Code for the choosing your randomization unit post - https://ianwhitestone.work/choosing-randomization-unit/

simulations.ipynb

To anyone reading this who came from the blog post, there's two things in this code that I didn't actually talk about (because they'll be leveraged in a future post and I was too lazy to remove them):

1. In the run_simulation function I have these arguments:

    conversion_dependent_on_spu=True,
    # x is conversion rate and y is sessions per user (spu)
    conversion_spu_func=lambda x,y: x*2/3 + x*1/3*np.exp(-0.4*y)

This lets me specify a decay function that maps the relationship between sessions per user and conversion rate - i.e. something like this:

2. All the user level metrics. These will be used in conjunction with ☝️ to explore if you get an increased false positives when doing user level randomization + session level metrics.