Model differences between Sales+"Revenue-per-kpi", vs just Revenue

[Jon77Ruler]

Hi,

I was wondering what the differences are in terms of modelling when I have both sales units and revenue, vs. just revenue?

If I set a non-revenue KPI pointing at sales units, then supply the actual revenue-per-KPI in each time unit derived from revenue/sales, does that allow the model better opportunity to fit vs. say setting a revenue KPI and not providing the sales-units data?

It feels like having both would allow the model to better understand response and the variation in revenue from sale in each time period... vs. just having revenue data where that variance isn't known.

I generally get good models from each, I just wondered which is preferable where the choice is available? Clearly the former gives more accurate CPIK values for one thing!

[cpulavarthi]

Hello @Jon77Ruler,

Thank you for contacting us!

You are right that there is a distinction between using revenue as the KPI vs. using a non-revenue KPI with revenue_per_kpi data. Meridian estimates the causal effect of media on revenue in both cases, but the KPI is used as the model response variable.

• If you use a non-revenue KPI and provide revenue per KPI data, then it is assumed that the revenue per KPI is constant within each geo and time period and that each incremental KPI unit generates the same amount of revenue.

• If you use revenue as the KPI, then no such assumption is not needed.

The assumption of constant revenue per KPI is a strong one, but there might be cases where it is preferred. For example, the relationship between price and sales units is generally monotonic (though not necessarily linear), whereas the relationship between price and revenue may be non-monotonic. If price level fluctuates considerably, then it may be more convenient from a modeling standpoint to use sales units as the KPI. (That said, another option is to use revenue as the KPI and convert "price" into a categorical variable.)

The choice comes down to which model structure you prefer. If in doubt, the more straightforward option is to use revenue as the KPI.

Do reach out if you have any further queries regarding this.

Google Meridian Support Team

[jmp1989]

If you use a non-revenue KPI and provide revenue per KPI data, then it is assumed that the revenue per KPI is constant within each geo and time period and that each incremental KPI unit generates the same amount of revenue.

Can you dumb this down for me or restate it for me? My real world scenario is:

• I have transactions and I have revenue per transaction.
• The revenue per transaction is generally $20 throughout the year although it becomes ~$40 in November and December.
• Historically we run 2 additional media channels during this time. We know with certainty that these channels are average at incremental transactions.
• What I am attempting to get across - I'm concerned that these 2 media channels while being average come across as being 'extremely good' just due to the time of the year in which they are run due where sales are 2x. Is this handled properly or am I overthinking this?

[cpulavarthi]

Hello @jmp1989,

Thank you for reaching out!

If you use a non-revenue KPI and provide revenue per KPI data, then it is assumed that the revenue per KPI is constant within each geo and time period and that each incremental KPI unit generates the same amount of revenue.

The statement is saying that it is assumed that each KPI unit generates a constant amount of revenue. This constant can vary by geo and/or time period, so for example, it could be $40 in Nov-Dec and $20 in other months. This is important because regression model directly estimates the incremental KPI units of each channel, which must be converted to revenue units to report an ROI. The assumption is also critical for response curves, which translate the incremental KPI units to revenue at various channel spend levels.

It might make sense to consider using revenue as the KPI in this case. Think about one of your other marketing channels besides the two mentioned. Suppose you ran a campaign on this channel in November, and then ran the same campaign again in May. If you think these campaigns would generate roughly the same incremental revenue, then revenue might be the better KPI unit. If you think these campaigns would generate roughly the same incremental transactions, then transactions might be better.

Regardless of which KPI you use, it might make sense to control for the revenue per transaction if this can be thought of as the "price" paid by the customer. The "price" likely has a significant effect on the KPI volume, and it is highly correlated with at least two of your campaigns. The KPI may not be linear in price, so you can treat it as a categorical variable using a binary indicator (high price vs. low price) or one-hot encoding.

Other considerations:

• You may want to increase the number of knots during periods of KPI changes that are seasonal and not necessarily attributable to media execution.
• Setting custom priors that reflect your domain knowledge and beliefs about these channels can help capture the causal impact more accurately.

Given below are some documentation links that can help you on setting knots and priors:

• ROI Priors and Calibration
• Set custom ROI priors using past experiments
• Set custom priors when outcome is not revenue
• Set knots

Do contact us if you have any further questions or suggestions for us!

Google Meridian Support Team