MMM vs GBHMMM
Why Geo levels Bayesian Mix models are to most advanced MMM for performance Marketing.
In order to allocate budgets per medias channels you typically have two choices:
- Using MTA
- Using MMM
Both are separate methods, usually MTA is a more tactical approach to segment on keyword, ad-set or banner level, meanwhile MMM is used more on strategic level to see what are the main drivers that impact sales with internal & exogenous factors combined with digital advertisement.
What is interesting, Is that you can sometimes have models that performs mix between MTA and MMM in the case of Neustar they offer a mix between both.
Typically with MMM you want to measure uplift/downlift in a response variable: Visits, sales or reach for example by modelling– In performance advertisement usually the main KPI is sales — Modelling is a pretty robust method when you want to isolate all the factors that could impact your sales (price, promotion, placements, products, competitors) and econometrics factors (weather, GDP, PIB, unemployment rate) by doing so you can have a solid picture of your incremental return on investment.
Why MMM is not good enough? While MMM seems to be a pretty solid when there is a significant change in the investment strategy (on&off) or very solid correlations between investment/sales. However usually when the model computes in a National level we might not be able to drive statically significance and any fluctuation can be random.
By applying a GEO level Bayesian Marketing Mix modelling we can have a much more robust model.
In a nutshell:
(1) In the descriptive analysis having more inputs of data and more variablity in the spend will improve your model.
(2) The marketing spend at the geo level generally has a wider range than at the national level, which is critical to MMM as insufficient variation often leads to extrapolation issues.
(3) If you have done GEO experiments, (Vaver & Koehler, 2011) the geo-level model will outperform a national level model — Independent variation in marketing spend across geos from the experiments offers the possibility to improve MMM results by eliminating or reducing ad targeting bias and increasing the effective sample size.
Conclusion:
If you have the opportunity to query your media investment per geos, then will you should definetly give it a go! :)
Cheers, Gus.
