Technical Summary: From Tokens to Policy: Causal and Interpretable Heterogeneous Treatment Effects Identification

1. Problem Statement

Heterogeneous Treatment Effect (HTE) identification is essential for optimizing interventions, yet existing approaches face a fundamental trade-off between expressivity and interpretability. Traditional methods either model treatment effects flexibly without identifying primitive drivers (lacking interpretability) or identify interpretable subgroups over curated structured covariates (lacking expressivity and potentially missing unmeasured drivers).

A critical gap exists in causal interpretation. HTE is causally explained by direct effect modifiers (interactors) that interact with the treatment. However, observed variables often include:

• Indirect modifiers: Ancestors of direct modifiers operating via mediation.
• Proxy modifiers: Descendants of direct modifiers.
• Common-cause modifiers: Variables sharing a common ancestor with direct modifiers.

Naive estimation on these proxies can yield spurious characterizations that fail to generalize under new policy regimes. Furthermore, in modern data regimes involving multi-modal and unstructured data (e.g., satellite imagery, sensor streams), direct modifiers are rarely measured directly but are entangled within complex pre-treatment observations. The challenge is to identify the minimal, sufficient, and causal set of direct modifiers from these high-dimensional, entangled representations without relying on prohibitive human annotation.

2. Methodology: Neural EXposure Interaction Search (NEXIS)

The authors propose NEXIS, an iterative procedure that reframes HTE identification as a Markov-blanket discovery problem over a learned, sufficient, and aligned pre-treatment representation.

2.1 Theoretical Framework

The method relies on three core assumptions to ensure causal identifiability:

1. Measurement Sufficiency: The measured pre-treatment variables $X$ (e.g., survey data + satellite imagery) entangle all heterogeneous treatment effect information.
2. Representation Sufficiency: A learned representation $Z$ (derived from $X$) preserves all HTE information.
3. Principal Alignment: Each latent direct effect modifier $W_{dir, k}$ is summarized by a distinct, dominant coordinate $Z_{j_k}$ in the representation space, with no residual signal scattered across other coordinates.

Under these assumptions, the set of principal proxy coordinates $S^*$ forms a Markov blanket for the Conditional Average Treatment Effect (CATE). Identifying $S^*$ allows for the recovery of the causal interventional effect $\tau_{do}(W_{dir})$.

2.2 The Algorithm

NEXIS operates as a forward-backward selection procedure on a dictionary of representation coordinates (e.g., sparse autoencoder features):

• Forward Step: Selects the coordinate $j$ that carries the strongest residual heterogeneity signal conditional on the currently selected set $S$. This is determined via a CATE-equivalence test ($H_0: E[\tau | Z_{S \cup \{j\}}] = E[\tau | Z_S]$).
• Backward Step: Re-tests all coordinates currently in $S$. If a coordinate becomes redundant given the rest of the set (i.e., the null hypothesis cannot be rejected), it is removed.
• Termination: The process repeats until the selected set stabilizes.
• Error Control: The procedure uses Bonferroni correction (or FDR) to control the Family-Wise Error Rate (FWER), ensuring high-probability recovery of the true set $S^*$.

2.3 Implementation Details

• Representation Learning: The authors utilize Sparse Autoencoders (SAEs) trained on top of frozen foundation-model encoders (e.g., SigLIP for images, Prithvi-EO for geospatial data). This creates a dictionary of interpretable, sparse features.
• Interpretation: Selected features are labeled post-hoc using Vision-Language Models (VLMs) by contrasting top-activating inputs against random inputs.
• Statistical Testing: The default implementation uses a linear treatment-interaction test, though doubly-robust Generalized Covariance Measure (GCM) variants are available for non-linear settings.

3. Key Contributions

1. Causal HTE Characterization: The paper introduces the first procedure to identify a causal and interpretable HTE characterization from complex, unstructured observations. It moves beyond pointwise predictions to identifying the specific "primitive drivers" (direct modifiers) of heterogeneity.
2. NEXIS Algorithm: A principled, iterative forward-backward selection algorithm with provable recovery guarantees. The authors prove that under Principal Alignment and mean-faithfulness, NEXIS consistently recovers the set of principal proxies $S^*$ with high probability.
3. Experimental Power Paradox: The authors identify and demonstrate a structural failure mode in marginal screening: as sample size or effect size increases, marginal tests accumulate non-direct modifiers (false discoveries), causing precision to collapse. NEXIS mitigates this via conditional reasoning.
4. Real-World Deployment: The method is deployed on two anti-poverty programs in Africa, augmenting standard survey data with satellite imagery to uncover environmental effect modifiers previously invisible to policymakers.

4. Results

4.1 Semi-Synthetic Validation (CelebA)

Using a benchmark based on the CelebA dataset with known ground-truth direct modifiers (wearing hats, wearing eyeglasses):

• Precision vs. Recall: While marginal screening methods achieve high recall, their precision collapses as sample size or effect magnitude increases (the "Experimental Power Paradox"). NEXIS maintains both high recall and high precision, consistently recovering the true set $S^*$.
• Ablations: The method is robust to variations in SAE sparsity, feature views (sparse vs. dense), and statistical test choices (linear vs. GCM).

4.2 Application 1: Youth Opportunities Program (Uganda)

• Context: A cash grant program for youth groups in post-conflict Northern Uganda.
• Findings: NEXIS identified that individual demographics (age, sex) were invariant to the program's impact. Instead, impact heterogeneity was driven by:
  • Ethnolinguistic groups: Reflecting post-conflict market connectivity (e.g., conflict-affected areas showed dampened effects).
  • Environmental factors: Presence of perennial rivers and vegetation spatial heterogeneity reduced program impact, likely by providing alternative subsistence options (fishing, bush subsistence) that reduced the incentive for skilled trade adoption.
• Comparison: Marginal screening on the same data yielded dozens of spurious "discoveries," whereas NEXIS isolated a parsimonious set of conditionally independent modifiers.

4.3 Application 2: LEAP 1000 Programme (Ghana)

• Context: A cash transfer program for households with newborns in Northern Ghana.
• Findings: No demographic modifiers were found. Impact heterogeneity was driven by:
  • Ephemeral waterways: Households near seasonal streams saw effects 6–8 times larger than the average. The mechanism is hypothesized to be the complementarity between cash and seasonal irrigation capabilities.
  • Closed-canopy forests: Households near forest patches also saw amplified effects, likely due to the ability to layer cash onto existing non-timber forest livelihood strategies.
• Validation: Temporal analysis of satellite imagery confirmed that agricultural land use expanded near waterways post-intervention, supporting the complementarity hypothesis.

5. Significance and Claims

The paper claims a paradigm shift from static HTE estimation on survey data to extensive exploration across unstructured modalities. By leveraging modern representation learning (SAEs on foundation models) and rigorous causal discovery (Markov blanket identification), the authors argue that identifying causal, interpretable HTE characterizations is now "within reach."

The significance lies in translating raw, complex measurements (tokens) directly into prescriptive policy guidelines. The authors demonstrate that integrating external data sources (satellite imagery) can reveal environmental effect modifiers that standard surveys miss, offering concrete, novel explanations for why interventions succeed or fail in specific contexts.

Modesty and Limitations:
The authors explicitly state that their causal guarantees rely on assumptions (Measurement Sufficiency, Principal Alignment, Test Validity) that cannot be fully verified without auxiliary experiments. Consequently, they frame their real-world findings as "strong candidate hypotheses" intended to guide future program iterations rather than established causal facts. They acknowledge limitations regarding multilevel statistical corrections and the granularity of environmental measurements (community-level vs. individual-level GPS).