How the Model currently works

The objective of the model is to create response curves, which are important outputs of the Marketing Mix Modeling (MMM). Each response curve estimates the relationship between the promotional tactics and business performance. In the case of Dupixent, business performance is measured with a few key metrics: 

• NBRx for new to brand prescriptions fulfilled,
• CBRx for continuation on brand, i.e. refills of prescriptions filled,
• TRx for total prescriptions filled.

The model is trained by fitting the historical data per week on NBRx and CBRx (dependent variables) with independent variables: promotional tactics and other variables, such as brand awareness and seasonality (e.g. Labor Day, Christmas), that constitute the baseline. Promotional tactics include sales force detailing, samples, TV GRPs, and other advertising and promotion to patients or physicians

Due to the small amount of data available for training (a few hundred points), the model uses a Bayesian approach that has proven robust for this type of marketing mix analysis. Based on the data (dependent and independent variables cited above) as all machine learning approaches, the Bayesian approach allows to also integrate business hypotheses into the model, thus giving a more accurate modeling (since more information is provided). Business hypothesis notably include thresholds (minimum number of an action to get a result) and saturation points (number of actions above which no additional effect is obtained).

After the model is trained, we are then able to understand the contribution of each promotional tactic of interest and extrapolate its impact to NBRx and CBRx at different levels of spend, thus creating the response curves. These response curves are then used as inputs in scenarios for optimizing NBRx, CBRx, TRx, Net Revenue, or spend.