{"id":"56fec6e0-c492-43df-a0f1-ad5f30e2d807","arxiv_id":"2607.15168","paper_version":1,"verdict":"CONDITIONAL","confidence":"HIGH","novelty_score":7.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":5,"one_line_summary":"Indirect variational inference treats a variational approximation as an auxiliary model and inverts its binding function, correcting the bias of variational estimates in nonlinear earnings dynamics.","lead":"This paper shows that a fast machine-learning estimation method, variational inference, can be biased for earnings-dynamics models, and introduces a corrective procedure called indirect variational inference (IVI) that recovers the true parameters. The method is tested on simulated and real U.S. earnings data, where it removes most of the bias while remaining faster than standard simulation-based estimation.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Appendix B.2's fixed-q binding function does not pass through the pseudo-true value: fixing the variational parameter changes the profiled objective, so the fixed-q IVI estimator is biased and does not share the standard IVI fixed point.","rationale":"The paper's central claim is that IVI is consistent and asymptotically normal even when the variational family is misspecified. That claim requires the binding function to be sufficiently regular and invertible, which the reader rightly identifies as unproven. My concern is more specific and more damaging: even granting the one-to-one assumption, the fixed-q binding function introduced in Appendix B.2 does not pass through the pseudo-true value. The reason is that the pseudo-true value is defined by maximizing the profiled ELBO over both ϑ and ϕ, whereas b_fixed-q holds ϕ fixed. These objectives differ, so their maximizers generally differ. This invalidates the paper's assertion that the fixed-q estimator and the standard IVI estimator share the same fixed point, and it means the fixed-q implementation is not consistent as stated. This is not an attack on the main fixed-point estimator used in the simulations; that estimator still faces the reader's unproved one-to-one condition. But the fixed-q error is a concrete, checkable mathematical flaw rather than a missing condition. A simple simulation in the paper's own linear-Gaussian benchmark can settle it. Therefore the manuscript should not be accepted without either correcting or removing the fixed-q claim and supplying a valid consistency argument for the implementation actually recommended. The reader's CONDITIONAL verdict is appropriate; my concern reinforces it with a more specific revision requirement.","tokens_in":34362,"tokens_out":12250,"duration_ms":100710,"concrete_test":"Use the Section 5 linear-Gaussian benchmark (N=30,000, T=6, ρ0=0.9, σe=0.23). First compute τ, the probability limit of the mean-field VI estimator under the DGP, and then compute ϕ0 as the variational optimum at the pseudo-true value τ. Next estimate b_fixed-q(τ) = argmax_{rϑ} E_{Pτ}[E_{rϑ,ϕ0}] by simulating panels at τ with ϕ held at ϕ0, exactly as in Appendix B.2. Appendix B.2 predicts b_fixed-q(τ)=τ. If the estimated b_fixed-q(τ) differs from τ by more than Monte Carlo error, the equality asserted in B.2 is false and the fixed-q IVI estimator does not share the standard IVI fixed point.","verdict_should_be":"CONDITIONAL","load_bearing_attack":"The most concrete load-bearing problem is in Appendix B.2, not merely the unproved one-to-one assumption. The paper defines b_fixed-q(ϑ) = argmax_{rϑ} E_{Pϑ}[E_{rϑ,ϕ0}(y)] with ϕ0 held fixed, and then claims 'by construction' that b_fixed-q(ϑ0)=ϑ0, because at ϑ=ϑ0 the objective is the population variational objective evaluated at fixed ϕ0, whose maximizer is ϑ0. This is false. The pseudo-true value ϑ0 is defined as the maximizer of the profiled objective max_ϕ E_{Pϑ0}[E_{rϑ,ϕ}], not of the objective with ϕ fixed at ϕ0. Replacing the inner maximization by a fixed ϕ0 generically changes the objective, so its argmax over rϑ is not ϑ0. Consequently b_fixed-q(ϑ0)≠ϑ0 in general, and the statement that the fixed-q and standard IVI estimators 'share the same fixed point ϑ0' is not established. If one uses the fixed-q implementation, the equation being solved is different from b(ϑ)=ϑ0, so the resulting estimator is asymptotically biased. This is a definite internal error in a proposed implementation of IVI, not just a missing regularity condition.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes indirect variational inference (IVI), which uses a variational estimator as an auxiliary model in indirect inference. The motivation is that standard VI maximizes a penalized log-likelihood and is generally inconsistent for the true parameter when the variational family is misspecified. IVI aims to correct this bias while retaining VI's tractability: it never evaluates the likelihood. The paper presents a population-level characterization of IVI, a fixed-point implementation, and claims root-N consistency and asymptotic normality for the true parameter. It then applies VI and IVI to a sequence of earnings-dynamics models — linear Gaussian, nonlinear non-Gaussian, with heterogeneity, and with MA(1) transitory shocks — using simulated data and a PSID application. The simulations show that mean-field VI can be badly biased, that flexible Gaussian variational families do better, and that IVI brings estimates close to the truth in the designed experiments. The PSID application reports a nearly linear conditional mean, a U-shaped conditional volatility, heterogeneous transitory variances, and an MA(1) coefficient of about 0.26.","tokens_in":34712,"tokens_out":7993,"duration_ms":71248,"significance":"If the central claims are correct, IVI is a useful methodological contribution: it offers a way to remove or reduce variational approximation bias without likelihood evaluation, which is attractive for nonlinear panel models. The paper also gives a clean analytical decomposition of the ELBO as penalized log-likelihood in Eq. (4.1), and an explicit mean-field penalty for the linear Gaussian case in Eq. (5.15). The simulation designs are relevant to current earnings-dynamics research, and the PSID application illustrates the type of empirical questions the method can address. The scalability comparison in Appendix D is useful. However, the theoretical section as written contains questionable identification statements and an incorrect claim in Appendix B.2, so the paper is not yet publishable in its current form. The main contribution is still potentially publishable after a careful revision.","major_comments":[{"comment":"The notation in this section conflates the true parameter with the pseudo-true value. As written, Eq. (4.8) states b(ϑ0)=ϑ0, but under the standard reading where ϑ0 is the true value this is contradicted by Figure 3, where the binding function at ρ0=0.90 returns 0.76 (the mean-field VI pseudo-true value). The correct population identity for indirect inference is b(ϑ*)=\\bar{ϑ}_0, where \\bar{ϑ}_0 is the probability limit of the VI estimator under the true model. The fixed-point equation being solved should therefore be b(ϑ)=\\bar{ϑ}_0 (or, in finite samples, b(ϑ)=\\hat{ϑ}_VI), not b(ϑ)=ϑ0. Because Eqs. (4.9), (4.10), and the asymptotic statement (4.13) all build on this identity, this is a load-bearing issue that must be corrected.","section":"§4.2, Eqs. (4.8)–(4.9)"},{"comment":"The claim that b_fixed-q(ϑ0)=ϑ0 'by construction' is false. The pseudo-true value ϑ0 is defined as the maximizer over rϑ of max_ϕ E_{Pϑ0}[E_{rϑ,ϕ}], whereas b_fixed-q maximizes E_{Pϑ0}[E_{rϑ,ϕ0}] with ϕ0 held fixed. Replacing the inner maximization by a fixed ϕ0 changes the objective generically, so its argmax need not equal ϑ0. Consequently the fixed-q estimator solves a different equation and is asymptotically biased in general. This is not merely a missing regularity condition: it is an internal error in a proposed implementation. The claim that fixed-q and standard IVI 'share the same fixed point' should be removed unless a genuine condition is proved that guarantees the argmax of the fixed-q objective at ϑ0 is ϑ0.","section":"Appendix B.2"},{"comment":"The one-to-one assumption on the binding function is stated but never established or tested for the high-dimensional models used in the simulations and PSID application. Figure 3 only displays a one-dimensional slice of the binding function for ρ. In the fixed-point implementation (4.10), convergence requires a contraction property and the solution being found must be the unique root of b(ϑ)=target. No evidence is provided on injectivity of the full binding function, multiplicity of fixed points, or sensitivity to starting values. The paper should either prove structural conditions for injectivity for the models considered, or provide numerical diagnostics (e.g., multiple random starts, Jacobian rank checks, or a grid in a lower-dimensional projection) that support the required identification assumption.","section":"§4.2–§4.3"},{"comment":"The central theoretical claim — root-N consistency and asymptotic normality of IVI for the true parameter — is presented as a consequence of combining Westling and McCormick (2019) and Gourieroux et al. (1993), but no theorem with explicit sufficient conditions is stated. Appendix A gives an expansion and says 'we follow the approach in Westling and McCormick (2019) and assume their conditions are satisfied.' This is too informal for a paper whose abstract claims consistency and asymptotic normality. The paper should either state a theorem with verifiable conditions (or cite a precise theorem that covers the VI pseudo-true value in this setting) and then verify those conditions for the earnings-dynamics models, or substantially weaken the claimed theoretical contribution and present IVI primarily as a simulation-based bias-correction procedure.","section":"§4.3 and Appendix A"}],"minor_comments":[{"comment":"The starting value in the fixed-point algorithm is written as ϑ0, which is ambiguous given the true/pseudo-true notation issue. In the finite-sample algorithm it should be the VI estimate \\hat{ϑ}_VI, not a population quantity.","section":"Eq. (4.10)"},{"comment":"The number of outer IVI iterations, the number of ELBO draws, and the neural-network width vary across experiments (e.g., outer iterations 10, 25, 50; ELBO draws 1 vs 40). The paper does not report how sensitive the conclusions are to these tuning choices. A short sensitivity analysis would address concerns about overfitting to the simulation designs.","section":"Table 6"},{"comment":"The statement that the mean-field penalty in Eq. (5.15) is 'an increasing function of |ρ|' is not proved and is only illustrated. This is a small point, but an analytical check or a footnote derivation would be helpful.","section":"§5.4"},{"comment":"The figure illustrates invertibility only along a one-dimensional slice. The text should acknowledge this explicitly and refer to the additional checks requested in the major comments.","section":"Figure 3"},{"comment":"The bootstrap standard errors are based on 100 replications. Given the nonlinearity and the inner optimization, it would be helpful to know the stability of the fixed-point iterations across bootstrap replicates; a small subsection or note in Appendix E would be enough.","section":"§8.1"}],"recommendation":"major_revision","confidential_remarks":"The paper has a solid empirical and computational core, and the IVI idea is attractive. The main issues are concentrated in Section 4 and Appendix B: the identification equation needs to be stated with unambiguous notation, the fixed-q implementation contains a definite error, and the asymptotic claim needs a proper theorem or a clearly qualified statement. These are fixable within the scope of the manuscript, so I do not recommend rejection. However, the central theoretical claims must be repaired before the paper can be accepted."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Two things you should know. The core idea is good: using the variational objective as an auxiliary model in just-identified indirect inference is a natural and useful way to debias VI without evaluating the likelihood, and their simulations show it works. But one of their proposed implementations, the fixed-q variant in Appendix B.2, is wrong as stated, and the paper's asymptotic claims are more asserted than proved.\n\nCredit where it's due. The just-identified formulation, the penalized-likelihood decomposition in (4.1), and the closed-form mean-field penalty in (5.15) are real contributions. The simulation coverage is serious: linear Gaussian, nonlinear non-Gaussian, heterogeneity, MA(1), and T=40. The results consistently show IVI correcting the VI bias. The PSID application is careful, with bootstrap standard errors and honest comparison to Arellano et al. (2017).\n\nNow the soft spots. The stress-test is correct about Appendix B.2. They define b_fixed-q(ϑ) = argmax over rϑ of E_{Pϑ}[E_{rϑ,ϕ0}], and claim b_fixed-q(ϑ0)=ϑ0 'by construction.' That is not by construction. The pseudo-true ϑ0 is the maximizer of the objective with the inner ϕ optimized out, not with ϕ held at ϕ0. Fixing ϕ0 changes the profiled objective, so the argmax over rϑ is generally not ϑ0. The fixed-q estimator therefore does not share the IVI fixed point and is asymptotically biased unless an extra argument is supplied. This is a concrete error in a proposed implementation, not just a missing regularity condition.\n\nSecond, even for the standard IVI estimator, Section 4.3 relies on 'assume' for the Westling-McCormick conditions and on the one-to-one binding function, which is only illustrated on a 1-D slice. That's a genuine gap between the theoretical claim in the abstract and what's shown. It may be acceptable if framed as a methods paper with simulation evidence, but the abstract currently overclaims.\n\nMinor: simulation tables report point estimates without Monte Carlo error bars; with a single simulated data set (or a few) it's hard to judge precision. There is no code or data package. For a methods paper, that's a requirement, not a nicety.\n\nBottom line: the IVI idea is valuable and the empirical application is credible, but the paper needs revision on the fixed-q issue and a more honest statement of the theory. I'd send it to referees, not desk reject. If the fixed-q variant is withdrawn or fixed, and the theory is properly caveated, this could be a solid publication. Useful for labor economists and econometricians who care about estimating nonlinear state-space models on large panels.","headline":"Clever, mostly working method; the fixed-q variant has a real hole and the theory overreaches, but worth refereeing.","tokens_in":35199,"tokens_out":3396,"would_cite":true,"duration_ms":29384,"reading_group":"yes","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":["62F12","62P20"],"pacs":[],"model":"deepseek-v4-flash","headline":"A new indirect-inference wrapper can make variational estimators consistent for the true parameters of nonlinear earnings models, with no likelihood evaluation.","keywords":["Earnings dynamics","Variational inference","Indirect inference","Nonlinear state-space models","Posterior distributions","Mean-field approximation","Binding function","PSID"],"falsifier":"Simulate data from the paper's nonlinear heterogeneity DGP, compute b(ϑ) on a fine grid around the variational pseudo-true value, and check whether b is globally one-to-one. If two distinct ϑ values produce the same variational estimate, or if the Jacobian of b becomes singular or changes sign at the converged IVI value, the fixed-point iteration can converge to the wrong parameter and the central consistency claim fails. A cheaper check is to run IVI from multiple random starting values and see whether the fixed point is unique.","tokens_in":34242,"feed_emoji":"📈","tokens_out":6325,"duration_ms":156042,"temperature":0.7,"pith_summary":"The paper argues that variational inference, a tractable approximation to intractable posterior integrals, can be made reliable for estimating nonlinear earnings dynamics models. Used naively, variational inference maximizes a penalized likelihood whose penalty distorts estimates whenever the variational family is too rigid; in the paper's linear benchmark, a mean-field (diagonal) approximation shrinks persistence from the true 0.90 to a biased 0.76. The proposed fix, indirect variational inference (IVI), treats variational inference as an auxiliary model and inverts its binding function using fixed-point iteration, restoring consistency and asymptotic normality for the true parameters without ever computing the likelihood. In simulations spanning linear, nonlinear, heterogeneous, and serially correlated specifications, IVI corrects substantial variational biases, including severely understated kurtosis of transitory shocks. The empirical application to PSID earnings data yields a nearly linear conditional mean with persistence close to one, a U-shaped conditional volatility, heterogeneous transitory variances, and positive serial correlation in transitory shocks.","feed_headline":"No likelihood needed: new estimator corrects variational bias","feed_subtitle":"Simulated and PSID earnings data show indirect variational inference recovers true parameters where plain VI is biased.","key_machinery":"The load-bearing object is the binding function b(ϑ)=argmax_rϑ E_{rϑ,ϑ}, the large-sample limit of the variational estimator when data are drawn from parameter ϑ, together with the ELBO identity that rewrites the variational objective as the log-likelihood minus a Kullback-Leibler penalty. IVI leverages the fixed-point property b(ϑ0)=ϑ0 and inverts b through the damped iteration ϑ^(k+1)=ϑ^(k)−κ(b(ϑ^(k))−ϑ̂_VI), or through gradient descent on ||b(ϑ)−ϑ̂_VI||². The mechanism is that any distortion introduced by the variational family is absorbed into b, which is estimated by simulation and then inverted, so the corrected estimator targets the true parameter even when the variational family is m","core_discovery":"The paper's central claim is that variational approximation error can be converted from a bias into a solvable equation. For any parameter value ϑ, define the binding function b(ϑ) as the probability limit of the variational estimator when data are simulated from the model at ϑ. The true parameter ϑ0 satisfies b(ϑ0)=ϑ0; if b is one-to-one, then ϑ0 is the unique solution of b(ϑ)=ϑ̂_VI, the variational estimate on the observed sample. The paper shows that solving this equation through a damped fixed-point iteration produces a root-N consistent, asymptotically normal estimator of the true value, with a sandwich variance that can be estimated without computing the likelihood. In the paper's simu","pith_inferences":["Because IVI only needs a binding function, the same correction could be applied to richer approximate posteriors, such as normalizing flows or importance-weighted bounds, provided their binding functions remain invertible; this is an extension beyond the paper's Gaussian-family implementations.","The one-to-one assumption on b is the natural stress point. The paper verifies invertibility only along a one-dimensional slice of the AR(1) parameter; in the high-dimensional PSID model, a reader should test global invertibility, for example by estimating the Jacobian of b over a grid or running IVI from multiple starting values.","A practical robustness check follows from the paper's own convergence diagnostics: monitor the residuals ||b(ϑ^(k))−ϑ̂_VI|| across iterations and require them to decrease persistently toward zero; failure indicates the contraction or invertibility condition may not hold.","The success in earnings dynamics suggests IVI could be imported to other economic latent-variable settings with intractable likelihoods, such as discrete-choice dynamics or network formation models, where variational approximations are already used but bias correction has been missing."],"forward_implications":["Nonlinear earnings dynamics models with heterogeneous, serially correlated, and non-Gaussian shocks become estimable in large panels, since variational inference is differentiable and scalable and IVI removes the bias caused by the variational family.","Mean-field (diagonal) variational posteriors, a common default in machine learning, can substantially attenuate persistence and understate transitory volatility; researchers should either use flexible Gaussian families or apply IVI.","IVI inherits indirect inference's asymptotic theory: root-N consistency and asymptotic normality for the true parameter, with a variance inflation factor 1+1/M reflecting simulation noise, even when the variational family is misspecified.","On the PSID, the bias-corrected estimates change economic conclusions: conditional volatility is U-shaped rather than flat, transitory shocks are leptokurtic and serially correlated, and persistence depends on both the state and the sign of the shock.","The method scales to longer panels: in the paper's T=40 simulation, IVI corrects variational bias where plain VI still understates shock kurtosis."],"fun_headline_variants":["Earnings models: variational bias corrected by IVI","No likelihood: indirect variational inference fixes bias","Turn variational bias into a solvable fixed-point","Variational bias? Solve it without the likelihood","IVI: root-N consistent estimator, no likelihood needed"],"cache_read_input_tokens":2304,"weakest_assumption_plain":"Everything rests on the binding function b(ϑ) being one-to-one, so that the equation b(ϑ)=ϑ̂_VI has a unique solution; the paper states this as a key identification assumption but only demonstrates it on a one-dimensional slice of the AR(1) parameter, not for the high-dimensional PSID model.","fun_headline_variants_meta":{"raw":{"variants":["Earnings models: variational bias corrected by IVI","No likelihood: indirect variational inference fixes bias","Turn variational bias into a solvable fixed-point","Variational bias? Solve it without the likelihood","IVI: root-N consistent estimator, no likelihood needed"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000191,"raw_usage":{"total_tokens":1155,"prompt_tokens":694,"completion_tokens":461,"prompt_tokens_details":{"cached_tokens":256},"prompt_cache_hit_tokens":256,"prompt_cache_miss_tokens":438,"completion_tokens_details":{"reasoning_tokens":388}},"tokens_in":438,"tokens_out":461,"duration_ms":5750,"temperature":1.0,"reasoning_tokens":388,"cache_read_input_tokens":256,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-01T23:55:41.925600+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Simulate data from the paper's nonlinear heterogeneity DGP, compute b(ϑ) on a fine grid around the variational pseudo-true value, and check whether b is globally one-to-one. If two distinct ϑ values produce the same variational estimate, or if the Jacobian of b becomes singular or changes sign at the converged IVI value, the fixed-point iteration can converge to the wrong parameter and the central consistency claim fails. A cheaper check is to run IVI from multiple random starting values and see whether the fixed point is unique.","supporting_citations":[],"review_version":1}