Commentary

Clearing The Confusion Surrounding Media Data Fusion

In recent years, we have witnessed a proliferation of data integration and fusion techniques. In large part, the increased popularity of these techniques is driven by a need for more accurate reach and frequency estimates for advertising schedules that rely on many different types of media (radio, television, Internet, magazines, and newspapers). This article doesn't focus on any one particular approach to fusion or data integration. Rather, it focuses on the rationale for using such approaches, the assumptions underlying these approaches, and the best way they can be evaluated in terms of their utility and accuracy.

By fusion, I mean the integration of data from two separate studies. For example, a fused database could result from combining television data from Nielsen with radio data from Arbitron. The data integration could be done at the respondent level or at the aggregate level, using a modeled approach. Generally, when data are combined at the respondent level, a person in one study (the recipient) is assigned the data of a person in the other study (the donor). The assignment of data from one person to another is usually done on the basis of how well the two people match on key demographic factors that are important to advertisers (age, sex, income, race, etc.).

So, if Arbitron data were fused to the Nielsen database, one would first try to match each Arbitron respondent with a given Nielsen respondent on the basis of demographic similarity. Then the radio data of the respondent in the Arbitron study would be appended to the television data of the respondent in the Nielsen study. The primary rationale for fusion is the perceived impossibility of collecting primary ratings data (i.e., studies serving as media currency) for more than one medium at a time. And since neither Nielsen nor Arbitron collects much more than basic demographic data in their studies, planning on the basis of consumer behavior is impossible. Fusion is promoted as a magical elixir, especially when one of the databases used in the fusion contains thousands of consumer measures, as is the case with the single-source database of Mediamark Research Inc. (MRI).

There are three primary advantages to fusing databases:

>> Multimedia Planning Using Demographic Targets. Users can more accurately calculate reach and frequency estimates for demographic targets for a schedule that includes different types of media. Accuracy is improved because you don't have to assume random duplication. In other words, you don't have to assume that people's behavior with respect to one medium is unrelated to their behavior with respect to another medium.

>> Multimedia Planning on the Basis of Consumer Behavior. Users are able to develop a multimedia plan based on the estimated cost-efficiency of reaching people who engage in a specific consumer behavior (e.g., buy a given type of product or brand), rather than the estimated cost-efficiency of reaching people who have certain demographic characteristics in common (e.g., women ages 25 to 49). The former approach is much preferred because demography is often not as predictive of media behavior as we would like.

>> Extensive Profiling of Individual Media. Users are able to provide a more detailed profile of a particular medium based on numerous measures of consumer behavior. For example, people who watch "Monday Night Football" are not only disproportionately young and male, they are also disproportionately more likely to consume beer, entertain friends and relatives at home, barbecue, and go to bars. Still, do fused databases provide good or bad information? If we know little or nothing about the relationships between fused variables, we can be blissful in our ignorance. Consider two ratings services, one for television and one for magazines, that capture only some basic demographics along with the ratings. We can produce a fused database and then show the duplication between Magazine "A" and TV Program "B."

Three possible fusion results are shown in Table 1. We are able to preserve the ratings of the TV program and of the magazine in each of the three fusions. So, for all three fusions, the number of people reading Magazine "A" is 300 and the number of people watching TV Program "B" is 500. However, the duplication between the two is dramatically different for the three fusions. For Fusion #1, there's no duplication between the two media vehicles and the net reach is 800: None of the 500 people who watch Program "B" reads Magazine "A," and none of the 300 people who read Magazine "A" watches Program "B." Fusion #2 shows 100 percent duplication between the two vehicles, so the net reach is 500. This is because everyone who reads Magazine "A" also watches Program "B."

So the inclusion of Magazine "A" in the media schedule did not increase the total net reach beyond that of Program "B," which is 500. In Fusion #3, the net reach is 650, midway between the two extremes (800 and 500) seen in the first two fusions. This is because half (150) of the 300 readers of Magazine "A" did not watch Program "B." The addition of Magazine "A" to the media schedule adds 150 people, bringing the net reach of the two vehicles to a grand total of 650.

The data in Table 1 underscores the question raised by most fusions: If we have no information about the correct duplication, how do we know whether the duplication estimate from the fused database is accurate? Some experts say that preserving the individual ratings is sufficient, but we have accomplished that in all three fusions. The absence of any data showing direct relationships between two media precludes us from knowing whether the duplication estimates are accurate. To illustrate, we can fuse "watching a golf program," say, with "reading a golf magazine." If we used MRI data, we would find that household income, individual employment, income, and education level are relatively strong predictors of viewing golf programs and of reading golf magazines, and that they should assume importance in the fusion algorithm. If we had two studies, one with information about reading golf magazines, the other containing information about watching golf programs, and we matched respondents on these variables, the estimated duplication level should be more accurate than any random matching of respondents.

How much of an improvement in estimating the duplication did we get based on matching respondents on key demographic variables? Using MRI data, we find that readers of golf magazines are 4 to 5 times more likely to watch golf shows than the general adult population. This is a strong relationship, and even if the ratings levels for TV were off from Nielsen's, it would perilous to disregard the duplication levels as a measure of the true relationship between media behaviors.

This is especially important given that demography is not always a strong predictor of consumer behavior. The best predictor of watching golf programs in MRI's study barely obtains a 200 index. Any fusion algorithm that relies extensively on demographics will often fail to replicate a single-source measure.

Julian Baim is executive vice president, chief research officer, Mediamark Research Inc. (julian.baim@mediamark.com)

Net Reach of Two Media Vehicles As A Function of Estimated Duplication

Total Magazine A Audience:300

Total Program B Audience: 500

 

Number
Reading
Magazine A

Number
Not Reading
Magazine A

Net Reach

Percent Of
Magazine A
Readers Viewing
Program B

 

Fusion #1

View Program B

 

0

500

800

0.0%

Do Not View Program B

 

300

 

 

 

Fusion #2

View Program B

 

300

200

500

100.0%

Do Not View Program B

 

0

 

 

 

Fusion #3

View Program B

 

150

350

650

50.0%

Do Not View Program B

 

150

 

 

 

Next story loading loading..