Predictors of chronic pain
area corresponds to 1 pain site; hence, the maximum number of
pain sites was 45. Based on these 45 pain sites, 23 anatomical
regions were determined and a total pain index, ranging from 0
to 23, was considered (22). Using a slightly modified definition
developed by MacFarlane et al. (33), widespread pain (WSP)
was defined as pain in at least 2 sections in 2 contralateral
limbs and the axial skeleton and marked equally on the front
and back of the chart. MacFarlane et al. defined WSP in the
limbs to be present “if there are at least 2 painful sections (in
2 contralateral limbs)”, a definition that does not require pain
to be marked equally on the front and back of the chart (33).
Therefore, the current study uses a stricter definition of WSP
than the American College of Rheumatology criteria (34). In ad-
dition to WSP, the following categories were defined: 1: No pain
(NP) if the participants reported zero anatomical sites with pain
(i.e. pain index = 0) and answered “no” to the question “Do you
frequently (‘usually’) have pain lasting more than 3 months?”
and did not report on pain intensity above (this group served
as the reference; category = 0); 2: Local pain (LP) if the partici-
pants reported 1–2 anatomical sites; 3: Regional Pain-Medium
(RP-Medium) if the participants reported 3–6 anatomical sites
with pain; and 4: Regional Pain-Heavy (RP-Heavy) if the
participants reported 7–17 anatomical pain sites with pain, but
which did not fulfil the WSP criteria. WSP, a common clinical
entity, depends both on the number of pain sites and their spatial
distribution. As discussed elsewhere (14), a minority of subjects
with RP-Heavy could have a higher number of pain sites than
some of the subjects with WSP (14); however, in this study
the 95% confidence intervals (95% CI) for the number of pain
sites clearly differed between RP-Heavy and WSP (Table SI 1 ).
Pain sensitivity. Pain sensitivity was assessed by all participants
using the Pain Sensitivity Questionnaire (PSQ), which consists
of 17 items that each describe a daily life situation (35). The
PSQ asks all participants to rate how intense pain in each situa-
tion would be for them on an NRS (ranging from 0 = not painful
at all to 10 = worst pain imaginable). Whereas 14 of the items
relate to situations that are assessed as painful by a majority of
healthy subjects, 3 items describe situations that are usually not
rated as painful (e.g. taking a warm shower). These 3 items are
interspersed between the items to serve as non-painful sensory
reference for the individuals and were not considered when
calculating the final score. The items cover a range of pain in-
tensities, a variety of pain types (e.g. hot, cold, sharp, and blunt),
and body sites (head, upper extremity, and lower extremity). The
mean of the 14 items mentioned above (relate to situations that
are assessed as painful by most healthy subjects) was calculated
(range 0–10). In this study, we used the Swedish adaptation of
the PSQ (22). The Cronbach’s alpha was 0.93 (22). However, the
Swedish version of the PSQ has so far not been validated. Since
reliability, content, structural validity, and hypothesis testing
regarding (36) the PSQ were quite consistently good across 3
investigated languages and cultures (35, 37, 38), there is good
reason to believe the same for the Swedish version.
Data analysis
All statistical analyses were performed using IBM SPSS Statis-
tics (version 23.0; IBM Inc., New York, USA). The sampling
weights for unequal possibilities of sample selection have been
reported elsewhere (14) Two-sided statistical tests were used
and p < 0.05 was regarded as significant. Distributions and de-
scriptive statistics were examined for all variables for the total
http://www.medicaljournals.se/jrm/content/?doi=10.2340/16501977-2519
1
185
sample at T0 and T1. Means and standard deviations (SDs)
were calculated for continuous variables, and frequencies with
percentages (n; %) were calculated for categorical variables.
The prospective 2-year follow-up analysis of each pain out-
come (i.e. pain intensity, spread and sensitivity) was performed
through a series of generalized linear models (GLM) used as
prediction models with baseline variables as predictors. GLM is
a flexible generalization of ordinary linear regression analyses
that allow for response variables that have error distribution
models other than a normal distribution (39). In the case of pain
intensity and pain sensitivity, which served as linear-response
data, the identity link function was used with maximum likeli-
hood estimation (MLE) and results are presented as parameter
estimates (B) with Wald 95% CIs. In the case of the outcome
of the pain spreading categories, which served as an ordinal
response variable, the ordinal logit function was used, yielding
odds ratios (OR) and Wald 95% CIs. The significance of the
estimated effects in all GLM analyses was evaluated using the
Wald test (39). For analytical purposes in the prediction models,
the categorical variables were dichotomized and entered as
follows: sex (female = 1), country of birth (abroad = 1) citizen-
ship (other = 1), marital status (married = 1), educational level
(university, i.e. higher education=1), employment status (unem-
ployment = 1), and existence of a certain comorbidity (yes = 1).
To determine the variables to include in the multivariable
model, separate univariable analyses were performed with each
independent variable 1 at a time (single predictor model). Only
variables with p < 0.20 in the single model were included in the
multivariable model. Next, we examined for multicollinearity
between the significant independent variables as derived by
univariable analyses by performing linear regression and by
examining tolerance and the variance inflation factor (VIF). A
tolerance of less than 0.20 or 0.10 and/or a VIF of 5 or 10 and
above indicates a multicollinearity problem (40). In the case
of multicollinearity, we also performed Pearson correlation (r)
analysis to test for bivariate correlations between the continu-
ous variables and phi coefficient (41), a measure of association
between the binary variables. High bivariate correlation coeffi-
cients ≤ 0.7 or phi (Φ) ≤ 0.3 (39, 40) indicate risk of collinearity;
in that case, only 1 of the 2 highly correlated variables were
included. Hence, citizenship as well as anxiety were excluded
from all multivariable models due to high correlation with birth
country (Φ = 0.45) and depression (Φ = 0.70), respectively. After
exclusion of the above variables, tolerance and variance inflation
revealed that there was no serious indication of multicollinearity.
Birth country and depression were kept in the model as they
exhibited more pronounced parameter estimates in univariable
analysis than citizenship and anxiety.
Two multivariable prediction models are presented. In model
1, all selected baseline (T0) variables according to the p-value
criteria of p < 0.20 with respect to the single predictor model (with
exception of highly correlated variables) were included in 1 mul-
tivariable model. In model 2, all variables from the multivariable
model 1 along with the 3 pain characteristics for each outcome of
interest were simultaneously controlled for. Hence, pain intensity,
spread of pain and sensitivity at T0 were entered in the multiva-
riable model 2, and only complete cases were included. These
2 models address different, but interrelated questions: 1: Which
socio-demographic features and comorbidities at baseline (T0)
predict pain at follow-up (T1)? And: 2: How do these predictive
associations change when differences in pain present already at
baseline are adjusted for? Model 1 thus provides estimates of
cumulative associations with pain, whereas the adjustment for
pain at baseline in model 2 implies that associations with changes
in pain occurring during follow-up are estimated.
J Rehabil Med 51, 2019