Using Mixture Distributions to Predict Radio Listening

• Activity Number: 365
• Type: Contributed
• Date/Time: Tuesday, August 6, 2013 : 10:30 AM to 12:20 PM
• Sponsor: Survey Research Methods Section
• Author(s): William Waldron*+
• Companies: Arbitron
• Keywords: Mixture Distribution ; Survey

Abstract

Through a combination of paper surveys and radio panels, Arbitron records and publishes radio ratings every month or quarter for about 300 media markets in the United States. One current area of applied research is to investigate the feasibility of modeling radio listening based on a series of demographic variables for a particular marketplace. The nature of radio listening data violates several assumptions inherent in the linear regression model. Due to varying amounts of heavy listeners, radio data exhibit a certain amount of skewness, which infringes the normality assumption. The constant variance assumption is also problematic. Furthermore, the listening data is nonnegative and can be zero-inflated for certain radio formats (i.e., soft rock, country, etc.). We will proceed to build a mixture distribution regression model. We will consider the available mixture distributions, link functions, mixture constants and also decide how many components to use in the model. Finally, we will compare the resulting predictions to those generated by the linear regression model as well as the log-transformation model.