Professional Documents
Culture Documents
– Matched pairs
• Non-central hypergeometric distribution
• Test for no association: McNemar test
• Estimating the odds ratio
– Conditional MLE
– Mantel-Haenszel estimate
• Confidence interval
– 1:M matching
• fixed number of controls per case
– Conditional MLE
– Mantel-Haenszel estimate
– Test for no association: Chi-square
• Variable number of controls per case
Matched Case-Control Studies
• Matched case-control study: a fixed
number of cases are identified and each
identified case is matched with one or
more than one controls on the basis of
important confounding variables (e.g. ,
age, sex).
• Matched case-control study has more
power than unmatched case-control
study.
Matched Pair: Example
• This example is a matched pair case-control that
studies the association of oral contraceptive use
with the disease, thromboembolism (blood clots
in the veins with inflammation in the vessel
walls). The cases were 175 women discharged
alive from 43 hospitals after initial attack of
thromboembolism. The controls were matched
with their cases for hospital, time of
hospitalization, race, age, martial status, parity
and pay status.
Matched Pair: Original data
sta y x sta y x
1 1 1 16 1 1
1 0 1 16 0 0
2 1 1 17 1 1
2 0 1 17 0 0
3 1 1 18 1 1
3 0 1 18 0 0
4 1 1 19 1 1
4 0 1 19 0 0
5 1 1 20 1 1
5 0 1 20 0 0
6 1 1 21 1 1
6 0 1 21 0 0
7 1 1 22 1 1
7 0 1 22 0 0
8 1 1 23 1 1
8 0 1 23 0 0
9 1 1 24 1 1
9 0 1 24 0 0
10 1 1 25 1 1
10 0 1 25 0 0
11 1 1 26 1 1
11 0 0 26 0 0
12 1 1 27 1 1
12 0 0 27 0 0
13 1 1 28 1 1
13 0 0 28 0 0
14 1 1 29 1 1
14 0 0 29 0 0
15 1 1 30 1 1
15 0 0 30 0 0
Matched Pair: Original data
• Q: Can we treat the matched pair case-control data
as case-control data, i.e., ignoring the matching
and proceed the analysis with the following 2x2
table?
x
1 0
1 67 108
y
0 23 153
Matched Pair: Original data
• A: No. The reason is that the control
sample is not a random sample of the
control population due to the fact the
selection of controls are dependent of
cases.
Matched Pair: Example
• The 2x2 table that are usually used for the analysis
of matched pair data is
x0
1 0
1 10 57
x1
0 13 95
Case exposed
n11 n10
Case unexposed
n01 n00
Matched pair: Conditional MLE
of Odds Ratio
• Q: How can we extract information from
the 2x2 table on previous slide to estimate
the odds ratio measuring the association
strength of the exposure with the disease?
• A: NOT obvious.
Matched pair: Conditional MLE
of Odds Ratio
• In order to derive conditional MLE for the
odds ratio, we view the data from each
pair as a 2x2 table of diseaseXexposure,
and consider the probability of observing
each table conditional on the row total and
column total. The conditional maximum
likelihood would be the product of such
conditional probabilities.
Matched pair: Conditional MLE
of Odds Ratio
Exposure
+ - + - + - + - Total
Case 1 0 1 0 0 1 0 1 1
Control 1 0 0 1 1 0 0 1 1
Total 2 0 1 1 1 1 0 2 2
# of
such n11 n10 n01 n00
tables
Matched Pair: Non-central
Hypergeometric Distribution
•Let the following 2x2 table represent one of the
four tables on previous slide
Exposed Unexposed
Diseased
a b n1
Disease-free n0
c d
m1 m0 N
Matched pair: Non-central
Hypergeometric Distribution
• The probability of observing the 2x2 table on the
previous slide, conditional on all the marginal totals
remaining fixed, n1 , n0 , m1 , m0 is
pr (a | n1 , n0 , m1 , m0 ; )
n1
a
n0
m1 a
a
n1
u
n0
m1 u
u
1 0
pr (1 | 1,1,2,0; )
1 1
1 2 1
1
1 1
u 2 u
u
0 1
pr (0 | 1,1,0,2; )
1 1
0 00
0
1
1 1
u 0 u
u
0 1
pr (1 | 1,1,1,1; )
1 1
1 11
1
1 1
u 0 u
u
1
max(0,11)u min(1,1)
Matched pair: Conditional MLE
of Odds Ratio
• The probability of observing
0 1
1 0
pr (0 | 1,1,1,1; )
1 1
0 11
0
1
1 1
u 0 u
u
1
max(0,11)u min(1,1)
Matched pair: Conditional MLE
of Odds Ratio
• Let
1
1
n10 n01
CL( ) (1)
1 1
• Remark: the data from concordant pairs do not
contribute to the likelihood function, that is, the
data of concordant pairs contains no information
of the odds ratio.
Matched pair: Conditional MLE
of Odds Ratio
• The conditional MLE of the odds ratio is obtained
by maximizing (1) with respect to . That is
n10
ˆCMLE
n01
Matched pair: Confidence Interval
of Odds Ratio
• Two steps:
1. Obtain the confidence interval for
L , U ˆ Z / 2 s(ˆ ), ˆ Z / 2 s(ˆ )
ˆ Z / 2ˆ (1 ˆ ) / n10 n01 , ˆ Z / 2ˆ (1 ˆ ) / n10 n01
2. Use the relationship
1
to convert L , U
to the CI for
L
L ,U , U
1 L 1 U
Matched pair: CMLE and CI of
Odds Ratio
data match11;
set match11;
y1=2-y;
run;
proc phreg data=match11;
strata sta;
model y1 = x /
details ties=discrete rl;
run;
Matched pair: CMLE and CI of
Odds Ratio
Analysis of Maximum Likelihood Estimates
N
ak d k / N k n11 0 n10 1 n01 0 n00 0 / 2 n10
ˆMH k 1
N
nn11 0 n10 0 n01 1 n00 0 / 2 n01
bk ck / N k
k 1
Matched pair: M-H Estimate of
Odds Ratio
McNemar's Test
+ - + - + - Total
1 0 1 0 1 0 1
Case
Control M 0 M-1 1 0 M M
# of
such n1M n1M 1 n10
tables
Exposure
+ - + - + - Total
Case 0 1 0 1 0 1 1
Control M 0 M-1 1 0 M M
n0 M n0 M 1 n00
1:M matching: Conditional MLE
of Odds Ratio
• First we consider the conditional probability of
observing the first table and that of observing the
last table. We will show that both conditional
probabilities are equal to 1.
• The 2M remaining tables may be paired into
sets of two, each having the same marginal total
of exposed. For example, the table with both the
case and two controls positive is paired with the
table with three controls positive and the case
negative.
1:M matching: Conditional MLE
of Odds Ratio
• The probability of observing
0 1
0 M
pr (1 | 1, M ,0, M 1; )
1 M
00
0
0
1
u
M
2 u
u
1
M
00
0
1
0
1 M 0
0 00
1:M matching: Conditional MLE
of Odds Ratio
• The 2M remaining tables may be paired
into sets of two, each having the same
marginal total of exposed. For example,
the table with both the case and two
controls positive is paired with the table
with three controls positive and the case
negative.
1:M matching: binary
exposure
• More generally, we pair together the
following two tables, and calculate their
respective conditional probability.
1 0 1 1 0 1
For m=1,2,…,M.
1:M matching: Conditional MLE
of Odds Ratio
• The probability of observing
1 0
m-1 M-m+1
pr (1 | 1, M , m, M m 1; )
1 M
1 m 1
1
1
u
M
m u
u
1 M
1 m 1
1
m
1
0
M
m 0
0 1 M
1 m 1
1
m M m 1
1:M matching: Conditional MLE
of Odds Ratio
• The probability of observing
0 1
m M-m
pr (0 | 1, M , m, M m 1; )
1 M
m 0
0
0
1
u
M
m u
u
1 M
m 0
0
M m 1
0
1
0
M
m 0
0 1 M
1 m 1
1
m M m 1
1:M matching : Conditional MLE
of Odds Ratio
• The conditional likelihood function of observing
all 2x2 tables is
m
n1m 1 n0 m
M
M m 1
CL( ) (2)
m1 m M m 1 m M m 1
1:M matching : Conditional MLE
of Odds Ratio
• The conditional MLE of the odds ratio, ˆ , is
obtained by maximizing (2) with respect to .
Therefore, ˆ is the solution of the equation
M M
n1m1 n0m m
n1m1 m M m 1
m 1 m1
M M
(M m 1)n1m1 / M 1 (M m 1)n1m1
ˆMH m 1
M
m 1
M
mn0m / M 1 mn0m
m 1 m 1
Variable number of controls per
case
• We will discuss this case in the context of
conditional logistic regression, which is the
topic of the lecture that follows.