Sei sulla pagina 1di 22

INTERACTION (EFFECT MODIFICATION)

Interaction between two variables is said to exist when the association


between two variables (generally measured by the odds ratio or relative
risk) is different at different levels of a third variable.
For example, the odds ratio that measures the association between
cigarette smoking and lung cancer may be smaller among individuals
who consume large quantities of beta carotene in their food when
compared to the analogous odds ratio among persons who consume
little or no beta carotene in their food.
There is a chi-squared test that assesses heterogeneity in the odds ratios
that is produced with the output from the casecont module in PEPI.

EXAMPLE OF INTERACTION IN FOURFOLD TABLE:


(DATA AGGREGATED FROM SAHAI AND KHURSHID, TABLE
6.5)
MI

140 mmHg

Yes
No
Total

290
7110
7400

SBP
< 140 mmHg

Total

270
12,440
12,710

560
19,550
20,110

OUTPUT FROM CASECONT (NOTE THAT USE OF


CASECONT NOT LIMITED TO CASE-CONTROL
STUDIES)
CASECONT - Analysis of 2 X 2 Tables for Case-Control Studies

DATA

Cases
Controls

Exposed
290
7110

TABLE 1
Not exposed
270
12440

ANALYSIS OF TABLE 1:
Total cases = 560
Total controls = 19550
Proportion of cases exposed = 0.518
0.364

Chi-square (1 DF)

Proportion of controls exposed =

= 55.641

P = 0.000

[ 8.70E-14 ]

Continuity corrected chi-sq. (Yates) = 54.980

P = 0.000

[ 1.22E-13

Upton's adjusted chi-square

P = 0.000

[ 8.71E-14

]
= 55.638

]
Odds ratio = 1.88 [Low-bias indicator of
1.87]
90% confidence interval
95% confidence interval
99% confidence interval
Adjusted O.R. (0.5 added in each cell)

O.R. in the population =


=
=
=
=

1.63 to 2.17
1.58 to 2.23
1.50 to 2.35
1.88

Yule's Q = 0.31
Phi = 0.05
Lambda (prediction of exposure status from "caseness") = 0.00
(prediction of "caseness" from exposure status) = 0.00

PREVIOUS DATA STRATIFIED BY AGE (SAHAI AND KHURSHID,


TABLE 6.5)

Data Showing Relationship Between Systolic Blood Pressure (SBP) and


Incidence of Myocardial Infarction (MI) Stratified by Age (< 60 / 60
Age 60
MI

140 mmHg

Yes
No
Total

OR=.95

SBP
< 140 mmHg

Total

90
1150

60
730

150
1880

1240

790

2030

Age < 60

MI

140 mmHg

Yes
No
Total

OR=1.87

SBP
<140 mmHg

Total

200
5960

210
11,710

410
17,670

6160

11920

18,080

Output From CASECONT: AGE60

DATA

Cases
Controls

Exposed
90
1150

TABLE 1
Not exposed
60
730

ANALYSIS OF TABLE 1:
Total cases = 150
Total controls = 1880
Proportion of cases exposed = 0.600
0.612

Proportion of controls exposed =

Chi-square (1 DF)

Continuity corrected chi-sq. (Yates) =


Upton's adjusted chi-square
=

Odds ratio = 0.95


0.94]
90% confidence
95% confidence
99% confidence
Adjusted O.R.

0.038
0.080

0.080

P = 0.777

P = 0.845
P = 0.777

[Low-bias indicator of O.R. in the population =

interval
interval
interval
(0.5 added in each cell)

=
=
=
=

0.71 to 1.28
0.67 to 1.36
0.60 to 1.51
0.95

Yule's Q = -0.02
Phi = -0.01
Lambda (prediction of exposure status from "caseness") = 0.00
(prediction of "caseness" from exposure status) = 0.00

Output From CASECONT: AGE<60


DATA

Cases
Controls

Exposed
200
5960

TABLE 2
Not exposed
210
11710

ANALYSIS OF TABLE 2:
Total cases = 410
Total controls = 17670
Proportion of cases exposed = 0.488
0.337

Chi-square (1 DF)

Proportion of controls exposed =

= 40.410

P = 0.000

Continuity corrected chi-sq. (Yates) = 39.743

P = 0.000

[ 2.06E-10 ]
[2.90E-10

Upton's adjusted chi-square

P = 0.000

[2.06E-10

]
= 40.408

Odds ratio = 1.87 [Low-bias indicator of O.R. in the population =


1.86]
90% confidence
95% confidence
99% confidence
Adjusted O.R.

interval
interval
interval
(0.5 added in each cell)

=
=
=
=

1.58 to 2.22
1.53 to 2.29
1.44 to 2.43
1.87

Yule's Q = 0.30
Phi = 0.05
Lambda (prediction of exposure status from "caseness") = 0.00
(prediction of "caseness" from exposure status) = 0.00

SUMMARY ANALYSIS OF TABLES 1 to 2

Mantel-Haenszel chi-square (DF = 1)


[ 1.22E-07 ]
continuity corrected (DF = 1)

= 27.987

= 27.512

P = 0.000

P = 0.000
[ 1.56E-07

Mantel-Haenszel odds ratio


90% confidence interval
95% confidence interval
99% confidence interval

=
=
=

1.57

1.36 to 1.81
1.32 to 1.86
1.25 to 1.96

Maximum-likelihood estimate of uniform odds ratio


90% confidence interval (Cornfield-Gart)
95% confidence interval (Cornfield-Gart)
99% confidence interval (Cornfield-Gart)

=
=
=
=

1.59
1.37 to 1.85
1.33 to 1.90
1.26 to 2.01

Heterogeneity of O.R.'s: chi-sq (DF: 1) = 11.57


P = 0.001 [ 6.69E04 ]
Standardized rate ratio (standard: exposed group)
= 1.44
Standardized rate ratio (standard: unexposed group)
= 1.72

Effect Modification/Interaction vs. Confounding


In the presence of confounding, the association between
exposure and disease is the same, or similar, at each
level of the third variable, but the crude and adjusted
OR /RR differ.
In contrast, when effect modification is present, the
association between each exposure and outcome is
different for different strata. The difference can be in
direction or magnitude .

Statistical testing can be used to assess whether effect


modification is present. The null hypothesis is stated in terms
of homogeneity of the relative risks or odds ratios.
H0: Stratum specific OR similar to summary OR
H1: Stratum specific OR not similar to summary OR
If the number of strata is large and the data are thinly spread
out, this statistic will not approximate Chi square.

Test for Homogeneity of Stratified Estimates


(see Appendix C p. 463-464)
General form

R R
k i

k21 =
Vi
i =1

Ri=Stratum specific measure, i.e OR, RR

R = Estimated common measure (RR OR etc)


Vi=variance of measure
K=number of strataChi square has k-1 degrees of freedom

For OR:

log
OR

log
OR
i

k21 =
Var (log ORi )
i =1

Log

OR =Summary ORMH

1 1 1 1
Var (log OR)= + + +
a b c d

Assessing Confounding and Effect Modification


CONFOUNDING
Compare crude vs.
adjusted OR or RR
No statistical testing*

EFFECT
MODIFICATION
Compare stratum specific
OR or RR
Statistical testing

*However, if the adjusted OR/RR is no longer significantly


different from 1.0 after adjusting for confounding, than it can
be said that the exposure /disease association was confounded
by the third factor.

Assessing effect modification requires looking at


stratum specific OR or RR, not comparing crude and
adjusted OR and RR.
If effect modification is present, it is inappropriate to
use the adjusted measure; the stratum specific estimates
should be used instead.

In the presence of confounding, an adjusted measure


appropriately controls for a third factor, and focuses on
the uniform (or close to uniform) impact that the
exposure is having on the outcome of interest.
In contrast, when effect modification exists, the third
factor is critical to understanding the differential
impact that the exposure is having on the outcome of
interest.

ILLUSTRATION OF DISTINCTION BETWEEN


CONFOUNDING AND INTERACTION
GI Upset

No GI Upset

Total

Ate Potato
Salad

35

65

100

Did Not Eat


Potato Salad

55

60

Total

40

120

160

OR=5.92

PEOPLE ATE AT TWO SHIFTS: IS SHIFT A


CONFOUNDER?

Yes

Developed GI Upset
No

Total

Yes

32

48

80

No

72

80

Total

40

120

160

Yes

Ate Potato Salad


No

Total

Yes

50

30

80

No

50

30

80

Total

100

60

160

Ate at Shift 2

OR = 6.00

Ate at Shift 2

OR = 1.00

IS THERE AN INTERACTION BETWEEN EATING AT


SHIFT 2 AND THE ASSOCIATION BETWEEN EATING
POTATO SALAD AND OCCURRENCE OF GI UPSET?
Shift 1 Eaters
GI Upset

No GI Upset

Total

Ate Potato
Salad

45

50

Did Not Eat


Potato Salad

27

30

Total

72

80

GI Upset

No GI Upset

Total

Ate Potato
Salad
Did Not Eat
Potato Salad

30

20

50

28

30

Total

32

48

80

OR=1.00

Shift 2 Eaters

OR = 21.00

CASECONT - Analysis of 2 X 2 Tables for Case-Control Studies


Sunday, 8th September 2002.
-------------------------------------------------------------------------

DATA

Cases
Controls

Exposed
5
45

TABLE 1
Not exposed
3
27

-------------------------------------------------------------------------ANALYSIS OF TABLE 1:
Total cases = 8
Total controls = 72
Proportion of cases exposed = 0.625
0.625

Proportion of controls exposed =

Chi-square (1 DF)
= 0.000
P = 1.000
Continuity corrected chi-sq. (Yates) = 0.000
P = 1.000
Upton's adjusted chi-square
= 0.000
P = 1.000
** WARNING: 1 cell has an expected frequency of <5.
Odds ratio = 1.00 [Low-bias indicator of O.R. in the population =
0.73]
Adjusted O.R. (0.5 added in each cell) = 0.95
Yule's Q = 0.00
Phi = 0.00
Lambda (prediction of exposure status from "caseness") = 0.00
(prediction of "caseness" from exposure status) = 0.00
------------------------------------------------------------------------DATA

Cases
Controls

Exposed
30
20

TABLE 2
Not exposed
2
28

-------------------------------------------------------------------------ANALYSIS OF TABLE 2:
Total cases = 32
Total controls = 48
Proportion of cases exposed = 0.938
0.417

Proportion of controls exposed =

Chi-square (1 DF)
= 22.222
]
Continuity corrected chi-sq. (Yates) = 20.056
]
Upton's adjusted chi-square
= 21.944
]

P = 0.000

[ 2.43E-06

P = 0.000

[ 7.52E-06

P = 0.000

[ 2.81E-06

Odds ratio = 21.00 [Low-bias indicator of O.R. in the population =


13.33]
90% confidence interval (approximate)
= 5.04 to 115.43
95% confidence interval (approximate)
= 4.12 to 143.76
99% confidence interval (approximate)
= 2.82 to 211.82
Adjusted O.R. (0.5 added in each cell) = 16.96
Yule's Q = 0.91
Phi = 0.53
Lambda (prediction of exposure status from "caseness") = 0.27
(prediction of "caseness" from exposure status) = 0.31
-------------------------------------------------------------------------SUMMARY ANALYSIS OF TABLES 1 to 2
Mantel-Haenszel chi-square (DF = 1)

= 15.960

P = 0.000

[ 6.47E-05

continuity corrected (DF = 1)

= 14.404

P = 0.000

[ 1.48E-04

]
]
Mantel-Haenszel odds ratio
90% confidence interval
95% confidence interval
99% confidence interval

=
=
=
=

5.57
2.49 to 12.49
2.13 to 14.57
1.57 to 19.71

Unconditional max. likelihood estimate of odds ratio


90% confidence interval (Cornfield-Gart)
=
95% confidence interval (Cornfield-Gart)
=
99% confidence interval (Cornfield-Gart)
=

= 7.38
2.76 to 21.04
2.36 to 24.71
1.75 to 33.32

Heterogeneity of O.R.'s: chi-sq (DF: 1) = 10.82


P = 0.001 [ 1.00E03 ]
Standardized rate ratio (standard: exposed group)
= 5.44
Standardized rate ratio (standard: unexposed group)
= 7.15
-------------------------------------------------------------------------------

Potato Salad-GI Upset OR's by Shift


25

21.0

Odds Ratio

20
15
10

5.9

5.6

1.0
0
1

E
D
UD
FIE
R
T
C
R
ST

Shift

EXAMPLE WITH MORE THAN TWO STRATA


Table 7-2 Szklo p. 262
Case control study of OC use and Myocardial Infarction in
Women

OC
No OC

Cases
29
205
234

Controls
135
1607
1742

Crude OR=1.7
Age associated with OC use, and age associated with MI
Stratify by age 25-29, 30-34, 35-39, 40-44, 45-49
Stratum specific OR
25-29: 7.2
30-34: 8.9
35-39: 1.5
40-44: 3.7
45-49: 3.9
Is there significant interaction (effect modification)?
(Are the stratum specific OR significantly different from one
another?)

Szklo p. 464.
Chi square with 4 degrees of freedom (5 stratum -1)=6.113
P>.10
No significant interaction. Can use adjusted OR. However
The P value from the test for homogeneity is strongly
dependent on sample size. When data are stratified, some
tables are sparse.
Since the OR in the younger women (25-34) appear to be much
higher than those in older women (35-49), it may be
appropriate to calculate two separate OR for the younger and
older women to reflect this difference. It may not be
appropriate to use the adjusted average OR of 3.97. This
becomes more salient when issues of prevention are involved.

Potrebbero piacerti anche