Curse of Correlation, how to measure marketing performance

Many times when organizations run digital campaigns, managers tend to look at Return on Ad spend (ROAS) numbers reported from attribution tools. Most of these attribution systems have simple rule based attribution which is either last click or last view attribution. The ROAS calculated using these techniques is flawed and this may sound controversial. 

Lets take a simple example: a user sees an ad o a social media platform then while on a streaming platform sees another advertisement. Then the same user after a few days goes online and searches for that product, sees a google search ad, clicks on it and makes a purchase. Now giving all the credit to Google for the purchase using the attribution system will be flawed. Some marketers may use multi touch attribution system. Usually these systems use some form of machine learning algorithm to solve for a logistic regression problem. As we all know that regression still is a measure of correlation and not really causation. Now the biggest question is that is measuring attribution a useless exercise. My response to that would be a Big NO! 

The solution to this curse of correlation is finding a good measure of causation. One of the best way of measuring causal impact is by experimentation which uses the principles of randomized control trials. In digital advertising world there are multiple ways of measuring performance. Different channels have their own experimentional tools for calculating incrementality. If the marketer does not trust first party tools then one can always leverage Geo-based experiment testing. With the above mentioned approach you can always evaluate the performance of each channel and calculate experimental Return on Ad Spend. This ROAS will be causal and not really correaltional. 

Now the big question how do we make sure that Attribution and experimentation based approach point to the same direction. The solution to that problem is calibration. We can take the inputs of experiments (RCT ROAS) and put them as inputs to the Multi touch attribution system. At Diagnolytics we have a Bayesean regression based approach for calibration. Reach out to us if want to know more.  

Last but the not the least I want to end my article by stating that marketing measurement is hard and there is no perfect solution. The general consensus is that in case of such complicated systems it is best to rely on experimental based measurement to get a sense of the true performance.