Under the Hood Part 1: The fundamentals of Automated MMM
Christopher Engman, CMO/CRO at Proof, explains the fundamentals of how regression analysis works in marketing. Christopher will anser the following; Why is the ROI bigger when using MMM than in a short sighted simple attribution model? What does delayed effects look like? Why is Marketing Mix Modeling the golden standard for quantifying marketing effects and optimizing marketing spend and related staffing? And finally how does the mathematics conceptually work behind MMM?
In a lot of meetings that myself and my colleagues are having with customers and potential customers, we get a lot of questions on what’s under the hood. How does it really work? Why does marketing mix modeling deliver the results it does? Why is it considered the golden standard for marketing and PR and sales analytics? I’ll describe the conceptual mathematics behind multiple regression modeling. Also called marketing mix modeling in the marketing world.
Assume that you have a sales on the y-axis, basically your revenue, and time on the a-axis. Normally a sales curve looks like the shape of a snake, varying up and down over time. Hopefully with a positive trend, but not necessarily. You normally have a horisontal baseline and this is how much you will sell if you fire all salespeople and you stop doing all marketing, you will still sell. Now what happens in reality if you stop doing anything around marketing and sales is that this curve will not remain flat, it will typically go down. Also there’s another reason why your baseline might go down and that is if you do very short-sighted marketing and sales, then your brand and PR etc is weakening, which makes your conversion become more and more expensive over time. Whereas if you’re investing in a good mix of branding, PR and performance marketing your baseline is going up.
Why is regression modeling or multiple regression modeling working?
With different spend levels, let’s say you have a total of 15 different tactics, X1 to X15 including retargeting, events, trade shows, Google search ads etc. What you’re trying to do with multiple regression modeling or marketing mix modeling is to create an approximation of the sales curve using these 15 variables. You will never find an approximation that is exactly like the sales curve. It might end up looking very similar. The mathematics behind regression modeling is trying to find the least error between the real sales curve and the approximated sales curve. Your sales curve, the y-axis, can be approximated by a combination of these 15 variables or at least those of the 15 that are significantly creating an effect. Then you might end up seeing that 11 out of the 15 are having a significant effect. This will allow you to approximate your sales curve by combining these variables and a variety of time delays. This mathematical formula, when run over time is going to create an approximated curve. In statistical terms this is called slope, we call it a proof multiplier. If you put in one dollar, how many dollars do you get back? And the time function is describing how the effect is spread out over time.
One important thing to know when you’re trying to quantify marketing, using simpler methods like Google Analytics and various customer acquisition cost models, is that you’re not taking the delayed effect into consideration. You might have very successful customer acquisition cost numbers but for some reason, the cost for customer acquisition is increasing. That’s a clear signal that you’re not analyzing marketing and PR holistically enough, you’re looking with too narrowly on the short-term effects.
Three important outcomes with Marketing Mix Modeling
First, marketing mix modeling can let you know what the slope is or what is the multiplier of the money. This number is typically higher in the marketing mix model than it is in a very short sighted model. Can you guess why? In the short sighted model you only see the value of what is immediately happening, whereas a lot of marketing has a spread out effect. The second thing you get out of marketing mix modeling is what is the delay function? When I put in $1 extra in retargeting, the effect is normally spread out over time. Retargeting doesn’t have a very long spread out effect, whereas a phenomenally good video might have an spread out effect over several years, in the same way with certain PR pieces. Various marketing activities can have very different delay functions. So when doing marketing mix modeling we can also start to quantify and look at a very short sighted marketing in the same model as we’re looking at very long sighted marketing. That cannot be done in traditional simpler models. The third thing you get from a multiple regression model or marketing mix model is the standard error. Meaning how much does the output vary from the average value? Events for instance, have a very high volatility because you can do the same thing between different events and one might still turn out much better. It’s very dependent on which people are there, timing and other things. In turn, the volatility on events is very high whereas the volatility and the standard error for retargeting is typically way lower. It is more predictable.
To summarize, from marketing mix modeling you get three important outcomes. First, what is the slope? So how much do I get back if I put $1.00 in? The second one is what is the delayed function looking like? And the third one is what is the volatility of the standard error of that variable?