{"id":"84f700ab-5e0b-44bf-9aaf-accf3af2f431","arxiv_id":"2412.02439","paper_version":1,"verdict":"CONDITIONAL","confidence":"HIGH","novelty_score":5.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":5,"one_line_summary":"Using causal machine learning on IllustrisTNG, environment is estimated to suppress star formation by up to ~100x at z=0 but to boost it by ~10x at z~1 and more at higher redshifts.","lead":"This study applies causal machine learning to the IllustrisTNG simulation to separate the effect of a galaxy's environment on its star formation rate from the effect of its internal properties. It finds that dense environments suppress star formation by up to about a factor of 100 today, but they actually boost it in the early universe, and it argues that common snapshot-based controls are inadequate.","discovery_kind":"new_application","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Causal estimates rely on the untested DAG; host halo mass is a plausible omitted confounder that could change the headline suppression and reversal.","rationale":"The reader identified the DAG/no-unobserved-confounder assumption as the weakest point; my stress-test agrees and sharpens it to a specific, testable omitted variable: host-halo mass (and, more broadly, large-scale environment). This is the single most load-bearing concern because the paper's entire contribution is the causal interpretation of the SFR-density relation. If this assumption fails, the headline numbers are biased and the sign reversal could be an artifact of omitted confounding rather than a genuine causal effect. The paper's internal diagnostics (weight balance, external consistency with prior work) provide supporting evidence but cannot rule out omission of a variable that is correlated with both the treatment and the outcome; the observed balance on subhalo mass does not guarantee balance on unobserved or omitted variables. The proposed test is feasible because the simulation data is complete and the omitted variables (host-halo mass, large-scale overdensity) are publicly available in the TNG catalog and merger trees. The reader's 'CONDITIONAL' verdict remains appropriate: the scientific advance is real and well-presented, but the causal claims should be conditional on passing this sensitivity check and ideally on eventual code/data release so others can perform it. No change to the verdict is needed, but the conditionality should be explicit about the omitted-confounder sensitivity analysis.","tokens_in":43069,"tokens_out":9829,"duration_ms":101712,"concrete_test":"Re-estimate the causal dose-response curves including the full host-halo mass history (e.g., FoF M200 along the main branch) as an additional time-varying covariate in the denominator of Eq. 3, alongside the subhalo mass history. If the z=0 CDRC changes by more than ~30% in the maximal suppression factor, or if the positive enhancement at z~1-3 vanishes or reverses sign, then the DAG's claimed sufficient adjustment set is false and the headline causal effect is not identified. A complementary run conditioning on large-scale overdensity (e.g., density smoothed at 10 Mpc) would test the same concern for cosmic-web-scale confounding.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claim that environment has a causal effect of ~100x suppression at z=0 and a reversal at z>1 rests entirely on the correctness of the DAG in Fig. 2a and the resulting claim that adjusting for subhalo (halo) mass history is sufficient to eliminate confounding (Section 2.1, Appendix D). The paper acknowledges the no-unobserved-confounder assumption is untestable, but the simulation itself contains variables that are omitted from the DAG yet plausibly cause both the 10th-nearest-neighbor density (treatment) and SFR (outcome). In particular, host-halo (FoF) mass is strongly correlated with environment (Extended Data Fig. 1) and independently drives environmental processes such as ram-pressure stripping and tidal forces that directly affect SFR. If host-halo mass (or large-scale overdensity, tidal anisotropy, or the broader cosmic web) is a common cause of E and SFR, sequential ignorability fails and the IPW weights in Eq. 3 do not remove confounding. The reported factors (~100 suppression, ~10 enhancement) would then be biased, not causal. The 'causal model (stellar mass)' comparison in Fig. 4c does not settle this, because stellar mass history is an effect of the treatment and outcome, not an independent confounder; substituting it for halo mass does not test the DAG's sufficiency.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"This paper applies the causal-inference framework to the IllustrisTNG100-1 simulation to estimate the causal effect of the local environment (10th-nearest-neighbour density) on the star-formation rate (SFR) of galaxies. The authors build a structural causal model from semi-analytic galaxy-formation theory, identify halo mass as the sole time-varying confounder, and estimate causal dose-response curves using inverse probability weighting of marginal structural models with random forests. They report that at z=0 the environment suppresses the average SFR by up to ~100x, that the effect reverses sign at z≳1 (boosting SFR by ~10x at z~1 and more at higher z), that ignoring halo mass underestimates the effect by ~2x in intermediate densities, that conditioning on a single snapshot of stellar mass is insufficient and adverse, and that stellar mass history is an adequate observational proxy for halo mass in the causal model.","tokens_in":43353,"tokens_out":5851,"duration_ms":59749,"significance":"If the causal interpretation holds, this is a substantive step forward: it replaces correlational SFR-density relations with a quantitative causal curve, provides a physically motivated DAG for the nature-nurture problem, and proposes stellar mass history as a viable observational proxy. The paper is methodologically careful in several respects: the estimation pipeline is described step-by-step, covariate balance and weight diagnostics are reported (Figs. D7-D8), confidence intervals are obtained by bootstrap of the full pipeline, and the untestability of the no-unobserved-confounding assumption is acknowledged explicitly in Appendix D. The qualitative agreement with the independent TNG300 analysis of Hwang et al. adds credibility. The main limitation is that the headline numbers rest entirely on the correctness of the assumed DAG; the manuscript does not yet provide any quantitative sensitivity analysis to unobserved confounding, which is standard when a claim of 100x causal effect is made.","major_comments":[{"comment":"The claim that the adjustment set consisting solely of the halo-mass history is sufficient hinges on the untestable no-unobserved-confounder assumption. The simulation itself contains an obvious candidate common cause that is omitted from the DAG: the host (friends-of-friends) halo mass. Extended Data Fig. 1 shows that the 10th-nearest-neighbour density is strongly correlated with host halo mass, and ram-pressure stripping, tidal stripping, and strangulation—processes that directly suppress SFR—are driven or modulated by the host halo mass rather than by the subhalo mass tracked in the DAG. Since the subhalo mass history is itself affected by the environment (tidal stripping), it is not a valid proxy for the host-mass process. Concretely, I would like to see either (i) the CDRCs recomputed with the host-halo mass (or its history) added to the weighting model, or (ii) a quantitative sensitivity analysis that reports how strong an unobserved confounder would have to be to overturn the ~100x suppression and the high-redshift reversal. The authors' statement in Appendix D that it is 'difficult to think of a physical process or variable' is not a sufficient response to this concrete alternative.","section":"2.1, Fig. 2a, Appendix D"},{"comment":"The positivity/overlap support for the continuous treatment is not demonstrated in the region where the headline effect is largest. The weight distributions in Fig. D8 have means near one, but the mean of a stabilized weight distribution can hide a small number of extremely large weights (the authors trim at the 1st and 99th percentiles precisely because extreme weights arise). Because the highest-density bins contain few galaxies and the treatment grid extends to the 99th percentile, the ~100x suppression at log(Σ10) ≳ 2.5 may be determined by a handful of units with very large weights. Please report the effective sample size before and after trimming, the maximum weights, the distribution of weights in the highest-density treatment bin, and the sensitivity of the CDRC to the trimming thresholds (e.g., 0.5th/99.5th and 5th/95th). If the estimate is not robust to these choices, the headline magnitude should be re-scaled or qualified.","section":"5.2, Appendix D, Fig. D8"}],"minor_comments":[{"comment":"In the definition of the environmental history, the text says 'N is the number of treatments and j = k', but the sum runs from k=0 to j, which contains j+1 terms. Please clarify whether N = j+1 or define the normalization accordingly.","section":"Eq. (1)"},{"comment":"The text contains the typo 'eF AM' for the eFAM method; please correct it.","section":"Section 3.3.3"},{"comment":"The first sentence of Appendix C contains the typo 'enviroment' for 'environment'; please correct it.","section":"Appendix C"},{"comment":"Please report the numerical values of the average absolute correlation coefficients (AACC) in the text, not only in Fig. D7, and state explicitly whether the post-weighting values fall below the 0.1 threshold cited from [304].","section":"Appendix D"},{"comment":"The statements that data and code 'will be made available upon request' should be replaced by a persistent repository link, which is the standard expectation for reproducibility in this journal.","section":"Data and code availability"}],"recommendation":"major_revision","confidential_remarks":"The gap between the strong quantitative causal claims (factor ~100 suppression and sign reversal) and the absence of any sensitivity analysis to unobserved confounding is the central issue. The manuscript is transparent about the untestable assumption, but for a journal-level causal claim, a concrete robustness check (e.g., adding host FoF mass to the weighting model, or a formal sensitivity bound) is required. I see no indication of citation manipulation or scope mismatch; the paper fits the astro-ph.GA readership. Given the careful diagnostics and external consistency with TNG300, the result is likely salvageable, hence major revision rather than rejection."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"The paper does something genuinely new: it treats the SFR–density relation as a causal question and applies inverse-probability weighting of marginal structural models, with a DAG built from semi-analytic theory, to TNG100 galaxies tracked along merger trees. That is a real step beyond the usual mass-matched or binned comparisons, and the estimation is careful—covariate balance, weight diagnostics, and bootstrapped intervals are all there. The qualitative result, a suppression at z=0 and a reversal at z>1, is also consistent with the independent TNG300 study by Hwang et al., which gives me some confidence it is not purely a methodological artifact.\n\nThe soft spots are real, though. The central numbers—the factor ~100 suppression and the factor ~10 boost—rest on the untestable assumption that the DAG is correct. The most obvious candidate for an omitted confounder is host-halo (FoF) mass, which the paper itself mentions as a driver of ram-pressure stripping and tidal effects. If host-halo mass causes both the local density and the galaxy's SFR, then the IPW weights in Eq. 3 do not remove that bias. The paper acknowledges the no-unobserved-confounder assumption in Appendix D but does not quantify how much the results would change if host-halo mass or large-scale overdensity were added to the weighting model. That is the central check. Also, the 'for the first time in the overall context of galaxy formation' claim is overstated given the cited work by McGibbon & Khochfar and Bluck et al., and code/data are only available on request, which hampers replication.\n\nI would send this to peer review. The question is important, the method is appropriate, and the qualitative reversal is corroborated. The referees should ask for a sensitivity analysis that includes host-halo mass or a large-scale density variable in the adjustment set, a toned-down novelty claim, and a commitment to release code and data. If those are addressed, the paper makes a useful contribution.","headline":"A serious causal-inference treatment of the SFR–density relation with a plausible central result, but the weight-bearing DAG and the omitted host-halo mass need a sensitivity analysis before the causal magnitudes are taken at face value.","tokens_in":43854,"tokens_out":3845,"would_cite":true,"duration_ms":35086,"reading_group":"yes","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"Dense environments suppress galaxies' star formation by a factor of 100 at z=0, while at z>1 they boost it.","keywords":["galaxy formation and evolution","star formation rate","environment","causal inference","inverse probability weighting","marginal structural models","IllustrisTNG","nature versus nurture"],"falsifier":"Compute the same stabilized IPW weights from the same DAG but with placebo confounder histories (e.g., shuffled halo-mass histories): if the causal dose-response curves shift by more than the bootstrap uncertainty, the adjustment set is not actually removing confounding. Alternatively, estimate the curves on an independent large-volume cosmological simulation: if the z=0 suppression or the z>1 reversal disappears, the claim is specific to TNG100-1 rather than to galaxy formation.","tokens_in":42866,"feed_emoji":"🌌","tokens_out":5651,"duration_ms":57853,"temperature":0.7,"pith_summary":"The paper aims to settle whether galaxy formation is driven by 'nature' (internal, halo-mass-driven processes) or 'nurture' (external environment), and to estimate the causal, not merely correlational, effect of environment on star-formation rate. Using causal machine learning on 18,629 galaxies from the IllustrisTNG simulations traced from z~6 to z=0, the authors find that dense environments suppress star formation by up to a factor of about 100 at z=0, while at z>1 the same kind of environment boosts star formation by a factor of about 10 at z~1 and more at higher redshifts. They also show that ignoring halo mass (nature) underestimates the environmental effect in intermediate-density environments by a factor of about 2, and that the common practice of controlling for snapshot stellar mass is not only insufficient but actively harmful, whereas stellar-mass history is an adequate proxy for nature. The work matters because it offers a framework for extracting causal statements from data with feedback loops, a problem that extends beyond galaxies to any evolving system.","feed_headline":"Dense environments quench star formation by 100x","feed_subtitle":"Causal analysis of 18,629 simulated galaxies shows the effect flips: at z>1 dense environments boost star formation.","key_machinery":"The load-bearing object is a hand-built causal directed acyclic graph (DAG) of galaxy formation and evolution, in which halo mass $H_k$ is the time-varying confounder of environment $E_k$ (treatment) and star-formation rate $\\mathrm{SFR}_k$ (outcome), with treatment-confounder feedback. The estimation machinery is inverse probability weighting of marginal structural models (IPW of MSMs), with generalized propensity scores learned by random forests; the stabilized weights $w_j = \\prod_{k=0}^{j} f(E_k|\\bar{E}_{k-1})/f(E_k|\\bar{E}_{k-1},\\bar{H}_{k-1})$ create a pseudo-population in which treatment is independent of confounders, and weighted outcome models produce causal dose-response curves. The DAG is what makes halo mass the sufficient and necessary adjustment set; without it, conditional adjustment either leaves confounding or introduces over-adjustment bias.","core_discovery":"The central discovery is that the SFR-density relation is genuinely causal and time-dependent: at z=0 environment quenches star formation, with the average SFR falling by a maximal factor of ~100 as density increases, while at z≳1 the causal effect reverses and dense environments accelerate star formation by a factor of ~10 at z~1 and by larger factors out to z~3. The reversal is not a side effect of massive galaxies living in dense regions; it survives adjustment for halo mass and is interpreted as environment-driven accelerated evolution, connecting to galaxy downsizing. The paper further establishes three methodological results: halo mass (nature) contributes causally and ignoring it biases the environmental effect low by ~2 in intermediate densities; conditioning on stellar mass at a single snapshot fails and is worse than no adjustment, because stellar mass can be a collider; and stellar-mass history recovers the halo-mass adjustment, making the causal effect estimable with observable quantities.","pith_inferences":["A natural next test is to run the same estimator on another large-volume cosmological simulation; if the sign reversal and the ~100 suppression are not reproduced, the causal DAG rather than the physics would be implicated.","Splitting the sample into central and satellite galaxies and estimating separate dose-response curves could reveal whether the flattening at the highest densities is a central-galaxy artifact, which the authors themselves flag as future work.","The authors' weight diagnostic could be turned into a direct falsification: if removing later halo-mass history from the denominator changes the z=0 curve substantially, the DAG's assumed lag structure is misspecified.","Applied to observational surveys, the framework suggests that only surveys with reconstructed stellar-mass histories, not single-epoch mass-matched samples, can recover environmental causality."],"forward_implications":["At z=0, environment's causal effect is negligible below $\\log \\Sigma_{10}\\sim1$, then becomes negative, saturating or weakening in the densest regions; the maximal suppression is a factor of ~100.","At z>~1 the causal SFR-density relation reverses: the same environments that quench locally boost star formation, with the effect growing with redshift to factors above 100 by z~3.","Snapshot stellar-mass control, the standard literature approach, does not separate nature and nurture and can induce selection bias instead of removing confounding.","Stellar-mass history is a sufficient observational proxy for halo mass, so the causal effect can in principle be estimated on real galaxies with reconstructed histories.","Because the framework handles feedback loops, the same causal-model-plus-IPW strategy can be transferred to other dynamical systems, such as the Earth's climate."],"supporting_citations":[{"why":"Supplies the IllustrisTNG simulation whose galaxy sample is analyzed.","marker":"[104]"},{"why":"Provides the TNG100-1 public data release and merger trees used to trace galaxies.","marker":"[109]"},{"why":"Supplies the inverse-probability-weighting of marginal structural models for time-varying treatments.","marker":"[134]"},{"why":"Provides the d-separation criterion that identifies halo mass as the sufficient adjustment set.","marker":"[131]"},{"why":"Supplies the random-forest algorithm used for propensity-score estimation and outcome modeling.","marker":"[141]"},{"why":"Shows that snapshot stellar-mass binning is insufficient, motivating the need for histories.","marker":"[57]"},{"why":"Gives the simulated reversal of the SFR-density relation at high redshift that the paper reproduces.","marker":"[161]"}],"fun_headline_variants":["Environment quenches galaxies 100x, but flips at high z","Causal AI shows environment flips galaxy star formation","Nature vs nurture: environment quenches 100x, boosts at z>1","Causal ML: Dense environs quench star formation, then boost","Environment causally quenches SFR 100x, flips to boost at z>1"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The entire causal estimate rests on the hand-built causal diagram being correct: halo mass is the only time-varying confounder, and no unobserved variable causes both environment and star formation; the paper itself states this is untestable.","fun_headline_variants_meta":{"raw":{"variants":["Environment quenches galaxies 100x, but flips at high z","Causal AI shows environment flips galaxy star formation","Nature vs nurture: environment quenches 100x, boosts at z>1","Causal ML: Dense environs quench star formation, then boost","Environment causally quenches SFR 100x, flips to boost at z>1"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000669,"raw_usage":{"total_tokens":3117,"prompt_tokens":1078,"completion_tokens":2039,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":694,"completion_tokens_details":{"reasoning_tokens":1935}},"tokens_in":694,"tokens_out":2039,"duration_ms":15160,"temperature":1.0,"reasoning_tokens":1935,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-11T23:26:51.031531+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Compute the same stabilized IPW weights from the same DAG but with placebo confounder histories (e.g., shuffled halo-mass histories): if the causal dose-response curves shift by more than the bootstrap uncertainty, the adjustment set is not actually removing confounding. Alternatively, estimate the curves on an independent large-volume cosmological simulation: if the z=0 suppression or the z>1 reversal disappears, the claim is specific to TNG100-1 rather than to galaxy formation.","supporting_citations":[],"review_version":1}