Net Revenue vs Gross Revenue

[mmiikkoo]

Hello, as the documentation does not mention it anywhere, I want to double check whether the input should use Gross Revenue or Net Revenue (with the costs deducted)?

Also if I use Revenue as the KPI, I suppose the revenue_per_kpi field can be skipped.

Thank you.

[cpulavarthi]

Hello @mmiikkoo,

Thank you for contacting us!

From the modeling perspective, either the gross or the net revenue can be used. However, the causal interpretation would be slightly different.

• Using gross revenue answers "how much incremental gross revenue is obtained due to ad spend?"
• Using net revenue answers "how much incremental net revenue is obtained due to ad spend?"

You should choose the one that better suits your needs. One caveat, however, of using net revenue, is that the ad cost shouldn't be subtracted from the gross revenue. Otherwise the model will lose the causal interpretation (since it will be using ad spend to explain non-ad revenue).

The default value of the revenue_per_kpi argument while loading data is None. If kpi corresponds to revenue, then an array of ones is passed automatically. Therefore, you definitely can skip the revenue_per_kpi field while using Revenue as the target KPI.

Feel free to reach out if you have any further questions.

Thank you

Google Meridian Support Team

[SteliosSake]

Hello @cpulavarthi

Thanks a bunch for your input, since this was a question I had as well!

This leads me to the following question:

• If we are modeling Y as gross revenue in the google meridian model (have not subtracted ad spend or discounts), then is the following definition of variables correct?

Note: The granularity of data points is weekly, the holidays have been added to the models with knots (I passed to the model the knots were each knot is associated with a holiday, so the knot is equal to the index of the data point, each data point is a week in which a holiday is present). For the "media" variable I am using the media's spend (no impressions available) and I am using the national model.

Variable definition:
coord_to_columns = load.CoordToColumns(
time='week_start',
controls=['direct_to_site_demand', 'natural_search_demand', 'unspecified_demand', 'last_month_spend', 'Economy Index'],
kpi='gross_demand',
media=['channel 1', 'channel 2, 'channel 3',...
],
media_spend=['channel 1', 'channel 2, 'channel 3',...
],
organic_media=['social_organic_demand'],
non_media_treatments=['discounts'] (the discounts variable is not in percentage format like 0.2, it is in currency like 1.298.120 euros)
)

loader = load.CsvDataLoader(
csv_path=(path to data),
kpi_type='revenue',
coord_to_columns=coord_to_columns,
media_to_channel=correct_media_to_channel,
media_spend_to_channel=correct_media_spend_to_channel,
)
data = loader.load()

I am asking this question because I did not know how to utilize the discounts variable (control, non-media treatment or even as a media channel to see how effective it is) at the start and whether it is better to utilize gross demand (without subtracting discounts and ad spend) or demand after subtracting discounts. I would really appreciate your input regarding the question I asked at the start of the comment and some feedback regarding the different scenarios for analysis (gross demand with discounts as control, or with discounts as non-media treatment, or any other that you think is important to be discussed).

Thanks in advance for all your help, time and effort!

[cpulavarthi]

Hello @SteliosSake,

Thank you for contacting us!

Your code loading the data into the Meridian InputData object is correct. It is correct to consider the discount as a non-media treatment variable. The difference between controls and non-media treatments is that controls aren't intervenable whereas non-media treatments are. What it means is that the marketing team has a certain level of control on non-media treatments and they can adjust them to optimize their spend allocations and get desired outcomes. For example, competitor sales is a control because we won't have any say in how much sales a competitor may have. On the other hand, discounts, pricing, promotional offers, etc., can all be considered as non-media treatments.

Regarding the format of the discount variable itself, it is fine even if it's not in the percentage format. The data loader will scale the column and the model will learn patterns in that format itself. Your final result analysis will revert all of these internal transformations and display your results in the format and scale of your input data.

I am however unable to understand what you mean by "gross demand" in your query. How are you defining it?

Feel free to reach out if you have any further questions regarding this.

Google Meridian Support Team

[SteliosSake]

Hello @cpulavarthi !!

First of all, I really appreciate your quick reply! I didn't expect it and I thank you for it.

Regarding your answer, reply with the following numbered points:
(I have provided an appendix section with the configurations of each analysis I did at the end of the post so you can have more context for each one).

1. The definition of the KPI in the analysis I presented you already is 'gross demand', because it is defined as: Revenue plus discounts.

2. I am glad that it is correct to consider the discounts variable as a non-media treatment. I did a lot of experimentation with this variable. I considered it a control, non-media treatment and a paid-media channel in different analyses. Below I present you the experimentation scenarios I did and I would like your input or thoughts on them, because I trust your judgement more than my own. In your opinion, is there a "best" analysis out of those? Or you could say that all of them are correct but they just explain different things?

   • Analysis 1 - Discounts as a control: I thought of adding as a control because you could say that it is a confounding variable (affecting both the treatments and the KPI) with the KPI of the analysis being net demand (Revenue minus discounts). I saw that this debiases the causal estimates of the treatments, hence why I chose to consider the discounts as a control. But I don't know whether it is correct in terms of causal inference or how it can be "explained" as part of the baseline. Do you have any thoughts or feedback on that?
   • Analysis 2 - Discounts as non-media treatment: I know that this choice for the discount variable makes sense and I did it with the KPI being the gross demand (Revenue plus discounts). This variable choice (discounts as a non-media treatment) is sound in terms of causal inference and in terms of the explainability of the baseline:
     - Causal Inference: because the KPI is gross demand and contains the discounts, which are now a treatment, if we used net demand as the KPI it would create a problem because the discounts are subtracted from the KPI and hence we are already accounting for a treatment variable.
     - Baseline: it makes sense because the baseline is calculated by setting the discounts to the minimum value, making a sensible scenario of calculating the gross demand that the company can generate on its own with minimum discounts.
   • Analysis 3 - Discounts as a paid media treatment: While analysis 2 helped with understanding the contribution of the discounts on the KPI (gross demand) it did not have a way of suggesting whether the company is doing way too many discounts or way to less. So I treated the discounts as a paid media channel so that it can be added in the optimization of the spending. The results of this analysis indicated that the discounts would account for almost 40% of the incremental gross demand, the baseline being 37% and the rest contributions were appointed to the marketing channels. The total spend for media was 500k, where 200k was the sum of spend for all true media channels and 300k was the considered "media" channel of the discounts. I know that this may not be a great idea but I would like to know your thoughts on this. If you need to execute an analysis like this, how would you go about it?

3. The 'social_organic_demand' variable is a variable that I have considered it t be both a control variable (in analysis 1) and an organic media variable (analysis 2 and 3). The variable itself is in the form of demand, not in the form of impressions, views or anything else. Does this create a problem in the analysis? Is it more correct to be considered a control due to it being demand? or is it okay to be treated as organic media?

Thanks in advance for all your support and your time!

Appendix (configurations for each analysis, all models are national and there are no impression data):

• Analysis 1
  coord_to_columns = load.CoordToColumns(
  time='week_start',
  controls=['social_organic_demand', 'discounts', 'direct_to_site_demand', 'natural_search_demand', 'unspecified_demand', 'last_month_spend', 'Economy Index'],
  kpi='net_demand',
  media=['channel 1', 'channel 2, 'channel 3',...
  ],
  media_spend=['channel 1', 'channel 2, 'channel 3',...
  ]
  )

• Analysis 2
  coord_to_columns = load.CoordToColumns(
  time='week_start',
  controls=['direct_to_site_demand', 'natural_search_demand', 'unspecified_demand', 'last_month_spend', 'Economy Index'],
  kpi='gross_demand',
  media=['channel 1', 'channel 2, 'channel 3',...
  ],
  media_spend=['channel 1', 'channel 2, 'channel 3',...
  ],
  organic_media=['social_organic_demand'],
  non_media_treatments=['discounts'] (the discounts variable is not in percentage format like 0.2, it is in currency like 1.298.120 euros)
  )

• Analysis 3
  coord_to_columns = load.CoordToColumns(
  time='week_start',
  controls=['direct_to_site_demand', 'natural_search_demand', 'unspecified_demand', 'last_month_spend', 'Economy Index'],
  kpi='gross_demand',
  media=['discounts', 'channel 1', 'channel 2, 'channel 3',...
  ],
  media_spend=['discounts', 'channel 1', 'channel 2, 'channel 3',...
  ],
  organic_media=['social_organic_demand']
  )

[SteliosSake]

Hello again @cpulavarthi ,

I know my questions are long but I would really appreciate some assistance with regards to question 2. out of the three I presented. All in all I don't know how to handle the discounts in my analysis. It boils down to:

• If I use discounts as a control (saying that its something that we can't change so not intervenable), then the KPI should be revenue without the discounts?
• If I use discounts as a non-media treatment (or media treatment) then the KPI should be revenue plus the discounts? since it is better in terms of causal inference? since we know that and I quote "One caveat, however, of using net revenue, is that the ad cost shouldn't be subtracted from the gross revenue. Otherwise the model will lose the causal interpretation (since it will be using ad spend to explain non-ad revenue)."

[cpulavarthi]

Hello @SteliosSake,

Thank you for sharing detailed notes of your analyses and your thought process.

We recommend treating discounts as a non-media treatment variable. By definition, a control variable must be non-intervenable. Please check #846 which has a detailed discussion on this.

Do reach out if you have any further questions or suggestions regarding Meridian.

Google Meridian Support Team