STATISTICS AND RESEARCH DESIGN
Effect modification
James R. Miller
Indianapolis, Ind
I
n a previous article on confounding, we used a Table I. Occurrence of mandibular asymmetry among
hypothetical prospective cohort study to demon- girls, boys, and all participants from a study in which
strate how sex might confound the effect of condylar
sex produced confounding (hypothetical)
fracture on the development of mandibular asymmetry.1
The data from that study are shown in Table I. The data No
were stratified by sex, and the relative risk for all partic- Asymmetry asymmetry Total Risk RR
ipants was initially calculated by simply combining the Girls
data from the girls and boys, producing an unadjusted rela- Condylar fracture 10 40 50 0.2000 19.61
tive risk of 16.23. This unadjusted relative risk for all partic- No condylar fracture 50 4850 4900 0.0102
ipants was confounded by sex, which was evident because Boys
the relative risk for all participants did not fall in the range Condylar fracture 15 135 150 0.1000 19.61
established by the stratum-specific relative risks for girls No condylar fracture 25 4875 4900 0.0051
and boys. That is, 16.23 was not somewhere between All participants
19.61 and 19.61 for girls and boys, respectively. By statis- Condylar fracture 25 175 200 0.1250 16.23
tically adjusting for the confounding influence of sex, the No condylar fracture 75 9725 9800 0.0077
adjusted relative risk for all participants was 19.61, the only RR, Relative risk.
possible average of the 2 stratum-specific relative risks.
This hypothetical study was obviously contrived to
make it easier to understand confounding. In a real cohort study of children 8 to 11 years old. After further
study, it would be highly unlikely that the relative risks discussion, he decides to restrict the study to girls to
for girls and boys would be exactly the same. A real study avoid the issue of confounding by sex. However, the
would most likely have produced relative risks for girls investigator is concerned that age might be a
and boys that were different to a greater or lesser degree. confounding factor, so he decides to collect data from
If the stratum-specific results in a study are different, the girls in 2 age groups, 8 to 9 and 10 to 11 years old.
questions are: how different are they, and is this differ- After review and approval by an institutional review
ence important? board, the investigator selects a random sample of
These questions bring us to the topic of this article: 20,000 girls who were either 8, 9, 10, or 11 years old,
effect modification (sometimes called interaction).2,3 3 years ago, and who had no reference to mandibular
As with the previous article on confounding, let’s asymmetry in their dental records at that time. He then
begin with a hypothetical study to illustrate effect queried the database again to determine how many of
modification and then discuss it in more detail. We get these girls developed mandibular asymmetry in the
a call from another investigator who is interested in subsequent 3 years, leading to the present. The data for
our study, which was conducted on children who were this hypothetical study are presented in Table II.
7 years old at the beginning of the study. He wants to These data indicate that the stratum-specific relative
see whether the strong relationship we found between risks for the 2 age groups of girls are quite different. How-
condylar fracture and the development of mandibular ever, the relative risk for all girls in this study, obtained by
asymmetry exists in older children. This investigator is simply adding the data from the 2 strata, would appear to
from a country that has good dental records on all be a reasonable average of the stratum-specific relative
children, so he is considering doing a retrospective risks. Based on our previous discussion of confounding,
one might initially think that this unadjusted relative
State oral health director, Indiana State Department of Health, Indianapolis, Ind. risk for all participants is probably not confounded by
Address correspondence to: James R. Miller, Oral Health Program, 2-F, Indiana
State Department of Health, 2 North Meridian St, Indianapolis, IN 46204; age and might represent a good estimate of the true effect
e-mail, JaMiller@[Link]. of condylar fracture on the development of mandibular
Am J Orthod Dentofacial Orthop 2014;145:839-41 asymmetry among all girls in this study.
0889-5406/$36.00
Copyright Ó 2014 by the American Association of Orthodontists. However, at this point, one must ask: How likely is it
[Link] that these seemingly quite different stratum-specific
839
Downloaded from [Link] at Users for NHS Scotland October 01, 2016.
For personal use only. No other uses without permission. Copyright ©2016. Elsevier Inc. All rights reserved.
840 Statistics and research design
Table II. Occurrence of mandibular asymmetry among girls 8 to 9 and 10 to 11 years old, and all participants from a
study in which age produced effect modification (hypothetical)
Asymmetry No asymmetry Total Risk RR
Girls (age, 10-11 y)
Condylar fracture 40 60 100 0.4000 33.06
No condylar fracture 120 9780 9900 0.0121
Girls (age, 8-9 y)
Condylar fracture 20 80 100 0.2000 19.80
No condylar fracture 100 9800 9900 0.0101
All participants
Condylar fracture 60 140 200 0.3000 27.03
No condylar fracture 220 19,580 19,800 0.0111
RR, Relative risk.
relative risks are just random variations from a single true we discussed the possibility of a different third variable
relative risk for all participants? Could these divergent (sex) confounding this same effect. In these 2 articles, we
relative risks actually indicate that age somehow modifies covered 3 possibilities for a third variable. It can produce
the effect of condylar fracture on the development of neither effect modification nor confounding; it can pro-
mandibular asymmetry? In essence, one is asking whether duce only effect modification; or it can produce only con-
there could be different biologic responses to condylar founding. The fourth possibility is that it can produce both
fracture according to age. This question naturally arises: effect modification and confounding. For more informa-
How different do the stratum-specific relative risks need tion on this possibility, I refer you to the references. In
to be to indicate a possible effect modification? The each hypothetical cohort study, we discussed 1 third vari-
answer is that these relative risks need to be significantly able for the sake of clarity. In a real cohort study, an inves-
different from a statistical standpoint. Recognizing effect tigator might need to consider many “third variables”
modification is important because it indicates a biologic simultaneously; this is where multivariable statistical anal-
difference in effect according to the strata of the effect- ysis would be helpful. Also, if a randomized controlled trial
modifying variable. The stratum-specific relative risks were feasible, it would greatly reduce the possibility of
from Table II should be reported because of the likely ex- confounding, generally leaving only effect modification
istence of effect modification. to consider. However, using a randomized controlled trial
Generally speaking, one checks for effect modifica- to study the effect of condylar fracture on the development
tion before considering confounding. If the stratum- of mandibular asymmetry is not feasible. If an investigator
specific results are significantly different, then there actually wanted to study the influence of condylar fracture
might be effect modification, and the stratum-specific on the development of mandibular asymmetry, he would
results should be reported. If the stratum-specific re- likely use a case-control study design with multivariable
sults are not significantly different, one still should analysis instead of a cohort study because of the resources
consider the possibility of confounding. If the unad- needed to conduct a cohort study.
justed result for all participants, obtained by simply I was pleased to be asked to write these articles on ef-
combining the stratum-specific data, does not fall fect modification and confounding. Yes, these concepts
within the range of the stratum-specific results, then can be challenging. However, understanding clinical
confounding most likely exists. With confounding, research and being able to critically evaluate studies
one has a choice of calculating the adjusted result for are fundamental to good clinical practice. I encourage
all participants or just reporting the stratum-specific re- you to continue your journey for greater understanding.
sults. There are 2 reasons that one might report It is worth the effort for both you and your patients.
stratum-specific results: because they are different,
indicating probable effect modification; or because KEY POINTS
confounding most likely exists and one has chosen to
present the stratified results. 1. Observational studies, such as cohort and case-con-
In this article, we discussed the possibility of a third var- trol studies, can have variables that produce effect
iable (age) modifying the effect of an exposure variable modification, confounding, or both.
(condylar facture) on the occurrence of a malocclusion 2. Confounding distorts the results of a study because
(mandibular asymmetry), whereas in the previous article of the structure of the data, but this distortion can
June 2014 Vol 145 IssueDownloaded
6 American
from [Link] at Users forJournal of Orthodontics
NHS Scotland October 01, 2016. and Dentofacial Orthopedics
For personal use only. No other uses without permission. Copyright ©2016. Elsevier Inc. All rights reserved.
Statistics and research design 841
often be mitigated in either the design or the anal- REFERENCES
ysis phase of a study. 1. Miller JR. Confounding. Am J Orthod Dentofacial Orthop 2014;145:
3. Effect modification indicates an underlying biologic 706-8.
process that an investigator would want to discover 2. LaMorte WW. Confounding and effect measure modification.
and report, frequently in the form of stratum- Available at: [Link]
BS704-EP713_Confounding-EM/. Accessed June 18, 2014.
specific results. 3. Hennekens CH, Buring JE. Analysis of epidemiological studies:
4. Studies should be carefully evaluated before their evaluating the role of confounding. In: Mayrent SL, editor. Epidemiology
results are adopted into clinical practice. in medicine. Boston: Little, Brown and Company; 1987. p. 287-323.
American Journal of Orthodontics andfrom
Downloaded Dentofacial Orthopedics
[Link] June 2014 Vol 145 Issue 6
at Users for NHS Scotland October 01, 2016.
For personal use only. No other uses without permission. Copyright ©2016. Elsevier Inc. All rights reserved.