{"id":"032cc1f8-ae35-42c9-b86e-46c0cdd63352","arxiv_id":"2501.04724","paper_version":1,"verdict":"REJECT","confidence":"HIGH","novelty_score":3.0,"correctness_risk":"high","formal_verification":"none","parameter_count":6,"one_line_summary":"Using LiNGAM on 922 breast cancer patients, the study estimates that adjuvant anti-HER2 therapy lengthens the time to last recurrence-free assessment by 169 days and skin/nipple involvement shortens it by 351 days.","lead":"This paper applies a causal discovery algorithm called LiNGAM to a breast cancer imaging dataset to estimate how treatments and tumor features affect the time to the last recurrence-free assessment. The authors report that anti-HER2 therapy adds about 169 days while skin or nipple involvement subtracts about 351 days, though the analysis ignores censoring and has no code released.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The headline effect rests on treating a censored follow-up time ('days to last recurrence-free assessment') as an uncensored linear outcome; under administrative or informative censoring, LiNGAM's regression-based estimates are biased and the 169/351-day effects are not causal.","rationale":"The reader identified the same load-bearing weak assumption, and I agree with that assessment. The paper's abstract and Results claim causal effects on local recurrence-free survival in days, but the variable actually modeled is 'days to last local recurrence free assessment,' which for a censored cohort is not a well-defined event time. The LiNGAM linear SEM assumes the observed outcome is a direct linear function of causes plus independent non-Gaussian noise; censoring or administrative follow-up makes the observed outcome a min(T,C) or an interval endpoint, so the independence tests and regression-asymmetry logic in Section 2.3 do not identify the structural coefficient. This alone invalidates the central numeric claims. The paper itself (Section 2.3) concedes that hidden common causes exist, and then Section 2.4 applies backdoor adjustment to a graph with latent confounding, which is a second unsupported step. The refutation p-values near 0.9 in Table 2 do not repair this, because they test stability of the biased estimator, not unbiasedness. The reader's rejection stands; my pass does not change it.","tokens_in":8875,"tokens_out":4527,"duration_ms":47375,"concrete_test":"Obtain the Duke dataset codebook and first determine the exact definition of 'Days to last local recurrence free assessment' along with the associated recurrence/event indicator; report the censoring proportion and whether the variable is administrative follow-up time rather than time-to-recurrence. Then reanalyze the same preprocessed data with standard survival models (Cox proportional hazards and an accelerated failure-time model) for Adjuvant Anti-Her2 Neu Therapy and Skin/Nipple involvement, using recurrence as the event and last follow-up as censoring, adjusted for the same covariates. If the survival-model estimates are null, reversed, or not significant, the LiNGAM 169/351-day claim is an artifact of censoring.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claim fails or becomes uninterpretable if the outcome is censored. The paper's target, 'Days to last local recurrence free assessment (from the date of diagnosis),' is a follow-up/assessment time, not a recurrence time: for patients with recurrence it is the last recurrence-free evaluation before the event, and for event-free patients it is administrative censoring. Section 2.3 feeds this variable directly into the linear SEM x_i = sum_{k(j)<k(i)} b_ij x_j + e_i, with no censoring or event indicator in the model. Censoring is a selection mechanism that induces dependence between the observed 'days' and both treatment and follow-up duration, so the LiNGAM/RCD pairwise regressions do not recover b_ij even if all other assumptions hold. The Table 1 coefficients (-169.638, 351.6813) are therefore not causal effects on local recurrence-free survival. A second, independent problem: Section 2.3 admits the no-hidden-common-cause assumption is violated in many pairs and switches to LiNGAM RCD, whose graph contains bidirected arrows, yet Section 2.4 applies the backdoor criterion and DoWhy as if the graph were a DAG with no unobserved confounding; latent confounders are not blocked by the chosen adjustment sets. The effect estimates are unsupported on both grounds.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The manuscript applies the LiNGAM family of causal discovery methods (DirectLiNGAM and LiNGAM RCD) to the Duke MRI Breast Cancer dataset to estimate causal effects of treatment and clinical features on the time to the last recurrence-free assessment. The authors report that Adjuvant Anti-Her2 Neu Therapy increases this time by 169 days and that Skin/Nipple involvement decreases it by 351 days. They attempt to validate these effects using the backdoor criterion and three refutation techniques implemented in DoWhy. The central methodological choices are to treat the outcome as an uncensored continuous variable in a linear non-Gaussian structural equation model, and to combine a graph learned with methods that allow hidden confounders with a backdoor adjustment procedure designed for DAGs.","tokens_in":9254,"tokens_out":5861,"duration_ms":53358,"significance":"If the results were valid, they would provide actionable causal estimates from observational data and could illustrate a useful workflow for other clinical datasets. The paper has several strengths: it uses a publicly available dataset, explicitly documents preprocessing steps, applies established causal discovery algorithms, and includes a suite of refutation checks. However, the central effect estimates are not credible because the outcome is a censored follow-up time that is modeled as an uncensored continuous variable, and because the backdoor analysis is applied to a graph that the authors themselves state contains bidirected edges representing latent common causes. These problems directly undermine the headline claims, so the paper does not currently contribute a trustworthy causal finding.","major_comments":[{"comment":"The outcome variable 'Days to last local recurrence free assessment (from the date of diagnosis)' is a censored time-to-event variable. For patients who experience a local recurrence, this is the time of the last recurrence-free assessment before the event; for patients who do not, it is the time of the last follow-up assessment (administrative censoring). The linear SEM in Section 2.3 treats this variable as an uncensored continuous outcome with no censoring indicator and no survival model. Censoring induces a selection mechanism: longer follow-up is associated with treatment choices and with event-free status, so the regression residuals are not independent of treatment even under the null. The LiNGAM coefficients reported in Table 1 (-169.638 and 351.6813) are therefore biased and cannot be interpreted as causal effects on local recurrence-free survival. The paper provides no justification for ignoring censoring and no sensitivity analysis that accounts for it.","section":"Section 2.3, Table 1"},{"comment":"The authors state that the absence-of-hidden-common-causes assumption is violated in many pairs and therefore switch from DirectLiNGAM to LiNGAM RCD, whose output contains bidirected edges representing latent common causes. In Section 2.4, they nonetheless apply the backdoor criterion and estimate effects with DoWhy 'with the causal structure from LiNGAM' as though the graph were a DAG. The backdoor criterion is defined only for directed acyclic graphs without bidirected edges; adjustment sets chosen from a graph with bidirected edges do not block paths through unobserved common causes. The reported backdoor estimates and the refutation results therefore do not support the claim that hidden confounding has been controlled, and the causal claims in Section 3 rest on an identification procedure that is invalid under the authors' own stated model.","section":"Sections 2.3 and 2.4"},{"comment":"Feature selection is performed with Lasso on the same outcome variable that is later used in the causal model, and no correction is made for the selection step. The selected features and the causal coefficients are estimated from the same data, so the reported effects are in-sample quantities with no clear uncertainty quantification. The 'Use a subset of data' refutation does not address this problem, because it refits the same selected model on a subset rather than repeating feature selection. As a result, the reported effect sizes may be substantially overfit and do not provide evidence of out-of-sample or replicable causal effects.","section":"Sections 2.2 and 3"}],"minor_comments":[{"comment":"The text cites 'Javidian and Valtorta (2018)', 'Hitchcock and Pearl (2001)', and 'Sharma and Kiciman (2020)', but none of these works appear in the reference list; full citations should be added.","section":"Section 2.4 and references"},{"comment":"The statement 'We already know from Lasso regression that the outcome variable is a linear combination of the extracted features' does not justify linearity, because Lasso assumes linearity rather than testing it; the correlation-matrix evidence is also weak. Linearity should be stated as an assumption or checked with a more direct diagnostic.","section":"Section 2.3"},{"comment":"The variable names include a 'No' suffix (e.g., 'Adjuvant Anti-Her2 Neu TherapyNo'), which makes the sign interpretation potentially confusing. The text says the therapy increased days by 169, while the table reports -169.638 for the 'No' level; please clarify the coding so that readers can verify the direction of each effect.","section":"Table 1 and Section 3"},{"comment":"The description of the feature-selection threshold is vague ('importance seems to be similar to Random 5'); the exact rule used to exclude features should be specified.","section":"Section 2.2"},{"comment":"Mean/mode imputation for missing clinical variables is known to be problematic for linear models; even if this is not a primary concern, the paper should discuss the potential bias introduced by single imputation.","section":"Section 2.1"},{"comment":"The text uses 'local recurrence-free survival' for what is actually 'days to last local recurrence-free assessment'; this terminology overstates the clinical endpoint and should be corrected or justified.","section":"Abstract and Section 3"}],"recommendation":"reject","confidential_remarks":"The methodological issues are fundamental and not local: the outcome is a censored time-to-event variable modeled as uncensored, and the backdoor analysis is applied to a graph with acknowledged hidden confounding. A proper revision would require a different outcome model (e.g., survival analysis) and a different identification strategy, which would change the paper's core results and conclusions. I do not see evidence of misconduct; the problems appear to stem from a mismatch between the clinical outcome and the causal-inference toolbox. The authors may benefit from collaboration with a statistician experienced in survival analysis and causal inference under hidden confounding."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Colleague,\n\nQuick take: the paper applies DirectLiNGAM/LiNGAM RCD to the Duke MRI breast cancer dataset and reports that no adjuvant anti-Her2 therapy reduces days to last local recurrence-free assessment by about 169 days, while skin/nipple involvement reduces it by about 351 days. The clinical findings are unsurprising — anti-HER2 therapy is known to benefit HER2-positive patients, and skin/nipple involvement is an established poor prognostic marker. The novelty is just a new application of an existing method.\n\nWhat the paper does well: it is transparent about its pipeline. It states the LiNGAM assumptions, runs Shapiro-Wilk to check non-Gaussianity, and explicitly reports that the no-hidden-common-cause assumption is violated in many pairs (Section 2.3), which is why the authors switch to LiNGAM RCD. The refutation suite (random common cause, subset, placebo) is a reasonable checklist. The writing is readable and the method description is accurate enough for a reader to see what was done.\n\nThe soft spots are not minor. First, the outcome variable is 'Days to last local recurrence free assessment (from the date of diagnosis)' — a follow-up/assessment time that is administratively censored for event-free patients and truncated by the event for others. Feeding this directly into a linear SEM as a continuous target, with no censoring indicator or survival model, means the regression coefficients are biased under any informative censoring. The 169- and 351-day effects are not causal effects on recurrence-free survival.\n\nSecond, the paper admits hidden common causes exist in many pairs and therefore uses LiNGAM RCD, which outputs bidirected arrows. But Section 2.4 then applies the backdoor criterion and DoWhy as if the graph were a DAG. Latent confounders are not blocked by the chosen adjustment sets. The backdoor refutations and p-values are therefore decorative; they compare estimates within the same misspecified model.\n\nThere are smaller problems: feature selection uses Lasso on the same outcome, which can bias subsequent causal estimates; mean/mode imputation is questionable for clinical variables; and no code or data are provided to reproduce the analysis. The stress-test note you shared holds up on reading.\n\nWho gets value from this paper? Possibly someone teaching a course on why causal discovery assumptions matter — it is a clean example of a pipeline that follows the recipe but ignores the outcome type. I would not cite it, and I would not bring it to a reading group. As a referee, I would reject: the central causal claims are unsupported, and the reanalysis needed (proper time-to-event handling, hidden confounding addressed) would be a different paper.\n\nRecommendation: desk reject, or reject after quick review. Not worth a serious referee's time beyond what you have already given.","headline":"A clear-eyed application of LiNGAM to a breast cancer dataset that collapses on a load-bearing flaw: the outcome is a censored follow-up time, not an uncensored linear target, and the backdoor analysis is run on a graph the authors themselves admit contains latent confounders.","tokens_in":9690,"tokens_out":1416,"would_cite":false,"duration_ms":15621,"reading_group":"no","serious_thinker":"yes","would_accept_peer_review":false},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"Applying LiNGAM causal discovery to a 922-patient breast cancer dataset, this paper estimates that adjuvant anti-Her2 therapy lengthens the last local recurrence-free assessment by 169 days on average, while skin/nipple involvement…","keywords":["breast cancer","causal inference","LiNGAM","DirectLiNGAM","local recurrence-free survival","adjuvant anti-Her2 therapy","skin-nipple involvement","observational data"],"falsifier":"Run a time-to-event analysis on the same 922-patient cohort, treating recurrence or death as the event and the last recurrence-free assessment as censoring, and compare Cox or competing-risk estimates for adjuvant anti-Her2 therapy and skin/nipple involvement; if the adjusted hazard ratio for the therapy is not clearly below 1, the claimed 169-day benefit is an artifact of treating censoring time as an outcome. A randomized comparison of Her2-positive patients with and without the therapy that fails to reproduce a roughly 169-day difference would also falsify the central claim.","tokens_in":8702,"feed_emoji":"🎗️","tokens_out":9291,"duration_ms":83150,"temperature":0.7,"pith_summary":"This paper argues that causal effects can be extracted from routine observational breast-cancer records, not just from randomized controlled trials. Using the Linear Non-Gaussian Acyclic Model (LiNGAM) on more than 40 features from 922 patients, it reports that adjuvant anti-Her2 therapy adds 169 days on average to the time until the last local recurrence-free assessment, while skin/nipple involvement removes 351 days. The authors present these as direct causal effects, and they claim the estimates survive refutation tests that add random common causes, use data subsets, and replace the treatment with a placebo. If the paper is right, clinicians gain quantitative, individualized guidance for which therapies matter and which anatomical presentations signal higher recurrence risk.","feed_headline":"Anti-Her2 therapy adds 169 recurrence-free days, model says","feed_subtitle":"Skin/nipple involvement costs 351 days; the two numbers come from causal analysis of 922 breast cancer patients.","key_machinery":"The key machinery is DirectLiNGAM, a causal-discovery algorithm for linear non-Gaussian acyclic structural equation models. Each variable $x_i$ is written as a linear combination of earlier variables $x_j$ plus an independent non-Gaussian error $e_i$; the algorithm finds the causal order by exploiting the fact that regressing an effect on its cause leaves a residual independent of the cause, while regressing in the reverse direction leaves a dependent residual unless errors are Gaussian. Once the ordering is fixed, adaptive Lasso removes weak connections. The paper also uses the LiNGAM RCD extension to allow hidden common causes, and checks the resulting causal effects with the backdoor criterion and three refutation techniques.","core_discovery":"The central claim is that the causal graph learned from the data contains a direct edge from Adjuvant Anti-Her2 Neu Therapy to the outcome 'Days to last local recurrence free assessment (from the date of diagnosis),' with an effect of +169 days, and a direct edge from Skin/Nipple Involvement with an effect of −351 days. The authors state these estimates in terms of the encoded variables: no skin/nipple involvement increases the interval by 351.68 days, and not receiving adjuvant anti-Her2 therapy decreases it by 169.64 days. They interpret the magnitude as clinically meaningful for Her2-positive patients and for patients with skin or nipple involvement, and they describe the refutation results as showing robustness to confounders and sampling variation.","pith_inferences":["The paper leaves implicit that its outcome is the last assessment date rather than a confirmed recurrence time; a survival-analysis companion would either confirm the 169-day estimate or reveal it as a censoring artifact.","The same causal-discovery pipeline could be applied to the twenty other breast cancer datasets cited in the paper to test whether these effect sizes replicate across institutions and imaging protocols.","Because the model estimates average effects, a natural next step is estimating conditional effects by receptor status, stage, and age, which could tell clinicians which Her2-positive patients gain the full 169 days and which high-risk patients need more than standard monitoring."],"forward_implications":["For Her2-positive patients, adjuvant anti-Her2 therapy should be a priority in treatment planning, since an average gain of 169 days to the last recurrence-free assessment is a concrete benefit.","Skin/nipple involvement is a prognostic marker for early recurrence: the 351-day average reduction points to patients who need intensified treatment or closer follow-up.","Causal discovery on observational data can complement RCTs, widening the evidence base for treatment decisions beyond narrowly selected trial populations.","The methodology, if correct, transfers to other observational oncology datasets and can estimate causal effects where trials are impractical or too expensive."],"supporting_citations":[{"why":"supplies the 922-patient cohort and the 40+ clinical, imaging, and treatment features used in every analysis.","marker":"[14]"},{"why":"introduces the LiNGAM linear non-Gaussian acyclic model that gives the paper its causal identification principle.","marker":"[17]"},{"why":"defines DirectLiNGAM, the specific algorithm used to learn the causal graph and estimate the reported effects.","marker":"[18]"},{"why":"provides the acyclicity and conditional-independence background that motivates the causal-discovery assumptions.","marker":"[19]"},{"why":"supplies the clinical evidence on nipple-sparing mastectomy and local recurrence used to frame skin/nipple involvement as a high-risk feature.","marker":"[3]"}],"fun_headline_variants":["Her2 therapy boosts recurrence-free time by 169 days, causal model finds","Skin/nipple involvement cuts recurrence-free time by 351 days: causal analysis","Causal AI: Her2 therapy +169 days, skin/nipple -351 days","From 922 patients: Her2 therapy adds 169 days, skin involvement costs 351"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The load-bearing premise is that 'Days to last local recurrence free assessment (from the date of diagnosis)' is an uncensored continuous outcome that enters a linear non-Gaussian causal model directly; if the real endpoint is censored time-to-recurrence or responds nonlinearly, the 169-day and 351-day causal estimates are biased and the paper's causal claims collapse.","fun_headline_variants_meta":{"raw":{"variants":["Her2 therapy boosts recurrence-free time by 169 days, causal model finds","Skin/nipple involvement cuts recurrence-free time by 351 days: causal analysis","Causal AI: Her2 therapy +169 days, skin/nipple -351 days","From 922 patients: Her2 therapy adds 169 days, skin involvement costs 351"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000749,"raw_usage":{"total_tokens":3283,"prompt_tokens":838,"completion_tokens":2445,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":454,"completion_tokens_details":{"reasoning_tokens":2355}},"tokens_in":454,"tokens_out":2445,"duration_ms":19046,"temperature":1.0,"reasoning_tokens":2355,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-10T22:14:12.483394+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Run a time-to-event analysis on the same 922-patient cohort, treating recurrence or death as the event and the last recurrence-free assessment as censoring, and compare Cox or competing-risk estimates for adjuvant anti-Her2 therapy and skin/nipple involvement; if the adjusted hazard ratio for the therapy is not clearly below 1, the claimed 169-day benefit is an artifact of treating censoring time as an outcome. A randomized comparison of Her2-positive patients with and without the therapy that fails to reproduce a roughly 169-day difference would also falsify the central claim.","supporting_citations":[{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"supplies the 922-patient cohort and the 40+ clinical, imaging, and treatment features used in every analysis."},{"cited_title":"Shimizu, P","cited_arxiv_id":null,"evidence_quote":"introduces the LiNGAM linear non-Gaussian acyclic model that gives the paper its causal identification principle."},{"cited_title":"Di- rectlingam: A direct method for learning a linear non- gaussian structural equation model","cited_arxiv_id":null,"evidence_quote":"defines DirectLiNGAM, the specific algorithm used to learn the causal graph and estimate the reported effects."},{"cited_title":"Spirtes, C","cited_arxiv_id":null,"evidence_quote":"provides the acyclicity and conditional-independence background that motivates the causal-discovery assumptions."},{"cited_title":"De La Cruz, A","cited_arxiv_id":null,"evidence_quote":"supplies the clinical evidence on nipple-sparing mastectomy and local recurrence used to frame skin/nipple involvement as a high-risk feature."}],"review_version":1}