Role of Pain Catastrophizing in the Effects of Cognitive Behavioral Therapy for Chronic Pain in Different Subgroups: An Exploratory Secondary Data Analysis Using Finite Mixture Models

ABSTRACT

Introduction

Providing effective treatment for debilitating chronic pain is a challenge among many populations including military service members. Cognitive behavioral therapy for chronic pain (CBT-CP) is a leading psychological pain treatment. Pain catastrophizing is a pivotal mediator of pain-related outcomes. The purpose of this study was (1) to identify patient subgroups who differ in response to CBT-CP and (2) to explore the characteristics that define these patient subgroups. The overall goal was to obtain a better understanding of factors that may influence response to CBT-CP.

Materials and Methods

This study was a secondary analysis of data from a clinical trial of 149 U.S. active duty service members with chronic pain. Participants underwent group-based CBT-CP for 6 weeks and completed pre- and posttreatment assessments. Finite mixture models were employed to identify subgroups in treatment response, with pain impact score as the primary outcome measure.

Results

We identified two classes of nearly equal size with distinct pain impact responses. One class reported improved pain impact scores following CBT-CP. This improvement was significantly associated with lower (better) baseline depression scores and greater improvement in posttreatment pain catastrophizing. In contrast, the other class reported slightly worse mean pain impact scores following CBT-CP treatment; this response was not related to baseline depression or change in pain catastrophizing.

Conclusions

Our findings demonstrate that a sizable proportion of individuals with chronic pain may not respond to group-based CBT-CP and may require a more individualized treatment approach.

INTRODUCTION

Chronic pain poses a significant challenge to both individuals and health-care systems in the United States.¹ This is especially true for active duty service members (ADSMs) and the Military Health System.^2,3 Among U.S. Army service members, the prevalence of chronic pain is estimated to be 31%-44%,^2,4 and chronic pain is a leading cause of disability and release from active duty due to health reasons.⁴

A holistic approach that integrates biological and psychological elements in the treatment plan is required to manage chronic pain effectively. Such a strategy transcends the boundaries of merely addressing the biological aspects of pain to also incorporate essential psychological interventions as vital components in the treatment toolkit.⁵ Cognitive behavioral therapy for chronic pain (CBT-CP) is a widely used and accepted psychological treatment that aims to lessen pain intensity and psychological distress by improving a person’s ability to cope with their pain.⁶ A systematic review of studies involving civilian populations found that CBT-CP had small effects on pain intensity and disability and moderate effects on mood and pain catastrophizing.⁵ Less is known about the effects of CBT-CP among ADSM.

Pain catastrophizing is characterized by a negative cognitive and emotional response to pain that is marked by rumination, magnification, and feelings of helplessness.⁷  Rumination is the tendency to constantly think about pain, magnification is the tendency to expect the worst outcomes of pain, and helplessness is the tendency to perceive little control over pain.⁷ Pain catastrophizing has been recognized as a process variable that mediates positive shifts in pain-related outcomes, such as pain intensity and pain interference.^8,9 Given this influential role, pain catastrophizing has emerged as a primary treatment target for chronic pain.¹⁰ In addition, Yoshino and colleagues have posited that negative emotions of any kind play an important role in the treatment effects of CBT-CP.¹¹ Previous studies investigating the predictors of efficacy for various chronic pain management approaches found that in addition to pain catastrophizing, baseline depression and anxiety also have negative effects on pain management.^8,11,12 This suggests that the effect of CBT-CP may vary depending on individuals’ physical, psychological, or cognitive traits.

Past studies have used latent class analysis (LCA)^13,14 to examine responses to CBT; however, LCA usually does not emphasize the relationship between predictor and outcome variables. To our knowledge, this is the first study to classify ADSMs’ responses to CBT-CP using finite mixture models (FMMs). FMM explicitly incorporates relationships between predictor and outcome variables to classify latent classes, thereby enabling the identification of relational heterogeneous subgroups within the data. This approach informs tailored treatment strategies for these distinct groups.

Given the prevalence of chronic pain among U.S. ADSMs, they represent a group in need of specialized attention. While a growing body of evidence demonstrates the potential of CBT-CP to enhance a person’s functional capacity and overall adjustment to chronic pain,¹⁵ it is less clear which specific groups of ADSMs would benefit the most.⁶ Against this backdrop, our study delineates two specific aims: (1) to identify patient subgroups among ADSMs with chronic pain that differ in response to CBT-CP and (2) to explore the characteristics that define these patient subgroups to gain a better understanding of individual differences in response to CBT-CP using FMM.

METHODS

Study Design and Participants

This study was a secondary analysis of data from a clinical trial that used a pretest–posttest design to evaluate the effects of a brief CBT-CP program conducted in an interdisciplinary pain management center within a military treatment facility.¹⁶ Our analyses included 149 ADSMs who had engaged in the CBT-CP intervention and completed the pretreatment and posttreatment Pain Assessment Screening Tool and Outcomes Registry (PASTOR) surveys. These surveys included demographic questions as well as validated instruments designed to capture a comprehensive and holistic record of health in the context of chronic pain.

Intervention

Participants engaged in a treatment program that included CBT-CP in 60- to 90-min group sessions once per week for 5-6 weeks. Each session was led by a psychologist, physician, or nurse practitioner. The program was designed to introduce self-management, improve pain management self-efficacy, reduce functional limitations, and reduce negative impacts of pain. The CBT-CP sessions included psychoeducation/goal setting, behavioral skills (activities, pacing, sleep improvement, and relaxation training), cognitive coping, relapse prevention, and flare management.

Measures

Participants completed PASTOR upon entering and exiting the treatment program (i.e., pretreatment and posttreatment). PASTOR is an electronic, web-based survey on a battery of patient-reported outcomes, adapted from the National Institutes of Health’s Patient-Reported Outcomes Measurement Information System (PROMIS).^17,18 Widely utilized across pain specialty clinics within the Military Health System, PASTOR facilitates a comprehensive, multidimensional assessment tailored for patients experiencing chronic pain.¹⁷ All patient-reported measures included in PASTOR have established validity and reliability and have been validated in the military population.^17,18

Demographic and pain-related characteristics

Demographic characteristics were self-reported by participants and included age, biological sex (survey options: male or female), race (survey options: Asian, Black, White, or other), ethnicity(survey options: Hispanic/Latino or not), education level, marital status, and income. This information was gathered to provide comprehensive profiles of the participants. Pain-related characteristics collected included pain location, type, duration, and persistence. If pain was present in multiple locations, we used the primary complaint at the time of the participant’s initial appointment at the interdisciplinary pain management center to determine the pain location.

Patient-reported measures from PASTOR

Pain Catastrophizing Scale.

The Pain Catastrophizing Scale (PCS) is a self-report questionnaire that assesses a person’s negative cognitive–affective response to anticipated or actual pain.^7,19 Participants were asked to reflect on past painful experiences and indicate the degree to which they experienced each of 13 thoughts and feelings when they were in pain.¹⁹ The PCS uses a 5-point Likert rating to measure how often the respondent experiences each thought and feeling, from 0 (not at all) to 4 (all the time). The overall PCS score ranges from 0 to 52, with higher scores representing greater catastrophic thinking. The three PCS subscales are rumination (four items), magnification (three items), and helplessness (six items).⁷ The PCS has strong internal consistency (Cronbach’s alpha = 0.92) and good criterion-related validity in outpatient participants and the adult community.²⁰

Defense and Veterans Pain Rating Scale.

The Defense and Veterans Pain Rating Scale is a self-report questionnaire that assesses pain intensity via a numeric rating scale, using visual cues and word descriptors to anchor pain ratings to perceptual experiences and limitations imposed by pain.²¹ The average pain intensity over the previous 7 days is rated across an 11-point scale from 0 (no pain) to 10 (as bad as it could be; nothing else matters).²¹

PROMIS measures.

PASTOR includes the following eight PROMIS measures: pain interference, physical function, depression, anxiety, anger, fatigue, sleep-related impairment, and satisfaction with social roles and activities.^22–24 Computerized adaptive testing is used to reduce the survey burden. The total score for each PROMIS measure is converted to a t-score, a standard score with a mean of 50 and a standard deviation of 10 for the referent general U.S. population; a higher PROMIS t-score represents a stronger association with the concept being measured. The PROMIS measures have been validated in a broad sample of individuals living with chronic conditions, including ADSMs.¹⁷

Pain Impact Score.

We used the Pain Impact Score (PIS) as a primary study outcome. The PIS, based on recommendations by the National Institutes of Health Task Force on Research Standards for Chronic Low-Back Pain, is defined as a combination of three measures: pain intensity, pain interference, and reverse physical function.²⁵ The total PIS has a range of 8-50, with higher scores indicating higher pain impact. Previous studies have estimated the threshold for clinically meaningful important difference in PIS, when using a 5- or 7-item Likert scale in populations with pain, as a change of three to seven points in total PIS.^26–28

Statistical Analysis

Our focus was on understanding how changes in PCS score after treatment influenced changes in PIS after treatment, while adjusting for the pretreatment PIS and depression. All analyses were carried out using R software (version 3.3.0, R Foundation for Statistical Computing, Vienna, Austria). We conducted two-tailed tests, with P-values less than .05 considered to be statistically significant. For demographic characteristics, we computed descriptive statistics, including mean, standard deviation, and number and/or percentage. To explore distinct groups (latent classes) within our data based on similar patterns of response to the CBT-CP, we applied FMMs, using the R package flexmix.²⁹

In pursuit of our aim to explore latent classes based on changes in PCS score and their impact on changes in PIS, we conducted a preliminary examination of potential confounders that could influence this relationship. Among the various factors considered in these preliminary analyses, baseline depression was found to have a significant association with PIS change; this is consistent with previous pain outcomes research.¹¹ We similarly adjusted pretreatment PIS to control for its influence on change in PIS. By adjusting for these pretreatment scores, our analyses could provide a more accurate understanding of the latent classes and highlight the specific role of pain catastrophizing in the dynamics of CBT-CP and treatment efficacy.

We chose FMMs because they offer a method for identifying subgroups within datasets that are not otherwise apparent.³⁰ This approach helped us to better understand the varying effects of treatment across different subsets of participants, while potentially offering insights into why some individuals might benefit more from certain treatments than others. Building on this foundation, we used several criteria to determine the optimal number of subgroups with distinct pain impact responses, as well as the clinical relevance of the resulting latent classes. We used the Bayesian information criterion and its adjusted version, the adjusted Bayesian information criterion, to determine which model best fit the data; for both metrics, lower values indicate a better fit to the data.³¹ In tandem with these statistical criteria, we consulted with clinical experts to ensure that the selected model was both mathematically robust and clinically useful. The best fitting number of latent classes was then retained for subsequent analyses. To ensure the reliability of the parameters obtained, we performed a refit of the model, using the parameter estimates obtained from the initial fit as new starting values to enable a more refined and stabilized estimation process. Latent groups were identified based on the posterior probabilities calculated from the FMMs.

Finally, we described the baseline PASTOR outcomes (PCS, Defense and Veterans Pain Rating Scale, and PROMIS measures) of each resulting latent group and examined the differences between the groups using chi-squared tests/Fisher’s exact tests and/or t-tests when appropriate. Some participants completed their baseline PASTOR assessments before the addition of clinical characteristics (specifically pain location, pain duration, and pain persistence). This timing discrepancy accounts for some of the missing data observed for these variables. After confirming that the patterns of missing values between classes are not influenced by temporal factors, we opted to exclude missing values from our analysis. This decision helps mitigate bias and maintain the validity of our findings, reflecting our commitment to accurate data interpretation.

RESULTS

Study Population

The participants’ demographic characteristics are reported in Supplementary Table S1. A total of 149 ADSMs with chronic pain were included in the analysis. The demographic subgroups with the highest representation were male (75.2%), White (63.1%), non-Hispanic/Latino (84.6%), aged between 25 and 34 years (45%), and married (69.1%).

Model Selection

Table I presents the model fit indices for models with one through four latent classes. The process of model selection necessitates a balance between statistical fit and substantive significance. Using the Bayesian information criterion and adjusted Bayesian information criterion, we initially considered the one-class model as the best statistical fit. However, the one-class solution did not align with the primary objective of our study, which was to identify distinct and meaningful subgroups in treatment response. To ensure the clinical relevance of our findings, we next consulted with clinical experts and reviewed empirical literature,^32–34 both of which indicated that a single-group model may overlook clinically significant subgroups. The threshold for clinically meaningful important difference in PIS, identified in previous studies as ranging from 3 to 7 in populations with pain,^26–28 also informed our decision: in the two-class model, class 1 demonstrated a mean PIS change exceeding 3, which aligns with clinically meaningful improvement, while class 2 showed a worsened impact on PIS change. However, the one-class model did not demonstrate a clinically meaningful improvement (mean PIS change = −0.52 ± 5.8). Therefore, despite a less optimal statistical fit, we selected the two-class model for its ability to provide a meaningful distinction between subgroups, particularly in their response to treatment.

TABLE I.

Finite Mixture Model Selection Results

			Class size (%)
Classes	BIC	aBIC	1	2	3	4
1	915.10	940.12	100.0
2	932.04	982.08	50.3	49.7
3	928.50	1003.56	32.9	55.0	12.1
4	951.08	1051.15	40.9	12.8	34.2	12.1

Abbreviations: BIC = Bayesian information criterion; aBIC = adjusted Bayesian information criterion.

We also considered three- and four-class models, but ultimately decided that the incremental gains in clinical interpretability did not justify the additional complexity. As indicated by the similar coefficients across these solutions, these models did not demonstrate clinically meaningful differentiation among the additional classes. Our decision to adopt the two-class model was therefore based on a comprehensive evaluation of both statistical indicators and clinical utility. This approach acknowledges the possibility that the true underlying structure of the population may be more nuanced than statistical measures alone can reveal and emphasizes the importance of incorporating clinical expertise into the model selection process.

FMM Analyses

Table II shows a standard regression model (standard one-class model) contrasted with the FMM that separates participants into two classes. It presents the coefficients and standard errors of the effects of change in PCS score on change in PIS, adjusted for pretreatment PIS and pretreatment depressive symptoms. The single-component FMM, which is equivalent to multiple linear regression, showed that change in PCS score (coefficient = 0.14, P = .001), pretreatment depression (coefficient = 0.19, P < .001), and pretreatment PIS (coefficient = − 0.45, P ≤ .001) were all significantly associated with change in PIS. That is, variables in the single-component FMM that were found to be significantly associated with greater posttreatment PIS improvement included greater posttreatment PCS score improvement, lower pretreatment levels of depression, and higher pretreatment PIS.

TABLE II.

Finite Mixture Models Representing the Association between PCS Score Change,^a Pretreatment Depression, Pretreatment PIS, and PIS Change^b (N = 149)

				Two-class model
	Standard one-class model			Class 1 (n = 75) Improved impact			Class 2 (n = 74) Worsened impact
Variable	Coefficient	SE	P	Coefficient	SE	P	Coefficient	SE	P
Intercept	3.49	2.14	.10	2.54	3.04	.40	6.83	3.62	.06
PCS score change	0.14	0.04	.001	0.22	0.06	<.001	0.08	0.06	.19
Pretreatment depression	0.19	0.04	<.001	0.30	0.06	<.001	0.02	0.08	.82
Pretreatment PIS	−0.45	0.06	<.001	−0.66	0.12	<.001	−0.18	0.09	.04

Bold font indicates statistical significance at P < .05.

Abbreviations: PCS = Pain Catastrophizing Scale; PIS = Pain Impact Score.

^aPCS score change = posttreatment PCS score − pretreatment PCS score.

^bPIS change = posttreatment PIS − pretreatment PIS.

The two-class FMM presented distinctly different results. In the two-class model, the class 1 participants showed similar results to the single-component FMM. Among the 50.3% (75/149) of the participants who composed class 1, significant associations were found between posttreatment PIS change and change in PCS score (coefficient = 0.22, P ≤ .001), pretreatment depression (coefficient = 0.3, P ≤ .001), and pretreatment PIS (coefficient = − 0.66, P ≤ .001). However, among the class 2 participants, who composed 49.7% (74/149) of the sample, no significant relationships existed between posttreatment PIS change and posttreatment PCS score, pretreatment depression or pretreatment PIS.

Summary of the Classes

Supplementary Table S2 reports the descriptive statistics and tests of the significant differences between classes in the two-class model. The average PIS change for the entire sample was −0.52 ± 5.82. As shown in Supplementary Fig. S1, class 1 participants demonstrated improvement in mean PIS (−3.20 ± 6.76), whereas class 2 participants displayed a worsening in mean PIS (2.20 ± 2.75). This was a notable difference between classes in mean change in posttreatment PIS (P < .001). However, when examining posttreatment changes in the mean PCS score, both classes showed improvements, with PCS scores of −2.85 ± 9.47 for class 1 and −3.42 ± 9.43 for class 2, and the improvements were not statistically significantly different (P = .72). The two classes had a marginally significant difference in pretreatment PCS scores, with class 2 reporting a mean PCS score more than four points higher than class 1 (27.0 ± 14.1 vs. 22.8 ± 14.3; P = .07) and a pretreatment score for the PCS magnification subscale significantly higher in class 2 than class 1 (P = .02). There were no statistically or clinically significant differences between the classes in other baseline PROMIS outcomes.

Almost three-quarters of participants reported experiencing pain for at least 1 year. In addition, it was noteworthy that over 72% of participants reported experiencing pain on a daily or near-daily basis. However, there was a marginal significant difference in pain persistence between the classes, with class 2 having greater pain persistence but lower frequency of missing values than class 1. In both class 1 and class 2, musculoskeletal pain emerged as the predominant pain type, as reflected by the International Statistical Classification of Diseases and Related Health Problems, 10th revision (ICD-10) codes used as the primary diagnosis by clinicians for more than two-thirds of the initial pain clinic encounters. The back was the most common pain location, reported by nearly three-quarters of participants.

DISCUSSION

To our knowledge, this is the first study to classify ADSMs’ responses to CBT-CP using FMM, a technique not previously applied in this context, although LCA has been used.^13,14 LCA is a statistical method that identifies classes of individuals based on their responses to pain interventions.³⁵ Unlike LCA, which generally does not emphasize the relationship between predictor and outcome variables, FMM explicitly incorporates these relationships to classify latent classes, thereby enabling the identification of relational heterogeneous subgroups within the data. This approach informs tailored treatment strategies for these distinct groups. Notably, the one-class model, closely resembling class 1 yet vastly different from class 2, underscores the importance of employing sophisticated statistical approaches like FMM. This differentiation illustrates the risk of treating the patient sample as homogeneous, which could obscure significant variances in treatment responses that are critical for effective pain management.

By applying a novel statistical approach based on FMMs, we categorized ADSMs’ responses to CBT-CP into two distinct classes. These classifications were based on the effects of pretreatment depression level and posttreatment change in pain catastrophizing on the adjusted pretreatment PIS. Notably, slightly over half of the ADSMs were categorized into a class that experienced posttreatment improvement in mean pain impact, which was associated with lower pretreatment depression with posttreatment improvement in PCS score. These findings were not unexpected, and in this group, pain catastrophizing may have functioned as a mediator of PIS improvement. What was unexpected was the identification of a sizable and distinct class with marginally higher baseline PCS scores who did not experience posttreatment improvement in mean PIS, and for whom neither depression nor change in PCS score was associated with posttreatment PIS change.

The findings in class 1 of the two-class FMM are consistent with findings from prior research that examined the role of pain catastrophizing as a process variable of chronic pain treatment.^8,9,12 The findings in this class are also consistent with the well-documented relationship between pain and other psychological conditions, such as depression and psychological distress, and the frequent co-occurrence of chronic pain and comorbid depression in clinical practice.³⁶

The findings of class 2 of this model, whose mean baseline PCS score was 27.0 (vs. 22.8 in class 1), are consistent with those of Scott and colleagues, who found that a pretreatment PCS score of 24 points or greater predicted higher follow-up pain intensity and a poorer likelihood of returning to work 1 year after completing a multidisciplinary rehabilitation program.³⁷ The authors underscored the importance of targeted pain intervention strategies for individuals who have pretreatment PCS scores of 24 or higher.³⁷ This threshold is higher than that reported by Schaaf and colleagues, who found that a baseline PCS score of 20 or greater was associated with twice the likelihood of future military medical disability 1 year after referral to pain specialty care compared to individuals with lower baseline PCS scores.³⁸ Despite the lack of statistically significant difference in baseline PCS between the classes in our model (P = .07; see Supplementary Table S2), it is possible that a PCS “threshold effect” may have influenced the response to CBT-CP of some individuals in class 2.

Our study reinforces previous research findings on the adverse impact of high baseline PCS scores on treatment outcomes.³⁸ ADSMs with chronic pain are often treated with group-based cognitive behavioral therapy, and those with high pain catastrophizing might require a more tailored approach to effectively address the heightened catastrophizing. This personalized approach could be integrated within group settings, leveraging the benefits of group therapy while addressing specific patient needs. For example, Schütze and colleagues have advocated for pain management to be fine-tuned to specific dimensions of pain catastrophizing rather than address it as a single construct.³⁹ While multimodal pain therapies (those involving both psychological and physical treatments) may have the strongest effect size in reducing catastrophizing, past reviews⁵ on this topic have been limited by the heterogeneity of studies. However, tailoring CBT-CP in a nuanced manner continues to hold substantial promise for reducing the impact of pain on treatment-resistant populations. Approaches to treatment tailoring might also identify other factors, such as pain acceptance, resilience, or pain self-efficacy, as potential targets of intervention.⁴⁰ These constructs not only diversify the therapeutic targets within pain management but also align with a more holistic approach to treating chronic pain, which may be particularly beneficial for patients who show limited response to traditional CBT-CP solely focused on reducing pain catastrophizing.

Limitations

This study’s relatively small sample size may limit its generalizability. While our results provide valuable initial insights, future studies with larger and more diverse cohorts are necessary to validate and expand upon our conclusions and determine if the observed relationships hold true across different populations and settings. Including measures of pain acceptance, resilience, or pain self-efficacy in future studies could help further examine the differences the two classes identified depending on the response to CBT-CP, providing deeper insights into the specific variables that influence the treatment outcomes. A second limitation to this study is that it was not a randomized controlled trial but a secondary analysis of clinical trial data; this may have introduced biases related to the selection of the dataset and the retrospective nature of the analysis. Future prospective randomized controlled trials could establish causal relationships with greater confidence.

While our investigation focused on the influence of pain catastrophizing and depressive symptoms on pain impact (measured via the PIS), other variables might also significantly influence the PIS. Consequently, observed changes between PCS score and PIS within the context of CBT-CP should not be interpreted exclusively as cause-and-effect relationships. To clarify the specific contributions of pain catastrophizing and depressive symptoms, further research should employ a multivariate approach that controls for potential confounding factors.

Despite these limitations, this study raises several important implications for both research and clinical contexts, particularly for individuals with high pain catastrophizing. The study used innovative statistical approaches to build on foundational research; its results ultimately both provide insights into who may benefit more from CBT-CP and illuminate potential risk factors. Our findings call for further, more diversified studies to refine and evaluate intervention strategies, particularly for people with pronounced pain catastrophizing. The insights already gained, however, prompt a call for clinicians to adopt more individualized, nuanced pain management approaches for patients with high levels of catastrophizing, as they may not respond to a uniform group-based model.

CONCLUSION

This investigation underscores the value of applying FMMs to identify subgroups of patients who do and who do not respond to CBT-CP. Our findings identified a subgroup with higher pretreatment PCS magnification scores and poor response to CBT-CP; this subgroup may require a more individualized approach to treatment. While these findings hold crucial implications for enhancing the treatment of all patients with chronic pain, they bear particular significance for ADSMs as a population with high prevalence of chronic pain.