Unlock Causal Inference: 3 Obstacles, 8 Insights & 1 Resource
Unlock Causal Inference: 3 Obstacles, 8 Insights & 1 Resource
Why it can be hard to get started and how it differs from other modeling mindsets
Reading time: 3 min
Causal inference can be difficult to get started with. This post is not an introduction to causal inference but lays out obstacles, insights, and a resource for getting started with causal inference.
Obstacles: I found it helpful to figure out why it’s so hard to get started.
Insights: While learning about causal inference, I had many “AHA” moments that I wanted to share with you.
Resource: My favorite (free) resource on causal inference.
Why Causality Is Hard To Get Into
I’ve had multiple attempts to break into causal inference. Each time, I’ve learned a bit more, but it always feels like a grind. Here are some obstacles that make it hard to get into causal inference:
There are many different approaches to causal inference, like the potential outcomes framework, structural causal models, natural experiments, and difference-in-differences. These different approaches often come from different communities, have different notations, and sometimes even different terms. Confusing.
Learning causal inference is not like adding another model to your toolbox. It’s also not just an “add-on” to a regression model. It’s a unique mindset of modeling (see also my book Modeling Mindsets) that comes with a different understanding of what a model is.
Causal inference is neglected in education about statistics and machine learning. It’s getting better, but in my Bachelor's + Master's in statistics, there was almost no causal inference. Given that causal inference is not just an add-on it can be hard to update your own idea of modeling.
The most coherent body of work is, in my opinion, the work from statistician Judea Pearl. This includes do-calculus, structural causal models, and a strong emphasis on directed acyclic graphs. However, I found the “Book Of Why” difficult to learn from. Many like his writings, so just find out what resources connect best with you.
Insights To Adjust Your Mindset
These insights helped to better understand causal inference:
#1 Start learning about directed acyclic graphs (DAGs) and the implications of blocking the "flow" between variables. High return on investment when getting started with causal inference.
#2 Causal inference means starting from a causal model; causality isn't added to an estimated model as an afterthought.
#3 You can't always estimate a causal effect from observational data.
#4 If a causal effect can be estimated, it might be done with a regression or classification model, but that's not always possible.
#5 The estimation model used for the causal effect of a covariate Xj doesn't automatically work for covariate Xk, even if Xk appears in the model. You might need a separate model for the causal effect of Xk.
#6 Depending on which other covariates you adjust for, you might measure the direct or total effect of the covariate of interest.
#7 An important part of causal inference is to identify which covariates have to be in your model and which may not.
#8 The causal relationships between variables can, to some degree, be detected through conditional (in)dependencies. But, usually, ambiguity about e.g. direction of causality remains.
A Free Resource To Get Started With
Long lists of resources can be useful, but I usually prefer to just get one strong recommendation.
For getting started my recommendation is the course “Introduction to Causal Inference” by Brady Neal.
Why this course?
It’s free
Super clean slides
Multiple formats: videos, slides, and in-progress book
Good overview of the many different causal inference approaches
For the record, I’m not getting any money to promote the course. It’s just a great resource.