Graduate Student Profiles

Craig McCabe's Thesis Animation Experiments

In working on the Niger measles project with CIDD, one of my early impressions was that the raw incidence data was quite chaotic and difficult to interpret when viewed in an animated choropleth map. The raw data had a "spiky" appearance, with large fluctuations in incidence from week to week:

This caused the individual districts in Niger to flash from high to low values in a manner similar to lights on a Christmas tree. This made me think about what effects temporal data manipulations might have on the perception of these patterns, specifically, the use of aggregation and moving-window averages to reduce the "visual chaos". Would these smoothing approaches enhance the user's ability to perceive patterns and trends in the data? An additional factor was the climate, size, and population density of the districts of Niger:

The largest districts, in the north, were mostly desert and sparsely populated, while the small districts on the southern border with Nigeria were densely populated. To account for the population distribution, I implemented an alternate map representation which converted each district to a proportional circle, then scaled its area based on the population value (calculated using Landscan 2007 grids). To communicate adjacency relationships between districts, I included connecting lines between districts that shared a border. The two map representations are shown below (Choropleth on top, Schematic on bottom)

A few examples of the experiment animations and map-reading tasks are provided below. The experiment was administered to 96 participants, most of whom were undergraduate students who were taking at least one class in geography. The participants were split into 3 groups, based on the three types of temporal manipulations: 1) Raw weekly data, 2) 2-week aggregated data, 3) 5-week moving window average (an average of the current week plus the 2 weeks before and after). Each group then viewed the same animations and completed the same four map-reading tasks.

Task 1 - The Animated Square

The first map-reading task employed a simple animated square, representing data for a single year epidemic for a single district in Niger. Participants were asked to view the animation (which repeated 3 times), then on the next screen select a time-series graph that matched the data that was animated (among 4 choices, only one of them is correct). Examples of each smoothing type for the same underlying data are provided below:

Raw Animation | Graphs
Aggregate Animation | Graphs
Moving-Window Average Animation | Graphs

(Click here for the correct response)

Task 2 - Pick the Peak Infection Week

The second map-reading task asked participants to determine during which week the peak rate of infections occurred (for a region of interest ranging from 6 contiguous districts to the whole country). Participants viewed data in both choropleth and schematic map forms, with animation durations of 1 year. Again, participants were allowed to view each animation 3 times before providing an answer (between 1-52 for the raw and averaged group, and 1-26 for the 2-week aggregate group). Some map examples for just the Raw data group are provided below:

Raw Choropleth (6 Districts) | Raw Schematic (6 Districts)
Raw Choropleth (All Districts) | Raw Schematic (All Districts)

(Click here for the correct responses)

Task 3 - Compare Epidemic Severity Between 2 Regions

The third map-reading task asked participants to determine which of two circled, contiguous regions (each consisting of 6 districts) exhibited the higher cumulative infection rate for the time period. Rates were reported according to the rate of infection per population for each individual district. Map examples for just the Aggregate data group are provided below:

Aggregate Choropleth
Aggregate Schematic

(Click here for the correct response)

Task 4 - Characterize Epidemic According to 3 Semantic Differential Scales

The fourth and final map-reading task asked participants to characterize the spatio-temporal behavior according to three semantic differential scales. These scales employ bipolar adjectives (e.g. Warm/Cold) and asks the participants to assess where on the continuum to rate the animation. The following scales, relevant to infectious disease behavior and patterns, were used (participants selected a number between 1 and 10 for each scale):

A. (1) Chaotic … Ordered (10)
B. (1) Stationary … Moving (10)
C. (1) Clustered … Dispersed (10)

Similar to Tasks 2 and Task 3, participants viewed animations using both choropleth and schematic map representations. Because there are no "correct" answers for this task, the responses were analyzed for statistical consistency among and between groups. Six examples are provided below - they all use the same underlying year of epidemic data, but differ in the smoothing approach and map type:

Choropleth Maps: Raw | Aggregate | Averaged
Schematic Maps: Raw | Aggregate | Averaged

Back to Craig McCabe's profile

Return to Grad Highlights