By using this Site, you consent to the use of cookies. For more information, please see our Privacy Policy.


How Accurate is Attribution Modeling?

Volume 5, Issue 3 - March, 2015

Andy Dubickas, Sr. Solutions Consultant, Visual IQ

There are many different approaches to marketing attribution, ranging from the overly basic to the incredibly sophisticated. All of them will “get you an answer,” but the questions is, how accurate is that answer?

It’s widely acknowledged that last click and subjective, rules based attribution (such as even-weighting, time decay, linear, and opener/advancer/closer models) are inaccurate methods for measuring marketing effectiveness. And while the advancement of data-driven, algorithmic attribution is an exciting development, there are still limitations to model accuracy. Let’s take a closer look at the varying degrees of accuracy in three algorithmic approaches:

Common Modeling
Common attribution models examine customer conversion paths and use techniques such as logistic regression or estimators to infer which media elements are driving the most consumer response. Since massive sample sizes are required to infer causal relationships, these models categorize interactions into larger groups, such as by channel or by campaign. As a result, common modeling techniques can only provide high-level directional insights into the contributions of each channel or campaign, with no granular level of optimization ability at the publisher, placement or keyword level. Since accurate attribution relies on having large amounts of granular-level data to determine how channels and touchpoints work together to produce outcomes, much is left to be desired when it comes to the accuracy of this approach.

Path-based algorithmic models compare the behaviors of consumers who took similar conversion paths, with one touchpoint added or removed. This approach enables marketers to see how a particular touchpoint impacted consumer behavior, and translate that into a fractional amount of credit given to that touchpoint. As with common models, a massive data set must be used in order to understand how individual placements, sizes, creatives, etc., impact consumer response. If comparisons cannot be made at these low levels of granularity, then shrinkage estimators such as Bayesian techniques or Markov chains are used to group media together and model at a higher level of granularity (such as grouping all of the publishers within a campaign, or grouping all of the placements from a particular publisher, and assuming the same performance for all). However, performance can differ wildly among publishers within campaigns, and even more so among placements within publishers, so to assume all media is performing the same will cause results to be off track.

Path-based attribution also leads to questions about accuracy when optimizing media based on the recommendations it provides. The consumer-decision journey may consist of millions of interactions, yet path-based models are based on the behaviors of consumers who took similar conversion paths. Since you can’t force a consumer to take a pre-defined journey or interact with media in a certain way, the optimization recommendations that are put into market may not drive the anticipated results.

Multi-dimensional is the most accurate of the algorithmic modeling approaches. It uses the full dataset of both converters and non-converters to perform billions of A/B comparisons between those exposed and those not exposed to different dimensions (attributes) of media. The result is a breakdown of every single touchpoint at the most granular level (ad size, placement, publisher, creative, offer, etc.). Since the data set used is exponentially larger than any other approach, the need for estimators like those used in common modeling and path-based approaches is eliminated. The outcome is an in-depth understanding of how each touchpoint and each attribute of each touchpoint changes the propensity of a consumer to take a desired action, so more specific and impactful optimization decisions can be made.

While algorithmic attribution offers a sophisticated approach to measuring marketing performance, the accuracy of results it provides will vary depending on the model behind it. Understanding these caveats will not only help you select the right solution, but also ensure you’re getting the complete, granular insight you need to maximize your marketing ROI.

Learn more about our cross channel marketing suite of products.

Subscribe to the IQ Advisor

Click here to subscribe to The Visual IQ monthly e-newsletter.