Spatial Clusters of Cancer Incidence: Analyzing 1940 Census Data Linked to 1966-2017 Cancer Records

Abstract

Purpose

A life course perspective to cancer incidence is important for understanding effects of the environment during early life on later cancer risk. We assessed spatial clusters of cancer incidence based on early life location defined as 1940 US Census Enumeration District (ED).

Methods

A cohort of 260,585 individuals aged 0–40 years in 1940 was selected. Individuals were followed from 1940 to cancer diagnosis, death, or last residence in Utah. We geocoded ED centroids in Utah for the 1940 Census. Spatial scan statistics with purely spatial elliptic scanning window were used to identify spatial clusters of EDs with excess cancer rates across 26 cancer types, assuming a discrete Poisson model.

Results

Cancer was diagnosed in 66,904 (25.67%) individuals during follow-up across 892 EDs. Average follow-up was 50.9 years. We detected 15 clusters of excess risk for bladder, breast, cervix, colon, lung, melanoma, oral, ovary, prostate, and soft tissue cancers. An urban area had dense overlap of multiple cancer types, including two EDs at increased risk for 5 cancer types each.

Conclusions

Early environments may contribute to cancer risk later in life. Life course perspectives applied to the study of cancer incidence can provide insights for increasing understanding of cancer etiology.

Introduction

A life course perspective, the evaluation of risk factors and sequences of events arising over the lifespan, offers a distinct perspective for understanding chronic disease etiology [1–5]. A key objective of this framework is to explore how the dynamic and complex interactions of experiences during gestation, childhood, adolescence, and early adulthood affect health at later ages [3, 6]. There is mounting evidence that the environment in early life can affect adult health, including risk of cardiovascular disease, obesity, diabetes, respiratory conditions, and cancer [7–14]. When considering the role of the environment during childhood or early adulthood on cancer risk, using the life course perspective is particularly important due to the potentially long latency periods between environmental exposures and cancer diagnosis [15]. Longitudinal databases that compile records across the lifespan allow for retrospective assessment of the role of the early life environment on later life cancer risk [16]. Direct measurement of environmental exposures is not possible using most retrospective datasets and investigating the spatiotemporal patterns of cancer clustering is one approach to evaluate the importance of the early life and young adult environment on cancer risk.

Spatial methods that detect geographic clustering of cancer offer unique opportunities to identify shared exposures of individuals with cancer in small areas, which will facilitate hypothesis generation about the potential role of environmental exposures or shared behaviors in local areas. Spatial scan statistics, which search for areas with higher than expected rates or densities of events, areas for possible disease clusters, are frequently used for identifying areas of increased cancer risk [17–19]. However, studies employing this method traditionally examine the geographic location of cases at time of cancer diagnosis; cancer registries routinely collect this information. While important for acute exposures, these designs are not effective for understanding the spatial patterning of diseases with long latency periods between exposure and disease onset. In addition, studies typically explore risks for a single cancer site when it is recognized that exposure to carcinogens or shared health behaviors, such as smoking, can lead to increased risk for multiple types of cancer. Spatial analyses using residential history to examine later life cancer risk may provide new insights into the role of environment on cancer risk for multiple cancer types.

The present study tests for spatial clusters of cancer incidence based on residential location in early life, with residence defined by the household Enumeration District (ED) from the 1940 US Census.

Methods

Data for this study were extracted from the Utah Population Database (UPDB). The UPDB holds linked birth, death, historic US Census, and driver license records, which together provide longitudinal data describing an individual’s residential location. The UPDB is linked to the Utah Cancer Registry (UCR), an original member of the Surveillance, Epidemiology, and End Results (SEER) program. Incident cancer cases from 1966 to 2017 for 26 cancer types were identified in the UCR or by death records (see Supplementary Table 1 for a detailed description of the cancer groupings). Because EDs were based on population size, the geographic area is variable. As a result, densely populated areas occupy a smaller geographic space. We created a point file representing estimated geographic-weighted 1940 US Census ED centroids (N=1012) in Utah using US Census maps (20) and ArcGIS software. Map images were geo-referenced using Township, Range, and Section boundaries, street intersections, or other consistent landmarks. For EDs with small populations (n≤10), we pooled EDs within a 3 mile radius. Remaining EDs with less than 11 individuals were excluded from the analysis.

A cohort comprised of all individuals residing in Utah age 0 – 40 years at the time of the 1940 Census (Ncohort=343,521) was selected from the UPDB. All members of the cohort were cancer free in 1940. Follow-up time was measured as years from 1940 -to first primary cancer diagnosis, death, or last known date of residence in Utah. We excluded individuals with an absence of any follow-up information (Nexcluded=82,936). Individuals were then linked to ED by Census records (1940 only). Our final cohort size included 260,585 individuals who contributed 13,263,895 person-years in 892 EDs.

SaTScan software (version. 8.0) with a purely spatial elliptic scanning window was used to identify spatial clustering of EDs with high rates of cancer [21]. The elliptic scanning window was selected in order to better estimate the shape of non-circular clusters [22]. The numbers of cancer diagnoses in an ED were assumed to be Poisson-distributed according to the underlying population at risk. The Gini coefficient was used to select the upper limit of the population at risk as 3%, which reflected a purposeful intention to identify small areas of increased cancer risk [23]. Separate analyses were run for each cancer type. The most likely clusters were identified by Monte Carlo hypothesis testing; the likelihood ratio value from the dataset was compared with the values from 999 random datasets generated under the null hypothesis of no spatial clustering [17]. We considered a p-value of 0.05 to determine statistical significance. Results that remained significant after Bonferroni correction for multiple testing (p=0.0019) are indicated with an asterisk. This study was approved by the Institutional Review Board (IRB_00088870) and the Resource for Genetic and Epidemiologic Research at the University of Utah.

Results

The demographic characteristics of the cohort are displayed in Table 1. The average age of individuals in our study was 18.22 years at time of 1940 Census (range 0 – 40 years) and 50.71% of the sample was male. Average follow-up time was 50.9 years (range 0.58–78 years). Average age of cancer diagnosis was 69 years (range 11 – 107 years). Cancer was diagnosed in 66,904 (25.67%) individuals during follow-up. The average population of an ED was 292.26 individuals (range 11–1298).

Table 1:

Cohort Statistics

NCohort	260,585
NPerson-Years	13,263,895
NEnumerationDistrict	892
Population per Enumeration District	292.26 (11.00 – 1298.00)
Follow-up (years)	50.90 (0.58 – 78.00)
Age (years) in 1940	18.22 (0 – 40)
Age (years) at cancer diagnosis	69.28 (11 – 107)
Sex
Male	132,144 (50.71%)
Female	128,541 (49.29%)
Cancer Diagnosis
Yes	66,904 (25.67%)
No	193,681 (74.33%)

^*Rows are N(%) or Mean(Range)

Table 2 shows the relative risk estimates for cancer sites with statistically significant geographic clusters. Figure 1 shows the statistically significant clusters. We identified 15 clusters with increased risk of cancer. Cancer sites with increased risk were bladder, breast, lung, melanoma, prostate, cervix, colon, ovary, oral, and soft tissue. We did not find evidence of spatial clustering of increased risk of brain, esophagus, Hodgkin’s lymphoma, non-Hodgkin’s lymphoma, kidney, liver, lymphocytic leukemia, myeloid leukemia, myeloma, nervous system, pancreas, small intestine, stomach, testis, thyroid, or uterus cancers.

Table 2:

Results of Spatial Scan Analysis of Cancer Incidence in Utah: 1966–2017

Cancer Site	Type	Person-Years	Observed	Expected	RR	LLR	Test Statistic	P-Value
Bladder	M	282,423	96	56.47	1.73	11.72	11.72	0.0078
Breast (A)*	M	284,049	307	221.30	1.40	15.15	14.29	0.00069
Breast (B)	S	291,750	305	227.30	1.35	12.28	10.64	0.022
Lung (A)*	M	385,527	193	117.09	1.68	21.28	20.07	0.0000039
Lung (B)*	S	241,292	128	73.28	1.77	17.05	16.70	0.000083
Lung (C)	S	117,734	70	35.76	1.97	12.93	11.19	0.013
Lung (D)	S	260,143	124	79.01	1.59	11.16	9.66	0.047
Melanoma	M	346,820	155	102.14	1.54	12.15	11.91	0.0047
Prostate (A)	M	341,599	466	364.63	1.29	13.31	12.55	0.0036
Prostate (B)	S	392,368	520	418.83	1.25	11.71	11.48	0.010
Cervix	M	357,090	71	35.28	2.07	14.44	12.50	0.0036
Colon	M	293,390	223	163.49	1.38	9.96	9.76	0.041
Ovary	M	321,383	64	32.77	2.00	11.97	10.37	0.019
Oral	M	377,735	110	67.75	1.65	11.45	9.92	0.037
Soft Tissue	M	101,812	15	3.03	5.11	12.20	9.76	0.032

M = most likely cluster

S = secondary cluster

RR = relative risk

LLR= log likelihood ratio

^*Indicates results which remained significant after Bonferroni correction (p=0.0019)

Fig. 1.

Locations of spatial clusters of cancer rates based on 1940 enumeration district (ED) in Utah. Results of Discrete Poisson model

Figure 1 shows the map of EDs included in the analysis and highlighted areas indicating increased risk for cancer. Shaded circles indicate EDs with increased risk for one or more cancers. In the Northern region of Utah, two clusters were observed for increased risk of prostate cancer (RR_prostateA= 1.29, p=0.0036; RR_prostateB=1.25, p=0.010). One cluster with increased risk of colon cancer was also observed in this region (RR_colon=1.38, p=0.041).

In Southern Utah, one lung cancer cluster was observed (RR_lungD=1.59, p=0.047). Overlapping with the lung cluster was a cluster with increased risk for oral cancers (RR_oral=1.65, p=0.037). A cluster of increased risk for melanoma was observed in southwestern Utah (RR_melanoma=1.54, p=0.0047).

Figure 1 shows the map of Salt Lake City, an urban area of the state. In this area we found overlapping clusters for lung (RR_lungA=1.68, p=3.9×10–6; RR_lungB=1.77, p=8.3×10–5; RR_lungC=1.97, p=0.013), bladder (RR_bladder=1.73, p=0.0078), soft tissue (RR_softtissue=5.11, p=0.032), cervix (RR_cervix=2.07, p=0.0036), ovary (RR_ovary=2.00, p=0.019), and breast (RR_breastA=1.40, p=6.9×10–4) cancers. This was the area of the highest overlap, with two EDs part of five elevated risk clusters each (dark red circles). We observed one additional breast cancer cluster that did not overlap with any other cancer types (RR_breastB=1.35, p=0.022).

Discussion

In this study, we detected areas of increased risk of cancer in later life based on geographic location in childhood and early adult life, defined as 1940 US Census Enumeration District. Shared space in this context is presumed to be a proxy for environmental exposures, economic and social conditions or shared health behaviors. Spatial scan analysis showed 15 clusters of increased cancer risk and individuals in these shared spaces were significantly more likely to develop cancer later in life, with a range of 1.25–5.11 times increase in incidence rate. In an urban area of Salt Lake City, Utah, we found dense overlap of multiple cancer sites. The relative risks in the overlapping clusters varied from 1.40 (breast) to 5.11 (soft tissue) and two EDs in our analysis were each part of five different cancer clusters. To our knowledge, this is the first study to examine risk of cancer in multiple sites from cancer records linked to historical census data based on residential location.

Spatial scan statistics are commonly used to assess cancer clustering, however previous analyses using spatial scan statistics have largely investigated spatial clustering based on case location at time of cancer diagnosis [18, 19, 24]. Studies that have examined spatial clustering of cancer based on birthplace have been limited to childhood cancers [25, 26]. Similar to our results, these studies found no evidence of spatial clustering of leukemia or lymphoma by birthplace. However, because long latency periods exist between exposure to carcinogens and clinical onset of cancer, follow-up to cancer diagnosis could take decades, making adult cancer incidence an important area of study as well [15, 27]. Our study benefits from the use of data from an early time period (1940 census) and therefore we are able to observe cancer incidence with adequate follow-up. Our results are also strengthened by our large cohort size and ability to examine multiple cancer types at once based on high quality SEER cancer diagnosis data.

Some of our findings make intuitive sense, given Utah’s economy, elevation, and topography, particularly for cancer clusters found in rural areas of the state. We found two prostate cancer clusters in the Northern region of Utah, an area with a large agricultural economy; this is consistent with previous findings of increased risk of prostate cancer among men in agricultural industries [28, 29]. We found a cluster of increased risk for melanoma in an area of Utah within the Mojave Desert which experiences year-round sunshine, a known risk factor for melanoma [30]. Furthermore, clusters for lung and oral cancers were found in a region with a large mining economy. Excess lung cancer risk among miners is well documented and smoking rates among miners are high [31, 32]. The overlapping clusters in Salt Lake City may suggest shared environmental or health behaviors that lead to multiple types of cancers. Multiple cancers can share similar mutational signature from common exposures such as tobacco smoke [33]. Social norms within a neighborhood may lead to shared health behaviors that include a range of cancers [34, 35]. Environmental exposures may also lead to heterogeneous phenotype expression. For example, exposure to arsenic is associated with bladder, skin, lung, liver, and prostate cancer [36]. A combination of chemical and behavioral data for this area could be useful for further investigation of risk factors, particularly in EDs with elevated risks for multiple cancers.

The majority of the identified clusters have moderately high RR. Of the 15 clusters identified, 10 have a RR greater than 1.5, with the highest estimated risk for soft tissue cancer (RR=5.1). While there are known genetic syndromes that lead to increased risk of soft tissue sarcomas, such as Li-Fraumeni syndrome, there is no clearly defined etiology for this cancer type. Some carcinogens, e.g. 2, 3, 7, 8- Tetrachlorodibenzo-para-dioxin and beta-particle emitters, have been associated with increased risk of soft tissue sarcomas [37]. Diesel pollution has been shown to increase the risk of lung, urinary bladder, and ovarian cancers [38, 39]. The EDs with increased risk for 5 cancer types, including the aforementioned, with RR>1.5 are in an area of Salt Lake City near railways in 1940 and may have had high levels of diesel pollution. Residential selection may also contribute to the observed clusters.

While the links between specific exposures and cancer are still being discovered, the results of this study are informative for medical and public health practice. Environmental and occupational exposures can predispose individuals to cancer at multiple sites, and risk varies by timing and dose of exposure [27, 40]. Therefore, neighborhoods are useful analytical units to describe the pattern of effects of other causal factors of cancer [41]. Public health policies or interventions which establish relationships between neighborhoods of increased cancer risk and screening centers could help to alleviate some cancer burden in those communities. In addition, medical histories which include geographic information may provide clues for cancer prevention among clinicians. Consideration of environmental factors along with personal behaviors (and the interactions between the two) informs cancer prevention strategies at the individual and population level.

Our study is hypothesis generating and has some notable limitations. We lacked exposure data from 1940, including data on personal behaviors such as smoking, and therefore cannot assess the association between specific environmental exposures and cancer risk. The inclusion of contaminant or biomarker data in future work could provide valuable information for the understanding of cancer risk in adulthood. Our study is limited by the inclusion of two-time periods, early life exposure and later life cancer risk. Future studies using a life course framework to study cancer risk should include multiple periods across the lifespan. In addition, we did not have the actual geocoded residential addresses in 1940 for our entire cohort and therefore we evaluated spatial clustering at the ED level. However, use of aggregated data reduces ability to detect small areas of increased risk that may have significant clusters [42], biasing our results towards the null hypothesis. No familial clustering was considered in this analysis and it is possible that genetic risks shared within a family who live in close proximity may be masquerading as environmental clustering. Individuals represented in the UPDB have lower racial and ethnic diversity than the general US population, however the approach described here could be applied to other study populations outside of Utah to enable stratification by racial/ethnic group.

The central strength of this study is our ability to link historical census data to cancer records. We took advantage of data held in the UPDB and were able to observe longitudinal risk of cancer from neighborhood in early life. While this study was descriptive in nature, environmental exposures in early life could contribute to the observed clustering of one or more cancer types. With the growing availability of data across the life span, life course research which links individuals through time can provide novel opportunities to improve understanding of cancer risk.