Advances in glioblastoma care have resulted in a larger proportion of patients surviving beyond 2 years after diagnosis. It is not clear how long-term survivors should be counseled with respect to future prognosis, or what factors influence that prognosis. The conditional probability of survival was evaluated from multiple time points in patients with glioblastoma, using Surveillance, Epidemiology, and End Results (SEER) data.
Patients diagnosed with glioblastoma from 1998 to 2008 who were treated with radiation-containing regimens were identified within SEER data. Conditional survival probabilities from multiple survival points were calculated. Cox proportional hazards models were constructed to identify predictors of survival from diagnosis and from 1 and 2 years after diagnosis.
A total of 10,022 patients with glioblastoma met study inclusion criteria; median survival was 12.61 months. Conditional probability of surviving an additional 2 years ranged from 19.8% at diagnosis to 65.9% at 5 years after diagnosis. The proportion of patients surviving 12 months from time of diagnosis as well as from 6, 12, and 18 months after diagnosis was significantly higher in patients diagnosed in 2005 through 2008 than those diagnosed in 1998 through 2004. Of demographic and treatment-related factors evaluated, only age was associated with hazard of death at diagnosis and 1 and 3 years after diagnosis (P < .0001 at each time point).
Prognosis is an issue of great importance for patients with glioblastoma and physicians who counsel them. Discussions of prognosis most frequently concern survival from time of diagnosis, and a great deal of data from clinical trials1 and analyses of population-based data2 exist to inform these conversations. Conversely, relatively little data exist concerning the prognosis of patients who have survived a year or more after initial diagnosis. It is apparent in clinical practice that patients with glioblastoma who survive beyond 2 years represent a group with a relatively favorable prognosis for ongoing survival, but this observation has not been fully quantified.
Conditional probability of survival is the likelihood of surviving into the future based on previous survival to a specified time point. To calculate the conditional probability of surviving to time B from earlier time A, the proportion of patients in the initial cohort surviving to time B is divided by the proportion that survived to time A. Davis and colleagues published the first large-scale examination of conditional survival in glioblastoma by evaluating 7234 patients diagnosed between 1979 and 1993 who were identified in the Surveillance, Epidemiology, and End Results (SEER) database.3 They evaluated survival in patient groups defined by brain tumor type and age, but did not report the significance of any other demographic or treatment-related variables. In addition, that analysis included patients without microscopic confirmation of disease and patients in whom microscopic confirmation was obtained only after death. Polley and colleagues evaluated conditional survival in a smaller but better-characterized group of patients by pooling results from multiple single-arm phase 2 clinical trials.4 Analyses of clinical trial populations are advantageous in that extensive patient-specific clinical information is available, but the relatively small number of patients involved results in wide confidence intervals. Furthermore, selection factors associated with the collection of clinical trial populations may potentially lead to patient groups that are not fully representative of glioblastoma patients at large.4 We use the SEER data to evaluate conditional survival in the patients with glioblastoma diagnosed from 1998 to 2008 and treated with regimens including external beam radiation.
MATERIALS AND METHODS
Patient Identification and Selection
The SEER Program currently collects cancer incidence and survival data on 17 geographically defined cancer registries in the United States, which together cover approximately 28% of the US population. SEER data is rigorously quality-controlled both with respect to casefinding and the evaluation of the reliability of coding by central and hospital registry personnel.5 We requested and received permission from the National Cancer Institute to use the most recent SEER data set, released in April 2011.6 The SEER data was used to identify adult patients with glioblastoma, as indicated by International Classification of Disease for Oncology, third edition (ICD-O-3) codes 9440-9442, diagnosed from 1998 to 2008. Because changes to the SEER treatment coding system began that year, 1998 was chosen as the initial year of analysis. Patients included in survival analysis must have received microscopic confirmation of disease prior to death and must have been treated with external beam radiation therapy as part of their initial tumor-directed treatment, given the need for tissue diagnosis and the inclusion of external beam radiation in the standard-of-care treatment regimen for newly diagnosed glioblastoma throughout the analysis period. Patients with other cancers were excluded from analysis.
Outcome and Statistical Analysis
The primary outcome for this study was conditional survival from various time points after glioblastoma diagnosis. Conditional survival was calculated from observed (overall) survival rather than relative survival, which is the ratio of observed survival to the expected survival of a comparable group in the general population, because observed survival figures are more intuitive and more readily comparable to clinical trial results. Furthermore, the high mortality rate of glioblastoma makes the difference between these measures minimal, because death from non–tumor-related causes is relatively infrequent.
Conditional survival point estimates with 95% confidence intervals (CIs) for selected survival durations, beginning at diagnosis, 6 months after diagnosis, and at 1, 2, 3, 4, and 5 years after diagnosis were calculated in SEER*Stat.7 Additional conditional survival analysis was performed with the patient population divided into 2 groups by diagnosis period: 1998 to 2004 and 2005 to 2008. Because no information on chemotherapy use is captured in SEER, the division of diagnosis periods was chosen to reflect times before and after the widespread use of concurrent and adjuvant temozolomide chemotherapy as part of the standard treatment for newly diagnosed glioblastoma. Two-year conditional survival statistics from selected survival intervals were generated for patient subpopulations defined by the age groups 18 to 39 years, 40 to 59 years, 60 to 79 years, and ≥80 years.
Cox proportional hazards (PH) models were used to identify factors associated with survival from time of diagnosis and from 1 and 3 years after diagnosis via milestone analyses. Patient demographic characteristics of interest as predictors of conditional survival in Cox models included age, sex,8, 9 race,10, 11 and marital status,12 each of which has previously been shown to be associated with survival among patients with glioblastoma. Treatment-related factors of interest included gross total resection versus other surgery13, 14 and diagnosis period (1998-2004 or 2005-2008). Determination of gross total resection in SEER is derived from operative notes rather than being determined radiographically. An initial univariate Cox PH model evaluated age as 4 subgroups. Subsequent multivariable models included age group along with single demographic or treatment characteristics of interest. The PH assumption was evaluated individually for each demographic and treatment variable of interest via log-log plots.15 Factors significant at the P < .05 level in this analysis were carried forward into the initial multivariable model evaluating predictors of survival from time of diagnosis. A backward stepwise procedure was then performed until all factors remaining in the model were significant at the P ≤ .05 level. This same model was then used to evaluate survival from 1 and 3 years after diagnosis. No interaction terms were included in the models. Two-sided statistical tests were used. All statistical analysis of the SEER case listing data was performed in JMP, version 9.0 (SAS Institute, Cary, NC).
A total of 11,763 adults with microscopically-confirmed glioblastoma initially treated with external beam radiation were identified. After application of the exclusion criteria, 10,022 patients were included in the final analysis group. Patient characteristics are displayed in Table 1. Median survival time for the entire patient group was 12.61 months. A total of 8139 patients were confirmed to have died, and 1883 were censored at the time of analysis. The size of the analysis group diminished with increasing time from diagnosis, both due to patient loss and due to data maturity requirements for some analyses (ie, only patients diagnosed early in the 1998 to 2008 period had the potential to be 10-year survivors at time of analysis). Of the original group of 10,022 patients, 4684 were alive at 1 year, 1474 were alive at 2 years, 198 were alive at 5 years, and 8 were alive at 10 years.
Table 1. Patient Demographic, Tumor, and Treatment Characteristics
% (N = 10,022)
Median (interquartile range)
Extent of resection (1998+)
Gross total resection
Conditional survival from diagnosis, 6 months, and 1, 2, 3, 4, and 5 years after diagnosis is displayed in Table 2. The percentage of patients surviving for an additional 6 months or 1 year initially declined before ultimately increasing in long-term survivors. Figure 1 displays the conditional probability of surviving an additional 12 months at various time points by diagnosis period. A higher proportion of patients diagnosed from 2005 to 2008 survived 12 months from diagnosis and from 6, 12, and 18 months after diagnosis, with no significant difference seen in survival from 24 or 30 months. Table 3 displays the percentage of patients surviving an additional 2 years by age group and time since diagnosis.
Table 2. Conditional Probability of Survival for Additional Periods by Time Since Glioblastoma Diagnosis
Time (no. at risk)
Additional Survival Period
6 mo (7642)
1 y (4684)
2 y (1474)
3 y (660)
4 y (349)
5 y (198)
Table 3. Conditional Probability of Surviving an Additional 2 Years at Various Survival Time Points, by Age Group
NC indicates not enough intervals to produce statistic (not calculated).
In the initial univariate Cox PH model, greater age was highly associated with mortality. Relative to the 18- to 39-year-old subgroup, hazard ratio (HR) of death was 1.86 (95% CI, 1.69-2.05) in the 40- to 59-year-old subgroup, 3.25 (95% CI, 2.95-3.59) in the 60- to 79-year-old subgroup, and 6.44 (95% CI, 5.60-7.39) in the ≥80-year-old subgroup. In models including age group, demographic factors significantly associated with survival from diagnosis included sex (HR of death, 1.09; 95% CI, 1.04-1.14; P < .0001 for men relative to women), marital status (HR, 1.08; 95% CI, 1.03-1.14; P = .0010 for unmarried relative to married patients), and race (HR, 0.87; 95% CI, 0.78-0.98; P = .02) for Asian patients relative to white patients. No significant differences in survival were seen between white patients and any other racial/ethnic group. Treatment-related factors associated with survival included diagnosis period (HR, 1.36; 95% CI, 1.29-1.42; P < .0001 for patients diagnosed in 1998 to 2004 relative to those diagnosed in 2005 to 2008) and extent of resection (HR, 1.35; 95% CI, 1.29-1.41; P < .0001 for patients receiving anything other than gross total resection relative to those receiving gross total resection). Age group remained significant at the P < .0001 level in all models.
In the multivariable Cox PH model of survival from diagnosis, all factors that had been significant when analyzed with age alone remained significant in the combined model, with the exception of race/ethnicity. The significance of the survival advantage of Asian patients over white patients attenuated to a trend (P = .08), and race/ethnicity was dropped from the model. Table 4 displays the results of the final multivariable model for survival from diagnosis and at 1 and 3 years after diagnosis. Age is the only factor significantly associated with survival at each time point. Sex, extent of resection, and treatment period were significantly associated with survival at diagnosis and at 1 year, but not at 3 years, after diagnosis.
Table 4. Multivariable Cox Proportional Hazards Analysis of Factors Predictive of Survival From Diagnosis, 1 Year After Diagnosis, and 3 Years After Diagnosis
Time of Diagnosis
HR (95% CI)
HR (95% CI)
HR (95% CI)
CI indicates confidence interval; HR, hazard ratio; ref, reference value.
Extent of resection
Gross total resection
Analysis of conditional probability of survival offers additional perspective on survival in glioblastoma. In this study, we found that long-term survivors do not share the dire prognosis of patients who have newly diagnosed glioblastoma. Although only 6.2% of patients survived to 5 years, among those who did, the conditional probability of surviving an additional 5 years was almost 45%. Even at 2 years, patients had a 20% chance of living an additional 5 years. These data provide a population-based perspective on conditional survival among long-term glioblastoma survivors that may be useful in counseling this patient group.
A number of demographic and treatment factors were associated with survival in univariate and multivariable analysis. Age at diagnosis proved to be a statistically and clinically significant predictor of mortality at all time points. This result contrasts with that of Polley and colleagues, who saw a significant effect of age only when considering survival from time of diagnosis.4 However, their analysis of patients enrolled in clinical trials included Karnofsky performance status and progression status at 1 year in the multivariable model. These data are not available in the SEER data set, and it is possible that Karnofsky performance status and/or likelihood of progression are associated with age, thus diluting the effect of age when all factors are evaluated together. Although the association between age and outcome in glioblastoma is well known, the nature of the association is unclear. Older patients may have more comorbidity or receive less aggressive care than younger patients; however, some data suggest that differences in survival based on patient age are due more to biological factors than clinical factors.16 In the present analysis, marital status was associated with survival from time of diagnosis, but not with survival at 1 year or 3 years after diagnosis. The role of marital status in glioblastoma survival is not clear, but studies in other tumor types suggest that married patients may present for care sooner and have stronger social networks, which is presumably important in providing care throughout the course of disease. Female sex was a positive prognostic factor at diagnosis and at 1 year, but had no significant effect at 3 years. The prognostic implications of sex on glioblastoma survival have been previously reported and are not well understood.8, 9 The statistical significance of demographic characteristics other than age at diagnosis, but not at later time points in the study population, may indicate that they influence quality of care or the patient's support networks, but do not fundamentally impact the course of disease.
Two treatment-related factors, extent of resection and diagnosis period, were prognostic of survival, both from diagnosis and from 1 year after diagnosis. The positive prognostic implications of gross total resection are well recognized. Notably, the use of operative notes rather than imaging to determine extent of resection in SEER may lead to overestimation of gross total resection rates by including some patients who truly underwent subtotal resection in the gross total resection group. To the extent that this affects our analysis, it would be expected to minimize the apparent survival benefit of gross total resection. The influence of diagnosis in or after 2005 on survival in the multivariable Cox model is also unsurprising, given that patients treated in this period were likely to receive treatment with concurrent and adjuvant temozolomide chemotherapy, which has been proven to increase median survival time and proportional survival at time points up to 5 years after diagnosis.17 It is likely that factors other than temozolomide have also contributed to recent increases in glioblastoma survival, as demonstrated by an analysis of multiple recent trials in which median survival was longer than in the temozolomide-containing arm of the phase 3 temozolomide trial.18 Although differences in study size and patient selection between trials make direct survival comparisons difficult, there are multiple plausible explanations for a true recent increase in overall survival, including novel agents (eg, bevacizumab) and changes in the pattern of care (eg, additional resection or reirradiation at recurrence). It could be hypothesized that advances in treatment of newly diagnosed glioblastoma, such as concurrent temozolomide or improved surgical interventions, underlie the superior 12-month survival from time of diagnosis, 6 months, and 1 year after diagnosis in 2005 through 2008 relative to 1998 through 2004 (Fig. 1), whereas improvements in treatment of recurrent glioblastoma are largely responsible for the superior 12-month survival from 18 months after diagnosis. The lack of improvement in 12-month conditional survival from 24 or 30 months after diagnosis between 1998 and 2004 and 2005 and 2008 suggests that true long-term survival is more a product of tumor biology than specific treatment regimen, at least with respect to currently available treatment options.
Although the use of SEER allows the analysis of survival in a much larger group of patients than would be possible by other techniques, this approach has known limitations. The most significant limitation is that diagnoses are assigned at individual treatment sites without central review by trained neuropathologists. It is likely that some long-term glioblastoma survivors identified in SEER may truly have had other diagnoses, such as anaplastic astrocytoma or anaplastic oligodendroglioma, which are associated with longer survival.19 Although the significance of pathologic misclassification in these data cannot be determined, the prevalence of misclassification would not be expected to exceed that seen in routine practice. The entire time period analyzed in this report occurred after the widespread recognition of the importance of oligodendroglial features in high-grade glioma and the corresponding increase in oligodendroglioma diagnosis rates.20 Thus, misclassification of high-grade oligodendroglioma or oligoastrocytoma as glioblastoma is likely to be less of an issue in the present analysis than it would have been at earlier times. Finally, our point estimates of survival up to 4 years are quite similar to those reported by Polley and colleagues in their analysis of 7 pooled glioblastoma clinical trials, suggesting that the survival figures that form the basis for conditional survival calculation in our analysis are not artificially optimistic due to patient misdiagnosis.4
An additional issue of importance when working with a cancer registry is proper patient cohort selection. It is essential to choose a patient subpopulation comparable to the larger patient group onto whom the results will be applied. Because glioblastoma is a World Health Organization grade IV tumor that can arise from lower-grade tumors, we chose to limit our analysis to patients who had tumor-directed surgery as part of their initial course of treatment, to avoid inadvertently classifying survival time from a predecessor lower-grade tumor as glioblastoma survival time. Furthermore, we limited analysis to patients who had external beam radiation therapy as part of their initial tumor-directed treatment, as external beam radiation was, and remains, part of the standard of care for glioblastoma. Although both of these decisions result in restriction of the patient population analyzed, they create a patient group representative of the vast majority of glioblastoma patients seen in routine practice or enrolled in clinical trials. Finally, cancer registries do not contain all of the clinical parameters that would be ideal, and the data available cannot be independently confirmed for accuracy. In this analysis, it would be beneficial to be able to evaluate performance status and presence or absence of tumor progression at each interval. These factors must be taken into account on a patient-by-patient basis using the best judgment of the clinician, and the prognostic information quoted in this report likely applies best to patients in the midst of some significant period of clinical stability
In summary, whereas the overall mortality rate of glioblastoma is extremely high, analysis of conditional survival may present cause for optimism when counseling survivors. This contemporary population-based analysis confirms and extends the findings of previous analyses based on clinical trial data and older population-based approaches. As data from ongoing and recently completed large phase 3 trials mature in the future, it may be possible to more fully evaluate patient-specific clinical and pathological factors to produce individualized estimates of conditional survival and better describe the functional state of long-term survivors treated with current therapies.