Forecasting Survey Panel Turnover Using Discrete Time Survival Analysis

• Activity Number: 649
• Type: Contributed
• Date/Time: Thursday, August 13, 2015 : 8:30 AM to 10:20 AM
• Sponsor: Survey Research Methods Section
• Author(s): Jun Ruan* and Frank Fasinski
• Companies: Nielsen Audio and Nielsen Audio
• Keywords: Turnover forecast ; Discrete time survival analysis ; Time-varying covariate

Abstract

Panelist turnover forecasting is the key to successfully managing the size and the distribution of a survey panel and is crucial for budget control and operational efficiency. Traditional predictive modeling approaches, such as logistic regression and decision tree, enable us to predict whether panelists will turnover within a certain period, but as the period gets shorter, the prediction error will increase and go beyond a reasonable level. This paper proposes a survival analysis approach to forecast the lifetime function of panelists with differing demographics. In particular, a discrete time survival model was developed to allow for flexible time interval selection, and a time-varying seasonal covariate was adopted to control the seasonal differences of turnover behavior. The model was applied to the Nielsen Audio PPM (Portable People Meter) panel data, and was implemented to support production decision-making including sample selection volumes. The estimated parameters indicated that panel turnover is influenced by multiple demographic factors and displayed a strong seasonal pattern. The out-of-sample forecast accuracy measured by Mean Absolute Percentage Error was 6%.