Business Case: Spillover Analysis

Online Marketing: Consideration of the "Spillover Effect" in Attribution due to View Contacts by Means of Competing Risk Procedures


The aim of attribution analysis is to find out the contribution of individual marketing channels (SEO, SEM, newsletters, affiliates, etc.) to a customer's purchase decision. This way the budget for online marketing can be optimally distributed to the various channels across all customers.

Common practice is to concentrate exclusively on the click contacts of the customer journeys. However, this approach ignores the influence of view contacts, such as the insertion of a banner advertisement. These view contacts often trigger additional click contacts within a certain time frame: they don’t result in a click on the displayed banner itself, but induce a “spillover effect” into non-display channels at a later point.

For example: Benni is surfing the Internet. In a forum, he is shown a banner advertising a fashion provider’s new t-shirt collection. Ten minutes later, after Benni has left the forum and checked his email, he googles the fashion provider. A Google ad takes him to the fashion company's website, and he buys three t-shirts. Under normal circumstances, this purchase would be completely attributed to the SEA channel, although the unclicked banner ad is the relevant trigger for the purchase.

The task is to prevent the misallocation of the marketing budget due to the disregard of displayed-but-not-clicked view contacts. To achieve this, the spillover effect between the view and the subsequent click must be taken into account during attribution.


The analysis is based on customer journey data. This means detailed data about the different contact points of a potential customer on the website: essentially the customer's transaction data from their first visit at any point on the website to the completion of their order. In addition to the usual click contacts, all contact points with banner advertising displayed but not clicked on are also required. The decisive factor for the analysis is the information on the chronological order of the contacts, as well as the characteristics of the banner advertising (such as format and type), the day of the week, and the time of the contacts.


After the appropriate preparation and validation of the data, the attribution model can be extended to include spillover analysis.

The spillover analysis is based on the Competing Risk method, which was originally developed as an extension to survival time analysis in biometrics. In this procedure the relevant click channels that are related to previously-displayed banner advertising are identified. The concrete contribution of the views is measured for these channels, and the time frame within which the views "work" after their insertion is determined. It is also determined which characteristics of the views are responsible for the spillover effect and its strength. The identification of cause-and-effect relationships makes it possible to calculate the potential for triggering a subsequent click ("trigger potential") for each banner ad delivered.


The results of the spillover analysis are taken into account in the automated attribution calculation. This corrects the systematic underestimation of banner advertising. The difference between the attribution results with and without views is explicitly displayed. In this way, campaigns can be specifically evaluated with regard to their spillover performance. Our experience shows that the increase in the share in the channel display can range from 5% to 20%.