Data analysis
Descriptive statistics for demographic and clinical characteristics for
both groups and the total group were used. Working hours per week,
and sick leave were analyzed only for patients who were at work.
Differences after 5 and 12 weeks were calculated for between-groups
comparisons and within-groups results according to the “intention-totreat
principle”. These results have been published previously (13).
Because of the unbalanced structure of our repeated-measures design
and the assumed correlation of observations in longitudinal data-sets
we used a linear mixed models approach for calculating differences
between baseline and our final follow-up at 52 weeks. This method
uses both fixed and random effects in the same analysis. It handles
naturally unbalanced data as, for example, uneven spacing of repeated
measures, and allows analysis of the relationship of predictor covariates
with the dependent variable. It also accounts successfully for the
observed pattern of dependences in those measurements. Appropriate
covariates were identified in a univariable regression analysis and from
literature. Before starting the analysis, the baseline SPADI score and
all identified covariates were centred by subtracting
the group mean.
In a first step a fixed effects model was run and in a
second step random effects were added. Insignificant
covariates were then stepwise removed from the model.
Model fit was assessed with the help of the Bayesian
Information Criterion (BIC) and the -2Log likelihood.
Costs recorded in the shoulder log-book were valued
using published prices for medical costs. Productivity
costs were calculated by applying the friction costs
method (24). Depending on data distribution betweengroup
differences in outcomes of total costs were
analyzed by Student’s t-tests for unpaired observations
or the Mann-Whitney U test.