AMERICAN JOURNAL OF INDUSTRIAL MEDICINE 53:352–360 (2010)
Follow-Up of Neck and Shoulder Pain
Among Sewing Machine Operators:
The Los Angeles Garment Study
Pin-Chieh Wang, MS, PhD,1,2Robert J. Harrison, MD, MPH,3Fei Yu, PhD,4
David M. Rempel, MD, MPH,5and Beate R. Ritz, MD, PhD1?
reported neck/shoulder pain in sewing machine operators.
were immigrants. We examine the influence of individual and work-related factors
on changes in neck/shoulder pain during follow-up employing linear mixed models
with time-spline functions.
Results We observed a dramatic decline (72%) in self-reported pain intensity in the
first month of follow-up, followed by a small increase from the first to fourth month (4%
per month). Workers who perceived and reported their physical workload as high or
worked overtime experienced less overall pain reduction. Higher baseline pain intensity,
being of Hispanic ethnicity (vs. Asian), and taking cumulative daily rest time during work
in the first month, but not thereafter.
Conclusion Our findings indicate that some work-related factors may be of clinical
relevance for reducing neck/shoulder pain. Having lower physical workloads and less
overtime work should be considered when treating patients or planning workplace
interventions for managing work-related musculoskeletal disorders in this underserved
immigrant population. Am. J. Ind. Med. 53:352–360, 2010. ? 2009 Wiley-Liss, Inc.
KEY WORDS: immigrant; sewing machine; garment industry; recovery; neck and
shoulder pain; work organization; musculoskeletal disorder; ergonomic; psychosocial;
sewing machine operator; garment industry
California is the home to the largestgarment production
center in the United States, with the majority of the garment
shops located in the Los Angeles Basin. Altogether these
shops employ over 144,000 sewing machine operators, the
majority of whom are minimum-wage, unrepresented,
immigrant women [Chan et al., 2002]. Our previous study
found high prevalence of work-related musculoskeletal pain
for this population [Wang et al., 2007]. In summary, the
prevalence of moderate or severe musculoskeletal pain in
the neck/shoulder region was 24.0% and for distal upper
? 2009 Wiley-Liss,Inc.
1Department of Epidemiology, School of Public Health, University of California, Los
4Department of Biostatistics, School of Public Health, University of California, Los
5Division of Occupational and Environmental Medicine,Department of Medicine,Univer-
Contract grant sponsor: Centers for Disease Control/National Institute for Occupational
*Correspondence to:Beate R.Ritz,Department of Epidemiology,UCLA, School of Public
extremity it was 15.8%. Elevated prevalence of upper body
pain was associated with age less than 30 years, female
gender, Hispanic ethnicity, being single, having a diagnosis
of a musculoskeletal disorder or a systemic illness, working
more than 10 years as a sewing machine operator, using a
single sewing machine, work in large shops, higher work–
rest ratios, high physical exertion, high physical isometric
loads, high job demand, and low job satisfaction.
The persistency of and recovery from work-related
musculoskeletal pain may depend on factors similar to or
different from those contributing to the occurrence of work-
related musculoskeletal disorders (WMSDs). Most studies
published in the WMSDs literature have addressed occur-
on the recovery from pain symptoms or illness related to
WMSDs. Factors contributing to WMSDs recovery or lack
thereof are largely unknown, and existing studies thus far
have focused on recovery from low back pain.
Work tasks and physical activity modifications such as
task reallocation and recreational physical activity may be
important for improvement of neck and arm pain [Jonsson
et al., 1988]. There is also evidence that older age may be
associated with slow recovery from shoulder disorders and
that psychosocial work-related factors including perception
delay recovery from shoulder pain [Bonde et al., 2003].
Furthermore, modeling and describing typical recovery
patterns may encourage more realistic expectations of
recovery in patients, thereby alleviating anxiety about the
length of recovery time [Ferguson et al., 2001]. Knowledge
about individual and work-related factors that can influence
recovery from pain may be useful to both general medicine
treatment and intervention strategies to specific patients.
Previously we presented findings from a randomized
(chairs) to garment workers and showed how they influenced
neck/shoulder pain over a 4-month period [Rempel et al.,
shoulder pain intensity, while controlling for our interven-
tions. To accomplish this, we constructed a measure of
proportional change of pain intensity over a 4-month period
and assessed how it was influenced by demographic/
individual factors (including age, gender, ethnicity, and a
history of musculoskeletal problems) and/or work-related
of daily rest time during work, and psychosocial factors (job
strain, social support, job dissatisfaction).
Thesourceforthiscohortofworkersreporting neck and
we enrolled sewing machine operators from 13 garment
shops in Los Angeles, California. All employees were
eligible for participation if they performed sewing machine
operations for more than 20hr per week, were not in a
probationary period, did not have an active workers’
compensation claim, had worked for at least 3 months, and
were notplanning toquit their jobs withinthe next6months.
Between October 2003 and April 2005, 560 subjects
completed the baseline interview. Within 1–2 months from
the baseline interview, 480 (86.5%) of the subjects from
11 shops remained in the study and we randomly assigned
them to one of the three interventions: (1) a basic
modification group received a footrest, a small table-top
storage box, a side table, a task lamp, and reading glasses;
(2) an intermediate modification group received a conven-
tional height adjustable task chair that can swivel in addition
to the items given to the first group; and (3) the best
modification group received an ergonomic chair custom
designed for sewing machine operators with a curved seat
these 480 subjects, 277 subjects self-reported neck/shoulder
pain at baseline interview. A total of 30 subjects withdrew
before receiving their assigned intervention package leaving
us with 247 subjects who stayed in the study follow-up and
responded to at least one follow-up survey.
In our previous study [Rempel et al., 2007; Wang et al.,
2008], we demonstrated that neck/shoulder and back/hip
pains were reduced in both the intermediate and best
modification groups compared to the basic intervention
group, and that the interventions provided to the best
modifications group were more effective in reducing neck/
shoulder pain than those given to the intermediate inter-
In accordance with National Institutes of Health (NIH)
policy, approval for all study procedures was obtained from
the University of California, Los Angeles (UCLA) and
the Committee on Human Research at the University of
California, San Francisco (UCSF), and all participants
provided written informed consent.
The baseline interview was conducted for all subjects
1–2 months ahead of the implementation of our interven-
tions. After randomization, we administered four follow-up
standardized surveys spaced 1 month apart, thus, we
followed most workers over a 4-month period. During the
follow-up, 62 monthly interviews (5% of all monthly inter-
views) could not be conducted (Fig. 1). Specifically, there
were 7 subjects who missed the interview in the first month,
Follow-Up of Neck and Shoulder Pain353
18 in the second month, 13 in the third month, and 32 in the
During the baseline interview we collected information
about demographic factors (gender, age, ethnicity, weight,
height, medical history of musculoskeletal problems, and
smoking behavior) and work-related psychosocial factors.
Data on work-related ergonomic factors, such as number of
sewing tasks performed, number of sewing machines
operated, number of workdays per week, total duration of
daily rest time during work, and self-reported musculoske-
letal pain (pain frequency and pain intensity), were collected
at baseline and again each month during follow-up. The
assessment of psychosocial factors has been described in
detail in a previous publication [Wang et al., 2007]. Briefly,
of questions selected from Karasek’s Job Content Question-
naire (JCQ) [Karasek, 1997]. Answer options ranged from
1 (‘‘strongly disagree’’) to 4 (‘‘strongly agree’’). The total
item scores for all questions related to that factor using the
from 0 to 100. To implement Karasek’s demand-control
model, as recommended, we derived a variable for ‘‘job
strain’’ from the factors ‘‘job demand’’ and ‘‘job control,’’
adopting a linear function by subtracting the score of job
control from job demands such that the contributions of low
control and high demands were equally weighted [Land-
sbergis et al., 1994]; that is, a subject scoring 3 for demands
and 2 for control receives the same value for job strain as
someone who scores 2 for demands and 1 for control. We
to assign subjects to high and low levels of job strain.
Musculoskeletal symptoms experienced in the past
4 weeks were assessed by asking each subject to self-report
pain intensity (0 to 5-point scale with verbal anchors of ‘‘a
little painful’’ for 1 and ‘‘very painful’’ for 5) [Dickinson
et al., 1992]. To examine recovery patterns for neck and
of pain intensity over time; specifically, we calculated a pain
baseline pain value for each worker. This allowed us to
examine how the factors we measured influenced pain
reduction or increase over the 4 months of observed period.
Statistical Methods and Analysis
intervention effects and effects of other factors of interest on
pain score change over follow-up time [Singer, 1998]. The
linear mixed effects models included the following compo-
nents: fixed effects, random effects, and non-independent
covariance structures. The estimates for fixed effects were
assumed to be constant for all subjects. They included time-
independent estimates for individual and psychosocial
factors collected at baseline, such as age and gender,
and time-varying estimates for factors collected over
follow-up time, such as the number of sewing tasks
performed and changes in work schedule. To account for
at the first month of follow-up. The specification of a
linear spline function allows us to estimate the nonlinear
trajectory of pain score change over follow-up time. We
controlled for the type of intervention package each
worker received in all models. In addition, because our
outcome variable—neck/shoulder pain recovery pattern—
longitudinally, we specified a first-order autoregressive
FIGURE 1. Flowchartofsubjectsincludedinthisanalysis.
354 Wang et al.
[AR(1)] covariance structure for all linear mixed effects
for the potential correlations among outcome measures at
different time points within each subject, and to specify the
correlations of two consecutive outcome measures between
increasing time intervals.
models to examine the main effect of intervention over
follow-up time while controlling for potential confounders.
The first or our ‘‘basic’’ model included a time factor in
addition to the intervention effect term to establish the
unadjusted pattern of pain score change over time for all
subjects. We adjusted for baseline pain intensity in our
‘‘simple’’ second adjusted model, and in the third ‘‘fully’’
adjusted model, we added 12 additional covariates to
the second model. These covariates were chosen based on
reports in the literature and because in our previous cross-
sectional study we found that they were important [National
Institute for Occupational Safety and Health, 1997; Thomas
et al., 2005; Wang et al., 2007]; these were gender, age,
ethnicity, medical history of musculoskeletal problems,
overtime work, number of tasks performed, number of
sewing machines operatedin the past 4weeks, total rest time
per day, job strain, social support, job dissatisfaction, and
perceived physical workload. Information about all of these
factors was collected at baseline, except for workdays
perweek and numberof tasksperformed inthe past 4weeks,
which were collected longitudinally throughout follow-up.
For the fully adjusted model described above we assume
that each factor of interest has a constant effect on pain
reduction over follow-up time. Thus, we further investigated
the interaction between follow-up time and each factor
separately, adjusting for all other covariates already included
in the fully adjusted model. In other words, we examined
whether each of these factors altered recovery from neck/
shoulder pain over follow-up time. Because our outcome
variablewas defined as the proportion of pain at baseline, this
variable was by definition set to 100% at baseline for each
subject. Thus, in our interaction models, the difference in
intercept (or baseline pain) between categories of each factor
cannot be estimated, instead the slope for each category was
estimated. The difference between these interaction models
and the fully adjusted model can be viewed as follows: in the
uptimeisassumedtobeparallelfor eachcategoryofa factor,
while in the interaction models the pattern of pain reduction
that factor. All statistical analyses were performed using
statistical software SAS version 9 [SAS Institute, Inc., 2002].
Of the 247 subjects reporting neck/shoulder pain during
the baseline interview majorities were female (65.2%),
Hispanic, or Asian (75.7% and 21.9%, respectively) and the
mean age was 38 years (Table I). Only 28 workers (11.3%)
to the interview. About half (54%) of all subjects reported
taking less than 50min of rest time in a workday. The
three time-varying variables (workdays per week, number of
sewing tasks performed, and number of sewing machines
operated in the past 4 weeks) did not change appreciably
a sudden drop in reported pain intensity between baseline
interview and the first follow-up survey, and the pain
intensity remained constant from the first to the fourth
survey (Table III).
Our analyses revealed that the patterns of pain score
(results not shown). We observed a 72% reduction for the
mean pain score from baseline to the first month and a non-
discernible increasing trend in pain score from the first to
fourth month (4% per month) of follow-up adjusting for
baseline pain intensity, intervention type, four individual
TABLE I. Frequency DistributionsofDemographicCharactersforSubjects
*Data included frequency in number [N] and %, and means with standard deviation
Follow-Up of Neck and Shoulder Pain355
level factors (age, gender, ethnicity, and history of muscu-
number of sewing tasks performed, number of sewing
work), and four psychosocial factors (high job strain, high
social support, high job dissatisfaction, and high perceived
Higher baseline pain intensity, being Hispanic (vs.
Asian), work rest periods of 35min or more per day, and
perceived low physical workload were associated with
greater mean pain score reduction in the first month, while
our estimates of differences were imprecise thereafter (see
Table IV for unadjusted models and Table V for adjusted
models). Workers who took 35min of rest or more during
work per day experienced approximately a 20% more
reductionin meanpain scorecompared totheir counterparts.
Workers who reported their physical workload as being low
tended to have a 16% more reduction in mean pain score in
the first month and a 2.5% more from the second to
fourth month. No differences were observed in estimated
TABLE II. Frequency DistributionsofTime-VaryingVariablesAmong Sewing Machine OperatorsWithNeck/Shoulder Pain at Baseline and Four Follow-Up
2.2(1.2) 2.1(1.2)2.1(1.4) 2.1(1.1)2.0(1.2)
356Wang et al.
pain score change when stratifying according to the
following factors: age, gender, medical history of MSD,
workdays per week, number of task performed, number of
machine operated, job stress, social support, and job
Although none of the interaction terms between any of
these factors and time affected pain score changes over time,
we observed that the mean pain score reduction was about
24% less for workers who reported high physical workload
compared to workers who reported low physical workload
and about 18% less for workers working overtime compared
to workers who did not work overtime. No differences were
observed in the overall pain reduction during our 4 months
follow-up when stratifying according to any of the other
11 factors examined.
There is little research evaluating the follow-up of neck/
shoulder pain in a work place setting. To the best of our
work-related factors. Within the setting of an ergonomic
intervention study in garment workers, we employed a
repeated measures approach and a multivariate model to
examine recovery patterns based on longitudinal pain scores
[Cole and Hudak, 1996; Bonacich and Appelbaum, 2000;
Ferguson et al., 2001]. This approach allows for a sensitive
repeated measurements in the same subject. We observed a
non-monotonic trend for self-reported neck/shoulder pain
Meandifference95%CI Meandifference 95%CIMeandifference95%CI
Follow-Up of Neck and Shoulder Pain 357
during the follow-up period. Specifically, pain scores
dropped dramatically in the first month and then steadied
or increased again with time. This drop can be interpreted in
several ways. First, this may suggest that neck and should
pain does improve considerably within a 1-month period in
on the jobs. Second, it could have been due to postural
improvements after the introduction of our ergonomic
interventions in this workforce. Yet, the drop occurred in
all subjects similarly no matter whether we assigned them to
a control group or to one of two types of interventions.
to a possible bias caused by changes in data collection
procedures between baseline and follow-up surveys. Specif-
ically, while wewere allowed to conduct the baseline survey
in a private interview room, the follow-up surveys had to be
administered at each subject’s workstation in order to
interfere minimally with the flow of work in the garment
shops. Even though the generally loud background noise in
the shops minimized the possibility of being overheard,
subjects may have been more concerned in these public
settings, and, thus, underreported pain.
Nevertheless, we identified a number of factors to be
associated with the extent of pain reduction in the
first monthly follow-up survey independent of the inter-
vention effects. These include baseline pain intensity,
ethnicity, total duration of daily rest time during work, and
perceived physical workload. We furthermore found that
TABLE V. Adjusted*Time^Effect Interaction ModelThat Comparedthe Difference ofLongitudinal Changein Mean Pain ScoreAcross Groups (Categorical
Meandifference95%CI Meandifference 95%CIMeandifference95%CI
16.4 (1.4,31.3) 2.5(?5.4,10.3)23.8 (7.2,40.5)
358Wang et al.
pain score during the 4-month follow-up period.
A randomized controlled trial examined follow-up of
lateral epicondylitis in a general practice [Haahr and
Andersen, 2003] and assigned 266 patients randomly to
receive minimal occupational intervention or the usual
approach in general practice. The minimal occupational
intervention involves some education about the disorder,
instruction on how to perform some exercises, and
encouragement to remain active. After 1 year of follow-up,
both intervention and control group received more than 80%
ofimprovement inpain condition while intervention was not
followed by anydifference in pain score or perceivedoverall
development compared with the control group. This study
also reported that poor overall improvement in 1 year of
follow-up was related to holding a manual job, experiencing
highlevelof physical strainatwork, and reporting highlevel
work such as push/pull activities, working with the hands
above shoulder level, and working with vibrating tools,
experienced poorer neck or shoulder pain than the general
working population [Grooten et al., 2007]. These findings
support our own results that perceived high physical work-
are associated with poorer neck and shoulder pain over time.
Work-related psychosocial factors such as job strain,
2001; Andersen et al., 2003; Helliwell and Taylor, 2004].
Furthermore, these factors may modify the state of disorders
of the upper body region [Chard et al., 1988; Houtman et al.,
1994; Cole and Hudak, 1996; Toomingas et al., 1997;
Westgaard, 1999; Bonde et al., 2003]. Our previous cross-
sectional analyses of these data found high job demands and
low job satisfaction to be related to a higher prevalence of
data were inconclusive with regard to whether these factors,
measured at baseline only, also influence changes in neck/
shoulder pain over time.
High baseline pain intensity, Hispanic race (vs. Asian),
total duration of daily rest time during work of 35min or
more allowing for the muscle to rest, and perceived lower
physical workload were all associated with larger pain
reductions within the first month offollow-up, and the initial
differences remained similar during the additional 3 months
of follow-up except for ethnicity. We acknowledge that self-
reporting of pain and work-related factors may have
introduced some reporting bias. During follow-up especially
between the second and fourth months, we observed that
workers with higher work loads and job strain, and those
surveys more quickly to be able to return to work. This may
have compromised data quality for pain measures during
follow-up in a differential way, that is, under-reporting of
pain may have occurred more often in the most highly
exposed workers; the resulting misclassification is likely to
have weakened effect estimates for these factors on pain
severity. Also, the data presented here were collected in a
convenience sample of 13 garment shops located in Los
Angeles Basin willing to cooperate in an ergonomic
intervention study. This type of convenience sampling may
result in effects different from those one would estimate in a
the shops and employees are different from shops choosing
not to participate in such a study. In fact, due to the rigorous
most likely represented a more stable garment worker
population working in shops willing to participate in
intervention research. Finally, several of the garment shops
due to the impact of unexpected changes in international
trade policies. Thus, we were unable to retain subjects in our
extended follow-up exploring a longer term follow-up of
these types of pain are warranted and would need to target
relatively stable working populations.
obtained detailed information for work-related and psycho-
socialfactors over time, which allowed us touse amultilevel
modeling approach and to relate repeated measures of key
exposure factors to outcomes over time while adjusting
for time-varying confounders. The use of a linear mixed
model improves efficiency compared to traditional uncondi-
tional regression modeling [Greenland, 1992; Witte et al.,
1994]. Since subjects may have different pain perception
thresholds those reporting higher pain intensities at baseline
would be able to improve more than subjects reporting a
baseline pain intensity standardizes the range of pain scores
for all subjects by assigning an equal pain level to every
subject at baseline.
pain at baseline and perceiving their work as physically very
demanding (high physical workload) or working overtime
(more than 5 days per week on average) experienced less
workers who perceived their work as requiring a low
physical effort or who worked 5 days per week or less. In
addition, higher baseline pain intensity, being of Hispanic
ethnicity (vs. Asian), and taking cumulative daily rest time
lead to a larger pain reduction in the first month, and the
difference remained the same over follow-up except for
ethnicity. These findings may be useful for guiding medical
neck/shoulder pain among this underserved immigrant
Follow-Up of Neck and Shoulder Pain 359
ACKNOWLEDGMENTS Download full-text
This study was supported in part by a grant (1 R01
OH007779) from the Centers for Disease Control/National
Institutes for Occupational Safety and Health. The authors
wish to thank the Los Angeles Garment Worker Center and
for their cooperation and support of this study. No funding
source or sponsor had any role in the design and conduct of
of the data; or preparation, review, or approval of the
Andersen JH, Kaergaard A, Mikkelsen S, Jensen UF, Frost P, Bonde JP,
Fallentin N, Thomsen JF. 2003. Risk factors in the onset of neck/
companies. Occup Environ Med 60:649–654.
Ariens GA, van Mechelen W, Bongers PM, Bouter LM, van der WG.
Ind Med 39:180–193.
Bonacich E, Appelbaum RP. 2000. Behind the label: Inequality in the
Los Angeles Apparel Industry. Berkeley, CA: University of California
Bonde JP, Mikkelsen S, Andersen JH, Fallentin N, Baelum J, Svendsen
a cohort of Danish industrial and service workers. Occup Environ Med
Chan J, Janowitz I, Lashuay N, Stern A, Fong K, Harrison R. 2002.
Preventing musculoskeletal disorders in garment workers: Preliminary
results regarding ergonomics risk factors and proposed interventions
among sewing machine operators in the San Francisco Bay Area. Appl
Occup Environ Hyg 17:247–253.
Chard MD, Sattelle LM, Hazleman BL. 1988. The long-term outcome
Cole DC, Hudak PL. 1996. Prognosis of nonspecific work-related
musculoskeletal disorders of the neck and upper extremity. Am J Ind
Dickinson CE, Campion K, Foster AF, Newman SJ, O’Rourke AM,
Thomas PG. 1992. Questionnaire development: An examination of the
Nordic Musculoskeletal questionnaire. Appl Ergon 23:197–201.
Ferguson SA, Gupta P, Marras WS, Heaney C. 2001. Predicting
recovery using continuous low back pain outcome measures. Spine
multiple associations, with an application to an occupational cancer-
mortality study. Stat Med 11:219–230.
Grooten WJ,Mulder M, Josephson M, Alfredsson L, Wiktorin C. 2007.
The influence of work-related exposures on the prognosis of neck/
shoulder pain. Eur Spine J 16:2083–2091.
A randomized trial with one-year follow-up in 266 new cases treated
practice. Rheumatology (Oxford) 42:1216–1225.
Helliwell PS, Taylor WJ. 2004. Repetitive strain injury. Postgrad Med
Houtman IL, Bongers PM, Smulders PG, Kompier MA. 1994.
Psychosocial stressors at work and musculoskeletal problems. Scand
J Work Environ Health 20:139–145.
Jonsson BG, Persson J, Kilbom A. 1988. Disorders of the cervicobra-
chial region among female workers in the electronics industry: A two-
year follow up. Int J Ind Ergon 3:1–12.
Karasek R. 1997. The Job Content Questionnaire and User’s Guide.
Lowell: University of Massachusetts Lowell.
Landsbergis PA, Schnall PL, Warren K, Pickering TG, Schwartz JE.
1994. Association between ambulatory blood pressure and alternative
formulations of job strain. Scand J Work Environ Health 20:349–
National Institute for Occupational Safety and Health. 1997. Muscu-
loskeletal disorders and workplace factors: A critical review of
epidemiologic evidence for work-related musculoskeletal disorders of
the neck, upper extremity and low back (NIOSH Publication No. 97-
141). Cincinnati, OH: Department of Health and Human Services.
randomized controlled trial evaluating the effects of new task chairs on
shoulder and neck pain among sewing machine operators: The Los
Angeles Garment Study. Spine 32:931–938.
SAS Institute, Inc. 2002. STAT Software. Changes and Enhancements
through Release 9.1. [9.01]. Cary, NC: SAS Institute, Inc.
Singer J. 1998. Using SAS PROC MIXED to fit multilevel models,
hierarchical models, and individual growth models. J Educ Behav Stat
Thomas E, van der Windt DA, Hay EM, Smidt N, Dziedzic K, Bouter
LM, Croft PR. 2005. Two pragmatic trials of treatment for shoulder
disorders in primary care: Generalisability, course, and prognostic
indicators. Ann Rheum Dis 64:1056–1061.
Toomingas A, Theorell T, Michelsen H, Nordemar R. 1997. Associa-
tions between self-rated psychosocial work conditions and muscu-
loskeletal symptoms and signs. Stockholm MUSIC I Study Group.
Scand J Work Environ Health 23:130–139.
Wang PC, Rempel D, Harrison R, Chan J, Ritz B. 2007. Work-
organizational factors associated with the prevalence of upper body
musculoskeletal disorders among sewing machine operators. Occup
Environ Med [Epub ahead of print].
Wang PC, Ritz B, Janowitz I, Harrison R, Yu F, Chan J, Rempel DM.
2008. A randomized controlled trial of chair interventions on back and
hip pain among sewing machine operators: The Los Angeles Garment
Study. J Occup Environ Med 50:255–262.
pain. Scand J Work Environ Health 25:19–24.
Witte JS, Greenland S, Haile RW, Bird CL. 1994. Hierarchical
regression analysis applied to a study of multiple dietary exposures
and breast cancer. Epidemiology 5:612–621.
360 Wang et al.