{"id":"ba5526a7-d0fa-496a-ad09-896a674017e1","arxiv_id":"2603.19573","paper_version":2,"verdict":"CONDITIONAL","confidence":"LOW","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":0,"one_line_summary":"Potential-outcomes estimands and asymptotic theory for jointly estimating within- and between-cluster spillover effects under randomized saturation designs, with a Kenya cash-transfer application.","lead":"This paper develops causal estimands and estimators for both within-cluster and between-cluster spillover effects in randomized saturation designs, dropping the usual assumption that clusters do not interfere. Field experiments with geographic or social links across clusters can use the theory to separate local and cross-cluster spillovers.","discovery_kind":"extension","skeptic_critique":{"model":"grok-4.5","headline":"Between-cluster estimands rest on saturation-only exposure mapping that may miss heterogeneous unit-level cross-cluster links.","rationale":"The reader correctly isolates the exposure-mapping assumption as the load-bearing modeling step on which identification of the between-cluster estimands rests. The corrupted manuscript prevents verification of the subsequent proofs and the Kenya application, but that is already reflected in the CONDITIONAL/LOW-confidence verdict; the scientific soft spot remains the same. No stronger internal inconsistency is visible from the abstract and readable fragments, so the verdict needs no adjustment.","tokens_in":12530,"tokens_out":401,"duration_ms":15679,"concrete_test":"Simulate units on a lattice with known distance-decaying interference, form clusters as contiguous blocks, generate outcomes under the true network, then apply the paper’s saturation-based estimators; if the estimated between-cluster effect is biased relative to the true average spillover from neighboring clusters by more than Monte-Carlo error, the exposure mapping is misspecified for realistic cross-cluster dependence.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The strongest claim (identification + consistency/AN of within- and between-cluster spillover estimators under two-stage randomization) requires that potential outcomes for unit i are functions of own treatment and the vector (or low-dimensional summary) of other clusters’ saturations alone. When true interference is unit-to-unit and driven by unmodeled structure (geographic adjacency, social ties that cross cluster boundaries), the paper’s between-cluster estimands average over the wrong exposure distribution; the two-stage design therefore fails to identify the scientifically relevant spillover, and the proposed estimators are consistent for the wrong target. This modeling choice is introduced when the potential-outcome indexing and estimands are defined (framework section after the introduction) and is not relaxed by the subsequent asymptotic theory.","agreement_with_reader":"agree"},"referee_report":{"model":"grok-4.5","summary":"The paper studies causal inference under randomized saturation designs (RSDs) when interference may occur both within and between clusters. RSDs first randomize cluster-level treatment saturations and then randomize unit-level treatment within clusters. Prior work typically rules out between-cluster spillovers; this manuscript formulates potential outcomes that allow both within-cluster and between-cluster spillover effects, defines corresponding estimands, and develops design-based estimators with asymptotic normality and variance estimation. An application reanalyzes a cash-transfer RSD in Kenya for household expenditure. The central claim is that, under the stated exposure structure and two-stage randomization, the proposed within- and between-cluster spillover estimands are identified and the estimators support valid inference.","tokens_in":12708,"tokens_out":1212,"duration_ms":23611,"significance":"If the theory holds under the paper’s exposure mapping, the contribution is practically important: many field RSDs (cash transfers, public health, education) use geographic or administrative clusters that are not isolated, so ignoring between-cluster spillovers can misstate both direct and spillover effects. Extending the RSD toolkit beyond pure within-cluster interference is a clear gap relative to the existing literature. Strengths include an explicit potential-outcomes formulation, design-based identification from the two-stage randomization, and an empirical reanalysis rather than pure theory. The value of the contribution hinges on whether the between-cluster estimands, which appear to be indexed by saturations (or low-dimensional summaries of other clusters’ saturations), are scientifically relevant when true cross-cluster interference is unit-to-unit and network-driven.","major_comments":[{"comment":"The load-bearing modeling choice is how between-cluster interference enters the potential outcomes. From the framework after the introduction, potential outcomes appear to depend on own treatment and (within- and) between-cluster saturations, or a low-dimensional summary of other clusters’ saturations, rather than arbitrary unit-level cross-cluster links. When true spillovers are driven by geographic adjacency or social ties that cut across cluster boundaries, the paper’s between-cluster estimands average over the wrong exposure distribution: the two-stage design identifies those estimands, but they need not equal the scientifically relevant unit-to-unit spillover. The manuscript should state this exposure mapping as an explicit assumption, give conditions under which saturation-based exposures are adequate (e.g., exchangeability within distance bands), and discuss what is not identified","section":null},{"comment":"Relatedly, the Kenya cash-transfer application is presented as motivation and illustration, but the report of results should speak directly to whether between-cluster spillovers are substantively large relative to within-cluster effects and to pure no-interference analyses. If the reanalysis only shows that the method can be run, without comparing magnitudes, precision, or policy conclusions under alternative exposure mappings (e.g., distance-weighted neighbors vs. cluster-saturation summaries), the empirical section does not yet demonstrate that allowing between-cluster spillovers changes applied conclusions. A short sensitivity or alternative-exposure analysis would make the application load-bearing rather than decorative.","section":null},{"comment":"The asymptotic theory for estimation and inference is claimed under the two-stage design, but the regularity conditions for between-cluster dependence need to be stated carefully. With geographic proximity, dependence across clusters is not sparse in the usual cluster-independence sense; variance estimators that treat clusters as independent (or only weakly dependent through saturations) can understate uncertainty. The manuscript should clarify the dependence structure assumed for the CLT and variance estimation (e.g., mixing over space, fixed number of saturation levels with many clusters, or network sparsity) and whether the proposed variance estimator remains conservative under local cross-cluster dependence. Without that, the inference claim is incomplete for the leading geographic example in the abstract.","section":null}],"minor_comments":[{"comment":"The abstract and introduction correctly emphasize that existing RSD work assumes away between-cluster spillovers; a short related-work paragraph contrasting exposure mappings in the interference literature (e.g., partial interference vs. network interference) would help readers place the contribution.","section":null},{"comment":"Notation for saturations, within-cluster exposures, and between-cluster exposures should be introduced in one place and used consistently in estimand definitions and estimator formulas to avoid ambiguity between design probabilities and realized exposures.","section":null},{"comment":"In the application section, report sample sizes (clusters and units), the realized saturation design, and standard errors alongside point estimates so readers can assess precision of between-cluster effects.","section":null},{"comment":"Several passages in the extracted manuscript are hard to parse (garbled characters in the source dump); ensure the camera-ready PDF has clean equations, theorem statements, and table captions before resubmission.","section":null}],"recommendation":"major_revision","confidential_remarks":"The full-text extraction supplied for review is heavily corrupted (large stretches of replacement characters; an erroneous arXiv header pointing to astro-ph.SR appears in the dump). My assessment is therefore based on the abstract, readable fragments, and the design-based structure typical of this literature, plus the reader/stress-test notes. I am moderately confident about the conceptual contribution and the exposure-mapping concern; I am less confident about the fine print of the asymptotic proofs. If the journal can supply a clean PDF, a second pass on the theorems would be warranted. Scope fit for a methods journal in causal inference / statistics is good if the exposure assumption and variance theory are tightened."},"author_rebuttal":null,"desk_editor":{"model":"grok-4.5","letter":"This paper does something useful and overdue: it takes randomized saturation designs, which the literature has mostly analyzed under no between-cluster interference, and writes down potential-outcome estimands, estimators, and design-based asymptotics that allow both within- and between-cluster spillovers. That is a real methodological contribution for people who run or reanalyze two-stage cluster experiments, especially in development settings where geography or markets cross cluster boundaries.\n\nWhat it does well is stay inside the design-based tradition. Estimands are defined from the two-stage randomization, the theory claims consistency and asymptotic normality with estimable variances, and there is a Kenya cash-transfer reanalysis as a concrete application. The framing of the gap is accurate; prior RSD work largely assumed away the between-cluster piece. For readers who already work with interference under partial interference or exposure mappings, this is a natural next step rather than a rehash.\n\nThe soft spot is real but proportionate. The between-cluster estimands rest on potential outcomes that depend on own treatment and the vector (or low-dimensional summary) of other clusters’ saturations. If true cross-cluster interference is unit-to-unit and driven by unmodeled adjacency or social ties, those estimands average over the wrong exposure distribution and the estimators are consistent for the wrong scientific target. That is a modeling choice introduced when the potential outcomes are indexed; the subsequent asymptotics do not relax it. It is the standard exposure-mapping trade-off, not a hidden circularity or a broken proof. The corrupted full-text extract we saw also means I cannot personally check the variance formulas or the application tables, so confidence on the details stays moderate.\n\nWho it is for: causal-inference methodologists and applied people who already use RSDs and need language for between-cluster leakage. It is not field-reorganizing, but it is important within the subfield. I would send it to peer review; a serious referee can pressure-test the exposure mapping and the application. Worth engaging if you work on interference or experimental design under clustering.","headline":"Solid methods extension of RSD theory to between-cluster spillovers; the main soft spot is the saturation-only exposure mapping, not the design-based asymptotics.","tokens_in":13302,"tokens_out":507,"would_cite":true,"duration_ms":4782,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":["62K15","62G05"],"pacs":[],"model":"grok-4.5","headline":"Randomized saturation designs can identify both within-cluster and between-cluster spillover effects when units interact across clusters.","keywords":["randomized saturation design","spillover effects","within-cluster interference","between-cluster interference","potential outcomes","causal inference","two-stage randomization","cash transfer experiment"],"falsifier":"In a setting with known cross-cluster network ties, check whether the proposed between-cluster estimators recover the true spillover when interference depends on those specific ties rather than only on cluster saturations; systematic bias under that alternative would falsify the claim.","tokens_in":13396,"feed_emoji":"📊","tokens_out":574,"duration_ms":5433,"temperature":0.7,"pith_summary":"Randomized saturation designs first assign treatment probabilities to whole clusters and then treat units inside those clusters. Prior work typically estimated only within-cluster spillovers by assuming no interference between clusters. This paper argues that assumption is often false: when people or households interact across cluster boundaries, between-cluster spillovers exist and should be estimated rather than ignored. Using the potential-outcomes framework, the authors define clear within-cluster and between-cluster spillover estimands that are identified from the two-stage randomization, construct consistent and asymptotically normal estimators with estimable variances, and re-analyze a cash-transfer experiment in Kenya. The result matters because many field experiments are already run as saturation designs; the same data can now be used to recover both kinds of spillover without assuming clusters are isolated.","feed_headline":"Saturation designs can measure spillovers across clusters","feed_subtitle":"Two-stage randomization identifies both within- and between-cluster effects without assuming isolated groups","key_machinery":"Potential-outcomes indexing of units by their own treatment and by the saturation levels of their own and neighboring clusters, which yields identifiable within- and between-cluster spillover estimands whose estimation theory follows from the two-stage design.","core_discovery":"Under a potential-outcomes formulation that allows interference both inside and across clusters, the within-cluster and between-cluster spillover effects are identified by the two-stage randomization of a randomized saturation design; the corresponding estimators are consistent and asymptotically normal, and their variances can be estimated so that valid inference is possible.","pith_inferences":[],"forward_implications":[],"fun_headline_variants":["Within- and between-cluster spillovers identified by saturation designs","Two-stage designs estimate cross-cluster spillover effects","Randomized saturation designs capture both within and between spillovers","Estimating cluster spillover effects without isolation assumptions","Saturation designs measure interference inside and across clusters"],"cache_read_input_tokens":128,"weakest_assumption_plain":"Between-cluster interference is assumed to enter potential outcomes only through cluster saturation levels (or a low-dimensional summary of them), not through arbitrary unit-to-unit cross-cluster links.","fun_headline_variants_meta":{"raw":{"variants":["Within- and between-cluster spillovers identified by saturation designs","Two-stage designs estimate cross-cluster spillover effects","Randomized saturation designs capture both within and between spillovers","Estimating cluster spillover effects without isolation assumptions","Saturation designs measure interference inside and across clusters"]},"model":"grok-4.5","effort":"low","cost_usd":0.009534,"raw_usage":{"total_tokens":2099,"prompt_tokens":669,"num_sources_used":0,"completion_tokens":81,"cost_in_usd_ticks":95340000,"prompt_tokens_details":{"text_tokens":669,"audio_tokens":0,"image_tokens":0,"cached_tokens":256},"completion_tokens_details":{"audio_tokens":0,"reasoning_tokens":1349,"accepted_prediction_tokens":0,"rejected_prediction_tokens":0}},"tokens_in":669,"tokens_out":81,"duration_ms":10298,"temperature":1.0,"reasoning_tokens":1349,"cache_read_input_tokens":256,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-07-13T22:00:08.120599+00:00","model_set":{"reader":"grok-4.5"},"falsifier":"In a setting with known cross-cluster network ties, check whether the proposed between-cluster estimators recover the true spillover when interference depends on those specific ties rather than only on cluster saturations; systematic bias under that alternative would falsify the claim.","supporting_citations":[],"review_version":1}