Scribd
Upload
36 views5 pages

Some Applications Involving Polytomous Data: (A) Matched Pairs: Nominal Response

The document describes two models for analyzing polytomous or ordinal response data from matched pairs: 1) A nominal response model that models the log odds of each response category as a function of the treatment effect. This model is identical to the Bradley-Terry model for paired comparisons. 2) An ordinal proportional odds model that models the cumulative log odds of each response category as a function of the treatment effect. This model accounts for the ordered nature of the categories. The document discusses methods for constructing likelihood-based inferences while eliminating nuisance parameters, such as using quasi-likelihood based on estimating equations. Optimal weights for the estimating equations are derived.

Uploaded by

juntujuntu
Copyright
© Attribution Non-Commercial (BY-NC)
We take content rights seriously. If you suspect this is your content, claim it here.
Available Formats
Download as PDF, TXT or read online on Scribd
36 views5 pages

Some Applications Involving Polytomous Data: (A) Matched Pairs: Nominal Response

The document describes two models for analyzing polytomous or ordinal response data from matched pairs: 1) A nominal response model that models the log odds of each response category as a function of the treatment effect. This model is identical to the Bradley-Terry model for paired comparisons. 2) An ordinal proportional odds model that models the cumulative log odds of each response category as a function of the treatment effect. This model accounts for the ordered nature of the categories. The document discusses methods for constructing likelihood-based inferences while eliminating nuisance parameters, such as using quasi-likelihood based on estimating equations. Optimal weights for the estimating equations are derived.

Uploaded by

juntujuntu
Copyright
© Attribution Non-Commercial (BY-NC)
We take content rights seriously. If you suspect this is your content, claim it here.
Available Formats
Download as PDF, TXT or read online on Scribd
You are on page 1/ 5

Some Applications Involving Polytomous Data

(a) Matched pairs: nominal response

Suppose the response is polytomous with k categories. Denote the logarithmic response probabilities for the control member

= + =

( 1 , 2 , K

,k

)
k

and the logarithmic response probabilities for the treatment member

( 1

+ 1,2 +

,K , k +

).

The probability of observing response category j for control subject is

exp ( j )

exp ( ) ,
r =1 r

while the probabilities for the treatment subject are

exp (

r =1

exp ( r +

r )

Let R1 and R 2 be the response for the control and treatment

1, 2 , K , k } . Let the indicator Z 1 groups, respectively, R1 , R 2 {


and Z 2 be the indicator vectors for control and treatment groups, respectively,

[0

1]

the ith component as R 1 = i and

= [0

1]

the jth component


1

as R

= j . Thus, the conditional distribution given

Z = Z 1 + Z 2 = [0
is

0 1
ith

0 1 0
jth

0]

exp(i ) P(R1 = i | Z ) = exp(i ) exp( j + j


k r =1

exp(r )
r =1

exp(r )
r =1

exp( + ) ) exp( )
r =1 r r

exp( j + j )
k

exp(r + r )

exp(r )
r =1

exp(i + i )
k r

exp(
r =1

+ r ) .

exp( i ) + exp( j )
and is only dependent on the

exp( j )

i j , which is independent of

treatment effect . Thus, if Yij is the number of ordered pairs responding

(i , j ) , the symmetric total

Y = Yij + Y ji = m ij is just

the number of vector sums Z that have values in positions i and j. Thus, given Y = m ij ,

Yij ~ B (m ij , ii ), logit ( ij ) = j i , i < j .


The conditional likelihood based on Yij involves the treatment effect . given Y = m ij only

Note:
The above model is formal identical to Bradley-Terry (1952) model used for ranking individuals in a paired competition.

Note:
A curious and unusual feature of the above model logit( ij ) = j i is that the model matrix X corresponding to the formula j i is singular, for example, if k = 3 ,

logit ( 12 ) 2 1 1 ( ) logit = 3 13 1 = 1 logit ( ) 0 23 2 3

1 0 1

0 1 = X 1 , 2 1 3

rank ( X ) = 2 .
(b) Ordinal responses

Consider the proportional-odds model for the comparison of 2 multinomial responses in which the categories are ordered.
G1 Y1 Sampled Not-sampled s1 Y1 = X 1 G2 Y2

Gk Yk s k Yk = X k sk

Total
m1

s 2 Y2 = X 2 s2

m2 m = s

s1

Let

Y ~ M (m 1 , 1 ), X ~ M (m 2 ,
in which the cumulative probabilities

),

logit (r1 j ) = j

r1 j , r2 j

satisfy

logit (r2 j ) = j , j = 1, 2, K , k 1 .
Let

s j = X j + Yj

for the response category totals and

Z 1 j = Y1 + Y2 + L + Y j , Z2 j = X1 + X 2 + K + X j , s j = s1 + s 2 + L + s j = Z j = Z 1 j + Z 2 j
Note that the treatment effect and parameters. To eliminate the nuisance parameter, the conditional distribution of .

Z ij ~ B (m i , rij ) . In this model, we are interested in

j s

are considered as the nuisance

Z 1 j given s j is the noncentral hypergeometric distribution


P (Z 1 j m1 m 2 z1 j z s z 1j 1j j , = e , j = 1, K , k 1 = z1 j | s j ) = b m1 m 2 r r s r r =a j
. The aim is to construct the conditional distribution based on Unfortunately, there is no joint conditional distribution of depends only on

Z1j .

Z 1 j that Z 1 j given
nuisance

. The conditional distribution of

s = (s 1 , s 2 , K , s k
parameters

depends

on

both

the

1 , 2 , K , k 1

and the parameter of interest .

Thus we could not eliminate the nuisance parameters using conditional likelihood. However, we might use the following method similar to quasi-likelihood. Let

1 j ( ) = 1 j (e ) = E (Z 1 j | s j ) =
4

P1 ( ) P0 ( ) .

Since the score function U ( ) (the first derivative of the log-likelihood function) is a linear function of the data, we can use

U ( ) =

w [Z
k 1 j =1 j

1j

1 j e ,

( )]

as a score function, where. We should choose w j so that U ( )

U () E = Var[U ()] . ( ) w = w , s has j zero mean and j j


However, the choice of weights is normally not critically important. At worst, a poor choice of weights may lead to a small loss of efficiency. For example, we may choose

wj = 1

or

wj = s j (m s j ) . By the standard theory for linear estimating


equations, the score function with the optimal weights is

U ( )

j + j +1 Z e 1 j 1 j , j =1
k 1

( )]

where

j =

1 j

2 j

, ij = ij ( ) ij 1 ( ) and = j .
j =1

, which is the solution of the above score function, The estimate c


has asymptotic variance

Var c

( )

k 1 j 3 (m 1 ) 1 m j =1

You might also like