Adult glioma incidence trends in the United States, 1977–2000

Abstract

BACKGROUND

Several authors have reported an increase in the incidence of brain tumors, especially among the elderly. A more complete understanding of adult glioma incidence trends might provide indications of risk factors for gliomas and contribute to the search for improved therapies.

METHODS

The authors used the Surveillance, Epidemiology, and End Results (SEER) registry public use data tapes, which included data on patients with cancer diagnosed between 1973 and 2000. For 3 histologies as well as for 12 histology categories combined, the authors used Poisson regression to model incidence as a function of year of diagnosis, age at diagnosis, race (white or African American), and gender. They used cubic splines to fit age at diagnosis and year of diagnosis and tested for all pair-wise interactions.

RESULTS

The interaction between year of diagnosis and age at diagnosis was significant in all four groups modeled. In glioblastoma, there was also a significant interaction between gender and age at diagnosis. In anaplastic astrocytoma, there was a significant interaction between gender and year of diagnosis. In oligodendroglioma, there was a significant interaction between race and gender. In the 12 histology categories combined, there was a significant interaction between gender and age at diagnosis.

CONCLUSIONS

The results in the current study were consistent with other published reports that showed an increase in the incidence of brain tumors using SEER data. Although others have observed increasing incidence trends among the elderly, the authors formally tested and found a statistically significant interaction between age at diagnosis and year of diagnosis. Cancer 2004. © 2004 American Cancer Society.

The incidence rate of central nervous system (CNS) tumors in 2000 was 6.7 per 100,000 persons as reported from the Surveillance, Epidemiology, and End Results (SEER) registry1 and gliomas account for approximately 51% of all CNS tumors.2 Several authors have reported an increase in the incidence of brain tumors3–8 and many have attributed the observed changes to developments in diagnostic imaging or changes in the classification system. However, other factors may explain at least part of these changes.2 Little is known about the risk factors for brain tumors9 and, more specifically, gliomas. The article by MacDonald10 disclosed a failure of current treatments to improve survival in patients with malignant gliomas. Perhaps, a better understanding of the distribution of glioma cases could provide indications of etiologic factors and contribute to the search for improved therapies. We used data from SEER to present estimates of glioma incidence by year of diagnosis, age at diagnosis, gender, and race. We modeled the incidence trends over year of diagnosis and tested for interaction effects to develop a more complete understanding of how glioma incidence has changed over the past several decades.

MATERIALS AND METHODS

We used the SEER registry public use data tapes provided from nine registries (Atlanta, Connecticut, Detroit, Hawaii, Iowa, New Mexico, San Francisco–Oakland, Seattle–Puget Sound, and Utah), which included data on patients with cancer diagnosed between 1973 and 2000. SEER collects information on cancer incidence and survival and also provides information on patient demographics, primary tumor site, morphology, and stage at diagnosis.11 We excluded data collected before 1977 from our analysis because of changes in the coding system. Incident cases diagnosed between 1973 and 1976 were classified according to the Manual of Tumor Nomenclature, incident cases diagnosed between 1977 and 1991 were classified according to the International Classification of Diseases for Oncology (ICD-O), and incident cases diagnosed between after 1992 were classified according to the the second revision of the International Classification of Diseases for Oncology (ICD-O2).3, 6, 12–14 Many cases reported using the Manual of Tumor Nomenclature lacked the detail necessary to be updated to an ICD-O2 code.3 We included only white and African-American subjects because the number of subjects from other racial groups was too small for comparative analysis. We excluded patients diagnosed after age 85 because ages > 85 are combined, thus preventing accurate age-specific incidence rate calculations. We excluded patients < 20 years of age because pediatric brain tumors are distinct from adult brain tumors.15

Twelve histology categories of adult glioma were included in our analyses: glioma not otherwise specified (NOS; SEER code 9380); mixed oligoastrocytoma (SEER code 9382); astrocytoma NOS (SEER code 9400); anaplastic astrocytoma (AA; SEER code 9401); protoplasmic astrocytoma (SEER code 9410); gemistocytic astrocytoma (SEER code 9411); fibrillary astrocytoma (SEER code 9420); oligodendroglioma (OL) NOS (SEER code 9450); anaplastic oligodendroglioma (AO; SEER code 9451); glioblastoma (GBM) NOS (SEER code 9440); giant cell glioblastoma (SEER code 9441); and gliosarcoma (SEER code 9442).

We first assessed changes in unadjusted incidence over year of diagnosis by fitting linear models to the unadjusted log incidence that included year of diagnosis as the only covariate. This was done separately for each of the 12 histology categories.

We next used Poisson regression to model incidence as a function of year of diagnosis, age at diagnosis, race (white or African American), and gender for three of the histology categories separately and for all 12 histology categories combined (all combined). GBM, AA, and OL histology categories were selected for this second analysis because they had a sufficient number of cases (> 1000) and because they tended to be more homogenous than the other large histology category. Astrocytoma NOS had a sufficient number of cases, but this group appeared to have age at diagnosis and survival characteristics similar to the AA and GBM cases. Due to the ambiguous composition of this category, we believed inference on this group would be problematic and omitted it from this analysis. GBM and OL are also NOS categories and may also have an ambiguous composition. however, these two groups had more homogenous age and survival profiles. In addition to assessing the main effects of year of diagnosis, age at diagnosis, race, and gender, we also tested for all pairwise interactions. Based on plots of the observed incidence versus age at diagnosis and versus year of diagnosis as well as from plots of the residuals from the first linear models, we decided to use cubic splines16 for age at diagnosis and year of diagnosis in the Poisson regression models. We used the Akaike Infamation Criterion (AIC) statistic17 to compare models with differing numbers of knots and found that four knots for age at diagnosis and one knot for year of diagnosis consistently had the lowest AIC. Knots were placed at the quintiles of age at diagnosis and the median of year of diagnosis. To facilitate presentation and interpretation, we selected the same set of main effects and interactions for the three histology categories and for all combined. Interactions between age at diagnosis and race and between year of diagnosis and race were not significant for any of the three histology categories or for all combined. Therefore, the final model included main effects for year of diagnosis, age at diagnosis, race, and gender and interaction terms between race and gender, between gender and age at diagnosis, between gender and year of diagnosis, and between age at diagnosis and year of diagnosis. We used graphic methods, residual plots, and plots of the predicted incidence and the observed incidence to assess the fit of our model. Multiple statistical tests were performed and we used a type I error rate of 1% to determine statistical significance.

RESULTS

Table 1 summarizes the demographic information by histology category and for all combined. Patients diagnosed with gliosarcoma tended to be the oldest, and patients diagnosed with protoplasmic astrocytoma tended to be the youngest. Approximately 95% of the patients in these groups were white, and 52–62% were male.

Table 1. Population Description

Histologic category	No. of cases	Median age (yrs)	Median survival (mo)	White (%)	African-American (%)	Male (%)	Female (%)
NOS: not otherwise specified; AA: anaplastic astrocytoma; AO: anaplastic oligodendroglioma; GBM: glioblastoma.
Glioma, NOS (9380)	648	58	10	92.4	7.6	60.6	39.4
Mixed oligoastrocytoma (9382)	593	42	59	96.6	3.4	56.0	44.0
Astrocytoma, NOS (9400)	5202	55	12	94.0	6.0	56.8	43.2
AA (9401)	1549	51	14	96.3	3.7	55.3	44.7
Astrocytoma, protoplasmic (9410)	59	34	92	93.2	6.8	52.5	47.5
Astrocytoma, gemistocytic (9411)	355	50	19	96.9	3.1	62.5	37.5
Astrocytoma, fibrillary (9420)	463	47	23	95.9	4.1	55.7	44.3
Oligodendroglioma (9450)	1356	42	105	96.5	3.5	57.7	42.3
AO (9451)	248	48.5	36	94.8	5.2	57.7	42.3
GBM (9440)	11,667	63	7	95.9	4.1	57.6	42.4
Giant cell GBM (9441)	94	56.5	9	93.6	6.4	57.4	42.6
Gliosarcoma (9442)	193	64	7	93.8	6.2	62.7	37.3
Total	22,427	59	10	95.4	4.6	57.4	42.6

Table 2 shows the results from fitting linear models to assess changes in the unadjusted log incidence values over year of diagnosis. Eight histology categories had significant increases in incidence. The AO histology category was estimated to have the fastest increases in incidence, i.e., 17.26% per year. Two histology categories had significant decreases in incidence. Astrocytoma NOS had the fastest estimated decrease, i.e., 7.84% per year. Figure 1 shows the observed incidence, the incidence predicted from the linear model, and a smoothing line18 by year of diagnosis. Except for the protoplasmic astrocytoma histology category, the linear model fits the observed data well.

Table 2. Results from Linear Model Fit to Unadjusted Log Incidence Data

Histologic category	Estimated change per Year of diagnosis (%)	P value
NOS: not otherwise specified; AA: anaplastic astrocytoma; AO: anaplastic oligodendroglioma; GBM: glioblastoma.
Glioma, NOS (9380)	0.63	0.43
Mixed oligoastrocytoma (9382)	4.06	0.0001
Astrocytoma, NOS (9400)	−7.84	< 0.0001
AA (9401)	5.53	< 0.0001
Astrocytoma, protoplasmic (9410)	−4.48	0.04
Astrocytoma, gemistocytic (9411)	−3.11	0.003
Astrocytoma, fibrillary (9420)	2.60	0.03
Oligodendroglioma (9450)	6.11	< 0.0001
AO (9451)	17.26	< 0.0001
GBM (9440)	2.38	< 0.0001
Giant cell GBM (9441)	−2.20	0.18
Gliosarcoma (9442)	8.56	< 0.0001

Figure 1. Observed incidence and predicted incidence from the linear model on unadjusted log incidence.

Figures 2 and 3 show the predicted incidence from Poisson regression models plotted by gender and race over age at diagnosis for 3 years of diagnosis (1980,1990, and 2000) and plotted over year of diagnosis for 3 age at diagnosis groups (30–34, 50–54, 70–74), respectively. Figures 2 and 3 show GBM, AA, OL, and all 12 histologies combined (ALL), each analyzed separately. We present only the incidence predicted from the model because these estimates have better precision than the observed rates and are, thus, more stable. Residual analyses and plots of observed and predicted values indicated that these models fit the data well. These plots show that whites tended to have a higher incidence than African Americans and males tended to have a higher incidence than females. More specifically, white males had the highest incidence followed by white females, African-American males, and African-American females.

Figure 2. Estimated incidence (per 100,000 person-years) from fitted Poisson regression models by age, race, and gender for the years 1980, 1990, and 2000. ALL: all 12 histologies combined; OL: oligodendroglioma; AA: anaplastic astrocytoma; GBM: glioblastoma.

Figure 3. Estimated incidence (per 100,000 person-years) from fitted Poisson regression models by year of diagnosis, race, and gender for the age groups 30–34, 50–54, and 70–74. ALL: all 12 histologies combined; OL: oligodendroglioma; AA: anaplastic astrocytoma; GBM: glioblastoma.

Figure 2 shows that the incidence was higher in each selected year of diagnosis for GBM, AA, OL, and for all combined. The incidence tended to increase in GBM, AA, and all combined until approximately an age at diagnosis of 70 and then tended to decrease. However, OL tended to have a higher incidence for the younger ages and a lower incidence for the older ages. Figure 2 also shows that for GBM, AA, and all combined, the incidence increased more slowly over age at diagnosis in 1980 and 1990 than in 2000. For OL, the incidence decreased faster over age at diagnosis in 1980 and 1990 than in 2000.

Figure 3 shows that for GBM, the incidence began to increase in 2000 with the fastest increases in the oldest ages. For AA, the incidence reached a maximum in the early 1990s for the two younger age at diagnosis groups and started to decline in 2000. However, the incidence of AA began to increase in 2000 for the oldest age at diagnosis group. For OL, the incidence began to increase in 2000 with the fastest increases in the younger age at diagnosis groups. In all combined, the incidence was fairly constant across year of diagnosis in the two younger age groups, and in the older age group reached a peak in the early 1990s and appears more level after 1990.

Figures 2 and 3 indicate an interaction between year of diagnosis and age at diagnosis. However, other interactions may be less evident. Table 3 shows the P values of all pairwise interactions in each model, as well as the interactions that were significant in GBM, AA, OL, and in all combined. The interaction between year of diagnosis and age at diagnosis was significant in all four groups. In GBM, there was also a significant interaction between gender and age at diagnosis. In AA, there was a significant interaction between gender and year of diagnosis. In OL, there was a significant interaction between race and gender. In all 12 histology categories combined, there was a significant interaction between gender and age at diagnosis.

Table 3. P values for Interaction Terms in Poisson Regression

Interaction term	Histologic category
	GBM	AA	OL	ALL
GBM: glioblastoma; AA: anaplastic astrocytoma; OL: oligodendroglioma; ALL: all 12 histologies combined.
Gender–race	0.82	0.080	0.0034	0.19
Gender–age	0.0009	0.45	0.33	0.0008
Gender–year	0.28	0.0062	0.80	0.49
Age–year	< 0.0001	0.018	0.0016	< 0.0001

Figures 4 and 5 show the ratio of incidence rates of males and females by year for 3 age at diagnosis categories and by age at diagnosis for 3 years, respectively. For the GBM category, the ratio is variable over the range of ages, indicating the gender and age at diagnosis interaction. For the AA category, the ratio increases over year of diagnosis, indicating the gender and year of diagnosis interaction. For OL, the ratio is much higher for African Americans than for whites, indicating the gender and race interaction. In all combined, similar to GBM, the ratio tends to be higher for the older ages at diagnosis, indicating the gender and age at diagnosis interaction

Figure 4. Male-to-female incidence ratios from Poisson regression models by race and year of diagnosis for the age groups 30–34, 50–54, and 70–74. ALL: all 12 histologies combined; OL: oligodendroglioma; AA: anaplastic astrocytoma; GBM: glioblastoma.

Figure 5. Male-to-female incidence ratios from Poisson regression model by race and age for years of diagnosis 1980, 1990, and 2000. ALL: all 12 histologies combined; OL: oligodendroglioma; AA: anaplastic astrocytoma; GBM: glioblastoma.

DISCUSSION

The current study presents model-based estimates of incidence that are more precise and stable than the observed incidence rates from SEER. Our analyses effectively pool information across the available data to compensate for groups with few observations. We found that the incidence of gliomas increased between 1977 and 2000, especially among the oldest patients included in our analyses. Previous reports have also found an increasing incidence of brain tumors in the elderly and have attributed the increases, at least in part, to improvements in diagnostic modalities, specifically the introduction of computed tomography (CT) scans in the 1970s andthe introduction magnetic resonance imaging scans in the 1980s.3–5, 7, 19 Our analysis shows, as does the report by Davis and McCarthy,20 that brain cancer incidence may be stabilizing. In addition, this suggests that glioma incidence increases may be artifactual.

Our analyses also formally assessed the statistical significance for all pairwise interactions between age at diagnosis, year of diagnosis, race, and gender. Although other reports have described interaction effects while discussing their results, these effects are typically not formally modeled or tested. We consistently found a significant interaction between year of diagnosis and age at diagnosis, as differences in incidence rates between the young and old were typically greater in the later years of diagnosis than in the earlier years. Legler et al.3 examined imaging procedures from Medicare part B participants and found an increasing use of CT scans in older subgroups. This finding, coupled with their observation that physicians are increasingly more willing to use aggressive procedures with older patients, could explain why changes in incidence over time are most dramatic in the oldest ages.

Also, for all gliomas combined, as well as in the GBM category, we found a significant interaction between gender and age at diagnosis. This finding is consistent with other reports in the literature. Greig et al.,6 who studied all malignant brain tumors, and Lonn et al.,16 who investigated gliomas, both reported that incidence rates were greater in men than in women, but that the gender differences were greater in the older age groups.

Other interactions were also found in the AA and OL histology categories. In the AA category, there was a significant interaction between gender and year at diagnosis. In the OL group, there was a significant interaction between gender and race. However, making inference about incidence rates over time, particularly within specific histology categories, is fraught with peril. Wrensch et al.9 noted that incidence trends are sensitive to changes in diagnostic procedures, the availability of neurosurgeons, access to medical care, and changing attitudes towards the elderly. In addition, incidence trends are only interpretable if they are based on consistent definitions throughout the time period and yet there are many reporting and classification difficulties for malignant gliomas.9, 20 The improvement of diagnostic imaging may have caused many gliomas that would have been classified as NOS in earlier years to be placed in more specific histology categories in the later years.5, 20 Our analysis showed large increases in the specific histology categories, and much more modest increases in all gliomas. Finally, SEER data are not necessarily representative of the entire U.S. population and trends noted in the SEER data are not necessarily reflective of trends in the population as a whole.

In summary, our study is consistent with other reports published showing an increase in the incidence of brain tumors using SEER data. However, although others have only observed and described interaction effects, we formally tested the hypothesis that the increase occurred among the elderly by showing that there is a significant interaction between age at diagnosis and year of diagnosis. Also, by focusing on the estimates of incidence from our model, we were able to show more stable and precise estimates of glioma incidence rates over time than would be possible from studying observed incidence rates.