{"id":"65285640-ea17-4773-9c50-f95900e9d83a","arxiv_id":"2508.21699","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":4.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":3,"one_line_summary":"In a stochastic multi-output setting, fixed-proportion Leontief production yields curved isoquants and non-constant returns to scale, supporting its use in the Thanzi la Onse health system model.","lead":"This paper shows that a simple Leontief production function, which fixes input proportions per treatment, generates smooth non-linear relationships between resources and treatments when placed inside a stochastic multi-disease health system model. That result supports using data-frugal production functions in large health system models such as the Thanzi la Onse model for Malawi.","discovery_kind":"new_application","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Eq. (5) assumes y2 is exogenous to inputs w,c, but in TLO y2 is an endogenous output sharing the same resources; the claimed CES-like richness may not transfer.","rationale":"The paper's strongest claim is that embedding a Leontief production technology in the multi-output, stochastic, agent-based TLO model removes the restrictive textbook properties (constant returns, piecewise-linear isoquants) and yields a CES-like rich structure. The mathematical derivation of Eq. (5) is correct for the stated toy problem: if y2 is an exogenous random variable with fixed density g independent of w,c, then E[y1] is indeed a concave function whose level sets can be smooth and whose returns to scale can be non-constant. The load-bearing question is whether Eq. (5) describes the TLO model's actual production process. The reader's weakest assumption pinpoints exactly this: y2 is treated as an independent draw, but in the TLO model y2 (competing treatments) is an endogenous output generated from the same resources, and disease burden responds dynamically to past treatments. Thus the distribution of y2 will generally depend on w,c and on y1 itself. If that dependence is present, then Eq. (5) is not the correct aggregate production function, and the non-linear isoquants and increasing returns shown in Figures 4-5 may be artifacts of the exogenous-shock assumption rather than genuine properties of the Leontief framework within TLO. This concern is more fundamental than the discrete-versus-continuous issue: even with a continuous distribution, the exogeneity gap undermines the transfer from toy model to TLO. The paper has independent support for the elementary mathematics, and the qualitative phenomenon of aggregation smoothing kinks is plausible, but the central claim about the TLO model specifically is not demonstrated. The reader's CONDITIONAL verdict is appropriate: authors should qualify the claim, provide evidence that competing outputs are exogenous or that the results survive endogenization, and engage the aggregation literature. No change to the reader's verdict is needed.","tokens_in":4840,"tokens_out":8423,"duration_ms":107051,"concrete_test":"Use the TLO model (or a stripped-down two-disease ABM with its allocation rules) to simulate a grid of (w,c) levels, holding everything else fixed. Record realized y1 and y2 across stochastic iterations. (i) Regress y2 on w and c (and on lagged y1) and test for nonzero coefficients; if y2 responds to inputs, the independence assumption in Eq. (5) fails. (ii) Estimate E[y1|w,c] nonparametrically from the simulations and compare the shape of its isoquants and the local returns-to-scale elasticity to Eq. (5) evaluated with the empirical marginal g(y2). If the estimated isoquants have kinks or the returns-to-scale elasticity is constant, the paper's claim that the Leontief framework becomes a CES-like rich structure in the TLO model is not supported.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central demonstration (Eq. 5) computes E[y1] by integrating the residual Leontief function over a density g(y2) that is taken to be independent of the inputs w,c and of y1. But in the TLO model y2 is not an exogenous shock: it is the quantity of treatment delivered for competing diseases, generated by the same agent-based simulation and consuming the same resources w,c. Its distribution will generally shift with input levels (more resources allow more y2 treatment) and with the history of y1 (disease burden responds to prior treatment). Hence the correct object is E[min((w - Y2/b1)/a1, (c - Y2/b2)/a2) | w,c], with Y2 and y1 jointly determined; Eq. (5) is a special case that the paper does not justify. Without independence, the smoothed isoquants and non-constant returns may be artifacts of the simplifying assumption rather than properties of the TLO production system. The paper needs either to show that Y2 in TLO is (approximately) exogenous to w,c, or to re-derive the aggregate production function with the feedback and demonstrate the properties still hold.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"This paper argues that the Leontief production function, normally associated with constant returns to scale and piecewise linear isoquants, acquires richer properties when embedded in a multi-output, stochastic health system model such as Thanzi la Onse. The authors analyze a two-output case: y1 is produced with fixed input coefficients a1, a2 from w and c, while a second output y2 consumes fixed amounts 1/b1 and 1/b2 of the same resources. Treating y2 as random with density g(y2), expected y1 is the integral in Eq. (5), and numerical examples show curved isoquants and locally non-constant returns. A three-output extension with correlated y2 and y3 via an Ali-Mikhail-Haq copula illustrates further curvature. The conclusion is that HSMs can use the data-parsimonious Leontief form without inheriting the restrictive properties of a single-output Leontief technology.","tokens_in":5103,"tokens_out":6531,"duration_ms":74884,"significance":"If the claims hold for the actual TLO model, this is a practically important observation: it suggests that a fixed-proportion micro-technology, which is much easier to calibrate from available health-system data than a CES or translog, can generate flexible aggregate production relationships in an agent-based, stochastic, multi-disease setting. The paper's derivation is transparent and self-contained, and the figures are generated by a supplied Mathematica notebook, which supports reproducibility. The main risk is that the smoothing result rests on treating the competing output y2 as exogenous with a fixed distribution, an assumption that is not obviously satisfied when y2 is produced within the same simulation. If that assumption can be justified or the result re-derived endogenously, the paper would make a useful conceptual contribution to health system modeling.","major_comments":[{"comment":"The central aggregation step, Eq. (5), integrates the residual Leontief function with respect to a density g(y2) that is independent of inputs w and c and of y1. This exogeneity assumption is not defended. In the TLO model y2 is not an exogenous random draw: it is the quantity of treatment delivered for competing diseases, generated by the same agent-based simulation and consuming the same resources. Its distribution will generally shift with w and c (more resources permit more y2 treatment) and with the history of y1 (treated disease burden changes future demand). The correct object would be E[min((w-Y2/b1)/a1, (c-Y2/b2)/a2) | w,c] with Y2 jointly determined; Eq. (5) is a special case. Unless the authors can show that Y2 is approximately exogenous in TLO, or re-derive the aggregate production function with feedback and verify that the curved isoquants and non-constant returns survive, t","section":"Section 3, Eq. (5)"},{"comment":"The abstract states that once the Leontief technology is incorporated into an agent-based model, constant returns to scale are no longer present. This is only true at finite scale. Holding y2 fixed in Eq. (4), doubling (w,c) more than doubles y1 for small expansions, but as (w,c) grow large the fixed deductions y2/b1 and y2/b2 become negligible and E[y1] behaves asymptotically as min(w/a1, c/a2), i.e., constant returns are restored in the limit. The unqualified claim in the abstract and conclusion should be corrected to: non-constant returns at finite scale, with CRS as an asymptotic property. This is not merely a wording issue, because the practical question of returns to health-system investment depends on the scale of expansion relative to the size of competing demands.","section":"Abstract; Section 3"},{"comment":"The smoothness of the isoquants in Figures 4 and 5 (and the claimed loss of piecewise linearity) is obtained by integrating a continuous density g(y2); the Technical notes state Uniform[0,1] throughout. The TLO model, however, is agent-based and generates integer counts of treatments, so Y2 has a discrete distribution. With discrete Y2, E[y1] is a sum of min(affine) terms and remains a concave piecewise-linear function of (w,c); its level sets retain kinks, although the kink structure can be richer than in the one-output case. The paper does not establish that the continuous-density case approximates the discrete agent-based system. The authors should either justify a continuous approximation of Y2 or qualify the claim that piecewise linear level sets disappear.","section":"Technical notes; Section 3"}],"minor_comments":[{"comment":"The CES formula appears to have a typo: (1-a)^rho should presumably be (1-a)c^rho; as printed, the function is not a valid CES form.","section":"Eq. (2)"},{"comment":"The sentence stating that w and c are reduced by 1/b1 y1 and 1/b2 y1 should read y2, not y1.","section":"Text after Eq. (4)"},{"comment":"The phrase 'Figure illustrates this' appears without a figure number.","section":"Section 3, Figure 3"},{"comment":"The caption for panel (b) repeats 'With positive second output y2' from Figure 3; it should describe the correlation between y2 and y3 or otherwise distinguish the three-output case.","section":"Figure 5"}],"recommendation":"major_revision","confidential_remarks":"The mathematical example is internally coherent and the reproducibility of the figures is a plus. The decisive issue is whether the exogeneity assumption in Eq. (5) can be defended in TLO. If the authors can add even a small numerical experiment using TLO's actual resource allocation, or a formal extension with endogenous Y2, the paper would be much stronger and could be acceptable. Otherwise the title and abstract promise more than the derivation supports."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"The useful core is simple and correct: if you average a fixed-proportion production function over an independent random competing output, the expected output has smooth, non-linear isoquants and non-constant returns to scale at finite input levels. Eq. (5) does exactly that, and the math checks out under the stated conditions. That is worth knowing for anyone who wants to defend a Leontief technology in a stochastic multi-output setting.\n\nThe soft spots are about scope and framing. The biggest one is the independence assumption. In the TLO model, y2 is not an exogenous draw: it is an endogenous output consuming the same resources, and its distribution will shift with input levels and disease dynamics. Eq. (5) treats g(y2) as fixed. The paper needs to either justify approximate exogeneity in TLO or re-derive the result with feedback. The abstract's claim that 'these properties are no longer present' in the agent-based model goes well beyond what the two-output toy establishes.\n\nSecond, the aggregation literature matters here. Houthakker (1955) showed that averaging Leontief-type activities over a distribution yields smooth aggregate production functions. The paper doesn't cite that, which makes the novelty claim stronger than it should be. The mechanism is not new; the specific health-system application is.\n\nThird, the paper says 'non-constant returns to scale' but doesn't mention that constant returns are restored asymptotically as w and c grow, because the fixed-support random draw becomes relatively negligible. That's a real qualification, not a nitpick, and it should be in the abstract.\n\nReproducibility is weak: the Mathematica notebook is referenced but no URL or parameters are given. The discrete-count issue is minor for the illustration but matters for the TLO claim.\n\nWho is this for? Health system modelers who want a conceptual justification for using Leontief technology without inheriting textbook restrictions. It delivers that only under the independence assumption. A serious referee should engage it, mainly to demand the exogeneity discussion and the Houthakker citation. With those revisions, the paper would be a useful conceptual note. My own verdict: conditionally accept after major revision, and I'd bring it to a reading group if the authors address the feedback loop.","headline":"A correct but overclaimed note: averaging a Leontief function over an independent random competing output does smooth isoquants and break constant returns, but the independence assumption is load-bearing and the paper overstates the transfer to the full TLO model.","tokens_in":5624,"tokens_out":2259,"would_cite":false,"duration_ms":30356,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":["91B38"],"pacs":[],"model":"deepseek-v4-flash","headline":"A fixed-proportion Leontief technology stops being rigid once random competing demands are averaged over: expected output gains curved isoquants and non-constant returns to scale, making the simplest production function rich enough for whol","keywords":["Leontief production function","health system modeling","returns to scale","isoquants","agent-based model","Thanzi la Onse","Malawi","CES production function"],"falsifier":"Compute E[y1] in Eq. (5) with a discrete, count-valued distribution for y2, as the agent-based model actually generates, instead of a continuous density on [0,1]; if isoquants stay visibly kinked at realistic scales, the CES-like smoothness is an artifact of continuity. Separately, run the model with disease feedback on and off and compare realized mean treatments across input combinations: if the realized surface departs from Eq. (5) as feedback strengthens, the fixed-density assumption is the source of the claimed flexibility.","tokens_in":4674,"feed_emoji":"📈","tokens_out":11905,"duration_ms":122844,"temperature":0.7,"pith_summary":"This paper argues that the Leontief production function — the fixed-recipe technology in which each treatment needs a set amount of staff time and consumables — is not the rigid tool of economics textbooks once it is placed inside a multi-disease, agent-based health system model. The reason is that competing treatments consume the same resources, and the volume of competing demand arrives at random: averaging over that randomness turns the 'min' of two linear constraints into a smooth, curved relationship between inputs and treatments. The authors prove this in the two-output case with the expected-value formula Eq. (5), and show that adding a third, correlated output makes the level sets richer still. If the claim is right, health system modelers can reproduce realistic non-linear responses to investment — non-constant returns to scale, curved isoquants — using a production function whose parameters can be measured from existing data, rather than the CES form that an all-disease model cannot estimate.","feed_headline":"Fixed-ratio production turns non-linear inside agent-based models","feed_subtitle":"With demand varying per disease, one simple averaging step gives non-linear returns and curved input trade-offs.","key_machinery":"The load-bearing object is Eq. (5): E[y1] = ∫ min((w − y2/b1)/a1, (c − y2/b2)/a2) g(y2) dy2, the expectation of the fixed-proportion production function after a random competing output has consumed resources. Two mechanisms act together: the competing output y2 enters as a negative input, breaking constant returns to scale, and integration over its density g smooths the kink of the min function, replacing the piecewise-linear isoquant with a curved level set. A third output, correlated with the second through a joint distribution (in the technical notes, an Ali–Mikhail–Haq copula), adds further curvature.","core_discovery":"The paper's claim is that the inflexibility of Leontief technology is an artifact of studying one output in isolation. With a competing output y2 consuming the same inputs, the first output is the min of two residual constraints; y2 acts as a negative input, so doubling inputs more than doubles y1. When y2 is random with density g, the expected output in Eq. (5) is non-linear in inputs: isoquants curve and returns to scale vary, 'analogous to the properties of a more general production function such as the CES.' A third, correlated output pushes the level sets further from piecewise-linear. Conclusion: inside a multi-output, stochastic, agent-based model, fixed-proportion technology is a ric","pith_inferences":["The smoothing mechanism is general: the expectation of a min of affine functions over any continuous distribution is concave and smooth, so any multi-output stochastic setting with shared capacity — hospital wards, operating theatres, or manufacturing lines — would show the same aggregation-restores-flexibility effect; the mathematics is not health-specific.","The independence assumption in Eq. (5) is an idealization: in the live model, disease burdens and hence treatment demand respond to past care, so the distribution of y2 is not truly fixed. The direction of bias is testable — if past treatment of the competing disease lowers future demand for it, realized expected output will be more responsive to inputs than Eq. (5) predicts.","Smoothness depends on continuity: with the discrete counts an agent-based model actually produces, E[y1] stays a concave piecewise-linear function and isoquants keep kinks; whether those kinks matter at realistic scales is a quantitative question the paper's uniform-density figures do not settle.","A direct test: fit a small simulator with two or three diseases and compare its realized input–output surface against the CES-like shape predicted by Eq. (5); agreement would justify estimating aggregate substitution elasticities from micro-level fixed-proportion parameters without econometric estimation."],"forward_implications":["A health system model can use rigid per-treatment input requirements at the micro level and still generate smooth, non-linear aggregate production relationships — curved isoquants and non-constant returns — when many stochastic disease demands share the same staff and consumables.","The returns to health investment seen in the model will generally not be constant: doubling all resources changes expected treatments by more or less than double, depending on the distribution of competing demands.","The practical payoff is parameter frugality: calibrating a Leontief technology needs only per-treatment input requirements, which existing data sources can supply, whereas a CES function needs factor shares and substitution elasticities that an all-disease model cannot estimate.","As the model scales up to more outputs, the production structure becomes richer rather than more constrained: correlated random demands push the aggregate isoquants further from piecewise linearity.","The framework gives modelers a defensible way to examine the returns to health system strengthening: the non-linear expected-output surface is exactly the object that maps resource budgets into treatments delivered."],"supporting_citations":[{"why":"Presents the Thanzi la Onse model for Malawi, the all-disease agent-based setting in which the Leontief technology is embedded and whose properties are at issue.","marker":"Hallett et al., 2025"},{"why":"Supplies the production-economics concepts — production function, isoquants, returns to scale — that define the paper's terms and its comparison.","marker":"Rasmussen, 2012"},{"why":"Cited with the CES lineage as the flexible functional-form benchmark whose data requirements motivate the Leontief choice.","marker":"Solow, 1956"},{"why":"The original CES formulation, the flexible alternative the paper contrasts with Leontief and argues cannot be parameterized at all-disease scale.","marker":"Arrow et al., 1961"},{"why":"Extends the CES analysis and is part of the same comparator the paper argues is infeasible to estimate for a whole health system.","marker":"McFadden, 1963"},{"why":"Provides the copula used in the technical notes to generate correlated competing outputs in the three-output illustration of richer isoquants.","marker":"Ali et al., 1978"}],"fun_headline_variants":["Fixed proportions yield non-linear returns in multi-disease models","Leontief's rigidity dissolves in stochastic multi-output systems","Fixed input ratios become flexible inside agent-based health models","When fixed-proportion tech gets a stochastic twist, it mimics CES","Multi-disease dynamics turn rigid Leontief into a rich production curve"],"cache_read_input_tokens":2688,"weakest_assumption_plain":"The argument assumes the demand for competing treatments follows one fixed chance pattern that does not change when more staff or supplies are added or when more of the main treatment is produced; in a real disease model, treatment history changes future demand, so that fixed pattern is an approximation.","fun_headline_variants_meta":{"raw":{"variants":["Fixed proportions yield non-linear returns in multi-disease models","Leontief's rigidity dissolves in stochastic multi-output systems","Fixed input ratios become flexible inside agent-based health models","When fixed-proportion tech gets a stochastic twist, it mimics CES","Multi-disease dynamics turn rigid Leontief into a rich production curve"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000519,"raw_usage":{"total_tokens":2323,"prompt_tokens":687,"completion_tokens":1636,"prompt_tokens_details":{"cached_tokens":256},"prompt_cache_hit_tokens":256,"prompt_cache_miss_tokens":431,"completion_tokens_details":{"reasoning_tokens":1551}},"tokens_in":431,"tokens_out":1636,"duration_ms":14388,"temperature":1.0,"reasoning_tokens":1551,"cache_read_input_tokens":256,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-05T14:02:37.747495+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Compute E[y1] in Eq. (5) with a discrete, count-valued distribution for y2, as the agent-based model actually generates, instead of a continuous density on [0,1]; if isoquants stay visibly kinked at realistic scales, the CES-like smoothness is an artifact of continuity. Separately, run the model with disease feedback on and off and compare realized mean treatments across input combinations: if the realized surface departs from Eq. (5) as feedback strengthens, the fixed-density assumption is the source of the claimed flexibility.","supporting_citations":[],"review_version":1}