Guidelines for Examining Unusual Patterns of Cancer and Environmental Concerns - Audio Description

Presentation slides on a screen recording. Blue text on a white background. Guidelines for examining unusual patterns of cancer and environmental concerns. Spatial clusters training. National Center for Environmental Health. Agency for Toxic Substances and Disease Registry. Logos CDC ATSDR a sound icon. Welcome to the spatial analysis portion of the cancer cluster presentation. I am Frank Curiero and in this training we'll be covering the basics for analyzing a spatial cluster. New slide training overview bullet points clustering versus cluster detection data and software examples considerations. We will cover several topics in this training starting with an important definition to lay the groundwork for this training. We will discuss the differences between clustering and cluster detection. We will go over the different software and data used in this training and cover some examples so you can see a preview of what a real life spatial cluster analysis might look like. And finally, we will conclude the presentation and give you some parting considerations and thoughts. New slide titled spatial autocorrelation bullet points. Spatial autocorrelation is the degree of similarity of nearby features. Positive means that features nearby one another have similar values. Negative signifies nearby features that have dissimilar values. Five images with distributions of small green and blue shapes on a scale from dispersed with a minus sign and clustered with a plus sign. Text image from https/estop.gis.com/n/cap gi.com/n/ a rc cm a p a t s t/ o ls/ s p a t i a l hy-en s t a t i s t i cs hyphen t o l b o x/ s p a t i a lyen a u t o c o r e l a t io n.htm htm. An important definition is that of spatial autocorrelation, which is the degree of similarity of nearby features. In the context of cancer cluster analysis, this would mean how similar the number of cancer cases or cancer rates are in each geographic unit. Spatial order correlation can either be positive or negative with values between minus1 and one. If the autocorrelation is close to one, this would mean that the area is very tightly clustered with all geographic units being near each other having very similar cancer rates. In the image below, let's assume that the shaped colors represent cancer rates with similar cancer rates demarcated by the same color. As you can see on the right side of the image, all the light green, dark green, and blue units are proximal to each other, representing near similar units, and we would expect a high autocorrelation. An autocorrelation of minus1 would mean all units are perfectly distributed with no similar features near each other like shown in the left side of the image where none of the colors are touching. New slide titled clustering versus cluster detection. Bullet points. Clustering generally refers to the concept of spatial correlation aka a spatial autocorrelation. Cluster detection generally refers to a set of methods that can be used to detect spatial clusters. An important distinction when performing an analysis would be to understand the differences between clustering and cluster detection and what they mean for your analysis. For census track level cancer rates, clustering would mean tracks that are closer together or adjacent have cancer rates that are more similar than tracks that are not close together or not adjacent. Spatial order correlation techniques are used to assess clustering such as man's eye that can assess clustering over the entire map and local versions such as local man's eye to assess where on the map clustering is stronger or weaker. The local versions are often referred to as local indicators of spatial association. or LISA LISA. Cluster detection is defined as a geographically bounded sub area on your map, such as a collection of adjacent tracks whose combined cancer rate is anomalous compared to what is going on elsewhere on the map. Anomalous can be on the high end like a hot spot or the low end like a cold spot. Similar descriptions hold when dealing with spatial point level data such as address level data. New slide titled examples bullet points. Does clustering exist? Examples. Qualitative brackets visualization. Use cororopath maps to visually examine the data. Quantitative bracket spatial autocorrelation. Use local man's eye to visually check for high and low areas of clustering and global clustering. Do spatial clusters exist? Cluster detection. Use SAT scan to check for significant cancer clusters. For this training, we will be running through a couple examples outlining techniques described in the guidelines. Visualizing data is an essential part of understanding your data and we will guide you through one way to visualize your cancer data using qualitative methods. First, we will be looking at whether clustering exists using quantitative methods, showing the use of local Moranzi to visually check for high and low areas of clustering. Last, we will show an example of cluster detection using the software SAT scan to find statistically significant cancer clusters. New slide titled examples bullet points software free and open-source software tool that serves as an introduction to spatial data science. SAT scan free cluster detection software that analyses spatial, temporal, or space-time scan statistics data used in this training. Pennsylvania prostate cancer census tract level 2010 to 2019 brackets tracking network Georgia lung cancer 5,000 population area groups 2006 to 2018 tracking network we're going to show a few examples of these techniques for these examples we will be using two different software that are freely available for download and satscan is an open- source tool with great documentation that will allow you to perform a large var variety of qualitative and quantitative analysis. SATscan is a cluster detection software that can analyze spatial, temporal, and spac-time data using the scan statistic methodology. While these are the tools being covered in this presentation, there are likely similar tools in your preferred platform such as Python or R. The data sets used for this study come from the data that is available on the CDC environmental public health tracking portal. We will use the aggregated data to show examples of these tools. Pennsylvania prostate cancer rates and standardized incidence ratios, SIRS, are available at the census track level for a 10-year aggregation 2010 to 2019. Georgia lung cancer rates and SIRS are available at the 5,000 population aggregation, which is part of the tracking network subcount aggregations. A point to stress. All analyses, results, and interpretations are for the sole purpose of demonstrating the software and processes that can be involved in a cancer cluster analysis. They should only be viewed in this context. New slide titled, "Does clustering exist? Qualitative coroplith maps. A map with various orange and yellow areas. Bullet points. Use color to display different sin values across spatial units. Visualize how a variable varies across geographic area. We will first be looking at whether clustering exists using qualitative and quantitative methods showing the use of local man's eye to visually check for high and low areas of clustering. First, visualizing data is an essential part of understanding your data and we will guide you through one way to visualize your cancer data using cororathth maps. Corroplath maps help you display differences in values across spatial units. Corpleth maps are thematic maps that use color to indicate the value of individual geographic units. For example, in a cancer analysis, you could map the cancer rates by census track or measure the level of a coariant per census track for comparisons. A closeup of a map with different areas in different colors. Text prostate cancer rate Pennsylvania 2010 to 2019. Census tract image source geodenter local autocorrelation tutorial. https colon/geenter.git hub.io/w k6a l o c a u t o/lab 6a.html. For example, this coral map was created using GOTA. Age adjusted prostate cancer rates by census track in Pennsylvania were categorized using manual brakes for the categories with the higher rates being displayed in darker colors. New images. Three maps with varying green shaded areas. First for 2006 to 2010, second from 2010 to 2014. Third from 2014 to 2018. Text lung cancer rate Georgia 5,000 population areas. link https/www.cdc.gov/ncce/tracking/topics/geographies.htm. This chloroplath map shows lung cancer rates in Georgia by the 5,000 population areas that the CDC environmental public health tracking program created to increase population counts within small geographies. If interested in learning more about that, the link is provided below. It shows the differences in rates over 5-year periods, although some of the years of data overlap. This map helps to show where rates may have increased or decreased over the 15-year period. A small shaded green section of each map is highlighted. Gota allows you to click and explore on multiple connected maps. Here we have highlighted one geography that has moved from the third rate level to the highest rate level that we have categorized manually. It also allows you to zoom in and take a look at the neighboring areas. New slide titled does clustering exist? Quantitative/spatial autocorrelation. Local mirin's eye bullet points. A measure of spatial autocorrelation of individual spatial units. GIS software will produce a map highlighting areas of high high low high and high low clustering. Next we will look at measures of spatial order correlation using local Moran's eye. When performing a local Moranzi, we are interested in not just whether cancer rates are high or low, but the relationship between the neighboring geographic areas. GIS software will produce a map highlighting areas in reds and blues that are called high high, low, low high, and high low. When we say high high, this means that there are geographic areas that have high values next to other geographic areas with high values. So if the area of interest has a high cancer rate, we should ask are the surrounding areas high as well or low. This will result in one of the four combinations stated. New bullet points titled local versus global local Moran's I evaluates the statistical significance of all combinations of high and low cancer at a local level brackets i.e. block, county, tracted, etc. Close brace global Moran's eye gives you the overall spatial autocorrelation levels between minus1 to one. Along with local Moranz eye, we will be doing a global Moranz eye. While a local Moranzi is interested in individual geographic units, the global Moranzi looks at the entire area to check for overall autocorrelation of the data. The value will be between minus1 and one like a typical correlation value. New slide. A table and two maps titled connectivity maps of census tracks in Pennsylvania created in geod data. The table's two columns are titled property and value. The map above shows a large shaded green area with a small set of white sections within and one dark green section. The map below shows all the same sections shaded dark green. Link https/geodatacenter.github.io/workbooks/4a_cont io/workbooks/4a_contig weights/lab4a.html. When starting this analysis in Gota, it is important to first identify a waiting matrix which identifies the nearest neighbors of each census track or geography. More can be found at the link provided below. Once the logo's eye is run, several maps are produced. The one on the slide is a connectivity map. It allows you to check which tracks are connected to each other. On the top image, you can see the neighboring tracks highlighted and the connection lines in the image below. More details about how to do this are on the GEOTA website. New slide titled Univariat Moran's eye analysis, Pennsylvania. Two maps and a graph. The top map has white sections and a few areas shaded in blues and reds. The lower map has the same shape and sections with shaded green areas rather than blue and red. The graph shows a diagonal line across a Y and X-axis with clusters of blue dots. Link https/geodata center.gitub.io/workbook a local_autolab 6.html. This slide shows the results of the local Moran's eye. The map on the top is showing the high high, low low, etc. spots that we mentioned previously. The red areas are tracks that have high rates and have neighbors whose rates are also high. The areas in blue are showing tracks that have low rates and also have neighbors with low rates. The lower map with green colors is the statistical significance map for the top map. Levels of significance, for example, 0.05 or 0.01 are shown in darker greens. It is important when looking at the significance map to note what level of significance you're interested in. GIOTA offers some important guidance to consider because of the multiple comparisons where a p value of 0.05 may not be appropriate to use. The graph on the right side of the slide shows a Moran scatter plot and the global spatial order correlation value of 268 at the top indicating fairly low positive spatial order correlation. Close-ups of the two maps and the graph. An outlined square sits in the top right quadrant of the graph. The matrix table allows you to look at specific geographies that may be of interest. Here we have highlighted some of the outlying points in the high high quadrant. Three maps with varying blue and red shaded areas. First for 2006 to 2010. The second from 2010 to 2014. Third from 2014 to 2018. Similarly, these are some results in Georgia where we are showing the importance of looking at changes over time. Here we can see some areas where the rates are high. Neighboring geographies are also high which point to some places to potentially examine the data further. New slide titled do spatial clusters exist? Cluster detection bullet points. SAT scan. SAT scan is a free software that analyses spatial, temporal of space-time cluster detection analyses. It is designed for any of the following interrelated purposes. Geographical surveillance, random distribution, evaluate cluster significance, time periodic disease surveillance. After running our quantitative and qualitative clustering analysis, we can move on to cluster detection techniques. For this analysis, we'll be using SATScan, which uses Caldorf's spatial scan statistics. SATscan helps to detect spatial or space-time disease clusters to see if they are statistically significant. The test essentially assumes whether disease is randomly distributed over space, over time, or over space and time. SAT scan can also perform prospective disease surveillance and can be applied to a variety of data types. New text bullet points. How do I use SAT scan? The SAT scan website contains a very helpful tutorial that can lead you through the basics of identifying cancer clusters. Link https/www.satscan.org/tutorials.html. An image with text and text boxes. SATscan tutorial number one purely spatial plus scan statistic for cancer incidents. If you are interested in a more thorough explanation, SASCAN has great documentation that can guide you through the process. We're using SATSCAN today because it is free, widely used, and has a lot of applications to other fields. New slide titled SAT scan analysis of PA prostate cancer. A screen grab of a window titled SAT scan with a map showing small red dots and shaded circular red areas for clusters. Bullet points input data case counts protract population protract bracket male 19 tracked centrid coordinates info on missing/suppressed data text SAT scan identified several areas for further evaluation brackets for demonstration purposes only. This slide shows output of running SAT scan for Pennsylvania prostate cancer at the census track level. cases are aggregated over the 10-year period 2010 to 2019. This analysis was run for prostate cancer rates and the data input included case totals per track, population male 18 plus population per track and track centrid coordinates. Also included is information for which tracks had their data suppressed due to small numbers, cases or population. Satscan can take this into account. As note, these images are for demonstration purposes only. The map shows 20 areas for further evaluation. SASCAN identified the rates within these areas as significantly higher than the rates outside the areas at a p value set of 0.05. Depending on your desired output, SATscan provides options for different types of graphical and textbased results. SATscan can provide HTML and KML files for quick and automatic mapping of results in Google Maps and Google Earth, as shown here, or shape file format, so users can import results into a GIS for more flexible mapping. The textbased output option from SATSCAN provides detailed information about each identified cluster. For example, as highlighted here, information about the Pittsburgh area cluster includes total number of cases, total population, a measure of relative risk, and the cluster p value, which you see is highly significant. This automatic HTML/KML mapping interface from SATscan allows a user to click on any cluster to see this information, all of which can be saved in textbased output files from SATSCAN. Additional text age adjusted analysis bullet points expected counts based on SIRIR why adjusted analysis. The previous analysis only adjusted for population at risk the male population 18 plus and we can understand the rationale for this. We expect more cases in areas where we have larger populations. The same can be said for known risk factors. For example, age is a risk factor for prostate cancer and we would expect more cases in areas with older male populations. So we can rerun the SAT scan analysis adjusting for the geographic distribution of age and identify clusters that cannot be explained by the male age population. To run this adjusted analysis in SAT scan, you'll need the age category by case breakdowns for each track as well as the population by age category. SATSCAN generates an expected count based on the cases age distribution. This is the same calculation that goes into calculating the standardized incidence ratio or SIR. And if those expected counts are available outside of SATSCAN, they can be brought in to run the adjusted analysis as was the situation here. After adjusting for the age distribution, SATSCAN identifies six significant clusters. Compared to the unadjusted analysis or the analysis that just adjusts for the population, the age adjusted analysis shows that the total number of clusters decreases from 20 to six. Some cluster areas decreased in size and some cluster areas increased in size. You can dive deeper into these interpretations, providing further insight into how the age distribution may have explained or concealed clusters of prostate cancer. New slide titled SAT scan analysis of GA lunch cancer. Brackets age adjusted three maps of the same area with red dots and shaded circled. First, seven age adjusted clusters for 2006 to 2010 gia lunch cancer. Second 8 adjusted clusters for 2010 to 2014 GIA lung cancer. Third, seven age adjusted clusters for 2014 to 2018 GIA lung cancer. We ran a similar analysis for the Georgia lung cancer. For these data, we had lung cancer outcome for the three time periods 2006 to 2010, 2010 to 2014, and 2014 to 2018, as well as expected counts based on age adjusted SIRS. Results from the age adjusted cluster detection analysis show similar numbers of clusters and in some consistent areas over the three time period. New slide titled summary bullet points. Does clustering exist? Qualitative visualize the data. Quantitative/spatial correlation gives you further insight into where high and low concentrations might be. Do clusters exist? Cluster detection identifies statistically significant clusters. In summary, we've shown a couple of different methods for analyzing spatial clusters, but these are not comprehensive. The important thing to note is that each of these steps build upon each other and can highlight various patterns within your data. The qualitative analysis gave us our initial look into the data and got us thinking about where higher rates may be occurring. We also identified areas where higher rates were near other areas with high rates. Finally, in our cluster detection analysis, we discovered statistically significant clusters where the observed values were higher than expected. New slide titled considerations bullet points for both cluster detection and clustering analyses. It is important to note that data analytic methods such as those described in the presentation are just one component of an assessment. For cluster detection, it bold is important to consider underlined statistical significance along with underlined practical significance. Random clustering is also possible for both cluster detection and clustering analyses. It's important to note that the data analytic methods such as those described in the presentation are just one component of an assessment for cluster detection. It's important to consider statistical significance along with practical significance. Identifying spatial clusters is more of a beginning in cancer cluster analysis than an end. This is particularly important as random clustering is always possible and not due to any particular cause. New slide titled additional data considerations or next steps. Bullet points. Population demographics, sociodemographic and environmental characteristics, behavioral risk factors, residential history, cancer characteristics such as sight and latency. Understanding the population demographics, sociodemographics, and environmental characteristics, for example, of identified clusters can be important as well as a multitude of other factors such as residential history, cancer site, and cancer latency. The methods in SAT scan allow for adjusted cluster detection analyses, that is to search for clusters that cannot be explained by other factors like known risk factors as was demonstrated. Such factors are not only restricted to those that are sociodemographic but can include for example environmental factors for clustering. Similar comments apply with interpretation to look into areas identified as being significantly similar. Two final things to add. If this information will be shared to the public, it needs to be placed in the proper context and should include a health communicator in discussions about the best way to relay the information. As a final consideration, the clustering analysis conducted in Gota and the cluster detection analysis performed in Satscan were for demonstration purposes only. These software packages have great supporting resources with web-based material, tutorials, and manuals. You are encouraged to utilize these resources when performing your own analysis. New slide text. Thank you. For more information, contact n/dr1-800- CDC-info brackets 232-4636 TTY1-888-232-6348 www.atsdr.cdc.gov www.cdc.gov gov. Follow us on Twitter at CDC Environment. The findings and conclusions in this report are those of the authors and do not necessarily represent the official position of the Centers for Disease Control and Prevention and the Agency for Toxic Substances and Disease Registry, Logos, CDC, ATSDR. Are