Estimating Production Functions
Erol Taymaz
http://users.metu.edu.tr/etaymaz
Dipartimento di Scienze Economiche e Statistiche (DiSES) / UNISA
September 23, 2025
Why production functions?
- Production functions are essential to understand how firms behave
- Profit maximization, \(\pi = PQ - wL -
rK\) where \(Q = f(K, L)\)
- \(w/P = f_L\) and \(r/P = f_K\)
- Production functions are main building blocks in many economic
models
- Analyzing growth dynamics at the country, industry or firm level
- Level and growth rate of (total factor) productivity
- Increasing returns, economies of scale
- Complementarity / substitutability between inputs
- Economics of scope
- R&D and innovation
- Resource allocation
Production functions
- \(Q* = f(K, L), Q ≤ Q*\)
- \(Q = f(K, L, E, M, S)\)
- \(Q = f(K, L, T)\)
Type of production functions
- Production functions
- Production frontiers
Some concepts
- Average product / partial productivity, \(AP = Q/L\)
- Marginal product, \(MP =
∂Q/∂L\)
- Output elasticity, \(ε_L = (∂Q/Q)/(∂L/L) =
(∂lnQ/∂lnL) = ∂q/∂l = MP/AP\)
- Scale elasticity, returns to scale, \(ε =
\Sigma ε_i\)
- Economies of scale
- Marginal rate of technical substitution (slope of the isoquant)
\(MRTS_{ij} = -MP_j/MP_i\)
- For profit maximization, \(w/r = MRTS =
-dK/dL|_{Q=Q*}\)
- Elasticity of substitution \(σ_{KL} =
(dKL/dMRTS)/(KL/MRTS)\)
Marginal rate of technical substitution

Elasticity of substitution

Some concepts - CD case
Production function \(Q =
e^{beta_0}K^{\beta_K}L^{\beta_L}\)
or \(q = \beta_0 + \beta_Kk +
\beta_Ll\)
- Labor elasticity of output, \(ε_L = ∂q/∂l
= \beta_L\)
- Perfectly competitive markets and constant returns to scale, \(ε_L = wL/pQ\)
- What if \(ε_L > wL/pQ\)
- Capital elasticity of output, \(ε_K =
∂q/∂k = \beta_K\)
- Constant returns to scale, \(ε = ε_K + ε_L
= 1\)
- \(q = \beta_0 + \beta_1k + (1 -
\beta_1)l\)
- \(q-l = \beta_0 +
\beta_1(k-l)\)
- Marginal rate of technical substitution \(MRTS = -dK/dL = -f_L/f_K = -MPL/MPK = -
(\beta_L/\beta_K)(K/L)\)
- Elasticity of substitution, \(σ_{KL} =
1\)
- Profit maximization, \(w/r =
dK/dL\)
- If the wage rate increases by 1% (w/r increases by 1%), K/L
increases by 1% and \(wL/rK\) remains
constant
Estimation of production functions
- Parametric methods
- Non-parametric methods
- Data envelopment analysis
Estimation methods
- OLS (Ordinary least squares)
- \(q_{it} = X_{it}\beta +
ε_{it}\)
- \(E[ε | X] = 0\)
- \(ε \sim N(0, \sigma^2)\)
- MLE - minimize the sum of squares of error terms
- GMM (Generalized method of moments)
- \(q_{it} = X_{it}\beta +
ε_{it}\)
- Moment conditions, \(E[X ' ε] =
0\)
- \(E[X ' (q - X \beta)] =
0\)
Estimation issues
- Data
- Measurement errors
- Unobserved firm-specific effects
- Functional form
- Technological change
- Endogeneity/omitted variable bias
- Selection
- Heterogeneity
Variables
- Output: \(Q\)
- Inputs: \(K, L, E, M, S, R\&D,
\omega\)
- Prices - sectoral level, firm level
Type |
Function |
Output |
Level |
|
|
|
|
KL |
Q = f(K, L) |
Value added |
Country, Sector |
KLM |
Q = f(K, L, M) |
Gross output |
Sector, Firm |
KLEM |
Q = f(K, L, E, M) |
Gross output |
Sector, Firm |
KLEMS |
Q = f(K, L, E, M, S) |
Gross output |
Sector, Firm |
CES, CD and Leontieff functions

Taylor series expansion
Production function in log form, \(q = f(k, l)\)
1st order Taylor series expansion of \(f(k, l)\) at point \((a,b)\)
\(q ≈ f(a, b) + f_k(a, b)(k - a) + f_l(a,
b)(l - b)\)
\(q ≈ \beta_0 + \beta_Kk +
\beta_Ll\)
2nd order Taylor series expansion of \(f(k, l)\) at point \((a,b)\)
\(q ≈ f(a, b) + f_k(a, b)(k - a) + f_l(a,
b)(l - b) +\)
\(½[f_{kk}(a, b)(k -a)^2 + 2f_{kl}(a, b)(k
-a)(l - b) + f_{ll}(a, b)(l - b)^2]\)
\(q ≈ \beta_0 + \beta_Kk + \beta_Ll +
\beta_{KK}k^2 + \beta_{LL}l^2 + \beta_{KL}kl\)
CD vs translog
Taylor series expansion, 1st, 2nd and 3rd order

CD functions
- Extensively used in theoretical models (with constant returns to
scale)
- Emprical support for CD but
- There is as a “publication bias”, the mean elasticity is 0.3.
- Gechert, Sebastian, Tomas Havranek, Zuzana Irsova and Dominika
Kolcunova (2022), “Measuring capital-labor substitution: The importance
of method choices and publication bias”, Review of Economic
Dynamics, 45: 55-82.
- CD is a special case of the translog function
- Test \(H_0\): \(\beta_{KK} = \beta_{LL} = \beta_{KL} =
0\)
- It is rejected in most of the cases
Technological change

CD function and technological change
Deterministic technological change
- Total factor productivity growth, \(Q =
AK^{\beta_K}L^{\beta_L}, A = A_0e^{\lambda t}\)
- \(q = a_0 + \lambda t + \beta_Kk +
\beta_Ll\), \(\lambda\) = TFP
growth rate
- Hick neutral technological change
- Increase in capital efficiency, \(K =
Ke^{\gamma t}\)
- \(q = a_0 + \gamma t + \beta_Kk +
\beta_Ll\), \(\lambda\) = TFP
growth rate
- Increase in labor efficiency, \(L =
Le^{\omega t}\)
- \(q = a_0 + \omega t + \beta_Kk +
\beta_Ll\), \(\lambda\) = TFP
growth rate
- All of them together
- \(q = a_0 + (\lambda + \gamma + \omega) t
+ \beta_Kk + \beta_Ll\), \(\lambda\) = TFP growth rate
We cannot identify TFP, capital efficiency and labor efficiency
effects in a CD function. They are all observationally equivalent.
CD functions and technological change
Stochastic technological change
- Production function, \(q_{it} =
\beta_Kk_{it} + \beta_Ll_{it} + \omega_{it}\)
- Technological change, \(\omega_{it} = \rho
\omega_{it-1} + ε_{it}\)
Translog function and technological change
Deterministic technological change
\(q = \beta_0 + \beta_Kk + \beta_Ll +
\beta_{KK}k^2 + \beta_{LL}l^2 + \beta_{KL}kl\)
\(+ \beta_Tt + \beta_{TT}t^2 + \beta_{TK}tk +
\beta_{TL}tl\)
Stochastic technological change
\(q = \beta_0 + \beta_Kk + \beta_Ll +
\beta_{KK}k^2 + \beta_{LL}l^2 + \beta_{KL}kl +
\omega_{it}\)
\(\omega_{it} = \rho \omega_{it-1} +
ε_{it}\)
Technological change and substitution

Technological change and substitution

Technological change and substitution

Technological change and substitution

Technical efficiency

Stochastic production frontiers
\(q = \beta_0 + \beta_Kk + \beta_Ll + ε -
u\)
\(u\) is non-negative technical
efficiency term, \(TE = exp^{-u}\),
therefore \(0 ≤ TE ≤ 1\)
\(u \sim N^+(\mu, \sigma^2)\)
- \(\mu = 0\)
- \(\mu = Z\delta\)
Production frontiers and efficiency
Technical efficiency vs plant size in Turkish motor vehicle
industry

Taymaz, Erol and Gulin Saatci (1997), “Technical Change and Efficiency
in Turkish Manufacturing Industries”, Journal of Productivity
Analysis, 8: 461-475.
Endogeneity/omitted variable bias
- Cobb-Douglas production function
- \(Q =
e^{\beta_0}K^{\beta_K}L^{\beta_L}e^{\omega + ε}\)
- \(q = \beta_0 + \beta_Kk + \beta_Ll +
\omega + ε\)
- Profit maximization
- \(\pi = PQ - wL - rK\)
- Decision variables?
- Does the firm observe \(\omega\)
before its decision?
Endogeneity/omitted variable bias
- The firm does not observe \(ε\)
before K and L decisions
- The firm observes \(\omega\) before
L and after K decisions
- Assume competitive markets
- \(\pi = Pe^{\beta_0}\bar
K^{\beta_K}L^{\beta_L}e^{\omega} - wL - rK\)
- \(L = [(p/w) \beta_Le^{\beta_0 +
\omega}\bar K^{\beta_K}]^{1/(1-\beta_L)}\)
- \(L\) and \(\omega\) are positively correlated
- Omitted variable bias: \(\beta_L\)
will be over estimated
Endogeneity/omitted variable bias
How to solve the endogeneity bias?
- Fixed effects, if assume \(\omega_{it} =
\omega_i\)
- If there are measurement errors, fixed-effect estimator can generate
higher biases than OLS
- Fixed effects may explain most the of the variation in output,
especially if the panel is short
- Returns to scale parameter tend to be low
- First-order conditions for inputs
- Needs additional assumptions on the structure of factor markets and
firm behavior
- Instrumental variables
- Find instruments that are correlated with \(L\) but not \(\omega\)
- Input prices? Usually low or no cross-sectional variation
- Lagged inputs? If \(\omega\) is
auto-correlated…
Endogeneity/omitted variable bias
- Proxy variables to control for unobservables (control function
approach)
- Olley & Pakes - use investment, define \(i = f(k, \omega)\), then \(\omega = f^{-1}(k, i)\)
- Levinsohn & Petrin - use intermediate inputs as proxy
- Ackerberg, Caves & Frazer - addresses the identification issues
in the first stage
- Wooldridge - one step GMM estimation
- Dynamic panel approach
- GMM and measurement errors (revenue based markup estimation)
- De Ridder, Grassi and Morzenti
- If there is a measurement error, variance is high, productivity if
measured as \(\omega + \eta\), and
estimates are not consistent if \(\omega\) follows a non-linear AR
process.
- First purge the measurement error (purging equation includes fixed
effects, price and controls for the mark up like the firm’s market
share)
Blundell & Bond (2000)
\(q_{it} = \beta_Ll_{it} + \beta_Kk_{it} +
\gamma_t + \eta_i + v_{it} + m_{it}\)
\(v_{it} = \rho v_{it-1} +
e_{it}\)
\(e_{it} \ sim N(0, \sigma_e), m_{it} \sim
N(0, \sigma_m)\)
Both employment (l)and capital (k) are potentially correlated with
the firm-specific effects (\(\eta\)),
and with both productivity shocks (e)and measurement errors (m).
\(q_{it} = \rho q_{it-1} + \beta_Ll_{it} -
\rho \beta_Ll_{it-1} + \beta_Kk_{it} - \rho \beta_Kk_{it-1} + \gamma_t -
\rho \gamma_{t-1} + \eta_i (1 - \rho) + e_{it} + m_{it} - \rho
m_{it-1}\)
\(q_{it} = \pi_1 q_{it-1} + \pi_2l_{it} -
\pi_3l_{it-1} + \pi_4k_{it} - \pi_5k_{it-1} + \gamma_t^* + \eta_i^* +
\omega_{it}\)
Blundell & Bond (2000)
Take the first difference to eliminate firm specific effects
\(\Delta q_{it} = \pi_1 \Delta q_{it-1} +
\pi_2 \Delta l_{it} - \pi_3 \Delta l_{it-1} + \pi_4 \Delta k_{it} -
\pi_5 \Delta k_{it-1} + \Delta \gamma_t^* + \Delta
\omega_{it}\)
Moment conditions
\(E[x_{it-s} \Delta \omega_{it}] =
0\)
where \(x_{it} = (l_{it}, k_{it},
q_{it})\) for \(s \ge 3\)
- GMM estimation: Use suitably lagged levels as
instruments
- First-differenced GMM estimator have poor finite sample properties
(bias and imprecision)
- System-GMM estimation: Add level equation
- \(E[\Delta_{i,t-s}(\eta_i^* + \omega_{it}]
= 0\) for \(s = 2\)
De Ridder, Grassi and Morzenti (2025)
Case 1. Quantity data, no persistence in
productivity
\(q_{it} = \alpha v_{it} +
\omega_{it}\)
The firms sets v after observing \(\omega\)
GMM estimator: \(E[\omega_{it}v_{it-1}] =
0\)
\(v_{it-1}\) is a valid instrument if
\(E[v_{it} v_{it-1}] ≠ 0\)
Case 2. Measurement error
\(\tilde q_{it} = \alpha v_{it} +
\omega_{it} + \eta_{it}\)
GMM is still consistent but has a higher variance if \(E[\eta_{it}v_{it-1}] = 0\)
De Ridder, Grassi and Morzenti (2025)
Case 3. Revenue as a proxy for output output
\(r_{it} = q_{it} + p_{it}\)
GMM is not consistent if \(E[p_{it}v_{it-1}] ≠ 0\)
Under imperfect competition, they are likely to be correlated
Case 4. Persistent productivity
\(q_{it} = \alpha v_{it} +
\omega_{it}\)
\(\omega_{it} % |rho \omega_{it-1} +
ε_{it}\)
GMM estimator: \(E[e_{it}v_{it-1}] =
0\) and \(E[e_{it}\omega_{it-1}] =
0]\)
\(\omega_{it} = q_{it} - \alpha
v_{it}\) and \(e_{it} = q_{it} - \alpha
c_{it} - \rho \omega_{it-1}\)
If there is a persistence in productivity, there may be two
solutions, one is unbiased, the other one is biased!
Selection
- If less efficient firms exit (low \(\omega\) value), and if capital intensive
firms stay in the market even for low \(\omega\), then, conditional on being in the
market, there is a (negative) correlation between K and \(\omega\)
- Solution
- Use balanced panel data
- Two stage estimation
- Use a method that takes into account the selection process (like
OP)
- If selection is based on fixed effects (\(\eta_i\)), panel GMM is fine
Heterogeneity
Do all firms use the same technology?
- Differences in products, processes, patented technology, etc.
- \(q_{it} = \beta_{0,it} +
\beta_{K,it}k_{it} + \beta_{L,it}l_{it} + \omega_{it}\)
- Putty-clay technology
- \(q_{it} = \beta_{0,i} + \beta_{K,i}k_{it}
+ \beta_{L,i}l_{it} + \omega_{i}\)
Good practice
- Check returns to scale parameter, is it around 0.9-1.1?
- Check output elasticities
- Check the sensitivity of your assumptions
- Check different models
Good practice
Food and Beverages Industry in Belgium

Source: Ilke Van Beveren (2010), “Total Factor Productivity Estimation:
A Practical Review”, Journal of Economic Surveys, 1-38.
Good practice
Food and Beverages Industry in Belgium

Source: Ilke Van Beveren (2010), “Total Factor Productivity Estimation:
A Practical Review”, Journal of Economic Surveys, 1-38.