In the past several years, there has been significant progress in reducing colorectal cancer (CRC) incidence and mortality rates in most US population groups [1]. Despite this progress, an unequal cancer burden is borne by blacks, relative to whites, and by individuals of lower socioeconomic position. These groups have higher incidence and mortality rates, lower survival rates, and greater percentages diagnosed at advanced stage [2,3]. Differences in CRC survival have been consistently observed in these groups even after adjusting for stage at diagnosis, a significant prognostic factor [4,5]. Survival disparities have been attributed to differences in individual and area-based socioeconomic factors, differences in access to and receipt of quality treatment, and post treatment follow-up, [3,6-9], and/or from differences in comorbidity [4].

Geographic disparities in survival have also been observed in several international and US studies for several cancer sites including colorectal cancer [10-23]. Knowing whether cancer survival varies geographically is especially relevant since area-based physical, social and behavioral factors can assist or impede patient survival. It is also relevant because health care is often delivered locally, and, therefore, the identification of areas with significantly better or worse survival outcomes may, for example, reflect access to and quality of care. Some factors offered as potential explanations for geographic variation in survival include regional or local distributions of patient characteristics, (race/ethnicity) tumor characteristics (stage), modifiable lifestyle factors (smoking, diet, exercise) [24], area-based characteristics (poverty) and treatment. Because geographic differences in survival can result from individuals with similar prognostic factors living in the same areas (e.g. areas with a high rates of late stage CRC or areas with high poverty) it is important to adjust for well known prognostic factors. Such an approach allows one to determine whether the adjustment factors may be contributing causes of the survival disparities. In several studies, geographic disparities in cancer survival persisted despite adjusting survival time for age and stage at diagnosis, two important prognostic factors [21,22].

While previous research has examined a range of determinants of CRC survival, only one study, to our knowledge, has explored geographic variation [22]. This study observed several statistically significant areas in California and Los Angeles County with shorter or longer CRC survival after adjusting patient survival time for age and stratifying by stage. Black patients were more likely to reside in the areas with significantly worse survival, and these lower survival areas were more likely to have high levels of poverty. Additional studies that consider geographic variation in CRC survival are needed to increase our understanding of local community influences that may affect a person's chance of surviving cancer and allow for targeted interventions to population groups at greatest risk of poor outcomes.

In this study we use a recently developed extension of the spatial scan statistic for analyzing time-to-event data to identify whether the survival of CRC patients diagnosed in New Jersey varies by place of residence after adjusting survival times for disease and patient factors. We sought to answer several questions. First, are there areas in New Jersey where CRC survival is significantly longer or shorter than the statewide experience after adjusting survival times for significant prognostic factors (sex, age and stage at diagnosis) and, if yes, what are the approximate locations of these areas? We specifically adjusted patient survival times for age and stage at diagnosis in the initial analysis because we wanted to avoid survival effects resulting from statewide geographic variation in these important prognostic factors, which have been previously documented in New Jersey [25]. Second, if significant geographic differences in CRC survival are detected and we additionally adjust survival times for race/ethnicity and area socioeconomic deprivation, do these geographic survival disparities persist and to what extent are they attenuated?

Our findings suggest that survival of CRC patients diagnosed in New Jersey varies by place of residence after adjusting for disease and patient characteristics. Geographic analysis based on age and stage adjusted survival times detected several areas in the northern part of the state where CRC survival outcomes were significantly different than expected. Survival disparities persisted in some areas even after adjusting patient survival times for race/ethnicity and area socioeconomic deprivation, as defined by census-tract poverty.

Regional demographics and patient characteristics provide some evidence that the initial results based on age-stage adjusted survival might reflect geographical concentrations of patients who can be presumed at greater or less risk of poor outcomes regardless of age and stage at diagnosis – blacks and persons living in poor areas may be at greater risk of poor outcomes compared with whites and persons living in wealthy areas (Table 3)[29,40,48,49]. Areas detected with the best survival, for example, were found in predominately high income white neighborhoods in Morris, Somerset and Bergen counties; whereas areas with the worst survival were found in mostly low income, racially diverse neighborhoods in several large cities in Essex and Union counties. There are numerous characteristics of poor neighborhoods that could impede patient survival such as high unemployment, poor education, health impairing environmental exposures, substandard housing and limited access to resources and information.

Our age-stage adjusted survival estimates were consistent with findings of Huang et al. (2007) who completed a similar study using CRC data from California. They found several statistically significant areas in Los Angeles (LA) County with shorter or longer CRC survival after adjusting patient survival time by age and stratifying by stage. The shorter survival areas in LA County, like those detected for New Jersey, had both a higher percent of black cases and a higher percent of cases living in impoverished areas than the longer survival areas. In another study that used similar methods, but analyzed prostate cancer survival, Gregorio et al. (2007) also noted significant geographic variation after applying age-stage adjusted survival times.

Adjustment for patient's race/ethnicity and area socioeconomic deprivation further reduced survival disparities for several areas, but two areas remained significant with worse-than-expected outcomes. These results suggest race/ethnicity and area socioeconomic deprivation does affect outcomes in some locations in New Jersey, while one area in particular remains unexplained. Further research is needed to identify the causal factors that mediate this relationship.

It is unclear why areas of worse than expected CRC survival remain unexplained after adjusting for area socioeconomic deprivation. Possibilities include a local problem of access to health care or a pattern of care at one or more hospitals. Such patterns have been documented in the Dartmouth Atlas of Health Care project [50]. For stage III colon cancer, Etzioni et al. (2008) found that the likelihood of receipt of chemotherapy was influenced by referral patterns, hospital volume, and the presence of a cancer program approved by the American College of Surgeons' Commission on Cancer [51]. There may be other individual factors that contribute to these survival differences. For example, it could be related to modifiable risk factors (e.g smoking, diet, exercise) or comorbidity that are often themselves geographically structured [24]. Also since these areas are ethnically diverse, with substantial immigrant populations; it is possible that language barriers may affect access or coordination of care [52]. It is also possible that the adjustments for race/ethnicity and socioeconomic deprivation are inadequate or incomplete.

We have identified significantly divergent areas of CRC survival in New Jersey after adjusting for important prognostic factors, including age and stage at diagnosis. The next focus of investigation could be comparing their differences in comorbidity status as well as medical care in terms of access, utilization and quality. Such analysis could be completed for persons 65 years and older using the SEER-Medicare linked database which includes registry data and Medicare claims for covered health care services, including hospitalizations. Schootman et al. (2009) used this database to examine geographic patterns of breast cancer survival in several geographic areas[23]. It would also be important to examine other determinants of CRC risk and survivorship such as diet and exercise. Interviewing cancer survivors in these locations about their experiences combined with medical chart reviews may lead to clarification of groups at greatest risk of dying from CRC and provide explanations to geographical patterns of CRC survival. Perhaps future analysis might reveal that areas with longer than expected survival may highlight protective effects such as social support and/or clinical advancements that warrant replication in other places.

Applying the spatial scan statistic as we did in this study to adjusted survival times allows a better understanding of the extent to which geographic patterns of CRC survival can be explained by important risk factors. Documenting risk of death (observed vs expected) by geographic area after each covariate adjustment and tracking changes in risk provides a useful approach, analogous to the methods used in non-spatial statistics (e.g., Cox regression). While traditional non-spatial analysis provides greater clarity as to the precise contribution of each risk factor, this complementary approach has the advantage of highlighting specific geographic locations.

Using the 'nested circles' approach to map areas detected from the spatial scan statistic provides a greater degree of information about the risk of death from CRC among significant localized excess contained within a broader region of deficit. Typically, studies using the spatial scan statistic document only the areas that are significant, have the highest likelihood ratios, and do not overlap. As Boscoe et al. explain, this approach tends to identify large geographic areas with large populations, but small elevations in risk because these areas have the highest statistical power[53]. Changing the maximum scanning window size in the software can control for this, but the optimal size is not obvious. Also, selecting the final maximum window size on previous analyses can lead to pre-selection bias. In our study, areas E and D would have been the only areas detected if we followed typical practice and set the scanning window to a maximum of 50% of the cases and did not allow for geographic overlap. This has important public health implications if the areas detected will be used for targeted intervention.

For cancer control and prevention activities it is also important to acknowledge two further caveats related to the interpretation of our results. First, the significantly better or worse survival areas detected in our study may not be circular and are based on an identification procedure that relies on circular scanning windows. While circles are effective at highlighting general areas or regions of concern the boundaries of these areas are always approximate. Other shapes could have been employed (e.g. elliptical) and this would have likely resulted in somewhat different boundaries. Second, it is also important to consider the variation within the detected clusters and to remember that risk within these areas may not be evenly distributed. Future work using the spatial scan statistic should consider displaying maps of smoothed survival rates (descriptive or model based) beneath the statistically significant clusters [54]. Doing so would provide additional information that could be helpful to assist with the generation of hypotheses about underlying causes of survival disparities. Approaches for mapping geographic variation of patient survival have been demonstrated by Banerjee et al. (2003) and Lawson [55,56].

The results of this study need to be evaluated in light of a number of important limitations. First, by using CRC-specific survival we are presuming accuracy of the underlying cause of death on death certificates [57]. The extent to which misclassification of underlying cause of death occurred and the impact it had on our findings cannot be determined. It has been reported, however, that when deaths from colon cancer or rectum cancer are combined the accuracy of coding CRC as an underlying cause of death on death certificates is around 93 percent [58]. However, little is known about whether the accuracy of cause of death on death certificates varies by geography.

Another limitation is related to our inability to assess competing risks. Competing risks occur when there are at least two possible ways a patient can die (patients in our study could die from CRC or some other cause). When using cause-specific survival the assumption is deaths from a specified cause (e.g. CRC) are independent of deaths from other causes – thus we are assuming the absence of competing risks. If the independence assumption is not met, a bias could result because cases who are censored are more likely to die than non-censored cases [59,60] The main reason we could not assess competing risks was because we did not have comorbidity information or individual risk factors (e.g smoking). Instead we conducted sensitivity analyses of our geographic results using the following censoring scenarios: [60] (1) all patients previously censored because of deaths from other causes are assumed to die of CRC (all cause-survival); (2) all patients previously censored because of deaths from other causes survive to the end of follow-up (December 31, 2006); (3) a randomly selected subset of 5% (or 10%, 20%, 30%, 40%, 50%) of patients previously censored because of deaths from other causes are assumed to die of CRC. These scenarios allow us to consider what Kleinbaum and Klein refer to as "worst-case violations of the independence assumption" [60]. For each scenario, we detected survival disparities in approximately the same locations as the CRC-specific analysis; however scenario (1) detected a significant area of shorter than expected survival in the southwestern part of the state. Similarities between the different censoring scenarios and our results suggest minimal bias related to the independence assumption.

A third potential limitation is the geographic distribution of cases censored because they were lost to follow-up (e.g. migration). While there were more cases lost to follow-up inside the shorter and longer survival areas compared to outside each of these areas, the differences were minimal (the maximum difference was around 2 percent) and would likely have nominal impact on our findings. Also a review of the proportion of cases censored because of non-CRC deaths inside and outside the shorter survival areas deaths indicated no significant differences.

Further limitations include a lack of important patient information available from the NJSCR. For example, incomplete CRC treatment data at the NJSCR limited our ability to use this information to determine its impact on survival differences. Furthermore, the NJSCR only collects information on first course of treatment. Patient insurance data was not included because it was only required by the NJSCR starting in 1999. And the NJSCR and the SEER programs do not collect information about co-morbidity or lifestyle risk factors that may be associated with cancer incidence and prognosis (e.g. obesity, smoking, diet, and alcohol).

One additional caveat concerns how missing data excluded from our study may bias our findings [61,62]. Cancer registry data is often not missing at random, but varies by age, race/ethnicity, socioeconomic status and geography[63,64]. Among cases excluded due to incomplete geocoding (2%) there were no statistically significant differences by race/ethnicity, but these cases were more likely to be older than 75 years of age and missing stage information. Because there were so few cases with incomplete geocoding it is unlikely to have influenced our results. The greatest number of cases excluded were due to missing stage of disease at diagnosis (approximately 9.6%). Cases typically lack stage information due to medical decisions, lack of information in the medical record due to a superficial workup, or because they were obtained from death certificates only (DCO). Stage which is a proxy for prognosis is often missing in population based registries which, not only make the geographic picture incomplete, but may introduce bias [63]. Treatment information as well as information on stage is related to socioeconomic status [65], and missing stage has been shown to be higher among blacks than whites in central cancer registry data [64]. In our study we found statistically significant differences among cases without stage information by race/ethnicity (Whites, 9.3% versus blacks, 10.8%), age at diagnosis (<75, 8.6% versus >75, 17.8%) and area poverty (Low poverty, 9% versus high poverty 10.8%). Survival estimates for cases missing stage were more similar to regional stage disease than local or distant stage disease. Because the profile of cases missing stage were more similar to those at greatest risk of dying from CRC, our survival estimates could be slightly conservative if a substantial number of the cases were local or regional stage and were geographically distributed non-randomly.

Despite several potential limitations, this study is strengthened by the use of histologically confirmed CRC cases followed for up to 10 years from a large population-based SEER cancer registry with a large population (8.7 million people) and socioeconomic and racial diversity. Further strengths are high quality geocoding and the completeness of patient vital status follow-up (95.5%) since the NJSCR uses both active and passive methods.

In summary, we observed significant differences in age and stage at diagnosis adjusted survival by geographic location among the over 25,000 residents of New Jersey diagnosed with localized or regional stage CRC from 1996 through 2003. Further adjustment for race/ethnicity and area poverty reduced geographic survival disparities but did not completely explain them. These findings suggest that, in areas where adjustment changed the geographic survival patterns and reduced the risk of death, these factors may be contributing causes of the disparities. Conversely, geographic disparities that persist after adjustment likely indicate areas of unexplained, and potentially amendable, variation. Further studies need to focus on identifying specific pathways by which local factors and area socioeconomic deprivation explain geographic survival disparities.

Our use of the recently developed exponential based spatial scan statistic to examine geographic variation in CRC survival demonstrates how researchers and public health practitioners can apply this method to monitor cancer survival disparities, evaluate the effectiveness of statewide or locally based interventions and generate hypotheses about the underlying causes of geographic disparities in cancer survival. To our knowledge the exponential based statistic has only been applied to cancer survival data, but its usefulness for analyzing time-to-event makes it suitable for other applications including disease remission, cure or cessation of behavior, or hospital discharge time.

The authors declare that they have no competing interests.

KAH conceived the study and wrote the manuscript. Both KAH and XN performed the analyses. FPB provided feedback on the study design, helped interpret the results and reviewed drafts of the manuscript. All authors read and approved the final version of the manuscript.