WHAT ARE THE KEY INPUTS TO OMNICHANNEL CUSTOMER ENGAGEMENT ANALYTICS?
Below are the high-level categories of inputs to customer engagement analytics (we will leave external factors and spend for another day):
Sales
Promotional Activity
Customer Attributes
The devil is in the data quality details and the first two are straightforward. Other than lining them up longitudinally, there is not much left to do. The sales are what they are and you interacted with the customers in the way that you did. It is the last bullet – customer attributes – that we want to focus on.
First some modeling background: it is easy to say (and we all want to say it) that engagement analytics are performed at the individual customer level. In practice, however, that is not typically the case. It is difficult to make accurate predictions by an individual customer because the individual lacks enough cross-promotional data, and execution variation is limited. Predictions on next best action (NBA), for example, are actually done on cohorts or groupings of customers with similar characteristics to overcome these challenges. The final recommendation is for the individual customer, but it is based on the group that shares similar characteristics with that individual. In other words, if the group behaves this way, then all the individuals within that grouping will follow the same pattern. The trick to doing effective customer engagement analytics is to get the cohorts as fine-grained as possible.
To explain, think about what would happen if you used one of the most common classifiers, decile, as the sole group attribute to predict Health Care Provider (HCP) behavior. We will assume Decile 10 (D10) as our example. Everyone in D10 is either a high brand or high market prescriber depending on what you look at. This is a shared attribute across the people in the group, but you will have a lot of HCPs in the group, and they do not all respond the same way to promotional activity. If you were to base your NBA recommendations solely on the decile attribute, you would have enough variation and interactions to model, but applying the group findings to all the individuals within the group would lead to many incorrect recommendations.
The logical next step would be to add in some other readily available attributes such as geography and specialty to get somewhat better predictions and recommendations. It is more likely that HCPs within a geography/decile/specialty group vs. a decile only group will behave similarly. However, much like decile only, adding in these high level attributes will only marginally improve individual recommendations since the groups within each are still quite large.
So what do you do? The “trick” as we called it previously is to get as many relevant customer attributes as you can to define the group while keeping enough data and variation to make accurate predictions. With the right mix of customer attributes you have a high confidence that the individuals in the group behave the same and your recommendations are the best ones possible at the individual customer level.
Getting that right mix of customer attributes is complicated for several reasons. First, there are a lot of potential customer attributes to go after and getting the data for those attributes can be difficult, time consuming, and expensive. On the expensive side, formulary status and office associations for HCPs would make the list. On the time consuming side, it is important to consider all the attributes that can be gleaned from claims data, including line of therapy, persistence, and treatment patterns.
Second, you have to determine which attributes matter. Running engagement models with unnecessary attributes can have significant performance implications. There are many upfront analyses that need to be performed to understand the relevance of each potential attribute individually and in conjunction with others.
Once you have identified the right mix of attributes and have gotten to a fine-grained level for your cohorts, the payoffs in analytic accuracy (confidence) and better predictions can be substantial. Instead of saying that all HCPs within a group of potentially tens of thousands behave the same, you are now saying that 10-30 HCPs within a group defined by a dozen relevant attributes behave the same.
Having a robust set of HCP features available is the most important input to engagement analytics, but it can be the most difficult to master. Spending the time to get the HCP attributes right for Customer Engagement analytics provides the assurance that you are making accurate predictions and delivering usable results to drive meaningful customer interactions.
----
SENTIER was founded in 2017 to disrupt analytic insight generation and delivery in Pharma. Through VELOCITY, the industry leading Analytic Operations as a Service (AOaaS) platform, our customers are able to maximize the ROI on AI and ML activities by allowing non-data scientists to deliver answers as frequently as the business demands. We also support our clients with advanced marketing & sales analytics and support services designed to prepare models and data pipelines for production. SENTIER believes that the ethical application of solutions will benefit patients and health care providers as well as the Pharma and Biotech companies we work with.
