Production Function Estimation and

Related Risk Considerations

Richard E. Just and Rulon D. Pope

There has been considerable interest in estimations of input effects on the probability
distribution of output. Most empirical and theoretical analyses utilize multiplicative

Downloaded from http://ajae.oxfordjournals.org/ at Penn State University (Paterno Lib) on May 9, 2016
stochastic specifications which are analyzed and found unduly restrictive, particularly
since inputs that marginally reduce risk are not allowed. A more general stochastic
specification is proposed, free of these a priori restrictions. The proposed functional form
estimation is discussed and demonstrated with nitrogen-response data and common
log-linear production functions. Though nitrogen is risk-increasing, the marginal variance
contribution is smaller when compared to estimates based upon multiplicative
specification. Finally, stochastic specification error effects are analyzed.

Key words: fertilizer response functions, production functions, risk.

Increasingly, risk considerations are neces- yields to be estimated within cells. A number
sary in the analysis of the agricultural sector. of other authors (e.g., de Janvry and Fuller)
Risk is affected not only by price and other have recognized that more efficient estimation
market-related phenomena but also by many is possible by utilizing continuous response
technological innovations and government functions. Apparently thus far, however, no
policies related to input use. For example, one has given adequate attention to the effect
much of the controversy surrounding pesticide of inputs on risk. This is true of both empirical
regulations relates to their reducing risk (see and theoretical studies. As shown in this pa-
Turpin and Maxwell). In such a case, intelli- per, virtually all empirical and theoretical
gent public policy formulation should consider studies make implicit, if not explicit, assump-
not only the marginal contribution of pesticide tions to the effect that inputs increase risk.
use to the mean of output but also the marginal Examples of such theoretical studies are such
reduction in variance of output. Other exam- notable works as Stiglitz, Batra, Magnusson,
ples in which increased input use appears to Crawford, Rothenberg and Smith, Bardhan,
reduce variability are "overcapitalization," and Feldstein. Examples of empirical works
frost protection such as use of smudge pots, include such notables as de Janvry, Wolgin,
and possibly irrigation. Consider, for example, Sadan, and Moscardi and de Janvry.
overcapitalization in grain harvesting. The use The purpose of this paper is to examine the
of large (and fast) harvesting equipment, as implications of traditional econometric pro-
opposed to smaller (and slower) equipment, duction function studies when the above is-
usually leads to less variability of output (be- sues are important and, then, to demonstrate
cause of random weather conditions which some useful generalizations. Attention will be
can destroy a ripe crop before harvest). focused on neoclassical log-linear production
The problem of investigating stochastic as- functions. First, the restrictions associated
pects of production response is apparently an with popular formulations of stochastic pro-
old one (Fuller, Day, and Anderson). The duction functions are developed. A more gen-
traditional approach in agricultural economics eralized stochastic specification for produc-
to evaluating the impact of inputs on risk in tion function estimation is then posed, and its
production has been to use experimental data. generality in reflecting the risk effects of input
This method, of course, allows the variance of use is demonstrated. The functional form pro-
posed is general enough to encompass the de-
Richard E. Just is an associate professor of agricultural and re- terministic implications of all the functions
source economics, University of California, Berkeley, and Rulon
D. Pope is an assistant professor of agricultural economics, Uni- used in the above studies but avoids the risk-
versity of California, Davis. related restrictions they impose. A methodol-
Just and Pope Production Functions and Risk 277

ogy for estimating the generalized specifica- ignored or is incorrectly estimated as an in-
tion is discussed. Finally, application of the crease in risk as suggested by (1) and (2).
methodology is exemplified with the well- Incorrect conclusions of this type also can be
known fertilizer response data used in a previ- ?btained in evaluation of policies which affect
ous study by Day. inputs that increase risk because of the lack of
flexibility in (1).
. Consider, further, the marginal effect of
Shortcomings of Popular Production input use on marginal productivity variability.
Specifications Here one finds
!Iist~ric~lly, a popular enconometric spec-
ay _ ~y

ification m production function estimation has aXi - ---X;-'

Downloaded from http://ajae.oxfordjournals.org/ at Penn State University (Paterno Lib) on May 9, 2016
been ~2
V(~)= Xl V(y),
(I) y = A(J)X;'}, \ aX i

aV(ay/ aXi)
wher~ y is output,. Xi is a factor input (Xi> 0), aXi
and e ISa stochastic disturbance with E(e) = 0
V(e) > o. '
Consider the marginal effect of input use on
production variability. One finds that
assuming ~ < 1 or equivalently that E(y) is
V(y) = A
2
(J)x 2"')V(e'),
i concave in Xi' Thus, the marginal effect of
increasing input use is always to reduce the
and, hence, that variability of the marginal product (unless ex-
2
pected production is not concave in X).
(2) - av(y)- = ~ 2 A ( nXlai n )
E
V(e > 0,
)
It appears, however, that in many realistic
aXi Xi i=l sit~a~ions this constraint may be overly re-
stnctive. For example, it seems that in some
assuming ~ > O. Thus, the marginal effect of
increasing input use must always be to in- cases the variability of the marginal productiv-
ity of land increases when other inputs are
crease the variability of output when the rele-
held fixed. As one farms on a more extensive
vant CXi > 0; and CXi must be positive if marginal
?as~s with the same inputs, "fire fighting" abil-
productivity is positive.
~ty IS re~uced (Radner and Rothschild). That
Pragmatically, the implications of using the
IS, one IS more subject to adverse weather
usual formulation in (1) are as follows. Con-
c.onditions, etc., during such critical opera-
sider the evaluation of a policy which limits
nons as harvesting and planting. On the other
the use of some input such as pesticides.
hand, an input such as smudge pots in frost
Using (1), a reduction in pesticide use would
imply by (2) a reduction in variability of out- p~otection is likely to have a decreasing mar-
ginal productivity variability because survival
~u~. In reality, however, a reduction in pes-
o~ the crop becomes more probable, and mar-
ticide use may lead to more variable produc-
g~~al productivity presumably tends in proba-
tion. Under risk aversion, the true utility loss
bility toward zero with increased input use (in
associat~d with higher risk (at the lower input
the relevant range).
level) WIll be greater than when the risk effect
Of course, many functional forms other than
is incorrectly estimated as a reduction in vari-
the Cobb-Douglas have been used economet-
ability [as implied by (1)].1 Similarly, if one
rically in production-function estimation. Ex-
were to promote the expansion of irrigated
acreage and thereby reduce risk, the estimated amples are the transcendental function (Hal-
ter, Carter, Hocking), the generalized power
benefits (again assuming risk aversion) would
production function (de Janvry), and the
be greater than when the reduction in risk is
translog function (Christensen, Jorgenson,
I Although it is well known that variance has limitations as a
Lau). However, all of these functions includ-
measure, of ri~k (Borch), risk is associated with variance in the ing Kmenta's approximation of the CES func-
above discussion because 1/2 av(y)/axi = cov (Y, iJy/aXi ) and tion are estimated in practice by specifying a
cov (y, iJy/ axi ) determines whether a risk a verter uses more or
l~ss of the i~put Xi than the risk-neutral producer under produc-
log-linear disturbance: y = f(X)e\ E(e) = 0,
non uncertainty (Horowitz). where f(X) represents the particular functional
278 May 1979 Amer. J. Agr. Econ.

form. And each of these cases has the same It can be affirmed readily that the signs of
shortcomings discussed above for the Cobb- neither (4) nor (5) are determined a priori even
Douglas because when h (X) follows one of the popular produc-
E), tion function forms (Cobb-Douglas, translog,
V(y) = .f(X)V(e etc.) with parameters unconstrained. For ex-
aV(y) = 2[(X).!i(X)V(eE) > 0, ample, if h(X) follows a Cobb-Douglas form
aX i and the ith input is associated with a negative
parameter, then hi(X) < 0; and thus the case of
aii = .!i(X)et, vl- aiJ
= .!i2(X)V(e E),
a risk-reducing input is exemplified.
It is further interesting to note that the popu-
lar specification in (1) is simply a special case
av(ay/axi) of (3). This is apparent because, regardless of

Downloaded from http://ajae.oxfordjournals.org/ at Penn State University (Paterno Lib) on May 9, 2016
aX i specific functional forms,

when ii > 0, iii < O. y = r(X)e E" = [(X) + h 1/ 2(X)e,

E(e*) = E(e) =0
A Reasonable Stochastic Specification
under Risk where[(X) === h 1l2(X) ===r(X)E(e E" ) , and e === e"
- E(e E' ) . Thus, the proposed functional form
The results of the previous section point to an can be no more restrictive in any sense so long
important problem with the traditional as explicit specifications admit the possibility
stochastic specification of production func- that[(X) = h 1/ 2(X).
tions. Namely, if any input has a positive ef-
fect on output, then a positive effect on vari- An Estimation Procedure
ability of output is also imposed. The argu-
ments of the previous section imply that the For empirical purposes, suppose both[ and h
effects of input on output should not be tied to follow a popular log-linear form-Cobb-
the effects of input on variability of output a Douglas or translog (Christensen, Jorgenson,
priori. To attain this generality, it seems that Lau; Just and Pope). Thus,[(X) = h 1/ 2(X) is a
an adequate production-fnnction specification potential special case. Estimation with log-
should include two general functions-one linearity can be accomplished as follows.
which specifies the effects of input on the First, rewrite (3) for observation t explicitly
mean of output and another which specifies including the parameters of [ and h as
the effects of input on the variance of output.
Such a function is given by (6) Yt = [(Xt, a) + e*t,
E(e*t) = 0, E(e*te* T) = 0 for t =1= T,
(3) y = [(X) + h 1 / 2(X)e, E(e) = 0, V(e) = 1.
where e*t = h / 2(Xt, {3)et, E(et)
1
= 0, E(eteT) = 0
for t =1= T. One can then consider (6) as a
Thus, E(y) = [(X), V(y) = h (X), so that the nonlinear, heteroscedastic regression of y on
effects on mean and variance of output can be X. It can be shown, following the methods of
independent. (For alternatives to this spec- Malinvaud, that nonlinear least squares (NLS)
ification, see Just and Pope.) applied to (6) leads to consistent estimators of
With the specification in (3), one can easily a, the parameters of I. and of [(Xt, a) itself
verify that under a broad range of conditions (see Just
av(y) = h.(X)
and Pope).
aX i I
If one is interested only in estimation of [,
this may be as far as one needs to proceed.
(4) - aii = fz(X) + ~h-l/2(X)hi(X)e, and However, there are several important reasons
for carrying the estimation procedure beyond
this point. An obvious reason is to learn more
v('~) = h i (X)
2
about the effect of input use on risk. But even
aXi 4h(X) , if risk is not important, there are several
aV(ay / aX i) econometric shortcomings if one obtains only
(5) the above results. First, because of hetero-
aX i
scedasticity, hypothesis testing about the im-
= hi(X)[h(X)hii(X) - hl(X)]2 portance of various variables cannot generally
2h 2(X) be performed. But, secondly, by taking ac-
 Just and Pope Production Functions and Risk 279

count of the heteroscedasticity, it is possible e*t (with the above functional form and condi-
to gain more efficiency in estimation (at least tions developed by Christensen, Jorgenson,
asymptotically). Lau; and Just and Pope)." Hence, a weighted
To accomplish this, one can consider two NLS regression of Yt on X t in (6) with weights
additional stages of estimation. Using the h- 1I2(Xt , /3) can attain asymptotic efficiency in
(consistent) estimate of a, say, ix, one can estimation of a. In other words, in a third
consistently estimate !(Xt , a) by!(Xt , ix) and, stage one would find an NLS estimate of a for
thus, e*t or h 1i2(Xt, (3) e, can be estimated by the model
E* t = Y - teXt, ix),
under a broad range of conditions. But note where
that
Y*t = Yth-1/2(Xt, /3),

Downloaded from http://ajae.oxfordjournals.org/ at Penn State University (Paterno Lib) on May 9, 2016
E[(e*t)2] = E[h(Xt, (3)e/] = nix; (3). f*(X t , a) = !(Xt , a)h- 1/2(X t , /3).
Recalling the work of Hildreth and Houck and It has been rigorously shown elsewhere by
others on the random coefficient regression Just and Pope that the estimator for a thus
problem (see Theil), this suggests a regression obtained is consistent, asymptotically effi-
equation based on the relationship cient, and unbiased under conditions indicat-
ed above so long as fourth moments of Et exist
(7) (e*t)2 = E[(e*tF]u t
when the regression equation in (7) is used
= h (X t , (3)ut , directly.
where E(ut) = 1 by the definition of expecta- In summary, the regression method thus in-
tions. That is, f3 can be estimated by regres- volves the following steps:
sing (e" t)2 on X t (in a nonlinear framework) or, (a) An NLS regression of Yt on !(Xt , a) =
because e*t consistently estimates E* t» regres- exp [(In Xt)a] obtaining, say, ix.
sing (e*t)2 on X t leads to the same regression (b) An OLS regression of InlE*d = InlYt -
results asymptotically. Taking logarithms, this teXt, ix)1 on In Xt obtaining, say, f3.
(c) An NLS regression of y", = Yth- 1/2(Xf, /3)
can be accomplished by ordinary least squares
(OLS) regression of Inle*tl on In X t because (7) = v. expr- ~ (In X t )I/3J onf*(Xt , a) = exp
implies
r(ln Xt)'a - ~ (In X t )I/3J obtaining, say, ix.
(8) Inlctl = f30 + ~ (In X t )' f3 + u*t,

E(u*t) = 0 Production Function Estimation with

Experimental Data
where (3o = E(ln Ut), u", = In u, - E(ln Ut),2 and The steps of estimation in (a), (b), and (c)
(9) In h (X t, (3) = (In X t)' f3. require some modification when data are gen-
erated experimentally by a cross-section of
Furthermore, OLS yields a consistent es- time series where perhaps the same random
timator for f3 under the same conditions re- phenomena, such as weather, affect all con-
quired for consistency in the first-stage NLS temporaneous observations." The investiga-
problem. tion of data with time or cross-section effects
Given consistent estimation of f3 in the re- has been discussed extensively in the litera-
gression problem in (8), it is possible to com- ture on variance components (e.g., Hoch;
pute a nonlinear generalized least-squares es- Wallace and Hussain). A general conclusion is
timator for the regression in equation (6) in
essentially the same way as proposed by Zell- 3 One can also note in this case that a negative estimate of the

variance of c, will not result as in the Hildreth and Houck treat-

ner and Hildreth and Houck. Where /3 is the ment of random coefficient regression because of the log-linear
estimator obtained above, one obtains a con- specification of h.
sistent estimator h (X t , /3) for the variance of 4 Alternatively or additionally, one similarly could consider plot

effects across years. The possibility of plot or cell-associated

random terms, however, will not be discussed explicitly in this
2 Note that f30 is defined to take account of the nonzero expecta- paper since they are not necessary in the empirical cases consid-
tion of InIE*,I. This modification is required to obtain consistency ered explicitly; furthermore, the necessary generalizations follow
in estimation of f3 using the linear regression equation in (8). For a simply from the discussion in this paper. A further generalization
proof of the consistency of this approach, see Just and Pope. of the model could also be developed by considering random
Kelejian uses a similar approach. coefficients as proposed by Hildreth and Houck and Feldstein.
280 May /979 Amer. J. Agr. Econ,

that time effects can be investigated either by cross-section of plots, however, it does not
using dummy variables representing different seem reasonable to specify a plot effect be-
time periods or by using a variance compo- cause the plots were in very close proximity of
nents procedure. However, Maddala has one another (a test against this specification
shown that the variance components approach was not possible because plot attributes were
is an efficient way of combining the informa- not recorded with the data). However, it is
tion obtained from using dummy time vari- likely that time effects are important because
ables with that from using least squares with- different plots are affected by the same
out dummies. Furthermore, Wallace and Hus- weather conditions each year.
sain show that asymptotic efficiency is at- The functional forms which are investigated
tained only with the variance components ap- are the Cobb-Douglas and the translog spec-
proach when the regressors repeat from time ifications. That is, in each case one set of
estimates is obtained in which both f and h

Downloaded from http://ajae.oxfordjournals.org/ at Penn State University (Paterno Lib) on May 9, 2016
period to time period as they do with the data
used here. Although these results were de- follow the Cobb-Douglas form,
rived under linearity, it is clear that similar
results would hold in the model considered in (11) Yti = AZticq + BZt/l( Eli + Wt);
this paper if any dummy terms are included another is obtained where bothf and h follow
additively. the translog specification:
To obtain a variance components modifica-
tion of the estimation procedure outlined (12) Yti = AZti(l:l ex~ ~ 0:2 ln Z ti}
2

above, equation (3) can be augmented by ad-

ding an error term w,, as well as time and plot
subscripts t and i, to existing variables in the + BZt/l ex p{ ~ e; ln2Zti}(Eti + Wt)·
equation. The model thus becomes
Note that the above equations are both special
(10) Yti = !(Xli, a) + h 1/ 2(Xti , /3)(Eli + wt ) , cases of (9).
Following the estimation methodology of
with E( Eti) = 0, E( W t) =
0, E( Eli ETj) = 0ti)ij, the previous section, the results are presented
E(WtWr) = Str(T, E(EliWr) = 0, and where (T is in tables 1, 2, and 3 for corn and oats. In each
the variance of Wt and Su is the Kronecker case the dependent variables are measured in
delta. A generalization of the estimation pro- bushels per acre, and the independent vari-
cedure in the previous section which attains
able is pounds of nitrogen per acre (see Day
consistency for this model is given in the ap- for the data and a further description). Table 1
pendix of this paper. contains the ordinary nonlinear least-squares
An Application in Fertilizer Response estimates of equation (AA).5 Table 2 contains
5 The standard errors reported in parentheses in table 1 should
To demonstrate the approach of this paper, thus be interpreted with caution because disturbances in (A.4) are
the notable data set investigated by Day has not unifonnly distributed under the assumptions in (A.I) and
(A.2).
been selected. Briefly, his yield data were
generated for several crops by controlled ex- Table 1. First-Stage Estimates of the Deter-
periments varying fertilization over several ministic Component of Production
fixed levels. Observations were generated by
recording time series of data over a cross- Constant Fertilizer
Term Coefficients"
section of plots. Using these data, Day at- Functional
tempts to determine the distribution of yield at Crop Form A Ql Q2
each level of fertilization among the family
of Pearson-type distributions. His data are use- Com Cobb-Douglas 14.759 .3053
(2.195) (.0445)
ful for exemplifying the methodology of this
Com Translog 20.916 .0634 .0806
paper because problems of multicollinearity ( 14.384) (.4698) (.1561)
are not encountered and because concentra- Oats Cobb- Douglas 10.900 .4195
tion on a single input in investigating risk ef.. (1.248) (.0317)
fects is possible. Oats TransJog 6.194 .8804 -.1568
Because the data possess both time-series (2.597) (.2356) (.0653)
and cross-section characteristics, the method- Note: Figures in parentheses are standard errors, estimated under
ology of the previous section is applicable. the assumption of homoscedasticity.
Even though the data were generated with a a See equations (II) and (12) for functional forms.
Just and Pope Production Functions and Risk 281

Table 2. Second-Stage Estimates of the Stochastic Component of Production

Fertilizer Standard
Constant Term Coefficients' Error

Functional a= /30
Crop Form f3*o f30 131 f3.l f3*o - 130
Com Cobb- Douglas 74.169 80.605 .1269 11.524
(22.107) (25.037) (.0462)
Corn Translog 178.61 193.98 - .1852 .1058 11.621
(251.48) (273.21) (.4908) (.1653)
Oats Cobb- Douglas 42.272 57.428 .1999 2.789
( 12.686) (17.708) (.0395)

Downloaded from http://ajae.oxfordjournals.org/ at Penn State University (Paterno Lib) on May 9, 2016
Oats Translog 5.879 7.489 .7846 -.1693 3.652
(8.151) ( 10.499) (.3909) (.1079)
Note: Figures in parentheses are standard errors, estimated under the assumption of homoscedasticity in equation (8); see equation (8)
for the interpretation of f5* 0 and 130.
a See equations (II) and (I2) for functional forms.

the second-stage results corresponding to LettingA = AE(eOt ) and Et = eOt - E(e Ot), (13)
(A.2).6 Table 3 presents the nonlinear gener- can be rewritten as
alized least-squares results corresponding to
Yt = AX/o + BXt{3 1Et , E(€t) = 0,
equation (A.4) with covariance matrix esti-
mated on the basis of table 2. where
Consistent with Fuller's findings, the results (14) B = A = A[E(eOt)]-t,
in table 2 imply that fertilizer has a positive
marginal effect on yield variability in the rele- (15) e. = at·
vant range. This result obviously holds for all Thus, the Cobb-Douglas case examined in this
positive fertilizer levels in the Cobb-Douglas paper has a special case which includes the
formulations; for the translog cases, the mar- common specification in (13). If (13), in fact,
ginal effect on corn yields is positive if the holds, one expects to find, among other things,
fertilization rate is greater than 5.757 pounds at = f3t; hence, the usual case is apparently
per acre, and the marginal effect on the oats' not supported." If one considers the above
yield is positive if the fertilization rate is less
7 By orthogonality of regression residuals to coefficient esti-
than 102.964 pounds per acre (active nutri-
mates in the first stage, the distribution of coefficients in the
ents). It is interesting that, with translog gen- second stage is independent of that for the first stage. A test of
erality, the marginal effect on yield variability equality of coefficients (other than the constant term) in the first
is increasing for corn but decreasing for oats. two stages, based on this fact and the usual least squares statistics,
leads to rejection of equality at the 2.5% level. Unfortunately,
Another notable result is that the elasticity however, least squares statistics are not applicable, so limited
for variability is much lower in each case than confidence can be placed in the test.
the elasticity for the mean of yield. For exam- Table 3. Third-Stage Estimates of the Deter-
ple, the at estimates are considerably higher ministic Component of Production
than the f3t estimates in the Cobb-Douglas
cases. This result has important implications Constant Fertilizer
for the applicability of the usual stochastic Term Coefficients-
Functional
specification in production-function estima- Crop Form A at a2
tion, i.e., the one with log-linear disturbances.
Consider the commonly used specification Com Cobb- Douglas 3.2839 .3532
(dropping the variance components specifica- (.7476) (.0885)
tion for convenience): Com Translog .00023 8.2887 -3.1903
(.00052) (1. 7738) (.6948)
(13) Yt = !(Xt)eE( = Ax/~teOt, E(Dt ) = O. Oats Cobb- Douglas 6.0863 .3101
(1.0067) (.0531)
Oats Translog .9187 1.6013 -.4260
(.6265) (.4473) (.1442)
6 The standard errors reported in table 2 are generated by ordi-

nary nonlinear least squares. However, they may not apply Note: Figures in parentheses are standard errors, estimated under
exactly even in an asymptotic sense because the disturbances in the assumption of homoscedasticity.
(A.3) may not be uncorrelated or uniformly distributed. a See equations (11) and (12) for functional forms.
282 May 1979 Amer. J. Agr. Econ.

derivation in the translog case, it is again pursue further the misspecification problems
found that the usual specification with a log associated with log linearity [if (3), in fact,
linear disturbance is the special case of the holds]. To investigate this problem, the re-
specification of this paper where f32 = a2 holds gression results in table 4 have been derived
in addition to the conditions in (14) and (15). under the usual assumption of log linearity.
It is further interesting to compare the stan- Comparing with table 1, the resemblance of
dard error estimates in tables 1 and 3 (although statistics, aside from those pertaining to the
they only apply asymptotically). One finds in constant term, is remarkable. This is true with
all cases except constant terms that standard respect to standard error estimates as well as
error estimates are larger in table 3 than in coefficient estimates. Furthermore, the differ-
table 1. But the estimator in the third stage is ence in constant term estimates is explained
asymptotically efficient and thus possesses by the fact that only those in table 4 are esti-

Downloaded from http://ajae.oxfordjournals.org/ at Penn State University (Paterno Lib) on May 9, 2016
lower standard errors (asymptotically) than mated in logarithmic terms (see Zellner,
the first-stage estimator, i.e., the true standard Kmenta, Dreze for a discussion of related
error estimates in table 1 should be higher than problems). Thus, the methodology associated
those in table 3. Of course, the standard error with table 4 apparently is subject to the same
estimates in table 1 are not applicable under criticism made of the homoscedasticity as-
the general stochastic specification used in ob- sumption above in comparing tables 1 and 3.
taining table 3 estimates. But, unfortunately, it That is, use of the common but simple ap-
is often the case in econometric practice to proach of log linearity in disturbances (with
assume homoscedasticity when risk aspects of homoscedasticity) when, in fact, (3) is appli-
a study are presumed unimportant, or infor- cable, apparently can lead to much smaller
mation to the contrary is not readily apparent. standard error estimates than are justified
The comparison of tables 1 and 3 illustrates (asymptotically). This is true because the
the potential danger of assuming homoscedas- table 3 estimates are asymptotically efficient
ticity if, in fact, it is not applicable. In other and their standard errors are much larger.
words, if functional forms such as (3) are ap- Thus, any hypothesis testing based on tables 1
plicable in reality, the standard practice of as- and 4 could be very misleading.
suming homoscedasticity could correspond to It appears, however, that there are even
the first-stage estimation of this paper. But, if more serious problems of bias associated with
so, the associated statistics apparently would log linear stochastic specification when, in fact,
be misleading and imply much more precise (3) is applicable. In logarithmic terms, and
estimation than is warranted. assuming log linearity off, the specification in
(3) becomes
Implieations of Assuming Log
Linear Disturbances (16) in Yt = in f(X t , a) + u,
= (in Xt)'a + u.,
Because the more common approach is to as- where
sume homoscedasticity with log-linear distur-
bances(rather than additivity as in the first- u, = inf 1 + Eth(Xt , f3) J.
stage case of this paper), it is interesting to L f(X t , a)

Table 4. OLS Estimates under the Assumption of Homoscedastic Log-Linear Disturbances

Constant Fertilizer
Term Coefficients
Functional
Crop Form In A a1 a2 R2

Corn Cobb-Douglas 2.5667 .3246 .219

(.1310)3 (.0413)
Corn Translog 2.8830 .0961 .0785 .220
(.6417) (.4556) (.1560)
Oats Cobb-Douglas 2.4897 .3855 .308
(.1337) (.0390)
Oats Translog 1.8026 .8286 -.1354 .316
(.4557) (.2838) (.0859)

a Figures in parentheses are standard errors.

Just and Pope Production Functions and Risk 283

Thus, where X = (In Xl' ... , In X r)' and V = effect of input on expected output and another
(u b . . . , ur)', the OLS estimate a* in (16) has explaining the effects of input on variability of
expectation output.
(e) Such a production function is given by
E(a*) = a + (X'X)-IX'E(V) 01= a
(3), and a three-step NLS procedure (with or
since E(V) < 0 by Jensen's inequality and without variance components) is proposed.
concavity of u, in Et, Use of a maximum likelihood procedure
(without variance components) to attain
E[Ut(Et)] < UtE(Et) = O. efficiency in estimating h( . ) is also discussed.
Of course, 0 LS also is inconsistent because (f) An application indicates that fertilizer
the magnitude of bias does not generally de- has a variance-increasing effect on yield; but
pend on T (see Kmenta for a similar discus- the marginal variance contribution is much

Downloaded from http://ajae.oxfordjournals.org/ at Penn State University (Paterno Lib) on May 9, 2016
sion). Taking expectations of a Taylor-series smaller than the standard log-linear distur-
expansion of u., bance approach would indicate.
(g) Empirical results indicate that estima-
1 E(El)h 2 (X t, f3)
E(Ut) = tion assuming the homoscedastic log-linear
2 [2(Xt , a) disturbance case can lead to standard error
estimates that are misleading and indicate
1 E(Et4)h4(X t, f3) much greater precision in estimation than is, in
4 f4(X t , a) fact, obtained.
it also is evident that the bias is necessarily
small only when instability in production is [Received June 1977; revision accepted
small (relative to expected production). One October 1978.]
must then question the use of log linear esti-
mates off if the arguments in the first part of References
this paper hold.
Anderson, Jock. "Sparse Data, Climatic Variability, and
Yield Uncertainty in Response Analysis." Amer. J.
Conclusions and Summary Agr. Econ. 55(1973):77-82.
Bardhan, P. "Variations in Forms of Tenancy in a Peasant
As argued in the first part of this paper, tradi- Economy." J. Devel. Econ. 4(1977):105-18.
Batra, R. "Resource Allocation in a General Equilibrium
tional production-function formulations are
Model of Production Under Uncertainty." J. Econ.
uninformative with respect to risk. The flexi- Theory 8(1974):50-63.
bility given by (3), however, is within tradi- Borch, K. "A Note on Uncertainty and Indifference
tional neoclassical orthodoxy with respect to f Curves." Rev. Econ. Stud. 36(1969): 1-4.
(as far as mean response) but appears superior Christensen, L. R., D. W. Jorgenson, and L. J. Lau.
for risk analysis. For firm decision problems "Transcendental Logarithmic Production Fron-
and public policy, it seems the generalization tiers." Rev. Econ. Statist. 55(1973):28-45.
suggested by (3) potentially offers greater Crawford, R. "Implications of Learning for Economic
flexibility in describing stochastic technolog- Models of Uncertainty. " Int. Econ. Rev.
14(1973):587-600.
ical processes and related behavior.
Day, Richard. "Probability Distributions of Field Crop
In summary, the major conclusions of this Yields." J. Farm Econ. 47(1965):713-41.
paper are as follows: de Janvry, Alain. "The Generalized Power Production
(a) Traditional production-function formu- Function." Amer. J. Agr. Econ. 54(1972):234-37.
lations are very restrictive and, in particular, Feldstein, M. "Production Uncertainty with Uncertain
imply that changes in inputs are directly re- Technology: Some Economic and Econometric Im-
lated to changes in the variance of output. plications." Int. Econ. Rev. 12(1971):27-36.
(b) This restriction contradicts other re- Fuller, Wayne. "Stochastic Fertilizer Production Func-
search findings. tions for Continuous Com." J. Farm Econ.
47(1965): 105-19.
(c) Hence, traditional production-function
Halter, A. N., H. O. Carter, and J. G. Hocking. "A Note
estimates may be of little use in evaluating on the Transcendental Production Function." J.
policies-particularly those which can have a Farm Econ. 39(1957):966-74.
risk-reducing effect on output. Hildreth, C., and J. Houck. "Some Estimates for a Linear
(d) A simple reasonable production- Model with Random Coefficients." J. Amer. Statist.
function specification must apparently contain Assoc. 63(1968):584-95.
at least two components-one explaining the Hoch, Irving. "Estimation of Production Function Pa-
284 May 1979 Amer. J. Agr. Econ.

rameters Combining Time Series and Cross-Section (10). In this case, a first-stage regression, such as sug-
Data." Econometrica 30(1962):34-53. gested by (6), is still possible and consistency for a is
Horowitz, I. Decision Making and the Theory ofthe Firm. obtained where C ti is defined by
San Francisco: Holt, Rinehart & Winston, 1970.
e*fi = h l / 2(Xfi , ~)(efi + wr)'
Just, Richard E., and Rulon Pope. "Stochastic Represen-
tation of Production Functions and Econometric Im- However, one finds in place of (7) a regression equation of
plications." J. Econometrics 7(1978):67-86. the form
Kelejian, H. "The Estimation of Cobb-Douglas Type
(A.1) (crY = h(XIi , ~)(l + a) + Uri;, E(uw) = 0,
Functions with Multiplicative and Additive Errors: A
Further Analysis." Int. Econ. Rev. 13(1972):179-82. because E[(e*rYJ = h(XIi , m(l
+ a), where t = 1, ... , T,
Kmenta, J. Elements of Econometrics. New York: Mac- and i = I, ... , N. Furthermore, since e*lie*(j, i j, no '*
millan Co., 1971. longer has a zero expectation [as with e.e; t 7, corre- '*
Maddala, G. "The Use of Variance Components Models sponding to (7)], one must also consider
in Pooling Cross-Section and Time-Series Data."

Downloaded from http://ajae.oxfordjournals.org/ at Penn State University (Paterno Lib) on May 9, 2016
Econometrica 39(1971):939-53.
(A.2) E*fiC(j = h I12(X r;, mh 1l2(X t b mO"
+ Utih E(urij) = 0,
Magnusson, Gudmunder. Production Under Risk: A The-
oretical Study. Uppsala: Almqvist and Wiksells, '*
for i j since E( e* ti v E* 0) = h 1/2 (Xr;, ~)h 1/2 (Xtj, mO", where
1969. t= 1, ... ,T;i= 1, ... ,j-l;and)= 1, ... ,N.Itisstill
Malinvaud, E. "The Consistency of Nonlinear Regres- the case, however, that E(E*tiC1j) = 0 if t 'T so only the '*
sions." Annals of Math. Statist. 41(1970):956-69. relationships in (A.l) and (A.2) must be considered.
Moscardi, Edgardo, and Alain de Janvry. "Attitudes To- Comparing with the earlier derivation and, in particular
ward Risk among Peasants: An Econometric Ap- equation (8), it is no longer possible even with log-linear
proach." Amer. J. Agr. Econ. 59(1977):710-16. functions to combine (A.1) and (A.2) into a linear
Radner, Roy, and Michael Rothschild. Notes on the Allo- second-stage equation replacing (8) because E*liE*(j can
cation of Effort. Center for Research in Management possibly be negative; thus, its logarithm does not exist. An
Sciences Tech. Rep. No. OW-I, University of Cali- alternative second-stage regression equation would com-
fornia, Berkeley, 1974. bine equations (A.1) and (A.2) directly:
Rothenberg, T. J., and K. Smith. "The Effect of Uncer-
(A.3) e*fiE*tj = [Bij~o
tainty on Resource Allocation in a General Equilib-
+ (I - cS;J~JhlI2(Xfi' mh 112(X u. ~) + Utij,
rium Model." Quart. J. Econ. 85(1971):440-59.
Sadan, Ezra. "The Investment Behavior of a Farm Firm where t = I .... , T; i = I, ... .]; andj = I, .... N; ~*() =
Operating under Risk." Amer. J. Agr. Econ. 1 + 0", i30 = 0". cS;j is the Kronecker delta, and the E*tk are
52( 1970):494-504. replaced by their first-stage estimates. Also, for identifica-
Stiglitz, Joseph E. "Incentives and Risk Sharing in tion purposes, any multiplicative constant term in h must
Sharecropping." Rev. Econ. Stud. 41(1974):219-55. be combined with ~*() and i3o. The addition of variance
Theil, Henri. Principles of Econometrics. New York: components in estimation thus suggests nonlinear estima-
John Wiley & Sons, 1971. tion in the second stage if asymptotic efficiency is to be
Turpin, F., and J. Maxwell. "Decision-Making Related to obtained in the third stage. That is, one cannot choose to
Use of Soil Insecticides by Indiana Com Farmers." disregard the observations generated by (A.2) and still
J. Econ. Entomol. 69(1976):359-62. identify 0";8 when (A.2) is used, the log-linear representa-
Wallace, T. D., and Ashiq Hussain. "The Use of Error tion cannot be used because e*tiE*(j is possibly negative. It
Components Models in Combining Cross-Section can be shown following Just and Pope that NLS estimates
with Time-Series Data." Econometrica 37(1969): of (A.2) are consistent under the general specification in
55-72. (9) so long as X satisfies conditions for consistency in the
Wolgin, J. M. "Resource Allocation and Risk: A Case first stage and Eli has bounded fourth moments.
Study of Smallholder Agriculture in Kenya." Amer. Thus, a third-stage, nonlinear, generalized least-squares
J. Agr. Econ. 57(1975):622-30. regression of Yri on Xfi, using the equation
Zellner, Arnold. "An Efficient Method of Estimating
(AA)
Seemingly Unrelated Regressions and Tests for
Aggregation Bias." J. Amer. Statist. Assoc. is possible where the covariance matrix is estimated by
57(1962):348-68. predicted values from the second-stage regressions corre-
Zellner, A., J. Kmenta, and J. Dreze. "Specification and sponding to (A.3). It can be shown that the a estimator
Estimation of Cobb-Douglas Production Function thus obtained is consistent and asymptotically efficient
Models." Econometrica 34(1966):784-95. along the same lines as in the case considered by Just and
Pope without error components.

Appendix

Generalized Estimation
8 This is true so long as h contains a multiplicative constant term
(so that homoscedasticity is a special case), in which case the
Consider multistage estimation of the error components estimate of f3/o would be an estimate of the multiplicative constant
version of the production function model specified by times I + cr.