You are on page 1of 6

Nathan Vos

June 5, 2014
The assumption of independence of observations is violated here because
it appears to be the case that most judges sample each combination of food,
though probably not each ordering of each combination. For example, if a judge
has a predisposition to A, he will likely prefer it to B, to C, and to D. So we have
positive correlation between observations. This is not too concerning because
this positive correlation reduces variances of contrasts, and so contrast variances
that would be obtained by the model can be seen as upper bounds.
The first two mosaic plots below give strong indication that order of presentation of foods is quite important in determining preference. In all matches, the
first food was preferred by a majority of those sampling it. These plots strongly
indicate that order presented is more important in determining preference than
food quality for these foods.
The last two mosaic plots both indicate that B seems to be somewhat preferred to the other foods relative to foods when accounting for order presented,
that is when considering the prefence between A,B,C and D when each is presented first, food B has the highest proportion of preference, and a similiar
phenomenon occurs when considering all foods when presented second.
(Analysis continues after plots.)

Preference s or t
BA

AC

CA

AD

DA

BC

Prefer.t

NoPref

Prefer.s

AB

CB

BD

DB

CD

DC

Preference 1st or 2nd


BA

AC

CA

AD

DA

BC

Prefer2nd

NoPref

Prefer1st

AB

CB

BD

DB

CD

DC

Preference when presented first


B

Prefer2nd

NoPref

Prefer1st

Preference when presented 2nd


B

Prefer2nd

NoPref

Prefer1st

I assume here that no overdispersion is present. The model I fit is a cumulative logit model:
logit(

jst )

= j

(F irsts + Secondt )

There are two 4 level factor covariates: first, indicating which food was tried
first, and second, indicating which food was tried second in that given test.
Interactions are not included as a chi squared test for lack of fit (calculated using
the model deviances) indicated that the main eects model was appropriate. I
computed the chi square test statistic on 24-8=16 df for lack of fit by computing
the deviances for the fitted model and the saturated model, and got a pvalue of
.30, indicating a lack of fit was not found.
One way to interpret the model is that the fitted value for observation j =
a, s = S, t = T is the logit of the estimate of the probability a judge would
strictly prefer (j = 1) or prefer/be indierent to (j = 2) the food S to food
T when food S is served first. It should be noted that the interpretation for
the sign of levels of factor First are dierent from factor Second. Specifically, a
positive coefficient for F irstS can be interpreted as weaker overall preference for
food S, when presented first, compared to overall preference for the baseline food
(A), when presented first , and this is due to the minus sign convention in the
model above. Analagously, a positive coefficeint for SecondS can be interpreted
as stronger overall preference for food S, when presented second, comparted to
overall preference for baseline when presented second. This is because jst gives
5

an estimate of the proportion that prefer the first food presented.

1
2
3
4
5
6
7
8
9
10
11
12

Pair
AB
AB
AC
AC
AD
AD
BC
BC
BD
BD
CD
CD

Order
AB
BA
AC
CA
AD
DA
BC
CB
BD
DB
CD
DC

Prefer.s
23
6
27
14
35
11
34
16
29
15
26
14

NoPref
8
8
5
6
1
4
1
3
2
5
5
5

Prefer.t
11
29
11
22
6
27
6
23
9
22
11
24

First
A
B
A
C
A
D
B
C
B
D
C
D

Second
B
A
C
A
D
A
C
B
D
B
D
C

Table 1: Taste Preference Data

1|2
2|3
FirstB
FirstC
FirstD
SecondB
SecondC
SecondD

Estimate
0.705
1.215
-0.247
0.488
0.333
0.268
-0.025
-0.349

Std. Error
0.290
0.294
0.295
0.273
0.269
0.265
0.276
0.283

z value
2.429
4.135
-0.837
1.788
1.239
1.010
-0.090
-1.232

Table 2: Coefficients

Pr(>|z|)
0.015
0.000
0.403
0.074
0.215
0.313
0.928
0.218

Prefer1st
23
29
27
22
35
27
34
23
29
22
26
24

Prefer2nd
11
6
11
14
6
11
6
16
9
15
11
14

You might also like