Residential Proximity to Petroleum Storage Tanks and Associated Cancer Risks: Double Kernel Density Approach Vs

Science of the Total Environment 441 (2012) 265–276

Contents lists available at SciVerse ScienceDirect

Science of the Total Environment

journal homepage: www.elsevier.com/locate/scitotenv

Residential proximity to petroleum storage tanks and associated cancer risks: Double Kernel Density approach vs. zonal estimates

Marina Zusman a,⁎, Jonathan Dubnov b,c, Micha Barchana b, Boris A. Portnov a a Department of Natural Resources & Environmental Management, Graduate School of Management, University of Haifa, Mount Carmel, Haifa 31905, Israel b School of Public Health, Faculty of Social Welfare & Health Sciences, University of Haifa, Mount Carmel, Haifa 31905, Israel c Haifa District Health Ofﬁce, Ministry of Health, Palyam Ave. 15, Haifa 31999, Israel

HIGHLIGHTS

► Lung and NHL cancer risks near a petroleum storage site have been assessed. ► NHL and lung cancer ASRs for small census areas near the site have been estimated and compared. ► Double kernel density (DKD) analysis have been used as an alternative tool for risk assessment. ► ASRs detected no association between site proximity and cancers while DKD analysis detected such an association. ► As we conclude, DKD analysis is a more sensitive method for risk assessment when the number of census areas is small. article info abstract

Article history: Background and aims: The relationship between exposure to petroleum products and cancer is well- Received 24 April 2012 established in occupational studies carried out among employees of transportation and oil-producing indus- Received in revised form 21 September 2012 tries. However, question remains whether living near petroleum storage facilities may represent a cancer risk. Accepted 22 September 2012 In the present study, we examined cancer incidence rates associated with residential proximity to the Kiryat Available online 10 November 2012 Haim industrial zone in Northern Israel, using different analytical techniques and adjusting for several potential confounders, such as road proximity, population density, smoking rates and socio-demographic attri- Keywords: butes. Double Kernel Density (DKD) approach Age-Standardized Rates (ASRs) Methods: Both traditional zonal approaches and more recently developed Double Kernel Density (DKD) tools Lung and NHL cancer incidence were used to estimate relative risks of lung and NHL cancers attributed to residential proximity to the petro- Residential proximity leum storage site. Petroleum storage sites Results: Zonal approaches based on comparing ASRs across small census areas (SCAs) did not detect any sig- niﬁcant association between residential proximity to the industrial zone and the two types of cancers under study (P>0.2). In contrast, the DKD approach revealed that the relative density of both lung and NHL cancers declined in line with distances from the industrial zone, especially among the elderly (Lung: t>−12.0; Pb0.01; NHL: t>−9.0; Pb0.01), adjusted for proximity to main roads, population density, smoking rate, average income, and several other potential confounders. Conclusions: Living near petroleum storage sites may represent signiﬁcant cancer risk which cannot always be detected by traditional zonal approaches commonly used in epidemiological studies, especially if the number of census areas available for the analysis is small. © 2012 Elsevier B.V. All rights reserved.

1. Introduction However, there are two main disadvantages associated with ASR calculation and use, viz.: information loss attributed to aggregation In cancer epidemiology, calculations of Age-Standardized Rates and averaging, and potentially different weight gains for different (ASRs) are commonly used for quantifying health-response variables “standard populations” (Schoenbach and Rosamond, 2000). In addi- (Ahmad et al., 2001; Curtin and Klein, 1995). tion, calculating and comparing ASRs may not always be efﬁcient, if a researcher deals with a very small number of statistical areas, e.g., 10 census tracts or less. If the number of census areas is small, ASR ⁎ Corresponding author at: Department of Natural Resources & Environmental Man- comparison may lead to a low statistical power and often inconclu- agement, Graduate School of Management, University of Haifa, Mount Carmel, Haifa 31905, Israel. Tel.: +972 4 828 8532; fax: +972 4 824 9971. sive results (Curtin and Klein, 1995; Hollenbeck et al., 2006; Li and E-mail addresses: [email protected], [email protected] (M. Zusman). Lian, 2010).

Another issue associated with exposure estimates based on ASR 2.1. Cancer morbidity and mortality in Israel calculations is dependence on the way in which geographic boundaries of statistical areas are delineated, reflecting the phenomenon According to the Israel Ministry of Health, cancer is the leading known as the “modifiable areal unit problem” or MAUP (Kloog et cause of mortality in Israel, with especially higher rates in the Haifa al., 2009). district in northern Israel, where large industrial enterprises, chemi- The Double Kernel Density (DKD) technique is an alternative cal storage facilities and major thoroughfares are located (Barchana investigation approach, which can help to address some of the et al., 2007; ICBS, 2008a; IMEP, 2008). drawbacks associated with ASR calculation and analysis. The DKD Several previous studies attempted to assess the effect of industri- approach is a non-parametric method of extrapolating point data al and traffic pollution on respiratory diseases and various types of over an area of interest without relying on fixed boundaries for cancer, using air pollution data obtained from air quality monitoring aggregation (Donthu and Rust, 1989; Kloog et al., 2009; Nakaya stations (see inter alia Dubnov et al., 2007; Portnov et al., 2009) and and Yano, 2010; Portnov et al., 2009; Sanvicente-Añorve et al., by calculating proximities to main thoroughfares (Paz et al., 2009). 2003). Although this technique has been commonly used in ecolog- Thus, a recently published study on non-Hodgkin lymphoma (NHL) ical research and criminology (Blundell et al., 2001; Harada, and in the Haifa Bay area found that NHL rates were significantly associat- Shimada, 2005; Xie and Yan, 2008), its use in epidemiological ed with living near heavy traffic roads (Paz et al., 2009). studies has been rather infrequent (Kloog et al., 2009; Portnov et In 2007–2008, the Israel Ministry of Environmental Protection al., 2009). carried out a comprehensive environmental survey of air pollutants The advantage of this technique, compared to zonal approaches, is in the Haifa Bay area, most of which are not routinely monitored by that DKD estimates do not depend on the geographic delineation of local air quality monitoring stations. The survey revealed exceeding statistical areas and produce a smooth, continuous “disease risk” sur- concentrations of several hazardous substances, including benzene, face for the entire study area, which can then be used to calculate re- methylene chloride, acetaldehyde, acrolein and benzo(a)pyrene sponse variables for multivariate statistical models (Carlos et al., (IMEP, 2009). These toxic substances (some of which are known as 2010; Kloog et al., 2009; Portnov et al., 2009). carcinogens or suspected carcinogens) may cause health effects in In the present study, we use two different analytical approaches the general population and especially among populations at risk to investigate cancer risks associated with proximity to the Kiryat (i.e., elderly, children, pregnant women, patients with chronic dis- Haim industrial zone in Northern Israel, which hosts several major eases) and lead to the development of several malignancies (NCI, petroleum-storage facilities. In particular, we compare the results 2003). of zonal ASR calculations with DKD estimates. As our study indicates, However, whether these hazardous substances (or residential cancer incidence rates in the study area tend to decrease in line with proximity to their sources) are linked to the elevated cancer morbid- distance from the industrial zone, especially among the elderly ity observed the Haifa Bay area has not been confirmed (Barchana et (Pb0.001), controlled for local income levels, housing densities, per- al., 2007; ICBS, 2008a; IMEP, 2008). cent of Jewish population, smoking rates, proximity to the sea and the main roads, and several other potential confounders. However, 3. Materials and methods such a distance-related decline was detected only by the DKD approach, whereas the results of ASR-based zonal calculations were 3.1. Study population found to be inconclusive. Together with its immediate suburbs – Krayot, Tirat Carmel,and Nesher – the City of Haifa hosts about 600,000 residents. The Greater 2. Background studies Haifa Metropolitan Area (GHMA) has a mixed population of Jews (81%), Arabs (10%) and other ethnic minorities (9%) (MH, 2010). Empirical evidence accumulated to date points out that occupa- The Kiryat Haim industrial zone, residential neighborhoods around tional and environmental exposure to crude oil emissions and refined which form our study area, is located in the northern part of the oil products (by leaked or spilled chemicals from petroleum storage city and covers the area of some 4.2 km2. 13 small census areas or tanks) may cause different health disorders, including anemia, pe- SCAs (similar in size to census blocks in the U.S.A.) are located in ripheral neuropathy, damage to the respiratory and central nerve the radius of 2 km from the industrial zone (which can be considered systems, and skin diseases (Lewtas, 2007; Wilkinson et al., 1999). as a potential exposure area), and host the total population of some According to other studies, exposure to petroleum emission was 53,000 residents (see Fig. 1). found to be linked to lung cancer, leukemia, and non-Hodgkin lym- In 2009, the Israel Ministry of Environmental Protection identified phoma (ATSDR, 1999; Clapp et al., 2008). The types and severity of several industrial plants in the Kiryat Haim (KH) industrial zone health effects associated with exposure to petroleum oil products (including Oil Refineries Ltd., Gadot Chemical Tanks and Terminals were found to depend on the amount and the length of exposure Ltd., and Gadiv Petrochemical Industries), which air pollution emis- (IARC, 1998; Sexton et al., 1995). sions exceeded the local air quality standards. Some of these plants Although several studies on the exposure to petroleum and pe- use toxic materials in their production lines, while others store and troleum emissions as cancer risk factors have been carried out to market such materials (IMEP, 2009). date, most of these studies have been focused on the occupational exposure of petroleum industry workers (Christie et al., 1991; 3.2. Cancer data Jarvholm et al., 1997). A marked limitation of such studies is the likely presence of the “healthy worker effect”, due to which the estimat- Data on cancer incidence rates for the present study were ed morbidity tends to be lower than among a general population, obtained from the Israel National Cancer Registry established in which includes children and the elderly (Kirkeleitetal.,2008). Con- 1960. Since 1982, all medical facilities such as hospitals, laborato- currently, studies of exposure to chemicals from petroleum industry ries, and other medical private and public institutions, report to and, particularly from petroleum storage facilities in community- the registry. The registry's completeness makes up for about 95% based studies, are relatively scarce and often attribute elevated mor- cancer cases (Barchana et al., 2007). bidity and mortality rates to industrial pollution in general (Liu et al., From all the patients diagnosed with cancer during the period of 2008; Pan et al., 1994; Tsai et al., 2009; Yang et al., 1997, 1999; Yu et 2000–2006 and living within the distance of up to 2 km from the al., 2006). KH industrial zone, 98 primary lung cancer cases and 53 cases of M. Zusman et al. / Science of the Total Environment 441 (2012) 265–276 267

Fig. 1. Map of the study area showing the Kiryat Haim industrial zone, small census areas (marked by 7-digit numbers) and residential locations of lung and NHL cancer patients (small dots and triangles). non-Hodgkin lymphoma (NHL) cancer were identiﬁed in the data- 3.3. Study phases base and used in the analysis. These 151 cases include all cases of lung and NHL cancers recorded within the 2-km radius from the in- The analysis was performed in the following two stages. During the dustrial zone. According to previous studies (see inter alia ATSDR, ﬁrst stage, ASRs of lung and NHL cancers were calculated for 13 Small 1999; Christie et al., 1991; Clapp et al., 2008; Jarvholm et al., 1997), Census Areas (SCAs), into which the entire study area is divided, sepa- these types of cancer may potentially be related to exposure to petro- rately for each cancer type. The rates were age-standardized to the leum emissions. “standard” Israel population (ICBS, 2008b), to enable comparison. 268 M. Zusman et al. / Science of the Total Environment 441 (2012) 265–276

Next, the SCAs covered by the analysis were divided into three distance Table 2 bands, according to the SCAs' centroid distance to the industrial zone – A non-parametric test of differences in the relative densities of different types of cancer as a function of distances from individual residences to the Kiryat Haim industrial zone. 0–500 m, 500–1000 m and more than 1000 m (see Fig. 1), – so as to compare ASRs observed in each distance band. During the second Cancer Age Distance No of Mean rate χ2 a P-value stage of the analysis, density of each cancer type (adjusted for age) type group range (m) cancer (per 100,000) was calculated and mutually compared using the DKD approach (as de- cases tailed in the next subsection). Lung b65 0–500 17 27.95 132.95 b0.001 500–1000 3 9.67 1000–1500 11 16.98 65+ 0–500 29 408.34 92.84 b0.001 3.4. Kernel Density and Double Kernel Density estimates 500–1000 18 130.74 1000–1500 20 223.86 As previously noted, the Kernel Density (KD) estimation is a non- NHL b65 0–500 9 16.66 139.95 b0.001 – parametric smoothing technique that calculates density value for 500 1000 5 11.85 1000–1500 10 26.23 each point of space (e.g., density of cancer patients) by transforming 65+ 0–500 13 327.81 244.86 b0.001 geographically referenced points (such as residential locations of in- 500–1000 7 131.05 dividual cancer patients) into a continuous surface of a disease's in- 1000–1500 9 98.06 cidence density (Portnov et al., 2009). According to this technique, a Kruskal–Wallis test. points lying near the center of a given raster cell's search areas are weighted more heavily than those lying near the edge (McCoy and Johnston, 2001). In our study, DKDs of lung and NHL cancer patients were calculated in several steps. During the first phase of the DKD calculation, the were examined (i.e., 100, 200, 300, 400 and 500 m), the search radius residential addresses of all cancer patients (both lung and NHL) were of 300 m was appeared to provide the best results and KD estimates geo-coded, that is, converted into X,Y coordinates. The information based on it were used in the multivariate analysis. about patients' addresses was obtained from the INCR database (see During the last phase of analysis, we normalized the KD of cancer Barchana et al., 2007). Next, distance from the place residence of patients by the total number of people residing in each 50×50-m ras- each cancer patient to the border of the industrial zone was estimated ter cell, which is comparable in size to a single building. This normal- using the “spatial join” tool of the ArcGIS9.x™ software (ESRI, 2009). ization technique is known as the Double Kernel Density (DKD) or At the next step, we generated separate KD maps for each type of can- Relative Kernel Density (RKD). The normalization was needed since cer under study, that is, lung and NHL. the concentration of cancer patients near a pollution source may po- One clarification is important. As mentioned previously, the KD al- tentially be due to high population densities near it (attributed to e.g., gorithm calculates the density of individual observations (such as res- availability of affordable housing near a pollution source (WHO, idential locations of cancer patients in our study) within a 2012)). The numbers of residents in each 50 by 50-m raster cell pre-defined search radius from a target point. The size of search radi- were calculated using the data from the general population registry. us (which is a user-specified variable) determines the degree of The calculations were performed separately for two age groups of smoothness of the surface. Larger values of the search radius produce cancer patients — younger than 65 and older than 65, to account for a smoother density raster but might lead to over-smoothing of the age effects. kernel surface, while a small value might lead to a spiked kernel over a location (Portnov et al., 2009). Since the size of the search 3.5. Exposure estimates radius may influence the results of the analysis, we examined several search radii for kernel density estimation. From the five radii that In order to convert continuous KD surfaces of cancer cases (see previous subsection for more detail) into discrete observations, suit- able for a multivariate analysis, we generated 1000 randomly dis- tributed “reference” points that covered the entire study area (see Appendix 2 and Fig. 1). The points were generated using Hawth's Table 1 Analysis Tool in ArcGIS9.x™. The reference points were then “spa- Non-parametric test of differences in cancer ASRs as a function of distance from SCAs' tially joined” with KD contours of both types of cancer under study, centroids to the Kiryat Haim industrial zone. that is, assigned values of the lung and NHL density contours being Cancer Age Distance No No of Population ASR χ2 a P-value closest to each reference point [For a more detailed description of type group range (m) of cancer estimate this procedure see Portnov et al., 2009]. SCAs cases (per 100,000) In the absence of actual measurements of petroleum emissions, aerial proximity from each “reference point” of the boundary of the Lung b65 0–500 4 17 14,050 20.35 0.62 0.732 500–1000 2 3 8950 8.70 industrial zone was calculated and used as a surrogate measure of ex- 1000– 7 11 22,780 14.37 posure (the limitation of such estimation is discussed in the conclud- 1500 ing section). The calculations were performed using “spatial join” tool 65+ 0–500 4 29 2620 38.70 0.36 0.834 in the ArcGIS9x™ software, which calculates areal distances between – 500 1000 2 18 1280 31.52 different geographic features based on their spatial location (McCoy 1000– 7 20 3480 27.08 1500 and Johnston, 2001). NHL b65 0–500 4 9 14,050 15.06 2.62 0.270 500–1000 2 5 8950 2.05 3.6. Control variables 1000– 7 10 22,780 14.35 1500 65+ 0–500 4 13 2620 20.64 0.30 0.859 In addition to the main explanatory variable (that is, distance to 500–1000 2 7 1280 5.73 the industrial zone), the analysis covered the following additional ex- 1000– 7 9 3480 11.46 planatory variables: distance to the main roads (m.), distance to the 1500 sea shore (m.), population density (per km2), smoking rate (in the a Kruskal–Wallis test. absence of individual data estimated as the average smoking rate in M. Zusman et al. / Science of the Total Environment 441 (2012) 265–276 269 the SCA, %), average income (NIS per capita), percent of Jewish popu- nearest major road was calculated and used in the multivariate lation and housing density (people per room). Descriptive statistics of analysis. We also included distance to the sea shore as a control these variables are reported in Appendix 1. variable since there is evidence that sea winds may transfer emis- Some of these variables deserve comment. According to sions and contaminants affecting human health (Bidoli et al., Nerriere et al. (2005), emissions from motor vehicles are found 1998; Busby, 2004). In particular, distances from the place of resi- contributing to the cancer morbidity. Therefore proximity to the dence of each cancer patient to the nearest major road and to the sea

Fig. 2. Estimated relative densities of lung and NHL cancers in the study area (number of cancer patients per 100,000 residents in each location), A — lung cancer; B - NHL cancer. 270 M. Zusman et al. / Science of the Total Environment 441 (2012) 265–276

Fig. 2 (continued).

shore were estimated as nearest aerial distances, using the “spatial join” we included in the multivariable analysis the percent of Jewish popu- tool of the ArcGIS9.x™ software. lation in the SCA as an additional explanatory variable. High population density and housing density (measured as the The population density data were obtained from the population number people per km2 and the number of people per room in the registry, whereas housing density and percent of Jewish population household, respectively) are also known as potential risk factors for data were obtained from the Israel Central Bureau of Statistics cancer incidence as surrogate measures of income and poverty (ICBS, 2008b). Smoking rates data were obtained from the Israel Na- (Fuller-Thomson et al., 2000). Lastly, in order to control for ethnicity, tional Health Interview Survey (INHIS) carried out by the Israel M. Zusman et al. / Science of the Total Environment 441 (2012) 265–276 271

120

100

Relative density (mean) 20

0 0 - 200 200 - 300 - 400 - 500 - 600 - 700 - 800 - 900 - 1000- 1100- 1200- 1300- 1400- 300 400 500 600 700 800 900 1000 1100 1200 1300 1400 1500 Distance (meters) Lung cancer NHL cancer

Fig. 3. Changes in the relative density of lung and NHL cancer cases (per 100,000) as a function of distance from the Kiryat Haim industrial zone.

Center for Disease Control (ICDC, 2007). While population density data 4. Results were available, as previously mentioned, for 50×50-m raster cells (see Section 3.4 on Kernel and Double Kernel Density estimates), housing 4.1. ASR differences between proximity bands densities and smoking rates were available for Small Census Areas (SCAs) only. Therefore, each cancer patient was assigned the values of Tables 1 and 2 report result differences in cancer incidence as a described above variables based on the values observed in areal units function of distance from the industrial zone, calculated separately the patient's house falls inside (either raster cell for population density using ASRs (Table 1) and DKD estimates (Table 2). or SCAs for housing density and smoking rates). As Table 1 shows, no significant differences in ASR values are detected among the two types of cancer under study (χ2 b2.62; P>0.270). This may be explained by a small number of SCAs (13), 3.7. Statistical analysis which is likely to reduce the statistical power of the test. In contrast, Table 2, which reports results calculated using the DKD method, Statistical analysis was performed in two stages. First, we exam- shows significant differences in the relative density across three dis- ined differences of cancer incidence rates across three proximity tance groups for both cancers under study: lung cancer (b65 yo: bands (0–500, 500–1000 and 1000+ m) from the industrial zone. χ2 =132.95, Pb0.001; 65+yo: χ2 =92.84, Pb0.001 respectively) Next, we ran a multivariate regression analysis to investigate the fac- and NHL (b65 yo: χ2 =139.95, Pb0.001; and 65+yo: χ2 =244.86, tors affecting cancer ASRs. The multivariable analysis was performed Pb0.001). As Fig. 3 confirms, DKDs of both cancers indeed drop separately for two age groups (“younger than 65” and “65 and older”) in order to adjust for the age factor. The analysis was performed in the SPSS 17.0™ software.

Table 4 Relative density of cancer cases (per 100,000) vs. distance to the Kiryat Haim industrial Table 3 zone (method — OLS regression; distance variable — linear distance term). Cancer ASRs vs. distance to the Kiryat Haim industrial zone (method — multivariate re- — a — a gression; proximity variable — linear distance term). Variable Age group younger than 65 Age group 65 and older Bb (tc)Bb (tc) Variable Age group — younger than 65a Age group — 65 and oldera

b c b c A. Lung B (t )B(t ) ⁎⁎ ⁎⁎ Constant 620.43 (10.88 ) 6037.63 (9.58 ) ⁎⁎ ⁎⁎ A. Lung D_industry (meters) −1.11E−02 (−5.31 ) −2.68E−01 (−11.61 ) Constant 406.74 (0.73) 152.65 (0.18) R2 0.206 0.334 ⁎⁎ ⁎⁎ D_industry (m) −1.97E−03 (−0.17) −1.54E−03 (−0.09) F 36.851 70.924 R2 0.638 0.477 No of refer. points 1000 1000 F 1.259 0.651 No of SCAs 13 13 B. NHL ⁎⁎ ⁎⁎ Constant 611.25 (9.53 ) 4912.85 (9.73 ) ⁎⁎ ⁎⁎ B. NHL D_industry (meters) 2.94E−02 (12.49 ) −1.38E−01 (−7.45 ) Constant 292.99 (0.82) 458.13 (0.88) R2 0.399 0.438 ⁎⁎ ⁎⁎ D_industry (m) 6.55E−03 (0.87) −9.53E−03 (−0.87) F 94.244 110.511 R2 0.691 0.605 No of refer. points 1000 1000 F 1.597 1.093 a Adjusted for distance to main roads, population density, smoking rate, average in- No of SCAs 13 13 come, % of Jewish population and housing density. a Adjusted for distance to main roads, population density, smoking rate, average in- b Regression coefficient. come, % of Jewish population and housing density. c t-Statistic (in the parentheses). b Regression coefficient. ⁎ Indicates a two-tailed 0.05 significance level. c t-Statistic (in the parentheses). ⁎⁎ Indicates a two-tailed 0.01 significance level. 272 M. Zusman et al. / Science of the Total Environment 441 (2012) 265–276

Table 5 As Table 5 shows, the “non-linearity adjusted” models appear to Relative density of cancer cases (per 100,000) vs. distance to the Kiryat Haim industrial provide a better fit and a higher degree of generality, compared to — — zone (method OLS regression; distance variables linear and quadratic distance the original models (R2 =0.228–0.609 vs. R2 =0.206–0.434; and terms). F=29.208–154.277 vs. F=36.851–108.714, respectively; see Tables 4 Variable Age group — younger than 65a Age group — 65 and oldera and 5). Bb and (tc)Bb and (tc) Characteristically, the models reported in Table 5 show, in general, the same tendencies as the “base” models reported in Table 4, viz. A. Lung ⁎⁎ ⁎⁎ fi Constant 627.51 (10.92 ) 4951.76 (8.66 ) asigni cant decline in DKD of both lung and NHL cancer patients in ⁎⁎ ⁎⁎ D_industry −4.32E−02 (−6.25 ) −1.03E+00 (−14.97 ) line with increasing distances to the industrial zone. Characteristi- ⁎⁎ ⁎⁎ D_industry^2 2.19E−03 (4.91 ) 4.85E−04 (10.92 ) cally, this effect is stronger for older patients (65+yo) than for R2 0.228 0.475 ⁎⁎ ⁎⁎ younger ones. Although in the younger group of NHL patients, “prox- F 29.208 89.428 ” N of refer. points 1000 1000 imity to industrial zone remains to be positively associated with the distance variable, this association is no longer statistically significant B. NHL (P=0.85). ⁎⁎ ⁎⁎ Constant 502.21 (8.39 ) 5049.25 (10.26 ) ⁎⁎ D_industry 1.40E−03 (0.19) −5.05E−01 (−8.53 ) ⁎⁎ ⁎⁎ 4.3. Sensitivity test D_industry^2 1.53E−05 (3.29 ) 2.39E−04 (6.25 ) R2 0.499 0.489 ⁎⁎ ⁎⁎ F 98.691 94.610 To estimate change in the relative density of lung and NHL N of refer. points 1000 1000 cancers in response to plausible changes in industry proximity, a a Adjusted for distance to main roads, distance to the sea, population density, sensitivity test was performed. The test was based on the models smoking rate, average income, % of Jewish population, housing density and the location reported in Table 5. In the calculation, the values of all other predic- of old industrial zone. tors (except for distance to the industrial zone) were set to their b Regression coefficient. c mean values observed in the study area as a whole. The results of t-Statistic (in the parentheses). ⁎ Indicates a two-tailed 0.05 significance level. the test are reported in Table 6. As the table shows, the distance ⁎⁎ Indicates a two-tailed 0.01 significance level. from the industrial zone increases from 0 to 1000 m, the relative density of lung cancer patients is estimated to decrease, ceteris paribus, by 48% in the younger cohort (b65 yo) and by nearly 81% in the older population group (65+yo). For the NHL cancer, the corresponding decrease is estimated to reach 2% and 57%, respectively with distancing from the industrial zone, especially in the 0–700 m (see Table 6). distance band. 5. Discussion

4.2. Multivariate analysis In the present study we estimated cancer incidence rates in the residential neighborhoods adjacent to the Kiryat Haim industrial Tables 3 and 4 report factors affecting cancer incidence of lung zone area in Northern Israel using two different methods (ASR cal- and NHL cancers estimated for the two age groups separately culation and Double Kernel Density method). The ASR calculation (ASRs in Table 3 and relative density in Table 4). The models are method did not detect any significant association between ASRs estimated using multivariate regression analysis and adjusted for of either cancer and industrial zone proximity, which may be distance to the main roads, population density, smoking rate, av- explained by a small number of SCAs available for the analysis erage income, percent of Jewish population and housing density (13), which naturally reduces the model power and generality. In explanatory factors. Table 5 reports additional models which contrast, differences in relative density of both lung and NHL can- account for non-linearity of relations between cancer DKD and in- cer patients with distancing from the industrial zone emerged as dustrial zone proximity and include a quadratic term of proximity statistically significant when DKD estimates were used. This find- and several other explanatory variables discussed, in some detail, ing corresponds to those of previous studies (Kloog et al., 2009; below. Portnov et al., 2009) in which kernel density approach was found Multiple regression analysis of ASR estimates shows, as expected, to perform better in detecting potential cancer risks than tradi- a very low degree of generality (Fb1.597ns) and, in none of the tional zonal approaches based on ASR rates in census designated models, the industrial zone proximity emerged as statistically signif- areas. icant (P>0.5; see Table 3). One of the possible explanations why ASR calculation was found In contrast, as Table 4 shows, the distance to the industrial zone ap- to be less relevant may stem from a single standardized rate it pears to be significantly and negatively associated with DKD of cancer assigns to all people living in a specificzone(Yu et al., 2006). patients in both age groups of lung cancer patients (b65 yo: t=−5.31, This kind of summary measure can hide patterns that might have Pb0.01 and 65+yo: t=−11.61, Pb0.01) and in older NHL patients important health implications (Schoenbach and Rosamond, 2000). (65+yo: t=−7.45, Pb0.01). In contrast, the “b65 yo” group of NHL As previously noted, in the present study, only 13 census tracts patients shows, somewhat surprisingly, a significant positive association were available for ASR calculations. In addition, two SCAs located (t=12.49, Pb0.01), indicating that, ceteris paribus, DKD of NHL patients within a distance of 1500 m from the industrial zone had no cancer in this age group tends to increase with increasing distance from the in- cases, which reduced the number of valid observations even dustrial zone. further. This peculiar trend necessitated an analysis of regression resid- Emissions from industry can cause different health disorders uals. The analysis reveals a clear clustering of residuals at the distance depending on such factors as the amount and the length of expo- of approximately 1500 m from the industrial zone (see Fig. 4). To ac- sure (IARC, 1998; Sexton et al., 1995), and abandoned industry count for this non-linearity we added the quadratic term of proximi- sites can also continue to contaminate through soil, ground ty, as well as “distance to the sea” and “location of the old industrial water and air pathways. Such exposure can lasts for decades slow- zone” to the “base” model (see Table 4). The results of the repeated ly contributing to increasing risks of diseases, including cancer. As analysis are reported in Table 5. we have observed in our study (see Figs. 1–3), both lung and NHL M. Zusman et al. / Science of the Total Environment 441 (2012) 265–276 273

Fig. 4. Spatial clustering of unstandardized residuals from the lung cancer model. Note: based on Table 4 (65+ age group).

cancer density rates appeared to peak at the distance of some The results of the present study correspond to the ﬁndings from 1400–1500 m from the existing industrial zone. Today, no other studies which observed the association between exposure to apparent pollution sources are located there. However, there are petrochemicals and the risk of lung and NHL cancers. Thus, in their some indications that this area was an industrial and commercial 2004 study, Belli et al. (2004) showed an increase in the risk of zone during the period of the British Mandate (that is, be- lung cancers among residents living within 2-km distance range fore 1948), which warrants further investigation in follow-up from the petrochemical plant in Brindisi, Italy after adjusting for studies. smoking habit, occupation and educational level. In a separate 274 M. Zusman et al. / Science of the Total Environment 441 (2012) 265–276

Table 6 particular, ASR method was found to be insufﬁciently sensitive Changes in the relative density of lung and NHL cancers in response to plausible for estimating differences in cancer incidence rates providing a changes in industry proximity (sensitivity test). very small number of statistical subdivisions available for the Distance Age group — younger than 65 Age group — 65 and older study (13), thus demonstrating a potential weakness of the range (m) Relative density (cases % change Relative density % Change ASR-based approach. In contrast, Kernel Density approach was per 100,000) (cases per 100,000) found to be more sensitive and informative than ASR calculations.

A. Lung cancer The latter method can thus be recommended when a researcher 200 43.81 872.61 has a small number of geographic units for a cross-sectional 400 35.17 −19.72 666.77 −23.59 study, because of its sensitivity. However, further studies with ad- 600 26.53 −24.56 460.93 −30.87 ditional individual-level factors and more precise exposure assess- 800 17.90 −32.55 255.09 −44.66 ments may need to be carried out in order to control for potential 1000 9.26 −48.26 49.25 −80.69 confounders. B. NHL cancer 200 17.76 479.62 400 18.04 1.57 378.70 −21.04 Disclaimer and competing interest declaration 600 18.32 1.55 277.78 −26.65 800 18.60 1.53 176.87 −36.33 The content of this paper is the responsibility of its authors and − 1000 18.88 1.50 75.95 57.06 does not necessarily reﬂect the view of the Israel Ministry of Health. Note: Based on models in Table 5; the mean values of control variables are set constant The authors declare hereby that they have no competing ﬁnancial to the average values observed in the study area as follows: distance to main roads= interests. 76.30 m; distance to the sea=1116 m; population density (per km2)=50,363.00; smoking rate (%)=11.00; average income (NIS per capita)=2248.05; % of Jewish population=98.23; and housing density (people per room)=0.93. Acknowledgments

The first author thanks the Environment and Health Fund (Jerusa- lem, Israel) for the Doctoral Fellowship support. The authors are also study, Taiwanese researchers reported that women living in the grateful to two anonymous reviewers for their helpful comments and municipalities characterized by higher levels of petrochemical pol- suggestions. lution had a statistically significant higher risk of developing lung cancer than those who lived in municipalities with the lowest petrochemical air pollution levels (Yang et al., 1999). In another recent study, de Roos et al. (2010) observed an elevated Appendix 1. Descriptive statistics of the research variables odds ratio of NHL although insignificant among residence near the petroleum industries. However, the results of the present study are different from the results of Patel et al. (2004) who did not find the association between gasoline vapor exposure from underground storage tanks and lung or NHL cancers but among leukemia patients. It Variables N. Minimum Maximum Mean Std. cases deviation should be noted, however, that compared to our study, most of the above studies analyze the exposure to chemicals in residential areas Lung cancer Men 68 –– –– neighboring petrochemical plants but not around petroleum-storage Women 30 –– –– facilities. In addition, most of these studies used zone-based ap- Age 98 32.00 91.00 69.12 11.83 proaches for estimating the exposure and calculating the relative Distance to industrial 98 103.33 1514.12 723.70 456.39 risk of health outcomes. zone, m Our findings are also consistent with the tendency from NCI's SEER Distance to the sea, m 98 121.73 2193.96 984.13 574.23 Distance to the closest 98 0.02 286.62 56.36 68.92 program report, according to which starting at 65 years of age, ap- main road, m proximately one in two men and one in three women will develop Average income, NIS 98 1320.92 3368.76 2112.18 573.07 cancer (Hayat et al., 2007). per capita A marked limitation of the present study is that it is an ecological- Jewish population, % 98 93.63 99.63 97.49 1.93 Population density, 98 18,000.00 79,000.00 50,163.27 14,561.42 type study in which socio-demographic attributes (such as smoking per km2 and socioeconomic status) were obtained for geographic units and Housing density, people 98 0.79 1.05 0.93 0.06 not for individuals. A well-known limitation of such studies is that per room such studies cannot measure causality but can only be indicative of Smoking rate, % 98 0.00 0.19 0.11 0.05 possible health risk effects. NHL cancer In addition, proximity to storage tanks as a proxy for exposure Men 30 –– –– may also introduce an “anisotropy” bias because chemicals may not Women 23 –– –– spread uniformly away from the source. The anisotropy of spatial dis- Age 53 34.00 97.00 66.64 13.27 tribution of chemicals released from petroleum storage tanks may Distance to industrial 53 104.41 1567.71 807.05 456.40 zone, m also depend on the height of pollution source and wind direction, in Distance to the sea, m 53 121.46 1826.38 735.14 494.97 addition to the properties of the specific type of chemical and the Distance to the closest 53 0.02 233.53 64.76 60.54 amount released. This potential anisotropy effect should thus be in- main road, m vestigated in follow-up studies. Average income, NIS 53 1320.92 3368.76 1947.59 416.82 per capita Jewish population, % 53 93.63 99.63 97.27 1.62 Population density, 53 24,000.00 73,000.00 44,943.40 11,478.10 6. Conclusions per km2 Housing density, people 53 0.80 1.05 0.95 0.05 The results of the study reveal that different estimates of relative per room Smoking rate, % 53 0.00 0.19 0.13 0.04 risks depend on different exposure assessment methods used. In M. Zusman et al. / Science of the Total Environment 441 (2012) 265–276 275

Appendix 2. Randomly generated reference points (see text for explanation)

References Agency for Toxic Substances and Disease Registry [Internet]. Public Health State- ment for Total Petroleum Hydrocarbons (TPH). Atlanta; 1999 [updated 2011 Ahmad O, Boschi-Pinto C, Lopez AD, Murray CJL, Lozano R, Inoue M. Age standardiza- March 3]. ATSDR. Available from: http://www.atsdr.cdc.gov/phs/phs.asp?id= tion of rates: a new WHO standard. World Health Organization: Geneva; 2001. 422&tid=75. 276 M. Zusman et al. / Science of the Total Environment 441 (2012) 265–276

Barchana M, Liptshitz I, Fishler Y, Green M. Geographical mapping of malignant dis- Lewtas J. Air pollution combustion emissions: characterization of causative agents and eases in Israel 2001–2005. Israel National Cancer Registry: Jerusalem; 2007. mechanisms associated with cancer, reproductive, and cardiovascular effects. Belli S, Benedetti M, Comba P, Lagravinese D, Martucci V, Martuzzi M, et al. Case– Mutat Res 2007;636(1–3):95-133. control study on cancer risk associated to residence in the neighborhood of a pet- Li L, Lian ZW. Application of statistical power analysis—how to determine the right rochemical plant. Eur J Epidemiol 2004;19(1):49–54. sample size in human health, comfort and productivity research. Build Environ Bidoli E, Franceschi S, Montella M. Cancer mortality by urbanization and proximity to 2010;45(5):1202–13. the sea coast in Campania Region, southern Italy. Tumori 1998;84(4):460–6. Liu CC, Chen CC, Wu TN, Yang CY. Association of brain cancer with residential exposure Blundell GM, Maier J, Debevec EM. Linear home ranges: effects of smoothing, sample to petrochemical air pollution in Taiwan. J Toxicol Environ Health A 2008;71(5): size, and autocorrelation on kernel estimates. Ecol Monogr 2001;71:469–89. 310–4. Busby C. Nuclear pollution, childhood leukaemia, retinoblastoma and brain tumours in McCoy J, Johnston K. Using ArcGIS spatial analyst. Redlands: ESRI; 2001. Gwynedd and Anglesey Wards near the Menai Straits, North Wales 2000–2003. Municipality of Haifa. Research and Statistical Information Department: Neighborhood Aberystwyth: Green Audit; 2004 [Jan. Bangor Report for HTV]. Profile, 2008. Haifa, Israel: MH; 2010. Carlos H, Shi X, Sargent J, Tanski S, Berke EM. Density estimation and adaptive band- Nakaya T, Yano K. Visualising crime clusters in a space–time cube: an exploratory data widths: a primer for public health practitioners. Int J Health Geogr 2010;9:31–9. analysis approach using space–time kernel density estimation and scan statistics. Clapp RW, Jacobs MM, Loechler EL. Environmental and occupational causes of cancer: Trans GIS 2010;14(3):223–39. new evidence 2005–2007. Rev Environ Health 2008;23(1):1-37. National Cancer Institute. Cancer: questions and answers cancer and the environment: Christie D, Robinson K, Gordon I, Bisby J. A prospective study in the Australian petro- what you need to know, what you can do, what you need to know about cancer. leum industry. II. Incidence of cancer. Br J Ind Med 1991;48(8):511–4. Bethesda, MD: NCI; 2003. Curtin LR, Klein RJ. Direct standardization (age-adjusted death rates). Hyattsville: Nerriere E, Zmirou-Navier D, Desqueyroux P, Leclerc N, Momas I, Czernichow P. Lung Maryland: National Center for Health Statistics; 1995 [Statistical Notes; No.6]. cancer risk assessment in relation with personal exposure to airborne particles in De Roos AJ, Davis S, Colt JS, Blair A, Airola M, Severson RK, et al. Residential proximity to four French metropolitan areas. J Occup Environ Med 2005;12:1211–7. industrial facilities and risk of non-Hodgkin lymphoma. Environ Res 2010;110(1): Pan BJ, Hong YJ, Chang GC, Wang MT, Cinkotai FF, Ko YC. Excess cancer mortality 70–8. among children and adolescents in residential districts polluted by petrochemical Donthu N, Rust RT. Estimating geographic consumer densities using kernel density es- manufacturing plants in Taiwan. J Toxicol Environ Health 1994;43(1):117–29. timation. Market Sci 1989;2:191–203. Patel AS, Talbott EO, Zborowski JV, Rycheck JA, Dell D, Xu X, et al. Risk of cancer as a re- Dubnov J, Barchana M, Rishpon S, Leventhal A, Segal I, Carel R, et al. Estimating the ef- sult of community exposure to gasoline vapors. Arch Environ Health 2004;59(10): fect of air pollution from a coal-fired power station on the development of 497–503. children's pulmonary function. Environ Res 2007;103(1):87–98. Paz S, Linn S, Portnov BA, Lazimi A, Futerman B, Barchana M. Non-Hodgkin lymphoma ESRI [Internet]. ArcGIS Desktop Help 9.3. [cited 2009]. Available from: http://webhelp. (NHL) linkage with residence near heavy roads—a case study from Haifa Bay, esri.com. Israel. Health Place 2009;15(2):636–41. Fuller-Thomson E, Hulchanski JD, Hwang S. The housing/health relationship: what do Portnov BA, Dubnov J, Barchana M. Studying the association between air-pollution and we know? Rev Environ Health 2000;15:109–33. lung cancer incidence in a large metropolitan area using a kernel density function. Harada Y, Shimada T. Examining the impact of the precision of address geocoding on Soc Econ Plann Sci 2009;43(209):141–50. estimated density of crime locations. Comput Geosci 2005;32:1096–107. Sanvicente-Añorve L, Salgado-Ugarte IH, Castillo-Rivera M. The use of kernel density Hayat MJ, Howlader N, Reichman ME, Edwards BK. Cancer statistics, trends, and multi- estimators to analyze length-frequency distributions of fish larvae. In: Browman ple primary cancer analyses from the Surveillance, Epidemiology, and End Results IH, Skiftesvik AB, editors. The big fish bang: proceeding of the 26th annual larval (SEER) program. Oncologist 2007;12:20–37. fish conference; 2003; Institute of Marine Research. Bergen: Norway; 2003. Hollenbeck JR, DeRue DS, Mannor M. Statistical power and parameter stability when p. 419–30. subjects are few and tests are many: comment on Peterson, Smith, Martorana, Sexton K, Callahan MA, Bryan EF. Estimating exposure and dose to characterize health and Owens (2003). J Appl Psychol 2006;91:1–5. risks: the role of human tissue monitoring in exposure assessment. Environ Health International Agency for Research on Cancer. IARC monographs on the evaluation of Perspect 1995;103(3):13–29. the carcinogenic risk of chemicals to humans: occupational exposures in petro- Schoenbach VJ, Rosamond WD. Understanding the Fundamentals of Epidemiology: an leum refining. Lyon, France: IARC; 1998. evolving text. Fall 2000 ed. North Carolina: Chapel Hill; 2000. Israel Central Bureau of Statistics. Statistical abstract of Israel: emissions from fuel com- Tsai SS, Tiao MM, Kuo HW, Wu TN, Yang CY. Association of bladder cancer with resi- bustion, by type of consumer. Jerusalem, Israel: ICBS; 2008a. dential exposure to petrochemical air pollutant emissions in Taiwan. J Toxicol Israel Central Bureau of Statistics. Statistical abstract of Israel 2007: population, by pop- Environ Health A 2009;72(2):53–9. ulation group, religion, sex and age. Jerusalem, Israel: ICBS; 2008b. World Health Organization. Environmental health inequalities in Europe. Assessment Israel Center for Disease Control. The Israel National Health Interview Survey, report. Regional Office for Europe: WHO; 2012. 2005–2006. Jerusalem, Israel: ICDC; 2007. Wilkinson P, Thakrar B, Walls P, Landon M, Falconer S, Grundy C, et al. Israel Ministry of Environmental Protection [Internet]. Industrial pollution. [updated Lymphohaematopoietic malignancy around all industrial complexes that include 2009]. Available from: http://www.sviva.gov.il. major oil refineries in Great Britain. Occup Environ Med 1999;56(9):577–80. Israel Ministry of Environmental Protection. Environmental bulletin 2008. Jerusalem, Xie Z, Yan J. Kernel density estimation of traffic accidents in a network space. Comput Israel: IMEP; 2008. Environ Urban Syst 2008;32:396–406. Jarvholm B, Mellblom B, Norrman R, Nilsson R, Nordlinder R. Cancer incidence of Yang CY, Cheng MF, Chiu JF, Tsai SS. Female lung cancer and petrochemical air pollution workers in the Swedish petroleum industry. Occup Environ Med 1997;54(9): in Taiwan. Arch Environ Health 1999;54(3):180–5. 686–91. Yang CY, Chiu HF, Chiu JF, Kao WY, Tsai SS, Lan SJ. Cancer mortality and residence near Kirkeleit J, Riise T, Bratveit M, Moen BE. Increased risk of acute myelogenous leukemia petrochemical industries in Taiwan. J Toxicol Environ Health 1997;50(3):265–73. and multiple myeloma in a historical cohort of upstream petroleum workers ex- Yu CL, Wang SF, Pan PC, Wu MT, Ho CK, Smith TJ, et al. Residential exposure to petro- posed to crude oil. Cancer Causes Control 2008;19(1):13–23. chemicals and the risk of leukemia: using geographic information system tools to Kloog I, Haim A, Portnov BA. Using kernel density function as an urban analysis tool: estimate individual-level residential exposure. Am J Epidemiol 2006;164(3): investigating the association between nightlight exposure and the incidence of 200–7. breast cancer in Haifa, Israel. Comput Environ Urban Syst 2009;33:55–63.