Hello @aadityachouhan-ah,

Thank you for contacting us!

You've identified that "referrals" likely encompass a mix of organic word-of-mouth, brand awareness, and your incentive program, with only the latter being intervenable.

If your primary interest was understanding the causal effect of the incentive program on your KPI, I would suggest modeling the incentive program (not referrals) as a non-media treatment. In that scenario, paying close attention to the prior (see default prior of contribution_n) and the default value of non_media_baseline_values in ModelSpec would be important. It's also necessary that your incentive program varies across the data (e.g., different incentives across geographies and time) to enable the model to estimate its causal effect. Without such variation, estimation is impossible.

If your primary interest was to improve the inference of the treatment variables, then you would consider referrals as a control variable. The question of whether referrals make sense as a control variable hinges on whether they act as a confounder or a strong predictor, rather than a mediator (see control variables). It would make sense as a control variable if it’s a strong predictor (e.g., the KPI is revenue-based and referrals drive a very large proportion of revenue) or if it’s a confounder (e.g, referrals cause the business to spend more on advertising).

However, if referrals are a mediator variable—meaning they are caused by more advertising spend—then including them as a control variable in your model wouldn't be appropriate. It's plausible that increased advertising could lead to a rise in word-of-mouth and brand awareness, which, in turn, drives referrals.

Ultimately, you'll need to determine which variable type (confounder, predictor, or mediator) referrals are most likely to be, considering the realities of your situation and the dynamics of your referral ecosystem.

Based on what you've shared, my intuition, with limited knowledge of your specific business context, is that referrals might lean towards being a mediator. In this case, you would not include it as a variable in Meridian. Ultimately, you'll need to determine which variable type (confounder, predictor, mediator) referrals are most likely to be, considering realities of your business and the dynamics of your referral ecosystem.

Do reach out for any further questions or suggestions regarding Meridian.

Google Meridian Support Team