A Summary of “Viewer Preference Segmentation and Viewing Choice Models for Network Television”
In 1990 broadcast television network advertising revenue was approximately $25.5 billion a year, with General Motors alone spending $598.4 million on network ad time (Rust et al. 1992). As such, both television networks and advertisers have a common interest in understanding why people choose to watch television, and why they prefer one program over another. In the pursuit of this understanding, it is useful to group television consumers into segments if possible in order to make decisions based on the shared preferences of large amounts of people. Segments can assist networks and advertisers with the forecasting necessary in deciding who will watch what program. In this way, advertisers can spend their ad dollars more effectively, and networks can maximize profits by designing shows that the most valuable demographics will consistently want to watch.
The purpose of the paper is to design models that can address the following questions accurately:
1) Do identifiable segments exist?
2) How do viewers decide when to start and stop watching television?
3) How do consumers decide which television programs to watch (Rust et al. 1992)?
Previous research on this topic primarily falls into three general types. The first type is “structuring viewing alternatives”, which attempts to group television programs according to their similarities differences. These groups of programs can be defined a priori, without supporting data. They can also be grouped according to scientifically gathered viewer data. Both of these two methods must make the assumption that these homogenous program categories actually exist in the first place. The third method of grouping does not rely on this assumption, but uses multidimensional scaling in order to show the relative differences among different shows (Rust et al. 1992).
The second type of previous research is “segmenting television viewers”, which is similar to the previous type in that viewer segments are often formed a priori from demographic data. Empirically derived segmentation on the other hand assumes that watching a television program gives a certain benefit to the consumer, and that similar programs will provide similar benefits (Rust et al. 1992).
The third type of previous research is “viewing choice models”, which attempts to predict which shows consumers will watch. These models are the most difficult to construct and as a result are much less common than the other two types. One phenomenon that complicates matters and needs to be considered here is “audience flow”, which is the way previous conditions (such as TV on or off, or what the consumer was watching previously) affect consumers’ future choices (Rust et al. 1992).
Rust et al. (1992) contend that their model is superior to previous research because: it combines the three types of existing models into one comprehensive model, it estimates the preference functions in addition to simply identifying them, it is open to the possibility that people’s preferences can be different for reasons other than “simple socio-demographic differences”, and finally that their model includes the decision to watch television in addition to simply deciding which program to watch.
The Rust et. al (1992) model shows the relationship between programs using Nielsen viewer data and multi-dimensional scaling. They are arranged by preference of similar audiences rather than seeming to be similar programs. Sample data was retrieved from 11,501 viewers and tested against the model. The results showed various ways in which the actual data differed from assumptions often made a priori.
Once the characteristics of programs are defined, they can then be matched with the preferences found in viewers. The closer the program is to the preference shown via multi-dimensional scaling, the more likely the viewer will watch. The results showed that this was in fact the case, was statistically significant, and the model produced similar results when applied to another randomly selected section of the data (Rust et al. 1992).
Lastly, the model was used to predict when people will turn their televisions on or off in the first place. It was found that different portions of the populace will watch television at different times. For example, the segment dubbed “Western” was less likely to watch on the weekend.
This model shows advantages over previous research, and provides a valuable tool for analyzing an important body of information.
Rust, Roland T., Wagner A. Kamakura, and Mark I. Alpert (1992), “Viewer Preference Segmentation and Viewing Choice Models for Network Television,” Journal of Advertising, 21 (March), 1-18.
1. In the Rust article, what assumptions did the Rust model make when segmenting television programs? When segmenting television viewers?
2. In what ways would the Rust model apply or not apply if it was used to segment data on webpages in place of television shows?