{"id":"11247ee4-797f-4697-8c44-ef2fc41f953f","arxiv_id":"2505.00310","paper_version":2,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":5.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":1,"one_line_summary":"Pretraining the R-learner by using the outcome model's active set to weight penalties in the CATE lasso reduces error and raises power when prognostic and predictive factors share support.","lead":"This paper proposes a way to borrow information from the risk (prognosis) model when estimating heterogeneous treatment effects: features that predict the outcome are given weaker penalties when fitting the treatment-effect model. In simulations with overlapping predictive features, this lowers estimation error and increases power, but the benefit depends on overlap actually being present.","discovery_kind":"new_application","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The paper's only safeguard against harmful pretraining is R-loss tuning of α, yet no theory or failure-mode analysis shows that the R-loss selects a safe α when nuisance fits are poor; in the high-dimensional, low-SNR regimes the method targets, this safeguard is untested.","rationale":"The reader's weakest assumption identifies the same core issue: the method relies on a noisy outcome-model active set and on R-loss tuning to prevent harm from pretraining. My stress-test focuses on the second half of that assumption—the R-loss as a selector of α—because it is the only mechanism that prevents active-set noise from translating into over-penalization of true effect modifiers. The paper's own Discussion flags this dependence, and the simulations cover mostly favorable settings (n=500, p=150, SNR=2) rather than the high-dimensional, low-SNR regimes the introduction motivates. Table 2 provides concrete evidence of occasional degradations. This does not refute the central claim in its intended 'synergies present' settings, but it does mean the claim requires the conditional qualification that the R-loss tuning actually works; the current evidence is simulation-only. Since the reader already returned CONDITIONAL, this concern does not change the verdict, but it sharpens the required condition: code release and a systematic failure-mode analysis of R-loss tuning should be part of the revision.","tokens_in":22144,"tokens_out":4462,"duration_ms":48888,"concrete_test":"Rerun the §3.1 DGP in the fully disjoint-support and weak-signal regimes (p=500, n=200, SNR=0.5, Toeplitz correlation 0.5), and record the α chosen by cross-validated R-loss versus the α minimizing test MSE over 200 replications. If the median selected α is below 0.9 in the disjoint case, or if the R-loss argmin is more than 0.1 away from the MSE argmin in a partially shared setting, the tuning safeguard is broken and the headline claim needs qualification.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claim is that pretraining the R-learner with the outcome model's lasso active set (Eq. 10) improves CATE estimation when supports overlap. For this to hold, two conditions must hold together: (i) the active set reliably identifies prognostic features, and (ii) the R-loss, used in Algorithm 1 to tune α, is a valid proxy for test MSE. The paper provides simulation evidence for (ii) in favorable settings (n=500, p=150, SNR=2, Figures 2 and 4), but no theory. The Discussion explicitly concedes that 'if the estimation error for nuisance parameters is large, the R-loss becomes less reliable for such hyperparameters.' In the high-dimensional, weak-signal settings the paper motivates (p≥n, SNR=0.5), cross-fitted lasso nuisance estimates are noisy and the active set is unstable—the paper itself introduces uniR-lasso for this reason in §3.3. If the R-loss surface is flat or has a minimum at an α that over-penalizes true effect modifiers, pretraining will degrade MSE and power rather than improve them. Table 2 already shows several configurations where +pt performs worse than base (e.g., Setup A, n=500, p=30, SNR=0.5: 1.04 vs 0.96), so the method does not uniformly dominate even in designed settings.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes a 'pretraining' strategy for the R-learner: first fit a lasso for the mean outcome m(X), then use its active set to construct feature-specific penalty weights (Eq. 10) in an adaptive-lasso fit of the CATE tau(X), with a tunable pretraining strength alpha selected by the R-loss (Algorithm 1). The motivating premise is that prognostic factors often overlap with predictive factors. The authors demonstrate the approach in linear settings, compare it with a univariate-guided variant (uniR-lasso), extend it to basis expansions and generalized random forests, and evaluate MSE, support recovery/FDR, and TOC-based power in simulations. A DR-learner extension appears in the supplement. The paper includes no formal theory and no real-data application, and some configurations in Table 2 show pretraining degrading performance even when support overlap is present.","tokens_in":22395,"tokens_out":10600,"duration_ms":108127,"significance":"The central idea is practically relevant and clearly presented. If the gains under shared support held in the high-dimensional, low-SNR regimes that motivate the paper, the method would be a useful addition to the CATE toolbox. The favorable linear simulations (Figures 3-6) are extensive, and the construction is not circular: the penalty weights come from the outcome model m(X), not from the CATE target tau(X), while the R-loss is used only as a tuning criterion. The extension to uniR-lasso, basis expansions, and GRF is a reasonable broadening of the contribution. However, the evidence is mixed: the nonlinear benchmarks in Table 2 contain cases where pretraining increases MSE despite designed support overlap, and the Discussion concedes that the R-loss safeguard can fail when nuisance fits are poor. A theoretical guarantee is not required for a simulation-based paper, but the counterexamples make it important to characterize when pretraining is safe. The current abstract-level claim is broader than the evidence supports.","major_comments":[{"comment":"In Setup A, tau(X) is linear in X1-X5 and m(X) is a nonlinear function of exactly X1-X5, so the support overlap the method is designed to exploit is present. Yet the pretrained R-lasso reports higher normalized MSE than the base R-lasso in several rows, for example n=500, p=30, SNR=0.5 (1.04 vs 0.96) and n=1000, p=30, SNR=0.5 (0.64 vs 0.48). These results contradict the abstract's claim that 'in settings where these synergies are present' the method yields lower estimation error, and they show that the R-loss-based selection of alpha does not always prevent harmful pretraining even with maximal support overlap. The paper should either characterize the regimes in which pretraining is safe or substantially qualify the central claim.","section":"Table 2, Setup A"},{"comment":"The Discussion explicitly states that 'if the estimation error for nuisance parameters is large, the R-loss becomes less reliable for such hyperparameters.' This is the key safeguard in Algorithm 1, but the paper offers no simulation or theory showing how often this failure occurs in the p>=n, SNR=0.5 regimes used to motivate the work. In those regimes cross-fitted lasso nuisance estimates are noisy, and the active set in Eq. (10) is unstable; Table 2 provides configurations, including those cited above, where the R-loss-selected alpha appears to over-penalize true effect modifiers. Without a failure-mode analysis or a formal condition under which the R-loss surface has a safe minimum, the main claim rests on an untested safeguard.","section":"Section 5 / Algorithm 1"},{"comment":"The penalty weights are constructed from the lasso active set of m(X), and Section 3.3 itself acknowledges that the lasso is unreliable for support recovery under correlated features, introducing uniR-lasso as a remedy. The main Algorithm 1 and the headline simulations of Sections 3.1-3.2 nonetheless use the plain lasso active set, and no sensitivity analysis is reported with respect to the choice of lambda1 for m(X), the degree of covariate correlation, or the strength of the outcome signal. Because the active set is the only channel through which prognostic information enters the CATE model, the paper should directly demonstrate how false positives and false negatives in this set affect the pretrained estimator.","section":"Eq. (10) / Section 3.3"}],"minor_comments":[{"comment":"The table reports point estimates only, without standard errors, confidence intervals, or repetition counts, even though Table 1 includes intervals. The reporting format should be made consistent so that the reader can assess whether the observed degradations are systematic or noise.","section":"Table 2"},{"comment":"The algorithm says alpha and lambda2 are chosen by 'cross-validated R-loss' but does not specify the grid or whether alpha selection uses a separate validation split from the data used for the final tau fit. Figure 2's 'Train Set R-Loss' suggests in-sample tuning, so the exact procedure should be stated for reproducibility.","section":"Algorithm 1"},{"comment":"The column headings such as '+xgb(m)' and '+xgb(m,tau)' are not self-contained; the caption should define each model explicitly, especially because Figure 9 uses a different ordering of the same models.","section":"Table 2 caption"},{"comment":"The caption should state the fixed values of the non-varying parameters in each panel (p for the left panel and n for the right panel), since the text mentions only the varying quantity.","section":"Figure 6 caption"},{"comment":"The abstract's claim of 'reduced false discovery rates' is supported in Figure 7 by FDR on coefficient support recovery, not by an inferential procedure for detecting effect modifiers. The text should clarify this distinction.","section":"Abstract / Figure 7"}],"recommendation":"major_revision","confidential_remarks":"The paper has a plausible and well-motivated idea, and the favorable-setting simulations are convincing. The obstacles to acceptance are the Setup A counterexamples in Table 2 and the lack of analysis of when the R-loss tuning safeguard fails. These are fixable: the authors could add a targeted simulation study varying correlation, outcome signal strength, and nuisance estimation quality, and align the abstract and Section 5 with the observed boundary of usefulness. A real-data application would also strengthen the applied claims, though I would not make it a prerequisite. If the authors can either show that the harmful configurations are identifiable ex ante or provide a clear fallback rule, I would support publication."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Two things to know. First, this is a genuinely useful methodological contribution: it takes the adaptive-lasso idea from prediction (Craig et al.) and ports it to CATE estimation by using the outcome model's lasso active set to construct penalty weights in the R-learner, then tunes the pretraining strength alpha on the R-loss. The extensions to generalized random forests and xgboost basis expansions are natural and broaden the scope. Second, the paper is honest about its limits, but those limits are real: there is no theory, and the safety of the R-loss tuning for alpha is established only by simulation in favorable settings.\n\nWhat the paper does well: the simulations are reasonably extensive. The linear shared-support experiments show clear gains in MSE and power when overlap is present, and the paper reports settings where pretraining hurts (e.g., Table 2, Setup A, n=500, p=30, SNR=0.5), which is a point in its favor. The discussion explicitly concedes that the R-loss becomes less reliable when nuisance estimation error is large. That is the right frame.\n\nThe soft spots. The key mechanism, Equation (10), is essentially adaptive lasso weighting, and the only safeguard against over-penalizing true effect modifiers is the R-loss as a proxy for test error. The stress-test worry about low-SNR, high-dimensional regimes is not fully answered: although Table 1 includes p=n and SNR=0.5 and shows gains, there is no analysis of when the selected alpha is wrong or how flat the R-loss surface is. Table 2 does not report uncertainties, so we cannot tell whether some apparent gains are noise. The absence of code and any real-data application also makes it hard to judge practical reliability. These are proportionately minor-to-moderate issues for an applied statistics paper, not fatal flaws.\n\nWho is this for? People building CATE estimators in biomedical or policy settings where prognostic and predictive factors plausibly overlap. They will get a clear recipe and some evidence it works in designed settings. The paper deserves a serious referee, but I would send it back with requests: release code, add error bars or intervals to the main tables, and add either a real-data example or a simulation where the R-loss is deliberately misled (e.g., poor nuisance fit, correlated design, weak signal) to show how often pretraining helps or hurts.","headline":"A useful, honest transfer-learning recipe for the R-learner, with convincing simulations when supports overlap and a thin theory/safety case; deserves a careful referee.","tokens_in":22936,"tokens_out":2730,"would_cite":true,"duration_ms":26861,"reading_group":"yes","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":["62D20","62J07"],"pacs":[],"model":"deepseek-v4-flash","headline":"Pretraining a treatment-effect model on outcome-model side information improves CATE estimation when prognostic and predictive factors overlap.","keywords":["heterogeneous treatment effects","conditional average treatment effect","R-learner","pretraining","shared support","adaptive lasso","false discovery rate","prognostic and predictive factors"],"falsifier":"In a simulation with known fully disjoint support between $m(X)$ and $\\tau(X)$, strong covariate correlation, and weak outcome signal, check whether the R-loss-selected $\\alpha$ falls well below 1 or the pretrained model's test MSE exceeds the baseline R-lasso's; if either happens, the claim that pretraining only helps and never hurts under R-loss tuning is falsified.","tokens_in":21922,"feed_emoji":"🎯","tokens_out":8369,"duration_ms":72553,"temperature":0.7,"pith_summary":"This paper claims that heterogeneous treatment effect (CATE) estimation can be improved by pretraining the effect model with side information from the outcome model, exploiting the common real-world overlap between prognostic and predictive factors. Within the R-learner framework, the authors fit the mean outcome with a lasso, then use the resulting active set to set feature-specific penalty weights for the CATE model, with a tunable pretraining strength $\\alpha$. They argue that when the supports of $m(X)$ and $\\tau(X)$ overlap, this cross-task learning yields lower estimation error, better support recovery, lower false discovery rates, and higher power for detecting heterogeneity. The same logic is extended to nonlinear settings via basis expansions, random forests, and a sparser univariate-guided lasso variant.","feed_headline":"Pretraining on risk signals sharpens treatment-effect fits","feed_subtitle":"In the R-learner, variables that predict outcomes also guide penalties for the effect model, cutting error and false discoveries.","key_machinery":"The load-bearing mechanism is the pretrained R-learner: the residualized loss $\\mathbb{E}[(Y - m(X)) - (W - e(X))\\tau(X)]^2$, estimated with plug-in cross-fitted nuisance models, combined with an adaptive-lasso penalty schedule. The penalty weight for covariate $j$ is $1/\\alpha$ exactly when the outcome lasso estimate is zero, and $1$ when it is nonzero; $\\alpha$ is the pretraining strength, tuned by the empirical R-loss. This machinery converts the shared-support assumption into differential shrinkage: covariates that never showed up for prognosis are shrunk harder, so the effect model is biased toward prognostic covariates unless the data overrule it. The same schedule is translated to nonlinear learners through feature-importance-weighted sampling and tree-based basis expansions.","core_discovery":"The central claim is that the R-learner's nuisance model for the mean outcome is not merely a nuisance: its selected support carries usable information about which covariates modify the treatment effect. Concretely, the paper proposes fitting $m(x)$ with a lasso, forming the active set $\\hat{A} = \\{j : \\hat{\\beta}_j \\neq 0\\}$, and then estimating $\\tau(x)$ by minimizing the R-loss with adaptive lasso penalty weights $w_j = 1/\\alpha$ if $j \\notin \\hat{A}$ and $w_j = 1$ otherwise. The pretraining strength $\\alpha$ is tuned on the same R-loss, which the simulations show is concordant with test MSE. When $m$ and $\\tau$ share support, this pretrained R-lasso dominates the plain R-lasso in MSE, support recovery, FDR, and power; when support is disjoint it is designed to do no harm. The paper demonstrates the same transfer principle for generalized random forests and for gradient-boosted basis expansions, and shows that a univariate-guided sparse regression version improves FDR further.","pith_inferences":["A testable diagnostic emerges for applied work: estimate the overlap between the outcome model's active set and the effect model's selected support on a holdout sample; when that overlap is estimated to be small, one can pre-commit to $\\alpha$ near 1 and avoid over-regularization.","The pretraining idea could be extended to borrowing side information from external datasets, fitting $m$ and $e$ on a large related cohort and transferring the penalty weights to a smaller target cohort, which the paper only mentions as future work.","The reliance on the R-loss for tuning $\\alpha$ is a latent fragility: if nuisance estimation error is large, the R-loss may mis-tune $\\alpha$, and the paper's simulations do not cover that regime; a practical safeguard would be to validate $\\alpha$ by cross-fitting and to examine sensitivity of the final effect estimate to $\\alpha$."],"forward_implications":["In high-dimensional biomarker studies where risk and treatment-response genes overlap, the pretrained R-lasso should find effect modifiers with less data and at a lower false discovery rate than the standard R-lasso.","When the shared-support assumption fails, the R-loss tuning is expected to drive $\\alpha$ toward 1, so the procedure degrades gracefully to the plain R-lasso rather than imposing a fixed penalty.","The same pretraining strategy transfers to nonlinear CATE learners: randomized forests with importance-weighted feature sampling and lasso on gradient-boosted basis expansions both gain accuracy when prognosis and effect modification share support.","The univariate-guided variant (UniR-lasso) inherits the pretraining gains and, in simulations, improves support recovery and FDR over the lasso-based pretrained model.","The R-loss can serve as a practical tuning instrument for pretraining strength, since its training-set minimum aligns with the test-set MSE minimum across overlap levels."],"supporting_citations":[{"why":"Supplies the R-learner loss and its oracle property, the framework the pretraining procedure modifies.","marker":"Nie & Wager 2021"},{"why":"Introduces the pretraining-and-lasso idea, including the pretraining strength $\\alpha$, which the paper adapts to the CATE setting.","marker":"Craig et al. 2024"},{"why":"Provides the adaptive lasso whose feature-specific penalty weights the paper uses for the CATE model.","marker":"Zou 2006"},{"why":"Implements the lasso and elastic-net path used in all the paper's linear simulations.","marker":"Friedman et al. 2010"},{"why":"Supplies UniLasso, the univariate-guided sparse regression the paper turns into UniR-lasso for better FDR.","marker":"Chatterjee et al. 2025"},{"why":"Provides generalized random forests, the base for the feature-importance pretraining variant.","marker":"Athey et al. 2019"},{"why":"Provides the grf software package used for causal forest experiments.","marker":"Tibshirani et al. 2024"},{"why":"Defines the TOC/AUTOC metric used to assess power for detecting heterogeneity.","marker":"Yadlowsky et al. 2025"}],"fun_headline_variants":["Risk signals pretrain effect estimates, cut false hits","Prognostic support sharpens CATE, lowers FDR","Pretrain effect model on risk to boost power","Shared support: pretraining improves heterogeneous effects","Leverage prognosis to sharpen treatment effect detection"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The entire gain rests on the lasso's chosen outcome predictors being a trustworthy stand-in for the covariates that actually modify the treatment effect; when that overlap signal is absent or noisy, pretraining can penalize exactly the wrong covariates, and the paper itself notes in the Discussion that it only has simulation evidence that the R-loss tuning repairs this.","fun_headline_variants_meta":{"raw":{"variants":["Risk signals pretrain effect estimates, cut false hits","Prognostic support sharpens CATE, lowers FDR","Pretrain effect model on risk to boost power","Shared support: pretraining improves heterogeneous effects","Leverage prognosis to sharpen treatment effect detection"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000489,"raw_usage":{"total_tokens":2420,"prompt_tokens":968,"completion_tokens":1452,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":584,"completion_tokens_details":{"reasoning_tokens":1378}},"tokens_in":584,"tokens_out":1452,"duration_ms":10304,"temperature":1.0,"reasoning_tokens":1378,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-16T04:45:20.704075+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"In a simulation with known fully disjoint support between $m(X)$ and $\\tau(X)$, strong covariate correlation, and weak outcome signal, check whether the R-loss-selected $\\alpha$ falls well below 1 or the pretrained model's test MSE exceeds the baseline R-lasso's; if either happens, the claim that pretraining only helps and never hurts under R-loss tuning is falsified.","supporting_citations":[{"cited_title":"(2006), ‘The adaptive lasso and its oracle properties’,Journal of the American statistical association101(476), 1418–1429","cited_arxiv_id":null,"evidence_quote":"Provides the adaptive lasso whose feature-specific penalty weights the paper uses for the CATE model."}],"review_version":1}