Study design

This is a prospective study using data from the Danish physical activity cohort with objective measurements (DPhacto). The study protocol for the cohort is reported in detail elsewhere.31 Data collection took place from April 2012 to May 2014 at 15 Danish companies, including workers in four occupational sectors (cleaning, manufacturing, transportation and office work/administration). The initial contact and recruitment of companies were performed in collaboration with a large Danish union. The companies were selected to represent blue-collar occupations with different levels of physical demands at work.

Baseline data collection consisted of a brief web-based questionnaire, a standard health examination and objective diurnal measurements of physical activity and heart rate (presented elsewhere).32 33 Prospective self-reported data on musculoskeletal pain, sick leave due to pain and work ability was collected repeatedly over 12 months using text messages.

Study population

The inclusion criterion for participation was current employment within any of the recruited work places. Exclusion criteria were holding a managing position or being pregnant or a student/trainee. In addition, workers not responding to the baseline questionnaire and/or the prospective measurements were excluded.

Among the 2107 invited workers, 1119 agreed to participate and 32 of them were excluded due to holding a managing position (n=17) or being pregnant (n=2) or a student/trainee (n=13). Of the remaining 1087 eligible workers, 782 responded to the questionnaire and 748 took part in the prospective measurements from baseline. Thus, the final study sample consisted of 748 workers (blue-collar, n=620; white collar, n=128). Descriptive characteristics of the study population are shown in table 1.

Repeated assessment using text messages

Text messages (SMS) were used to assess self-reported pain intensity in the neck–shoulder region, days on sick leave and work ability using the commercial software ‘SMS-Track’ (https://sms-track.com/). Starting at baseline, data on NSP and sick leave were collected at 4-week intervals (14 waves), while data on work ability were collected at 12-week intervals (four waves in total) during the 1-year study. The SMS were sent on Sundays, with a reminder the following day.

Neck–shoulder pain

Pain intensity in the neck–shoulder region (NSP) the past month was assessed using an 11-point numeric rating scale (NRS), which ranges from 0 (‘no pain’) to 10 (‘worst pain imaginable’). The worker responded to the question ‘Rate the worst pain you have experienced in your neck/shoulders within the past month?’ The NRS is a reliable and valid instrument for assessing pain intensity34 and is recommended as an outcome in clinical trials.35

Sick leave

Sick-leave due to pain was assessed using a single-item from the validated Outcome Evaluation Questionnaire36: ‘Within the past month, how many days have you been absent from work due to pain in muscles or joints?’ with response categories ranging from 0 to 31 days. Based on a recent meta-analysis, self-reported sick leave demonstrates good test–retest reliability and reasonably high convergent validity against records.37

Work ability

Work ability was assessed using a validated single item22 from the work ability index.38 The worker responded to the question ‘Please rate your present work ability?’ with response categories ranging from 0 (unable to work) to 10 (‘work ability as its best’). A score ≤7 denotes poor work ability.39

Assessment of possible confounders and effect modifiers

Theoretical assumptions and empirical evidence were used to select possible confounders and effect modifiers which were accounted for in the statistical analyses. The following variables were measured at baseline as previously described30: age (years) and gender (male or female) based on civil registration number, body mass index (BMI) based on objectively measured height and weight, occupational sector (manufacturing, cleaning, transportation and administration/office work within the same workplaces) and seniority in the current job (years). Physical work load was measured by the question ‘How physically demanding do you normally consider your present work?’ using a 10-point (1–10) response scale modified from Borg,40 with higher values indicating higher physical demands. Multisite pain was measured based on the Standardised Nordic Questionnaire for the analysis of musculoskeletal symptoms41 asking about pain intensity (NRS, scale 0–10) during the past 3 months in seven different anatomical areas (ie, neck/shoulders, elbows, hands/wrists, lower back, hips, knees and feet/ankles). A cut-point of >2 was used to indicate the occurrence of pain, whereby the number of pain sites was determined.42 Since the relationship between NSP, sick leave and work ability may depend on the level of physical demands at work, physical work load was considered both a confounder and an effect modifier, while the other variables were solely included as possible confounders.

Statistical analyses

Growth trajectories (latent classes) of the intensity in NSP were identified using LCGA in Latent Gold V.5.1 (Statistical Innovations Belmont, MA, USA). The LCGA procedure and the resulting trajectory classes of NSP are described in a previous study on the same population.30 In brief, the LCGA assigns individuals to latent classes based on maximum posterior probabilities. That is, using growth parameters (ie, intercept, slope and residual variance) reflecting change in an observed variable (ie, pain intensity) over time, LCGA assigns individuals to latent classes (categorical variable) assuming homogeneity within class and heterogeneity between classes.29 43

The LCGA models were performed using Time (14 waves over 1 year) as a continuous linear predictor and NSP intensity as a continuous dependent variable. Missing values were considered as missing at random and included in all models without imputations. The optimal number of classes was determined based on appropriate model fit indices, that is, Bayesian information criterion (BIC), entropy and bootstrap log likelihood ratio (BLRT), which were obtained in consecutive LCGA models with 1–10 a priori class solutions. Then, the models were evaluated based on the estimated growth parameters and clinical distinction between the identified classes. The identified trajectories of NSP were used as an independent variable in further statistical analyses of associations with sick leave and work ability over the same year.

The association of NSP trajectories with sick leave (SMS at 4-week intervals over 1 year) and work ability (SMS each quarter) during the same year was determined using generalised estimation equation (GEE) models with an auto regressive first order covariance structure to account for weaker correlations with increasing distance between time points.

The association of NSP trajectories with sick-leave (days/month) was tested using GEE models with a Poisson distribution and a log link function. Fixed factors were NSP trajectory class (categorical variable), time (continuous variable, 14 time points) and the interaction between trajectory class and time. The interaction term was kept in the model if it was significant (p<0.05). We tested both linear and quadratic time trends, but decided on a linear model since the quadratic model did not improve model fit. The primary GEE models were constructed in three steps: (model 1) unadjusted, (model 2) adjusted for age, gender and BMI, and (model 3) additionally adjusted for occupational sector (four categories) and physical work load (continuous variable). Then, we tested if the results of the primary models were consistent when accounting for baseline pain intensity and comorbidity of multisite pain, as they may be associated with both NSP trajectories and the outcomes. Thus, secondary models were estimated with additional adjustment for the intensity of NSP (model 4) and the number of pain regions (model 5) at baseline (past 3 months). Finally, to test for potential effect modification by physical work load, model 3 was re-run with an interaction between NSP trajectory class and physical work load (model 6).

The association of NSP trajectory class with work ability (dependent variable, ordinal scale) was tested using GEE models with a multinomial distribution and a cumulative logit link function. Fixed factors were NSP trajectory class (categorical variable), time (linear continuous variable, four time points), and the interaction between trajectory class and time, which was kept only if reaching significance (p<0.05). The models were constructed with and without adjustment for covariates, as explained above.

Additional GEE models were constructed to investigate the within-person association of temporal fluctuations in the intensity of NSP with sick leave and work ability (outcomes). To partition the within and between subjects variances in the predictor (intensity of NSP), the mean pain intensity score across all time points was determined for each individual. Then, the person mean pain score was subtracted from each repeated pain rating and used as an independent variable (within subject effect), while adjusting for the person mean pain as a covariate (between subjects effect).43 Thus, these GEE models were constructed with three fixed factors: (1) time (14 waves), (2) the person mean pain intensity across waves and (3) the difference from the mean pain intensity during each wave for the individual (within subject effect). The within-subject effect would then reflect the population averaged association with the outcome for each unit change in pain intensity per month for the individual. Model specifications for sick leave and work ability were the same as above. The models were estimated with and without adjustment for potential confounders as explained above.

The GEE models were estimated using SPSS software V.22 (IBM, USA). For each model, we derived the exponential estimate, that is, relative risk (RR) and OR for sick leave and work ability, respectively, and 95% CIs. P values <0.05 were considered significant.

Patient and public involvement

No patients or public were directly involved in setting the research questions and outcomes, design and conduct of this study, or interpretation of the results. Results will be disseminated to the participants at http://www.nfa.dk/.

Characteristics of the study population

Characteristics of the study population are shown in table 1. The study sample consisted of both males and females, and most were blue-collar workers. The workers were on average of middle age, slightly overweight, and had been in their current job for 14 years. The mean intensity of NSP across all time points was 2.6 (scale 0–10). Most of the workers rated high work ability (>7, scale 0–10) across the study period, while the average worker accumulated 6 days on sick leave due to pain over 1 year, although with a considerable dispersion between individuals. Compliance to the repeated measurements (text messages) in the study was high; on average, the workers had 1.2 missing responses to pain and sick leave (14 waves) and 0.4 missing responses to work ability (four waves) (table 1). The amount of missing data increased over time. During the second wave, the response rates were 96%, 96% and 93% for pain, sick leave and work ability, respectively, while the response rates dropped to 87%, 88% and 90% during the last wave.

Identified latent trajectories of NSP

Based on model fit indices (BIC, entropy and BLRT) and distinction between classes obtained from consecutive LCGA models, a six-class solution was chosen.30 The growth pattern and prevalence (%) of the six identified trajectory classes of NSP were characterised as follows (see also online supplementary figure A): class 1, asymptomatic (11%); class 2, very low NSP (10%); class 3, low recovering NSP (18%); class 4, moderate fluctuating NSP (28%); class 5, strong fluctuating NSP (24%) and class 6, severe persistent NSP (9%). The trajectory classes with lower intensities of NSP (classes 1–3) did not differ in total days on sick leave or mean work ability over 1 year; and the occurrence of sick leave was very low, while work ability was high across the three classes. Thus, classes 1–3 were merged into a single category (low NSP 39%), which was used as a reference in further analyses.

Association between NSP trajectory classes and sick leave over 1 year

Estimates of the association of NSP trajectories with sick leave due to pain (days/month) are shown in table 2. There was no significant time effect on sick leave (unadjusted RR=1.02, 95% CI 0.99 to 1.06). Thus, the interaction between NSP trajectory class and time was discarded from the models.

Based on the fully adjusted model (table 2, model 3) referencing low NSP, the RR of sick leave increased for moderate (RR=3.1, 95% CI 1.8 to 5.5), strong fluctuating (RR=7.6, 95% CI 3.9 to 14.7) and severe persistent NSP (RR=13.8, 95% CI 6.7 to 28.5). On average, the estimated number of days on sick leave per month was 0.1 (95% CI 0.1 to 0.15) for low NSP, 0.3 (95% CI 0.2 to 0.4) for moderate NSP, 0.8 (95% CI 0.5 to 1.2) for strong NSP, and 1.5 (95% CI 1.0 to 2.4) for severe persistent NSP. Predicted values of sick leave for each wave over 1 year are shown in figure 1.

• Download figure
• Open in new tab
• Download powerpoint

Figure 1

Mean 1-year trajectories of days on sick leave obtained in the fully adjusted model in each trajectory class of neck–shoulder pain (NSP). The x-axis represents the 14 pain ratings over 1 year. The y-axis represents the mean predicted number of days on sick leave per month.

Additional adjustment for multisite pain showed similar estimates for moderate NSP (RR=3.0, 95% CI 1.7 to 5.4), strong fluctuating NSP (RR=7.2, 95% CI 3.5 to 14.5) and severe persistent NSP (RR=13.1, 95% CI 6.1 to 28). Also, adjustment for baseline pain intensity revealed stronger estimates for moderate NSP (RR=3.7, 95% CI 2.1 to 6.8), strong fluctuating NSP (RR=10.9, 95% CI 5.1 to 23.7) and severe persistent NSP (RR=24.5, 95% CI 10.4 to 57.9), although with wider CIs.

There was no significant interaction between NSP trajectory class and physical work load on sick leave.

Association between NSP trajectory classes and work ability over 1 year

Estimates of the association between trajectories of NSP and work ability (measured each quarter) are shown in table 2.

There was a small time effect indicating reduced work ability over the year (unadjusted OR=1.05, 95% CI 0.99 to 1.11). There was no interaction between NSP trajectory class and time, whereby this interaction was discarded from the model.

NSP trajectory class was inversely associated with work ability (table 2). Based on the fully adjusted model referencing low NSP (table 2, model 3), the likelihood of a 1-unit reduction in work ability increased for moderate (OR=2.4, 95% CI 1.8 to 3.2), strong fluctuating (OR=8.1, 95% CI 5.9 to 11.2), and severe persistent NSP (OR=12.9, 95% CI 8.5 to 19.7). The median scores (IQR) of work ability during the year were 9.5 (1.7) for low NSP, 9.0 (2.0) for moderate NSP, 7.8 (2.0) for strong NSP and 7.1 (2.5) for severe persistent NSP.

Additional adjustment for multisite pain showed slightly weaker estimates for moderate NSP (OR=2.2, 95% CI 1.7 to 2.9), strong fluctuating NSP (OR=6.4, 95% CI 4.6 to 9.0) and severe persistent NSP (OR=10.4, 95% CI 6.7 to 16.0). Further, adjustment for baseline pain intensity showed similar estimates for moderate NSP (OR=2.6, 95% CI 1.9 to 3.5), strong fluctuating NSP (OR=9.1, 95% CI 6.3 to 13.1) and severe persistent NSP (OR=15.4, 95% CI 9.3 to 25.5).

There was no significant interaction between NSP trajectory class and physical work load on work ability.

Within-person association of NSP with sick leave and workability

Due to the high work ability scores and low prevalence of sick leave in the classes with lower intensities of NSP, these analyses included only workers assigned to the trajectory classes strong fluctuating and severe persistent NSP (n=248).

The within-person associations of pain intensity with sick leave due to pain and work ability are shown in table 3.

Within-person fluctuations in the intensity of NSP were positively associated with sick leave (adjusted RR=1.11, 95% CI 1.02 to 1.21). That is, higher intensity of NSP was associated with more days of sick leave during a particular month at the individual level (table 3 and figure 2A).

• Download figure
• Open in new tab
• Download powerpoint

Figure 2

Association between temporal fluctuations in neck–shoulder pain (NSP) intensity and the outcomes sick leave and work ability. The x-axis represents the difference in pain intensity scores from the person mean pain intensity across time points. The y-axis represents the predicted number of days on sick leave per month (A) and the predicted cumulative probability of poor work ability (B), as defined by the cut-point ≤7 (scale 0–10).39

A similar within-person association was found between fluctuations in pain intensity and work ability (adjusted OR=1.13, 95% CI 1.04 to 1.21). For example, increasing intensity of NSP was associated with higher probabilities of reduced work ability (table 3). This association is illustrated in figure 2B as the estimated probability of reporting poor work ability (≤7 on the 0–10 scale).39

Trajectories of NSP, sick leave and work ability

The trajectory classes of NSP used in this study were distinguished using LCGA, resulting in six distinct trajectories of NSP.30 This corroborates a study by Lövgren et al (2014) which used Growth Mixture Modelling to identify six trajectory classes of NSP in nursing students entering working life. In contrast, the severity of NSP in the current sample of workers was much higher, perhaps due to the large proportion of blue-collar workers in this study.1 The high prevalence of strong fluctuating (24%) and severe persistent NSP (9%), with mean pain intensities of 5 and 7 (scale 0–10), respectively, is noteworthy.

Trajectory class of NSP was strongly associated with the number of days on sick leave over the 1-year study period. Particularly, the fully adjusted model indicated an increased RR of sick leave in the trajectory class with severe persistent NSP (mean 1.5 days/month), compared with low NSP (mean 0.1 days/month). This result is in line with previous prospective studies showing a positive association between NSP intensity and sick leave.4 6 This result persisted with adjustment for potential confounding by multisite pain, which was associated with NSP and sick leave in previous studies.19 44 45 Thus, severe persistent NSP appears to be strongly associated with sick leave due to musculoskeletal pain, regardless of multisite pain and other personal and occupational factors.

The four trajectory classes of NSP were also associated with poor work ability over the year (table 2). For instance, the probability of reporting reduced work ability was 13 times higher for the trajectory class with severe persistent NSP (median work ability 7.1 on the 0–10 scale), compared with low NSP (median work ability 9.5), regardless of inclusion of potential confounders in the model (table 2, model 3). In agreement, previous studies have found that intense NSP is associated with reduced work ability in workers,3 46 although none of these examined pain trajectories. Interestingly, including baseline intensity of NSP (ie, past 3 months) as an additional covariate in model 3 did not reduce the estimated association for sick leave or work ability. In fact, this adjustment resulted in even stronger estimates for work ability, which clearly indicates that the LCGA-based trajectory classes of NSP have a predictive value beyond that explained by past pain intensity assessed at a single time point.

The observed consistent associations between the identified trajectories of NSP and the outcomes sick leave and work ability support the prognostic value of LCGA-based trajectories of NSP among workers, and suggests that such sub-populations can be of clinical and occupational relevance. Thus, this study supports using LCGA to identify distinct sub-populations of workers with different patterns of NSP. Further, the increased risk of sick leave and poor work ability for severe persistent NSP points to the need for interventions and preventive strategies aiming at reducing severe persistent NSP in working populations.

Effect modification by physical work load

High physical work load is a known risk factor for incident NSP and has been associated with a poor prognosis.47 Thus, the association for NSP trajectories with sick leave and work ability was expected to be modified by the level of self-reported physical work load at baseline. However, we could not confirm any interaction between trajectory class of NSP and physical work load neither for sick leave nor work ability. Still, it is possible that more precise technical measurements of physical work exposure would have yielded different results.

Association of temporal fluctuations in NSP with sick leave and work ability

NSP is often referred to as a recurrent and fluctuating condition.11 However, the temporal fluctuations in NSP have rarely been investigated in detail, and few, if any, studies have examined whether fluctuations in NSP are associated with sick leave and work ability. The frequent repeated measures allowed us to assess the within-person temporal association between fluctuations in the intensity of NSP and sick leave (every fourth week) and work ability (every 12th week). We found that the within-person fluctuations in pain intensity were significantly associated with both sick leave and work ability, which persisted in the fully adjusted model (table 3). That is, when the individual pain score increased, the likelihood of sick leave and poor work ability also increased. Thus, not only could we address the differences between workers in temporal patterns of NSP, but also whether the temporal fluctuations in NSP at the individual level are of predictive value.

Methodological discussion

As this study is limited to self-reported measures of pain and the outcomes sick leave and work ability, one cannot overlook the possibility of bias. Regarding self-reported sick leave, meta-analytic evidence indicates reasonably high convergent validity against organisation and register-based records, although with a slight tendency for under-reporting.37 Thus, there is a risk for underestimation of sick leave in this study. The question about days on sick leave the past month did not distinguish between work days and non-work days, which may have resulted in less precise estimates. Also, although both outcomes sick leave and work ability were assessed prospectively over 1 year, the NSP trajectories were determined during the same time period and thus causal inferences should be made with caution. Still, it seems most likely that pain preceded the occurrence of sick leave, rather than the reverse relationship. The compliance to text messages was very high on average (table 1). Still, missing data increased slightly over time, which might have introduced some uncertainties in the models.

We addressed several relevant factors as confounders or effect modifiers (physical work load) of the association of pain trajectories with sick leave and work ability. However, as the causes of sick leave and poor work ability are likely multifactorial,4 19 21 the focus on pain trajectories as predictors is a potential limitation because it does not inform about other potentially important prognostic factors. Also, the possibility of residual confounding by non-measured factors cannot be ruled out. For instance, we did not measure comorbidity of chronic conditions that might be associated with both NSP trajectories and the outcomes.

Since this study was conducted in a non-random sample with a predominance of blue-collar workers, it is important to verify the study findings in other working populations. The general notion of NSP as a fluctuating and reoccurring condition11 is corroborated by our data, but this is rarely taken into account in observational studies on NSP. The inclusion of residual variance in the LCGA model allowed us to distinguish trajectory classes with more or less fluctuating patterns of NSP. Temporal fluctuations were more prominent in the NSP trajectories with moderate and strong pain, which may have contributed to the lower risk of sick leave and poor work ability compared with the trajectory class with severe persistent NSP.

Since the trajectories with lower intensities of NSP, that is, including asymptomatic, very low NSP and low recovering NSP, did not differ regarding their low occurrence of sick leave and poor work ability, we decided to merge them into a single reference category in the prediction models. Still, it is possible that these three classes differ in other prognostic outcomes.