The Translog Cost Function Specification
• The non-homothetic translog cost function can be envisaged as a
second-order Taylor’s series approximation to an arbitrary cost
function.
n
1 n 1 n
lnC = ln 0 + i ln Pi + ij ln Pi ln Pj + Y lnY + YY (lnY ) + iY ln Pi lnY
2
i =1 2 i =1 2 i =1
where ij = ji
• For cost function to be well-behaved, among other things, it must be
HOD 1 in prices, given Y. It implies following restrictions on the above
cost function: n n n n
i =1
i = 1; ij = ji = iY = 0.
i =1 i =1 i =1
The Translog Cost Function Specification
• A number of additional parametric restrictions can be imposed on the
translog cost function. Homotheticity requires that: iY = 0 i = 1, , n.
• CRTS requires that iY = 0; YY = 0; and Y = 1.
• Finally, the translog reduces to the CRTS Cobb-Douglas function when
in addition to all the above restrictions, each of the ij = 0, 1, , n.
• We can estimate the translog cost function directly, however,
efficiency gains can be realized by estimating the optimal, cost
minimizing input-demand equations, transformed into cost share
equations.
Cost share equations:
lnC Pi C C
= . Shepherd's lemma: = Xi
lnPi C Pi Pi
lnC Pi X i
= which is cost share equation
lnPi C
lnC Pi X i n n
= = i + ij ln Pj + ij lnY where Pi X i = C
lnPi C j =1 i =1
n
Pi X i
Defining the cost shares, Si , it follows that Si = 1.
C i =1
This adding up condition of the share equation system has
implications for econometric estimation.
Let us again expand a four-input (K,L,E,M) cost function. The cost share
equations become:
SK = K + KK lnPK + KL ln PL + KE ln PE + KM ln PM + KY lnY
SL = L + LK lnPK + LL ln PL + LE ln PE + LM ln PM + LY lnY
SE = E + EK lnPK + EL ln PL + EE ln PE + EM ln PM + EY lnY
SM = M + MK lnPK + ML ln PL + ME ln PE + MM ln PM + MY lnY
Imposing the 6 symmetry restriction, the number of parameters drops to 18.
KL = LK ; KE = EK ; KM = MK ; LE = EL ; LM = ML ; EM = ME .
HOD 1 in input prices is required for the cost function:
n n n n
i =1
i = 1; ij = ji = iY = 0.
i =1 i =1 i =1
Imposing these (6) restrictions reduces the
K + L + E + M = 1
number of estimated paramters from 18 to 12.
KK + KL + KE + KM = 0 If CRTS restrictions are also imposed then the
KL + LL + LE + LM = 0 number of free paramerter would be reduced
KE + LE + EE + EM = 0 to 9.
CRTS restriction:
KM + LM + EM + MM = 0
iy = 0, YY = 0, Y = 1
KY + LY + EY + MY = 0
KY = 0, LY = 0, EY = 0
The Translog Cost Function Specifications
• To implement the share equation system, we add a typical random disturbance
term to each equation.
• The share equation possesses a special property that the sum of the dependent
variables (the cost shares) over the all equations equals 1.
• This adding up feature of the share equation has several econometric
implications.
1) Since shares sum to unity and only n – 1 equations are linearly independent, for each
observation the sum of the disturbances across equations must always equal zero implying
that the disturbance covariance matrix is singular (its determinant is zero).
2) Because shares sum to unity at each observation, when six symmetry restrictions are not
imposed, the arithmetic of equation-by-equation OLS parameter estimates always obey
the column adding up conditions
The Translog Cost Function Specifications
3. It also implies that the OLS residuals ei across equations will sum to zero at
each observation, i.e., eK + eL + eE + eM =0
4. Because the disturbance covariance and residual cross-products matrices
will both be singular, ML estimation that minimizes the determinant of E’E,
will not be feasible because the determinant will be zero for any set of
parameters.
5. The most common procedure to handle this singularity problem is to drop
an arbitrary equation and then estimate the remaining n – 1 share
equations by ML.
• The parameter estimates, log-likelihood values and estimated SE will be invariant to
which equation is deleted.
The Translog Cost Function Specification
6. If one were estimating the translog share equations using highly aggregated
data and believed that input prices were endogenous, to avoid
simultaneous equations problem, one could employ 3SLS provided that
appropriate instruments were available. Estimates invariant to which
equation is deleted.
7. For the translog function, hypothesis testing can be undertaken using any
of the three test procedures: a) Wald test; b) the likelihood ratio test; and
Lagrange multiplier test.
8. For the singular share equation system, these test procedures must be
based on direct estimation of n – 1 share equations.
9. If instrumental variables estimation in 3SLS is employed, then the LR test
statistic is not appropriate. However, the Wald test and the LM test could
be used.
The Translog Cost Function Specification
• In our four-equation symmetry-constrained translog example, one could
impose homogeneity restrictions, delete the M share equation and directly
estimate 12 free parameters in K, L and E equations.
• Six parameters of the deleted equation could then be obtained by rearranging
the homogeneity restrictions in terms of directly estimated parameters as
follows:
M = 1 − K − L − E
KM = −( KK + KL + KE )
LM = −( KL + LL + LE )
EM = −( KE + LE + EE )
MY = −( KY + LY + EY )
MM = −( KM + LM + EM ) or MM = KK + LL + EE + 2( KL + KE + LE )
where we substitute for unknown param by known param.
Elasticities of Substitution and Price Elasticities
• Allen partial elasticities of substitution turn out to be equal to:
ij + Si S j
ij = , where i , j = 1,, n, but i j.
Si S j
• In the translog form is restricted to be Cobb-Douglas, all ij terms will
simultaneously be equal to zero and ij will equal unity. Similarly,
ii + Si2 − Si
ii = 2
where i = 1, , n.
Si
• The price elasticities are given by
ij + Si S j ii + Si2 − Si
ij = , where i , j = 1,, n, but i j. ii = where i = 1, , n.
Si Si
• Parameter estimates and fitted shares should replace gammas and S.
Elasticity of Substitution and Price Elasticity
• The translog form specified above is a highly general, non-homothetic
form, which implies that RTS are not constrained a priori.
• RTS are computed as the inverse of the elasticity of cost w.r.t. output:
1 lnC
= where CY = and where for the translog cost function
CY lnY
CY = Y + KY lnPK + LY lnPL + EY lnPE + MY lnPMK + + YY lnY
• One potential problem with estimation of scale economies is that the
alpha_Y and alpha_YY parameters do not appear in the share
equations, and so these param cannot be estimated by share eq. alone.
• To estimate RTS, it is usually necessary to add the translog cost function
to the share equation system to be estimated. −1
1 lnC 1
• At the point of approximation, the returns to scale is = = =
CY lnY Y
C
• Cost is monotonically increasing in prices if 0.
Pj
C lnC C
= 0
Pj lnPj Pj
C Pj C C
= . By Shephers' lemma = Xi
P C P Pj
j j
X i Pj C
= P
C j
C C
Since C and Pj are always positive 0 iff Si > 0. In other words, 0 when
Pj Pj
n
Si = i + ij lnPj + iY lnY 0
j =1
lnC n
Monotonicity in output requires that = Y + YY lnY + iY lnPi 0.
lnY i =1
Curvature Condition for the translog
• The concavity condition requires that the Hessian matrix of second-partials of C
with respect to factor prices is negative semi-definite (n.s.d).
2C 2C
P 2 Pi Pj
= 2
i
HWW n.s.d
C C2
Pj Pi Pi 2
• In terms of the estimated parameters of the translog cost model, the concavity
requires the following:
KK + K2 − K KL + K L KE + K E KM + K M
KL + K L LL + L − L LE + L E
2
LM + L M
HWW = n.s.d
+ LE + L E EE + E − E EM + E M
2
KE K E
KM + K M LM + L M EM + E M MM + M − M
2
Translog Variable Cost Function
• Unlike the assumptions for the long run equilibrium, all inputs may
not adjust instantaneously to their long run equilibrium values.
• An important extension of the long run model is where at least one of the inputs is fixed.
• More specifically, the capital input may not adjust fully to their full
equilibrium level during a particular point in time.
• Therefore, researchers adopt frameworks that distinguish variable
from quasi-fixed inputs where the later adjust only partially to their
long run values during one time period.
• Following Marshallian tradition, the researchers have distinguished
the short run cost functions (where at least one input is fixed) from
the long run cost functions (where all inputs are at their full
employment values).
Translog Variable Cost Function
• Variable costs are a function of the variable inputs, the level of
output(s), and quantities of the fixed factors.
• A KLEM translog variable cost function with K fixed and L, E, and M as
variable inputs can be written as
• where VC = PL L + PE E + PM M.
Translog Variable Cost Function
• To be consistent with economic theory, parametric restrictions must be adapted to
this framework so that VC function is HOD 1 in variable input prices, given K and Y.
• Estimation equation of the translog VC function can be obtained by using Shephard’s
Lemma, as with LR cost function where lnK replaces ln PK. The share equations for
VC inputs turn out to be
Technical Progress and Factor Demands and Costs
• Under CRTS the effects of technical progress are shown either in
increased output or in reduced factor demand and costs.
• These effects are called multifactor productivity (MFP) or total factor
productivity (TFP).
• MFP research has mostly assumed that tech progress is disembodied.
• This type of technical progress is implemented empirically by simply
inserting the variable (t), or (ln t), for time into the cost function
specification.
• For example, linear (t) and quadratic forms of t ([Link], [Link], t2) are
added to the translog cost function, and by Shephard’s Lemma, t also
appears as a regressor in the cost-minimizing share equations.
Technical Progress and Factor Demands and Costs
• The translog cost function with disembodied tech. change is written as
n
1 n 1 n
lnC = ln 0 + i ln Pi + ij ln Pi ln Pj + Y lnY + YY (lnY ) + iY ln Pi lnY
2
i =1 2 i =1 2 i =1
n
1
+ t t + tt t 2 + it lnPi . t + tY lnY . t where ij = ji
2 i =1
• Homogeneity of degree 1 in factor prices, given Y and t, implies
n n n n n
i =1
i = 1; ij = ji = iY = it =0.
i =1 i =1 i =1 i =1
• Cost share equations become:
lnC Pi X i n
= = Si = i + ij lnPj + ij lnY + it . t
lnPi C j =1
Technological Change
• Technological change refers to changes in production process through
application of scientific knowledge.
• May lead to leftward shift in the production function and inward shift in the cost function.
• The technological change may be disembodied or embodied.
• Disembodied technical change occurs through improved methods of
utilizing existing resources for higher output per unit of inputs.
• Embodied technological change occurs through changes in input quality.
• Technological change can affect input productivities and factor
utilization differently in a multiple input production process.
• Distinction is made between neutral and biased technological change.
Biased and Neutral Technical Change
• J.R. Hicks (1963) defined neutral and biased technological change in
terms of marginal rates of technical substitution (MRTS).
• Assume we have only two factors capital (K) and labor (L).
• We can classify inventions according to their initial effects on the ratio
of MPK and MPL.
• If MPK/MPL ratio increases, intervention is labor saving and capital using.
• If MPK/MPL ratio is unchanged, intervention is neutral.
• If MPK/MPL ratio decreases, intervention is labor using or capital saving.
it 0, i − saving; it = 0, i − neutral; it 0, i − using.
• One can restrict cost function to be Hicks neutral by removing all it
terms from the share equations. One can also conduct cost share
neutrality test by the LR test.
How to Obtain MFP from Translog Cost Function?
• The partial derivative −C t is interpreted as the shadow value of
time. Under CRTS, − lnC t is the rate of MFP growth.
• Consider a model with CRTS imposed, i.e., in which Y = 1, YY = 0,
and iY = 0, i = K , L, E , M.
• Solving out the translog with symmetry and homogeneity in prices
restriction imposed as well as the above CRTS will result in a new
dependent variable, c, involving the log of average, not total cost
(where lnc = lnC–lnY). [Note: also see comments in Ex.9, page 497-98]
lnc PK PL PE
= t + tt t + Kt ln + Lt ln + Et ln .
t PM PM PM
• Substitute parameter estimates and values of variables to obtain an
annual time series of predicted MFP.
The Translog Profit Function
• The translog profit function, (P), is given by
1 1
ln = 0 + i ln Pi + ij ln Pi ln Pj + t t + tt t 2 + it ln Pi . t
i 2 i j 2 i
where Pi denotes the price of the ith unit of input or output.
• Total profit equals = Pi X i = PY Y − wi X i
i i
• Symmetry and linear homogeneity of the profit function in prices implies:
ij = ji i , j; i = 1; ij = 0 j and it = 0.
i i i
• Differentiating w.r.t. lnPi yields the system of equations in the form of shares.
Pi X i
Si = = i + ij ln Pj + it t
Pi Xi j
ln d P Xw
• If Pi = inputi price, denoted by w, then: = . =−
lnw dw
• If Pi = output price, then: ln = d . P = PY
ln P dP
The Translog Profit Functions
• The monotonicity condition requires that it be non-decreasing in
output prices and non-increasing in input prices.
• It can be verified by examining the predicted profit shares.
• At the point of approximation, these share are equal to the
corresponding i .
• If Xi is an output, i 0.
• If Xi is an input, i 0.
The Translog Profit Function
• For convexity, the Hessian matrix needs to be positive semi-definite. It
implies that all principal minors must have non-negative determinants.
• The modified Hessian is given be
ii + i2 − i ij + i j
HPP = ij + i j jj + j − j
2
• The convexity and monotonicity conditions can be checked or
imposed.