If Applied Econometrics Were Easy, LLMs Could Do It!

Summary

Can AI do applied econometrics and causal inference? Can LLMs pick up on the nuances and social norms that dictate so many of the decisions made in applied work and reflect them in response to a prompt? LLMs bring to the table incredible capabilities and efficiencies and opportunities to create value. But there are risks when these tools are used like Dunning-Kruger-as-a-Service (DKaaS), where the critical thinking and actual learning begins and ends with prompt engineering and a response. We have to be very careful to recognize as Philip Tetlock describes in his book "Superforecasters" that there is a difference between mimicking and reflecting meaning vs. originating meaning.  To recognize that it’s not just what you know that matters, but how you know what you know. The second-handed tendency to believe that we can or should be outsourcing, nay, sacrificing our thinking to AI in exchange for misleading if not false promises about value, is philosophically and epistemically disturbing.

AI vs. Causal Thinking

This is a good article, from causal lens: Enterprise Decision Making Needs More Than Chatbots. 

"while LLMs are good at learning and extracting information from a corpus, they’re blind to something that humans do really well – which is to measure the impact of one’s decisions." 

In a recent talk Cassie Kozrykov puts it well: "AI does not automate thinking!"

   

Channelling Judea Pearl, understanding what makes a difference (causality)requires more than data, it also requires something not in the data to begin with. So much of the hype around AI is based on a tools and technology mindset. As Captain Jack Sparrow says about ships in Pirates of the Caribbean, a ship is more than sails and rudders, those are things a ship needs. What a ship really is, is freedom. Causal inference is more than methods and theorems, those are things causal inference needs, but what it really is, is a way of thinking. And in business, what is required is an alignment of thinking. For instance, in his article The Importance of Being Causal, Ivor Bojinov describes the Causal Data Analysis Review Committee at LinkedIn. It is a common best practice in learning organizations that leverage experimentation and causal inference. 

If you  attended very many of those reviews you begin to appreciate the amount of careful thinking required to understand the business problem, frame the hypothesis, and translate it to an analytical solution....then interpret the results and make a recommendation about what action to take next. Similarly a typical machine learning workflow requires up front thinking and problem framing. But unlike training an ML model, as Scott Lundberg describes (see my LI Post: Beyond SHAP Values and Crystal Balls), understanding what makes a difference is not just a matter of letting an algo figure out the best predictors  and calling it a day, there is an entire garden of forking paths to navigate and each turn requires more thinking and a vast difference in opinions among 'experts' about which direction to go.

As I discussed in a past post about forking paths in analysis: 

"even if all I am after is a single estimate of a given regression coefficient, multiple testing and researcher degrees of freedom may actually become quite a relevant concern...and this reveals the fragility in a lot of empirical work that prudence would require us to view with a critical eye"

Sure you could probably pair a LLM with statistical software and a data base connection and ask it to run a regression, but getting back to Jack Sparrow's ship analogy, a regression is more than just fitting a line to data and testing for heteroskedasticity and multicollinearity (lets hope if LLMs train on econometrics textbooks they don't weight the value of information by the amount of material dedicated to multicollinearity!!!) and the laundry list of textbook assumptions. AI could probably even describe in words a mechanical interpretation of the results. All of that is really cool, and something like that could save a lot of time and augment our workflows (which is valuable) but we also have to be careful about that tools mindset creeping back on us. All those things that AI may be able to do are only the things regression needs, but to get where we need to go, to understand why, we need way more than what AI can currently provide. We need thinking. So even for a basic regression, depending on our goals, the thinking required is currently and may always be beyond the capabilities of AI.

When we think about these forking paths encountered in applied work, each path can end with a different measure of impact that comes with a number of caveats and tradeoffs to think about. There are seldom standard problems with standard solutions. The course of action taken requires conscious decisions and the meeting of minds among different expert judgements (if not explicitly then implicitly) that considers all the tradeoffs involved in moving from what may be theoretically correct and what is practically feasible. 

In his book, "A Guide to Econometrics" Peter Kennedy states that "Applied econometricians are continually faced with awkward compromises" and offers a great story about what it's like to do applied work: 

"Econometric theory is like an exquisitely balanced French recipe, spelling out precisely with how many turns to mix the sauce, how many carats of spice to add, and for how many milliseconds to bake the mixture at exactly 474 degrees of temperature. But when the statistical cook turns to raw materials, he finds that hearts of cactus fruit are unavailable, so he substitutes chunks of cantaloupe; where the recipe calls for vermicelli he used shredded wheat; and he substitutes green garment die for curry, ping-pong balls for turtles eggs, and for Chalifougnac vintage 1883, a can of turpentine."

What choice would AI driven causal inference make when it has to make the awkward compromise between Chalifougnac vintage 1883 and turpentine and how would it explain the choice it made and the thinking that went into it? How would that choice stack up against the opinions of four other applied econometricians who would have chosen differently? 

As Richard McElreath discusses in his great book Statistical Rethinking:

"Statisticians do not in general exactly agree on how to analyze anything but the simplest of problems. The fact that statistical inference uses mathematics does not imply that there is only one reasonable or useful way to conduct an analysis. Engineers use math as well, but there are many ways to build a bridge." 

This is why in applied economics so much of what we may consider as 'best practices' are as much the result of social norms and practices as they are textbook theory. These norms are often established and evolve informally over time and sometimes adapted to the particulars of circumstances and place unique to a business or decision making environment, or research discipline (this explains the language barriers for instance between economists and epidemiologists and why different language can be used to describe the same thing and the same language can mean different things to different practitioners). A kind of result of human action but not human design, many best practices may seldom be formally codified or published in a way accessible to train a chatbot to read and understand. Would an algorithm be able to understand and relay back this nuance? I gave this a try by asking chatGPT about linear probability models (LPMs), and while I was impressed with some of the detail, I'm not fully convinced at this point based on the answers I got. While it did a great job articulating the pros and cons of LPMs vs logistic regression or other models, I think it would leave the casual reader with the impression that they should be wary of relying on LPMs to estimate treatment effects in most situations. So they miss out on the practical benefits (the 'pros' that come from using LPMs) while avoiding the 'cons' that as Angrist and Pischke might say, are mostly harmless. I would be concerned about more challenging econometric problems with more nuance and more appeal to social norms and practices and thinking that an LLM may not be privy to.

ChatGPT as a Research Assistant

Outside of actually doing applied econometrics and causal inference, I have additional concerns with LLMs and AI when it comes to using them as a tool for research and learning. At first it might seem really great if instead of reading five journal articles you could just have a tool like chatGPT do the hard work for you and summarize them in a fraction of the time! And I agree this kind of summary knowledge is useful, but probably not in the way many users might think. 

I have been thinking a lot about how much you get out of putting your hands on a paper or book and going through it and wrestling with the ideas, the paths leading from from hypotheses to the conclusions, and how the cited references let you retrace the steps of the authors to understand why, either slowly nudging your priors in new directions or reinforcing your existing perspective, and synthesizing these ideas with your own. Then summarizing and applying and communicating this synthesis with others. 

ChatGPT might give the impression that is what it is doing in a fraction of the time you could do it (literally seconds vs. hours or days). However, even if it gave the same summary you could write verbatim the difference couldn't be as far apart as night and day in terms of the value created. There is a big difference between the learning that takes place when you go through this process of integrative complex thinking vs. just reading a summary delivered on a silver platter from chatGPT. I’m skeptical what I’m describing can be outsourced to AI without losing something important. I also think there are real risks and costs involved when these tools are used like Dunning-Kruger-as-a-Service (DKaaS), where the critical thinking and actual learning begins and ends with prompt engineering and a response. 

When it comes to the practical application of this knowledge and thinking and solving new problems it’s not just what you know that matters, but how you know what you know. If all you have is a summary, will you know how to navigate the tradeoffs between what is theoretically correct and what is practically feasible to make the best decision in terms of what forking path to take in an analysis? Knowing about the importance of social norms and practices in doing applied work, and if the discussion above about LPMs is any indication, I'm not sure. And with just the summary, will you be able to quickly assimilate new developments in the field....or will you have to go back to chatGPT. How much knowledge and important nuance is lost with every update? What is missed? Thinking!

As Cassie says in her talk, thinking is about:

"knowing what is worth saying...knowing what is worth doing, we are thinking when we are coming up with ideas, when we are solving problems, when we are being creative"

AI is not capable of doing these things, and believing and even attempting or pretending that we can get these things on a second-handed basis from an AI tool will ultimately erode the real human skills and capabilities essential to real productivity and growth over the long run. If we fail to accept this we will hear a giant sucking sound that is the ROI we thought we were going to get from AI in the short run by attempting to automate what can't be automated. That is the false promise of a tools and technology mindset.

It worries me that this same tools and technology based data science alchemy mindset has moved many managers who were once were sold the snake oil that data scientists could simply spin data into gold with deep learning, will now buy into the snake oil that LLMs will be able to spin data into gold and do it even cheaper and send the thinkers packing! 

Similarly Cassie says: "that may be the biggest problem, that management has not learned how to manage thinking...vs. what you can measure easily....thinking is something you can't force, you can only get in the way of it."

She elaborates a bit more about this in her LinkedIn post: "A misguided view of productivity could mean lost jobs for workers without whom organizations won't be able to thrive in the long run - what a painful mistake for everyone."

Thunking vs. Thinking

I did say that this kind of summary info can be useful. And I agree that the kinds of things that AI and LLMs will be useful for are what Cassie refers to in her talk as 'thunking.'  The things that consume our time and resources but don't require thinking. Having done your homework, the kind of summary information you get from an LLM can help reinforce your thinking and learnings and save time in terms of manually googling or looking up a lot of things you once knew but have forgotten. If there is an area you haven't thought about in a while it can be a great way to help get back up to speed. And when trying to learn new things, it can be leveraged to speed up some aspects of your discovery process or make it more efficient, or even help challenge or vet your thinking (virtually bouncing ideas back and forth). But to be useful, this still requires some background knowledge and should never be a substitute for putting your hands on a paper and doing the required careful and critical thinking.

One area of applied econometrics I have not mentioned is the often less glamorous work it takes to implement a solution. In addition to all the thinking involved in translating the solution and navigating the forking paths, there is a lot of time spent accessing and transforming the data and implementing the estimation that involves coding (note even in the midst of all that thunking work there is still thinking involved - sometimes we learn the most about our business and our problem while attempting to wrangle the data - so this is also a place where we need to be careful about what we automate). Lots of data science folks are also using these tools to speed up some of their programming tasks. I'm a habitual user of stack-exchange and git hub and constantly recycle my own code or others' code. But I burn a lot of time somedays in search of what I need. That's the kind of thunking that it makes since to enlist new AI tools for!

Conclusion: Thinking is Our Responsibility

I've observed two extremes when it comes to opinions about tools like ChatGPT. One is that LLMs have the knowledge and wisdom of Yoda and will solve all of our problems. The other extreme is that because LLMs don't have the knowledge and wisdom of Yoda they are largely irrelevant. Obviously there is middle ground and I am trying to find it in this post. And I think Cassie has found it:

"AI does not automate thinking. It doesn't! There is a lot of strange rumblings about this that sound very odd to me who has been in this space for 2 decades"

I have sensed those same rumblings and it should make us all feel a bit uneasy. She goes on to say:

"when you are not the one making the decision and it looks like the machine is doing it, there is someone who is actually making that decision for you...and I think that we have been complacent and we have allowed our technology to be faceless....how will we hold them accountable....for wisdom...thinking is our responsibility"

Thinking is a moral responsibility. Outsourcing our thinking and fooling ourselves into thinking we can get knowledge and wisdom and judgment second-handed from a summary written by an AI tool, and to believe that is the same thing and provides the same value as what we could produce as thinking humans is a dangerous illusion when ultimately, thinking is the means by which the human race and civil society ultimately thrives and survives. In 2020 former President Barak Obama emphasized the importance of thinking in a democracy: 

"if we do not have the capacity to distinguish what's true from what's false, then by definition the marketplace of ideas doesn't work. And by definition our democracy doesn't work. We are entering into an epistemological crisis." 

The wrong kind of tools and technology mindset, and obsequiousness toward the technology, and a second-handed tendency to believe that we can or should be outsourcing, nay, sacrificing our thinking to AI in exchange for misleading if not false promises about value, is philosophically and epistemically disturbing.

LLMs bring to the table incredible capabilities and efficiencies and opportunities to create value. But we have to be very careful to recognize as Philip Tetlock describes in his book Superforecasters, that there is a difference between mimicking and reflecting meaning vs. originating meaning.  To recognize that it’s not just what you know that matters, but how you know what you know. To repurpose the closing statements from the book Mostly Harmless Econometrics: If applied econometrics were easy, LLMs could do it.

Additional Resources:

Thunking vs Thinking: Whose Job Does AI Automate? Which tasks are on AI’s chopping block? Cassie Kozrykov. https://kozyrkov.medium.com/thunking-vs-thinking-whose-job-does-ai-automate-959e3585877b

Statistics is a Way of Thinking Not a Just a Box of Tools. https://econometricsense.blogspot.com/2020/04/statistics-is-way-of-thinking-not-just.html 

Will There Be a Credibility Revolution in Data Science and AI? https://econometricsense.blogspot.com/2018/03/will-there-be-credibility-revolution-in.html 

Note on updates: An original version of this post was written on July 29 in conjunction with the post On LLMs and LPMs: Does the LL in LLM Stand for Linear Literalism? Shortly after posting I ran across Cassie's talk and updated to incorporate many of the points she made, with the best of intentions. Any  misrepresentation/misappropriation of her views is unintentional.