This paper, published in 2018 by Sokbae Lee and Bernard Salanié, tackles one of the most persistent challenges in econometrics: how to accurately measure the impact of a "treatment" (like a job training program or a new drug) when people choose to participate based on complex, hidden factors.
TL;DR
The authors provide a new mathematical framework to identify treatment effects in scenarios where individuals choose between multiple options based on several different criteria. By moving beyond the restrictive "monotonicity" assumptions of the past, they show how to use continuous instruments and multidimensional calculus to uncover the true impact of treatments even when selection processes are messy and non-linear.
What Problem Does This Paper Address?
In social sciences and medicine, we rarely have perfect experiments. Instead, we deal with selection bias: people who choose a treatment are often different from those who don't in ways we can't see (unobserved heterogeneity).
Standard tools like the Local Average Treatment Effect (LATE) rely on a "monotonicity" assumption. In a binary world (Treatment vs. No Treatment), this means that if an external factor (an "instrument") encourages people to take a treatment, it shouldn't simultaneously discourage others.
However, real-world choices are often multivalued (e.g., choosing between no college, a 2-year degree, or a 4-year degree) and driven by multiple hurdles. For example, a student might need to pass both a math and a language exam to enter a program. If a policy change makes the math exam easier but the language exam harder, some students will move into the program while others move out. This creates "two-way flows" that break standard econometric models. This paper provides a way to identify treatment effects in these complex, non-monotonic environments.
Methods
The authors introduce a model where treatment assignment is governed by a vector of threshold-crossing rules.
- Multiple Thresholds: Instead of one hidden variable, they allow for a vector of unobserved characteristics ($V$) and a vector of thresholds ($Q$) that depend on external instruments ($Z$).
- Measurability: They assume that the treatment a person receives is a "measurable function" of whether these hidden variables cross their respective thresholds.
- Identification via Differentiation: The core technical innovation is the use of multidimensional cross-derivatives. They prove that by taking the $J$-order derivative (where $J$ is the number of thresholds) of the observed data with respect to the instruments, they can isolate the "Marginal Treatment Effect" (MTE).
- Control Function Approach: If the thresholds are known or can be identified, they can be used as control variables to handle the multidimensional "hidden" factors that usually bias our results.
Datasets
This is a theoretical econometrics paper, so it does not use a specific empirical dataset for training or evaluation. Instead, the authors use analytical examples to demonstrate the framework's utility:
- The Double Hurdle Model: A scenario where two conditions must be met for treatment.
- Entry Games: A model of firm competition where the number of firms entering a market depends on mutual profitability.
- Dynamic Treatments: A "decision tree" model where a student's or worker's path through time is governed by a sequence of thresholds (e.g., grades or performance reviews).
Key Insights and Ideas
- Embracing Two-Way Flows: The paper’s biggest "aha moment" is showing that we don't need to assume everyone reacts to an instrument in the same direction. We can mathematically "untangle" the different groups of people moving in and out of treatment cells.
- Partitioned Choices: Often, researchers only see aggregated data (e.g., "enrolled in training" vs. "not enrolled"), even though the "training" category contains many different programs. This framework allows researchers to identify the effects of those underlying programs even when they aren't explicitly labeled in the data.
- The "Index" of Treatment: They define the "index" of a treatment based on the highest-order term in its selection polynomial. If this index is non-zero, the treatment effect is point-identified through their differentiation method.
Key Results
- Point Identification: The authors prove that the joint density of unobserved heterogeneity and the conditional expectation of counterfactual outcomes (what would have happened under a different treatment) are point-identified.
- Aggregated Effects: They show how to derive more common metrics like the Average Treatment Effect (ATE) and the Average Treatment Effect on the Treated (ATT) from their more granular MTE results.
- Double Hurdle Identification: They provide new, specific identification theorems for the "double hurdle" model, which is a staple in labor and health economics but has historically been difficult to identify nonparametrically.
Limitations and Future Work
- Continuous Instruments: The method relies heavily on having "enough" continuous instruments that provide variation across all thresholds. In practice, many instruments are discrete (like a policy change that happens in some states but not others), which would require different techniques.
- Identifying Thresholds: The paper assumes the threshold functions ($Q$) are either known or can be identified first. Identifying these is highly model-specific and can be a significant hurdle in itself.
- Dimensionality: As the number of thresholds ($J$) increases, the requirement for high-order derivatives makes the model more sensitive to "noise" in the data, which could make estimation difficult with smaller samples.
Summary updated