An estimate of participants' intelligence quotient (IQ) was obtained using a two-test short-form of the Wechsler Adult Intelligence Scale-III (WAIS-III, Wechsler, 1997a), the Digit-Symbol Coding and Information subtests. Estimated full-scale IQ obtained using this dyadic short form of the WAIS-III has been shown previously to correlate (r² = 0.82) with IQ-values obtained using the full 12-subtest scale (Sattler and Ryan, 2001), and this combination of subtests was chosen based on its relatively short administration time. In the Digit-Symbol Coding subtest, participants were presented with nine digit-symbol pairs followed by a list of digits only. Participants were required to fill in the symbol corresponding to each digit as quickly as possible on a standard form. For the Information subtest, the participant was required to answer a series of factual general knowledge questions about common events, objects, places, and people. Raw scores for both subtests were converted to age-adjusted scaled score equivalents using the WAIS-III Administration and Scoring Manual (Wechsler, 1997a). Full-scale IQ was then estimated from the sum of the scaled scores using the extrapolation tables taken from Sattler and Ryan (Sattler and Ryan, 2001).

The Logical Memory subtests I and II of the Wechsler Memory Scale-III (WMS-III, Wechsler, 1997b) were used to assess short- and long-term verbal memory and recognition memory. Two different stories (A and B) were read to the subject, and immediately afterwards the subject was asked to retell the story from memory.

After an interval of ~25 min, the participant was again asked to recall as many details as possible from both stories A and B. For recognition memory, the participant was required to give “Yes” or “No” answers to a set of 30 questions relating to stories A and B, which includes a mixture of correct and incorrect statements about the story content. Participants were scored on the accuracy of the story recall (“story” and “theme” units) and number of correct responses to recognition questions. Raw scores were converted to age-adjusted scaled score equivalents using the WMS-III Administration and Scoring Manual (Wechsler, 1997b).

The spatial span subtest of the WMS-III was used to measure short-term spatial memory capacity. The test was administered using the spatial span board, which consists of 10 cubes with the numbers 1–10 printed on the sides of the cubes facing the examiner. For spatial span forward, the examiner tapped the cubes in a specified sequence pattern and then asked the participant to tap the same sequence. For spatial span backward (reverse), the participant was asked to tap the sequence in reverse order. The sequence length increased until the participant could no longer replicate the sequence correctly. Raw scores were converted to age-adjusted scaled score equivalents using the WMS-III Administration and Scoring Manual (Wechsler, 1997b).

The Continuous Performance Test—Identical Pairs (CPT-IP), adapted from the MATRICS (Measurement and Treatment Research to Improve Cognition in Schizophrenia) Consensus Cognitive Battery (Nuechterlein and Green, 2004), was used to measure attention and vigilance. This computerized test required the participant to monitor a series of 4-digit numbers as they appeared briefly on the computer screen. The participant responded to the sequential appearance of identical pairs of numbers by clicking a mouse as quickly as possible. The number of correct responses, the number of incorrect responses and the reaction time were recorded automatically. Incorrect responses were categorized as “false alarms” (i.e., responses to similar, but not identical, numbers presented in sequence) or random responses. The Dprime (range 0–4.2), an indication of the rate of hits to false alarms, was also calculated as an index of attention. The average reaction times to both hits and false alarms during the CPT-IP trial were used as tentative measures of psychomotor speed.

Executive function was assessed using the Wisconsin Card Sorting Test—Computerized Version 4 (WCST-CV4). This test measured the respondent's ability to adapt to changing schedules of reinforcement (i.e., the “rules” about the task) and thus assesses cognitive flexibility, a key component of executive functioning (Berg, 1948). The computerized version used in this study presented the participant with four key cards and a stimulus card. The cards were matched according to one of three categories: color, shape or quantity of items on the card. Matching is achieved by placing the cursor on the key card selected and left clicking the mouse. The participants were not told how to match the cards, but they were given verbal and visual feedback on whether each match was “right” or “wrong.” The category by which the cards were to be matched changed without warning after presentation of every ten stimulus cards; the subject was required to recognize the changed “rules” and identify the new presentation pattern. Raw scores and demographically (age and education) corrected standard scores were computed for five outcome measures: percentage errors, percentage perseverative (repetitive) responses, percentage perseverative errors, percentage non-perseverative errors and, percentage conceptual level responses. Raw scores were also computed for the number of categories completed, the number of trials to complete the first category, number of failures to maintain set and a learning-to-learn score (indicative of conceptual efficiency across consecutive categories).

Data were analyzed using the Statistical Package for the Social Sciences (SPSS, version 18, IBM Corp., USA) for Windows. The statistical analyses were performed in two distinct phases; the first included both patient and control participant data, while the second included patient data only.

In this phase of analysis, hypotheses (1) and (2), as stated in the Introduction, were tested. The first level of analysis was the computation of descriptive and inferential statistics. This was followed by correlation analyses of cognitive outcome measures and participant characteristics aimed at identifying potential covariates. The variables included in the correlation analyses were: age, gender, number of years of education, smoking status and time since last nicotine consumption, time since last caffeine consumption, and depression and anxiety scores. These variables can predict cognitive performance (see for example, review by Moore et al., 2009), effect of alcohol (Rohrbaugh et al., 1988), nicotine (Foulds et al., 1996; Mancuso et al., 1999), and caffeine (Brice and Smith, 2001; Kelemen and Creeley, 2001). We intended controlling for correlates of the outcomes when testing our hypotheses. This would allow us to examine the extent to which the hypothesized effects emerged, having held constant important correlates of the outcome being investigated.

Hierarchical multiple regression analyses were used to test the hypotheses; groups were coded as −1 (patient group) and +1 (control group), age was mean-centered, and an interaction term [group (−1/+1) × age (mean-centered)] was calculated. The purpose of centring was to minimize multicollinearity (arising, in this case, from entering both the variables and their interaction into the regression equation). Participant-characteristic variables that were correlated with the cognitive outcomes were entered into the first block of the regression model. Values for significance levels and β coefficients quoted in the text are for the overall model. Interaction plots for estimated marginal means (2 × 2 ANCOVA with patient vs. control and age median-split at 45 years) were used for visualization of the directionality of effects and for determining the nature of the interaction between variables (data not shown).

The second phase of the study tested hypothesis (3), as stated in the Introduction. As in Phase I, an interaction effect between pain variables and age was also hypothesized, and this was tested in the analyses. The pain variables investigated were present pain intensity, pain-related disability score, number of painful areas, and pain chronicity (number of months since diagnosis of pain). Additional patient characteristics investigated as potential covariates were the presence or absence of medication (and different medication sub-groups), total number of medications, percentage pain relief from medications, and self-assessed impairment of cognitive function. Descriptive and inferential statistics relating to pain information (for patient participants only, n = 38) were calculated. Correlation matrices, using the patient group data only, were constructed to identify potential covariates. Each of the pain variables was centered, and interaction terms with mean-centered age were calculated. Hierarchical regression analyses were again used to test the hypothesis.

Prior to testing the hypothesized effects, group differences (patient vs. control) were explored. Results indicated consistently lower outcome scores in the patient group, or higher scores on reverse scales (number of false alarms and number of random responses in the CPT-IP, and raw scores for % errors, % perseverative responses, % perseverative errors, % non-perseverative errors, number of trials to complete the first category and failure to maintain set in the WCST, see Table 3). Significant differences between groups, where observed, are indicated in Table 3.

A correlation matrix of participant characteristic and cognitive outcome variables was used to identify potential covariates (see Table S1). Gender and smoker/non-smoker classifications were correlated with a number of cognitive measures (gender correlated with immediate verbal memory and reaction time; cigarette smoking correlated with immediate verbal memory), while, as expected, a large number of the outcomes were positively correlated with years of education and negatively correlated with measures of depression and state anxiety. Interestingly, the duration since last consumption of caffeine was positively correlated with scores measuring attentional performance. Thus, an increase in the time since participants had last consumed caffeine was associated with higher scores, an effect inconsistent with the recognized acute effect of caffeine as a CNS stimulant (Brunye et al., 2010).

The hypotheses predicted an effect of participant group (patient or control) and an interaction effect of group and age, and were tested through a series of individual hierarchical regressions for each of the cognitive outcomes. Significant covariates of the dependent variable (identified by correlation analysis) were entered in the first block, group (patient vs. control, coded −1/+1) and age (mean-centered) were entered in block 2, and the interaction term (pain × age) was entered in block 3. Statistics for cognitive variables predicted by group or by the interaction term (at levels close to or below the p = 0.05 level of statistical significance) are presented in Table 4.

In the regression model predicting estimated full scale IQ (FSIQ), years of education, PHQ-9 severity score and STAI-S total score were entered as controls in block 1. PHQ severity score and STAI-S total score were found to be correlated; however, the observed variance inflation factors (VIFs) were <10, suggesting that the assumption of minimal multicollinearity was not violated and that scores can be assumed to reflect independent constructs. Years of education significantly contributed to the FSIQ model (β = 0.37, p < 0.001), as did the presence of pain (β = 0.33, p = 0.04). Interaction plots showed that FSIQ scores were decreased in the presence of pain. Neither age nor the interaction term made any significant contribution to the model (see Table 4).

Immediate verbal memory was measured as story-unit and theme-unit recall scores and a learning-slope score. Presence of a chronic pain diagnosis, age, and their interaction term made no contribution to the story-unit or theme-unit regression models. There were no significant predictor variables of the theme-unit model, and only gender significantly contributed to the story recall unit model (β = 0.24, p = 0.04). In the case of learning slope, the contribution of STAI-S total score (entered in the first block) approached significance (β = −0.27, p = 0.06). There was a significant contribution of age (β = −0.29, p = 0.02); the interaction term of pain and age was also associated with an effect close to statistical significance (β = 0.20, p = 0.08, Table 4), with interaction plots suggesting that older pain patients performed worse than controls of a similar age and worse than younger pain patients.

Delayed verbal memory was also measured with story-unit and theme-unit recall scores, as well as with a recognition-memory index and with percentage retention. In a manner similar to the case of immediate verbal memory, neither pain condition, age, nor their interaction significantly contributed to either story-unit or theme-unit models, and none of the additional predictor variables significantly affected these models. Furthermore, no significant predictors of percentage retention were identified (though the effect of age was close to significance: β = −0.21, p = 0.08). Recognition memory, however, was significantly predicted by both age (β = −0.27, p = 0.01) and the pain-age interaction term (β = 0.30, p = 0.006, Table 4), with poor performance particularly evident in older pain patients.

Spatial memory was assessed in forward and reverse trials of the spatial span task. Total spatial span and forward trial scores were not predicted by any of the three main independent variables of interest (pain, age, pain × age interaction), or by any of the control measures entered into the model. Effects close to the level of statistical significance were associated with the presence of pain (β = 0.21, p = 0.07) and years of education (β = 0.23, p = 0.06) in the reverse scores model (Table 4).

Attention was measured using the CPT, the outcome variables of which were the numbers of hits, false alarms and random responses. The number of hits in the CPT was significantly predicted by age (β = −0.36, p = 0.002) and there was also a trend for a contribution of chronic pain (β = −0.36, p = 0.06, Table 4). The control variable of depressive-symptom score (PHQ severity) was also a significant contributor to the model (β = −0.59, p = 0.002, Table 4). The slopes of the interaction plot indicated a reduced number of hits in the control group compared with the patient group; however, the group means in fact indicate an increased number of hits in the control participants compared with the patient group. Further analysis revealed a strong negative association between PHQ severity score and pain condition (r² = −0.83, p < 0.001). Therefore, despite tolerance and VIF-values being within accepted limits, the relationship between PHQ severity score and pain condition may have affected the overall predictive capability of the model. Age (β = 0.28, p = 0.02) and the age × pain interaction term (β = −0.25, p = 0.02) made significant contributions to the regression model for number of false alarms (Table 4). The number of random responses was influenced by age (β = 0.36, p = 0.001) and there was a near-significant effect for the age × pain interaction term (β = 0.20, p = 0.06, Table 4). In the case of the model fitted for D-prime scores, age (β = −0.34, p = 0.003) and PHQ severity scores (β = −0.33, p = 0.07) emerged as predictors. There were no significant effects of pain or pain-age interaction.

The reaction times to both hits and false alarms in the CPT task were considered measures of psychomotor speed. There were no main effects of age or pain. Though the individual contributions of age and presence of pain were non-significant, the pain-age interaction term significantly contributed to the model for “hit” reaction time (β = 0.24, p = 0.03, Table 4). This interaction effect was approaching significance in the case of “false alarm” reaction time (β = 0.21, p = 0.08), though the overall model was not significant. The interaction plots suggest that reaction time increased with the presence of pain in the younger participants but was reduced in the presence of pain in the older group.

The independent variables of pain, age, and age × pain interaction did not predict any of the WCST standard scores or the secondary outcomes (number of categories completed, number of trials to first category, failure to maintain set and learning-to-learn score). The only significant predictor was age (number of categories completed: β = − 0.28, p = 0.02), whereby older people achieved lower scores than younger participants.

Table 5 shows descriptive statistics for variables unique to the pain patient group and Table 6 provides an overview of patients' medication status. Seventy-nine percent of patients were receiving long-term medication for the treatment of pain. Opioids were the most commonly prescribed class of medication, followed by anticonvulsants. The “other” category included classes of analgesics taken by a small number of participants within the sample. These treatments included: antispasmodics, benzodiazepines and other sedative hypnotics, local anesthetics, and transient receptor potential vanilloid 1 (TRPV1) ligands (capsaicin), as well as adjunctive therapies such as acupuncture and spinal-cord stimulation. In cases where a drug had more than one indication (for example, benzodiazepines used as a muscle relaxant, as an anxiolytic, or as treatment for insomnia), the precise reason for which it was prescribed was not queried. More than half (55%) of the patient participants were taking three or more different drugs to manage their pain. The average subjectively-reported relief provided by pharmacological treatments was 37%.

A correlation matrix of cognitive outcome variables and general participant characteristics was again constructed to identify significant correlations to be entered as control variables in the regression models (patient data only). Similarly to Phase I, years of education correlated with a large proportion of the outcome measures (see Table S2). The only other significant correlations were between: duration since last consumption of nicotine and the spatial span forward trial; depression score and number of hits, false-alarm reaction time and WCST failures to maintain set; and STAI-S total score and false alarm reaction time. Correlations between cognitive outcomes and patient-specific variables, including medications and patients' subjective assessment of the effect of their pain on cognitive function were also investigated (see Table S3). Anticonvulsant treatment was found to correlate significantly with learning-slope performance (r_pb = 0.37, p < 0.05) suggesting lower scores in those receiving anticonvulsant medications. On the other hand, antidepressant treatment was associated with higher scores on the attention task (r_pb = −0.38, p < 0.05 D-prime) and improved WCST performance as indicated by a decrease in the number of trials required to complete the first category (r_pb = 0.34, p < 0.05). Self-assessment of the effect of pain on cognitive function was positively correlated with the failure to maintain set in the WCST (r_pb = 0.45, p < 0.01). All significant correlations were entered as control variables in the hierarchical regression analyses. The hypothesis stated that pain variables, and the interaction of these variables with age, would predict differences in cognitive outcomes, and this was again tested through a series of hierarchical regressions. Significant covariates of the dependent variable were entered in the first block. Each pain variable (present pain intensity, average pain intensity, pain-related disability score, pain chronicity and the number of painful areas) was tested individually. Both the pain variable and age were centered and entered in block 2, and the interaction term was entered in block 3.

Pain chronicity was found to contribute significantly to the models of FSIQ (β = 0.31, p = 0.05), immediate verbal memory (β = 0.47, p = 0.003) and delayed verbal memory (β = 0.40, p = 0.02). The association was positive in each of these cases, suggesting that performance improved with an increasing duration of pain. The contribution of chronicity was also close to significance in the hit response reaction time model (β = 0.36, p = 0.07), which would suggest an increase in reaction time with increased pain chronicity.

The number of painful areas was found to be an effective predictor of immediate verbal memory (story-unit recall: β = −0.34, p = 0.05; theme-unit recall: β = −0.37, p = 0.04) and delayed verbal memory (β = −0.35, p = 0.04), whereby “more painful sites” was associated with poorer performance. The interaction of the number of painful areas and age was a significant contributor to the model of the spatial span reverse trial (β = −0.36, p = 0.03, ANOVA p = 0.06) and the number of random responses on the CPT (β = 0.30, p = 0.03).

Chronic pain grade significantly predicted D-prime scores of attention (β = −0.43, p = 0.002), as did disability score (β = −0.28, p = 0.03). The interaction term of disability score by age also made a significant contribution to the memory retention model (β = −0.47, p = 0.006), while the interaction of present pain intensity and age had an effect close to significant on the number of hits in the CPT (β = −0.31, p = 0.07). The negative association indicates that worse pain was associated with a lower performance level on this task. In general, therefore, aspects of verbal memory and attention appear to be related to a variety of pain symptom and disability variables. When the total number of medications, and individual medication classifications, were entered as variables of interest, no main effects were observed on any of the cognitive outcome measures.