of view of λ or β.<br />
probability<br />
0.35<br />
0.30<br />
0.25<br />
0.20<br />
0.15<br />
0.10<br />
0.05<br />
0.00<br />
prior distribution for λ<br />
1<br />
prior( λ)<br />
∝<br />
λ<br />
0 0.5 1 1.5 2 2.5 3 3.5 4 4.5 5<br />
lambda
Informed priors<br />
An informed prior may be based on actual data or be<br />
purely subjective<br />
A conjugate prior has the same functional form as the<br />
likelihood function and leads to a posterior<br />
distribution belonging to the same distribution as the<br />
prior
Conjugate priors and their associated likelihoods<br />
and posterior distributions
Step 2. Likelihood functions<br />
The likelihood function calculates the probability of<br />
observing the data for a given value of the prior<br />
estimate of a parameter<br />
The shape of the likelihood function embodies the<br />
amount of information contained in the data<br />
If the information is limited, the likelihood function<br />
will be broadly distributed, whereas if the information<br />
is significant, the likelihood function will be tightly<br />
focused around a particular parameter value
Likelihood functions<br />
There are a number of useful probability distribution<br />
functions that can be used as likelihood functions,<br />
depending on the circumstances<br />
These include the Binomial, Poisson, hypergeometric and<br />
These include the Binomial, Poisson, hypergeometric and<br />
negative binomial
Step 3. Posterior distributions<br />
The posterior distribution is the revised estimate<br />
of the parameter we are investigating and is<br />
obtained simply by multiplying the prior<br />
distribution and the likelihood function<br />
Since the individual probabilities calculated by the<br />
likelihood function are independent of each other,<br />
the resulting posterior probabilities need to be<br />
normalised<br />
This ensures that the area under the curve of a<br />
continuous distribution equals one and that the<br />
probabilities for a discrete distribution all add up<br />
to one
What form does <strong>Bayes</strong>’ theorem assume<br />
when dealing with distributions<br />
where:<br />
pr(H i |D) = Posterior probability of the hypothesis after the analysis of<br />
data. We have a distribution of posterior probabilities (one for each of the i<br />
values of the distribution)<br />
pr(H i ) = Prior probability of the hypothesis (one for each of the i values of<br />
the distribution)<br />
Σ j pr(H j )pr(D|H j ) = Sum of all probabilities of the data given all the<br />
hypotheses. (*)
What form does <strong>Bayes</strong>’ theorem assume<br />
when dealing with distributions<br />
Note:<br />
Σ j pr(H j )pr(D|H j ) = Sum of all probabilities of the data given all the<br />
hypotheses. It is very difficult to know all probabilities of data given all the<br />
hypotheses.<br />
Nevertheless, since this expression is the same denominator for each<br />
pr(H i ǀD):<br />
pr(H i ǀD)∝pr(H i )pr(DǀH i )<br />
and Σ j pr(H j )pr(DǀH j ) becomes a normalization constant to make the sum of<br />
all numerators = 1
How to apply bayesian inference<br />
Let’s imagine that we test a flock of 50 sheep with a<br />
test with 93% sensitivity and 95% specificity<br />
We get no positive results<br />
What is the probability distribution of the number of<br />
What is the probability distribution of the number of<br />
(undetected) infected animals in the flock?
How to apply bayesian inference<br />
Tested= 50<br />
Positive = 0<br />
Sensitivity of the test= 93%<br />
Specificity of the test= 95%<br />
=Uniform(0,50)<br />
→Pr(H i )=1/51<br />
Infected animals<br />
in the sample Prior Binomial likelihood Posterior Nurmalized posterior<br />
0 0,019608 1 0,019608 0,93<br />
1 0,019608 0,07 0,001373 0,0651<br />
2 0,019608 0,0049 9,61E-05 0,004557<br />
3 0,019608 0,000343 6,73E-06 0,00031899<br />
49 0,019608 2,56924E-57 5,04E-59 2,38939E-57<br />
50 0,019608 1,79847E-58 3,53E-60 1,67257E-58
How to apply bayesian inference<br />
Tested= 50<br />
Positive = 0<br />
Sensitivity of the test= 93%<br />
Specificity of the test= 95%<br />
=Binomial(0,n,93%,false)<br />
Infected animals<br />
in the sample Prior Binomial likelihood Posterior Nurmalized posterior<br />
0 0,019608 1 0,019608 0,93<br />
1 0,019608 0,07 0,001373 0,0651<br />
2 0,019608 0,0049 9,61E-05 0,004557<br />
3 0,019608 0,000343 6,73E-06 0,00031899<br />
49 0,019608 2,56924E-57 5,04E-59 2,38939E-57<br />
50 0,019608 1,79847E-58 3,53E-60 1,67257E-58
How to apply bayesian inference<br />
Tested= 50<br />
Positive = 0<br />
Posterior=Prior x Likelihood<br />
Sensitivity of the test= 93%<br />
Specificity of the test= 95%<br />
Infected animals<br />
in the sample Prior Binomial likelihood Posterior Nurmalized posterior<br />
0 0,019608 1 0,019608 0,93<br />
1 0,019608 0,07 0,001373 0,0651<br />
2 0,019608 0,0049 9,61E-05 0,004557<br />
3 0,019608 0,000343 6,73E-06 0,00031899<br />
49 0,019608 2,56924E-57 5,04E-59 2,38939E-57<br />
50 0,019608 1,79847E-58 3,53E-60 1,67257E-58
How to apply bayesian inference<br />
Tested= 50<br />
Positive = 0<br />
Sensitivity of the test= 93%<br />
Specificity of the test= 95%<br />
Normalized=Posterior/Sum<br />
Infected animals<br />
in the sample Prior Binomial likelihood Posterior Nurmalized posterior<br />
0 0,019608 1 0,019608 0,93<br />
1 0,019608 0,07 0,001373 0,0651<br />
2 0,019608 0,0049 9,61E-05 0,004557<br />
3 0,019608 0,000343 6,73E-06 0,00031899<br />
49 0,019608 2,56924E-57 5,04E-59 2,38939E-57<br />
50 0,019608 1,79847E-58 3,53E-60 1,67257E-58
How to apply bayesian inference<br />
Tested= 50<br />
Positive = 0<br />
Sensitivity of the test= 93%<br />
Specificity of the test= 95%<br />
Normalized=Posterior/Sum<br />
Infected animals<br />
in the sample Prior Binomial likelihood Posterior Nurmalized posterior<br />
0 0,019608 1 0,019608 0,93<br />
1 0,019608 0,07 0,001373 0,0651<br />
2 0,019608 0,0049 9,61E-05 0,004557<br />
3 0,019608 0,000343 6,73E-06 0,00031899<br />
49 0,019608 2,56924E-57 5,04E-59 2,38939E-57<br />
50 0,019608 1,79847E-58 3,53E-60 1,67257E-58
The result is:<br />
<strong>Bayes</strong>ian inference<br />
1<br />
0,9<br />
0,8<br />
0,7<br />
Probability<br />
0,6<br />
0,5<br />
0,4<br />
0,3<br />
0,2<br />
0,1<br />
0<br />
0 1 2 3 4 5<br />
Infected animals in the sample
If we have<br />
positives?
Second order bayesian inference<br />
Let’s imagine to test the same flock and to get three<br />
positives<br />
What is the probability of having 0, 1, 2, 3 true<br />
positives and the remaining positives as false positives?<br />
What is the probability of having 0, 1, 2, 3, 4, ...<br />
infected animals in the flock?
Second order bayesian inference<br />
Tested= 50<br />
Positive = 3<br />
Sensitivity of the test= 93%<br />
Specificity of the test= 95%<br />
Uniform(0,50)<br />
Uniform(0,3)<br />
Binomial likelihood<br />
Posterior<br />
True positives<br />
True positives<br />
Infected animals 0 1 2 3 0 1 2 3<br />
in the sample Prior 0,25 0,25 0,25 0,25 0,25 0,25 0,25 0,25 Normalized posterior<br />
0 0,019608 0,219875 0 0 0 0,001078 0 0 0 0,258188132<br />
1 0,019608 0,015229 0,24538 0 0 7,47E-05 0,001203 0 0 0,306020881<br />
2 0,019608 0,001053 0,034685 0,186289 0 5,16E-06 0,00017 0,000913 0 0,260715921<br />
3 0,019608 7,28E-05 0,003674 0,040322 0,072187 3,57E-07 1,8E-05 0,000198 0,000354 0,136512804<br />
4 0,019608 5,02E-06 0,000346 0,005816 0,021276 2,46E-08 1,69E-06 2,85E-05 0,000104 0,032224319<br />
49 0,019608 0 0 2,67E-53 1,05E-49 0 0 1,31E-55 5,17E-52 1,23862E-49<br />
50 0,019608 0 0 0 8,27E-51 0 0 0 4,05E-53 9,70674E-51
Second order bayesian inference<br />
Binomial(TP,Inf,Se,False) x<br />
Tested= 50<br />
Positive = 3<br />
Sensitivity of the test= 93%<br />
Specificity of the test= 95%<br />
Binomial(Tested-Inf-(Pos-TP),Tested-Inf,Sp,False)<br />
Binomial likelihood<br />
Posterior<br />
True positives<br />
True positives<br />
Infected animals 0 1 2 3 0 1 2 3<br />
in the sample Prior 0,25 0,25 0,25 0,25 0,25 0,25 0,25 0,25 Normalized posterior<br />
0 0,019608 0,219875 0 0 0 0,001078 0 0 0 0,258188132<br />
1 0,019608 0,015229 0,24538 0 0 7,47E-05 0,001203 0 0 0,306020881<br />
2 0,019608 0,001053 0,034685 0,186289 0 5,16E-06 0,00017 0,000913 0 0,260715921<br />
3 0,019608 7,28E-05 0,003674 0,040322 0,072187 3,57E-07 1,8E-05 0,000198 0,000354 0,136512804<br />
4 0,019608 5,02E-06 0,000346 0,005816 0,021276 2,46E-08 1,69E-06 2,85E-05 0,000104 0,032224319<br />
49 0,019608 0 0 2,67E-53 1,05E-49 0 0 1,31E-55 5,17E-52 1,23862E-49<br />
50 0,019608 0 0 0 8,27E-51 0 0 0 4,05E-53 9,70674E-51
Second order bayesian inference<br />
Tested= 50<br />
Positive = 3<br />
Sensitivity of the test= 93%<br />
Specificity of the test= 95%<br />
Prior 1 x Prior 2 x Likelihood<br />
Binomial likelihood<br />
Posterior<br />
True positives<br />
True positives<br />
Infected animals 0 1 2 3 0 1 2 3<br />
in the sample Prior 0,25 0,25 0,25 0,25 0,25 0,25 0,25 0,25 Normalized posterior<br />
0 0,019608 0,219875 0 0 0 0,001078 0 0 0 0,258188132<br />
1 0,019608 0,015229 0,24538 0 0 7,47E-05 0,001203 0 0 0,306020881<br />
2 0,019608 0,001053 0,034685 0,186289 0 5,16E-06 0,00017 0,000913 0 0,260715921<br />
3 0,019608 7,28E-05 0,003674 0,040322 0,072187 3,57E-07 1,8E-05 0,000198 0,000354 0,136512804<br />
4 0,019608 5,02E-06 0,000346 0,005816 0,021276 2,46E-08 1,69E-06 2,85E-05 0,000104 0,032224319<br />
49 0,019608 0 0 2,67E-53 1,05E-49 0 0 1,31E-55 5,17E-52 1,23862E-49<br />
50 0,019608 0 0 0 8,27E-51 0 0 0 4,05E-53 9,70674E-51
Second order bayesian inference<br />
Tested= 50<br />
Positive = 3<br />
Sensitivity of the test= 93%<br />
Sum(Posterior) / Sum<br />
Specificity of the test= 95%<br />
Binomial likelihood<br />
Posterior<br />
True positives<br />
True positives<br />
Infected animals 0 1 2 3 0 1 2 3<br />
in the sample Prior 0,25 0,25 0,25 0,25 0,25 0,25 0,25 0,25 Normalized posterior<br />
0 0,019608 0,219875 0 0 0 0,001078 0 0 0 0,258188132<br />
1 0,019608 0,015229 0,24538 0 0 7,47E-05 0,001203 0 0 0,306020881<br />
2 0,019608 0,001053 0,034685 0,186289 0 5,16E-06 0,00017 0,000913 0 0,260715921<br />
3 0,019608 7,28E-05 0,003674 0,040322 0,072187 3,57E-07 1,8E-05 0,000198 0,000354 0,136512804<br />
4 0,019608 5,02E-06 0,000346 0,005816 0,021276 2,46E-08 1,69E-06 2,85E-05 0,000104 0,032224319<br />
49 0,019608 0 0 2,67E-53 1,05E-49 0 0 1,31E-55 5,17E-52 1,23862E-49<br />
50 0,019608 0 0 0 8,27E-51 0 0 0 4,05E-53 9,70674E-51
Second order bayesian inference<br />
Tested= 50<br />
Positive = 3<br />
Sensitivity of the test= 93%<br />
Sum(Posterior) / Sum<br />
Specificity of the test= 95%<br />
Binomial likelihood<br />
Posterior<br />
True positives<br />
True positives<br />
Infected animals 0 1 2 3 0 1 2 3<br />
in the sample Prior 0,25 0,25 0,25 0,25 0,25 0,25 0,25 0,25 Normalized posterior<br />
0 0,019608 0,219875 0 0 0 0,001078 0 0 0 0,258188132<br />
1 0,019608 0,015229 0,24538 0 0 7,47E-05 0,001203 0 0 0,306020881<br />
2 0,019608 0,001053 0,034685 0,186289 0 5,16E-06 0,00017 0,000913 0 0,260715921<br />
3 0,019608 7,28E-05 0,003674 0,040322 0,072187 3,57E-07 1,8E-05 0,000198 0,000354 0,136512804<br />
4 0,019608 5,02E-06 0,000346 0,005816 0,021276 2,46E-08 1,69E-06 2,85E-05 0,000104 0,032224319<br />
49 0,019608 0 0 2,67E-53 1,05E-49 0 0 1,31E-55 5,17E-52 1,23862E-49<br />
50 0,019608 0 0 0 8,27E-51 0 0 0 4,05E-53 9,70674E-51
And the result is:<br />
Second order bayesian inference<br />
0,35<br />
0,3<br />
0,25<br />
Probability<br />
0,2<br />
0,15<br />
0,1<br />
0,05<br />
0<br />
0 1 2 3 4 5 6 7 8 9 10<br />
Number of infected animals