sample_posterior_predictive doesn't scale X data

[kylejcaron]

In the docs example for MMMs, the line of code to sample the posterior predictive is

mmm.sample_posterior_predictive(X, extend_idata=True, combined=True)

When X is on the scale of dollars, the data wont get automatically scaled from mmm.sample_posterior_predictive, so it causes the results to be off. The example doesn't run into this problem because channel spend is simulated as X ~ U(0,1). I personally got pretty confused adapting the code. to a real dataset and seeing the posterior predictive plot look so off and had to dig through the codebase to figure out what was going on.

I think the following is a more general approach that would lead to less confusion for now in the docs

mmm.sample_posterior_predictive(mmm.preprocessed_data['X'], extend_idata=True, combined=True)

An alternative to changing the docs could be to scale the X data in the mmm.sample_posterior_predictive method

[williambdean]

Yes, I am addressing this in PR #444 to scale the new data based on the data the model was fit.

The Fourier features and control channel transformations will be handled appropriately after that PR as well

That PR will be merged in after the custom prior and time varying parameters get added. So there are a few items scheduled before this will be fixed

[geirrlod]

Hi,

Also find this a bit confusing as well when I work on real non scaled spend. Do you think PR444 will be pushed out in the next couple of weeks, or should we try to find a way around the problem?

[kylejcaron]

@geirrlod I made this quick change to get it working for me for now

mmm.sample_posterior_predictive(mmm.preprocessed_data['X'], extend_idata=True, combined=True)

also there's this event tomorrow from pymc-labs that you might be interested in here, it sounds like time varying covariates is a main bottleneck for this update, maybe they'll discuss that tomorrow?

edit: AHHH sorry I accidentally hit close issue - just reopened!