Double machine learning for causal inference in a multivariate sample selection model

summary

The gist

Double machine learning (DML) estimators of the average treatment effect (ATE) and the average treatment effect on the treated (ATT) in a multivariate sample selection model with multiple ordinal…

In short

The authors propose Double Machine Learning (DML) estimators for Average Treatment Effect (ATE) and Average Treatment Effect on the Treated (ATT) within a complex multivariate sample selection model involving multiple ordinal selection equations. The analysis confirms that these DML estimators effectively mitigate significant biases that arise when multivariate selection is ignored, providing more accurate causal parameter estimates.

Key concepts

Multivariate Sample Selection Model (MSSM)
This model describes a situation where the outcome you are interested in can only be observed if certain conditions related to multiple selection variables are met. It involves several 'selection equations' that determine whether an observation is included in the final sample, making standard causal inference difficult.
Neyman-orthogonal scores
These are specific mathematical scores derived from the data, based on an observation vector including treatment and selection variables. They are designed to be useful for estimating causal parameters like ATE and ATT by being orthogonal (uncorrelated) to nuisance functions, which helps isolate the causal effect.
Double Machine Learning (DML)
This is a statistical technique that uses machine learning methods twice. The process involves partitioning the data and using regression to estimate various nuisance functions. By combining these estimates in a specific way, DML estimators are proposed to provide robust and unbiased estimates of the causal parameters.
Average Treatment Effect (ATE) and ATT
ATE measures the average effect of a treatment across all individuals in the population. ATT measures the average effect of a treatment specifically on those individuals who actually received it. The paper focuses on estimating these two key causal quantities within their selection model.

Terminology used across episodes

This episode discusses

The paper

Double machine learning for causal inference in a multivariate sample selection model · Read on arXiv

Sofiia Dolgikh sidolgikh, Bodan Potanin bogdanpotanin

HSE University

We propose double machine learning (DML) estimators of the average treatment effect (ATE) and the average treatment effect on the treated (ATT) in the multivariate sample selection model with multiple ordinal selection equations. We establish asymptotic normality of the proposed estimators and study finite sample properties on simulated data. Specifically, the results of the analysis suggest that without addressing multivariate sample selection, the estimates of the causal parameters may be highly biased. However, the proposed estimators allow us to avoid these biases. Moreover, we apply the estimators to study the effect of female education on the log-wage ratio of spouses.

Transcript

Introduction to the show: ident: Economics. The day in economics: theory, mechanisms and the evidence beneath them.

Ruth: I'm Ruth.

Warren: And I'm Warren, here with Iris.

Ruth: We're looking at the paper: "Double machine learning for causal inference in a multivariate sample selection model".

Warren: Here's the gist of the paper. Double machine learning (DML) estimators of the average treatment effect (ATE) and the average treatment effect on the treated (ATT) in a multivariate sample selection model with multiple ordinal selection equations are proposed, establishing asymptotic normality and showing that addressing multivariate sample selection prevents highly biased causal parameter estimates The results of the analysis suggest that without addressing multivariate sample selection, the estimates of the causal parameters may be highly biased.

Ruth: Model Specification and Identification The paper introduces a multivariate sample selection model (MSSM) with multiple ordinal selection equations where an outcome is observable only under specific conditions The main notations used throughout the paper are provided below and are similar to those used by Potanin (2026).

Warren: The observation of the outcome is conditional on a vector of selection variables, denoted as Z̃i = 1, where the selection criterion is as follows Z̃i = I (Zi = z∗) = 1, if Zi = z∗0, otherwise, (1).

Ruth: Identification of ATE and ATT in this MSSM requires several assumptions, including Assumption 1 regarding multivariate ordinal selection mechanisms Assumption 2 concerning the conditional exogeneity of treatment and exclusion restrictions and Assumption 3 about the common support of treatment.

Paper summary: Ruth: So we're looking at this paper today called "Double machine learning for causal inference in a multivariate sample selection model". Basically, they're proposing new ways to estimate average treatment effect on the average treatment effect and average treatment effect on the treated within a multivariate sample selection model that has multiple ordinal selection equations.

Warren: And the main claim is that without properly handling this multivariate selection, your causal parameter estimates can be highly biased. But this paper suggests their double machine learning estimators allow you to avoid those biases.

Iris: It sounds like they are tackling the issue where an outcome isn't observable unless you meet several different criteria, not just one simple condition. This is a more complex setup than what we usually see in standard selection models, which is important for real-world data.

Ruth: Exactly. The paper sets up this multivariate sample selection model with multiple ordinal selection equations where the outcome only shows up under specific conditions defined by a vector of selection variables. They lay out the necessary assumptions for making those average treatment effect and average treatment effect on the treated parameters identifiable in this setup.

Warren: And they then build these Neyman-orthogonal scores based on observation vectors that include the outcome, the treatment, the covariates, and some nuisance functions called mu to facilitate these estimations.

Iris: The methodology involves using these scores to derive double machine learning estimators for both average treatment effect on the average treatment effect and average treatment effect on the treated. It's an extension of methods used in simpler binary selection models, like the one proposed by Bia et al., but adapted here for this more complex model.

Ruth: So, they are using a process where they partition their sample and get estimators for various nuisance functions through regression techniques before averaging those scores across the sample to get the final estimates for both average treatment effect on the average treatment effect and average treatment effect on the treated.

Warren: And to check if this works, they conduct simulation results. They show that these double machine learning estimators are more accurate than estimators that only address one selection equation or none at all. For instance, they report that the mean absolute error for the average treatment effect on the average treatment effect drops from zero point two hundred fifteen in a small sample to zero point one hundred twenty-eight in a large sample when using double machine learning, while an estimator without addressing selection bias doesn't shrink with the sample size.

Paper summary: Iris: That difference in performance, that shrinking error, is what makes it compelling for causal inference. It shows that they are actually capturing some of the underlying causal structure better than simpler approaches. But I wonder about the assumptions they made regarding how these selection equations interact—that’s a big area for scrutiny in this kind of model.

Ruth: Right, Iris, that's exactly what we need to unpack later. The core message is that ignoring multivariate sample selection leads to potentially highly biased causal parameter estimates, and this double machine learning approach gives us a way to avoid those biases in these more realistic scenarios. This whole paper is about building tools that work when the selection process is complicated by multiple criteria.

Warren: And they test it out on a real-world application too, showing they apply these estimators to study the effect of female education on the log-wage ratio of spouses. That shows it has some practical relevance beyond just theoretical identification.

Iris: The real implication here is moving beyond single selection equations when modeling observable outcomes like wages or household spending, where multiple factors can determine whether that outcome is even seen by an observer. This method gives us a way to handle that complexity without falling into the traps of simple biased estimates.

Ruth: So, to wrap up for this part, the paper "Double machine learning for causal inference in a multivariate sample selection model" proposes double machine learning estimators for average treatment effect on the average treatment effect and average treatment effect on the treated in models with multiple ordinal selection equations. It establishes their asymptotic normality and shows finite sample properties through simulation data.

Warren: The key finding is that addressing multivariate sample selection prevents highly biased causal parameter estimates, which is a significant improvement over methods that ignore this selection structure. This work also applies these estimators to study the effect of female education on the log-wage ratio of spouses, which gives it some practical grounding.

Iris: And the method itself relies on constructing Neyman-orthogonal scores from observation vectors and then using double machine learning to estimate those scores across the sample. It's a sophisticated way to handle the identification in these nested selection structures, though we should keep an eye on how robust those initial assumptions are when dealing with such multiple criteria.

Ruth: Right, so we've seen how they set up this multivariate selection model, what they claim their double machine learning estimators can do—avoiding bias where simpler methods fail—and the fact that they tested it on a wage ratio study. It really puts the dots together for us on why this kind of modeling is necessary for complex economic questions.

Paper summary: Warren: And for those of you listening who are thinking about running these kinds of estimations, the simulation results show a real improvement in accuracy, with an error falling from zero point two hundred fifteen to zero point one hundred twenty-eight when using double machine learning versus methods that don't account for the selection process.

Iris: And we need to remember that the paper itself points out its limitation, which is about the assumptions required for identifiability in this model, especially concerning how those multiple selection equations nest within each other. That's where future work needs to focus if we want to push this into even more complex settings.

Ruth: So, we've covered the paper summary, and now we’re moving toward what this means for causal inference in general. It shows that when the selection process is complicated by multiple criteria, standard methods break down and you need a more robust estimator like the double machine learning one to get reliable results.

Warren: And when you apply this to things like education and wages, it suggests that we can get much cleaner estimates of how factors influence outcomes even when those factors are themselves subject to complex selection rules. It makes the causal picture clearer in these intricate scenarios.

Iris: The authors are doing a lot of heavy lifting here by proving asymptotic normality and showing finite sample performance on simulated data, which gives us more confidence in trusting these estimators when we actually try to use them on real datasets.

Ruth: So, to conclude our discussion of "Double machine learning for causal inference in a multivariate sample selection model," the authors propose double machine learning estimators that handle multiple ordinal selection equations. They show these estimators avoid the high bias you get from ignoring this complex selection process, and they give us a way to study things like education's effect on spouse's wages more reliably.

Warren: And if you're thinking about applying this, remember the performance gain they showed in simulations: accuracy improves significantly when you account for the multiple selection structure. It’s about getting better numbers on your causal parameters without having to throw out your entire analysis setup.

Iris: And what this means for the field is that we have a new way to handle these nested selection problems by building scores and then using double machine learning to estimate them, providing a path forward when single selection equations aren't enough.

Ruth: We’re going to take a quick pause here before we talk more about how this work fits into the broader landscape of causal inference research on arXiv.

Warren: That's right, we'll keep talking about the implications of this paper coming up in the next segment.

Conclusion: Ruth: "Double machine learning for causal inference in a multivariate sample selection model." That’s the title of this piece we’re looking at today.

Warren: Yeah, it sets out how you can get better estimates of those averages when the rules for seeing an outcome are complicated by several different conditions.

Iris: The authors show that if you ignore these multiple selection rules, your causal parameter estimates can be pretty heavily biased.

Ruth: Exactly what they’re saying is that the way people choose to observe something matters a lot in these models.

Warren: They propose this double machine learning approach as a way to avoid those biases when dealing with messy selection processes.

Iris: They establish that their estimators are asymptotically normal, which means they behave predictably as your data gets bigger, and they also test them on simulated data to see how they perform in the real world.

Ruth: And the simulation results show that these double machine learning estimators actually get more accurate than simpler methods when you account for all those selection equations.

Warren: The error for the average treatment effect on the average treatment effect drops from about zero point two one five in a small sample to zero point one two eight in a large sample with this method.

Iris: It’s showing that double machine learning isn't just theoretical; it actually delivers better numbers when you have this kind of complex selection structure.

Ruth: So, the main implication is that for real-world data where outcomes depend on several factors at once, you need a more robust way to figure out those causal effects.

Warren: It suggests that we can get cleaner pictures of how things actually work when the rules for observation are intricate.

Iris: But we still have to keep an eye on the assumptions they made about how these multiple selection equations interact, because that’s where you need to be careful with identification.

Ruth: Right, so while this method seems powerful for handling complex selection, we still need to make sure the underlying assumptions about those nested rules hold up when we apply this to new situations.

Warren: That’s exactly what we’ll be digging into next when we talk about where this research fits in the bigger picture of causal inference on arXiv.

More episodes

← Home