In the past articles & videos, we have seen how MMM via FB Robyn can help us answer some of the key questions in the post-ATT world. While we no longer have user-level data, Robyn helps us establish incrementality that is independent of the platform-reported numbers. It gives us an understanding of our effective CPAs and how much budget we should allot across channels to get the maximum output.
We have also seen how we can run the models on R via Robyn to get the outputs; however, there has not been much discussion on what variables we can & should include in order to get better accuracy so we can rely on the models for key business decisions.
Data collection is definitely one of the most important steps in an MMM; furthermore, it can be difficult to collect data in the same format across multiple different sources, making it one of the most time-consuming steps.
There are various factors & variables we need to prepare for before we start running the code on Robyn. Let’s discuss these now:
- Dependent variable: The dependent variable will be the primary KPI/metric that the MMM will measure against. It should be important to your business & can vary across different verticals.
Examples of typical dependent variables:
- Sales/Purchases
- Trials
- Account sign-ups
- Purchase conversion value/ revenue
- Independent variables: The independent variables are the factors that impact the dependent variable. These factors can be broken down into media, non-media marketing, and macroeconomic factors.
Independent variables depend on the type of activity:
2.1. Media activity: To accurately represent media activity, it’s best to use exposure metrics as direct inputs into the model. It should be the most relevant metric since we should not use installs or clicks. Installs or clicks are dependent on other factors as well (for ex. ad relevance, ad quality, app store page content, etc.) & our aim is to collect metrics that reflect “eyeballs” for all the channels.
Examples of typical media activity:
- Spend
- Impressions
2.2. Organic Variables: Robyn allows us to specify organic_vars to model the impact of marketing activities on our customers. The best part is that organic variables are expected to have similar carryover (adstock) and saturating behavior as paid media variables. The transformation techniques we use for paid (such as geometric or Weibull transformation for adstock), are also applied to organic variables.
Examples of typical organic variables:
- Reach / Impressions on blog posts
- Impressions on organic & unpaid social media
- Newsletters
- Push notifications
- App store ranking
2.3. Non-media marketing activity: Non-media marketing activity can be measured using dummy variables to indicate whether or not an offer was given. If running multiple promotions, ensure separate dummy variables for each promotion separately (e.g., free trial for 14 days vs. free shipping).
Examples of typical non-media marketing variables:
- Discounts
- Promotions
2. 4. Macro-economic factors: Macroeconomic factors tend to impact a wide population rather than just a few select individuals. These indicators are closely monitored by governments, businesses, and consumers alike.
Examples of macroeconomic variables:
- Gross domestic product
- Unemployment rates
- Inflation
2.5. Seasonality & Holidays: While seasonality & holidays are automatically captured in Robyn via Prophet which calculates the impact of seasonality and holidays by decomposing the time series data, one can collect the data for seasonality and holidays using external data sources & use them in the dataset.
Now, If you have successfully collected the data required for modeling based on the above pointers, congratulations! You have taken a great first step on your Robyn journey.
If you want a full walkthrough of how to set up a Media Mix Model using Robyn, check out our video mini-series here:
