Multigroup latent class model of musculoskeletal pain combinations in children/adolescents: identifying high-risk groups by gender and age

This cross-sectional analytical study was performed in the city of Tabriz, Iran. Tabriz is the capital of East Azerbaijan province in North-western Iran and the fourth major city in Iran with a population of about 1.58 million. School children aged 11–15 years and were in grades 6–8 were included in the study. A three-stage sampling method was utilized to acquire a representative sample of schoolchildren. In the first stage, five educational districts were chosen as strata. In the second stage, a total of 40 schools were selected randomly from these districts, which comprised four girl’s schools and four boy’s schools. Finally, within each school, the study participants were selected randomly by simple random sampling scheme. The sample size of the study was calculated to be approximately 1105 cases, based on the main outcome of the study, the prevalence of MSP (the minimum about 7% to reach maximum sample size), considering 95% confidence, a precision about 20% and using sample size formula on single proportion in the descriptive studies. Taking into account the design effect about 1.5, the sample size increased to 1611 cases. The samples were allocated to each district and school based on a proportion to size scheme.

Before any participation, permission was taken from the education department of Tabriz as well as the school authorities involved. Additionally, the parental permission and informed consent for school children was also taken asking for them to contribute in the study. The study was approved by institutional review board of Tabriz university of medical sciences (code: IR.TBZMED.1394.1166). The research was conducted in accordance with the Declaration of the World Medical Association and the Helsinki Declaration on human subjects testing.

The data were collected using a checklist consisting of demographic and MSP data after confirmation by ethics committee. The demographic variables included gender (girls: 1, boys: 2) and age (years). The data on MSP included neck, shoulder, upper back and low back complaints during the preceding month was evaluated using a pre-shaded manikin picture showing the pain location and the modified standardized Nordic Musculoskeletal Questionnaire (NMQ) [16]. The response alternatives were: “no” and “yes”. The Farsi version of the NMQ with an established reliability and validity was used in the study [17].

Data were analysed using LCA utilizing SAS 9.4 PROC LCA software to find the clusters of individuals by similar pattern of reported MSP [18]. The reports of symptoms were dichotomized as neck pain (no = 0, yes = 1), shoulder pain (no = 0, yes = 1), low back pain (no = 0, yes = 1), and upper back pain (no = 0, yes = 1). Simple LCA was used to estimate the probability and class membership. The study participants were assigned to the latent class clusters with the highest posterior probability as calculated based on Bayes’ theorem. A successful model fit was evaluated by a Chi-square Goodness of Fit (GOF) statistic p > 0.05. The conditional independence assumption was intuitively fulfilled since the measures of symptoms in anatomical sites were independently assessed. Multigroup LCA was also used to assess various features of data taking into account the effect of gender (girls = 1 and boys = 2) and age (≤ 13 years = 1 and ≥ 14 years = 2) as well as gender–age combinations (girls ≤13 years = 1, girls ≥14 years = 2, boys ≤13 years = 1, and boys ≥14 years = 4) as the grouping variables. The multigroup models were fitted based on three number of classes obtained in the first step taking into account the effect of gender, age as well as gender–age as the grouping variables in the model. To achieve an appropriate number of classes, a sequence of models were fitted with two, three and four number of classes, and the optimal model was selected by Akaike’s Information Criterion (AIC) [19] and Bayesian Information Criterion (BIC) [20]. A smaller AIC and BIC for a particular model proposes a preferable model. A 5% decrease in these measures was used to reach the conclusion. In addition, model interpretability was also considered (e.g. on the basis of the item-response probabilities, each class was distinguishable from others and no class was trivial in size) so that a meaningful label was assigned to each class. In multiple-group LCA, both the class memberships and item response probabilities can change across groups, and therefore measurement invariance can be empirically verified across groups. The parameters were estimated by the maximum likelihood procedure using the EM algorithm. The Newton-Raphson technique was used to estimate the parameters. The convergence index used was the maximum absolute deviation (MAD) to achieve the iteration limit. Missing data on the latent class and latent status indicators were permitted and treated under the assumption that data were missing at random (MAR). Any record with missing data on grouping variables specified in the model was eliminated from the analysis.

A total of 1611 school children participated in the study, of which the percentage of gender–age combinations were 29.8%, 23.1%, 22.9% and 24.2% for girls ≤13 years, girls ≥14 years, boys ≤13 years and boys ≥14 years of age categories, respectively.

The mean weight of the participants were 49.7 (SD 11.36) kg; [girls ≤13 years: 48.1 (SD 10.5) kg, girls ≥14 years: 47.2 (SD 11.7) kg, boys ≤13 years: 51.1 (SD 10.1) kg and boys ≥14 years: 53.2 (SD 12.2) kg] and BMI, 19.9 (3.83) kg/m² [girls ≤13 years: 19.6 (SD 3.8) kg/m², girls ≥14 years: 20.0 (SD 4.2) kg/m², boys ≤13 years: 19.8 (SD 3.3) kg/m² and boys ≥14 years: 20.3 (SD 4.0) kg/m²]. More details on characteristics of the physical leisure activity, use of schoolbags and psychological trait among the schoolchildren are presented in [5, 6].

The prevalence of neck, shoulder, low back, and upper back symptoms were 27.9% (95% Confidence Interval (CI): 25.8–30.2%), 20.7% (95% CI: 18.8–22.8%), 34.3% (95% CI: 32.0–36.7%) and 19.0% (95% CI: 17.1–21.0%), respectively. Also 31.5% (95% CI: 29.3–33.9%) of participants reported pain in any of the four body regions.

Table 1 shows that prevalence of MSP in gender–age categories in different body regions. The results for pain at one location showed a significantly lower prevalence of shoulder symptoms in boys aged ≥14 years and higher prevalence of low back symptoms in boys and girls aged ≤13 years (p < 0.05) than in other gender–age categories. The results for pain at 2 locations showed significantly lower prevalence of neck/shoulder and shoulder/upper back symptoms in boys aged ≥14 years as well as higher prevalence of shoulder/low back and low back/upper back symptoms in boys and girls aged ≤13 years (p < 0.05) than in other gender–age categories. The results for pain at 3 locations and pain at all 4 locations showed significantly lower prevalence of symptoms in boys ≥14 years of age category (p < 0.05) than in other groups.

A smaller AIC (32.36) and BIC (107.75) for a model by three classes support this model compared to the models with two and four number of classes. Additionally, each class was distinguishable from others on the basis of item-response probabilities, so that a meaningful label was assigned to each class. The first class included those participants (60.2%) with small/moderate value (around 0.3 or lower) of probability of pain in all anatomic locations which was assigned as moderate pain class. The second class consisted of participants (7.1%) with high value (> 0.65) of probability of pain in all anatomic locations, which was named as major pain class. A small probability of pain (around 0.1 or lower) was observed in the third class with 32.7% of participants as minor pain class (Table 2).

The results of gender–specific LCA modeling are presented in Table 3. As can be seen from this table, the minor pain class consisted of 30.7% of girls and 34.5% of boys. With the same pattern of difference, the moderate pain class involved 57.9% and 61.6% of girls and boys, respectively. While the major pain class had a higher rate of pain in girls (11.4%) compared to boys (4.0%), the test of invariance across gender groups showed no significant difference between girls and boys in the pain clusters (p > 0.05). The findings indicated that the likelihood of shoulder and upper back pain in the major pain class was considerably higher in boys than girls.

Table 4 shows the results of age–specific LCA modeling. Based on the test of invariance, there was significant difference across children ≤13 years and ≥ 14 years of age categories in membership probability (p < 0.05). The minor pain class involved a higher percentage of children aged ≥14 years (61.4%) compared to those aged ≤13 years (29.1%). In the moderate pain class, children aged ≤13 years (63.9%) were more probable to experience pain than those aged ≥14 years (33.3%). The major pain class consisted of 7.0% and 5.3% of children aged ≤13 years and ≥ 14 years, respectively. The results showed considerable differences in the neck and low back pain in the moderate pain class as well as neck pain in the major pain class across age categories.

The results of gender and age (combined)-specific LCA modeling are presented in Table 5. The results indicated that boys (77.8%) and girls (59.7%) aged ≥14 years were the most probable groups for experiencing MSP in the minor and moderate pain classes, respectively. Girls (30.0%) and boys (34.7%) aged ≤13 years were the most probable groups to experience MSP in the major pain class. However, the test of invariance across gender–age categories showed no significant difference between gender–age categories in the pain clusters (p > 0.05). Nevertheless, the item response probabilities showed considerable differences across gender–age categories. In the major pain class, girls aged ≥14 years were more likely to experience pain in all four anatomical areas than other groups. In the moderate pain class, the likelihood of neck and shoulder complaints in boys aged ≥14 years and also low back pain in girls aged ≥14 were higher than in other categories.