TL;DR
Causal inference usually requires observing causes and effects together, but what if they are measured in entirely different samples? This paper introduces SplitUP, a novel method that consistently estimates causal effects from unpaired data across many experimental environments with few repetitions. By leveraging cross-fold sample splitting, SplitUP eliminates the persistent measurement-error bias that causes traditional two-sample instrumental variable methods to fail in high-dimensional regimes.
What Problem Does This Paper Address?
Imagine trying to find the causal effect of gene expression ($X$) on a cellular phenotype ($Y$). To measure $X$, you might have to destroy the cell; to measure $Y$, you have to let it live and grow. Consequently, you can never measure $X$ and $Y$ in the exact same cell—your data is completely unpaired, as illustrated by the missing values ($\times$) in the data matrix below:
However, you can observe these systems across many different experimental conditions or environments ($m$), such as different gene knockouts. This allows you to treat the environment as an "instrumental variable" (IV) to bypass hidden confounding ($U$), as diagrammed in their causal setup:
While two-sample IV (TS-IV) methods exist for unpaired data, they assume a large number of observations per environment. In modern biology (like single-cell perturbation screens), we face the opposite: many environments but few repetitions per environment. In this high-dimensional regime, standard TS-IV estimators break down because the noise in estimating the instrument-covariate relationship creates a severe measurement-error bias that does not vanish as the number of environments grows.
Methods
The authors propose SplitUP, a Generalized Method of Moments (GMM) estimator designed specifically for the many-environment, few-repetition regime.
- Cross-Fold Sample Splitting: To fix the measurement-error bias in the denominator of the estimator, SplitUP splits the instrument-covariate sample into $K \ge 2$ disjoint folds. It computes fold-wise covariances and uses cross-moments (multiplying independent views of the covariance) to ensure the expectation is unbiased.
- SplitUP (analytic): They derive a elegant closed-form solution for the infinite-split limit ($H \to \infty$). This analytical version subtracts a diagonal "self-inner-product" correction from the standard plug-in quadratic form, removing the bias without adding random splitting variance.
- Sparsity Handling ($\ell_1$-regularization): When the number of covariates $d$ is large (potentially even larger than $m$), they introduce an $\ell_1$-penalized version of the cross-moment GMM, followed by a post-selection refitting step to provide valid confidence intervals.
Datasets
The paper evaluates the method using extensive synthetic datasets modeled after real-world biological screening scenarios (such as Perturb-seq genomics data). They simulate both categorical instruments (representing discrete experimental conditions) and continuous instruments across three distinct settings:
- Finite-dimensional instruments with sparse effects.
- High-dimensional instruments with dense effects.
- High-dimensional instruments with sparse effects.
Key Insights and Ideas
- The Unpaired Bias Flip: In standard paired data, having many weak instruments biases your estimator toward the confounded OLS solution. In unpaired data, the high-dimensional instrument setting creates a measurement-error effect that acts as an attenuation bias, pulling the estimate toward zero instead.
- Independence via Splitting: By ensuring that the covariance matrices multiplied in the denominator come from independent data splits, the noise terms become uncorrelated, and their product zeroes out in expectation.
- Sparsity as an Enabler: Causal effects can be identified even when the number of covariates exceeds the number of environments, provided the underlying causal effect vector is sparse.
Key Results
SplitUP significantly outperforms traditional two-sample IV as the number of environments ($m$) increases, maintaining a low Mean Absolute Error (MAE) even with as few as 2 to 4 observations per environment:
In high-dimensional settings with dense effects, standard TS-IV and standard UP-GMM remain asymptotically biased, whereas SplitUP converges to the true causal effect as $m \to \infty$. The authors also document a fascinating "peaking" phenomenon where TS-IV error spikes dramatically at intermediate sample sizes due to near-singularity of the plug-in Gram matrix before recovering.
Limitations and Future Work
- Beta-min Condition: The support recovery and post-selection inference rely on a "beta-min" condition, which assumes that all non-zero causal effects are bounded safely away from zero—an assumption that can be restrictive in real-world applications.
- Linearity Assumption: The current framework assumes a linear structural equation model between covariates $X$ and outcome $Y$. Extending this to non-linear causal relationships remains an open challenge.
- Instrument Validity: The method strictly requires the exclusion restriction (environments affect $Y$ only through $X$). If there is direct pleiotropy (environments directly affecting $Y$), the estimator will be biased.
Summary updated