This paper, published in arXiv (and later in the Journal of the Royal Statistical Society), explores a fundamental tension in causal inference: the more we try to account for hidden biases by adding variables, the harder it becomes to actually compare treated and control groups.
TL;DR
In observational studies, researchers often add high-dimensional covariates to satisfy the "unconfoundedness" assumption, but this paper proves that doing so makes the "overlap" assumption—the requirement that treated and control groups are comparable—mathematically much harder to satisfy. Using information theory, the authors show that as the number of variables grows, strict overlap forces the treatment and control groups to have nearly identical covariate means, creating a "curse of dimensionality" for causal inference.
What Problem Does This Paper Address?
To estimate a causal effect from observational data (where we don't have a randomized trial), we usually rely on two pillars:
- Unconfoundedness: We’ve measured every variable that affects both the treatment and the outcome.
- Overlap: Every individual, regardless of their characteristics, had a non-zero chance of being in either the treated or the control group.
There is a natural urge to include as many variables (covariates) as possible to satisfy the first pillar. The logic is: "The more I control for, the less likely I am to have hidden bias."
However, the authors point out a "double-edged sword." As you add more variables, it becomes easier to perfectly predict who received the treatment. If a combination of 100 variables can perfectly tell you that a person was treated, then there is no "overlap"—there is no comparable person in the control group. This paper formalizes this "curse of dimensionality," showing that the "strict overlap" assumption (where the probability of treatment is bounded away from 0 and 1) is a much more restrictive constraint than researchers typically realize.
Methods
The authors move away from looking at overlap as a "local" property of individual data points and instead treat it as a "global" restriction on the probability distributions of the treated and control groups.
- Information Theory Framework: They re-frame the strict overlap assumption as a bound on the likelihood ratio between the treated and control covariate distributions.
- f-Divergences: They use $f$-divergences (mathematical measures of how different two distributions are, such as Kullback–Leibler or Chi-squared divergence) to quantify the discrepancy between groups.
- Rukhin’s Theorem: They adapt a result from information theory (Rukhin, 1997) to prove that if the likelihood ratio is bounded (i.e., strict overlap holds), then the "distance" between the treated and control distributions must be small.
- Stochastic Processes: They model the covariates as a sequence to analyze what happens as the number of variables ($p$) approaches infinity.
Datasets
The paper is purely theoretical and does not use empirical or "real-world" datasets for benchmarking. Instead, it uses mathematical examples (such as Gaussian distributions and stationary ergodic processes) to illustrate how the bounds behave under different correlation structures.
Key Insights and Ideas
- The Vanishing Mean Difference: The most striking insight is that under strict overlap, the average difference between covariate means must shrink as the number of variables increases. Specifically, the average imbalance scales at a rate of $1/\sqrt{p}$.
- The Correlation Trade-off: For overlap to hold in high dimensions, the variables must either be highly redundant (highly correlated) or the groups must be almost perfectly balanced to begin with. You cannot have many independent, informative variables and still maintain overlap.
- The Classifier Bound: The authors prove that if strict overlap holds, no statistical classifier can accurately distinguish between the treated and control groups. If you can build a "treatment vs. control" model with high accuracy, you have effectively proven that your study violates the overlap assumption.
- Trimming is a Trap: A common fix for poor overlap is "trimming"—throwing away units with extreme treatment probabilities. The authors show that in high dimensions, this might require throwing away nearly the entire sample to achieve a desired level of overlap.
Key Results
- Theorem 1: Provides an explicit upper bound on the Euclidean distance between the mean vectors of the treated and control groups. This bound is determined by the overlap constraint ($\eta$) and the "operator norm" (a measure of correlation) of the covariance matrix.
- Corollary 1: Shows that if the variables aren't too correlated, the average absolute difference in means across all $p$ variables must go to zero as $p$ grows.
- Proposition 4: Demonstrates that as the dimension $p$ increases, the proportion of the population that satisfies a strict overlap bound ($\tilde{\eta}$) likely drops toward zero, making standard trimming procedures potentially non-viable.
Limitations and Future Work
- Strict vs. Weak Overlap: The paper focuses on "strict overlap" (probabilities bounded away from 0 and 1). While this is the standard assumption for many estimators, the implications for "weak overlap" (where probabilities can get very close to 0 or 1) are less clear.
- Sparsity Assumptions: The authors note that if the treatment assignment only depends on a small subset of variables (sparsity), the curse of dimensionality is mitigated. However, in many social science or medical contexts, treatment assignment is "dense" (affected by many small factors), where these results are most punishing.
- Practical Diagnostics: While the paper provides the theory, it suggests that more work is needed to develop practical, finite-sample tests to tell researchers when they have crossed the line into "no overlap" territory. (The authors mention they are addressing this in separate, related work).
Summary updated