Racial differences in pain during 1 year among women with metastatic breast cancer

A hazards analysis of interval-censored data


  • Liana D. Castel PhD,

    Corresponding author
    1. Cecil G. Sheps Center for Health Services Research, University of North Carolina at Chapel Hill, Chapel Hill, North Carolina
    • Center for Health Services Research, University of North Carolina at Chapel Hill, 725 Martin Luther King Jr. Blvd. CB# 7590, Chapel Hill, NC 27599
    Search for more papers by this author
    • Fax: (919) 640-1001.

  • Benjamin R. Saville MS,

    1. Department of Biostatistics, University of North Carolina at Chapel Hill, Chapel Hill, North Carolina
    Search for more papers by this author
  • Venita DePuy MStat,

    1. INC Research, Inc, Raleigh, North Carolina
    Search for more papers by this author
  • Paul A. Godley MD, PhD,

    1. Cecil G. Sheps Center for Health Services Research, University of North Carolina at Chapel Hill, Chapel Hill, North Carolina
    2. Division of Hematology/Oncology and the Lineberger Cancer Center, University of North Carolina at Chapel Hill, Chapel Hill, North Carolina
    3. Program on Ethnicity, Culture and Health Outcomes, University of North Carolina at Chapel Hill, Chapel Hill, North Carolina
    Search for more papers by this author
  • Katherine E. Hartmann MD, PhD,

    1. Institute for Medicine and Public Health, Vanderbilt University Medical Center, Nashville, Tennessee
    2. Department of Obstetrics and Gynecology, Vanderbilt University Medical Center, Nashville, Tennessee
    Search for more papers by this author
  • Amy P. Abernethy MD

    1. Department of Medicine, Division of Medical Oncology, Duke University Medical Center, Duke University, Durham, North Carolina
    Search for more papers by this author



Longitudinal tumor-specific studies of cancer pain across the disease trajectory provide insight into the course of pain. Information on pain predictors refines our understanding of patients with greatest distress and need.


The authors studied 1124 women with metastatic breast cancer and bone metastases, all of whom received standard treatment in an international clinical trial conducted from October 1998 to January 2001. The Brief Pain Inventory (BPI) was administered repeatedly during the course of 1 year. Hazard models were fitted to identify baseline and time-dependent covariates as predictors of pain worsening within cumulative 80-day intervals during the year.


Increased severe pain hazards were associated with non-Caucasian race (hazard ratio [HR] = 2.52; 95% CI, 1.69–3.76), restricted performance status (HR = 1.73; 95% CI, 1.13–2.64), and radiation therapy in a previous interval (HR = 2.86; 95% CI, 1.61-5.09). Estimated cumulative rates for not yet reaching a BPI score of 7 or above ranged from 0.79 (0.72–0.85) in the first interval to 0.64 (0.55–0.74) in the last interval for non-Caucasian women, whereas these rates ranged from 0.91 (0.89–0.93) to 0.84 (0.81–0.87) for Caucasian women.


By using a time-to-event hazards analysis for cancer symptom data, the authors demonstrated that non-Caucasian race predicted poorer pain control among women with metastatic breast cancer. Disparity findings from cross-sectional studies were confirmed. Pain management strategies should take race into account as a risk factor for worsening pain outcomes, and further investigation should seek to uncover and resolve the reasons for this obvious disparity. Cancer 2008. © 2007 American Cancer Society.

Pain is a key dimension of suffering and degradation of quality of life associated with cancer. 1 Pain is of particular concern in cancer; etiologies include the cancer itself, cancer procedures, treatment side effects, inactivity, and weight change. Chronic or recurrent pain affects about 30% of all patients with cancer, and 60% to 90% of patients with advanced cancer. 2, 3

Most cancer pain prevalence studies have involved only cross-sectional data. 4 Tumor-specific studies of patients' experiences of pain over time are needed, with consideration of how various factors affect pain throughout the illness trajectory. 5, 6 Repeated-assessment pain data are harder to acquire because of the expense of large-scale longitudinal data collection; clinical trials are, thus, a potentially rich source of information. However, although pain data may be gathered at multiple time points, analyses of trial outcomes data are often limited to the use of 1 measurement at a single time point or to percentage-change-from-baseline approaches 7 (in which the analysis uses only the first and last pain assessments, ignoring any interim data). Such approaches do not optimize longitudinal information available within this type of data.

Assessment of predictors of poor pain control provides critical information to refine pain management strategies. Age, race, tumor type, genetics, psychosocial context, and culture have all been found to affect pain and analgesic efficacy. 8 Younger breast cancer patients have been found to be at higher risk for post-treatment pain, 9, 10 and investigation is needed into the effects upon longitudinal pain outcomes of events and characteristics that occur or change over the course of disease, such as radiation, surgery, chemotherapy, analgesia, disease progression, and changes in performance status.

Our objective was to conduct exploratory analyses to describe worsening pain outcomes during the course of a year by using predictive models to identify baseline and time-dependent clinical and demographic risk factors associated with differences in pain severity outcomes. We defined the outcome as time to first reach a pain score of 7 or above on the Brief Pain Inventory (BPI) 0–10 pain severity or interference scales. 11 On the basis of existing evidence for cross-sectional racial disparities, 12–14 we hypothesized that, compared with their Caucasian counterparts, non-Caucasian women would have higher hazards of pain severity and pain interference in daily activities.


Overview of Study Design

This was a secondary analysis of a prospectively collected dataset from a multicenter randomized controlled trial (Novartis protocol 4244603010) of 2 bisphosphonate drugs: intravenous zoledronic acid (4 or 8 mg) versus intravenous pamidronate disodium, an adjunct to standard therapies, in the treatment of multiple myeloma and breast cancer patients with cancer-related bone lesions (no placebo arm was used). 15 The trial assessed patient outcomes during the course of a year by using the BPI to measure pain at multiple visits. No pain differences were found between treatment arms. We limited our analysis to breast cancer patients and conducted proportional hazards analyses 16 for reaching a score of 7 or above versus not yet reaching a score of 7 or above on the BPI severity and interference scales over time, identifying clinical and demographic baseline and time-dependent predictors.

Informed consent was obtained from each patient, and the trial was carried out under approval from each institution's ethical review board. Our secondary analyses were conducted under approval from the University of North Carolina at Chapel Hill Public Health Institutional Review Board.

Patients and Procedures


The original study's intent-to-treat study population comprised men and women with stage III multiple myeloma (n = 510) or stage IV breast cancer with at least 1 lytic or mixed bone metastasis (n = 1130). For the present analysis, we wanted to focus our investigation on gaining information about the experience of pain in a single tumor type: breast cancer. Thus, we excluded the multiple myeloma patients as well as 6 men, limiting the sample to women with breast cancer (n = 1124). All patients had to have at least ambulatory Eastern Cooperative Oncology Group (ECOG) performance status at enrollment. The trial was conducted at 207 centers worldwide, from October 1998 to January 2001. See the original trial publication for more details. 15

Outcomes: pain severity and interference

Patients completed the BPI questionnaire at baseline, at Months 1 and 2, and at alternate months thereafter up to Week 51. The BPI was administered in person before the patient saw a physician or received study medication. Severity was measured as average pain, pain right now, worst pain, and least pain, all 4 of which were answered on a 0–10 scale, with 0 = no pain and 10 = worst pain imaginable. The severity composite score was calculated as the arithmetic mean of the 4 severity items; the mean was calculated at assessments where at least 4 of the 7 individual items were completed. The BPI also includes a 7-item pain Interference scale, which consists of the same 0-10 response scale to the question, “Describe how, during the last 7 days, pain has interfered with your: general activity, mood, walking ability, normal work, relations with other people, sleep, and enjoyment of life.” The arithmetic mean of the 7 interference items was used to measure pain interference. The BPI has been administered and assessed for validity in several languages including Spanish, French, Japanese, Chinese, Italian, Hindi, German, Greek, and Vietnamese. 17–23 The BPI was administered in the patient's native language in any case where a validated translation existed. When no such translation existed, then it was administered with translation assistance.

There are statistical and clinical reasons for using a cutpoint of 7 on the BPI. In a measure validation study of the BPI, Serlin and colleagues found optimal cutpoints that formed 3 distinct levels of pain severity on a 0-10 numerical rating scale as follows: 1-4: mild; 5-6: moderate; 7-10: severe. 24 In addition to their importance for measure validation studies, designations of severity based on cutpoints have been used to establish clinically meaningful changes used to measure therapeutic effectiveness as well as to create clinical practice guidelines. Hence, a patient-reported pain score of 7 or above is often a red flag to clinicians that a change in pain management is necessary. 25

Predictors: clinical and demographic variables

In addition to pain assessments, subjects provided demographic, clinical, and other outcome information through interviews, written questionnaires, and physical examinations. Clinical and demographic characteristics included performance status (measured by ECOG performance status as active vs restricted), age in decades, education (some college vs no college), employment status (full-time vs other), geographic region (North America, South America, Europe, or Other), antineoplastic therapy on study entry (chemotherapy plus hormonal therapy vs hormonal therapy only), experience of a previous skeletal-related event (SRE), and time from documented initial bone metastasis to randomization. SRE refers to a composite outcome defined in the trial as experiencing 1 or more of the following: pathologic fractures, spinal cord compression with vertebral compression fracture, need for surgery to treat or prevent pathologic fractures or spinal cord compression, or need for radiation to bone. 26 Race was categorized in the case report form as “Caucasian” (n = 991), “Black” (n = 64), “Oriental” (n = 17), or “Other” (n = 52). For the present analyses, we dichotomized this variable as Caucasian (n = 991) versus Non-Caucasian (n = 133), collapsing the last 3 categories.

The following time-dependent characteristics were included: active versus restricted performance status, hospital admission, surgery, chemotherapy treatment, and radiation treatment in the preceding 80-day interval. This time lag was used to ensure that observed associations in the model results would be only between outcome events and those time-dependent predictor events that had preceded them.

Statistical Analysis

Overall statistical strategy

We conducted the analyses using SAS version 9.1 (SAS Institute Inc, Cary, NC). We first assessed the extent of missing data at baseline and at each study visit when pain was assessed. Next, we conducted preliminary exploratory analyses of the race variable by chi-square tests, t-tests, and Wilcoxon rank-sum tests to assess baseline differences by race for each predictor and outcome variable. We then visually examined Kaplan-Meier curves under a continuous time assumption to determine which factors (from among groups of demographic and clinical characteristics) would be selected as predictors for modeling. To further identify explanatory variables, we used a piecewise exponential framework 27 to fit proportional hazards models for pain outcomes.

Time to first occurrence of pain score 7 or above

We modeled the effects of the predictor variables on outcomes by using a piecewise exponential model (in PROC GENMOD, an SAS statistical software procedure [SAS Institute, Cary, NC]) under a categorical time assumption using Poisson regression. The GENMOD procedure is used to fit generalized linear models, allowing the user to specify the underlying data distribution (in this case Poisson). This approach incorporated a proportional hazards framework for each interval and enabled us to obtain parameter estimates accounting for interval-censoring in the data. 27

Intervals were based on outcome event occurrence and assigned as every 80 days after randomization with a total of 5 intervals numbered 0 through 4, with the last interval beginning 320 days after randomization and ending at 400 days (57 weeks) after randomization. Interval 0, also called the “first interval”, was the referent and was 80 days in length.

For any interval in which a patient remained in the study but did not reach the outcome, we set time at risk at 80 days. A patient's last interval was when she reached the outcome of a score of 7 or above, dropped out before the 400-day mark, or died before the 400-day mark. For those who reached the outcome, we set an event indicator variable to 1 in their last interval, and we defined their time at risk in the last interval as the time from the beginning of the interval to the date of the report of a score 7 or above. For patients who dropped out early or died, their time at risk in the last interval was set equal to the number of days from the beginning of the last interval to the date of dropout or death. If a patient never reached the outcome in any interval, then we censored that observation, such that the event indicator variable was set to 0 for all intervals, and time at risk within each interval was assigned as 80 days (or, in the cases of dropout or death, the time at risk in the last interval would be the number of days from the beginning of their last interval to the date of dropout or death).

To account for within-subject correlation across intervals, we used generalized estimating equation (GEE) methods to adjust standard errors and confidence intervals (CIs) around the estimated model parameters. Liang and Zeger (1986) introduced GEEs as a method of dealing with correlated data in generalized linear models. 28 In longitudinal data, multiple responses to the same question over time are correlated with each other, causing the variance around each estimate to appear smaller; GEE is used to make sure that the variance around each estimate is adjusted for this correlation. Our model relied on the assumption that hazards were proportional for different levels (eg, yes vs no) of each predictor within a given time interval.

Sensitivity analyses

We conducted 2 sensitivity analyses to compare results by various scenarios as follows. 1) In cases where BPI data were partly missing, instead of quantifying time-to-event as the number of days from trial enrollment to the date of the event, we quantified it by using the number of days from trial enrollment to the date of the preceding nonmissing BPI measurement. 2) Instead of counting deaths as censored, for the sensitivity analyses, we considered death as an instance of the worst possible outcome (severe pain). 29 For individuals who died, we assigned a BPI score of 10 at the time of death.


Baseline demographic and clinical characteristics of the overall sample have been previously reported. 15, 30 Average age was 57.5 years (±12.6 standard deviations [SD]). The majority of the sample was Caucasian. Of the non-Caucasian patients worldwide, 81% were in the United States. Of the 438 patients who did not complete the last scheduled visit (Week 51 ± 2 weeks), 132 (30%) patients discontinued the study because of adverse events, and 113 (26%) patients discontinued because of death. Survival over the 400-day period differed slightly by race (P = .06). Of those patients enrolled in the trial at any given visit, the proportion completing BPI assessments was ≥82%.

Table 1 shows the baseline characteristics and BPI scores by race at randomization. Caucasian patients had slightly higher mean age. Non-Caucasian patients had higher baseline pain severity and interference scores, with median severity scores 0.63 points higher for non-Caucasians and mean interference scores 0.86 points higher.

Table 1. Baseline Characteristics by Race*
Categorical baseline variablesCaucasian race, n = 991Non-caucasian race, n = 133Test statistic: baseline differencesP
  • *

    Values are expressed as number (percentage) unless otherwise indicated.

  • Mantel-Haenszel chi-square test.

  • t-statistic (normal distribution), Wilcoxon rank sum statistic (non-normal distribution).

Geographic region    
 European Union212 (21)5 (4)6.08.014
 North America661 (67)112 (84)  
 South America38 (4)1 (<1)  
 Other80 (8)15 (11)  
Baseline performance status    
 Active (0 or 1)841 (85)111 (83)0.10.756
 Restricted (2)147 (15)21 (16)  
 Missing3 (<1)1 (<1)  
College education    
 Yes247 (25)34 (26)0.06.804
 No581 (58)76 (57)  
 Missing163 (16)23 (17)  
Employed full-time    
 Yes138 (14)19 (16)0.14.712
 No694 (70)91 (68)
 Missing159 (16)23 (17)  
Previous skeletal-related event    
 Yes599 (60)78 (59)0.19.661
 No389 (39)55 (41)
 Missing3 (<1)0 (0)  
Continuous variables    
 Age, y, mean ± SD57.9±12.754.8±11.0−3.05.003
 Time from first bone metastasis to randomization in days median, [IQ range]103 [36-473]124 [36-606]0.20.654
Continuous outcomes    
 BPI Composite Score-Severity, median [25%–75% interquartile range]2.75 [1.25-4.50]3.38 [1.75-5.75]3.59.0003
 BPI Composite Score-interference, median [25%–75% interquartile range]3.00 (0.57-5.42)3.86 [1.43-6.86]3.08.0020

Figure 1 is a Kaplan-Meier curve showing differences by race in hazards of reaching a score ≥7 on the BPI severity scale. Differences by race were statistically significant for pain severity (P < .0001) and for pain interference (P = .003) (interference curve not shown). The curves were generated during exploratory analyses under an assumption of time being continuously distributed; the interval-censored nature of these data as a result of the trial visit schedule is shown in Figure 2.

Figure 1.

Kaplan-Meier estimator of distribution function (exploratory analysis, continuous time assumption) for first occurrence of a score of ≥7 on pain severity scale, by race.

Figure 2.

Distribution of visit windows around scheduled BPI visit dates is depicted. BPI indicates Brief Pain Inventory.

We fit piecewise exponential models to data structured by 80-day intervals and truncated at 400 days postrandomization. Table 2 shows the parameters resulting from the model for pain severity. The intercept represents the log incidence density, or failure rate for reaching a score ≥7, for all referent categories (including non-Caucasian race) at the referent (lowest) interval. 27 The parameter coefficients reflect each predictor's association with increased or decreased likelihood of outcome. We tested race-by-covariate interaction terms for all models, but none were significant. As shown in Table 2, the baseline predictor non-Caucasian race presented the greatest hazard for first reaching a ≥7 severity score. Inactive performance status and radiation treatment in the preceding 80-day interval were the most hazardous time-dependent characteristics.

Table 2. Pain Severity Model: Baseline and Time-Dependent Predictors for Hazard of First Reaching a Score of 7 or Above
ParameterEst.SEHR (emath image)Wald 95% CLs for HRZP > |Z|
  • SE indicates standard error; HR, hazard ratio; CLs, confidence limits.

  • SEs and related values were adjusted by using generalized estimating equation methods.

  • *

    Time-dependent covariates are events that occurred in the preceding 80-day interval.

Intercept−7.3860.4650.00 −15.89<.0001
Non-Caucasian race0.9250.2032.521.69–3.754.55<.0001
College education−0.2970.2010.740.50–1.10−1.48.139
Employed full-time−0.3690.2790.690.40–1.19−1.32.186
Previous skeletal-related event0.2500.1751.280.91–1.811.43.154
Age in decades0.0690.0671.070.94–
Admitted to hospital*0.2530.2171.290.84–1.971.17.242
Inactive performance status*0.5470.2171.731.13–2.642.52.012
Radiation therapy*1.0520.2932.861.61–5.083.59<.0001
Interval 1: 81–160 days−1.8080.2760.160.10–0.28−6.54<.0001
Interval 2: 161–240 days−1.3940.2480.250.15–0.40−5.62<.0001
Interval 3: 241–320 days−1.3730.2650.250.15–0.43−5.19<.0001
Interval 4: 321–400 days−1.3900.3350.250.13–0.48−4.15<.0001
Interval 0 (referent): 0–80 days

For the pain interference model, the baseline predictor non-Caucasian race (hazard ratio [HR] = 1.62, 95% CI = 1.09–2.41) presented the greatest hazard for reaching a ≥7 interference score. Full-time employment at baseline (HR = 0.64, CI = 0.42–0.96) and older age (HR for each 1-decade increase in age = 0.83, CI = 0.75–0.93) were associated with lesser hazard for the outcome. Inactive performance status (HR = 3.69, CI = 2.72–5.00), hospital admission since the last study visit (HR = 1.68, CI = 1.20–2.36), and radiation treatment (HR = 2.98, CI = 1.85–4.79) in the preceding 80-day interval were the most hazardous time-dependent variables for pain interference.

Model-based estimates of cumulative probabilities for not yet reaching a score ≥7 in each interval were weighted by using the sample mean values of predictors for each respective race and interval. CIs were obtained by using a linear Taylor series expansion described by Koch. 31 Figure 3 displays results for the severity outcome. For pain interference, results were similar to those for severity, with estimated cumulative rates for not yet reaching a score ≥7 ranging from 0.68 (95% CI, 0.59–0.76) in the first interval to 0.54 (95% CI, 0.44–0.64) in the last interval for non-Caucasian women, and ranging from 0.81 (95% CI, 0.78–0.83) to 0.71 (95% CI, 0.67–0.74) for Caucasian women. For both severity and interference, the CIs for the probability estimates did not overlap between Caucasian and non-Caucasian women in any given interval.

Figure 3.

Model-based 400-day cumulative population probabilities by race shows first occurrence of a score ≥7 on the BPI severity scale. Values are derived from the model estimates (Table 2) and represent cumulative per-group probabilities of reaching each interval without yet having reported a score of 7 or above on the BPI.

All results were robust in direction, magnitude, and statistical significance with regard to the 2 sensitivity analyses' various assumptions in 1) counting of deaths and 2) handling of missing data.


This study is one of the first to examine whether patients' experiences of pain over the course of metastatic cancer differ by race; it also serves as a practical analysis model for future longitudinal studies that seek to collect repeated pain data from cancer patients in their homes or in outpatient clinical settings. Our findings demonstrate that race is a longitudinal, as well as a cross-sectional, risk factor for pain; non-Caucasian women have higher hazards of pain severity and pain interference in daily activities. This confirms published clinical evidence that non-Caucasians are at highest risk for undertreatment of pain including inadequate dosing and poor access to analgesics, 13, 14, 32 as well as reports and extensive reviews that conclude that minority patients have been found consistently to be at greater risk of having undertreated pain and worse pain outcomes 12, 14, 26, 33 and greater breast cancer mortality. 32 Interestingly, pain differed by race even when employment status (a possible proxy for socioeconomic status) was included in the models.

Although matched for disease severity at baseline, during the course of a year, subtle differences in more advanced bony metastatic disease might have led to worse pain despite seemingly equal disease status (in this sample, survival did not differ significantly by race up to 400 days, but it was significantly worse for non-Caucasians when analyses were carried out to 700 days). Future research should investigate whether non-Caucasian women with metastatic disease are treated less aggressively with analgesics, and if so, the underlying sources of this problem should be identified.

Besides race, other predictors for greater pain severity hazards were an inactive performance status (a known predictor of pain and other health-related quality of life outcomes) 30, 34 and just before radiation treatment. Other predictors of interference hazards were younger age, not being employed full-time, inactive performance status, preceding hospital admission, and preceding radiation treatment. Aside from age, these factors are all indicators of more aggressive disease status – and serve as a resounding clinical reminder that more aggressive disease is a substantial predictor of worsening pain. Patients are more likely to be inactive, unemployed, hospitalized, undergoing radiation, or suffering from skeletal morbidities when their metastatic breast cancer is more rapidly advancing. Younger age has previously been demonstrated to be a predictor of greater postoperative breast cancer pain. 9, 10 Our findings showed an association between younger age and higher levels of pain interference; pain may interfere more with life tasks for younger women with breast cancer because of the competing duties of this group (eg, sick role plus mother, employee, or caretaker). The role of age in predicting not only pain levels but also adequacy of analgesia should be explored further in patients with breast cancer and other tumor types. Cleeland et al. found in a sample consisting of 21% breast cancer patients that older age (70 years or above) predicted inadequate analgesia in metastatic cancer. 13 Thus, age is an important factor that relates to different pain outcomes in different ways, and it should be considered in future research that assesses both pain and adequacy of analgesia.

Neither surgery nor chemotherapy was found to be associated with higher pain hazards in a given interval, but radiation therapy had the highest association with greater pain severity and interference, compared with all of the other covariates assessed. A likely explanation is that radiation therapy is a harbinger of painful bone metastases that require palliative irradiation for pain and tumor control, whereas chemotherapy and surgery are more commonly used for visceral disease and local recurrence (in addition to bone metastases), which may not be as painful as bone metastases. Radiation therapy is often used when pain is not alleviated by other interventions 35; thus it may have functioned as a proxy for disease progression in this analysis. Because this patient population was selected on the presence of at least 1 bone metastasis, the exaggeration of effect for an intervention directed at bone metastases, ie, radiation, is not surprising.

Hospital admission (found here to be associated with interference only) may have a greater impact on pain interference in daily activities than on pain severity because once patients with pain are admitted to a hospital, they may receive more analgesic treatment and pain management than they would outside the hospital. However, their daily activities would be disrupted by the admission; in addition, analgesics may cause enough functional limitation (eg, sedation from opioids or anticonvulsants for neuropathic pain) to register higher pain interference.

One novel aspect, and strength, of the present study is that it improves upon previous change-from-baseline analyses 7 by using repeated-assessment data collected during the trial. Another strength of this study is that the piecewise exponential model takes into account the interval-based framework of data. Two main difficulties in analysis of data from repeated-assessment studies are 1) complication of analyses by dependence among repeated observations made on the same experimental unit (the patient in this case) and 2) imbalanced or partly incomplete data. 27 The present analysis addresses these challenges by 1) using GEE methods to account for the interdependence of multiple observations for each patient and 2) evaluating the sensitivity to handling of missing data due to death or early discontinuation.

One limitation of the present study is that gaps of at least 1 week in information exist between BPI assessments; thus all pain events of interest are not captured. A patient could have experienced a severe pain or interference event between 2 assessments, but the exact date of the event would be unknown. Ideally, pain would be assessed at shorter intervals to gather more complete information. Even when that is not possible on a practical level, gaps in coverage of time should be avoided, ie, if a pain assessment instrument asks about the “last 7 days”, then the measure should be administered every 7 days. The impact of this limitation on the present study is that pain might have increased or decreased in time periods not covered by the assessments. The statistical model helped accommodate these gaps, but more complete outcome information would be desirable. Although more and complete data are desirable, this is not always practical. The method described here improves on past analytic approaches and better accommodates gaps between visits and missing data, factors inherent to studying a phenomenon like pain that occurs in real people and is measured imperfectly.

This analysis was limited also by the inclusion/exclusion criteria of the clinical trial, which was designed to answer research questions about the relative efficacy of 2 bisphosphonates. Notably, no information on specific dosage and timing of analgesic treatment was available (although all patients, regardless of treatment arm, received analgesic treatment that followed the standard of care in their country). Another issue is that our sample of non-Caucasian women was relatively small and consisted entirely of participants in a clinical trial. Given trial inclusion criteria, this sample may not be representative of the larger populations of women with breast cancer. Oversampling of minorities, as well as collection of better information on race and ethnicity, is warranted in future studies.

Adequate analgesia depends in part on patients' accurate reporting, and patients are sometimes reluctant to report pain, resulting in potential inaccuracies. Minority patients may be less willing to report pain to their physicians because of social, cultural, or religious issues, such as fear of being labeled a drug addict or feeling that suffering is deserved. 36–38 However, we found that estimated probabilities of not yet reaching severe pain were divergent by race over time (Fig. 3), even while the survival between races remained similar over the 400-day time period. This finding suggests that even with similar survival probabilities, non-Caucasian patients are at risk for experiencing more pain.

Oncologists bear responsibility for providing adequate analgesia and effective pain management over time in non-Caucasian women with metastatic breast cancer. The medical profession is obligated both to determine the scope of cancer pain treatment disparities and to devise remedies that address this disparity—starting with educating physicians about this problem.

Future research should explicitly collect and model the longitudinal effects of factors explored here, as well as other psychological, sociocultural, healthcare-level, and clinical characteristics known to affect pain. Specific factors that should be explored in future longitudinal studies include tumor characteristics, genetic characteristics, pain management preferences, and adequacy of analgesia (which may be assessed by using an algorithm described by Cleeland et al. 13).This information should be used to inform individual prognoses and to design clinical, educational, and patient-physician communication-oriented interventions aimed at improving pain management and reducing disparities. Clinicians should use information about known risk factors to inform more aggressive and earlier intervention among non-Caucasian women with metastatic breast cancer because, compared with Caucasian women, they are at risk for more pain and greater worsening of pain over the course of this disease.


The authors thank Robert DeVellis, Gary Koch, and William Miller for advice on analyses and concepts and for editorial assistance. We thank also Jean-Francois Baladi, Jens Grueger, Kevin Schulman, and Kevin Weinfurt for assistance in acquiring data.