Predicting depression in rheumatoid arthritisThe signal importance of pain extent and fatigue and comorbidity.код для вставкиСкачать
Arthritis & Rheumatism (Arthritis Care & Research) Vol. 61, No. 5, May 15, 2009, pp 667– 673 DOI 10.1002/art.24428 © 2009, American College of Rheumatology ORIGINAL ARTICLE Predicting Depression in Rheumatoid Arthritis: The Signal Importance of Pain Extent and Fatigue, and Comorbidity FREDERICK WOLFE1 AND KALEB MICHAUD2 Objective. To determine the incidence of self-reported depression (SRD) in rheumatoid arthritis and to identify and rank clinically useful predictors of depression. Methods. We assessed 22,131 patients for SRD between 1999 and 2008. We collected demographic, clinical and treatment data, household income, employment and work disability status, comorbidity, scales for function, pain, global, and fatigue, the Regional Pain Scale (RPS), the Symptom Intensity (SI) scale (a linear combination of the RPS and the fatigue scales) and linear combinations of the Health Assessment Questionnaire, pain and global severity. We used logistic regression analyses with multivariable fractional polynomial predictors, and Random Forest analysis to determine the importance of the predictors. Results. The cross-sectional prevalence of self-reported depression was 15.2% (95% conﬁdence interval [95% CI] 14.7– 15.7%) and the incidence rate was 5.5 (95% CI 5.3–5.7) per 100 patient years of observation. The cumulative risk of SRD after 9 years was 38.3% (95% CI 36.6 – 40.1%). Almost all variables were signiﬁcant predictors in logistic models. In Random Forest analyses, the SI scale, followed by comorbidity, best predicted self-reported depression, and no other variable or combination of variables improved prediction compared with the SI scale. Conclusion. Pain extent and fatigue (SI scale) are the dominant predictors of SRD. These variables, also of central importance in the symptomatology of ﬁbromyalgia, are powerful markers of distress. A strong case can be made for the inclusion of these assessments in routine rheumatology practice. In addition, actual knowledge of comorbidity provides important insights into the patient’s global health and associated perceptions. INTRODUCTION Depressive symptoms and depression are more common in persons with rheumatoid arthritis (RA) (1) and other chronic illnesses (2) compared with healthy subjects, and add substantially to the distress of a chronic illness (3). In patients with RA, social disadvantage (lower education level, unmarried status, lower income), female sex, and illness severity are among common groups of variables that contribute to depression, as they do across all chronic illnesses. In patients with RA, speciﬁc measures that contribute to depressive illnesses include pain, functional loss, and impact on daily activities (1,4,5). 1 Frederick Wolfe, MD: University of Kansas School of Medicine and National Data Bank for Rheumatic Diseases, Wichita, Kansas; 2Kaleb Michaud, PhD: University of Nebraska Medical Center, Omaha, and National Data Bank for Rheumatic Diseases, Wichita, Kansas. Address correspondence to Frederick Wolfe, MD, National Data Bank for Rheumatic Diseases, 1035 North Emporia, Suite 288, Wichita, KS 67214. E-mail: fwolfe@ arthritis-research.org. Submitted for publication August 25, 2008; accepted in revised form January 16, 2009. Among the problems faced by researchers regarding RA and depressive illness are to identify all of the factors that contribute to depressive illness, and to quantify the relative importance of these factors. Relative importance is crucial, otherwise one is left with a multitude of factors, but also the impossibility of adequately understanding the contribution and clinical relevance of these factors. However, to a large extent, useful determination of relative importance of depressive illness not been done. Most studies of RA and depression have not examined a full range of variables, have not dealt with nonlinear relationships, and have foundered on differences between causal models and clinical models. Causal factors related to depression in the absence of chronic illness include traumatic experiences, genetic factors, temperament, interpersonal relations (6), family vulnerability (7), and hypothalamic⫺pituitary⫺adrenal (HPA) axis dysregulation (8). However, these factors are almost impossible to measure clinically. The goal of this study was to identify and quantify useful clinically important predictors of depressive symptoms. We used 2 major methods to identify predictors of self-reported depression, logistic regression using frac667 668 tional polynomial predictors (9), and Random Forest analysis (10,11) to identify the predictors and to rank them quantitatively. In addition, we have reported the incidence rate, cumulative incidence, and cumulative risk of selfreported depression in order to place the predictive data in proper perspective and to measure the burden of selfreported depression in RA. PATIENTS AND METHODS Patient population. We studied participants in the National Data Bank for Rheumatic Diseases (NDB) longitudinal study of rheumatic disease outcomes. NDB participants are diagnosed by US rheumatologists and are recruited from their practices. Participants indicated in their rheumatologist’s ofﬁce that they were willing to participate in long-term survey research (12). Patients are followed prospectively with semiannual, detailed, 28-page questionnaires, as previously described (13–15). Approximately 8% of patients discontinue participation per year. This report utilized NDB data in a longitudinal cohort analysis of 29,524 adult participants (ages 18 –103 years), of whom 22,131 had RA and 3,717 had a noninﬂammatory rheumatic disease, 1,002 had systemic lupus erythematosus (SLE), and 2,674 had ﬁbromyalgia. In contrast to unselected patients from rheumatology practices, 7,597 patients with RA were enrolled speciﬁcally as part of a treatment safety registry. Safety registry patients (a subset of the 22,131 RA study patients) were enrolled in their rheumatologist’s ofﬁce at the time they started a speciﬁc therapy. Diagnostic groups were mutually exclusive. Three hundred thirty-three patients with both RA and SLE or with both SLE and ﬁbromyalgia were excluded from the study. Patients with RA could have other conditions with the exception of SLE. Patients were enrolled continuously beginning in 1999 and ending in January 2008. Rheumatic disease diagnoses were made or conﬁrmed by the patient’s rheumatologist. Noninﬂammatory rheumatic diseases included diagnoses such as osteoarthritis, back pain syndromes, tendonitis, etc., excluding ﬁbromyalgia. Study variables were assessed at entry into the NDB and at every subsequent semiannual questionnaire. Study variables. Demographic variables included age, sex, ethnicity, education level, marital status, and total household income. Work and disablity status variables were based on patient self-report. Speciﬁcally, we asked patients which of the following categories best characterized their current work status: paid work, housework, student, unemployed, disabled, or retired. Validation studies have demonstrated the reliability of the work and disability assessments (16). Questions relating to depression and mood included self-reported depression and the Medical Outcomes Study Short Form 36 (SF-36) mood and mental component summary (MCS) scales (17). We assessed self-reported depression by a single question in each semiannual survey (paraphrasing from a table of questions), “Have you had a problem with depression in the last 6 months.” As evi- Depression Prediction in RA Table 1. Prevalence of self-reported depression in RA and other rheumatic disease* Group N Percentage (95% CI) Fibromyalgia SLE RA NIRD 2,674 1,002 22,131 3,717 37.9 (36.1–39.8) 32.5 (29.6–35.4) 15.2 (14.7–15.7) 14.7 (13.6–15.8) * RA ⫽ rheumatoid arthritis; 95% CI ⫽ 95% conﬁdence interval; SLE ⫽ systemic lupus erythematosus; NIRD ⫽ noninﬂammatory rheumatic disease. dence of validity, self-reported depression was signiﬁcantly associated with the SF-36 mood and MCS scale. The area under the receiver operating characteristic (ROC) curve for mood was 0.826 and Kendall’s tau-a was 0.164; the area under the ROC curve for the MCS scale was 0.823 and Kendall’s tau-a was 0.166. Comorbidity was measured by a patient-reported composite comorbidity score (range 0 –9) comprised of 11 present or past comorbid conditions including pulmonary disorders, myocardial infarction, other cardiovascular disorders, stroke, hypertension, diabetes mellitus, spine/hip/ leg fracture, depression, gastrointestinal ulcer, other gastrointestinal disorders, and cancer (18,19). For the purposes of this study, self-reported depression was omitted from the scale so that it would only assess nondepressive comorbid conditions. As a consequence, the range in this study was 0 – 8. Questions related to RA severity included the Health Assessment Questionnaire (HAQ) disability index (20,21), visual analog scales (VAS) for pain and patient global severity, the Patient Activity Scale (PAS) (22), the Regional Pain Scale (RPS) (23), and the Symptom Intensity (SI) scale (24). The PAS is formed by multiplying the HAQ score by 3.33 and then dividing the sum of the VAS pain, the VAS global, and the HAQ score by 3. This yields a 0 –10 scale. The PAS is a composite patient measure of RA activity. The RPS is a self-administered count of the number of painful nonarticular regions with a range of 0 –19. The SI scale measures the combination of fatigue and pain extent. Derived from the fatigue VAS and the RPS, the SI scale combines these 2 measures in a continuous form according to the following formula: (VAS fatigue ⫹ [RPS/2])/2. This yields a scale with a 0 –9.75 range. To ascertain death status, we searched the National Death Index (25) annually for deaths between 1998 through 2007 and also received reports of deaths from family and friends. Statistical analyses. The cross-sectional prevalence of self-reported depression was calculated at a randomly selected observation, assuming a Poisson distribution (Table 1). Differences between patients reporting and not reporting depression were examined at a randomly selected observation using t-tests and chi-square tests as appropriate (Table 2). Kaplan-Meier survival estimates were used to determine the annual rate of self-reported depression, after excluding patients who reported depression on entry (Figure 1). The Wolfe and Michaud 669 Table 2. Demographic and clinical characteristics of depressed and nondepressed patients with RA* Categories and variables Demographic Age, years Sex, % male Education, years Married, % White, % RA duration, median years Psychological status SF-36 mood scale SF-36 MCS score Antidepressant use, % Income and work status Total income Employed, % Disabled (self-report), % Comorbidity index (range 0–9) RA severity variables Global severity (0–10) Pain (0–10) HAQ DI (0–3) Patient activity score (0–10) Fatigue (0–10) Regional pain score (0–19) Symptom intensity scale (0–10) Treatment, % patients Biologic therapies DMARDs Prednisone NSAIDs Opioids Depressed (n ⴝ 3,364)† Not depressed (n ⴝ 18,767)† 57.2 ⫾ 13.0 15.7 13.0 ⫾ 2.3 68.7 95.6 12.2 61.6 ⫾ 13.5 24.1 13.2 ⫾ 2.3 72.1 95.9 12.6 52.3 ⫾ 19.6 38.2 ⫾ 10.8 35.9 76.4 ⫾ 16.4 52.5 ⫾ 10.1 8.5 $39,258 ⫾ $28,609 26.5 31.4 1.9 ⫾ 1.6 $45,200 ⫾ $29,091 33.0 12.9 1.3 ⫾ 1.3 5.1 ⫾ 2.5 5.6 ⫾ 2.8 1.5 ⫾ 0.7 5.2 ⫾ 2.1 6.5 ⫾ 2.7 8.6 ⫾ 5.8 5.4 ⫾ 2.3 3.5 ⫾ 2.5 3.8 ⫾ 2.8 1.1 ⫾ 0.7 3.6 ⫾ 2.2 4.3 ⫾ 2.9 5.1 ⫾ 4.9 3.4 ⫾ 2.3 40.6 70.6 43.4 60.3 39.6 36.2 76.8 38.1 60.3 21.4 * Values are the mean ⫾ SD unless otherwise indicated. RA ⫽ rheumatoid arthritis; SF-36 ⫽ Short Form 36; MCS ⫽ mental component summary; HAQ ⫽ Health Assessment Questionnaire; DI ⫽ disability index; DMARDs ⫽ disease-modifying antirheumatic drugs; NSAIDs ⫽ nonsteroidal antiinﬂammatory drugs. † By self-report. Differences between groups are signiﬁcant at P ⱕ 0.5, except for NSAID use (P ⫽ 0.989). cumulative hazard of self-reported depression was determined by the Nelson-Aalen estimator. Differences between categories and ordered groups were tested by the log rank test for ordered trends. To further validate the study mea- Figure 1. Variable importance of predictors of depression in rheumatoid arthritis in Random Forest analyses ascertained by the mean decrease in accuracy criterion. SI scale ⫽ Symptom Intensity Scale; RPS ⫽ Regional Pain Scale; PAS ⫽ Patient Activity Scale; HAQ ⫽ Health Assessment Questionnaire disability index. sures, we used time-varying Cox regression analysis to assess the association of self-reported depression, comorbidity, and SI scale with mortality, after adjusting for age and sex. Random Forest analysis was used to determine variable importance and out-of-bag error rates (10,11) using the R statistical package. The out-of-bag error rate is a robust measure of misclassiﬁcation error. Variable importance is described by the mean decrease in accuracy criterion (Figure 2), and represents a ranking of variables in terms of their importance as predictors. Figure 2 shows the 16 best predictors. We also examined the Gini index, but did not display those results in the ﬁgure, as the mean decrease in accuracy is thought to be a better measure. “. . .the Gini Index reﬂects the overall goodness of ﬁt, while the predictive accuracy depends on how well the model actually predicts. The two are related, but they measure different things. Breiman argues that the decrease in predictive accuracy is the more direct, stable and meaningful indicator of variable importance” (personal communication) (26). Classiﬁcation tree analysis used the rpart recursive partitioning R analysis programs, and the 1 stan- 670 Depression Prediction in RA SF-36 MCS score. They tended to be younger, female, with lower household income, less employment, and greater work disability. The ethnicity of the total RA sample was as follows: non-Hispanic white 95.9%, black 1.9%, Asian 0.4%, Native American 0.4%, Hispanic 1.1%, and other 0.24%. With respect to illness characteristics, they had more comorbidity and worse scores on all RA severity scales, and they used more opioids, biologic agents, and corticosteroids but less disease-modifying antirheumatic drugs (Table 2). Figure 2. The cumulative risk (solid line) of self-reported depression in rheumatoid arthritis over 9 years of followup. Estimated cumulative risk at 9 years was 38.3% (95% conﬁdence interval [95% CI; shaded area] 36.6 – 40.1%). dard error rule was applied to truncate trees created by rpart (27). In addition, we examined the effect of predictor variables on self-reported depression by fractional polynomial logistic regression using Stata’s MFP (multivariable fractional polynomial models) procedure (28). We used fractional polynomials because the association of some of the study variables with self-reported depression was nonlinear. To be consistent with the Random Forest analyses, we used the same predictor variables. Because of colinearity with PAS and SI scales, we removed the HAQ and fatigue from the analyses. All data analyses were carried out using Stata, version 10.0 (28), and the R statistical package (29). The level of statistical signiﬁcance was set at 5%, and all tests were 2-tailed. RESULTS Prevalence and incidence of self-reported depression. The cross-sectional prevalence of self-reported depression in 22,131 patients with RA was 15.2% (95% conﬁdence interval [95% CI] 14.7–15.7%) (Table 1). Among the 14,534 patients with RA who were not members of the drug safety registry, the cross-sectional prevalence was 15.4% (95% CI 14.8 –15.9%). These percentages were similar to the 14.5% noted in patients with noninﬂammatory rheumatic disorders, but considerably less than was found in patients with SLE (32.5%) and ﬁbromyalgia (37.9%). Among patients who did not report depression on entry into the study cohort, the annualized incidence rate for self-reported depression was 5.5 (95% CI 5.3–5.7) per 100 patient years. The estimated cumulative risk of self-reported depression at 9 years of followup was 38.3% (95% CI 36.6 – 40.1%) (Figure 1). Characteristics of patients reporting and not reporting depression. Patients with RA reporting and not reporting depression at a randomly selected observation point differed in essentially all characteristics, as shown in the univariate analyses of Table 2. Depressed patients had considerably greater use of antidepressants (35.9% versus 8.5%) and had worse scores on the SF-36 mood scale and Predictors of self-reported depression. To better understand these differences, we performed 2 multivariable analyses using the nonpsychological variables shown in Table 2 as predictor variables. In particular, we were interested in which variables best contributed to identiﬁcation and prediction of self-reported depression. Random Forest analyses of variable importance (Figure 2) show that the SI scale, followed by comorbidity, the RPS, the PAS, the patient’s global, and fatigue, are the most important variables as judged by reduction in the misclassiﬁcation rate. The out-of-bag misclassiﬁcation rate was 30.4%. We also applied recursive partitioning to this set of variables. In these analyses, only the SI scale was required for identiﬁcation of self-reported depression, and no other variables were signiﬁcant in the model by the 1 standard error rule. Values of the SI scale ⬎3.625 best classiﬁed patients as depressed or not depressed. In the second multivariable approach, we used fractional polynomial logistic regression to identify nonpsychological predictor variables of self-reported depression using the variables from Table 2. All variables were significant predictors of self-reported depression except for ethnicity and household income. Interestingly, duration was a signiﬁcant predictor in this model after nonlinear transformation, but not in untransformed form. Graphic examination of the effect of this predictor suggested that the effect of duration on self-reported depression increased for the ﬁrst 5 years of RA, after which increasing duration had no effect. Although it was not a primary importance measure selected for this study, when using the mean decrease in the Gini coefﬁcient as a criterion, duration of RA ranked fourth, after SI scale, age, and PAS. The effect of key predictor variables on the cumulative risk of self-reported depression. The effect of the 2 major predictors on the cumulative hazard of self-reported depression is shown in Figures 3 and 4. Figure 3 shows the cumulative hazard function for 4 categories of comorbidity. Although the risk increases with comorbidity classiﬁcation, there is a striking increase in the risk associated with ⱖ4 comorbid conditions. The best overall predictor of self-reported depression was the SI scale. The SI scale values for the quartiles for the SI scale are Q1 0 –1.5, Q2 1.75–3, Q3 3.25–5, and Q4 5.25–9.75. Increasing quartiles were associated with a striking increase in the cumulative risk of self-reported depression, as shown in Figure 4. In time-varying Cox regression analyses, the hazard ratios (HRs) and P values for the association of mortality and the key study variables for self-reported depression were Wolfe and Michaud Figure 3. The effect of each of 5 comorbidity categories on the cumulative hazard of self-reported depression. The comorbidity scale categories were signiﬁcant by the log rank test for ordered trend, P ⬍ 0.001. HR 1.3, P ⫽ 0.002, for comorbidity categories were HR 1.3, P ⬍ 0.001, and for SI scale quartiles were HR 1.4, P ⬍ 0.001. DISCUSSION There are several important ﬁndings in the results of this study. The ﬁrst is that the cumulative risk of self-reported depression in patients with RA is substantial, reaching almost 40% at 9 years of followup. By contrast, the crosssectional percentage of patients with self-reported depression was only 15.2%. This indicates that many patients with RA will experience self-reported depression over the course of RA, but that self-reported depression will also tend to remit or be intermittent in most instances. This is in accord with the observations of Katz and Yelin, who noted that only 4% of patients had persistent depression (“every year”) (30). The prevalence of depressive symptoms or self-reported depression depends on the method of ascertainment and Figure 4. The effect of each quartile of Symptom Intensity (SI) scale on the cumulative hazard of self-reported depression. The SI scale quartiles were signiﬁcant by the log rank test for ordered trend, P ⬍ 0.001. Q1 ⫽ 0 –1.5; Q2 ⫽ 1.75–3; Q3 ⫽ 3.25–5; Q4 ⫽ 5.25–9.75. 671 the setting. In a sample of 75,858 patients who visited primary care physicians for any reason, the prevalence of clinically signiﬁcant depressive symptoms was found to be 20.9%, but the percentage of patients citing depression as a reason for visit was 1.2% (31). From telephone interviews conducted during 1985 in a national probability sample of 1,232 noninstitutionalized US residents age ⱖ65 years, 9.9% had high depressive symptoms by the Center for Epidemiologic Studies Depression scale (ⱖ16) (32). In a meta-analysis of RA depression studies, the increase of depression in patients compared with healthy controls was found to be moderate (effect size 0.21) (1). The results of the current study were similar to those of Katz and Yelin, who also surveyed a longitudinal sample of patients with RA. Using the Geriatric Depression Scale, they found an annualized prevalence of depressive symptoms to be 15–17% (30). Self-reported depression is not a usual measure in research studies, but it is a reasonable one, representing patients’ perceptions and depressive symptoms, and it is related to conventional measures of mood. As noted in Patients and Methods, self-reported depression and the SF-36 mood and MCS scales had ROC curve association values of 0.826 and 0.823, respectively, and we found that it was associated with mortality. It should be clear, however, that self-reported depression is a measure of depressive symptoms, not of major depression. It should also be noted that our self-reported depression question might not identify depressed patients who are being successfully treated for depression and now considered themselves not depressed. Most of the literature about single-item depression questionnaires reviews them with respect to major depression. In 197 terminally ill cancer patients, a single-item questionnaire correctly identiﬁed all depressed patients (33). A short, 2-question questionnaire had a very high sensitivity and speciﬁcity in patients in multiple sclerosis (34), and a Veterans Administration single-item screening questionnaire was 88% speciﬁc and 78% sensitive (35). However, in another study, the Yale single-item screening questionnaire had a sensitivity of 65.3% and a speciﬁcity of 87.3% (36) compared with the Beck Depression Index cutoff (37). Clearly, the cutoff level is important; in the current study we allowed the patient to determine the cutoff (present or absent). The major ﬁnding of this study, however, concerns selfreported depression predictors. The most important predictor was the SI scale (Figure 2). The SI scale is a linear combination of the fatigue VAS and the RPS, each of which separately was an important predictor of self-reported depression. Widespread pain, the extent of which is measured by the RPS, is increasingly recognized as a central rheumatic disease variable. An entirely subjective variable and a key ﬁnding in ﬁbromyalgia, we also found that it was associated with mortality. The SI scale is associated with severe RA, and with physical and mental distress (24,38). Of particular importance was the result of the recursive partitioning analyses, which suggested that once the results of the SI scale were known, other variables did not increase the ability to correctly classify patients as depressed or not depressed. However, the result was opti- 672 mized for best overall classiﬁcation, but not primarily for self-reported depression sensitivity. In clinical practice, other variables will also be of interest, particularly comorbidity. Comorbidity was the second most important predictor of self-reported depression. It, too, was associated with mortality. Next in rank were the RPS, PAS, and patient global and fatigue scales (Figure 2). As noted above, the duration of RA was the fourth most important variable after SI scale, age, and the PAS when the Gini index method of assessing importance was used. However, the importance indicated by this method appeared to be determined by the association of self-reported depression and duration only within the ﬁrst 5 years. Therefore it appears that the default method is more clinically valid and meaningful (Figure 1). This is in accord with Breiman’s comments favoring the reduction in accuracy method (27). Although many factors contribute to self-reported depression, the most important factors are severity measures, as perceived and reported by the patients, followed by comorbidity. Further insight into the effect of RA severity comes from the predictive value of the use of opioids, disabled status, and reduced income, which is often the result of RA (39). The component played by demography can be seen in the contribution of younger age, reduced income, and marital status; however, ethnicity and education were not important predictors in these analyses. Ethnicity has been shown to be important in other studies, in which depression was increased among Hispanics and blacks, but decreased in Asians (40). However, there were not enough minority subjects in the current study to adequately investigate ethnicity. Symptoms may also contribute to self-reported depression indirectly. Pain and fatigue may contribute to depression beyond their physical/subjective manifestations. For example, they could have disruptive effects on social relationships and life activities that might, in turn, contribute to depression. However, we were not able to evaluate this in the current study. In speaking of prediction, we note that there are 2 types, prediction with respect to causal factors, and prediction with respect to case identiﬁcation. The model we have presented here is a clinical model that is concerned with case identiﬁcation, not causality. Not only are we not able to measure traumatic experiences, genetic factors, temperament, interpersonal relations (6), family vulnerability (7), and HPA axis dysregulation (8), but we cannot always clearly distinguish cause and effect in the variables included in our model. Depression may increase pain and fatigue, just as pain and fatigue may increase the likelihood of depression (4,41). In some cases, self-reported depression may have started before covariate assessment. However, there are still useful lessons here for the clinician. Pain extent (widespread pain) and fatigue, particularly in combination as in the SI scale, are very important clinical variables that should be assessed in the clinical interview. Although most medical records contain information on comorbidity, actual physician familiarity with the extent of an individual patient’s comorbidities will provide important insights into the patient’s global health and associated perceptions. Depression Prediction in RA AUTHOR CONTRIBUTIONS Dr. Wolfe had full access to all of the data in the study and takes responsibility for the integrity of the data and the accuracy of the data analysis. Study design. Wolfe, Michaud. Acquisition of data. Wolfe, Michaud. Analysis and interpretation of data. Wolfe, Michaud. Manuscript preparation. Wolfe. Statistical analysis. Wolfe. REFERENCES 1. Dickens C, McGowan L, Clark-Carter D, Creed F. Depression in rheumatoid arthritis: a systematic review of the literature with meta-analysis. Psychosom Med 2002;64:52– 60. 2. Chapman DP, Perry GS, Strine TW. The vital link between chronic disease and depressive disorders. Prev Chronic Dis 2005;2:A14. 3. Noel PH, Williams JW Jr, Unutzer J, Worchel J, Lee S, Cornell J, et al. Depression and comorbid illness in elderly primary care patients: impact on multiple domains of health status and well-being. Ann Fam Med 2004;2:555– 62. 4. Fishbain DA, Cutler R, Rosomoff HL, Rosomoff RS. Chronic pain-associated depression: antecedent or consequence of chronic pain? A review. Clin J Pain 1997;13:116 –37. 5. Evers AW, Kraaimaat FW, Geenen R, Bijlsma JW. Determinants of psychological distress and its course in the ﬁrst year after diagnosis in rheumatoid arthritis patients. J Behav Med 1997;20:489 –504. 6. Kendler KS, Kessler RC, Neale MC, Heath AC, Eaves LJ. The prediction of major depression in women: toward an integrated etiologic model. Am J Psychiatry 1993;150:1139 – 48. 7. Kendler KS, Gardner CO, Prescott CA. Clinical characteristics of major depression that predict risk of depression in relatives. Arch Gen Psychiatry 1999;56:322–7. 8. Ehlert U, Gaab J, Heinrichs M. Psychoneuroendocrinological contributions to the etiology of depression, posttraumatic stress disorder, and stress-related bodily disorders: the role of the hypothalamus-pituitary-adrenal axis. Biol Psychol 2001; 57:141–52. 9. Royston P, Sauerbrei W. Multivariable model-building: a pragmatic approach to regression analysis based on fractional polynomials for modeling continuous variables. Hoboken (NJ): John Wiley & Sons; 2008. 10. Breiman L. Statistical modeling: the two cultures. Stat Sci 2001;16:199 –231. 11. Breiman L. Random forests. Mach Learn 2001;45:5–32. 12. Wolfe F, Michaud K. A brief introduction to the National Data Bank for Rheumatic Diseases. Clin Exp Rheumatol 2005;23: S168 –71. 13. Wolfe F, Michaud K. Biologic treatment of rheumatoid arthritis and the risk of malignancy: analyses from a large US observational study. Arthritis Rheum 2007;56:2886 –95. 14. Nadareishvili Z, Michaud K, Hallenbeck JM, Wolfe F. Cardiovascular, rheumatologic, and pharmacologic predictors of stroke in patients with rheumatoid arthritis: a nested, case– control study. Arthritis Rheum 2008;59:1090 – 6. 15. Wolfe F, Michaud K. Prevalence, risk, and risk factors for oral and ocular dryness with particular emphasis on rheumatoid arthritis. J Rheumatol 2008;35:1023–30. 16. Wolfe F, Allaire S, Michaud K. The prevalence and incidence of work disability in rheumatoid arthritis, and the effect of anti-tumor necrosis factor on work disability. J Rheumatol 2007;34:2211–7. 17. Stewart AL, Hays RD, Ware JE. The MOS short-form general health survey: reliability and validity in a patient population. Med Care 1988;26:724 –35. 18. Michaud K, Wolfe F. The development of a rheumatic disease research comorbidity index for use in outpatients patients with RA, OA, SLE and ﬁbromyalgia (FMS) [abstract]. Arthritis Rheum 2007;56 Suppl:S596. Wolfe and Michaud 19. Michaud K, Wolfe F. Comorbidities in rheumatoid arthritis. Best Pract Res Clin Rheumatol 2007;21:885–906. 20. Fries JF, Spitz PW, Kraines RG, Holman HR. Measurement of patient outcome in arthritis. Arthritis Rheum 1980;23:137– 45. 21. Wolfe F. A reappraisal of HAQ disability in rheumatoid arthritis. Arthritis Rheum 2000;43:2751– 61. 22. Wolfe F, Michaud K, Pincus T. A composite disease activity scale for clinical practice, observational studies and clinical trials: the Patient Activity Scale (PAS/PAS-II). J Rheumatol 2005;32:2410 –5. 23. Wolfe F. Pain extent and diagnosis: development and validation of the regional pain scale in 12,799 patients with rheumatic disease. J Rheumatol 2003;30:369 –78. 24. Wolfe F, Rasker JJ. The Symptom Intensity Scale, ﬁbromyalgia, and the meaning of ﬁbromyalgia-like symptoms. J Rheumatol 2006;33:2291–9. 25. Doody MM, Hayes HM, Bilgrad R. Comparability of national death index plus and standard procedures for determining causes of death in epidemiologic studies. Ann Epidemiol 2001;11:46 –50. 26. Berk RA. Data mining within a regression framework. In: Maimon O, Rokach L, editors. Data mining and knowledge discovery handbook. New York: Springer; 2005. p. 231–55. 27. Atkinson EJ, Therneau TM. An introduction to recursive partitioning using the RPART routines. Rochester (NY): Mayo Foundation; 2000. 28. Stata Corporation. Stata statistical software: release 10.0. College Station (TX): Stata Corporation; 2007. 29. The R Development Core Team. R: a language and environment for statistical computing. Version 2.7.1. Vienna: R Foundation for Statistical Computing; 2008. 30. Katz PP, Yelin EH. Prevalence and correlates of depressive symptoms among persons with rheumatoid arthritis. J Rheumatol 1993;20:790 – 6. 673 31. Zung WW, Broadhead WE, Roth ME. Prevalence of depressive symptoms in primary care. J Fam Pract 1993;37:337– 44. 32. Stallones L, Marx MB, Garrity TF. Prevalence and correlates of depressive symptoms among older U.S. adults. Am J Prev Med 1990;6:295–303. 33. Chochinov HM, Wilson KG, Enns M, Lander S. “Are you depressed?” Screening for depression in the terminally ill. Am J Psychiatry 1997;154:674 – 6. 34. Mohr DC, Hart SL, Julian L, Tasch ES. Screening for depression among patients with multiple sclerosis: two questions may be enough. Mult Scler 2007;13:215–9. 35. Corson K, Gerrity MS, Dobscha SK. Screening for depression and suicidality in a VA primary care setting: 2 items are better than 1 item. Am J Manag Care 2004;10:839 – 45. 36. Avasarala JR, Cross AH, Trinkaus K. Comparative assessment of Yale Single Question and Beck Depression Inventory Scale in screening for depression in multiple sclerosis. Mult Scler 2003;9:307–10. 37. Beck AT. Depression inventory. Philadelphia: Center for Cognitive Therapy; 1978. 38. Wolfe F, Michaud K. Severe rheumatoid arthritis (RA), worse outcomes, comorbid illness, and sociodemographic disadvantage characterize RA patients with ﬁbromyalgia. J Rheumatol 2004;31:695–700. 39. Wolfe F, Michaud K, Choi HK, Williams R. Household income and earnings losses among 6,396 persons with rheumatoid arthritis. J Rheumatol 2005;32:1875– 83. 40. Jackson-Triche ME, Greer Sullivan J, Wells KB, Rogers W, Camp P, Mazel R. Depression and health-related quality of life in ethnic minorities seeking care in general medical settings. J Affect Disord 2000;58:89 –97. 41. Fiﬁeld J, Tennen H, Reisine S, McQuillan J. Depression and the long-term risk of pain, fatigue, and disability in patients with rheumatoid arthritis. Arthritis Rheum 1998;41:1851–7.