Professional Documents
Culture Documents
Econometrics of Planning and Efficiency
Econometrics of Planning and Efficiency
Managing Editors:
J.P. Ancot, Netherlands Economic Institute, Rotterdam, The Netherlands
A.J. Hughes Hallett, University of Newcastle, U.K.
Editorial Board:
F.G. Adams, University of Pennsylvania, Philadelphia, U.S.A.
P. Balestra, University of Geneva, Switzerland
M.G. Dagenais, University of Montreal, Canada
D. Kendrick, University of Texas, Austin, U.S.A.
J.H.P. Paelinck, Netherlands Economic Institute, Rotterdam, The Netherlands
R.S. Pindyck, Sloane School of Management, M.I.T., U.S.A.
H. Theil, University of Florida, Gainsville, U.S.A.
W. Welfe, University of Lodz, Poland
For a complete list of volumes in this series see final page of this volume.
Econometrics
of Planning
and Efficiency
edited by
Jati K. Sengupta
Professor of Economics and Operations Research,
University of California, Santa Barbara
Gopal K. Kadekodi
Professor of Economics, Institute of Economic Growth, Delhi, India
for the United States and Canada: Kluwer Academic Publishers, P.O. Box 358,
Accord Station, Hingham, MA 02018-0358, USA
for the UK and Ireland: Kluwer Academic Publishers, MTP Press Limited,
Falcon House, Queen Square, Lancaster LA 1 1RN, UK
for all other countries: Kluwer Academic Publishers Group, Distribution Center,
P.O. Box 322, 3300 AH Dordrecht, The Netherlands
Copyright
List of Contributors vu
PART I: METHODOLOGY 1
G. K. KADEKODI
Institute ofEconomic Growth, Delhi, India
T. K. KUMAR
University ofHyderabad, Hyderabad, India
and
1. K. SENGUPTA
University ofCalifornia at Santa Barbara, California, u.s.A.
1. INTRODUCTION
2. GENERAL METHODOLOGY
3. ECONOMIC THEORY
5. OTHER CONTRIBUTIONS
Books
1. Prices in the Trade Cycle, Vienna, Springer, 1935, (XII and 204 pp., 60
graphs, mathematical appendix, French and German resume).
2. The Variate Difference Method, published jointly by the Cowles Commission
for Research in Economics and the Department of Economics and Sociology,
Iowa State College in 1940, Principia Press, Bloomington, Indiana (175 pp.,
three figures, 48 tables, appendix, etc.).
3. Econometrics, published by John Wiley & Sons, New York City, and
Chapman & Hall Ltd., London 1952. (Second printing 1955, XIII and 370
pp.), pocket book ed. 1965. Japanese translation: Bun Ga Do Publishers,
Tokyo,Japan 1961.
4. Mathematics and Statistics for Economists, published by Holt, Rinehart &
Winston Inc., New York City 1953. Second Printing 1954. English edition:
Constable & Co. Ltd. London. Japanese translation: Chikurashobe Publishing
Co., Tokyo, Japan (XIV and 363 pages). French Translation: Mathematiques
et Statistiques pour les Economistes (XV and 488 pp.), Paris: Dunod 1965,
nouveau tirage 1969, 508 pages. Korean translation: Il-Jo-Ka publishers,
Introduction to Economic Mathematics, 315 pp., translation by Moon Sik
Kim. Second Edition (with C. B. Millham), XX and 485 pp., 1970. Spanish
translation: Matematica y estadistica para economistas, Interamaericana, Mexico
1973.
5. Handbuch der Okonometrie (Handbook of Econometrics), has been included
in the series Encyclopaedie der Rechts- und Staatswissenschaften (Springer
Verlag, Berlin, 1960, XI and 328 pp.). Russian translation: Vvedenie v
ekonometriyu Izdatelstove Statistika, 360 pp., Moscow 1965.
6. Introduction a La Econometria, 128 pp., has been published by the Central
University of Quito and the Junta Nacional de Planfication y Coordinacion
Economics, Quito, Ecuador. Portugese translation: Elementos de Econometrica,
117 pp., 1965, published by Enio Matheus Guazelle & Cia., Sao Paulo, Brazil.
7. Topics in Econometrics, A course of Lectures delivered at the Institute of
Agricultural Research Statistics, June - August 1965, Indian Council of
Agricultural Research, New Delhi, India 1966, 125 pp.
11
Articles
1952c 'Die Anwendung der variate difference Methode auf die Probleme der
gewogenen regression und der MultikoUinearitat', Mittellungsblatt /iir
Mathematische Statistik 4, pp. 159 ff.
1952d 'Complementarity and shifts in demand', Metroeconomica 4, p. 1.
1952e 'Econometrica', A Revista de Perquisas Economica-sociales 2, pp. 361 ff.
1953a 'Econometrics', Scientia 88, pp. 250 ff.
1953b 'Econometrie', Scientia 88, Supplement, pp. 148 ff.
1953c 'Econometria', El Trimestre Economico 20, pp. 75 ff.
1953d 'The Definition of Econometrics', Econometrica 21, pp. 31 ff.
1953e (With 0. H. Brownlee), 'The Production Functions Derived from Farm
Records - A Correction', Journal ofFarm Economics 35, pp. 123 ff.
1954a 'The Teaching of Econometrics', Econometrica 22, pp. 77.
1954b 'Ein Beitrag zur Nicht - Statischen Wertheorie', Zeitschrift /iir Nationalo-
konomie 14, pp. 358 ff.
1954c 'The teaching of econometrics, UNESCO - The University of Social
Sciences, Economics, Paris, pp. 282 ff.
1954d 'The definition of econometrics', (Japanese translation), Bulletin of the
Bureau of Statistics, Office of the Prime Minister Tokyo, Japan, Vol. 3, No.
6, pp. 57 ff.
1954e 'The Use of Mathematical Methods of Econometrics and Economic
Statistics', International Social Science Bulletin, UNESCO, Vol. 6, No.4,
pp.640ff.
1955a 'Stochastic Linear Programming with Applications to Agricultural Econom-
ics', Symposium on Linear Programming, National Bureau of Standards,
Washington, D.C., Vol. 1, pp. 197 ff.
1955b 'The distribution of the variance of variate differences in the circular case',
Metron 17, p. I ff.
1955c 'Programmazione lineare stocastica con applicazioni a problemi di economia
agraria', Giornale degli Economisti, pp. 3 ff.
1955d (With M. M. Babbar and E. Heady), 'Programming with consideration of
variation in input coefficients', Journal ofFarm Economics 37, pp. 33 ff.
1955e 'Einige Grund - Probleme der Okonometrie', Zeitschrift Fuer die Gesamte
Staats wissenschaft 111, pp. 601 ff.
1956 'Statistik und Okonometrie', Statistische Viertel-Jahrschrift 9, pp. 92 ff.
1957a 'Les programmes lineaires stochastiques', Revue d' Economie Politique 67,
pp. 601 ff.
1957b 'La theorie probabiliste de Carnap et son Application aux problemes de
l'Econometrie', Economie Appliques 10, pp.19 ff.
1957c 'Mathematics and Statistics for Economists', Metroeconomica 8, pp. 146 ff.
1957d 'Game Theory, Linear Programming and Input-output Analysis', Zeitschrift
fuer Nationaloekonomie 17, p. 1 ff.
1957e 'La Teoria dei Giochi, La Programmazione Lineare e I' Analisi delle
Interdependenze Structurali', Industria, pp. 505 ff, 653 ff.
1957f 'Strategische Spieltheorie und ihre Anwendung in den Sozialwissenschaf-
ten', Allgemeines Statistiches Archiv, pp. 242 ff.
1957g 'Makroekonomische Modelle fUr die Osterreichische Wirtschaft', Zeitschrift
fuer Nationaloekonomie 17, pp. 262 ff.
14
KARL A. FOX*
Iowa State University, Ames, Iowa 5OOlI, u.s.A.
PART I.
THE MATHEMATICS AND ECONOMICS OF OPTIMIZATION OVER TIME:
EVANS, ROOS, AND TINTNER, 1922-1942
Tintner's own articles, and (4) the influence of his articles upon the
thinking and writing of others. From the vantage point of 1986, I
should also have known enough about recent (and future?) develop-
ments to say that Tintner's articles had received less (or more?)
recognition than was their due.
This task should have been tackled by a professional historian of
economic thought. After a generation of neglect following World War
II, the field has enjoyed a resurgence symbolized in part by the journal
History of Political Economy, established in 1969. A few young
people with strong training in mathematical economic theory and
econometrics have entered the field very recently and during the next
few years they may contribute the historical articles I sought, but did
not find, in 1986. Let me describe some of the problems of under-
standing Tintner's contributions.
From 1936 through 1942, Tintner published nearly a dozen articles
in which he used the calculus of variations to formulate dynamic
versions of various branches of static economic theory (Tintner, 1936,
1937, 1938a, b, 1939, 1941a, b, 1942a, b, c; Tintner, 1946 is a
logical sequel to the series). He cited Griffith C. Evans (1930) and
Charles F. Roos (1934) as sources of inspiration and examples of the
fruitful application of calculus of variations to dynamic economics.
I had met Evans in 1937 (when I was 20) and I had seen Roos at a
session of the Econometric Society about 1954. I knew that Roos had
joined with Ragnar Frisch in 1928 to seek Irving Fisher's support in
organizing a new society for the advancement of economic theory in
its relation to mathematics and statistics. The enthusiasm of the two
younger men convinced Fisher, and the Econometric Society held its
organizational meeting in December 1930. For this role if nothing
else, Roos deserves a prominent place in the history of econometrics.
As of 1937 I knew that Evans was a respected professor of mathe-
matics at Berkeley who taught courses in calculus of variations and
occasionally in mathematical economics. Many years later I saw a
ten-story building on the Berkeley campus called Evans Hall and
formed the impression that Evans' long chairship of the Mathematics
Department had contributed greatly to the department's international
stature. during the 1970s and early 1980s at least, Evans Hall also
sheltered the mathematically-oriented wing of Berkeley's Economics
Department.
Why was Tintner so impressed by Evans and Roos? I leafed
through the two books specifically cited by Tintner (Evans, 1930 and
25
ment firm, and then organized the Institute for Applied Econometrics,
whose name was later shortened to the Econometric Institute, Inc.
Roos had served as the first Secretary of the Econometric Society
during 1931-1936; he later served as the Society's Vice-President
(1947) and President (1948). A full-page photograph of him appears
as the frontispiece of the July 1952 issue of Econometrica.
It seems clear that Tintner's enthusiasm as of 1936-1942 for Roos'
work was scientifically appropriate. But I would like to have known
much more about the interactions among Evans, Roos, and several
other Americans who had been trained to the PhD. level in mathe-
matics, including Harold Hotelling, Harold T. Davis, Edwin B. Wilson,
and Irving Fisher. Wilson and Fisher had both written their doctoral
dissertations under the guidance of 1. Williard Gibbs, probably the
greatest scientist the United States had produced prior to 1900.
Fisher, of course, became primarily an economist and university
professor. Roos became a full-time econometrician active (from 1937)
in the private sector. Hotelling moved through economics to mathe-
matical statistics. Collectively, had these six or more trained mathe-
maticians stimulated enough graduate students and younger colleagues
to provide a readership for Tintner's work? I don't believe the needed
history has been written.
My impression is that Tintner's 1936-1942 articles received much
less recognition than their scientific merits justify. Moreover, econo-
mists who failed to cite Tintner also failed to cite Evans and Roos.
Samuelson (1947) says a great deal about dynamical theory and
functional equations but makes no mention of Evans, Roos, or
Tintner. This may have been an oversight, as in his enlarged edition
Samuelson (1983, p. xxv) includes Griffith C. Evans along with Frisch,
Tinbergen, Hotelling, Wilson, Leontief, and Georgescu-Roegen as his
'then living elders' whose works he had pored over during the 1930s
and early 1940s.
Baumol's Economic Dynamics (1951) cites two articles by Tintner
(1941a and b); the references are in an appendix to Chapter 5
supplied by Ralph Turvey: 'Uncertainty and the equilibrium of the
firm'. Turvey also supplied an appendix to Chapter 8 entitled 'A note
on functionals' but makes no reference to Evans or Roos. I scanned a
collection of articles on modern growth theory published in the 1950s
and 1960s, all of which involved optimization over time; I found no
references to Evans, Roos, or Tintner.
Roos (1927b, p. 280) wrote: "In a previous paper I have shown
28
"... the new theory [based on topology and convexity K.F.] claims
to be more suggestive, more general, more realistic, and more
rigorous than the earlier theories based on differential calculus and,
in light of the results to which it leads, it concludes that differential
calculus should never be used in studying equilibrium conditions
and maximum efficiency situations".
PART II.
AGRICULTURAL ECONOMISTS AS WORLD LEADERS IN
APPLIED ECONOMETRICS, 1917-1933
During the 1950s and 1960s many young economists shared a com-
mon set of misconceptions about econometrics: (1) econometrics
consisted of the limited information maximum likelihood (LIML)
method of estimation; (2) structural equations estimated by LIML
were right, all others were wrong: (3) all structural equations con-
tained disturbances and the disturbances in all possible pairs of
equations were significantly correlated; (4) all economic variables were
measured without error; and (5) the history of econometrics began in
the 1940s with solution of the identification problem, which had been
formulated by E. J. Working in 1927.
Not being curious about history, they did not realize that E. J.
Working's article was part of a very extensive literature; they had
found an island and missed a continent. A little further searching in
the same journal (QJE) would have disclosed articles by Holbrook
Working (1925) and Mordecai Ezekiel (1928) reflecting a sophis-
ticated understanding of the identification problem posed by E. J.
Working and noting a solution for it in most practical problems
involving agricultural products. Serious effort would have led them to
Holbrook Working (1922) who found substantial levels of error in
important variables and to F. V. Waugh (1923) who encountered
errors and disturbances in the same equation. A semester's explora-
tion of the related literatures might have converted them to a mature,
realistic, and professional conception of the place of quantitative
methods in economic research. Only the most rigid could have drawn
back at the last moment, saying "if the world does not conform to
LIML specifications, so much the worse for the world".
The coincidence in 1985 of the 100th anniversary of the American
31
USDA Graduate School during 1921-1922 and soon found that the
dominant intellect among his 40 or 50 students was a young Census
Bureau employee named Mordecai Ezekiel. Ezekiel had no farming
experience, and he had a B.S. degree in technical agriculture (not
economics). But he was brilliant, so Tolley offered him a job; Ezekiel
accepted because he admired Tolley. Louis Bean was a Harvard
M.B.A., but he had particular talents that BAE could use. Some
agricultural economists were recruited at the B.S. level and learned on
the job; among them was O. V. Wells, whom Tolley referred to as the
only other person he knew who could think as fast as Ezekiel.
The BAE during the 1920s was in a class by itself as a home for
applied econometrics; the land grant universities shared some of its
advantages. The size, economic organization, and political importance
of U.S. agriculture created a demand for comprehensible results that
could be relied on (within carefully-stated limits of applicability) by
farmers and policy-makers. These results were in the public domain
and they were supposed to reflect the state of the arts in the relevant
basic sciences. Analogous conditions were not met during the 1920s
in any agency, university department, or institute staffed by general
economists.
Additional historical and institutional background is given in Fox
(1987c). I tum now to some contributions of agricultural economists
and their close associates to applied econometrics. The contributions
involve multiple regression techniques, errors in variables, identifica-
tion problems, and examples of substantive research. Space permits
little more than an enumeration of these.
The first problem tackled by Tolley and Ezekiel (1923) was the design
of efficient and accurate methods for computing linear multiple regres-
sion and correlation measures by the method of least squares. Their
version of the Doolittle method greatly extended the accessibility of
least squares regression techniques to economists.
The second problem tackled was that of curvilinear multiple regres-
sion for any number of variables. Its solution is credited exclusively to
Ezekiel (1924), though the need for it arose during the joint research
effort reported in Tolley et al. (1924). Faced with the problem of
estimating a production function in which the output would be
expected to show diminishing returns to each of three inputs, Ezekiel
35
Sewall Wright. We have already noted that he had been familiar with
the demand-supply identification problem since 1915. Wright and
Ezekiel were well acquainted, and Wright's 1921 article, 'Correlation
and Causation', was cited by Ezekiel, Elliott (1927), Henry Schultz,
and others during the mid-twenties. Wright's method of 'path co-
efficients' was capable of quantifying the implications of causal
hypotheses for each separate path in very complicated systems. In his
1925 USDA bulletin, Com and Hog Correlations, Wright reduced a
system of 510 observed correlations to a central system of 14 path
coefficients connecting corn production, corn price, the summer price
of hogs, the winter price of hogs, and the amount of hog breeding in
successive years. Wright (1934) devoted several pages to applications
of path analysis to supply-demand models.
Concluding Remarks
articles from the former journal and five from the latter. It should not
surprise us to find parts of our history in them.
Eight or nine of our references are to bulletins of the U.S.
Department of Agriculture and various state agricultural experiment
stations. I believe few general economists are aware of the existence of
such publications, let alone their vast number and in many cases their
high quality and high prestige among the scientific communities for
which they were written. Great numbers are footnoted in Taylor and
Taylor (1952), most of them published during the years with which we
are concerned. An important part of our history is contained in them.
It is perhaps most astonishing to find one of the earliest and
greatest articles on causal ordering and identification in the U.S.
Department of Agriculture's Journal of Agricultural Research (Wright,
1921a). The article was at first rejected not on its merits but because
the editor believed that Wright, an animal geneticist, had no business
writing an article on statistics; it was reinstated through the interven-
tion of a maize geneticist in the Bureau of Plant Industry.
The formidable analysis of corn and hog correlations reported in
Wright (1925) was apparently done by him during World War I
(1917-1918) as a member of a committee appointed to assign hog
production quotas to the various states. Later, he wrote it up and
submitted it for publication as a USDA bulletin; this manuscript was
also rejected by the editorial office as out of Wright's field. Still later,
the manuscript was published through the intervention of Henry A.
Wallace, whose father, Henry c., had meanwhile become Secretary of
Agriculture. (My comments in this and the preceding paragraph are
based on a 1977 letter from Wright to the author.)
The editors' quibbles seem quite ludicrous when we realize that
Wright was soon recognized as a world-class scientist and one of the
'Big Three' co-founders (with R. A. Fisher and 1. B. S. Haldane) of the
field of population genetics. Wright (1921b) had a major impact on
animal breeding and Wright (1968-1978) must rank among the most
impressive scientific treatises ever written by a single author.
We have called attention to two traditions in theoretical and applied
econometrics which could greatly enrich the teaching and practice of
economic analysis today if they were widely known. The econometric
paradigm which coalesced during the 1940s and early 1950s incor-
porated an extremely narrow range of ideas. Errors in variables were
excluded from the paradigm for mathematical convenience, and the
exclusion was perpetuated in a set of computing instructions carefully
42
REFERENCES
GOPAL K. KADEKODI
Institute ofEconomic Growth, Delhi, India
Abstract. This paper deals with a theoretical framework to develop a pricing system
for energy products. While doing so either in a market or planned economy, issues
of equity and exhaustibility are accounted along with some efficiency measures. The
dynamic path of optimal energy prices so evolved calls for corrections on marginal
cost pricing rules. Both interest rate changes and tax (or subsidy) measures are
required in a judicious way all along the period of energy production out of a stock
of exhaustible resources. In the long run, energy extraction is equivalent to capital
investment while only for short-run planning it can be treated under the con-
servation ethic namely, resource conserved is equivalent to capital investment
underground.
1. INTRODUCTION
Jati K. Sengupta and Gopal K. Kadekodi (eds.), Econometrics of Planning and Efficiency.
ISBN-13: 978-94-010-8146-7
© 1988 Martinus Nijhojf, Dordrecht
52
so called Hotelling's r percent rule implies that the royalty must rise
along the optimum price path at a discount rate r. This will be the
case irrespective of whether the extraction cost is constant or varying.
In a welfare framework, as against revenue maximisation, the inter-
temporal choice of extraction rates implies pricing of such resources
covering not only the marginal extraction cost but also a scarcity rent.
This rental or degradation cost as coined by Solow and Wan (1976)
reflects the impact of current rate of extraction on future resource
cost. Under a back-stop technology and increasing marginal extraction
cost, Heal (1976) shows that contrary to the usual notions (not
attributable to Hotelling), the scarcity rent decreases over time and
reaches zero by the time the resource is depleted and a switching is
done to a substitute or additional resource.
In addition to exhaustibility and economic efficiency arguments, two
specific welfare aspects are also important in pricing of resources.
First, the distribution of income among people and the demand
patterns are uneven at any given point of time, and is also varying
over time. Therefore, the utility of income over time and across people
can affect the demand pattern and hence the time preference or
discount rates. Secondly, the social marginal utility of income from the
production of exhaustible resources may have to be adjudged while
determining the optimum price path. Both these distributional equity
factors affect the social welfare benefits of the resource. In a price
determination model then, the objective itself can be social welfare
rather than profit.
In this paper a welfare maximisation model is formulated with an
intention of deriving optimum price paths for energy production,
keeping exhaustibility and distributional equity into consideration.
Section 2 develops a price determination model with efficiency,
exhaustibility and equity considerations. The next section is devoted to
interpret the results and analyse the price paths under static and
dynamic conditions with alternative assumptions on distributional
patterns. Finally, comparative static analyses of the price scenarios for
changes in cost of energy production and discount rate are examined
for an optimal set of tax, subsidy and interest rate policies.
Let there be only two commodities called consumption good (Qc) and
53
rL
Lagrangian principle as:
+AI [ R, - r Q. dt ] + (1)
+ A, [r (P. - m) • Q. e-. dt - F ]
for energy is a function of its price, the basic optimisation can be done
with respect to energy price P, given its demand structure.
The optimisation process then implies that at each time t the
following condition be satisfied by the energy price to be so
determined. 3
au -e)/ aQ,
N
Jy
-a
P,
. f (y) e dy - At -a-
P,
+
(2)
(P, - m) a
Q
+ 11.2
1 [
Q, e -or = O.
ap,,+]
The total demand for consumption and energy products are the sum
of individual demands. Fourth, it may be noted that the individual's
price elasticities are also those for the aggregates.
aQ, P, aq, P,
'YJ,=-' -=-. - etc. (5)
ap, Q, ap, q,
Finally, an average social marginal utility of income measured in terms
of the shadow price of profit or surplus be defined as:
au 1
B =
J
y
-. -
ay A2
. f(y) dy. (6)
[f [f :: :,
r
q,J(y) dy ] f(y) dy ]
(7)
au 1
N f qr - ay
- • fey) dy
Az
---"-------':.------=:.--_--
B . Qr
This measure expresses an association between social marginal utility
of income and the demand pattern for energy products, resulting from
an income distribution fey) and a price structure (P" Pc). Specifically,
R r is a ratio of weighted average demand for energy over all income
classes (with weights being the social marginal utilities of different
income classes) to the average demand ~ corresponding to equal
social marginal utilities of income. Furthermore, B . R r can be inter-
preted as a weighted average social marginal utility of all income
classes, with consumption of energy as weights.
Specifically, if we assume that (i) the income distribution to be log-
normal with V(log y) = oZ, (ii) the commoditywise income elasticities
of demand to be constants (say a c and a r ), (iii) the elasticity of social
marginal utility of income to be a constant e, then the expressions (6)
and (7) can be written as: 5
B = e 1120(1 + 0)0 2 (6a)
(7a)
It can then be seen that for increased income inequality, Le. OZ
increasing, the average social marginal utility of income (for the
society as a whole) increases whereas the income distributional
characteristic decreases. In the limit we have the following: For an
extreme inequality when OZ -> 00 and B -> 00 note that Rr -> e.
Similarly, for a perfect equality of income, both Band R r tend to
unity. Any commodity which is of the nature of an essential good,
is expected to have a high distributional characteristic (due to low
values of a r and a c ). Luxury goods will have low distributional
characteristics.
56
Using (6a) and (7a) and the definitions of price elasticities, one can
rewrite (2) as:
A2 Q,BR,e- ol + A1 rJ, Q, -
P,
(2a)
(leO I + m) . 17,
P, = --"l-+-n---B-'--'-R'-'--- (8)
'" ,
The price path follows the Hotelling's rule with the scarcity value of
energy increasing at the discount rate o. The non-negativity of prices
is assured by imposing an assumption that 1 + rJ, - B . R, < O. The
initial price P, at t = 0 depends upon the intrinsic or opportunity
value of the resource in reserve before extraction, as well as the
income distributional and demand characteristics,
Following the same approach, the optimal price of consumption
goo,ds can be derived by maximising the welfare function in (1) subject
to budgetory conditions. The price of consumption goods can be
shown to be:
n
'Ire
~
Qc (B . Rr -1)
Pc = -----=-=------ . P" (9)
17
,
[B .R+ 11 ++ 17c
c
]
17,
The price path for the exhaustible energy resource can be analysed
under different scenarios, assumptions and comparative static para-
metric variations. Clearly, both Rotelling's basic result of inter-
temporal 0 percent increase in prices and Feldstein's static pricing
rule incorporating income distributional and social marginal utility
characteristics are special cases of this general pricing rule.
In the case of perfect competition in the resource market, the
optimising pricing rule (8) simplifies to:
limit Pr = A e bt + m. (10)
1],- 00
Clearly, the resultant price is much more than the price under the
competitive case.
Removal of budgetary condition amounts to a modified Rotelling's
rule which also accounts for income distributional characteristics. 6
A Of
P
r
= I e I I
B. R, 'YJr' (12)
58
1+ - B . R ] 1],
oo
+ Aeo1f' dt = _H_O
"0 [
f
(m 17, '. (15)
o A 17,
Since an explicit solution to the energy scarcity rental is not easy to
obtain from Equation (15), its behaviour is deduced using implicit
function rules. Equation (15) and the shadow price of the resource A
can be expressed as:
(16)
_ ~ [ 1 + 17, - B . R, ] 1], = 0
A 17,
q, . A (m + J. e")"-' e& dt
1+ n - B . R ] 'I, -
-Ro . B ./' ,
1
[
rJ, > 0.
J:
-------"'------'-----------='---
(19)
A . q, (m +J. e&)"-' e& dt
NOTES
1. See Dasgupta and Heal (1978), Solow (1974), Solow and Wan (1976),
Herfindahl (1967), Northaus (1974), and Pindyck (1978) for detailed discussion
of these issues.
2. In order to keep the model specification in a manageable form we are not
introducing the profit constraint on the production of consumer goods; but can
be added if necessary without loss of any generality.
3. Similar condition can be derived for the price of consumer goods also.
4. The utility U (PC' P" y) be maximized subject to the budget constraint Pc • qc +
P, . q, = y.
5. See Kadekodi (1985) for details. These assumptions made here are not far from
reality at least for highly aggregated commodities considered here.
6. Here the income distributional characteristic and average social marginal utility
of income are to be redefined without A,2'
62
REFERENCES
Ahmed, E. and Stem, N. H. (1981), 'On the evaluation of indirect tax systems: an
application to India', (mimeo), Development Economic Research Centre, Univer-
sity of Warwick.
Dasgupta, P. and Heal, G. M. (1978), Economics of Exhaustible Resources,
Cambridge University Press.
Farzin, Y. H. (1984), 'The effect of the discount rate on depletion of exhaustible
resources', Journal ofPolitical Economy 92: 841-851.
Feldstein, M. S. (1972), 'The pricing of public intermediate goods', Journal of Public
Economics 1: 45-72.
Heal, G. (1976), 'The relationship between price and extraction cost for a resource
with a backstop technology', Bell Journal ofEconomics 7: 371-378.
Herfindaul, O. C. (1967), Depletion and Economic Theory, University of Wisconsin
Press.
Hotelling, H. (1931), 'The economics of exhaustible resources', Journal of Political
Economy 39: 137-175.
Kadekodi, Gopal K. (1985), 'A welfare approach to energy pricing: a case study for
India', Energy Journal 6.
Murty, M. N. (1983), 'Efficiency and distributional equity and optimal structure of
prices for public electricity supply in India', Institute of Economic Growth,
Working paper No. E/86/83.
Nordhaus, W. D. (1974), 'Resources as a constraint for growth', American
Economic Review 64.
Pindyck, R. S. (1978), 'The optimal exploration and production of non renewable
resources', Journal ofPolitical Economy, Vol. 5.
JEFFREY B. NUGENT*
University ofSouthern California, Los Angeles, U.S.A.
Thirty-five years ago there was considerable hope that foreign aid
could playa vital role in the economic development of less developed
countries (LDCs). This hope was buoyed by the following considera-
tions: (1) LDCs were conceived of as being constrained almost exclu-
63
Jati K. Sengupta and Gopal K. Kadekodi (eds.), Econometrics of Planning and Efficiency.
ISBN-13: 978-94-010-8146-7
© 1988 Martinus Nijhojf, Dordrecht
64
Bloc A
I. Production Relationships
Production Function (1a) Xs=Min [ : ' XIs +M1s ... XiS +Mks ... X ns + Mus ] for each s, s = 1, ... , 12.
us a ls ais ans
for Gross Output
Production Function (1b) vs = ys[ar sR-Prks
s
+ (1 - a r)K-PrkSj-V/PrkS
ss for each s.
for Value Added
Fixity of Input-Output (2) XiS + M ks _ Xis + Mks = dis + mks for each i, s, = 1, ... , 12.
ais = X - X X
Coefficients s s s
mks - 0
mks [ Pdis ] 0",,1.
Substitution between (3) - -= ----=-0 -- for each k, s.
Imported and Domestic dis dis Pmks
Input Components
Domestic Value Added (4) dvs = ( I- ~ ais ) = ( 1- ~ dis - ~ mks ) for each S = 1, ... , 12.
Per Unit of Gross Output
IS
Ps - L~ d PdIS - L mksPmks p*
i k =_s
Effective Price of Value (7) Pvs = drs for each s.
Added 1- L diS - L mks
k
Sectoral Profits (9) TIs = 1lsX s = PvsdvsXs - L ResPre = PvsdvsXs - RsPrs for each s.
Sectoral Profit Rate (10) 1l~ = (lls/Ks) for each s, s = 1, ... , 12.
Household Class c
III. Factor Demand and Factor Supply
c
Variable Supply of (24a) R~ - R~o r w;;r l~Pr' [ :c~~c, ]~Y'd [ PrclPrco ]~pr,o.
Labor Type c co c,o Pr/Pljo
Prco
Pc,"
BlocB
IV. Factor Income Generation and Expenditure Allocation
Disposable Income of (26) Yc = (1 - iyc)prcR~ + L TR hc for each c, where h = c!, ... , C6' n, G, F
Household Class c h
+I I tmks(Pm ks - tmks)Mks +
k s
+I I td;c(Pdjc -tdjc)Cd;c+
e
+I td;c(Pd;G - tdjG)Cd;G+
+I td;(Pd; -td;)ld;+
+I tmkn(Pmkn - tmkn)Imk +
k
+I I tm kc (Pm kc - tmkC> Cm kc +
k e
+I tmkc(PmkG -tmkG)CmkG+
k
---l
.....
-...l
N
Table 1 (continued)
P,d iE "'-D'i2
E .
Sectoral Exports (29) Ei = f io [ pte. Pews ] WI
for each i.
Consumption Expenditure (30) CiePC;e = YiePC;e + cie [ ~. - ~ yj'Pc;,.] for each i and each c.
of Household Class c
of Goods Type i
Ie d it Pmks
into Domestic and
Imported Components
GDP, fGDI' [ _
fR ' ] fiR
Investment Function [
(41a) 1=10 GDP ] fRo'
o
-.J
W
Table 1 (continued)
-.)
~
Equation (49) defines the aggregate demand for imports of type k (in
real terms) and, finally, Equation (50) aggregates over all commodity
types to obtain total imports.
Table 2 (continued)
Note. All the equation numbers referred to are those given in Table 1.
(a) the level of foreign savings (Sf), (b) the supplies of labor of the
different region and income-skill groups, (c) the consumption func-
tions of these region and income-skill classes, (d) the investment
function, and (e) the parameters of the sectoral production functions
which indeed play such an important role in generating the factor
demands and hence the incomes of the different income-skill classes.
While the alternatives identified cover only a tiny fraction of the
sensitivity tests that could be performed, they include many of the
80
Row A B C D E F
-
GDP I GDP I GDP I GDP I GDP I GDP I
1 6583 1589 6583 1589 6583 1589 6583 1589 6583 1589 6583 1589
2 6589 1647 6590 1631 6590 1631 6590 1591 6590 1589 6589 1621
3 6584 1590 6584 1590 6584 1590 6585 1591 6585 1589 6592 1597
4 6673 1711 6683 1687 6685 1681 6710 1620 6722 1589 6684 1687
5 6515 1552 6513 1558 6511 1562
6 6605 1672 6613 1653 6614 1652
7 6752 1674 6756 1663 6761 1653
8 6843 1799 6858 1765 6865 1750
9 6148 1313 6132 1366 6121 1400..
10 6226 1424 6219 1447 6211 1473
11 6661 1588 6678 1622 6680 1616
12 Sno 1748 6780 1720 6784 1709
13 6568 1581 6568 1582 6567 1583
14 6650 1697 6658 1674 6660 1670
15 6592 1592 6593 1592 6593 1591
16 6687 1717 6698 1693 6702 1687
17 6581 1590 6580 1590 6580 1590 6581 1588 6581 1589 6582 1589
18 6603 1644 6606 1630 6611 1616 6614 1596 6615 1589 6617 1606
19 6407 1437 6400 1472 6396 1486
20 6432 1481 6427 1503 6424 1515
21 6831 1756 6838 1727 6842 1716
22 6852 1820 6863 1777 6869 1759
23 6443 1588 6445 1578 6443 1588 00
.....
00
N
Table 3 (continued)
Notes. All figures are in millions of Egyptian pounds at 1976 prices. For explanation of modelling assumptions and parameter
values see Table 2 and text. Blank entries indicate simulation not performed.
83
Table 4. Index of Income Inequality: Ratio of the Real Wage Rate of the Highest
Income Group in the Urban Sector to that of the Lowest Income Group in the Rural
Sector for Alternative Assumptions About Levels of Foreign Savings, Functional
Forms and Parameter Values
A B C D E F
Table 4 (continued)
Footnotes: I Odd numbers rows pertain to the base level of foreign savings, i.e.,
654 million Egyptian pounds. Even numbered rows assume a 10% higher level of
foreign savings, i.e., 719.4 million Egyptian pounds. 2 For explanation of other rows
and column numbers see Table 2.
decrease in the real wage rate differential between these two groups
relative to that of the base run, the latter ratio indicates a substantial
widening of the wage differentials, again relative to the base run, i.e.,
what actually transpired according to the 1976 Egyptian SAM.
By reflecting on the characteristics (assumptions) of these runs
which yield the extreme values of these relative wage rate indexes, it
becomes easy to understand how these results arise.
Given our primary objective of measuring the effects of external
assistance, we once again concentrate on comparisons with respect to
the different levels of foreign assistance represented by the value of
the ratio in a particular column of an odd-numbered row with that of
the corresponding cell of the even-numbered row immediately below
it. As the reader can easily see, such comparisons show that an
increased level of foreign savings almost invariably raises the degree of
income inequality. The reason for this derives from the fact that the
investment goods sectors, sectors 7 and 8, 'other industry' and 'con-
struction', the demand for, and hence the output of, which are
increased by virtue of the sizeable increases in real gross investment
induced by the additional foreign savings, (according to the Egyptian
SAM utilized in these simulations) employ only urban workers. The
increase in the demand for urban labor, even if most of it were in the
form of the lowest skilled class for that region, has the effect of raising
the real wage rates of urban labor relative to those of rural labor and
hence also the inequality index employed here. The only exceptions to
this rule occur between the cells of row 45 and the corresponding
ones of row 46. The difference in these cases is attributable to the fact
that the elasticity of substitution (ark) is assumed to be considerably
lower in these sectors (i.e., all urban sectors) than in the rural sectors,
requiring a greater reduction (or less of an increase) in the real wages
of urban labor in order to increase employment and hence output in
these sectors than in the rural ones.
While the results of Table 3 seemed to offer grounds for greater
optimism with respect to the effects of foreign savings on both income
and investment and hence the rate of growth of income, the results of
Table 4 indicate that, unless certain offsetting policies or programs are
adopted, increased levels of foreign savings would have the effect of
increasing the degree of income inequality. These results, of course,
call attention to the need for specific programs and policies for
offsetting the inequality-increasing effect of higher Sf
The reader should be reminded, however, that increased intergroup
87
inequality in real wage rates need not imply overall increased income
inequality among households both because intragroup income inequal-
ity may be reduced by fuller employment of those in that income class
as a whole and because fuller employment of the low income surplus
labor group may well allow household incomes of this class to be
increased substantially even without increases in wage rates. Likewise,
even if the increased real wage inequality result would carry over to
inequality in real disposable income, as indeed our results (not shown)
indicate that it does, increased income inequality need not imply
decreased absolute income or increased poverty among the poorest
group (the lowest income group in the rural sector, income class 4).
Indeed, the results for real disposable income of the rural poor (not
shown) demonstrate that this is not the case. Real disposable income
of the lowest income group, invariably increases, though generally only
modestly, with increased external assistance.
NOTES
*' The author expresses his apprecIation for the programming assistance of
Charles Williams who is the co-author and co-designer of the CGE modelling
package used in this paper and to fle editors for their useful comments and
suggestions. Gerhard Tintner's pioneering work in modelling some of the
issues discussed here, much of which was done at the University of Southern
California, was also an inspiration to this work.
1. See, e.g., Chenery and Strout (1966), Landau (1971), and Weisskopf (1972a,
b).
2. Among the most important studies of this type are Houthakker (1965), Griffin
(1970), Griffin and Enos (1970).
3. See especially Bauer (1981) and Tendler (1975).
4. One should also bear in mind that foreign aid may take many different forms,
each form possibly giving rise to different kinds of effects. For a relatively
comprehensive but outdated textbook analysis see Mikesell (1968).
5. For a nice example of the case study approach as well as numerous references
to other case studies see Tendler (1975).
6. This kind of an issue has a rich tradition in economics. For a recent and
especially relevant example see Chichilnisky (1980).
7. See, e.g., Adelman and Robinson (1978), Dervis et ai. (1982), and de Melo
(1982).
8. I is virtually exogenous in these cases since from Equations (41a), with the
elasticity of investment with respect to the interest rate set to zero, it can be
influenced only by the rate of change in GDP which, because of the fixed
labor supply and other assumptions of these cases, is minimal, i.e., 0.1 %.
9. This result pertains to the static results only. If the foreign aid takes the form
of loans, future net savings may be reduced, thereby giving the present value
of the net saving stream an ambiguous sign.
90
10. The real wage rates are the nominal wage rates divided by the class-specific
price indexes.
REFERENCES
JATI K. SENGUPTA
University ofCalifornia, Santa Barbara, California, U.S.A.
1. INTRODUCTION
lati K. Sengupta and Copal K. Kadekodi (eds.), Econometrics of Planning and Efficiency.
ISBN-13: 978-94-010-8146-7
© 1988 Martinus Nijhojf, Dordrecht
94
U = !U f
j-l
Uij = 1, uij ~ 0,
(1.2)
2. ROBUST SOLUTIONS
(2.4)
0-
E(~) = (1 - U21)b2 + (1 - u21)b20 22 • 0-2! 1 + 2]
30 22
a22 (a22)3 (a22)2
Since it follows from (2.4) that E(x;) > x;o, for i = 1, 2, it is clear
that it pays to have information on the probability distribution of Xl
and Xz. For a specific choice of the allocation ratios (uy I' ug l ), the gain
in higher profits must of course be evaluated against any higher risk
due to higher variance of profits. The trade-off between expected
profit and its risk is explored in the active approach by varying the
allocations lkJ E U. This may be called the mean variance efficiency
frontier. The selection of a point on the mean variance frontier as the
terminal decision may be made by the DM either on the basis of his
attitudes to risk or, by the risk premium prevailing in the current
capital market.
As a specific example consider that the vector c alone in the
objective function z = c' X is random and the risk averse DM
minimizes the risk R(z) = (x' Vx)/2 i.e.,
Min R(z) = (+)x'Vx
S.t. Ax=b (3.1)
m'x=r
where c is distributed with mean vector m and the covariance-matrix
V and x = x( lkJ) is conditional on a specific choice of the allocation
matrix lkJ E U. Note that we have deliberately ignored the non-
negativity constraints on x. For a fixed positive value of rand
b = b( lkJ), the optimal solution x* and the associated mean variance
efficiency frontier may then be directly calculated as
'* = g/2 when a~(,**) = a~* = kgh + b' Q-t b - g2/ 4. (3.3)
Secondly, let us consider two allocation processes Ut , ~ belonging
to U which transform the resource vector b to b(l), b(2) respectively.
Let a~(i) = a~(b(i» be the corresponding varaince frontier (3.2) for
i = 1, 2. If the two variance frontiers intersect and it holds that
a~(l) < a~(2) to the right Oeft) in a local neighborhood, then the
allocation U t E U may be said to dominate ~ E U in terms of lower
risk. Since the set U is compact, there must exist an allocation process
denoted for example by up E U, which is not dominated in some local
neighborhood Dt i.e., a~(P) ~ a~(i) for U j E U, i = 1, 2, ... ,
P - 1. In this case the allocation\ up E U is efficient in D t in the
sense that it minimizes the optimal risk for a given value of ,. Thirdly,
the allocation matrices U E U of the active approach may be used to
select m out of n outputs (n > m), so that we have a total of K
selections where K is the combination of n outputs taking m at a
time. Let x(j), j = 1, 2, ... , K denote a particular selection with
parameters VU), mU), AU) and bU). Then we can set-up the
following optimization problem:
Min A(S)
n
and
Max ,u(s)
n
L Pi = 1, Pi ~ 0, i = 1,2, ... , m
i=1
99
where the errors Eij , Ei are small perturbations around the mean
° °
values. Clearly the conditional variance V(l IqO) of l given q = (qJ)
can be evaluated in principle and if this is too high, the DM has to
select another terminal decision vector different from qo.
The active approach introduces allocation mechanisms through the
resource matrices u E U so that each hies) is changed to f3i(S) =
hi (s Iu) and hence bi changed to f3i' Thus for each allocation mecha-
nism u = up it introduces a sequence of optimal solutions denoted as
(qO( t), XO( t», for which a sequence of conditional variances VO( t) =
V (l o( t) I q-0(t» can be calculated as in (4.3). Let T be a finite set of
selections t = 1, 2, ... , T of the allocation mechanism uy E U. The
mean variance efficiency frontier is then specified by (l O(t), VO( t);
t = 1, 2, ... , T). Knowledge of this frontier may be utilized by the
DM in several ways. First, he may apply the method in (3.3) if there is
a region over which VO(t) is a strictly convex function of XO(t).
Secondly, the observed risk premium in the market may be used to
determine an equilibrium point on the mean variance frontier, just as
the capital market line in portfolio theory identifies a market equi-
librium. Thirdly, he may evaluate the higher conditional moments of
the distribution of l O(t) over and above the mean and the variance
and decide on the final solution vector. This will of course depend on
his risk attitudes and the form of the conditional probability distribu-
tion of the payoff l O( t). If he has any subjective goal e.g., A, then the
criterion of minimizing the distance d 2(t) = (A - XO(t»2/VO(t) over
100
lack of input and output prices. Let Xij be the quantity of input
i = 1, 2, , m and jj be the single output for the j-th DMU where
j = 1,2, , n. Then the DEA model may be simply represented by
the following LP model:
subject to
m
(5.1)
L f3iXij ~ Yj; j = 1,2, ... , n
i=1
By letting the reference DMU index k vary over 1, 2, ... , n one may
determine n optimal weights f3*(k), k = 1, 2, ... , n for the inputs.
For the more general case of x outputs (s ~ 1), the DEA model
becomes \
subject to
m s
L f3i Xij ~ L a,y,j; j = 1,2, ... , n (5.2)
i=1 ,-1
L a,Y,k = 1
,=1
the dual:
n
max z = L Yj).,j
J. j=1
subject to
n (5.3)
L xij).,j ~ Xik ; i = 1,2, ... , m
j-l
).,j ~ 0; j = 1, 2, ... , n.
Hence there must exist an optimal feasible vector f3* = f3*(k) for any
given k.
The primal and dual problems may be written in vector-matrix
terms as
min gk = x~f3
fJ
subject to X' f3 ~ Y (5.4)
f3~0
and
max Z =Zk = y' )"
J.
output for j-th DMU. One aspect of the DEA method not yet fully
explored is the manner in which the stochastic variations in input-
output data affect the parameter f3 * used in efficiency measurement.
Two methods may be suggested for defining a robust solution vector
such as f3*. One is to replace the objective function of (5.4) by
min g = x' f3, where x is the mean vector of inputs such as Xi =
(lin) k;_1xij' The LP model then takes the form:
min g=x'f3; R ={f3:x'f3 ~ y;f3 ~ O} (5.6)
{JER
Let /3* be the optimal solution of (5.6) and assume that there are data
variations (i.e., Ed around the mean i.e., xk = X + Ek within a certain
neighborhood N e such that the optimality of /3* is preserved for all
Ek E Ne • Then the solution /3* is robust in Ne and it may be used to
rank the n DMUs in order of efficiency. Several implications of
characterizing such a robust solution /3* have been discussed by
Sengupta [4, 5J and applied empirically.
The second method is to apply a minimax procedure to estimate the
production frontier from the observed data set (X, Y), on the analogy
of a regression procedure based on the least squares (LS) approach.
Let ej(f3) = k ~_I f3iXij - Yj be the error in Yj associated with a vector
f3, then the Chebyshev method of estimation consists in finding a
vector point f3 0 at which the minimax level L(f3°) of loss is achieved,
where
L = L (f3 0) = min max Ie/f3) I. (6.1)
(J 1 <;j <; n
So long as the errors ej(f3) are bounded, the solution f3 0 also known
as the Chebyshev solution always exists, even when the inequalities
ej (f3) ~ 0, j = 1, 2, ... , n are mutually inconsistent. The objec-
tive function (6.1) is closely related to that of the least absolute
value (LAV) method of estimation wI¥ch minimizes the los function
k;=1 Iej(f3) I with respect to f3. Let f3 be an LAV estimator. It is
known [3J that the LAV estimator has a significantly smaller standard
error than the LS estimator for any r~gression model with high-
kurtosis disturbances and that the error (f3 - f3) in the LAV estimator
is approximately distributed as a normal variate with mean zero and
covariance matrix k2 . (X'Xr 1 where k 2/n is the variance of the
median of a sample of size n from the error distribution and X is the
data matrix.
104
Min f3m+t
m
(6.4)
n m
Xi = Ui L Xij' Ui ~ 0, L Ui = 1
j=1 i=1
(6.6)
where it is assumed that the variance-covariance matrix Ve is non-
singular. Thus if all the vector points x in the cluster are very close to
the mean, the value of D2 will be close to zero. The higher the value
of D2, the farther the points are scattered away from the mean level.
The role of the variance-covariance matrix Jt: is more subtle however.
It has two practical implications. One is that it acts as a filtering
device, whereby with noisy data any difference Xi - Xi is deflated or
corrected in terms of its standard errors, so that the squared distance
between x and X can be more reliably computed. Secondly, if the
underlying probability distribution of the errors e can be assumed to
be normal (e.g. under certain conditions by the central limit theorem),
then the random quantity D2 in (6.6) is known to follow a chi-square
distribution with m degrees of freedom if the variance-covariance
matrix Jt: is known. Thus one may choose a level of significance a
and determine a suitable positive scalar C = ca such that x defines a
confidence region
(6.7)
in the sense that for any set of points x belonging to region R we have
Prob(x E R) = a. By varying a one can therefore vary the size of
the neighborhood around X. This formulation now represents the
industry manager's efficiency problem comprising n units or enter-
prises in an explicit form. Thus he may choose points which are close
to X in the sense defined by (6.7), thus generating new solutions which
are either the same as before or different. In the first case we have
robust solutions within the neighborhood defined by (6.7). In the
106
(6.8)
where xa , xb are two suitable fixed points such that for any point x
belonging to N(x; x), the original solution vector ~* associated with
mean inputs x remains optimal. Now let B be the intersection or-the
two sets R and N(x; x), where xa , xb are fixed and a is varied in (6.7)
such that B is a nonempty set. It is easy to show [51 that such a
construction is always feasible for a given solution vector ~*. Thus a
robust solution vector could be specified through suitable variations in
the allocation ratios U i •
An important implication of the active approach formulated in this
way is that it can be used to characterize a nonempty core of the many
player game associated with (6.4). Assume that the n DMUs form a
set N = {I, 2, . . . , n} of n players who can form coalitions (or
mergers) of different sizes. Let S be a coalition of less than n players
and N is the grand coalition of all the n players. For the LP problem
Max z = I y.k
J J
jES
Aj ~ 0 j E S
We may form the characteristic function v(S) for the coalition seN
as
v( S) = max I YjA j
yES
Aj ~ 0, j E S
where xi(S) is a particular allocation of input i form the total amount
~ j E S xij the coalition has. It is easy to show that this input-allocation
107
4. CONCLUDING REMARKS
minimax model
Min flo
kEK
subject to a 2( k) ~ flo
#(k) ~ r
where flo and r are scalar positive quantities. By using a distance
function
d(k) = i(k), V(k)i(k) - wfl*'(k)x(k)
this may be put in the form
Min Max d(k)
p' i
REFERENCES
Abstract. The paper first groups some key results in the choice of E, V efficient
portfolios by homogeneous programming. The optimal portfolio and the risk share
accruing to the financial intermediary are determined by the rate of interest. This
portfolio solves a two-person game. Existence theorems are established for Nash-
type equilibria in imperfect capital markets. This is done first for financial institu-
tions acting on given monetary policy parameters. The result is then generalized to
allow interaction with the instruments of monetary policy.
Jati K. Sengupta and Gopal K. Kadekodi (eds.), Econometrics of Planning and Efficiency.
ISBN-13: 978-94-010-8146-7
© 1988 Martinus Nijhojf, Dordrecht
110
where
K == {kIProb(f(x) ~ k) ~ a;x EX}
t:
and, for normal distributions for instance,
where
X={x~OIAx~b}
m = (Ex* - v*b)/ax*
so that the risk attitude m can, in principle, be measured objectively
as the coefficient of variation of net returns.
Empirically plausible orders of magnitude found for m were 1.5 for
Midwest farmers and 0.5 for sharemarket investors, ct. Moeseke
(1965a) and Moeseke and Hohenbalken (1974).
113
3. OPTIMAL PORTFOLIOS
X={x~Olux~l} (3.1 )
where
?(x, m) == Ex - max
M == {m ~ 0 IEx - max ~ r}
and r is the rate of interest on deposits. To see this, recall that by the
duality theorem of homogeneous programming any solution x* to
max (Ex - max) = max ?(x, m)
x x
satisfies
Ex* - max* = A* = ?(x*, m) (3.3)
where A* is the optimal value of the dual variable A assigned to the
single budget constraint ux ~ 1, i.e. the budget dollar's marginal value
to the investor for given m.
Define
(a2) that
G(m+) = r
for a unique m + such that 0 < m + < liz (explicit proofs in Moeseke,
1968, 1980) and that, if x+ is a maximizer of ¢(x, m+), the pair
(x+, m+) solves (3.2).
We have shown elsewhere (Moeseke, 1980) that (3.2) is a two-
person convex game with kernel ¢ defined on the strategy spaces
X X Nt where Nt =
[0, liz]. Both X and Nt are convex, compact and
¢ is concave in x, linear in m. Hence there is a solution (x+, m+),
where m+ is unique, such that
min max ¢ = max min ¢ = ¢(x+, m+) = r.
M x x M
4.1. Modell
(3.4) by
Gi(milx-i );: max ¢i(Xi , mil X-i),
Xi
where
Xi ;: {Xi ~ 0 I ux ~ 1},
and in (a2) write Gi (m i IX-i)' Define
M, E [ 0, m~ m, l
PROPOSITION 4.1. There exist sequences x*, m* such that all
investors hold optimal portfolios, i.e. portfolios satisfying
¢i(xi, mi I X!i) = min max ¢i(Xi , mil X!i), all i. (4.1)
Mi Xi
4.2. Model II
PROPOSITION 4.2. There exist sequences r*, m*, x* such that all
investors hold optimalportfolios, i.e. portfolios satisfying
~i(xi, mi I r*, X~i) = min max ~i(Xi' mil r*, X~i)' (4.3)
Mi Xi
PROPOSITION 4.3. There exist sequences x*o, x*, m*, r*, 13* such
that all investors hold optimal portfolios, i.e. portfolios satisfying
(4.6)
all i.
if
,{,.' =- Y'i
Y'i Xi' Xi', m ,i I r,- 13-,x-
,{,. (0, - )
i
= min max ¢i(X?, Xi' mil r, i3, X-i), all i.
Mi Xi
2. See however Hogan et ai. (1979) for an econometric study of the effects of
banking regulations.
120
The correspondence [3 sending (r, (3, i) into (f, xo" x') is convex
usc and by the Kakutani theorem there is an (r*, xo*, x*), where
r* = ~* and xo* = {3*, satisfying (4.6). QED.
REFERENCES
INTRODUCTION
Jati K. Sengupta and Gopal K. Kadekodi (eds.), Econometrics of Planning and Efficiency.
ISBN-13: 978-94-010-8146-7
© 1988 Martinus Nijhojf, Dordrecht
124
subject to x( N) = v( N) where x( S) @ L Xi
;ES
Notice that this problem is well defined and soluble whether or not
(N, v) is super-additive. Indeed, the concept may be further extended
by .weighting the coalitional excesses which appear in the absolute
values. The relation of this extremal principle to the core is rendered
in the following:
THEOREMl.
x(n)-v(N) S"N
Proof
Now
= (n-1)
k-1 v(N), since x(N) = v(N).
So
S~N X(S) = 1+
I (n-1) 1 + ... + (n-1)
k-1 + ... + (n-1)]
n-1 v(N).
But
So
L xeS) = 2n - 1 v(N).
S"N
126
Hence
x(N)=v(N)
+ 2 L Ix(S)- v(S)1
x(S) < v(S)
= Kn + 2 L Ix(S)- v(S)I.
x(S)<v(S)
Thus, the left side equals K n iff {Xi: x(S) < v(S)} = 0, i.e. X is in the
core of (N, v). Q.E.D.
THEOREM 2.
A* ~ K n l(2 n -1)
with equality iff x* is in the core of (N, v).
Proof Sum the inequalities for 1 S I = k to obtain
L Ix(S)- v(S)I.
ISI=k
Hence
n
L L Ix(S)-v(S)1
k-l ISI=k
127
or
Thus
,*
I\. ~ 2nK
n . h
WIt .
equality
-1
iff L Ix*(S)-v(S)I=Kn
S';;N
Xi
*= 2-f3 [ /A-i + f3w(N) - M ]
n
where
/A-i = L w(S)m(S), f3 = L
n-I
m(k)
(2)
k= '
n
S?- i k=1 1
M = L /A-i
i=1
128
Proof Considering the Lagrangian L(x, ).) and setting its gradient
equal to zero, we have
and
0= x(N) - w(N).
Thus
Jli - ). = L x(S)m(s)
S ~ i
n-l
n-l
= L m(k) L x(S)
k=l S-k
S ~i
=
n-l
L m( k)
k=l
( n
k =1)1 Xi +
n-l
L m( k)
k~2
(
n
k
=2)
2 L. Xj
I""
129
+ L
n-I
k=2
m(k) n
k
= w(N)
(2)
2
= L
n-I
k=1
m(k) n
k
(2)
= 1
Xi + L
n-I
k=2
= w(N)
(2)
m(k) n
k 2
where
and
rJ =
n-I
L
n
m(k) k
(2)
= w(N).
k=2 2
* _ Ii
Xi -
1 [
Mi + f3w(N) - M ]
n . Q.E.D.
Notice that this solution, in common with the Shapley value, is linear
in the w( S). Thus a wide class of solutions, which are additive over
general games, has been specified. The Shapley value, as will be
shown, is the special case in which the coalitional weights
We obtain
<l>i = -1 L (n-l)-1
_ [w(S).:... w(S - {i})]
n S~i s 1
xi = _1_ [M + (n -
i
l)w(N) - M]
n -1 n
Proof Since both Shapley value and (GS)-solution are additive over
games, it suffices to match them over all 'Rlh unit vector' games i.e.
games with
)J.i -
-1 m(r)w(R),
0
,
ER
$R
and
n
Then
x'!' = mer)
f3
11 -- (r/n),
(rln),
ER
I fER
= mer)
f3
I (n - r)/n,
-rln,
iER
i fE R.
131
~ (n -1 )-1, ER
n r-1
<p.=
I
_~(n-1)-I, itER.
n r-1
Setting x1 = <Pi'
f3= L
n-I
m(k) n- (2) = L
n-I (
n- 2)-1 (n- 2) =n-1.
k=1 k-1 k=1 k-1 k-1
Thereby we obtain the above, new, expression for <Pi of the Shapley
value. Q.E.D.
EXTREMALPmNCWLESFOR
INCOMPLETE CHARACTERISTIC FUNCTIONS
s n-s
min A, A ~ w(S) - v(S) + v(S) - xeS) (HMCH)
x(N)=v(N) n n
s _ n- s
-
n
v(S) +
n
v(S), v(S) =v(N) - v(N - S),
REFERENCES
[1) Lucas, W. F. (1968), 'A game with no solution', Bull. Am. Math. Soc. 74:
237-239.
[21 Schmeidler, D. (1963), 'The nucleolus of a characteristic function game',
Research Memorandum 23, Mathematics Dept., The Hebrew University of
Jerusalem.
(3) Charnes, A. and Kortanek, K. O. (1970), 'On classes of convex and pre-
emptive nuclei for n-person games', in H. W. Kuhn (ed.), Proc. 1967 Princeton
Symposium on Mathematical Programming, Princeton, N.J.
[4] Charnes, A. and Keane, M. (1969), 'Convex nuclei and the shapley value',
Center for Cybernetic Studies Research Report 12, The University of Texas,
Austin.
[5] Charnes, A. and Keane, M. (1970), 'Convex nuclei and the shapley value
(abstract)" Proc. Int. Congress ofMathematicians, Nice.
[61 Seiford, L. (1977), 'Entropic solutions and disruption solutions for n-person
games', Ph.D. Dissertation Department of Mathematics, University of Texas,
Austin.
133
[7] Gately, D. (1974), 'Sharing the gains from regional cooperation: a game
theoretic application to planning investment in electric power', Int. Economic
Review 15: 195-208.
[8] Littlechild, S. C. and Vaidya, K. G. (1976), 'The propensity to disrupt and the
disruption nucleolus of a characteristic function game', Int. Jour. Game Theory
5: 151-161.
[9] Chames, A, Rousseau, J., and Seiford, L. (1977), 'Complements, mollifiers
and the propensity to disrupt', Int. Jour. Game Theory 7, 37-50.
[10] Michener, H., Yuen, K., and Sakurai, M. (1981), 'Experimental games', Int.
Jour. Game Theory 10: 75.
[11] Chames, A and Golany, B. (1983), 'Homocores, cores and operational
inefficiency in superadditive n-person games', Int. 1. Systems Sci. 14: 877-
893.
[121 Chames, A, Rousseau, J., and Seiford, L. (1982), 'Mollifiers for games in
normal form and the Harsanyi-Selten valuation function', Int. Jour. Game
Theory 11: 163-174.
Econometric Estimation of Decision Models
Under Uncertainty
T. V. S. RAMAMOHAN RAOI
Indian Institute of Technology, Kanpur
Jati K. Sengupta and Gopal K. Kadekodi (eds.), Econometrics of Planning and Efficiency.
ISBN-13: 978-94-010-8146-7
© 1988 Martinus Nijhojf, Dordrecht
136
p= a - f3X+u,
137
where
p = price per unit sold,
X = quantity sold at any point of time, and
u = random variable.
That is, the price which can be obtained on the market for any
given volume of sales is uncertain.
However, once the firm chooses the quantity X of output pro-
duced,8 the costs of production are well-defined. Let
C=F+mX, where
C = total cost of producing X units,
F = fixed cost, and
m = marginal cost 9 of producing a unit of X.
Now, for illustrative purposes, consider the decision process of type
(iv) detailed in Section 1. For this case the following decision proce-
dure is optimal:
(a) Fix a price p*,
(b) Fix a quantity X* which the firm will produce during each unit
of time, and
(c) Allow the actual X sold at price p* to be determined by the
market during each unit of time. 10
Hence, the observations generated in the actual operation of such a
decision process consist of:
(a) A fixed price p*,
(b) Actual quantity sold, X, which varies over time, and
(c) The total cost C. This will be mostly constant if X* is unaltered
over the time period under consideration.
The econometric estimation problem consists of obtaining usable
estimates of a, f3 as well as F and m. This is the crux of the identifica-
tion problem encountered in dealing with decision making situations
under uncertainty.
Clearly, there is no direct theoretical resolution 11 of the identifica-
tion problem. An indirect approach to the problem of parameter
estimation has to be developed. This is the main task of the rest of the
present study.
case. Such a firm generally waits until it gets to know the realized
value of the random variable u. It decides p and/or X only after this
information is available. That is, for any specific time period, it
chooses
2fJX = (a + u) - m, and p = m + fJx.
However, note that over time the value of u varies and this is the
primary source of variation in the observations generated for this firm.
Further, as noted earlier, an increase in u causes both X and p to rise.
The observed X, and consequently the total cost of production,
varies over time. Hence, the regression procedure can be utilized to
obtain estimates of F and m which are the parameters in the cost
curve. In the sequel it will be presumed that an estimate of the
parameter m is available.
For purposes of concreteness in the estimation process it will be
postulated that the random variable u is uniformly distributed over
the interval (0, 2A) where A is an unknown parameter. 12 Reconsider
the choice of X where u = 2fJX - (a - m).
Then, by the method of moments, it can be shown that
A = 2fJX - (a - m),
where
x = sample mean of the observed values of X.
Similarly, fJ can be obtained from the equation
p =m +fJx,
where
p = sample mean of the observed values of p.
Consequently, given an estimator of m,
"
fX
... -~_ . -_.
10'6 ..... - 1.
- m=2·0
"
"
10'3 m~2'5
J ·014
,
10'0
400 450 500
X "- ..... m=2'0
"
A ........ ,/
"- ....
·44 ·010 ~ ,
'9
,,
,, '008
400 450 500
X
,
" m=2·0
·36 ,I
"- , ,
·2 B I--_-..L.._ _-'---_---L-_ X
400 450 500
Fig.1a.
140
"
Q'
13
I'
"
"
(3 , 11
m=2·0
\ ,,-" "
~"
,
,
,,
~"
·012 m=2·0 I 9 ..... "
....
I
\/
....
·010 I 7 P
I
I 6 7 8
I
I "A
I
'5
I'
- "
•
.....
P .....
"
·006 m~2·0
'4
6 7 8
\...... " .....
'"
,." ..... m:02·5
1·
·2
6 7 8
is
Fig. lb.
charging the same price at a lower marginal cost only if the market
demand is substantially higher.
Consider a price setting firm next. That is, it corresponds to the type
(ii) decision structure alluded to earlier. For this case it would be
necessary to define the process of fixing the price.
Let the firm be risk neutral. 16 That is, the firm will be presumed to
maximize expected profits. To proceed with the analysis systematically
note that the
Expected revenue = p* E,
and
Expected cost = F + mE,
where
E = (a - p* + A )/(J,
and
p* = parametrically chosen price.
Clearly, E represents the expected market demand at the price chosen
by the firm.
From these expressions the choice of p* implied by the expected
profit maximization can be written as
2p*=(a +A +m).
Further, as in the previous section, it can be shown that
~ = (p* - m)/X,
A=8~,
and
a =2p*-A -m,
where
8 2 =(3/n)L(X-Xf
Figure 2 provides a numerical illustration of the sensitivity of the
142
11-0
10'6
1\
f3
'014
....
10-2 X
"A
400 450 500 """ .... ' ..... , m:2-0
·22 "I.
\
\
'010
"" .... ...
" "-
\
\
m:2·0
·008
400 450 500
X
,~
·18 , , ....
'e
m:2·5
.14 I I I
X
400 450 500
Fig. 2a.
"
cr .-
,- f3"
ni:.2·0 .- '"
,
,-
,
11 \,--'" "012 m=2·0 I
, , .-
,-
\/
9
,-
,-•
m=2·5
\ ·010 I
,,
7
•
P
I \
m=2·5
6 7 8
"
A
·006
.
P
6 7 8
·25 m:::2'0 ;0"
\... ..... ;0 . . . . . .
·15 ;0
K
.. K
·05 p"
6 7 8
Fig.2b
Expected revenue
= (p*u*l2f3A)(a - p* + O.5u*) + (p*X*I2A)(2A - u*),
where
Expected cost
= (u*I2Af3) (Ff3 + am - p*m) + (mu*2/4Af3) +
+ (1I2A)(F+ mX*)(2A - u*).
Maximizing the expected profits with respect to p* and X* and
proceeding as in the above case it can be shown that 20
p= (p* -
m)/X,
a =2p*- m -A,
and
A = f3(x* - X).
Figure 3 conveys an approximate idea regarding the sensitivity of
the estimates to variations in X, p*, and m. The following observations
are pertinent: 21
(a) As X approaches X* and full utilization of capacity is recorded
at each point of time the market uncertainty is eliminated. Hence, the
estimate of A becomes zero in the limit. The simulated results confirm
this expectation. However, note that the stability of the market
demand curve can signify either the emergence of a competitive
market organization or the possibility of significant monopoly power
to the firm. /3 reduces with X but appears to tend to zero only for
high values of m.
(b) Similarly, it was noted that as m tends to p* the estimates of A
and 13 become zero and the value of a tends to p*. This is also in
agreement with apriori expectations.
(c) However, it was found that all the three parameters - viz.,
a, 13, and A - are far more sensitive to X and p* than they are to
variations in the marginal cost itself. 22
145
II
()(
II
"
·014
8'5
500 X
·010
400 500 X
O'---~-:----'-----l--X
400 450 500
Fig.3a.
A few commodities would be such that the firm will have to sell all the
output produced at the going market price. Only some agricultural
commodities and style goods will belong to this classification. How-
ever, this case must be considered for logical completeness in a
discussion of decision models under uncertainty.
A quantity setting firm fixes X* to be produced and sold on the
market. In this process the firm is postulated to accept whatever price
it can obtain on the market. The expected profit for such a firm can be
written as
En = (a +A - m)X* - PX*2 - F,
146
1\
ot. ,
11'5 m=2'0
,,
I
I
I
\1' {3
1\
., "
10'5 I
/
I
/
·014 m=2·0
,I \ I
9'5 I m=2'5 \l/
p.
·012
I
,
I
/
1\ 7 8 9 /
A '010
2·0
·008
..
p
7 8 9
.......
1·5 m~2'O
\ ........
.....
1·0
.. JI ....
·5 p •
7 8 9
Fig.3b.
where
()2 = (3/n) L (p - pf
As before, if the price at each point of time equals the marginal cost,
then both the estimates of {3 and A would be zero. Further, the
estimate of a would be equal to m so that the market environment
can be described as competitive with none of the firms experiencing
any market uncertainty.
Notice that, as in the previous sections, it has so far been assumed
that an independent estimate of the parameter m is available. How-
ever, since X* and C are practically invariant over the sample, an
estimate of m cannot be obtained by a regression procedure. Instead,
an estimate of F will have to be obtained from a knowledge of the
capital costs, life of machines, and the value of m calculated on this
basisP
The sensitivity of the parameter estimates to the observed variations
in X*, p, and m can again be examined.
The following aspects may be noted:
(a) Neither A nor a depend upon the observed X*. Variations in
X* can only alter the estimated value of {3. Quite clearly, ~ decreases
with an increase in X*.
(b) Similarly, A is invariant with respect to changes in p and m.
(c) Both a and ~ vary directly with p.
Given these observations it may be concluded that the estimation
problem in the context of type (iii) decision procedures is the most
uninteresting of the various cases examined so far.
"
~
m=2·0
... ........ ........ .
8
",
,
\ ... y
f3"
4
·016
m=2·5
0 X .... I
-..::
"
A ·012 ""
8,5 ·008 I
X
400
4-5
Fig.4a.
~
m::.2·0 /
9·5
/
\../
y ",
~
;'
7·5 ;'
0·145
3·5~-~--~-p· 0·125
"
A
Fig.4b.
Expected revenue
= (p*u*12A{3) (a - p*) + (p*u*2/4A{3) +
+ (p*12A) (X* + I) (2A - u*),
where
u* = {3 (X* + I) - a + p*.
151
H = 1 - 2( m + i)l(p* + i)
P=AHI(X*- X),
a = p* + ~(X + I) - A,
and
A is estimated from the equation
A(m - p*)2 = (p* + i)2[(m - p*) + p(X* + 1)1.
1\
_A--,
.. ___ e
't
m=2·0
7·5
1\
6·5 f3
,0118
5·5 I
m::.2·0
20 30
A
1\ ·0114
~/
1i)::.2·5-- " " •
5·0 ,0110
20 30
4·0
3·0 - ... (
m=2·0
20
- 30 1
Fig.5a.
II
tX
8'0
...... -f--"'-"
m=2·0
7·0
II
[3
5'0 I
J.
·08 .16
1\
A
.110
·08 -16 .•
5·0
4·0
m=2·0
NOTES
variations in the costs can still be utilized to obtain an estimate of the marginal
cost.
20. Notice that in this case there are three equations involving the observed
decisions, viz., X*, p*, and X. This system is exactly determined and the
second moment is no longer necessary to obtain the estimators.
21. Figure 3(a) was drawn by fixing X* = 500, and p* = 6.5. In Figure 3(b) while
X* was fixed at 500, X was chosen to be 400.
22. It would have been instructive if a comparison of the estimated degree of
uncertainty with or without capacity limits can be obtained. However, the
estimates are not comparable.
23. It should be remarked that this procedure will not be adequate in the context
of multi-product firms. The system will remain under determined unless the
proportions in which the outputs will be produced can be determined
endogenously and utilized for estimation of the marginal cost parameters.
24. In the context of power generation, where the output cannot normally be
stored, the power plants generally choose a policy of fixed operations levels as
well as tariffs a priori. The possibility of the firm holding inventory to smooth
out sales over time will be examined in the next section.
25. In Figure 4(a) the assumed value for p* is 6.5. Similarly, in Figure 4(b) X* is
fixed at 500 while X takes the value 400.
26. This is the tradition of the Mills' model of decisions under uncertainty.
27. The assumption that there can be no backlogging of orders is being main-
tained. It should be easy enough to see that these assumptions can be relaxed
further without having to alter the basic analytical framework of this section.
28. In Figure 5(a), X* = 500, X = 400, p* = 6.5, and i = 0.08. Similarly, in
Figure 5(b), 1= 25 while i is allowed to vary.
REFERENCES
JAMES H. GAPINSKI
Florida State University, Tallahassee, U.S.A.
and
T. KRISHNA KUMAR
University ofHyderabad, Hyderabad, India
Abstract. After briefly reviewing the recent history of estimating the elasticity of
factor substitution by nonlinear least squares from the CES production function, this
note activates the elasticity estimates generated by a previous Monte Carlo study and
processes them to yield relative frequency distributions of the elasticity estimator.
The note then deals with the issue of estimating the 'location parameter' of the
underlying sampling distribution. Mean, mode and median are compared as
estimators of the 'location parameter'. It is shown that all the estimators are biased,
and biased towards unity - the Cobb-Donglas function. Among all these point
estimators median estimator is shown to have the least percent bias. The paper
highlights the limitations of CES production function specification and Nonlinear
Least Squares estimation of the elasticity of factor substitution. But if one must use
CES specification and NLS estimation it shows how to cumulate evidence from
comparable studies.
Jati K. Sengupta and Gopal K. Kadekodi (eds.), Econometrics of Planning and Efficiency.
ISBN-13: 978-94-010-8146-7
© 1988 Martinus Nijhojf, Dordrecht
158
function. Since NLS extracts the a estimate directly from the function
without recourse to marginal conditions, it has obvious appeal to
researchers interested in the production structures of planned econo-
mies. Weitzman [15], for example, pressed it into service during his
1970 investigation of the Russian growth performance. Three years
later Asher and Kumar [1] relied on it in their look at the growth
records of six countries, half (namely, Hungary, Russia, and Yugoslavia)
being planned in orientation. NLS also appeals in those instances
where the marginal conditions, while valid, are difficult to invoke
either because the data on prices have limitations or, more funda-
mentally, because the economic units under examination have a non-
competitive nature. Such consideration led Ryan [11] to adopt it in his
1973 work on British manufacturing industries. Even in the absence
of special circumstances, NLS appeals as the 1970 inquiry by Tsurumi
[14] into the production side of Canadian manufacturing attests.
Support for NLS in the CES setting continues and can be found in the
new generation of econometrics texts, notably in the 1980 volume by
Judge et aL [5].
ESTIMATION PROCEDURE
and
(2)
where, of course, a = 1/(1 + p). Data on capital, K i , and labor, L i ,
consist of annual figures for the US private domestic economy and
cover a twenty-year period. The parameters in the deterministic
portions of each specification carry the assignments y = 1.00, P =
0.03, ~ = 0.40, a = 0.35 and 1.35, with {3 = 0.80 and 1.20; hence
they give rise to four parameter combinations depending upon the
values of a and {3. The disturbances Ui and Wi obey lognormal and
normal distributions respectively. A total of 150 sets of 20 values for
ui are drawn randomly from the lognormal along with 150 sets of 20
values for Wi from the normal fashioning 150 samples of data on
Equation (1) and an equal number on Equation (2) for any parameter
combination. Applying NLS to these data therefore yields 150 Sj from
each model for each parameter combination. The Error Sum of
Squares (ESS) is a nonlinear function of the unknown parameters.
Hence the estimation by the method of least squares does not yield
a closed-form solution as in the ordinary least squares. The NLS
method developed by Marquardt [10] is essentially a gradient or
steepest descent method for minimization of ESS. It starts with an
initial set of estimates of the parameters (guesses) and expresses the
ESS as a Taylor Series linear approximation around the initial guesses.
The second round estimates are obtained by adding the optimally
determined differences or increments to the initial guesses. The
optimal differences are those that minimise the linearised ESS using
the standard matrix algebra of OLS. The second round estimates are
used as reference points to arrive at a revised linear approximation
of the ESS. This procedure is repeated iteratively until either the
parameter estimates converge or until the ESS reaches a minimum
level as determined relative to pre-specified tolerance limits. To help
shorten the convergence process and to give the regression package
every opportunity for producing reliable Sj' the true parameter values
are used as the initial estimates. 2
Eight Sj series emerge from this exercise, and their relative frequency
161
0< Sj < 0.25 0.329 0.075 0.252 0.035 0.345 0.013 0.283 0.000
0.25 < ~ < 0.50 0.200 0.150 0.343 0.141 0.183 0.169 0.276 0.098
0.50 < Sj < 0.75 0.057 0.100 0.084 0.071 0.063 0.078 0.152 0.146
0.75 < Sj < 1.00 0.336 0.088 0.273 0.129 0.345 0.104 0.262 0.085
1.00 < ~ < 1.25 0.050 0.213 0.014 0.329 0.014 0.325 0.028 0.341
1.25 < ~ < 1.50 0.014 0.113 0.007 0.094 0.014 0.078 0.000 0.122
1.50 < ~ < 1.75 0.000 0.063 0.000 0.012 0.000 0.091 0.000 0.037
1.75 < Sj < 2.00 0.000 0.025 0.000 0.024 0.000 0.013 0.000 0.061
2.00 < Sj < 2.25 0.007 0.038 0.007 0.012 0.007 0.000 0.000 0.024
2.25 < Sj < 2.50 0.000 0.025 0.000 0.035 0.014 0.013 0.000 0.012
2.50 < Sj < 2.75 0.000 0.025 0.007 0.012 0.000 0.026 0.000 0.012
2.75 < ~ < 3.00 0.007 0.013 0.000 0.012 0.000 0.026 0.000 0.000
3.00 < Sj < 3.25 0.000 0.000 0.007 0.012 0.000 0.000 0.000 0.000
3.25 < Sj < 3.50 0.000 0.013 0.000 0.000 0.000 0.013 0.000 0.012
3.50 < Sj < 3.75 0.000 0.013 0.007 0.000 0.000 0.026 0.000 0.012
3.75 < Sj < 4.00 0.000 0.000 0.000 0.024 0.014 0.000 0.000 0.000
4.00 < Sj < 4.25 0.000 0.013 0.000 0.012 0.000 0.013 0.000 0.024
4.25 < ~ < 4.50 0.000 0.013 0.000 0.024 0.000 0.000 0.000 0.000
4.50 < ~ < 4.75 0.000 0.013 0.000 0.024 0.000 0.000 0.000 0.012
4.75 < Sj < 5.00 0.000 0.013 0.000 0.000 0.000 0.013 0.000 0.000
0< Sj < 5.000 1.000 1.000 1.000 1.000 1.000 1.000 1.000 1.000
......
S Total counts for the four parameter combinations are 140,80, 143, and 85 respectively in model (1) and 142,77,145, and 82 0'1
W
respectively in model (2).
164
WHERE DO WE STAND?
.....
0\
VI
~
~
~
take that value which separates evenly all the estimates, one half being
less and the other being more than it. 6
The variation between estimates of different studies can be due to
random errors, measurement errors, differences in the range of varia-
tion of the explanatory variables, etc. The practical significance of our
analysis is that if the variation between the estimates is only due to
random disturbances then the median of the values can be taken as
the cumulated evidence provided by all the estimates. But to the
extent that the variation is not entirely due to random disturbances
one must correct for the variation attributable to other factor's before
using the median estimator. Special tools of meta analysis, such as
those developed by Hunter et al. for a simple correlation coefficient
[41, are needed for this purpose.
NOTES
1. In Thursby's article [12] (p. 297) any ~ to exceed 30 from the KG regression
technique is treated as equaling 30.
2. Equation (1) is fitted in logs, and both equations are fitted subject to parametric
transformations prompted by the inability of the regression program to impose
the parametric inequalities implicit in forms (1) and (2). Those transformations,
which include the replacement of p by (1 - 0)/0, enable the a estimates and
their standard errors to be calculated directly by the program. The numerical
method of estimation is that of Marquardt [10). More particulars concerning the
experimental design can be found in [7) (pp. 258-260) [8) (pp. 563-564).
3. For the record, the ESSQ surface remains flat when S exceeds one regardless of
whether 0= 0.35 or 1.35 [81 (p. 566).
4. A bias toward one is virtually a tradition in elasticity estimation. Griliches [3)
(pp. 286-297), for instance, discussed it while Thursby [121 (p. 299)
demonstrated it at the start of the eighties in the context of nonlinear models.
5. For a detailed treatment of point estimation of location parameter under a
variety of distributional assumptions the reader may see Lehman [9).
6. There are other variants. For example one may take the mean value of all
estimators found to be significant. The reader may refer to Hunter et af. [4] for
some interesting illustrations of cumulating evidence across studies.
REFERENCES
JATI K. SENGUPTA
University ofCalifornia, Santa Barbara, California, u.S.A.
1. INTRODUCTION
This condition (2.2) holds if and only if Yj = F(xj). The function F(x)
represents a nonparametric production frontier such that it is every-
where not greater than any other non-decreasing concave function.
The observed data set D = (Dj : j = 1, 2, .. , , n) is said to be
consistent with the frontier function F(x) if it satisfies the conditions
(2.2). The data consistency problem thus reduces to solving the linear
programming (LP) problem:
(2.3)
where aj = x/Yj is the input coefficient vector for each j and x is any
one of the n input vectors. However the observed data set D may fail
this consistency test in at least two ways. One arises when the non-
negativity conditions are dropped e.g. the inputs and outputs are
allowed to vary over negative and positive domains such as dummy
variables or, the framework of statistical designs. A second situation
occurs when the data set is subject to a stochastic generating mecha-
nism. In this case, a part of the data set D may satisfy the consistency
requirement with a probability p, while the remaining part may fail the
consistency test i.e. may allow no feasible solution to the LP problem
(2.3). If the above probability p is very low (e.g. less than 0.25), then
the frontier function F(x) would have a very low probability of
realization. We may thus define as follows.
where
Sp = {p: p = min uj and (2.5) holds}
1 ~j" n
(2.4) i.e. if for any j we get f3*'xj > Yj' 13* ~ 0, then the unit j is not
efficient i.e. j f£ D. It is thus clear that the statistical divergence of the
regression model and nonparametric model can be tested in terms of
the contamination of the efficiency subset D. Third, the probabilistic
nature of the consistency (or efficiency) hypothesis implies that the ad
hoc procedure of replacing the random variables by their mean values
and solving for the LP model at the mean may not have a very high
probability of realization in many situations. For example, consider the
LPmodel
Min[f3l + 132: af3l + 132 ~ 7, bf3l + 132 ~ 4,131' 132 ~ 01
where a, b are uniformly distributed as {I ~ a ~ 4} and {1/3 ~ b
~ I}. On using the mean values of a and b in the LP model above
we get f3j = 18/11, f3i = 32/11. If we compute the probability of the
event that this optimal solution is feasible with respect to the original
problem, then we get
Prob{(a, b): af3j + f3i ~ 7, bf3j + f3i ~ 4 and f3j, f3i ~ o}
=Prob{(a,b):a ~ 5!2,b ~ 2/3}=0.25.
Thus this optimal solution (f3j, f3i) is infeasible with a probability as
high as 0.75. It is clear that in this case the optimal solution at the
mean is more often infeasible than not. This type of criticism applies
equally well to the LP method of efficiency estimation by Timmer
(1971) who used the mean inputs in the objective function as:
G(x)=min[,8'x:,8'aj ~ 1,,8 ~ 0,j=I,2, ... ,n] (2.7)
where x = (Xi) is the mean input vector. On the basis of these
definitions above one can state some theoretical results.
Max Y = L Yj (2.8a)
n
sub,iect
J 1)) ~ Vi' i = 1, 2
to \L a··y (2.8b)
j=1
(2.8c)
where Y is aggregate output and ~, Jt; are the two current inputs for
the industry as a whole where j = 1, 2, ... , n refers to plants (or
units) with a capacity of Yj assumed given. The necessary first order
conditions are:
1- L f3iaij
~
<: ° when
Yj=Yj
Yj E [0, Yj)'
Yj =0
(2.9)
The variables f31' f32 are the shadow prices of the two current inputs,
the optimal values of which denote the marginal productivities of the
inputs in the industry function. On taking the most realistic situation
of the three cases in (2.9), we obtain the dual LP model for deter-
mining the industry production function
Min C = L f3i ~ subject to f3 E R(f3) (2. lOa)
where
(2. lOb)
Y= ff G(a)
[(aj, az) da j daz = g({3j' {32)
(2.11)
where the functions g({3j' {32), h j({3j, {32), and ~({3j, {32) represent
output and the two inputs corresponding to any given set of feasible
values of {3j and {32 belonging to the utilization region G(a j , az).
Assuming invertibility and other regularity conditions we may solve
for {3j and {32 from (2.11):
/3j = h~l(V;, liz), /32 = hzl(V(, liz)
and substituting these values in the first equation of (2.11) we obtain
the macro production function:
(2.12)
Two remarks may be made about this derivation. One is that there are
two types of aggregation problem as pointed out by Seierstad (1985),
one given by the capacity distribution [( aj , az) within the utilization
region G(·) as above and the other by the empirically given distribu-
tion of only those micro units which are identified by some optimal
solution of LP model (2.10a). The first type of aggregation implies that
the capacities of all micro units with the same input coefficients can be
added together for computing the total capacity of the industry. The
second one constructs a core subset of efficiency units from the
optimal solutions of the LP models (2.10) when j belongs to a subset
177
subject to (2.15)
For a = 0.50 this optimal vector f3* = f3*0.5 reduces to the linear
efficiency measure.
Proo/Since
we obtain
g = v'f3 + qJf3'~f3
Yj =/3'xj -Ej ; Ej ~ 0, j E In
for a subset 15 of the data set and examine how robust is the optimal
efficiency vector f3* in (2.22). As an empirical application we refer to
the educational production function studies for California reported
elsewhere (Sengupta, 1987b) for 25 public school districts with four
inputs, where two tests were made. In one the sample size n was
varied as n = 9, 12, 20, 25 and the mean LP model (2.22) was run.
The optimal values f3*' = (0.247, 0.523, 0.626, 0.179) were the same
in each case. In the second case the sample set of n = 25 units was
first arranged in a decreasing order of efficiency, as measured by the
optimal solution of the LP model (2.22) with n = 25 and then the
sample set is divided into six groups with 4 in each group except the
last group which contained 5 units. Thus the first group contained the
100% efficiency units in the sense ~1=1 f3"tXij = Yj' the second group
contained the next four efficient (but less than 100% efficient) units
and so on. For each of the first three groups the optimal efficiency
vector /3* turned out to be the same as before i.e. /3*' = (0.247,
0.523, 0.626, 0.179). This shows the relative insensitivity of the
optimal efficiency vector /3 * for such data variations. Since the
objective function in (2.22) may be interpreted as the criterion of
minimizing the sum of absolute deviations of errors Le. Min ~ j I Cj I =
Min[~j ~i {f3;X;j - Yj}] = n Min ~; f3i X; since ~i f3iXij ~ Yj' the
estimate /3 *'x is in the form of a generalized median and hence less
sensitive to outlying observations.
• =
C
~ for r = 0,1, ... , k - 1
r n(tr+ 1 - tr)
186
and hence
~'fort, ~ t < t'+1>r=O,1, ... ,k-1
lH(t) =
{°
C,-l for tk = b
otherwise.
(3.1)
00 and hn -+ °
spacing number hn defined by 2hn = t,+l,n - t'on is such that if n -+
then nhn -+ 00. For most applied situations these
regularity conditions would hold and hence the sample histogram
estimate lH of the population density f(l) in its histogram form fH(I)
can be easily applied. Also by using the multinomial distribution the
likelihood function L(~, nl , .... , n,-ll Co, cI , . . . , C,-l) = L can be
written as
k-l
L = n (c,) (3.3)
,=0
where
Yo if X =X~)
YI if X = X~)
E(yj x) = (3.7)
Ok = 11 if X = X~) (3.8)
o otherwise
then the conditional regression model (3.7) can be rewritten in a linear
form
(3.9)
Since this is in the form of a regression model, the importance of the
189
modal efficiency set xtO) may be directly tested from the explanatory
power (i.e. R 2 value) of the regression equation
y = YoOo + constant + error.
Thus if R2 is low (high), the modal efficiency set is less (more)
important in the estimation of the production frontier.
NQw we consider the divergence of the two density functions JH( e)
and tH(e) measured for eX3:.JTlple by (3.4). Consider a data set D for
which the density function tH(e) corresponds to the case of minimum
absolute derivations (MAD):
Hence for any subset 15 for which (3.9) holds, we would have
robustness of the type associated with the MAD solution. The close-
ness of g! and g* can also be measured by the rate at which the
divergence measure d 12 defined in (3.4) tends to zero in the limit.
4. GENERAL IMPLICATIONS
their mode and then compare the closeness of these two densities in
terms of this location measure. If the underlying distributions are
normal (or tend to normality by the conditions of the central limit
theorem), then this will amount to a comparison of their means and
hence the usual statistical tests for the difference of means can be
performed. As an empirical example consider the educational produc-
tion function with four inputs reported by Sengupta (1987b) where
only nondegenerate optimal LP solutions (i.e., 11 out of 25) are taken:
LPs
(unit k) Pi P! P! In Remark
Clearly if we exclude the two outliers (i.e. units 12 and 23), the mean
of grs is very close to the MAD solution g* which is identical with
the LP model for unit 5.
A second implication of the conditional regression model (3.7) is
that one can apply the theory of optimal statistical design in this
framework. For instance, write the optimal basis equations as
X*(k)f3*(k) = y(k)
where B*(k) is denoted by the square matrix X*( k) of o!der m.
Assuming nondegeneracy o( the optimal basis we obtain f3*(k) =
X*-ly(k), where the vector f3*(k) has nonnegative elements. We now
assume that the output vector y( k) is random as follows:
y(k) = JA,(k) + u(k)
where u( k) is the error vector assumed to be normally distributed
with zero mean and a variance-covariance matrix V fixed for all k =
191
to obtain the best estimator P(O) say. But since V js not estimable due
to lack of degrees of freedom, this estimator f3(0) is obtained by
maximizing the determinant of [X*(kl'X*(k)J. As Chernoff (1972)
has shown the design associated with f3(0), also called the D-optimal
design is equivalent under certain regularity and continuity conditions
to the so-called A-optimal design based on the Chebyshev criterion
which minimizes the maximum variance of prediction among the
linear unbiased estimators of {3*( k).
As an example consider three DMUs each with one unit of output
and the following two inputs:
2 4 3 3
2 1 2 1.6
192
DMU's
Thus the best LP estimator !J(O) is given by the vector (:~~) where the
two units DMU I and DMU2 are in the optimal basis. It turns out that
in this case the optimal basis X(O) associated with !J(O) turns out to be
the modal efficiency set defined in Section 2; this is also identical to
the single LP model associated with MAD.
We now summarize our main results.
1. The data consistency problem in nonparametric theory of the
production frontier has to allow for stochastic feasibility, since
some input output data may fail to maintain feasibility with a high
probability.
2. The stochastic nature of the data aggregation problem when there
are input constraints leads us to consider a nonparametric theory
of distribution of efficiency. According to this theory the distribu-
tion of efficiency may be analyzed in relation to that associated
with the mean absolute deviation (MAD) approach.
3. Under certain general conditions the MAD approach has some
robustness properties, which can be utilized to characterize suitable
subsets (called the efficiency subset) of the overall data set. Such
characterizations are useful in running regressions with (0, 1)
dummy variable regressors and testing for the contamination of the
efficiency subset.
REFERENCES