Rise of the internet has created opportunities to specifically target consumers. Counterintuitively, this has resulted in overexposure of users to repetitive ads. In this talk, I will show why this is a result of causality problems using a case study. I will demonstrate how we apply instrumental variables and pymc3 to solve these problems.
First, I will give a short introduction about the problem statement, explaining why analysing historical data results in overexposing users to ads. Secondly, I will give a short introduction to instrumental variables, why the method is so suitable to this specific setup and why it shouldn't be used in general. Afterwards, I will go into a specific case study conducted at Greenhouse where we apply instrumental variables to learn the truth about repetitive ads. I will show why statistical power is a challenge in these applications. I will discuss the options for making (parametric) assumptions to face these issues and will show how Bayesian statistics can help in finding a balance between these assumptions for the different parameters in your model. If there is some time left, I can explain how and why bootstrapping can be applied for constructing posterior distributions of the results.