{"id":"38bdd619-f4f9-4cb1-80c8-65a3d14ac77f","arxiv_id":"1908.08600","paper_version":4,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":7.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":3,"one_line_summary":"The optimal bid in a second-price ad auction equals the conditional average treatment effect of the ad, and the paper's Thompson-sampling algorithm learns that bid and therefore the ad effect while reducing experimentation cost.","lead":"This paper proposes an adaptive bidding algorithm, BITS, that estimates the causal effect of online ads by learning the optimal bid in real-time auction markets. In second-price auctions the paper shows the optimal bid equals the average ad effect, so a single algorithm serves both profit and measurement.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Assumption 1 is the load-bearing link: if unobserved common values correlate B_CP with {Y(1),Y(0)} given x, Propositions 1–2 and BITS's read-off of CATE from b* break.","rationale":"The reader's weakest assumption is exactly the one I would stress-test. Propositions 1 and 2 are derived cleanly under their assumptions, and the simulation design is transparent: the bid grids contain the true optimal bid, the DGP satisfies Assumption 1 by construction, and the MSE and pseudo-regret comparisons are well specified. The off-grid issue is real but secondary: Section 4 explicitly assumes b*(x) lies in the arm grid, so if it does not, exact recovery of CATE fails even under Assumption 1; however, this can in principle be addressed by refining the grid or moving to continuous bid optimization, whereas Assumption 1 is a structural identification condition. The abstract promises a regret bound and a real-data application that the main text does not deliver; Section 5.1 only cites Gopalan et al. (2014) for a logarithmic regret result, and the simulations are the only cost evidence. These are genuine gaps, but they do not bear on whether the bid-to-CATE mapping is correct. The load-bearing step is the second equality in the proofs of Propositions 1 and 2, which is exactly Assumption 1. In RTB, competing bids are generated by algorithms using user-level value predictions; unless x contains the full common information set, B_CP and the potential outcomes share a common driver, and the payoff-relevant object is not CATE(x) but a conditional expectation given the auction event. A simulation that injects an unobserved common factor directly tests the sensitivity of the CATE estimate to this violation. I therefore agree with the reader's conditional verdict and would not change it: conditional acceptance pending the confounder sensitivity check, real-data validation, and a delivered regret analysis.","tokens_in":42613,"tokens_out":8190,"duration_ms":90848,"concrete_test":"Simulate the SPA non-contextual DGP in (36) with an unobserved common factor W: log Y(1) = 0.809 + gamma*W + epsilon_1, log Y(0) = 0.22 + gamma*W + epsilon_0, log B_CP = 0.4 + gamma*W + epsilon_CP, with W ~ N(0,1) independent of the epsilons, and run BITS exactly as specified, omitting W. For gamma in {0, 0.25, 0.5, 1}, record the posterior of the optimal bid and the final CATE estimate from (27) over 1,000 epochs. If the estimated ATE shifts away from the true value 1 by more than the simulation MSEs reported in Table 2 as gamma grows, Assumption 1 is quantitatively load-bearing and the conditional verdict should require a robustness check; if the shift is negligible for plausible gamma, the concern is weakened.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claim is that b*(x) is a known transform of CATE(x), so a bandit over bid arms recovers the causal effect. This identification lives entirely in the second equality of the proofs of Propositions 1 and 2 (Section 3): Assumption 1 is used to replace E[Y(1)-Y(0) | B_CP <= b; x] with CATE(x) = E[Y(1)-Y(0) | x] in equations (6) and (7). If, after conditioning on x, competing bids remain correlated with potential outcomes through an unobserved common value (for example, user purchase propensity or advertiser value not fully captured by x), then the integrand in (6) becomes E[Y(1)-Y(0) | B_CP = u; x] - u, which varies with the conditioning event. The optimal bid then solves a fixed point b* = m(b*) with m(u) = E[Y(1)-Y(0) | B_CP = u; x], not b* = CATE(x), and the SPA estimator in (27) is biased; the FPA estimator in (29) inherits the same problem through chi. Assumption 3 (i.i.d. data) is also strong for RTB, and the grid-containment assumption in Section 4 is practically restrictive, but Assumption 1 is the load-bearing condition: without it, the alignment between the economic and inference goals that the whole design exploits does not hold. The paper motivates Assumption 1 by arguing that x can be made large, but provides no diagnostic, sensitivity analysis, or real-data check, so the central claim rests on this assumption untested.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"This paper proposes an online experimental design for estimating conditional average treatment effects (CATEs) of advertising bought through real-time bidding, using multi-armed bandit learning over bid arms. The authors show that in second-price auctions the optimal bid equals max{0, CATE(x)} (Proposition 1), and in first-price auctions it equals max{0, chi^{-1}[CATE(x)]} with chi(b)=b+F_CP(b|x)/f_CP(b|x) (Proposition 2). They introduce Bidding Thompson Sampling (BITS), a Thompson-sampling algorithm with a Bayesian Gibbs sampler that augments missing potential outcomes and censored competing bids, and they report simulations for SPA and FPA, contextual and non-contextual versions, comparing BITS with A/B, explore-then-commit, and off-the-shelf TS. The abstract promises an order-optimal regret bound and validation on RTB auction data, but the body contains no regret-bound theorem and no real data analysis; the simulations are the only empirical evidence.","tokens_in":43036,"tokens_out":7084,"duration_ms":66094,"significance":"The central identification idea is attractive and, under Assumptions 1 and 2, the proofs of Propositions 1 and 2 are correct: they justify using a bandit over bids to recover CATEs while controlling experimentation cost. The paper is also transparent about the tension between economic and inference goals, and its simulation comparison is informative for the SPA case. Its strengths are the explicit derivation linking bidding to causal effects and the detailed Bayesian estimation procedure. However, the contribution is narrower than advertised: the order-optimal regret bound and the RTB empirical validation claimed in the abstract are absent, the empirical evaluation sets every true optimum exactly on the bid grid, and the load-bearing conditional-independence assumption (Assumption 1) is not tested or subjected to sensitivity analysis.","major_comments":[{"comment":"The abstract states that the paper derives an order-optimal regret bound and uses data from RTB auctions, but no regret-bound theorem appears anywhere in the manuscript, and Section 6 is entirely simulation-based. Section 5.1 merely cites the logarithmic regret bound of Gopalan et al. (2014). These two advertised contributions must either be supplied or removed from the claims.","section":"Abstract and Section 5.1"},{"comment":"The second equality in both proofs applies Assumption 1 to replace E[Y(1)-Y(0)|B_CP <= b; x] with CATE(x). This is the load-bearing step that makes the optimal bid a function of CATE alone. If B_CP and the potential outcomes remain dependent conditional on x (e.g., through an unobserved common value such as user purchase propensity), then for an SPA the optimal bid solves b* = E[Y(1)-Y(0)|B_CP <= b*; x], not b* = CATE(x), and the estimators in (27) and (29) are biased. The paper motivates Assumption 1 by arguing that x can be made large, but provides no diagnostic, sensitivity analysis, or real-data check. A formal sensitivity analysis or an empirical test of this assumption is needed before the central claim can be accepted.","section":"Section 3, equations (6)-(7) and proofs of Propositions 1 and 2"},{"comment":"The manuscript assumes the bid grid contains b*(x) for each x, and every simulated design places the true optimum exactly on a grid point (e.g., ATE = 1 with grids containing 1 for SPAs; b* = 0.5, 0.75, 1.00, 1.25, 1.50 in the FPA contextual grids). Because the SPA estimator (27) and the FPA estimator (29) are finite convex combinations of the discrete grid labels, BITS cannot represent a CATE that falls between grid points. The simulation evidence therefore conditions on the favorable grid-containment case. Off-grid experiments (e.g., dropping the true optimal bid from the grid) and a discussion of sensitivity to grid resolution are necessary to support the generality of the recovery claim.","section":"Section 4 and Section 6"},{"comment":"In the FPA simulations, BITS's MSE (0.010, 0.023, 0.011) is larger than that of A/B (0.004, 0.003, 0.003) and ETC (0.007, 0.007, 0.007) for all three grids, so BITS does not outperform the conventional methods on the inference goal in FPAs. The paper acknowledges this in the text. The abstract's unqualified claim that the method outperforms commonly used methods should be restricted to the SPA setting or to the economic (regret) objective.","section":"Table 3 and abstract"}],"minor_comments":[{"comment":"The phrase 'pursuit of the economic goal ... may come at the cost of performance of the economic goal' appears to be a typo; the second occurrence should be 'inference goal'.","section":"Section 6.2.1, paragraph after Table 3"},{"comment":"The claim that the difference between the two stopping rules is 'minimal in simulations' is not supported by any reported results; either present the simulation evidence or remove the statement.","section":"Section 5.7.2"},{"comment":"The reference 'Amemyia' should be 'Amemiya' and 'Sidney' should be 'Sydney'.","section":"References"},{"comment":"Given the title's promise of causal inference, the paper should discuss how to produce uncertainty quantification (e.g., posterior credible intervals) for the CATE estimates; currently only point estimates and MSE are reported.","section":"Title and overall framing"},{"comment":"The i.i.d. assumption on (Y(1), Y(0), B_CP) is acknowledged to be strong for RTB, but the paper does not discuss possible relaxations or robustness checks for serial dependence; adding such a discussion would help.","section":"Assumption 3"}],"recommendation":"major_revision","confidential_remarks":"The mismatch between the abstract and the body is large: the regret bound and RTB data promised in the abstract are absent, and the 'outperforms' claim is contradicted by Table 3 for FPAs. I do not see this as a matter of author intent; it is an over-claiming problem. I recommend major revision rather than rejection because the SPA identification result is correct and useful, and the suggested fixes (adding off-grid simulations, sensitivity analysis for Assumption 1, and correcting the abstract) are within the scope of a revision. The editor should ask the authors to reconcile the abstract with the actual content."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Quick take: the centerpiece — in a second-price auction the advertiser's optimal bid is her CATE, and in a first-price auction it is chi^{-1}(CATE) — is correct, clean, and worth having. I traced the proofs: under Assumptions 1 and 2 the equalities hold. Prior learn-to-bid work either assumed the advertiser knows her value or set Y(0)=0, so the nonzero-baseline, treatment-effect-as-valuation framing is a genuine step forward. The authors also report honestly: when their BITS algorithm loses on MSE to A/B in the FPA simulations, they say so without spinning it.\n\nThe soft spots, in proportion. First, the arXiv abstract still promises an order-optimal regret bound and RTB data, and the manuscript delivers neither; the v2 abstract says \"simulations\" and that is accurate. No regret analysis appears anywhere, and the Gopalan et al. (2014) citation covers a related but not identical setting, so the order-optimal claim is unbacked. Second, Assumption 1 is load-bearing, and the stress-test concern is right: if unobserved common values move both competing bids and potential outcomes given x, the second equality in the proofs collapses and the CATE read-off breaks. The authors argue x can be made large but give no sensitivity analysis or external check. That is a real gap, though it is a standard unconfoundedness condition and plausible in data-rich RTB. Third, the simulations are friendly: the DGP is exactly the lognormal model the estimator assumes, and every true optimal bid sits on the bid grid. Off-grid optima and misspecification are untested.\n\nThe circularity worry does not land on reading the paper: the bid-CATE link is derived from auction payoffs, not assumed, and the bandit learns bids from rewards. The FPA map is a genuinely new object.\n\nWho this is for: people doing advertiser-side causal inference in RTB, and anyone working on auction-embedded experimentation. It should go to serious peer review. The referee should require the abstract to match the content and should push for off-grid and sensitivity exercises, but the core idea holds.","headline":"The SPA bid = CATE identity is correct and citable, but the abstract overpromises (regret bound and real data are absent) and the load-bearing private-values assumption is untested.","tokens_in":43514,"tokens_out":5478,"would_cite":true,"duration_ms":55330,"reading_group":"yes","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":["62L05","91B26","62F15"],"pacs":[],"model":"deepseek-v4-flash","headline":"This paper shows that in second-price auction advertising, the optimal bid equals the causal effect of the ad, so learning to bid is learning the treatment effect.","keywords":["causal inference","real-time bidding","multi-armed bandits","Thompson sampling","conditional average treatment effect","second-price auctions","first-price auctions","advertising experiments"],"falsifier":"In a data set where the highest competing bid and both potential outcomes are observed for the same impressions, compute the regression of $Y(1)-Y(0)$ on $B_{CP}$ within cells of $x$; a nonzero slope would directly contradict Assumption 1 and invalidate the identity $b^*(x)=\\mathrm{CATE}(x)$.","tokens_in":42423,"feed_emoji":"🎯","tokens_out":5517,"duration_ms":53032,"temperature":0.7,"pith_summary":"The paper's central claim is that in real-time bidding auctions, the causal effect of showing an ad can be read off the advertiser's optimal bid. In second-price auctions the two objects are equal, $b^*(x)=\\max\\{0, \\mathrm{CATE}(x)\\}$, while in first-price auctions $b^*(x)=\\max\\{0,\\chi^{-1}[\\mathrm{CATE}(x)]\\}$ for the bid-shading function $\\chi$; this makes the treatment effect identifiable from bidding behavior alone. To exploit it, the authors propose BITS, a Thompson-sampling bandit whose arms are bids, updated by a Gibbs sampler that imputes missing potential outcomes and censored competing bids. Simulations show BITS recovers conditional average treatment effects across contexts while generating lower cumulative regret than A/B tests, explore-then-commit, and off-the-shelf Thompson sampling. A sympathetic reader would care because the result gives advertisers a way to run cost-efficient experiments where ad exposure is controlled by an auction rather than by the experimenter.","feed_headline":"In ad auctions, optimal bid equals the ad's causal effect","feed_subtitle":"A bid-learning bandit recovers ad effects while cutting experimentation costs.","key_machinery":"The load-bearing object is the auction payoff function rewritten in terms of the treatment effect: $\\pi(b,Y(1),Y(0),B_{CP}) = \\mathbf{1}\\{B_{CP}\\le b\\}(Y(1)-Y(0)-\\text{payment}) + Y(0)$. Under the private-values conditional independence assumption, the expected payoff becomes a function of $\\mathrm{CATE}(x)$ plus a truncated payment term, so maximizing it yields the optimal-bid identities in Propositions 1 and 2. The algorithm BITS is a Thompson-sampling contextual bandit whose arms are bid values; each round a Gibbs sampler with data augmentation imputes the missing potential outcomes and the censored or missing highest competing bid, then computes the posterior probability that each bid is optimal, and reads off $\\mathrm{CATE}(x)$ as the best bid in a second-price auction or as the adjusted best bid in a first-price auction.","core_discovery":"The paper establishes an identity between what an advertiser should bid and what she wants to measure. In a second-price auction, under conditional independence of potential outcomes and the highest competing bid given context $x$, the optimal bid equals $\\mathrm{CATE}(x)$: $b^*(x)=\\max\\{0, \\mathrm{CATE}(x)\\}$. In a first-price auction, the optimal bid is $b^*(x)=\\max\\{0, \\chi^{-1}[\\mathrm{CATE}(x)]\\}$, where $\\chi(b)=b+F_{CP}(b|x)/f_{CP}(b|x)$, so the treatment effect is recovered by adding the reversed hazard rate of the competing bid to the winning bid. Therefore the causal effect of advertising is identifiable from the bidder's own optimal bidding behavior, and a multi-armed bandit that learns optimal bids, called BITS, delivers both profit-efficient bidding and treatment-effect estimates in one online procedure.","pith_inferences":["An implication the authors leave implicit is that in second-price auctions, bid logs alone may be sufficient for incremental ad measurement: advertisers or platforms could reconstruct treatment effects from the winning bid without serving separate ghost ads.","A natural extension would be to replace the lognormal parametric assumption for competing bids with a flexible hazard estimator in first-price auctions, since the treatment-effect estimate inherits any misspecification of $F_{CP}$ through the $\\chi$ adjustment.","The private-values assumption could be tested empirically in publisher-side data where both potential outcomes and the highest competing bid are observed; such a test would tell practitioners how much unobserved common-value variation matters in real RTB markets.","Budget constraints and strategic multi-advertiser learning are left out of the model; incorporating them would make the optimal bid depend on remaining budget and on competitors' learning dynamics, breaking the simple $b^*$-to-CATE link."],"forward_implications":["In second-price auctions, any bandit procedure that identifies the best bid directly estimates the conditional average treatment effect, because the best bid equals the treatment effect and no separate outcome regression is required.","In first-price auctions, estimating the treatment effect requires also estimating the reversed hazard rate of competing bids, which is identifiable from the lower and upper bounds on the competing bid observed when losing and winning.","The BITS algorithm achieves logarithmic regret in the stochastic bandit setting, so it protects the advertiser's profit during the experiment while still collecting the losing-auction observations needed to see the no-ad outcome.","Simulations under the paper's data-generating process show BITS attains lower cumulative regret than A/B tests, explore-then-commit, and off-the-shelf Thompson sampling in both auction formats, and lower mean squared error for treatment-effect estimation in second-price auctions.","The experiment can be stopped by a posterior-probability threshold, which gives a Bayes-factor interpretation and lets the advertiser end the test once the best bid, and hence the treatment effect, is learned with a chosen confidence."],"supporting_citations":[{"why":"Establishes the fundamental missing-data problem that makes potential outcomes censored by auction outcomes.","marker":"Holland (1986)"},{"why":"Supplies the Bayesian Tobit Gibbs sampler used to impute censored highest competing bids in second-price auctions.","marker":"Chib (1992)"},{"why":"Supplies the Bayesian Probit data augmentation used for fully censored competing bids in first-price auctions.","marker":"Albert and Chib (1993)"},{"why":"Provides the regret definitions and stochastic bandit framework used to evaluate BITS.","marker":"Bubeck and Cesa-Bianchi (2012)"},{"why":"Shows Thompson sampling has logarithmic regret in complex online problems, which the paper relies on for BITS.","marker":"Gopalan et al. (2014)"},{"why":"Introduces cross-learning in contextual bidding bandits and the independence assumption that motivates cross-arm learning here.","marker":"Balseiro et al. (2019)"},{"why":"Provides a learning-to-bid baseline under valuation uncertainty in second-price auctions, which the paper contrasts with its own nonzero $Y(0)$ setting.","marker":"Weed et al. (2016)"},{"why":"Provides a learning-to-bid-without-knowing-value baseline; the paper relaxes its $Y(0)=0$ assumption that blocks causal inference on ads.","marker":"Feng et al. (2018)"},{"why":"Serves as the baseline bid-randomization IV design for RTB ad effects that BITS improves upon with adaptivity and auction-theoretic identification.","marker":"Lewis and Wong (2018)"}],"fun_headline_variants":["Optimal bid reveals ad's causal effect in RTB","Bid optimally to learn ad causal effects","In RTB, optimal bid is the ad's effect","BITS: bandit learns ad effects from bids","Ad causal effect? Just bid the optimum"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"Everything rests on Assumption 1: after conditioning on the observed context, a user's ad-response outcomes must be independent of the highest competing bid; if unobserved common-value factors move both, the optimal-bid-to-CATE identity and the resulting estimates collapse.","fun_headline_variants_meta":{"raw":{"variants":["Optimal bid reveals ad's causal effect in RTB","Bid optimally to learn ad causal effects","In RTB, optimal bid is the ad's effect","BITS: bandit learns ad effects from bids","Ad causal effect? Just bid the optimum"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000992,"raw_usage":{"total_tokens":4168,"prompt_tokens":875,"completion_tokens":3293,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":491,"completion_tokens_details":{"reasoning_tokens":3219}},"tokens_in":491,"tokens_out":3293,"duration_ms":25281,"temperature":1.0,"reasoning_tokens":3219,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-14T11:34:35.954827+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"In a data set where the highest competing bid and both potential outcomes are observed for the same impressions, compute the regression of $Y(1)-Y(0)$ on $B_{CP}$ within cells of $x$; a nonzero slope would directly contradict Assumption 1 and invalidate the identity $b^*(x)=\\mathrm{CATE}(x)$.","supporting_citations":[{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Supplies the Bayesian Probit data augmentation used for fully censored competing bids in first-price auctions."}],"review_version":1}