The Scientific Status of the Labour Theory of Value

W. Paul Cockshott and Allin Cottrell∗

March 25, 1997

Contents prior data. For instance, the australopithecine long

predated Darwin, but its existence—as the “miss-
1 General criteria of scientificity 1 ing link”—was only subsequently deduced on the
basis of the theory of evolution. Such predictions
2 Three formulae 2 involve applying some general law to pre-given data
to produce derived data whose validity can then be
3 Some practical concerns 3 tested. The derived data need not be in the future,
or even unknown, provided that they are not made
4 Comparison of information gain 5 available as input data.
5 Non-equalization of profit rates 8
Elegance or simplicity
In this paper we wish to argue that the labour A second criterion is elegance, captured in William
theory of value is a scientific theory in the strongest of Occam’s dictum that entities should not be multi-
sense of the empirical sciences. We first elaborate plied without cause. Given two theories, predicting
upon what we take to be criteria of scientificity, and available data equally well, science opts for the sim-
then show that these are in practice met by the pler. Ptolemy plus epicycles might run Newton a
labour theory of value. close race, had not simplicity favored the latter.
This favoring of the simple echoes the demands
of prediction. The better prediction extends our
1 General criteria of scientificity
knowledge furthest on the basis of the least pre-
Criterion of testability given information. A theory with added epicycles,
special cases and fudge factors may predict obser-
Scientific theories tell us something about the ma- vation well. . . since the observations are woven into
terial world. They are means by which we can both its very weft.
explain and predict aspects of reality. They are not
the only form of explanation: myth, story telling Information gain
and religion also explain things, but with more lim-
ited predictive power. A scientific law compresses many, even a potential
A theory which makes no predictions is never sci- infinity of, observations into a simple formula. It
entific, whilst one whose predictions have now been is a machine for gaining information. In this guise
invalidated has forfeited any prior claims to science. we revisit prediction and simplicity. The informa-
A scientific theory makes testable predictions. Note tion gained from a law is given by the information
that by “prediction” we don’t just mean forecasting yielded from the law’s application less the informa-
the future, though this can be part of it. An histor- tion put in. We can put in information either in
ical science can formulate temporal “postdictions” the data-set to which the law is applied, or in the
as logical predictions. Scientific predictions make formula in which the law is encoded.1 Although the
statements about the world that are not available as encoding of the formula should, in principle, be in-
cluded in any measure of the information cost of a
∗ Department of Computer Science, University of Strath-

clyde, and Department of Economics, Wake Forest Univer- 1 Strictly the formula and input data should be expressed

sity, respectively. This paper was prepared for the fourth in a form suitable for interpretation by the Universal Turing
mini-conference on value theory at the Eastern Economic As- Machine. A good exposition of the concept of the Universal
sociation meetings, April 3–6, 1997. Turing Machine is provided by Penrose (1989).

1
theory, for the value theories that we will consider 2 Three formulae
the formula lengths are essentially the same, though
the input data lengths are not. This section deals with the size of the formula
Let us refer to the application of a law as the needed to construct three variant predictors of
function application p = L(d), where L stands for prices.
the law, d for the data to which it is applied, and
p the predictions produced. For a law to be valid
the information content of p must exceed the sum Vertically integrated labour coefficients
of the information content in L and in d. The first These are obtained (in principle—but see section 3
criterion said that the predictions must not be in- for a discussion of practical issues) as the n-vector
cluded in the input data; they must be new infor- v that solves
mation, p must not just be a repetition of d. The
theory may predict the input data but it must pre-
dict more besides. Our second criterion, simplicity, v = Av + λ
ensures that the law itself does not simply tabu-
late its predictions. The information gain criterion where A is an n × n matrix of technical coefficients
incorporates both criteria in a stronger form: the and λ is an n-vector of direct labour-hours per phys-
sum of the information in the input data and the ical unit of output. The required data are then the
encoding of the law must be less than the predic- n2 technical coefficients and the n direct-labour co-
tions produced. Only then is there real information efficients. The solution vector is typically obtained
gain (Solomonoff, 1964). using an iteration

Randomness v(0) = λ
Random processes exhibit no lawful behavior, and v(n + 1) = Av(n) + λ
thus cannot be predicted. Information theory states
that a sequence of numbers is random if there ex- which terminates when the largest element of the
ists no formula shorter than the sequence itself, ca- difference vector v(n + 1) − v(n) shrinks below some
pable of generating the sequence. A non-random set tolerance.
sequence, by contrast, has a shorter generator. 1, 1,
2, 3, 5, 8 is arguably random while 1, 1, 2, 3, 5, 8, 13,
21, 34, 55, 89, 144, 233, 377, 510, 887, 1397 is not, Sraffian prices of production
since the latter can be generated by the Fibonacci
formula These are obtained as the n-vector p that solves2

p = (1 + r)Ap + w
f (n) = f (n − 1) + f (n − 2); f (0) = 1; f (1) = 1
where the scalar r represents the general rate of
of shorter length. A scientific law captures the non- profit and w is an n-vector of (direct) wages paid per
random behavior of reality. If we view the law plus unit of output in each industry. The data require-
its input data as a formula, the formula must be ments here are essentially the same as for the ver-
shorter than the data its application produces. tically integrated labour coefficients, although one
Our object in this paper is to use this general per- more piece of information is required, namely the
spective to assess the relative merits of three vari- profit rate (or the real wage, from which the profit
ant formulae for predicting observed prices, namely rate can be inferred). The calculation proceeds in a
“standard” Marxian values (or vertically integrated similar manner, for instance via the iteration
labour coefficients), Sraffian prices of production,
and prices of production as interpreted by the “Tem-
poral Single System” (TSS) school (e.g. McGlone p(0) = w
and Kliman, 1996). We are aware that these theo- p(n + 1) = (1 + r)[Ap(n)] + w
ries may be conceived as serving discursive purposes
other than the prediction of actual prices, but we 2 If wages are taken to be paid in advance then the follow-

treat them here as predictive formulae. ing equation is respecified as p = (1 + r)(Ap + w).

2
TSS prices of production a vector of hours of direct labour performed in each
industry we have a vector of wage bills.
TSS prices of production are obtained as the sum
This raises a concern: if prices are written into
of the (predetermined) price of the inputs plus an
the data we are using to calculate labour values in
aliquot share of the aggregate profit. Writing u for
the first place, is there not something circular about
such prices we have
turning around and claiming to predict or explain
prices on the basis of these values? In fact this ob-
u = (1 + r)[Ap(−1) + w]
jection is mistaken, but since it has some force prima
where p(−1) denotes the given vector of input prices facie we shall address it in detail. It will be useful to
as of the beginning of the period and the other sym- distinguish two points: the use of a wage-bill vector
bols are as defined above. The data-requirement for as a proxy for a labour-hours vector, and the use of
calculating this vector is clearly greater than that monetary intersectoral flows (and output figures) in
for the vertically integrated labour coefficients and place of in-kind flows. The first of these issues does
Sraffian prices, since in addition to the A matrix and create a real problem (but not a very serious one,
the wage vector we also require a vector of given in- in our view), while the second does not.
put prices.
Wage bills versus labour hours
3 Some practical concerns
By using wage-bill data as a proxy for labour hours
Section 2 skated over some issues that have to be one is in effect computing a vector, not of vertically
faced when the above-mentioned formulae are ac- integrated labour coefficients as such, but of verti-
tually put to use. This section raises two such cally integrated wage-cost coefficients. If the wage
issues. The first, which can be dealt with quite were uniform across industries this would not mat-
briefly, concerns the distinction between fixed and ter at all, but the existence of inter-industry wage
circulating capital. The “textbook” presentation of differentials creates a complication. The question
the calculation of prices of production (whether in is, what is the relationship between such wage dif-
the Sraffian or the TSS variant) in effect assumes ferentials on the one hand, and intersectoral differ-
a pure circulating capital system, which suppresses ences in the “value-creating” power of labour on the
the distinction between capital stocks (on which, other? Here are two polar possibilities:
presumably, the rate of profit “ought” to be equal-
ized) and the flow consumption of capital (as mea- 1. Intersectoral wage differentials have nothing to
sured in the “A” matrix of an input–output table). do with differential value-creation: they are
When this simplifying assumption is dropped, it be- an arbitrary outcome of market or other social
comes apparent that the price of production calcu- forces. In this case, clearly, the “values” calcu-
lation requires additional data—namely the stocks lated using the wage-bill proxy will be inaccu-
of capital in each industry—while the vertically in- rate as a representation of “true” values. Fur-
tegrated labour coefficients emerge as distinctly the ther, it is likely that the figures thus obtained
most parsimonious predictor. will be better correlated with market prices
The second question demands a fuller discussion than the unknown “true” values—since wage-
since it goes to the heart of the project of predicting cost is presumably relatively closely related to
prices using empirical labour values. In the stan- actual capitalist pricing practice—leading to an
dard presentation of the principle of calculation of over-statement of the predictive power of the
Marxian values one starts from data on (a) inter- labour theory of value.
sectoral flows of product in natura and (b) direct
labour-hours performed in each sector of the econ- 2. Intersectoral wage differentials are an accurate
omy. Given the statistics that are available for cap- reflection of differential value-creating power.
italist economies, however, one has no choice but Wage differentials reflect the cost of training,
to use data that are expressed in monetary terms. while relatively highly trained labour creates
Instead of in-kind figures for, say, the quantity of more value per clock hour—or more precisely,
coal used by the steel industry or the quantity of transfers to the product the labour-content of
aluminium used by the aircraft industry, we have the training alongside the actual creation of
figures for the money value of the purchases of each new value. In this case the industry wage-bill
industry from each other industry. And instead of figures are actually a better approximation to

3
 what one is trying, theoretically, to measure Proof
than simple direct clock hours would be.
Consider an economy characterized by the following
arrays:
As we have said elsewhere (e.g. Cockshott, Cot-
trell and Michaelson, 1995), surely the truth lies U An n × n matrix of intersectoral product
somewhere between these poles. Intersectoral wage flows in kind, such that uij represents the
differentials will in part reflect “genuine” differences amount of industry j’s output used as in-
in value productivity, and partly reflect extraneous put in industry i.
factors. In any case, if the input–output structure of q An n × 1 vector of gross outputs of the
the economy exhibits fairly strong interdependence industries, in their natural units.
then the vertically integrated wage-cost coefficients l An n×1 vector of direct labour-hours per-
for any given sector will comprise a broad mix of formed in each industry.
labour from different sectors so that the effects of
the extraneous factors will tend to cancel out.
It will be useful also to define an n × n diagonal
matrix Q such that
Product flows: Quantities versus monetary magni- (
tudes qi for i = j
Qij =
0 for i 6= j
The necessity of working with monetary magnitudes
rather than in-kind product flows is in part a result The standard calculation of labour-values pro-
of the degree of aggregation of the actually available ceeds as follows. First calculate the n × n matrix of
input–output tables for capitalist economies. That technical coefficients as A = Q−1 U and the n-vector
is, in order to construct a meaningful input-output of direct labour input per unit of physical output as
table in natura it is necessary that the data be fully λ = Q−1 l. The n-vector of unit values (vertically
disaggregated by product, but many of the indus- integrated labour coefficients) is then given by
tries as defined in the actual tables produce a wide
range of different products. There can be no mean- v = (I − Q−1 U )−1 Q−1 l = (I − A)−1 λ
ingful number for the quantity of output of “Aircraft and the n-vector of aggregate values of the sectoral
and Parts” or “Electronic Components and Acces- outputs is
sories”, or for the in-kind flow of the product of the
latter industry into the former. The practical solu- V = Qv = Q(I − A)−1 λ (1)
tion is to present the aggregate monetary values of
flows of this sort. We now construct the monetary counterpart to
the above arrays. Let the n-vector p represent the
But this does not create a problem, if one is in-
prices of the commodities and the scalar w denote
terested in comparing the aggregate monetary value
the (common) money wage rate.3 Let us also define
of the output of the industries with the aggregate
an n × n diagonal matrix P such that
labour-value of those same outputs. The point is (
this: The vector of aggregate sectoral labour val- pi for i = j
ues calculated from a monetary table will agree with Pij =
0 for i 6= j
the vector calculated from a physical table, up to a
scalar, regardless of the price vector and the (com- Corresponding to each of the initial “real” arrays
mon) wage rate used in constructing the monetary above there is a monetary version as follows:
table. Or in other words, the vector of sectoral
labour values obtained is independent of the price Û = U P Matrix of money-values of inter-
vector used. One might just as well (if it were prac- sectoral product flows
tically possible) use an arbitrary vector of account- q̂ = P q Vector of money-values of gross
ing prices or weights to construct the “monetary” outputs
table. The fact that actual prices are used in the
published data does not in any way “contaminate” l̂ = wl Vector of industry wage-bills
the value figures one obtains; no spurious goodness 3 Having addressed the issue of intersectoral wage differ-

of fit between values and prices is induced. entials above, we abstract from it here.

4
From these we can construct counterparts to the de- Recall that (1) specified V = Q(I − A)−1 λ. Com-
rived “real” arrays. First the n × n diagonal matrix paring these two equations we see that V̂ = wV on
Q̂, whose diagonal elements are pi qi , is given by condition that
(
q̂i for i = j (I − A)−1 = P (I − P −1 AP )−1 P −1 (7)
Q̂ij = = QP (2)
0 for i 6= j That this condition is indeed satisfied may be seen
The counterpart to the matrix of technical coeffi- by taking inverses on both sides of (7). On the left,
cients is we simply get (I − A); on the right we get

 = Q̂−1 Û = (QP )−1 U P [P (I − P −1 AP )−1 P −1 ]−1

= P −1 Q−1 U P = P −1 AP (3) = P (I − P −1 AP )P −1
= (P − AP )P −1 = I − A
The elements of  represent the dollars’ worth of
input from sector j required to produce a dollar’s
worth of output in sector i. Finally, the counterpart This means we have proved that V̂ = wV ,
to λ is the n-vector λ̂ which is to say that the aggregate sectoral values
obtained from the monetary data agree—up to a
λ̂ = Q̂−1 ˆl = (QP )−1 wl scalar, namely w, the common money wage rate—
with those that would be obtained from the data
= wP −1 Q−1 l = P −1 wλ (4) in natura, if these were available. The aggregate
value vector is independent of the price vector used
whose elements represent the direct labour cost per in forming the monetary tables.
dollar’s worth of output in each sector.
Now here is the issue: suppose we are not privy
4 Comparison of information gain
to the information on product flows in kind and
labour-hours, and have at our disposal only the in- We can now return to our main theme, the assess-
formation given in the monetary tables. On this ment of the three versions of value theory in terms
basis we can calculate a vector v̂, of predictive power or information gain. As we have
shown, the labour theory of value has the simplest
v̂ = (I − Â)−1 λ̂ formula, the Sraffian price of production formula-
tion has a slightly more complex formula, and the
While vi represented the vertically integrated labour
TSS price of production theory has the most com-
hours per physical unit of output of commodity i,
plex formula (counting only the data input). A more
the v̂i that we are able to obtain from the monetary
complex formula can be justified if the increase in
tables represents the vertically integrated labour
prediction gained exceeds the additional formulaic
cost per dollar’s worth of output of commodity i.
complexity.
If we then multiply up by the money-value of the
The general equation for the information content
gross outputs of the industries we obtain the vector
or entropy of a sequence of symbols is
of vertically integrated labour costs for the indus-
tries. X
H =− πi log2 (πi )
i
V̂ = Q̂v̂ = Q̂(I − Â)−1 λ̂ (5)
We are interested in the relationship between (1), where H is the entropy of the sequence in bits, πi is
the aggregate sectoral values that could be obtained the probability of the ith symbol occuring, and we
in principle from the data in natura, and (5), the sum over the whole sequence of symbols.4 Let us
corresponding figures obtained by using the mone- identify our symbols as being the ratio p/e where p
tary data. is a market price and e is an estimator of market
On the basis of the correspondences (2), (3) and price, either a labour value or some form of price
(4) we can rewrite (5) as of production. Defining our symbols in this way we
4 For discussion of information and entropy see for instance

V̂ = QP (I − P −1 AP )−1 P −1 wλ (6) Reza (1961), Chaitin (1987).

5
obtain the conditional entropy of the market price for the same year.5 The BEA give figures for plant
vector P given the estimator vector E. Thus and equipment at a higher level of aggregation than
X that employed in the i/o table. We therefore had
H(P |E) = − πi log2 (πi ) to merge some rows and columns of the i/o ta-
i ble to ensure that each industry had a distinct fig-
ure provided for the value of plant and equipment.
where πi = Pr(p/e ∈ xi ), the probability that a The resulting table has 47 columns and 61 rows.
randomly selected unit of output has a p/e ratio The columns—which constitute a square submatrix
that falls within the ith small interval (the width of along with the first 47 rows—represent the aggre-
these intervals being determined by the precision of gated industry groups. The remaining rows consist
measurement). of:
In general the best estimator will be that which
gives the lowest conditional entropy of the market • Inputs for which there is no corresponding
price vector, since this means that most of the in- industry output listed such as “Educational
formation content of the latter vector is already en- and social services, and membership organiza-
coded in the estimator. Given the total information tions” or “Noncomparable imports” (a total of
content of the market price vector H(P ), we can 9 rows).
calculate the mutual information or “transinforma-
tion” (Reza, 1961) as I(P ; E) = H(P ) − H(P |E). • “Compensation of employees”, which we treat
This figure represents the reduction in uncertainty as equivalent to variable capital.
regarding P that is achieved by conditioning on the • “Other value added”, which we treat as being
predictor E; other things equal we would like this profit;
to be as large as possible, but we also have to factor
into our assessment the required information input. • “Finance”—we treat this as corresponding to
In this light the marginal predictive efficiency of a interest and include it in our measure of profit;
theory is given by
• “Real estate and royalties”; and
∆I(P ; E)
, • “Indirect business tax and nontax liability”.
∆I(input)
The last two items create some analytical prob-
the ratio of incremental transinformation to incre-
lems. At the aggregate level they are arguably part
mental information input, relative to a simpler the-
of surplus value, yet both indirect taxes and rents
ory. This marginal predictive efficiency gives a mea-
appear as costs from the point of view of the firms
sure of the elegance of the theory.
paying them, and insofar as there is any tendency
To perform these calculations we thus require (1) for equalization of the rate of profit, we would ex-
a figure for the total information content of the mar- pect this to be net rather than gross of indirect taxes
ket price vector, and (2) the probability distribution and rent payments. Pending a more satisfactory so-
for the ratio p/e. We estimate the information con- lution, we have in the present study simply netted
tent of the market price vector by making an as- these out of all our calculations (i.e. excluded them
sumption about the accuracy of the source data, from our measures of both costs and profits).
namely that figures are given to 3 digits or 9.96 bits The BEA figures are for fixed capital; we as-
of accuracy. The information content of the market sumed that in addition industries held stocks of
price vector would then be 47 industries each con- working capital amounting to one month’s prime
tributing 9.96 bits, or about 470 bits. The exact costs (where prime costs included wages in the TSS
number of digits assumed is relatively unimportant, case but not in the Sraffian case).
since a higher H(P ) will “benefit” (i.e. raise the It should be noted that modeling capital stocks is
value of H(P ; E) for) all the methods of estimation the logical dual of modeling turnover times. We are
equally. assuming that for the aggregate capital, turnover of
circulating capital is one month. This assumption is
The data 5 Specifically, the stock data came from files wealth14–

wealth17 on diskette 4 of the BEA’s series of diskettes for

Our data were drawn from the 1987 US input– Fixed Reproducible Tangible Wealth in the United States,
output table along with BEA capital stocks figures 1925–1989.

6
based upon the heroic simplification that there ex- of social labour time accounted for by each sector
ist 12 production periods per year corresponding to in the case of labour values): in this case the TSS
monthly salary payments, and that the total stocks prices are still ahead at a correlation of 0.989, but
of goods held in the production, wholesale and re- labour values (0.987) did better than Sraffian prices
tail chain amount to one month’s sales. That is to (0.985).
say, we assume that the turnover time of variable
capital is one month with wages paid in advance, Conditional entropies
and that circulating constant capital is purchased
simultaneously with labour. (In the calculation of For each industry we then computed the ratios of
Sraffian prices we assume wages are paid at the end the market price of output to each of the estima-
of the month.) A more sophisticated study would tors, giving the ratio vectors market price/value,
look at company accounts for firms in each sector to market price/TSS price and market price/Sraffian
build up a model of the actual stocks of working cap- price. The entropies of these vectors were computed
ital. Industries operating just-in-time production as follows.
will have considerably lower stocks and thus faster
turnover; for other industries one month’s stocks 1. A Gaussian convolution (with standard devia-
may be an underestimate. tion 0.08) was run over the observations, with
each observation weighted by the total money
output of the industry in question. This re-
Correlations sulted in a tabulation of the probability den-
The iterative procedures described earlier were used sity function relating prices to the various es-
to compute the total value of output, industry by timates. The theoretical basis for the convolu-
industry, using the labour value and Sraffian mod- tion is that each discrete observation is in fact
els. The TSS estimate of total values was derived in the mean figure for an entire industry: we as-
one pass without transforming the inputs. This gave sume that the different producers in the indus-
three estimates for the aggregate price vector; the try exhibit a normal distribution of their price–
correlations between these estimates and observed value ratios around this mean. A Gaussian con-
prices are shown in Table 1. volution substitutes for each individual obser-
vation a normal distribution having the same
mean and the same integral. The density func-
Table 1: Correlations between sectoral prices and tions are shown in Figure 1.
predictors, for 47 sectors of US industry
2. The entropy function H = −πi log2 πi was in-
tegrated over the range 0 to 3 with 1000 in-
Observed price tegration steps. Taking 1000 integration steps
Labour values 0.983 corresponds to a maximum possible value of the
integral of log2 1000 = 9.96, the assumed accu-
TSS prices 0.989
racy of the data in the tables. The interval [0,3]
Sraffian prices 0.983 was chosen for integration as the probability
densities for all estimators fall to zero outside
this range.
As can be seen, all the estimates are highly cor-
related with market prices, with the TSS estimates The resulting conditional entropies of market
performing marginally better than the other two. prices, with respect to each of the estimators, are
The reported correlations are unweighted; each in- shown in Table 2. The fourth row of the table shows
dustrial sector is simply counted as one observa- the total amount of information about prices for the
tion. We also calculated the correlations between 47 industries that each of the theories is able to
the logs of the variables—in line with Shaikh’s ar- produce. In terms of information output the TSS
gument (1984) for a multiplicative error term in the model outranks the Sraffa model which outranks the
price–value relationship. The ranking of the cor- labour value model.
relations remained unchanged (labour values 0.980, In terms of information input the ranking is the
TSS prices 0.986, Sraffian prices 0.980). And we other way round, as the Sraffa model requires an
calculated weighted correlations, the weights sup- additional 47-element vector of capital stocks, and
plied by the predictor variable (e.g. the proportion the TSS model requires a further 47-element vector

7
Figure 1: Probability density functions for observed price/estimator (E1 = labour values, E2 = TSS prices,
E3 = Sraffian prices)

0.01
E1
0.009 E2
E3
0.008
0.007
0.006
pdf(x) 0.005
0.004
0.003
0.002
0.001
0
0 0.5 1 1.5 2 2.5 3
x = p/e

of input prices. What is the additional information timate of the “production period”—the distinction
contained in each of these vectors? If we assume between the current and previous period becomes
them to be accurate to 3 figures, then each of these blurred. The figures used cover the months Jan-
vectors contains 468 bits. But it is probably unrea- uary to December 1987. Assuming a one-month
sonable to require that figures for prices and capi- advance of variable and circulating capital, the pe-
tal stocks be provided with this degree of accuracy. riod for which the input price data should, ideally,
Since the motivation of the TSS model, at least, is have been obtained is December 1986 to Novem-
to reproduce the Marxian technique for transform- ber 1987. Since these two periods are substantially
ing values into prices of production, we will assume overlapping, the actual price data used must be a
that the capital stock figures need only be as accu- very good approximation to the price data that we
rate as the output prices generated by labour val- should have used. Taking this into account, the
ues alone—roughly 1.9 bits per figure or 90 bits per TSS theory appears to have negative predictive effi-
vector. The efficiency with which this additional in- ciency. What we get out is an estimate of this price
formation is used in the Sraffian theory is low: for vector that is less accurate than the one we start
every one bit added to the input data about 0.15 out with. The information gain (actually, loss) here
bits of information is added to the output data. is thus H(Pt |Pt−1 ) − H(Et ) ≥ −354 bits. The con-
struction of a predictor using untransformed input
It should be noted that there is a significant dif- prices produces a net information gain only if the
ference between the additional information in the correlation between the price vector for 1.xii.86 to
two cases. In the TSS case, the extra information 30.xi.87 and that for 1.i.87 to 31.xii.87 is less than
required is about the variable being predicted by the correlation between the TSS estimator and the
the model, namely price. Ideally we should supply latter.
the prices lagged by one production period as in-
puts to the calculation. Using input–output table
data this is not possible so we in fact used current 5 Non-equalization of profit rates
relative market prices. Given that the price data in
the tables reflect the average prices prevailing over Most accounts of the theory of value entail the ex-
a year—considerably longer than any plausible es- pectation that Sraffian prices of production should

8
Table 2: Entropy calculations (P = observed price, E1 = labour values, E2 = TSS prices, E3 = Sraffian
prices)

(P |E1 ) (P |E2 ) (P |E3 )

Conditional entropy 8.028 7.538 7.734
Potential entropy 9.966 9.966 9.966
Transinformation, I(P ; E)
– per prediction 1.938 2.428 2.232
– for all 47 industries 91.09 114.12 104.90
Subtract what is already known 0 −468 0
about prices
Net prediction 91.09 −354 104.90
Additional information input 0 90 90
Information gain per additional – −3.93 0.153
bit input

Table 3: Profit rates and organic composition, BEA fixed capital plus one month’s circulating constant
capital as estimate of capital stock (C). Summary statistics weighted by denominator in each case.

s/C c/v s/v

Mean 0.292 1.948 0.569
Standard deviation 0.221 3.042 0.500
Coefficient of variation 0.756 1.562 0.878
s/C and c/v s/C and v/c
(weighted by C) (weighted by C)
Correlation coefficient −0.454 0.780

9
predict actual prices substantially better than the a rate of profit half way between that predicted by
simple labour theory of value (LTV).6 We find this the simple labour theory of value and that predicted
is not so. The Sraffian predictor performs about by the price of production theory. The gas utilities
the same as the LTV under the correlation metric, have a rate of profit of 20% on an organic com-
and only slightly better under the entropy metric.7 position of 10.4; the labour theory of value would
This finding demands some explanation. The fact predict a profit rate of 7% and the production price
that the Sraffian predictor is not clearly ahead of theory 32%. In each case the industry is regulated,
the LTV is comprehensible in terms of the fact that and of course the regulatory system builds in the
profit rates, counter to Sraffian theory, tend to be assumption that the utilities should earn an aver-
lower in industries with a high organic composition age rate of profit. Second, there are industries of
of capital. high organic composition in which rent plays a ma-
This is shown in both Table 3 and Figure 2. The jor role. At an organic composition of 16.4, the
table displays the correlation coefficient between the crude petroleum and natural gas industry has a rate
rate of profit and organic composition, and also of profit substantially in excess of that predicted by
between the profit rate and the inverse of organic the labour theory of value, and approximating more
composition, across the 47 sectors. The former closely that predicted by an equalization of the rate
coefficient—at −.454—is statistically significant at of profit. But an industry like this would, on the
the 1% level. If, however, prices corresponded to basis of the Ricardian theory of differential rent, be
the simple LTV we would expect to find a posi- expected to sell its product above its mean value,
tive linear relationship between profit rate and the and hence report above average profits. In a sim-
inverse of organic composition (in other words, the ilar position we find the oil refining industry with
relationship between profit rate and organic compo- an organic composition of 9.4. Oil production and
sition would be inverse, rather than negative linear), oil refining have similar rates of profit, at 31% and
so the second coefficient is perhaps more telling: at 32%. Since the industry is vertically integrated, this
0.780 it has a p-value or marginal significance level would indicate that the oil monopolies chose to re-
< 0.0001. port their super profits as earned pro-rata on capi-
Figure 2 shows three sets of points: tal employed in primary and secondary production.
In both cases, however, the super profit can be ex-
1. the observed rate of profit, measured as s/C plained by differential rent.
(where C denotes capital stock);
2. the rate of profit that would be predicted on Sensitivity to turnover time
the basis of Volume I of Capital, i.e. s′ v/C,
where s′ is the mean rate of exploitation in the As mentioned above, we do not currently have inde-
economy; and pendent data on turnover times across the sectors,
hence our figures for sectoral capital stocks are not
3. the rate of profit that would be predicted on entirely satisfactory. The most we can do in the
the basis of prices of production (mean s/C). present paper is examine the sensitivity of the re-
sults to the (common) assumption about turnover
It can be seen that the observed rates of profit
time. Table 4 replicates Table 3 under the alter-
fall close to the rates that would be predicted by
native assumption that industry capital stocks are
the Volume I theory. The exception is for a few
composed of BEA fixed capital plus 2 months’ worth
industries with unusually high organic compositions
of wages plus 3 months’ worth of circulating con-
> 10.
stant capital. The correlations indicative of a neg-
But what are these industries? It transpires that
ative or inverse association between profit rate and
they fall into two categories, each arguably “excep-
organic composition are still statistically significant,
tional”. First there are the regulated utilities, elec-
and apparently robust with respect to this sort of
tricity supply and gas supply. Electricity supply
change.
has an organic composition of 23.15, and displays
6 The exception being Farjoun and Machover (1983).
7 Correlation, which depends upon squared errors, lays References
more emphasis on a few big errors whereas entropy, which
depends upon log errors, lays more emphasis on a large num-
ber of small errors—hence the possibility of a difference in Chaitin, G. J., 1987, Information, Randomness
assessment according to the two metrics. and Incompleteness, Singapore: World Scientific.

10
Figure 2: Relationship between profit rates and organic composition, BEA fixed capital plus one month’s
circulating constant capital as estimate of capital stock (log scales)

10
vol. 1 rate of profit
general rate of profit
empirical data +

+ +
1 +
++ +
++
++
+ + ++ +
profrate +++ + +
+
++ + +
++ ++ + ++ + + +
+ +
0.1 ++
+
+ +

0.01
0.1 1 10 100
orgcomp

Table 4: Profit rates and organic composition, BEA fixed capital plus 3 months’ circulating constant capital
and 2 months’ wages as estimate of capital stock (C)

s/C c/v
Mean 0.239 2.218
Standard deviation 0.133 3.146
Coefficient of variation 0.558 1.418
s/C and c/v s/C and v/c
(weighted by C) (weighted by C)
Correlation coefficient −0.457 0.650

11
Cockshott, W. P., A. Cottrell and G. Michaelson,
1995, “Testing Marx: some new results from UK
data”, Capital and Class, 55, Spring, pp. 103–29.
Farjoun, E. and M. Machover, 1983, Laws of Chaos,
London: Verso.
McGlone, T. and A. Kliman, 1996, “One system
or two? The transformation of values into prices
of production versus the transformation prob-
lem”, in A. Freeman and G. Carchedi (eds) Marx
and Non-Equilibrium Economics, Aldershot:
Edward Elgar, pp. 29–48.
Penrose, R., 1989, The Emperor’s New Mind , Ox-
ford: Oxford University Press.
Shaikh, A., 1984, “The transformation from Marx
to Sraffa”, in A. Freeman and E. Mandel (eds)
Ricardo, Marx, Sraffa, London: Verso, pp. 43–
84.

Reza, F. M., 1961, An Introduction to Information

Theory, New York: McGraw-Hill.
Solomonoff, R. J., 1964, “A formal theory of induc-
tive inference”, Information and Control, vol. 7,
no. 1, March, pp. 1–22.

12