2.1. Data for GC incidence and geographic characteristics

We obtained 5-year age-standardized incidence rates of GC across 249–252 districts (“Si-Gun-Gus”) for 2009–2013 and 2014–2018. This statistics is a type of the Cancer Registration Statistics (CRS) created from the Korea Central Cancer Registry (KCCR) and are available in the South Korean Statistical Information Service (KOSIS, https://kosis.kr/) [19]. District-specific age-standardized incidence rates are calculated for 24 types of cancer over five-year periods, with age standardization by using the administrative population as of midyear in July 1, 2000. The district is the second level of the administrative divisions in South Korea. Three hierarchical structures of administrative division include 17 metropolitan cities and provinces (“Si-Do”; Population range: 624,395–12,522,606), 252 districts (“Si-Gun-Gu”; 10,153–660,302), and 3,554 neighborhoods (“Eup-Myeon-Dong”; 116–96,266) in 2015. Given small numbers of new cancer cases in some districts resulting in confidentiality concerns, district-level incidence rates are available as 5-year aggregation. GC incidence was defined according to the International Classification of Diseases 10th edition (ICD–10) code of “C16”. All cancer cases in South Korea were originally registered according to the International Classification of Diseases for Oncology, 3rd edition (ICD-O–3) and subsequently converted to the ICD–10 [18].

We also obtained 23 district-specific characteristics which are known as being associated with GC in previous studies (S1 Table). These geographic characteristics represent eight categories of demography, socioeconomic status, lifestyle, medical status, healthcare infrastructure, medical accessibility, health screening, and physical environments (S2 Table and S2 Text). Previous studies of spatial regression showed that GC incidence was positively associated with large elderly population, socioeconomic deprivation, unhealthy behaviors, and/or limited healthcare access. We downloaded district-specific proportions and summaries of demography, socio-economic status, GC screening participation, and healthcare infrastructure from the KOSIS in 2010 and 2015 at the middle of two 5-year periods of cancer incidence. GC screening rate was the proportion of individuals aged ≥ 40 years who underwent GC screening either gastrointestinal series or endoscopy among the eligible population in a given year. District-level lifestyle, medical status, medical accessibility, health screening, and physical environments are available in the Korea Community Health Survey (KCHS) website (https://chs.kdca.go.kr/).

Because district boundaries had changed due to the consolidation or separation of districts over the 10-year study period, we applied a uniform boundary based on the year 2015 that is available in the Statistical Geographic Information Service (https://sgis.kostat.go.kr/). Specifically, we combined or split the districts that do not retain the same boundaries to those in 2015, and recomputed cancer incidence rates and geographic characteristics for modified districts. While we computed population-weighted averages for combined districts, we assigned the same proportions as the original or the counts multiplied by population or area fraction for separated districts. In addition, we excluded nine districts which are islands or located on the coastline without neighbors, resulting in 243 districts for our spatial clustering analyses.

This study used anonymized, aggregated public data, and ethical review was waived by the Institutional Review Board of the National Cancer Center (IRB number: NCC2025–0361).

2.2. Study design and period rationale

This ecological study examined spatial clusters of gastric cancer incidence and their relationships with geographic characteristics across 243 districts in South Korea. The study periods for 2009–2013 and 2014–2018 were determined based on the availability of district-level cancer incidence data aggregated into 5-year intervals. These periods also correspond to phases before and after the observed decline in gastric cancer incidence around 2012, allowing us to explore temporal and spatial differences in incidence patterns. District-level geographic characteristics were applied as ecological proxies to represent population-level exposures. We used geographic characteristics in 2010 and 2015 as the mid-year of two study periods based on data availability.

2.3. Clustering analysis

We assessed the spatial clusters of age-standardized GC incidence rate across districts by applying two commonly applied clustering analysis methods: local Moran’s I and Getis-Ord-Gi*. While local Moran’s I identifies spatial clusters based on significantly similar or dissimilar values between a target and neighboring areas, Getis-Ord Gi* relies on the spatial concentration of significantly high or low values in a target and neighboring areas. We identified the high- and low-risk areas of GC incidence as the overlapping districts clustered as high- and low-risk areas from two approaches, as suggested in previous studies [20,21].

The detailed description of two approaches is provided in the “Details of statistical methods to assess spatial autocorrelation and spatial clusters” of the Supplementary Materials (S1 Text). In brief, the local Moran’s I method defines the spatial cluster by using local Moran’s I statistic computed as the spatial autocorrelation of GC incidence in each target district with those in neighboring districts [22]. Here, we used the Queen’s contiguity method to identify neighboring districts which share either edges or corners with a target district [23]. Specifically, spatial weights of 1 and 0 were assigned to neighboring and all the other districts, respectively, and row-standardized to make each row sum of spatial weight matrix equal to 1. We chose the Queen’s method, instead of the distance-based method, to effectively define neighboring districts given highly irregular shapes and sizes of districts in South Korea (minimum and maximum area size: 3 and 1,815 km², respectively, in 2015). This neighborhood definition was commonly applied in epidemiological studies of environments [24]. The districts with significantly high positive z-scores of local Moran’s I statistics imply a cluster where a target district has similarly high or low GC incidence as its neighborhood. These clustered areas include high-high and low-low clusters, indicating districts with high and low GC incidence, respectively, in target and neighboring districts compared to the overall mean. On the other hand, the Getis-Ord-Gi* approach assesses the local sum of weighted GC incidence for each target district together with the neighboring districts and compares to the sum of weighted GC incidence for all districts [25]. As Getis-Ord-Gi* statistics are considered as equivalent to z scores, the districts with significantly positive and negative z-scores, defined as hot and cold spots, indicate the clusters with higher and lower GC incidence than would be expected. Finally, we defined the “high-risk area” as the overlapping districts of the high-high cluster identified by the local Moran’s I and the hot-spot identified by the Getis-Ord-Gi* [20,21]. Likewise, the “low-risk area” indicates the overlapping districts of the low-low cluster and the cold-spot.

Furthermore, we examined the sensitivity of our cluster detection to neighborhood definition and inferential criteria. Three alternative spatial weight specifications included Rook contiguity, k-nearest neighbors, and distance band approaches (S3 Table). For k-nearest neighbors, we applied seven neighbors that gave the maximum global Moran’s I statistic beyond the average contiguous neighbors of five (S1 Fig). The distance band of 35 kilometers was chosen based on a previous study of GC clustering analysis in South Korea [26]. K-nearest neighbors also allowed the clustering analysis using all 252 districts without excluding nine districts as islands or on the coastline. In addition, as conservative inferential criteria, we applied permutation-based inference with a significance threshold of p-value < 0.05 and Benjamini–Hochberg false discovery rate with a threshold q-value < 0.10.

2.4. Analysis for geographic characteristics and GC clusters

We examined the difference in each geographic characteristic between high-and low-risk areas using Student t-test or Mann Whitney U test in each period. The statistical significance was defined as p-value < 0.00217 (i.e., 0.05/23 geographic variables) with multiple comparison adjustment. In addition to the absolute mean difference, we computed the average relative difference as the difference in averages of each geographic characteristic between high- and low-risk areas divided by the average in the low-risk area. Then, to examine the differences in relevant geographic characteristics between high- and low-risk areas and between the two periods, we computed the differences of the average absolute and relative differences between the two periods. We also presented the standardized mean difference (SMD) that is the average difference divided by pooled standard deviation to compare the difference by minimizing the impact of the spatial correlation of geographic characteristics (S3 Text) [27]. In addition, we applied an interaction model including the status of clusters and periods, and evaluated the statistical significance in the differences of characteristics between high- and low-risk areas in the later period compared to the early period.

In order to examine the robustness of our findings for geographic characteristics and GC incidence, we performed three sensitivity analyses. First, given the long latency of gastric carcinogenesis, we applied geographic characteristics in the early period for 2010–2011 to the incidence in the later period for 2014–2018. Second, we used age-specific GC incidence rates, instead of the binary indicator of high- vs. low-risk areas, and conducted regression analyses on each geographic characteristics within each period as well as for their differences between periods. Lastly, we explored the role of geographic characteristics in the four types of districts where risk cluster status changed over the two periods: high-risk to non-significant (NS), NS to low-risk, low-risk to NS, and NS to high-risk areas. There were no districts that showed cluster shifts of high- to low-risk or low- to high-risk areas.

Data processing and statistical analyses were carried out in ArcGIS 10.5 (ESRI, Redland, CA, USA), R 4.4.1, (R Core Team, Vienna, Austria), and SAS 9.4 (SAS Institute Inc., Cary, NC, USA).

3.1. Data exploration for GC incidence and geographic characteristics

The median 5-year age-standardized incidence rate for GC across 243 districts decreased from 83.4 to 68.8 per 100,000 people between 2009–2013 and 2014–2018. The districts that showed high GC incidence slightly changed between the two periods: the incidence was high in the districts located in the central inland region of South Korea for 2009–2013 and in Southern inland region for 2014–2018 (Fig 1). Both periods gave significant spatial autocorrelation of GC incidence across districts (global Moran’s I statistic = 0.52 and 0.47 in 2009–2013 and 2014–2018, respectively). Most district-specific characteristics showed significant changes between the two periods (Table 1). From 2009–2013–2014–2018, the average district-specific proportions of older adults aged ≥ 65 years, urban dwellers, and the proportion of individuals with higher educational attainment increased by 2.4, 1.8, and 5.0%, respectively. The proportion of current smokers declined, whereas the proportion of heavy drinkers increased. Although fewer people eat breakfast, low-salt preference became more common. While medical accessibility and health screening rates including GC screening participation improved, the prevalence of chronic diseases such as hypertension, diabetes, and dyslipidemia also increased.

Download:

• PNG
  larger image
• TIFF
  original image

Table 1. Means and standard deviation of 23 geographic characteristics across 243 districts between 2009–2013 and 2014–2018.

https://doi.org/10.1371/journal.pone.0349384.t001

Download:

• PNG
  larger image
• TIFF
  original image

Fig 1. Maps for age-standardized incidence rates of gastric cancer across 243 districts in South Korea for 2009–2013 and 2014–2018.

https://doi.org/10.1371/journal.pone.0349384.g001

3.2. Spatial clusters of GC incidence rates between two time periods

S2 Fig shows the locations of spatial clusters for GC incidence defined by local Moran’s I and Getis-Ord-Gi*. Both methods identified similar districts as clusters: high-high and hot-spot clusters in the central inland, and low-low and cold spot clusters in the northern region. These similar cluster locations between the two approaches gave high- and low-risk areas in the central and northern regions, respectively (Fig 2). The locations of the two types of clusters were generally consistent between the two time periods. The low-risk area was expanded from 26 districts for 2009–2013 to 37 districts for 2014–2018, while the high-risk area was slightly reduced from 31 to 28 districts. The average age-standardized incidence rates for GC were 95.9 and 73.4 per 100,000 people in high- and low-risk areas, respectively, for 2009–2013. These rates decreased in both high- and low-risk areas to 79.2 and 61.1 per 100,000 people for 2014–2018 (Table 2). The locations of high- and low-risk areas were largely consistent in our sensitivity analyses using alternative spatial weight specifications and inferential criteria (S1, S3, and S4 Figs).

Download:

• PNG
  larger image
• TIFF
  original image

Table 2. Geographic characteristics in high- and low-risk areas and their differences by 2009–2013 and 2014–2018.

https://doi.org/10.1371/journal.pone.0349384.t002

Download:

• PNG
  larger image
• TIFF
  original image

Fig 2. Maps of high- and low-risk areas of age-standardized incidence rates for gastric cancer across 243 districts in South Korea for 2009–2013 and 2014–2018.

https://doi.org/10.1371/journal.pone.0349384.g002

3.3. Geographic characteristics between GC clusters and time periods

Table 2, S4 Table, and Fig 3 present the comparison of district-level geographic characteristics between the high- and low-risk areas as well as between the two periods. In the earlier period for 2009–2013, the high-risk area showed significantly lower proportions of educational attainment and dyslipidemia diagnosis and higher proportion of GC screening participation than those in the low-risk area. In the later period for 2014–2018, educational attainment and GC screening gave the consistent pattern of significantly lower and higher proportions, respectively, in the high-risk area. In addition, demographic, lifestyle, and environmental characteristics became significantly different between high- and low-risk areas (Table 2 and S4 Table). The high-risk area showed the higher proportion of older adults, lower population density, and the lower proportion of urban-dwelling population. While proportions of regular walking and urban forest coverage within residential areas were lower, the proportions of self-reported obesity and frequent breakfast consumption were higher.

Download:

• PNG
  larger image
• TIFF
  original image

Fig 3. Average relative differences in geographic characteristics between high- and low-risk areas by 2009–2013 and 2014–2018.

Relative differences were calculated as (high-risk − low-risk) / low-risk × 100. Positive and negative relative differences indicate higher levels in high- and low-risk areas, respectively. The ‘+’ and ‘−’ signs on x-axes indicate anticipated directions in the association with GC incidence.

https://doi.org/10.1371/journal.pone.0349384.g003

When we compared absolute and relative differences in high- vs. low-risk areas between the two periods, the patterns of positive and negative differences in demographic characteristics were generally similar, but the quantity of differences were more pronounced for 2014–2018 than for 2009–2013 (S4 Table and Fig 3). The positive difference in older adults and negative difference in urban-dwelling population and urban forest coverage within the residential areas, indicating higher and lower proportions in the high-risk area, were greater with 16–33% changes (relative difference: 45% for 2009–2013 to 78% for 2014–2018, –24 to –40%, and –48 to –75%, respectively). Interestingly, some behavioral characteristics showed the reverse direction between the two periods from less to more healthy patterns in the low-risk area. The positive difference for low-salt preference and the negative difference for heavy drinking in the early period became negative and positive differences in the later period. Sex ratio, hypertension diagnosis, and unmet healthcare needs also gave the reverse patterns between the two periods. Although GC screening showed higher rates in high-risk areas in both periods, these differences were consistent between two periods, resulting in a little change (–0.3%). The interaction analysis by risk area and period also showed no significant interaction for GC screening. The patterns of absolute and relative differences were consistent in SMDs.

In our sensitivity analysis using geographic characteristics prior to GC incidence, the patterns were overall consistent with mostly < 5% relative differences compared to those in the original analysis using later-period geographic characteristics (S5 Table). Regression analyses using district-level GC incidence rates also showed the negative and positive associations with educational attainment and GC screening, respectively, in both periods (S6 Table). When we used the differences in geographic characteristics and GC incidence between two periods, the negative association of educational attainment remained. In contrast, the association with GC screening participation disappeared. For four types of districts where risk cluster status changed over the two periods, 15 and 17 districts of high-risk and NS areas for 2009–2013 became NS and low-risk areas, respectively, indicating decreased risks of GC incidence. In contrast, 6 and 12 districts in low-risk and NS areas changed to NS and high-risk areas, respectively. However, these districts did not show significant differences in geographic characteristics between the two periods (S5 and S6 Figs).