A comparison of charlson and elixhauser comorbidity measures to predict colorectal cancer survival using administrative health data
Cancer survival is related to features of the primary malignancy and concurrent presence of nonmalignant diseases (comorbidities), including weight-related conditions (obesity, weight loss). The Charlson and Elixhauser methods are 2 well-known methods that take comorbidities into account when explaining survival. They differ in both the number and categorization of comorbidities.
Cancer, comorbidity, and survival data were acquired from inpatient administrative hospital records in 574 colorectal cancer patients. Robust Poisson regression was used to analyze 2- and 3-year survival according to cancer features and comorbidities classified by the Charlson and Elixhauser methods. Data for weight-related conditions (body mass index, weight loss) and performance status were acquired upon a new patient visit to the regional cancer center. Discrimination was assessed with the concordance (c) statistic.
A base model (age, sex, stage) had excellent discrimination (c-statistic, 0.824 [2-year survival] and 0.827 [3-year survival]). The addition of Charlson comorbidities did not outperform the base model (c-statistic, 0.831 [2-year survival] and 0.833 [3-year survival]). Elixhauser comorbidities added higher discrimination compared with the base model, both in stage and overall (c-statistic, 0.852 [2-year survival] and 0.854 [3-year survival]; P < .01). The greatest increase in the c-statistic contributed by the addition of the Elixhauser comorbidities occurred in stage II patients (increased from 0.683 to 0.838). Overall, the Elixhauser comorbidities outperformed the Charlson comorbidities (P < .05). The use of self-reported weight and performance status data significantly increased discrimination by the Elixhauser method in 2-year but not 3-year survival.
The Elixhauser method is a superior comorbidity risk-adjustment model for colorectal cancer survival prediction. Cancer 2011. © 2010 American Cancer Society.
Nonmalignant diseases (comorbidities) are common and have been shown to contribute to cancer survival.1-4 Cancer patients who are older than 70 years typically have 3 comorbidities,5 of which cardiovascular disease and hypertension are among the most common.6 In addition to cancer site and stage and patient variables such as sex, age, and performance status, weight-related conditions (eg, underweight, obesity, and involuntary weight loss) predict cancer survival.7, 8 Survival models excluding any of these relevant explanatory variables would result in biased estimates. Inpatient administrative health data is a common source used to identify comorbidities recorded with International Classification of Disease (ICD) diagnostic codes. The Charlson Index9 is the most common measure to assess comorbidity in cancer patients, with approximately 1000 citations since 2005. It was originally based on the risk of 1-year death for 19 conditions in 559 hospitalized internal medicine patients. The Elixhauser method, a more recent approach encompassing 31 conditions, was developed and validated by predicting in-hospital mortality, length of stay, and hospital charges using administrative data encompassing all adult acute-care inpatient hospitalizations that occurred in 1992 in California (n = 1,779,167).10 Elixhauser has been suggested to be a superior risk-adjustment model in patients with cardiac and respiratory conditions11, 12 and with osteoarthritis.13 Of the 70 citations using Elixhauser in cancer research since 2005, only 1 compared Elixhauser with Charlson.14
Weight loss and obesity are included in the Elixhauser but not the Charlson measure. Involuntary weight loss is an independent prognostic factor for poor treatment response and reduced survival in cancer patients.7 Obesity (defined as a body mass index [BMI] ≥30 kg/m2) is regarded as a risk factor for all-cause mortality in the general population15; however, recent reports suggest that in wasting diseases (including cancer and renal disease), obese patients have longer, not reduced, survival compared with those with a lower BMI.8, 16, 17 It is also presently uncertain how body weight variables should be included in cancer survival models and whether the manner in which they are captured in administrative data and included in the Elixhauser method correspond to what is currently known about their impact on survival.
We studied a population-based cohort of stages II-IV colorectal cancer patients to further understand optimal risk-adjustment procedures in cancer survival research. Building on a base model that included age, sex, and stage, we added either Charlson or Elixhauser comorbidities to determine the better risk-adjustment method. Evaluation of treatment efficacy, both in clinical and research settings, requires identification of superior risk-adjustment models. Consequently, treatment variables are not included in risk-adjustment models. We also evaluated whether adding performance status and replacing Elixhauser weight-related conditions with more cancer-appropriate weight variables derived using clinical data improved model performance.
Ethical approval was obtained from the Alberta Cancer Board Research Ethics Board.
Patient Population, Demographics, and Cancer-Related Variables
All Alberta residents with stages II-IV colorectal cancer (ICD-O: C18-C20 excluding appendix cancer)18 who visited a new patient medical oncology clinic at the Cross Cancer Institute (Edmonton, Alberta, Canada) between June 2004 and March 2006 were included. This single tertiary-cancer facility in northern Alberta (population 1.8 million) handles almost all referrals (>95%) for consideration of chemotherapy and radiation therapy. Age, sex, stage, tumor site, treatment type, and date of death were obtained from the Alberta Cancer Registry. The first clinic visit date was our index date and the point in time at which stage was established and from which survival time was quantified. Surviving patients were followed for more than 3 years. Cancer stage was defined using the American Joint Committee on Cancer (6th edition) stages I, II, III, and IV.19 The study end date was June 9, 2009.
Height, weight, and weight history were patient-reported variables obtained from the Patient Generated Subjective Global Assessment (PGSGA).20 This is a clinical tool used for nutritional screening of cancer patients and is completed at this clinic visit. Self-reported height, weight, and weight history have been found to be reliable measures in healthy adults21, 22 and cancer patients.23 The PGSGA also includes a patient-reported version of the Eastern Cooperative Oncology Group (ECOG) performance status score.
Comorbidities were obtained from inpatient hospitalization administrative data provided by the provincial Ministry of Health. This data set includes all inpatient hospitalizations that occurred in any Alberta hospital and includes up to 16 ICD-10 diagnostic codes for each hospitalization. Hospitalizations from the year prior to the index date were used to assess comorbidities.
Charlson and Elixhauser methods
Comorbidity burden was quantified using validated Charlson and Elixhauser coding algorithms available for ICD-10 codes.24 Binary variables indicating the presence or absence of each Charlson and Elixhauser comorbidity were created. Patients not hospitalized in the year prior to the index date were considered to have no comorbidities. Cancer comorbidity categories were excluded, and those with ≤3 occurrences were omitted for statistical reasons.
Augmented Elixhauser method
Clinical data were used to augment the Elixhauser method. All comorbidities unrelated to body weight were derived using administrative data. Two weight-related variables were created using PGSGA data for this method: abnormal weight loss and abnormal BMI. These variables were chosen because this is a weight-losing population (mean 6-month weight loss, 6.7% ± 7.8%), and risk of death in cancer patients increases at lower BMI values.1, 25, 26 Abnormal weight loss was defined as a loss of ≥20% within the past 6 months (ie, grade 3 weight loss from the National Cancer Institute Common Terminology Criteria for Adverse Events, v3.0). Abnormal BMI was defined as <20 kg/m2, a value consistent with a diagnosis of cachexia.27 This method also included self-reported ECOG performance status from the PGSGA.
Analysis was conducted in 3 stages. In each stage, regression models were fitted and comorbidities were grouped using the following methods: 1) Charlson, 2) Elixhauser, and 3) augmented Elixhauser. Two- and 3-year survival (robust Poisson regression) and overall survival (Cox regression) were the outcome variables. Because >10% of the population had the outcome, robust Poisson regression was used instead of logistic regression to avoid overestimation of risk ratios.28, 29 Reporting both Cox and robust Poisson regression (at 2 and 3 years) alongside univariate data is justified on the grounds of establishing parameter stability. Two- and 3-year survival times were chosen based on the number of events (ie, events were too infrequent at 1 year) and the finding that this period is clinically relevant to this population.
First, unadjusted hazard ratios (HR) and incidence rate ratios (IRR) were calculated for each Charlson and Elixhauser comorbidity variable using Cox and robust Poisson regression models; corresponding 95% confidence intervals were also determined.
Second, adjusted Cox and robust Poisson regression models were fitted. Models included variables with well-accepted associations with survival (ie, age, sex, and stage), as well as all the comorbidity variables using Charlson, Elixhauser, and augmented Elixhauser approaches. For each comorbidity, adjusted HR, IRR, and 95% confidence intervals were calculated. Age was treated as a continuous variable,30 and performance status was dichotomized by combining scores 0-1 and 2-4.7, 23
Finally, robust Poisson regression models were compared to determine which of the methods—Charlson, Elixhauser, or augmented Elixhauser—best predicted 2- and 3-year survival. Each regression model encompassed 1 of the 3 comorbidity methods as well as age, sex, and stage. We also evaluated a base model that included only age, sex, and stage. We also chose to include robust Poisson regressions because concordance (c) statistics for Cox regression models are not validated.
To determine which comorbidity method best predicted survival, the change in c-statistic was tested using the Stata roccomp command.31 This command tests for differences between 2 receiver operating characteristic (ROC) curves,32 a graph of the true- and false-positive rates. The c-statistic is equivalent to the area under the ROC curve33—in this case, the probability that someone who died had a higher predicted probability of dying than someone who did not die in the specified time frame. A c-statistic of 0.5 indicates the model predicts the outcome as well as chance, 0.7 to <0.8 is acceptable, 0.8 to <0.9 is excellent, 0.9 to <1.0 is outstanding, and, 1.0 is perfect prediction.34 Other comorbidity measure validation studies have also used this method.11, 12, 14 To ensure that our model was not overfitting, we internally validated our robust Poisson regression models using 10-fold cross-validation.35 We took all the predicted probabilities from the 10 models and generated a single ROC curve for each model.
SPSS version 17.0 (SPSS Inc., Chicago, IL) and Stata version 10.0 (Stata Corp., College Station, TX) were used for statistical analysis. All tests were 2 sided, and α was set at 0.05.
Demographic and Cancer-Related Variables
We included 574 patients, 85% of whom were hospitalized in the year prior to the index date in a total of 764 hospitalizations. Population characteristics are presented in Table 1. No differences in age, sex, and performance status were observed between hospitalized and nonhospitalized patients.
Table 1. Population Characteristics (n=574)
| Age, y 64±12||Range: 32-90|
|Primary tumor site (%)|
| Rectosigmoid junction||12.2|
|Cancer stage (%)|
|BMI Category (%)|
|Performance status score (%)|
| Colon/rectosigmoid junction (%)|
| Surgery and chemotherapya||57.1|
| Surgery only||31.4|
| Chemotherapya only||4.2|
| Chemotherapy,a radiation, and surgery||2.6|
| Other, including immunotherapy||2.1|
| None or refused||2.6|
| Rectum (%)|
| Chemotherapy,a radiation and surgery||57.1|
| Chemotherapya and surgery||12.2|
| Surgery only||9.5|
| Radiation and surgery||6.8|
| Chemotherapya and radiation||6.1|
| Chemotherapya only||2.0|
| Other, including immunotherapy||1.4|
| None or refused||4.8|
Charlson and Elixhauser comorbidities
Overall, 25.6% had ≥1 Charlson comorbidities, and 7.5% of those patients had ≥3 comorbidities. Table 2 presents the frequency of each Charlson comorbidity and its unadjusted and adjusted HR and IRR. The following Charlson comorbidities were dropped from statistical analyses because of low counts (ie, ≤3 patients affected): peripheral vascular disease, rheumatic disease, hemiplegia/paraplegia, moderate/severe liver disease, and HIV/AIDS. In the unadjusted Cox regression models, congestive heart failure, dementia, and renal disease were significant predictors of survival, but only renal disease was significant in the adjusted model.
Table 2. Prevalence of Charlson Comorbidities and Relation With All-Cause Mortality in Stages II-IV Colorectal Cancer Patients (n=574)
|Myocardial infarction||31 (5.4)||1.1 (0.69, 1.9)||0.98 (0.54, 1.8)||1.1 (0.68, 1.8)||0.96 (0.61, 1.5)||1.1 (0.76, 1.7)||1.1 (0.75, 1.5)|
|Congestive heart failure||19 (3.3)||2.2 (1.3, 3.8)a||1.9 (0.99, 3.8)b||2.0 (1.4, 2.9)a||1.6 (0.87, 3.0)||1.6 (1.1, 2.3)a||1.2 (0.70, 2.1)|
|Cerebrovascular disease||7 (1.2)||0.27 (0.038, 1.9)||0.21 (0.024, 1.8)||0.44 (0.072, 2.7)||0.85 (0.11, 6.4)||0.35 (0.057, 2.2)||0.67 (0.095, 4.7)|
|Dementia||6 (1.1)||3.1 (1.4, 7.0)a||1.7 (0.61, 4.7)||2.6 (1.8, 3.9)a||2.0 (0.83, 4.9)||2.5 (2.3, 2.8)a||2.3 (1.1, 5.1)a|
|Chronic pulmonary disease||37 (6.4)||1.4 (0.89, 2.2)||0.94 (0.58, 1.5)||1.3 (0.86, 1.9)||0.84 (0.55, 1.3)||1.4 (1.0, 1.9)a||0.98 (0.69, 1.4)|
|Peptic ulcer disease||5 (0.9)||1.3 (0.42, 4.1)||1.5 (0.48, 4.7)||0.62 (0.11, 3.6)||0.63 (0.14, 2.8)||1.0 (0.34, 2.9)||0.96 (0.56, 1.6)|
|Mild liver disease||6 (1.0)||0.30 (0.042, 2.1)||0.51 (0.071, 3.7)||0.52 (0.086, 3.1)||0.78 (0.32, 1.9)||0.41 (0.069, 2.5)||0.62 (0.24, 1.6)|
|Diabetes without complications||65 (11.3)||1.1 (0.73, 1.6)||1.1 (0.72, 1.5)||1.1 (0.74, 1.5)||1.0 (0.75, 1.4)||1.1 (0.80, 1.5)||1.0 (0.80, 1.3)|
|Diabetes with complications||7 (1.2)||1.9 (0.69, 5.0)||0.83 (0.19, 3.6)||1.8 (0.94, 3.5)||0.62 (0.34, 1.1)||1.4 (0.75, 2.7)||0.70 (0.42, 1.2)|
|Renal disease||11 (1.9)||2.5 (1.2, 5.0)a||3.3 (1.2, 9.1)a||2.0 (1.3, 3.2)a||1.9 (1.2, 3.0)a||1.6 (1.0, 2.5)a||1.5 (0.99, 2.2)b|
Overall, 46.5% had ≥1 Elixhauser comorbidity, and 26.9% of those patients had ≥3 comorbidities. Table 3 presents the frequency of each Elixhauser comorbidity and its unadjusted and adjusted HR and IRR. The following Elixhauser comorbidities were dropped from statistical analyses because of low counts: peripheral vascular disorders, paralysis, peptic ulcer disease, rheumatoid arthritis, coagulopathy, drug abuse, alcohol abuse, psychosis, and HIV/AIDS. In the unadjusted Cox regression models, congestive heart failure, cardiac arrhythmias, renal failure, and fluid and electrolyte disorders were significant predictors of survival. Cardiac arrhythmias, uncomplicated hypertension, and fluid and electrolyte disorders were significant predictors in the adjusted Cox regression model. Unlike other comorbidities, uncomplicated hypertension was positively associated with survival. Pairwise correlations between the comorbidities did not reveal any collinearity issues in different conditions within the Elixhauser and Charlson measures.
Table 3. Prevalence of Elixhauser Comorbidities and Relation With All-Cause Mortality in Stages II-IV Colorectal Cancer Patients (n=574)
| Congestive heart failure||19 (3.3)||2.2 (1.3, 3.8)a||1.7 (0.87, 3.4)||2.0 (1.4, 2.9)a||2.1 (1.1, 3.8)a||1.6 (1.1, 2.3)a||1.6 (0.88, 2.8)|
| Cardiac arrhythmias||52 (9.1)||1.6 (1.1, 2.3)a||1.6 (1.0, 2.5)a||1.5 (1.1, 2.1)a||1.4 (0.98, 1.9)b||1.2 (0.91, 1.7)||1.1 (0.79, 1.4)|
| Valvular disease||10 (1.7)||1.8 (0.83, 3.7)||0.58 (0.21, 1.6)||1.6 (0.84, 3.0)||0.65 (0.32, 1.3)||1.3 (0.67, 2.3)||0.63 (0.34, 1.2)|
| Pulmonary circulation disorders||11 (1.9)||1.7 (0.80, 3.6)||1.4 (0.61, 3.2)||1.4 (0.74, 2.8)||0.92 (0.49, 1.7)||1.6 (1.0, 2.5)a||0.99 (0.63, 1.6)|
| Hypertension, uncomplicated||140 (24.4)||0.92 (0.69, 1.2)||0.60 (0.43, 0.84)a||0.89 (0.66, 1.2)||0.62 (0.46, 0.85)a||0.88 (0.69, 1.1)||0.68 (0.53, 0.88)a|
| Hypertension, complicated||12 (2.1)||2.0 (0.98, 4.0)b||0.60 (0.12, 2.9)||1.9 (1.1, 3.0)a||1.1 (0.55, 2.1)||1.5 (0.90, 2.4)||0.98 (0.58, 1.7)|
| Other neurological disorders||11 (1.9)||1.3 (0.57, 2.9)||1.1 (0.47, 2.7)||1.4 (0.74, 2.8)||1.3 (0.65, 2.6)||1.4 (0.79, 2.4)||1.2 (0.79, 2.0)|
| Chronic pulmonary disease||37 (6.4)||1.4 (0.89, 2.2)||1.0 (0.61, 1.7)||1.3 (0.86, 1.9)||0.93 (0.61, 1.4)||1.4 (1.0, 1.9)a||1.1 (0.79, 1.5)|
| Diabetes, uncomplicated||65 (11.3)||1.1 (0.74, 1.6)||1.1 (0.71, 1.6)||1.1 (0.74, 1.5)||1.1 (0.78, 1.6)||1.1 (0.80, 1.5)||1.2 (0.86, 1.5)|
| Diabetes, complicated||7 (1.2)||1.9 (0.69, 5.0)||0.45 (0.095, 2.2)||1.8 (0.94, 3.5)b||0.52 (0.28, 0.96)a||1.4 (0.75, 2.7)||0.60 (0.36, 1.0)b|
| Hypothyroidism||35 (6.1)||1.1 (0.64, 1.8)||1.0 (0.59, 1.8)||0.98 (0.59, 1.6)||0.70 (0.47, 1.0)b||1.1 (0.72, 1.6)||0.90 (0.63, 1.3)|
| Renal failure||11 (1.9)||2.5 (1.2, 5.0)a||4.9 (0.87, 27.7)b||2.0 (1.3, 3.2)a||2.0 (1.1, 3.6)a||1.6 (1.0, 2.5)a||1.6 (1.0, 2.6)a|
| Liver disease||8 (1.4)||0.78 (0.25. 2.4)||1.3 (0.39, 4.3)||1.2 (0.48, 2.9)||1.7 (0.74, 3.9)||0.94 (0.38, 2.3)||1.3 (0.59, 3.0)|
| Obesity||16 (2.8)||0.71 (0.32, 1.6)||1.4 (0.57, 3.4)||0.78 (0.33, 1.8)||1.3 (0.62, 2.9)||0.62 (0.26, 1.5)||0.96 (0.47, 2.0)|
| Weight loss||6 (1.0)||0.62 (0.15, 2.5)||0.46 (0.11, 2.0)||0.52 (0.086, 3.1)||0.75 (0.099, 5.7)||0.41 (0.069, 2.5)||0.45 (0.060, 3.4)|
| Fluid and electrolyte disorders||29 (5.1)||2.2 (1.4, 3.4)a||2.0 (1.2, 3.3)a||1.7 (1.1, 2.4)a||1.2 (0.78, 2.0)||1.7 (1.3, 2.2)a||1.4 (1.0, 1.9)a|
| Blood loss anemia||6 (1.0)||0.66 (0.16, 2.6)||0.26 (0.061, 1.1)b||0.52 (0.086, 3.1)||0.26 (0.078, 0.88)a||0.83 (0.27, 2.6)||0.49 (0.19, 1.2)|
| Deficiency anemia||27 (4.7)||1.4 (0.82, 2.3)||1.3 (0.75, 2.2)||1.2 (0.70, 1.9)||1.2 (0.67, 2.0)||1.2 (0.81, 1.8)||1.3 (0.94, 1.8)|
| Depression||17 (3.0)||1.5 (0.79, 2.8)||1.8 (0.88, 3.6)||1.5 (0.89, 2.5)||2.2 (1.2, 4.2)a||1.5 (0.99, 2.2)b||2.1 (1.3, 3.4)a|
| ECOG performance status scores 2-4||134 (23.3)||2.1 (1.6, 2.7)a||2.5 (1.9, 3.3)a||1.9 (1.5, 2.4)a||1.6 (1.3, 1.9)a||1.7 (1.4, 2.1)a||1.5 (1.2, 1.7)a|
| BMI <20||38 (6.6)||1.7 (1.1, 2.7)a||1.6 (0.94, 2.6)b||1.8 (1.3, 2.5)a||1.3 (0.91, 1.8)||1.4 (1.0, 1.9)a||1.0 (0.75, 1.4)|
| 6-Month weight loss ≥20%||26 (4.5)||1.9 (1.2, 3.1)a||1.7 (1.0, 3.0)a||1.7 (1.2, 2.5)a||1.3 (0.92, 1.9)||1.5 (1.0, 2.1)a||1.2 (0.88, 1.6)|
Augmented Elixhauser comorbidities
Only 1.0% of patients had a diagnosis of weight loss recorded in the administrative data; however, 4.5% reported a weight loss ≥20% based on self-reported data. Weight loss derived using administrative data did not significantly predict survival. However, weight loss derived using the 20% cut point from the self-reported data was significant in both the adjusted and unadjusted Cox regression models. Approximately, 6.6% of patients had a self-reported BMI <20 kg/m2. This was a significant survival predictor in the unadjusted Cox regression model and tended toward significance in the adjusted model. Performance status was a significant survival predictor in the unadjusted and adjusted Cox regression models.
Robust Poisson Regression Model Discrimination
The c-statistics for the Elixhauser and Charlson comorbidity measures adjusted for age, sex, and stage for 2- and 3-year survival are reported in Table 4. A base model (age, sex, and stage) with no comorbidities already demonstrated excellent discrimination (ie, c-statistic, >0.8) for 2- and 3-year survival. Adding Charlson comorbidities as individual binary variables generated a model that was not different from the base model (2-year survival, P = .14; 3-year survival, P = .17). The addition of Elixhauser comorbidities to the base model, however, increased the c-statistic by 0.027-0.028 compared with the base model (2-year survival, P = .0051; 3-year survival, P = .0017). The c-statistic for the Elixhauser comorbidities was 0.021 higher than that for the Charlson comorbidities and significantly different (2-year survival, P = .018; 3-year survival, P = .016). The addition of the augmented Elixhauser variables to the base model also had higher discrimination than the base model alone (2-year survival, P = .0003; 3-year survival, P = .0001). The augmented Elixhauser method achieved a higher c-statistic than the standard Elixhauser method for 2-year survival (P = .026) but not for 3-year survival (P = .13). The c-statistics for models with either Charlson9 or Elixhauser36 weighted scores added to the base model were not significantly different from those of the base model alone. The 10-fold cross-validated c-statistics ranged from 0.804 to 0.839 for all models, including the base model and the Charlson and Elixhauser (normal and augmented) methods, for both 2- and 3-year survival.
Table 4. Elixhauser and Charlson Comorbidity Method Discrimination for 2- and 3-Year All-Cause Mortality in Stages II-IV Colorectal Cancer Patients
|Base model (age, sex, stage)||0.824*||0.827*|
As stage was associated with substantial discrimination power in this population (c-statistics: 2-year survival, 0.777; 3-year survival, 0.790), we tested the differences in comorbidity measure performance adjusted for age and sex separately by stage (Table 5). For all stages and survival times, the Elixhauser and augmented Elixhauser methods had significantly higher discrimination compared with the base model (ie, age, sex). The greatest increase in the c-statistic contributed by the addition of the Elixhauser comorbidities was for 2-year survival in stage II patients (increased from 0.683 to 0.838). Overall, the augmented Elixhauser had higher c-statistics than the standard Elixhauser method; however, this did not reach statistical significance except for 3-year survival in stage IV. Although c-statistics vary by stage, these findings were also true within stage, including a lack of discrimination by the Charlson model and statistically increased discrimination by the Elixhauser method.
Table 5. Elixhauser and Charlson Comorbidity Method Discrimination for 2- and 3-Year All-Cause Mortality by Stage in Colorectal Cancer Patients
|Base model (age, sex)||0.683*||0.664*||0.783*||0.692*||0.571*||0.585*|
As cancer is largely a disease of the elderly, who frequently possess multiple comorbidities, it is essential that the best measure is used to control for additional conditions that may affect survival. Emerging evidence, largely from noncancer populations, suggests that the Elixhauser method is a superior comorbidity risk-adjustment model compared with the Charlson method.11, 12 Despite these findings, the Elixhauser method has not been popular in cancer studies, perhaps because of few reported comparisons with the Charlson measure and concerns about the inclusion of too many explanatory variables (ie, 30+), thus requiring a fairly large sample size.
We found that the Elixhauser measure is a superior comorbidity risk-adjustment method with a significantly higher c-statistic for both 2- and 3-year survival compared with the Charlson measure (Table 4). Some reasons for this may be the lack of comprehensiveness of the Charlson comorbidities and that the Charlson measure was designed to predict 1-year survival. Survival studies with colorectal cancer patients are generally focused on 2- to 5-year survival. The study by Baldwin et al is the only published comparison of the Charlson and Elixhauser methods in cancer patients.14 A strength of that study is that it included both outpatient and inpatient administrative data; however, we believe our analysis better addresses the question of whether Elixhauser is a better comorbidity risk-adjustment model than is Charlson in cancer patients. Our population encompassed patients of all ages (ie, not just those ≥66 years) at different stages of disease (ie, stages II-IV, not just III), used different survival times to determine when comorbidities exert their influence, used all-cause mortality, and, most importantly, compared unbiased estimates (c-statistic) to identify the best prediction model. Although Baldwin et al examined whether the different comorbidity measures significantly add to a base model containing standard risk-adjustment variables including age, sex and race, they did not conduct statistical comparisons from which concrete conclusions could be drawn about which method is best.
Although differences in c-statistics are small, the finding that Elixhauser provides more comorbidities that significantly predict survival is clinically relevant (ie, 3 from Elixhauser vs 1 from Charlson using multivariate Cox regression). As observed elsewhere, patients with uncomplicated hypertension are likely to be healthier, which explains the positive impact on survival.10
We also tested whether these results were representative of a comorbidity index recently developed for Elixhauser.36 The resulting c-statistic was comparable to the model incorporating the Charlson binary variables. Various Charlson indices likewise performed poorly, as has been reflected by others.37
Overall, the c-statistics for all-cause mortality for the different comorbidity measures were similar to or slightly higher than values reported in other populations.11, 12, 24, 38 It will also be of interest to determine how the 2 approaches compare with other cancer types. It is important to do this analysis by cancer stage because stage by itself already exerts substantial survival discrimination. This analysis clearly revealed that the addition of comorbidities results in a larger improvement in survival prediction for stage II patients, and the magnitude of this effect decreased for stages III and IV. This would be expected, as stage II patients have substantially higher 5-year cancer-specific survival rates compared with stages III and IV patients and therefore are more likely to die from a condition other than their cancer than would a person with more advanced disease. The increased importance of comorbidity in less aggressive compared with more aggressive cancers has been previously suggested.39 Regardless of stage, however, the Elixhauser method was better than the Charlson method.
Despite the likely underreporting of weight loss and obesity in administrative data noted in this study and others,10, 24, 38, 40, 41 the replacement of administrative data with clinical weight data for obesity and weight-loss Elixhauser comorbidities did not improve the predictive performance of the method. Two conditions significantly related to survival (congestive heart failure and renal failure) are well-known causes of cachexia and wasting,27 and it may be that the addition of weight loss data adds little to an analysis in which these conditions are already accounted for. Underreporting of weight-related data in administrative data may be a result of poor recording of these features in medical records, the ICD codes available to identify weight-related conditions (ie, abnormal weight loss [ICD-10: R83.4]) having no specific defining cut points, and/or the lack of incentive to record this information.
Testing comorbidity measure performance in different cancer cohorts is worthy of future work. This will ensure that the best risk-adjustment model is used to test the relationships between emerging variables and cancer survival. The lack of significance of most comorbidities from the adjusted models shown in Table 3 suggests a good predictive model for survival may only need to include a few comorbidities. Now that we have established the value of the Elixhauser method, it can now be used to risk-adjust in future cancer studies evaluating the effects of various treatment modalities on survival.
Our study is the first to directly compare the Charlson and Elixhauser comorbidity measures for survival prediction in colorectal cancer patients. Choosing optimal risk-adjustment methods is essential to ensuring the report of unbiased relationships between an independent variable and the outcome of interest. Standardized and comprehensive risk-adjustment protocols to control for differences in baseline status will have important future applications in many domains of oncology and will allow for better comparison between studies.
We thank Lisa Martin and Charlotte King for assistance with the databases used for this study and Dr. Sunita Ghosh for statistical advice.
CONFLICT OF INTEREST DISCLOSURS
Supported by the Canadian Institutes of Health Research; Alberta Cancer Research Institute, Alberta Cancer Foundation, and Alberta Health Services.