Prognostic Factors and Treatment Effect Modifiers for Physical Health, Opioid Prescription, and Health Care Utilization in Patients With Musculoskeletal Disorders in Primary Care: Exploratory Secondary Analysis of the STEMS Randomized Trial of Direct Access to Physical Therapist–Led Care

Abstract Objective The aims of the study were to identify prognostic factors associated with health care outcomes in patients with musculoskeletal (MSK) conditions in primary care and to determine whether characteristics associated with choice of care modify treatment effects of a direct-access physical therapist–led pathway in addition to general practitioner (GP)–led care compared to GP-led care alone. Methods A secondary analysis of a 2-parallel-arm, cluster randomized controlled trial involving general practices in the United Kingdom was conducted. Practices were randomized to continue offering GP-led care or to also offer a direct-access physical therapist–led pathway. Data from adults with MSK conditions who completed the 6-month follow-up questionnaire were analyzed. Outcomes included physical health, opioid prescription, and self-reported health care utilization over 6 months. Treatment effect modifiers were selected a priori from associations in observational studies. Multivariable regression models identified potential prognostic factors, and interaction analysis tested for potential treatment effect modifiers. Results Analysis of 767 participants indicated that baseline pain self-efficacy, pain severity, and having low back pain statistically predicted outcomes at 6 months. Higher pain self-efficacy scores at baseline were associated with improved physical health scores, reduced opioid prescription, and less health care utilization. Higher bodily pain at baseline and having low back pain were associated with worse physical health scores and increased opioid prescription. Main interaction analyses did not reveal that patients’ age, level of education, duration of symptoms, or MSK presentation influenced response to treatment, but visual trends suggested those in the older age group proceeded to fewer opioid prescriptions and utilized less health care when offered direct access to physical therapy. Conclusions Patients with MSK conditions with lower levels of pain self-efficacy, higher pain severity, and presenting with low back pain have less favorable clinical and health care outcomes in primary care. Prespecified characteristics did not modify the treatment effect of the offer of a direct-access physical therapist–led pathway compared to GP-led care. Impact Patients with MSK conditions receiving primary care in the form of direct-access physical therapist–led or GP-led care who have lower levels of self-efficacy, higher pain severity, and low back pain are likely to have a less favorable prognosis. Age and duration of symptoms should be explored as potential patient characteristics that modify the treatment response to a direct-access physical therapist–led model of care.

6][37][38] Firstly, the distribution of key variables between patients with complete and incomplete observations was analysed.There were differences in several variables including level of education, age, paid employment and PSEQ (supplementary Table 2).There was no clear pattern of 'missingness' between observations (supplementary Table 3).T-tests for the continuous outcome of physical health (PCS) and chi-squared tests for outcomes opioid use or healthcare utilisation (analysed as a binary outcome 'Yes' or 'No'), were used to compare the distribution of key outcome variables between complete and incomplete cases and provide information around the nature of missing data.There was evidence of a differences between those with incomplete and complete data for the variable pain duration and the outcome healthcare utilisation (p=0.04), and paid employment and the outcome opioid prescription(p=0.036).In addition, relationships between incomplete variables and potential predictor variables were examined using t-tests or chi-squared tests.There was strong evidence (p=0.002) of a difference in the level of deprivation for those with complete or incomplete observations for the variable level of education.These findings helped justify the Missing At Random (MAR) assumption used for multiple imputation in this secondary analysis.

Multiple imputation:
All variables used in analysis modelling, as well as predictive variables (deprivation level) were imputed using a multiple imputation by chained equations approach (MICE). 36 total of 40 datasets were imputed based on an established 'rule of thumb' 37,67 that suggests imputed sets should be at least equivalent to the total percentage of missing data.Continuous variables (PSEQ, physical component score, mental component score, and physical component score at 6 months) were imputed using linear regression.Although PSEQ was non-normally distributed it was imputed without transformation as only means and variance were important for subsequent regression analyses and this method has been reported to introduce limited bias. 68A logistic regression was used to impute binary variables (work status, health literacy) and an ordered logistic regression model was used for ordered categorical variables (global assessment of change, general health at baseline, level of education, bodily pain at baseline and pain duration).A Poisson regression was used for healthcare utilisation imputation.Complete variables included in the imputation modelling were opioid prescription, widespread pain, sex, deprivation, area of pain, comorbidities, age, and treatment arm.A passive approach was used for interaction analyses which has been shown to be comparable to active imputation. 69All variables were imputed together.A total of 40 data sets were imputed based on a 'rule of thumb' established in literature 37,67 that suggests imputed sets should be at least equivalent to the total percentage of missing data.

Imputation model checking:
The multiple imputation model was checked by firstly creating histograms and Kerneldensity plots of the continuous variables in the imputed data sets and checking that values were plausible and, although variable, matched the distribution of the observed data set.PSEQ data had imputed observations outside the expected scale range, but mean and standard deviation of the imputed data sets were acceptable.Given the prior knowledge of its under-dispersed distribution, this was somewhat expected and not corrected for as skewed variables imputed with a conditional normal distribution are known to produce acceptable estimates for means, variances and regressions which is the purpose of these analyses. 68Categorical variables were tabulated to compare proportions against observed data.Kolmogorov-Smirnov tests were carried out on continuous data to compare the distribution of observed and imputed data sets.Abayomi et al suggested any variable with a significant test (p<0.05) is of potential concern. 70No continuous variables in the model had a Kolmogorov-Smirnov test exceeding this value.
VIF results: After computing the VIF, bodily pain at baseline had a VIF >10.Further exploration revealed PCS at baseline impacted the coefficients of multiple variables, inconsistent with the model, and removing this variable resulted in all remaining variables with a VIF <10.
Works Cited: 28.Bishop A, Ogollah RO, Jowett S, et al.STEMS pilot trial: A pilot cluster randomised controlled trial to investigate the addition of patient direct access to physiotherapy to usual GP-led primary care for adults with musculoskeletal pain.BMJ Open.
Approached to participate in the trial (n=5 GP practices)

Analysis Supplementary Table 1 .
Baseline characteristics/potential prognostic factors and measures Based on the American College of Rheumatology's definition 66Supplementary Table6.Variables with most missing data as a percentage of total data collected at 6 months General HealthSelf report, single categorical question "In general, would you say your health is Excellent, Very Good, Good, Fair, Poor" (STEMS questionnaire) Pain self-efficacy Pain Self Efficacy Questionnaire MSK pain presentation Self report, coded body manikin completion (STEMS questionnaire).Areas coded as low back pain labelled 'low back pain', all other conditions labelled as 'all other MSK presentations'.Supplementary Table 2. Healthcare Utilisation by frequency at 6 months (complete cases) Supplementary Table 3. Outcomes for healthcare utilisation by procedure at 6 months (complete cases) Supplementary Table 7. Distribution of the three outcomes for missing and complete observations Supplementary Table 8.Missing data patterns for key variables with the greatest proportion of incomplete observations