Welcome to my Thinking World…

Abstract:

Consumer-file marketing companies collect and sell customer’s demographic and behavioral information wherever they can: through loyalty cards, website visits, social media posts, etc. These ancillary data are tremendously valuable and have a variety of applications in the field of television and digital media measurement that include targeting households or individuals for marketing campaigns, adjusting for nonresponse bias, and projecting estimates from a sample to an underlying universe. These potential advantages, however, depend on the completeness of the information contained in the data. Today, most applications involving the use of ancillary data exclude incomplete records from the substantive analysis. Exclusion as a missing data handling strategy has been known to result in inaccurate conclusions depending on the nature of the mechanism assumed to have generated the missing data. In this paper, we demonstrate a model-based alternative to exclusion, called multiple imputation (MI), which is used to fill in the missing values. A correctly specified imputation model preserves important characteristics of the data (e.g., variances, correlations, interactions) and can mitigate any bias resulting from exclusion-based missing data handling strategies. Our paper demonstrates MI in the context of a substantive analysis of the relationship between a completely observed measure of the number of advertisement impressions served across television networks and an incomplete measure of the number of children per household. We include likely correlates of the suspected cause of the missing data in the imputation model to evaluate whether their inclusion improves the quality of the imputed values. These correlates are measures pulled from a second, more completely observed, ancillary dataset containing television and digital media audience measurements. This paper concludes with a discussion of our findings, which are likely to have important implications for the use of ancillary data in television and digital media audience measurement.

Recommended Citation:

Craft, M., Raghunath, A., & Rao, K. (2023). Does Ancillary Data about Ancillary Data Help Treat Missing Responses in Ancillary Data? American Association for Public Opinion Research, Philadelphia, PA.

Attached Documents/Links:

• AAPOR 2023 Program
• For a copy of this presentation, please send me a comment with your email address in the box below or an email to kumarrons at gmail dot com.

Abstract:

The number of US households with access to computers at home has continued to grow. According to the 2011 Computer and Internet Use report published by US department of Commerce, 77% of US homes have computers in their home, compared to 62% in 2003. Many households, however, do not have multiple computers dedicated to each member living in the house. As such, sharing of computers amongst household members can be a prevalent phenomenon in home computer usage. Understanding this within-house computer sharing phenomenon and identifying the mostly likely person behind the computer screen can be of interest to market researchers and practitioners, particularly those interested in studying effective ways to target online ads based on users, online activities. For survey researchers who are attempting to recruit hard-to-reach individuals like teens and young adults, understanding of computer sharing could help establish contact at times when those individuals are more likely to be behind the computer. Despite its prevalence, within-house computer sharing has barely received any research attention. This study hopes to break through the barriers preventing the light of scientific inquiry into this phenomenon. Read more

Abstract:

With declining response rates, surveys increasingly rely on weighting adjustments to correct for potential nonresponse bias. The resulting increased need to improve survey weights faces two key challenges. First, additional auxiliary data are needed to augment the models used to estimate the weights. Depending on the properties of these auxiliary data, nonresponse bias can be reduced, left the same, or even increased. Thus, the second challenge is to be able to evaluate the alternative weights, when the assumption of “different estimates means less bias” may not hold. Ideally, data need to be collected from as many nonrespondents as possible to provide direct estimates of nonresponse bias. Read more

Abstract:

This study builds on a previously published panel recruitment experiment (Rao, Kaminska, and McCutcheon 2010), extending that analysis to an examination of the effectiveness of pre-recruitment factors such as mode and response inducements on three post-recruitment panel participation effects: attrition rates, survey completion rates, and panel data quality. The panel recruitment experiment, conducted with the Gallup Panel, netted 1,282 households with 2,042 panel members. For these recruited members, we collected data on panel participation and retention, and use it for analysis in this study. Read more

Abstract:

In this study, for multiple websites, we estimate duplicated audience reach between home and work Internet access locations. By employing a probability-based matching technique, we use metered-panel based data from non-overlapping home and work panels for creating a virtual overlapping home-work panel. Read more