2. summary
! binomial proportion
" the likelihood function
! the Bernoulli likelihood function
" the prior/posterior distribution
! beta distribution
" estimation ---- uncertainty <-HDI
" prediction ---- p(y)<-(z+a)/(N+a+z)
" comparison ---- best model <-p(D|M)
13. belief in detail (prior)
! beta distribution : beta(θ|a,b)
" two parameters
! mean :
! standard deviation :
14. belief in detail (prior)
! beta distribution : beta(θ|a,b)
" guess the values of a and b
! from (observed) data
" ex) a=b=1, a=b=4, etc…
! m = a/(a+b), n=(a+b)
a = mn, b = (1-m)n
! from mean and standard deviation
a 1 & b 1 a<1 &/or b<1
17. belief in
detail(posterior)
! supposition : N flips, z heads
! prior distribution : beta(θ|a,b)
! posterior distribution :
beta(θ|z+a, N-z+b)
one of the beauties of
mathematical approach to
Bayesian inference!
18. belief in detail
(updated parameters)
! probability distribution :
" prior : beta(θ|a,b)
N flips, z heads
" posterior : beta(θ|z+a, N-z+b)
! mean :
" prior : a/(a+b)
" posterior : (z+a)/[(z+a)+(N-z+b)]
= (z+a)/(N+a+b) !!!
19. belief in detail
(updated parameters)
! probability distribution :
" prior : beta(θ|a,b)
N flips, z heads
" posterior : beta(θ|z+a, N-z+b)
! mean :
" prior : a/(a+b)
" posterior : (z+a)/[(z+a)+(N-z+b)]
= (z+a)/(N+a+b) !!!
20. belief in detail
(updated parameters)
! probability distribution :
" prior : beta(θ|a,b)
N flips, z heads
" posterior : beta(θ|z+a, N-z+b)
! mean :
" prior : a/(a+b)
" posterior : (z+a)/[(z+a)+(N-z+b)]
= (z+a)/(N+a+b) !!!
0 z/N a/(a+b)
1-α α
α = N/(N+a+b)
22. estimation
! uncertainty of the prior distribution
" From the posterior distribution
! HDI : the highest density interval (chapter 3)
HDI
L : broad
R : narrow
prior dist.
L : more uncertain
23. estimation
! reasonable credibility of a value concerned
" From the posterior distribution
! ROPE : region of practical equivalence
coin flipping
θ = 0.5
credible?
ROPE = [0.48,0.52]
if
95% HDI ∩ ROPE =
then
θ is incredible
24. estimation
! reasonable credibility of a value concerned
" From the posterior distribution
! ROPE : region of practical equivalence
coin flipping
θ = 0.5
credible?
ROPE = [0.48,0.52]
if
95% HDI ∩ ROPE =
then
θ is incredible
includes many extra
assumptions!
29. comparison
to compare the models,
p(θ|D,M) = p(D|θ,M)p(θ|M)/p(D|M)
posterior likelihood prior evidence
prior
probability
observation
posterior
probability
31. comparison
uniform strongly peaked uniform strongly peaked
N = 14, z = 11 N = 14, z = 7
p(D|M)=0.000183>p(D|M)=6.86×10-5 p(D|M)=1.94×10-5<p(D|M)=5.9×10-5
32. comparison
! both are important
" the prior distribution
" the likelihood function
" in detail, see chapter 4
The best model (so far)
is not a good model.
33. summary
! binomial proportion
" the likelihood function
! the Bernoulli likelihood function
" the prior/posterior distribution
! beta distribution
" estimation ---- uncertainty <-HDI
" prediction ---- p(y)<-(z+a)/(N+a+b)
" comparison ---- best model <-p(D|M)