February 16, 2011
Harley Norrgren studied Statistics at University College London and after a two year tenure is currently heading up Infectious Media’s Analytics Team. Being with Infectious throughout their entire working relationship with RTB he’s had the opportunity to watch the space develop since the start of RTB’s European adoption.
Regardless of what you think about attribution modelling, we can all agree that whilst practically useful, last click models tend to see attribution as a yes or no decision. Is a model that negates all exposure¹ history, apart from the ultimate exposure, necessarily painting the most accurate and useful picture of which advertising spend is actually generating conversions?
Thanks to the increasing digitalisation of our industry, exposure level data is readily available for building much more sophisticated and informative models, which can inform more efficient budget allocation and increase ROIs for clients. Using statistical methods we can isolate the uplift on probability of conversion that each individual exposure has and therefore assign attribution on an ROI rather than binary basis: exchanging the notion of a single exposure driving a conversion with an each exposure has a small effect approach. Furthermore these models can be implemented at the start of a campaign and updated throughout the campaign at predetermined intervals, providing simultaneous analysis and reactive attribution which would reward the efforts of effective advertising rather than competing media buyers simply gaming some system for what is essentially a random conversion attribution.
Approaches like these do take more time and require more skill to implement than last click models but there are three compelling arguments for their use:
To illustrate my point I’ve pulled some data from our platform on user behaviour under different retargeting conditions with a view to compare their behaviours against unexposed users.
We have chosen to examine the effect of aggressive retargeting versus less aggressive retargeting on a user’s propensity to make a return visit within 24 hours of the impression. Both types of retargeting will be compared against the natural return rate rather than against each other using a simple Chi2 test.
The aggressive retargeting (AR) group will be defined as users exposed to an advert within the same hour as the first site visit.
The less aggressive retargeting (LAR) group will be defined as users exposed to an advert between 24 and 48 hours after the first site visit.
For the AR group the results can be summarised as follows:
|Didn’t Return||Returned in 24hrs||Return Rate|
|Didn’t See an Impression||268,530||26,691||0.099397|
|Saw an Impression Within 1 Hour||949||84||0.088514|
For the LAR group the results can be summarised as follows:
|Didn’t Return||Returned between 24 and 48 hrs||Return Rate|
|Didn’t See an Impression||268,530||8,947||0.033318|
|Saw an Impression between 24 and 48 hours||2,456||123||0.050081|
We found that there was no significant uplift on the return rate (p-value = 0.3354) for the AR group, but there was a significant uplift on the return rate (p-value = 1.327E-5) for the LAR group despite the overall return rate being lower. A last click model would have given attribution to the AR group and suggested that it performed better than the LAR group, but this is plainly not the case. An exposure level attribution model could discriminate between significant and insignificant exposures, assigning attribution to where behaviours were driven and painting a much more realistic picture of advertising effectiveness.
Furthermore, tying in campaign setup with attribution model insights generated on the client’s side has become easier than ever before: granular campaign targeting is available through most platforms so making the transition could be easily achieved on the buy side. On the client’s side, implementing and updating a multivariate attribution model based upon maybe billions of rows of data is no simple task. Yet the industry is already starting to rely on big data to inform their advertising decisions and when clients want the improved results others may be achieving they’ll be playing catch up if they don’t start experimenting now. So my advice is to start small and when the time comes for de-facto bespoke attribution modelling you’ll be ready for it.
Of course this is quite a small insight into a vast array of possible analyses that could inform attribution model specification and I’d be keen to hear your opinions on this.
¹ For more information on exposures and interactions, please see Measurement: The Elephant in the Attribution Room where we discuss measurement in attribution models.