Location via proxy:   [ UP ]  
[Report a bug]   [Manage cookies]                
Scribd
Upload
35 views

Some Applications Involving Binary Data: (A) Comparison of Two Binomial Probabilities

This document discusses two approaches to comparing binomial probabilities from clinical trial data: 1. The full likelihood approach fits a linear logistic model to estimate the treatment effect (Δ) using the data from all centers. 2. The conditional likelihood approach estimates Δ by maximizing the conditional likelihood given the total number of successes and failures at each center. This approach only estimates one parameter (Δ). When combining data from multiple centers, the full likelihood is difficult to use as it involves estimating many parameters. The conditional likelihood approach allows consistent and efficient estimation of the treatment effect by only depending on one parameter. A Mantel-Haenszel test can be used to test if the treatment effect is equal to zero based on the

Uploaded by

juntujuntu
Copyright
© Attribution Non-Commercial (BY-NC)
Available Formats
Download as PDF, TXT or read online on Scribd
35 views

Some Applications Involving Binary Data: (A) Comparison of Two Binomial Probabilities

This document discusses two approaches to comparing binomial probabilities from clinical trial data: 1. The full likelihood approach fits a linear logistic model to estimate the treatment effect (Δ) using the data from all centers. 2. The conditional likelihood approach estimates Δ by maximizing the conditional likelihood given the total number of successes and failures at each center. This approach only estimates one parameter (Δ). When combining data from multiple centers, the full likelihood is difficult to use as it involves estimating many parameters. The conditional likelihood approach allows consistent and efficient estimation of the treatment effect by only depending on one parameter. A Mantel-Haenszel test can be used to test if the treatment effect is equal to zero based on the

Uploaded by

juntujuntu
Copyright
© Attribution Non-Commercial (BY-NC)
Available Formats
Download as PDF, TXT or read online on Scribd
You are on page 1/ 7

Some Applications Involving Binary Data

(a) Comparison of two binomial probabilities

Motivating Example:

In a small-scale clinical trial, the following table might represent the contribution of one of the smaller centres to the study. Success Treatment Control
Y =2 X =1
X +Y = 3

Failure
3Y =1

m1 = 3 m2 = 4 m = 7

4 X =3

There are two approaches to make the inference about treatment effect. The first one is based on the full likelihood while the second one is based on the conditional likelihood. (i) Full likelihood approach

Suppose Y ~ B (m 1 , treatment group and

is the number of successes in the


2

X ~ B (m

is the number of

successes in the control group, where

logit ( 1 ) = + logit ( 2 ) =

and

is a good and useful measure of the treatment effect. Then,

we can fit the linear logistic model and use the methods in chapter 4. Then, we have

= logit ( ) logit ( ) 1 2 1 = log 1 1 2 log 1 2

1 (1 2 ) = log (1 ) 1 2
Y X 1 m2 m1 = log X Y 1 m m1 2 Y (m 2 X ) = log ( ) X m Y 1 2 3 = log 1 1 = 1.792
and the large-sample variance of
Var )= ( =
is

1 1 1 1 + + + Y X m1 Y m2 X

1 1 1 1 + + + 2 1 1 3 17 = 6
2

Denote the profile deviance,

2l , , , D (y , ) = 2l
where

is the maximum profile likelihood estimate of for


100 (1 )% confidence interval for is

given . Then, a

) < }. { : D ( y , ) D (y ,
For example, a 90% confidence interval for is

)< { : D ( y , ) D (y ,
Thus, (ii) Since

2 0 .1

= 2 . 71 = ( 0 . 8 , 4 . 95 ) .

0 ( 0 .8, 4 .95 ) , the treatment effect is not significant.


Conditional likelihood approach

Y ~ B (m1 , 1 ), X ~ B (m 2 , 2 ), logit ( 1 ) = + , logit ( 2 ) =


we can use conditional likelihood given X + Y = 3 is

H (m, s, ) ,

1 (1 2 ) = ( ) ( ) m = 3 , 4 , s = 3 , 4 where and 2 (1 1 ) . Then,


= logit ( 1 ) logit ( 1 ) 1 2 = log log 1 1 1 2 1 (1 2 ) = log (1 ) 1 2 = log ( )
and the log-likelihood function is

y lc ( ) log b 3 4 j j = a 3

= y log ( ) log P0 e

[ ( )]

= y log ( ) log [P ( )] 0 j

where

P0 ( ) =

3 4 j j 3 j j = a 3 4 3 4 3 4 2 3 4 3 = + + 0 3 1 2 2 1 + 3 0 = 4 + 18 + 12 2 + 3
b

The estimate

of satisfies

P0' e e l c ( ) = y =0 P0 e = = c c y=
Thus,

P0' e c e c
c 0

( ) P (e )
c

( ) ( )

can be found. In this example,

= 1 . 493 . For

the variance of
That is,

, we can use the large sample approximation .

1 = Var c I ( )
since

( )

1 l ( ) E c
2

1 Var (Y | X + Y = 3 )

2 2 P0' e e l c ( ) = E y E | X + Y = s1 = 3 P0 e = Var (Y | X + Y = 3 )

( ) ( )

the variance
Thus,

of H (m , s ,

ar V c =

( )

1 [Var (Y | X + Y = 3 )] = c
P e c c P1 e 2 c c P e P e 0 0

( ) ( )

( ) ( )

(b)

Combination of several

2 2 tables

Suppose there are n centres participating in the clinical trial. Then, the following tables can be obtained, i = 1, 2 , K , n , Success Treatment Control
Yi Xi s1i

Failure
m1i Yi m2i X i s 2i m1i m2i

We may consider

Yi ~ B (m1i , 1i ), X i ~ B (m2i , 2i ), logit ( 1i ) = i + , logit ( 2i ) = i .


5

The main difficulty based on the full likelihood is that it contains

n + 1 parameters 1 , 2 , K , n , to be estimated. In such


circumstances, maximum likelihood needs not be consistent or efficient for large n . However, if we used the conditional likelihood based on Yi | X i + Yi = s1i , the likelihood then depends on only one parameter . The conditional log-likelihood for is

l c ( ) =

i =1

y log = ( ) P 0
i 0

{y
i =1

log ( ) log [P0 (

)]}

where

{y log [P (e )]} (1 2 i ) = 1i 2 i (1 1i ) ,
i =1

1i = logit ( 1i ) logit ( 1i ) = log 1 1i 1i (1 2 i ) = log (1 ) = log ( ) 1i 2i


and

2i log 1 2i

m 1i m 2 i P0 ( ) = j j=a s1i
b

Thus, the maximum conditional likelihood estimate is the solution of

lc ( ) = 0 . In addition, provided the total conditional Fisher

information is sufficiently large, standard large-sample likelihood theory applies to the conditional likelihood. To test H 0 : = 0 , we can use the score statistic under H 0 ,
6

l ( ) = U = c = 0
In addition,

m1i ( x i + y i ) y i m1i + m 2 i . i =1
n

Var (U ) =

m1i m 2 i ( xi + y i )(m1i + m 2 i xi y i ) . (m1i + m 2 i )2 (m1i + m 2 i 1) i =1


n

Then, the test statistic is

Z =

1 2 Var (U )

for one-sided hypothesis. The test is known as the Mantel-Haenszel test.

Note:
There is the probability that varies from centre to centre. That is,

might not be constant over different centres. Such interactions


require careful investigation and detailed plausible explanation.

You might also like