Analytic Forensics Logo 







MMM (Marketing Mix Modeling) depends on many complex (non-linear) effects.  (To access a discussion of how Marketing Mix Modeling works, click here MMM.)   While the process of Marketing Mix Models is based on looking for simple trends within a very complex system, if we oversimply our models will not be accurate enough and will miss important dynamics.

The major non-linear marketing effect is ad stock, which is how the effects of advertising is distributed over time.   (To access a discussion of Ad Stock, click here Ad Stock).

Similar to ad stock is the element of baseline decay.  Baseline decay is simply how well would sales for a product do if the company stopped all advertising and marketing.   If a product is going to be in a grocery shelf anyway, and if it is the least expensive item on that shelf, likely it will continue to do well with or without advertising, with very little decay.  Same for an iconic brand or one with a very loyal customer base.  But new products, or more expensive products likely would decay quickly without advertising/marketing support.  The graph below shows 4 examples of actual decay curves in Marketing Mix Models I’ve worked on.

The results of Marketing Mix Models are usually some ROI measures. For example, TV might have an average ROI of “2.3” which means you get on average $2.30 dollars of sales for every dollar you spend on TV.  However, presenting ROI that way gives people the impression that the effects of adding a dollar more of any media or marketing is a straight line.  It isn’t, it is a curve (often a complex curve, see the next two graphs below).

For one, almost all advertising and marketing have a “Threshold Level”.  In other words, you need to spend at least a given amount on that advertising before it starts working well.   This might be because very cheap advertising buys are rarely cost effective.  Yes, you can spend very little and put an ad on a little watched network at 3AM, but it costs so little because the audience is very small and not very engaged.  It might also be a function of “reach and frequency”, it is generally believed that if a person sees an ad 3 times they remember it far better than seeing it just once (much more than 3 times better).  But Marketing Mix Modeling doesn’t really answer “why” threshold levels exist, we just have statistical evidence that they do in almost every case.   The below graph shows the threshold levels I was able to detect for one marketing factor in one Marketing Mix Model I created.

While threshold levels likely exist for every media and product, they are rarely put in Marketing Mix Models. For one thing, most large advertisers rarely spend less than the threshold level.  Without historical examples, we can’t calculate threshold levels.  And if future advertising never spends below the threshold levels there is no need to model the threshold levels.

But at the other end of the graph above, we also have “diminishing returns”.    Once you spend enough to get past the threshold level every incremental dollar you spend gets less and less effective.   Eventually (in theory) if you spend enough money you will reach a point that spending more money gets you no extra sales.   Rarely do advertisers reach the point of no effect, but often they reach the point in spending where every dollar spent in a specific media gets them less than a dollar back in sales (a waste of money at that point). 

Unlike threshold levels most companies have lots of history of spending spikes that help us calculate “diminishing returns”.  And so most marketing mix models need to include some way to calculate these diminishing returns.

Another very important factor in marketing mix models is “synergy”.  By this I mean TV might do well, Print might do well, but a mix of TV and Print does better per dollar than doing TV or Print alone (see the graph below).   A truth is that is there is at least some synergy between every pair of marketing or advertising factor.  In other words, having lots of different marketing and media options always does better than focusing on only one or two.

There are lots of examples out there of companies who decided that “spot TV” or “Digital banners” were their most cost effective media, so to save money they decided to put their entire marketing budget in that one most effective media.  In every case I’ve seen, this has terrible results.   Not only do these experiments get above the point of diminishing returns on that “most effective” media, but the client loses all the synergy effects of the other media.  The result is less sales rather than more. 

But synergy could be a minor factor or a major one.  Coupons usually do wonderfully when there is synergy with an awareness media (like TV) but do poorly without that support.  On the other hand, print and TV synergies are usually minor.  And if you have a model with 10 media factors, you would need to calculate 45 synergy effects to catch them all.   So often in marketing mix models you don’t calculate all the synergies, but only include the synergies that you can calculate, and that you need to calculate due to possible changes in future spending.

There are lots of non-linear effects that I haven’t mentioned yet (and are rarely used in models). For example, reach and frequency patterns might be important.  Studies have shown that people are more effected by the 3 viewing of an ad then by the first, but the effects of additional frequency diminishes by the 20th viewing etc.  On the other hand, having a break in frequency (a period of a month without seeing the same ad for example), often “revitalizes” the effectiveness of a repeat viewing of the ad. Rarely is these time series effects added to a marketing mix models. But if the client wants to use the model to test flighting strategies, then these time series effects can be calculated and included.

The art of marketing mix models is to focus on what things the client is likely to change about their advertising and what history they have that will allow you to calculate non-linear effects.  Then to create a model with enough of these complications so that the model dose what the client needs but doesn’t get overloaded with complications that have limited use and effects.   It is balance.

Home Page


Analytic Forensics Logo




Text Box:                                                                                                                                                To Contact Us:







                                                                                                                      Phone:                                                     212-529-5337

                                                                                                                      Voice Mail:                                               917-838-7966


                                                                                                                      Address:                              23 East Tenth Street #304

                                                                                                                                                                New York City, NY     10003