Identifying Effects of Multivalued Treatments

Sokbae Lee, Bernard Salanié

Abstract: Multivalued treatment models have typically been studied under restrictive assumptions: ordered choice, and more recently unordered monotonicity. We show how treatment effects can be identified in a more general class of models that allows for multidimensional unobserved heterogeneity. Our results rely on two main assumptions: treatment assignment must be a measurable function of threshold-crossing rules, and enough continuous instruments must be available. We illustrate our approach for several classes of models.

Introduction

In social sciences and medicine, researchers rarely have access to perfect experimental data. Instead, they must contend with selection bias, where individuals choose to participate in a treatment based on unobserved factors. Traditional econometric tools like the Local Average Treatment Effect (LATE) rely on a "monotonicity" assumption. In a binary setting, this implies that an instrument (an external factor) encouraging treatment should not simultaneously discourage others.

However, real-world choices are often multivalued—such as choosing between no college, a 2-year degree, or a 4-year degree—and are driven by multiple hurdles. For instance, 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 enter the program while others exit. These "two-way flows" break standard models. This paper provides a framework to identify treatment effects in these complex, non-monotonic environments.

Theoretical Framework: Threshold-Crossing Rules

The authors propose a model where treatment assignment is governed by a vector of threshold-crossing rules. Unlike simpler models that rely on a single hidden variable, this framework allows for a vector of unobserved characteristics ($V$) and a vector of thresholds ($Q$) that depend on external instruments ($Z$).

Figure 1: Partition of the unobserved heterogeneity space (V1, V2) into treatment regions D based on threshold-crossing rules.

As illustrated in the figure above, the space of unobserved heterogeneity is partitioned into different treatment regions ($D=0, 1, 2$). The treatment a person receives is a "measurable function" of whether their hidden variables cross specific thresholds. The core technical innovation is the use of multidimensional cross-derivatives. By taking the $J$-order derivative (where $J$ is the number of thresholds) of the observed data with respect to the instruments, the authors isolate the Marginal Treatment Effect (MTE).

Key Insights and Identification

Embracing Two-Way Flows

The paper’s primary contribution is showing that researchers do not need to assume everyone reacts to an instrument in the same direction. By mathematically "untangling" the different groups moving in and out of treatment cells, the model identifies effects even when monotonicity fails.

Partitioned Choices

Often, researchers only see aggregated data (e.g., "enrolled in training" vs. "not enrolled"), even though the "training" category contains many distinct programs. This framework allows for the identification of the effects of those underlying programs even when they are not explicitly labeled, provided the underlying threshold structure is understood.

Point Identification

The authors prove that the joint density of unobserved heterogeneity and the conditional expectation of counterfactual outcomes are point-identified. This allows for the derivation of common metrics like the Average Treatment Effect (ATE) and the Average Treatment Effect on the Treated (ATT) from granular MTE results.

Applications

While theoretical, the paper demonstrates the framework's utility through several analytical examples:

  • The Double Hurdle Model: A scenario where two conditions must be met for treatment, a staple in labor and health economics.
  • Entry Games: A model of firm competition where market entry depends on mutual profitability.
  • Dynamic Treatments: A decision-tree model where a subject's path is governed by a sequence of thresholds, such as grades or performance reviews over time.

Limitations and Future Work

The method relies heavily on the availability of continuous instruments that provide variation across all thresholds. In practice, many instruments are discrete (e.g., state-level policy changes), which may require different estimation techniques. Additionally, as the number of thresholds ($J$) increases, the requirement for high-order derivatives makes the model more sensitive to data noise, potentially complicating estimation in smaller samples.

Summary updated

Abstract

Multivalued treatment models have typically been studied under restrictive assumptions: ordered choice, and more recently unordered monotonicity. We show how treatment effects can be identified in a more general class of models that allows for multidimensional unobserved heterogeneity. Our results rely on two main assumptions: treatment assignment must be a measurable function of threshold-crossing rules, and enough continuous instruments must be available. We illustrate our approach for several classes of models.
BibTeX
@inproceedings{lee2018identifying,
  author = {Sokbae Lee and Bernard Salanié},
  title = {{Identifying Effects of Multivalued Treatments}},
  booktitle = {Econometrica},
  year = {2018},
  doi = {10.3982/ecta14269},
}
Figure