{"id":"0e7c8183-8bda-4d3a-a261-36dbe99a10fa","arxiv_id":"1908.05824","paper_version":1,"verdict":"ACCEPT","confidence":"MODERATE","novelty_score":7.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":4,"one_line_summary":"A stochastic choice function is generated by some drift-diffusion model if and only if its stopping-time distribution equals the hitting-time distribution of Brownian motion with the data-revealed drift and boundary, which are uniquely identified.","lead":"The paper derives a condition that pairwise choice-and-response-time data must satisfy to be explainable by the drift-diffusion model, and shows the model's drift and stopping boundary are then uniquely identified. It also constructs a nonparametric test that checks whether observed response-time distributions match the model's predicted hitting times.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"No significant objection identified: Theorem 1 is internally sound under the stated DDM definition; the noted limitations affect empirical scope, not the characterization.","rationale":"The paper's strongest claim is Theorem 1. The proof is standard and internally coherent: Fudenberg-Strack-Strzalecki's equation (4) is the likelihood-ratio identity for Brownian motion with symmetric absorbing boundaries, valid for time-dependent boundaries; optional stopping with finite mean decision time gives δE[τ]=E[sgn(Zτ)b(τ)], which yields the revealed drift formula; substitution gives the revealed boundary. Sufficiency is constructive: if the observed F equals the hitting-time distribution with the revealed parameters, then the DDM with those parameters reproduces both F and p. I checked for hidden assumptions: the odds-ratio identity does not require optimality of the boundary, the optional sampling theorem applies under the paper's finite-mean assumption, and the relabeling that makes δ positive is consistent with the definition. The reader's weakest assumption—the absence of nondecision time and trial-level heterogeneity—is real as a scope limitation. The paper explicitly defines DDM with a fixed environment and a single boundary, so the characterization is not internally invalidated by these omissions. The econometric section is less fully verified: Assumption 5 is high-level and no Monte Carlo evidence is provided, but Theorem 3 is stated conditionally and the variance construction is detailed. This supports the reader's medium correctness risk rather than a change of verdict. I therefore see no load-bearing objection to the central claim and recommend no change.","tokens_in":20608,"tokens_out":13420,"duration_ms":146493,"concrete_test":"Run a Monte Carlo size study: simulate n=5000 i.i.d. choices from the DDM in Definition 1 with a smooth nonconstant boundary, e.g. b(t)=a+c exp(-λt), and δ>0; estimate δ and b with the proposed B-spline estimator; compute the test statistic with J=10 and large S; repeat 1000 times and compare rejection rates to the nominal 5% level. This checks the finite-sample size of the test and indirectly validates the odds-ratio identity and the plausibility of Assumption 5 in a primitive DDM environment.","verdict_should_be":"UNCHANGED","load_bearing_attack":"No significant objection identified. The central characterization (Theorem 1) is a conditional equivalence: for a fixed pair, if (p,F) is generated by a DDM with a symmetric boundary, nonzero drift, no nondecision time, and no trial-level heterogeneity, the revealed drift and boundary recover the true parameters via the odds-ratio identity and optional stopping; conversely, if F equals the hitting-time distribution generated by the revealed parameters, the same DDM reproduces p through the identity p/(1-p)=exp(2δb(t)). I checked the key steps: the odds-ratio identity is the pathwise likelihood-ratio for Brownian motion with symmetric absorbing boundaries and is valid for time-dependent boundaries; optional stopping requires only finite mean decision time, which the paper assumes; and the sufficiency argument is constructive. The restrictions flagged by the reader (no nondecision time, no trial heterogeneity, common symmetric boundary) are explicit in Definition 1 and footnote 2; they limit external applicability to standard Ratcliff DDM data but do not undermine the theorem's internal validity. Theorem 3's asymptotic test rests on high-level Assumption 5 that is not verified for primitive DDM conditions, but this is a standard division of labor and a reason for moderate confidence rather than a flaw in the stated conditional result. No load-bearing internal inconsistency was found.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper develops a characterization and a statistical test for the drift-diffusion model (DDM) with general time-dependent, symmetric stopping boundaries. For a fixed pair of alternatives, the analyst observes the joint distribution of choices and response times, summarized by the stopping-time CDF F_xy and the conditional choice probability p_xy. The paper defines a revealed drift and a revealed boundary from the choice imbalance and mean decision time, and proves (Theorem 1) that the pair (p_xy, F_xy) admits a DDM representation if and only if F_xy equals the hitting-time distribution generated by the revealed drift and boundary, with uniqueness of the representation up to the volatility scale. Theorem 2 extends this to all binary menus under common boundary and utility-additive drifts. The paper then proposes a nonparametric estimator of the drift and boundary based on B-spline estimation of p_xy, and a moment-based test that compares sample moments of decision times with moments simulated from the estimated DDM. The asymptotic chi-square distribution of the test statistic is stated in Theorem 3 under high-level regularity conditions.","tokens_in":20770,"tokens_out":23281,"duration_ms":217846,"significance":"If the results are correct, this is a substantial methodological contribution: it gives the first general necessary and sufficient condition for a DDM with arbitrary time-dependent boundaries, shows that the drift and boundary are uniquely identified from choice/RT data, and provides a nonparametric test that avoids likelihood computations. The proof of Theorem 1 is elegant, using the pathwise odds-ratio identity for Brownian motion and optional stopping, and it correctly highlights that the mean choice probability and mean RT do not summarize the model's content. The paper is explicit about the scope conditions (no nondecision time, no trial-level heterogeneity, common symmetric boundary), which limit external applicability to standard Ratcliff-type DDM data but do not undermine the conditional characterization. The econometric section is ambitious and attempts a careful decomposition of sampling, estimation, and simulation variance.","major_comments":[{"comment":"The estimator in Section 5.1 is inconsistent with the identification formula in Section 4. Equation (3) defines the revealed drift as \\tilde\\delta_xy = \\sqrt{\\bar I_xy / (2 \\bar T_xy)}, but Section 5.1 defines \\hat\\delta := \\sqrt{\\bar I / \\bar\\tau}, omitting the factor 1/2. Likewise, the boundary estimator \\hat b(t) = (1/\\hat\\delta) \\ln[\\hat p(t)/(1-\\hat p(t))] omits the factor 1/2 that appears in equation (4). Under a true DDM with parameters (\\delta,b), these estimators converge to (\\sqrt{2}\\delta, \\sqrt{2}b). There is no rescaling of time and space that maps a DDM with parameters (\\sqrt{2}\\delta,\\sqrt{2}b) into one with (\\delta,b) while preserving the hitting-time distribution, so the simulated moments \\hat m_S in Section 5.2 do not converge to the true moments under the null. Consequently the statistic \\hat A in Theorem 3 will not be asymptotically \\chi^2_J even in large samples. Lemma 5 in Appendix D uses the functional \\delta(I,\\tau)=\\sqrt{I/\\tau}, which is the inconsistent functional; with the correct estimator the functional is \\sqrt{I/(2\\tau)} and the influence functions in Lemma 5 and in the proof of the test's asymptotic distribution must be re-derived. This error is fixable but it affects the core estimation and testing procedure.","section":"Section 5.1 and Appendix D, Lemma 5"},{"comment":"Theorem 3 is stated under the high-level Assumption 5, but the paper does not verify this assumption for the DDM class. In particular, condition (d), (J+1)E[1(\\tau_i < 1/(J+1))\\psi_{i\\delta x}^2] \\geq C, requires enough mass of the decision-time distribution near zero. For a DDM with constant positive boundary b, the density of \\tau near zero is exponentially small, of order exp(-b^2/(2t)), and whether condition (d) holds depends on the analyst's chosen transform G and on the unknown boundary. The brief remark after Assumption 2 that the condition can be weakened by assuming b(t) constant on known intervals near zero and at large \\tau is not a substitute for a primitive verification. Without such verification, the asymptotic size control claimed in Theorem 3 is conditional on an unverified regularity condition. I recommend either proving these conditions for the DDM class or stating them as explicit primitive restrictions in the theorem.","section":"Appendix C, Assumption 5"}],"minor_comments":[{"comment":"Condition (iii) appears to contain a typo: \\tilde\\delta_{xy} + \\tilde\\delta_{yz} = \\tilde\\delta_{xyz} should be \\tilde\\delta_{xy} + \\tilde\\delta_{yz} = \\tilde\\delta_{xz}.","section":"Section 4.2, Theorem 2"},{"comment":"The display in the proof of Lemma 1, p*(\\delta(x,y),b,\\alpha) = p*(1/\\alpha \\delta(x,y), b/\\alpha, \\alpha), should have the last argument equal to 1, not \\alpha, after the rescaling.","section":"Lemma 1 proof"},{"comment":"There is a typo in the definition of \\hat\\tau_s: \"Browning motion\" should be \"Brownian motion\".","section":"Section 5.2"},{"comment":"The proof section is titled \"Proof of Theorem 4\" but the theorem in the main text is numbered Theorem 3; the numbering should be made consistent.","section":"Appendix D"},{"comment":"The sentence beginning \"choice data where pxy(t) and ¯T xy are any 2 given constants is only consistent with one possible distribution of stopping times F xy However...\" is grammatically garbled and should be rewritten for clarity.","section":"Section 4.1, discussion after Theorem 1"}],"recommendation":"major_revision","confidential_remarks":"The factor-2 inconsistency between the identification formula and the proposed estimators is a genuine, load-bearing error in the econometric section, but it is local in the sense that the characterization theorem is sound and the estimator can be corrected. I recommend major revision rather than rejection, with the requirement that the authors re-derive the estimator, the influence functions, and the asymptotic test under the correct formulas, and that they address the primitive verification of Assumption 5."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"The headline: Theorem 1 is the real result—an if-and-only-if characterization plus unique identification for DDM with general time-dependent boundaries, and the paper proves it cleanly. The econometric test is secondary but credible.\n\nWhat's new: Baldassi et al. gave necessary conditions for constant boundaries; this gives iff for general boundaries and identifies drift and boundary from observables. Equations (3) and (4) are simple but powerful—you can read DDM parameters directly off choice data without likelihood or simulation. The sufficiency argument via the FSS odds-ratio identity is checkable and works. I verified the proof's key steps: optional stopping with finite mean decision time, and the condition p/(1-p)=exp(2δb(t)). No internal inconsistency.\n\nSoft spots, in descending order:\n- The DDM in Definition 1 has no nondecision time and no trial-level variability. Ratcliff's standard model includes both. So Theorem 1 is a conditional characterization: if data are generated by the bare DDM, it recovers and tests it. Applied to real RT data, rejection is ambiguous. The authors are explicit about this; it limits scope, not correctness.\n- Theorem 3's asymptotics lean on Assumption 5, a high-level Frechet differentiability/rate condition not verified for primitive DDM. This is normal for semiparametric series tests, and the appendix gives a detailed proof sketch for the variance estimator, but there is no Monte Carlo evidence, so finite-sample behavior of the chi-square approximation is unknown.\n- Theorem 2's across-menu conditions are stated quickly via Sincov; no estimation or testing for multiple alternatives is carried out. That is left for future work.\n- The test drops one interval because drift uses τ information; that is honest overidentifying restrictions, but it means not all DDM restrictions are tested.\n\nMinor: no data or code shipped; some typos like “satsified” are harmless. Self-citation to FSS for the odds-ratio identity is legitimate—it is a published, standard result.\n\nWho it's for: anyone doing structural estimation of DDM or axiomatizing sequential sampling. I'd cite it for the identification formulas. It deserves a serious referee—the theory is sound, and the test is a reasonable starting point even if empirical validation is missing.","headline":"A clean iff characterization with unique identification for time-dependent-boundary DDM; the econometric test is credible but unvalidated and scope-limited.","tokens_in":21373,"tokens_out":1825,"would_cite":true,"duration_ms":19326,"reading_group":"yes","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":["62G10","62G20","60J65","91B06"],"pacs":[],"model":"deepseek-v4-flash","headline":"For binary choice data, the drift-diffusion model has a unique representation if and only if observed decision times match the hitting-time distribution implied by the revealed drift and boundary; otherwise the model is rejected.","keywords":["drift-diffusion model","sequential sampling","response times","stochastic choice function","nonparametric identification","testable implications","Brownian motion","choice imbalance"],"falsifier":"Take a fixed pair with nonzero average choice imbalance and compute $\\tilde{\\delta}_{xy}$ and $\\tilde{b}_{xy}(t)$ from the observed choice probabilities and decision-time distribution. Simulate many hitting times of $B_t$ with drift $\\tilde{\\delta}_{xy}$ and boundary $\\tilde{b}_{xy}(t)$, and compare interval frequencies of the simulated times with the observed decision times using the paper's $\\chi^2$ statistic. If the statistic exceeds the $\\chi^2_J$ critical value at the chosen level, the DDM representation is rejected; conversely, a simulated DDM dataset should yield the equality $F_{xy}(t)=F^*(t;\\tilde{\\delta}_{xy},\\tilde{b}_{xy})$ at every $t$.","tokens_in":20338,"feed_emoji":"⏱️","tokens_out":9576,"duration_ms":82968,"temperature":0.7,"pith_summary":"The paper asks whether observed binary choices and response times can be explained by a drift-diffusion model: Brownian motion with drift toward one alternative and a possibly time-dependent stopping boundary. It proves that for any pair of options a DDM representation exists exactly when the observed distribution of decision times equals the hitting-time distribution of Brownian motion with a specific 'revealed' drift and boundary recovered from the choice probabilities. The revealed drift is the square root of average choice imbalance divided by twice the mean decision time, and the revealed boundary is the time-t log-odds of choosing one option divided by twice that drift; when a representation exists it is unique. This makes the model testable without likelihood or simulation-based estimation, and it gives the drift and boundary a direct behavioral reading. The paper then constructs a nonparametric estimator of the drift and boundary and a chi-square test that compares sample moments of decision times with moments simulated from the estimated model.","feed_headline":"Choices and response times identify every drift-diffusion parameter","feed_subtitle":"Revealed choice imbalance and decision time give the unique drift and boundary; mismatch rejects the model.","key_machinery":"The machinery is the pair of 'revealed' quantities computed directly from the stochastic choice function. The load-bearing identity is $p_{xy}(t)/(1-p_{xy}(t))=\\exp(2\\delta b(t))$, which pins down the boundary from choice log-odds, together with Doob's optional sampling theorem applied at the hitting time, giving $\\delta E[\\tau]=E[\\operatorname{sgn}(Z_\\tau)b(\\tau)]$ and hence the drift formula $\\tilde{\\delta}_{xy}=\\sqrt{\\bar{I}_{xy}/(2\\bar{T}_{xy})}$. The boundary $\\tilde{b}_{xy}(t)$ follows the time path of the log-odds of choice. Theorem 1 reduces the DDM question to a single distributional equality; Theorem 2 enforces cross-menu consistency; and the test statistic compares sample moments of decision times with moments from simulated hitting times of Brownian motion with the estimated drift and boundary.","core_discovery":"The central result, Theorem 1, is a necessary and sufficient characterization for a fixed pair $x,y$: if the revealed drift $\\tilde{\\delta}_{xy} = \\sqrt{\\bar{I}_{xy}/(2\\bar{T}_{xy})}$ is nonzero, then $(p_{xy},F_{xy})$ has a DDM representation if and only if $F_{xy}(t)=F^*(t;\\tilde{\\delta}_{xy},\\tilde{b}_{xy})$ for all $t\\ge 0$, where $\\tilde{b}_{xy}(t)=(\\ln p_{xy}(t)-\\ln(1-p_{xy}(t)))/(2\\tilde{\\delta}_{xy})$ and $F^*$ is the hitting-time distribution of Brownian motion with that drift and boundary. Whenever such a representation exists it is unique up to the volatility normalization. Across all pairs, Theorem 2 adds that the same boundary must be used in every menu and the drifts must satisfy $\\tilde{\\delta}_{xy}+\\tilde{\\delta}_{yz}=\\tilde{\\delta}_{xz}$, the Sincov equation that makes drift a utility difference. The econometric test then checks whether the observed stopping-time distribution matches the simulated hitting-time distribution from the estimated drift and boundary.","pith_inferences":["The characterization's reliance on no nondecision time suggests a natural extension: introduce a minimum response-time parameter and test whether the shifted distribution satisfies the same equality; the revealed formulas would change but the logic would carry over.","Because the theorem only needs the joint distribution of choice and time, any model that predicts the same choice probabilities and mean times as a DDM but a different full distribution of response times will be rejected, which sharpens earlier partial characterizations that only used marginal choice data.","The cross-pair additive-drift condition is exactly the Sincov equation, which forces the drift to be an additive utility index; a multi-menu dataset that passes the test therefore reveals an interval-scale utility representation from choice times alone.","If real data come from a DDM with trial-to-trial variability in drift or start point, the test will tend to reject; simulating such heterogeneous data and applying the test would quantify how much power is lost to misspecification."],"forward_implications":["The drift-diffusion model with general time-dependent boundaries becomes falsifiable from choice and response-time data: compute the revealed drift and boundary, then reject the model if observed stopping times deviate from the predicted hitting-time distribution.","Whenever the model fits, the drift and boundary are uniquely identified, so researchers can recover the structural parameters directly from the stochastic choice function without computing likelihoods or running Monte Carlo fits.","The drift has a transparent behavioral meaning—it grows with how lopsided and how fast choices are—and the boundary is the time path of the choice log-odds, so boundary shape is observable rather than latent.","In multi-alternative data, the model requires the same boundary in every pair and drifts that add across pairs; both conditions are directly checkable and amount to a test of utility-difference representation.","The proposed test has asymptotically chi-square critical values, so a researcher can perform a formal rejection decision at a chosen significance level once the drift and boundary are estimated nonparametrically."],"supporting_citations":[{"why":"Supplies the identity $p_{xy}(t)/(1-p_{xy}(t))=\\exp(2\\delta b(t))$ and the optional-sampling relation used to derive the revealed drift and boundary.","marker":"Fudenberg, Strack, and Strzalecki (2018)"},{"why":"Provides an earlier partial characterization of DDM with constant boundaries that the paper's joint distribution condition strengthens from necessary to necessary-and-sufficient.","marker":"Baldassi, Cerreia-Vioglio, Maccheroni, and Marinacci (2018)"},{"why":"Gives the uniform convergence rate for the spline estimator of the choice probability used in Lemma 4.","marker":"Belloni, Chernozhukov, Chetverikov, and Kato (2015)"},{"why":"Provides the asymptotic variance formula for semiparametric estimators that underlies construction of the test's variance estimator.","marker":"Newey (1994)"},{"why":"Supplies the many-weak-moments lemma used to show the test statistic is asymptotically chi-square.","marker":"Newey and Windmeijer (2009)"},{"why":"Provides the influence-function result used for the uniform expansion of the boundary estimator in Lemma 6.","marker":"Ichimura and Newey (2018)"},{"why":"Identifies the Sincov functional equation used to derive the additive-drift consistency condition across menus.","marker":"Aczél (1966)"}],"fun_headline_variants":["Unique drift and boundary recovered from choice probabilities","Statistical test decides if your data fits drift-diffusion","Necessary and sufficient condition for drift-diffusion fits","Choices alone can verify or reject drift-diffusion models","Characterizing drift-diffusion: a test from observed choices"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The characterization assumes the data are generated by a single DDM with one common boundary across all menus, no nondecision time, and no trial-to-trial variation in drift or starting point; if real response times contain a sensorimotor delay or the drift varies across trials, the revealed drift and boundary are misspecified and the tested equality can fail even when a more flexible DDM generated the data.","fun_headline_variants_meta":{"raw":{"variants":["Unique drift and boundary recovered from choice probabilities","Statistical test decides if your data fits drift-diffusion","Necessary and sufficient condition for drift-diffusion fits","Choices alone can verify or reject drift-diffusion models","Characterizing drift-diffusion: a test from observed choices"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000762,"raw_usage":{"total_tokens":3371,"prompt_tokens":921,"completion_tokens":2450,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":537,"completion_tokens_details":{"reasoning_tokens":2369}},"tokens_in":537,"tokens_out":2450,"duration_ms":17946,"temperature":1.0,"reasoning_tokens":2369,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-14T13:03:33.390914+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Take a fixed pair with nonzero average choice imbalance and compute $\\tilde{\\delta}_{xy}$ and $\\tilde{b}_{xy}(t)$ from the observed choice probabilities and decision-time distribution. Simulate many hitting times of $B_t$ with drift $\\tilde{\\delta}_{xy}$ and boundary $\\tilde{b}_{xy}(t)$, and compare interval frequencies of the simulated times with the observed decision times using the paper's $\\chi^2$ statistic. If the statistic exceeds the $\\chi^2_J$ critical value at the chosen level, the DDM representation is rejected; conversely, a simulated DDM dataset should yield the equality $F_{xy}(t)=F^*(t;\\tilde{\\delta}_{xy},\\tilde{b}_{xy})$ at every $t$.","supporting_citations":[],"review_version":1}