Demographic variation in incidence of adult glioma by subtype, United States, 1992-2007

Background We hypothesized that race/ethnic group, sex, age, and/or calendar period variation in adult glioma incidence differs between the two broad subtypes of glioblastoma (GBM) and non-GBM. Primary GBM, which constitute 90-95% of GBM, differ from non-GBM with respect to a number of molecular characteristics, providing a molecular rationale for these two broad glioma subtypes. Methods We utilized data from the Surveillance, Epidemiology, and End Results Program for 1992-2007, ages 30-69 years. We compared 15,088 GBM cases with 9,252 non-GBM cases. We used Poisson regression to calculate adjusted rate ratios and 95% confidence intervals. Results The GBM incidence rate increased proportionally with the 4th power of age, whereas the non-GBM rate increased proportionally with the square root of age. For each subtype, compared to non-Hispanic Whites, the incidence rate among Blacks, Asians/Pacific Islanders, and American Indians/Alaskan Natives was substantially lower (one-fourth to one-half for GBM; about two-fifths for non-GBM). Secondary to this primary effect, race/ethnic group variation in incidence was significantly less for non-GBM than for GBM. For each subtype, the incidence rate was higher for males than for females, with the male/female rate ratio being significantly higher for GBM (1.6) than for non-GBM (1.4). We observed significant calendar period trends of increasing incidence for GBM and decreasing incidence for non-GBM. For the two subtypes combined, we observed a 3% decrease in incidence between 1992-1995 and 2004-2007. Conclusions The substantial difference in age effect between GBM and non-GBM suggests a fundamental difference in the genesis of primary GBM (the driver of GBM incidence) versus non-GBM. However, the commonalities between GBM and non-GBM with respect to race/ethnic group and sex variation, more notable than the somewhat subtle, albeit statistically significant, differences, suggest that within the context of a fundamental difference, some aspects of the complex process of gliomagenesis are shared by these subtypes as well. The increasing calendar period trend of GBM incidence coupled with the decreasing trend of non-GBM incidence may at least partly be due to a secular trend in diagnostic fashion, as opposed to real changes in incidence of these subtypes.


Background
In a previous descriptive study of international variation in the incidence of adult primary brain cancer, we found that: 1) White populations had the highest incidence rates, North American Blacks had intermediate rates, and populations of eastern and southeastern Asian origin had the lowest rates; 2) incidence rates increased less steeply with age in the latter populations; 3) the male/female incidence rate ratio was between 1.4 and 1.5 among populations throughout the world; and 4) the male/female rate ratio was higher for peri-and postmenopausal ages than for pre-menopausal ages [1]. The lower incidence rate among non-Whites and the male excess had previously been noted by other investigators [2][3][4][5][6].
Although we examined brain cancer as a whole (due to limitations of the data we were unable to examine variation by morphologic type), our results were most likely driven by variation in the incidence of glioma, the predominant morphologic type of adult primary malignant brain tumors [2,6]. However, because gliomas themselves are heterogeneous, we now hypothesize that race/ethnic group, sex, and/or age variation in adult glioma incidence differs by morphologic subtype. To test these hypotheses, we utilized data from the population-based cancer registries of the National Cancer Institute's Surveillance, Epidemiology, and End Results (SEER) Program. We also tested the hypothesis that calendar period trends in adult glioma incidence varied by subtype.
The three main categories of adult glioma according to traditional pathological classification are astrocytoma, oligodendroglioma, and mixed oligoastrocytoma [7][8][9]. Each of these categories is further sub-classified by grade (grade II, grade III, or grade IV, the latter grade reserved for glioblastoma [GBM], which is grade IV astrocytoma) [8,9]. Secondary GBM develops through progression from a lower grade astrocytoma, whereas primary GBM appears to arise de novo, with no evidence of a lower-grade precursor [9].
There is considerable variability in the diagnosis of the traditional glioma categories and grades among pathologists, across geographic regions, and over time [10]. Nevertheless, epidemiologists and neuropathologists have reached a consensus that a pathology report of GBM is likely to be valid (although a small proportion of GBMs may be incorrectly classified as grade III [anaplastic] astrocytomas), whereas valid classification of non-GBM subtypes requires centralized pathologic review [10]. Thus, adult gliomas can validly be categorized broadly as GBM versus non-GBM based on pathology report diagnoses (which are utilized by SEER cancer registries), but non-GBM cannot be sub-divided further without compromising validity. Primary GBM, which constitute 90-95% of GBM [9], differ from non-GBM with respect to DNA copy number [11], gene expression [11], DNA methylation [12][13][14], and isocitrate dehydrogenase (IDH) 1 or 2 gene mutation status [11][12][13][14][15][16], providing a molecular rationale for the two broad glioma subtypes of GBM and non-GBM.
For valid classification of adult glioma as GBM versus non-GBM, we made two necessary exclusions. First, we excluded cases with the non-specific code 9380 because such cases could not be classified as GBM versus non-GBM. Second, we excluded cases that were not microscopically confirmed because valid distinction between GBM and non-GBM requires microscopic examination of tumor tissue and cannot be accomplished by imaging alone [19][20][21].
Given these exclusions, we sought to minimize selection bias due to differences in the proportion of cases excluded according to age, calendar period, sex, or race/ ethnic group. First, we required that the proportion of cases that were both microscopically confirmed and had a specific morphology code (a code other than 9380) be greater than 90% in each included 5-year age group, resulting in restriction of the age range to 30-69 years (Table 1). Overall, in this age range 92.3% of cases were both microscopically confirmed and had a specific morphology code. Then, we examined this proportion within this age range according to period, sex, and race/ ethnic group and found that the proportion did not vary meaningfully by these characteristics, although it was slightly lower among Blacks (89.8%) and Asians/Pacific Islanders (89.9%) than among other race/ethnic groups (Table 1). Overall, in the age range 30-69 years, 4.6% of cases were excluded due to having morphology code 9380 and 3.1% were excluded due to lack of microscopic confirmation (3.8% of GBM cases and 2.4% of non-GBM cases). We included a total of 15,088 GBM cases and 9,252 non-GBM cases in our analyses.

Data analysis and statistical methods
We downloaded sex-, age-, race-, Hispanic-ethnicity-, and calendar-year-specific numbers of cases and personyears at risk from the SEER website. We classified age into eight five-year age groups ranging from 30-34 to 65-69. We divided calendar-years into four four-year periods (1992-1995, 1996-1999, 2000-2003, and 2004-2007). Our race/ethnic group categories were non-Hispanic White, Hispanic White, Black, Asian/Pacific Islander, and American Indian/Alaskan Native.
For each of our two glioma subtypes, we calculated sex-and age-specific incidence rates (for each race/ethnicity group) by dividing the number of cases by the number of person-years at risk. We calculated sex-specific, age-standardized incidence rates (ASRs) and 95% confidence intervals (CIs) (for each race/ethnicity group) by direct standardization of the age-specific incidence rates to the 2000 U.S. Standard Population (Census P25-1130). We used Poisson regression to calculate adjusted rate ratios (RRs) and 95% CIs. We modeled age as a categorical variable or as the natural logarithm of age (log e (age)), where age was defined as the midpoint of the five-year age group. We modeled calendar period as a categorical or interval variable, the latter being used to calculate the p-value for period trend. Results for models with and without adjustment for cancer registry did not meaningfully differ; we only present the unadjusted results. To determine a p-value for heterogeneity, we entered the appropriate cross-product term into the Poisson model and conducted a likelihood ratio test for its addition, with the appropriate degrees of freedom. To further explore the relationship between glioma subtype incidence and age, we calculated adjusted RRs for 5-year age groups using Poisson regression and used the calculated RRs to perform weighted linear least squares regression of log 10 (RR) versus log 10 (age). We also performed weighted linear least squares regression of log 10 (age-specific incidence rate) versus log 10 (age). We performed Poisson regressions using Proc Genmod of SAS version 9.1; statistical tests were two-sided with α = 0.05.

Race/ethnic group
In Table 2, we tested the hypothesis that race/ethnic group variation in incidence differs according to glioma subtype. The reference group for the RRs was non-Hispanic Whites. Except for American Indians/Alaskan Natives, 95% CIs for ASRs and RRs were relatively tight due to large numbers of cases. For each subtype, ASRs and RRs varied by race/ethnic group similarly for males and females. For GBM, rates varied about three-to-fourfold, with highest rates among non-Hispanic Whites, followed by Hispanic Whites, Blacks, Asian/Pacific Islanders, and American Indians/Alaskan Natives. As with GBM, rates for non-GBM were highest among non-Hispanic Whites, followed by Hispanic Whites, with rates among Blacks, Asian/Pacific Islanders, and American Indians/Alaskan Natives each about 40% of the rates among non-Hispanic Whites. The main differences in race/ethnic variation between GBM and non-GBM were that Blacks had lower RRs, and Asian/Pacific Islanders and American Indians/Alaskan Natives had higher RRs, for non-GBM than for GBM. Overall, race/ethnic group variation was less for non-GBM than for GBM. The difference in race/ethnic variation between GBM and non-GBM was highly significant among both males and females (p-value for heterogeneity < 0.0001 in each sex).

Sex
In Table 2, we also tested the hypothesis that sex variation in incidence differs according to glioma subtype. For each race/ethnic group and glioma subtype, males had a higher incidence rate than females. For each race/ ethnic group, the male/female RR was higher for GBM The p-value for heterogeneity for the variation in male/female RR by race/ethnic group was 0.0013 for GBM and 0.0011 for non-GBM, indicating that the male/female RR varied by race/ethnic group for each glioma subtype. However, when we excluded Hispanic Whites, among whom we observed the lowest male/ female RR for both GBM and non-GBM, the p-value for heterogeneity was no longer significant (p = 0.76 for GBM and p = 0.85 for non-GBM).
In Table 3 we tested the hypothesis that the male/ female RR is higher for peri-and post-menopausal ages than for pre-menopausal ages. The male/female RR did not significantly vary among these menopausal age groups (p-value for heterogeneity = 0.38 for GBM and 0.087 for non-GBM). Because we observed in our previous study that the male/female RR was higher for peri-and post-menopausal ages than for pre-menopausal ages for adult brain cancer as a whole [1], most of which is glioma, in the current study we also examined GBM and non-GBM combined (data not shown) and did observe a borderline-significantly higher male/female RR for peri-and post-menopausal ages than for premenopausal ages (p-value for heterogeneity = 0.047).

Age
In Poisson models including age (categorical variable with 5-year age groups), sex, and a cross-product term between age and sex, adjusted for period and race/ethnic group, the p-value for the cross-product term was 0.37 for GBM and 0.12 for non-GBM, indicating homogeneous age effects in males and females for GBM and non-GBM, respectively. However, the difference in age effect between GBM and non-GBM was highly significant (p-value for heterogeneity < 0.0001).
To further explore the hypothesis that age variation in incidence differs according to glioma subtype, for each subtype we calculated adjusted RRs for age by 5-year age group using Poisson regression, adjusting for period, race/ethnic group, and sex. We then used these adjusted RRs in weighted least squares regressions of log 10 (RR) versus log 10 (age) and found reasonable fits, with the GBM incidence rate increasing proportionally with approximately the 4 th power of age (slope = 4.17; 95% CI: 3.80, 4.54; adjusted-R 2 = 0.99), and the non-GBM incidence rate increasing proportionally with approximately the square root of age (slope = 0.48; 95% CI = 0.25, 0.72; adjusted-R 2 = 0.78), showing the marked difference in age effect between GBM and non-GBM. We then calculated unadjusted RRs by 5-year age group and subtype using Poisson regression and used the unadjusted RRs in weighted least squares regression of log 10 (RR) versus log 10 (age). We found slopes and adjusted-R 2 s similar to those found when using adjusted RRs, showing that it would be valid to plot log 10 (age-specific incidence rate) versus log 10 (age) without adjustment, which we show in Figure 1.
Finally, we modeled age as log e (age) in Poisson models (Table 4) and found the period-, sex-, and race/ethnicgroup-adjusted regression coefficients (intersections of "Total" row and "Total" column) to be identical to the corresponding slopes observed in the weighted least squares regression models that used the adjusted RRs for 5-year age groups. The p-value for heterogeneity for the sex difference in the log e (age) effect was 0.33 for GBM and 0.18 for non-GBM, confirming the sex homogeneity; the p-value for heterogeneity for the difference in log e (age) effect between GBM and non-GBM was highly significant (p < 0.0001).
In Table 4, we also tested the hypothesis that the log e (age) effect differs according to race/ethnic group for each subtype. For GBM, the slope for log e (age) for each race/ethnic group was approximately 4. However, there was some evidence for heterogeneity (p-value for heterogeneity = 0.021), with Asians/Pacific Islanders and American Indians/Alaskan Natives having smaller slopes than the other race/ethnic groups. For non-GBM, the slopes by race/ethnic group ranged form 0.14 to 0.95, with a p-value for heterogeneity of 0.0005, with the smallest slopes again observed among Asians/Pacific Islanders and American Indians/Alaskan Natives.  Figure 1 Sex-specific log 10 -log 10 plots of age-specific incidence rates versus age for GBM and non-GBM, respectively, age 30 to 69 years, SEER, 1992-2007. The logarithm 10 of the age-specific incidence rate was plotted against the logarithm 10

Calendar period
In Table 5 we tested the hypothesis that calendar period variation in incidence differed according to glioma subtype. For GBM, we observed a significant trend of increasing RR with more recent period; however, the magnitude of the increase in incidence rate was not large (6% increase between 1992-1995 and 2004-2007).
For non-GBM, we observed a significant trend of decreasing RR with more recent period, with a 16% decrease in incidence rate between 1992-1995 and 2004-2007. The difference in period effect between GBM and non-GBM was highly significant (p-value for heterogeneity < 0.0001). Overall, for GBM and non-GBM combined, we observed a slight trend of decreasing incidence rate with more recent period (3% decrease between 1992-1995 and 2004-2007).

Discussion
Although we observed statistically significant differences between GBM and non-GBM in the variation of incidence rates according to age, race/ethnic group, and sex (we will discuss period separately), only the difference in age effect, which has been observed by others [2,5,22], was sizeable, with the incidence rate of GBM rising steeply with age and the incidence rate of non-GBM rising comparatively slightly with age. In spite of the statistically significant differences, the commonalities between GBM and non-GBM with respect to race/ethnic group variation and sex were more notable than the somewhat subtle differences.

Age
The GBM incidence rate increased in proportion to approximately the 4 th power of age, similar to what was found in an analysis of SEER data among Whites in 1973-1982 [22]. Because secondary GBM cases are known to have a younger age distribution than primary GBM cases [9], the increase in primary GBM incidence with age (after removing secondary GBM) must be somewhat steeper, rising toward the 5 th power, which would make the age-incidence curve for primary GBM similar to that of many carcinomas, which typically increase in incidence approximately in proportion to between the 4 th and 6 th power of age [23]. On the other hand, the non-GBM incidence rate increased in proportion to approximately the square root of age, an ageincidence pattern that is quite unusual [23]. The previous SEER analysis among Whites in 1973-1982 found slopes of 1.7 for astrocytomas and 1.0 for oligodendrogliomas [22]. The pronounced difference in age-incidence curves between GBM (mainly primary GBM) and non-GBM suggests a fundamental difference in the genesis of these glioma subtypes, as has been suggested previously [22]. Mathematical models of carcinogenesis suggest that development of cancer types with steep log(incidence)log(age) slopes, such as primary GBM, involve a large number of stable genetic or epigenetic changes (slope plus one; thus 5-6 changes for primary GBM) [24,25] or rapid expansion of a premalignant clone [26], whereas development of cancer types with shallow log(incidence)-log(age) slopes, such as non-GBM, involve a small number of stable genetic or epigenetic changes (as few as one or two) or slow expansion of a premalignant clone.
Indeed, GBM exhibits greater molecular complexity than non-GBM, with the implication that the path to GBM involves a greater number of molecular alterations [7]. Furthermore, the steeper age-incidence curve for GBM may indicate a greater role for age-related processes such as immunosenescence [27] or reduced efficiency of DNA repair mechanisms [28]. Given that among malignancies (brain or otherwise), non-GBM (as well as secondary GBM) are unique in exhibiting a high frequency (as high as 80%) of IDH mutations [15,16,29,30], which appear to be early events in non-GBM development [16], one can speculate that these mutations drive an atypical path to malignancy. For both GBM and non-GBM, the log(age) slope was similar for males and females. However, within the context of the large difference in age-incidence curves between GBM and non-GBM, we observed subtle, but statistically significant, variation in age-incidence curves among race/ethnic groups. For both subtypes, Asians/ Pacific Islanders and American Indians/Alaskan Natives had the smallest slopes. We observed a similar result for Asians/Pacific Islanders in our previous work on brain cancer as a whole [1].

Race/ethnic group
For race/ethnic group variation, we observed an important commonality between GBM and non-GBM. For each subtype, compared to non-Hispanic Whites, the incidence rate among Blacks, Asian/Pacific Islanders, and American Indians/Alaskan Natives was substantially lower (one-fourth to one-half for GBM; about two-fifths for non-GBM). However, secondary to this primary effect, race/ethnic group variation in incidence was less for non-GBM than for GBM, a difference that was highly statistically significant but only moderate in magnitude.
There is evidence for race/ethnic group differences in genetic pathways to glioma [31][32][33]. Furthermore, genome-wide association studies have identified several genetic susceptibility regions for glioma [34,35]. Given the genotype variability across race/ethnic groups [36], it is possible that variation in the frequency of susceptibility alleles across race/ethnic groups explains at least some of the race/ethnic group variation in glioma incidence, including the race/ethnic group heterogeneity in the relationship between glioma incidence and age. The commonality between GBM and non-GBM in race/ethnic group variation suggests that at least some of the susceptibility loci that may help explain race/ethnic group variation in glioma incidence would be the same for GBM and non-GBM, although some susceptibility loci appear to show specificity with respect to glioma subtype [37][38][39].

Sex
As with race/ethnic group variation, we observed an important commonality between GBM and non-GBM for sex. For each subtype, the incidence rate was higher for males than for females; this male excess of glioma is well known [2,3,5,6]. However, we did find the male/ female RR to be somewhat higher for GBM (1.6) than for non-GBM (1.4), a result that was highly statistically significant. We previously suggested that the male/ female difference in brain cancer incidence is biologically based [1], and that an explanation should be sought in genetic differences between males and females, sex hormones, and/or female reproductive factors [40]. Now we would add that any explanation should take into account the difference in the male/ female RR between glioma subtypes.
For GBM and non-GBM, respectively, we observed similar male/female RRs among race/ethnic groups, with the exception of Hispanic Whites, among whom the male/female RR was anomalously low. Further work is required to determine whether the latter result has a biological basis or stems from an unidentified bias. It is noteworthy that the quality of the SEER Hispanic ethnicity variable was found to be moderate to substantial, but not excellent [41].
In our previous work, we found that for adult brain cancer as a whole (most of which are glioma), the male/ female RR was higher in the peri-and post-menopausal age range than in the pre-menopausal age range [1]. However, in the current study, we found no evidence for an association between male/female RR and age for GBM and only suggestive evidence (p = 0.09) for a higher RR in the post-menopausal age range for non-GBM. We also found that the result for all glioma appeared to have been confounded by glioma subtype, with non-GBM both having a lower male/female RR and occurring at younger ages compared to GBM, which would explain the result from our previous work.

Calendar period
We observed trends of increasing GBM incidence rate and decreasing non-GBM incidence rate with more recent period, with an overall trend for GBM and non-GBM combined of slightly decreasing rate. Although the latter trend was statistically significant (p = 0.037), we do not conclude that the overall incidence of glioma decreased over the 16 year period of observation because the small 3% decrease between 1992-1995 and 2004-2007 could easily be explained by an unrecognized bias. Furthermore, given the well-known inter-observer variability in glioma diagnosis [10,[42][43][44][45], the increasing trend of GBM incidence coupled with the decreasing trend of non-GBM incidence may at least partly be due to a secular trend in diagnostic fashion, as opposed to real changes in incidence of these two subtypes.

Limitations
Due to the known inter-observer variability in the detailed morphologic classification of non-GBM in particular [10], we did not attempt to sub-classify these tumors into the traditional categories of astrocytoma, oligodendroglioma, and mixed oligoastrocytoma or by grade. Furthermore, we were unable to distinguish primary from secondary GBM and were unable to take into account known molecular heterogeneity within the two broad categories of GBM [46] and non-GBM [47]. Another limitation was possible selection bias due to differential diagnosis of microscopically confirmed glioma that had a specific morphology code (not the nonspecific code 9380), according to demographic factors. However, our restriction to ages 30-69 years, an age range with a high proportion of cases both microscopically confirmed and with a specific morphology code, tended to minimize the potential for any such bias appreciably distorting the results. This restriction was at the expense of generalizability to a broader age range.

Conclusions
We found statistically significant differences between GBM and non-GBM in the variation of incidence rates according to age, race/ethnic group, sex, and period. The difference in age effect, which was substantial, suggests a fundamental difference in the genesis of primary GBM (the driver of GBM incidence) versus non-GBM. However, the commonalities between GBM and non-GBM with respect to race/ethnic group and sex variation were more notable than the somewhat subtle, albeit statistically significant, differences. For each subtype, the incidence rate among Blacks, Asian/Pacific Islanders, and American Indians/Alaskan Natives was substantially lower (one-fourth to one-half for GBM; about two-fifths for non-GBM) than the rate among non-Hispanic Whites, and the rate among females was about twothirds the rate among males. These similarities suggest that within the context of a fundamental difference in the genesis of these subtypes, some aspects of the complex process of gliomagenesis are shared by these subtypes as well. The increasing calendar time trend of GBM incidence coupled with the decreasing calendar time trend of non-GBM incidence may at least partly be due to a secular trend in diagnostic fashion, as opposed to real changes in incidence of these two subtypes.