Encoding Multi-Channel Lift Tests for PyMC MMM Calibration

1

Hey,
I’m wondering what the best way of encoding a multi-channel lift test to calibrate a model. e.g. one experiment evaluating Facebook, Google, Snapchat etc.

  • Does PyMC have under the hood functionality to manage this a bit like Robyn does? e.g. In Robyn, you encode the test details along with (‘Facebook’ + ‘‘Google’ + ‘Snapchat’) and I believe it apportions out the result based on spend.
  • Is it better to encode them as three individual channels with outcomes divided out based on a rule?

Thank you!

2

cc @cetagostini @juanitorduz

3

We don’t have a native way to support this, nevertheless a simple way to make it will be create a custom deterministic with the sum of channels index A,B,C over time, and force this variable with potential to certain value.

Something like this could work:

period_mask = slice(99, 120) # select days between 100 and 120
with mmm.model:
  pm.Deterministic("variable_sum_of_g_s_f", mmm.model["channel_contribution"][:, 1:4].sum(axis=-1)) # assuming channel contribution shape date, channel and channels 1,2,3 are facebook google snapchat.
  tol = 1e-8
  pm.Potential(
       "hard_constraint",
       pm.math.switch(
           pm.math.all(pm.math.abs(mmm.model["variable_sum_of_g_s_f"][period_mask] - target_values) < tol),
           0.0,
           -np.inf,
       )
  )

Once is done, you can run your model :slight_smile:

2 Likes