WPS4536
Policy ReseaRch WoRking PaPeR 4536
Heterogeneous Technology and Panel Data:
The Case of the Agricultural Production Function
Yair Mundlak
Rita Butzer
Donald F. Larson
The World Bank
Development Research Group
Sustainable Rural and Urban Development Team
February 2008
Policy ReseaRch WoRking PaPeR 4536
Abstract
The paper presents empirical analysis of a panel of variables. The empirical results differ from those reported
countries to estimate an agricultural production function in the literature for cross-country studies, largely in
using a measure of capital in agriculture absent from most augmenting the role of capital, in combination with
studies. The authors employ a heterogeneous technology productivity gains, as a driver of agricultural growth. The
framework where implemented technology is chosen results indicate that total factor productivity increased at
jointly with inputs to interpret information obtained in an average rate of 3.2 percent, accounting for 59 percent
the empirical analysis of panel data. The paper discusses of overall growth. Most of the remaining gains stem from
the scope for replacing country and time effects by large inflows of fixed capital into agriculture. The results
observed variables and the limitations of instrumental also suggest possible constraints to fertilizer use.
This paper--a product of the Sustainable Rural and Urban Development Team, Development Research Group--is part
of a larger effort in the department to examine the determinants of agricultural growth. Policy Research Working Papers
are also posted on the Web at http://econ.worldbank.org. The author may be contacted at DLarson@worldbank.org. This
study was supported by the Bank's Research Support Budget under the research project, "The Contributions of Governance
to Growth in Agriculture" (RP0 94759).
The Policy Research Working Paper Series disseminates the findings of work in progress to encourage the exchange of ideas about development
issues. An objective of the series is to get the findings out quickly, even if the presentations are less than fully polished. The papers carry the
names of the authors and should be cited accordingly. The findings, interpretations, and conclusions expressed in this paper are entirely those
of the authors. They do not necessarily represent the views of the International Bank for Reconstruction and Development/World Bank and
its affiliated organizations, or those of the Executive Directors of the World Bank or the governments they represent.
Produced by the Research Support Team
Heterogeneous technology and panel data:
The case of the agricultural production function
Yair Mundlaka, Rita Butzerb and Donald F. Larsonc
________________________________________________________________________
JEL classifications: C230, D240, Q110, O130
Keywords: Agriculture, development, economic growth, panel data, productivity
Acknowledgment
We thank Yacov Tsur for helpful comments. This study was funded by the Bank's Research
Support Budget under the research project, "The Contributions of Governance to Growth in
Agriculture" (RP0 94759).
aCorresponding author. E-mail address: mundlak@agri.huji.ac.il. Professor Emeritus, Hebrew University of
Jerusalem.
bPh.D. Candidate, Economics, The University of Chicago.
cSenior Research Economist, World Bank.
Heterogeneous technology and panel data:
The case of the agricultural production function
Introduction
The focus of empirical analysis of agricultural production functions, similar to that
of production functions in general, has changed over the years.1 Following the work of
Cobb and Douglas (1928), research centered on questions about the efficiency of the factor
markets. Research interest subsequently shifted to issues related to changes in factor
demand, leading to interest in the elasticity of substitution, to factor augmentation, to a
search for the proper algebraic form, to issues related to aggregation, and more recently to
issues related to the variability in income and in productivity growth. The statistical
aspects of the analysis were affected by the recognition of the endogeneity of the inputs,
raised by Marschak and Andrews (1944), and by the methods of accommodating this
complication in the estimation. This led to the use of panel data, to the dual approach to the
estimation, and to the use of instrumental variables.
A common assumption in much of the work was that of a homogeneous
technology, implying that a common production function generated observations used in
the analysis. In reality, firms face the practical problem of choosing which technology to
employ jointly with inputs so that technology is heterogeneous in that there is more than
one function associated with the data. In this paper we examine the consequences of this
extension in estimating an aggregate agricultural production function using panel data,
where the underlying functions are all of the Cobb-Douglas form. We model the
technology choice as conditional on predetermined variables referred to as state variables.
This approach provides a different view of the empirical results of most of the
aforementioned subjects. The root of the difference is the recognition that the observations
in the sample represent moves between functions as well as movements along a given
function. The empirical formulation allows for the dependence of parameters of the
1For the development of the work on agricultural production functions, see Mundlak (2001).
function, just as the inputs, on the state variables, and in this sense the observed, or
implemented, technology is endogenous.
The paper is oriented toward the understanding of the role of inputs and technology
in agriculture. Although the subject has a long history, it is still relevant, and is related,
among other things, to the interest in structural changes that take place in the process of
economic development. This topic of inquiry has real world implications because most of
the poor households in developing countries live in rural areas and depend on agriculture
for their livelihood.2 An increase in agricultural productivity directly contributes to the
welfare of the rural area. Changes in factor demand due to changes in technology affect the
intersectoral flow of resources, primarily labor and capital, which constitutes the essence of
the structural changes in the process of development. Understanding this process has
policy implications concerning what is worth doing and what is worth avoiding.
The analysis is related to the macroeconomic literature on the determinants of
economic growth and productivity for overall economies. The models used vary in the
parsimony of the parametric specification and in the choice of what parameters are to be
estimated and those to be imposed. The heterogeneity of approaches reflects the inability to
obtain empirically reliable and robust results that follow from basic economic reasoning.
This feeds the search for better specifications and for the appropriate ways of handling the
data. Most growth studies include a set of core Solow-Swan variables related to human
capital investment, physical capital investment, initial income conditions and either
population or labor growth.3 Studies of differences in productivity levels use similar
variables. There is less of a consensus on the broader set of state variables, though a
growing interest has arisen in the role of certain state variables such as culture, geography,
institutions, and market integration.4 Defined very broadly, institutions include the
humanly devised rules that shape economic incentives and that are particularly related to
the protection of property, the enforcement of contracts, and the dissemination of
information (North 1990; Acemoglu, Johnson, and Robinson 2005). There is no unity in
2According to the World Bank (2007), this includes 2.1 billion people living on less than $2 a day.
3See for instance the review by Durlauf, Johnson, and Temple (2005) and Durlauf and Quah (1999).
4See, for example, Sachs and Warner (1997), Hall and Jones (1999), Rodrik, Subramanian, and Trebbi
(2004), and Presbitero (2006).
2
how the state variables are to be used in the analysis. In some models they are assumed to
be the sole causes of accumulation and productivity changes (Hall and Jones 1999), while
in others they are added to inputs in the estimation of the production function. They are
also used as instrumental variables to deal with endogeneity of inputs, institutions, and
measurement errors.
Finding a set of variables that adequately describes prevailing economic incentives
is challenging for several reasons. First, while the objective of most research is to get at the
long-run determinants of economic growth, much of the variation in economic activity,
and thus in the data, is associated with short-term fluctuations, which may be linked to
quasi-fixed constraints or gaps caused by misplaced expectations. Second, economic
growth theory is open-ended, and the set of potential determinants of economic incentives
is large relative to the panel datasets available for studying macroeconomic growth.5
As a practical matter, many empirical models deal with the prevalence of short-
term variability in the data by averaging across periods, and thereby affect the empirical
results.6 Even so, some authors argue that this type of transformation is not arbitrary but
rather necessary to eliminate nuisance variations and potential biases.7 As we show, the
short-term variations are essential for identifying the production function, and thereby to
obtain the appropriate weights for calculating total factor (TF) and total factor productivity
(TFP). On the other hand, weights based on country averages provide distorted estimates
of productivity.8
In terms of end results, part of the debate involves the importance of total factor
productivity relative to accumulations of human and physical capital in determining
patterns of growth. For example, Mankiw, Romer, and Weil (1992) and Henderson and
Russell (2005) find that accumulations largely account for growth, while Klenow and
Rodriguez-Clare (1997) and Easterly and Levine (2001) place greater emphasis on
productivity growth.
5In their review of applied macroeconomic growth studies, Durlauf, Johnson, and Temple (2005) compile a
set of 145 variables that have been used as growth determinants.
6For a review of applied techniques based on growth rates, see Durlauf, Johnson, and Temple (2005).
7See, for example, Pritchett (2000).
8See, for example, Sala-i-Martin, Doppelhofer, and Miller (2004).
3
Similarly, a variety of conclusions are reached concerning the core set of state
variables. Hall and Jones (1999) emphasize the role of social capital and Rodrik,
Subramanian, and Trebbi (2004) assert the dominance of institutions over other
determinants. In other papers, trade, monetary policy, and cultural factors related to
religion, language, and colonial heritage are viewed as key determinants.9 But in their
review of the recent literature, Durlauf, Johnson, and Temple (2005, p. 558) conclude that
"(e)ven when the study of growth is viewed in terms of a collective endeavor, the various
papers cannot easily be distilled into a consensus that would meet standards of evidence
routinely applied in other fields of economics."
In terms of this literature, we deal with a sectoral production function, representing
a lower level of aggregation, but many of the issues still remain here as well. Agricultural
output has grown as a result of changes in technology and in resource allocation where the
role of labor declined and that of capital increased. Regarding the decomposition of the
output growth to TF and TFP, we rely on the empirical estimates of the production
function parameters, and as such the outcome depends on the quality of the estimates.
When the parameters of the production function depend on the state variables, the relative
contribution of the TF and the TFP is also endogenous; hence TFP can not be considered
to be the trigger of growth but rather is a result of it. The dependence of the parameters on
the state variables accounts for the wide spread in the empirical results reported in the
literature.10
The inputs used in the analysis are land, capital, fertilizer, and labor. The capital
variables, constructed for this analysis, revise and update an earlier series that was used in
Mundlak, Larson, and Butzer (1999). The state variables consist of variables representing
technology, institutions, incentives, and physical environment. A substantial portion of the
empirical section of the paper is devoted to finding a robust set of variables that adequately
accounts for agricultural output differences not explained by factors of production.
9For example, Frankel and Romer (1999) and Sachs and Warner (1997) discuss the role of trade. Examples
related to exchange rates include Dollar (1992) and Barro and Lee (1994). Barro and McCleary (2003),
among others, write about the role of religion.
10Mundlak, Larson, and Butzer (1999) summarize the empirical agricultural production functions based on
cross-country data that had appeared prior to the writing of the paper. More studies have appeared since, but
they only confirm the existence of diversity of the results.
4
I. The Model
The underlying premise is that producers at any time face more than one technique
of production, and their economic problem is to choose the techniques to be employed
together with the choice of inputs and outputs. The outline of the approach follows
Mundlak (1988, 1993). Let X be the vector of inputs and Fj(X) be the production function
associated with the jth technique, where Fj is concave and twice differentiable, and define
the available technology, T, as the collection of all possible techniques, T = {Fj(X);
j=1,...,J}. Firms choose the implemented techniques subject to their constraints and the
environment. We distinguish between constrained (K) and unconstrained (V) inputs,
X=(V,K), and assume for simplicity, without a loss of generality, that the constrained
inputs have no alternative cost. Prices for inputs (W) and output (P) are given by the
markets. The optimization problem calls for a choice of the level of inputs to be assigned
to technique j so as to maximize profits. To simplify the presentation, we deal with a
comparative statics framework and therefore omit a time index for the variables. The
extension to the intertemporal version is conceptually straightforward.
Ignoring the analytic details, we turn to characterize the solution and its
implication. Let s=(K,P,W,T) be the vector of state variables of this problem and write the
solution as: Vj*(s), Kj*(s), to emphasize the dependence of the solution on the state
variables. The optimal level of inputs Vj*, Kj* determines the intensity of implementing the
jth technique. To the extent that the implementation of a technique requires positive levels
of some inputs, when the optimal levels of these inputs are zero, the technique is not
implemented. The optimal output of technique j is Yj* = Fj(Vj*, Kj*), and the implemented
technology (IT) is defined by IT(s) = {Fj(Vj,Kj); Fj(Vj*,Kj*) 0, FjT}.
The empirical analysis is based on observations generated by production functions
that are implemented. The aggregate production function expresses the aggregate of
outputs, produced by a set of micro production functions, as a function of aggregate inputs.
This function is not uniquely defined because the set of micro functions actually
implemented, and over which the aggregation is performed, depends on the state variables
and as such is endogenous. The aggregate production function is written as:
5
Yj (s) F(X*,s) =(s)
* (I.1)
This production function is defined conditional on s, but changes in s imply
changes in X* as well as in F(X*,s), and this is summarized by the reduced form (s). It is
therefore meaningless in this framework to think of changes in X, except by `error', which
are not instigated by changes in s, or more precisely by a change in the implemented
techniques. This means that whenever the implemented technology is affected by some
state variables, it is impossible to reveal a stable production function from a sample of
observations taken over points with different state variables. Thus, in general, the
aggregate production function is not identifiable.
For (I.1) to be a production function in the usual sense, we need to introduce an
allocation error to identify that portion of the applied inputs that is disjoint from s. With
this in mind let =X X*; E(X*) = E(s) = 0; we elaborate further on the allocation error
in the next section. With this modification we write the empirical the production function
as:
jPjYj F(X,s) F(s,) (I.2)
The function F(X, s) can be approximated by a Cobb-Douglas-like function where
the coefficients vary with the state variables and possibly with the inputs:
y = (s) + (s,)x + u (I.3)
where y is the ln output, x is ln X, is redefined in terms of logs, = x x*, (s, ) and (s)
are the slope (vector) and intercept of the function respectively, and u is a stochastic term.
Variations in the state variables affect (s) and (s, ) directly, as well as indirectly
through their effect on inputs. The elasticity of output with respect to a given state variable
is
y/si = (s,)(x*/si)+[(s)/si + x*((s,)/si)] (I.4)
6
The first term on the right hand side shows the output response to a change in
inputs under constant technology. The remaining terms show the response of the
implemented technology to a change in the state variables. This part is contained in the
unexplained productivity residual in the standard productivity analysis under the
assumption of constant technology. On the other hand, the elasticity with respect to the
allocation error is
y/ = (y/x)(x/) = (s,) (I.5)
The main message of this discussion is that to obtain a consistent estimate of the
slope we need to estimatey / . Of course the allocation error is unobserved, but panel
data can help us to deal with this problem.
II. The Statistical Model11
To relate the discussion to the literature on empirical production functions, we start
with the generic Cobb-Douglas model, and thus suppress the dependence of on , which
is the quadratic component of the function. The formulation is based on the micro model,
but it is oriented toward macro data analysis by the introduction of additional state
variables. The production function implemented under state s is:
Y = (s)X (s) m0+u0
e (II.1)
where m0 is an idiosyncratic term known to the firm but not to the econometrician, u0 is a
random term whose value is unknown when the production decision is made.12 Eeu0 e ,
and without a loss in generality, it is absorbed in (s). The expectation of output,
conditional on X and s, isY e = (s)X (s) m0
e , and the choice criterion is:
e(X | s) = max(Ye -WX) (II.2)
x
11The discussion is based on Mundlak and Hoch (1965) and Mundlak (1996).
12The formulation does not allow for delayed response to the transmitted error m0 as discussed in
Chamberlain (1982).
7
where now W is the price of X measured in units of Y. The first order condition from the
vantage of the econometrician, who is blind to idiosyncratic behavior, yields the optimal
input X* conditional on s
(s)Ye / X*=W (II.3)
whereas actual input is given by X = X *em1+u1 , where m1 summarizes the idiosyncratic
behavior and u1 is a random term. The actual input, X, differs from X*, partly due to
optimization error, and partly due to the econometrician's failure to read the firms'
decision correctly. Let = ln , b = ln , y = ln Y, etc., substitute y = ye + u0 , and write the
first order condition and the production function in log form:
y - x = -b(s) + w + m1 + u1 + u0 (II.4)
y - (s)x = (s) + m0 + u0 (II.5)
Solve for x:
x = x *+
x* = -c(s)[(w - (s) - b(s) - m0]
= -c(s)(u1 + m1) (II.6)
c(s) = [1- (s)]-1
The system of equations (II.5) and (II.6) extend the standard analysis by the
inclusion of s. Both s and m0 affect jointly x* and y, and thus cause a bias in the OLS
estimation of the production function. The inclusion of the state variables in the model is
likely to reduce the direct impact of m0 on output. We return to this subject in section VI
below. To isolate the joint role of s and m0 we make two initial assumptions.
Assumption 1: (s) =
which also implies b(s)=b. To simplify the analysis, we linearize (s):
8
Assumption 2: (s) = s.
A common approach to overcome the impact of m0 on the estimates is to utilize the panel
structure of the sample. Let zit be the i,t-th observation of a raw vector z for country i and
year t, i =1,...,N, and t=1,...,T. We rewrite the jth input demand, and without a loss in
generality we normalize x by -c(s), b(s) is absorbed in (s), and note that xjit-xjit* is the
composite = ujit + mji + mjt :
jit
xjit = wjit - sit + m0i + m0t - (II.7)
jit
yit = xit + sit + m0i + m0t + u0it (II.8)
where u0it ~ IID (0, 00). ujit ~ IID (0, jj), E(u0uj) = 0; where m0t and m0i are the time
effect and country effect on the production function (respectively), and the matrices are: xit
is 1xk, as is it, sit is 1xh, is hx1.13 Let zi.. and z.t denote the averages of zit over t and i
respectively, and let z.. be the overall mean. Let W(it) be a projection matrix defined by its
operation on a vector z: W(it)z = (zit - zi.- z.t + z..) . Then the system reduces to:
W(it)xit = W (it)[wit - sit - it ] (II.9)
W (it)yit = W (it)[xit + sit + u0it] (II.10)
Observed state variables are to be included in the regression as exogenous variables
and thus cause no identification problem. State variables that are not included, observed or
unobserved, are part of the error term and as such lead to OLS biased estimates. When,
however, s consists only of country and time dummies, W(it)s vanishes and OLS of (II.10)
yield consistent and efficient estimates of . The precision of the estimate increases with
the variance of W(it) and decreases with the variance of W(it)u0.
In contrast to the within-country-time transformation, the between transformations
amplify the transmitted impact of s and m0 on the estimates. This is the case for the
13To simplify the presentation, we assume here that all inputs obey the first order condition in (II.3) and thus
j=1,...,k. More accurately, some inputs are determined by longer term contracts (such as `fixed' inputs) and
could be thought of as exogenous, but even in this case, they may be affected by s and m0.
9
between-country, defined by the projection matrix B(i)x=(xi.. x..), and the between-time,
defined by the projection matrix B(t)x=(x.t x..). The transformed between-system is
B(.)xit = B(.)[wit - sit + m0(.) -it ] (II.11)
B(.)yit = B(.)[xit + sit + m0(.) + u0 ] (II.12)
it
B(.) is either B(i) or B(t) for the between-country or between-time transformation
respectively. By construction, W(it), B(i), and B(t) are orthogonal. Note that B(.)s does not
disappear, and as such it is part of the equation disturbance and leads to the bias in the OLS
estimates. The impact of changes in s on y is summarized in equation (I.4).
The regression coefficients of interest can be written in a generic form for a
projection matrix P as: b = [(x,s)`P(x,s)]-1(x,s)`Py , where P can be any one of the
projection matrices of interest listed above with rank not smaller than the rank of the
composite matrix (x,s), where x and s are matrices built by stacking the i,t rows of the
corresponding vectors for all i and t. It is to be noted that the three regressions mentioned
here, within-country-time, between-country, and between-time, constitute a canonical set
in the sense that regressions obtained from any other linear transformation of the data, such
as pooled, within-time (country dummies), or within-country (time dummies), or by time
differencing are matrix weighted combinations of the three canonical regressions.
Some authors use first differences to eliminate the i-effect and thus eliminate the
bias in the b(i) estimates.14 This approach is less efficient than that of the within
transformation. Both estimators are linear in the observations, and under the Gauss-
Markov condition, the within estimator is the efficient one. It should be noted in this
connection that the transformation by the projection matrix P changes the variance matrix
of the disturbance. Thus, for the vector u0, var u0 = 00I, and var Pu = 00P, indicating
heteroscedasticity. However, as indicated in the appendix, when P is a projection matrix,
then GLS of the transformed equation is equal to OLS, and hence it is the best linear
unbiased estimate.
14For instance, see Lau and Yotopoulos (1989), Mairesse (1990), and Griliches and Mairesse (1998).
10
As the data generated by W(it) is cleaned from the time and country effects, they
should best represent the more stable technology, referred to here as the core implemented
technology.
III. Data15
Output and inputs
We estimate a cross-country agricultural production function where agricultural
output depends on inputs, agricultural technology, and the state of the economy. In this
analysis we use a measure of agricultural capital which revises and updates the previously
constructed data set from Larson, Butzer, Mundlak, and Crego (2000). The inclusion of
agricultural capital is one of several aspects which differentiate this study from most
studies of agricultural production functions based on a panel of countries.
Agricultural output is measured as agricultural GDP in 1990 US dollars. We choose
the GDP variable rather than the more often used agricultural production because it comes
from the national accounts used for the construction of the fixed capital variable. Inputs to
agricultural production include land, capital, labor, and fertilizers. Hectares of agricultural
area are used for the measure of land. This includes arable land, land under permanent
crops, and permanent pastures. Agricultural labor is defined as the economically active
population in agriculture. Fertilizer consumption is often viewed as a proxy for the whole
range of chemical inputs. The data on agricultural capital consists of two components:
fixed capital, consisting primarily of structures and equipment, and capital of agricultural
origin, consisting of livestock and trees. The two components differ in the method of
construction, and also in terms of markets and pricing.
Technology
As the available technology is unobserved, what we can do in empirical analysis is
to identify variables associated with variations in the implemented technology. In the case
of agriculture, there is a natural variable to measure the level of technology for a given
15For a description of the construction of the data, see Mundlak, Larson, and Butzer (1997). While the
methodology is the same, sources have been updated and in some cases revised. Details can be obtained from
the authors.
11
crop; this is the yield or output per unit of land. The yield has been the main criterion for
the introduction of the modern varieties of cereals and other crops, termed as the green
revolution, beginning around the middle of the last century. The higher is the yield of the
modern varieties, or the larger is the area devoted to these varieties, the larger is the
average yield. Extending this concept to aggregate output, we construct an aggregate peak
yield variable. For each country and each commodity, the maximum of the past yields is
computed, thereby reflecting the potential output from the implemented technology in any
given year. Country-specific Paasche indices (1990=1) are constructed of these peak
commodity yields, weighted by land area. A Paasche index is used since changing the
composition of output changes the relevance of existing technologies.
The most common variable used in empirical studies as a carrier or representative
of technology is some measure of human capital, mostly schooling. The basic idea is that
higher levels of education are conducive to technological progress.16 We include the
average schooling years of the total labor force, taken from Barro and Lee (2000).17
Empirical studies show the relevance of various public goods that are associated
with productivity, such as infrastructure in transportation and communication, measures of
public health, and research and extension.18 In this study we do not attempt to determine
the contribution of these variables individually, but rather allow for the overall effect of the
group on the estimation. We do this by selecting the per capita output in the country as a
comprehensive measure of capital and technology (Mundlak and Hellinghausen 1982). We
measure it as the ratio of the country per capita output to that of the United States and refer
to it as a development indicator. This variable replaces the need for introducing a dummy
variable to differentiate between developed and developing countries as some studies do.
16However, the causality could go in either direction in that economic progress generates a demand for
schooling. Therefore, the interpretation of a schooling variable in empirical analysis is somewhat ambiguous.
17Education data are reported for every five years through the World Bank website
(http://devdata.worldbank.org/edstats/). Data for other years are obtained through linear interpolations.
18See for instance Evenson and Kislev (1975), Antle (1983), Binswanger et al. (1987), Lau and Yotopoulos
(1989), and Craig, Pardey, and Roseboom (1997).
12
Institutions
It is assumed that the physical, legal, and regulatory infrastructure and institutions
support overall, including agricultural, development. We measure this influence with two
variables obtained from the Freedom House political rights and civil liberties. The
measure of political rights reflects the electoral process, political pluralism and
participation, and functioning of the government. The civil liberties measure includes
aspects of freedom of expression and belief, associational and organizational rights, rule of
law, and personal autonomy and individual rights. Both measures are on a scale of 1 to 7,
where 1 represents the most free and 7 the least free. If these variables matter, they are
expected to be correlated with development and reflected in the development indicator.
They are nevertheless introduced here explicitly because of our interest in trying to isolate
the effects of institutions on agricultural productivity. Hence, the expected contribution of
these variables in the present analysis is over and above that of the development variable.
Incentives
We introduce two measures of incentives to allow for the direct effect of incentives
on productivity over and above their indirect effect through resource allocation and
accumulation. The measures are the terms of trade between the agricultural sector and the
overall economy, obtained as the relative price (agricultural GDP deflator to total GDP
deflator, lagged one period), and its variability, calculated as a moving standard deviation
from the three previous periods.19 Note that this measure confounds in it the various taxes
or subsidies, direct or indirect, applied to the sector. The variability in agricultural prices
reflects the market risk faced by agricultural producers. In addition to the sector-specific
risk, there is an economy-wide market risk, that of price volatility for the economy as a
whole, measured by the rate of inflation. This is calculated as the rate of change in the total
GDP deflator.
19In an earlier paper, Mundlak, Larson, and Butzer (1999) used the price ratio of agriculture to that of
manufacturing. The reason for the change is that the alternative to agricultural resources is not limited to
manufacturing, but includes also services, which may, in fact, have become more important than
manufacturing, particularly in developed countries.
13
Physical environment
Agricultural production depends on the physical environment or natural conditions.
We represent the environment by using two variables: potential dry matter (PDM) and a
factor of water availability (FWA).20 The first variable is intended to measure the
theoretical potential production of dry matter. The production of dry matter requires
moisture. Arid areas may have a large value for PDM, but actual production is small due to
water deficit. The relative water availability is measured by the ratio of actual transpiration
to potential transpiration. These two variables are country specific and do not vary with
time.
IV. Sample Description
The sample was determined by the data availability and the preference for a
balanced data panel in order to simplify the analysis. It consists of annual data from 30
countries21 for a 29-year period (1972-2000). The information conveyed by the sample is
summarized in Table 1. The first column presents the average annual growth rate of the
variables over the sample period. Agricultural output grew at a rate of 5.43 percent,
whereas agricultural capital grew at a higher rate of 5.77 percent. Agricultural labor
declined at the average rate of 0.6 percent. Thus, the average labor productivity increased
at the average rate of 6.03 percent, and the ratio of capital to labor increased at the average
rate of 6.37 percent. The growth rate of capital of agricultural origin (4.94 percent) is lower
than that of fixed capital (5.80 percent). Fertilizer grew on average at the rate of 1.87
percent, whereas the agricultural area grew at the rate of 0.01 percent, implying a growth in
the fertilizer-land ratio. The growth of agricultural output took place in spite of unfavorable
prices as indicated by the decline in the terms of trade of agriculture at the average rate of
1.26 percent. These results signal an increase in productivity. The technology measures
show a growth rate of schooling of 1.67 percent and 1.41 percent for peak yield.
20The measures are based on Buringh, van Heemst, and Staring (1979) and were used in Mundlak and
Hellinghausen (1982) and Binswanger et al. (1987).
21Countries included in the study are: Australia, Austria, Canada, Cyprus, Denmark, Egypt, Finland, France,
Greece, India, Indonesia, Italy, Kenya, Republic of Korea, Malawi, Mauritius, Morocco, Netherlands,
Norway, Pakistan, Peru, Philippines, Sri Lanka, Sweden, Republic of Tanzania, Tunisia, Turkey, United
Kingdom, United States, and Uruguay.
14
As mentioned above, the institutions measures are ordinal, and thus, growth rates
would be meaningless. To give a picture of how institutions have evolved over the time
period studied, we looked at the averages and medians of the indices for each year. For our
sample of countries, the average measure of civil liberties is 3.27 in 1972 and 2.57 in 2000.
For political rights, the averages are 3.17 and 2.37 respectively. Our sample covers
countries on both ends of the spectrum, with a few countries advancing from 6 and 7
("non-free") to 1 and 2 ("free"),22 while there are 7 countries which remain at 1 throughout
the time period.23
The qualitative nature of the above results is consistent with the common
knowledge on agricultural development in the sample period. The results are highlighted
here for two reasons: first to characterize the sample and second to show that the data are
subject to a great deal of variability over time and across countries. This variability
provides an insight into the relationships between the different variables of interest. To
describe the variability we decompose the total sum of squares to the three orthogonal
components (within-country-time, between-country, and between-time). Thus, SS total =
SS(xit - x..) is decomposed to SSW(it) = SS(xit - xi. - x.t + x..), SSB(i) = SS(xi. - x..), SSB(t)
= SS(x.t - x..), where, for any variable z, we use the notation: SS(z) = z 2.
it
i t
To standardize the results, we divide the components by the total sum of squares so
that the numbers in Table 1 show the percentage of each component in the total sum of
squares. The between-country differences account for most of the variability in output and
more so in the inputs; about 89 percent of the output variability is due to the between-
country differences. Thus, a regression which allows for a country effect, without any
quantitative variables, would yield an R2 of 0.89, so that the unexplained residual from
country averages accounts for only 11 percent of the total sum of squares of output. If we
add the time effect, the R2 rises to 0.98. Similarly, the between-country variability accounts
22The terms "free, partially free and non-free" are used by Freedom House as classifications corresponding
to the indices of political rights and civil liberties.
23If we restrict the sample to countries which have at least one value not equal to 1, the range of averages
increases to 4.09 and 3.14 for civil liberties and 4.10 and 2.95 for political rights (in 1972 and 2000
respectively).
15
for 95 percent to almost 100 percent of the total variability in land, labor, livestock, and
fertilizers. The situation is similar when the output and inputs are measured per worker.
The relative importance of the country and time components is different for the
variability of the state variables. The between-country component is important in
schooling, development, political rights, and civil liberties and less important in the other
variables. In part, this difference among the state variables reflects the way the variables
are measured. Schooling, development, political rights, and civil liberties are measured in
units that allow cross-country comparisons, and interestingly, the relative importance of
the between-country component in the total sum of squares is similar to that of output. We
can relate this discussion to the determinants of the inputs as shown in equation (II.6). It
seems that schooling, development, and the institutional variables can be identified with s
in that they are associated with the technology level and also affect the input level. They
seem to have a strong correlation with the country effect. On the other hand, the price
variables are indices, and as such, do not allow cross-country comparisons. They have a
strong deviation component and perhaps are associated with the allocation error. A strong
between-time effect is represented by the peak variable.
To sum up, the relative importance of the between-country component is dominant.
This can lead to the erroneous conclusion that the within analysis has little to contribute.
As a matter of principle, this conclusion is not well-founded because the precision of the
estimated coefficients depends not only on the spread in the regressors but also on the
variance of the equation disturbance which usually contains a component that is time
invariant. Consequently, the variance of the within component is considerably smaller than
the total variance. This is validated below where we show that the within estimates are
meaningful empirically and informative substantively.
V. Empirical Results
Our ultimate interest is the estimation of the role of the inputs in production, or
simply the production elasticities. To do this we have to eliminate the jointness effect, or
the transmitted effect consisting of the state variables and of the country and time
idiosyncratic variables. To accomplish this we present here two models: the first is a pure
16
production function where inputs are the only regressors, and the second is an extended
function which contains also the state variables described in section III.
We organize the empirical results of each model in three blocks. The first block
presents the within-country-time estimates, b(it). The working hypothesis is that these
estimates are based on observations taken from the core technology. The second block
presents the between-time estimates, b(t), representing the time-series component, common
to all countries, and as such it captures the impact of changes over time in the available
technology. The last block presents the between-country estimates, b(i), summarizing the
between-country variability. The estimates are based on the locus of points that go across
the different techniques implemented by the countries which, in principle, operate under
the same available technology.
The general form of the estimated equation is:
yit = 0 +W (it)(xit + sit ) + B(i)(xitbc + sitbc) + B(t)(xitbT + sitbT ) +0it (V.1)
where 0 is the intercept, and 0 = m0 + m0 + u0 .
it i t it
Due to the orthogonal structure of the regressors, it is possible to estimate the three
blocks separately. We can do it by estimating (II.10) and (II.12), or by estimating (V.1).24
The difference will be in the dependent variable, and consequently in the value of R2 and
in the degrees of freedom used in the derivation of the t-score. We present the results from
both approaches for reasons to be discussed below.
Inputs only
The results are presented in Table 2. The dependent variable for columns termed
`block' is the transformed variable, Py, where P=W(it), B(t), and B(i) respectively,
associated with (II.10) and (II.12). The R2 and the t-score in the independent-block column
are obtained from this regression. The R2 appearing in the joint-block column is the
proportion of the total variability of y (not of Py) explained by this regression. The ratio of
the two values of R2 reflects the proportion of the block sum of squares in the total sum of
24We refer to estimations of equations (II.10) and (II.12) as `independent-block', while the estimation of
equation (V.1) is termed `joint-block'.
17
squares of output, referred to as the weight. Thus, the R2 of the joint-block equation (V.1),
0.9566, is equal to the weighted average of the independent-block values for R2. The t-
score in the joint-block regression is obtained from the estimation of equation (V.1). It is
clear that the contribution of the within variables to the explanation of total output (y) is
relatively small and that the t-score of the coefficients is lower than that obtained from the
independent-block estimation where the dependent variable is W(it)y. On the other hand,
the difference between the two versions is smaller for the between-country estimates. This
is a demonstration of the consequences of an implicit or explicit preference for using the
between estimates, which in the case of panel data, would be the between-country
estimates. The within estimates are avoided by working with country averages. However,
the within variables provide information for identifying the production function and are
less contaminated by variables leading to inconsistency in the estimates.
The key question of this analysis is whether the coefficients of the variables
common to the three canonical regressions are the same, aside from sampling error. A
casual inspection of the results indicates that they are quite different, confirming the basic
initial hypothesis that the regressions summarize the combined effect of changes in inputs
and technology, and therefore the within and between regressions summarize different
processes.
To introduce uniformity in the results of the various models we impose constant
returns to scale on the within estimates. This constraint is imposed only on the within
estimates, because the between estimates are subject to the jointness effect and therefore do
not present pure input elasticities. The sum of the within elasticities without this constraint
is 1.25, and the difference between the input elasticities in the constrained and
unconstrained models is absorbed mostly in the land elasticity.25 The Wald test of constant
returns to scale in the within regression is not rejected at the 4 percent level.
The sum elasticities of the two types of capital is 0.42, and the elasticity of land is
0.33. With sum elasticities of 0.75 for capital and land, there is little scope left for labor
25To save space we do not present the unconstrained results.
18
and fertilizer.26 This is the most important substantive result which indicates that
agriculture is capital-cost-intensive. The elasticity of fertilizers, 0.13, is considered to be
high for several reasons. First, we deal with the aggregate agricultural production function,
whereas fertilizer is used only on plant products. Thus the corresponding elasticity related
to the plant products should be higher than that obtained for the aggregate product. Second,
note that the dependent variable is value added, and in a competitive economy the
elasticity of fertilizer, whose cost is allowed for in the computation of value added, should
be nearly zero (an outcome of the envelope theorem). A higher value for the fertilizer
elasticity is likely to signal constraints on the supply of fertilizer causing the shadow price
to exceed the official price used in the national accounts to compute the cost of fertilizer.
The labor elasticity appears low, and we return to this subject later on. In Mundlak, Larson,
and Butzer (1999), the early literature on cross-country studies was reviewed, and it stands
out that our results differ from those reported in that literature. In part, it may be due to the
fact that we use a complete capital series, which was absent from the other studies, and in
part it is due to the fact that we use the within-time-country estimates, whereas the reported
results are mostly cross-country in nature and resemble the between-country regression in
the present study.
Turning to the between regressions, we note that the values of R2 are by far higher
than that of the within equation, and it is particularly high for the time-series component as
given in the between-time regression. The sum of the capital elasticities is 0.85 in the
between-time regression and 0.32 in the between-country regression. The between-time
elasticity is particularly high, and this suggests that the pace of the implementation of
changes in the available technology was strongly constrained by the level of the capital
stock in agriculture. Similarly, the land coefficient in the between-time regression is high,
but its t-score is low. A high value for the land coefficient suggests an increase in the
shadow price of land associated with the increase in productivity, while at the same time
there was little increase in the time series of agricultural area (Table 1).
26These values are of the same order of magnitude obtained in an earlier study (Mundlak, Larson, and Butzer
1999) for a different sample of countries and different time period.
19
What is striking in the between-country regression is the low elasticity of land,
0.02, and the high elasticity of fertilizer, 0.41. This suggests that the techniques used by the
more productive countries were land-saving and fertilizer-using. The subject is taken up in
section VI.
The reduced form
The country and time effects are estimated as residuals, and the question is to what
extent they can be replaced by the state variables. The potential list of pertinent state
variables is of unknown length, but we can only deal with observed variables. To get an
idea on the relevance of our set of state variables, we estimate the reduced form of output,
equation (I.1), which in view of equation (I.4) amounts to a quasi-supply function, in the
sense that it allows for changes in the supply function. The state variables are decomposed
to their orthogonal components, and their impact on the estimates of the various blocks is
determined accordingly. The within estimates are determined by the interaction term,
W(it)s, and the between regressions are determined by B(.)s. The time behavior of the state
variables is demonstrated in Figure 1 which presents plots of the annual averages of B(t)s
over time. It is seen that schooling and peak yield show a positive trend and are highly
correlated (0.995). On the other hand, the relative prices and the price variability show a
negative trend. Civil liberties and political rights are also subject to negative trend, which
means overall improvements over time of these attributes. Inflation is fairly stable except
for a big jump around 1990.
The OLS results organized by blocks appear in Table 3.27 The values of the R2 are
not high for the within and the between-country regression in contrast to the time series.
This means that our variables capture well the changes over time, and less so for the cross
section. The R2 for the model as a whole is 0.613. Comparing these values to those in
Table 1, where the country effect by itself accounts for 89 percent of the output variance,
indicates that our set of state variables is far from coming close to being a perfect substitute
27To save space, we present here t-scores only from the independent-block estimates.
20
for the country effects. The D.W. test statistic (2.171) is reported only for the between-time
regression, where it is relevant.28
It is obvious that the coefficients in the three blocks are different and therefore
represent different processes. Specifically, the relative price is positive and significant in
the within and the between-country regressions and negative in the between-time
regressions.29 The price variability has a negative coefficient in the within and in the
between-country regressions but not in the between-time regression. Schooling is positive
in all the three regressions, as is the development indicator. Civil liberties has the expected
negative sign in the within and between-time regression but not in the between-country
regression, while political rights has a negative coefficient in the between regressions, but
not in the within regression. Most of these results are carried over to the production
function with the state variables, referred to as the extended model.
To sum up, it is important to note that even though the within regression has a low
R2, it presents a supply function with expected signs. This is achieved by the interaction
terms W(it)s which are expected to cause input variations and to have a smaller impact on
the technology choice. This is not the case for the between regressions.
The extended production function
The OLS estimates of the extended model appear in Table 4. The R2 of this model
is 0.9694 as compared with 0.9566 in Table 2. This means that 29 percent of the
unexplained error of the model in Table 2 was reduced by the introduction of the state
variables. An F-test indicates that this addition is significantly different from zero.
Constant returns to scale is imposed on the within inputs elasticities. The sum elasticities
without this constraint is 1.22, and the difference between the constrained and
unconstrained elasticities is mostly in the elasticity of land.
An examination of the input elasticities shows that the big picture presented in
Table 2 has not changed in a dramatic way. Specifically, for the within estimates, the sum
28We have also computed a principal components version. The results are not different in a substantive way
from the OLS estimates and are therefore not reported here.
29A similar result is reported in Binswanger et al. (1987).
21
elasticities of capital is 0.37 and the elasticity of land is 0.45, and this again leaves little
scope for fertilizer and labor. The input coefficients in the between-regressions are also
quite similar to those observed in Table 2. What then is the contribution of the state
variables to the within regression? The answer is the rise of R2, or the reduction of the
equation variance. To see where it comes from we first review the impact of the state
variables in the within block. The main contribution comes from the price block and the
development indicator. The price coefficient is positive and that of the price variability is
negative. This is consistent with a positive response to price changes and a negative one to
risk. In interpreting the role of price, we note that w which appears in the theoretical model
discussed above is the vector of real factor prices, which is unobserved. The relative price
in the regression is the terms of trade of agriculture, which is the denominator of the
components of the vector w. As we deal with the log of the price variables, the
denominator of the vector w can be separated from the nominal factor prices, and it is
introduced here explicitly into the equation. Thus the positive sign of the relative price
coefficient in the within regression is interpreted as a positive supply response of the
implemented technology. The price elasticity of productivity is 0.29, and that of the price
variability is -0.31. These are quite sizable values. Using a somewhat different
formulation, to which we return below, Fulginiti and Perrin (1993) reports a price elasticity
of productivity of 0.13.30
The development indicator, which reflects the overall infrastructure of the
economy, as well as the institutional and technological environment, seems to be the most
robust variable. This indicates that the more productive is the economy as a whole the
higher is the productivity of agriculture. The civil liberties variable has the anticipated
(negative) sign, whereas schooling has a weak negative impact. Note that the variables in
the within regression are the interaction after the main effects were extracted, and therefore
the sign indicates the correction to the influence of the variables as given by the main
effects.
The productivity response observed in the between regressions is not always
consistent with that of the within regression. In the between-country regression, the sign of
30See also Hu and Antle (1993) and Binswanger et al. (1987) for price response in different formulations.
22
both the price and the price variability is the same as in the within regression. However, for
the between-time estimates, the price coefficients have the opposite signs, as the
productivity growth was associated with a decline in price and in a rise of the price
variability. This reflects a downward trend in the relative price associated with productivity
rise in world agriculture.
The price variability coefficient has a negative sign in the between-country
regression, indicating negative response to risk. Inflation has a negative coefficient in the
between-country regression.
The magnitude and the sign of the development indicator are robust across the three
equations. Schooling has a weak positive impact in the between regressions.31 The impact
of political rights and civil liberties is ambiguous and weak. The two physical environment
variables vary across countries but are time invariant. The sign of the water availability is
positive, as expected. The sign of potential dry matter is negative. This is inconsistent with
our earlier results and indicates that in this sample the high PDM countries were less
productive.
Stability of results
What happens when we remove the assumptions made in section II above?
Assumption 2 of the linearity of (s) is not crucial and can be ignored here. Assumption 1
(constant ) however, is more crucial. One way to find out the validity of this assumption
is to run the regression for subperiods. Table 5 presents results for two subperiods, 1972-
1985 and 1985-2000.32 A comparison of the two periods, and with the results in Table 4 for
the whole period, indicates some changes but the qualitative nature of most of the main
results is preserved. The strength of the capital elasticities is preserved even though there is
some change in the composition of the two components of capital. The fertilizer elasticity
is 0.14 and 0.13, and the interpretation of this value remains intact. The labor elasticity
increased in the second period. The rise in some of the elasticities reduces the elasticity of
31Because of the strong correlation between the time averages of schooling and peak yield, the latter variable
is not included in the between-time regression.
32We present here only the t-scores from the independent-block estimation. This is sufficient to establish the
point of the dependence of the estimates on the period chosen.
23
land, but still the sum elasticities of capital and land exceed 0.5. The rise in the elasticity of
labor in the second period may reflect the decline in the agricultural labor force.
Turning to the state variables, schooling has the wrong sign as before, but the peak
yield is positive and significant in the first period, which experienced a stronger rise in
yields. Finally, the role of price is positive and significant, whereas the price variability is
negative. Thinking of the within estimators as representing the core technology, we see
that the core technology is not detached from the economic environment and therefore is
not invariant to the sample.
The between-time regressions also preserve the important result of high elasticity
of capital. The sum elasticities of the two capital components are 0.94 for the first period,
0.70 for the second period as compared with 0.83 for the period as a whole. The between-
country estimates of the input elasticities show little change. The main changes are in the
coefficients of the state variables.
There are two possible approaches to incorporate the variability of (s,x) in the
analysis.33 The first requires knowledge of the factor shares and consists of estimating the
elasticities from a system of factor shares and the production function.34 This approach has
problems of its own which are related to the interpretation we can give to the factor shares.
The second approach is to write out (s,x) as a linear function of s and x which leads to a
quadratic production function in s and x.35 Such a function is blessed with many terms
which are intercorrelated and thus create a problem for the extraction of reliable results.
Note that the system reported in Table 4 already has a very high R2, and there is little scope
for squeezing in many additional terms. One possibility is to be selective with the number
of quadratic terms. For instance, Fulginiti and Perrin (1993) used the heterogeneous
technology framework to estimate such an equation including quadratic terms of the inputs
with some of the state variables (they refer to the state variables as technology-changing
variables) for a sample of 16 developing countries for the period 1961-1985. A key issue in
33See Mundlak (1988, 2001).
34See Mundlak, Cavallo, and Domenech (1989).
35The dependence on x alone yields the translog function (Christensen, Jorgenson, and Lau 1973). We
suppress here the dependence on x and concentrate the discussion on the dependence on s.
24
that study was to obtain a positive supply response to prices, and the outcome is an
elasticity of 0.13 for output price and -0.09 for wages.
VI. The Role of the Effects
Ordinarily, panel data analysis starts with the estimation of the coefficients of the
quantitative variables, and this is followed up with the introduction of discrete qualitative
variables, namely the effects. This natural course of action emphasizes the role of the
quantitative variables and diverts attention from the information embedded in the effects.
To clarify this point we can reverse the order and start the analysis by examining the role
of the effects. In our case this calls for the decomposition of the output sum of squares and
the computation of the R2 of an equation consisting solely of country and time dummies. In
our sample, such an equation explains about 98.5 percent of the total output sum of
squares, as shown in Table 1. All this, to be sure, is without an inclusion of any input in the
equation. But the inputs are there, because there is a strong correlation between the inputs
and the effects, and this goes back to the optimization described in equations (II.1) and
(II.2). As long as m0 affects the decision on input demand, the estimated effects reflect
input variations. For a similar reason they also reflect variations in the state variables.
The introduction of inputs to the empirical equation yields significant coefficients
but has little impact on the degree of explanation. The reason for this weak impact is that
the inputs are subject to strong country and time effects. These effects, however, do not
exhaust the input variability so that W(it)x does not vanish, and it is this remaining
variability that provides the information for the estimation of the coefficients.
To express the relationship between the effects and the regressors we rewrite
equation (V.1):
yit = 0 + xit + sit + B(i)[xitxc + sitsc]+ B(t)[xitxT + sitsT ]+0it (VI.1)
where xc = bc - , xT = bT - sc =bc - , sT = bT - .
The within estimator provides an estimate of and , whereas the between-
country and between-time estimators provide estimates of bc, bc, bT, and bT respectively.
Thus the 's are the bias of the between estimators.
25
The relationship between the effects and the regressors is summarized by the
following equations
m0 = B(i)[xit xc + sitsc] +0 (VI.2)
i i
m0 = B(t)[xitxT + sitsT ]+0 (VI.3)
t t
where and are the error terms.
0i 0t
An estimate of m0i and m0t is obtained from a regression of (II.8) with country and
time dummies. The values of R2 for the country regression (VI.2) are 0.554 with the state
variables alone, 0.830 with the inputs alone, and 0.885 for both groups. Similar regressions
for the time effect (VI.3) yield 0.982, 0.995, and 0.998 respectively. From this we learn
that the state variables account for most of the time effect and less so for the country effect.
The inputs account for a larger proportion of the country effect, but still less than of the
time effect. Technical change is the main event which evolved over time, and the set of the
state variables seems to be strongly correlated with it. The weaker relationship between the
state variables and the country effect indicates that there is a scope for introducing
additional state variables that are correlated with the country effect.
What are the implications of the estimates of equations (VI.2) and (VI.3)? The first
one is that it provides a set of variables that account for the effect. The set is not unique,
and we have already alluded to the long list of potential state variables used in the literature
to account for growth and productivity. Thus one would have to provide a rational for
preferring one set to an alternative one. The statistical analysis alone is insufficient to do
the task. The second implication is related to the estimation itself. Suppose that we have a
deterministic solution for the two equations, which means that we can replace the
unobserved effects with observed variables. How would it affect the estimation? The
answer is given by equation (VI.1). The estimate of would be the within estimator, since
the unobserved effects represented by moi and mot and reflected in the would vanish.
This means that explaining the effects can tell us something about how transformations of
the data affect bias related to unobservables, but it should not change our choice of
estimator.
26
VII. Evaluation
As there are big differences between the three estimators, it is desirable to get a
sense of reality and check how our estimates relate to the real world. We do it at a general
level, starting with the calculation of the TFP. Using the growth rates in Table 1 and the
within elasticities from Table 4 we obtain that TF increased at an average annual rate of
2.23 percent whereas TFP increased at an annual rate of 3.2 percent which accounts to 59
percent of output growth.
Using the elasticities, we compute the marginal value productivity, or shadow
price, as the product of the average value productivity and the corresponding elasticity.
Because the distribution is quite skewed, Table 6 presents the results of the median and of
the mean. We note that the shadow rate of return on fixed capital is quite high, and this is
consistent with the high growth rate of this input. Figure 2 presents the time path of the
median shadow prices from which we learn that the capital deepening resulted in
convergence to around 0.14. The shadow price of capital of agricultural origin is lower,
and this may be related to the way the variable was constructed. The shadow wage of labor
is relatively low which explains the migration of labor out of agriculture. The decline in
the labor force and the rise of capital caused the shadow wage to grow at the annual rate of
5.4 percent, considerably higher than that of TFP. There is a problem in comparing the
shadow wages to published wages. Published wages refer to payment for actual work,
whereas the labor data refer to the available labor force which is not fully occupied due to
the seasonality of farm work. This issue is discussed in some detail in a study on Asian
agriculture (Mundlak, Larson, and Butzer 2004). The shadow price of fertilizer increased
over time at a higher rate than that of TFP in spite of the fact that the fertilizer-land ratio
has increased constantly over time. Recall that the output is value added, and thus it is net
of fertilizer cost. It is likely that the price at the farm gate is higher than the price used in
national accounts, but still there may be a gap reflecting the rise in demand due to the shift
to fertilizer-intensive crops. Finally, the rent per hectare of land in 1990 dollar is 568 at the
mean and 271 at the median. To get from this to the value of land, we assume a
depreciation rate of 0.05 and subtract it from the shadow price of capital. We then
capitalize it by dividing the rent on land by the net rate of return to capital to obtain 2185
and 2464 1990 dollars per hectare at the mean and median respectively. There is of course
27
considerable variability in the sample as in reality. To sum up this evaluation, it seems that
our results have a realistic flavor which would not be the case if we repeated the
calculations with the between estimates.
VIII. Perspective
It is useful to relate the model briefly to the discussion in the literature on panel
data.
1) Identification: The identification of the production function depends largely on
the allocation error. The more the firms deviate from the first order conditions, the more
accurate the estimates will be.
2) Consistency: In the absence of state variables, or under a weaker assumption
where W(it)s = 0, the OLS estimates of the within equation are consistent and those of the
between equations are not. Some authors use first differences to eliminate the i-effect in
order to eliminate the bias in the b(i) estimates.36 As indicated above, this approach is
inefficient.
3) Sample size: Increasing the sample size does not eliminate the bias caused by the
jointness effect; it only reduces the sampling error. This is true regardless of whether the
sample is increased through N or T (the number of countries or years).
4) Input spread: The decomposition of the sum of squares of the inputs show that
SSB(i) is dominating, and that SSW(it) is relatively small. It is, therefore, claimed that the
within estimator does not utilize important information. This is true but not the whole
truth, because SSW(it) also constitute a small fraction of the total SS of output. Thus, there
is less information, but there is less to be explained by this information. We have
demonstrated that the within estimator provides meaningful and statistically significant
results.
5) Fixed or random effects: The foregoing discussion is invariant to the assumption
about the nature of the idiosyncratic variables, or effects. Under the random effect model
the GLS estimator is a matrix-weighted average of the within and between estimators
36For instance, see Lau and Yotopoulos (1989) and Griliches and Mairesse (1998).
28
(Maddala 1971), and it is therefore inconsistent. The source of the bias is the jointness
effect.
6) Measurement error: The within estimator is more sensitive to measurement
errors.37 This statement assumes implicitly that the measurement error is unaffected by the
transformation, so that its relative contribution to the within SS is larger than to the
between SS. This possibility is not ruled out, but it should be noted that there is good
reason to believe that part of the measurement error is country (or firm) specific, and by
the same token it is time specific, and is thus eliminated by the within transformation.38 It
is impossible to generalize on the relative importance of the measurement error in the
universe of all panels. What we learn from this study is that the most sensible results come
from the within transformation, and this transformation is consistent with the theory
formulated above.
7) Diversity of results: Concern has been expressed from the fact that there is a
great deal of diversity in the results obtained in production function estimates from panel
data depending on how the data are pooled (Griliches and Mairesse 1998; Mairesse 1990).
The diversity is a problem when the working hypothesis is that the estimates should be
invariant to way the data are pooled. The general model presented here indicates that one
should expect diversity, and in fact the diversity serves as a starting point for the
construction of more meaningful models.
8) Instrumental variables: The use of instrumental variables was suggested as a way
to overcome the bias in the estimates of panel data (Hausman and Taylor 1981). In the
present framework the scope for the use of instrumental variables is rather limited because
variables which are associated with the choice of inputs are assumed also to affect the
choice of the function itself. In other words, the instrumental variables fall in the category
of state variables in the present framework. The same argument also rules out the GMM
estimator.
37 See Griliches and Mairesse (1998).
38 For a fuller discussion, see Mundlak (2001).
29
9) Input ratios: In the Cobb-Douglas a difference between the log of two inputs,39
say j and g, xjit xgit, eliminates the terms m0 and, in the absence of state variables, can
serve as an instrumental variable. In this approach, when the production function is
estimated in terms of average productivity, and constant returns to scale is imposed, the
estimation is free of the jointness bias. Unfortunately, under heterogeneous technology,
this is no longer the case.
IX. Summary and Conclusions
The paper presents an estimate of the agricultural production function from a panel
of countries.
1) Framework: In the world of heterogeneous technology, the implemented
techniques and inputs are jointly determined conditional on the state variables that are
assumed to specify the economic environment. Because of variability in the state variables,
the production function of a sector is an aggregate of micro production functions. It is
approximated by a Cobb-Douglas function with parameters that depend upon the state
variables.
2) State variables enter as exogenous variables in the empirical equation, and the
estimation is straightforward when they are observed. In contrast, unobserved state
variables become part of the production function shock, thus creating a correlation between
the inputs and the productivity shocks, similar in nature to the transmission of the
idiosyncratic variables in panel data. Due to the structure of the problem, the only way to
identify the production function is through allocation errors, namely, through input
variations that are unaffected by the omitted state variables or the idiosyncratic
productivity shock.
3) The sum of squares of the panel data is decomposed into the three orthogonal
components. Most of the variability in output and inputs comes from between-country
variations, whereas the within-country-time variations account only for a small proportion
of the total sum of squares. Estimates obtained from between-country variations are
39See Mundlak (1996).
30
popular because they are based on a wide spread in the regressors. They are, however,
biased. On the other hand, the within-country-time variations of the inputs reflect largely
allocation errors and thus produce consistent or low-bias estimates.
4) When not all state variables are observed, the choice of regression matters. We
provide a practical example and present estimates obtained under the assumption of
constant slopes, for the canonical set of regressions, between-country, between-time (time-
series component), and within-country-time variations. There are great differences in the
estimates of the three canonical regressions. The elasticity of capital from the within
regression is 0.37, as compared to 0.27 from the between-country regression, and 0.83
from the between-time regression. The latter suggests that capital was a constraint in the
implementation of new capital-intensive techniques, in spite of the fact that capital grew
faster than all other inputs and output. These numbers are indicative of the differences in
the results obtained from the three regressions, and similar differences exist for the other
variables. The elasticity of fertilizer from the within regression is 0.1, although it should be
close to zero because output is measured by GDP, which is net of fertilizer costs. This
indicates that the shadow price of fertilizer was higher than the market price. However, the
value obtained for the fertilizer coefficient from the between-country regression is 0.44. If
this were a true elasticity, it would mean that 44 percent of GDP should be attributed to
fertilizer; clearly, this is absurd. In contrast, it is likely that land is a dominant factor of
agricultural production. In this case, the elasticity of land in the between-country
regression is 0.03 as compared to 0.45 in the within regression. These comparisons provide
substantive evidence on the superiority of the within estimator. This is true even in
comparison to linear estimates obtained from pooled data, since these are weighted matrix-
averages of the three canonical regressions and reflect the bias of their components.
5) Agriculture: The new techniques were capital and fertilizer intensive. This is
reflected in the growth rates of these inputs. On the other hand, the techniques were labor
saving; this is consistent with the decline in the size of the labor force in agriculture. The
land elasticity is high for the within and the time component, and low for the between
countries. Thus, the more productive countries use land-saving and fertilizer-using
techniques. The land elasticity reflects the terms of trade of agriculture for the period,
which, on the whole, enjoyed important improvements in productivity. The decomposition
31
of output growth shows that TFP accounted for 59 percent of the output growth of 5.43
percent.
6) State variables: The relevance of the state variables was tested by estimating the
reduced form, or a quasi-supply function. They account for most of the variability of the
between-time output, and only slightly for the within output. Still the within regression
provides a supply function with the right signs and significant coefficients. Turning to the
production function, the relative price of agriculture has a positive impact, and its
variability has a negative impact in the within regression. The development indicator
indicates that the agricultural productivity is positively correlated with the strength of the
economy as a whole. The indicator is assumed to represent total capital, physical and
human, and the institutional infrastructure. Some of the variables which are confounded in
the indicator were introduced explicitly into the regression, but their contribution was
marginal.
7) Accounting for the effects: The country and time effects account for most of the
variability in the data. It is shown that the effects are embedded in the country and time
means of the inputs and the state variables. The state variables are particularly important in
capturing the time effect, and less so for the country effect. This suggests that there is a
scope for trying out additional state variables to account for the cross-country variations.
8) Stability of results: The estimates are sensitive to the economic environment.
This is demonstrated by estimating the regressions for two sub-periods. Even though the
estimates change, the main message is preserved.
10) The results are consistent with the changes that take place in the process of
growth. Agricultural productivity rises, there is a shift to capital and fertilizer-intensive and
labor-extensive techniques. The response in resource allocation leads to growth in the
marginal productivities of labor, fertilizer, and land, and a slight decline in the marginal
productivity of capital as a result of the fast growth of the capital-labor ratio. From the
point of view of growth, policies should encourage the process to continue.
32
Appendix
Equivalence of OLS and GLS estimators:
Following Zyskind (1967), Rao (1973), Baltagi (2006), we state the following theorem:
Theorem: Let y=X + u, where u ~ (0, V).
OLS and GLS of are equal if and only if there exists a matrix B such that VX = XB.
We apply the theorem as follows:
Consider the transformation Py = (PX) + Pu, where Pu ~(0, 2P)
Check the condition of the theorem: VPX = 2PX = PX2
Thus the matrix 2I is the matrix B of the theorem.
This theorem explains why we continue to apply OLS to equations which are premultiplied
by projection matrices.
33
Bibliography
Acemoglu, Daron, Simon Johnson, and James Robinson, "Institutions as the Fundamental
Cause of Long-Run Growth" (pp. 385-472), in Philippe Aghion and Steven N.
Durlauf (Eds.), Handbook of Economic Growth Volume 1A (Amsterdam and San
Diego: Elsevier, North-Holland, 2005).
Antle, John M., "Infrastructure and Aggregate Agricultural Productivity: International
Evidence," Economic Development and Cultural Change 31:3 (1983), 609-619.
Baltagi, Badi H., "An Alternative Derivation of Mundlak's Fixed Effects Results Using
System Estimation," Econometric Theory 22:6 (2006), 1191-1194.
Barro, Robert J., and Jong-Wha Lee, "International Data on Educational Attainment:
Updates and Implications," CID Working Paper No. 42 (2000).
Barro, Robert J., and Jong-Wha Lee, "Sources of Economic Growth," Carnegie-Rochester
Conference Series on Public Policy 40 (1994), 1-46.
Barro, Robert J., and Rachel M. McCleary, "Religion and Economic Growth Across
Countries," American Sociological Review 68:5 (2003), 760-781.
Binswanger, Hans P., Yair Mundlak, Maw-Cheng Yang, and Alan Bowers, "On the
Determinants of Cross-Country Aggregate Agricultural Supply," Journal of
Econometrics 36:1/2 (1987), 111-131.
Buringh, P., H. D. J. van Heemst, and G. J. Staring, "Computation of the Absolute
Maximum Food Production of the World," Wageningen: Netherlands Department
of Tropical Soil Science, 1975; also published in H. Linneman et al. (Eds.) Moira
Model of International Relations in Agriculture (Amsterdam: North-Holland
Publishing Co., 1979).
Chamberlain, Gary, "Multivariate Regression Models for Panel Data," Journal of
Econometrics 18:1 (1982), 5-46.
Christensen, Laurits R., Dale W. Jorgenson, and Lawrence J. Lau, "Transcendental
Logarithmic Production Frontiers," Review of Economics and Statistics 55:1
(1973), 28-45.
Cobb, Charles W., and Paul H. Douglas, "A Theory of Production," American Economic
Review 18:1 supplement (1928), 139-165.
Craig, Barbara J., Philip G. Pardey, and Johannes Roseboom, "International Productivity
Patterns: Accounting for Input Quality, Infrastructure, and Research," American
Journal of Agricultural Economics 79:4 (1997), 1064-1076.
Dollar, David, "Outward-Oriented Developing Economies Really Do Grow More Rapidly:
Evidence from 95 LDCs, 1976-1985," Economic Development and Cultural
Change 40:3 (1992), 523-544.
Durlauf, Steven N., and Danny T. Quah. "The New Empirics of Economic Growth" (pp.
235-308) in John B. Taylor and Michael Woodford (Eds.), Handbook of
Macroeconomics Volume 1A (Amsterdam; New York and Oxford: Elsevier
Science, North-Holland, 1999).
34
Durlauf, Steven N., Paul Johnson, and Jonathan R. W. Temple, "Growth econometrics"
(pp. 555-677), in Philippe Aghion and Steven N. Durlauf (Eds.), Handbook of
Economic Growth Volume 1A (Amsterdam and San Diego: Elsevier, North-
Holland, 2005).
Easterly, William, and Ross Levine, "It's Not Factor Accumulation: Stylized Facts and
Growth Models," World Bank Economic Review 15:2 (2001), 177-219.
Evenson, Robert E., and Yoav Kislev, "Investment in Agricultural Research and
Extension: A Survey of International Data," Economic Development and Cultural
Change 23:3 (1975), 507-521.
Frankel, Jeffrey A., and David Romer, "Does Trade Cause Growth?" American Economic
Review 89:3 (1999), 379-399.
Fulginiti, Lilyan E., and Richard K. Perrin, "Prices and Productivity in Agriculture," The
Review of Economics and Statistics 75:3 (1993), 471-482.
Griliches, Zvi, and Jacques Mairesse, "Production Functions: the Search for Identification"
(pp. 169-203), in Steinar Ström (Ed.), Econometrics and economic theory in the
20th century: the Ragnar Frisch Centennial Symposium (Cambridge University
Press, New York, 1998).
Hall, Robert E., and Charles I. Jones, "Why Do Some Countries Produce So Much More
Output Per Worker Than Others?" Quarterly Journal of Economics114:1 (1999),
83-116.
Hausman, Jerry A., and William E. Taylor, "Panel Data and Unobservable Individual
Effects," Econometrica 49:6 (1981), 1377-1398.
Henderson, Daniel J., and R. Robert Russell, "Human Capital and Convergence: A
Production-Frontier Approach," International Economic Review 46:4 (2005), 1167-
1205.
Hu, Fan, and John M. Antle, "Agricultural Policy and Productivity: International
Evidence," Review of Agricultural Economics 15:3 (1993), 495-505.
Klenow, Peter J., and Andres Rodriguez-Clare, "The Neoclassical Revival in Growth
Economics: Has it Gone Too Far?" (pp. 73-103), in Ben S. Bernanke and Julio
Rotemberg (Eds.), NBER Macroeconomics Annual 1997 (Cambridge, MA: MIT
Press, 1997).
Larson, Donald F., Rita Butzer, Yair Mundlak, and Al Crego, "A Cross-Country Database
for Sector Investment and Capital," World Bank Economic Review 14:2 (2000),
371-391.
Lau, Lawrence, and Pan A. Yotopoulos. "The Meta-Production Function Approach to
Technological Change in World Agriculture," Journal of Development Economics
31:2 (1989), 241-269.
Maddala, G. S., "The Use of Variance Components Models in Pooling Cross Section and
Time Series Data," Econometrica 39:2 (1971), 341-358.
35
Mairesse, Jacques, "Time -Series and Cross-Sectional Estimates on Panel Data: Why are
They Different and Why Should They Be Equal?" (pp. 81-95), in Joop Hartog,
Geert Ridder, and Jules Theeuwes (Eds.), Panel data and labor market studies
(Amsterdam: North-Holland, 1990).
Mankiw, N. Gregory, David Romer, and David N. Weil, "A Contribution to the Empirics
of Economic Growth," Quarterly Journal of Economics 107:2 (1992), 407-437.
Marschak, Jacob, and William H. Andrews, Jr., "Random Simultaneous Equations and the
Theory of Production," Econometrica 12:3/4 (1944), 143-205.
Mundlak, Yair. "Endogenous Technology and the Measurement of Productivity" (pp. 316-
331), in Susan M. Capalbo and John M. Antle (Eds.), Agricultural Productivity:
Measurement and Explanation (Washington, D.C.: Resources for the Future,
1988).
Mundlak, Yair, "On the Empirical Aspects of Economic Growth Theory," American
Economic Review 83:2 (1993), 415-420.
Mundlak, Yair, "Production Function Estimation: Reviving the Primal," Econometrica
64:2 (1996), 431-438.
Mundlak, Yair, "Production and Supply" (pp. 3-85), in Bruce L. Gardner and Gordon C.
Rausser (Eds.), Handbook of Agricultural Economics Volume 1A, Agricultural
Production (Amsterdam; London and New York: Elsevier Science, North-Holland,
2001).
Mundlak, Yair, Domingo Cavallo, and Roberto Domenech, "Agriculture and Economic
Growth, Argentina 1913-84," Research Report 76 (Washington, D.C.: International
Food Policy Research Institute, 1989).
Mundlak, Yair, and Rene Hellinghausen, "The Intercountry Agricultural Production
Function: Another View," American Journal of Agricultural Economics 64:4
(1982), 664-672.
Mundlak, Yair, and Irving Hoch, "Consequences of Alternative Specifications in
Estimation of Cobb-Douglas Production Functions," Econometrica 33:4 (1965),
814-828.
Mundlak, Yair, Donald F. Larson, and Rita Butzer, "The Determinants of Agricultural
Production: A Cross-Country Analysis," Policy Research Working Paper 1827
(Washington, D.C.: The World Bank, 1997).
Mundlak, Yair, Donald F. Larson, and Rita Butzer, "Rethinking within and between
Regressions: The Case of Agricultural Production Functions," Annales d'Economie
et de Statistique 55-56 (1999), 475-501.
Mundlak, Yair, Donald F. Larson, and Rita Butzer, "Agricultural Dynamics in Thailand,
Indonesia and the Philippines" Australian Journal of Agricultural and Resource
Economics 48:1 (2004), 95-126.
North, Douglass C., Institutions, Institutional Change and Economic Performance
(Cambridge; New York and Melbourne: Cambridge University Press, 1990).
36
Presbitero, Andrea F., "Institutions and Geography as Sources of Economic
Development," Journal of International Development 18:3 (2006), 351-378.
Pritchett, Lant, "Understanding Patterns of Economic Growth: Searching for Hills among
Plateaus, Mountains, and Plains," World Bank Economic Review 14:2 (2000), 221-
250.
Rao, C. R., Linear Statistical Inference and its Applications (New York: John Wiley &
Sons, 1973).
Rodrik, Dani, Arvind Subramanian, and Francesco Trebbi, "Institutions Rule: The Primacy
of Institutions over Geography and Integration in Economic Development,"
Journal of Economic Growth 9:2 (2004), 131-165.
Sachs, Jeffrey D., and Andrew M. Warner, "Fundamental Sources of Long-Run Growth,"
American Economic Review 87:2 (1997), 184-188.
Sala-i-Martin, Xavier, Gernot Doppelhofer, and Ronald I. Miller, "Determinants of Long-
Term Growth: A Bayesian Averaging of Classical Estimates (BACE) Approach,"
American Economic Review 94:4 (2004), 813-835.
World Bank, World Development Report 2008: Agriculture for Development
(Washington, D.C.: The World Bank, 2007).
Zyskind, George, "On Canonical Forms, Non-Negative Covariance Matrices and Best and
Simple Least Squares Linear Estimators in Linear Models," The Annals of
Mathematical Statistics 38:4 (1967), 1092-1109.
37
TABLE 1: GROWTH RATES AND THE DECOMPOSITION OF THE SUM OF SQUARES
Average Annual Decomposition of the Sum of Squares
Growth Rate (expressed as a percentage of total)
Variable (%) SSB(t) SSB(i) SSW(it)
Output
GDP 5.43 9.00 89.47 1.54
Inputs
Capital 5.77 8.84 88.62 2.54
Fixed capital 5.80 6.60 91.03 2.37
Capital of agricultural origin 4.94 5.50 91.94 2.55
Livestock 3.59 3.09 95.72 1.18
Orchards 5.77 4.13 93.35 2.53
Agricultural area 0.01 0.00 99.93 0.07
Labor -0.60 0.07 99.03 0.90
Fertilizer 1.87 1.01 96.47 2.53
Technology
Schooling 1.67 7.32 88.05 4.63
Peak yield 1.41 79.37 6.30 14.33
Development indicator 1.10 95.46 3.44
Institutions
Political rights 1.57 79.28 19.16
Civil liberties 1.17 82.67 16.16
Prices
Relative prices -1.26 25.01 23.14 51.85
Price variability 11.06 27.34 61.60
Inflation 3.12 8.50 88.37
Per Labor Output and Inputs
GDP 7.93 90.12 1.95
Capital 6.71 90.41 2.88
Fixed capital 4.59 92.87 2.54
Capital of agricultural origin 6.74 89.80 3.46
Livestock 3.58 94.17 2.25
Orchards 7.53 88.35 4.11
Agricultural area 0.09 99.13 0.78
Fertilizer 1.11 97.72 1.18
38
TABLE 2: PRODUCTION FUNCTION, INPUTS ONLY
Within time-country Between time Between country
Variable Estimate t-score Estimate t-score Estimate t-score
Block Joint Block Joint Block Joint
Inputs
Fixed capital 0.34 18.83 8.75 0.70 77.47 5.95 0.27 19.16 16.79
Capital of agricultural origin 0.08 4.11 1.91 0.15 13.62 1.05 0.05 5.81 5.09
Agricultural area 0.33 5.01 17.12 1.32 0.02 2.52 2.21
Fertilizer 0.13 5.51 2.56 -0.07 -5.37 -0.41 0.41 24.37 21.35
Labor 0.12 3.17 1.47 -1.10 -21.67 -1.67 0.17 24.57 21.53
Sum of estimates 1.00 0.93
Summary statistics Joint
regression
Panel R-squares R-square Weight R-square Weight R-square Weight R-square
0.957 0.015 0.384 0.090 0.997 0.895 0.962
Durbin-Watson statistic 1.088
39
TABLE 3: REDUCED FORM PRODUCTION FUNCTION, STATE VARIABLES ONLY
Independent block regressions
Variable Within time-country Between time Between country
Estimate t-score Estimate t-score Estimate t-score
Technology
Schooling 0.26 4.89 1.20 12.86 0.54 3.50
Peak yield 0.10 0.82 11.27 8.25
Institutions
Civil liberties -0.03 -2.85 -0.48 -14.61 2.38 20.68
Political rights 0.02 2.00 -0.04 -1.00 -1.42 -17.56
Development indicator 0.96 12.60 1.79 20.38 2.50 8.73
Price block
Relative prices 0.34 9.77 -1.72 -18.86 12.46 25.01
Price variability -0.39 -3.40 2.09 10.49 -26.78 -16.77
Inflation 0.00 1.44 -0.05 -6.76 -0.30 -5.14
Environmental
Potential dry matter -0.76 -3.91
Factor of water availability -0.54 -3.50
Joint
regression
Summary statistics R-square Weight R-square Weight R-square Weight R-square
Panel R-squares 0.613 0.015 0.280 0.090 0.977 0.895 0.582
Durbin-Watson statistic 2.171 .
40
TABLE 4: EXTENDED PRODUCTION FUNCTION
Independent block regressions
Within time-country Between time Between country
Variable Estimate t-score Estimate t-score Estimate t-score
Block Joint Block Joint Block Joint
Inputs
Fixed capital 0.31 17.47 8.70 0.59 49.46 2.81 0.13 7.29 6.19
Capital of agricultural origin 0.06 2.89 1.44 0.24 22.18 1.26 0.14 12.08 10.25
Agricultural area 0.45 0.34 1.01 0.06 0.03 3.16 2.68
Fertilizer 0.10 4.39 1.16 -0.24 -9.79 -0.56 0.44 21.78 18.48
Labor 0.09 2.34 2.19 -0.78 -9.01 -0.51 0.19 11.88 10.08
Sum of estimates 1.00 0.93
Technology
Schooling -0.13 -2.51 -1.25 0.14 2.56 0.15 0.14 3.31 2.81
Peak yield -0.02 -0.19 -0.09 0.23 0.62 0.52
Institutions
Civil liberties -0.02 -1.75 -0.87 0.000 -0.002 0.000 0.03 0.81 0.69
Political rights 0.01 0.70 0.35 -0.16 -15.37 -0.87 0.03 1.19 1.01
Development indicator 0.63 9.04 4.50 0.58 19.32 1.10 -0.19 -2.38 -2.02
Price block
Relative prices 0.29 9.13 4.54 -0.07 -2.42 -0.14 0.97 5.97 5.07
Price variability -0.31 -3.13 -1.56 0.24 4.11 0.23 -1.84 -3.35 -2.84
Inflation 0.002 0.76 0.38 0.004 1.62 0.09 -0.07 -3.99 -3.39
Environmental
Potential dry matter -0.66 -12.31 -10.45
Factor of water availability 0.22 4.96 4.21
Joint regression
Summary statistics R-square Weight R-square Weight R-square Weight R-square
Panel R-squares 0.969 0.015 0.490 0.090 0.999 0.895 0.975
Durbin-Watson statistic 1.895
41
TABLE 5: EXTENDED PRODUCTION FUNCTION BY SUB-PERIODS
Within time-country Between time Between country
1972-1985 1985-2000 1972-1985 1985-2000 1972-1985 1985-2000
Independent block regression Estimate t-score Estimate t-score Estimate t-score Estimate t-score Estimate t-score Estimate t-score
Inputs:
Fixed capital 0.38 10.29 0.43 14.38 0.80 50.84 0.76 42.72 0.11 4.78 0.18 6.64
Capital of agricultural origin 0.20 5.54 0.07 2.83 0.14 9.85 -0.06 -4.54 0.14 8.39 0.16 10.67
Agricultural area 0.15 0.06 7.55 12.76 -6.55 -28.98 0.01 0.07 0.01 0.55
Fertilizer 0.14 4.38 0.13 3.83 0.40 9.08 0.10 7.18 0.42 17.01 0.42 14.61
Labor 0.13 1.41 0.31 4.92 0.54 2.08 -1.12 -5.71 0.29 13.21 0.21 10.17
Sum of estimates 1.00 1.00 0.98 0.97
Technology:
Schooling -0.04 -0.48 -0.22 -2.19 -0.29 -2.50 0.26 2.30 -0.002 -0.04 -0.05 -0.85
Peak yield 0.67 4.61 -0.54 -2.83 -0.64 -2.53 -1.80 -2.00
Institutions:
Civil liberties -0.04 -3.14 -0.01 -0.74 0.04 1.74 -0.08 -10.64 0.07 1.93 -0.32 -6.76
Political rights 0.01 1.02 -0.01 -0.96 -0.02 -1.06 0.16 13.61 0.01 0.26 0.18 5.28
Development indicator 0.26 2.72 0.48 5.22 -0.50 -6.30 0.74 19.85 0.34 2.81 -0.50 -4.79
Price block:
Relative prices 0.23 5.82 0.39 7.24 0.64 14.09 -0.39 -10.61 0.08 0.63 -0.36 -1.62
Price variability -0.26 -2.62 -0.53 -3.25 0.31 9.35 -0.33 -3.40 1.77 3.70 -0.84 -1.05
Inflation 0.04 0.70 -0.001 -0.67 -0.40 -13.78 0.003 2.00 0.08 0.67 0.003 0.27
Environmental:
Potential dry matter -0.60 -7.69 -0.61 -8.98
Factor of water availability 0.18 2.57 0.21 3.22
Summary Statistics Weight R-square Weight R-square Weight R-square Weight R-square Weight R-square Weight R-square
R-squares 0.012 0.546 0.013 0.569 0.031 1.000 0.020 0.999 0.957 0.977 0.967 0.973
Durbin-Watson statistic 3.443 2.499
Joint regressions 1972-85 1985-2000
R-square 0.973 0.968
Number of observations 420 480
42
TABLE 6: PRODUCTIVITY EVALUATION
Growth decomposition Marginal products
Growth rate Elasticity Contribution Mean Median Growth rate of
median value
Factors
Fixed capital 5.80 0.31 1.80 0.31 0.16 -0.37
Capital agricultural origin 4.94 0.06 0.30 0.06 0.04 0.49
Area 0.01 0.45 0.00 568 271 5.43
Fertilizer 1.87 0.10 0.19 1,468 1,097 5.24
Labor -0.60 0.09 -0.05 911 307 6.03
Total Factors 2.23
Productivity 3.20
Output 5.43
43
FIGURE 1: STATE VARIABLES
SCHOOLING PEAK YIELD CIVIL LIBERTIES
.3 .20 .4
.15 .3
.2
.10 .2
.1
.05 .1
.0 .00 .0
-.05 -.1
-.1
-.10 -.2
-.2
-.15 -.3
-.3 -.20 -.4
72 74 76 78 80 82 84 86 88 90 92 94 96 98 00 72 74 76 78 80 82 84 86 88 90 92 94 96 98 00 72 74 76 78 80 82 84 86 88 90 92 94 96 98 00
POLITICAL RIGHTS DEVELOPMENT INDICATOR RELATIVE PRICE
.5 .08 .3
.4
.2
.3 .04
.2 .1
.00
.1
.0
.0
-.04
-.1 -.1
-.2 -.08 -.2
-.3
-.4 -.12 -.3
72 74 76 78 80 82 84 86 88 90 92 94 96 98 00 72 74 76 78 80 82 84 86 88 90 92 94 96 98 00 72 74 76 78 80 82 84 86 88 90 92 94 96 98 00
PRICE VARIABILITY INFLATION
.08 2.0
.06 1.6
.04 1.2
.02 0.8
.00 0.4
-.02 0.0
-.04 -0.4
72 74 76 78 80 82 84 86 88 90 92 94 96 98 00 72 74 76 78 80 82 84 86 88 90 92 94 96 98 00
44
FIGURE 2: MEDIAN MARGINAL PRODUCTS
FIXED CAPITAL CAPITAL OF AGRICULTURAL ORIGIN
.20 .048
.19
.18 .044
.17
.040
.16
.15
.036
.14
.13 .032
.12
.11 .028
72 74 76 78 80 82 84 86 88 90 92 94 96 98 00 72 74 76 78 80 82 84 86 88 90 92 94 96 98 00
AGRICULTURAL AREA FERTILIZER
700 2000
1800
600
1600
500
1400
400 1200
1000
300
800
200
600
100 400
72 74 76 78 80 82 84 86 88 90 92 94 96 98 00 72 74 76 78 80 82 84 86 88 90 92 94 96 98 00
LABOR
1200
1000
800
600
400
200
0
72 74 76 78 80 82 84 86 88 90 92 94 96 98 00
45