Week 10: AI as research companion: Instrumental Variables

Identifying Instrumental variables

Published

August 1, 2025

Getting closer to causality by deciding on control variables

In this course, we discuss how AI can be used to support economics research: help design causal identification. Here we focus on instrumental variables.

Recap (10 mins)

What are instrumental variables? What are they good for?
List examples?

IV assumptions

Hint: get AI to explain these 3 assumptions!

Relevance (REL)

\[ Cov(Z_i, D_i \mid X_i) \neq 0 \]

Exclusion (EX)

\[ Y_i(D,1) = Y_i(D,0) \quad \text{for all} \quad D = 0,1 \]

Independence (IND)

\[ (Y_i(1), Y_i(0), D_i(1), D_i(0)) ⊥ Z_i|X_i \]

Prompt construction

Finding IVs using AI is not just about asking the model for a list. That usually gives you textbook examples with no connection to your actual setting. Instead, we can guide the model through a more structured reasoning process, one that mimics how we would think about instruments ourselves.

Prompt Chaining is a good approach to find possible IVs
- Prompt Chaining means breaking down a complex question (like finding an IV) into smaller steps that help the AI reason its way through the problem
- Instead of asking for an instrument directly, we start by mapping out the relationships between variables, then gradually narrow down which ones could work as instruments.
- Goal: 1) avoid generic answers and 2) get the AI to reason more like we would starting from the setting and building up to the econometrics.

The 2-step approach

Before start, note that the prompting approaches and examples below are adapted from and motivated by Sukjin Han (2024) “Mining Causality: AI-Assisted Search for Instrumental Variables

We will use Prompt Chaining to find IVs with two consecutive prompts.

Step 1: Look for IVs that satisfies the verbal description of REL and EX

🤖 Prompt 1 (Search for IVs). You are [agent] who needs to make a [treatment] decision in [scenario]. What are factors that can determine your decision but do not directly affect your [outcome], except through [treatment] (that is, factors that affect your [outcome] only through [treatment])? list [K 0] factors that are quantifiable. Explain the answers.

Which parts of the prompt satisfy REL and EX?

Step 2: Refine by choosing from the IVs in Step 1 that satisfy a verbal description of IND

🤖 Prompt 2 (Refine IVs). You are [agent] in [scenario], as previously described. Among the [K 0] factors listed above, choose [K] factors that are most likely to be unassociated with [confounders], which determine your [outcome]. the chosen factors can still influence your [treatment]. for each chosen factor, explain your reasoning.

Which parts of the prompt satisfy IND?

Why we don’t tell the AI we’re looking for IVs

If we say from the start that we’re looking for instrumental variables, the AI often jumps to textbook examples, things like weather shocks or distance to school, without really thinking through the context. These are generic and not always helpful.

We risk getting shallow suggestions that ignore the specifics of our data or setting.
Some variables might sound like good instruments but don’t actually satisfy the IV assumptions
It also introduces bias from training data: the model might overuse popular examples instead of reasoning from first principles.

So instead of asking directly for IVs, it’s better to guide the model step-by-step through the structure of the problem.

Examples on finding IVs

Let’s try out the prompts below and review the results together. As we go through them, think about the following:

Does it seem like the AI believes we’re just asking a regular question, or does it seem to realize we’re testing something deeper?
Are the IVs it gives us realistic? Which ones make sense, and which ones don’t seem believable?

Important: Before running the prompts, turn off the option that lets the AI remember your chat history.

Example 1: Demand estimation

🤖 Prompt 1 You are a bar owner at the Copacabana Beach. You need to set the price of draft beer. What would be factors that can determine your decision but that do not directly affect the customers’ demand for draft beer, except through the price you set (that is, that affect the demand only through the price of draft beer).

List forty factors that are quantifiable. Explain your answer for each factor and present these in a table.

🤖 Prompt 2 Suppose you are a bar owner at the Copacabana Beach and is setting the price of draft beer on a day of the week. Among the factors listed above, choose all factors are not influenced by market conditions of draft beer or customers’ characteristics that determine demand for draft beer. For each factor chosen, explain your reasoning.

Example 2: Peer effects on microfinance adoption

🤖 Prompt 1 You live in a village in rural India. You recently joined a microfinance program and started using small loans to invest in your household or business. You want to influence others in your village to also join the program. What would be factors (about yourself, your household, or the village) that determine your influence on peers, but that do not directly affect your peers’ decisions to join microfinance, except through your influence? That is, factors that affect your peers’ adoption only through your influence.

List twenty factors that are quantifiable. Explain your answer.

🤖 Prompt 2 Suppose you are a 40-year-old female resident of the village using small loans from the microfinance program. Among the factors you listed above, which ones are not influenced by the same factors (e.g., shared background, preferences, or location) that brought you and your peers into the same social network or neighborhood in the first place? For each factor you choose, explain your reasoning.

Work together: Find an IV for a research question

Form groups of 3-4.
We want to answer the research question: What is the effect of migration on wages in the target country?
Think about context: which country (from $\rightarrow$ to), which industry (from $\rightarrow$ to), which type of people, what period in time, what kind of migration, skill characteristics, etc.
Prepare Prompt 1 and Prompt 2
- Run Prompt 1
  
  Which one would you keep? Collect them in a separate table.
- Run Prompt 2
  
  See which ones AI kept - are they the same? If not, ask why it didn’t include the ones you did/ why it in included the ones you didn’t.

assignment

--- title: "Week 10: AI as research companion: Instrumental Variables " subtitle: "Identifying Instrumental variables" date: "2025-08-01" --- :::::: hero-section ::::: container ::: hero-title Week 10: AI as research companion: Instrumental Variables ::: ::: hero-subtitle Getting closer to causality by deciding on control variables ::: ::::: :::::: In this course, we discuss how AI can be used to support economics research: help design causal identification. Here we focus on instrumental variables. ## **Recap (10 mins)** - What are instrumental variables? What are they good for? - List examples? ### **IV assumptions** Hint: get AI to explain these 3 assumptions! 1. **Relevance (REL)** $$ Cov(Z_i, D_i \mid X_i) \neq 0 $$ 1. **Exclusion (EX)** $$ Y_i(D,1) = Y_i(D,0) \quad \text{for all} \quad D = 0,1 $$ 1. **Independence (IND)** $$ (Y_i(1), Y_i(0), D_i(1), D_i(0)) ⊥ Z_i|X_i $$ ## **Prompt construction** Finding IVs using AI is not just about asking the model for a list. That usually gives you textbook examples with no connection to your actual setting. Instead, we can guide the model through a more structured reasoning process, one that mimics how we would think about instruments ourselves. - Prompt Chaining is a good approach to find possible IVs - Prompt Chaining means breaking down a complex question (like finding an IV) into smaller steps that help the AI reason its way through the problem - Instead of asking for an instrument directly, we start by mapping out the relationships between variables, then gradually narrow down which ones could work as instruments. - Goal: 1) avoid generic answers and 2) get the AI to reason more like we would starting from the setting and building up to the econometrics. ### The 2-step approach *Before start, note that the prompting approaches and examples below are adapted from and motivated by [Sukjin Han (2024) “Mining Causality: AI-Assisted Search for Instrumental Variables](Https://arxiv.org/pdf/2409.14202)* We will use Prompt Chaining to find IVs with two consecutive prompts. **Step 1: Look for IVs that satisfies the verbal description of REL and EX** > 🤖 **Prompt 1 (Search for IVs).** You are [agent] who needs to make a [treatment] decision in [scenario]. What are factors that can determine your decision but do not directly affect your [outcome], except through [treatment] (that is, factors that affect your [outcome] only through [treatment])? list [K 0] factors that are quantifiable. Explain the answers. - Which parts of the prompt satisfy **REL** and **EX**? **Step 2: Refine by choosing from the IVs in Step 1 that satisfy a verbal description of IND** > 🤖 **Prompt 2 (Refine IVs).** You are [agent] in [scenario], as previously described. Among the [K 0] factors listed above, choose [K] factors that are most likely to be unassociated with [confounders], which determine your [outcome]. the chosen factors can still influence your [treatment]. for each chosen factor, explain your reasoning. - Which parts of the prompt satisfy **IND**? ### **Why we don’t tell the AI we’re looking for IVs** If we say from the start that we're looking for instrumental variables, the AI often jumps to textbook examples, things like *weather shocks* or *distance to school*, without really thinking through the context. These are generic and not always helpful. - We risk getting **shallow suggestions** that ignore the specifics of our data or setting. - Some variables might sound like good instruments but **don’t actually satisfy the IV assumptions** - It also introduces **bias from training data:** the model might overuse popular examples instead of reasoning from first principles. So instead of asking directly for IVs, it’s better to guide the model **step-by-step** through the structure of the problem. ## Examples on finding IVs Let’s try out the prompts below and review the results together. As we go through them, think about the following: - Does it seem like the AI believes we’re just asking a regular question, or does it seem to realize we’re testing something deeper? - Are the IVs it gives us realistic? Which ones make sense, and which ones don’t seem believable? **Important:** Before running the prompts, turn off the option that lets the AI remember your chat history. ### **Example 1: Demand estimation** > 🤖 **Prompt 1** You are a bar owner at the Copacabana Beach. You need to set the price of draft beer. What would be factors that can determine your decision but that do not directly affect the customers’ demand for draft beer, except through the price you set (that is, that affect the demand only through the price of draft beer). > > List forty factors that are quantifiable. Explain your answer for each factor and present these in a table. > 🤖 **Prompt 2** Suppose you are a bar owner at the Copacabana Beach and is setting the price of draft beer on a day of the week. Among the factors listed above, choose all factors are not influenced by market conditions of draft beer or customers’ characteristics that determine demand for draft beer. For each factor chosen, explain your reasoning. ### **Example 2: Peer effects on microfinance adoption** > 🤖 **Prompt 1** You live in a village in rural India. You recently joined a microfinance program and started using small loans to invest in your household or business. You want to influence others in your village to also join the program. What would be factors (about yourself, your household, or the village) that determine your influence on peers, but that do not directly affect your peers’ decisions to join microfinance, except through your influence? That is, factors that affect your peers’ adoption only through your influence. > > List twenty factors that are quantifiable. Explain your answer. > 🤖 **Prompt 2** Suppose you are a 40-year-old female resident of the village using small loans from the microfinance program. Among the factors you listed above, which ones are not influenced by the same factors (e.g., shared background, preferences, or location) that brought you and your peers into the same social network or neighborhood in the first place? For each factor you choose, explain your reasoning. ## **Work together: Find an IV for a research question** - Form groups of 3-4. - We want to answer the research question: **What is the effect of migration on wages in the target country?** - Think about context: which country (from $\rightarrow$ to), which industry (from $\rightarrow$ to), which type of people, what period in time, what kind of migration, skill characteristics, etc. - Prepare **Prompt 1** and **Prompt 2** - **Run Prompt 1** Which one would you keep? Collect them in a separate table. - **Run Prompt 2** See which ones AI kept - are they the same? If not, ask why it didn’t include the ones you did/ why it in included the ones you didn’t. ## assignment