By Annmarie Turpin SVP Analytics, Ocean Media
In the eternal quest to answer performance marketing’s most burning question – which marketing channels drive the highest ROI – the prevailing approaches fall in to one of two camps: multi-touch attribution or media mix modeling. Diametrically opposed in approach (bottom up vs. top down), each approach brings its own value and drawbacks.
The promise of multi-touch attribution (MTA) in which every trackable touchpoint is incorporated and considered is undeniably appealing. Such a detailed and data-driven approach seems to offer a high degree of precision. Perhaps the greatest threat to this promise hinges on trackability of touchpoints.
Specifically, the walled gardens of data within digital (i.e. Facebook who does not share the clickstream data necessary to be incorporated into an MTA) and the largely un-trackable machine of offline marketing (i.e. linear television and radio) can leave major gaps in the analysis. Further challenges to the accuracy of MTA modeling lie in the accuracy of establishing identity, a formidable challenge which is being exacerbated by new terms of privacy and depreciation of cookies. While some brands are content to continue the model with missing data, these data gaps leave opportunities for misattribution and inaccuracies.
When exploring an MTA approach or partner, advertisers should seek out partners will capabilities and data sets with coverage across all screens and all types of media. Additionally, great care should be taken to understand the methodology and accuracy of the identity graph driving the solution and data to be used.
Alternatively, one might consider the media mix (MMM) approach for attribution in which large amounts of input data are collected over a long period of time but often aggregated less granularly than MTA – either daily or weekly vs. touchpoint data. These models are more easily executed, are more comprehensive in their coverage of the entire marketing portfolio and have been widely adopted and consistently applied across time, which facilitates longitudinal tracking of results. MMM carries its own drawbacks as well.
The greatest drawbacks of MMM could be the following: 1) the models are static in their output only changing when the model is re-run. Models are infrequently re-run as the results are unlikely to change substantially before six months has elapsed. 2) A MMM output represents an average of the past and may not provide an accurate depiction of the most recent performance of any given channel. 3) As such, results might be considered more directional than MTA – hammer approach vs. scalpel approach. And lastly, 4) MMM relies on correlations for attribution, which can be spurious.
To summarize, neither approach is surefire, and every execution of each approach will have its own unique pitfalls. As stated by George Box, “all models are wrong, some are useful”. With this in mind, the process of channel level attribution should be considered a journey and not a destination. Making space for multiple approaches in answering the question allows for cross-validation and for reliable trends to emerge. Tracking changes in the output using the same methodology and comparing across approaches over time builds a holistic and representative depiction of channel-specific attribution. Perhaps the winningest strategy to understanding attribution is one that is agnostic to approach, critical of the inputs, and favors the collection of many well-built models as is affordable and feasible.