Low back disorders are the most common, costly and disabling of musculoskeletal health problem.¹ They are one of the most common reasons for visits to primary care,² and the single largest category of workers’ compensation claims in most compensation systems.³ While some cases of low back pain are transient and resolve or considerably improve over several weeks,^4,5 symptoms can often be recurrent or chronic.^6,7 Studies have shown that over 65% of back pain patients in primary care continued to experience at least mild pain one month after seeking care, and approximately 33% reported continuing pain of at least moderate intensity after 12 months.^5,8

The course of back disorders can be characterised using measures of different aspects of health,⁹ such as pain, disability, activity limitations and/or participation.¹⁰ For instance, the course has been studied by assessing the prevalence of back pain at different time points, or by evaluating changes in the amount of pain across time in pre-classified patient groups.^10–12 Transitions between different states such as resolution, improvement, aggravation and recurrence have also been used to study courses.¹³

Participation in work is also a key outcome in studies of back pain. Duration of claim in workers’ compensation systems^10,14,15 and time to return to work^10,16 have also been used as end-points in studies evaluating the duration of back-pain episodes, or to identify prognostic factors or effective interventions. However, measures based on first return to work may underestimate the percentage of people impaired^17,18 because workers may continue to experience pain after returning to work, or may experience recurrent episodes of pain and disability. While some studies have shown a relation between pain and disability and time to claim closure,¹⁹ other factors also may influence the decision to return to work.¹⁰

Von Korff described the course of back pain as highly variable, occurring in transient, recurrent, and chronic phases.⁷ Few longitudinal studies have been carried out to classify the different pathways of back pain.^13,20 Dunn et al used longitudinal latent class analysis to identify four groups of back pain patients with distinctive recovery patterns in a primary care setting.²⁰ It would be of interest to explore the course of back pain in a population of injured workers to determine similarities and differences to those among the general population.

The purpose of this study is (1) to identify and summarise the course of pain in a cohort of injured workers off work because of back pain and (2) to describe the distribution of the patterns across other commonly used proxy measures of recovery.

METHODS

Study design

The original sample for this study was a cohort of 1825 injured workers having lost-time claims with the Ontario Workplace Safety & Insurance Board for soft tissue injuries of the upper limb, lower limb or the back. The Board is the principal provider of workers compensation in Ontario and covers approximately 65% to 70% labour force participants.^21,22 Claimants were recruited shortly after they registered their claims and interviewed at up to five points in time over the course of one year. The baseline interview was conducted shortly after injury, with follow-up interviews at 4, 10, 16 and 52 weeks post injury. The focus of this study is on the results from the 885 cases in the cohort who were workers filing “new” claims for back injury (that is, not a reopened claim) and were still off work at the time of recruitment. More details of this study sample can be found elsewhere.^14,15

Measures

Main outcome: pain intensity

Pain intensity was measured using a subscale of the Chronic Pain Grade,²³ a composite measure derived from self-reported information on three 0–10 numeric rating scales for the worst and average back pain since the beginning of current episode (at baseline) or since the last interview (at follow-ups), and back pain severity of this moment. It was rescaled to range from 0 to 100, with 0 representing no pain and 100 representing the worst pain. Properties of the original Chronic Pain Grade have been established.²³ The subscale used here with altered time frames has been used in previous research.^14,15

Other measures of recovery

Three other measures of recovery were compared with patterns of pain intensity. Disability was measured by using the Roland-Morris Disability (R-M Disability) Questionnaire.^24,25 The disability score was rescaled from 0 to 100, with 0 representing no disability and 100 representing the highest disability. Two dimensions of the SF-36,²⁶ a generic, health-related quality-of-life measure, were used in this study: Role Physical (SF-36 RP) score and Mental Health (SF-36 MH) score. Both SF-36 scores were rescaled to range from 0 to 100, with 0 representing worst health and 100 representing best health. The length of time on workers’ compensation was assessed by the cumulative duration on 100% benefits because of back injury within one year. Duration on benefits or length of claim, while not necessarily equivalent to the total time away from work since the injury occurred, is often used as a proxy for it.²⁷ Associations between length of claim and other key measures of health have previously been demonstrated.¹⁹ In addition to duration on benefits, binary measures were created to indicate whether a respondent was receiving 100% benefits or not at each interview time. This measure is a more accurate estimate of the percentage of respondents who were off work at a given time point.

Baseline individual characteristics

Individual characteristics recorded at baseline included demographics, past history of back injuries and other health conditions (such as bronchitis, arthritis, heart trouble, cancer, balance problems, etc).

Statistical analysis

SAS version 9.1 (SAS Institute, Cary, NC, USA) was used to perform all statistical analyses. The methods used to identify back pain recovery patterns required at least three reports of pain intensity. Therefore, only workers with at least three measures of pain intensity over the one-year follow-up were included in the analysis (n = 678, 77% of study sample). A comparison of included and excluded respondents was undertaken to evaluate potential bias due to exclusion. To identify recovery patterns, a two-step classification method was used. The assumption of this method is that a certain number of distinct recovery patterns of back pain exist among injured workers, and injured workers can be classified into one and only one of the distinct pattern groups based on their profiles of back pain over the one-year period.

First, the longitudinal change of pain intensity, often called the “slope”, was calculated for each individual across their corresponding repeated measures in the one-year follow-up by using simple linear regression. The slopes indicated the overall trend of pain development for each individual. Then, workers were stratified into three groups based on the magnitude and precision of these “slopes” of pain intensity: those with increasing pain (slope >0 and p⩽0.10), those with decreasing pain (slope <0 and p⩽0.10), and those with constant or fluctuating pain (p>0.10).

In the second step, individual patterns within each of the three groups were further classified into relatively homogeneous subgroups. If the group size from the first step was less than 30, no further classification was attempted. Otherwise, K-means cluster analysis using the FASTCLUS procedure in SAS was applied to the repeated pain intensity scores within each group. This method assigned each observation to one disjoint cluster based on the shortest Euclidean distance from the cluster centre. For the observations with missing values, the procedure computes an adjusted distance to the cluster centres using the non-missing values.²⁸ The K-means algorithm requires specification of the number of clusters, a priori. We started by examining one cluster, then sequentially increasing the number of clusters until pre-specified criteria were met.

There is no conclusive standard for determining the optimal number of clusters.²⁹ The pseudo-F statistic, the ratio of the between-cluster mean square to the within cluster mean square, is one criterion often used.³⁰ A higher F-statistic value reflects more distinct but homogeneous subgroups. In the end, the number of clusters we defined in this study was based on a trade-off between the statistical criteria, the interpretability and clinical relevance of the results.³⁰

The derived clusters were subsequently compared at baseline, week 4, 16 and 52 with respect to other measures of recovery and baseline subject characteristics using ANOVA, Fisher’s exact test, or log-rank test.

As respondents’ reports of the average and worst pain since the last interview may be affected by missing responses in the previous interview (that is, different recall periods among respondents), we performed an additional sensitivity analysis; re-running the two-step classification method on only those subjects who had all five interviews (39% of our study sample). We compared the clusters from the whole study sample to those from the subjects with complete data to determine whether there was evidence of systematic bias in the analysis. We found no significant differences between samples, in either the distribution of clusters, or the average pain intensities (results not shown, but available on request). Therefore only those results with the full sample are reported here.

RESULTS

Subject characteristics and pain intensity overview

Of the 679 respondents who had at least three interviews over the one-year follow-up period, one respondent was further excluded because the corresponding pain intensity reports were zero across all five interviews. Table 1 reports the baseline characteristics of the included sample and their average pain intensity at each time point. A comparison of baseline characteristics among included and excluded respondents showed that those included in the analysis were more likely to be female (p<0.001), had higher average disability scores (p = 0.10), and lower average SF-36 Role Physical scores (p = 0.05). No other notable differences between included and excluded participants were found.

The average pain intensity across the five interviews for both excluded and included study respondents were not statistically significantly different (p>0.1). For respondents included in the study sample, 17% had three observations over the one-year follow-up, 43% completed four interviews, and almost 39% completed all five evaluations. Overall, average pain intensity steadily decreased between baseline and 52 weeks. However, the corresponding standard deviation increased over time, suggesting large variation in the magnitude of the improvement between individuals. The average difference in pain intensity between week 16 and week 52 was less than five points, indicating improvements in pain were slower after the first three months after injury.

The patterns of longitudinal change in pain intensity

Figure 1 describes the five recovery patterns identified from the two-step cluster analysis via the mean pain intensity within each cluster. Further information on the process and the derivation of these patterns is provided in the Appendix. The increasing pain cluster consisted of nine (1%) workers who had a linear increasing pattern of pain over follow-up. The 282 (42%) workers showed limited change in pain with mean values ranging from 65 to 80, and this cluster was called the continuous high pain cluster. Approximately 33% of the study sample had substantial decreases in pain intensity during the first third of the follow-up period and fluctuations in pain towards the end; this cluster was called the fluctuating pain cluster. Workers who experienced decreasing pain were classified into two clusters: the 84 workers in the cluster of large reductions in pain (around 12% of the study sample) had the greatest decrease in pain and had the lowest mean pain intensity at 52 weeks; in the cluster of moderate reduction in pain were the other 81 workers. (Individual trajectories in pain intensity for 10 randomly selected respondents from each of the clusters were examined. These figures are presented in the supplementary file.)

Group comparisons of subject characteristics

Table 2 presents comparisons in baseline characteristics across the clusters, using univariate analysis (ANOVA, χ² test or Fisher’s exact test as appropriate). We combined the cluster of increasing pain and the cluster of continuous high pain, because of the small number of respondents with increasing pain (n = 9) and the minimal statistical and clinical differences between these two clusters (less than 20 points) at all follow-up time points. Significant differences were found between clusters on all characteristics at baseline except for gender. Individuals in continuous high pain (cluster 1) were older and more likely to have experienced other health conditions at baseline compared to those in fluctuating pain (cluster 2), moderate reduction in pain (cluster 3) and large reduction in pain (cluster 4). No significant differences were found between these last three clusters. Respondents in fluctuating pain (cluster 2) and large reduction in pain (cluster 4) were less likely to have had a previous back injury than those in continuous high pain (cluster 1) and moderate reduction in pain (cluster 3).

Comparisons between pain intensity and other measures of recovery

Pain intensity was most highly correlated with the Roland-Morris Disability (R-M Disability) scores at baseline (r = 0.50) and 52 weeks (r = 0.82). Similar strength of correlation was found between pain intensity and SF-36 Role Physical (SF-36 RP) scores at 52 weeks (r = −0.71), and between pain intensity and SF-36 Mental Health (SF-36 MH) scores at 52 weeks (r = −0.53) as well.

Table 3 compares, among clusters, mean scores of pain intensity, R-M disability, SF-36 RP, SF-36 MH, the proportion of respondents on 100% benefit at time of each interview, and the probability of receiving 100% benefits for longer than specified period of time (represented by the survivor function of cumulative duration on 100% benefits). Mean pain intensity was the lowest in cluster 2 (fluctuating pain) at baseline, 4 weeks and 16 weeks. However, at 52 weeks mean pain intensity was the lowest in cluster 4—large reduction in pain. The variances of pain intensity in cluster 2 were highest across all time points, compared to other clusters. The change of the average R-M disability scores in each cluster closely followed the path of pain in corresponding cluster, showing significant differences between each other.

No statistically significant differences in SF-36 RP were present across clusters at baseline. However, respondents in fluctuating pain (cluster 2) showed relatively worse physical health condition at baseline than those in large reduction in pain (cluster 4) although their initial pain and disability was lowest, compared to other groups. Over time, the four clusters developed distinctive recovery patterns in SF-36 RP similar to those found in pain intensity.

Respondents in fluctuating pain (cluster 2) and large reduction in pain (cluster 4) had significantly better mental health than respondents in continuous high pain (cluster 1) and moderate reduction in pain (cluster 3) at baseline. Nevertheless the development of SF-36 MH varied among clusters. While the respondents in fluctuating pain (cluster 2) reached high SF-36 MH scores, there was not much improvement in mental health for those in continuous high pain (cluster 1). The change of mental health in the other two clusters—moderate reduction in pain (cluster 3) and large reduction in pain (cluster 4)—followed their corresponding patterns of pain intensity.

The proportions of workers who were currently off work (receiving 100% benefits) at each point in time were significantly different (p<0.001) across four clusters at 4, 16 and 52 weeks. Cluster 2 (fluctuating pain) showed the highest proportion of respondents who were off benefits (back at work) at each of above interview times, compared to other clusters. By the end of the one-year follow-up, almost all respondents with fluctuating pain, moderate reduction in pain and large reduction in pain were back at work (only 4% still on 100% benefits).

All clusters were different (p<0.001) regarding the cumulative duration on 100% benefits over the one-year follow-up (fig 2, table 3). Cluster 2 (fluctuating pain) showed the shortest cumulative duration among other clusters, with the average at 47 days (95% CI 41 to 53 days). The survival curves of the clusters of continuous high pain (cluster 1) and moderate reduction in pain (cluster 3) were virtually indistinguishable until 138 days post-injury. Respondents in cluster 3 (moderate reduction in pain) had the longest average cumulative duration on benefits (150 days, 95% CI 136 to 163 days).

DISCUSSION

The primary objectives of this study were to identify different courses of back pain in a cohort of injured workers and to compare these patterns across other commonly used measures of recovery. Using a two-step cluster analysis we described four distinctive patterns of back pain intensity over a one-year period, with the most typical patterns of recovery being recurrent or chronic. We feel our study not only expands on the existing knowledge of recovery courses of back pain at the individual level, but also provides greater detail about the development of back pain across a 52-week period by comparing several commonly used measures of recovery.

Instead of looking at the change of mean pain intensity among the total study sample or examining changes among respondents in predefined groups at baseline, the longitudinal clustering methods allowed us to study the individual response profiles and also gain a deeper insight into the variation and the speed of change in back pain intensity over time. Previous studies have applied cluster analysis to look for patterns in anxiety severity,³¹ psychiatric treatment,³² and smoking habits.³³ Because cluster analysis does not account for the temporal sequence of values, some have suggested it may not be appropriate for summarising the different patterns of changes over time.³⁰ However, by using a two-step longitudinal cluster analysis we have overcome some of these limitations. Using the individual slopes of pain intensity takes into account the direction and magnitude of change between repeated observations at different time points. Our approach also accommodates missing and irregular observations, provided these occur at random. Imputing the missing values within each of the groups stratified by slopes did not fundamentally change the individual trajectories we found in this paper. Further, the recovery patterns we found were reproduced among random subsamples (75% of total sample) from our study population.

Our results support previous studies^7,16,34,35 that suggest the course of back pain is highly variable and that most respondents with back pain still experience a certain degree of pain even one year after injury. Three of our recovery patterns, continuous high pain, fluctuating pain and large reduction in pain (clusters 1, 2 and 4), are similar to those found by Dunn et al in a primary care setting.²⁰ However, the proportions of our sample which stayed in continuous high pain or fluctuating pain were over double the corresponding proportions from the study by Dunn et al. One explanation for this difference is the wide variety of baseline pain in Dunn’s primary population sample, compared to the relatively homogenous level of pain found in our sample of injured workers. Also note that all workers in the analysis presented here were off work because of their pain, indicating a certain level of severity that may not be seen across all cases presented to primary care. The cluster identified as fluctuating pain showed more variation than the corresponding cluster from the general population. Furthermore, despite a trend towards improvement in both clusters of reduction in pain (clusters 3 and 4), full recovery from back pain was less frequently observed compared to the general population. Based on our findings we suggest that long-lived difficulties from high pain levels are more commonplace in the individuals with work-related back injuries requiring time away from work than those found in a general population.

The groups of recovery patterns also differed across baseline variables such as age, baseline SF-36 mental health score and past history of back injuries. Old age^13,36,37 and past history of back pain/injuries^4,6,36 have previously been postulated as risk and prognostic factors of persistent disabling back pain. Other researches have also reported that psychosocial problems and high levels of distress were predictive of poor recovery of back pain.³⁷ Further research should better elucidate the characteristics of injured workers and specifically investigate predictors of different back-pain recovery patterns.

This study confirms that there was a strong agreement between the trajectories of disability levels and those of pain intensities among pattern groups, consistent with the result found by Dunn et al, but in contrast to other research.^38,39 This finding suggests that change in pain is reflected in disability, contrary to research³⁹ concluding that the effect of back pain on disability becomes relative as patients get used to pain over the course.

Almost 33% of our study sample reported fluctuating pain over our study period. This group had the earliest return to work and the shortest cumulative duration on temporary benefits. Compared to other groups, we observed high variation in mean levels of pain intensity at each interview in this group. Examining this issue further, we found that at least 10% of workers remained at high levels of pain (pain intensity of 50 or more out of 100) at each follow-up time. Although the plot of average pain intensity (fig 1) suggested that respondents with fluctuating pain might have multiple episodes of back pain, only a little over 7% of this group had been on temporary benefits more than once in the follow-up period.

When comparing different measures of recovery, pain intensity, R-M disability, SF-36 RP scores and the proportion of respondents receiving benefits all displayed similar patterns across clusters up to 16 weeks. However, between 16 and 52 weeks, measures of pain intensity, R-M disability and SF-36 RP scores all deteriorated among the group with fluctuating pain, while at the same time they showed improvement in the cluster with a large reduction in pain (cluster 4). This same pattern was not observed for the percentage of respondents who were receiving benefits. This suggests that a proportion of the group with fluctuating pain continued to work although their back pain recovery had either stalled or reversed. Taken together, the patterns of temporary benefits and return-to-work among the cluster of fluctuating pain suggest that both these measures may not accurately measure recovery from back pain, once initial return to work (or cessation of benefits) has taken place. For example, work absence data are often extracted from registries that contain minimal or no information about the pain, the disability or the clinical characteristics of the worker.⁴⁰ Provision of workers’ compensation or time to return to work may be influenced by many other factors,^10,40 such as legislation, the system of payment, the type of work available for returning to work. Therefore, while these measures reflect the cost of back pain at a societal level they may not reflect the ongoing burden of back pain among injured workers, as reflected by individual reports of pain and disability.

The results of this study should be interpreted given the following limitations regarding the data collection and analytic approach. First, claimants with reopened claims and people who had returned to work at the time of recruitment were excluded from the study. Although clinical and third party payer experience suggest that the natural history of patients from reopened cases is different from that of initial cases,¹⁴ their exclusion here limits the generalisability of the findings. Second, data were not available for all participants at each time point, given our longitudinal study. Therefore, some patterns of recovery may be overlooked due to the scarcity of data. However, the correspondence found between the trajectories and individual characteristics concerning the development of back pain argues for the validity of the trajectories identified in our study. Although we found a substantial proportion of our study sample reported fluctuating pain (33%) we have limited information on the responsiveness of the pain measure used, and its ability to pick up meaningful fluctuations in pain among individuals. For example, the inclusion of average pain reports in our current measure may mask small, but significant fluctuations in pain. Finally, K-means cluster analysis is highly dependent on the choice of the clustered variables, the initial seeds, and the distance to evaluate the proximities of the points to one another. This implies that there may be some relatively important variability among workers in a given cluster.³⁰ In the implementation of cluster analysis, criteria for decision making at each step was carefully assessed and validated. The recovery patterns were further confirmed by the plots of actual patterns of change for a randomly selected subset of workers in each cluster (supplementary file), as well as previous studies^13,20 in the general population.

CONCLUSION

The patterns of recovery found in this study have important implications for clinical practice concerning back pain, especially among injured workers. Our results suggest wide variation in the patterns of recovery, as assessed by self-reports of pain, with the most common recovery patterns being continuous high pain and fluctuating pain. We found that membership in different recovery groups differed in baseline age, weekly earnings and past history of back injuries. Further research should be undertaken to identify factors predicting which recovery pattern individuals may follow, using information at the time of injury/claim/consultation. Our findings suggest that, while all measures of recovery were similar across groups up to 16 weeks, after initial return-to-work, or cessation of benefits, administrative measures of percentage of respondents back at work, or no longer on benefits, may not accurately reflect an individual’s pain intensity or disability.

APPENDIX: TWO-STEP CLUSTER ANALYSIS PROCEDURE

In the two-step cluster analysis, the workers were first divided into three groups according to the individual pain intensity slopes and p values calculated by simple linear regression (table A1). Only a little over 1% of the workers experienced deterioration over the one year period after injury. Although almost 25% of workers had improvement overtime, the majority of them experienced multiple episodes or chronic back pain. Because the sample size of the first group was small and related to workers with a consistent trend of increasing pain, the focus of cluster analysis was on the second and third group.

In the second step, the cluster model with highest pseudo-F statistic and at least 30 individuals in each cluster were selected as the result. Two clusters were identified within each of group 2 (decreasing pain) and group 3 (continuous pain) based on the clustering criteria (table A2).