Christensen

Background
Background on Rune H B Christensen

Psychometric and Statistical Models in the R packages
sensR and ordinal
Rune H B Christensen
February 9th 2012
Rune H B Christensen (DTU)
The sensR and ordinal packages

Outline
Education:
Engineer from DTU in 2008 Statistics and Data Analysis
Research interests:
Sensometrics
Likelihood methods
Mixed effects models
Computational statistics
Applied statistics (food science, biology, . . . )
R-packages:
sensR with Per Bruun Brockhoff
ordinal
binomTools with Merete Kjr Hansen
DTU Informatics, IMM

Section for Statistics
Technical University of Denmark
rhbc@imm.dtu.dk
PhD: Sensometrics: Thurstonian and Statistical Models

November 2008 April 2012
Psychoco 2012
1 / 64
Outline

The sensR package
The sensR package
The sensR package
The ordinal package overview
Implementation in the ordinal package
Assessment of estimation accuracy
Cumulative link mixed models (CLMMs)
2 / 64
Psychoco 2012
4 / 64
Outline
Psychoco 2012
Psychoco 2012
3 / 64
The sensR package
The sensR package
Replicated
Regression analysis
Psychometric protocols supported in sensR
The sensR package
X
X
X
X
X
X
X
X
X
X

The sensR package
Psychoco 2012
5 / 64
Sample size
Simulation
X
X
X
X
X
X
X
X

The sensR package
Likelihood CI
Power
X
X
X
(X)
X
(X)
X
X
X
X
X
X
Psychoco 2012
6 / 64
Psychoco 2012
8 / 64
plot
ROC
AUC
lla
ne
ou
rescale
psyfun
psyinv
pc2pd
pd2pc
isc
e
Illu
str
ati
on
Tr
an
sfo
r
d': sensory difference
findcr
clm2twoAC
SDT
samdiffSim
discrimSim
a1
a2
What is the probability of a correct answer?
Beyond the basics:

glm family objects for Thurstonian models:
twoAFC(), threeAFC(), duotrio(), triangle()
X
X
X
X
X
X
The Thurstonian model, 3-alternatives
ma
tio
n
Sa
mp
le
siz
e
Po
we
r&
d 0,
CI
,t
est
s
Basic functions in sensR
Similarity test
Examples for papers
discrimPwr
d.primePwr
discrimSS
d.primeSS
twoACpwr
Difference test
Statistical methodology for sensory discrimination tests and its

implementation in sensR
discrim
AnotA
samediff
twoAC
betabin
d 0 estimation
Vignettes:
X
X
X
X
X
X
isc
rim
in
at
Development on R-Forge:
https://r-forge.r-project.org/projects/sensr/
Duo-Trio, Triangle
2-AFC, 3-AFC
A-not A
Same-Different
2-AC
A-not A w. Sureness
io
n
On CRAN since July 2008:

www.cran.r-project.org/packages=sensR
Estimation and inference in Thurstonian models for sensory discrimination
It depends on the question 3-AFC or Triangle.

Psychoco 2012
7 / 64
The sensR package
The sensR package
Psychometric functions
Psychometric functions: Inverse link functions
1.0
f3-AFC (d 0 ) =
0.9
2AFC
3AFC
0.8
Duotrio
pc
d0 = X
0.5
0.4
d'
Psychoco 2012
(z d 0 )(z ) dz = (d 0 / 2)
psyphy: mafc(m=3): F () = m1 (1
only depends on no. alternatives.

(z d 0 )(z )2 dz
Family objects:
twoAFC(), threeAFC(), duotrio(), triangle()
Problem: d 0 0
0.3
ftriangle (d ) = 2
pc = fpsy (d 0 )
Triangle
n h
i
h
io
p
p
z 3 + d 0 2/3 + z 3 d 0 2/3 (z ) dz
0
fduo-trio (d 0 ) = 1 (d 0 / 2) (d 0 / 6) + 2(d 0 / 2)(d 0 / 6).

0
y binom(pc , n)
0.7
0.6
f2-AFC (d 0 ) =
A GLM:
9 / 64
1
m )()

Introduction
Psychoco 2012
10 / 64
The ordinal package
Outline
Regression models for ordinal data via cumulative link models

1
The sensR package
On CRAN since March 2010:

www.cran.r-project.org/packages=ordinal
Development on R-Forge:
https://r-forge.r-project.org/projects/ordinal/
Vignettes:
Analysis if ordinal data with cumulative link models (32 pages)
clm tutorial (18 pages)

clmm tutorial (9 pages)
Psychoco 2012
11 / 64
Psychoco 2012
12 / 64
Introduction
What is a cumulative link model (CLM)?
Ordinal data the wine data
Ordinal data: large, medium, small

Human assessments subjective judgements
Introduction
(preference, grades)
Objective:
How does perceived bitternes depend on temperature and contact?
Grouped continuous, e.g., age (15-24, 25-34, 35-50)

Table: The wine data (Randall, 1989), N=72
CLM:
ij = P (Yi j ) = F (j x T
i )
A regression model for an ordered variable

(Agresti, 2002; Greene and Hensher, 2010)
Intuitively:
A logistic regression model for J 2 (ordered) categories

Introduction
Psychoco 2012
13 / 64
Interpretation of the cumulative link model
temperature
contact
judges
predictor
predictor
random

Introduction
Latent bitterness follows a linear

model:
2
Si = + x T
i + i , i N (0, )
= + (tempi ) + i
warm
j 1 Si < j Y = j
cold
cold
1
Psychoco 2012
= + (tempi ) + i
We only observe a grouped

version of Si :
P(Y = 2|cold)
14 / 64
Latent bitterness follows a linear

model:
We only observe a grouped

version of Si :
Psychoco 2012
2
Si = + x T
i + i , i N (0, )
warm
Values
1, 2, 3 ,4, 5
less more
cold, warm
no, yes
1, . . . , 9
Interpretation of the cumulative link model

Y:
Type
response
Temperature and contact between juice and skins can be controlled when
cruching grapes during wine production.
A linear model that respects the ordered categorical nature of the

response
Variables
bitterness
15 / 64
P (Yi j ) = F (j x T
i )
Psychoco 2012
15 / 64
Introduction
A teaser Fitting cumulative link models with clm
Introduction
Likelihood ratio tests of CLMs
> data(wine)
> fm1 <- clm(rating ~ contact + temp, data=wine, link="probit")
> summary(fm1)
formula: rating ~ contact + temp
data:
wine
> fm2 <- update(fm1, ~.-temp)

> anova(fm1, fm2)
Likelihood ratio tests of cumulative link models:
link
threshold nobs logLik AIC
niter max.grad cond.H
probit flexible 72
-85.76 183.52 5(0) 1.53e-13 2.2e+01
formula:
link: threshold:
fm2 rating ~ contact
probit flexible
fm1 rating ~ contact + temp probit flexible
Coefficients:
Estimate Std. Error z value Pr(>|z|)
contactyes
0.8677
0.2669
3.251 0.00115 **
tempwarm
1.4994
0.2918
5.139 2.77e-07 ***
--Signif. codes: 0 '***' 0.001 '**' 0.01 '*' 0.05 '.' 0.1 ' ' 1
no.par
AIC
logLik LR.stat df Pr(>Chisq)
fm2
5 210.05 -100.026
fm1
6 183.52 -85.761 28.529 1 9.231e-08 ***
--Signif. codes: 0 '***' 0.001 '**' 0.01 '*' 0.05 '.' 0.1 ' ' 1
Threshold coefficients:
Estimate Std. Error z value
1|2 -0.7733
0.2829 -2.734
2|3
0.7360
0.2499
2.945
3|4
2.0447
0.3218
6.353
4|5
2.9413
0.3873
7.595
16 / 64

Overview
convergence
slice
drop.coef
Carefully designed printing
C:
Psychoco 2012
18 / 64
lla
ne
ou
clm
clmmC
clm.fit
clm.control
clmm.control
Convergence assessment
isc
e
M
Efficient computational methods
Fit
tin
g
Extensive model framework
Psychoco 2012
17 / 64
Functions (exported) in ordinal
What is unique about the implementation in ordinal?
clm2
clmm2C
clm2.control
clmm2.control
Di
str
ibu
tio
ns
Psychoco 2012
im
pl.

Overview
Fo
rm
er
[pdqrg]gumbelC
[pdg]lgammaC
gnormC
glogisC
gcauchyC
Implementations in C
Psychoco 2012
19 / 64
Overview
Methods for clm objects
Overview
An extended CLM Framework

Standard CLM:
Extractor and Print
Inference
Checking
coef
fitted
logLik
nobs
vcov
AIC, BIC
extractAIC
anova
drop1
add1
confint
profile
predict
step, stepAIC
slice
convergence
print
summary
model.frame
model.matrix
update
F (j x T
i )
Extended CLM:
T
g(j ) w T
i j x i
exp(z T
i )
threshold effects
nominal effects
scale effects
CLMM (Mixed effects):

fixed
random
F (j X Z b )

Overview
Psychoco 2012
20 / 64
Thresholds: impose restrictions

Y:

Overview
Y:
The cumulative link model:
P (Yi j ) = F (j (tempi ))
warm
j ordered, but otherwise not

restricted
require symmetry?
P(Y = 2|cold)
require symmetry?
P(Y = 2|cold)
require equidistance?
cold
cold
P (Yi j ) = F (j (tempi ))
j ordered, but otherwise not
restricted
require equidistance?
21 / 64
warm
Psychoco 2012
Psychoco 2012
22 / 64
Psychoco 2012
22 / 64
Overview
> fm.equi <- clm(rating ~ contact + temp, data=wine,

link="probit", threshold="equidistant")
> summary(fm.equi)
> fm.equi <- clm(rating ~ contact + temp, data=wine,

link="probit", threshold="equidistant")
> fm.flex <- clm(rating ~ contact + temp, data=wine,
link="probit")
> anova(fm.flex, fm.equi)
formula: rating ~ contact + temp

data:
wine
link
threshold
nobs logLik AIC
probit equidistant 72
-87.24 182.47 4(0) 1.40e-08 3.2e+01
Likelihood ratio tests of cumulative link models:
Coefficients:
contactyes
0.8571
0.2645
3.241 0.00119 **
tempwarm
1.4891
0.2882
5.166 2.39e-07 ***
--Signif. codes: 0 '***' 0.001 '**' 0.01 '*' 0.05 '.' 0.1 ' ' 1
formula:
link: threshold:
fm.equi rating ~ contact + temp probit equidistant
fm.flex rating ~ contact + temp probit flexible
fm.equi
fm.flex
threshold.1 -0.5865
0.2326 -2.522
spacing
1.2415
0.1284
9.668

Overview
Psychoco 2012
23 / 64
Nominal effects: relax restrictions
contact:no
contact:yes
3|4
2|3
1|2
Nominal effects:
ij = F (j 1 (tempi )2j (contacti ))
This is:
partial proportional odds
(Peterson and Harrell Jr., 1990)
Threshold

Overview
Psychoco 2012
24 / 64
> fm3 <- clm(rating ~ temp, nominal=~contact, data=wine, link="probit")

> summary(fm3)
ij = F (j 1 (tempi )2 (contacti ))
4|5
no.par
AIC logLik LR.stat df Pr(>Chisq)
4 182.47 -87.237
6 183.52 -85.761 2.9515 2
0.2286
Fitting nominal effects with clm

Overview
Treatment coding:
j : for contact: no
2j : contact: yes no
Psychoco 2012
25 / 64
formula: rating ~ temp

nominal: ~contact
data:
wine
link
probit flexible 72
-85.33 188.65 5(0) 3.73e-13 3.9e+01
.....
1|2.(Intercept) -0.7829
0.3178 -2.464
2|3.(Intercept)
0.7521
0.2707
2.779
3|4.(Intercept)
2.1323
0.3674
5.804
4|5.(Intercept)
2.7544
0.4512
6.105
1|2.contactyes
-0.8229
0.5650 -1.457
2|3.contactyes
-0.8892
0.3431 -2.592
3|4.contactyes
-1.0094
0.3797 -2.659
4|5.contactyes
-0.5818
0.4800 -1.212
Psychoco 2012
26 / 64
Overview
Including scale effects

Y:
Outline
5
Model for latent bitterness:
Si = + 1 (tempi ) + 2 (contacti ) + i ,
The sensR package
i N (0, (tempi ))
warm
(Cox, 1995)
ij = F
j 1 (tempi ) 2 (contacti )
1 (tempi )
cold
1

Psychoco 2012
29 / 64

ML estimation of CLMs
Approaches to ML estimation of CLMs
Objective: Optimize the log-likelihood function
Conventional approaches:
`(, ; y ) =
n
X
30 / 64
IRLS for multivariate GLMs (Fahrmeir and Tutz, 2001)

vglm in VGAM (Yee, 2010)
wi log i
General purpose optimization (quasi-Newton)

polr from MASS using optim (Venables and Ripley, 2002).
i=1
i = ij i,j 1
Psychoco 2012
ij = F (j x T
i )
Approach in ordinal:
Accurately
CLM-specific Newton-Raphson algorithm
Reliably
Fast
Psychoco 2012
31 / 64
Psychoco 2012
32 / 64
IRLS for multivariate GLMs:
Implementation the approach
Solve for = [ T , T ]T :
XTW X = XW z
Key aspects of the implementation in clm:

A novel matrix expression of CLMs (this is key!)
Size is important here!
ML Estimation via a Newton-Raphson algorithm

is q + p = r
X is nq r
W is block diagonal with n blocks of q q in total nq nq
Parameters updated in an R-environment
A large computational problem

Psychoco 2012
33 / 64
A novel matrix expression of CLMs

Psychoco 2012
34 / 64
A few details on matrix the expression
From:
ij = F (j x T
i )
Step 1: Change index j k ; j = Yi k + 1 where k = 1, 2
To:
k = F (B k + o k )
ik = ik x T
i
ik = F (ik )
k = 1, 2
Step 2: Generate design matrices:
B k and o k are fixed generate them once!
k = F (k )
Why?
k = Ak X + o k
Step 3: Concatenate design matrices:
It leads to a fast and simple algorithm

Gradient is simple and fast
k = B k + o k
Hessian is simple and fast

Covers extended model framework
Psychoco 2012
35 / 64
Psychoco 2012
36 / 64
Generating matrices in R
A (modified) Newton-Raphson algorithm

The Newton step:
Initialize environment:
> rho <- new.env(parent = parent.frame())
(i+1) = (i) h
Generate o k from y (a factor):

> A <- 1 * (col(matrix(0, n, nlevels(y))) == c(unclass(y)))
> rho$o1 <- c(1e5 * A[, nlevels(y)])
> rho$o2 <- c(-1e5 * A[,1])
Step halving: /2 in case of overshoot

Stop when:
Generate Ak :
> A1 <- A[, -(ntheta + 1), drop = FALSE]
> A2 <- A[, -1, drop = FALSE]
max |g ()| <
NR step is in right direction (log-likelihood is concave)

Quadratic convergence
> rho$B1 <- cbind(A1, -X)

> rho$B2 <- cbind(A2, -X)
Gradient and Hessian are easy and fast to compute

Psychoco 2012
37 / 64
The negative log-likelihood

`(; y ) =
wi log i
i=1
38 / 64
g (; y ) = C T $
= F (1 ) F (2 )
T
C T = BT
1 11 B 2 12
k = B k + o k
are n n diagonal
> clm.nll <- function(rho) { ## negative log-likelihood

with(rho, {
eta1 <- drop(B1 %*% par) + o1
eta2 <- drop(B2 %*% par) + o2
fitted <- pfun(eta1) - pfun(eta2)
if(all(fitted > 0))
-sum(wts * log(fitted))
else Inf
})
}
Psychoco 2012
The gradient

n
X
(default: = 106 )
Why is NR good for CLM estimation?
Assign B k = [Ak , X ] to rho:
H ( (i) )h = g ( (i) )
A simple cross product

> clm.grad <- function(rho) { ## gradient of the negative log-likelihood
with(rho, {
p1 <- dfun(eta1)
p2 <- dfun(eta2)
wtpr <- wts/fitted
dpi.psi <- B1 * p1 - B2 * p2
-crossprod(dpi.psi, wtpr)
})
}
Psychoco 2012
39 / 64
Psychoco 2012
40 / 64
The Hessian
T
T
H (; y ) = B T
1 21 B 1 B 2 22 B 2 C 3 C
How does this estimation routine handle the extended model framework?
are n n diagonal
Simple cross products

> clm.hess <- function(rho) { ## hessian of the negative log-likelihood
with(rho, {
dg.psi <- crossprod(B1 * gfun(eta1) * wtpr, B1) crossprod(B2 * gfun(eta2) * wtpr, B2)
-dg.psi + crossprod(dpi.psi, (dpi.psi * wtpr / fitted))
})
}

Psychoco 2012
41 / 64
Structured thresholds
Psychoco 2012
42 / 64
Nominal effects
T
ij = j w T
i j x i
ij = g(j ) x T
i
Step 1: W is design matrix for a single factor or covariate

Step 2: Define:
D k = Ak : W
Step 1: Define Jacobian J for the transformation: = J

Step 2: Redefine = [T , T ]T and B k = [Ak J T , X ]
nqs
nq
ns
Step 3: Redefine B k = [D k , X ]
Result:
The model can still be written as: k = B k + o k
Result:
The algorithm does not change!
The model can still be written as: k = B k + o k
The log-likelihood, the gradient and the Hessian apply unchanged!
The algorithm does not change!

The log-likelihood, the gradient and the Hessian apply unchanged!
> B1 <- cbind(A1 %*% tJac, -X)

> B2 <- cbind(A2 %*% tJac, -X)

> tmp1 <- lapply(1:ncol(NOM), function(x) A1 * NOM[,x])

> B1 <- do.call(cbind, tmp1)
Psychoco 2012
43 / 64
Psychoco 2012
44 / 64
Scale effects
Outline
g = [C 2 , C 3 ]T $

D ET
H =
E F
Result:
The log-likelihood, the gradient and the Hessian are slightly more
complicated
The algorithm changes slightly
Psychoco 2012
45 / 64
Accuracy of parameter estimates
Relative loglikelihood
link
niter max.grad
probit flexible 72
-85.76 183.52 5(0) 1.59e-13
Coefficients:
tempwarm contactyes
1.4994
0.8677
4|5
2.9413
1.344388
Has the model converged?
1.344382
2.503099
1|2
How accurate are these estimates?
Psychoco 2012
46 / 64
> slice.fm1 <- slice(fm1, parm = c(1, 6))

> par(mfrow=c(1,2))
> plot(slice.fm1)
formula: rating ~ temp + contact

data:
wine

Assessment of model convergence
> (fm1 <- clm(rating ~ temp + contact, data=wine, link="probit"))
1|2
2|3
3|4
-0.7733 0.7360 2.0447
1e11

3e11
5e11
Hessian:
ii = exp(Z )i
Gradient:
The sensR package
1e11
k = (B k + o k ),
3e11
In matrices:
j x T
i
T
exp(z i )
5e11
ij =
2.503102
2.503105
tempwarm
See vignette for more details.

Psychoco 2012
47 / 64
Psychoco 2012
48 / 64
Assessment of parameter accuracy
Robustness of starting values

Standard starting values can fail:
> convergence(fm1)
> data(iris)
> iris.polr <- polr(Species ~ Sepal.Length + Sepal.Width + Petal.Length +
Petal.Width, data=iris)
nobs logLik niter max.grad cond.H logLik.Error

72
-85.76 5(0) 1.59e-13 2.2e+01 <1e-10
1|2
2|3
3|4
4|5
tempwarm
contactyes
Error in polr(Species ~ Sepal.Length + Sepal.Width + Petal.Length + Petal.Width,

attempt to find suitable starting values failed
In addition: Warning messages:
1: glm.fit: algorithm did not converge
2: glm.fit: fitted probabilities numerically 0 or 1 occurred
Estimate Std.Err Gradient

Error Cor.Dec Sig.Dig
-0.7733 0.2829 1.86e-14 3.60e-16
15
15
0.7360 0.2499 1.38e-13 -4.53e-16
15
15
2.0447 0.3218 -1.59e-13 -8.48e-15
13
14
2.9413 0.3873 2.66e-15 -7.40e-15
13
14
1.4994 0.2918 -9.83e-15 -4.64e-15
14
15
0.8677 0.2669 -5.50e-15 -2.61e-15
14
14
Eigen values of Hessian:

61.616 53.876 32.283 17.241 13.393
2.825
The Method Independent Error Theorem (Elden et al., 2004)

Psychoco 2012
49 / 64

Psychoco 2012
51 / 64
> data(iris)
> data(iris)


This runs fine, though:
This runs fine, though:
> set.seed(1)
Petal.Width, data=iris, start = runif(6), Hess=TRUE)
> set.seed(1)
Petal.Width, data=iris, start = runif(6), Hess=TRUE)
and so does:
> iris.clm <- clm(Species ~ Sepal.Length + Sepal.Width + Petal.Length +
Psychoco 2012
51 / 64
Psychoco 2012
51 / 64
Comparing parameter estimates
Comparing parameter estimates
Estimate Std. Error

Sepal.Length
-2.465
2.394
Sepal.Width
-6.681
4.479
Petal.Length
9.429
4.737
Petal.Width
18.286
9.742
setosa|versicolor
5.292
550.912
versicolor|virginica
42.638
25.707
polr
0.0e+00
clm:
Value Std. Error

Sepal.Length
-2.464
2.393
Sepal.Width
-6.681
4.480
Petal.Length
9.427
4.734
Petal.Width
18.286
9.739
setosa|versicolor
3.629
0.013
versicolor|virginica 42.631
25.670
clm
2.0e05 1.0e05
polr:
> plot(slice(iris.clm, parm=1, lambda=5e-3, quad=FALSE))

> abline(v=iris.polr$zeta[1], col="red")
> mtext(c("polr", "clm"), at=c(iris.polr$zeta[1], iris.clm$alpha[1]),
line=1)
setosa|versicolor

Psychoco 2012
53 / 64
Assessing the accuracy of parameter estimates

Psychoco 2012
54 / 64
Assessing the accuracy of parameter estimates

> iris.clm2 <- update(iris.clm, gradTol = 1e-07)
> convergence(iris.clm2)
150 -5.95 19(0) 6.59e-11 4.0e+07 <1e-10
> convergence(iris.clm)
150 -5.95 18(0) 1.56e-07 4.0e+07 <1e-10
setosa|versicolor
Sepal.Length
Sepal.Width
Petal.Length
Petal.Width
setosa|versicolor
Sepal.Length
Sepal.Width
Petal.Length
Petal.Width

5.292 550.912 2.23e-08 6.75e-03
1
2
42.638 25.707 5.79e-12 -1.79e-05
4
6
-2.465
2.394 -1.41e-07 2.34e-07
6
7
-6.681
4.479 -4.00e-08 2.61e-06
5
6
9.429
4.737 -1.56e-07 -3.14e-06
5
6
18.286
9.742 -5.59e-08 -6.69e-06
4
6

5.286 550.920 4.55e-13 1.36e-07
6
7
42.638 25.707 7.13e-14 -1.11e-08
7
9
-2.465
2.394 -2.13e-11 1.45e-10
9
10
-6.681
4.479 1.07e-11 1.63e-09
8
9
9.429
4.737 -6.59e-11 -1.96e-09
8
9
18.286
9.742 -2.50e-11 -4.16e-09
8
10

1.329e+02 1.686e-01 6.959e-02 1.933e-02 1.367e-03 3.295e-06

1.329e+02 1.686e-01 6.959e-02 1.933e-02 1.367e-03 3.295e-06
Silent divergence is an important issue!

See also (Marschner, 2011) for similar issues with glm.
Psychoco 2012
55 / 64
Psychoco 2012
56 / 64
Outline
Including random effects

The sensR package
ij = F (j 1 (tempi ) 2 (contacti ))
warm
Judges perceive wine bitterness

differently
Add random effects for judges:
ij = F (j 1 (tempi ) 2 (contacti )
Judges use the response scale

differently
cold
1

Psychoco 2012
57 / 64


differently
warm

differently
cold
cold
b(judgei )), b N (0, b2 )
Psychoco 2012

differently
differently
58 / 64
warm
Psychoco 2012

58 / 64
Psychoco 2012
58 / 64


differently
warm

differently
warm

differently
cold
1

differently

Psychoco 2012
cold
58 / 64
Fitting cumulative link mixed models with clmm

Psychoco 2012
58 / 64
Cumulative link mixed models
> fm.ran <- clmm(rating ~ temp + contact + (1|judge), nAGQ=10, data=wine)

> summary(fm.ran)
k = F (B k Z v o k )
Cumulative Link Mixed Model fitted with the adaptive Gauss-Hermite

quadrature approximation with 10 quadrature points
V N (0, )
The log-likelihood function:
formula: rating ~ temp + contact + (1 | judge)

data:
wine
`(, ; y ) = log
link threshold nobs logLik AIC

niter
max.grad cond.H
logit flexible 72
-81.53 177.06 16(723) 3.23e-06 2.8e+01
Rr
p (y |v )p (v ) dv
Integration methods:
Random effects:
Var Std.Dev
judge 1.288
1.135
Number of groups: judge 9
Laplace approximation (Tierney and Kadane, 1986; Pinheiro and Bates, 1995;
Coefficients:
tempwarm
3.0619
0.5951
5.145 2.67e-07 ***
contactyes
1.8334
0.5124
3.578 0.000346 ***
.....
Adaptive Gauss-Hermite quadrature (AGQ) (Liu and Pierce, 1994)
Joe, 2008)
Gauss-Hermite quadrature (GHQ) (Hedeker and Gibbons, 1994)
A Newton-Raphson algorithm updates the conditional modes of the

random effects (Laplace and AGQ)
Psychoco 2012
60 / 64
Psychoco 2012
61 / 64
References
Estimation of cumulative link mixed models
References
Agresti, A. (2002). Categorical Data Analysis (Second ed.). Wiley.
Bates, D. and M. Maechler (2012). Matrix: Sparse and Dense Matrix Classes and Methods. R package version 1.0-3.
1 random term:
Cox, C. (1995). Location-scale cumulative odds models for ordinal data: A generalized non-linear model approach. Statistics in
medicine 14, 11911203.
implemented in C
Eld
en, L., L. Wittmeyer-Koch, and H. B. Nielsen (2004). Introduction to Numerical Computation analysis and MATLAB
illustrations. Studentlitteratur.
Laplace, GHQ, AGQ
Fahrmeir, L. and G. Tutz (2001). Multivariate Statistical Modelling Based on Generalized Linear Models (Second ed.). Springer
series in statistics. Springer-Verlag New York, Inc.
Greene, W. H. and D. A. Hensher (2010). Modeling Ordered Choices: A Primer. Cambridge University Press.
2 random terms:
Hedeker, D. and R. D. Gibbons (1994). A random-effects ordinal regression model for multilevel analysis. Biometrics 50,
933944.
sparse matrix methods from Matrix (Bates and Maechler, 2012)
Joe, H. (2008). Accuracy of laplace approximation for discrete response mixed models. Comput. Stat. Data Anal. 52 (12),
50665074.
exclusively in R
Liu, Q. and D. A. Pierce (1994). A note on gauss-hermite quadrature. Biometrika 81 (3), 624629.
Marschner, I. C. (2011, December). glm2: Fitting Generalized Linear Models with Convergence Problems. The R Journal 3 (2),
1215.
Laplace
Peterson, B. and F. E. Harrell Jr. (1990). Partial proportional odds models for ordinal response variables. Applied Statistics 39,
205217.
Pinheiro, J. C. and D. M. Bates (1995). Approximations to the nonlinear mixed-effects model. Jounal of Computational and
Graphical Statistics 4 (1), 1235.
Speed is an issue here:

The speed of clmm is really stunning. A tryout three-level model in
GLLAMM took 3 hours, in clmm about 15 minutes.
Randall, J. (1989). The analysis of sensory data by generalised linear model. Biometrical journal 7, 781793.
Tierney, L. and J. B. Kadane (1986). Accurate approximations for posterior moments and marginal densities. Journal of the
American Statistical Association 81 (393), 8286.
Venables, W. N. and B. D. Ripley (2002). Modern Applied Statistics with S (Fourth ed.). New York: Springer. ISBN
0-387-95457-0.
Yee, T. W. (2010, 1). The vgam package for categorical data analysis. Journal of Statistical Software 32 (10), 134.

Thank you
Psychoco 2012
62 / 64
Psychoco 2012
64 / 64
Thank you for listening!
Psychoco 2012
63 / 64

Christensen

Uploaded by

Copyright:

Available Formats

Christensen

Uploaded by

Document Information

Copyright

Available Formats

Share this document

Share or Embed Document

Sharing Options

Did you find this document useful?

Is this content inappropriate?

Copyright:

Available Formats

Christensen

Uploaded by

Copyright:

Available Formats

Background

Background on Rune H B Christensen

February 9th 2012

Rune H B Christensen (DTU)

The sensR and ordinal packages

DTU Informatics, IMM

PhD: Sensometrics: Thurstonian and Statistical Models

Rune H B Christensen (DTU)

The sensR and ordinal packages

The sensR package

The sensR package

The ordinal package overview

The ordinal package overview

Implementation in the ordinal package

Implementation in the ordinal package

Assessment of estimation accuracy

Assessment of estimation accuracy

Cumulative link mixed models (CLMMs)

Cumulative link mixed models (CLMMs)

Rune H B Christensen (DTU)

The sensR and ordinal packages

Rune H B Christensen (DTU)

The sensR and ordinal packages

The sensR package

The sensR package

Psychometric protocols supported in sensR

The sensR package

The sensR and ordinal packages

The sensR and ordinal packages

Rune H B Christensen (DTU)

The sensR and ordinal packages

d': sensory difference

What is the probability of a correct answer?

Beyond the basics:

The Thurstonian model, 3-alternatives

Basic functions in sensR

Examples for papers

Statistical methodology for sensory discrimination tests and its

Rune H B Christensen (DTU)

On CRAN since July 2008:

Estimation and inference in Thurstonian models for sensory discrimination

It depends on the question 3-AFC or Triangle.

Rune H B Christensen (DTU)

The sensR and ordinal packages

The sensR package

The sensR package

Psychometric functions: Inverse link functions

Rune H B Christensen (DTU)

fduo-trio (d 0 ) = 1 (d 0 / 2) (d 0 / 6) + 2(d 0 / 2)(d 0 / 6).

Rune H B Christensen (DTU)

The ordinal package

Regression models for ordinal data via cumulative link models

The sensR package

The ordinal package overview

On CRAN since March 2010:

Implementation in the ordinal package

Assessment of estimation accuracy

Cumulative link mixed models (CLMMs)

clm tutorial (18 pages)

Rune H B Christensen (DTU)