Spatial Clustering of MDS in the Seattle-Puget Sound Region

Abstract

OBJECTIVES

Incidence of myelodysplastic syndromes (MDS) has been described in the United States since its inclusion in the Surveillance, Epidemiology, and End Results program in 2001, and the Seattle-Puget Sound region of Washington State has among the highest rates of the registries. In this investigation, we described small-scale incidence patterns of MDS within the Seattle-Puget Sound region from 2002 to 2006 and identified potential spatial clusters to inform planning of future studies of MDS etiology.

METHODS

We used a spatial disease mapping model to estimate smoothed relative risks for each census tract and to describe the spatial component of variability in the incidence rates. We also used two methods to describe the location of potential MDS clusters: the approach of Besag and Newell and the Kulldorff spatial scan statistic.

RESULTS

Our findings from all three approaches indicated the most likely areas of increased MDS incidence were located on Whidbey Island in Island County.

CONCLUSION

Interpretation is limited because our data are based on the residential location of the MDS case only at the time of diagnosis. Nevertheless, inclusion of identified cluster regions in future population-based research and investigation of individual-level exposures could shed light on environmental risk factors for MDS.

INTRODUCTION

The myelodysplastic syndromes (MDS) are a group of clonal proliferative bone marrow disorders that result in dysmyelopoiesis and peripheral blood cytopenias.¹ The course of disease is characterized by gradual and cumulative damage inflicted by persistent cytopenias which may result in resistance to transfusions and an overall deterioration of the immune system.² In approximately 30% of patients, MDS transforms to acute myeloid leukemia (AML).

Despite the serious health outcomes of MDS, little is known about its causes. The few known risk factors include radiation or chemotherapy treatment for a previous malignancy, and exposure to benzene and possibly other solvents.^3;4 It has also been suggested that the initial event in MDS may be infectious.² There have been very few epidemiologic studies of MDS, primarily because of its previous non-inclusion in population-based cancer registries; however, recent changes in MDS reporting will facilitate identification of MDS for future studies. Starting in 2001, a requirement was put in place for reporting of MDS to cancer registries of the Surveillance, Epidemiology, and End Results (SEER) program of the United States. Since that time, the incidence of MDS was estimated from U.S. cancer registry data at approximately 3.3 incident cases per 100,000 persons per year for 2001-2003.⁵

There has been considerable geographic variability of estimated MDS incidence between the SEER registries, with the age-sex-race-adjusted annual incidence rate of MDS in 2001-2005 ranging from 2.6-2.7 per 100,000 (Hawaii, San Francisco-Oakland SMSA, and Connecticut) to 5.9-6.0 per 100,000 (Detroit Metropolitan and Seattle-Puget Sound).⁶ Differential reporting completeness likely plays a role in the geographic discrepancies, however, true regional differences in incidence are also possible and could result from differing lifestyle or exposure to environmental or infectious agents. Within-region small-scale geographic clustering may provide more specific clues regarding environmental/infectious risk factors for MDS, and has been investigated to a limited extent. A significant geographic cluster of MDS involving 41 MDS cases within 46 census tracts from 2001-2003 was detected in western Connecticut using a spatial scan statistic;⁷ no cause of this cluster was identified.

The aims of our study were to investigate geographic clustering of incident MDS and to identify the location of potential clusters, for cases reported to the SEER program in the Seattle-Puget Sound region of Washington State from 2002-2006. To this end, we identified potential spatial clusters using several methods and carried out regression analyses to investigate ecological associations with census variables. The overall purpose of our spatial analysis was to describe small-scale spatial incidence patterns of MDS within our region and to identify clusters that may indicate potential environmental or infectious causes of MDS, in order to inform future population-based studies of MDS etiology.

MATERIALS AND METHODS

Data Description

We obtained data from the Cancer Surveillance System of the SEER program on incident cases of MDS diagnosed in the Seattle-Puget Sound region in 2002-2006. Although MDS was reportable starting in 2001, we did not include 2001 in our analysis because the number of cases increased considerably from 2001 than 2002 (by 28%), indicating a possible lag in acclimating to the reporting requirement – a point that has also been noted at the national level.⁵ MDS cases were identified by ICD-O-3 code,⁸ and included the histologic subtypes refractory anemia (RA, ICD-O-3 9980), refractory anemia with sideroblasts (RAS, ICD-O-3 9982), refractory anemia with excess blasts (RAEB, ICD-O-3 9983), refractory anemia with excess blasts in transformation (RAEB-t, ICD-O-3 9984), refractory cytopenia with multilineage dysplasia (RCMD, ICD-O-3 9985), MDS with 5q deletion (5q- syndrome, ICD-O-3 9986), therapy-related MDS, NOS (ICD-O-3 9987), and MDS, NOS (ICD-O-3 9989).

The data set for the cluster investigation consisted of 1238 patients, among whom there was adequate residential information with which to assign census tract for 1225 patients; these patients comprised the study population for our analysis. Case counts within each of the 887 U.S. Census tracts in the study area were stratified by sex, age (<50, 50-54, 55-59, 60-64, 65-69, 70-74, 75-79, 80-84, ≥85), and race (white versus non-white). The case counts data were combined with U.S. Census data from 2000⁹ of population counts within each census tract stratified by age, sex and race in the same categories as the cases.

Statistical Analysis

We investigated clustering among all MDS cases and among first primary MDS cases only (i.e., MDS cases with no previous cancer diagnoses), because of the possibility that spatially-determined environmental and/or infectious exposures may be more etiologically relevant for first primary MDS than for MDS following a previous cancer. Nevertheless, we considered both analyses in line with the aims of our investigation, as the importance of spatially-dependent exposures for MDS following a previous cancer are not known.

Our study is essentially exploratory in nature and we consequently investigated complementary methods to investigate clustering of MDS cases and to identify the location of geographic clusters. The spatial modeling was carried out using the WinBUGS software,¹⁰ and cluster detection was conducted using SaTScan¹¹ and the R software version 2.4.1.

Spatial modeling

We examined both spatial clustering and regression using a popular disease mapping model.^12;13 The basic model for the case counts is defined as:

$$Y_i~\mathit{Poisson}\left(E_i{\theta }_i\right),$$ (1)

where Y_i is the number of cases in area i, E_i is the number of expected cases in area i and θ_i is the relative risk of disease associated with area i. The maximum likelihood estimate (MLE) ${\widehat{\theta }}_i=Y_i∕E_i$, corresponds to the standardized incidence ratio (SIR) and gives an estimate of the area-level relative risk in area i. The variance of the MLE is proportional to 1/E_i and so areas with small expected numbers can show large variability. To overcome this instability, we specified a random effects model in which we allow both global and local smoothing. Specifically the model was given by:

$$\log {\theta }_i={\beta }_0+U_i+V_i,$$ (2)

where global shrinkage was achieved through the independent random effects $V_i{~}_{\mathit{ind}}N(0,{\sigma }_V^2)$ and the U_i were random effects with a spatial structure. We modeled U = (U₁,…,U_n) with a so-called intrinsic conditional autoregressive (ICAR) prior:

$$U_i\mid U_j,j\in {\delta }_i~N({\stackrel{‒}{U}}_i,{\omega }_U^2∕m_i),$$

where δ_i was the set of neighbors (census tracts) of area i, m_i was the number of neighbors,${\stackrel{‒}{U}}_i$ was the mean of the spatial random effects of the neighbors, and ${\omega }_U^2$ was the conditional variance whose magnitude determined the amount of spatial variation. This model imposes smoothing by assuming that the spatial effect in a particular area is similar to the mean of the spatial effects in close-by areas, with the strength of similarity determined by the number of neighbors (so that the similarity imposed is stronger for an area with more neighbors). In this analysis, we defined census tracts i and j to be neighbors if they share a common boundary. With this definition, census tracts located on islands were initially defined (automatically within our GIS) to have neighbors on the mainland. Based on our beliefs of the similarities of certain areas located in the Seattle-Puget Sound region, census tracts located in San Juan County were redefined to only have neighbors in San Juan County. In addition, census tracts located in Island County with neighboring relationships that crossed the Admiralty Inlet (which lies between Whidbey Island and the northeastern mainland of the Olympic Peninsula) were broken. Figure S1 of the supplementary material shows the neighborhood structures that were examined in this study.

There are two variances in this model, one spatial and one non-spatial, but they are not directly comparable since ${\omega }_U^2$ is a conditional variance while ${\sigma }_V^2$ is a marginal variance. To make the variances comparable, we were able to approximately standardize to get an approximate spatial marginal variance, ${\sigma }_U^2$. As in Wakefield (2007),¹³ we placed the inverse gamma prior with parameters 1 and 0.026 on the total variance distribution, ${\sigma }_U^2+{\sigma }_V^2$, and a uniform prior on the proportion of variance that is spatial, p. This inverse gamma prior ensures that the residual relative risks, exp(U_i + V_i) , fall between 0.5 and 2, with 95% probability. Once fitted, model (2) may inform on the level of geographical residual risk in area i (through examination of U_i and V_i ), the absolute amount of residual variability through ${\sigma }_U^2+{\sigma }_V^2$, and the amount that is spatial via the estimate of p, and the location of “clusters”, i.e. areas with increased risk. Cluster locations were investigated via examination of maps of the posterior probability that the relative risk exceeded 1.2 in each area.

We further investigated the geographical distribution of cases by carrying out ecological spatial regression using several census-tract-level demographic variables from the 2000 U.S. Census;⁹ namely, median household income (scaled and continuous), education (proportion of population with a bachelor’s degree), race (proportion white, black, Asian, other), Hispanic ethnicity (proportion), housing density (scaled and continuous), and urbanicity (Rural-Urban Continuum Code [RUCA] code categorized as urban=1-3, suburban=4-6, rural=7-10). The regression was carried out via two methods. The first was a simple quasi-likelihood approach,¹⁴ which controls for excess-Poisson variability but does not account for residual spatial dependence in the residuals. The second method extends the disease mapping described above via

$$\log {\theta }_i={\beta }_0+X_i^T\beta +U_i+V_i,$$

where the X_i are a vector of area-level risk factors associated with census tract i and exp(β) are the corresponding ecological relative risks.

Cluster detection

We used two methods to investigate whether there were clusters present, one due to Besag and Newell,¹⁵ and one due to Kulldorff.¹¹ In each method, circles are centered upon census tract centroids, and the significance of the number of cases within each circle is determined. The methods differ in the manner in which the circles are defined; Besag and Newell uses the number of cases, and Kulldorff uses populations. They also differ in the way they report clusters, as we now describe.

Besag and Newell

In the method of Besag and Newell (1991),¹⁵ circles are defined by containing a number, k, of cases, and all circles of that size are drawn. In the version we implement we use each census tract as a circle center, regardless of whether the census tract contains a case or not. As the circle expands to contain the required number of cases, a census tract is included in the circle if its centroid lies within the circle, so that jagged circles result. The number of expected cases is then derived, based on the observed population, and a p-value is calculated assuming the cases within the circles follow a Poisson distribution by summing over areas. Hence, a cluster size, k, needs to be decided upon. We examined the distribution of the number of cases per area, and the 10%, 25%, 50%, 75%, and 90% quantiles of this distribution were 0, 0, 1, 2, 3, respectively. Given these values, cluster sizes of 5 and 10 cases were selected. There is a large multiple testing problem with the Besag and Newell method, and so we chose a significance threshold of 0.005 – lower than that conventionally chosen. Given that there are 887 geographic areas, this means that if all nulls were true (i.e. the cases are randomly distributed across the study region within each of the strata of the study population), the expected number of falsely identified clusters would be 4. There is an additional problem of non-independence of tests, which is not dealt with by the method.

Spatial scan statistic

In the spatial scan statistic approach to cluster detection described by Kulldorff,¹¹ circles are centered on the centroids of each census tract, and are defined in differing sizes defined by inclusion of different proportions of the underlying population, ranging between zero and the specified maximum of up to 50% of the population. For each circle, the observed and expected numbers of cases inside and outside the circle are calculated. A likelihood ratio statistic is calculated based on the null hypothesis that the relative risk of disease inside the circle is the same as that outside the circle and the alternative hypothesis that the relative risk of disease inside the circle is greater than that outside the circle (a Poisson likelihood is assumed). The maximum of these statistics is then used as the overall test statistic, and we assess its significance level using a Monte Carlo procedure in which we simulate data under the null.⁷ We performed 999 replications for each test. Hence, in contrast to the Besag and Newell method, the multiple testing and dependency issues are addressed via the Monte Carlo procedure.

RESULTS

Characteristics of the 1225 MDS cases and the general population in the study area are shown in Table 1. MDS cases were generally age 70 or older (71.3%) and white (90.9%). The age distribution of the general population was considerably younger than MDS cases. The general population was also more likely to be of nonwhite or unknown race compared to MDS cases. Of the MDS patients, 362 had been diagnosed with a previous cancer (29.6%), and 863 (70.4%) had not (first primary MDS).

Table 1.

Characteristics of MDS cases included in the investigation of spatial clustering in the Seattle-Puget Sound region of Washington State^a

	MDS Casesb N (%)	2000 Census Population N (%)
Total	1225	4,045,707
Age (years)
<50	63 (5.1)	2,991,622 (73.9)
50-59	111 (9.1)	470,017 (11.6)
60-69	178 (14.5)	259,535 (6.4)
70-79	394 (32.2)	203,610 (5.0)
80-84	256 (20.9)	65,417 (1.6)
≥85	223 (18.2)	55,506 (1.4)
Sex
Female	543 (44.3)	2,030,879 (50.2)
Male	682 (55.7)	2,014,828 (49.9)
Race
White	1114 (90.9)	3,253,688 (80.4)
Non-white or unknown	111 (9.1)	792,019 (19.6)
Year of diagnosis
2002	230 (18.8)	-
2003	234 (19.1)	-
2004	258 (21.1)	-
2005	255 (20.8)	-
2006	248 (20.2)	-
Previous cancer diagnosis
No (first primary MDS)	863 (70.4)	-
Yes	362 (29.6)	-

^aStudy area included 13 counties and 887 census tracts in Washington State that report to the Cancer Surveillance System of the Surveillance, Epidemiology, and End Results (SEER) program

^b13 of the total 1238 MDS cases were excluded from analyses due to missing information on residential location

All MDS

The proportion of the total variability in MDS that is spatial in nature was estimated to be 66% in unadjusted analyses (Table 2). From Table 2, we see that the 95% interval for the spatial residual relative risks is (0.81, 1.28), which is larger than the corresponding interval for non-spatial residual relative risks (0.95, 1.08). There was not a great deal of residual variability in relative risk estimates for MDS, indicating that there is not a large amount of unexplained residual variation in relative risk, spatial or otherwise. A 95% range for the residual relative risk is (0.80, 1.32). Table S1 in the supplementary material contains summaries of the total residual variability, and the decomposition into non-spatial and spatial components, both before and after the inclusion of census-tract-level demographic covariates in the model. The proportion of the total variability in MDS that is spatial in nature increased slightly to 72% after inclusion of the covariates. There was no evidence of any association between the demographic covariates and census-tract level MDS incidence rates, as evidenced by inclusion of 0 in the 95% confidence intervals for the log relative risks associated with the covariates. In general, the quasi-likelihood and Bayesian regression methods were in good agreement.

Table 2.

Parameter estimates, spatial and non-spatial random effects from the intrinsic conditional autoregressive (ICAR) model for MDS incidence

	2.5%	Median	97.5%
σ U a	0.12	0.21	0.31
σ V b	0.076	0.15	0.29
p c	0.24	0.66	0.92
Exp(U)	0.81	0.99	1.28
Exp(V)	0.95	1.00	1.08
Exp(U+V)d	0.80	0.99	1.32

^aσ_U is the spatial standard deviation of the residual log relative risk for MDS incidence

^bσ_V is the non-spatial standard deviation of the residual log relative risk for MDS incidence

^cp = σ_U²/ (σ_U² + σ_V²), i.e. the proportion of variability in census-tract level incidence rates that is spatial

^dExp(U+V) is the residual relative risks for MDS incidence

Figure 1a shows the posterior probability that the relative risk of MDS in a given census tract exceeds 1.2 based on the unadjusted analysis. There were 15 census tract areas with posterior probability greater than 0.65 for higher relative risks, and the figure shows that these areas were located in Island, Pierce, Skagit, and Thurston counties. The highlighted areas were essentially unchanged in the analysis that adjusted for the census variables.

Figure 1.

(a) Posterior probability that the relative risk of incident MDS exceeds 1.2 (from ICAR model). (b) Posterior probability that the relative risk of first primary MDS exceeds 1.2 (from ICAR model).

There were 12 unique clusters of incident MDS (p ≤ 0.005) identified using the Besag and Newell method for k=5 cases. These were located in Island, King, Mason, Pierce, and Thurston counties. There were 13 significant clusters for k=10 cases, and these were located in Island, Pierce and Thurston counties. With this significance level, we would expect to see 4 clusters due to random chance. Hence, for both choices of k we saw more clusters than expected. The results from investigation of MDS for k=10 cases are displayed in Figure 2a.

Figure 2.

(a) Potential clusters of incident MDS detected using the Besag and Newell method based on expected MDS cases with a cluster size of 10 cases. (b) Potential clusters of incident first primary MDS detected using the Besag and Newell method based on expected MDS cases with a cluster size of 10 cases

The most likely cluster of MDS identified using the spatial scan statistic included 91 cases and encompassed 40 census tracts located in Island, Skagit, and Snohomish counties, however this “cluster” was not significant at conventional levels (p = 0.116, Figure 3a).

Figure 3.

(a) Most likely cluster of incident MDS identified using a spatial scan statistic (b) Most likely cluster of incident first primary MDS identified using a spatial scan statistic

First primary MDS

In analyses limited to first primary MDS, the proportion of the total variability that is spatial in nature in the unadjusted (adjusted) analysis was estimated to be 63% (Table 3), indicating slightly less spatial dependence for first primary MDS than for all MDS. From Table 3, we see that a 95% interval for the spatial residual relative risks is (0.85, 1.29), which is larger than the corresponding interval for non-spatial residual relative risks (0.96, 1.06). As with all MDS, there was not a great deal of residual variability. A 95% interval for the residual relative risk is (0.83, 1.30). In adjusted analyses, as for all MDS, none of the demographic covariates was statistically significantly associated with census-tract-level incidence rates. There were 18 areas with relative risks exceeding 1.2 in unadjusted analyses of first primary MDS and they were located in Island and Skagit counties (Figure 1b).

Table 3.

Parameter estimates, spatial and non-spatial random effects from ICAR Model for first primary MDS incidence

	2.5%	Median	97.5%
σ U a	0.074	0.21	0.32
α V b	0.084	0.15	0.34
p c	0.17	0.63	0.88
Exp(U)	0.85	0.99	1.29
Exp(V)	0.96	1.00	1.06
Exp(U+V) d	0.83	0.99	1.30

^aσ_U² is the spatial standard deviation of the residual log relative risk for MDS incidence

^bσ_V² is the non-spatial standard deviation of the residual log relative risk for MDS incidence

^cp = σ_U²/ (σ_U² + σ_V²), i.e. the proportion of variability in census-tract level incidence rates that is spatial

^dExp(U+V) is the residual relative risks for first primary MDS incidence

For first primary MDS, there were there were 4 highlighted clusters with a significance level of 0.5% identified using the Besag and Newell method for k=5 cases, which were located in Island and King Counties. There were 11 highlighted clusters for k=10 cases, which were located in Island, King and Pierce Counties (Figure 2b). As for all MDS, there were a slightly greater number of clusters detected for first primary MDS than would be expected due to random chance with k=10 cases.

The most likely cluster of first primary MDS identified using a spatial scan statistic included 67 cases and 39 census tracts located in Island, Skagit, and Snohomish counties (Figure 3b). This “cluster” was of borderline significance (p = 0.073).

DISCUSSION

Our analyses did not indicate strong spatial dependence (clustering) either among all MDS cases or among first primary MDS. Despite little evidence of overall spatial dependence of MDS incidence across the study region, there was some suggestion of local clusters. We used three methods since each has the ability to pick up clusters of different types. For example, the Besag and Newell method may pick up remote rural clusters since it is case-based, rather than population-based. The Kulldorff method more effectively adjusts for multiplicity and dependence of tests. The disease mapping approach shrinks relative risk estimates, which is not desirable for a cluster detection method, but the use of the totality of areas provides more reliable estimates in each area. For all three cluster detection methods, the highlighted areas tended to overlap. All three methods identified the most likely clusters of MDS and first primary MDS in Island County. The Besag and Newell method also identified significant clusters in Pierce County for both MDS and first primary MDS. These results suggest that there may be localized regional environmental exposures causing increased incidence of MDS; however, our study was not designed to identify causal agents.

As an exploratory analysis, two other neighborhood structures were used in the ICAR analysis (see Figure S1). The first retained neighboring between census tracts located on islands and on the mainland. This dramatically reduced the proportion of the variability that was spatial in nature as well as the standard deviation of the spatial random effects. However, it did not substantially change the range of the residual relative risks for MDS incidence nor the locations of elevated relative risks. The second neighborhood structure used redefined census tracts located in San Juan County to only have neighbors in San Juan County, however census tracts located in Island County with neighboring relationships that crossed the Admiralty Inlet were left intact. This also reduced the proportion of the variability that was spatial in nature as well as the standard deviation of the spatial random effects; however, the effect was not as dramatic. The sensitivity of our results reveals the care that is required in neighborhood specification.

Cancer cluster investigations such as the one presented here are limited by the fact that the patient’s location is based on the residential location at the time of diagnosis. This location may be more or less important for MDS etiology depending on the length of residence in the home at the time of diagnosis, because long-term exposures or exposures in the distant past may be more relevant than recent exposure to development of MDS or other cancers. Nevertheless, the relevant timing of exposure is likely to differ by the specific exposure (e.g., the type of chemical, physical, or infectious agent).

The distribution of MDS cases in the Seattle-Puget Sound region likely reflects underreporting by some local hospitals and more complete reporting by others. However, active case finding methods employed by the SEER Cancer Surveillance System in Seattle-Puget Sound are likely to partially resolve localized discrepancies in reporting. Active case-finding methods include searching hospital disease index codes (ICD-9 codes), pathology reports, cytogenetics test results, and death certificates in order to find potential cases. Nevertheless, it is also possible that under certain circumstances MDS would not be detected even with active case finding. For example, patients who are diagnosed outside of a hospital setting and are not treated are likely to be missed. Incomplete ascertainment of MDS would affect the results of our spatial analysis of incident MDS if the ‘selection’ of MDS cases into the study population was differential according to local geographic area. Hypothetically, if MDS cases were underreported in all other areas outside of Island County, then our results would be spurious.

MDS comprises a heterogeneous group of histologies, for which risk factors may differ. A large proportion (34.3%) of the MDS cases in the Seattle-Puget Sound region from 2002-2005 were recorded as MDS, NOS (ICD-O-3 9989). Although cluster investigation for specific MDS histologic subtypes would be of interest for honing in on the nature of any increased disease risk, the high proportion of cases in our study with the MDS, NOS subtype raises concerns about the validity of the ICD-O-3 classification of these cases in SEER and therefore limits any investigation of similarities or differences between MDS subtypes in spatial clustering. Future studies of MDS in the Seattle-Puget Sound region would benefit from a centralized review of MDS case pathology.

Despite the limitations of the Cancer Surveillance System SEER data, this resource has allowed us to conduct a spatial analysis in order to identify local regions of interest for future investigation of environmental or infectious agents as risk factors for MDS in the Seattle-Puget Sound region. Our findings indicate increased MDS incidence and first primary MDS incidence in census tracts located in Island County, WA. Investigation into potential causes of identified clusters could potentially shed light on environmental risk factors for MDS.

Figure S1. (a) Neighborhood structure which retained neighboring between census tracts located on islands and on the mainland. (b) Neighborhood structure which redefined census tracts located in San Juan County to only have neighbors in San Juan County. (c) Neighborhood structure which broke neighboring relationships that crossed the Admiratly Inlet and which redefined census tracts located in San Juan County to only have neighbors in San Juan County (neighboring structure used in analysis).