(B) I had the great privilege, while being a student at UCLA, to have Professor Judea Pearl as a professor. A few years after I graduated from UCLA, Professor Pearl started a new field: causal inference. Simply said, causal inference attempts to establish the relationships between causes “x” and effects “y”:

• Would lowering price increase revenues?
• Did PR coverage drive sign-ups?
• Does a mobile app improve retention?

“x” is generally a product, a feature or an initiative – while “y” is generally a KPI.

Knowing the relationships between “x” and “y” enables to generate actionable insights in order to make the right courses of action to move from “x” to “y”.

The first application of causal inference for businesses is usually to be an alternative to A/B testing, when A/B tests are too costly or can hurt the experience of end-users

While A/B testing aims to establish any difference between two randomized groups “A” and “B”, and the causal effect of changing “x” on the outcome “y” on group “A” versus group “B”, causal inference techniques attempt to discover and mitigate the cofounding factors “c” (called also omitted variable bias in statistics) that affect the input of interests “x” with the outcome of interest “y”.

A simple example of a cofounder is “age” when researching for instance if “a customer live chat increases product sales” as young users are more likely to chat online than seniors.

Present causal inference techniques include:

• Controlled/fixed effect regression: Add the cofounder to the equation of the regression
• Regression discontinuity design: Focus in on a cut-off point that, within some narrow range, can be thought of as a local randomized experiment

• Difference-in-differences: Comparison of pre and post outcomes between treatment and control groups

• Instrumental variables: Use an instrument variable z to mitigate the effect of the omitted variable c – first z affects x – the second z affects y only through its effect on x

A few technology companies such as Coursera and Airbnb have documented some of their causal inference applications.

Coursera has in particular used causal inference for the following use cases:

• Control regression: effect of product quality on usage
• Regression discontinuity design: effect of adding subtitles to a course
• Difference-in-differences: effect of lowering the price on revenue
• Instrumental variables: having friends on the platform leading to platform retention

Airbnb has used causal inference for the following use case:

• Difference-in-differences: experimentation & measurement for search engine optimization

A recent development of causal inference for technology companies is the use of machine learning techniques for causal inference. Machine learning models can be leveraged to control many potential cofounders, for predicting and selecting those cofounders, and/or reducing the number of cofounders to a set that matters.
Present techniques to that end include double selection (based on Lasso regression model), and double debiased (generic ML model).

Following is video from Professor Pearl, about “The Mathematics of Causal Inference: with Reflections on Machine Learning”:

References

• Causal inference in statistics: an overview
• Causal inference @ Coursera:
  • Controlled Regression: Quantifying the Impact of Course Quality on Learner Retention
  • Instrumental Variables & Randomized Encouragement Trials: Driving Engagement of Learners
• Causal inference @ Airbnb:
  • Experimentation & Measurement for Search Engine Optimization
• Machine learning and causal inference:
  • Double/Debiased Machine Learning for Treatment and Causal Parameters

Note: The picture above are a few friends from Tahiti.

Copyright © 2005-2020 by Serge-Paul Carrasco. All rights reserved.
Contact Us: asvinsider at gmail dot com.