BMJ Publishing Group

Statistics Notes: Interaction Revisited: The Difference Between Two Estimates

Author(s): Douglas G. Altman and J. Martin Bland
Reviewed work(s):
Source: BMJ: British Medical Journal, Vol. 326, No. 7382 (Jan. 25, 2003), p. 219
Published by: BMJ Publishing Group
Stable URL: http://www.jstor.org/stable/25453491 .
Accessed: 27/11/2011 08:11

Your use of the JSTOR archive indicates your acceptance of the Terms & Conditions of Use, available at .
http://www.jstor.org/page/info/about/policies/terms.jsp

JSTOR is a not-for-profit service that helps scholars, researchers, and students discover, use, and build upon a wide range of
content in a trusted digital archive. We use information technology and tools to increase productivity and facilitate new forms
of scholarship. For more information about JSTOR, please contact support@jstor.org.

Digitization of the British Medical Journal and its forerunners (1840-1996) was completed by the U.S. National
Library of Medicine (NLM) in partnership with The Wellcome Trust and the Joint Information Systems
Committee (JISC) in the UK. This content is also freely available on PubMed Central.

BMJ Publishing Group is collaborating with JSTOR to digitize, preserve and extend access to BMJ: British
Medical Journal.

http://www.jstor.org
 Education and debate

Statistics Notes
Interaction revisited: the difference between two estimates
Douglas G Altaian, J Martin Bland

We often want to compare two estimates of the same two we can test Cancer Research
0.2206 (row 8). From these values the
UK Medical
quantity derived from separate analyses. Thus we might interaction and estimate the ratio of the relative risks
Statistics Group,
want to compare the treatment effect in
subgroups
in a (with confidence interval). The test of interaction is the Centre for Statistics
randomised as
two age groups. The term of to error: inMedicine,
trial, such for ratio d its standard z=-0.2726/
Institute for Health
such a is a test of interaction. In earlier Sta 0.2206=-1.24, which P=0.2 when we refer it to a
comparison gives Sciences, Oxford
tistics Notes we discussed interaction in terms of hetero table of the normal distribution. The estimated OX3 7LF

geneity of treatment effect1"3 Here we revisit interaction interaction effect is exp( - 0.2726)=0.76. (This value can Douglas G Altman
and consider the concept more also be obtained direcdy as 0.67/0.88=0.76.) The professor of statistics
generally. inm diane
The
comparison of two estimated
quantities, such as confidence interval for this effect is -0.7050 to 0.1598
means or on the log scale (row 9). Transforming back to the rela Department of
proportions, each with its standard error, is a Public Health
general method that can be applied widely. The two esti tive risk scale, we get 0.49 to 1.17 (row 12).There is thus Sciences,
mates no St George's
should be not obtained from the evidence to a different treatment effect
independent, good support
Hospital Medical
same are the results from in younger and older women.
individuals?examples School, London
in a randomised trial or from two The same is used for odds SW17 ORE
subgroups independ approach comparing
ent studies. The should be If the estimates ratios. means or coefficients is J Martin Bland
samples large. Comparing regression
are El and E2 with standard errors SE(El) and SE(is2), as there professor ofmedical
simpler is no transformation. The two esti statistics
log
then the difference
d=El-E2 has standard error mates must be independent: themethod should not be
' Correspondence to
SE(d)=^l[SE(Elf + SE(?2)2] (that is, the square root of the used to compare a subset with the whole group, or two D G Altman
sum of the squares of the separate standard errors). This estimates from the same
patients. doug altman
cancer org uk
formula is an of a well known relation that the There is limited power to detect interactions, even in
example
variance of the difference between two estimates is the a the results from several stud
meta-analysis combining
BMJ 2003,326 219
sum of the variances (here the variance is the ies. As this illustrates, even when the two
separate example
square of the standard error). Then the ratio z=d/SE(d) estimates and P values seem very different the test of

gives a test of the null hypothesis that in the population interaction may not be significant It is not sufficient for
the difference d is zero, by comparing the value of z to the relative risk to be significant in one subgroup and
the standard normal distribution. The 95% confidence not in another.
Conversely, it is not correct to assume
interval for the difference is d-1.96SE(d) to d+1.96SE(d). that when two confidence intervals
overlap
the two esti
We illustrated this for means and
proportions,3
mates are not significandy different6 Statistical analysis
although we did not show how to get the standard should be targeted on the question in hand, and not
error of the difference. Here we consider based on P values from
comparing comparing separate analyses.2
relative risks or odds ratios. These measures are
always
analysed on the log scale because the distributions of 1 Alunan DG, Matthews JNS Interaction 1 Heterogeneity of effects BMJ
the tend to be to normal 1996,313 486
log ratios those closer than of
2 Matthews JNS, Altman DG Interaction 2 Compare effect sizes not P
the ratios themselves. values BMJ 1996,313 808
3 Matthews JNS, Altman DG Interaction 3 How to examine heterogeneity
In a of non-vertebral fractures in ran
meta-analysis
BMJ 1996,313 862
domised trials of hormone the 4 Torgerson DJ, Bell-Syer SEM Hormone replacement therapy and
replacement therapy
estimated relative risk from 22 trialswas 0.73 (P=0.02) in prevention of nonvertebral fractures A meta-analysis of randomized
trialsJAMA 2001,285 2891-7
favour of hormone From 14 trials 5 Bland JM, Altman DG Loganthms BMJ 1996,312 700
replacement therapy.4
of women aged on average < 60 years the relative risk 6 Bland M, Peacock J Interpreting statistics with confidence Obstetrician
and Gynaecologist (in press)
was 0.67 (95% confidence interval 0.46 to 0.98; P=0.03).
From eight trials of women aged ^60 the relative risk
was 0.88 (0.71 to 1.08; P=0.22). In other words, in Calculations for comparing two estimated relative risks
women the estimated treatment benefit was a
younger 1
Group Group 2
33% reduction in risk of fracture, which was statistically
I RR
0.67 0.88
|
significant, compared with a 12% reduction in older (E2)
2_*log RR_-0.4005 (?,)_-01278
women, which was not But are the relative
significant 3 CI for
95% RR 0.46 to 0.98 O.f 1 to 1.08
risks from the subgroups significandy different from CI for log to to 0.0770
4_*95% RR_-0.7765 -0.0202_-0.3425
each other? We show how to answer this question of 0.4195
using 5_Width CI_0.7563 _
the summary data 0.1070
just quoted. 6_SE[=width/(2x1.96)]_0.1929_
Because the calculations were made on the scale, Difference between log relative risks
log
the two estimates is 7
comparing complex (see table). We d[=?i-5>] -0.4005-(-0.1278)=-0.2726
need to obtain the logs of the relative risks and their 8_SE(cO_V(0.19292+0.10702)=0.2206_
confidence intervals (rows 2 and 4).5As 95% confidence CI(flQ 9 -0.2726 ?1.96x0.2206
to
intervals are obtained as 1.96 standard errors either side _or-0.7050 0.1598_
of z=-0.2726/0.2206=-1.24
of the estimate, the SE of each log relative risk is 10_Test interaction_ (P=0.2)_
Ratio of relative risks (RRR)
obtained by dividing the width of its confidence interval II
RRR=exp(cf) exp(-0.2726)=0.76
by 2x1.96 (row 6). The estimated difference in log 12_CI(RRR)_exp(-0.7050)
to exp(0.1598), or 0.49 to 1.17
-
relative risks is d=E- E2= 0.2726 and its standard error
*Values obtained by taking natural logarithms of values on preceding row.

BMJ VOLUME 326 25 JANUARY 2003 bmj.com 219