Characterizing the Course of Low Back Pain: A Latent Class Analysis
Kate M. Dunn, Kelvin Jordan, and Peter R. Croft
From the Primary Care Sciences Research Centre, Keele University, Keele, United Kingdom.
Received for publication August 4, 2005; accepted for publication December 5, 2005.
Understanding the course of back pain is important for clinicians and researchers, but analyses of longitudinal
data from multiple time points are lacking. A prospective cohort study of consecutive back pain consulters from five
general practices in the United Kingdom was carried out between 2001 and 2003 to identify groups defined by their
pain pathways. Patients were sent monthly questionnaires for a year. Longitudinal latent class analysis was
performed by using pain intensity scores for 342 consulters. Analysis yielded four clusters representing different
pathways of back pain. Cluster 1 (‘‘persistent mild’’; n ¼ 122) patients had stable, low levels of pain. Patients in
cluster 2 (‘‘recovering’’; n ¼ 104) started with mild pain, progressing quickly to no pain. Cluster 3 (‘‘severe chronic’’;
n ¼ 71) patients had permanently high pain. For patients in cluster 4 (‘‘fluctuating’’; n ¼ 45), pain varied between
mild and high levels. Distinctive patterns for each cluster were maintained throughout follow-up. Clusters showed
statistically significant differences in disability, psychological status, and work absence (p < 0.001). This is the first
time, to the authors’ knowledge, that latent class analysis has been applied to longitudinal data on back pain
patients. Identification of four distinct groups of patients improves understanding of the course of back pain and
may provide a basis of classification for intervention.
classification; cohort studies; longitudinal studies; low back pain; primary health care; prospective studies; statistics
Abbreviation: RMDQ, Roland-Morris Disability Questionnaire.
Understanding the course of low back pain is important
for clinicians and researchers because it provides informa-
tion on the need for, and potential benefits of, treatment
(1, 2). It also helps patients learn what to expect in terms
of symptoms, the impact of the problem on their life, and the
interventions they may receive. Information on symptom
course may enable patients with nonspecific low back pain
to be classified into clinically meaningful subgroups. There
are currently no accepted methods for classifying these pa-
tients, who constitute 85–95 percent of those seeking care
for low back pain (3). Thus, it is difficult to select clearly
defined subgroups of patients for clinical trials, and the
potential effectiveness of treatments may be masked by
the heterogeneity of the patients studied.
It has been suggested that long-term pain conditions, such
as low back pain, follow recurrent or fluctuating patterns
(4, 5), and hypothetical time courses for these conditions
have been proposed (6). However, because of, in part, the
difficulties in measuring such symptoms and the repeated
measurements necessary (6), there is little empirical evi-
dence to support these models. Previous study of the course
of low back pain has tended to provide data on only the
proportion of persons who have recovered or are still experi-
encing symptoms at various time points (2, 4). The majority
of longitudinal studies of low back pain are not designed
to characterize symptom course but to collect information
at baseline, which is then used to predict an outcome at
later time points, commonly 3, 6, or 12 months (7–10). A
few studies have gathered more detailed information over
shorter time periods (11, 12), but studies collecting detailed
measurements over a longer period of time are lacking.
Such studies using repeated measurements are necessary
Correspondence to Dr. Kate M. Dunn, Primary Care Sciences Research Centre, Keele University, Keele, Staffordshire, ST5 5BG, United
Kingdom (e-mail: firstname.lastname@example.org).
754Am J Epidemiol 2006;163:754–761
American Journal of Epidemiology
Copyright ª 2006 by the Johns Hopkins Bloomberg School of Public Health
All rights reserved; printed in U.S.A.
Vol. 163, No. 8
Advance Access publication February 22, 2006
at Harbin Institite of Technology on June 1, 2013
diagnosis and treatment in general practice. J R Coll Gen Pract
29. Coste J, Delecoeuillerie G, Cohen de Lara A, et al. Clinical
course and prognostic factors in acute low back pain: an in-
ception cohort study in primary care practice. BMJ 1994;308:
30. Klenerman L, Slade PD, Stanley IM, et al. The prediction of
chronicity in patients with an acute attack of low back pain in
a general practice setting. Spine 1995;20:478–84.
31. Carey TS, Garrett JM, Jackman AM, et al. Recurrence and
care seeking after acute back pain: results of a long-term
follow-up study. North Carolina Back Pain Project. Med Care
32. Croudace TJ, Jarvelin MR, Wadsworth ME, et al. Develop-
mental typology of trajectories to nighttime bladder control:
epidemiologic application of longitudinal latent class analysis.
Am J Epidemiol 2003;157:834–42.
33. Delucchi KL, Matzger H, Weisner C. Dependent and problem
drinking over 5 years: a latent class growth analysis. Drug
Alcohol Depend 2004;74:235–44.
34. Muthe ´n B. Latent variable analysis. In: Kaplan D, ed. The Sage
handbook of quantitative methodology for the social sciences.
Thousand Oaks, CA: Sage Publications, 2004:345–68.
35. Pedersen PA. Prognostic indicators of low back pain. J R Coll
Gen Pract 1981;31:209–16.
36. Langeheine R, Pannekoek J, Van de Pol F. Bootstrapping
goodness-of-fit measures in categorical data analysis. Sociol
Methods Res 1996;24:492–516.
No definitive method of determining the optimal number
of clusters in a latent class analysis exists. One of the most
common is to examine the model fit likelihood ratio chi-
squared statistic, L2: the amount of the relation between the
monthly levels of pain that remains to be explained. The
larger the value of L2, the worse the model fit, and a p value
can be calculated to assess the goodness of fit. When data
of variables or categories compared with the number of
to determine the p value, and bootstrap p values are recom-
the pvaluebecomes nonsignificantatthe desiredsignificance
level. Other methods that are particularly useful when data
are sparse include information criterion statistics that take
into account the parsimony of the model, such as Akaike’s
Information Criterion, Bayes’ Information Criterion, and the
Consistent Akaike’s Information Criterion (23). The optimal
number of clusters occurs when the information criterion
value is at its lowest. A more subjective method is to assess
the percentage reduction in L2from the model with one clus-
ter and select the number of clusters from the model beyond
which this reduction is considered minor (23).
Latent Gold uses both the estimation-maximization and
Newton-Raphson algorithms to estimate model parameters
(24). A problem that sometimes occurs in latent class anal-
ysis is that a local maximum, rather than the global best
solution, is obtained. To avoid this situation, 1,000 repeated
runs were performed from random start values. Bootstrap
p values based on 500 replications were determined to as-
sess the model fit based on the L2statistic.
The goodness-of-fit statistics for the one-cluster to eight-
cluster models are shown in appendix table 1. These statis-
tics are based on the 188 subjects whose datawere complete.
Based on bootstrap p values for the model fit L2statistics,
the six-cluster model was the optimal assuming a 5 percent
significance level. However, the percentage reduction in
is gained by expanding beyond a four-cluster model. There
to the four-cluster model. The six-cluster model reduced
L2by only a further 4 percent. The information criterion
values suggested a seven-cluster solution based on Akaike’s
Information Criterion, a four-cluster model based on Bayes’
Information Criterion, and a three-cluster model based on
the Consistent Akaike’s Information Criterion.
Goodness-of-fit statistics in which data for all 342 sub-
jects were used suggested a four-cluster solution (based on
Bayes’ Information Criterion, Consistent Akaike’s Infor-
mation Criterion, and percentage reduction in L2) or a six-
cluster solution (bootstrap p value, Akaike’s Information
Criterion). On the basis of these results and the character-
istics and size of the clusters, the four-cluster solution was
selected as optimal.
APPENDIX TABLE 1.
Goodness-of-fit statistics for cluster models of primary care low back pain consulters (n ¼ 188), United
% reduction in
One-cluster (H0)1,044.957 716
<0.00012,384 2,423 2,435
Two-cluster 471.427 709
<0.0001 55 1,8251,886 1,905
Three-cluster 219.930 702
Four-cluster 181.041 6950.005 831,562 1,6691,702
Five-cluster 153.901 6880.023 851,5491,679 1,719
Six-cluster 132.268 6810.073 871,5421,694 1,741
Seven-cluster116.532674 0.17891,5401,715 1,769
Eight-cluster 105.368667 0.18 901,543 1,7401,801
* AIC, Akaike’s Information Criterion; LL, log-likelihood; BIC, Bayes’ Information Criterion; CAIC, Consistent Akaike’s Information Criterion.
Characterizing the Course of Low Back Pain761
Am J Epidemiol 2006;163:754–761
at Harbin Institite of Technology on June 1, 2013