Nonparametric Identification and Estimation of Production Functions Invariant to Productivity Dynamics

Abstract: Production function estimates underpin the measurement of firm-level markups, allocative efficiency, and the productivity effects of policy interventions. Since Olley and Pakes (1996), every major proxy variable estimator has identified the production function through a first-order Markov assumption on unobserved productivity; I show that misspecification of this assumption generates persistent upward bias in the materials elasticity that propagates into overestimated markups and inflated treatment effects. I replace the Markov restriction with conditional independence across three intermediate input demands, a static condition grounded in input market segmentation, and establish nonparametric identification from a single cross-section. I develop a GMM estimator and establish consistency and asymptotic normality. Monte Carlo simulations confirm that the proposed estimator is unbiased across Markov and non-Markov environments, while the standard estimator exhibits persistent bias of up to 63 percent of the true materials elasticity. In 502 Japanese manufacturing industries, the proposed method yields systematically lower markups than the standard method across the entire distribution (median 0.93 vs. 1.03), reducing the share of industries with markups above unity from 54 to 37 percent. In a difference-in-differences analysis of the 2011 Tohoku earthquake, the standard method overstates the productivity loss by 0.40 percentage points, roughly $3.6 billion (400 billion yen) per year.

• The paper demonstrates that production functions can be identified nonparametrically without assuming a first-order Markov productivity process, significantly reducing bias in output elasticities.
• It employs a cross-sectional conditional independence restriction among flexible input shocks combined with spectral methods and a three-block GMM estimator.
• Empirical and simulation evidence shows that the proposed estimator yields lower, more accurate markups and mitigates bias compared to conventional Markov-based methods.

Nonparametric Identification and Estimation of Production Functions Invariant to Productivity Dynamics

Introduction and Problem Statement

The paper "Nonparametric Identification and Estimation of Production Functions Invariant to Productivity Dynamics" (2604.04458) addresses fundamental limitations in the canonical identification strategy for firm-level production function parameters. Since Olley and Pakes (1996), nearly all state-of-the-art estimators (including Levinsohn and Petrin, Ackerberg–Caves–Frazer (ACF), Gandhi–Navarro–Rivers (GNR), and their dynamic panel analogues) impose that the unobserved productivity process follows a first-order Markov process. The Markov property is not ancillary—it is the identification device that pins down the output elasticity of flexible inputs via the structure of the productivity transition equation.

This paper demonstrates both theoretically and empirically that the first-order Markov restriction is not necessary for identification. More critically, it documents that Markov misspecification induces persistent, substantial bias in estimated output elasticities and all policy-relevant downstream quantities (markups, measured misallocation, treatment effects in event studies). The author proposes a new identification strategy that replaces the Markov restriction with a cross-sectional conditional independence restriction: mutual independence of input demand shocks across three distinct flexible inputs within a given cross-section. Using the spectral identification strategy of Hu and Schennach (2008), the paper gives a nonparametric identification result that holds without any assumptions on the evolution of productivity, and constructs a GMM estimator consistent under arbitrary productivity dynamics.

Identification Theory and Assumptions

The production technology is modeled by a canonical gross output function

$$y_{jt}=f_{t}(k_{jt},l_{jt},m_{jt},e_{jt},w_{jt},\omega_{jt})+\varepsilon_{jt}$$

where $k_{jt}$ and $l_{jt}$ are predetermined primary inputs (capital, labor), $m_{jt}$, $e_{jt}$, $w_{jt}$ are three observable flexible intermediate inputs (materials, electricity, water), $\omega_{jt}$ is the latent productivity, and $\varepsilon_{jt}$ the ex-post production shock.

The paper's main identifying condition is not on the law of motion of $\omega_{jt}$ (as in Olley–Pakes/ACF/GNR), but instead posits that, conditional on the latent $\omega_{jt}$ and observable state $k_{jt}$0, the input-specific demand shocks $k_{jt}$1 affecting materials, electricity, and water are mutually independent: $k_{jt}$2 No restriction is placed on the stochastic process for $k_{jt}$3: the identification argument is entirely cross-sectional and static. The author exploits the mutual independence to use spectral methods (Hu–Schennach decomposition), leveraging three noisy signals (input demands) of a common latent ($k_{jt}$4) to recover the conditional density and thus nonparametrically identify both the production function and the productivity distribution.

Regularity conditions (injectivity of the measurement models, distinct eigenvalues, normalization) are formally stated and interpreted. The conditional independence is motivated both economically (segmentation of input markets yields input-specific demand shocks beyond what can be absorbed with observed controls) and through microfoundations (e.g., cost minimization with input-specific markdowns).

The theory includes a complete characterization of observational equivalence: any two structures $k_{jt}$5 yielding the same observables can differ only by a location shift in $k_{jt}$6 that is an arbitrary continuous function of primary inputs $k_{jt}$7. This "location indeterminacy" is the only ambiguity created by the fully static assumption set.

To close this indeterminacy, two alternatives are developed:

1. Nonparametric point identification by exclusion restrictions: if at least one input's demand is driven solely by productivity and not by capital/labor, point identification is restored.
2. Parametric point identification by curvature: if the expected productivity conditional on $k_{jt}$8 exhibits sufficient nonlinearity (e.g., homothetic CES curvature), this regularity pins down the scale, and the entire vector of elasticities is point identified.

Estimation Methodology

Given identification, the author develops a three-block GMM estimator. The first block (A) leverages cross-equation orthogonality (elimination of $k_{jt}$9 across residuals); block (B) exploits cross-covariances implied by the independence structure; block (C) imposes moment conditions arising from parameterization of the curvature in expected productivity as a function of $l_{jt}$0. In the absence of the latter, the procedure consistently identifies the intermediate input elasticities but not $l_{jt}$1. With both, the full vector is identified.

The estimator is computationally tractable and written for panel data, but relies only on cross-sectional moment conditions, so is robust to arbitrary time-series patterns of productivity, including non-Markovian dynamics, regime shifts, aggregate shocks, and potential outcomes frameworks where treatment affects the law of motion for $l_{jt}$2.

Monte Carlo Evidence

High-dimensional simulations document that all current Markov-based estimators (ACF, GNR, variants) exhibit persistent bias when their Markov restriction is violated, even as $l_{jt}$3. For materials elasticity, the upward bias can reach $l_{jt}$4 (63% of true value) under dynamically misspecified environments, leading to grossly inflated markups and misleading treatment effect estimates. In contrast, the proposed estimator yields negligible bias across simulated AR(1), AR(2), and potential outcome dynamics, demonstrating robustness to dynamic misspecification.

Simulations also examine conditional independence violations (e.g., correlated utility demand shocks arising from common energy/water cost shocks): moderate violations induce upward bias in the materials elasticity, but the magnitude is quantitatively much smaller than the Markov misspecification bias. Thus, the empirical finding that the proposed estimator often yields lower markups than standard methods cannot be explained by conditional independence failure but instead points to Markov misspecification in mainstream approaches.

Empirical Evidence and Specification Testing

Empirical implementation uses 502 Japanese manufacturing industries, leveraging census panels containing physical inputs for materials, electricity, and water. The estimator yields systematically lower materials elasticities, and thus lower markups, than ACF across the entire industry distribution (median markup 0.93 vs. 1.03 under ACF). The share of industries with markups above unity drops from 54% to 37%. Empirical productivity effects of large policy shocks—such as the 2011 Tohoku earthquake—are substantially attenuated: the ACF estimator overstates productivity loss by $l_{jt}$5 percentage points (or about $l_{jt}$6 billion USD/year).

Specification diagnostics are developed:

• Pairwise comparison of exclusion-based and curvature-based routes to identifying capital and labor elasticities shows high concordance for capital but reveals consistent exclusion failure for labor, due to its marginal adjustability within the time structure of manufacturing.
• Parameters recovered via block C (homothetic regularity) are consistent across the sample when CES curvature is significant.
• Ancillary regressions document that productivity residuals under the proposed approach are more strongly associated with log wage rates, suggesting improved signal-to-noise over ACF.

The conditional independence restriction is empirically supported in the input triplet studied (materials/electricity/water), both by institutional context and by formal testable implications.

Implications and Future Directions

Theoretical

Dropping the Markov assumption in proxy variable approaches involves no loss of identification power for the production function or for the productivity distribution conditional on observed signals. Identification is achieved cross-sectionally using economic structure and measurement error arguments, not dynamic law-of-motion restrictions. The paper provides a complete characterization of what can and cannot be identified without Markovity and presents generic sufficient conditions for closing the remaining gaps. The framework robustly nests earlier methods but does not rely on scalar unobservability, thus accommodating persistent, input-specific nonparametric heterogeneity (e.g., markdowns, procurement frictions) commonly present in administrative data.

Empirical

Markov misspecification induces systematic, economically meaningful upward bias in output elasticities for flexible inputs (especially materials), which propagates directly to overestimated firm-level markups, measured misallocation, and treatment effects in policy evaluation. The consequences are first-order: reinterpretation of the magnitude and prevalence of market power, and thus policy evaluation based on previous mainstream estimates, is warranted. Because the estimator is invariant to productivity dynamics, it supports truly dynamic panel studies, structural break/event study designs, and nonstationary environments, with consistent estimation of production function parameters even when productivity is an object of policy intervention.

Future Work

Open directions include: (i) generalizing to settings with only two observable flexible inputs by exploiting rapid labor adjustment (as in certain services or industries with casualized labor); (ii) implementing over-identification-robust specification tests via block C or structural restrictions, to empirically evaluate input demand independence; (iii) extending to non-Hicks-neutral, factor-augmenting, or higher-order flexible functional forms; (iv) scaling the framework to address international patterns, exploring the sensitivity of global markup estimates to dynamic identification assumptions.

Conclusion

The paper establishes that the standard Markov assumption embedded in nearly all contemporary proxy variable production function estimators is not only unnecessary for structural identification, but is a major source of bias—often exaggerating fundamental parameters such as materials elasticity and industry markups. By instead invoking conditional independence among input-specific demand shocks for multiple flexible inputs, nonparametric identification is achieved from static, cross-sectional data alone. The consequences for both empirical measurement and structural policy evaluation are substantial: measured market power, productivity decompositions, and treatment effect estimates in event studies are all highly sensitive to the identification assumption. The framework substantially expands the reliability and scope of semi/nonparametric production function estimation, without the risk of misspecified productivity dynamics inhering in prior protocols.

Citation: "Nonparametric Identification and Estimation of Production Functions Invariant to Productivity Dynamics" (2604.04458)