Post on 27-Jan-2017
ESTIMATION OF RISK EFFECTS WITH SEEMINGLY
UNRELATED REGRESSIONS AND PANEL DATA
Guang H. Wan, William E. Griffiths
and Jock R. Anderson
No. 40 - September, 1989
ISSN
ISBN
0157-0188
0 85834 843 8
Department of EconometricsUniversity of New EnglandARMIDALE NSW 2351 AUSTRALIA
ESTIMATION OF RISK EFFECTS WITH SEEMINGLY
UNRELATED REGRESSIONS AND PANEL DATA
Guang H. Wan and William E. Griffiths
University of New England, Armidale, NSW 2351, Australia
:lock R. Anderson
The World Bank, Washington, D.C., USA
Abstract
In this paper, production functions in the form of seemingly unrelated
regressions (SUR) with errors that are heteroscedastic and that contain
cross-section and time components are proposed. These functions are dis-
tinguished from others in that they allow the risks (indicated by variances)
of outputs to change in any direction in response to input changes. The
SUR are then applied in the analysis of cross-section time-series data for
rice, wheat, and maize production in China.
1 Introduction
The importance of risk (variance of output) has long been recognised in theanalysis of production functions, particularly agricultural production functions.
See, for example, Anderson, Dillon and Hardaker (1977). It is recognised thatsome inputs, e.g., investment in improving environmental conditions, are in-
versely or negatively reiated So the variance of crop outputs; whereas a positiverelationship may exist between other inputs, e.g., areas sown with modern eul-tivars (el., Anderson et al. (1989)), and the output variabilities of agricultural
crops. As long as a decision maker is not risk neutral, the relationship between
the variance of output and each of the inputs is an important ingredient into anydecisions concerning optimal input allocation. When increasing the level of an
input leads to an increase in the variance of output, we say the the marginal risk
of that input is positive. Conversely, we say an input has a negative marginalrisk when increasing its level leads to a decline in the variance of output. :lust
and Pope (1978) show that these different relationships cannot be correctly han-
dled by the commonly-used functions, no matter whether the function has an
additive error or a multiplicative error and no matter whether the function is
linear or nonlinear. For example, the widely-used Cobb-Douglas, transcendental
and CES functions, restrict the marginal product and marginal risk to be of the
same sign, normally positive. Other restrictions of these functions are detailed
by Just and Pope (1978).
To relax these restrictions, Just and Pope propose models with heteroscedas-
tic disturbances such as
Y = f(X) + h(X)~, (1)
Y = f(X) + h(X,~), (2)
Y = f(X,~), (3)
where Y and X are output and input variables respectively, e is usually a vector
of random disturbances, and h, f represent functional forms. Since equations
(2) and (3) are rather too general for insightful discussion of their estimation,
Just and Pope (1978) focus on equation (!) and suggest a four-step procedure for
estimating/1), where both f and h are assumed to be log-linear in parameters.
Using a Cobb-Douglas function for f and h, Griffiths and Anderson (1982)
consider an error component version of equation (1) and develop and apply
corresponding estimation techniques.
This paper presents an extension of the model considered by Griffiths and
Anderson (1982) into seemingly unrelated regressions (SUR). The SUR specifi-
cation is relevant when a number of different outputs exist, such as the output
of different crops, and where the disturbance terms from the functions for each
crop are correlated. This specification, which includes components for tiIne
series and cross-sectional units, is outlined in section 2. Discussion of an econo-
metric estimation procedure is presented in section 3. Some empirical results
based on Chinese data are provided in section 4 and the paper is concluded with
a summary in section 5.
2 The model
If there are N cross-sectional firms over T time periods producing M crops, a
set of M nonlinear stochastic equations of the form
K K
t’~ = 7,n H X~m’~ +em o H X~ (4)k=l k=l
can be established, where m = 1,2,..., M, 1’~ is the NT x 1 vector of ob-
servations on the output variable, ~.~ is an NT x 1 disturbance vector, X.~
is the NT x 1 vector of observations on the k-¢h input variable for the m-th
equation and as,3s are parameters to be estimated. The symbols, o and 11,
denote component multiplication of matrices. Raising the vector X.~ to the
power of 3,,~ (or a,~) means that each element in the vector is raised to the
power 3,~ (or ~,~).Assumingi= 1,2,..., N andt= 1,2,..., T, let
.K
Xmkit,
Hm = diag(h,~l, h,~,..., hmyT), (6)
H = diao(H1, H~,..., HM), (7)
~m = Z~#r~. + Z~A.~ + u,,~, (8)
where
Z. = IN ® eT, (9)
Z~ = eN ® IT, (10)
and ® denotes the Kronecker operation; IN, I~, denote N x N and T x T unit
matrices, ear,e~ are N x 1 and T x 1 vectors of ones. The model can then bewritten as
K
= 1I x +
where the i-¢h element of the vector ~,~ = [/~,~l,/~m~,-",/~m~r]’ and the i-th
element of the vector A,~ = [)wni, ;k,~.,, .. ¯, X,~T]~ represent the error components
specific to the i-th firm and t-th period in the m-*h equation, respectively; the
NT × 1 vector v,~ = [V-~l, ~-~2, "" ’, v,~NT]’ contains the error component which
is random over time and space for the m-th equation. Further, define
and
Ul
UM
k=l k=l k=l
the SUR models can be written as
Y = Xc7 + u. (13)
Following Avery (1977) and Baltagi (1980), the three components of u (i.e.,
#, ~ and v) are, as seems reasonable, assumed to be stochastically independent
from each other and
However, it is assuIned that the 3 components can be correlated with their
counterparts in the equations for different crops. That is,
or in matrix notation,
O’tzml IN 0
0 cr~mtIT0 0
0
0 (17)o’~,ml I NT
for m and 1 =
subscripts and 4, s time subscripts.
By defining
�1
�2
CM
the covariance matrix for (13) can be expressed as
1, 2, ..- , M, where i,j are section subscripts, m,l equation
~2 = E(uu’) = HE(��’)H
(18)
(19)
where
~-~M1 ~M2 " "" ~-~MM
The typical element of fl denoted by ~,~z has the form
where A = IN ~ eTe~ and B = eye~ @ IT. Let
A B JNT
T N NT’]NT : eNTe~T,
then equation (21) can be alternatively expressed as (cf., B~tagi (1980))
+ ~ NT
where
(20)
(21)
(22)(23)
(24)
O’lrnl : O’vmI ÷ (25)
(26)(27)
and o’umt are the distinct characteristic roots of flint of multiplicity N- 1, T- 1,
1 and (N - 1)(T - 1), respectively. These eigenvalues of a,,~l can be computed
according to Nerlove (1971), if necessary.After obtaining the values of ~rl,~z, ~r;,~l, ~a~t and ¢~mt by equations (25)
to (27) for m, l = 1, 2,-.., M, Baltagi (1980) shows that
+ (2S)
where
all of dimension M x M. As shown later, this expression will be useful for
computing f~ - 1
Under the above model specification,/3s represent production elasticities andc~s "risk elasticities" or risk effects of inputs, where risk is defined as the variance
of Y. Since as can be of any sign, the proposed SUR are distinguished from
more conventional ones in that they allow risks of outputs to change in any direc-tion in response to input changes. Also, the three error components in the model
are al! heteroscedastic in the sense that the variances ofand H,~u,~ depend on the input levels. This implies that the magnitude of both
firm and time effects will be influenced by the measured input levels, a situation
which may be more realistic than otherwise.
3 The multi-stage estimation procedure
Given the covariance matrix of (13) in (19), it can be seen that to use nonlinear
least squares to estimate 7 = (7~, 7~,’" ’, 7M)’ and/3 = (!3t,/3~., -..,/3M)’, where
/3.,. = (/3,~i,/3.~,’" ", fl,~K)’, the objective function to be nfinimised is
where it’ = u’H-1.
However, H cannot be computed without estimates ofa = (al, as, " ¯, aM)’,
and ~ depends on the unknown variances of the error components. To estimate
these various quantities, we begin by minimising
---- ~ ~ y ~’ (34)~U mit -- ~m Xmkit
m=l i=l t=l k
and obtain ~ and ~. Since cross-equation erroris not considered here, the
estimation can be undertaken for each m separately. Due to the existence of
heteroscedasticity and cross-equation error, the estimates will be asymptoticMly
inefficient, but they are generally consistent. Therefore, the estimated residual
~it = Ymit - ~ ~ X~t will converge in distribution to ~t under appro-
priate assumptions.
The second step is to estimate a. To do so, rewrite equation (12) in a slightly
different form as
u~ = h~, (~ + ~, + ,~,). (35)
Squaring the above equation and taking logarithms yields
K
In u~t = ln(,m~ + ~mt + "~it)2 + 2 ~ a~ In X~t. (36)
Let
then
K
E(ln u~t) = a,~0 ÷ 2 ~ a,~ In X,~u~t, (39)
~,~, = ln(.,~ + ~.~, + ~,.~)~ - ~o. (40)
Thus
ln(g.~i ÷ Ant ÷ r’..{t)2 : ~,~o ÷ ~.~t (41)
and equation (36) reduces to
Combining the set of M equations,
(42)
(43)
is obtained, where a = (al,a2,’’’,aM)’, am = (a,~o,2a,n,’’’,2a,~K)’, { :
({1,(2,’’’,~MNT)’,~=diag(~,~2,’’’,~M), ~m is a NTx (K + I) matr~
with 1.0s in the first column and lnXm} in the other columns, ~ is defined
similarly to ~ with ~ (In ~ ~ ~ )’.
When umit is replaced by its consistent estimator ~i~, equation (42) can be
used for estimation of a~. However, properties of (~it have to be investigated
in order to discover the properties of the estimates and to employ an appropriate
estimation technique.
If #mi, A~t and ~mi* are assumed to be normally distributed, the random
variables defined as
(44)
(45)
where
become standard normal variables with zero means and unit variance. Moreover,2q,~it, m = 1, 2,.-., M are each X2 random variables with one degree of freedom.
Taking the logarithm of the square of equation (44) produces
In 2 = ln(/z,~ + Amt + v,~)2 - lno’~qmit
am0 + (mi* In ~ (46)~ -- crm ,
where the second equality is obtained by use of equation (.1l). This variable
is thus distributed as the logarithm of a distribution with one degree of
freedom. Furthermore, this relationship between In q~it and ~mit gives us the
opportunity to derive the properties of the ~,,u, from those of the In q2mit. Since
both a,~0 and In o’~ are constant and ~t is defined by equation (38), it can be
shown (Harvey (1976)) that
Var(lnq2m,,) = Vav((,~it) = 4.9348, (47)
E(lnq2m,t) -- -1.2704
a.~o In : (48)
According to equations (38) and (47), ~,~i, has zero mean and a constant
variance. Therefore, a,~ can be estimated by applying OLS to equation (42) for
m = 1, 2,..., M separately and this produces no bias or inconsistency. But, it
does result in inefficiency since the M sets of equations are related and each of
them has a composite error structure similar to that of (17) as shown below.
When i = j and/or t = s, q,,~t and qljs will be correlated. This implies that
Inqmit" and In q~s will be also correlated when i = j and/or t = s. It can be
shown that
~(~.~, ~,~,)= E [~ q~,, ~ ~-,] - ~.2~0~~.
Since
E(qmitqlj,) - ~,,,m, i=j and t¢ s, (50)
E(qmitqlj~) = ~ t = s and i ~ j, (51)
E(qmitq,j,) = ~ i=s and i=j, (52)
E(q~,tq~j,) = O, 1�s and iCj, (53)
where ~r.~t = a~,nt + o’;~.~i + ~rv,~t, the following can be derived (Griffiths and
Anderson (1982), Johnson and Kotz (1972)):
(54)
(55)
Thus, ~mit can be viewed as having an error components structure similar to
that of e,~t and (42) can be estimated via a generalised least squares procedure
that utilises this structure. Specifically, ~ can be more efficiently estimated
by modifying the procedure and formulae in Baltagi (1980). This modification
produces generalised least squares(GLS) estimates of ~, namely 5, where
(59)
The As in the above expression are similar to fls defined by (25) to (32) with
o’s replaced by 3s. Estimated As can be specified after calculating estimated 6s
according to (54) to (57). However, such a calculation requires the estimation
of the ~rs, as will be discussed in (68) to (72). Alternatively, one can obtain the
best unbiased estimates of As directly (el., Baltagi (1980)) from
where ~ = (~, ~:,..., (M) is an NT × M matrix of residuals, which can be
obtained in two ways: (a) applying OLS to equation (42) for m = 1, 2,..-, M
separately and calculating the corresponding residuals; or (b) performing least
squares with dummy variables (LSDV) on (42) for m = 1, 2,..., M separately
10
and computing the corresponding residuals (Amemiya (1971)). It is noted that
both sets of residuals can be used to replace ~ for estimating the As and theresulting a has the same asymptotic efficiency in each case. However, As es-
timated tiom the LSDV residuals are asymptotically more efficient than those
from OLS residuals (Prucha (1984)). Thus, LSDV is used in this study to obtain
¢.Referring to both Baltagi (1980)and Prucha (1984), it can be shownthat
^Once d is obtained, generalisedleast squares residuals ~mit can befound
fronlK
^ ^’2 2~^ in X,,~,. (64)~*’nit ~ In umit - ~m0 -k=l
It is now possible to find efficient estimates of/3, which correct for het-
eroscedastieity, error components and correlation across equations. This is thetask of the third step.
According to equations (46) and (48),
= ~.~it - 1.2704,
O,~it = ~/exp(~,~it - 1.2704). (65)
It can be shown that
E(qmit qlit)
(66)
11
where O’mt = O’umt + O’:.~Z + Cq.mt and P,~z is the correlation coefficient between
em and et, which can be estimated byN T
i=1 t=l
From equation (48),
^2O’m = exp(&mo + 1.2704). (68)
These results give
&mZ = tSmt&~&t. (69)
Now, ~r,mt, ~Xmt and o’~mz can be estimated by
T N-1 N2
&~mt - NT(N 1) E E E umit ~tjt (71)- htctt=l i=1 j=i+l
~T~,ml ~ ~ml -- ~ml -- ~’~ml, (72)
whereK
~’,~, : l~ x~,. (73)
If these computations are being made with a view towards using (54) to
(57), then the estimate for &m~ would be from OLS or LSDV, rather than
GLS, because (54) to (57) are required before GLS estimation is employed.
Substituting &~mt, ~,mt, and &~mZ into equations (25) to (27) enables the
computation of 5 via equations (28) to (32). According to equation (19),
(74)
where ~ can be obtained through equations (5) to (7) with h,~t replaced by
~m~t. Thus, to obtain the efficient estimates of/3, represented by /~, it is a
matter of minimising
~__ Ut4--1U
== i~,’h-li~, (75)
12
where it = 9-1u.
For the purpose of programming, it is necessary to find a transformation
of the error term, say pi~, such that iz’p’pi, = i~’~-lit. When ~ is of small
dimension, one of the methods is to find c and ^ such that p = A-~c’, where
c is an orthogonal matrix consisting of the characteristic vectors of ~2 and /x
is a diagonal matrix consisting of eigenvalues of ~. However, f~ is of order
(MNT × MNT), which could well exceed a dinaension of 200. In this case,
finding c and A from f~ requires solving a polynomial equation of degree of over
200. This is a difficult task and unreliable results may occur. To tackle this
problem, a two step procedure is developed: (a) Decomposing ~-1 according to
the suggestion of Baltagi (1980) gives
+ (76)
where ill, h~, ha and ~u can be calculated according to equations (25) to (27)
and (29) to (32) with ~rs replaced by their estimated counterparts. It is noted
that these matrices only have dimensions of M x M. (b) Let h:~ = P4P~ and
define
for i = 1, 2, 3, 4. Further, define
A (78)D1 -T NT’B ~NT (79)D= -N NT ’&+r (80)Da -NT ’
D~ = Q. (81)
Then, since the Dis are all idempotent and DiDj = 0 for i ~ j, equation (76)
can be written as4
= ~ (P~P~ ® DiDo)i----1
® (82)
13
Therefore, an equivalent operation for minimising ~ is to minimise ii’ii, where
To summarise, the estimation of seemingly unrelated regression models,
which carry risk implications and incorporate composite errors, can be carried
out using the following steps:
(1) Find 3 and 5" by using nonlinear least squares either to minimise
for ’m :-- l, 2,..., M, or to minimise u’u; denote the corresponding residuals by
{l.
(2) Obtain & by applying the GLS techniqne on the SUR models with error
components, where in fi~,i, is regressed linearly on the ln X,,~i,s; denote the
corresponding residues by ~.
(3) Use ~ to estimate q via (65) and then p,.t,~r .......via (67), (68). This
enables the estimation of ~r,~t via (69).
(4) Use &,~t and & to find ~z,,~ from (73) and subsequently &umt, &.x,,,~ and
b~,~ from (70) to (72). Meanwhile, H can be estimated via (6) and (7).
(5) Construct ~1, ~,-, ~a and ~ by replacing o-s in (25) to (27) and (32) by
their estimated counterparts computed in step (4).
(6) Find Pi from 5, for i-- 1,2,3,4 and then obtain ~-~ from (76)
(7) Use ~ from step (4) and ~-~ from step (6) to find 5’ and ~ by employing
nonlinear least squares to minimise u’ft-~ fi-~ fiI-~u.
4 Empirical application
Chinese survey data for 28 regions (i.e., firms) for a 4-year period from 1980 to
1983 are utilised to estimate the disturbance-related production functions, as
proposed in the preceding sections. The data, covering three crops (rice, wheat
and maize), comprise output (jin*), sown-area (mu~ ), organic fertitiser (yuan~),
chemica! fertiliser (yuan), machinery cost yuan), irrigation cost (~luan), labour
"1 jin = 0.5 kg.
?1 rnu = ~ ha./As at the end of June, 1989, 3.7 yuan approximately equalled US$1.
14
input (mandays) and other costs (yuan). Those variables in value terms are
deflated by a weighted index of agricultural prices in state and free markets.
The Marquardt-Levenberg-Nash approach was used to find the nonlinear
least squares estimates offls (Marquardt (1963), Nash and Walker-Smith (1987)).
Estimates for the mean output functions and the output variance functions are,
respectively, presented in Tables 1 and 2. Estimates are given for different error
structures and different estimators. In each case the results were obtained from
several different sets of starting values.
In Table 1, estimated coefficients for the SUR heteroscedastic models are
reported in the third column. For comparison only, results from assuming uit =
~it are also presented.
From Table l(a), it is seen that, among the eight variables included in the
model, four of them have coefficients with negative signs. That is, rice pro-
duction elasticities with respect to labour, chemical fertiliser, animal cost and
machinery cost are less than zero. Since rice is mainly planted in Southern
China, where substantial underemployment or over-supply of labour exists in
the rural areas, it may be possible that negative returns with respect to labour
starts occuring, particularly after the resumption of double cropping (after triple
cropping) since the late 1970s. The negative elasticity with respect to chemi-
cal fertiliser is consistent with the findings of Wiens (1982). Large increases in
the application of nitrogen without corresponding increases in potassium and
phosphorus might be one of the most important reasons for the negative elas-
ticity (Stone (1986)). The negative elasticity associated with machinery cost
is plausible as replacement of labour by machines "destroys" the traditional
tabour-intensive farming technique. This is particularly true with rice produc-
tion since rice requires fine soil preparation and flat land, but machine operation
cannot meet these requirments as well as labour does. As for the animal cost,
the negative sign is implausible. However, except for labour, all the negative co-
efficients have ninety-five per cent confidence intervals which include a positive
range.
[Table 1 (a) near here]
Among the remaining variables, all but irrigation are significant contributors
to rice output. An examination of the magnitudes of the estimates indicates
15
that sown area asserted the greatest positive impact on rice output, followed by
organic fertiliser. The insignificance of irrigation may result from the fact that
almost all the rice area sown is irrigated and thus the effect of irrigation cannot
be estimated precisely.
The estimates of the mean maize output function are tabulated in Table l(b).
From the asymptotic t-ratios, all positive estimates are statistically significant
at the five per cent level. On the other hand, the three negative coefficients for
labour, chemical fertiliser and animal cost are implausible, but they do have
ninety-five per cent confidence intervals which include a positive range. Unlike
the case of rice, machinery cost is positively related to maize production. A
possible explanation is that, for maize production, machine operation is mainly
involved with cultivation and planting. Thus, there is less post-harvest loss when
harvesting by machines. More importantly, timing of planting is more crucial
for maize production than for rice and the requirement for seedbed preparation
is not as great as for rice. Maize is mainly grown in the central and north of
China, where farming techniques are relatively poorer than in the south. In
other words, replacement of labour by machines is likely to create a positive
impact on maize output. Moreover, in the far north the excess labour problem
is less severe if it exists at all. This may also help explain the positive sign of
[Table 1 (b) near here]
Area sown is the dominant factor influencing maize output. The production
elasticity with respect to sown area is 0.68, followed by 0.16 with respect to
organic fertiliser and 0.15 with respect to other costs. The elasticity is only 0.02
for machinery cost and 0.016 for irrigation.
The wheat production function is the function that seems to be estimated
most successfully (Table l(c)). The only negative estimate is the elasticity
with respect to irrigation. Wheat is largely planted in the far north of China,
where water supply relys heavily on rainfall. It is noted that the negative value
has a small t-statistic. Thus, the true elasticity of wheat output with respect
to irrigation nfight be very small and its estimate could well turn out to be
nonpositive.
16
Contrary to both rice and maize, the coefficients of labour, chemical fertiliser
and animal cost are all positive, although the estimate associated with animalcost is not significant at the 5 per cent level. Chemical fertiliser has the smallest
positive elasticity and organic fertiliser has the largest elasticity. The elasticitywith respect to labour is not only positive, but substantia! relative to that for the
other inputs. This comes as no surprise since wheat is predominantly planted
in the far north of China where labour is relative scarce. The above-mentioned
reason could also explain the relatively large elasticity of machinery cost inwheat production.
[Table 1 (¢) near here]
Overall, Table 1 indicates that, where labour is relatively scarce, machinery
generates a positive and significant impact on crop yield. For example, when
labour input has negative returns in rice production, machinery creates a nega-
tive effect on production and the effect is significant at a 10 per cent level. In the
case of maize production, labour had no significant impact and machinery gen-
erated a limited, though a significant effect on yield (the coefficient is only 0.02),
whereas when the labour effect is significantly positive in wheat production, the
machinery effect becomes positive, significant and substantial (the elasticity is
0.08).
The parameters whose signs determine the signs of marginal risks are pre-
sented in Table 2. Although attention will be focused on the estimates given
by GLS, parameters estimated by other techniques are also shown in Table 2.
Goodness-of-fit for each equation is measured by the R: from separate OLS ap-
plied to each equation. The goodness of fit for the SUR system was calculated
according to
2 ~(2-1 @ INT)~RsuR = 1 -}:,(E_1 ® DhrT)g (84)
where E-1 is the variance covariance matrix of the SUR models, ~ is a MNTx 1
vector containing the GLS residuals, and DNT = INT -- JNT/NT. The FsUR
statistic is obtained based on R~un (Judge et al. (1985, p. 478)). The results
(not reported in the tables) were R~SUR = 0.55 and Fs[zR = 16.01. Noting
that the data used for estimation are basically cross-sectional (the time span
is relatively short), 0.55 indicates a reasonable goodness of fit. The Fsr~R is
17
statistically significant at any conventional level, which suggests the existence
of heteroscedasticity. This may imply the inadequacy of conventional functions
or the superiority of the heteroscedastic SUR models.
Machinery, organic fertiliser and other costs seem to have stabilising effects
on rice output (Table 2(a)). The coefficient for machinery is insignificant. It is
reasonable to have a negative &s since the major component of other costs is
expenditure on management. The significance of both positive ~7 and negative
&7 imply the importance of organic fertiliser in achieving a high and stable yield
in China’s rice production. The variance of production is positively related to
chemical fertiliser application, though there is a lack of statistical significance.
This is in line with the expectation of Hazell (1984), who suspected that, as
seed-fertiliser technology advances with the adoption of high yielding varieties,
increased use of chemical fertiliser may bring about higher production vari-
ability. As far as animal cost is concerned, the significantly positive sign is
implausible and thus needs further investigation. While irrigation is expected
to help stabilise production, the empirical result here does not seem support-
able. Noting that &5 is significant and of considerable magnitude, it seems that
better management of the irrigation system in China is urgently needed. The
positive d5 and negative ~ could well be the result of a malfanctioning of the
irrigation system due to a collapsed management of water resource and irriga-
tion facilities after the introduction of the agricultural production responsibility
system in late 1978. The impact of area sown on production risks basically
depends on the correlation coefficients among rice outputs of different seasons
and on management skills. In general, a positive relationship is expected. Fi-
nally, labour does not produce a significant impact on production risk. This is
primarily because the labour input in China was near "saturation" long before
1980. Thus its changes may not generate any effect on either mean output or
output risk.
[Table 2 (a) near here]
Contrary to the case of rice production, animal cost and irrigation were
estimated to be stabilising factors in maize production in China (Table 2(b)).
This may be due to the relative insensitivity of maize to water supply in timing,
quantity and frequency. In other words, irrigation can help to stabilise maize
18
production and, while there are irrigation problems in China, these may only
generate a very limited impact on maize yield variability. The variables other
than animal cost and irrigation are all positively related to maize production
variance. This is plausible for labour, area sown and chemical fertiliser for the
reasons discussed earlier. The positive signs of machinery, organic fertiliser and
other costs are implausible. However, all the positive estimates have ninety-five
per cent confidence intervals which include negative values.
[Table 2 (b) near here]
The relationship estimated between wheat output variance and inputs can
be found in Table 2(c). All the slope parameters are insignificant at the 5 per
cent level. The negative value for area sown is unexpected as is that for organic
fertiliser. The estimates associated with labour and other costs are not only
positive, but also quite large in magnitude. It should be stressed that all the
slope coefficients could well be zeros in accordance with the asymptotic ~-ratios.
[Table 2 (c) near here]
It is difficult to generalise findings from the estimates of the three equations
because (a) most of the estimates are not encouraging in terms of statisticzd
significance; and (b) the magnitudes of and especially the signs of parameters
are often inconsistent across equations. However, as far as the relationship
between the ’green revolution’ and production risks is concerned, the empirical
results indicate that there is a positive link between seed-fertiliser techno!ogy
and output variability. This may be due to the introduction of modern cultivars
which have a narrower genetic base than their predecessors (Hazell (1984)). The
nature of irrigation in the context of output variability crucially depends on the
reliability of the water supply. Taking into account the fact that the irrigation
systems in many parts of China are severely damaged, a nonnegative effect of
irrigation on output risk may be understandable. The machinery input possibly
brought about higher risks, which could arise from the poor quality of both
tools and operations.
The estimated matrices [~u-~t], [~vmt], [~,~t] and [~,~1] are given in Table 3.
A comparison of the corresponding values of [~x-~t] and [~u,~t] indicates that
the time effect may be negligible. Cross-equation covarianees are given by the
19
off-diagonal values of [~,~z]. If the assumption of a normal distribution for each
of the three (time, firm and random) errors holds, the diagonal elements of [~mZ]
should be close to 4.9348. See equation (47). Statistical tests (F statistics) are
undertaken and it is found that all three values are not significantly different
from 4.9348 at a five per cent level. In passing, it is noted that the negative
values on the diagonals of these matrices are possible and they can be set to
zero in practice if necessary (Fuller and Battese (1974, p. 72)).
[Table 3 near here]
The variance-covariance matrices of the mean output functions are given inTable 4. The lack of time effects is again seen by the small ratios of
The contemporary covariances across equations are all positive and substantial.
[Table 4 near here]
5 Summary
In this paper, SUR models which incorporate time-specific and firm-specific er-
ror components and permit the marginal variances of outputs to have either
positive or nonpositive signs are proposed. An estimation procedure is sug-
gested. This attempt is of empirical significance, particularly in agro-economic
research, since outputs of various agricultural activities tend to be influenced by
some common factors, notably weather and policy changes. Also, increases of
different inputs can either enhance or reduce output risks. Conventional SUR
models restrict the marginal risks to be positive.
Using combined time-series (4 years) and cross-section (28 regions) data
on Chinese rice, maize and wheat production, heteroscedastic SUR production
functions were estimated. The results indicate that, as chemical fertiliser, sown
area and irrigation costs increase, output variances generally rise. On the other
hand, organic fertiliser, machinery cost and the other costs may help stabilise
Chinese cereal production. The labour input does not create significant impacts
on either mean outputs or output variances. These results suggest the possible
superiority of the heteroscedastic SUR models over more conventional ones.
2O
It must be noted that most of the inputs considered in the models are not
necessarily siginificantly related to production risks. This is not to suggest that
these and other inputs are, in fact, unimportant or unnecessary in production
and its riskiness. It may, however, imply the importance of weather and govern-
inent intervention in agriculture in determining the variability of Chinese cereal
production.
21
References
Amemiya, T., 1971, The estimation of the variances in a variance-components
model, International Economic Review 12, 1-13.
Anderson, J.R., C.J. Findlay and G.H. Wan, 1989, Are modern cultivars more
risky ? a question of stochastic efficiency, in: Anderson, J.R. and P.B.R. Hazell,
eds., Variability in grain yields: implications for agricultural research and policy
in developing countries (:lohns Hopkins University Press, Baltimore) pp. 470-
482.
Anderson, J.R., J.L. Dillon and J.B. Hardaker, 1977, Agricultural decision analysis
(The Iowa State University Press, Ames, Iowa).
Avery, R.B., 1977, Error components and seemingly unrelated regressions, Econo-
metrica 45, 199-209.
Baltagi, B.H., 1980, On seemingly unrelated regressions with error components,
Econometrica 48, 1547-1551.
Fuller, W.A. and G.E. Battese, 1974, Estimation of linear models with crossed-error
structure, :Iournal of Econometrics 2, 67-78.
Griffiths, W.E. and :I.R. Anderson, 1982, Using time-series and cross-section data
to estimate a production function with positive and negative marginal risks,
Journal of the American Statistical Association 77, 529-536.
Harvey, A.C., 1976, Estimating regression models with multiplicative heteroscedas-
ticity, Econometrica 44, 461-465.
Hazell, P.B.R., 1984, Sources of increased instability in Indian and US cereal pro-
duction, American Journal of Agricultural Economics 66, 302-311.
Johnson, N.L. and S. Kotz, 1972, Distribution in statistics: continuous multivariate
distributions (John Wiley, New York) p. 226.
Judge, G.G., W.E. Griffiths, R.C. Hill, H. Lfitkepohl and T.C. Lee, 1985, The
theory and practice of econometrics, 2nd ed.(John Wiley and Sons, New York).
Just, R.E. and R.D. Pope, 1978, Stochastic specification of production functions
and economic implications, Journal of Econometrics 7, 67-86.
Marquardt, D.W., 1963, An algorithm for least squares estimation of nonlinear
parameters, Journal of the Society for Industrial and Applied Mathematics 11,
431-441.
Nash, J.C. and M. Walker-Smith, 1987, Nonlinear parameter estimation, vol 82
22
(Marcel Dekker, New York).
Nerlove, M., 1971, A note on error components models, Econometrica 39, 383-396.
Prucha, I.R., 1984, On the asymptotic efficiency of feasible Aitken estimates for
seemingly unrelated regression models with error components, Econometrica
50, 203-207.
Stone, B., 1986, Chinese fertiliser application in the 1980s and 1990s: issues of
growth, balance, allocation, efficiency and response, in: China’s economy looks
towards the year 2000, vol 1, selected papers submitted to the joint economic
committee, Congress of the United States (U.S. Government Printing Office,
Washington, D.C.) 453-496.
Wiens, T.B., 1982, The limits to agricultural intensification: the Suzhou experience,
in: China under the four modernisations, vol 1, selected papers submitted to
the joint economic committee, Congress of the United States (U.S. Government
Printing Office, Washington, D.C.).
23
Table 1: Parameter Estimates for the Mean Output Function(a) Rice
Error Structure
nit = ~it nit = (gi+At÷~it)hit
6.836 269.663(3.29) (3.37)
~1 -0.181 0.728(*tea) (-2.47) (S.TS)
~2 0.347 --0.125(Labour) (10.52) (-2.38)
~3 -0.007(Chemical Fertiliser) (-0.18)
~4 -0.031 -0.035(Animal Cost) (-1.61) (-1.43)
~5 -0.010 0.017(Irrigation) (-0.75) (0.73)
~6 -0.031(Machinery Cost) (-2.59)
~7 0.905(Organic Fertihser) (11.98)
0.379(5.46)
~s 0.120 0.098(Other Costs) (7.30) (3.00)
Note: Figures in brackets are asymptotic t-ratios.
Table 1: Parameter Estimates for the Mean Output Function(b) Maize
Error Structure
~t =~t u~t=(~+At+~)h~t325.277 415.365(4.02) (6.42)
~1 0.584 0.676(Area) (15.10) (13.24)
~2 0.012 -0.027(Labour) (0.32) (-0.70)
~a 0.0002(Chemical Fertihser) (0.01)
~4 -0.062 -0.017(Animal Cost) (-5.72) (-0.81)
~6 0.013 0.016(Irrigation) (1.79) (1.94)
~6 0.009(Machinery Cost) (0.88)
0.022(2.13)
~7 0.106(Organic Fertihser) (2.71)
0.161(4.55)
~s 0.330 0.147(Other Costs) (8.61) (5.36)
Note: Figures in brackets are asymptotic t-ratios.
Table 1: Parameter Estimates for the Mean Output Function(c) Wheat
Error StructureUit = bit Uit = (l~i + At + ~it)hit103.304 136.950(3.79) (3.75)
81 0.382 0.198(Area) (4.41) (1.96)
82 0.212 0.140(Labour) (4.95) (2.14)
83 -0.014(Chemical Fertiliser) (-0.82)
0.049(1.98)
84 -0.044 0.064(Animal Cost) (-2.37) (1.79)
85 0.011 -0.024(Irrigation) (0.80) (- 1.21)
86 0.074(Machinery Cost) (2.90)
0.084(3.29)
87 0.230(Organic Fertihser) (7.60)
0.261(4.09)
8s 0.148 0.187(Other Costs) (2.78) (3.13)
Note: Figures in brackets are asymptotic t-ratios.
Table 2: Parameter Estimates for the Output-Variance Function(a) Rice
Estimation TechniqueOLS LSDV GLS
22.432 - 18.813(14.90) - (7.66)
251 3.443 1.339 2.072(Area) (8.02) (0.67) (2.56)
2&2 -0.910 0.651 0.174(Labour) (-3.03) (0.44) (0.32)
253 0.303 1.017 0.515(Chemical Fertiliser) (2.82) (1.16) (1.92)
254 0.470 1.228 1.005(Animal Cost) (2.63) (1.36) (2.84)
2&~ 0.273 !.049 0.796(Irrigation) (1.69) (1.27) (2.40)
256 0.101 -0.179 -0.005(Machinery Cost) (1.08) (-0.33) (-0.03)
257 -2.003 -1.987 -1.850(Organic Fertiliser) (-5.59) (-1.22) (-2.84)
253 -0.450 -1.894 -1.433(Other Costs) (-1.89) (-1.66) (-3.17)
R2 0.637F-ratio 22.609
Note: Figures in brackets are asymptotic t-ratios.
Table 2: Parameter Estimates for the Output-Variance Function(b) Maize
Estimation TechniqueOLS LSDV GLS9.138 - 9.107(6.77) - (3.91)
2&I 0.377 1.675 1.056(Area) (0.83) (0.45) (1.34)
2&: 0.368 0.013 0.058(Labour) (1.03) (0.004) (0.11)
2&3 0.040 0.054 0.099(Chemical Fertiliser) (0.35) (0.08) (0.56)
2~4 --0.055 -0.785 -0.440(Animal Cost) (-0.39) (-0.70) (-1.77)
2&s -0.160 -0.436 -0.182(Irrigation) (-2.22) (-0.79) (-1.28)
2&6 0.282 0.390 0.349(Machinery Cost) (2.87) (0.51) (1.86)
2&7 0.340 0.338 0.437(Organic Fertiliser) (1..08) (0.16) (0.76)
2&s 0.250 0.484 0.261(Other Costs) (0.82) (0.29) (0.49)
R2 0.533F-ratio 14.684
Note: Figures in brackets are asymptotic t-ratios.
Table 2: Parameter Estimates for the Output-Variance Function
(c) Wheat
Estimation TechniqueOLS LSDV GLS8.310 7.226(6.99) (2.54)
(Area)
(Labour)
0.016 -0.709 -0.407(0.04) (--0.37) (-0.41)
0.886 -0.207 0.700(3.40) (--0.17) (1.26)
2&3 -0.331 0.290 0.018(ChemicM Fertiliser) (-3.01) (0.96) (0.09)
254(Animal Cost)
-0.005 0.003 -0.073(-0.03) (0.01) (-0.24)
(Irrigation)-0.002 0.184 0.056(-0.02) (0.42) (0.24)
25’6 0.444 -0.011 0.247(Machinery Cost) (4.14) (-0.03) (1.13)
2&7 -0.141 0.682 0.113(Organic Fertiliser) (-0.47) (0.67) (0.19)
2&g 0.482 1.270 0.878(Other Costs) (1.64) (1.58) (1.65)
R~ 0.596F-ratio 19.010
Note: Figures in brackets are asymptotic t-ratios.
Table 3: Covariance Matrices of Output-Variance Functions
-0.010 -0.043 -0.168:0.043 0.058 -0.046-0.168 -0.046 0.148
1.660 1.336 -0.6971.336 3.373 -0.402
-0.697 -0.402 3.321
5.293 -0.227 -0.088
[~~~z] -0.227 2.803 0.122-0.088 0.122 2.009
6.093 0.833 -0.576
0.833 5.035 -0.118-0.576 -0.118 3.981
Table 4: Covariance Matrices of Mean Output Functions
1570.784 619.503 -131845.170619.503 246.119 -277397.578
-131845.170-277397.578 76840813.036
15928.522 5809.082 840678.539[&~mZ] 5809.082 4593.849 271162.822
840678.539 271162.822 186857666.081
14630.159 3448.611 1982192.8153448.611 56.530 985374.936
1982192.815 985374.936 263829703.787
WORKING PAPERS IN ECONOMETRICS AND APPLIED STATISTICS
The Prior Likelihood and Best Linear Unbiased Prediction in StochasticCoefficient Linear Models. Lung-Fei Lee and William E. Griffiths,No. 1 - March 1979.
Stability Conditions in the Use of Fixed Requirement Approach to ManpowerPlanning Models. Howard E. Doran and Rozany R. Deen, No. 2 - March1979.
A Note on A Bayesian Estimator in an Autocorrelated Eeror Model.William Griffiths and Dan Dao, No. 3 - April 1979.
On R2-Statistics for the General Linear Model with Nonscalar CovarianceMatrix. G.E. Battese and W.E. Griffiths, No. 4 - April 1979.
Const~ction of Cost-Of-Living Index Numbers - A Unified Approach.D.S. Prasada Rao, No. 5 - April 1979.
Omission of the Weighted First Observation in an Autocorrelated RegressionModel: A Discussion of Loss of Efficiency. Howard E. Doran, No. 6 -June 1979.
Estimation of Household Expenditure Functions: An Application of a Classof Heteroscedastic Regression Models. George E. Battese andBruce P. Bonyhady, No. 7 - September 1979.
The Demand for Sawn Timber: An Application of the Diewert Cost Function.Howard E. Doran and David F. Williams, No. 8 - September 1979.
A New System of Log-Change Index Numbers for Multilateral Comparisons.D.S. Prasada Rao, No. 9 - October 1980.
A Comparison of Purchasing Power Parity Between the Pound Sterling andthe Australian Dollar - 1979. W.F. Shepherd and D.S. Prasada Rao,No. I0 - October 1980.
Using Time-Series and Cross-Section Data to Estimate a Production Functionwith Positive and Negative Marginal Risks. W.E. Griffiths andJ.R. Anderson, No. ii - December 1980.
A Lack-Of-Fit Test in the Presence of Heteroscedasticity.and Jan Kmenta, No. 12 - April 1981.
Howard E. Doran
On the Relative Efficiency of Estimators Which Include the InitialObservations in the Estimation of Seemingly Unrelated Regressionswith First Order Autoregressive Disturbances. H.E. Doran andW.E. Griffiths, No. 13 - June 1981.
An Analysis of the Linkages Between the Cons~ner Price Index and theAverage Minimum Weekly Wage Rate. Pauline Beesley, No. 14 - July 1981.
An Error Components Model for Prediction of County Crop Areas Using Surveyand Satellite Data. George E. Battese and Wayne A. Fuller, No. 15 -February 1982.
Networking or Transhipment? Optimisation Alte~at~ves for Ptanl. l,oc~! iwnDecisions. H.I. ~oft and P.A. ea~sidF0 ~o. 16 - ~ebruary 1985.
A Further Consideration of Causal Relationships Between Wages and Prices.
J.W.B. Guise and P.A.A. Beesley, No. 18 - February 1985.
A Monte Carlo Evaluation of the Power of Some Tests For Heteroscedasticity.
W.E. Gr±ffiths and K. Surekha, No. 19 - August 1985.
A Walrasian Exchange Equilibrium Interpretation of the Geary-Kh~nisInternational Prices. ~.~. Pra~a~a ~ao, ~o. 20 - O~to~er 1985.
Using Ourbin’s h-Test to Validate the Partial-Adjustment Model.H.E. Doran, No. 21 - November 1985.
An Inw~stigation into the Small Sample Properties of Covariance Matrixand Pre-Test Estimators for the Probit Model. William E. Griffiths,R. Carter Hill and Peter J. Pope, No. 22 - November 1985.
A Bayesian Framework for Optimal Input Allocation with an UncertainStochastic Production Function. William E. Griffiths, No. 23 -February 1986.
A Frontier Production Function for Panel Data: With Application to theAustralian Dairy Industry. T.J. Coelli and G.E. Battese, No. 24 -February 1986.
Identification and Estimation of Elasticities of Substitution for Firm-Level Production Functions Using Aggregative Data. George E. Batteseand Sohail J. Malik, No. 25 - April 1986.
Estimation of Elasticities of Substitution for CES Production FunctionsUsing Aggregative Data on Selected Manufacturing Industries in Pakistan.George E. Battese and Sohail J. Malik, No.26 - April 1986.
Estimation of Elasticities of Substitution for CES and VES ProductionFunctions Using Firm-Level Data for Food-Processing Industries inPakistan. George E. Battese and Sohail J. Malik, No.27 - May 1986.
On the Prediction of Technical Efficiencies, Given the Specifications of aGeneraLized Frontier Production Function and Panes Data on S~nple Firms.George E. Battese, No.28 - June 1986.
A General Equilibrium Approach to the Construction of Multilateral IndexNumbers. D.S. Prasada Rao and J. Salazar-Carrillo, No.29 - August1986.
Further Results on Interval Estimation in an AR(1) E~,r~on Model.W.E. Griffiths and P.A. Beesle¥, No.30 - August 1987.
Bayesian Econometrics and How to Get Rid of Those Wrong Signs.Griffiths, No.31 - November 1987.
H.E. Doran,
William E.
3o
Confidence Intervals for the Expected Average Marginal Products ofCobb-Douglas Factors With Applications of Estimating Shadow Pricesand Testing for Risk Aversion. Chris M. Alaouze, No. 32 -September, 1988.
Estimation of Frontier Production Functions and the Efficiencies ofIndian Farms Using Panel Data from ICRISAT’s Village Level Studies.G.E. Battese, T.J. Coelli and T.C. Colby, No. 33 - January, 1989.
Estimation of Frontier Production Functions: A Guide to the ComputerProgram, FRONTIER. Tim a. Coelli, No. 34 - February, 1989.
An Introduction to Australian Economy-Wide Modelling. Colin P. Hargreaves,
NO. 35 - February, 1989.
Testing and Estimating Location Vectors Under Heteroskedasticity.William Griffiths and George Judge, No. 36 - February, 1989.
The Management of Irrigation Water During Drought. Chris M. Alaouze,
No. 37 - April, 1989.
An Additive Property of the Inverse of the Survivor Function and theInverse of the Distribution Function of a Strictly Positive RandomVariable with Applications to Water Allocation Problems.Chris M. Alaouze, No. 38 - July, 1989.
A Mixed Integer Linear Progra.~ning Evaluation of Salinity and WaterloggingControl Options in the Murray-Darling Basin of Australia.Chris M. Alaouze and Campbell R. Fitzpatrick, No. 39 - August, 1989.
Estimation of Risk Effects with Seemingly Unrelated Regressions andPanel Data. Guang H. Wan, William E. Griffiths and Sock R. Anderson,No. 40 - September, 1989.