Attribution: Who Gets Credit - Part II

The Good, the Bad and the Ugly. Let’s choose the ‘Right One’

What we will learn:

• Linear Attribution
• Time-Decay Attribution
• Position based Attribution
• Custom Attribution Model
• Use cases and steps for creating a custom model
• Data Driven/Algorithmic Attribution

In this article I would like to continue sharing my experience and learnings during the fifth week of studying Digital Analytics Mini-degree at CXL Institute. This article is a continuation of the ‘Attribution - Who Gets Credit’ series. If you are interested in browsing through part 1 of this series, you can do so here (link). The instructor for this course was Russell McAthy. CEO of Ringside Data, a Marketing Data Platform. Russell has garnered vast experience in attribution modelling and business strategy.  

In this part I focus on discussing more advanced attribution methods along with the capabilities that google analytics (GA) provides with regards to various attribution models, how we can use the model comparison tool, how to make a custom attribution model and a glimpse of multi-channel funnels in GA.

Linear attribution (LA)

This model distributes the credit through every touch point and interaction equally. This basically means that every channel that participated in getting the visitor to convert shall be rewarded equally. This model provides the picture of which channels are repeatedly driving conversions, but inherently ignores that fact that in reality all channels are not supposed to be equal in the first place.

When is it useful:

This model is useful if your campaigns are designed to maintain contact and awareness with the customer throughout the entire sales cycle. In this case, each touch point is equally important during the consideration process.

Source: Google

Cons:

• Does not show what campaigns and channels actually ended up impacting the user consideration.
• Cannot identify the relative usefulness of a specific channel.

Time-decay Attribution (TDA)

This model dictates that the closer a touch point is to the conversion (over time scale), the more influence it has on the consumer’s decision making. As this model is based on the concept of exponential decay, the last interaction has the most value in this model but what this is also saying is we want to take into consideration how long back through that journey that previous interaction happened. In google analytics, the time decay model has a default half-life of 7 days, meaning that a touch-point occurring 7 days prior to a conversion will receive 1/2 the credit of a touch-point that occurs on the day of conversion. Similarly, a touch-point occurring 14 days prior will receive 1/4 the credit of a day-of-conversion touch-point.

TDA can be employed in two ways. One with a static model and another with a variable model. The static model assigns the last interaction a fixed percentage and then the previous interactions will get fixed percentages based on the number of interactions in total. 

Russell gave a simple example to explain this model:

“So it could be that you give the last interaction 60% of the value and then the next interaction back 30%and the next interaction back seven and the next interaction back three. Now statistically, that makes sense.However, if the previous interaction before the conversion was a matter of days and then the interaction before that was a few weeks,those two are not the same time decay that you would want to apply.”

Whereas a variable model would take into consideration the amount of time between every single interaction before assigning values.

When is it useful:

If you run one-day or two-day promotion campaigns, you may wish to give more credit to interactions during the days of the promotion. In this case, interactions that occurred one week before have only a small value as compared to touch-points near the conversion.

Source: Google

This model can be of choice for products with longer sales cycles and higher number of repeat purchasers. For e.g. paid search may have originally brought a user to a website, but it was the promotional email that made that user to come back and repurchase.

Cons: 

• Does not recognize the original source as being widely significant.
• Cannot recognize high influence touches if they happened very early in the consumer journey.

Position-based (PBA) aka bathtub model

This model not only accepts that the first and last interactions are highly important but also gives credit to the interactions in between and this was popularized by Google Analytics as being the most accurate. PBA is kind of a hybrid between FCA, LCA and linear attribution. As shown in the image above, the first interaction gets 40% of the value,the last interaction gets 40% of the value and the remaining 20% is equally distributed between the two other interactions within the model.

When is it useful:

If you most value touch-points that introduced customers to your brand and final touch-points that resulted in sales, use the Position Based model.

Source: Google

Cons:

• Naively assigns high credit to first and last interactions which can result in two very low-value touches being given too much credit.
• Gives low credit to post lead identification interactions.

Custom Attribution Model

Custom model is somewhat an uncharted territory in the sense that it can get dangerous for marketers as they are able to set each touch point to have a different value which can quickly lead to inaccurate analysis if marketers fail to choose a point that reflects the context and reality of a specific business or industry.

Once you have dabbled enough around with all the above baseline models, you can move on to creating your own customized attribution models. Using one of the methods above as a baseline for the model, you can then make decisions on a unique look-back window, user-engagement and credit rules. When attempting this, the inherent problem remains, the model will be based on instinct and essentially arbitrary assumptions as to how your purchase funnel works.

Example use-case for the requirement of creating a custom model

Need: Assume your organization is using a last click attribution model and you need to justify your high spend on display advertising because you think your display channels may be undervalued because of your consumer behavior but your head thinks otherwise. Hence you need to find out whether display as a channel is undervalued or overvalued in your marketing efforts. 

Hypothesis: “If a user performs a conversion on the website/app within 6 hours of viewing (i.e. just an impression but no click) a display ad then the display impression should get a higher credit as compared to other interactions in the conversion path.

Now we need to test this hypothesis by creating a custom attribution model. 

Steps to create a custom attribution model in Google Analytics

1. Navigate to the model comparison tool under conversions > multi-channel funnels in GA.
2. From the select model drop down and choose ‘create new custom model’.
3. Name your model
4. Use the baseline model drop-down menu to select the default model you want to use as a starting point for your custom model. The baseline model defines how credit is distributed to touch-points in the path before the custom credit rules are applied
5. Enable look-back window to specify a look-back window of 1-90 days.
6. Enable Adjust credit for impressions to customize how impressions are valued.
7. Enable Adjust credit based on user engagement to distribute credit proportionally based on engagement metrics.
8. Set Apply custom credit rules to On to define conditions that identify touch-points in the conversion path according to characteristics
9.  Identify and specify how these touch-points will be distributed conversion credit, relative to other touch-points.
10. Save and then Compare your attribution model with other models in the model comparison tool based ROI and CPA metrics and evaluate your marketing efforts.

An example from google:

Data-driven/Algorithmic Attribution (DDA)

The other side of the ‘custom model’ coin is actually the data-driven or algorithmic models. So this is where you use basic algorithms through to more complex algorithms where you're using logistic regression,all the way through to machine learning,where you're able to put data back through the system and give it the ability to train and learn what values are really important. These kinds of models are usually used at a very large scale of data since it requires a certain amount of data going back a certain amount of time. There are some interesting reads on making attribution models based on Markov models using R and GA data.

In GA basically there are two main parts to the data-driven attribution methodology: 

• Analyzing all of your available path data to develop custom conversion probability models.
• Applying a suitable algorithm to that probabilistic data set that assigns partial conversion credit to your marketing touch-points.

The main advantage of using data-driven models is that this model eliminates any assumptions and biases of a marketer and uses your conversion data to calculate the actual contribution of each ad interaction across the conversion path. 

Let’s look at a simple example provided by Google:

This shows that DDA uses all available path data including data from both converting and non-converting users. The resulting probability models show how likely it is for a user to convert at a specific touch-point in the path provided that a particular sequence of events occur.

DDA then applies to this probabilistic data set an algorithm based on a concept from cooperative game theory called the Shapley Value. For more details on Shapley Value in the context of GA attribution click here.

Phew! I think that was a lot to absorb, why don't we use the day off to take it all in. Now that we have a better understanding of marketing attribution, try to explore and apply attribution models to ROI, CPA and feast on how the numbers look different.