6.1 PREDICTING GAMBLING REFERENDA WITH NEURAL NETWORKS
On average, the ANN model predicted the voting outcome with 82 percent accuracy (correctly predicting four out five counties) on the test dataset (data that the ANN has not seen during model building process).
Using sensitivity analysis on the trained neural network model, the researchers identified
the most important variable in predicting gaming ballot outcomes,
the most dominant variable were the county’s religious inclination (i.e., percent church membership),
the county’s ethnic diversity (i.e., percent minority),
and whether the county was classified as a Metropolitan Statistical Area (MSA) by the U.S. Census.
Contrary to conventional wisdom, a county’s financial characteristic and age distribution were not found to be significant factors in determining ballot outcomes.