The New Age Market Mix Modelling. It is simple but not easy

Most marketers think that building MMM models is a simple task and hence easy to perform. But, as the adage goes-“the simple is usually never easy”. In this article, we will discuss why MMM models are not as easy to build as is commonly assumed.

First, comes context. Marketers are interested in measuring the impact of their marketing initiatives like advertising, pricing, consumer promotions, distribution, and trade promotions on metrics such as sales or market share.

They look to quantify the impacts of their inputs, by estimating their contribution to specified business metrics like sales, volume, value, and goodwill. Contribution is also called the effectiveness of the input. Effectiveness, hopefully, brings in large volumes of sales, rather quickly, but could cost a lot (e.g. effective media include TV, Print, and Outdoor Billboards )

They will then proceed to compute the Return on Investment for that input. We shall call this the efficiency of the input – the sales obtained for every rupee of investment. Efficiency usually brings in smaller volumes of sales but can be very inexpensive (examples of efficient media include email, postal mail, SEM, and sponsorship of events)

Both the contribution & ROI will then be used by marketers, to improve the impact of inputs, by moving money to the more productive initiatives, from the more unproductive activities.

‍

This appears, at first glance to be a straightforward optimization problem

• Decide on an objective function (maximize sales or market share )
• Specify the decision variables –  advertising, promotions, price, and distribution
• Pop in a few business constraints – most likely a budget and some capacity constraints of media channels
• Call on Newton & Raphson to deliver a solution
• Measure the lift

So what’s the big deal?

For one, there are more than fifty marketing variables to contend with. And they are all correlated amongst themselves. Ordinary regression is not capable of handling lots of variables without sufficient data and can certainly mess up signs, under the influence of multi-collinearity.

Relationships amongst them could be one-way, reciprocal, multi-way, or indirect. TV would impact sales and increased sales calls for higher budgets- this is reciprocity. And TV could drive website visitors who may then buy at a later point in time- this is indirect. TV could impact both search counts and website visitors, who may make a purchase at a later point in time- the multiway effect.

Relationships could be linear or non-linear, symmetrical or asymmetrical. Non-linearity is due to carry-over effects, or due to thresholds and saturation properties. Asymmetry is usually overserved in price changes, where a 10% increase acts very differently on sales from a 10% decrease.

Media and promotions do not work in silos but in synergy. Synergy is the premium marketers expect to get when they deploy TV in conjunction with Digital or Print with Outdoors. There is also the halo effect to exploit when two sister brands and advertised together.

Advertising has both long-term and short-term effects. Short-term sales are detected via the contribution. Long-term shows up in the increase in baseline sales. The ordinary regression model creates a constant baseline in the form of an intercept.

Advertising impact can be due to two sources, the creative and the media deployment. The two impacts are confounded and need to be separated. High-impact creatives need much less investment behind them. Today’s marketers are interested in this separation.  

Finally, in the digital world, campaigns are run consistently and in an “always on” mode. Digital applications spit out metrics such as Cost per Click, Click-through rate, and Cost per sale (ROI). A Marketing Mix modeling may produce results that are much smaller than quoted by digital applications. The MMM results and digital results must be reconciled.

Even if all these considerations are taken into account, models can still contain errors. While one may never get to the exact truth, it is imperative that we are not far from it either. George Box the famous statistician said it so succinctly - “all models are wrong but some are useful”.

‍