Abstract

INTRODUCTION

In the United States, approximately 15,780 incident cancers are diagnosed annually in individuals aged <20 years.¹ It is noteworthy that cancer is the second leading cause of death among those aged <15 years in the United States.² Although survival rates have improved substantially over the last several decades, incidence rates have either remained steady or increased, and the etiology of most childhood cancers remains largely unknown.^3,4 Etiologic investigations of childhood cancer face challenges, primarily because of its relative rarity; thus, collaboration across many hospitals and institutions is essential to obtain sufficient numbers of cases to conduct highly powered studies.

In the United States and Canada, the Children’s Oncology Group (COG) was created in 2000 through a merger of 4 pediatric clinical trials groups: the Children’s Cancer Group (CCG), the Pediatric Oncology Group (POG), the National Wilms Tumor Study Group, and the Intergroup Rhabdomyosarcoma Study Group.⁵ Currently, there are more than 200 COG member institutions in the United States and Canada and a smaller number outside of North America (eg Australia). A previous ecologic study of catchment reported that approximately 94% of children aged <15 years who were diagnosed with cancer in the United States were seen at a POG or CCG institution.⁶ Nevertheless, in a 1-to-1 record linkage study between CCG/POG cases and cases from the National Cancer Institute’s Surveillance, Epidemiology, and End Results (SEER) Program for an adjacent time period (1992–1997), Liu et al reported lower enrollment rates of approximately 71% for children aged <15 years and 24% for adolescents ages 15 to 19 years.⁷

In recent decades, case recruitment in the pediatric clinical trials groups was not conducive to nontherapeutic studies because of administrative burdens in gaining local institutional review board approvals, difficulty obtaining local institutional consent from families, and lower priority compared with clinical trials. The Childhood Cancer Research Network (CCRN) was established within COG in 2008 to ameliorate some of the logistic and regulatory challenges encountered. The CCRN is a collaborative effort to register all patients with childhood cancer (diagnosis at age <20 years) who are treated at a COG institution in the United States and Canada through the addition of an informed authorization process to the already established COG patient registration system.⁸ Newly diagnosed patients in COG and/or their parents are presented with 3 enrollment options: 1) refuse CCRN participation entirely, 2) register with the CCRN but refuse future contact for research purposes, and 3) both register with the CCRN and agree to possible future contact about participating in nontherapeutic research studies.⁸

Although the CCRN is presumed to be the sole source for conducting (nearly) population-based etiologic studies of childhood cancers in the United States, given the lack of a national pediatric cancer registry, case capture has not been examined before the current analysis. A pilot study of 1990 patients diagnosed between March 2002 and April 2006 suggests that CCRN enrollment rates are high once patients and/or their guardian are identified and approached for registration by staff at individual COG institutions, with 96% (n = 1901) providing consent to release of personal identifiers and potential future contact, 3% (n = 65) agreeing to release of identifiers only, and only 1% (n = 24) refusals.⁸ Although these percentages are encouraging, data are not available to confirm that all eligible patients and their parents are being approached for participation in the CCRN. Gaps in ascertainment of important subgroups based on patient age, sex, race/ethnicity, or geographic region also remain unknown. We assessed the completeness of enrollment in the CCRN registry during the first 3 years of operation (January 1, 2009 through December 31, 2011), both overall and by demographic and geographic characteristics. We hypothesized that the group ages 15 to 19 years would be under-represented in all geographic areas, given that many of these patients are seen by adult oncologists. In addition, we hypothesized that patients from rural geographic areas with limited access to a COG institution would be under-represented.

MATERIALS AND METHODS

To assess the catchment of the CCRN registry, we compared the observed numbers of pediatric cancer cases in the CCRN with the expected numbers based on data from SEER and the 2010 US Census Bureau.^9,10 Observed-to-expected ratios were calculated overall and according to International Classification of Childhood Cancer (ICCC) category,¹¹ geographic region, sex, race/ethnicity, and age at diagnosis. Analyses also were conducted of selected cancer subtypes to evaluate whether the case population in the CCRN is similar to the SEER population.

RESULTS

Of the 18,580 patients with childhood cancers identified from the CCRN, 44.2% were female, 25.5% were non-white, and 23.1% were Hispanic (Table 1). Observed-to-expected CCRN enrollment ratios were similar for His-panics and non-Hispanics, whereas the ratio for nonwhites was lower than that for whites. The enrollment ratio was lower in males than in females.

Complete data were available for 17,651 (95%) of the identified patients. Missing data included zip code (n = 316; 1.7%), race (n = 743; 4%), ethnicity (n = 613; 3.3%), age at diagnosis (n = 167; 0.09%), or sex (n = 93; 0.05%), and 111 patients (0.6%) were missing more than 1 item. Missing values were imputed as described above.

Across all age groups, substantially fewer patients were enrolled compared with the expected numbers (Table 1). When the analysis was restricted to patients aged <15 years, the 15,160 cases observed represented only 50% of the expected total. With the exception of infants, there was also a trend of decreasing percentage of expected cases with increasing age: The group ages 15 to 19 years had substantially lower observed enrollment compared with all younger age groups.

Figure 1 illustrates the observed-to-expected ratios for all childhood cancers combined across all ages and races by geographic area. For example, the shading for Minnesota indicates that the majority of that state had 75% to 100% of the expected numbers of cases; whereas the lightest shaded regions—areas of Nevada, Arizona, Kansas, northern Oklahoma, and Arkansas and regions of Georgia—indicate areas with the lowest observed-to-expected ratios.

Open in a separate window

Figure 1

Observed-to-expected ratios of pediatric cancer cases in the United States enrolled in the Children’s Oncology Group Childhood Cancer Research Network from January 2009 to December 2011 are illustrated according to location of residence at diagnosis. The 4 levels of solid shaded areas reflect areas with observed-to-expected ratios of <0.25 (lightest shading), from ≥0.25 to <0.5, from ≥0.5 to <0.75, and ≥0.75 (darkest shading), whereas white areas represent regions with insufficient numbers of expected cases to generate reliable estimates.

Table 2 provides observed-to-expected ratios by ICCC subtype and age group. Some variation was present with respect to ICCC subtype. For example, the ratios for renal tumors were closer to 1 overall, whereas those for retinoblastoma were lower. Ratios for the groups ages 1 to 4 years and 5 to 9 years at diagnosis tended to be higher than ratios for the other age groups. Comparison of the disease subtype and tumor location for the cancers with sufficient numbers suggested some differences between the CCRN data and the SEER registry data (Table 3). For acute leukemia, 85.2% of the CCRN cases were of lymphoid lineage, and 14.8% were acute myeloid leukemia. This is similar to the distribution in SEER, in which lymphoid and myeloid leukemias comprise 80.2% and 19.8% of childhood acute leukemia diagnoses, respectively. Likewise, the distributions of CNS tumors were similar in CCRN and SEER. In contrast, 53.3% of CCRN lymphoma cases were NHL (excluding Burkitt lymphoma), which is much higher than in SEER, in which NHL comprises 39.2% of childhood lymphoma diagnoses. Table 3 provides the location-specific subanalyses for neuroblastoma and osteosarcoma. Overall, the distributions were similar between the CCRN and SEER, with the exception of over-representation of mediastinum neuroblastoma (16.2% of CCRN cases vs 7.3% of SEER cases) and under-representation of peripheral neuroblastoma (3.8% of CCRN cases vs 18.7% of SEER cases).

DISCUSSION

In this study, we estimated the disparity between anticipated and actual enrollment of childhood cancers in the COG CCRN to quantify how representative registrants are of all childhood cancer cases in the US population and to characterize cases identified using the newly implemented consent system. We observed low observed-to-expected enrollment ratios for nearly all patient subgroups examined, although there was substantial variation by geographic location. When we analyzed subtype-specific and age-specific observed-to-expected ratios (Table 2), there were some differences with respect to disease subtype. For example, ratios were highest for leukemias, peripheral nervous system tumors, and renal tumors; whereas ratios were lowest for germ cell tumors, CNS tumors, and retinoblastomas. Overall, the groups ages 1 to 4 years and 5 to 9 years at diagnosis groups tended to have observed-to-expected ratios closest to 1, whereas the ratios were lowest for the oldest (15–19 years) and youngest (<1 year) age groups.

Although ratios were well below 1 for all cancer subtypes, ages, race/ethnicity groups, and sexes, there were some noteworthy trends. Age at diagnosis after infancy demonstrated a downward trend in enrollment, and the group ages 15 to 19 years had substantially lower ratios than all younger age groups. This is consistent with earlier reports^15,16 and is attributed to the finding that older adolescents are often treated by adult oncologists.¹⁵ This difference in enrollment by age group may be responsible in part for some of the trends by cancer subtype. For example, a majority of germ cell tumors are diagnosed in adolescent boys¹⁶; thus, the predicted ratios for this tumor type would be much lower.

The identification of sufficient numbers of patients with childhood and adolescent cancers for the conduct of well powered, informative research studies is a global issue. In the United States, where there is no national cancer registry, the high-quality SEER Program data have historically been used to monitor trends in incidence and survival, both overall and among demographic and clinical subgroups.^3,4,17–20 More recently, the National Program of Cancer Registries of the Centers for Disease Control and Prevention has partnered with SEER in assembling data covering >90% of the US population,²¹ which will allow for increased power to monitor rarer subgroups as additional annual data are amassed. Similarly, the European Union developed the Automated Childhood Cancer Information System to conduct cancer incidence and survival surveillance in the region; the original data spanned from 1970 to 1999^22,23 and are currently being updated (available at: http://accis.iarc.fr/callfor-data/index.php; Accessed July 17, 2013). In addition, the EUROCARE Project has more recent population-based survival data (1978–2007).²⁴ Some investigators, such as those in the United Kingdom and in Nordic countries, have been able to leverage national identification numbers to link data from birth, hospital, and cancer registries to conduct etiologic studies.^25,26 Unlike the CCRN, those registries have nearly complete case ascertainment; however, studies based on those data are limited in the questions that can be answered based on which data are present. In addition, linked biologic specimens are not generally readily available for these cases, with some exceptions (eg, dried blood spots in some Nordic countries²⁷). Thus, the collaboration of epidemiologists, bio-statisticians, and clinicians at treating institutions in cooperative networks like the COG remains the most promising avenue for large-scale nontherapeutic research, particularly because the networks also are needed for therapeutic studies.

Currently, the CCRN is the best source for conducting etiologic studies of childhood cancers in the United States. Because pilot data suggest that enrollment rates are high once patients and/or their guardian are identified and approached for enrollment,⁸ it is likely that the observed gap in ascertainment is because not all eligible patients and their families are being approached by COG institutions regarding CCRN participation. The differences observed by cancer subtype could be because of differential referral patterns—for example, patients with retinoblastoma being treated more often by an eye specialist—or perhaps because of increased recruitment of particular subtypes for which there were also concurrent open COG treatment protocols. The very low ratios in the group ages 15 to 19 years at diagnosis suggest that this subgroup requires particular attention in the future. In subgroup analyses of the most common cancers, the CCRN case distribution was similar to national rates with respect to leukemia and CNS tumors,¹¹ but not lymphoma, in which there was over-representation of NHL in the CCRN. It is noteworthy that lymphomas tend to present later in life compared with leukemias; thus, enrollment data for the lymphoma patient population may not be as representative because of declines in registration with increasing patient age.²⁸ Indeed, the observation that 66% of patients with Hodgkin lymphoma were in the group ages 15 to 19 years at diagnosis, whereas those with NHL comprised more patients aged <10 years at diagnosis, helps to explain this discrepancy.

Because of the low ratios, the CCRN cannot currently be considered a population-based pediatric cancer registry. The primary importance of being population-based lies in the theoretical framework of the classic case-control etiologic study—the only practicable study design for rare diseases like childhood cancer. In such a study, when the goal is to make inferences regarding exposures and disease in a given population, the controls must be drawn from the population that gave rise to the cases.^29,30 That is, anyone who would have been identified as a case had they had developed the disease in question must have a chance of being selected as a control. Hence, a population-based registry is ideal, because corresponding control participants may be sampled from the general population rather than identifying a complex sampling frame that defines the source population (for example, defining the catchment area of a particular hospital where all of the cases are being recruited).

Nevertheless, the CCRN remains a useful tool for the etiologic study of childhood malignancies. Although the CCRN is not population-based, it does not necessarily follow that the CCRN is biased in its sampling. Studies with underascertainment or low enrollment still can provide unbiased effect estimates as long as the participation rates are not associated with prognosis or the exposure of interest.^30,31 The assumption that study enrollment is not associated with an exposure of interest may have been more tenuous in COG-sponsored etiologic studies in the past, when exposures of interest were behavioral or environmental exposures measured through parental report, such as maternal alcohol consumption or diet during pregnancy—factors that carry sociocultural implications with a potential impact on a family’s decision to participate in a study.^30,32 However, a large number of current etiologic studies of childhood cancer that use the CCRN focus on biologic factors like common genetic polymorphisms and epigenetics, for which it is highly unlikely that participation would be differential. Furthermore, many new study designs are family based, including the case-parent–triad design, in which traditional controls are not selected.³³ These shifts in epidemiologic study design mean that the patients currently being enrolled in the CCRN are suitable for nontherapeutic studies.

The CCRN provides researchers with a substantial number of potential cases for diseases that are extremely rare. For example, in 3 years, the CCRN enrolled greater than 1000 children with a bone tumor, nearly 750 with a germ cell tumor or other gonadal tumor, and nearly 300 with retinoblastoma. Nearly all families (approximately 95%) have agreed to future contact regarding new research and, thus, are eligible to be contacted for participation in newly developed studies. Such studies would be extremely well powered compared with previous studies conducted on these malignancies in the United States, with sample sizes ranging from 24 to 278 patients.^34–36

Our study is the first to assess the completeness of CCRN enrollments. The use of the 2010 Census, and particularly the zip code tabulation area, to determine geographic areas allowed for improved matching of observed and expected populations, thereby improving the precision of the estimated expected counts. The SEER registry remains the best source for cancer rates in the United States. That we also were able to use multiple, recent, and overlapping years of enrollment in the CCRN as well as data from SEER means it is unlikely that our results are biased because of particularly low or high rates in a single year.

There are also some important limitations that must be considered. Of primary concern is the incomplete 1-to-1 mapping of case counts to expected rates by geographic area, because SEER does not universally capture every area of the United States; hence, there could be some accuracy lost in estimating the expected counts. We used region-specific rates only when an appropriate registry could be identified; otherwise, we used SEER averages.

Despite the overall low observed enrollment rates, the CCRN, if maintained and improved, will remain a valuable tool in the study of childhood cancer. Given that there is no comparable mechanism currently in place as an alternative, there is a need to continue enrollment to allow researchers to access willing patients as new etiologic questions are developed. The CCRN is an especially valuable source for extremely rare cancers, and particular attention should be paid to enrolling these patients. The CCRN currently captures minimal data (eg patient’s sex, age [birth date], race/ethnicity, treating institution, tumor histology/location, and zip code) because of limitations in personnel for collecting additional information at individual institutions. Etiologic investigations of childhood cancer using the CCRN rely on external funding to support and collect important environmental, biologic, and genetic information directly from the families. Ideally, in the future, the CCRN could capture additional etiologic information that could be useful in and of itself or as a basis for selecting patients for study (eg minimal family history, twinship, birth weight, place of birth, parental age, etc). Notably, there are plans in place in the COG to activate a new protocol, called EveryChild, which would include the CCRN current process along with biospecimen collection at the time of enrollment.

In conclusion, despite the overall low observed-to-expected ratios, the current results provide useful insight into the enrollment patterns by geography, age, sex, race, and histology in the completeness of case ascertainment of childhood cancer in the CCRN. This analysis demonstrates how the CCRN population differs from the entire population of childhood and adolescent cancers in the United States, so that researchers can adapt their enrollment strategies as needed to capture the patient population of interest. Steps must now be taken to improve the enrollment rates of the CCRN to maximize its utility as a research tool and fully realize its potential as a resource in the study of childhood cancer etiology.

Supplementary Material

Acknowledgments

Footnotes

References