Sponsoring Section/Society: ASA-BIOMETRICS
Session Slot: 2:00- 3:50 Tuesday
Estimated Audience Size: 100-200
AudioVisual Request: one projection equipment for using laptop, others overhead projector
Session Title: Effects of Smoothing on Health Effects Maps
The National Center for Health Statistics (NCHS) has published and the
National Cancer Institute (NCI) will soon publish mortality atlases in
which maps by cause, race, and sex depict geographic patterns of disease
mortality. The NCHS atlas maps both raw and smoothed rates
(age-adjusted) , while it was decided that the NCI atlas will not
include smoothed rates for several reasons, one of which is the fear of
masking true ``hot spots'' caused by localized risk factors.
In light of the opposing decisions by NCHS and NCI on inclusion of
smoothed rates, I am organizing an ASA session for 1998 entitled
``Effects of Smoothing on Health Effects Maps.'' A rule-of-thumb
subscribed to by many is that smoothing is helpful for viewing
geographical patterns, but is not appropriate when local variation is
important to monitor. The session would focus on elucidating
appropriate methods and uses of smoothing for health effects maps and
for comparing smoothed and unsmoothed procedures for computing rates in
terms of performance criteria.
Theme Session: Yes
Applied Session: Yes
Session Organizer: Pennello, Gene National Cancer Institute DCEG/BB
Address: Gene Pennello National Cancer Institute DCEG/BB Executive Plaza North, Room 403 6130 Executive Blvd MSC 7368 Bethesda MD 20892-7368
Phone: (301)496-3344
Fax: (301)402-0081
Email: pennellg@epndce.nci.nih.gov
Session Timing: 110 minutes total (Sorry about format):
Opening Remarks by Chair - 5 or 0 minutes First Speaker - 30 minutes (or 25) Second Speaker - 30 minutes Third Speaker - 30 minutes Discussant - 10 minutes (or none) Floor Discussion - 10 minutes (or 5 or 15)
Session Chair: Pickle, Linda W. National Center for Health Statistics
Address: National Center for Health Statistics 6525 Belcrest Road Rm 915 Hyattsville MD 20782
Phone: (301)436-7904 ext 148
Fax: (301)436-7955
Email: lwp0@cdc.gov
1. Bayesian Inference For Extreme Rate From Health Effects Maps
Cressie, Noel, Iowa State University
Address: 102F Snedecor Hall Department of Statistics Iowa State University Ames, IA 50011-1210
Phone: 515-294-3441
Fax: 515-294-4040
Email: ncressie@iastate.edu
Stern, Hal S., Iowa State University
Abstract: Health effects (e.g., disease-incidence rates) are often reported for small areas and displayed using a choropleth map, such as one might find in cancer atlases. Areas with extremely high incidence rates may be subject to further scrutiny in an attempt to identify possible risk factors. Hierarchical probability models for disease-incidence rates can be used to obtain the joint posterior distribution for an entire set of small-area rates while taking into account the spatial nature of the data. Optimal estimators of small-area incidence rates will differ depending on whether ensemble estimation is the goal or special emphasis is placed on other functions, such as extreme values (e.g., optimal estimation of the top ten rates and the small areas corresponding to them). A loss function for extremes will be developed and MCMC methodology will be used to carry out the subsequent Bayesian inference. The results will be compared to a simpler approach obtained from constrained Bayesian ensemble estimation.
2. Ranks of Extreme Rates by Summary Rates For Geographic Health Data
Pennello, Gene, National Cancer Institute DCEG/BB
Address: Gene Pennello National Cancer Institute DCEG/BB Executive Plaza North, Room 403 6130 Executive Blvd MSC 7368 Bethesda MD 20892-7368
Phone: (301)496-3344
Fax: (301)402-0081
Email: pennellg@epndce.nci.nih.gov
Abstract: Methods for summarizing disease rates by geographic area for purposes of mapping include the use of crude rates, age-standardized rates (direct and indirect), and empirical Bayes rates. Epidemiologists focus on the highest summary rates because they can suggest leads to disease etiology. This note uses simulation to compare summary rates by their ranking of the areas with the highest underlying risks of disease. Sensitivities and specificities at ranking the designated areas above and below cut points in rank are computed and used to produce receiver operating characteristic (ROC) curves. These ROC curves are compared with the mean square errors of the summary rates decomposed into the variance of the summary rate and the bias introduced by confounding age effects. Data were simulated assuming the Poisson-Gamma mixed model for disease counts and relative risks with parameters based on cancer mortality data for blacks and whites from the National Center for Health Statistics. The results were used in deciding on the method of choice for the maps in the upcoming atlas of cancer mortality from the National Cancer Institute.
3. Effects of Smoothing Mortality Data Using the Weighted Head-Banging Algorithm
Mungiole, Michael, National Center for Health Statistics
Address: National Center for Health Statistics 6525 Belcrest Rd., room 915 Hyattsville, MD 20782
Phone: 301-436-7904, x145
Fax: 301-436-7955
Email: mim4@cdc.gov
Simonson, Katherine Hansen, Sandia National Laboratories
Pickle, Linda W., National Center for Health Statistics
Abstract: Maps of smoothed data permit the reader to identify general spatial trends by removing the ``background noise'' of random error in the original data. We used the weighted head-banging algorithm (Mungiole et al. 1996) on mortality data to consider the effects of varying smoothing parameters in an attempt to obtain a more objective measure for determining the appropriate degree of smoothing. We also modified the smoothing algorithm to consider the effects of increasing the number of nearest neighbors being considered when smoothing areas along the perimeter of the contiguous United States. As was found in previous smoothing research with mortality data sets, weighted head-banging smoothed spikes and edge effects in the perimeter area data. Increasing the number of nearest neighbors in the smoothing process resulted in moderate changes to the areas along the perimeter, with the border more closely resembling areas in close proximity to it. When varying the degree of smoothing, it was found that the appropriate degree of smoothing was somewhat dependent on the amount of noise in the original data. The process used to obtain the smoothed data, including the choice of head-banging parameters, is discussed.
4. Should Spatial Maps Be Smoothed?
Simon, Gary, New York University
Address: Department of Statistics and Operations Research Stern School of Business, New York University 44 West Fourth Street New York NY 10012-1126
Phone: 212-998-0451
Fax: 212-995-4003
Email: gsimon@stern.nyu.edu
Abstract: The values on a spatial map are smoothed for at least two reasons, improved estimation and enhancement of visual effect. The improved estimation logic simply exploits local sameness and then averages to produce estimates of lower variance, much as one might do to a time series periodogram. We will consider the use of smoothing for visual effect. It will be argued that, on substantive issues alone, some maps should not be smoothed. Further, we will propose measures to assess visual smoothness and develop the statistical properties of these measures. It will be claimed that smoothing should not be done on maps for which the measures do not reach statistical significance. For those maps on which the measures do reach significance, it will be possibly to assess quantitatively the improvement in visual smoothness.
List of speakers who are nonmembers: None