Low back pain is common among adults of all ages, but

Transcription

1 SPINE Volume 40, Number 17, pp , Wolters Kluwer Health, Inc. All rights reserved. HEALTH SERVICES RESEARCH Trajectories of Symptoms and Function in Older Adults With Low Back Disorders Richard A. Deyo, MD, MPH, * Matthew Bryan, PhD, Bryan A. Comstock, MS, Judith A. Turner, PhD, Patrick Heagerty, PhD, Janna Friedly, MD, Andrew L. Avins, MD, MPH, Srdjan S. Nedeljkovic, MD, ** David R. Nerenz, PhD, and Jeffrey G. Jarvik, MD, MPH Study Design. Prospective cohort study. Objective. To determine whether there are distinct trajectories of back pain and function among older adults and to identify characteristics that distinguish among patients with substantially different prognoses. Summary of Background Data. Although the differential diagnosis and course of low back pain among older adults may differ from middle-aged adults, there is little evidence. Better understanding variability in recovery among older adults may help target patients for more intensive clinical interventions, plan resource use, and design clinical studies of more homogeneous patient groups. Methods. Patients aged 65 years or older with a new episode of care for back pain were recruited at 3 geographically diverse sites. Patients completed pain intensity and Roland-Morris Disability measures at baseline and 3, 6, and 12 months later. We used latent class analysis to identify distinct trajectories of pain and function From the * Departments of Family Medicine, Medicine, and Public Health and Preventive Medicine, and the Oregon Institute of Occupational Health Sciences, Oregon Health & Science University, Portland, OR ; Department of Biostatistics and Epidemiology, University of Pennsylvania, Philadelphia, PA; Departments of ; Biostatistics ; Psychiatry and Behavioral Sciences, and ; Rehabilitation Medicine, Comparative Effectiveness, Cost and Outcomes Research Center, University of Washington, Seattle, WA ; Division of Research, Northern California Kaiser-Permanente, Oakland, CA ; ** Department of Anesthesiology, Perioperative and Pain Medicine, Brigham and Women's Hospital and Spine Unit, Harvard Vanguard Medical Associates, Boston, MA ; Neuroscience Institute, Henry Ford Hospital, Detroit, MI; and Departments of Radiology, Neurological Surgery, and Health Services, Comparative Effectiveness, Cost and Outcomes Research Center, University of Washington, Seattle, WA. Acknowledgment date: August 14, First revision date: February 23, Second revision date: April 9, Acceptance date: April 26, The manuscript submitted does not contain information about medical device(s)/drug(s). The Agency for Healthcare Research and Quality (grant no. R01 HS019222) funds were received to support this work. Relevant financial activities outside the submitted work: board membership, consultancy, employment, patents, stocks, employment, grants, royalties, payment for lectures, travel/accommodations/meeting expenses. Address correspondence and reprint requests to Richard A. Deyo MD, MPH, Oregon Health & Science University, Mail Code FM, 3181 SW Sam Jackson Park Rd, Portland, OR 97239; deyor@ohsu.edu DOI: /BRS and then logistic regression to identify predictors of membership in the improving trajectories. Results. There were 3929 participants who completed outcome measures at every follow-up interval. Latent class analysis identified subgroups with low, intermediate, or high pain or disability scores who remained relatively stable over time. However, small subgroups showed dramatic improvement from baseline to 1 year (17% with major improvement in Roland score, pain intensity, or both). Shorter pain duration, higher patient confidence in improvement, and fewer comorbid conditions at baseline were each associated independently with favorable prognosis. Conclusion. Although most patients remained relatively stable over a year, latent class analysis identified small groups with major improvement in pain, function, or both. This technique may, therefore, be useful for studying back pain prognosis. Our results should help assemble more prognostically homogeneous groups for research, and the technique may help identify subgroups of patients with uniquely successful responses to investigational interventions. Key words: prognosis, latent class analysis, older adults, lumbar stenosis, Roland-Morris Disability Questionnaire, variability in recovery, cohort study, back pain prognosis, functional recovery, registry, BOLD registry. Level of Evidence: 3 Spine 2015;40: Low back pain is common among adults of all ages, but most research has focused on working-aged adults. 1 The differential diagnosis and course of back pain may be different among older adults than among middle-aged adults, yet little is known about the prognosis in older adults. Better understanding variability in recovery among older adults seeking back pain care might identify relatively homogeneous prognostic subgroups. Such information may help target some patients for earlier or more intensive clinical interventions. 2 At a population level, it may help plan resource use. For researchers, it could help in designing clinical studies by identifying more prognostically homogeneous patient groups. The Back pain Outcomes using Longitudinal Data (BOLD) project is a prospective cohort registry examining the course of low back pain in older adults. 3 Goals of the present study were to use latent class analysis in this cohort to determine September 2015

2 whether there are distinct trajectories of back pain and function and to identify features that distinguish among patients with substantially different prognoses. MATERIALS AND METHODS Design of the BOLD project is described in detail elsewhere 3 and summarized here. Activities were approved by institutional review boards at each participating site. Setting Patients were recruited from integrated health care systems: Kaiser Permanente, Northern California; Henry Ford Health System, Detroit; and Harvard Vanguard Medical Associates in the Boston area. The University of Washington was the Data Coordinating Center. Patient Recruitment Patients aged 65 years and older were identified from electronic medical records by having a visit for back pain in primary care or urgent/emergency care. Patients with any of 57 International Classification of Diseases, Ninth Revision, Clinical Modification ( ICD-9-CM ) 3 codes for back pain, degenerative spinal disorders, or vertebral fractures were included. Patients with a visit for back pain within the previous 6 months were excluded, so those enrolled were having a new episode of care for back pain. This could include patients with recurrent pain, chronic pain, or those with recent care only outside the participating health plan ( e.g., chiropractic care). Other exclusions were previous lumbar surgery, developmental deformities, inflammatory spondylopathy, spinal malignancy or infection, history of cancer, HIV infection, no telephone, non English speaking, plans to leave the health care system within 12 months, or severe mental impairment. Follow-up Enrollment Eligibility Screening 13,376 patients identified for screening through ICD-9 code 2,037 Ineligible (>3 weeks from index visit) 505 Not reached 3,499 Declined further screening 7,335 assessed for eligibility 1,953 Patients ineligible for one or more reasons: 666 Incomplete screening 403 Prior visit for back pain in last 6 months 399 Had a major medical comorbidity 372 Did not speak English 251 Had prior back surgery 211 Recent visit not back-related 62 Failed cognitive screening assessment 61 Other 5,382 Patients were eligible 143 Patients were not included 74 did not consent for further participation 69 did not complete the baseline questionnaire 5,239 Patients included in BOLD Registry 257 Patients exited study prior to 1-year follow-up 35 deaths 222 withdrawals 1,053 Patients had incomplete Roland score or pain rating scale through 1-year follow-up 3,929 Patients with complete follow-up for Roland score and pain rating scale through 1-year follow-up Data Collection and Measures Data were obtained from questionnaires and electronic medical records. Baseline interviews were conducted within 3 weeks of the index visit. Follow-up was by mailed questionnaire or telephone interview 3, 6, and 12 months after baseline interviews. Baseline data included demographics and duration of current back pain. Participants rated confidence that their back or leg pain would be much better in 3 months on a scale of 0 ( not at all confident ) to 10 ( extremely confident ). Comorbidity was assessed 4 using diagnosis codes from all outpatient visits and hospitalizations in the year prior to the index visit. Every assessment included the Roland-Morris Disability Questionnaire (Roland) 5 as a measure of function and ratings of average back pain intensity in the past week on 0 to 10 scales. Baseline ICD-9-CM diagnosis codes and some demographics were obtained from electronic medical records. Analysis ICD-9-CM diagnoses were grouped into 4 crude categories: axial back pain, back and leg pain or herniated disc, lumbar stenosis, and other (see Supplemental Digital Content Appendix 1, available at: ). Figure 1. Flow diagram of patients recruitment and exclusion. ICD-9 indicates International Classification of Diseases, Ninth Revision ; BOLD, Back pain Outcomes using Longitudinal Data. Although patients with stenosis may have both back and leg pain, we separated them from others with back and leg pain. The other category was heterogeneous, ranging from sprains to fractures, with small numbers of each diagnosis. They are included for completeness. We conducted latent class analyses 6 separately for pain and function. This technique assumes that there are distinct trajectories of symptom course and that individuals can be grouped into a small number of clusters representing each trajectory (the latent classes ). Each individual belongs to 1 cluster on the basis of the highest posterior probability of belonging to a particular cluster. 6, 7 In our case, latent class modeling seeks the smallest number of clusters that accounts for associations between pain and function levels measured at baseline, 3 months, 6 months, and 1 year. Therapy may affect outcomes, so we identified patients who underwent lumbar surgery. Numbers were too small Spine

3 TABLE 1. Baseline Patient Characteristics (N = 3929) Patient Characteristics n (%) Site Henry Ford 698 (18) Kaiser 2396 (61) Harvard 835 (21) Age, mean (SD), yr 73.6 (6.7) Sex (male) 1381 (35) Race* Caucasian 2960 (75) African American 569 (14) Asian 149 (4) Native American 18 (0) Pacific Islander 10 (0) Other 188 (5) Ethnicity (Hispanic)* 196 (5) Smoking status (current smoker)* 228 (6) Married* 2364 (60) Education* Less than high school 195 (5) High school, trade school, or some college 1979 (50) College graduate 1023 (26) Graduate degree 722 (18) Baseline diagnosis Axial back pain 2646 (67) Back and leg pain 834 (21) Lumbar stenosis 231 (6) Other 218 (6) Pain duration* < 1 mo 1347 (34) 1 3 mo 764 (19) 3 6 mo 253 (6) 6 12 mo 245 (6) 1 5 yr 556 (14) > 5 yr 759 (19) Expected improvement* 0 (no confidence) 793 (20) (15) (16) (Continued) TABLE 1. (Continued ) Patient Characteristics n (%) (28) 10 (Extremely confident) 831 (21) Quan Comorbidity Score* (61) (26) > (12) Surgical procedures postindex visit Within 3 mo 9 (0) Within 11 mo 34 (1) *Variables with missing values (maximum n = 82, 2.1% of respondents); percentages reported among those who were observed. to examine statistically, but we present the numbers in each latent class group. We did not examine nonoperative treatments given their variety ( e.g., prescription and over-thecounter medications/dosages, physical therapy of various types, complementary and alternative treatments, mental health interventions, injections); the almost infinite variations in combinations, sequences, and timing in relationship to our assessments; and evidence that most have small average treatment effects ( e.g., in Cochrane systematic reviews). We performed latent class analysis for participants who completed both the Roland and back pain measures at each assessment. Latent class analysis was generated using the traj command in Stata 13 statistical software, 8 allowing for a 3 polynomial to model the longitudinal trajectory in each latent group. We adjusted the analysis for study sites, given differences in demographics and management approaches. 9 Latent class models were fitted iteratively, starting with 1 cluster and adding additional clusters until the model obtained the largest number of clusters for which the number of participants in each group was at least 5% of the sample. We plotted mean trajectories for each latent class for each outcome and a point-wise 95% confidence interval about group means. Each cluster was classified as stable over time or as showing major recovery (moderate to high disability or pain at baseline decreasing to low disability or pain by 1-yr followup). All clusters could be classified into 1 of these 2 groups (none showed major worsening). Baseline characteristics of patients in each latent class were compared. We then used multivariate logistic regression with the outcome being major improvement or not to identify baseline factors independently associated with a major recovery trajectory (separately for physical function and pain). Participants assigned by latent class analysis to the major recovery groups were compared with individuals assigned to a stable group with moderate or high scores (on disability or pain). Latent class groups with low baseline September 2015

4 TABLE 2. Mean Pain and Roland Scores at Each Timepoint Measure and Time of Assessment Overall Sample (n = 3929), Mean (SD) Axial Back Pain (n = 2646), Mean (SD) Back and Leg Pain or Herniated Disc (n = 834), Mean (SD) Lumbar Stenosis (n = 231), Mean (SD) Other Diagnoses (n = 218), Mean (SD) Roland score (0 24) Baseline 9.26 (0.10) 9.3 (0.1) 9.5 (0.2) 9.8 (0.4) 7.7 (0.4) 3 mo 8.94 (0.10) 8.7 (0.1) 9.3 (0.2) 10.8 (0.4) 8.2 (0.4) 6 mo 8.66 (0.11) 8.6 (0.1) 8.6 (0.2) 10.2 (0.4) 7.8 (0.5) 12 mo 8.33 (0.11) 8.2 (0.1) 8.2 (0.2) 10.0 (0.5) 8.3 (0.5) P * < 0.01 < 0.01 < Back pain intensity rating (0 10) Baseline 4.99 (0.04) 5.1 (0.1) 4.8 (0.1) 4.6 (0.2) 4.5 (0.2) 3 mo 3.69 (0.04) 3.7 (0.1) 3.6 (0.1) 4.2 (0.2) 3.6 (0.2) 6 mo 3.64 (0.04) 3.6 (0.1) 3.5 (0.1) 4.3 (0.2) 3.6 (0.2) 12 mo 3.57 (0.04) 3.6 (0.1) 3.4 (0.1) 4.1 (0.2) 3.9 (0.2) P * < 0.01 < 0.01 < < 0.01 *Wald test for a linear trend in the means over time. Roland or pain scores were removed from these analyses because of little room for improvement. Regression models included demographic characteristics and baseline diagnosis, Roland score, pain rating, pain duration, confidence in pain improvement, and Quan comorbidity score, with additional adjustment for site-level differences. RESULTS Figure 2. Mean Roland score trajectories within latent class groups. Groups were identified on the basis of a latent class analysis that adjusted for health care system site. The gray region around each trajectory represents the point-wise 95% confidence interval at each observed timepoint. LM indicates low moderate; MH, moderate high; FR, functional recovery; CI, confidence interval. Patient Characteristics BOLD enrolled 5239 patients. By 1-year follow-up, 35 patients had died and 222 had withdrawn, leaving 4982 eligible for follow-up. This analysis is based on 3929 participants (79% of those eligible) who completed outcome measures at each follow-up timepoint ( Figure 1 ). Mean age was 73.6 years and 65% were female. Although all initiated a new episode of care for back pain, duration of current pain at baseline varied; about half reported duration less than 3 months ( Table 1 ). Few patients underwent surgery during follow-up (9 patients within 3 mo of index visit; 34 patients within 11 mo). Changes in Pain and Function Over the Year After the Index Visit On average, participants demonstrated modest improvement in Roland scores over 1 year: from a mean of 9.26 to 8.33 ( P < 0.01; Table 2 ). Pain ratings improved from a mean of 4.99 to 3.57 ( P < 0.01; Table 2 ). On average, patients with stenosis or other diagnosis showed no improvement in physical function, whereas patients with axial pain or back and leg pain/herniated disc diagnoses demonstrated small improvements ( Table 2 ). Regarding pain, patients with axial or back and leg pain/herniated discs improved more on average, but improvement was modest in all groups. Experts suggest 30% improvement in pain intensity or Roland scores as a clinically important difference Spine

5 TABLE 3. Baseline Characteristics of Patients in Each Latent Class Cluster of Roland Score Trajectories, and Comparison of -High/High Disability Group (Combined) Versus Functional Recovery Group (Excluding Patients in Low and Low Disability Groups) Low Low High High Functional Recovery P * n (% of total) 727 (18.5) 984 (25.0) 1253 (31.9) 725 (18.5) 240 (6.1) Site (overall % of sample), % NA Henry Ford (18%) Kaiser (61%) Harvard (21%) Age, mean (SD), yr 72.5 (6.2) 73.2 (6.5) 74.4 (6.8) 74.5 (7.0) 72.2 (6.5) < 0.01 Sex, % male < 0.01 Race, % 0.04 Caucasian African American Asian Pacific Islander American Indian Other or multiple Ethnicity, % Hispanic Current smoker, % Married, % Education, % < 0.01 Less than High school High school, trade school, or some college College graduate Graduate degree Baseline diagnosis, % < 0.01 Axial back pain Back and leg pain Lumbar stenosis Other Roland score, mean (SD) 2.9 (3.3) 4.7 (3.2) 11.1 (4.3) 16.8 (3.5) 15.1 (3.3) < 0.01 Pain rating, mean (SD) 3.3 (2.6) 3.9 (2.5) 5.5 (2.5) 6.8 (2.3) 6.1 (2.5) 0.65 Pain duration, % < 0.01 < 1 mo mo mo mo yr > 5 yr (Continued) September 2015

6 TABLE 3. (Continued ) Low Low High High Functional Recovery P * Confidence in improvement, % < (not at all confident) (extremely confident) Quan Comorbidity Score, % < > Surgical procedures postindex visit, N Within 3 mo NA Within 11 mo NA * P value for a Wald test for differences in the odds of being in the functional recovery group compared with the moderately high and high-stable disability groups on the basis of a logistic regression with the given variable as predictor adjusting for site-level differences. Site-level differences were not of interest as a predictor for functional recovery. Because of sparse data, the Asian, Pacific Islander, Native American, and Other or multiple categories were combined for the logistic regression model. The logistic regression was not performed because of insufficient number of surgical procedures. NA indicates not applicable. We, therefore, examined this metric at 1 year in addition to mean score changes. Overall, 20% of patients reported no back-related functional limitations ( i.e., Roland score of 0). Forty-one percent showed 30% or greater improvement in Roland scores, 34% had no meaningful change, and 25% were at least 30% worse. For pain rating, 18% reported no pain, 49% had 30% or greater improvement, 33% had no meaningful change, and 18% worsened. Latent Class Analysis of Functional Improvement Latent class analysis of Roland scores, adjusting for study site, revealed 5 distinct trajectories ( Figure 2 ). Four groups had relatively constant functional limitations over time at different levels of severity. The fifth group had moderate or high baseline functional limitations but improved substantially over time. Those in the functional recovery group began with a mean Roland score of 15.1 and improved to a mean of 2.9. This 81% average improvement far exceeds estimates of the 11, 13 minimal clinically important difference. Characteristics of Patients in Roland Trajectory Clusters There were only small differences in mean age across clusters ( Table 3 ). However, the group with high and stable functional limitations had the largest proportion of females, African Americans, Hispanics, smokers, and patients without a college degree. This group had the longest duration of pain, the highest proportion with stenosis diagnoses, the lowest confidence in improvement, and the highest level of comorbidity ( Table 3 ). The group with low and stable functional limitations had the largest proportion of males, Caucasian patients, college or graduate degrees, high confidence in improvement, and fewer comorbid conditions. Like the functional recovery group, a high proportion had axial back pain. The stable middle groups (low moderate and moderate high disability) were intermediate on most measures. Across the 4 stable trajectories, the proportions with pain for less than 1 month declined consistently with worsening Roland score and the proportions with pain for more than 5 years increased consistently. Confidence in improvement followed the same pattern. The functional recovery group had the smallest proportion of stenosis diagnoses (1.2%), the shortest pain duration (77.1% < 3 mo), and high confidence in improvement (68.7% > 5). Comparing the functional recovery group with the moderate or high stable functional limitation groups, all these differences were statistically significant ( P < 0.01). Baseline Characteristics Associated With Major Functional Improvement The multivariate logistic regression model compared patients with or without major recovery among those with moderate to high baseline disability. Shorter pain duration and greater Spine

7 TABLE 4. Multivariable Logistic Regression Models Predicting Functional Recovery and Pain Recovery Group Membership Covariates Prediction of Functional Recovery Group Prediction of Pain Recovery Groups OR (95% CI) P OR (95% CI) P Age (per 5-yr increase) 0.82 ( ) < ( ) 0.49 Male Sex (reference = female) 1.52 ( ) ( ) < 0.01 Race (reference = Caucasian) African American 0.50 ( ) 0.57 ( ) Other or multiple 0.76 ( ) 1.01 ( ) Ethnicity (reference = non-hispanic) 1.02 ( ) ( ) 0.90 Smoking status (reference = current nonsmoker) 0.45 ( ) ( ) 0.27 Marital status (reference = all nonmarried status categories) 1.09 ( ) ( ) 0.24 Education (reference = Less than high school) High school, trade school, or some college 0.87 ( ) 0.84 ( ) College graduate 1.10 ( ) 1.03 ( ) Graduate degree 1.52 ( ) 1.26 ( ) Baseline diagnosis (reference = Axial back pain) Back and leg pain 0.91 ( ) 0.95 ( ) Lumbar stenosis 0.30 ( ) 0.78 ( ) Other 0.63 ( ) 0.76 ( ) Baseline Roland score (per 1-point increase, 0 24 scale) 1.13 ( ) < ( ) < 0.01 Baseline pain rating (per 1-point increase, 0 10 scale) 0.93 ( ) ( ) < 0.01 Pain duration (reference = < 1 mo) < 0.01 < mo 0.67 ( ) 0.67 ( ) 3 6 mo 0.45 ( ) 0.47 ( ) 6 12 mo 0.19 ( ) 0.31 ( ) 1 5 yr 0.28 ( ) 0.23 ( ) > 5 yr 0.32 ( ) 0.19 ( ) Confidence in improvement (reference = 0, no confidence; 0 10 scale) < 0.01 < ( ) 1.22 ( ) ( ) 1.77 ( ) ( ) 2.17 ( ) 10 (extremely confident) 6.26 ( ) 3.98 ( ) Quan Comorbidity Score (reference = 0, 0 10 scale) ( ) 0.85 ( ) > ( ) 0.54 ( ) The logistic regression models were run on a subset of patients defined by the latent class analysis. The functional recovery model compared patients who were assigned to the moderate-high stable disability or high stable disability groups with the functional recovery group, with an indicator for those in the functional recovery group ( vs. the other 2 groups combined) used for the outcome. We excluded patients in the low stable disability and the low-moderate stable disability groups. The pain recovery model compared patients assigned to the moderate-high stable pain, the high stable pain groups (combined) with the severe pain recovery, and the moderate pain recovery groups (combined) using an indicator for those in the stable groups versus the recovery groups for the outcome. For each outcome (functional recovery, pain recovery), all listed variables were included in 1 multivariable logistic regression with additional adjustment for study recruitment site (results not shown). OR indicates odds ratio; CI, confidence interval September 2015

8 Figure 3. Mean back pain intensity trajectories within latent class groups. Groups were identified on the basis of a latent class analysis that adjusted for health care system site. The gray region around each trajectory represents the point-wise 95% confidence interval at each observed timepoint. LM indicates low moderate; MH, moderate high; SPR, severe pain recovery; MPR, moderate pain recovery. confidence in improvement remained among the strongest predictors of major functional recovery ( Table 4 ). Although likely related, these measures retained independent predictive power. Greater comorbidity was strongly associated with lower odds of functional improvement. Baseline diagnosis was not significantly associated with major recovery after adjusting for the other variables in the model. Being male, Caucasian, and nonsmoking were also significantly associated with major functional recovery. Increasing age was associated with lower likelihood of functional recovery ( P < 0.01), though the effect was small. Latent Class Analysis of Pain Improvement For back pain ratings, latent class analysis produced 6 distinct trajectories ( Figure 3 ). Four were basically stable lines showing minimal improvement but with varied starting points. However, 2 trajectories (labeled moderate pain recovery and severe pain recovery ) showed substantial improvements. These groups each improved approximately 6 points on average on the 11-point pain rating. For the severe pain recovery group, this represented a 74% improvement, far more than a minimal clinically important change Baseline characteristics of the various stable pain and substantial pain improvement groups were similar to those of patients in the comparable Roland groups ( Table 5 ). For multivariate logistic regression, analysis was restricted to participants who began with moderately high to high pain, contrasting those in a recovery group to the stable groups. As in the functional recovery group, shorter pain duration, greater confidence in improvement, less comorbidity, male sex, and Caucasian race were predictors of pain recovery versus stable scores ( Table 4 ); age and smoking were not. Among the total sample, 654 patients (17%) were in either the functional improvement group or the pain improvement groups. Among the 240 patients in the functional improvement group, 120 were also in a pain improvement group. These 120 comprised 22% of those in a pain improvement group, 18% of those in either the functional improvement group or a pain improvement group, and 3% of our overall sample. Thus, only 3% of patients showed major improvements in both pain and function. DISCUSSION Examining 1-year outcomes of the BOLD cohort leaves an impression of modest change over time. Mean improvement in Roland disability score was just 0.93 points (on a 24-point scale) and mean improvement in pain was just 1.4 points on an 11-point scale. However, latent class analysis revealed small subgroups with dramatic improvements in Roland scores, pain ratings, or both. Patient baseline demographics, pain duration, confidence in recovery, and level of comorbidity were associated with outcomes. Latent class analysis is relatively new in studying back pain. It has the virtue of distinguishing subgroups with substantially different prognoses that may be obscured by average changes alone. Distinguishing groups with major prognostic differences may be important for both clinical and research purposes, and our data emphasize the importance of carefully describing patients in clinical reports. For example, if a study enrolls an enriched group of patients with highly favorable prognoses, it may give a misleadingly favorable impression of treatment effect if there is no comparison group or a poorly selected one. Differences in study conclusions regarding treatment effectiveness may in part reflect enrollment of patients with substantially different prognoses. Our findings suggest characteristics of older patients that seem important in this regard. In addition to baseline pain and function, it may be particularly useful to assess and report patient's sex, race, education, clinical diagnosis, pain duration, smoking status, confidence in improvement, and extent of medical comorbidity. Our predictors of favorable trajectory are similar to some prognostic indicators noted in populations with younger ages and differing durations of pain. Recovery expectation was an important prognostic indicator in studies of patients with new episodes of back pain, 14, 15 as was pain duration. 16 A systematic review found that fear avoidance (not measured here) was an important predictor. 17 Lower education levels 14, 15, 18 and smoking 18 have also previously been associated with worse outcomes. 15 Mood disorders worsened prognosis in some studies 17, 19 but were not studied here. Differences among the clusters may be partially explained by differences in the proportion of participants at each site. Among the stable groups with moderate to high Roland scores, a larger proportion came from Henry Ford and fewer Spine

9 TABLE 5. Baseline Characteristics of Patients in Each Latent Class Cluster of Pain Score Trajectories, and Comparison of -High/High Pain Groups (Combined) Versus Pain Recovery Groups Combined (Excluding Patients in Low or Low Pain Groups) N (% of Total) Low Pain, Low Pain, High Pain, High Pain, Pain Recovery Severe Pain Recovery 244 (6.2) 1200 (30.5) 1426 (36.3) 525 (13.4) 264 (6.7) 270 (6.9) Site (overall % of sample), % NA Henry Ford (18%) Kaiser (61%) Harvard (21%) Age, mean (SD), yr 73.1 (6.3) 73.3 (6.5) 74.0 (6.7) 73.9 (6.8) 74.1 (6.9) 72.9 (6.5) 0.15 Sex, % male < 0.01 Race, % 0.02 Caucasian African American Asian Pacific Islander American Indian Other or multiple Ethnicity, % Hispanic Current smoker, % Married, % < 0.01 Education, % < 0.01 Less than high school High school, trade school, or some college College graduate Graduate degree Baseline diagnosis, % < 0.01 Axial back pain Back and leg pain Lumbar stenosis Other Roland score, mean (SD) 2.9 (4.3) 6.1 (5.1) 10.7 (5.5) 15.4 (4.6) 12.2 (5.9) 6.9 (5.9) < 0.01 Pain rating, mean (SD) 0.7 (0.9) 2.9 (1.8) 5.6 (1.9) 8.1 (1.4) 8.3 (1.3) 5.9 (2.0) < 0.01 Pain duration, % < 0.01 < 1 mo mo mo P * (Continued) September 2015

10 TABLE 5. (Continued ) N (% of Total) Low Pain, Low Pain, High Pain, High Pain, Pain Recovery Severe Pain Recovery 244 (6.2) 1200 (30.5) 1426 (36.3) 525 (13.4) 264 (6.7) 270 (6.9) 6 12 mo yr > 5 yr Confidence in improvement, % < (no confidence) (extremely confident) Quan Comorbidity Score, % < > Surgical procedures postindex visit, N Within 3 mo NA Within 11 mo NA * P value for a Wald test for differences in the odds of being in the functional recovery group compared with the moderately high and high stable disability groups based on a logistic regression with the given variable as predictor adjusting for site-level differences. Site-level differences were not of interest as a predictor for functional recovery. Because of sparse data, the Asian, Pacific Islander, Native American, and Other or multiple categories were combined for the logistic regression model. The logistic regression was not performed because of insufficient number of surgical procedures. NA indicates not applicable. P * from Harvard Vanguard relative to low and low moderate disability groups. Others have begun advocating more rigorous examination of back pain trajectories, given that definitive cures are rare outcomes. Some have used text messaging to obtain frequent (weekly) assessments. 20 Dunn et al 7 used latent class analysis to study back pain among working-aged adults but did not identify clusters with dramatic improvement or deterioration, perhaps due to smaller sample size (n = 342) and only 6 months of follow-up. Our study has some important limitations. We relied on ICD-9 codes for diagnoses in the absence of routine imaging. We did not consider the effects of treatments (for back pain or comorbid conditions), given the small number of operations and complexity of nonsurgical interventions. The group with stenosis may have been more likely to have had previous imaging (hence the diagnostic label). This could be a marker for patients who are doing poorly and could also influence patient behavior. There may be other important unmeasured prognostic variables. We analyzed participants with complete 1-year follow-up. Excluding patients who died, withdrew, or were lost to follow-up potentially introduces sample bias. Latent class analysis is a data-driven approach that identifies longitudinal trajectories that maximize the amount of outcome variation explained; the clusters are not necessarily those most meaningful from a scientific perspective. The recovery groups in our analysis showed average improvements well above the 30% improvements considered as clinically meaningful In our sample, 41% of participants had a 30% improvement in Roland score over the year of follow-up, but only 6% of participants were in the recovery group from the latent class analysis. At 12 months, 18% of patients had no pain and 20% had a Roland score of 0, indicating no functional limitation. Thus, some who were not in our recovery groups may still have had meaningful positive outcomes. In summary, most patients in this cohort of older adults showed only modest changes in function or pain over a year of follow-up. However, latent class analysis identified small groups with major improvement in pain, function, or both. This technique may, therefore, help in studying prognosis for Spine

11 this condition. Our results should help investigators assemble more prognostically homogeneous groups for research, and the method may help identify subgroups with uniquely successful responses to investigational interventions. Key Points We used a registry cohort of 3929 patients with low back pain aged 65 years and older to examine possible trajectories of pain and function over time. Using latent class analysis, we identified 5 distinct trajectories of functional change over time, with 4 relatively stable and 1 showing dramatic improvement. Similarly, there were 6 trajectories of pain intensity, of which 4 were relatively stable and 2 showed dramatic improvement. The favorable prognostic trajectories were associated with baseline diagnoses (fewer patients with lumbar stenosis), shorter duration of pain, higher patient confidence in recovery, fewer comorbid conditions, and certain demographic characteristics. Examining only the average course of pain or function in a group of patients may obscure subgroups with major improvement. Supplemental digital content is available for this article. Direct URL citation appearing in the printed text is provided in the HTML and PDF version of this article on the journal's Web site ( ). References 1. Paeck T, Ferreira ML, Sun C, et al. Are older adults missing from low back pain clinical trials? A systematic review and meta-analysis. Arthritis Care Res 2014 ; 66 : Hill JC, Whitehurst DG, Lewis M, et al. Comparison of stratified primary care management for low back pain with current best practice (STarT Back): a randomised controlled trial. Lancet 2011 ; 378 : Jarvik JG, Comstock BA, Bresnahan BW, et al. Study protocol: the Back Pain Outcomes using Longitudinal Data (BOLD) registry. BMC Musculoskel Disord 2012 ; 13 : Quan H, Sundararajan V, Halfon P, et al. Coding algorithms for defining comorbidities in ICD-9-CM and ICD-10 administrative data. Med Care 2005 ; 43 : Roland M, Morris R. A study of the natural history of back pain. Part 1: development of a reliable and sensitive measure of disability in low back pain. Spine 1983 ; 8 : Garrett ES, Zeger SL. Latent class model diagnosis. Biometrics 2000 ; 56 : Dunn KM, Jordan K, Croft PR. Characterizing the course of low back pain: a latent class analysis. Am J Epidemiol 2006 ; 163 : Jones BL, Nagin DS. Advances in group-based trajectory modeling and an SAS procedure for estimating them. Sociol Method Res 2007 ; 35 : Jarvik JG, Comstock BA, Heagerty PJ, et al. Back pain in seniors: the Back pain Outcomes using Longitudinal Data (BOLD) cohort baseline data. BMC Musculoskelet Disord 2014 ; 15 : Farrar JT, Young JP, Jr LaMoreaux L, et al. Clinical importance of changes in chronic pain intensity measured on an 11-point numerical pain rating scale. Pain 2001 ; 94 : Ostelo RW, Deyo RA, Stratford P, et al. Interpreting change scores for pain and functional status in low back pain: towards international consensus regarding minimal important change. Spine (Phila Pa 1976) 2008 ; 33 : Dworkin RH, Turk DC, Farrar JT, et al. Core outcome measures for chronic pain clinical trials: IMMPACT recommendations. Pain 2005 ; 113 : Patrick DL, Deyo RA, Atlas SJ, et al. Assessing health-related quality of life in patients with sciatica. Spine (Phila Pa 1976) 1995 ; 20 : Turner JA, Shortreed SM, Saunders KW, et al. Optimizing prediction of back pain outcomes. Pain 2013 ; 154 : Turner JA, Franklin G, Fulton-Kehoe D, et al. Worker recovery expectations and fear-avoidance predict work disability in a population-based workers compensation back pain sample. Spine 2006 ; 31 : Mallen CD, Peat G, Thomas E, et al. Prognostic factors for musculoskeletal pain in primary care: a systematic review. Br J Gen Pract 2007 ; 57 : Chou R, Shekelle P. Will this patient develop persistent disabling low back pain? JAMA 2010 ; 303 : Dionne C, Koepsell TD, Von Korff M, et al. Formal education and back-related disability. In search of an explanation. Spine 1995 ; 20 : Edwards RR, Klick B, Buenaver L, et al. Symptoms of distress as prospective predictors of pain-related sciatica treatment outcomes. Pain 2007 ; 130 : Axen I, Leboeuf-Yde C. Trajectories of low back pain. Best Pract Res Clin Rheumatol 2013 ; 27 : September 2015