Thread by @AleksanderMolak

😯 The greatest fear that drives people away from causality is not about complexity

It's about something much simpler

🧵 (1/n)

#causality #causalinference #python #machinelearning #datascience #casualtwitter

If you ask people who have some knowledge about causality, but they never worked in the field, you can often hear that using causal methods in the real world is "too risky".

🧵 (2/n)

The number one reason for this are the dreaded "causal assumptions".

🧵 (3/n)

The number one dreaded assumption is "no hidden confounding". There's a popular mantra about this assumption: "it's not verifiable". This is sometimes true. Most people take this mantra and conclude that "therefore we cannot use causality in the real world".

🧵 (4/n)

This conclusion does not seem to be correct. There are many ways to use causal methods to your benefit and there are many successful stories across industries that prove it.

🧵 (5/n)

The main reason why many of us draw wrong conclusions about causality is not flawed logic. It's an error in the assmptns. Many of us assume that causality is "binary". Either all assmptns are met all the time & we can use the mthds or not. What if we change the prspctve?
🧵 (6/n)

Today, I want to share with you a method that can help us draw useful conclusions from causal analyses, even if hidden confounding cannot be excluded.

🧵 (7/n)

Sensitivity analysis helps us answer a set of questions regarding the stability of a causal effect at hand.

Historically, there were various approaches to sensitivity analysis, but recent 3-4 years brought a significant progress in the field.

🧵 (8/n)

In their 2020 paper (lnkd.in/dYe4k6pN), @analisereal and C. Hazlett from UCLA proposed a flexible method** that does not require assumptions on the functional form of the treatment assignment nor on the distribution of the unobserved confounders

🧵 (9/n)

It also allows to easily calculate sensitivity parameters using standard regression

The best thing? The method is available as a Python package (also available in R and Stata)

🧵 (10/n)

⭕ PySensemakr (lnkd.in/dRSr3wNq) offers a comprehensive set of tools for sensitivity analysis, including stat and viz tools for basic & adv usage. Both - the docs and the repo (lnkd.in/dfFXS2X5) - offer basic tutorials to make you up & running in no time!
🧵 (11/n)

Not all scenarios can benefit from sensitivity analysis, but a broad class can and these are the ones where the fear of hidden confounding can hide meaningful benefits from us.

I am curious - what are your thoughts?

🧵 (12/n)

**This research line has been recently extended. Check nber.org/papers/w30302 for more details

🧵 (13/n)

💡 Go to causalpython.io and subscribe to get weekly curated causality content and updates on my upcoming causal book, where we discuss sensitivity analysis in a greater detail

🧵 (14/14)

🦩 Follow for more!