American Journal of Epidemiology
ª The Author 2011. Published by Oxford University Press on behalf of the Johns Hopkins Bloomberg School of
Public Health. All rights reserved. For permissions, please e-mail: [email protected].
Vol. 173, No. 8
DOI: 10.1093/aje/kwq461
Advance Access publication:
February 22, 2011
Practice of Epidemiology
Past Injury as a Risk Factor: An Illustrative Example Where Appearances Are
Deceiving
Gavin M. Hamilton, Willem H. Meeuwisse, Carolyn A. Emery, Russell J. Steele, and Ian Shrier*
* Correspondence to Dr. Ian Shrier, Centre for Clinical Epidemiology and Community Studies, Lady Davis Institute for Medical
Research, Jewish General Hospital, McGill University, 3755 Ch. Côte Ste-Catherine, Montréal, Quebec, Canada H3T 1E2 (e-mail:
[email protected]).
Initially submitted June 19, 2010; accepted for publication November 30, 2010.
Previous injury is believed to be a causal risk factor for subsequent injury. Using empirical data on circus artists
(n ¼ 1,281 artists) between 2004 and 2008 in Montreal, Canada, as a motivating example, the authors use patient
vector plots to demonstrate that a bias away from the null must always occur in the typical analyses cited as
evidence (i.e., survival analysis, Poisson regression), except in the improbable context where all subjects have the
same inherent risk independent of previous injury. In addition, using simulated data, the authors demonstrate that
a simple method that conditions on the individual will approximate conclusions from more complex analytical
methods. By using the typical analysis of the authors’ empirical data, Kaplan-Meier curves and Cox regression
suggested increasing injury rates for both the second and third injuries compared with the first injury. However,
conditional analyses using a matched population (i.e., time to first, second, and third injuries among artists with 3 or
more injuries) showed that injury rates were unchanged for both the second and third injuries compared with the
first injury. These results suggest that previous injury should not be evaluated as a causal risk factor unless one
conditions on the individual in some way.
bias (epidemiology); causality; survival analysis; wounds and injuries
Abbreviations: CI, confidence interval; HR, hazard ratio; RR, risk ratio.
Authors of systematic reviews have suggested that a previous injury increases the risk of future injuries (1–3). Most
of the original research cited was conducted in prospective
cohort studies and compared the injury rate ratios among
those with and without a previous injury, as defined by selfreport at baseline. Five specific studies were found that followed subjects after an initial index injury and prospectively
calculated subsequent injury rates (4–8). Highlighting the 3
most recent studies, in 2001, Orchard (6) reported that previous injuries (either within an 8-week time interval or not)
increased the risk of sustaining a muscle strain to the same
location for the hamstrings, quadriceps, and calf muscles. In
2006, Hagglund et al. (7) studied 12 Swedish male football
teams over 2 seasons and found that players who were injured in the first study year were at greater risk of any injury
(i.e., their subsequent injury) in the following season compared with noninjured players (hazard ratio (HR) ¼ 2.7,
95% confidence interval (CI): 1.7, 4.3). In addition, the injury risk increased with the number of injuries that occurred
in the previous season. In 2007 in an observational cohort
study including 5 female high school sports over 3 years,
Rauh et al. (8) found that the risk of a second injury to the
same location (subsequent injury) was greater than the risk
of the index injury (risk ratio (RR) ¼ 2.8, 95% CI: 2.6, 2.9).
Similarly, the injury risk to a different body part was also
increased (RR ¼ 1.7, 95% CI: 1.6, 1.8).
There are 2 plausible theories to explain the relation between previous injury and subsequent injury risk: 1) a causality theory (1, 7, 9–11) and 2) a ‘‘noncausal marker’’
theory (7, 12). In the causality theory, inadequate rehabilitation leads to increased risk of the previously injured tissue
due to incomplete healing and weakness, and to increased
risk of injury to the surrounding and anatomically unrelated
areas due to altered movement patterns, loss of balance, or
941
Am J Epidemiol. 2011;173(8):941–948
942 Hamilton et al.
other functional/psychological impairments. If true, injury
risk would return to normal after adequate rehabilitation. In
the ‘‘noncausal marker’’ theory, the previous injury is simply a marker for other traits that would cause an individual
to be at a higher risk of injury. Certain individuals might be
at higher risk of injury due to genetic factors, risk-taking
behavior, or other nongenetic injury-prone characteristics
(e.g., training, playing position, psychological). If true, the
previous injury does not causally increase the risk of a subsequent or new injury but rather is simply a marker for
a common cause of both previous injury and subsequent/
new injury; that is, confounding bias is present for the parameter of an increased injury risk as defined by the structural approach to bias (13, 14).
Although complex statistical techniques exist to analyze
data that include repeated events such as injury (15–19), the
underlying assumptions are complex (20), it can be difficult
to estimate marginal distributions of event times (17), and
one cannot easily assess when important model assumptions
are violated. The purpose of this study is to 1) illustrate that
bias must occur in analyses commonly cited in the literature
(typical analyses) to infer that previous injury increases the
risk of subsequent injury and 2) demonstrate a simple and
transparent method of analysis that allows those without
specific expertise in advanced analysis of repeated events
to approximate the causal effect of previous injury on the
risk of subsequent injury (i.e., to minimize bias). We use
data from a population of individuals with multiple injuries
(a prospective cohort study on circus artists (21)) and compare typical analyses examining subsequent injury risk with
analyses where one conditions on the individual. Although
a formal statistical proof of the method is beyond the scope
of this paper, we have included some simulations in a Web
Appendix, which is posted on the Journal’s Web site (http://
aje.oxfordjournals.org/), that compare our method with one
of the more complex methods that have been proposed to
address this issue (17, 19).
MATERIALS AND METHODS
This paper describes analyses to determine if previous
injury increases the risk of subsequent injury. We first describe the data set we used and then describe the different
types of analyses.
The motivating example represents a secondary analysis of
data collected to report on injury patterns and injury rates of
circus performers, and the details of the cohort study are
presented elsewhere (21). One author’s institution stated that
formal ethics approval was not required, because historical
data were gathered as part of routine business practice and
therefore fell under a category of quality assurance. For the
other institution, the Office of Medical Bioethics provided
approval. Subjects were included if they were 1) a circus
artist between August 2004 and March 2008 and 2) had at
least 3 performances of exposure data recorded. The circus
company performs shows with artistic and gymnastic elements that are considered highly athletic in nature. Exposure
data were recorded electronically for each performance, and
all therapists used an electronic database as their sole chart-
ing procedure for injuries and associated treatments. Duplicate injuries were removed, exacerbations were considered
as part of the original injury, and an injury was considered
healed if the artist did not have a treatment for 3 months. If an
event resulted in multiple injuries (i.e., multiple locations or
multiple injury types), each injury was recorded separately.
An injury was included in the current analysis if it 1)
occurred during a performance within the study period, 2)
was reported to a therapist, and 3) resulted in a completely
missed subsequent performance (time loss injury). We restricted ourselves to time loss injuries because these are
considered more likely to be causal factors for subsequent
injury. Finally, an injury was defined as fully healed on the
date the artist returned to full participation (22, 23).
Concerning analyses, we first used survival analysis on
the empirical data to obtain the (biased) estimate that has
been reported more typically in injury epidemiology studies
(7). In brief, we considered the artist to have no previous
injuries when he/she entered the cohort and compared the
Kaplan-Meier curves and hazards using Cox regression if
the proportional hazards assumption was considered reasonable (R package, ‘‘survival version 2.35-7’’). The time to
first injury among all artists (those with no injuries were
censored at end of follow-up) was compared with the time
to second injury among all artists with 1 injury, time to third
injury among all artists with 2 injuries, and so on.
Next, we used a simple patient vector plot (24) to illustrate why these types of analyses (survival analysis or rate
ratios) introduce a large bias when estimating the causal
effect of previous injury on the risk of subsequent injury.
In brief, these typical analyses compare the injury risk of an
entire group with the injury risk after removing individuals
with a particular predisposition toward injury (e.g., excluding subjects with a very low risk of injury). To remove this
bias, one can simply compare injury risk within the individual (subset analysis). Simply put, one conducts survival
analyses but restricts the population to those individuals
with a specified number of injuries (e.g., 3 or more, 5 or
more, and so on). In this subset approach analysis, the subsample of subjects is the same for time to first, second, and
third injuries. Essentially, this matched analysis conditions
on the individual, and therefore noninjury individual predisposition to risk is controlled for (assuming it is constant
over time). The Web Appendix shows the comparison of our
results with one of the more complex methods proposed in
the literature (17, 19). Although both the subset approach
and the more complex method can be used to estimate time
to injury, it is not appropriate to directly compare the numerical results as they are not using the same data nor
comparing the same distributions.
Finally, we illustrate the effect of our subset approach
using the empirical injury data from circus artists. We conducted the subset approach restricted to individuals with 3 or
more injuries. We also conducted the subset approach on
subjects with 5 or more injuries (time to first, second, third,
fourth, and fifth injuries) and 2 or more injuries to explore
for potential biases due to conditioning on individuals with
only a specific number of injuries.
All statistical analyses were conducted by using R Foundation for Statistical Computing, version 2.7.1 (25).
Am J Epidemiol. 2011;173(8):941–948
Past Injury as a Risk Factor
943
100
Artists Without a Previous Injury
% of Individuals Without a Subsequent Injury
Artists With 1 Previous Injury
Artists With 2 Previous Injuries
80
Artists With 3 Previous Injuries
Artists With 4 Previous Injuries
60
40
20
0
0
200
400
600
800
1,000
Artist Exposures
Figure 1. A typically used analysis of previous injury as a risk factor for subsequent injury is presented by using empirical data obtained from
Cirque du Soleil (Montreal, Quebec, Canada) for circus artists between 2004 and 2008. The solid line represents the time to first injury for all artists.
The dashed line represents time to second injury for all artists who had 1 injury, the dotted line represents time to third injury for artists with 2 injuries,
and so on. The time to injury decreases with each injury, suggesting that previous injury is a risk factor for subsequent injury.
RESULTS
Of the 1,281 artists that met our inclusion criteria, 180
artists (14.1%) incurred 3 or more time-loss injuries, and 65
artists (5.1%) incurred 5 or more time-loss injuries. In total,
there were 1,230 time-loss injuries that resulted in 1 completely missed performance.
Figure 1 shows the comparison of time to injury according to typical analyses to determine if a previous injury
increases the risk of subsequent injury. As previous authors
have shown, the time to injury is decreased (risk was increased) among those with a previous injury compared with
those without a previous injury, and the time to injury decreases further (the Kaplan-Meier curves are steeper) as the
number of previous injuries increases. Using time to first
injury as a reference, we found increased risk comparing the
time to first injury with the time to second injury (HR ¼
1.99, 95% CI: 1.72, 2.30), as well as time to third injury
(HR ¼ 2.41, 95% CI: 2.03, 2.87).
Causal inferences on this analysis are problematic because each line represents different groups of subjects,
and therefore the different risk profiles may occur simply
because previous injury is a marker for some other common
Am J Epidemiol. 2011;173(8):941–948
cause of both previous injury and subsequent injury. Figure 2
uses a patient vector plot (24) to schematically illustrate that
typical analyses for estimating risk with repeated events
data (both in survival analysis and Poisson regression) are
always biased away from the null. We have created a patient
vector plot assuming that each subject has never been previously injured and has a constant injury rate that would be
unaffected by previous injury; that is, the time to injury is
identical for the first and all subsequent injuries. The parameter of interest is injury risk, and an unbiased estimate would
show that injury risk does not increase with previous injury.
In the typical analysis illustrated in Figure 1, the time to
injury for those with no previous injury is simply an average
of the time to injury across all individuals (Figure 2A).
However, the time to third injury requires that 2 previous
injuries have occurred, which means that subjects 1 and 2
are excluded from this analysis because they have 0 and 1
injury, respectively (Figure 2B). Therefore, the average of
the time to injury for subjects 3–6 must be shorter than for
subjects 1–6, because the individuals that have been eliminated are those with the longest time to injury. The same
principle is applied for the analysis on time to fifth injury,
and one now averages only subjects 5 and 6 (Figure 2C).
944 Hamilton et al.
A)
Subject 1
Subject 2
Subject 3
Subject 4
Subject 5
Subject 6
Time
B)
Subject 1
Subject 2
Subject 3
Subject 4
Subject 5
Subject 6
Time
C)
Subject 1
Subject 2
Subject 3
Subject 4
Subject 5
Subject 6
Time
Figure 2. A schematic diagram to demonstrate that typically used analyses are always biased away from the null for the causal effect of previous
injury on subsequent injury risk. Subjects are each represented by 1 horizontal line, their exposures (time at risk) by the line segments, and their
injuries by an ‘‘X.’’ The space between segments represents periods when the subject is not exposed as the result of injury. In each of the panels,
we have simulated data so that each subject has a fixed injury rate that is unaffected by previous injury; that is, the time to injury is identical for each
subject’s injuries. In the typical analyses, the time to injury for all individuals includes all subjects (A), and this is calculated by taking the average of
all subjects. However, when we compare the subjects’ times to third injury, we are using only a subset of the population (excluding the subjects
represented by dotted gray lines who had 2 or fewer injuries). Because the excluded subjects had the longest time to injury, the remaining included
subjects must always have a shorter time to injuries compared with the overall average (B). Therefore, one must see an overall reduced time to
injury even though previous injury does not actually increase the risk of subsequent injury. The same principle applies for the time to fifth injury (C).
Using Poisson regression does not avoid the bias, because
one is still comparing individuals with an inherently higher
injury risk with individuals with an inherently lower injury
risk; there is no comparison of change in risk within an
individual following an injury.
Figure 3 shows the comparisons in our empirical data
when we directly measure the time between each injury
(i.e., our parameter of interest) for those individuals with
3 or more injuries (Figure 3A) and 5 or more injuries (Figure
3B). Unlike Figure 1, the lines that denote the time to injury
are now practically superimposed for both Figure 3A and
Figure 3B. For artists with 3 or more injuries, the hazard
ratios were 1.08 (95% CI: 0.88, 1.33) for time to second
injury and 1.08 (95% CI: 0.88, 1.32) for time to third injury
(using time to first injury as reference). The hazard ratio for
artists with 5 or more injuries (Figure 3B) could not be
calculated as the curves crossed at several points and the
proportional hazards assumption was violated. As a sensitivity analysis, we also ran the analysis for artists with 2 or
more injuries and obtained the same result (data not shown).
Therefore, in our subset approach for the analysis, the risk of
an artist’s subsequent injury appeared to be the same as it
was before their previous injury. These results suggest that
the typical analysis using our data shown in Figure 1 is
biased with respect to estimating previous injury as a causal
risk factor, and that the observed differences in our population are due to noncausal factors. Of course, a comparison of
the values across Figure 3A and Figure 3B would show that
subjects with 5 injuries have greater risk of injury compared
with subjects with 3 injuries. However, in both groups, the
risk of injury returned to baseline, and there was no increased risk with previous injury.
Am J Epidemiol. 2011;173(8):941–948
Past Injury as a Risk Factor
A)
B)
100
100
% of Individuals Without a Subsequent Injury
% of Individuals Without a Subsequent Injury
Time to First Injury
Time to Second Injury
Time to Third Injury
80
60
40
20
0
945
Time to First Injury
Time to Second Injury
Time to Third Injury
Time to Fourth Injury
Time to Fifth Injury
80
60
40
20
0
0
200
400
600
800
1,000
0
Artist Exposures
200
400
600
800
1,000
Artist Exposures
Figure 3. In these analyses using empirical data obtained from Cirque du Soleil (Montreal, Quebec, Canada) for circus artists between 2004 and
2008 (the same as Figure 1), the population under study is restricted to artists with either 3 or more injuries (A) or 5 or more injuries (B). In A, the
solid line, dashed line, and dotted line represent the times to first, second, and third injuries, respectively. The superimposition of the curves
suggests that previous injury did not increase the risk of injury in these artists, and that the observed pattern in Figure 1 is due to bias. In B, a similar
representation is repeated for artists with 5 or more injuries. The superimposition of the curves again suggests that previous injury did not increase
the risk of injury in these artists.
To further illustrate the bias, we plotted the time to first
injury stratified by the number of injuries that a subject had
(i.e., a different line for subjects with a total of 2 injuries, 3
injuries, and 5 injuries) (Figure 4). Although the results
appear similar to those of Figure 1, they illustrate a very
different point. First, Figure 1 compares the time to first
injury among all subjects, the time to second injury among
artists with 1 injury, the time to third injury among artists
with 2 injuries, and so on. However, Figure 4 compares time
to first injury for each of the groups (artists with 1 injury,
artists with 2 injuries, and so on). The time to first injury for
subjects with 2 injuries was slightly longer than the time to
first injury for subjects with 3 injuries, which was longer
than the time to first injury for subjects with 5 injuries.
These results suggest that subjects who had incurred
a greater number of injuries during the entire study period
had a shorter time to their first injury, or increased risk of
their index injury. Because a subsequent injury after the
index injury cannot possibly cause an increased risk for
the index injury, the differences in time to first injury must
represent a bias. If previous injury did cause an increase in
subsequent injury risk, one would expect to observe shorter
Am J Epidemiol. 2011;173(8):941–948
times between injuries in the previously injured group, but
the time to first injury should have been the same.
DISCUSSION
Our results illustrate that typical epidemiologic analyses
examining previous injury as a causal risk factor must be
biased. Using empirical data, as well as simulated data
where time to injury is unaffected by previous injury, we
demonstrate how and why this bias occurs and that conditioning on the individual is one simple and transparent
method that can minimize the bias.
To date, authors have examined only previous injury as
a ‘‘risk factor’’ and have not adequately considered adjusting for the bias created when different individuals have
different predispositions toward injury. For example, in
2006, Hagglund et al. (7) published a study of elite football
in 12 Swedish male teams over 2 seasons and found that the
risk of an injury was higher for individuals with 1–2 injuries
in the previous season compared with uninjured players in
the previous season (HR ¼ 2.2, 95% CI: 1.4, 3.6). Those
946 Hamilton et al.
100
Artists With 2 Injuries
% of Individuals Without a Subsequent Injury
Artists With 3 Injuries
Artists With 5 Injuries
80
60
40
20
0
0
200
400
600
800
1,000
Artist Exposures
Figure 4. Kaplan-Meier survival curves depicting time to first injury for different artist populations using the same empirical data obtained from
Cirque du Soleil (Montreal, Quebec, Canada) for circus artists between 2004 and 2008 as for Figures 1 and 3. The time to first injury is compared
among different populations that were stratified by their number of injuries (i.e., artists with 2 injuries, artists with 3 injuries, and artists with 5
injuries). The difference in curves suggests that individuals with a greater number of injuries have a shorter time to their first injury. Because the
second injury occurs after the first injury, it cannot cause a reduced time to first injury. This supports the conclusion that the analysis in Figure 1 is
biased.
who had 3–4 injuries or 5 injuries in the previous season
had an even higher risk of injury compared with uninjured
players in the previous season (HR ¼ 3.0, 95% CI: 1.8, 5.3
or HR ¼ 5.2, 95% CI: 2.9, 9.0), respectively. Using the same
type of analysis for our empirical data on circus artists,
we obtained a similar result for time to second injury
(HR ¼ 1.99, 95% CI: 1.72, 2.30) and for time to third injury
(HR ¼ 2.41, 95% CI: 2.03, 2.87). These results confirm that
previous injury is associated with higher injury rates. However, minimizing the bias by restricting the population and
conditioning on the individual so that we compare the results
of each individual with him or herself, we demonstrated that
previous injury is not associated with an increased risk of
subsequent injury for an individual in our sample.
Although our patient-vector plot uses a static cohort with
no previous injury to demonstrate that bias must occur, the
results are generalizable to all situations. In a large dynamic
cohort, the specifics are different, but the principles the
same; there are many more individuals, individuals enter
and leave the cohort at different times, the duration of each
injury may be different, and there may be periods of non-
exposure due to illness or other noninjury reasons. However,
the typical analyses still eliminate subjects with lower fixed
injury rates, resulting in the appearance of increased injury
risk with previous injury. We considered the index injury in
our cohort to be the first injury that is observed in the data
obtained, and that all subjects were considered free of previous injury before then. All adults who participate in sport
have likely been injured at least once in their lifetime, so no
one is truly ‘‘injury free.’’ To account for previous injury,
some authors have restricted (or stratified) analyses to subjects who have had no injury within a certain period of time
(e.g., 1 year). This practice just eliminates (or stratifies)
those individuals with the highest fixed injury rates; for
example, subjects injured every 6 months never enter the
study, and the problems we outlined still exist for the remaining subjects (or within the stratified subgroups). There
are additional problems related to this type of left-truncating
of data (26, 27) and interpretations of hazard ratios in general (28) that are often unrecognized in the injury epidemiologic literature. Although important, these are beyond the
scope of this paper and do not affect the general principles
Am J Epidemiol. 2011;173(8):941–948
Past Injury as a Risk Factor
we have outlined. Finally, in the simulated data shown in the
Web Appendix where previous injury was a causal risk factor, the typical analyses would still overestimate the true
causal relation because the above biases are external to
any method of simulating the data and, therefore, continue
to be present.
Although our results strongly suggest that previous injury
was not a causal risk factor in our data (i.e., support the
‘‘noncausal marker’’ theory over the causality theory) and
that our explanation of the effect using patient vector plots
shows that bias must have been present in previous studies
and will in fact always be present in the typical analyses,
this does not mean that previous injury is never a causal risk
factor for subsequent injury. The circus company that provided the data is a unique setting, and the injury rates and
types of injuries may not be similar to injuries in other
sports. In addition, the artists are closely monitored after
an injury occurs, and rehabilitation is strongly encouraged.
In considering a dynamic model of etiology in sport injury
that defines injury risk as the cyclic nature of changing risk
factors (29), our results suggest only that, after an injury, the
circus artists studied here appear to return to the baseline
risk that they had before their injury. In a different population where appropriate rehabilitation does not occur, results
could indeed indicate that previous injury is a causal risk
factor. Therefore, we suggest that our method of analysis
could be used to determine if adequate rehabilitation has
occurred (i.e., a type of quality assurance) within large-scale
studies across leagues and sports. In addition, the same
method should be encouraged for any study with multiple
event outcomes because the same principles apply (e.g., recurrent exacerbations in patients with chronic obstructive
pulmonary disease). Finally, our subset approach is helpful
when investigators wish to establish a causal relation between previous injury and subsequent injury. In some contexts, authors might be concerned only about whether
a previous injury predicts subsequent injury even if it is not
causally related. Of course, our theoretical analysis using
patient vector plots illustrates that previous injury must
predict subsequent injury under realistic assumptions.
Therefore, these specific contexts should be limited to
a comparison of the magnitude of the prediction under different scenarios.
Limitations
Our analysis is based on the assumption that there are
different underlying risks for different individuals or groups
of individuals. We believe this to be valid because, in most
sports, different positions (e.g., a football lineman vs. a defensive safety) are associated with different injury risks, and
different players have different levels of aggressive play. If
one were able to precisely control for all of these covariates
(which we believe is impossible due to unknown effect
modifiers and measurement error), the typical analyses
would yield an unbiased result.
The proposed analysis can only be done if one has longitudinal information on each subject, including precise data
on all subsequent injuries. Thus, it cannot be used with injury surveillance methods that do not link injuries to specific
Am J Epidemiol. 2011;173(8):941–948
947
individuals over time. As such, our theoretical analysis demonstrates that examining the role of previous injury using
data that do not link injuries to specific individuals does not
provide useful information about previous injury as a causal
risk factor because the analysis will always show an
increased risk. These analyses should not be included in
articles using these types of data. In addition, the analysis
cannot include subjects who were censored at the time of the
last injury, or a bias will be introduced (17, 30). For
example, Figure 3A specifically includes time to third injury
only among subjects with 3 injuries. It specifically excludes
subjects with 2 injuries who would normally be censored but
included in a survival plot for time to third injury.
The purpose of this article was to clearly illustrate the
epidemiologic concept underlying the bias with the typical
analyses. Although we restricted our population to obtain
a ‘‘matched-analysis,’’ we did not actually account for the
matching in the statistical model, because we felt it would
unnecessarily introduce another level of complexity. This
means that our variance calculation (which assumed an unmatched population) would be expected to overestimate the
true variance of the matched population, just as using an
unpaired t test with matched data would usually (but not
always) overestimate the variance. Although variance estimates in our analysis are inaccurate, the hazard ratio point
estimates, Kaplan-Meier curves, and the interpretation of
the results would not change. Using one of the more complex survival-analysis-for-repeated-events-data approaches
such as the one described in the Web Appendix is possible,
but these require specialized expertise to verify the underlying assumptions of the model and to include other limitations (15, 16, 20, 28). An alternative variance calculation for
our subset approach that might be more accessible is a bootstrap method (31, 32). We believe that the rough approximation obtained using our method is acceptable given that it
will be accessible to a much greater number of researchers
compared with the more complex methods.
Finally, our analyses were restricted to time-loss injuries
for performances in artists with at least 3 exposures and
defined an injury to be fully healed when the artist returned
to full participation, which may affect the generalizability
of the results. In addition, exposure information for
performances was the only exposure information recorded.
Although the numerical results would be different for other
injury definitions (e.g., medical attention injuries) and fully
healed injuries, as well as different levels of exposure (performance vs. training), this would not affect the conclusion
that the typical analysis is always biased away from the null.
Conclusion
Using both simulated and empirical data, we have demonstrated that the typical analysis examining previous injury
as a causal risk factor will always bias the results away from
the null unless it were possible to precisely measure and
condition on all other injury risk factors in the statistical
model. Although previous injury was not a causal risk factor
for subsequent injury in our empirical data, it may still be
a causal factor in other studies. Researchers who are interested in the causal effects of previous injury need to conduct
948 Hamilton et al.
some type of matched analysis in order to obtain approximate estimates of effect.
ACKNOWLEDGMENTS
Author affiliations: Roger Jackson Centre for Health and
Wellness Research, Faculty of Kinesiology, University of
Calgary Sport Medicine Centre, Calgary, Alberta, Canada
(Gavin M. Hamilton, Willem H. Meeuwisse, Carolyn A.
Emery); Department of Community Health Sciences, Faculty of Medicine, University of Calgary, Calgary, Alberta,
Canada (Willem H. Meeuwisse, Carolyn A. Emery);
Department of Pediatrics, Faculty of Medicine, Alberta
Children’s Hospital, Calgary, Alberta, Canada (Carolyn A.
Emery); Department of Mathematics and Statistics, McGill
University, Montreal, Quebec, Canada (Russell J. Steele);
and Centre for Clinical Epidemiology and Community
Studies, Jewish General Hospital, McGill University,
Montreal, Quebec, Canada (Ian Shrier).
This work was supported by the Canadian Academy of
Sport and Exercise Medicine. I. S. was supported by the
Senior Clinician Scientist Award from the Fonds de la
Recherche en Santé du Quebec. C. E. was supported by
a Population Health Investigator Award from the Alberta
Heritage Foundation for Medical Research, a New Investigator Award from the Canadian Institutes of Health Research, and a Professorship in Pediatric Rehabilitation
funded through the Alberta Children’s Hospital Foundation.
I. S. is also the consulting medical director for the circus
company, which could be considered to represent an apparent conflict of interest.
REFERENCES
1. Murphy DF, Connolly DA, Beynnon BD. Risk factors for
lower extremity injury: a review of the literature. Br J Sports
Med. 2003;37(1):13–29.
2. Emery CA. Risk factors for injury in child and adolescent
sport: a systematic review of the literature. Clin J Sport Med.
2003;13(4):256–268.
3. Maffey L, Emery C. What are the risk factors for groin strain
injury in sport? A systematic review of the literature. Sports
Med. 2007;37(10):881–894.
4. Mueller FO, Blyth CS. North Carolina high school football
injury study: equipment and prevention. J Sports Med. 1974;
2(1):1–10.
5. Rauh MJ, Margherita AJ, Rice SG, et al. High school cross
country running injuries: a longitudinal study. Clin J Sport
Med. 2000;10(2):110–116.
6. Orchard JW. Intrinsic and extrinsic risk factors for muscle
strains in Australian football. Am J Sports Med. 2001;
29(3):300–303.
7. Hägglund M, Waldén M, Ekstrand J. Previous injury as a risk
factor for injury in elite football: a prospective study over two
consecutive seasons. Br J Sports Med. 2006;40(9):767–772.
8. Rauh MJ, Macera CA, Ji M, et al. Subsequent injury patterns
in girls’ high school sports. J Athl Train. 2007;42(4):486–494.
9. Ekstrand J, Gillquist J. The avoidability of soccer injuries. Int
J Sports Med. 1983;4(2):124–128.
10. Arnason A, Gudmundsson A, Dahl HA, et al. Soccer injuries
in Iceland. Scand J Med Sci Sports. 1996;6(1):40–45.
11. Waldén M, Hägglund M, Ekstrand J. Injuries in Swedish elite
football—a prospective study on injury definitions, risk for
injury and injury patterns during 2001. Scand J Med Sci
Sports. 2005;15(2):118–125.
12. Taimela S, Kujala UM, Osterman K. Intrinsic risk factors and
athletic injuries. Sports Med. 1990;9(4):205–215.
13. Pearl J. Causality: Models, Reasoning and Inference. Cambridge,
United Kingdom: Cambridge University Press; 2000.
14. Hernán MA, Hernández-Dı́az S, Robins JM. A structural approach to selection bias. Epidemiology. 2004;15(5):615–625.
15. Box-Steffensmeier JM, De Boef S. Repeated events survival
models: the conditional frailty model. Stat Med. 2006;
25(20):3518–3533.
16. Kelly PJ, Lim LL. Survival analysis for recurrent event data:
an application to childhood infectious diseases. Stat Med.
2000;19(1):13–33.
17. Lawless JF, Cook RJ. Analysis of gap times. In: The Statistical
Analysis of Recurrent Events. New York, NY: Springer Verlag;
2007: 121–159.
18. Clement DY, Strawderman RL. Conditional GEE for recurrent
event gap times. Biostatistics. 2009;10(3):451–467.
19. Lin DY, Sun W, Ying Z. Nonparametric estimation of the gap
time distributions for serial events with censored data. Biometrika. 1999;86(1):59–70.
20. Cain LE, Cole SR. Survival analysis for recurrent event data:
an application to childhood infectious diseases. Stat Med.
2006;25(8):1431–1433;author reply 1433.
21. Shrier I, Meeuwisse WH, Matheson GO, et al. Injury patterns
and injury rates in the circus arts: an analysis of 5 years of data
from Cirque du Soleil. Am J Sports Med. 2009;37(6):1143–1149.
22. Fuller CW, Ekstrand J, Junge A, et al. Consensus statement on
injury definitions and data collection procedures in studies of
football (soccer) injuries. Clin J Sport Med. 2006;16(2):
97–106.
23. Fuller CW, Molloy MG, Bagate C, et al. Consensus statement
on injury definitions and data collection procedures for studies
of injuries in rugby union. Clin J Sport Med. 2007;17(3):
177–181.
24. Boers M, Berkhof J, Twisk JW, et al. A new graph and scoring
system simplified analysis of changing states: disease remissions in a rheumatoid arthritis clinical trial. J Clin Epidemiol.
2010;63(6):633–637.
25. R Development Core Team. R: a language and environment
for statistical computing. Vienna, Austria: R Foundation for
Statistical Computing; 2007. (http://www.R-project.org).
26. Guo G. Event-history analysis for left-truncated data. Sociol
Methodol. 1993;23:217–243.
27. Lawless JF, Fong DY. State duration models in clinical and
observational studies. Stat Med. 1999;18(17-18):2365–2376.
28. Hernán MA. The hazards of hazard ratios. Epidemiology.
2010;21(1):13–15.
29. Meeuwisse WH, Tyreman H, Hagel B, et al. A dynamic model
of etiology in sport injury: the recursive nature of risk and
causation. Clin J Sport Med. 2007;17(3):215–219.
30. Pena EA, Strawderman RL, Hollander M. Nonparametric
estimation with recurrent event data. J Am Stat Assoc. 2001;
96(456):1299–1315.
31. De Angelis D, Young GA. Bootstrap method. In: Armitage P,
Colton T, eds. Encyclopedia of Biostatistics. New York, NY:
John Wiley & Sons; 2005:426–433.
32. Shrier I, Steele RJ, Hanley J, et al. Analyses of injury count
data: some do’s and don’ts. Am J Epidemiol. 2009;170(10):
1307–1315.
Am J Epidemiol. 2011;173(8):941–948