{"id":"a04050d8-eb07-443b-bce4-e6893b946d48","arxiv_id":"2411.15980","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":2,"one_line_summary":"Using Empirical Bayes on three country panels, the authors estimate firm-specific Cobb-Douglas and CES technologies and report large heterogeneity in output elasticities plus a strong negative correlation between factor-neutral productivity and returns to scale.","lead":"This paper estimates production functions that let every firm have its own technology parameters, using an Empirical Bayes method on manufacturing data from Chile, Colombia, and Japan. It finds large differences across firms in both productivity levels and how strongly output responds to capital and labor, with a strong negative link between the two.","discovery_kind":"new_application","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The headline negative correlation may be an artifact of the quadratic productivity-dynamics assumption: if true TFP follows a persistent shock process, inputs correlate with omitted productivity and the estimated intercept-RTS correlation is biased, as the paper's simulation does not test it.","rationale":"The reader's weakest assumption identifies the same point I would stress: the productivity process is the linchpin. The entire estimation design in equations (4)-(5) removes transmission bias only if the quadratic trend captures all productivity variation that firms know when choosing inputs. This is a strong functional-form restriction, and with short panels it is exactly the kind of assumption that can generate spurious negative correlation between intercepts and slopes because the intercept is essentially the time-average residual level after fitting firm-specific trends. The paper's own simulation (Section 5.1) validates the mean and standard deviation of the parameters but not the correlation, and it considers only a fixed-effect DGP, not persistent shocks. The negative correlation is the central substantive finding, so it needs to be stress-tested with a DGP that violates the quadratic assumption. If the placebo produces a comparable negative correlation from independent true parameters, the manuscript's headline is not identified; if not, the concern does not land. Until then, the CONDITIONAL verdict is appropriate. I am not raising objections to the EB machinery itself, which has coherent fixed-point justification and reasonable simulation performance for means and dispersions, provided the identifying assumptions hold; the issue is specifically whether the headline correlation survives misspecification of productivity dynamics. The dominance explanation in Section 6.1 is consistent but does not provide external validation; it illustrates why a negative correlation is economically plausible, not whether the estimate is unbiased.","tokens_in":14115,"tokens_out":4221,"duration_ms":43752,"concrete_test":"Run a Monte Carlo placebo using the Japanese sample structure (N=5,588, T=7) in which true firm-specific intercepts and returns to scale are drawn independently, true productivity follows an AR(1) process with persistence calibrated to plant TFP moments, and inputs are chosen after observing the current productivity shock (as in OP/ACF). Estimate the paper's quadratic-trend EB model and compute the posterior correlation between the intercept and RTS across firms. If the estimated correlation is strongly negative while the true correlation is zero, the headline claim is an artifact of the assumed dynamics; if it is near zero, the concern is resolved.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claim is the strong negative correlation between factor-neutral productivity (intercept) and returns to scale (Table 4). The identification of firm-specific elasticities in equations (4)-(5) rests on the assumption, stated in footnote 11, that the firm-specific quadratic trend fully captures the productivity dynamics relevant to input choices and that residual innovations eta_it are observed only after inputs are chosen. If true firm productivity follows a persistent process (e.g., an AR(1) as in Olley-Pakes), firms will select inputs partly in response to current productivity shocks, so the residual in equation (5) is correlated with k_it and l_it. With only T=7-12 years per firm, the within-firm variation used to identify beta_i and gamma_i is thin, and the estimated slopes will absorb the omitted productivity component; the resulting relationship between firm-level intercepts and slopes can be strongly negative even when the true technology parameters are independent. Section 5.1's simulation does not address this: it generates productivity as a pure firm fixed effect with alpha1=alpha2=0, so the possibility that omitted persistent shocks generate the Table 4 correlations is never tested. The correlations are also reported without standard errors, but the dynamics misspecification is the more load-bearing threat.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper develops an empirical Bayes (EB) estimator for heterogeneous Cobb-Douglas production functions, treating intercept, factor elasticities, and error variance as firm-specific and estimating their joint distribution nonparametrically via a discretized fixed-point approach. Using balanced manufacturer panels from Chile (1986–1996), Colombia (1978–1989), and Japan (2013–2019), the authors report substantial dispersion in factor-neutral productivity and output elasticities, with a strong negative correlation between the estimated intercept and returns to scale (about −0.8 in all three countries, Table 4). These correlations are found to persist under CES and intensive-CD specifications. The paper compares means with ACF Translog estimates, documents that most heterogeneity is within sectors and size classes, and shows implications for total technology productivity (TTP) and markup dispersion.","tokens_in":14362,"tokens_out":3615,"duration_ms":37656,"significance":"If valid, the headline result—that factor-neutral productivity and returns to scale are strongly negatively correlated—implies that conventional one-number TFP measures omit a key dimension of inter-firm technology variation, with direct consequences for markup estimation and misallocation analysis. The paper demonstrates the viability of an EB approach for joint estimation of heterogeneous production parameters, and it offers a transparent comparison with standard ACF Translog estimates on the same samples. The central empirical claim, however, is only as credible as the identifying assumption that firm-specific productivity dynamics are exactly captured by a quadratic trend; this is the main vulnerability that the current manuscript does not resolve.","major_comments":[{"comment":"The identification of firm-specific elasticities β_i and γ_i rests on the assumption that the firm-specific quadratic trend fully captures the productivity dynamics relevant to input choice, with innovations η_it observed only after input decisions. If true productivity follows a persistent process (e.g., an AR(1) as in Olley–Pakes), then k_it and l_it will be correlated with the residual ϕ_it in Eq. (5), and with only T=7–12 observations per firm the within-firm variation used to separate elasticities from productivity is thin. The resulting estimated slopes can absorb the omitted productivity component, and the correlation between estimated intercepts and returns to scale may be spurious even when the true technology parameters are independent. The authors should address this threat, either by relaxing the timing assumption, providing an alternative identification strategy, or showing through Monte Carlo evidence that the estimated correlation is not induced by this misspecification.","section":"Section 4, Eqs. (4)–(5) and footnote 11"},{"comment":"The simulation does not validate the central claim concerning the negative correlation between intercept and returns to scale. The DGP in Eq. (6) sets α1=α2=0 and generates productivity as a firm fixed effect, which is nested inside the maintained quadratic-trend assumption. Consequently, the simulation cannot speak to the possibility that omitted persistent productivity shocks generate the Table 4 correlations. I recommend an additional Monte Carlo exercise in which the true productivity process is a persistent, input-correlated shock (for instance, an AR(1) with input choices reacting to current productivity) and in which the true correlation between the intercept and returns to scale is set to zero. Reporting the EB estimates of that correlation under such a DGP would directly assess the skeptical scenario.","section":"Section 5.1, Eq. (6) and Table 5"},{"comment":"The headline correlations between α0 and β, γ, and β+γ are reported without any standard errors, confidence intervals, or other measures of sampling or posterior uncertainty. Given the EB framework, one can compute the posterior distribution of these correlations or use a bootstrap that resamples firms and/or the estimated type distribution. Without such measures, the reader cannot judge whether the ≈−0.8 correlations are distinguishable from, say, −0.5 or −0.3, nor whether they are statistically significant at all. This is a load-bearing statistic for the paper, so it should be accompanied by an appropriate uncertainty measure.","section":"Table 4 and Section 4.2"},{"comment":"The estimation samples are restricted to firms that survive the entire sample period (Chile 1986–1996, Colombia 1978–1989, Japan 2013–2019), and the Japanese sample additionally excludes 10% of observations as univariate and multivariate outliers via the BACON algorithm. The paper explicitly acknowledges that it does not address the Olley–Pakes selection issue. If survival is correlated with the technology parameters or with productivity dynamics, the estimated joint distribution—and specifically the negative intercept–RTS correlation—could be a selection artifact rather than a population feature. The authors should at least discuss the likely direction of selection bias, and if feasible, estimate the model on an unbalanced sample or examine the sensitivity of the correlation to the outlier trimming threshold.","section":"Section 3, balanced-panel screen and outlier trimming"}],"minor_comments":[{"comment":"The name of the country is spelled inconsistently (“Colombia” in the text and tables versus “Columbia” in several places, e.g., Section 6.2 and Table 8); please standardize to “Colombia”.","section":"General"},{"comment":"The header of Table 9 appears to duplicate “α1” in the column names (one column is presumably α2); please correct the label.","section":"Table 9"},{"comment":"The grid intervals used for discretizing α0, β, γ, α1, α2, and s are not fully specified in the text (e.g., footnote 5 says “intervals which contain most of the parameters’ density” without giving the actual bounds). Adding the exact grid support and a brief sensitivity analysis to alternative grid sizes/bounds would improve reproducibility.","section":"Section 4.1"},{"comment":"The simulation reports bias and MSE for means and standard deviations of α, β, γ, but not for the correlation between the intercept and returns to scale; given that correlation is the paper’s central finding, it should be included in the simulation evaluation.","section":"Section 5.1"},{"comment":"The “absence of dominance” explanation for the negative correlation is intuitive, but it is presented as a theoretical rationalization rather than a direct test; the paper could clarify that this is one possible interpretation, not a validation of the empirical result.","section":"Section 6.1 and Figure 5"}],"recommendation":"major_revision","confidential_remarks":"The manuscript is technically interesting and the empirical Bayes approach is novel in this application, but the central empirical result depends on an identifying assumption that is not adequately stress-tested. The authors should be asked to extend the simulation to a persistent-shock DGP and to report uncertainty for the headline correlations. The comparison with ACF Translog is a useful sanity check but does not resolve the dynamics concern. I see no grounds for rejection, but the revision needs to be substantive, not merely editorial."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"The headline result here is a consistently strong negative correlation (about -0.8) between firm-level factor-neutral productivity and returns to scale, in Chile, Colombia, and Japan, under Cobb-Douglas, CES, and intensive specifications. That is a big claim if true: it would mean one-number TFP measures miss most of the technology structure, and it would connect directly to technology-menu theory. The paper is also genuinely new in applying the rational-expectations Empirical Bayes machinery from the authors' companion paper to production function estimation, and in computing TTP and markups that exploit the joint distribution of intercepts and elasticities.\n\nWhat the paper does well: the estimation logic is laid out clearly, the comparison with ACF Translog is helpful, and the consistency across three countries with different development levels is striking. The authors are honest about the method's limitations, including the identifiability requirement and the dispersion under-estimation in their own simulation.\n\nThe soft spots are real and, in my view, load-bearing. First, Table 4 reports the key correlations without any standard errors or uncertainty quantification. For a claim this surprising, that is a non-negotiable gap. Second, the simulation in Section 5.1 checks bias and MSE of means and standard deviations but never asks whether the EB procedure recovers the correlation between intercept and returns to scale. That is exactly the parameter the paper's whole argument rests on, so the absence is worrying. Third, the quadratic-trend assumption for productivity dynamics is strong; if true productivity follows a persistent process, inputs respond to it and the within-firm variation used to identify the heterogeneous slopes will absorb the omitted component, potentially generating a spurious negative intercept-slope correlation. The simulation sidesteps this by generating productivity as a fixed effect, so it cannot speak to the threat. These are not minor quibbles; they directly concern whether the main empirical fact is an artifact.\n\nThe dominance explanation in Section 6.1 is more tautology than explanation—absence of dominance implies negative correlation, but that does not tell us why firms select such technologies. I would not hold that against the paper's core empirical contribution, but it should be framed as a theoretical consistency argument, not a test.\n\nWho gets value from this: researchers working on production function estimation, productivity dispersion, and markups. It deserves serious peer review, but the authors need to add uncertainty quantification, a simulation or placebo exercise aimed at the correlation, robustness to alternative productivity dynamics, and replication files before this is publishable. The raw idea is good; the current support is not yet there.","headline":"A genuinely interesting empirical finding—negative correlation between productivity intercepts and returns to scale—that is not yet nailed down because the headline correlations come without standard errors and the simulation never tests them.","tokens_in":14886,"tokens_out":1597,"would_cite":false,"duration_ms":17788,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"Using manufacturing panels from Chile, Colombia, and Japan, this paper argues that firms differ not only in factor-neutral productivity but also in capital and labor output elasticities, and that these two dimensions are strongly…","keywords":["empirical Bayes","production function","factor elasticities","factor-neutral productivity","returns to scale","total factor productivity dispersion","markup estimation","technology heterogeneity"],"falsifier":"Simulate a panel of firms whose true elasticities are drawn independently of true productivity, let productivity follow an AR(1) process that managers observe when choosing inputs, and then apply this paper's estimator, which forces a quadratic productivity trend. If the estimated correlation between the intercept and returns to scale is strongly negative while the true correlation is zero, the central discovery is an artifact of the quadratic-trend assumption.","tokens_in":13871,"feed_emoji":"🏭","tokens_out":8961,"duration_ms":78128,"temperature":0.7,"pith_summary":"The paper tries to establish that firm-level production technology varies along two dimensions that standard estimates treat as one. Using manufacturing panels from Chile, Colombia, and Japan, it estimates a Cobb-Douglas production function in which each firm has its own factor-neutral productivity and its own capital and labor elasticities. The estimates show wide dispersion in both dimensions and a strong negative correlation, roughly $-0.8$, between the productivity intercept and returns to scale. If that correlation is real, one-number productivity rankings, markup estimates, and misallocation diagnostics built on a shared technology are systematically off, because they conflate low factor-neutral productivity with low efficiency.","feed_headline":"Firms with high productivity tend to have lower returns to scale","feed_subtitle":"A correlation around -0.8 across three countries rewrites how markups and productivity dispersion are measured.","key_machinery":"The central object is a nonparametric Empirical Bayes estimator built on a rational-expectations prior. The firm-level parameters $(\\alpha_0, \\alpha_1, \\alpha_2, \\beta, \\gamma, s)$ are discretized into a finite grid of $15^3 \\times 6^3 = 729{,}000$ configurations for the Cobb-Douglas case. For each firm, the likelihood of each configuration is computed from the firm's own time series, and a prior probability vector over configurations is updated by Bayes rule. The estimator chooses the fixed point of the prior-to-posterior map that is coherent (prior equals posterior) and stable; this fixed point is the maximum likelihood estimate of the discrete joint distribution. Iterating the map, using an EM-like step, yields the posterior distribution, and firm-specific parameters are obtained as posterior means. The key work of this machinery is to estimate the joint distribution of all technology parameters without restricting the correlation pattern among them, so the negative correlation between $\\alpha_0$ and $\\beta+\\gamma$ is an estimated feature rather than an imposed one.","core_discovery":"On the paper's own terms, the central discovery is that the joint distribution of Cobb-Douglas technology parameters is much wider than standard estimators suggest, and that its two main components pull in opposite directions. Across the Chilean, Colombian, and Japanese samples, the intercept $\\alpha_0$ has a standard deviation roughly six times larger than the Translog estimate, the capital elasticity $\\beta$ about twice as large, and the labor elasticity $\\gamma$ two to three times as large. The correlation between the intercept and returns to scale $\\beta+\\gamma$ is $-0.828$ in Chile, $-0.887$ in Colombia, and $-0.819$ in Japan, and similar magnitudes appear under a CES specification and an intensive Cobb-Douglas specification. The paper reads this negative correlation as absence of technology dominance: because no single firm's production function dominates all others, many firms with different techniques can coexist in one sector, and measured capital-labor differences need not indicate misallocation.","pith_inferences":["If the negative correlation is a technology-menu equilibrium rather than an estimation artifact, then resource reallocation toward high-factor-neutral-productivity firms may be much less productivity-enhancing than standard misallocation calculations suggest, since those firms tend to have lower returns to scale.","A natural next test is to estimate the same Empirical Bayes model on gross-output data with materials, or on longer panels with a higher-order productivity process; the paper's own identifiability requirement, more time periods than parameters, says the test will need longer panels and coarser grids.","The correlation pattern predicts a specific cross-sectional fact: within an industry, firms with high estimated intercepts should use different capital-labor ratios from firms with low intercepts, and that input-basket prediction can be checked directly on the same datasets without re-estimating the model."],"forward_implications":["One-number total factor productivity rankings overstate productivity dispersion: once returns to scale are allowed to vary, the 90/10 ratio of Total Technology Productivity is about 3 to 4 in all three countries, an order of magnitude smaller than the 90/10 ratio of factor-neutral productivity alone.","Markup estimates built on homogeneous or Translog production functions understate markup dispersion; the Empirical Bayes labor-markup distributions have 90/10 ratios roughly double the Translog ones in Chile and Colombia.","Firm size and industry sector explain at most about 10 percent of the variance in estimated technology parameters, so sector-level or size-level production functions cannot substitute for firm-level technology heterogeneity.","If firms differ in both factor-neutral productivity and returns to scale, observed variation in capital-labor ratios need not be misallocation; it can reflect different technologies rather than distorted input choices."],"supporting_citations":[{"why":"Supplies the foundational argument that production parameters should be treated as random across firms.","marker":"Marschak and Andrews (1944)"},{"why":"Provides the rational-expectations fixed-point result that the Empirical Bayes estimator relies on.","marker":"Dardanoni and Demichelis (2024)"},{"why":"Provides the control-function production-function estimator used as the comparison benchmark for average elasticities and Translog-based markups.","marker":"Ackerberg, Caves and Frazer (2015)"},{"why":"Provides the alternative flexible productivity-process model that the paper's quadratic-trend assumption is compared against.","marker":"Ackerberg, Hahn and Pan (2022)"},{"why":"Earlier random-coefficient Cobb-Douglas estimates that already reported heterogeneity and discussed the negative-intercept puzzle.","marker":"Mairesse and Griliches (1988)"},{"why":"Provides the Total Technology Productivity measure used to show that productivity dispersion collapses once elasticity heterogeneity is counted.","marker":"Bernard and Jones (1996)"},{"why":"Argues that Translog output elasticities cannot capture the full degree of heterogeneity in input shares and that markup estimation requires more technology heterogeneity.","marker":"Raval (2023)"},{"why":"Documents large measured productivity dispersion across producers as the stylized fact the paper reinterprets.","marker":"Syverson (2011)"},{"why":"Suggests latent omitted factors as a possible explanation for the negative correlation, which the paper's dominance interpretation is meant to address.","marker":"Li (2021)"}],"fun_headline_variants":["High-productivity firms show lower scale returns","Productivity and returns to scale: strong inverse link","Wide technology gaps: no dominant production method","Heterogeneous elasticities reshape markup measures"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The entire result depends on assuming that the part of productivity that affects input choices is exactly a quadratic function of time, so that any remaining productivity shock arrives only after inputs are chosen.","fun_headline_variants_meta":{"raw":{"variants":["High-productivity firms show lower scale returns","Productivity and returns to scale: strong inverse link","Wide technology gaps: no dominant production method","Heterogeneous elasticities reshape markup measures"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000161,"raw_usage":{"total_tokens":1183,"prompt_tokens":843,"completion_tokens":340,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":459,"completion_tokens_details":{"reasoning_tokens":282}},"tokens_in":459,"tokens_out":340,"duration_ms":3759,"temperature":1.0,"reasoning_tokens":282,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-12T13:39:59.266246+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Simulate a panel of firms whose true elasticities are drawn independently of true productivity, let productivity follow an AR(1) process that managers observe when choosing inputs, and then apply this paper's estimator, which forces a quadratic productivity trend. If the estimated correlation between the intercept and returns to scale is strongly negative while the true correlation is zero, the central discovery is an artifact of the quadratic-trend assumption.","supporting_citations":[{"cited_title":"and Andrews, W","cited_arxiv_id":null,"evidence_quote":"Supplies the foundational argument that production parameters should be treated as random across firms."},{"cited_title":"Rational Expectations in Empirical Bayes","cited_arxiv_id":"2411.06129","evidence_quote":"Provides the rational-expectations fixed-point result that the Empirical Bayes estimator relies on."},{"cited_title":"A., Caves, K","cited_arxiv_id":null,"evidence_quote":"Provides the control-function production-function estimator used as the comparison benchmark for average elasticities and Translog-based markups."},{"cited_title":"A., Hahn, J","cited_arxiv_id":null,"evidence_quote":"Provides the alternative flexible productivity-process model that the paper's quadratic-trend assumption is compared against."},{"cited_title":"and Griliches, Z","cited_arxiv_id":null,"evidence_quote":"Earlier random-coefficient Cobb-Douglas estimates that already reported heterogeneity and discussed the negative-intercept puzzle."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Provides the Total Technology Productivity measure used to show that productivity dispersion collapses once elasticity heterogeneity is counted."},{"cited_title":"(2023) Testing the production approach to markup estimation, Review of Economic Studies, 90, 2592--2611","cited_arxiv_id":null,"evidence_quote":"Argues that Translog output elasticities cannot capture the full degree of heterogeneity in input shares and that markup estimation requires more technology heterogeneity."},{"cited_title":"(2011) What determines productivity?, Journal of Economic Literature, 49, 326--365","cited_arxiv_id":null,"evidence_quote":"Documents large measured productivity dispersion across producers as the stylized fact the paper reinterprets."}],"review_version":1}