[1] 14w The fear-avoidance model of chronic pain: Validation and age analysis using structural equation modeling

[1] 305w The cognitive-behavioral, fear-avoidance (FA) model of chronic pain (Vlaeyen JWS, Kole-Snijders AMJ, Boeren RGB, van Eek H. Fear of movement/(re)injury in chronic low back pain and its relation to behavioral performance. Pain 1995a;62:363-72) has found broad empirical support, but its multivariate, predictive relationships have not been uniformly validated. Applicability of the model across age groups of chronic pain patients has also not been tested. Goals of this study were to validate the predictive relationships of the multivariate FA model using structural equation modeling and to evaluate the factor structure of the Tampa Scale of Kinesiophobia (TSK), levels of pain-related fear, and fit of the FA model across three age groups: young (640), middle-aged (41-54), and older (P55) adults. A heterogeneous sample of 469 chronic pain patients provided ratings of catastrophizing, pain-related fear, depression, perceived disability, and pain severity. Using a confirmatory approach, a 2-factor, 13-item structure of the TSK provided the best fit and was invariant across age groups. Older participants were found to have lower TSK fear scores than middle-aged participants for both factors (FA, Harm). A modified version of the Vlaeyen JWS, Kole-Snijders AMJ, Boeren RGB, van Eek H (Fear of movement/(re)injury in chronic low back pain and its relation to behavioral performance. Pain 1995a;62:363-72.) FA model provided a close fit to the data (v 2 (29) = 42.0, p > 0.05, GFI = 0.98, AGFI = 0.97, CFI = 0.99, RMSEA = 0.031 (90% CI 0.000-0.050), p close fit = 0.95). Multigroup analyses revealed significant differences in structural weights for older vs. middle-aged participants. For older chronic pain patients, a stronger mediating role for pain-related fear was supported. Results are consistent with a FA model of chronic pain, while indicating some important age group differences in this model and in levels of pain-related fear. Longitudinal testing of the multivariate model is recommended.

[1] 148w Pain-related fear is among the most salient predictors of chronic pain and associated disability (Vlaeyen and Linton, 2000;Keefe et al., 2004). The fear-avoidance (FA) model of chronic pain was proposed by Lethem et al. (1983) to explain why some musculoskeletal injuries can lead to longstanding pain, depression, and disability. Vlaeyen et al. (1995a) elaborated the FA model to suggest that fear of movement/(re)injury represents a response to pain that is influenced by catastrophizing (Fig. 1). This fear contributes to avoidance behaviors and subsequent disuse, depression, and disability. Research on this model has commonly employed the Tampa Scale for Kinesiophobia (TSK) to assess fear of movement/(re)injury (Kori et al., 1990). High levels of fear have been consistently associated with impaired physical function and greater self-reported disability (Vlaeyen et al., 1995a,b;Crombez et al., 1999). There has been some debate regarding the factor structure of the TSK (McNeil and Vowles, 2004).

[2] 132w Predictive relationships in the FA model have been independently supported through studies of chronic and acute pain patients, using correlational/regression analyses (Vlaeyen et al., 1995a,b;Crombez et al., 1999;Swinkels-Meewisse et al., 2003). The most consistent finding is the strong predictive power of pain-related fear for physical performance and perceived disability (Vlaeyen et al., 1995a,b;Asmundson et al., 1997;Crombez et al., 1999;Al-Obaidi et al., 2000;Verbunt et al., 2003). Structural equation modeling (SEM) has been applied to validate a modified FA model for prediction of pain severity (Goubert et al., 2004a). However, the full FA model, incorporating pain, disability, and depression, has not been uniformly validated. SEM offers significant benefits for model validation, including a confirmatory approach, estimation and adjustment for measurement error, and integration of both observed and latent variables (Byrne, 2001;Schumacker and Lomax, 2004).

[3] 147w Potential age differences in the FA model have received little attention. In one study (Crombez et al., 1999), age was significantly correlated (r = 0.40) with 1 of 3 scales of pain-related fear in a young chronic pain sample. Two additional studies in this report found no significant age -fear correlations. Another study found a small correlation between age and TSK fear (Goubert et al., 2004a). In acute low back pain patients under age 65, age did not correlate with two unique TSK factors (Swinkels-Meewisse et al., 2003). General age group comparisons for chronic pain have revealed multiple similarities, as well as important differences for older adults, including lower anxiety and a stronger relationship between pain severity and depression (Middaugh et al., 1988;Buckelew et al., 1990;Keefe and Williams, 1990;Sorkin et al., 1990;Corran et al., 1994;Cutler et al., 1994;Benbow et al., 1995;Turk et al., 1995;Gibson and Helme, 2000).

[4] 46w The goals of this study were to validate the multivariate FA model of chronic pain using SEM and compare the following across young, middle-age, and older chronic pain patients: factor structure of the TSK (confirmatory approach), levels of pain-related fear, and goodness-of-fit of the FA model.

[1] 248w Based on previous investigations of the factor structure of the TSK, four alternative models were evaluated for goodness-of-fit (Vlaeyen et al., 1995b;Clark et al., 1996;Goubert et al., 2004b;Heuts et al., 2004). The chi-square statistic was evaluated as a measure of exact model fit, but given its known limitations in relation to sample size and evaluation of model approximations, other indices of closeness of fit were employed. The root mean square error of approximation (RMSEA) with 90% confidence interval and p value for test of close fit (RMSEA < 0.05) were selected as primary indices, based on widespread use, good interpretive guidelines, and sensitivity to number of estimated parameters (Browne and Cudeck, 1993;Byrne, 2001;Tomarken and Waller, 2005). Based on published guidelines, RMSEA values less than 0.05 indicate close fit, less than 0.08 reasonable fit, and less than 0.10 mediocre fit, with p values greater than 0.50 indicating close fit (Byrne, 2001). Other fit indices examined and their criteria levels were goodness-of-fit index (GFI; good fit > 0.90, close fit > 0.95), adjusted goodness-of-fit index (AGFI; adjusted for degrees of freedom, good fit > 0.80), and comparative fit index (CFI; adequate fit > 0.90) (Bentler, 1990;Byrne, 2001). The 4 models were: model 1: 1-factor model including all 17 TSK items; model 2: 4-factor 12-item model proposed by Vlaeyen et al. (1995b); model 3: 1-factor 13-item model with omission of the inverse scored items (numbers 4, 8, 12, and 16); model 4: 2-factor 13-item model proposed by Clark et al. (1996).

[2] 262w Table 2 summarizes the goodness-of-fit indices for each of the four models. Consistent with published analyses from other chronic pain populations (Goubert et al., 2004b;Heuts et al., 2004), results suggested that model 4 (2-factor model of Clark et al. (1996)) provided the best fit in this sample. Goubert et al. (2004a,b) have suggested Harm (TSK-Harm) and Fear-avoidance (TSK-FA) as appropriate labels for the two factors. The goodness-of-fit indices indicated mediocre and not close fit of this model to the data. With bootstrapping of maximum likelihood estimates for nonnormal data (500 random samples), estimated biases for all standardized regression weights in the model were less than 0.01. As with the maximum likelihood chi-square test, the Bollen-Stine modified chi-square test based on bootstrapped estimates was rejected (p < 0.01). As with maximum likelihood estimates, this chi-square test is highly sensitive to sample size (Byrne, 2001). Underestimation of standard errors was minimal to moderate. The highest parameter bias (0.007) and error underestimation (44%) were for item 10 on the TSK-FA factor. This was the only item having a 90% confidence interval for its parameter estimate that contained zero. Because item 10 did not load well on either factor in model 4, the model was re-evaluated with this item deleted. The two models are nested and could be compared with a likelihood ratio test. This comparison indicated an improved fit (Dv 2 (11) = 68.1, p < 0.01) while the RMSEA value, in contrast, suggested a decline in overall fit (RMSEA = 0.089, 90% CI 0.078-0.100). Since the overall improvement was marginal, model 4 was retained.

[3] 96w Model 3 (1 factor, 13 items) provided the next best fit. Per Heuts et al. (2004), this model is nested within the 2-factor model of Clark et al., and the models can be compared with a likelihood ratio test. The goodnessof-fit of model 4 was significantly better than that of model 3 (Dv 2 (1) = 68.1, p < 0.001). The fear-avoidance (TSK-FA) and harm (TSK-Harm) scales of model 4 were significantly correlated (r = 0.66, p < 0.001). Internal consistencies of the two scales were moderate: TSK-FA Cronbach's alpha = 0.74, TSK-Harm alpha = 0.72.

[4] 156w Multigroup analysis, a set of procedures developed for testing invariance of a SEM model across groups, was employed for probability testing of the equivalence of the 2-factor Clark et al. model across the three age groups (Byrne, 2001;Schumacker and Lomax, 2004). In this analysis, the validity of the factor structure is tested simultaneously across the groups. Consistent with the full sample analysis, an exact fit of the model was rejected, but an adequate to close overall fit was indicated across groups: v 2 (192) = 440.9 (p < 0.001), GFI = 0.87, AGFI = 0.82, CFI = 0.86, RMSEA = 0.053 (90% CI 0.046-0.059), p close fit = 0.24. In sequential comparisons of models constraining measurement weights (factor loadings), structural covariances, and measurement residuals across the age groups, chi-square differences for nested model comparisons were nonsignificant (p > 0.08). Thus, invariance of the fit of the 2-factor, 13-item model across the three age groups was supported.

[5] 26w MANOVA of the three age groups on the TSK fearavoidance (TSK-FA) and Harm (TSK-Harm) scales revealed a significant main effect (F = 2.94, p < 0.05).

[6] 67w Step-down analysis revealed significant effects for both TSK-FA and TSK-Harm (see Table 3). Post hoc analyses indicated that older participants had significantly lower TSK-FA and TSK-Harm scores than middle-aged Model 1 = 1-factor 17-item model, model 2 = 4-factor, 12-item model of Vlaeyen et al. (1995b), Model 3 = 1-factor 13-item model (without reverse scored items), Model 4 = 2-factor, 13-item model of Clark et al. (1996).

[7] 46w participants. Because the TSK total score is frequently employed in clinical and research practice, a total score age group comparison was conducted with single-factor ANOVA. A significant effect was found, with older participants having lower TSK-Total scores than middleaged participants per post hoc comparisons (Table 3).

[8] 291w Means, standard deviations, and Pearson correlations for the measurement scales are presented in Table 4. Based on the cognitive-behavioral fearavoidance model of Vlaeyen et al. (1995a, Fig. 1) we developed a structural equation model of latent variables, with fear of (re)injury as a mediator between pain catastrophizing and pain disability, depression, and pain severity. Avoidance behavior, as proposed in the Vlaeyen et al. model, was not included in our model. SEM required selection of beginning and endpoints in the cyclical model. We designated pain severity as the endpoint for our model, consistent with contemporary views regarding the multiple physical and psychosocial influences on pain perception (Keefe et al., 2004). These influences on pain perception would be expected to be salient after protracted pain chronicity, as in our sample (mean 6.2 years). As shown in Fig. 2, the latent construct fear of (re)injury was specified by the two subscales of the TSK: TSK-FA and TSK-Harm. Depression was estimated by the CESD total score and the Negative Affect scale from the PANAS. Pain disability was specified by the PDI and the WAS. The Visual Analogue Scale (VAS), Present Pain Intensity (PPI), and Pain Rating Index total (PRI) subscales from the SF-MPQ were used to specify the latent construct of pain severity. The latent construct of pain catastrophizing was specified solely by the CSQ catastrophizing scale. The error variance for this measure was fixed at a value of 0.43 based on a reliability coefficient of 0.80 and standard deviation of 1.47. Coefficient alpha for CSQ Catastrophizing was 0.83 in our sample, and published values have ranged from 0.78 to 0.84 (Rosentiel and Keefe, 1983;Riley and Robinson, 1997;Robinson et al., 1997;Stewart et al., 2001). Test-retest reliability has been assessed at 0.77 (Stewart et al., 2001).

[9] 383w The first step of SEM analyses with latent variables is evaluation of the measurement model. Confirmatory factor analysis showed a very close fit of this model to the data: v 2 (26) = 31.5 (p = 0.21), GFI = 0.99, AGFI = 0.97, CFI = 0.99, RMSEA = 0.021 (90% CI 0.000-0.044), p close fit = 0.984. The Bollen-Stine chisquare test for acceptable fit for the model based on bootstrapped estimates (500 samples) was accepted (p = 0.27). All of the standardized path coefficients were within acceptable range. This indicates that measurement scales employed in the model can be considered valid operationalizations of the latent constructs of fear of (re)injury, catastrophizing, depression, pain disability, and pain severity. Interpretation of SEM results is strengthened by comparisons of data fit for alternative, theory-based models (Bollen, 1989;Tomarken and Waller, 2005). We next evaluated a modified model in which the latent variable catastrophizing was hypothesized to directly influence the latent variables of disability and depression, in addition to the influence mediated by fear of re(injury). This hypothesis was based on the substantial literature supporting the strong predictive value of catastrophizing for these variables (Keefe et al., 2004). This modified SEM model is shown in Fig. 3. SEM analysis indicated a close fit of this model by all indices: v 2 (29) = 42.00 (p = 0.06), GFI = 0.98, AGFI = 0.97, CFI = 0.99, RMSEA = 0.031 (90% CI 0.000-0.050), p close fit = 0.95. As the original model was nested within this adapted model, a likelihood ratio comparison was undertaken. The modified model (Fig. 3) was found to be a significantly better fit to the data: Dv 2 (2) = 106.1, p < 0.001. Bootstrapping of maximum likelihood estimates for nonnormal data (500 random samples) revealed minimum bias of parameter estimates, with expected underestimation of some error variances. Absolute value of estimated bias for all standardized path coefficients was less than 0.01. Bootstrapped estimates suggested a maximum 20% increase in standard error estimates, with the majority of the error terms being underestimated by 10% or less. The Bollen-Stine chi-square test for acceptable fit based on bootstrapped estimates (500 samples) was accepted (p = 0.10). Mean standardized path coefficients from the bootstrapped samples are shown in Fig. 3 and are consistent with a good model fit.

[10] 180w Potential age group differences in the adopted SEM model (Fig. 3) were evaluated through multigroup analysis as described in the confirmatory factor analysis section above. Simultaneous fitting of the model to the 3 age groups confirmed a close overall fit of the model: v 2 (87) = 111.7 (p = 0.04), GFI = 0.95, AGFI = 0.91, CFI = 0.99, RMSEA = 0.025 (90% CI 0.006-0.037), p close fit = 1.00. A sequence of planned, nested comparisons of constrained models was initiated to test invariance of model parameters across the three age groups, in the following order: measurement weights, structural weights (path loadings), structural covariances, structural residuals, and measurement residuals. The critical ratio of differences was nonsignificant when measurement weights were constrained (Table 5), supporting the invariance of these parameters across age groups. However, with measurement weights constrained as equal, the critical ratio for structural weights was significant (p < 0.05) indicating age group differences (Table 5). With measurement weights, and structural weights, covariances and residuals constrained equal, measurement residuals were also found to vary between groups (p < 0.01).

[11] 185w Subsequent multigroup analyses were conducted with pairings of age groups to identify between-group differences. Measurement and structural weights did not differ significantly for the following pairings: young vs. middle-aged, young vs. older. However, when the model fit was compared for middle-aged and older participants, measurement weights were invariant but structural weights differed (Dv 2 (7) = 16.5, p < 0.05). With measurement weights and structural weights, covariances and residuals constrained as equal, measurement residuals differed between the older participants and both middle-aged (p < 0.05) and younger (p < 0.001) participants. Structural weights for older and middleaged participants are shown in Fig. 4. Fear of (re)injury was found to play a stronger mediating role between castastrophizing and the depression and disability variables for older participants. This difference was especially notable for the relationship between catastrophizing and depression. For middle-aged participants, catastrophizing was a strong direct predictor of depression, whereas this relationship was mediated by fear of (re)injury to a much greater degree for older pain patients. Depression and disability had less predictive strength for pain severity among older participants, as compared to their middle-aged counterparts.

[1] 92w Through cross-sectional analyses of a large, heterogeneous chronic pain sample via structural equation modeling (SEM), we found a pattern of relationships consistent with a cognitive-behavioral fear-avoidance (FA) model of chronic pain. The 2-factor, 13-item structure of the Tampa Scale of Kinesiophobia (TSK) was supported as the best fit across three adult age groups. Older chronic pain patients (age 55 and older) were found to have lower levels of pain-related fear than middle-aged patients, and structural weights for the adopted SEM fear-avoidance model were found to differ significantly between these two age groups.

[2] 184w Our findings are consistent with the multivariate FA model and its application for chronic pain assessment and treatment. Though much of prior research has focused on low back and musculoskeletal pain (Keefe et al., 2004), our findings suggest that the model could apply to diverse chronic pain conditions. Both catastrophizing and fear of (re)injury are significantly associated with important dimensions of chronic pain. Our results are consistent with fear as a mediator between catastrophizing and perceived disability, depression, and pain, particularly for older adults. Negative appraisals of injury and pain predict levels of fear, while both variables predict increased pain and dysfunction ratings. This model is consistent with the notion that fear is a more important predictor of pain-related disability than pain itself (Waddell et al., 1993), with reports of pain being predicted by many factors. It is important to recognize, however, that SEM does not establish the validity or superiority of a model in relation to untested models (Tomarken and Waller, 2005). Such models can be equally or better fitted to the data and provide equally plausible explanations. Causal relationships in the FA

[3] 62w Table 5 Nested comparisons of sequentially constrained SEM models for multigroup age analyses SEM Model v 2 df Full unconstrained model 111.7 87 Parameters constrained (sequentially): Dv 2 Ddf p Measurement weights 7.12 10 0.71 Structural weights 24.88 14 0.04 Structural covariances 0.28 2 0.87 Structural residuals 5.16 8 0.74 Measurement residuals 43.39 18 0.001 model cannot be evaluated without longitudinal analyses.

[4] 90w The improvement of our SEM model by allowing catastrophizing to directly predict disability and depression was not surprising, given the evidence for castatrophizing as a powerful predictor of depression, disability, and pain (e.g., Sullivan and D'Eon, 1990;Turner et al., 2002;Jones et al., 2003;Keefe et al., 2004). This elaborated model may help to explain the complex relationships between these variables, as past studies have variably shown pain-related fear (Crombez et al., 1999;Woby et al., 2004) or catastrophizing (van den Hout et al., 2001;Denison et al., 2004) as superior predictors of disability.

[5] 162w Although various scoring schemes and factor structures have been suggested for the TSK in chronic pain, our confirmatory analyses support the 2-factor method of Clark et al. (1996) (subsequently named FA and Harm), with reverse items omitted, for a heterogeneous sample. This supports findings from different chronic pain populations (Goubert et al., 2004b;Heuts et al., 2004). We found the 2-factor structure to be invariant across three age groups. However, fit of this model was mediocre, suggesting need for further refinement of the TSK for use in heterogenous clinical populations. A recent study found problems with the TSK factor structure in a fibromyalgia sample (Burwinkle et al., 2005). Content overlap with the construct of catastrophizing was raised as a concern, though based on a 4-item factor scale. Pain-related fear and catastrophizing are known to be significantly correlated, consistent with prediction of the FA model (Fig. 1). Refinements in the assessment of pain-related fear may benefit from further content distinction of these important constructs.

[6] 126w Several authors have identified pain-related fear as one of the most promising areas of research in persistent pain (McCracken and Turk, 2002;Keefe et al., 2004). The FA model has led to graded in vivo exposure interventions for chronic pain. These approaches have been effective for reducing fear and associated behaviors (Vlaeyen and Linton, 2000;Vlaeyen et al., 2001). Recent research has suggested that brief education can reduce fear and catastrophizing, while in vivo graded exposure is required for changes in behaviors and long-term pain (de Jong et al., 2005). Our findings support such treatments that address both appraisals and reactions to pain. Early identification of pain-related fear (Fritz et al., 2001;Sieben et al., 2002;Boersma and Linton, 2005a) is likely to find increasing support in preventive treatment approaches.

[7] 111w An important omission in the FA model is the role of self-efficacy beliefs. These beliefs predict pain, physical functioning, and disability in chronic pain patients, and partially mediate the relationship between pain intensity and disability (Arnstein et al., 1999;Asghari and Nicholas, 2001). Recent studies have shown pain self-efficacy to be equally or more important in predicting disability than FA beliefs (Denison et al., 2004;Woby et al., 2004). Self-efficacy may act through perceived controllability, which produces attenuated activation in the anterior cingulate, insular, and secondary somatosensory cortices, the three areas most consistently linked with pain processing (Salomons et al., 2004). Future investigations should continue to evaluate these important relationships and related interventions.

[8] 151w Our findings are in agreement with prior studies showing both similarities and differences in the experience of chronic pain for older adults (Middaugh et al., 1988;Buckelew et al., 1990;Keefe and Williams, 1990;Sorkin et al., 1990;Corran et al., 1994;Cutler et al., 1994;Benbow et al., 1995;Turk et al., 1995;Gibson and Helme, 2000). We found a stronger mediating role for fear between catastrophizing and the dependent variables' depression and disability for older, as compared to middle-aged, patients. Higher pain catastrophizing has typically been associated with younger age (Santavirta et al., 2001;Turner et al., 2004), while the relationship between age and fear of (re)injury ranges from nil (Swinkels-Meewisse et al., 2003) to moderate (Crombez et al., 1999). Age has been found to moderate the relationship between catastrophizing and pain (Santavirta et al., 2001). It can be hypothesized that (re)injury risks are higher for older adults and that associated fears are fueled when catastrophic thinking occurs.

[9] 178w However, we found that older pain patients had lower fear relative to middle-aged patients, the most frequently represented group in pain clinics. These differences were present for both factors of the TSK, and the total score. There is evidence from clinical and community studies of more stoic beliefs and reactions to pain among older adults (Cook and Chastain, 2001;Yong et al., 2001), despite differences in nociception and pain perception that could make older adults more vulnerable to the negative impacts of pain (Gibson and Farrell, 2004). We found a weaker association between depression and pain severity among older adults, in contrast to some prior findings (Turk et al., 1995) while in agreement with others (Santavirta et al., 2001). These findings are consistent with reductions in the intensity and frequency of negative emotions, declines in emotional expressivity, and increases in emotional control with increasing age (Gross et al., 1997). Very little is known about how these differences interact with the experience of chronic pain. Longitudinal research will be needed to distinguish potential cohort and/or aging factors in pain stoicism.

[10] 160w There are several limitations to this study. First, we did not include a separate measure of activity avoidance, precluding confirmation of its role in the FA model. Though self-reported activity avoidance overlaps with the TSK FA scale, a separate measure should be included in future studies. Similarly a measure of pain vigilance (Vlaeyen and Linton, 2000;Goubert et al., 2004a) could enhance the analyses. A second potential limitation was the exclusion of a measure of medical comorbidity. Though such comorbidities have been found to predict decreased activity levels in older chronic pain patients (Farrell et al., 1995), a recent study found no predictive value for physical disability or performance (Weiner et al., 2004). Model comparisons by pain type could also prove valuable, as some evidence suggests the FA model is less applicable in upper extremity pain and fibromyalgia, where task persistence could play a greater role (Vlaeyen and Morley, 2004). The heterogeneity of our sample could have masked such potential differences.

[11] 159w Additionally, our age groupings were influenced by our sampling distribution. These groupings differ from those commonly used in age comparisons (<45, 45-64, P65), though there is variability (e.g., Middaugh et al., 1988) and the cutoff point for ''old age'' has been arbitrarily based on social policy (Butler, 1987). Our results could also be influenced by the unequal and relatively small sizes of the age group samples. Finally, our analyses were cross-sectional and retrospective. As the fear-avoidance paradigm was developed to explain the evolution of acute to chronic pain, prospective longitudinal analyses of the multivariate model will be required. Fear of pain has been found to predict low back pain and disability at 6-month follow-up (Picavet et al., 2002). Recent data also suggest that chronicity is a factor in the influence of fear on self-reported activities (Boersma and Linton, 2005b). The inter-relations among measures of disability, affect, fear/anxiety, and pain intensity require further analyses, particularly those that assess multidirectional relationships.

[12] 52w In conclusion, we have provided additional support for the multivariate FA model of chronic pain, while elaborating the complex inter-relationships, and adding to the growing evidence-base for age differences in the experience of chronic pain. The clinical practice of chronic pain assessment and management will benefit from further research in these areas.

[1] 192w Four hundred and eighty three chronic pain patients evaluated at a multidisciplinary, university-based pain management clinic were eligible for inclusion. While there were no exclusionary diagnoses, the vast majority of patients evaluated in this clinic have musculoskeletal and neuropathic diagnoses. All study measures were completed as part of standard clinical evaluations. The study was approved by the Human Investigation Committee of the university's Institutional Review Boards. Participants received no compensation for their participation. Nine participants provided incomplete data and were excluded from analyses. Five participants were deemed multivariate outliers based on elevated Mahalanobis distances with a conservative probability level of p < 0.001 (Tabachnick and Fidell, 2001). They were deemed to be outside of the target population and were excluded from analyses. The final sample included 469 participants. For age analyses, groupings were: 40 and under (n = 152), 41-54 (n = 198), and 55 and over (n = 119), based on age distribution within the sample. Demographic and pain characteristics are shown in Table 1, including age group comparisons. There were expected age differences for marital and employment status, and receipt of financial benefits. Young participants had shorter average pain duration.

[2] 123w Accurate estimation of measurement error in SEM requires at least two measures for each latent variable. If only one measure is used, error variance must be specified based on external data (Schumacker and Lomax, 2004). The following scales comprised the measurement model for the SEM analyses, as represented in Fig. 2: 2.2.1. Tampa Scale of Kinesiophobia (TSK; Kori et al., 1990) The TSK is a self-report measure of fear of movement and (re)injury. It consists of 17 items scored on a 4-point scale. Several studies of the TSK in chronic pain samples have supported a 13-item 2-factor structure, with omission of reverse-scored items (Clark et al., 1996;Goubert et al., 2004a,b;Heuts et al., 2004). Adequate psychometric properties have been documented (Vlaeyen et al., 1995b).

[3] 66w Pain Questionnaire (short-form) (MPQ; Melzack, 1987) The short-form MPQ is a widely used measure of the sensory, affective, and intensity dimensions of pain. It includes 15 pain descriptors that form the Pain Rating Index (PRI), a 10 cm visual analogue scale (VAS), and a 6-point present pain intensity (PPI) scale. Good consistency, validity, treatment sensitivity, and discriminant ability have been demonstrated (Melzack, 1987;Melzack and Katz, 2001).

[4] 61w 2.2.3. Pain Disability Index (PDI; Pollard, 1984) The PDI is a brief measure of perceived pain-related disability for seven areas of daily functioning. It has been found to be an internally consistent (a = 0.86) measure, with good concurrent, criterion-related, and discriminative validity (Pollard, 1984;Tait et al., 1990). The balance of research supports a single underlying factor (Jacob and Kerns, 2001).

[5] 71w All questionnaires were administered in electronic format, based on Microsoft Access 2000 (Microsoft Corporation, USA) platforms. We have previously cross-validated the SF-MPQ and PDI for chronic pain assessment in this format (Cook et al., 2004). Data were entered via 17-in., 1280 • 1024 pixel resolution, Accutouch resistance touchscreen LCD monitors (Elo Touchsystems, USA) and/or a standard two-button mouse, and with traditional keyboards, with setup as previously described (Cook et al., 2004).

[6] 78w All participants completed the study measures in clinic, immediately prior to initial clinical evaluations. Informed consent was obtained for participation in a broader study of chronic pain assessment, and participants were given a brief information session to familiarize them with the relevant computer components and response methods. A trained facilitator was available to provide limited procedural assistance to participants as needed, while providing for adequate privacy. Participants were encouraged to take brief rest or stretching breaks as needed.

[7] 183w Data were screened for integrity and assumptions of multivariate analyses. As previously noted, 5 participants were identified as multivariate outliers and were excluded from analyses. The multivariate distribution was found to be nonnormal, with a Mardia's coefficient of multivariate kurtosis of 9.3 (p < 0.001). While SEM parameters with maximum likelihood (ML) estimation are robust to nonnormality, fit indices and standard errors can be biased (Bollen, 1989;West et al., 1995). This can include overestimation of chi-square values, underestimation of CFI, and underestimation of standard errors (Byrne, 2001). Bootstrapping of ML estimates (n = 500 samples) (West et al., 1995) was employed to evaluate potential bias. Bootstrapping provides repeated resampling of the original sample (with replacement) to create a sampling distribution that is not dependent on the parametric assumption of normality. All SEM analyses were conducted with AMOS v. 5.0 for Windows (Arbuckle, 2003). Single factor analysis of variance (ANOVA) and multivariate analysis of variance (MANOVA) using Wilks' lambda criterion were employed for mean comparisons by age group. Tukey's honestly significant difference (HSD) test was employed for post hoc mean comparisons by age group.

[1] 55w Coping Strategies Questionnaire Catastrophizing Scale (CSQ-catastrophizing; Rosentiel and Keefe, 1983) The CSQ is a widely used self-report measure of cognitive and behavioral coping strategies in chronic pain. The 6-item catastrophizing subscale has been used extensively as a measure of pain catastrophizing, with adequate psychometric properties (Stewart et al., 2001;Sullivan et al., 2001;Turner and Aaron, 2001).

[2] 46w 2.2.5. Center for Epidemiological Studies -Depression scale (CESD; Radloff, 1977) The CESD is a 20-item self-report measure of depressive symptomatology. It has good internal consistency, and has been found to be appropriate for chronic pain patients, with good sensitivity and specificity (Radloff, 1977;Geisser et al., 1997).

[3] 48w 2.2.6. Positive and Negative Affect Schedule (PANAS; Watson et al., 1988) The PANAS is a brief, self-report measure comprised of two 10-item scales for positive and negative affect. The scales have been shown to be internally consistent, valid, and largely uncorrelated, with adequate stability (Watson et al., 1988).

[4] 46w The WAS is a single item self-rating of activity function used in our clinical assessments. Patients are asked ''what percent of want-to-do activities do you get done in an average day (0-100%)?'' Responses to this question correlated r = À0.37 with the PDI total disability score.