A Practical Guide to Instrumental Variables Methods with Heterogeneous Treatment Effects
summary
The gist
Instrumental variables (IV) methods represent one of the most popular approaches to identifying and estimating causal effects in economics The paper offers a nontechnical…
In short
The paper guides researchers in using instrumental variables (IV) methods for causal effects with varying treatment effects. It compares three main estimates: the linear IV regression, a 2SLS estimate, and the 'true' local average treatment effect. The authors recommend focusing on the 'true' LATE as it is often more substantive when treatment effects are heterogeneous.
Key concepts
- Linear IV Regression Estimand
- This is the standard way to estimate causal effects using IV methods, assuming a simple first stage relationship. It provides one type of causal effect but may not capture the full complexity of how treatment affects different groups in the population.
- 2SLS Estimand
- This method estimates a different causal effect, especially when there are multiple instruments or complex interactions in the model. It is useful for saturated specifications but might not align with what researchers ultimately want to measure.
- True Local Average Treatment Effect (LATE)
- This is the authors' preferred estimate. It accounts for heterogeneity by aggregating covariate-specific treatment effects based on how much each group responds to the instrument, making it more intuitive and often more meaningful in real-world applications.
Terminology used across episodes
This episode discusses
- A Practical Guide to Instrumental Variables Methods with Heterogeneous Treatment Effects · Paper Radio
- Testing Instrument Validity with Covariates
- Doubly Robust Estimators with Weak Overlap
- Quantifying the Internal Validity of Weighted Estimands
- Doubly Robust Estimation of Local Average Treatment Effects Using Inverse Probability Weighted Regression Adjustment
- Two-stage least squares with treatment-covariate interactions for treatment effect heterogeneity · Paper Radio
The paper
A Practical Guide to Instrumental Variables Methods with Heterogeneous Treatment Effects · Read on arXiv
Department of Economics, Brandeis University · Department of Economics, University College London · Department of Economics, LMU Munich
Transcript
Introduction to the show: ident: Economics. The day in economics: theory, mechanisms and the evidence beneath them.
Ruth: We're looking at the paper: "A Practical Guide to Instrumental Variables Methods with Heterogeneous Treatment Effects".
Warren: Here's the gist of the paper. Instrumental variables (IV) methods represent one of the most popular approaches to identifying and estimating causal effects in economics The paper offers a nontechnical, practical guide to the literature on IV methods with heterogeneous treatment effects, focusing on areas where existing applied work deviates from best practices motivated by recent theoretical literature Target Parameters The paper discusses differences between three leading estimands in applications of IV methods: the probability limit in the usual linear IV regression, the 2SLS estimand in a saturated specification with multiple interacted instruments, and the “true” local average treatment effect The authors argue that applied researchers should more often estimate this last parameter, ideally as the main object of interest or at least as a robustness check Motivating Example: Stratified RCTs with Imperfect Compliance The paper uses a stratified RCT with imperfect compliance to illustrate the differences between the usual linear IV regression estimand, the 2SLS estimand, and the “true” local average treatment effect The LATE regression is presented as a way to accommodate heterogeneity in both the first stage and the covariate-specific LATEs Estimand Estimation Procedures The paper details three ways to control for covariates in stratified RCTs with imperfect compliance: 1.
Ruth: The linear IV regression, which uses an additively separable first stage 2.
Warren: The 2SLS regression, referred to as the “saturate and weight” approach 3.
Ruth: The LATE regression, which augments the outcome equation with interactions between the instrument and centered covariates Relative Weights in Heterogeneous Treatment Effects The paper shows that when treatment effects are heterogeneous, the three estimands—βIV, βAI, and βLATE—are not the same The LATE regression is described as the most intuitive because it aggregates covariate-specific treatment effects in proportion to each cell’s estimated share among compliers Conversely, βIV and βAI introduce additional terms to their respective weights, complicating interpretation Parametric Misspecification and Machine Learning Solutions The paper addresses the possibility that estimates of βIV and βLATE may be biased due to parametric misspecification, such as omitting important interaction and higher-order terms Modern estimation approaches based on machine learning are recommended as a general solution to this problem Double/debiased machine learning (DML) is presented as an automated approach to constructing “rich” covariate specifications by flexibly approximating the relevant conditional means Assumption Violations and Testing The paper advises practitioners to systematically assess whether the assumptions underlying the LATE framework may be violated through testable implications Instrument validity can be tested using inequalities derived from the standard LATE assumptions, such as those involving moments of observed variables Robustness to monotonicity violations is addressed by suggesting that focusing on βAI may be a useful compromise under weak monotonicity Formal tests for these assumptions are implemented in the R package montest The paper concludes by recommending that practitioners explicitly target the “true” LATE rather than the usual IV estimand, consider flexible functional forms to avoid parametric misspecification, and use formal tools in response to possible assumption violations The authors emphasize that βLATE is arguably of greater substantive interest when it differs from βIV and βAI The paper also highlights the utility of double/debiased machine learning for flexibly approximating conditional means The final recommendation is to use covariate balancing and doubly robust estimators, together with their machine learning counterparts, for greater robustness The authors suggest that the results from double/debiased machine learning are generally consistent across different algorithms The paper also notes that the choice of estimation approach can affect the conclusions, as shown by comparing linear IV and IPW estimates in an application The authors stress that a cautious researcher should investigate whether a similar pattern arises in their specific setting The paper provides several R and Stata packages for implementing these advanced estimation techniques Table 3: Causal Effects of 401(k) on IRA Participation The table presents the causal effects of 401(k) on IRA participation under the original specification, flexible specification, and DML estimates The results show that in both specifications, βˆLATE is close to βˆIV and βˆAI Table 4: DML Estimates of Causal Effects of 401(k) on IRA Participation This table reports the DML estimates of causal effects for both βˆIV and βˆLATE across six different machine learning algorithms The short-stacking estimate, which combines all six algorithms by assigning nonnegative weights that sum to one, is broadly consistent with the individual learner estimates Table 2: Causal Effects of Medicaid on the Number of Outpatient Visits This table illustrates how the different estimands yield nearly identical point estimates and standard errors in a full sample setting However, when restricted to a subsample of three-person households, the estimated LATE is substantially larger than βˆIV and βˆAI The differences between these estimands follow directly from their distinct weighting schemes Table 3: Causal Effects of 401(k) on IRA Participation This table shows the causal effects of 401(k) on IRA participation under the original specification, flexible specification, and DML estimates The RESET p-Value for both specifications is zero, indicating a rejection of the “rich covariates” assumption in both cases Table 4: DML Estimates of Causal Effects of 401(k) on IRA Participation This table reports the DML estimates for βˆIV and βˆLATE across six machine learning algorithms The short-stacking estimate is broadly consistent with the individual learner estimates, suggesting that conclusions are not sensitive to the choice of algorithm Table 3: Causal Effects of 401(k) on IRA Participation This table reports the causal effects for both logit and probit specifications under the original specification, showing that βˆLATE is very similar to βˆIV The RESET p-Value remains zero for both specifications in this flexible specification Table 4: DML Estimates of Causal Effects of 401(k) on IRA Participation The DML estimates for βˆIV and βˆLATE are generally smaller than the initial linear IV estimate but larger than the corresponding LATE estimates The short-stacking estimate is broadly consistent with the individual learner estimates, suggesting that conclusions are not sensitive to the choice of algorithm Table 3: Causal Effects of 401(k) on IRA Participation The original specification shows βˆIV and βˆLATE as statistically significant at the 5% level The flexible specification shows that the RESET p-Value is not significant, suggesting that a more flexible specification can salvage a causal interpretation Table 3: Causal Effects of 401(k) on IRA Participation
Paper summary: Ruth: So we're looking at this paper called "A Practical Guide to Instrumental Variables Methods with Heterogeneous Treatment Effects" and it’s trying to give applied researchers a nontechnical way to handle instrumental variables when treatment effects aren't constant.
Warren: It seems the authors are focusing on three different ways people estimate causal effects using instrumental variables, specifically comparing the usual linear IV regression, something called the two-stage least squares or 2SLS in a saturated specification, and then they introduce what they call the true local average treatment effect <ref:2605.15115#pg2>.
Iris: What this paper claims is that applied researchers should more often be estimating that last one, the true local average treatment effect. They argue it's at least a robustness check for their main object of interest.
Ruth: They use an example of a stratified randomized controlled trial with imperfect compliance to show how these three estimands differ in practice. The LATE regression they present is framed as a way to account for heterogeneity in both the first stage and the specific local average treatment effects for different groups.
Warren: When we look at how they control for covariates in those stratified trials, they detail three approaches: using a linear IV regression with an additively separable first stage, using what they call the saturate and weight approach which is basically 2SLS, or finally the LATE regression which augments the outcome equation with interactions between the instrument and centered covariates <ref:2605.15115#pg2>.
Iris: The paper shows that when treatment effects are actually different across groups, those three things—the beta IV, the beta AI, and that true LATE—are not all the same thing. The LATE regression is described as being more intuitive because it aggregates those specific treatment effects in proportion to each cell's estimated share among compliers.
Ruth: Conversely, they say that the linear IV or 2SLS methods introduce extra terms into their weighting schemes, which makes them harder to interpret when you have these heterogeneous treatment effects <ref:2605.15115#pg2>.
Warren: Then they get into the problem of parametric misspecification, which is when you might be biased if you leave out important interaction and higher-order terms in your model. They suggest that modern estimation approaches based on machine learning are a good general solution for this kind of problem.
Iris: Specifically, they present double/debiased machine learning as an automated way to build richer covariate specifications by flexibly approximating the relevant conditional means.
Ruth: They also spend time addressing assumption violations, suggesting practitioners systematically test whether the assumptions underlying that LATE framework might be broken and how to do those tests. For instance, instrument validity can be checked using inequalities derived from standard LATE assumptions involving moments of observed variables.
Paper summary: Warren: They also discuss robustness to monotonicity violations, which is a controversial assumption in this context, suggesting that focusing on the beta AI might be a useful compromise under weak monotonicity issues. And they implement formal tests for these assumptions in an R package called montest.
Iris: The authors conclude by recommending that practitioners explicitly target the true LATE instead of just the usual IV estimand, and to use flexible functional forms to avoid those parametric misspecification issues. They also stress using tools like covariate balancing and doubly robust estimators, along with their machine learning counterparts, for better robustness.
Ruth: The paper highlights that the results from double/debiased machine learning are generally consistent across different algorithms when they are applied to things like the four hundred one(k) on IRA participation example they use.
Warren: They also show some specific results where the different estimands yield nearly identical point estimates and standard errors in a full sample setting, but they point out that when you restrict it to a smaller subsample, like three-person households, the estimated LATE becomes substantially larger than both the linear IV and AI estimates.
Iris: That difference between those estimands is directly tied to their different weighting schemes, which is what the paper emphasizes. They show that even in logit and probit specifications under their flexible specification, the beta LATE was very similar to the beta IV.
Ruth: The authors suggest that a more flexible specification can sometimes salvage a causal interpretation when you look at those results, like in the four hundred one(k) example where they find that the reset p-value is not significant in both original and flexible specifications.
Warren: And when they run those machine learning estimates for beta IV and beta LATE across six different algorithms, the short-stacking estimate which combines all of them by assigning non-negative weights that sum to one is broadly consistent with what each individual learner estimates.
Iris: It seems like a reasonable conclusion is that the choice of estimation approach can affect your final conclusions, as shown by comparing linear IV and IPW estimates in an application they ran.
Ruth: So, to wrap up this paper "A Practical Guide to Instrumental Variables Methods with Heterogeneous Treatment Effects," the main message is to move beyond just the usual IV estimand and explicitly target that true LATE when you suspect treatment effects are heterogeneous.
Warren: The authors suggest using flexible forms and those double/debiased machine learning estimators for greater robustness, especially when dealing with potential assumption violations.
Iris: It seems the paper provides a solid framework for how to test those underlying assumptions, which is something applied work often skips when it comes to LATE frameworks.
Ruth: It really points toward using these more flexible and robust tools when you're working with real-world data that doesn't fit the textbook linear model.
Conclusion: Ruth: So, this paper is titled "A Practical Guide to Instrumental Variables Methods with Heterogeneous Treatment Effects," and it's really focused on moving past just using one simple way to estimate things with instrumental variables.
Warren: Right, it’s about showing applied researchers how to handle those situations where the treatment effect isn't the same for everyone in the group.
Iris: Essentially, they are comparing three main ways people calculate causal effects with these methods: the usual linear instrumental variable regression, something called two-stage least squares in a complex setup, and this "true" local average treatment effect.
Ruth: The authors strongly suggest that applied researchers should actually be targeting that true local average treatment effect more often. They see it as something they should use as a main focus of their study or at least a very important check on their other numbers.
Warren: They use an example from a randomized controlled trial with imperfect compliance to really show how these three different calculations end up being different in the real world.
Iris: When you look at how they control for outside factors in those trials, they lay out three specific procedures: a linear instrumental variable regression using an additively separable first stage, this two-stage least squares method which is what they call the saturate and weight approach, and finally this LATE regression which adds interactions between the instrument and centered outside factors.
Ruth: They point out that when treatment effects are different across groups, those three different calculations—the beta IV, the beta AI, and that true LATE—aren't going to be the same number. The LATE regression is described as being more intuitive because it puts all those specific group treatment effects into proportion to each cell's estimated share of compliers.
Warren: Conversely, they say that the linear instrumental variable or two-stage least squares methods introduce extra terms into their weighting schemes, which makes them way harder to interpret when you have these heterogeneous treatment effects.
Iris: Then they tackle the issue of parametric misspecification, which is when your estimates for beta IV and beta LATE might get skewed because you leave out important interaction terms or higher-order terms in your model.
Ruth: They recommend modern estimation approaches based on machine learning as a way to handle that problem generally. Specifically, they present double/debiased machine learning as an automated way to build richer sets of outside factors by flexibly approximating the relevant conditional means.
Warren: They also spend time checking if the assumptions behind that LATE framework might be broken and how to test those assumptions systematically through inequalities derived from standard LATE requirements.
Iris: And they address robustness to monotonicity violations, which is a tricky assumption in this area, suggesting focusing on beta AI might be a useful compromise under weak monotonicity issues.
Ruth: The final recommendation is pretty clear: target that true LATE instead of just the usual instrumental variable estimate, use flexible forms to avoid those misspecification problems, and use tools like covariate balancing and double robust estimators with their machine learning counterparts for better toughness.
Warren: They show some specific examples where the different calculations end up being very close in a full sample setting, but when you restrict it to a smaller group, like three person households, the estimated LATE gets substantially larger than both the linear instrumental variable and AI estimates.
Iris: That difference between those calculations comes directly from their different weighting schemes which they highlighted earlier.
Ruth: They also show that even in logit and probit settings under their flexible specification, the beta LATE was very similar to the beta IV estimate, and they suggest a more flexible model can sometimes save a causal interpretation when you look at those results.
Warren: And when they run those machine learning estimates for both beta IV and beta LATE across six different algorithms, the short-stacking estimate which combines all of them is broadly consistent with what each individual learner estimates.
Iris: It looks like the choice of estimation approach can definitely affect your final conclusion, as shown by comparing linear instrumental variable and IPW estimates in an application they ran.
Ruth: So, the paper’s main message is to move past just using the standard instrumental variable estimate and explicitly target that true local average treatment effect when you suspect treatment effects are different for different people.
More episodes
- Daily Summary for 2026-11-19
- 2608.21274-Recommendation Quality and the Concentration of Consumption: Experimental Evidence from Netflix
- 2408.02757-A nonparametric test for diurnal variation in spot correlation processes
- 2306.16393-High-Dimensional Canonical Correlation Analysis
- 2605.17299-Geometric Brownian motion with intermittent entries and exits
- 2607.04753-Fooling Yourself: how narratives shape beliefs
- 2408.14671-Inference on Treatment Effects with Repeatedly Mismeasured High-Dimensional Controls
- 2201.06898-Difference-in-Differences Estimators for Treatments Continuously Distributed at Every Period
- 2501.15307-Influence Functions: A Unified Differentiation-Based Framework
- The day in conversation for 2026-11-19