Conditioned Choropleth Maps: Dynamic Multivariate Representations of Statistical Data. Research conducted by Dan Carr, Alan MacEachren, Duncan MacPherson, Erik Steiner and Mark Harrower.

Statisticians have long been interested in modeling relationships that exist among the factors that explain variance in a phenomenon of interest. For example, a simple regression model might specify that income is (in part) a function of average age, gender and the number of years of education. Statistical tools make it possible to explore such hypotheses numerically. However, if one wants to "see" these complex relationships visually, say on a map, there have been relatively few cartographic tools available. Researchers need to be able to more than "see" relationships - they need to explore the geospatial aspects of the relationships (which are not addressed using standard statistical methods). Conditioned choropleth (CC) maps were designed to fill this geospatial void.

CC maps address the problem of visually representing the geographic aspects of multivariate relationships. They do so by extending the concept of small multiples from depiction of multiple related variables separately (on individual maps) to the representation of subcomponents of multivariate relationships (with a portion of the relationship among three variables shown on each map).The implementations illustrated here add dynamic interaction to the basic CC map template.

Dynamic maps can incorporate 'dynamic information filtering' whereby users can filter/hide/reveal detail as needed and focus their attention on portions of the entire dataset, thereby keeping the maps legible. The dynamic CC maps presented here use a novel implement of information filtering: the mapped variable (e.g. income) is selectively revealed in the nine panels of the display according to two other 'conditioning variables'. These other two variables are not mapped directly, rather they are used to selectively reveal or hide parts of the original map. In other words, they dynamically 'filter' the original data allowing users to focus-in on specific numerical co-occurances. For example, the lower right panel in example #1 reveals portions of a lung cancer mortality map only for regions of the USA that have both low annual rainfall and low rates of poverty. What remains on the map are lung cancer mortality rates in those regions (the colored regions).

The Exploratory Power of Dynamic CC Maps?
Returning to example #1, the conditioned maps have shown that areas with high rainfall and high poverty ALSO have higher-than-average lung cancer rates. There is a correlation between these three variables. What is really interesting is that this statistical correlation also has a geographic correlation: these places form a regional "blob" in only one portion of the country (the 'red' areas in the south-central US). Using tools like conditioned choropleth maps, analysts can visually explore three datasets simultaneously looking for interesting spatial patterns and co-occurances.

Although conditioned choropleth maps can be static, they become much more powerful if the user can interact with them directly. Example #2 is one prototype for such a system. Users are free to choose the variable to map, which variables to apply as conditioning agents, and where the class breaks should be. Data are retrieved from an Oracle database to match these user choices. Once data are displayed, panning and zooming also allow users to customize the presentation of the maps. The technique of 'probing' allows users to click on a region in the map and retrieve any information associated with that place (e.g. rates for each of the three current variables). Turning on the highlighting feature enhances interactivity by reminding users of where a certain place (e.g. Pennsylvania) falls in the data distribution for the entire country (high/medium/low rates).

The dynamic CC Map described above was developed in Flash 5. We are using this version to support usability testing and design modifications. The current implementation, however, exhibits some severe performance problems when used outside our local network. An alternative CC Map prototype (but as yet less flexible verision) has been implemented in Java as an applet. This implementation supports dynamic interaction with a map of the U.S. by county.

For background on development and application of the CC Map concept, see:

Carr, D., Wallin, J.F. and Carr, A., 2000. Two new templates for epidemilogy applications: linked micromap plots and conditioned choroplth maps. Statistics in Medicine, 19: 2521-2538.

For discussion of Flash-based dynamic maps linked to an Oracle database, see:

Steiner, E., MacEachren, A.M. and Guo, D., 2001. Developing and assessing light-weight data-driven exploratory geovisualization tools for the web, Workshop on Geovisualization for the Web. ICA Commission on Visualization & Virtual Environments, Taupo, New Zealand, April 30 - May 3, 2001.