Problems of standard attribution modelling

Attribution modelling is like digital magic.

Introduction

Wow, so I’m reading a great piece by Funk and Nabout (2015) [1]. They outline the main problems of attribution modelling. By “standard”, I refer to the commonly used method of attribution modelling, most commonly known from Google Analytics.

Previously, I’ve addressed this issue in my digital marketing class by saying that the choice of an attribution model is arbitrary, i.e. marketers can freely decide whether it’s better to use e.g. last-click model or first-click model. But now I realized this is obviously a wrong approach — given that the impact of each touch-point can be estimated. There is much more depth to attribution modelling than the standard model leads you to believe.

Five problems of standard attribution modelling

So, here are the five problems by Funk and Nabout (2015).

1. Giving touch-points accurate credit

This is the main problem to me. The impact of touch-points on conversion value needs to be weighed but it is seemingly an arbitrary rather than a statistically valid choice (that is, until we consider advanced methods!). Therefore, there is no objective rank or “betterness” between different attribution models.

2. Disregard for time

The standard attribution model does not consider the time interval between touch-points – it can range anywhere from 30 minutes to 90 days, restricted only by cookie duration. Why does this matter? Because time generally matters in consumer behavior. For example, if there is a long interval between contacts A_t and A_t+1, it may be that the effect of the first contact was not very powerful to incite a return visit. Of course, one could also argue there is a reason not to consider time, because any differences arise due to discrepancy of the natural decision-making process of the consumers which results in unknown intervals. Ignoring time would then standardize the intervals. However, if we assume patterns in consumers’ decision-making process, as it is usually done by stating that “in our product category, the purchase process is short, usually under 30 days”, then addressing time differences could yield a better forecast, say we should expect a second contact to take place at a certain point in time given our model of consumer behavior.

3. Ignoring interaction types

The nature of the touch or interaction should be considered when modeling customer journey. The standard attribution model assigns conversion value for different channels based on clicks, but the type of interaction in channels might be mixed. For example, for one conversion you might get a view in Facebook and click in AdWords whereas another conversion might have the reverse. But are views and clicks equally valuable? Most marketers would not say so. However, they would also assign some credit to views – at least according to classic advertising theory, visibility has an impact on advertising performance. Therefore, the attribution model should also consider several interaction types and the impact each type has on conversion propensity.

4. Survivorship bias

As Funk and Nabout (2015) note, “the analysis does not compare successful and unsuccessful customer journeys, [but] only looks at the former.” This is essentially a case of survivor bias – we are unable to compare those touch-points that lead to a conversion to those that did not. By doing so, we could observe that a certain channel has a higher likelihood to be included in a conversion path [2] than another channel, i.e. its weight should be higher and proportional to its ability to produce lift in the conversion rate. Excluding information on unsuccessful interaction, we risk getting Type I and Type II errors – that is, false negatives and positives.

5. Exclusion of offline data

The standard attribution model does not consider offline interactions. But research shows multi-channel consumer behavior is highly prevalent. The lack of data on these interactions is the major reason behind exclusion, but the the same it restricts the usefulness of attribution modelling to ecommerce context. Most companies, therefore, are not getting accurate information with attribution modelling beyond the online environment. And, as I’ve argued in my class, word-of-mouth is not included in the standard model either, and that is a major issue for accuracy, especially considering social media. Even if we want to measure the performance of advertising channel, social media ads have a distinct social component – they are shared and commented on, which results in additional interactions that should be considered when modeling customer journey.

Solutions

I’m still finishing reading the original article, but had to write these few lines because the points I encountered were poignant. Next I’m sure they will propose solutions, and I may update this article afterwards. At this point, I can only state two solutions that readily come to mind: 1) the use of conversion rate (CVR) as an attribution parameter — it’s a global metric and thus escapes survivorship bias; and 2) Universal Analytics, i.e. using methods such as Google’s Measurement Protocol to capture offline interactions. As someone smart said, solution to a problem leads to a new problem and that’s the case here as well — there needs to a universal identifier (“User ID” in Google’s terms) to associate online and offline interactions. In practice, this requires registration.

Conclusion

The criticism applies to standard attribution modeling, e.g. to how it is done in Google Analytics. There might be additional issues not included in the paper, such as aggregate data — to perform any type of statistical analysis, click-stream data is a must have. Also, a relevant question is: How do touch-points influence one another? And how to model that influence? Beyond technicalities, it is important for managers to understand the general limitations of current methods of attribution modelling and seek solutions in their own organizations to overcome them.

References

[1] Funk, B., & Abou Nabout, N. (2016). Cross-Channel Real-Time Response Analysis. in O. Busch (Hrsg.), Programmatic Advertising: The Successful Transformation to Automated, Data-Driven Marketing in Real-Time. (S. 141-151). Springer-Verlag.

[2] Conversion path and customer journey are essentially referring to the same thing; perhaps with the distinction that conversion path is typically considered to be digital while customer journey has a multichannel meaning.