A Review of the Application of Spatial Survival Methods in Cancer Research: Trends, Modeling, and Visualization Techniques

Cancer is one of the leading causes of morbidity and death worldwide (1) with an estimated 19.3 million new cases diagnosed in 2020 and 10 million deaths (2), thus contributing to a very high economic burden (3). Over 80% of the countries within the World Health Organization (WHO) have publicly available cancer prevention and control strategies, including improving accessibility to early identification and treatment. However, there are continued challenges regarding evidence-based prioritization of areas for resource allocation (4).

Spatial epidemiology describes the distribution of disease indicators by place and may facilitate ecological analyses between these indicators and sociodemographic and other area-level characteristics. It has emerged as an increasingly important tool for informing and assisting resource-sensitive cancer prevention efforts (5, 6) and has been recommended by the WHO to support evidence-based decisions (5, 7).

The development of various sophisticated spatial analysis methods and technologies have provided greater opportunities for small-area disease mapping (6, 8, 9). For example, Lyseen and colleagues highlighted that the number of published articles on spatial patterns of disease incidence or prevalence rates increased between 2000 and 2012 (10). Another review specifically related to spatial analyses of cancer incidence or mortality found an increasing trend of articles published between 1979 and 2015 (11).

Spatial survival analysis uses statistical models to assess the association between time to event data (survival) and space (small geographical units). To date, most cancer atlases have focused on incidence and/or mortality patterns (10, 11). However, survival after cancer diagnosis is an important indicator of the accessibility and effectiveness of diagnostic, treatment, and support services. Hence, understanding small-area variations in cancer survival can provide unique insights into utilization disparities in diagnostic and treatment services.

While previous reviews (10, 11) have included some spatial papers with survival outcomes, to our knowledge, there has been no comprehensive review of the published peer-reviewed literature in this field. Here, we aimed to address this gap by scoping reviewing published articles on the spatial analysis of cancer survival. Given the typically more complex statistical methodology required for survival analyses and the unique considerations when reporting survival estimates, we also provide an overview of the methods and visualization techniques used in these studies to help guide other researchers in this rapidly emerging field.

The Preferred Reporting Items for Systematic Reviews and Meta-Analyses (PRISMA) guidelines were used to conduct this review (12).

The review question was structured using the Population, Exposure, Comparator and Outcome framework (13) to assess published articles on patterns in survival (Outcome) following a cancer diagnosis (Population) in small geographical areas (Exposure) and to summarize the methods and visualization techniques.

While there was no specific comparator group in this study, spatial analyses inherently compare discrepancies by geographic unit.

In this review, spatial survival analysis methods are considered to include mapping, clustering, or regression modeling of small-area data. The scope of visualization techniques in the included articles focused on survival outcome measures, including rates or regression coefficients, which were visualized by small geographical units. The term “articles” refers to published peer-reviewed studies.

No ethics approval was required since the data were extracted from previously published articles.

A systematic search for all indexed articles was carried out on September 21, 2022 using PubMed and Web of Science electronic bibliographic databases. No filter was applied to the year of publication. Selected Medical Subject Headings and keywords were used as follows: “cancer OR neoplasms” AND “proportional hazards model OR survival analysis OR survival OR excess mortality OR excess death OR mortality/epidemiology OR hazard ratio” AND “geographic mapping OR geographic OR spatial analysis OR spati* OR spatiotemporal analysis OR spatial pattern OR disease map.” The details of the search and number of articles accessed during the search are summarized in Supplementary Materials and Methods (Supplementary Table S1).

Articles were included if they met the following four criteria: (i) the study population included individuals diagnosed with cancer, (ii) the outcome measure was survival after cancer diagnosis, (iii) survival outcomes were estimated across small geographical areas, and (iv) quantitative estimates of cancer survival were reported and/or visualized at the small area level.

The scope of the review was limited to peer-reviewed original quantitative articles in English. Reviews, books, reports, editorial letters, commentaries, conference abstracts, and mass-media publications were excluded. The titles of the articles included in the reference sections of reviewed articles identified during the searches were examined for any additional potentially relevant articles.

After removing duplicate articles, all articles were initially screened using titles, and those deemed relevant were further screened on the basis of their abstracts. One author (H.M. Bizuyehu) screened full-text articles and those articles with abstracts and titles with insufficient information using the inclusion or exclusion criteria. Reasons for exclusion of articles by full-text review were noted. Another reviewer (P. Dasgupta) independently reviewed the full-text articles. Any differences during the screening process were discussed, and the senior author (P.D. Baade) was consulted when necessary. Automatic screening was not performed in this review.

The data extracted from each of the included articles were year of publication, journal name, total follow-up period (time from diagnosis to study end), geographical unit of analysis (type, number), software used for spatial analysis, statistical methods used for survival analysis, visualization options, country, data source, cancer type, and participants’ age and sex.

The extracted data from the included articles were summarized using numbers and proportions, with the total follow-up period summarized as means. A bar chart was used to present the location of articles by individual countries, WHO regions (Africa, Americas, South-East Asia, Europe, Eastern Mediterranean, and West Pacific) and income categories (low, lower middle, upper middle, and high).

The keywords supplied by the original authors from each article were combined and visualized using the online WordClouds software (https://www.wordclouds.com/). In the software, the most frequently utilized terms are visualized in large and bold font, whereas the least frequently cited words are visualized in smaller and normal font.

All data reported in this study were extracted from published studies. All relevant data from the published articles are provided within the article and Supplementary Data.

Of the 1,370 articles identified by the systematic database searches, 129 were duplicates (Fig. 1). The remaining articles (1,241) were screened first by title, then abstract, and finally by the full-text. Thirty-two articles reported and/or visualized survival results at small geographical units were included in the final review; these were selected from the 75 full-text articles reviewed. Approximately three-quarters (n = 31) of the 43 excluded articles provided survival results but did not report and/or visualize at small geographical units in those survival estimates. The remaining articles were excluded because they did not report survival results (n = 12).

Nearly all (31/32) studies were carried out using administrative data, primarily from cancer registries (n = 28; Table 1), with the same number using unit record (individual-level) data rather than aggregated data sources. Of the cancer registry-based studies (n = 28), nearly all (n = 27) used population-based cancer registries, of which four had publicly available data [i.e., the Surveillance, Epidemiology, and End Results Program (14–17)], while one article used hospital-based cancer registry data.

More than three-quarters (n = 25) of the articles examined only one specific cancer type, and the rest included multiple cancer types, or all cancers combined. The most frequently assessed individual cancer types were colorectal (n = 9), female breast (n = 4), ovarian (n = 3), and prostate cancer (n = 3). In more than half (n = 17) of the articles, the study cohort ranged between 10,000 and 60,000 patients with cancer, with nine and six articles having fewer or greater than this range, respectively (Table 1).

Approximately two-thirds of the included articles (n = 20) used data for both sexes combined (persons), while the remainder used data only for males (n = 4) or females (n = 8; Table 1). Twenty-two articles reported the specific age groups of the included participants, and all participants were adults ages 15 years and older. About 70% of the articles studied areas of Americas WHO region, and 90% were from high-income countries (Fig. 2A): the United States (n = 22), Australia (n = 5), Canada (n = 1), and the United Kingdom (n = 1), and the rest (n = 3) were from middle-high-income countries, China, Iran, and Malaysia (Fig. 2B). Four-fifths of studies (n = 26) were conducted at the subnational level, being either one or more administrative units (states, provinces, regional administrative units) within the country (n = 21) or some locations within the administrative units (n = 5; Table 1).

All articles reported the geographical units used for the spatial analysis, with the commonly used units being census tracts (n = 9), counties (n = 5), and participants’ latitude and longitude (n = 4). Of the 14 articles that reported the number of geographical units, over half (n = 8) used 400 units or more (Table 1).

The top five most common keywords, as provided by the original authors of the articles, were survival, cancer, colorectal cancer, geography, spatial model, and relative survival. Spatial survival, spatiotemporal, and accelerated failure time models were also frequently used keywords (Supplementary Fig. S1). Of articles that visualized and/or analyzed spatial survival outcome measures (n = 28), 26 reported the software/packages used for their data analysis, with R (n = 13), WinBUGS (n = 5), ArcGIS/ArcMap (n = 4), and SaTScan (n = 3) being among the most frequently used packages (Table 1).

The number of publications using spatial survival analysis increased over time (Fig. 3). The earliest publication was in 2007, with a higher number of publications from 2016 (n = 4) and onward, with the highest number reported in 2020 (n = 7). In 2022, two articles were accessed, and approximately four articles were projected to be published by the end of the year (Fig. 3). The mean total follow-up period (time from diagnosis to the end of the study) was 14 years (range, 5–41 years). Approximately four-fifths (n = 26) of the articles had a follow-up period of 10–20 years (Supplementary Fig. S2).

Nearly four-fifths (n = 25) of the articles used spatial survival modeling methods. The remaining mapped survival estimates in small geographical areas or estimated and mapped survival based on spatial proximity to cancer care services. Eight different spatial survival modeling methods were reported, with Bayesian spatial survival models being the most frequently employed (n = 7), followed by Cox proportional hazards additive models (n = 4), and accelerated failure time models (n = 3). Smoothing techniques for eight of nine spatial survival modeling methods were reported. Model-based smoothing was the most frequently used method (n = 15), followed by expanding windows (n = 5) and penalized splines (n = 2). Model-based smoothing methods assume that the spatial effect for each area is the average of their neighbors’ spatial effects. Particularly, the smoothing techniques reported in this review included: intrinsic conditional autoregressive priors (n = 7), locally weighted regression (LOWESS; n = 4), Leroux priors (n = 2), and ordinary kriging (n = 2). Most of the articles that used Bayesian spatial survival models smoothed estimates using intrinsic conditional autoregressive priors (n = 5), which assumes that the effects of areas are correlated with neighboring areas’ effects, while the remaining articles (n = 2) used the Leroux model, which assumes that the effects of areas have both a structured and unstructured component.

Over the most recent 5 years (2018–2022), published articles used five different spatial survival modeling methods, whereas the three remaining methods (adaptive spatial filtering, Bayesian spatiotemporal flexible parametric relative survival model, and spatial scan statistic for survival data) were used in articles published prior to 2018 (see Table 2 for descriptions of the models). Three methods—spatial survival models, Bayesian geoadditive models, and accelerated failure time models—were used in published articles within the last 3 years only (2019–2022). Bayesian spatial survival models have been used in articles published since 2008, half of which were published in the past 2 years (2021–2022).

Survival measures were visualized in 23 articles using HRs (n = 8), excess HRs (n = 5), 5-year survival rate (n = 3), median survival time (n = 3), spatial frailty (n = 2), excess OR (n = 1), and probability of exceedance for hazard (n = 1; Table 1). Eight of the remaining nine articles additionally visualized nonsurvival measures such as the distribution of health facilities and population density. One article did not visualize any spatial outcomes.

This review has highlighted, for the first time, the increasing utilization of spatial methods to describe disparities in cancer survival and the wide variety of available spatial survival modeling methods and visualization techniques. Several of these published articles have reported substantial geographic variation in cancer survival (18–21), suggesting inequities in diagnosis and/or treatment strategies. These types of spatial survival studies provide quantitative evidence that is often required to motivate efforts to understand why inequalities exist, and then ultimately to intervene to reduce these inequalities.

Although the use of spatial survival methods has increased over time, the application of these spatial methods to date has been restricted to only a few countries, particularly those with established cancer registries and favorable data release policies. The wide variety of spatial survival models and methods of visualizing the results provides many different and informative ways of interpreting the data, yet this tends to make comparisons across studies difficult.

The increasing number of articles over time could be related to the increasing number of cancer registries, with data having the required population coverage and geographical precision, as well as methodologic developments and technological advancements. Population cancer registries were the main data sources for the included articles, and their complete population coverage of diagnoses and subsequent follow-up is required for the accurate measurement of spatial variation in cancer survival (22). A recent study reported that a quarter (26%) of all countries worldwide do not have any cancer registry, with about an additional 32% having a cancer registry but without national coverage (i.e., limited to subnational or hospitals; ref. 23). However, the number and coverage of cancer registries worldwide has increased substantially over the past 50 years, and with the recent opening of Regional Hubs in Asia, Africa, the Caribbean, Latin America, and the Pacific Islands, it is likely that this expansion will continue (22). Hence, these methods have increasing potential for application to a broader range of populations.

Another reason for the limited application of spatial survival methods in those countries with population-based cancer registries could be the barriers regarding establishing and effectively maintaining population-based cancer registries, including linkage with complete population level mortality information, along with policies enabling the statistical data to be accessed and analyzed at the required geographical precision while ensuring confidentiality of the original record (24–26). In addition, researchers need to have the necessary statistical capability and computational capacity for the spatial survival models (27, 28). While the expansion of population-based cancer registries internationally is encouraging (22), the more widespread use of these spatial survival models internationally will be conditional on these barriers being removed.

A large variety of spatial survival methods and visualization options have been reported, with eight different spatial survival modeling methods being identified in this review. This highlights the range of statistical methods available for these types of analyses, and specific methods are often selected on the basis of the characteristics of the data being analyzed or the outcome measures required. However, these different methods do limit the ability to directly compare published spatial patterns between countries.

In addition, the detailed geographical information within these study data means that, even when deidentified, it is typically not possible to get ethics or data custodian approval to enter into a research collaboration. Further development of the increasingly popular federated learning methods (29) could potentially make the application of these spatial survival models across different research groups internationally possible without requiring the original data to be shared. In addition, platforms such as GitHub, along with synthetically generated example datasets or online technical reports such as that released through the Australian Cancer Atlas project (30), provide an excellent opportunity to share expertise and ensure that the same approaches can be used between research groups.

Because the correct interpretation spatial survival maps can be complex, it is important to increase the training and collaboration opportunities available for researchers and policy makers. Studies in China, Bangladesh, and the United Kingdom have highlighted a number of barriers to understanding spatial analysis results and applications in the decision-making process, including technical limitations, political complexities, and cultural sensitivities (29, 31, 32).

This is the first review to synthesize published articles on the spatial analysis of cancer survival. The review was conducted according to PRISMA guidelines with multiple databases that were searched with complex queries; however, it is possible that the search terms and inclusion criteria used could have unintentionally led to the exclusion of relevant articles. This may include relevant peer-reviewed articles published in languages other than English. In addition, reports in grey literature were not included as they are typically not peer-reviewed.

This review identified an increasing trend in publications on spatial survival analyses in cancer research. It highlighted a large variety of spatial survival modeling options and visualization techniques, each providing a different approach for reporting and interpreting the data. Although more than 25 years have passed because the WHO recommended using small-area estimated disease mapping in decision-making processes (5, 7), the current application for cancer survival outcomes is mostly limited to high-income countries. A greater worldwide uptake of these methods could be facilitated through enhanced cancer registry data, knowledge sharing, and collaboration between research groups.

J.K. Cameron reports grants from Australian Research Council during the conduct of the study; grants from National Health and Medical Research Council outside the submitted work. No disclosures were reported by the other authors.

No external funding was received for this study.

We appreciate Dr. James Retell for providing assistance in visual design.

The publication costs of this article were defrayed in part by the payment of publication fees. Therefore, and solely to indicate this fact, this article is hereby marked “advertisement” in accordance with 18 USC section 1734.