|
|
|
THE TOP 10 INSIGHTS From the IIR Conference on Marketing Mix Modeling, Oct 25-26,1999 1. Marketing Mix Models may be part of the problem, not the solution. This conference was an eye-opener. There were lots of indications models are being used to manage price competition in package goods all the way down to the store level. There were few indications of models being used to optimize advertising, or that they are being used in non-package goods fields. Package goods are not the most profitable sector of our current economy. Could all this be related?2. There is great emphasis on store level data and price elasticity. A number of speakers mentioned their greatest success was being achieved collecting marketing data and building models to predict sales store by store. The topic given the most attention by speakers and particularly by those asking questions was price elasticity: how much sales change when you change the price. 3. This can understate the importance of advertising. Paul Tarr of MSW pointed out when store level data is being studied to see what explains differences in sales between stores in the same market, TV advertising is a constant. Customers at all the stores have the same opportunity to see it. So whatever method of measurement and model fitting is used, advertising won’t explain the differences between the stores. When market level, regional level or national level models are constructed by aggregating store level results, which appears to be a common approach, the real importance of advertising continues to be underestimated. MSW was the only other ad-testing firm represented at the conference, putting us in the unusual position of working together to emphasize the importance of including valid measures of advertising’s effect. Paul Tarr did it with a study of 15 cases showing ad test scores were related to changes in market share. Further, the ad’s performance in these tests accounted for three times as much of the change in market share as GRPs. That backed up a point I was making. The quality of a commercial was much more important than how often it was aired. He also described his firm’s models using more marketing mix variables in a non-stationary Markov Chain to predict switching in Europe. That also showed ad quality was more important than ad weight. 4. Kraft is working on the problem. Jim Savage reported on a year of experimenting at Kraft with new approaches that focus on the consumer. One of the things they were looking for were more accurate estimates of advertising’s contribution. They tried using information about consumer diversity across stores in addition to traditional market mix elements. That provided insights into what drives volume with different consumer segments. (See 9 for more on that approach.) They also tried using household panel data in logit choice models, producing more insights. Results were mixed. Both store level and household panel models underestimated advertising’s contribution. However, when both were combined with market level data they produced good advertising estimates. They are going to be testing more store level demographics to see if it will help target promotion and advertising. They will also be doing more to understand what drives base volume, paralleling the work described below by Len Lodish. 5. Brand Health can help show what is important. Wharton’s Len Lodish, of "How TV Advertising Works" fame, reported success using a health related concept of "Brand Health" that has two new dimensions. First, it’s a brand that can resist attacks by competitors. He uses that ability to predict who is going to lose share when a new product is introduced. He then described a massive "Mind of the Consumer" model that fits all the product attribute and marketing variables for each product in a category in every store during every week to each product’s share of the market. He showed a number of examples where that had been used to filter out all the factors having short term effects, making it possible to study his second dimension, the brand’s "pure" underlying attractiveness to the consumer, and how it is changing over the long term. 6. Conflicting test results were reported. Single-market regression models were better than pooled market modeling data in a test reported by Art Christiani. They recovered known elasticities exactly. It received some vocal criticism in the Q&A session. It was said the functional form of the test favored single-market models, and his use of artificial data was reported by Ross Link of Marketing Analytics. He used simulators that start with the output of marketing mix models (elasticities) and combine them with marketing plans and cost data to forecast sales, profit and volume. They can also use optimization techniques to suggest changes. He found the best results modeling store group data and then pooling coefficients at a high level like a sales district, using a simulator loaded with data from the level you needed to forecast: store, store group, chain, or market.7. It is important to factor in the quality of the advertising: Yours truly gave a talk showing marketing models fit historic data better when they allow for the quality of advertising, and to show the dangers of models that only include expenditures on advertising, or GRPs. They assume all advertising has the same effect. It doesn’t. The main piece of new evidence I offered to show it doesn’t was from our eight years of syndicated testing of all Super Bowl commercials. It showed test results for 65 commercials that had identical real-world exposure – they appeared once on the Super Bowl and nowhere else. The top 20% reached and affected seven times as many people as the bottom 20%. Models that don’t allow for differences like that can’t hope to fit. We didn’t miss the opportunity to point out the best way to get the needed data is with continuing ad tracking studies - the type BRC specializes in. 8. 8. Watch out for that unmeasured third factor: Joe McHugh of Pepperidge Farm illustrated another type of error with a case history. The brand was on a roll; sales were up 20%. Pretesting indicated, and their marketing mix model confirmed, their new campaign was effective. So advertising was credited with the unusual gain until modelers noticed two abnormalities. They stopped advertising the end of the year, yet sales stayed up. And although sales were up over last year, they were not up much over two years ago, because sales had dropped last year. When everybody started "thinking like modelers" they found a third factor not reflected in their model. Distribution problems had reduced the amount available for sale last year, so this year’s gains reflected a return to normal distribution, not exceptional advertising. In summarizing the process he put understandable emphasis on two points. Validate your model against real data. When support levels change, rerun your model. 9. Marketing Mix Modeling can be done by Type of Consumer: Paul Flugel described this as Consumer Mix Modeling, an enhancement showing how effective each element in the marketing mix is on each type of respondent. The object is to waste less on efforts that reach those who will never buy, and devote more to those who are prime prospects. He does it by breaking out all relevant populations into his firm’s lifestyle/lifestage groups (Spectra Grids). That includes product customers, store customers, newspaper subscribers, TV programs, etc. 10. Another prominent academic stressed advertising: Columbia’s Sunil Gupta got off the conference’s best one liner describing his young MBAs as often wrong – but never in doubt. He also gave one of the most meticulous and technologically complete presentations, yet he objected to complexity, saying simple models work perfectly well. He has found arguments over logit vs. regression vs. other forms are not important. Studies he and his colleagues have been publishing in recent years have shown in the long run advertising reduces price sensitivity and promotion increases it. Sensitivity to both price and promotion are increasing. The nonloyal segment, where the effects are greatest, is growing. More price promotion and less advertising has led to decreased differentiation between brands. On the average, long-term effects offset short-term effects by about 40%. With short-term elasticities, depth>frequency>regular price. The effect is reversed over time. Most of the short term effect of discounts is on choice rather than quantity. The opposite may be true in the long term. Almost all of the effect of advertising is on choice. In summary, in the long run, advertising appears to help brands while promotions hurt – and the effects are "measurable". Those of us who feel models may have led packaged goods out of advertising and into poverty came away somewhat reassured. At least the leading thinkers and doers seemed to have it right. Don Bruzzone, December 1999 (This was an IIR conference - No proceedings - You will have to contact speakers directly for more.) |