Harvard University Press

Chapter Title: Jevons on the King-Davenant Law of Demand

Book Title: Statistics on the Table

Book Subtitle: The History of Statistical Concepts and Methods
Book Author(s): Stephen M. Stigler
Published by: Harvard University Press. (1999)
Stable URL: https://www.jstor.org/stable/j.ctv1pdrpsj.8

JSTOR is a not-for-profit service that helps scholars, researchers, and students discover, use, and build upon a wide
range of content in a trusted digital archive. We use information technology and tools to increase productivity and
facilitate new forms of scholarship. For more information about JSTOR, please contact support@jstor.org.

Your use of the JSTOR archive indicates your acceptance of the Terms & Conditions of Use, available at
https://about.jstor.org/terms

Harvard University Press is collaborating with JSTOR to digitize, preserve and extend access
to Statistics on the Table

This content downloaded from 37.166.22.103 on Mon, 08 May 2023 08:04:12 +00:00
All use subject to https://about.jstor.org/terms
 c ha pter 4

J
Jevons on the King-Davenant
Law of Demand

In studying the history of statistics, sometimes a question of conceptual

understanding can be illuminated by a seemingly trivial computation.
When writing in general terms, a scientist may be subject to a variety
of interpretations, or seem to endorse a variety of methods of proce-
dure. Yet a close numerical reworking of a statistical application by that
same scientist may be revealing. A surprising twist may appear, or an
unexpected limitation in scope of analysis may become apparent. Inves-
tigations into the exact numerical procedures a scientist followed can
be frustratingly difficult, since the details are frequently absent, or er-
rors of calculation or the typographer can confuse, leaving the historian
to attempt speculative reconstruction. But when the exact results can
be derived by a plausible procedure, by convincingly reconstructing in
minute numerical detail what the scientist actually did, there can be
an exhilarating sense of discovery and clearer understanding. For ex-
ample, by reenacting Mayer’s study of the geography of the Moon, or
Laplace’s numerical investigation of the lunar tides of the atmosphere,
or Quetelet’s calculation of the propensities for crime, we gain insight
into their conceptual understanding of statistics that could scarcely have
been achieved otherwise (Stigler, 1986a). Jevons and the King-Davenant
law of demand afford another example.
In his 1871 Theory of Political Economy, William Stanley Jevons
presented a short schedule relating the price of corn to different har-
vest sizes, a schedule published by Davenant in 1699 and sometimes
attributed to Gregory King (Jevons, 1871, pp. 148–153).1 The sched-

1. For the history of demand theory generally, and Davenant’s schedule in particular,
see George J. Stigler (1954). The King-Davenant schedule is discussed in detail by Creedy

80

This content downloaded from 37.166.22.103 on Mon, 08 May 2023 08:04:12 +00:00
All use subject to https://about.jstor.org/terms
 Jevons on the King-Davenant Law of Demand 81

ule was probably only loosely based on empirical evidence; William

Whewell (in 1850), Philip H. Wicksteed (in 1889), and G. Udny Yule
(in 1915) found, presumably independently, that the third difference of
the relative price is constant and the “data” can be exactly fit by a cubic.
The schedule, as presented by Jevons, is:
Quantity of corn Price
1.0 1.0
.9 1.3
.8 1.8
.7 2.6
.6 3.8
.5 5.5

Jevons fit this schedule by a function of the form

a
p=
(x − b)n
where we use p to denote relative price (relative to the price after an
average harvest), and Jevons used x to represent the size of the corn
harvest (as a fraction of the average harvest). Jevons’s full description of
his analysis was:
An inspection of the figures shows that n is probably about equal
to 2, and assuming it as such, I have found that the most probable
values of a and b are—

a = .824 b = .12.

The formula thus becomes

.824
price of corn = ,
(x − .12)2
5
or approximately, = .
6(x − 18 )2
The following numbers show the degree of approximation be-
tween the first of these formulae and the data of Davenant—

(1986a). Bostaph and Shieh (1987) have discussed this demand curve and examined its his-
torical context.

This content downloaded from 37.166.22.103 on Mon, 08 May 2023 08:04:12 +00:00
All use subject to https://about.jstor.org/terms
82 STAT IST I C S A N D S O C I A L S C I EN C E

Harvest 1.0 .9 .8 .7 .6 .5
Price (Davenant) 1.0 1.3 1.8 2.6 3.8 5.5
Price calculated 1.06 1.36 1.78 2.45 3.58 5.71

(Jevons, 1871, pp. 152–153)

Jevons presented no further explanation of his calculations, and John

Aldrich (1987) has speculated that Jevons had assumed n to be 2 and
used least squares to compute a and b, either by an iterative use of
a nonlinear fitting technique, or by transforming the data somehow
and using linear least squares. The importance of the point would be
that if Jevons did use least squares, this would stand as an (admittedly
lonely) counterexample to claims advanced in Chapter 3 that Jevons did
not use methods based on the calculus of probabilities in his empirical
work (as opposed, of course, to his appeals to probability concepts
in his logical and theoretical work, see, for example, Schabas, 1990,
chap. 4). Mary Morgan has also remarked on the scarcity of applications
of probability in Jevons’s empirical work (Morgan, 1990, p. 20). Did
Jevons or did he not ignore this central conceptual tool in his empirical
social investigations?
Aldrich’s suggestion that Jevons used nonlinear least squares is in-
triguing, but Aldrich was unable to reproduce Jevons’s results despite
several attempts, including weighted least squares and the application
of a Gauss-Newton algorithm to a nonlinear least squares formulation
(Aldrich, 1987, p. 251). Aldrich’s suggestion has been cited with appro-
bation (for example, White, 1989, p. 434), but no one has succeeded in
verifying it by producing Jevons’s results by a plausible (or even implau-
sible) use of least squares. Jinbang Kim has shown that Jevons’s results
produce a sum of squared deviations close to that of a nonlinear least
squares solution, but not how they could have been plausibly derived
via that route. Indeed, Kim has found that Jevons’s results are not close
to any of the several least squares solutions he tried (Kim, 1991).
I will show how Jevons could have derived the results using a simple
ad hoc method. The method does not entail appeal to least squares or
any other probability-based methodology (indeed, the components of
the hypothesized method are as ancient as any used in numerical work).
Consider, then, the problem of fitting the equation
a
p=
(x − b)2

This content downloaded from 37.166.22.103 on Mon, 08 May 2023 08:04:12 +00:00
All use subject to https://about.jstor.org/terms
 Jevons on the King-Davenant Law of Demand 83

to the six pairs of values for p and x. (I follow others in accepting the
hypothesis that the exponent “2” was derived either informally, as from
a graph, or perhaps from a suggestion of Whewell, made in 1862; see,
for example, Creedy 1986a). Begin by rewriting the equation as
p(x − b)2 = a.
Start by finding b. If we subtract two of these equations for different
pairs of (x, p) we will eliminate a and obtain an equation in b. As astute
a numerical scientist as Jevons would surely have known that a solution
based on a selection of two out of six pairs (x, p) would be most stable
(least susceptible to minor variations in the numbers) if the two extreme
pairs are selected. Using the pairs (1.0, 1.0) and (.5, 5.5) we then have
the following equation for b:
1.0(1.0 − b)2 = 5.5(.5 − b)2
or (taking square roots and solving)
r
c(.5) − 1.0 5.5
b= where c = = 2.3,
c − 1.0 1.0
or
b = .12,
exactly Jevons’s value for b. Now to find a, compute
p(x − .12)2 = a
for each of the six pairs and take the simple arithmetic mean. The result
to three places is
a = .824,
exactly Jevons’s value for a.
Thus using simple numerical techniques that involve nothing more
complicated than taking a single square root of a two-digit number and
a simple average of six values, we have recovered Jevons’s results with
an exactness that has thus far eluded more complicated attempts. Both
aspects of this numerical approach—the first akin to pinning down a
line’s slope by focusing on the extremes, and the second the use of simple
averages—were commonplace in the handling of astronomical observa-
tions as far back as the seventeenth century. It is of course speculative to

This content downloaded from 37.166.22.103 on Mon, 08 May 2023 08:04:12 +00:00
All use subject to https://about.jstor.org/terms
84 STAT IST I C S A N D S O C I A L S C I EN C E

claim that this is the procedure that Jevons actually used, but speculation
that agrees exactly with observation to the number of decimal places re-
ported is preferable to speculation that does not. Hence it appears that,
rather than nonlinear least squares, Jevons used nonleast squares, and
Jevons’s discussion of the King-Davenant schedule in his Theory of Po-
litical Economy is not a likely counterexample to the claimed absence of
probability-based methods by Jevons in the analysis of social data.
The matter of what algorithm Jevons used has bearing on important
issues. It would not have been beyond arithmetic feasibility for Jevons
to have used some variant of least squares in 1871. Legendre and Gauss
had used nonlinear least squares half a century earlier in astronomy
and geodesy, and Jevons would have learned of the technique from
his teacher Augustus De Morgan, who had written extensively about
the topic. Indeed, in 1874 Jevons included a tutorial on the simpler
applications of least squares in the chapter on “The Law of Error” in
his Principles of Science, with copious reference to the wider literature
on the topic. But to find him using the method with social data in 1871
would have been anomalous: it would require either that Jevons was
naively expropriating a method from astronomy and geodesy to use in
a setting where the conceptual basis for the earlier use of the method—
additive errors superimposed on a theoretically derived model—did not
apply, or alternatively that Jevons had succeeded in developing a new
conceptual basis for the use of least squares with social data.
Now, the naive use of inappropriate methodology is hardly unknown
in empirical social science, but Jevons was not naive—to the contrary,
he was extraordinarily conscious of methodologies and their bases. In
particular, he was acutely aware of the myriad of factors that made the
specific determination of the laws of supply and demand infeasible. In
his 1874 Principles of Science he wrote to this very point:

No one will be found to deny that there are certain uniformities of

thinking and acting which can be detected in reasoning beings, and
so far as we detect such laws we successfully apply scientific method.
But those who attempt thus to establish social or moral sciences,
soon become aware that they are dealing with subjects of enormous
complexity. Take, for instance, the science of Political Economy. If a
science at all, it must be a mathematical science, because it deals with
quantities of commodities. But so soon as we attempt to draw out the

This content downloaded from 37.166.22.103 on Mon, 08 May 2023 08:04:12 +00:00
All use subject to https://about.jstor.org/terms
 Jevons on the King-Davenant Law of Demand 85

equations expressing the laws of variation of demand and supply, we

discover that they must have a complexity entirely surpassing our
powers of mathematical treatment. We may lay down the general
form of the equations, expressing the demand and supply for two
or three commodities among two or three trading bodies, but all the
functions involved are of so complicated a character that there is
not much fear of scientific method making a rapid progress in this
direction. (Jevons, 1874a, 2:457–458)

Elsewhere that same year he noted, “The laws of supply and demand are
the best established part of the prevailing doctrines, although they have
never been defined or analysed with any numerical precision” (Jevons,
1874b, p. 85).
Indeed, I have argued extensively (Stigler, 1986a, chaps. 7–9) that the
development of a conceptual basis for the use of least squares in a setting
such as the King-Davenant schedule only occurred toward the end of
that century, with the development of regression and correlation—really
multivariate analysis—in the works of Francis Galton, Francis Edge-
worth, and G. Udny Yule. Early quantitatively inclined workers in social
science, as in physical science, would only apply probability-based meth-
ods when they perceived the data in terms such that a probability frame-
work was plausible: homogeneous except for the factor under study
and accidental variation. The Belgian statistician Adolphe Quetelet, as
a young and eager adventurer, had been willing to throw caution to the
winds and had proposed the determination of the population of Belgium
from a sample, using a probability-based method equivalent to ratio esti-
mation. His enthusiasm lasted until an older and wiser official, the Baron
de Keverberg, pointed to a vast array of important inhomogeneities in
birth, death, and marriage rates to which Quetelet had been oblivious,
and lifted the scales from Quetelet’s eyes. Where Quetelet had seen a
probability framework before, he suddenly saw none. A major portion
of Quetelet’s subsequent work may be viewed as attempting to overcome
this difficulty by cataloguing all manner of social influences and their
magnitudes.
Jevons, educated at University College and seasoned in the Aus-
tralian gold fields, was not so innocent as the young Quetelet. Even if, as
seems unlikely, Jevons had anticipated the later development of regres-
sion and correlation to some degree, it stretches credulity to suppose that

This content downloaded from 37.166.22.103 on Mon, 08 May 2023 08:04:12 +00:00
All use subject to https://about.jstor.org/terms
86 STAT IST I C S A N D S O C I A L S C I EN C E

he would apply the method to this questionable a set of numbers, a set

that no writer since Davenant seems to have taken literally as “data.”
Indeed, Jevons described the schedule as “a certain celebrated estimate
of the variation of the price of corn” (Jevons, 1871, p. 148), and em-
phasized the agreement he found between his “price calculated” and
Davenant’s estimates. As Thomas Kuhn has argued in a different con-
text (Kuhn, 1961), comparisons such as the one Jevons presented of the
fit of his curve to the Davenant-King “data” might best be interpreted
as defining the degree of fit to be expected from a theory, not as a test
of the theory, and for such a purpose, a probability-based conceptual
apparatus was not required.
This example, then, helps fill out a picture of a pioneering social
scientist at a crucial juncture in history. Jevons was too alert to the
subtle difficulties in applying probability-based methods to social data
to proceed heedlessly. This is not to claim that he had an explicit sense
of the precise nature of the conceptual barriers he faced; had that been
the case we could expect that he would have articulated the problem and
at least attempted to address it. But vaguely sensed dangers can be more
troubling than well-understood ones, and Jevons drew back, here and
in other situations where he analyzed social data statistically. This did
not paralyze him into inaction—he did produce a quick ad hoc fit, but
only to gain a rough sense of the sort of discrepancy one might expect
between a hypothesized theoretical curve and data. Such exploration
was to be a part of the process that would eventually define and at least
partially overcome the barriers.

This content downloaded from 37.166.22.103 on Mon, 08 May 2023 08:04:12 +00:00
All use subject to https://about.jstor.org/terms