Search:


Data Driven Attribution Modeling in Roivenue

 

DATA DRIVEN ATTRIBUTION MODELING IN ROIVENUE

 

Now is the time to bring your analytics to a whole new level - forget about non-direct last click, data driven attribution is here. In this section we will go through the process of model selection and how to switch the main model of the app, according to which channels will be evaluated. If there is a time to get really excited, the time is now.

The Attribution Analysis section is the place in Roivenue which gives users the ability to compare available attribution models and modify the main model of attribution which will be used to recalculate data in the Performance Monitor.

 

 

To define the displayed graphic, you can use the following functionalities:
- Metric : select metric which will be recalculated for each model. Please note that unpaid channels will not be displayed for ROMI, MIR, mROMI, mMIR.

- Attribution models : check the models you would like to display, up to six models can be selected.

 
 
 
 
There are three tabs for making comparisons between the selected models. These are:
 
 
 
 
 
1. Attributable [metric] - displays selected metric by channel for selected models on a horizontal column graph. This can be seen on the picture below. Each of the models have corresponding color and the displayed models are selected from attribution models dropdown menu.
 
 
 
 
 
 
 
model-comparison_orig (1)
 
 
 
 
3. Explore the numbers - compares values for each channel in a matrix. Each column is assigned to an attribution model and cells in this column are conditionally formatted based on their value compared against last touch - less than last touch is red, more is green.
 

*Sorting can be applied via dropdown menu on the top right corner of the screen. All results can be filtered by channel and date on the right hand side of the screen.
 
 
 
 
The final step of this analysis is the selection of an attribution model. Roivenue offers three data driven attribution models: Shapley Value, Markov 1st and Markov 2nd order. The question is which model suits your situation better, as each has a different logic and thus different strengths.
 
 The Shapley value evolved from attempts to optimize formation selection by ice hockey coaches. The calculation simply compares paths which resulted in conversions with the paths which did not, for each channel. The higher Shapley value is, the higher the ratio of successful participation to unsuccessful. Therefore, this model tells us whether the contribution of a single channel to the whole marketing mix is positive or negative, but it does not take into account a place on which the channel stood in the conversion path. For this reason, it is usually used as a supporting model to Markov chain models.
 
Markov models are probability models, which are based on calculation of likelihood of each possible transition from a touchpoint. This is calculated for for every channel. Markov models can be calculated on several orders, with each order signifying how many jumps over touchpoints are taken into account. In Roivenue, we work with order 1st and 2nd. As their mechanics are very similar, we can explain them on Markov 1st as it is a bit easier to comprehend.
 
 
 
 
 
 
 
 
 
 
 
The diagram below proposes an example schema of Markov 1st order.  It provides a scenario of channel A and its evaluation according to this model. There are several options for advancement from the current state to the next state. This advancement means the next action the customer will take. B, C, D, and E represent other marketing channels, the checkmark represents a conversion and the X represents no action. Each of these steps have a certain probability, so for example the probability that a customer will touch channel C after touching channel A is 20%.
 
The higher is the order of the calculation, the more transitions are taken into account. For example, Markov 3rd order takes into account three transitions, so one group of transitions would be Channel A - Channel B - Channel C - Conversion. Higher orders decrease probability for each transition and therefore decreases statistical strength of the calculation. That is why we work with Markov order chains up to 2nd order in Roivenue attribution modelling.
 
markov-chain-example_1
 
 
 
 
 
 
So, for the big question - which model is the best for you? 
 
 
 
 
 
 
 
 
 
 
 
 
main-model-selection-gif_orig (1)
 
 
 
 
 
 
 
Attribution coefficients are recalculated on a weekly basis. So at the beginning of each  week, Roivenue collects all conversion paths from the last complete week, calculates models and sets conversion coefficients. These will be updated each week. You only need to select your model once, Roivenue will use it for the selected business unit until further changes are applied. Please note that the selected model changes conversion distribution for the whole account, so your selection will impact data displayed to all users who have access to this account. Therefore, it is better to properly manage your user rights - limit administrator and operator access only to core members of your team.
 
 
 
 
 

 

*"Attribution Analysis" can be accessed via "Marketing" section in the left hand filter.
**In order for this page to be available, attribution must be properly set up and data collection must be up and running for at least four weeks.

 

attribution-gif_3_orig (1)

 

 

attribution-screenshot_orig (1)

 

 
2. Compare selected models - Compares selected models on a vertical graph. In case only one model is selected, it is compared against last touch. Under each channel there is an index, which describes relationship between the recalculated values. For example when there is an index of 0.8, it means that according to the selected attribution model, the channel is performing at 80% of its performance according to last touch. When there are two models checked in the model selection, these two models are compared against each other and the index is recalculated on these two models.
 
 
explore-the-numbers_orig (1)
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
Markov 1st order model can be explained by the following steps:
- Inputs are several channels (touchpoints) and a set of collected data about transitions between these touchpoints. We can call these channels A, B, or C.
Let's take channel A as our example touchpoint. Markov 1st order estimates the probability of direct transition from this touchpoint to any other touchpoint or a conversion. There will be some probability that a user visits channel B right after channel A, then there will be some other probability that the user will visit channel C after A and so on. One of the transitions will be a transition from channel A to a conversion.
First, the total number of conversions is calculated for all channels
Next, it removes one channel from the mix and calculates total number of conversions without this channel, based on previously calculated probability transition between the channels
The difference between the total number of conversions and the sum of conversions after channel A is removed equates to the number of attributed conversions to channel A according to Markov 1st order model
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
Markov 2nd order model follows the same logic, it just looks overa pair of touchpoints instead of one. When applied to the Marketing Attribution, probabilities of transition are calculated on pairs of channels. For example, what is the probability of transition from the pair Facebook and AdWords to Direct, Conversion, or None.
 
 
 
 
Well, that depends on volume of your traffic. Markov 2nd order provides superior insights about channel collaboration, but it is less robust for websites with less conversions. The reason for that is that it divides total conversions into higher number of groups (one group is not a single channel, but a pair of channels, so number of groups increase exponentially with each order of markov chains) and this causes each group to contain less data. And less data means less precision. The best practice is that if you have a lot of conversions (more than 500 a week) feel free to go for Markov 2nd order. If you are not yet on this level, stick Markov 1st order.

After picking one of the models to become the primary model of the app, enter the "model selection" menu, accessible via Change Main Attribution Model button.
 
 
 
 
 
 
 
 
After selecting one of the models and confirming your selection, this model becomes primary. What it means is that conversion, revenue, and profit distribution in the Performance Monitor and Dashboard are recalculated according to this model. Total numbers stay the same, what changes is the distribution of totals among the channels. It takes few minutes for your selection to take effect, as Roivenue needs to recalculate the data from last touch.