- © 2002 by American Society of Clinical Oncology
Effectiveness of Adjuvant Fluorouracil in Clinical Practice: A Population-Based Cohort Study of Elderly Patients With Stage III Colon Cancer
- From the Department of Medicine, Hospital of the University of Pennsylvania, Philadelphia, PA; and Department of Medicine, Sections of General Internal Medicine and Hematology-Oncology, Pritzker School of Medicine, and the Cancer Research Center, University of Chicago, Chicago, IL.
- Address reprint requests to Elizabeth B. Lamont, MD, MS, University of Chicago Medical Center, 5841 S Maryland Ave, MC 2007, Chicago, IL 60637; email: elamont{at}medicine.bsd.uchicago.edu
Abstract
PURPOSE: Although randomized controlled trials (RCTs) consistently show that adjuvant fluorouracil (5-FU) extends the survival of patients with stage III colon cancer, it is not yet known whether this benefit exists in populations underrepresented on clinical trials, particularly the elderly with medical comorbidity treated in the community. In this study, we ask the following: (1) What is the hazard of death associated with adjuvant 5-FU in the general population of elderly stage III colon cancer patients? (2) Does the hazard vary with patient age?
PATIENTS AND METHODS: In this prospective, nonrandomized, population-based cohort study of 3,357 elderly Medicare beneficiaries who had undergone resection of stage III colon cancer according to the Surveillance, Epidemiology, and End-Results registries, we use propensity score matching to compare the all-cause mortality of patients who received 5-FU to matched untreated patients.
RESULTS: 5-FU reduces the hazard of death by 27% (hazard ratio, 0.73; 95% confidence interval [CI], 0.65 to 0.82) across the 6 years of our data in a Cox model. At 5 years, 52.7% (95% CI, 49.6% to 55.6%) of the treated and 40.7% (95% CI, 38.1% to 43.4%) of the matched untreated are still alive. We find that these effects do not diminish with advancing patient age.
CONCLUSION: The survival benefit of adjuvant 5-FU that has been demonstrated in participants of RCTs is also evident in a population sample of elderly Medicare beneficiaries with stage III colon cancer treated in the community. Furthermore, this survival benefit does not appear to diminish with patient age. These findings support the continued use of adjuvant 5-FU in the general population of elderly patients with stage III colon cancer and suggest that oncologists in the community are practicing at a high level of effectiveness.
RANDOMIZED CONTROLLED trials (RCTs) conducted over the past decade show that treatment with fluorouracil (5-FU) after curative-intent resection of stage III colon cancer prolongs patient survival.1-7 Sargent et al8 recently pooled survival data from several RCTs and quantified the hazard reduction at 24% (hazard ratio [HR], 0.76; 95% confidence interval [CI], 0.68 to 0.85). That group also noted that there was no evidence of decreased 5-FU efficacy among the older patients studied. However, their study sample did not reflect the age distribution of stage III colon cancer patients in the United States, as only 15% were over age 70, whereas the median age for colon cancer diagnosis in the United States is 73.9 So although 5-FU appears quite efficacious among patients treated on clinical trials, two important questions remain regarding the clinical care of the average stage III colon cancer patient: (1) Is adjuvant 5-FU as effective in clinical practice among the elderly as it is in clinical trials? (2) Does the effectiveness of adjuvant 5-FU diminish with advancing patient age?
The general concern about the external validity of chemotherapy RCT findings results not only from the well-known tendency of trial participants to differ from the population from which they are drawn (eg, to have different decision-making styles, socioeconomic status, treatment expectations)10-13 but also from the underrepresentation of elderly patients and those with medical comorbidity and/or poor performance status.14,15 In such situations where RCTs are not possible (eg, ethical or logistical reasons), well-designed observational trials (ie, cohort studies, case-control studies) may function as effective surrogates for RCTs. Some authors, including Horwitz and Feinstein,16-20 have argued convincingly that when scientific principles inherent in RCTs (eg, strict eligibility criteria) are incorporated into the design of observational studies, the bias related to the nonrandom allocation of treatment is lessened, improving their scientific validity and therefore their ability to approximate results of RCTs.
In the current project, we apply rigorous analytic techniques to an observational data source, the Surveillance, Epidemiology, and End-Results (SEER)-Medicare data, to ask two critical questions relating to the clinical care of elderly patients with colon cancer: (1) What is the hazard of death associated with adjuvant 5-FU in the general population of elderly stage III colon cancer patients? (2) Does that hazard vary with patient age? In this project, we parallel an intention-to-treat analysis of an RCT: we compare stage III patients who received any dose of 5-FU after resection to observationally similar patients who did not, and examine differences in all-cause mortality.
PATIENTS AND METHODS
Data and Cohort Construction
We used data from the National Cancer Institute’s SEER-Medicare program to study the all-cause mortality of elderly Medicare beneficiaries with histories of stage III colon cancer. The SEER-Medicare database is a National Cancer Institute–sponsored individual-level linkage of the clinical data collected by the SEER registries with health services billing claims collected by Medicare for administrative purposes. These data are widely used by researchers studying outcomes, clinical epidemiology, and health services factors among elderly cancer patients.21
The SEER program collects information regarding the diagnosis and treatment of patients with cancer from 11 geographically diverse tumor registries in order to monitor trends in incidence and survival. It is estimated that approximately 14% of the American population with cancer9 is represented in these data and prior research has shown that, in the aggregate, patients in these registries are demographically representative of the general population.22 The SEER program collects detailed information about initial diagnosis and treatment, including date of diagnosis, site, histology, and stage of tumor at diagnosis, in addition to demographic information. Mortality data are provided through linkage to death certificates. The SEER program conducts annual audits of their data to ensure data quality and completeness, holding the standard of ascertainment at 98%.23
Medicare is a federally sponsored health insurance program administered by the Centers for Medicare and Medicaid Services whose beneficiaries include more than 96% of all United States citizens aged 65 and older.24 The Centers for Medicare and Medicaid Services maintains billing records of outpatient, inpatient, and other claims for all beneficiaries not enrolled in risk contract health maintenance organizations. Inpatient data are from the Medicare Provider Analysis and Review (MedPAR) file. Outpatient data are from the Outpatient Standard Analytic File (SAF) and the National Claims History file.
We empaneled an incidence cohort of all patients in the SEER file diagnosed with pathologically confirmed stage III adenocarcinoma of the colon; diagnosed at or after age 67; during the period January 1, 1993, to December 30, 1996; entitled to Medicare parts A and B during the observation period so that evaluation of both outpatient care (part B) and inpatient care (part A) was possible; and not enrolled in a health maintenance organization (for whom individual claims may not be submitted to Medicare because of capitated payment) during the observation period. Patients were excluded if they had Medicare International Classification of Diseases, Ninth Revision, Clinical Modification (ICD-9-CM) codes for colon cancer that preceded the SEER date of colon cancer diagnosis by more than 2 months; if they had not undergone cancer surgery; if their colon cancer was diagnosed at autopsy; or if they died within 90 days of diagnosis. Surviving patients were fixed right censored at December 31, 1998. The analytic sample consisted of 3,357 patients.
Variables
Information regarding demographic factors of age, race, sex, marital status, socioeconomic status (eg, census tract median income), and market structural factors (ie, urban v rural residence, and cancer registry) was obtained directly from the data set. We also created variables to indicate 5-FU exposure and to adjust for prior medical comorbidity. 5-FU use was ascertained by reviewing each patient’s National Claims History, Outpatient SAF, and MedPAR file corresponding to the observation period for the presence or absence of codes consistent with administration of 5-FU. The details of the method and the resultant construct validity of the measure have been published previously.25 Briefly, the procedure codes are examined for the presence of Healthcare Financing Administration Common Procedure Code (HCPC) J9190, indicating intravenously administered 5-FU. Patients with at least one HCPC J9190 during the observation period were considered to have been treated with 5-FU. Patients with IDC-9-CM codes V58.1, V662, V672, and 9925, indicating chemotherapy administration, in any diagnostic fields in the MedPAR files were also considered to have been treated with 5-FU, as prior investigations have shown that, for patients with histories of colon cancer, these codes invariably represent 5-FU administration (Warren, unpublished data).
To minimize confounding by unmeasured health, we used health-related data from a number of sources and controlled for health in a maximally flexible way to estimate baseline medical comorbidity for each patient. The Charlson comorbidity score is a convenient method of operationalizing co-occurring medical illness in cancer patients and is often used for risk adjustment. The score, ranging from 0 to 29, consists of a weighted sum of 17 major illnesses (eg, myocardial infarction, cerebrovascular accident, diabetes, liver disease, dementia, renal disease). All claims from the prior 2 years of inpatient and outpatient use were reviewed, and separate scores for each data source were developed and parameterized as a family of five indicator variables.26-29 Furthermore, the histologic data pertaining to tumor grade, pathologic data pertaining to number of lymph nodes containing tumor, and clinical data regarding treatment with postoperative radiotherapy (all of which are important prognostic factors) from the SEER registries were incorporated into all models.
Propensity Score Matching
In order to minimize the bias related to the nonrandom allocation of treatment, we developed a matching scheme that included variables that have been shown previously to be associated with chemotherapy use.25,30-42 We included variables within the following domains: oncologic (tumor grade, number of lymph nodes containing tumor, need for postoperative radiotherapy), medical comorbidity (Charlson scores), demographics (age, race, sex, marital status, socioeconomic status), market structural factors (urban v rural, with separate indicator variables for each registry), and temporal factors (year of diagnosis). These variables were included in a logistic regression model that determined the probability of adjuvant 5-FU use; this probability is referred to as the propensity score.43-47 By matching treated and untreated patients on their propensity score, we then reduce selection bias. The approach has been used successfully in the medical context in the past48-53 and recent work has shown that propensity scores can be used to observationally replicate RCT results, even in contexts where conventional regression performs poorly.54,55
To implement the propensity score matching, we first used logistic regression to model likelihood of receipt of 5-FU during the observation period, using the previously mentioned covariates as predictors. The resultant model had a c-statistic of greater than 0.83, indicating good discrimination between those who did and did not receive 5-FU. We then used the model to predict each patient’s probability of receiving 5-FU. The resultant probability is the propensity score. On the basis of their propensity score, patients who received 5-FU were then matched, with replacement, to patients who did not receive 5-FU. Individuals were matched within tight bounds on their propensity scores; that is, their predicted probability of receiving adjuvant 5-FU could vary by no more than 0.005 (0.5%) on a scale of 0 to 1.
Statistical Analyses
We evaluated the association of claims for 5-FU on the cohort’s survival in two ways. First, we examined unadjusted Kaplan-Meier curves to evaluate for crude differences between treated cases and matched untreated patients. Second, we used Cox regression to estimate an average difference in the hazard of death across a 6-year follow-up period. We stratified these results on matched pairs in order to appropriately adjust standard errors for the matching, and controlled for the replication of controls as clusters in the manner of Lin and Wei.56 All regression results control for propensity score using categorical variables for each decile of score among treated patients.43,54 The Kaplan-Meier curves without matching show an implausible substantial and instantaneous benefit of 5-FU. Conventional Cox modeling with all 67 variables included in the propensity score generation regression are, in this case, able to reproduce the 5-FU benefit demonstrated by propensity score matching, although without having performed the matching, we would have no assurance that it was able to perform so well.
All analyses were performed in STATA 7.0 (Stata Corp, College Station, TX). The research qualified for institutional review board exemption.
Sensitivity Analyses
Using the propensity score technique, we adjusted adequately for all factors included in the propensity regression. This approach relies on the assumption of selection on observables; that is, we assume that the decision to administer chemotherapy is made on the basis of only those variables that we can observe through our data: tumor grade, number of lymph nodes containing tumor, postoperative radiation therapy, patient age, comorbidity, race, sex, marital status, socioeconomic status, location of residence, and year of diagnosis. Our approach meets the criteria that Heckman et al57 articulated for the effective use of a propensity score: the data for cases and controls are drawn from the same source, outcome data are drawn from the same sources, a rich set of covariates is available to model selection, and cases and controls come from the same local areas so that any likely unobserved factors shaping behavior are also matched.57,58 The process of matching serves to ensure that the treated and untreated groups are equivalent in any known confounders that influence both 5-FU use and outcome; each patient has a matched untreated control that was equally likely (but for random factors) to have been treated.
Because there may be important but unmeasured covariates that were not represented in our data set and not identified in prior epidemiologic research, we sought to evaluate the sensitivity of our analysis to possible missing covariates in the following manner. After identifying variables that were highly associated with the receipt of chemotherapy (as quantified by the partial R2 in the propensity score regression) and removing them from the propensity regression, we then re-estimated the model to determine the stability of our estimates of 5-FU’s effectiveness in the setting of an intentionally inadequate model. We quantified the degree to which results varied between the intentionally misspecified model and the original model. Results that are relatively insensitive to the exclusion of factors known to be of major importance suggest that the degree of bias introduced by some other unidentified factor would be small.
In a second sensitivity test, we replicated all of our analyses examining the impact of age on the effectiveness of 5-FU. We found that our conclusions did not vary with different parameterizations of age, including quadratic, cubic, or multiple indicator variables approaches. We therefore presented the results with a linear age effect for simplicity.
RESULTS
Description of Cohort and Matching
Table 1 lists differences between 1,834 patients treated with 5-FU and those not treated with 5-FU using first unmatched analyses and then matched analyses. Of note, 32 treated individuals (1.7%) could not be matched to an untreated patient within our tight propensity score bounds and are excluded from the analyses. In the matched analyses, the treatment groups are comparable, with the only statistically significant difference being that patients treated with 5-FU are slightly more likely to live in urban regions than those untreated. The groups are also balanced across registries (data not shown).
Table 1. Cohort Characteristics
As shown in Fig 1, our study population approximates the central portion of the age distribution of patients with surgically resected stage III colon cancer in the United States. Patients in our cohort treated with 5-FU had a median age of 74, with an interquartile range from 70 to 78 years. The median age for all stage III colon cancer patients in the SEER registry was 74, with an interquartile range of 67 to 80 years.
Fig 1. Age distribution of study cohort. Comparison information for SEER distribution is all surgically resected, pathologically confirmed, stage III colon cancer patients who survived at least 90 days from diagnosis present in SEER (same restrictions as required for study population).
5-FU Dosing and Schedules
Details about 5-FU regimens can be adequately ascertained for the 1,482 of 1,802 patients (82.2%) whose 5-FU claims were not exclusively detected in the Outpatient SAF or MedPAR files. These 1,482 patients received 42,394 doses over the 13 months after diagnosis. The median dose was 1,000 mg, with 95% of doses either 500 (21%) or 1,000 (74%) mg. The median time between doses was 7 days. Patients received a median of 29 5-FU doses. Patients received a median of 0.58 5-FU doses per week after adjusting for survival.
Effectiveness of 5-FU in the Elderly
As shown in Fig 2, the Kaplan-Meier curves for the effects of 5-FU after matching on the propensity score show a clear survival benefit associated with 5-FU. Fully 89.6% (95% CI, 88.1% to 90.1%) of patients receiving 5-FU are alive 1 year after diagnosis, in contrast to 85.6% (95% CI, 83.9% to 87.2%) of matched patients who did not receive 5-FU. At 5 years, 52.7% (95% CI, 49.6% to 55.6%) of the treated and 40.7% (95% CI, 38.1% to 43.4%) of the matched untreated are still alive. 5-FU reduces the hazard of death by 27% (HR, 0.73; 95% CI, 0.65 to 0.82) across the 6 years of our data in a Cox model.
Fig 2. 5-FU in the elderly, matched sample results. Kaplan-Meier estimates for survival among patients who received 5-FU and propensity score–matched untreated patients. One- and 5-year cut points are shown for the reader’s convenience.
Does the Effectiveness of 5-FU Vary With Age Among the Elderly?
The effect of 5-FU appears to vary little with age; an interaction term between age and treatment is of neither clinical nor statistical significance (HR, 1.07 per additional decade of age; 95% CI, 0.85 to 1.35). For example, the hazard ratio for a 75 year-old is 0.73 (95% CI, 0.65 to 0.82); for an 85 year-old, 0.78 (95% CI, 0.62 to 0.99); these differences are not statistically significant.
Sensitivity Analyses
To evaluate the sensitivity of our analyses to possible excluded variables, we intentionally misspecified the model. The three most important predictors of 5-FU use were comorbidity, age, and cancer registry. We repeated our analyses, sequentially excluding each of these predictors from the propensity score generation algorithm (logistic regression). The effect sizes are listed in Table 2; in no case did intentional omission of a known important confounder lead to a 20% change in the estimated benefit of 5-FU treatment. These results suggest the magnitude of the effect that an important unmeasured covariate (ie, highly correlated both with receipt of 5-FU and with mortality) would need to have to introduce substantial bias into our estimation.
Table 2. Sensitivity of Results to Intentional Misspecification
DISCUSSION
Our results suggest that the survival benefit associated with adjuvant 5-FU that has been demonstrated in participants of multiple RCTs is also evident in elderly Medicare beneficiaries with stage III colon cancer, a population with a much broader range of age and comorbidity than typically is included in RCTs. Similarly, these results suggest that physicians treating stage III colon cancer with adjuvant 5-FU are doing so with a high degree of effectiveness throughout the population; typical practitioners are obtaining the same levels of effectiveness (a 0.73 HR) that were obtained in RCTs (a 0.76 HR), a difference that is statistically indistinguishable. Our results also suggest that the magnitude of the benefit of 5-FU does not vary with age among the elderly. We demonstrate this to be true across an age range more representative of the population of colon cancer patients than was typically present in previous RCTs. Together, these findings support the continued use of adjuvant 5-FU in the general population of elderly patients with stage III colon cancer. Moreover, our results suggest that adjuvant 5-FU would have benefited those untreated patients in the sample who were otherwise similar to treated patients.
There are limitations to the work. Our sensitivity analyses suggest that our results would not be immune to bias if there exists an unmeasured covariate that both (1) has as powerful an effect as comorbidity on both the decision to treat patients and mortality, and (2) was poorly correlated with our existing controls. Furthermore, our work has the well-described limitations of administrative claims data.59-65 Among these may be incomplete ascertainment of 5-FU treatment; such omissions would likely bias our results towards the null (ie, it would lead us to underestimate the effectiveness of 5-FU). Finally, although we have controlled extensively for comorbidity, the effectiveness of 5-FU might vary depending on what other diseases a patient has; future work should establish whether there are particular categories of comorbidity for which 5-FU may be more or less effective.
In summary, these results suggest that adjuvant 5-FU is similarly effective in unselected elderly patients with resected stage III colon cancer as it is efficacious in younger patients enrolled on clinical trials. Moreover, these results suggest that practicing oncologists treating patients in the community with adjuvant 5-FU provide the same survival advantage as those clinical researchers treating patients on experimental protocols, providing a reassuring counterexample to studies suggesting that community physicians may be suboptimal in their practice.66
Acknowledgments
Supported by National Institutes of Health grant nos. K12 AG-0048-09 (to E.B.L.), K07 CA93892-01 (to E.B.L.), and T32 GM07281 (to T.J.I.) and John Hartford Foundation grant no. 990590 (to E.B.L.).
ACKNOWLEDGMENT
We thank the Applied Research Program, National Cancer Institute; the Office of Information Services, and the Office of Strategic Planning, Centers for Medicare and Medicaid Services;Information Management Services, Inc; and the SEER program tumor registries in the creation of the SEER-Medicare database. We also thank Joan Warren, PhD,at the National Cancer Institute for assistance with our analyses.
Footnotes
-
This study used the linked Surveillance, Epidemiology, and End-Results–Medicare database. The interpretation and reporting of these data are the sole responsibility of the authors.
- Received March 15, 2002.
- Accepted June 12, 2002.
