Interpretation bias and its relationship with pain: a systematic review and meta-analysis.

The final search was conducted on March 5, 2024, across 5 databases (Web of Science, PsycINFO, CINAHL, MEDLINE, and the Cochrane Library Databases). This review was preregistered on PROSPERO (CRD42022354742). Web of Science Core Collection, PsycINFO, CINDAHL, MEDLINE, and the Cochrane Library Databases were systematically searched from inception using the following terms: (“ambiguous*” OR “interpret*” OR “cognitive” OR “information processing”) AND (“bias*” OR “selective”) AND (“pain*”). Backward citation searching of the reference lists of included studies was conducted, and the corresponding authors of included studies were contacted to identify further potentially eligible studies. Studies were eligible for inclusion if they were English-language studies, included original data, published in a peer-reviewed journal, and included human participants who experienced chronic or acute pain or were subjected to acute experimental pain. Studies were included where the chronic pain occurred in the context of a health condition (eg, rheumatoid arthritis, diabetes, or endometriosis), and in these cases, additional health condition was included as a moderator, where there were sufficient studies. If the same data were included across multiple articles, the largest sample was included. Studies required interpretation bias data to be available. Studies also needed to either include pain outcome data and its association with interpretation bias, or assess interpretation bias differences between pain and no-pain groups. There were no restrictions on the measures of interpretation bias to be included, so long as the measure included some form of ambiguous information that could be interpreted as pain-related or benign, and was labelled as interpretation bias. Longitudinal studies were included if they included a measure of interpretation bias and pain outcomes at a later timepoint. Intervention studies were included, if they used an intervention that was designed to specifically modify pain-related interpretation bias, and measured interpretation bias and/or pain outcomes postintervention. All references were screened by 2 independent reviewers at the title/abstract (κ = 0.699) and full-text stage, with any discrepancies resolved through discussion with a third reviewer. For all included studies, we extracted data pertaining to study design and context, participant age, sex, interpretation bias measure, baseline characteristics, pain characteristics, and data to facilitate quality ratings. In addition, we extracted data related to pain-related interpretation bias and the relationship between these biases and pain outcomes. For broader interpretation bias measures that included a pain/bodily threat subscale (eg, Adolescent Interpretation of Bodily Threat 26 and Bodily Sensations Interpretation Questionnaire 10 ), we included data only on the pain/bodily threat subscale, not the complete measure. In addition, effect sizes that demonstrate postintervention changes in interpretation biases and related changes in pain outcomes were extracted. Extracted data were independently checked by another author to ensure accuracy of information. If data were missing, the corresponding authors of the relevant study were contacted to request the data if available. Where studies included more than 1 measure of interpretation bias, we took the primary measure. For example, where both disability/health and pain interpretations were reported, we gave preference to the pain interpretation data. 46 , 48 , 55 Where there was not a clear primary outcome, we combined the outcomes with meta-analysis 25 , 26 (ie, likelihood + believability ratings), assuming a correlation of r = 0.5 where the association between interpretation bias measures was not reported. The meta-analysis was conducted using the Comprehensive Meta-Analysis software. 4 We used a random effects model for all primary analyses and a mixed effects model for moderator comparisons, with fixed effects between moderator groups and random effects within these groups. First, differences in interpretation bias between pain and pain-free groups were calculated using Hedge g , which is appropriate for small sample sizes. 61 Next, cross-sectional associations between interpretation bias and pain outcomes were combined as correlations ( r ). Forest plots were created for primary outcome results. Tau 2 estimates were pooled. To conduct an analysis, we required at least 3 comparisons. Heterogeneity was assessed through Cochrane Q (using P < 0.05 to indicate significant heterogeneity). Where this was significant, I 2 provided further information about the degree of heterogeneity (80%: considerable heterogeneity). 27 Where heterogeneity was detected and there were sufficient studies (ie, ≥3 studies per group), moderator analyses were conducted to test whether effects differ by pain status (chronic vs acute), interpretation bias measure, and study quality. Where studies were not able to be meta-analysed due to insufficient or disparate studies, they were reported narratively. This was the case for the longitudinal and intervention studies. To assess for publication bias, funnel plots were created, and Duval and Tweedie's trim and fill method 14 with a random effects model was applied. We also calculated Rosenthal's fail-safe N to estimate the number of missing studies required to render the effect nonsignificant. To assess for study risk of bias, the quality of each included study was coded using the Standard Quality Assessment Criteria for Evaluating Primary Research Papers from a Variety of Fields by Kmet et al. 35 This quality appraisal tool has the advantage of allowing evaluation of different types of articles (eg, trials and observational studies) within the same framework and has been used in other recent cognitive bias reviews. 1 , 52 , 59 The risk was assessed using the following criteria: (1) the objective/s were sufficiently described; (2) the study design was evident and appropriate; (3) subject selection was adequately described; (4) adequate description of subject characteristics; (5) outcome measures adequately defined; (6) appropriateness of sample size; (7) analytical methods described and justified; (8) appropriate estimates of variance reported; (9) confounding variables controlled for, where appropriate; (10) the results were reported in sufficient detail; and (11) the conclusions were supported by the results. Intervention studies were additionally assessed with the following criteria: (12) participants randomly allocated to conditions and randomisation method described, (13) blinding of investigators, and (14) blinding of participants. Each item is scored as yes (2), partial (1), or no (0), with the total sum divided by the total possible sum. 35 That is, interventions were scored out of 28 and other studies were scored out of 22. Although there are criticisms of aggregate scoring for assessing methodological quality, especially for interventions and clinical trials, 31 this information can provide a crude assessment of whether study quality moderates reported effects. 41 Quality ratings were independently double-coded with moderate interrater reliability (κ = 0.69), which indicates substantial agreement. Any discrepancies were resolved through consensus, involving a third author.

The database search resulted in k = 9147 articles ( k = 6190 after removal of duplicates). After title and abstract screening, 73 articles underwent full-text review to assess for eligibility. A total of k = 32 articles preliminarily met the inclusion criteria for the current review. Of the 14 authors contacted regarding 22 articles, 12 provided additional data for 20 articles (91%). Data could not be obtained for 1 article, leaving a final total of k = 31 articles reporting 33 studies with a combined sample of n = 4849. See Preferred Reporting Items for Systematic Reviews and Meta-Analyses flow diagram in Figure 1 , and a list of included studies in Supplementary File 1 (available at http://links.lww.com/PAIN/C265 ). Preferred Reporting Items for Systematic Reviews and Meta-Analyses flow diagram. Among the 33 included studies, 15 investigated group differences in interpretation bias between chronic pain and control samples, and 25 studies measured the association between interpretation bias and at least 1 pain outcome. Three studies investigated the effects of Cognitive Bias Modification for Interpretation (CBM-I) as an intervention, and 3 further studies were longitudinal. A range of interpretation bias measures were used including ambiguous scenario tasks 29 , 37 ( k = 13), homophone (word association) tasks 42 ( k = 11), the homophone task modified with pictures ( k = 1), the incidental learning task 34 ( k = 3), the Bodily Sensations Interpretation Questionnaire 10 ( k = 2), the sentence generation task 51 ( k = 2), and the ambiguous word priming task 62 ( k = 1). Sample sizes varied, ranging from n = 37 to n = 1070. Of the studies included, k = 19 included chronic pain samples, k = 5 acute pain, k = 4 included samples with chronic illness and persistent pain, and k = 5 included experimental pain samples. For further details of study characteristics, see Table 1 . Study characteristics. AIBT, Adolescent Interpretation of Bodily Threat (pain/bodily sensations subscale only); BPI-SF, Brief Pain Inventory-Short Form; BSIQ, Bodily Sensations Interpretation Questionnaire (other bodily symptoms subscale only); CBM-I, Cognitive Bias Modification for Interpretation; Correlation (I), correlation between pain interference and IB; Correlation (PCS), correlation between PCS and IB; Correlation (S), correlation between pain severity and IB; IB, interpretation bias; MPQ, McGill Pain Questionnaire; PPIQ, Pelvic Pain Impact Questionnaire; PSI, Physical Symptoms Inventory; PVAS, Pain Visual Analogue Scale; RPEQ, Recent Pain Experiences Questionnaire. A total of 15 studies compared interpretation bias in people with (n = 1944) and without (n = 1232) chronic pain or chronic illness with persistent pain. There were no studies that compared interpretation bias in acute pain vs control. People with chronic pain displayed a greater interpretation bias than people without pain, with a medium effect size ( g = 0.602, k = 15, P < 0.001, 95% confidence interval [CI] 0.374-0.831), displayed in Figure 2 . The pattern was the same if studies with chronic illness samples were removed ( g = 0.560, k = 12, P < 0.001). Duval and Tweedie's trim and fill method suggested that there were no missing studies, and fail-safe N indicated that 853 studies would need to have been missed to render the effect nonsignificant. There was significant heterogeneity, Q 14 = 109.87, P < 0.001, I 2 = 87.26%. Between-group differences (Hedge g ) in chronic pain vs pain-free samples. We conducted a moderator analysis to compare effects by the task type for the ambiguous scenario task and the word association task. The effect was significantly greater for studies using the word association task than the ambiguous scenario task ( Q 1 = 20.87, P < 0.001). People with chronic pain showed a greater interpretation bias than people without chronic pain for the word association task with a large effect ( g = 0.899, k = 6, P < 0.001, 95% CI 0.639-1.160) and for the ambiguous scenarios task with a small effect ( g = 0.217, k = 6, P = 0.001, 95% CI 0.083-0.350). We then conducted meta-regression to assess the effect of study quality, which was not significant ( b = −0.170; 95% CI −3.956 to 3.617; P = 0.930). We identified sufficient studies to assess the association between interpretation bias and pain severity, and pain interference in pain samples (acute pain, chronic pain, and chronic illness with persistent pain). We also identified sufficient studies to assess the association between interpretation bias and experimental pain outcomes of pain intensity, pain threshold, and pain tolerance for pain-free samples. Finally, we identified sufficient studies to assess the relationship between interpretation bias and pain catastrophising including all (pain and healthy with experimental pain) samples. Forest plots of correlation analyses are provided in Supplementary File 2 (available at http://links.lww.com/PAIN/C265 ). The overall correlation between interpretation bias and pain severity was significant but small ( r = 0.109, k = 25, n = 3525, P = 0.001, 95% CI 0.047-0.171), although there was significant heterogeneity ( Q 24 = 66.73, P < 0.001, I 2 = 64.04%). A random effects funnel plot of the standard error indicated potential for bias or missing studies, and Duval and Tweedie's trim and fill method suggested an additional 8 studies to the left of the mean, with an adjusted random effects estimate of r = 0.037 (95% CI −030 to 0.103). By contrast, fail-safe N indicated the need for 194 additional studies to render the effect nonsignificant. We investigated the pain type as a moderator of the relationship between pain severity and interpretation bias; however, the groups did not significantly differ in their correlation with interpretation bias Q 2 = 1.59, P = 0.453. Nonetheless, the correlation between interpretation bias and pain severity was significant for acute pain ( r = 0.137, k = 4, n = 724, P < 0.001, 95% CI 0.064-0.208) and chronic illness with persistent pain ( r = 0.161, k = 6, n = 1754, P = 0.002, 95% CI 0.057-0.262), but not for samples with chronic pain ( r = 0.069, k = 16, n = 1047, P = 0.224, 95% CI −0.042 to 0.177). Next, we investigated the task type as a moderator of the relationship between pain severity and interpretation bias, although the only 2 tasks that had sufficient studies for comparison were the word association task and the ambiguous scenario task. The correlation between interpretation bias and pain severity was significant for the word association task ( r = 0.151, k = 10, n = 2287, P < 0.001, 95% CI 0.077-0.224), but not the ambiguous scenario task ( r = 0.079, k = 10, n = 986, 95% CI −0.008 to 0.164), although the difference was not significant ( Q 1 = 1.58, P = 0.209). Finally, we conducted meta-regression to assess the effect of study quality on the correlation between pain severity and interpretation bias, which was not significant ( b = 0.808; 95% CI −0.214 to 1.830; P = 0.121). The overall correlation between interpretation bias and pain interference was significant and small ( r = 0.161, k = 18, n = 3002, P < 0.001, 95% CI 0.110-0.212), with significant heterogeneity ( Q 17 = 27.77, P = 0.048, I 2 = 38.79%). Duval and Tweedie's trim and fill method indicated 1 potential study missed to the left of the mean, with a revised random effects point estimate of r = 0.158 (95% CI 0.107-0.208). The fail-safe N indicated that 266 missing studies would need to be available to render the effect no longer significant. The moderator of the pain type was not significant Q 2 = 0.006, P = 0.997, with the correlation between interpretation bias and pain interference remaining significant for chronic pain ( r = 0.160, k = 12, n = 869, P < 0.001, 95% CI 0.091-0.228), chronic illness with persistent pain ( r = 0.166, k = 4, n = 1527, P = 0.018, 95% CI 0.029-0.298), and acute pain ( r = 0.161, k = 3, n = 606, P < 0.001, 95% CI 0.082-0.238) samples. Similarly, the moderator of the task type was not significant ( Q 1 = 0.030, P = 0.863). The correlation between pain interference and interpretation bias was comparable when interpretation bias was measured with the ambiguous scenario task ( r = 0.184, k = 5, n = 365, P = 0.001, 95% CI 0.081-0.283) and the word association task ( r = 0.172, k = 7, n = 2116, P < 0.001, 95% CI 0.091-0.252). The moderator of the study quality was also not significant ( b = 0.217; 95% CI −0.798 to 1.232; P = 0.675). The association between interpretation bias and pain catastrophising was significant and small ( r = 0.138, k = 11, n = 1145, P = 0.007, 95% CI 0.038-0.235), with significant heterogeneity ( Q 10 = 26.37, P = 0.003, I 2 = 62.08%). Duval and Tweedie's trim and fill method suggested that no studies were missed, but the fail-safe N was relatively small with only 45 studies needed to render the effect no longer significant. The moderator of the pain type was not significant ( Q 1 = 0.18, P = 0.668). The association between interpretation bias and pain outcomes while completing an experimental pain task (eg, cold pressor) was not significant for pain severity ( r = 0.146, k = 4, n = 288, P = 0.507, 95% CI −0.279 to 0.523), time taken to register pain (ie, pain threshold, r = −0.023, k = 5, n = 387, P = 0.660, 95% CI −0.124 to 0.079), or how long the pain could be withstood (ie, tolerance time, r = −0.065, k = 5, n = 388, P = 0.211, 95% CI −0.165 to 0.037). The predictive nature of interpretation biases and subsequent pain outcomes were assessed in 4 samples of individuals in 3 articles. 8 , 19 , 54 Three of those samples were people with chronic pain, 8 , 19 and another was a sample of people with acute pain. 54 Two samples were followed up after a month, 19 while the other 2 were followed up 3 8 and 6 months later. 54 Two of the 4 studies identified a relationship between interpretation bias and subsequent pain intensity measured 3 to 6 months later, with small effect sizes ( r = 0.14-0.22, n = 126-264). The 2 studies that failed to find a relationship between interpretation bias and pain severity had shorter follow-ups (1 month) and smaller samples (n = 46-59). Regarding pain interference, 3 studies reported this outcome and none of them found direct relationships between interpretation biases and subsequent pain interference ( r = 0.10-0.14). However, one of these studies found an association between interpretation bias and pain anxiety, which subsequently predicted pain interference. 54 In one study to measure disability, Chan et al. 4 did find a relationship between interpretation bias and disability for 1 of 2 measures of interpretation bias, with small effect sizes ( r = 0.10-0.21). Hence, there is some evidence for a small relationship between interpretation bias and subsequent pain severity over longer time intervals (3-6 months). Studies with shorter follow-ups had smaller sample sizes, and therefore, a lack of significant relationship may be due to a lack of power. There have been 3 studies 2 , 29 , 50 that have applied cognitive bias modification in the context of pain. One study was conducted in an experimental setting 29 and 2 were conducted with people with chronic pain. 2 , 50 All studies used some form of an ambiguous scenario task to modify interpretation bias, where participants were given feedback as to whether their response was “correct” when they endorsed benign responses and “incorrect” when they endorsed pain resolutions. Training was conducted in a single setting in 2 studies 2 , 29 and over 4 sessions in the other. 50 All 3 studies measured the impact of the intervention on interpretation bias, and all 3 showed that the intervention successfully modified interpretation biases. Each study used different outcomes to assess the efficacy of CBM-I. Jones and Sharpe 29 examined the impact of CBM-I on experimental pain outcomes using the cold pressor task, including hesitance and tolerance as coprimary outcomes, and pain threshold, pain level, and pain-related distress as secondary outcomes. Their results (n = 98) showed that participants receiving CBM-I demonstrated less hesitance (a proxy for pain avoidance) than those who received training toward pain-related interpretations, which we calculated as a small-moderate effect (partial η 2 = 0.056). Interpretation bias fully mediated the effect. There were no group differences in pain tolerance or other pain outcomes. An et al. 2 compared CBM-I with a sham training in 48 people with chronic pain and assessed response to negative emotions in relation to pain outcomes, but did not measure pain. Cognitive Bias Modification for Interpretation led to a greater reduction in negative emotions to pain-related images than the sham control, with a large effect (partial η 2 = 0.59), and interpretation bias mediated the effect. In a larger study (n = 288), Sharpe et al. 50 found that among those with chronic pain, CBM-I led to lower pain intensity and pain interference compared with sham training, with small effects at posttreatment (d = 0.29) and 2 weeks later (d = 0.39). Cognitive Bias Modification for Interpretation was also found to significantly reduce fear of movement compared with sham. Mediation was not established, and there were no other effects on other secondary outcomes including pain catastrophising and psychological symptoms. In summary, there is, to date, little available evidence regarding the efficacy of CBM-I on pain outcomes. Research suggests that interpretation bias can be modified, and the only study in chronic pain patients to assess pain outcomes supports the likely efficacy of CBM-I for pain. However, this result requires replication. In the laboratory, Jones and Sharpe 29 did not find an effect of CBM-I on pain outcomes of severity and tolerance, and only found benefits for hesitance. Nevertheless, they did find that interpretation biases mediated the impact of group (CBM training) on hesitance, showing promise for interpretation bias as the causal mechanism of action. All 33 studies clearly defined the objective and used an appropriate study design for that objective. Articles generally did well at describing the characteristics of participants (29/33 studies met criteria), having well-defined outcome measures (28/33), using an appropriate sample size (26/33), sufficiently describing the analytic methods (30/33), controlling for confounding variables (28/33), reporting results in sufficient detail (31/33), and drawing conclusions that were supported by the results (32/33). Articles did not do as well at providing estimates of variance for the main results (20/33) or at sufficiently describing how participants were selected (22/33), with details surrounding recruitment strategies often missing. All 3 intervention studies randomly allocated participants to conditions. Only 1 study reported blinding of experimenters, and 2 studies reported blinding of participants. Study quality was moderately correlated with recency of publication ( r = 0.356, P = 0.042), suggesting that study quality has improved over time.

This review demonstrated that people with chronic pain displayed an interpretation bias compared with people without pain, with a moderate effect size. Interpretation bias was most pronounced when interpretation bias was measured with the word association task. Interpretation bias was also associated with the degree of pain interference, but not consistently associated with pain severity. Finally, preliminary evidence suggests that interpretation predicts later pain severity, at least in larger samples and over moderate time frames (3-6 months). Furthermore, there was consistent evidence that interpretation bias can be modified through CBM-I, with some promising results that CBM-I could impact pain outcomes. We found a moderate effect size ( g = 0.602) providing evidence that people with chronic pain show a bias toward pain-related interpretations of ambiguous information, compared with pain-free individuals. This effect was comparable with the effect size ( g = 0.67) in an earlier meta-analysis of 4 studies. 47 This consistency is encouraging because often early studies overestimate effects. These effects are also consistent with interpretation bias in mental health domains such as social anxiety (large effects), 9 depression (moderate effects), 16 health anxiety (moderate effects), 13 and even paranoia (moderate–large effects). 60 Our results suggest that interpretation biases toward pain-related interpretations among those with pain compared with those without are robust and reproducible. The reliability of these findings is further bolstered by a fail-safe N indicating that more than 50 times as many studies would need to have been missed to render the effect nonsignificant. We were not able to assess this effect in acute pain, as we did not identify any studies that measured interpretation bias in acute pain that also included a control group. For other cognitive biases such as attentional bias, research has confirmed the presence of such biases in people with chronic pain compared with pain-free samples. 12 , 56 However, studies have not consistently found an association between degree of attention bias and pain outcomes. By contrast, our meta-analysis identified a small but significant association between interpretation bias and both pain interference and pain severity, although the latter effect was less robust and susceptible to publication bias. These findings suggest that interpretation bias may be more closely linked to the extent to which pain disrupts daily functioning, rather than its severity. The fear-avoidance model 63 , 64 suggests that it is the degree to which people interpret their pain sensations as harmful that prompts catastrophizing and, in turn, pain-related fear and distress that fuel the escalating cycle of pain and disability. Future research would benefit from measuring pain-related fear in studies of interpretation bias to further probe the mechanisms through which interpretation bias affects pain outcomes. We did find a small association between interpretation bias and pain catastrophising, suggesting that these constructs are distinct but potentially interinfluential. Although associations were found between interpretation biases in clinical samples and pain outcomes, we found no association between interpretation bias and experimental pain outcomes. These findings suggest that interpretation bias is not activated in pain-free samples by the anticipation of a painful experience. Given that there is no reason for participants in an experiment to interpret the experimental procedure as potentially harmful, this is perhaps unsurprising, but does raise questions as to whether interpretation bias can be meaningfully studied in the laboratory. Importantly, in the only clinical study of interpretation bias in people with acute pain, interpretation bias predicted pain severity and interference 3 months later. Interestingly, path analyses confirmed that the causal pathway was indirectly through increased pain anxiety. 54 Further assessment of interpretation biases in acute pain and their potential role in the development of chronic pain is warranted. Although robust differences in interpretation bias were observed between people with and without persistent pain, the effect was moderated by task. Specifically, the word association task gave rise to large effects, whereas the ambiguous scenario task led to small but significant effects. Similarly, the association between interpretation bias and pain severity was present when measured with the word association task but not the ambiguous scenario task, although both tasks were comparable in detecting relationships with interpretation bias and pain interference. While the previous meta-analysis of interpretation bias in pain 47 showed low heterogeneity of effects, only 3 homograph (eg, word association) and 1 incidental learning task studies were meta-analysed. Therefore, our findings of differences between word association and scenario-based tasks are not inconsistent. Furthermore, a meta-analysis of interpretation bias in social anxiety showed that effects were moderated by task factors. 9 The word association task is relatively simple where participants respond to ambiguous words with the first word that comes to mind. These words are fairly general (eg, back and relief). By contrast, the ambiguous scenario task presents a range of specific scenarios, followed by sentences that either resolve the scenario in a pain-related way or a benign way, which are then rated by participants on their similarity to the ambiguous scenario. As such, the scenarios may not be as relevant to the individual and their pain experience. Some researchers have attempted to address this issue by modifying the scenario task, for example, by including likelihood ratings as well as similarity ratings. 26 Although word association and scenario-based tasks have been favoured in the past decade since Schoth and Liossi's earlier review, 47 there are other methods for assessing interpretation bias that show promise but had insufficient studies to be tested as modifiers, such as the sentence generation task 51 and the incidental learning task. 34 Furthermore, novel tasks that assess interpretation to ambiguous sensations could be developed 65 which may yield more ecologically valid results. Compared with cross-sectional studies, there were fewer prospective studies. Although the results were mixed, they potentially support a small relationship between interpretation bias and subsequent pain outcomes. The 2 studies 8 , 54 that found small relationships between interpretation bias and subsequent pain outcomes in pain samples had larger samples (n > 120), while the 2 studies 19 that failed to find an effect were much smaller (n < 60) and potentially underpowered to detect small effect sizes. Furthermore, only 1 prospective study investigated acute pain and found evidence that interpretation bias at baseline was associated with pain severity 3 months later and indirectly with pain interference through pain anxiety. 54 Although these results require replication, the early data are encouraging. Similarly, we identified only 3 studies that used CBM-I in the context of pain. Interpretation bias was successfully modified across all 3 studies. 2 , 29 , 50 It is important to establish that interpretation bias can be modified, given the inconsistency with which CBM interventions have successfully modified cognitive bias in other literatures. 11 , 23 Importantly, 2 of the studies found that induced interpretation bias mediated group effects on outcomes, 2 , 29 confirming that interpretation bias is the likely mechanism of action, although the third study did not establish mediation. 50 Each of the 3 studies investigated the impact on different outcomes, which makes synthesis difficult. Nevertheless, all 3 studies found changes following CBM-I compared with a sham intervention in at least 1 outcome, suggesting that more research is indicated. It is worth noting the limitations of this meta-analysis. The moderator analysis by the task type was limited to the word association task and ambiguous scenario task because of the low number of studies. There are few studies exploring the reliability, structural, and construct validity of available tasks, nor research comparing them directly. Similarly, there were a limited number of longitudinal and intervention studies and they used different samples, time frames ,and outcomes, and as a result, no meta-analysis could be conducted. Because interpretation bias in CBM-I trials are typically measured in recognition tasks, interpretation bias was not measured before the intervention, relying on differences between groups at posttreatment. Despite this, most quality criteria were met by most studies, although reporting of variance estimates and descriptions of participant selection were often lacking, and should be included in future research. Analyses also indicated potential bias due to missing studies measuring pain severity and interference, although this concern can be partly alleviated by large fail-safe N numbers. Based on the results of this meta-analysis, we make the following recommendations to improve the assessment of pain-related interpretation bias. This review identified a gap in prospective data on interpretation bias and pain. There was only 1 study in acute pain, and no daily diary studies that would be helpful for establishing causality, which have been helpful in other fields such as within sleep psychology. 39 Interpretation bias predicted pain outcomes over longer, rather than shorter periods. Extending the follow-up duration could help determine whether interpretation bias during acute pain serves as a marker for future chronic pain. As with prospective studies, there were few studies of CBM-I and there were no studies of CBM-I in the management of acute pain. Future research is needed to confirm the efficacy of CBM-I. Finally, most studies measured interpretation bias as the only cognitive bias. Cognitive bias research in depression supports a combined cognitive bias hypothesis, 15 with studies in pain also indicating connections between these biases. 6 , 8 , 65 Research that explores the interactions of interpretation bias with other cognitive biases, such as attentional bias, as well as other psychological processes, such as imagery, 53 could provide a more complete understanding of cognitive processes and how they contribute to pain. In summary, this meta-analysis confirmed that people with chronic pain tend to interpret ambiguous information as pain-related more than people without chronic pain. This interpretation bias was greatest when assessed with the word association task, which may be favoured going forward unless more ecologically valid tasks or stimuli emerge. Interpretation biases were related to pain interference and, less robustly, to pain severity. There was emerging evidence to suggest that interpretation biases were associated with subsequent pain outcomes and could be modified to improve pain outcomes. Clinically, this research suggests that clinicians should consider the way in which people with chronic pain interpret ambiguous sensations, pain, or related situations that have the potential to amplify pain. Available interventions, 49 such as cognitive–behaviour therapy, can address the content of unhelpful interpretations, while CBM-I targets the automatic processes of interpreting information in a pain-related way. Future research could focus on testing the efficacy of administering CBM-I as an adjunct to other evidence-based treatments such as cognitive–behaviour therapy or even as a potential prevention or early intervention given highly scalable and low-cost nature.

One of the biggest challenges with chronic pain is determining why for some people, pain resolves, while for others, acute pain becomes chronic. Identifying the factors that exacerbate and maintain chronic pain can help establish modifiable treatment targets to prevent or treat chronic pain. This is important because chronic pain is not only highly prevalent globally 17 , 28 , 66 but also costly to the individual and society, surpassing the respective costs of heart disease, cancer, and diabetes. 21 Chronic pain is regarded as a biopsychosocial phenomenon, whereby complex interactions between biological, psychological, and social factors account for both the pain experience and disability. 22 One of the psychological factors that is believed to underlie chronic pain is the degree to which people interpret that pain indicates harm. For example, a twinge in the back could be experienced as painful and interpreted as something seriously wrong, or it could be dismissed as harmless. The interpretation of ambiguous information as pain-related or threatening is described as an interpretation bias. The fear-avoidance model of pain suggests that interpretation bias drives catastrophic thinking about pain leading to efforts to avoid further pain. 63 , 64 However, when pain is no longer associated with tissue damage, avoidance of pain can inadvertently perpetuate chronic pain and disability. Other models such as the threat-interpretation model of pain 58 also highlight how interpreting information as painful and threatening can lead to an unhelpful pattern of hypervigilance and attentional avoidance that promote and maintain chronic pain. A small review by Schoth and Liossi 47 found support for this premise, confirming that people with chronic pain showed a greater pain-related interpretation bias than those without pain ( g = 0.67). However, this review included 7 studies, only 4 of which were meta-analysed. Since their search in early 2015, research on pain-related interpretation biases has proliferated, more than tripling in the past decade. Moreover, the previous review had a limited scope, focusing solely on comparing interpretation bias in people with and without chronic pain, through cross-sectional research. Currently, it is not known whether pain-related interpretation biases differ between those with chronic and acute pain, or whether interpretation bias is associated with degree of pain severity or pain interference. There has also been no review to synthesize longitudinal research to determine whether interpretation bias may play a causal role on pain. Specifically, there is no synthesis of whether interpretation bias predicts later pain outcomes or whether interventions for modifying interpretation bias affect pain outcomes. This is important because the presence of pain-related interpretation bias could simply be an epiphenomenon, such that being in pain is associated with interpretation bias. Alternatively, if it were shown that interpretation biases predict later pain or, when modified, can reduce pain, then interpretation bias could be a novel treatment target in acute and/or chronic pain. We therefore aimed to systematically review and meta-analyse available data to determine whether: (1) pain-related interpretation biases differ between those with chronic and acute pain compared with those with no pain, (2) pain-related interpretation bias is associated with chronic, acute, or experimental pain outcomes, (3) whether interpretation biases predict subsequent pain; and (4) whether interventions for pain-related interpretation bias affect pain outcomes.

Supplemental digital content associated with this article can be found online at http://links.lww.com/PAIN/C265 .

The authors have no conflicts of interest to declare.