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

Introduction

Contributions

• Causal and interpretable framework

  More data doesn't help if we ask the wrong question. We distinguish variables correlated with treatment-effect heterogeneity from those that causally drive it, and formulate the search for the latter as a Markov-blanket discovery problem over a sufficient and aligned pre-treatment representation. We show when and how it is achievable end-to-end, from raw multi-modal measurements to interpretable policy hypotheses.
• New method

  We introduce NEXIS, a forward–backward Markov-blanket procedure based on conditional CATE-equivalence tests: a forward step followed by one terminal backward step. It comes with a selection guarantee, and is extensively tested and ablated on semi-synthetic experiments, providing practical guidelines for different data regimes.
• New candidate discoveries in development economics

  We deploy NEXIS on two anti-poverty program evaluations in Sub-Saharan Africa, extending both with satellite imagery to capture environmental modifiers. In both programs, the selected modifiers are mainly environmental factors, coarsely proxied or absent in survey data, while no individual-level demographic is selected. With no ground truth, these results are hypotheses for field validation, not established causal modifiers or policy prescriptions.

Experimental Power Paradox

Example: Consider a cash transfer program targeting mothers with a newborn, designed to increase the monthly household expenditure on infant care. Market access is a latent direct modifier — communities near markets can convert the cash into goods for the newborn more effectively. Road density, visible from satellite imagery, correlates with market presence and surfaces as an effect modifier, but only by common cause. A naive interpretation suggests building roads to amplify the program effect; the actual driver is market presence, so a road built where no market exists leaves the effect unchanged.

Method

From Tokens to Policy explaining the effect heterogeneity

Procedure:

1. Experiment. Design a controlled experiment, run and measure any candidate treatment interactors.
2. Represent. Represent such pre-treatment measurements in an interpretable dictionary with minimal or no human supervision, e.g., training a Sparse Autoencoder (SAE) on a frozen domain-specific foundation model.
3. Identify. Neural EXposure Interaction Search (NEXIS) retrieves the minimal and sufficient heterogeneous effect characterization by conditional CATE-equivalence tests on the representation coordinates, adapting forward–backward Markov-blanket discovery:

   • Forward step (IAMB’s grow phase): given the current selection S, add the coordinate with the smallest p-value if it passes the gate α/|S̄|, where S̄ are the coordinates not yet selected; repeat until no coordinate passes.
   • Terminal backward step (MMPC/HITON-PC subset rule), run once: keep a selected coordinate j only if its p-value is at most α/m given every subset of the other selected coordinates, with m the number of candidates.

   An optional interleaved backward step and a spectral-gap gate ρ are practical variants discussed in the paper’s appendix; neither is part of Algorithm 1, animated below. Interpret the selected coordinates a posteriori, e.g., by querying a VLM on the most and least activating observations.

   ---

   Theorem (Causal Identification)

   In a randomized experiment, with faithfulness and valid and consistent CATE-equivalence tests, assuming Principal Alignment of the representation, NEXIS’ output Ŝ_n recovers the principal proxies S* of the direct effect modifiers (selection guarantee):

   lim inf_n P(Ŝ_n = S*) ≥ 1 − α

   ---

   If we further assume Measurement and Representation Sufficiency, the principal proxies identify the causal and interpretable heterogeneity characterization (causal identification):

   τ^do(W^dir) = τ(W^dir) = E[τ | Z^S*]  almost surely
4. Policy. Update the policy accordingly, e.g., the treatment assignment of the next program iteration. Structural assumptions determine the causal reading, so discoveries call for cautious interpretation and experimental validation first.

Controlled Experiments — CelebA Benchmark

Case Study 1: Youth Opportunities Program (Uganda)

Program. The Youth Opportunities Program (YOP) is an anti-poverty program launched in 2008 in post-conflict Northern Uganda, awarding cash grants (approximately $382 per group member, plus optional vocational training) to self-organised youth groups, randomized across 535 groups. The original evaluation reports a positive average effect on productive economic capacity, captured through skilled employment and business asset accumulation. We ask why and how the impact varies.

Experiment. Our analytic sample has 2,082 individuals in 439 groups and 327 distinct communities (331 RCT sites) across four sub-regions (Karamoja, Teso, Lango and West Nile), 825 of them treated. Outcomes are skilled employment and log business assets at endline; pre-treatment covariates include age, sex, parental education, the group share of women and the community language group (7 ethnolinguistic clusters). We define the treatment T as grant receipt: 29 of the 265 groups assigned to the grant never received it and count as untreated. This exploratory as-treated analysis assumes non-receipt is as good as random.

Data fusion (ours). We extend the trial with Landsat-7 multispectral imagery (2005–2007 composite, immediately before the program), embedded via Prithvi-EO and summarised by a 1,024-atom TopK SAE trained on a held-out Uganda national grid. The 146 atoms active in at least 5 sites, 12 spectral-index summaries (NDVI, NDWI, MNDWI, NDBI, EVI, BSI; mean and standard deviation per tile) and the trial covariates form 170 NEXIS candidates. The map below shows the trial communities by treatment, outcome or covariate.

Results

Certified modifiers are selected by the forward step and survive the terminal backward step; candidate modifiers are selected by the forward step only. Both are hypotheses: exploratory findings to be investigated further, since whether the requirements of NEXIS’s guarantee hold in practice cannot be verified here.

Case Study 2: Livelihood Empowerment Against Poverty (Ghana)

Program. The Livelihood Empowerment Against Poverty 1000 (LEAP 1000) programme was launched in Ghana in 2015 to target early-childhood poverty and stunted nutrition, providing bimonthly cash transfers to extremely poor households with newborns. Its evaluation uses a regression discontinuity design on a proxy means test score, with difference-in-differences estimation between the 2015 baseline and the 2017 endline.

Experiment. The evaluation covers 2,331 households (1,185 treated) in 162 communities across five districts in two regions, Northern and Upper East (region boundaries as of 2015; the map shows today’s regions). The outcome is monthly adult-equivalent household consumption expenditure, and 24 pre-treatment survey covariates describe household composition, the household head, housing, water and sanitation, and livelihoods. The difference-in-differences estimate of the local average treatment effect on the treated (local ATT), for households near the eligibility cutoff, is +7.35 GH₵/month.

Data fusion (ours). We extend the evaluation with Landsat-8 2015 imagery, embedded via Prithvi-EO and summarised by a 4,096-atom TopK SAE trained on a held-out Ghana national grid. The 131 atoms active in at least 5 communities, 12 spectral-index summaries and the 24 survey covariates form 167 NEXIS candidates. Standard errors are clustered by community. The map below shows the GPS centroids of the 162 communities.

Results

Certified modifiers are selected by the forward step and survive the terminal backward step; candidate modifiers are selected by the forward step only. Both are hypotheses: exploratory findings to be investigated further, since whether the requirements of NEXIS’s guarantee hold in practice cannot be verified here.

Take Aways

Citation

@article{cadei2026nexis,
  title   = {From Tokens to Policy: Causal and Interpretable Heterogeneous Treatment Effects Identification},
  author  = {Cadei, Riccardo and Otchere, Frank and Tirivayi, Nyasha and Angeles Tagliaferro, Gustavo and Bargagli-Stoffi, Falco J. and Locatello, Francesco},
  year    = {2026},
  url     = {https://arxiv.org/abs/2606.17010},
  note    = {Under review}
}