Causal Inference Workshop and Causal AI
This will be two posts — one about my workshop on causal inference that starts Saturday September 21st at Mixtape Sessions, and another about Causal Artificial Intelligence. I want to be clear that my workshop isn’t about “Causal AI”, except that my workshop is about causal inference and covers a primer on causal graphs which is usually the framework these firms trying to build something integrating AI with causal inference style reasoning and estimation are using. But mainly I thought I’d just share some about it as it’s an area I’ve been wondering where it’s going as one of the main firms just made a large partnership with Google.
But the other reason is to tell you about the upcoming changes to my causal inference workshop that in an unrelated sense will have some AI. Namely, a chatbot and a transcription of the workshop. But you can read that below.
Causal Artificial Intelligence
Look close in that curve and you’ll see something about eight spots from the bottom called Causal AI, or “Causal Artificial Intelligence”. What is that? Causal AI is thought to be a possible frontier in AI in which causal inference works more closely with AI to make reasonable inference about cause and effect in non-experimental data. The approach, at this moment, tends to draw heavily on Judea Pearl’s graphical approach to causal inference, which as many readers know represents causal relationships as structural elements of a graph linking variables that are thought to have an a priori causal relationship.
This area is usually trying to engage with traditional machine learning data scientists by simultaneously helping people more broadly understand that quantitative prediction algorithms are not themselves causal, but that there is a scientific paradigm that can be. As causal AI is more often applications in industry, not academia, those in academia may feel further away from it than they truly are. But take this example in the context of a telecommunications company facing rising customer churn. A causal AI model will either try to example the effects of some known intervention, or increasingly, try to work backwards from “effects” to causes that are driving the departures. This is often done using a graph framework, each of which can imply independence between variables under different conditioning strategies. These conditioning strategies can be worked out trivially using free browser based software like daggity and more advanced deductions can be found in academic software like Causal Fusion. These are starting points, not ending points, but where it is going does seem to be wrapping causal inference and AI together, and most often by building on Pearl’s work.
There are a few academic review articles you can read on this if you want to get started. For instance, a 2023 Econometrics Journal article by Paul Hünermund and Elias Bareinboim entitled “Causal Inference and Data Fusion in Econometrics”, both formal and informal students of Pearl, or this PNAS by Elias and Pearl.
Causal AI Firms
causaLens is the leading firm at the moment working on this intersection of causal inference and AI technologies. They’ve raised $45 million in a previous funding round, and in addition to their work with clients, they have been focused on publicizing the importance of causal inference to industry and the public at large with several conferences on causal inference aimed at a broad audience. They have had Guido Imbens as keynote and you can find his speech here. They are headquartered in London and have around 65 employees. You can find their information on LinkedIn here.
Two days ago, they announced that they had a groundbreaking collaboration with Google Cloud to offer a new grounding service for quantitative data. The partnership between them represents a significant step in integrating causal reasoning with cutting-edge cloud computing and generative AI technologies. The collaboration will integrate CausaLens AI applications with Google Cloud’s Vertex AI and generative AI models. So generative AI, cutting edge cloud computing, and causal inference — the hope is that the partnership will aid businesses in modeling cause-and-effect relationships within their own nonexperimental datasets and ultimately improve decision making and profitability.
Another company is Geminos Software where I am an advisor. So it is a growing space, and it’s kind of interesting to watch it grow and see where it may be going, as it seems like the broader restructuring taking place around artificial intelligence integration into production processes may be sucking causal inference in too. Only time will tell whether these big society wide bets pay off, but they are definitely happening.
Causal Inference I at Mixtape Sessions Workshops
Now, moving along, let me tell you about my workshop. Starting this Saturday, September 21st, I will be starting up my “Causal Inference 1” workshop on my Mixtape Sessions platform.
It’s a four day workshop, lasting from 9am to 5pm CST with 15 minute breaks on the hour, and a 1 hour lunch break. Guests will use a Discord channel to converse with one another so that they can meet each other as well as have long conversations that are recorded so you can return to them later (unlike Zoom chat which disappears). I also am going to do something new this fall that I’ve not done before. I’m going to use a product called otter.ai which will attend the workshop too. Let me tell you about that first, and what to expect from it, before I then describe the itinerary of the workshop, and more or less manage your expectations.
What is Otter.ai?
Otter.ai is an AI-powered transcription service that changes and enhances how we can interact with spoken content (wikipedia explanation). Among the things it has is a Zoom sync option which I will use to help make the causal inference workshop next week more accessible multiple dimensions. Here’s what you can expect.
Transcription. For all three of my upcoming causal inference workshops (Causal 1, Causal 2 and Causal 3) this fall, Otter.ai will attend too and transcribe in real time the workshop. You will have access both during and after. And not only will you get a full transcript linked to the video, but Otter also generates useful summary of the day, including to-do lists or action points, once it is over. Here’s a little more of an explanation.
Chatbot. In addition to a machine learning transcription service, Otter.ai also has a small large language model chatbot that you can use to ask questions about the workshop both live (in your own window without having to ask questions to me when you don’t want to or post them necessarily to Discord). This chatbot is called Otter Chat and here’s a FAQ about it.
Flexible Attendance. My hope is that otter.ai will enhance guests accessibility of the content along multiple dimensions. They’ll be able to link the transcript with the video giving them two mediums to use to interact with me. They’ll get a chatbot that they can query during the workshop or afterwards. While the chatbot isn’t me, it’s also probably a good tutor for when you can’t reach me or Kyle Butts, and it may provide a little assistance with people who can’t quite remember something that was covered earlier, or maybe just don’t yet feel like they’ve found their wings. It gives them more of a pace with which they feel comfortable interacting with me and the content in a way that I hope complements the workshop. But it also provides some flexibility much needed for people in other parts of the world where 9am to 5pm central time is more like 4pm to midnight or worse their time. And I’m hoping it helps people for whom English is not their first language.
What about recordings?
You’ll still get the recordings. We will continue to store them on Vimeo for you, password protected, so that you can have those. But you’ll also get a transcript to help you after the fact. But just remember — Mixtape Sessions is not a shop. You can’t buy old recordings. We don’t “sell” recordings at all. Recordings belong to those who are registered for a workshop. It doesn’t require that you attend, because people are all over the world and it’s quite literally impossible for some to make it, but it does mean that the recordings belong to registered guests. And with otter.ai, you’ll still get your recordings — you just also get the otter.ai stuff.
Itinerary
There’s four days to the workshop. September 21st, Saturday, is 9am to 5pm CST. Then it goes Sunday the 22nd, Saturday 28th, and Sunday 29th. All the same times. Here’s what you can expect.
Day 1: Potential outcomes and design approach, randomization and selection bias, treatment effect heterogeneity, directed acyclical graphs, applications
Day 2: Unconfoundedness, including matching, weighting and regression
Day 3: Instrument variables
Day 4: Regression discontinuity design
Examples switch between Stata and R, but there are also python examples at my online book if you want to see that too.
Pricing
Same pricing as before: $595 is the sticker price with discounts for people who are on higher teaching loads or not tenure track ($95), students, predocs, RAs, postdocs, residents of middle income countries and folks in between jobs ($50), and residents of low income countries plus India ($1). You just need to email me at causalinf@mixtape.consulting and I can sort you into the correct pricing tier depending on your current status. Note the pricing is not based on citizenship, but rather, residence.
So that’s it. Tell your friends and hit me up if you think you qualify for you a discount, but I think this is a good opportunity, whether you’re a student, undergrad or grad, academic or non-academic, interested in causal inference with or without AI. I think with all the resources you get from coming, it’ll help. So come check it out!