Causal inference traditionally relies on "paired data"—observing both a cause ($X$) and an effect ($Y$) on the exact same individual unit. But what happens when measuring the cause inherently destroys the subject? In fields like biology, techniques such as single-cell sequencing require destroying a cell to measure its internal state ($X$), making it impossible to measure a subsequent outcome ($Y$) on that same cell.
This creates a scenario of unpaired data. A paper by Felix Schur and colleagues, presented at ICML 2024, tackles a provocative question: Is causal inference possible when we never observe $X$ and $Y$ together?
To solve this, the authors introduce SPLITUP, a novel estimator designed to identify causal effects in unpaired data settings across many experimental environments, overcoming the persistent biases that plague standard instrumental variable methods.
The Challenge: Unpaired Data and Hidden Confounding
The researchers address a highly specific and challenging scenario characterized by three factors:
- Unpaired Data: We possess one dataset containing the environment and the cause $(I, X)$, and another containing the environment and the effect $(I, Y)$. The environment $I$ could be a specific experimental condition, like a gene knockout. We never observe $(X, Y)$ jointly.
- Hidden Confounding: Unobserved variables ($U$) influence both $X$ and $Y$, rendering simple correlations misleading.
- High-Dimensional Instruments: Modern screening experiments (like CRISPR screens) often involve thousands of environments ($m$) but only a handful of repetitions or cells ($n$) per environment.
In classical econometrics, "Two-Sample Instrumental Variable" (TS-IV) methods handle unpaired data by treating the environment as an instrument. However, these methods rely on the assumption that the number of observations per environment approaches infinity. In the modern "many-instrument" regime (large $m$, small $n$), standard TS-IV estimators fail. The noise inherent in estimating the $X$ averages does not cancel out; instead, it creates a persistent measurement error that severely biases the final causal estimate.
The Solution: The SPLITUP Estimator
To overcome this bias, the authors propose a Generalized Method of Moments (GMM) approach called SPLITUP.
Cross-Moment Estimation
The core innovation of SPLITUP is its use of cross-moment estimation. Standard methods use a single estimate of the covariance between $X$ and the instrument, which introduces a biased "squared noise" term in the denominator of the estimator. SPLITUP fixes this by splitting the $X$ data into independent folds. By multiplying the estimates from different folds, the noise in fold A remains independent of the noise in fold B. Their product has an expectation of zero, effectively "cleaning" the bias.
The authors also derive a closed-form, analytic version of this estimator that averages over all possible data splits, maximizing statistical efficiency and removing the variance introduced by any single random split.
Handling Sparsity
In many practical applications, researchers might measure hundreds of covariates ($X$), but only a sparse subset actually causes the outcome ($Y$). The authors extend SPLITUP with $\ell_1$-regularization (Lasso-style) to identify these sparse causal effects, even when the number of covariates exceeds the number of environments.
Experimental Validation
The paper evaluates SPLITUP across both synthetic and real-world datasets:
- Synthetic Simulations: The authors generated data mimicking categorical environments and continuous instruments to test the high-dimensional regime. While standard TS-IV and TS-2SLS methods showed a flat-line bias that never disappeared regardless of total sample size, SPLITUP remained consistent, approaching the true causal effect as data grew.
- Avoiding the "Peaking" Phenomenon: The simulations revealed a "peaking" phenomenon where TS-IV error drastically increases when the number of environments roughly equals the number of covariates—a finding supported by random-matrix theory. SPLITUP avoids this failure mode entirely.
- The Causal Chambers Light Tunnel: For real-world validation, the team used a physical testbed with programmable LEDs and photodiode sensors. Using red-channel intensity as a hidden confounder and infrared LEDs as instruments, SPLITUP successfully recovered the known causal link between sensors from intentionally unpaired data, whereas baseline methods remained significantly biased.
Key Takeaways and Limitations
A striking theoretical result of this work is that SPLITUP is consistent as the number of environments ($m$) grows, even if the number of observations per environment ($n$) is as low as two. This is a major departure from standard IV theory.
The authors also highlight the "averaging trap": a common intuitive practice is to average $X$ and $Y$ within each environment and run a regression on those means. The paper proves this is mathematically equivalent to TS-IV and inherits its fatal attenuation bias in high-dimensional settings.
Limitations and Future Directions: * Linearity: The current framework assumes the structural causal effect of $X$ on $Y$ is linear, though the relationship between the instrument and $X$ can be non-linear. * Invariance: The method assumes the relationship between the instrument and $X$ remains identical across the two different samples. If environments behave differently in the $X$-sample versus the $Y$-sample, the estimator could fail.
Future work will look toward extending this framework to non-linear structural effects and developing robust uncertainty quantification (like confidence intervals) for high-dimensional, sparse settings. Ultimately, SPLITUP provides a vital new tool for researchers navigating the complex realities of modern, destructive experimental designs.
Summary updated