Participants
The randomized controlled trial involved 60 individuals with SAPS, recruited from March 2016 to June 2017. Patients were referred by an orthopedic physician to physiotherapy due to shoulder pain. We included patients with positive results for 3 out of 5 SAPS tests: Neer, Hawkins-kennedy, painful arc, pain or weakness resistant to external rotation and Jobe [20]. The exclusion criteria were: history of shoulder trauma or surgery; total rotator cuff or biceps brachii tendon rupture (imaging exam or self-report); sports activities involving the upper limbs; individuals with neurological disorders and alterations in cognitive function (e.g., stroke, epilepsy, multiple sclerosis, Parkinson’s disease, and peripheral neuropathy); shoulder pain for primary involvement in the cervical or thoracic region; systemic disease involving the joints (e.g., rheumatoid arthritis); carpal tunnel syndrome; and underwent physiotherapeutic treatment of the shoulder in the last 6 months [17]. A detailed description of the trial was presented by Hotta et al. [17] (clinicaltrials.gov: NCT02695524). A brief overview of the trial is presented below.
Randomization and interventions
Participants included in the study were randomized into two groups: periscapular strengthening and scapular stabilization exercises. Interventions were carried out for 8 weeks, three times a week, on non-consecutive days. Each session lasted 50 min and individuals were treated separately. Participants assigned to the periscapular strengthening group (PSG) performed only six exercises of periscapular strengthening (upper trapezius, middle trapezius, lower trapezius, and serratus anterior). Participants allocated to the scapular stabilization group (SSG) performed the same six periscapular strengthening exercises applied to PSG, and six scapular stabilization exercises, emphasizing retraction and depression of the scapula, were added to this group [17].
Assessment time points
Patients characteristics, outcome measures, primary and alternative mediators, and potential confounders were measured at baseline prior to randomization. The putative mediators were measured after 4 weeks of the beginning of the treatment. Outcomes were measured right after the end of the treatment (i.e. 8 weeks).
Primary outcome measures
Shoulder disability was assessed by The Shoulder Pain and Disability Index (SPADI). The SPADI is valid and reliable (Cronbach Alpha = 0.89) for the assessment of individuals with shoulder disorders (Martins et al., 2010). The score of the questionnaire ranges from 0 to 100 points, and higher scores indicate higher disability [21]. The minimal clinically important difference (MCID) considered for the questionnaire was 10 points [22].
Pain intensity was assessed by the 0–10 numerical pain rating scale [23, 24]. Changes of 15 to 20% from baseline values were considered clinically relevant [25].
Putative mediators
The primary mediators were scapular motion and position, measured through a digital inclinometer (Lafayette®, Lafayette Instrument Company, Ind., USA) and expressed in degrees. We measured scapular upward/downward rotations and anterior/posterior tilt at rest (scapular position), 90° and 180° of arm elevation (scapular motion). Scapular upward/downward rotations and anterior/posterior tilt measurements presented intra-rater reliability ranging from good to excellent, with standard error of measurement ranging 2 to 2.8 degrees [26, 27] and criterion validity ranging from good to excellent [26, 28]. Other mediators included muscle strength of serratus anterior, upper, middle, and lower trapezius [29]. The evaluation of the isometric strength was performed using a portable dynamometer that has excellent reliability [26, 27] (Lafayette®, Lafayette Instrument Company, Ind., USA).
Potential confounders
We assumed that both the intervention-mediator and intervention-outcome paths were not confounded due to randomization. Given the mediator cannot be randomized, we assumed that the mediator-outcome path might be confounded by pain duration, age, sex and baseline measures of mediators and outcomes. Therefore, we controlled the analyses in the outcome regression models for pain duration, age, sex and baseline measures of mediators and outcomes. Besides that, the outcome regression models for pain and disability, with scapular position and motion as mediators, were also controlled for muscle strength and all outcome models were controlled for the dominant side of complaint. The hypothesized causal pathways are presented in Figs. 1 and 2.
Sample size
The sample size for the main trial was estimated to identify clinically meaningful between-group differences in shoulder function, considering a minimal clinically important difference of 10 points in the global SPADI score, with alpha set at 0.05, power 80%, and a 20% sample loss. The minimum sample size for the original trial was 30 on each intervention arm. The trial had no drop-outs, with 60 participants (30 each group) completing the study.
We conducted a post hoc power calculation as suggest by the literature [30, 31] using the estimator for joint indirect effect developed by Vittinghoff and Neilands [32]. We performed two sample size estimations: (1) one assuming a large treatment-mediator and mediator-outcome effect (r = 0.65); (2) another assuming a moderate treatment-mediator and mediator-outcome effect (r = 0.3). The remaining variables were kept the same for both analysis, and were as follows: absence of exposure-mediator confounding (i.e. error term correlation coefficient = 0.0), given the design of the study (i.e. clinical trial); a moderate confounding for the mediator-outcome (r = 0.3) as suggested by Vittinghoff and Neilands [32]; with power set at 0.8. The coefficients were standardized.
Data analysis
Analyses were performed in R (The R Foundation for Statistical Computing). The causal mediation analysis was performed using the mediation package [33]. A model-based inference approach was used to estimate the average causal mediation effect (ACME), average direct effect (ADE) and the average total effect (Fig. 3) [33, 34]. Two regression models were created: the mediator model and the outcome model. As there was no total effect, we decided to conduct several univariate mediation models to verify where the causal pathway break down. The mediator model was constructed with treatment allocation as the independent variable and the putative mediator as the dependent variable. The outcome model was constructed with the treatment allocation and the putative mediator as independent variables and the outcome as independent variable. The outcome models were adjusted for potential confounders. Continuous mediators and outcomes that were normally distributed were modelled using linear models (lm). However, if the data was skewed or the assumptions of the linear model were violated, mediators and outcomes were modelled using robust linear models22 or generalized linear models (glm) with respective family and link function [35].
The mediate function was used to estimate the value of the mediator and outcome. The simulated potential values of the mediator and outcome was used to compute the ACME, ADE and average total effect. We used 1000 bootstrap simulations to generate 95% confidence intervals (95% CI) if linear assumptions of mediator and/or outcome models were not violated. Non-parametric bootstrap simulations were used if the linear assumptions of the mediator and/or outcome models were violated.
We performed sensitivity analysis for unmeasured confounding to assess how a hypothetical level of unmeasured confounding would impact on ACME. We used the medsens function to explore the level of confounding due to unknown confounders from the mediator ant outcome models. The level of confounding [ρ (rho)] is represented by the correlation between the error terms from the mediator and outcome models. The level of confounding ranged from − 1 to 1. A ρ = 0 suggests no correlation between error terms and can be interpreted as the absence of unmeasured confounding [34].