Thursday, March 16, 2006

All Hail the Semantic Differential

To my knowledge, the most extensive writing about the semantic differential came in 1957 with The Measurement of Meaning by Charles E. Osgood, George J. Suci, and Percy H. Tannenbaum. (Learn more about the semantic differential here.)

Factor analysis is one tool that can be used to help interpret results of the semantic differential. In our most recent study, we asked participants 31 questions about 32 brands. Factor analysis helps break down the 31 questions into a smaller number of dimensions. For example, the good-bad question appears to tap the same underlying attitude as pleasant-unpleasant. In our case, the factor analysis suggested that these 31 questions were measuring five basic dimensions.

In their seminal work, Osgood, Suci, and Tannenbaum found three persistent dimensions: evaluation (we call it valence), activity (we call it arousal), and potency (we call it dominance). In our brand study, the first two dimensions held up, but there is no reliable potency dimension.

When used in this way, factor analysis helps us understand how the meaning of things is represented in the mind. The focus is on the questions (i.e., the semantic differentials) and not so much on the things about which the questions are asked.

However, one might want to know something about that, too. Since we study strategic communication in this lab, we do care about brand attitudes. In addition to learning about the measurement of meaning, we want to know how people think about brands.

The second phase of data analysis was to see what mental representations these participants had for Apple, Dell, UPS, FedEx, Ford, BMW, Corona, Miller Lite, and the other brands we used. In this case, we want to do the opposite of factor analysis. We want to see how the brands group together in cognitive space.

To do this, we can use cluster analysis. This involved taking the average score for each of the 31 semantic differentials for each brand. So we stripped away everything about the brand. Coke was no longer a familiar red logo. It was no longer from Atlanta. Instead, it was 31 numbers that ranged from 1 to 7 -- the average score for good-bad, arousing-calm, and so on.

The question is whether those 31 numbers could partition the 32 brands into a meaningful space. Since we are supposed to be measuring meaning, is the meaning there? Cluster analysis shows which things go together mathematically. Did the math find the meaning?

The answer is a resounding yes. When the numbers were submitted to a cluster analysis, the results show that the semantic differentials did a pretty good job representing the brands in a meaningful way.

The cluster analysis revealed that FedEx was most like UPS. American Express was most like Citibank. Coke was most like Pepsi. Samsung, Sony, and Microsoft clustered together. That sounds obvious until you think about how this really occurred. The cluster analysis algorithm had no access to any information about these brands beyond whether they were good-bad, old-young, static-dynamic, passive-active, and so on. That's it.

And from that, the algorithm uncovered the credit card product category, the soft drink category, and the technology product category.

The results were not perfect. The clustering algorithm stubbornly put Nike with BMW instead of Adidas. Here, it appears the image of the brands was driving the association.

Likewise, Dell was right between UPS and the other technology brands. One can only speculate whether this is a spurious finding, or whether these participants mentally represent Dell differently because the brand is shipped rather than purchased at a store like Samsung and Sony.

For good or bad, product category leaders such as McDonald's and Budweiser did not cluster. Instead, they stood out. I can alternatively convince myself this is a good or bad thing.

Apart from the market leaders, the algorithm did not know what to do with KFC, Smirnoff, and Ford. This suggests that these brands lack a clear identity in the minds of college students, at least.

We tend to forget about the roots of the semantic differential. It is often used to measure attitudes and even persuasion. The instrument was intended to measure meaning. And these results show that it does a pretty good job of that. From these 31 adjective pairs, a mathematical algorithm could discover that soda pop is a unique kind of thing. To me, that's just way cool.

Findings such as these are why I go to work with a smile every day.


Anonymous n said...

saw this site on google when i was browsing for materials on semantic differential. cute.

7:11 PM  

Post a Comment

<< Home