TL;DR
This paper introduces a novel instrumental variable (IV) framework that identifies the causal effects of continuous or discrete treatments using only a binary instrument. By leveraging a "copula invariance" assumption, the authors provide a constructive method to estimate heterogeneous treatment effects across the entire population, overcoming the traditional requirement for instruments with high variation.
What Problem Does This Paper Address?
In causal inference, two major hurdles are endogeneity (unobserved factors influencing both treatment and outcome) and heterogeneity (the treatment effect varying across individuals).
Instrumental variables (IV) are the standard tool for these problems, but they often come with a restrictive trade-off. Existing methods typically require either:
- Strict outcome structures: Assuming "rank invariance" (individuals maintain their relative ranking across potential outcomes).
- Rich instrument variation: For continuous treatments (like hours of sleep or years of schooling), standard "control function" approaches usually require an instrument that is also continuous and has "large support" (lots of variation).
If you only have a binary instrument (e.g., a simple "treatment offered" vs. "not offered" in a trial with non-compliance), it is notoriously difficult to identify the global effects of a continuous treatment. This paper fills that gap.
Methods
The authors propose an intermediate route that imposes structure on the relationship between treatment assignment and potential outcomes rather than on the variables themselves.
- Local Gaussian Representation (LGR): The core of the approach is representing the joint distribution of potential outcomes and unobserved selection factors using a Gaussian copula. Crucially, this is nonparametric; the correlation parameter ($\rho$) is treated as a local function that can change at every point, meaning the model doesn't actually assume the data is Gaussian.
- Copula Invariance (CI): This is the key identifying assumption. It posits that the local dependence (the "shape" of the endogeneity) between the potential outcome and the selection unobservable does not change with the treatment propensity (the probability of being treated).
- Distribution Regression (DR): The identification results are "constructive," meaning they lead directly to an estimation recipe. The authors use distribution regression to estimate the entire conditional distribution of the outcome, allowing them to capture complex patterns of heterogeneity.
- Inference: They establish that their estimators are $\sqrt{n}$-consistent and asymptotically normal, providing a valid bootstrap algorithm for constructing confidence bands.
Datasets
The paper demonstrates the method using data from a field experiment by Bessone et al. (2021b):
- Context: A study of low-income adults in Chennai, India.
- Treatment ($D$): Sleep time per night (treated as a continuous variable).
- Instrument ($Z$): Randomized interventions (information, encouragements, sleep trackers, and environmental improvements).
- Outcome ($Y$): An index of individual well-being.
- Sample: $n = 226$ individuals who did not receive a secondary "nap" treatment.
Key Insights and Ideas
- Breaking the "Large Support" Requirement: The most significant "aha moment" is that you don't need a continuous instrument to study a continuous treatment. If you are willing to assume the form of the endogeneity is locally invariant (CI), a binary instrument is enough to identify global structural functions.
- CI vs. Rank Similarity: The authors show that CI is a distinct and often more flexible alternative to the "rank similarity" assumption used in Quantile Treatment Effects (QTE) literature. CI allows for richer patterns of treatment effect heterogeneity.
- Nuanced Selection Patterns: Unlike standard Roy models that assume joint normality, the CI framework can accommodate non-monotonic selection patterns. In the sleep study, they found negative selection: individuals with poorer underlying health conditions (unobserved) tended to sleep more, a pattern that would bias simpler models.
Key Results
- Heterogeneity in Sleep Effects: While standard Two-Stage Least Squares (2SLS) suggested moderate average effects of sleep on well-being, the authors' method uncovered significant heterogeneity. The effects were substantially larger at the lower tail of the well-being distribution.
- Policy Implications: This suggests that interventions to increase sleep are most beneficial for those with the lowest baseline well-being—a nuance missed by average-effect models.
- Robustness: The method identified that simpler estimators (ignoring endogeneity) would significantly underestimate the benefits of sleep because of the negative selection mentioned above.
Limitations and Future Work
- Assumption Strength: While "Copula Invariance" is flexible, it is still a high-level assumption. The authors provide some ways to test it if multiple instruments are available, but in the single binary instrument case, it must be justified by economic theory or context.
- Sample Size Sensitivity: In the empirical application, the small sample size ($n=226$) resulted in wide uniform confidence bands, meaning that while pointwise results were significant, the "global" patterns were less precisely estimated.
- Non-parametric Slopes: The authors focused on the Quantile Structural Function (QSF) rather than its derivative (the marginal effect) because estimating derivatives requires additional tuning parameters and results in slower convergence rates. Future work could refine these "slope" estimators within the CI framework.
Summary updated