In a recent engagement, Strong worked with a biopharmaceutical company that was looking to answer questions related to multi-touch attribution. This industry leader in discovering, developing, and delivering innovative medicines for diseases was interested in understanding the effectiveness of their marketing to healthcare providers (HCPs), specifically in the context of new-to-brand prescriptions (NBRx).
They had a proof-of-concept model, but it needed to be scaled up to run in a production environment. It also needed to be parameterized so that MTA could be run for different brands, franchises, and indications. This parameterized, scalable pipeline also needed to produce quality results enabling the company to make informed decisions about how to allocate budget across multiple channels.
Advertising efforts typically span multiple channels, each varying in frequency and efficacy. For instance, a company may incorporate prime-time commercials, daily emails to spotlight ongoing promotions, or weekly coupons sent via postal mail.
It's crucial, however, to recognize that effectiveness is not uniform across these channels or even among publishers within a single channel. This raises an important question for companies: which channels or publishers should they prioritize to maximize their marketing efforts?
In a perfect world, companies would carry out controlled experiments that take into account all potential confounds and run long enough to collect statistically significant results. This ideal solution is often cost prohibitive, but even when it is not, the frequency at which experimentation can occur is not reconcilable with business cycles. A more pragmatic solution would leverage data that companies already have to measure the past effectiveness of marketing and advertising campaigns to serve as a heuristic for making decisions now about future advertising spend.
Enter Machine Learning
Both media-mix modeling (MMM) and multi-touch attribution (MTA) measure the effectiveness of marketing channels with respect to a given outcome, but they differ in scale. MMM seeks to answer more macroscopic questions such as the following:
- How do external factors, such as seasonality or macroeconomic trends, impact the overall performance of various marketing channels?
- Which marketing channels are most effective in driving incrementality or lift in key performance metrics?
- What is the long-term impact of marketing activities on brand awareness, customer acquisition, and customer retention?
MTA on the other hand, attempts to answer more granular questions:
- How do individual touchpoints contribute to conversion within a specific user journey?
- How do different marketing channels interact with one another, impacting their effectiveness when used together?
- How does the effect of touchpoints vary across different customer segments or demographic groups?
Strong built a highly parameterized, fully unit-tested, scalable python package that performs MTA across various BFIs. At its core, we predict the probability of an NBRx based on a set of control and independent variables. We then used a secondary explainer model to assign a partial contribution to each of the control and independent variables.
On the backend, we calibrate explainer values to account for the sigmoid distortion in probability densities produced by models, and employ a novel adjustment scheme that allows us to attribute a positive partial probability contribution to each control and independent feature. Moreover, being controlled by a single configuration file and triggered by a command-line interface, the package allows for rapid experimentation of features, study-period lengths, market baskets, etc.
This significantly shortened the research loop for our client, empowering their existing data science teams to iterate at a more rapid pace. Through maintaining a continued relationship with our partner, we have provided enhancements to allow for greater experimentation and refinement of the underlying methodology to produce reliable results.