`media_prior` with impressions only (unknown spend)

[PhilippeMoussalli]

Hello,

First of all, thank you for open-sourcing such an amazing library !

I have some question regarding the model that I hope you can help me with :)

I'm considering running lightweights mmm on a dataset that only comprises of different channels with their impressions (no channel costs included). My end goal is to use get an estimate on the the media effects of each channel (I don't need to generated the ROI estimates).

My question is, would it be possible to conduct this sort of analysis without including the spend for each channel? I only have impressions with no costs but I know that spend is an obligatory prior in mmm.fit (media_prior).

In theory this should be possible since the three different marketing approach (ad-stock, hill_adstock and carryover) can be modelled based on spend or impressions as mentioned in the docs

Thank you in advance for your response.

[pabloduque0]

Hello @PhilippeMoussalli !

If you do not need ROI calculations you can only use impressions with no cost related data.

For what to pass the as the media_prior:

• Just pass an array of ones (giving the same prior to all channels). Ones or any value as you might need to adjust this, but as long as they all have the same values its okay.
• Pass manual priors based on your believes of what channels perform best (you might want to still scale this values using a scaler like we do with the costs in the example notebooks).
• Create some heuristic for generating a prior: Eg. If some of your channels are turned off (meaning you have 0/Nan impressions for that column) every now and then you can use a % of time on as a prior. With this you are really saying that you expect the channels that are 100% of time on to perform better than the ones that are only on 50% of the time (and therefore the reason behind why they are off some time). Might still want to scale this one as well. This is just an example but any other heuristic with a different logic of your choosing could work as well.

The only caveat is that some other aspects might be a bit tricky like for example the media optimisation, you will need some price estimation in order to run that. But if your main interest is media contribution/effect, you should be fine without costs/spend/prices.

Hope that helps!

[PhilippeMoussalli]

Thanks for the quick response @pabloduque0!

Indeed all those suggestions seem to be valuable. I will most likely pass in manual priors since I have an overall estimate on the effectiveness of each channel.

Maybe a follow up question based on that approach:

• Is is then plausible to generate interpret the response curves in terms of impressions as opposed to spend?

[pabloduque0]

Yes you can generate the response curve normally and either provide no prices or prices of 1. Although labels will say spend, in your case it will be impressions.