Transcript Model Evaluation and Model Selection

Model Evaluation
&
Model Selection
Modeling process
Problem identification / data collection
Identify scientific
objectives
Collect &
understand data
Draw upon existing theory/
knowledge
Model specification
Visualize model (DAG)
Write model mathematically using probability
notation & appropriate distributions
or
Write down (unnormalized) posterior
Derive full-conditional distributions
Program model
components in software
Construct MCMC algorithm
Model implementation
Model evaluation &
inference
Fit model (using MCMC & data)
Evaluate models (posterior predictive checks)
Use output to make inferences
Model selection
Motivating issues
• How well does my model(s) fit my data? [evaluation]
– Just because the MCMC procedure “went smoothly,” doesn’t mean
you have a “good model”
– Just because you got posterior stats for your parameters of interest
doesn’t mean you have a “good model”
– Check ability of model to “replicate” observed data
– Potentially check ability of model to “predict” observed data (via
cross-validation)
• Alternative model formulations = alternative
hypotheses about system. Which model? [selection]
–
–
–
–
Which model agrees the best with my data?
Which model is simpler to interpret?
Which model satisfies both criteria? Which model should I chose?
Combine alternative models? Model averaging
3
Lecture content
• Posterior predictive checks:
– Observed vs “predicted”
– Replicated data
– Bayesian p-values
• Model selection/comparison:
– Deviance
– Akaike Information Criterion (AIC) in the likelihood
framework
– Deviance information criteria (DIC)
– Posterior predictive loss (D)
The first question we should ask after fitting a model: Are the
predictions of the model consistent with the data?
1. Is our process model a reasonable representation?
2. Have we made the right choices of distributions to represent the
uncertainties?
Evaluate model fits
• Model evaluation and diagnostics are relatively under-developed in Bayesian
analysis
• We often rely on relatively qualitative and informal methods that are based on
ideas developed in “classical” analyses
• But, one particularly useful method for evaluating model fit is to compare
posterior predictions of “replicated” data with observed data.
• Examine the ability of the model to produce “replicated” data that are consistent
with the observed data
• Assume that “replicated” data (yrep) arise from same sampling distribution used
to define the likelihood of the observed data (y):
yi ~ P  y |  
yrepi ~ P  y |  
• Again, y is observed and yrep is not observed; thus, we obtain the posterior
predictive distribution for each yrep.
• Compare the posterior predictive distribution for each yrep to the corresponding
6
observed y value.
Posterior predictive checks
P( y new | y )   P( y new |  ) P( | y )d
Posterior predictive distribution
• It is called posterior because it is conditional on the
observed y and predictive because it is a prediction for an
observable ynew.
• It gives the probability of a new prediction of y
conditional on θ , which in turn is conditioned on the data
at hand, y.
The mechanics
• We have a scientific model g(θ,x) that predicts a response y. We
estimate the posterior distribution, P(θ|y). For any given value of
x, we can simulate the posterior predictive distribution ynew by
making a draw of θ= θ’ from P(θ|y) and estimating
ynew ~P(g(θ’|x),σ).
• In MCMC, this simply means making draws from the data
model because each draw is conditional on the current value of
the parameters. These draws define the posterior predictive
distribution in exactly the same way that draws allow us to
define the posterior probability of the parameters.
DAG: Back to hemlock trees
yijis the observed growth.
Data model
yi
μi is the true value of growth.
It is “latent”, i.e.
not observable.
 proc
Process model
i
Parameter model

g (bo , b1 , diam)  bo  b1diami
n
P(bo , b1 , | y, diam)   normal ( yi |g (bo , b1 , diami ), ) x
i 1
normal (bo | 0,.0001)normal (b1 | 0,.0001) gamma( | .001,.001)
model{
for(i in 1:length(y)){
mu[i] <- b0 + b1*diam[i]
y[i] ~ dnorm(mu[i],tau)
#posterior predictive distribution of y.new[i], unobserved trees
y.sim[i] ~ dnorm(mu[i],tau)
}
# Priors
b0 ~ dnorm(0,.0001)
b1 ~ dnorm(0,.0001)
tau ~ dgamma(.0001,.0001)
sigma<-1/sqrt(tau)
}
Marginal posterior distributions
Recall that by Monte Carlo Integration….
1
E ( | y ) 
K
and
K
k 


k 1
K
var( | y ) 
k 
2
(


E
(

|
y
))

k 1
K
Derived quantities
• Equivariance: Any quantity calculated from a random variable
becomes a random variable with its own probability distribution.
• These quantities may be of scientific interest in themselves (e.g.,
biomass using allometric equations, Shannon Diversity Index,
effect sizes…and so on).
• The derived quantity may involve model parameters, latent
processes, or data.
• Equivalence is also incredibly useful in calculating goodness of
fit of the model against observed data and in making forecasts
about yet-unobserved quantities.
Bayesian p-values
• Let T(y,) be a test statistic (e.g., mean, standard deviation, CV, quantile,
sums of squared discrepancy, etc.) associated with the observed data
• Likewise, let T(yrep,) be the corresponding test statistics associated with
the replicated data
• We can calculate the “tail” probability p:
p  P T  yrep |    T ( y |  ) y 
or
p  P T  yrep |    T ( y |  ) y 
• If p is very large (e.g., p >>0.5 or close to 1) or very small (i.e., p <<0.5
or close to 0), then the difference between the observed and simulated
data cannot be attributed to chance, indicating potential lack of fit.
R.A. Fischer’s ticks
A simple example: We want to know the average
number of ticks on sheep. We round up 60 sheep and
count ticks on each one. Does a Poisson distribution fit
the data?
60
P( | y )   P( yi |  )P( )
i 1
For each value in the MCMC chain, we generate a new
data set, ysim, by sampling from:
P( y
sim
|  ) P ( | y )
yi

A single mean
governs
the pattern
Data
Parameter
model{
#prior
lambda ~ dgamma(0.001,0.001)
Key part
for(i in 1:60){
y[i] ~ dpois(lambda)
y.sim[i] ~ dpois(lambda) #simulate a new data set of 60 points
}
cv.y <- sd(y[ ])/mean(y[ ])
cv.y.sim <- sd(y.sim[])/mean(y.sim[ ])
mean.y <-mean(y[])
mean.y.sim <-mean(y.sim[])
# find Bayesian P value--the mean of many 0's and 1's returned by the step function, one
for each step in the chain
pvalue.cv <- step(cv.y.sim-cv.y)
Step function=1 if ()>0
pvalue.mean <-step(mean.y.sim - mean.y)
# Sums of Squares
for(j in 1:60){
sq[j] <- (y[j]-lambda)^2
sq.new[j] <- (y.sim[j]-lambda)^2
}
fit <- sum(sq[])
fit.new <- sum(sq.new[])
pvalue.fit <- step(fit.new-fit)
} #end of model
0.15
0.00
Density
Real Data
0
2
4
6
8
10
Number of Ticks
Simple
Model
0.10 0.20
0.00
Density
Simulated Data
0
2
4
6
Number of Ticks
8
10
Posterior predictive check
Simple Model
fit <- sum(sq[])
fit.new <- sum(sq.new[])
pvalue.fit <- step(fit.new-fit)
} #end of model
p-value for CV=0.0015
P-value for mean=0.51
Remember, this is a twotailed probability, so
values close to 0 and 1
indicate lack of fit.
0
2
4
6
Real data
8
10
0
2
4
6
Simulated data
8
10

yi
Data
i
Parameter

Each sheep has its own mean
(a.k.a. random effect)
Hyperparameter
model{
# Priors
a~ dgamma(.001,.001)
b~ dgamma(.001,.001)
60
for(i in 1:60){
P(a, b,  | y ) 
P ( yi | i )P (i | a, b)
lambda[i] ~ dgamma(a,b)
i 1
y[i] ~ dpois(lambda[i])
y.sim[i] ~ dpois(lambda[i])
P(a) P (b)
}
cv.y <- sd(y[ ])/mean(y[ ])
cv.y.sim <- sd(y.sim[])/mean(y.sim[ ])
pvalue.cv <- step(cv.y.sim-cv.y) # find Bayesian P value--the mean of many 0's and
1's returned by the step function, one for each step in the chain
mean.y <-mean(y[])
mean.y.sim <-mean(y.sim[])
pvalue.mean <-step(mean.y.sim - mean.y)
Hierarchical Model

for(j in 1:60){
sq[j] <- (y[j]-lambda[j])^2
sq.new[j] <- (y.sim[j]-lambda[j])^2}
fit <- sum(sq[])
fit.new <- sum(sq.new[])
pvalue.fit <- step(fit.new-fit)
} #end of model
Hierarchical
Model
Posterior predictive check
Hierarchical Model
fit <- sum(sq[])
fit.new <- sum(sq.new[])
pvalue.fit <- step(fit.new-fit)
} #end of model
p-value for CV=0.46
p-value for mean=0.51
Remember, this is a twotailed probability, so
values close to 0 and 1
indicate lack of fit.
0
2
4
6
Real data
8
10
0
Simulated data
2
4
6
8
10
Posterior predictive checks
• Gelman, A., and J. Hill. 2009. Data analysis using regression
and multilevel / hierarchical models. Cambridge University
Press, Cambridge, UK.
• Link, W. A., and R. J. Barker. 2010. Bayesian Inference with
Ecological Applications. Academic Press.
• Kery, M. 2010. Introduction to WinBUGS for Ecologists: A
Bayesian approach to regression, ANOVA, mixed models and
related analyses. Academic Press.
Motivating issues
• How well does my model(s) fit my data? [evaluation]
– Just because the MCMC procedure “went smoothly,” doesn’t mean
you have a “good model”
– Just because you got posterior stats for your parameters of interest
doesn’t mean you have a “good model”
– Check ability of model to “replicate” observed data
– Potentially check ability of model to “predict” observed data (via
cross-validation)
• Alternative model formulations = alternative
hypotheses about system. Which model? [selection]
–
–
–
–
Which model agrees the best with my data?
Which model is simpler to interpret?
Which model satisfies both criteria? Which model should I chose?
Combine alternative models? Model averaging
26
Lecture content
• Posterior predictive checks:
– Observed vs “predicted”
– Replicated data
– Bayesian p-values
• Model selection/comparison:
– Deviance
– Akaike Information Criterion (AIC) in the likelihood
framework
– Deviance information criteria (DIC)
– Posterior predictive loss (D)
27
“Model selection and model averaging are deep
waters, mathematically, and no consensus has
emerged in the substantial literature on a single
approach. Indeed, our only criticism of the wide use
of AIC weights in wildlife and ecological statistics is
with their uncritical acceptance and the view that
this challenging problem has been simply resolved”
Link, W. A., and R. J. Barker. 2006. Model weights and the
foundations of multi-model inference. Ecology 87:2626
The problem of model selection
• Up until now, we have been concerned with the
uncertainties associated with a given model.
• What about the uncertainty that arises from our
choice of models?
• How do we decide which model is best?
• How do we make inferences based on multiple
models?
Parsimony==Ockham’s razor
William of Ockham (1285-1349)
“Pluralitas non est ponenda sine neccesitate”
“entities should not be multiplied unnecessarily”
“Parsimony: ... 2 : economy in the use of means to an end;
especially : economy of explanation in conformity with Occam's
razor”
(Merriam-Webster Online Dictionary)
Information theory and the principle of
parsimony
True model:
ye
( x 0.3) 2 1
Generated ten datasets sampling from normal distribution with
mean=0 and var=0.01.
Fit five models to the these ten datasets.
y   0  1 x
y   0  1 x   2 x 2
y   0  1 x   2 x 2   3 x 3
y   0  1 x   2 x   3 x   4 x
2
3
4
y   0  1 x    2 x 2   3 x 3   4 x 4   5 x 5
What creates noise in models?
Illustration of tradeoff
The Kullback-Leibler
distance
Interpretation of Kullblack-Leibler Information
(aka. distance between 2 models)
• Given truth represented by f and a model
approximating truth g, the K-distance measures the
information lost by using model g to approximate f.
Interpretation of Kullblack-Leibler Information
(aka. distance between 2 models)
650
Count
520
f(x)
Truth
390
260
650
130
520
0
0
5
10
GAMMA
15
20
g2(x)
390
260
130
650
0
0
520
390
10
WEIBULL
15
20
Approximations to truth
g1(x)
260
5
130
0
0
5
10
15
LOGNORMAL
20
Measures the (asymmetric) distance between two models. Minimizing the information
lost when using g(x) to approximate f(x) is the same as maximizing the likelihood.
Heuristic interpretation of K-L
Model comparison
• Within the classical modeling framework, we tradeoff
a measure of complexity (typically deviance) for a
measure of complexity (typically number of
parameters).
How do we know the truth?
Akaike’s Information Criterion
Akaike defined “an information criterion” that related
K-L distance and the maximized log-likelihood as follows:
^
AIC  2 ln( L(  | y ))  2 K
This is an estimate of the expected, relative distance between
the fitted model and the unknown true mechanism that
generated the observed data.
K=number of estimable parameters
Deviance
D( y,  )  2  log  P( y| )
• Deviance (deviance) is a built-in node in JAGS,
• thus, you can monitor deviance,
• look at its history plots (helpful to evaluating overall
model convergence and potential “problem” chains)
• compute posterior statistics
44
• Use the difference in AIC to compare
competing models.
 r  AICr  min( AIC )
min( AIC )  min( AIC1 ,..... AICn )
 r  AICr  min( AIC )
As a rule of thumb models having Δr ≤ 2 have sufficient support—
they should receive consideration in making inferences. Models having
Δr within about 3-7 have considerable less support, while models with
Δr ≥10 have essentially no support .
But there is a better way….
L( | y)  e
1
( r )
2
• The likelihood of a model given the data
decreases exponentially with increasing Δr . Note
that the likelihood of the best model = 1 and all
other likelihoods are relative to the likelihood of
the best model.
Likelihood ratio from AIC
e
e
1
(  1 )
2
1
( 2 )
2
 relative strength of evidence in data
for model 1 over 2
Akaike Weights
wr 
e
R
1
( r )
2
e
1
(   r ,i )
2
likelihood model r | data

total likelihood of all models | data
i 1
wr are Akaike weights, the likelihood of one of the candidate
models divided by the sum of the likelihoods of all of the
candidates. The wr for the best model does not equal 1. The wr sum
to 1.
The wr can be thought of as “probabilities.” This is a frequentist
interpretation derived from simulation. They are not “true”
probabilities. (Link, W. A., and R. J. Barker. 2006)
Interpretation of Akaike Weights
• wi is the weight of evidence in favor of model i being
the actual best K-L model given that one of the R
models must be the K-L best model of the candidate
set.
• “probability” that model i is the actual best K-L
model
• Last statement is quite controversial.
The raptors…moving towards model selection
Summary of problem and data: In most northern temperate regions, diurnal birds of
prey (raptors) migrate seasonally between their breeding and wintering grounds. Most
raptors are 0.
obligate
facultative
soaring migrants
that congregate
along major thermal
Identifyorscientific
problem/objectives;
understand
data; draw-up
and orographicexisting
updraft
corridors.
theory/knowledge.
We might wish to analyze the raptor survey data to understand how temperature
and wind speed affect the chance of observing birds of each species.
Data: Autumn migration counts of multiple species of raptors in NE U.S., conducted
during 2010.
yd,s = number of birds observed on day d for species s
95 days, 15 species
xd = total time of observation period (minutes)
Td = average air temperature (C) during observation period on day d
WSd = average wind speed (km/hr) during observation period on day d
Therrien et al. 2012. Ecology.
Visualize model via DAG
T
WS
x
Day
y

Day, species

Species

σ
σ
σ
Month
Population
Data (stochastic)
Latent process
Data parameters
Process parameters
Specify model
yd ,s ~ Poisson  d ,s  xd 
Likelihood:
Log link function (for log-linear Poisson regression):
T
WS
d ,s  log(d ,s )
x

d ,s  exp(d ,s )
Stochastic model for linear predictor (account for over-dispersion)

y
d ,s ~ Normal 1s  2s  Td  3s Wd  4s  Td Wd   m(d ) ,  2

Hierarchical priors for species-level effects parameters:


k ,s ~ N ˆk , 2k

k  1, 2,3, 4 parameters
Zero-centered hierarchical prior for month random effect:

 m ~ N 0,  2


Conjugate, relatively non-informative priors for root nodes:

ˆk ~ N  0,10000 
 ,  k ,  ~ gamma(0.01, 0.01)
σ
σ
σ
where σ2 = 1/τ for each σ2 term
Is over-dispersion needed?
T
WS
Log link function (for log-linear Poisson regression):
x
d ,s  log(d ,s )
y

d ,s  exp(d ,s )
Stochastic model for linear predictor (without over-dispersion)
d ,s  1s   2 s  Td  3s Wd   4 s  Td  Wd   m ( d )

Hierarchical priors for species-level effects parameters:

k ,s ~ N ˆk , 2k


k  1, 2,3, 4 parameters
Zero-centered hierarchical prior for month random effect:

 m ~ N 0,  2


Conjugate, relatively non-informative priors for root nodes:
ˆk ~ N  0,10000 
σ
σ
σ
 ,  k ,  ~ gamma(0.01, 0.01)
where σ2 = 1/τ for each σ2 term
Implement (code) models: Model 1
BUGS code shown here.
BUGS code shown here.
Implement models: Model 2 (no overdispersion)
BUGS code shown here.
57
Evaluate results/make inferences
Model 1: includes over-dispersion, example (temperature effects at sp level):

d ,s ~ Normal 1s  2 s  Td  3s Wd  4s  Td Wd   m( d ) ,  2

th e ta .s ta r[2 ,]
0 .0
0 .5
Eff ect of temperature on observation rate
[2,10]
[2,13]
[2,3]
[2,1] [2,2]
[2,4]
[2,5] [2,6]
[2,14]
[2,15]
[2,16]
[2,7]
[2,8]
[2,9]
[2,11]
[2,12]
Population-level parameter
-0 .5
ˆ2
Species ID
Do the posterior stats for the over-dispersion standard deviation term indicate the
presences of “significant” over-dispersion?
sig (σ)
mean
1.68
sd
0.07601
val2.5pc
1.538
median
1.678
val97.5pc
1.84
Replicated data (Model 1)
12
10
Predicted # of birds
Posterior mean & 95% CI for yrep
Things to look for:
Species 8 (northern goshawk)
R2 = 0.965
Coverage = 100%
14
○ Bias/accuracy: Do the points fall
around the 1:1 line, or is there some
prediction bias?
8
○ Coverage: Do most of the observed
values fall inside the 95% CIs for the
replicated values (the Yrep’s)?
6
4
2
0
0
2
4
6
8
10
12
14
○ Is the variability in the observed data
consistent with the variability in the
replicated values?
Species 10 (broad-winged hawk)
R2 = 1.0
Coverage = 100%
800
600
○ What percentage do you expect to
fall outside of the Yrep CIs?
○ Can also overlay plots the observed
Y values and the predicted Y values
(i.e., posterior means for Yrep and
95% CI) as functions of a covariate.
400
200
0
0
200
400
600
Observed # of birds (y)
800
○ Why we get a “perfect” fit and 100%
coverage when including overdispersion here?
Replicated data (Model 2)
12
10
Predicted # of birds
Posterior mean & 95% CI for yrep
Things to look for:
Species 8 (northern goshawk)
R2 = 0.082
Coverage = 97.9%
14
○ Bias/accuracy: Do the points fall
around the 1:1 line, or is there some
prediction bias?
8
○ Coverage: Do most of the observed
values fall inside the 95% CIs for the
replicated values (the Yrep’s)?
6
4
2
0
0
2
800
4
6
8
10
12
14
○ Is the variability in the observed data
consistent with the variability in the
replicated values?
Species 10 (broad-winged hawk)
R2 = 0.259
Coverage = 52.6%
600
○ Can also overlay plots the observed Y
values and the predicted Y values (i.e.,
posterior means for Yrep and 95% CI)
as functions of a covariate.
400
200
0
0
200
400
600
Observed # of birds (y)
○ What percentage do you expect to fall
outside of the Yrep CIs?
800
○ Why is the fit so much worse when we
don’t include an over-dispersion term?
Point estimates of deviance
• Can get a point estimate of deviance by plugging in a point estimate
of the parameters (e.g., ’s)
D ˆ ( y )  D( y, ˆ )

= point estimate
(usually posterior mean)
• But, this doesn’t account for uncertainty in the parameters.
• Compute an “expected” deviance, which may be used as an overall
measure of model fit.
• Compute “expected” deviance by “averaging” over the posterior
distribution of the parameters:
Dave ( y )  E( D( y ,  )| y ))   D( y ,  )  P ( | y )d 
• If we have L draws from the posterior, then an estimate of Dave(y) is:
L
1
Dˆ ave ( y )   D( y ,  l )
L l 1
Posterior mean
of deviance
Model complexity
• To account for model complexity, compute the effective number
of parameters. To do this we compute deviance in two ways:
• The posterior mean of the deviance.
• The deviance evaluated at the posterior mean values of
model parameters.
pD  Dˆ ave ( y )  D ˆ ( y )

• Why should first component be larger than the second
component?
• In some situations, the above solution for pD can lead to pD < 0
(i.e., a negative # of effective parameters), which renders DIC
and pD useless.
Computing DIC
• Thus, DIC is given by:
DIC  Dˆ ave ( y )  pD
The deviance of the model evaluated
at the means of the posterior
distribution of parameters
The effective number of
parameters
model fit (lower better) + model penalty (lower better)
• Lower DIC -> “better” model, but what is an “important difference” in DIC?
• Interpretation of DIC values same as AIC
• DIC difference of 1-2: “best” model deserves consideration
• DIC differences of 3-7: “considerably” more support for best model
• Differences can be affected by Monte Carlo error
• I look for differences > 10
# JAGS model
model{
for (i in 1:n){
mu[i]<-(alpha*x[i])/((alpha/gama)+x[i])
y[i]~ dnorm(mu[i],tau)}
tau~dgamma(0.001,.001)
alpha~dgamma(0.001,.001)
gama~dgamma(.001,.001)
In R
} # end of model
#In R
jm=jags.model("Bugs_light_example.R",data=data,mod.inits,n.chains=3,n.adapt = n.adapt)
update(jm, n.iter=5000)
#generate coda object for parameters and deviance.
zm=coda.samples(jm,variable.names=c("alpha", "gama", "c","sigma","deviance"),n.iter=5000)
dic.samples(jm,n.iter, type="pD")
#Mean deviance: 529.8
#penalty 2.97 =pd
# Penalized deviance: 532.8 =dic
# Another way:
summary(zm[,"deviance"])$stat[,]
#pd* = (1/2)*Var(deviance)
pd= 0.5*summary(zm[,"deviance"])$stat[2]^2
Why DIC?
• DIC, AIC, and BIC all have the same general form:
xIC = model fit + model penalty
• So, why focus on DIC? Why not use AIC or BIC?
• BIC and AIC require us to count the number of parameters
in a model, but an informative prior, or hierarchical priors
makes it impossible to count the number of “effective”
parameters
• Thus, Spiegelhalter et al. (2002) developed DIC, which
computes an “effective” number of parameters that (should)
capture the effect of “shrinkage” or “borrowing of strength”
due to informative priors or hierarchical priors
Some intuition for DIC
• The problem is parameters that are “free” to be influenced
by noise in the data. How free are they?
• If a prior on a parameter is very informative—the parameter
is not free to respond to the data, it does not contribute to
the effective number of parameters.
• If a prior is uninformative, the opposite is the case. It is free
to respond and contributes to the effective number of
parameters in the same way as in a likelihood analysis.
• If a parameter is part of a hierarchy, should it count to same
way as a parameter that is part of a simpler model?
Posterior predictive loss
• Posterior predictive loss (i.e., Dk) is fairly widely used within statistics, but
is not frequently used in ecology
• Dk provides an index of a model’s predictive ability by comparing observed
data to replicated data
• The “best” model(s) is the one that performs the best under a “balanced loss
function.”
• Similar to the DIC, the loss function penalizes for both departure from the
observed data (measure of model “fit”) and departure from what we expect
the replicated data to be (measure of “smoothness” – somewhat analogous
to DIC’s effective number of parameters).
• The loss function puts weights, which depend on k > 0, on the model fit
component (G) and a weight of 1 (one) on the smoothness or penalty
component (P); the value of k is determined by the user:
k
Dk 
GP
k 1
Posterior predictive loss (D)
• We often assume “k = ,” in which case Dk (call this D) is
given by:
D  G  P
• Under squared-error loss, G and P are given by:
N
G    repi  y i 
2
i 1
N
2
P    rep
i
i 1
where
• Thus, D is equivalent to:
𝐷∞ = 𝐺 + 𝑃
where
repi  E( yrepi | y )
2
 rep
 Var ( yrepi | y )
i
Computing D in practice
• In your code, simply monitor the squared deviation, (yrepi - yi)2, for each
observation i, and outside of the i loop, compute the sum of the squared deviations:
Dsum.species[s]<-sum(sqdiff[,s]) # in the species loop
Dsum is NOT D; D is computed after you’ve run the model (after convergence)
and it is approximated as the posterior mean (expected value) of Dsum.
69
30
20
10
tree.data$Observed.growth.rate
40
Back to the tree and the light
20
40
60
tree.data$Light
80
100
Model 1:mu[i]<-(alpha*x[i])/((alpha/gama)+x[i])
Model 2: mu[i]<-alpha*x[i]
# JAGS model
model{
for (i in 1:n){
mu[i]<-(alpha*x[i])/((alpha/gama)+x[i])
y[i]~ dnorm(mu[i],tau)
yrep[i]~dnorm(mu[i],tau))
}
tau~dgamma(0.001,.001)
alpha~dgamma(0.001,.001)
gama~dgamma(.001,.001)} # end of model
In R
zm=coda.samples(jm,variable.names=c("alpha","gama","sigma","deviance",yrep"),n.iter=5000)
y<-tree.data$Observed.growth.rate # assign response variable
ntree=nrow(tree.data)
yrep.stats=summary(zm[,paste("yrep[",1:ntree,"]",sep="")])$stat
G <- sum((yrep.stats[,1]-y)^2)# sq diff (yrep.mean-y)
P <- sum((yrep.stats[,2])^2)# var for each yrep
Dinf <- G + P
Interpreting D values
• For a “poor” model, we expect large predictive variance (large P)
and poor fit (large G).
• Better models have a lower Dk (or lower D) associated with a
smaller P and/or smaller G
• But, as we start to “overfit” (e.g., model with lots of parameters),
G will continue to decrease (better fit), but P will start to increase
(i.e., variance will be inflated due to multi-collinearity between
parameters).
• The model with the smallest Dk (or D) is preferred.
• But, how small is “small”?
72
Linear vs non-linear
pd
Mean(dev)
P
G
Dinf
DIC
R2
Nonlinear
2.94
529.8 4551.13 4236.13 8787.27
532.7
0.91
Linear
2.05
574.2 8019.67 7612.01 15631.6
576.3
0.82
Consider multiple comparison indices
Average Deviance
DIC
pD
Countable # unknowns
D
G
P
2
R (obs vs pred)
Coverage
CI width of replicated data
Model 1
2069
2576
507
1500
9895
201
9693
1.00
100%
4.36
Model 2
11210
11270
61
75
682802
677141
5661
0.30
85%
4.44
Indices computed across
all species.
Some conclusions:
• Posterior predictive checks suggest we should “pick” Model 1 (with overdispersion) over Model 2 (Model 1 has much lower DIC, D, G)
• But, the results suggest that Model 1 is over-fitting the data – it has a very high
P, the coverage of the replicated data is too high (should be ~95%), and the
uncertainty in the replicated data (e.g., width of 95% CIs) is similar to Model 1.
• Perhaps we should evaluate a third model that lies somewhere between the
complexity of Model 1 and Model 2? E.g., explore incorporation of day (nested
in month) random effects without observation-level over-dispersion?
74
• In the index of Gelman et al. 1995: “Model
selection, why we do not do it”
• In the index of Gelman et al. 2004: “Model
selection, why we avoid it.”
– Gelman, A., J. B. Carlin, H. S. Stern, and D. B.
Rubin. 2004. Bayesian data analysis. Chapman and
Hall / CRC, London.
When should we avoid model selection?
• When we have a firm, theoretical / mechanistic basis
for a particular model formulation.
• When our objectives for insight determine the form of
the model.
• When we want to make forecasts and must include
known influences on future behavior of the system.
Concluding remarks
• Use multiple approaches for comparing / selecting between models
• Selection criteria may depend on scientific objectives:
• Use model to learn (heuristic tool)
• Use model to predict under novel conditions
• Other topics not covered
• Model averaging & Bayes factors
• E.g., can use posterior model weights to average derived or
predicted quantities obtained from each model
• Can use BIC to approximate the Bayes factors (BF)
• See Link & Barber (2006) or Kass (1993)
• Evaluation of model assumptions
• Appropriate choice of distributions?
• Appropriate model structure
• E.g.: linear vs non-linear; choice of covariates; random vs
fixed effects; hierarchical vs non-hierarchical priors, etc
77
References
• Posterior predictive loss: Gelfand & Ghosh (1998) Model choice: A
minimum posterior predictive loss approach. Biometrika 85:1-11.
• Bayes factors & BIC: Link & Barker (2006) Model weights and the
foundations of multimodel inference. Ecology 87:2626-2635.
• Model checking & improvement: Gelman, Carlin, Stern, Rubin (2004)
Bayesian Data Analysis. Chapman & Hall/CRC. (Chapter 6)
• Bayes factors: Kass (1993) Bayes factors in practice. The Statistician
42:551-560.
• Elements of hierarchical Bayes, including Bayes factors, DIC, D:
Carlin et al. (2006) Elements of hierarchical Bayesian inference, in
Hierarchical Modelling for the Environmental Sciences: Statistical
Methods and Applications, J.S. Clark & A.E. Gelfand (eds.) Oxford
Univ Press.
78
Quantiles vs. HPD intervals
HPDI
quantiles
Quantiles vs. HPD intervals
• HPDI: The longest horizontal line that can be placed
within the distribution such that the area between
the vertical dashed lines and beneath the
distribution curve = 1 –alpha.
• Equal tailed intervals: quantiles of distribution (1alpha/2).
Likelihood Ratio Test
• Ratios of llikelihoods (R) follow a chi-square
distribution with degrees of freedom equal to
the difference in the number of parameters
between models A and B.
R  2[ L(Y | M A )  L(Y | M B )]
the Likelihood Ratio Test
Chi-square probability
Difference in loglikelihood
Chi-sq=3.84
θ
Likelihood profile