{"id":"f4b50aea-03de-464e-9415-9193bb52ab1c","arxiv_id":"2507.05846","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":4.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":7,"one_line_summary":"A job-based 'hidden complexity' measure derived from O*NET skills and BLS occupations predicts US industry wages, productivity growth, and county GDP growth better than export-based 'revealed complexity'.","lead":"The paper builds a new measure of economic complexity for US industries and counties from the skills and occupations those industries employ, calling it 'hidden complexity'. It reports that this measure correlates with wages, productivity growth, and county GDP growth, while the standard export-based measure does not, suggesting a more direct capability-based indicator.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The headline comparison between hidden and revealed complexity rests on unmatched samples: model [1] uses 72 goods industries while hidden-complexity models use 74–220 industries, and no joint regression is reported.","rationale":"The reader's weakest-assumption analysis correctly identifies the non-nested comparison in Section 4.2 as the central vulnerability. The paper does provide a plausible construction of a job-based complexity measure and shows several interesting descriptive patterns, including the ability to cover services and non-exporting regions. However, the strongest claim in the abstract and Section 4.2 depends on a comparison between models estimated on different samples: 72 goods industries for the revealed measure versus 74–220 industries for the hidden measure. Without a joint regression or a matched-sample estimate, one cannot tell whether the difference in significance reflects the measures themselves or merely the sample composition. This is addressable with a straightforward econometric check, which is why the appropriate verdict remains conditional rather than rejection or acceptance. The concrete test proposed would either validate the headline or require the authors to weaken the comparative claim to a statement about coverage rather than predictive superiority.","tokens_in":12206,"tokens_out":3640,"duration_ms":45556,"concrete_test":"Reestimate Table 1 and Table 2 on exactly the 72 industries for which both QHidden and QRevealed can be computed, with identical control variables, and estimate specifications that include both complexity measures simultaneously. If QHidden remains statistically significant and QRevealed remains insignificant in the joint model on the common sample, the headline comparison survives; if not, the claim that the revealed measure is not associated with wages or productivity is unsupported. In addition, rerun Table 3 columns [1] and [3] on the common 2,079-county sample with both fitness measures included, to test whether the hidden fitness adds explanatory power beyond the exogenous export-based fitness.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claim of Section 4.2 is that the hidden, job-based measure is positively associated with wages and labor productivity growth, whereas the revealed, export-based measure is not. This claim is supported in Table 1 by comparing model [1], where QRevealed is tested on 72 goods-producing industries, with models [2]–[4], where QHidden is tested on 74 goods-producing industries, 146 service industries, and 220 total industries. The same structure appears in Table 2. Because the samples differ in composition and size, the comparison does not isolate the predictive power of the two measures. The revealed measure could be insignificant in model [1] simply because the goods-only sample is small or because services, where the hidden measure appears strong, are excluded from the revealed benchmark. No specification includes both QHidden and QRevealed in a single regression, and no common-sample estimate is reported. This is the load-bearing weakness of the paper's headline result: the superiority of the hidden measure is asserted from non-nested models rather than from a direct test. A secondary concern is that QHidden is a weighted average of job fitness scores that are themselves strongly wage-correlated, so its wage association is not surprising; however, the sample-matching issue is the decisive obstacle to the central comparison. The county-level results in Table 3 have the same structure, since model [3] loses about 1,000 counties relative to models [1] and [2], and the excluded counties are described as less developed.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"This paper constructs a 'hidden' job-based economic complexity measure from a four-layer network: O*NET skills to occupations, BLS occupations to industries, and BLS industries to US counties, with UN COMTRADE exports used for the standard 'revealed' measure. The hidden industry complexity is the employment-weighted average of job fitness scores from the economic fitness algorithm, and county fitness is the sum of hidden complexities over industries with wage location quotient above one. The authors regress industry wage levels and labor productivity growth (2017–2022) on the hidden and revealed complexity measures, and county GDP per capita growth on three fitness measures. They report that the hidden measure is significantly associated with wages and productivity growth while the export-based revealed measure is not, and that all county fitness measures are positively associated with growth.","tokens_in":12538,"tokens_out":5727,"duration_ms":66857,"significance":"If the central comparison were established, the paper would offer a capability-based complexity measure that covers services and non-exporting counties and avoids some numerical pathologies of the standard fitness algorithm. The paper is transparent about data construction and provides falsifiable predictions, and the use of O*NET/BLS data is a constructive step. However, the headline finding of hidden-measure superiority is currently supported only by unmatched-sample comparisons, and the county-level results actually show significant coefficients for all three measures; these issues must be resolved before the paper's claims can be accepted.","major_comments":[{"comment":"The central claim that the hidden measure is associated with wages and productivity growth while the revealed measure is not rests on a comparison of non-nested regressions with different samples: model [1] uses 72 goods-producing industries for QRevealed, whereas models [2]–[4] use 74 goods-producing, 146 service, and 220 total industries for QHidden. This design does not isolate the predictive power of the two measures, because the revealed coefficient could be insignificant in the smaller goods-only sample even if the measures were equally predictive on a common sample. Please report a common-sample regression on the 72 industries for which both measures are available, a specification including QHidden and QRevealed jointly, and a goods-only hidden-complexity model so that samples are matched.","section":"Section 4.2, Tables 1 and 2"},{"comment":"The conclusion states that 'the revealed complexity shows no statistical significance,' but Table 3 model [3] reports a positive and strongly significant coefficient on the exogenous export-based fitness (2.691, s.e. 0.544). The sentence is therefore internally inconsistent with the reported county-level results unless it is explicitly restricted to the industry wage and productivity regressions of Tables 1 and 2. In addition, model [3] drops about 1,000 counties relative to models [1] and [2], and the text acknowledges that the excluded counties are on average less developed; this sample selection further weakens any comparison of the three county-level measures. Please clarify the scope of the claim and provide comparable-sample estimates.","section":"Section 4.4 and Table 3"},{"comment":"The binarization thresholds (skill importance above the skill average, IWQ>1, WLQ>1, RCA>1), the choice of the 2017–2022 window, and the exclusion of 30 industries are not subjected to sensitivity analysis. Because the headline result is a comparison of significance across measures, it would be important to show that the conclusion is robust to reasonable variations in these choices, at least for the main specifications in Tables 1 and 2.","section":"Sections 2.1–2.3 and 4.2"},{"comment":"Because QJB_i is a weighted average of job fitness scores that in [31] are explicitly designed to predict wages, the strong association in Table 1 may partly reflect a mechanical link rather than a distinct complexity channel. Please compare the hidden measure against simpler occupational skill or wage aggregates (e.g., average occupational wage or average skill level) in the same regressions, to demonstrate that the complexity aggregation adds predictive content beyond its inputs.","section":"Section 3.2 and Table 1"}],"minor_comments":[{"comment":"There is a typo: 'Morever' should be 'Moreover'.","section":"Section 4.3"},{"comment":"The AIC and BIC values are not comparable across models with different numbers of observations; please state this explicitly or restrict the information-criterion comparison to models estimated on the same sample.","section":"Tables 1 and 2"},{"comment":"The figure excludes four outliers without specifying the selection rule; please state the rule or show the outliers in an inset.","section":"Figure 2"},{"comment":"The notation '*** significant at 1‰ level' is nonstandard; consider using '0.1%' for clarity.","section":"Tables 1 and 2"},{"comment":"The text reports 435 occupations in OEWS versus 439 in O*NET; the matching procedure between these classifications should be described.","section":"Section 2.2"}],"recommendation":"major_revision","confidential_remarks":"To the editor: The manuscript's central claim is not yet supported due to the sample-matching problem in Tables 1 and 2, and the conclusion overstates the county-level results. The issues are fixable with additional regressions, so I recommend major revision rather than rejection. I also note that the underlying job-fitness measure comes from [31], co-authored by one of the present authors; the manuscript should clarify the incremental contribution beyond that paper, but this is not a reason for rejection."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"What you should know: this paper builds a plausible new measure of industry and county \"hidden complexity\" from O*NET job-skill data, and it shows that measure tracks wages and productivity growth. The part I do not yet believe is the headline claim that it beats export-based revealed complexity, because that claim comes from comparing regressions on different samples rather than from a head-to-head test.\n\nThe construction is genuinely useful. The authors compute a job fitness from the skill-occupation network, average it over BLS employment weights to get industry complexity, and sum over counties. That gives a complexity measure that covers services and counties with little manufacturing, and it avoids the degenerate, multimodal distributions that plague the endogenous fitness algorithm. The regressions are run on public data with sensible controls, and the examples where hidden and revealed diverge (Pharma, Aerospace, Oil and Gas) are informative. This is a real extension of Lo Turco and Maggioni and Buyukyazici et al., not a repackaging.\n\nThe soft spot is exactly where the stress-test note points. Tables 1 and 2 compare model [1] with QRevealed on 72 goods industries against models [2]-[4] with QHidden on 74, 146, and 220 industries. The revealed measure is insignificant in its small goods-only sample; the hidden measure is significant on the goods-only sample too. But no specification includes both measures, and no common-sample regression is shown. So the sentence \"whereas the classic revealed measure is not\" is stronger than the evidence. The revealed measure could be insignificant for other reasons, and hidden could be adding nothing once revealed is controlled for. That is fixable: run the 72-industry sample with both measures, and maybe a horserace on the goods sample. The county-level Table 3 has the same issue in milder form: the exogenous fitness model loses about a thousand counties, described as less developed, so the comparison across models is not clean. Also, thresholds and sample exclusions get no robustness checks, and no code or data are posted.\n\nThis is a paper worth engaging with. The measure itself is a contribution, and the comparison can be repaired. I would send it to peer review and ask for a matched-sample joint test before the headline claim can stand.","headline":"A useful new job-based complexity measure, but the headline claim that it beats export-based revealed complexity rests on unmatched sample comparisons and needs a joint test before I'd trust it.","tokens_in":13026,"tokens_out":2709,"would_cite":true,"duration_ms":32443,"reading_group":"yes","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"A job-based complexity measure built from occupational skills predicts wages and productivity growth where export-based complexity does not.","keywords":["hidden complexity","job-based complexity","economic complexity","occupational skills","fitness algorithm","labor productivity","regional growth","human capital"],"falsifier":"Run one regression of log 2017 average compensation and one of 2017–2022 labour productivity growth on the 72 industries where both $Q^{\\mathrm{Hidden}}$ and $Q^{\\mathrm{Revealed}}$ are available, including both measures and the same controls; the paper's comparative claim survives only if the revealed coefficient stays statistically insignificant while the hidden coefficient stays significant. A county-level analogue would regress GDP per capita growth on both hidden and export-based fitness for the subset of counties where both exist.","tokens_in":12001,"feed_emoji":"💼","tokens_out":5723,"duration_ms":62662,"temperature":0.7,"pith_summary":"The paper argues that the capabilities behind economic complexity can be measured directly from the skills that jobs require, rather than inferred from what a country exports. It builds a job-based 'hidden complexity' score for 220 US industries from occupational skill data, then aggregates it to US counties. The central finding is that this hidden score is positively associated with industry wages and labour productivity growth, and with county GDP per capita growth, while the traditional export-based 'revealed' measure shows no significant association. If right, complexity analysis could extend to services and to regions with little manufacturing, where export-based measures fail.","feed_headline":"Job skills beat exports for measuring economic complexity","feed_subtitle":"Occupation-based scores predict wages, productivity growth, and county GDP where export-based complexity does not.","key_machinery":"The central object is a four-layer network—skills, jobs, industries, and counties—connected by three bipartite matrices $M^{(1)}$, $M^{(2)}$, and $M^{(3)}$. On the skill-job layer, the paper runs the Economic Fitness and Complexity algorithm: job fitness is a complexity-weighted sum of required skills, and skill complexity is a nonlinear function of the fitness of the jobs that require it, iterated to a fixed point. High-fitness jobs require many rare, complex skills. Industry complexity $Q^{\\mathrm{JB}}_i$ is the employment-weighted average of job fitnesses over the jobs in that industry, using an Industry Wage Quotient threshold to define the job-industry links; county fitness $F^{\\mathrm{JB}}_c$ is the sum of the complexities of the industries in which the county has a Wage Location Quotient above one. The machinery does the work of replacing an unobservable capability layer with an observable human-capital layer, then lets the same algorithmic logic used for exports operate on skills instead of outputs.","core_discovery":"On the paper's own terms, the discovery is that capabilities—treated in the literature as an unobservable layer between territories and activities—can be approximated by the skill content of occupations. The authors compute a fitness for each occupation by applying the Economic Fitness and Complexity algorithm to the skill-occupation network, so that a job is complex if it requires many rare, complex skills. The Job-Based Complexity of an industry is the employment-weighted average fitness of its jobs, and the Job-Based Fitness of a county is the sum of the complexities of the industries in which it has a wage-location-quotient advantage. Across 220 industries, this hidden complexity is positively and significantly related to 2017 average compensation and to 2017–2022 labour productivity growth; the revealed, export-based complexity, computable for only 72 goods-producing industries, is not significant in either regression. At the county level, the job-based fitness is positively and significantly related to real GDP per capita growth over 2017–2022, with diversification separately controlled.","pith_inferences":["If the hidden measure is the truer capability signal, then industries where the two measures disagree—such as Pharmaceuticals and Aerospace, scored high by hidden and low by revealed complexity—deserve a re-examination of how complexity rankings are used in policy.","The paper does not test international portability, but because the job-based measure needs only employment and occupational data, it could in principle be computed for subnational units and service economies that lack export data.","The wage plateau at high hidden complexity suggests diminishing returns to skill complexity; a direct extension would split industries by complexity quartile and test whether the wage elasticity falls at the top.","The paper does not include both complexity measures in a single regression, so a natural next step is to test whether hidden complexity retains its predictive power once revealed complexity is held constant."],"forward_implications":["Hidden complexity can be computed for service industries and for counties with little or no manufacturing, extending complexity analysis beyond goods trade.","Job-based complexity is a statistically significant predictor of 2017 wage levels across goods-producing industries, services, and all industries combined, with wages first rising and then plateauing at higher complexity.","Job-based complexity predicts 2017–2022 labour productivity growth out of sample, for both goods and services, while export-based complexity does not.","County job-based fitness is positively associated with 2017–2022 real GDP per capita growth, while diversification alone is negatively or insignificantly related to growth.","The hidden measure produces a smoother, better-behaved distribution of county fitness values than the endogenous fitness algorithm, avoiding the multimodal and zero-clustered values that complicate regression analysis."],"supporting_citations":[{"why":"Supplies the Economic Fitness and Complexity algorithm whose fixed point defines job fitness and industry complexity.","marker":"[8]"},{"why":"Establishes the fitness-of-jobs approach on the skill-job network that the paper extends to industries and counties.","marker":"[31]"},{"why":"Supplies the Location Quotient normalization that the Industry Wage Quotient adapts to wages.","marker":"[32]"},{"why":"Defines Revealed Comparative Advantage, which motivates the threshold-based binarization of the trade and wage networks.","marker":"[33]"},{"why":"Provides the wage-based Location Quotient formula used to build the industry-county network.","marker":"[34]"},{"why":"Provides the convergence criteria used to stop the fitness algorithm iterations.","marker":"[36]"},{"why":"Defines the exogenous fitness from export data that the paper compares against hidden fitness.","marker":"[37]"},{"why":"Provides the HS-to-NAICS concordance used to map export-based product complexity onto industries.","marker":"[38]"}],"fun_headline_variants":["Job skills beat exports for economic complexity","Hidden job complexity predicts wages and growth","Occupation skills outperform exports in complexity","Export complexity fails where job skills succeed"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The paper's comparative claim that revealed complexity is not associated with wages or productivity growth rests on comparing a regression on 72 goods-producing industries with regressions on 74 to 220 industries, without ever including both complexity measures in the same regression or matching the sample.","fun_headline_variants_meta":{"raw":{"variants":["Job skills beat exports for economic complexity","Hidden job complexity predicts wages and growth","Occupation skills outperform exports in complexity","Export complexity fails where job skills succeed"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.00102,"raw_usage":{"total_tokens":4257,"prompt_tokens":853,"completion_tokens":3404,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":469,"completion_tokens_details":{"reasoning_tokens":3354}},"tokens_in":469,"tokens_out":3404,"duration_ms":22652,"temperature":1.0,"reasoning_tokens":3354,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-06T19:16:40.336075+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Run one regression of log 2017 average compensation and one of 2017–2022 labour productivity growth on the 72 industries where both $Q^{\\mathrm{Hidden}}$ and $Q^{\\mathrm{Revealed}}$ are available, including both measures and the same controls; the paper's comparative claim survives only if the revealed coefficient stays statistically insignificant while the hidden coefficient stays significant. A county-level analogue would regress GDP per capita growth on both hidden and export-based fitness for the subset of counties where both exist.","supporting_citations":[{"cited_title":"A new metrics for countries’ fitness and products’ complexity","cited_arxiv_id":null,"evidence_quote":"Supplies the Economic Fitness and Complexity algorithm whose fixed point defines job fitness and industry complexity."},{"cited_title":"Mapping job fitness and skill coherence into wages: an economic complexity analysis","cited_arxiv_id":null,"evidence_quote":"Establishes the fitness-of-jobs approach on the skill-job network that the paper extends to industries and counties."},{"cited_title":"The location quotient approach to estimating regional economic impacts","cited_arxiv_id":null,"evidence_quote":"Supplies the Location Quotient normalization that the Industry Wage Quotient adapts to wages."},{"cited_title":"Tariff protection in industrial countries: an evaluation","cited_arxiv_id":null,"evidence_quote":"Defines Revealed Comparative Advantage, which motivates the threshold-based binarization of the trade and wage networks."},{"cited_title":"Economic development and wage inequality: A complex system analysis","cited_arxiv_id":null,"evidence_quote":"Provides the wage-based Location Quotient formula used to build the industry-county network."},{"cited_title":"On the convergence of the fitness-complexity algorithm","cited_arxiv_id":null,"evidence_quote":"Provides the convergence criteria used to stop the fitness algorithm iterations."},{"cited_title":"Dynamics in the fitness-income plane: Brazilian states vs world countries","cited_arxiv_id":null,"evidence_quote":"Defines the exogenous fitness from export data that the paper compares against hidden fitness."},{"cited_title":"A concordance between ten-digit us harmonized system codes and sic/naics product classes and industries","cited_arxiv_id":null,"evidence_quote":"Provides the HS-to-NAICS concordance used to map export-based product complexity onto industries."}],"review_version":1}