• http://cashwithatrueconscience.com/rbblog Ryan Biddulph

    I dig the idea of qualifying each marketing channel. Serious analysis with a bottom line approach helps you see what’s working and what’s holding you back, channel wise. Then you can rock it out as you trim the channel fat to focus on the prospering strategies.

    I found this post on Kingged.com.

  • Nico N

    With all respect, this comparison is pretty flawed and confuses basic principles of statistics. First of all, I understand the distinction between top-down and bottom-up models, but your choice of “machine learning” for the third group is rather unfortunate. Machine learning is a method and can be used for all kind of models, including the first two mentioned approaches (see random forests). So it’s not a separate group (because that would be like saying a BMW is not a car). BTW: A more traditional way to cluster the first two is to differentiate between a) aggregate data vs. b) individual-level data models.
    Second, please note that both aggregate-data and individual-level models can be descriptive, explanatory, or predictive. This completely depends on the actual type of model, estimation technique, and set up one eventually uses (compare a VAR* with a structural VAR for time-series models). There are also very different advantages between individual-level models and aggregate-data models. E.g., the former are less prone to some crucial inference biases, while the latter typically can distinguish between short- and long-term effects (at least in case of time-series models, which are the typical applications for marketing-mix modelling). Again, depending on model set up.
    *VAR=Vector Autoregression

  • Kohki Yamaguchi

    Hi Nico, to your points on what’s theoretically possible using different methodologies, I agree wholeheartedly. But at the same time, this article is meant to be accessible to a wider audience, so any sort of simplification results in loss of information. :-)

  • Nico N

    Kohki, thanks for your reply. I am completely with you regarding simplification to make the boring and complex world of statistics more appealing to a wider audience. However, the classification suggested above is simply wrong and I am particularly concerned that this makes the rounds (for example, Datalicious already reposted this classification on their blog in Australia). I already see our clients coming to us asking for an “attribution” model because they do not want an “econometric” model.
    Since Marketingland.com has such a large audience, it is particularly important not to mix up basic concepts (in terms of education). To make it really simple, we could say that on the one hand there are
    (1) top-down models (aggregate-data) and
    (2) bottom-up models (individual-level data).
    On the other hand,there are models that use
    a) real measured market data (survey / purchase data) or
    b) simulated market data
    to provide insights into marketing-channel effectiveness. Of course, the different forms can be mixed and combined.