Journal of Econometrics 14 (1980) 1I 48.

0 North-Holland Pubhshmg Company

ECONOMIC MONETARY AGGREGATES

An Application of Index Number and Aggregation Theory*

William A. BARNETT

The debate o\er what should be counted as money is between people

who do not know and people who do not know that they do not know.
John Kenneth Galbraith (1979. p. 90)

1. Introduction

Monetary policy is related to the behavior of indices of the quantity, ‘price’

and velocity of money. Yet, for such aggregates to be useful. they must have
meaning and must be measurable. This raises troublesome methodological
questions. What is money? Is it a ‘good’ whose quantity can be measured, or
is it just a vector of different characteristics (liquidity, means of payment,
etc.)‘? Do currency and time deposits possess identical ‘moneyness’ so that
their quantities can be aggregated linearly and with equal weights to acquire
a meaningful quantity aggregate? If money is a meaningful good, then what
is its price? Can more than one monetary aggregate jointly have meaning‘?
We shall explore these issues using aggregation theory and index number
theory.
Although not previously applied to money demand, the literature on
aggregation theory exists prrci.se/j, for the purpose of providing rigorous and
unique answers to the sort of questions listed above. We apply aggregation
theory to the construction and estimation of economic aggregates for
passbook accounts across institution types and then to nested aggregation
over transaction balances.’ We show that passbook accounts at different

*The views expressed herein arc solely those of the author and do not necessarily represent
the crews of the Roard of Governors of the Federal Reserve System. This research was partially
supported by National Science Foundation Grant SOC 76-84459. 1 have benefited from
comments on thus research by William Brainard. Kenneth Clements, David Humphrey. Donal
Donovan, Donald Hester, Franc0 Modigham. Robert Rasche, Henri Theil, Peter Tinsley, and
participants at the University of Chicago’s Econometrics Symposium and at the 1979 European
Econometric Society Meetings in Athens.
‘Economic aggregates are also called functional. true. or exact aggregates. The other well
known class of aggregates consists of statistical indrces, which approximate functional
aggregates.
 institution types are close substitutes and hence can be aggregated linearly.
But aggregation by simple summation is rejected, since the coefficients of the
linear function are unequal. Substitutability between passbook savings and
transaction balances is found to be low. Hence nonlinear aggregation is
required to approximate the economic aggregate over savings and transac-
tion balances.* We provide similar empirical results for time deposits and
large certificates of dcpnsit.
Since the beginning of this century a highly respected and increasingly
sophisticated literature has been under development on statistical index
number theory. While aggregation theory results in exact aggregator fun-
ctions depending upon unknown (but estimable) parameters, statistical index
number theory results in ptrrrrn~rte~_fT(‘~~approximations to aggregator fun-
ctions. Index number theory provides the basis for the index numbers
published by nearly every governmental agency in the world (other than the
central banks).” In the latter sections of this paper, we explore the impli-
cations of statistical index number theory for the construction of monetary
quantity index numbers. and we advocate the use of the TGrnquistVTheil
Divisia index to measure the quantity of money.
During the past decade there has been much concern about the apparent
destabilization of velocity.’ In fact the problem arose primarily because of
the long-run substitution effect resulting from rising own rates on unre-
gulated monetary assets relative to the own rates on rate-regulated monetary
assets. But the value of an LYY~~~o~~~~c aggregate (by its definition) cr~r~r~ot
change as a result of internal substitution effects. Hence the money market
substitution effects destabilizing velocity should be completely internalized by
proper aggregation over the money market.
When any reputable index number formula is used, we find that the
velocity of money is increasingly stabilized as the level of aggregation is
increased, but the velocity of the usual simple sum index is destabilized by
aggregation beyond an intermediate level. Furthermore, on direct empirical
grounds. Barnctt and Spindt (1979, 1980) have shown that the Tiirnquist-Theil
Divisia index tlr~rnirltrrt~s the simple sum index, regurdless of the indexes’
monetary components or of the final targets. The primitive simple sum index
of money supply is scrrrrl~~ defective. as should be no surprise to anyone
who is familiar with any portion of the past half century of literature on
aggregation and index number theory.

‘Our results also suggesl that passbook accounts at commercial banks provide greater
services than passbook account> in sawngs and loans or mutual savings hanks, since the
economic aggregate wclghts passbook accounts at commercial banks more heavily than
pasbhook accounta at sawn~s and loans or at mutual savings banks. Similarly transaction
balances are far mow heavily weighted than passbook accounts.
“See Barnett ( 19x0).
%x, e.g.. En&x, Johnaon and Paulu (1976) and Goldfeld (1976).
2. 0 bjectives

Suppose we should wish to construct a monetary quantity index defined

over the components of M,. In economic aggregation theory, economic
agents (consumers and producers) then must be able to treat M2 as the
quantity of a meaningful single good in their decisions. Economic agents
must be able to select their desired aggregate quantity of M2 without regard
to its composition. The allocation of M, over its component elements then
could be accomplished in a later second stage decision, conditionally upon
the prechosen aggregate level of M,. Varying the relative quantities of
currency and time deposits within M, while holding the aggregate M2 level
constant must not affect tastes or technology over any other goods. If this
condition is satisfied, economic agents can possess stable preferences and
technology over M, and other goods. If M2 is not a good in this
fundamental sense, then tastes or technology over M, and other goods will
appear to shift whenever the relative proportions of the components of M,
change.
If the concept of money has meaning, then it follows that an aggregate of
monetary assets must exist which is treated by the economy us $ it were a
single good. which we thereby can call ‘money’. Such an aggregate is a
function (of its component monetary quantities) which is separable from the
economy’s structure. Thllf concept of money is the subject of aggregation
theory and is the concept relevant to policy, since both aggregation theory
and policy postulate the appearance of a monetary aggregate as a PJICUI-
ir~g,firl stably defined variable in the economy’s structure. Without the
appropriate separability conditions, any aggregate is inherently arbitrary and
spurious and does not define an economic variable.
It can be shown that when a functional quantity aggregate exists for a
consumer, that aggregate itself must possess the known properties of a utility
function. and that utility function must possess certain additional special
properties (homotheticity and weakly separable nesting within the
consumer’s full utility function). Then, when the aggregate quantity index is
held constant, the ‘utility of money’ is necessarily held constant inde-
pendently of its composition. As has been observed by Samuelson and
Swamy (1974: p. 568): ‘The fundamental point about an economic quantity
index, which is too little stressed by writers, Leontief and Afriat being
exceptions, is that it must itself be a cardinal indicator of ordinal utility.’ The
conditions for economic aggregation over goods demanded by firms are the
analogous conditions on production functions. and the form of aggregator
functions and index numbers is identical for firms and consumers. To
simplify the discussion, we present the theory and results in terms of
consumer decisions.”
‘The utility approach to consumer money demand modeling is currently the basis for rapidly
expanding empirical research in the literature. Consider, for example, Chetty (1969). Bisignano
 Irving Fisher (1922) provided a list of desirable properties for economic
price and quantity indices. Ragnar Frish (1930) proved that when the
number of goods exceeds unity, no index number formula can satisfy all of
those properties. However, Samuelson and Swamy (1974, p. 566) have shown
that if price and per capita quantity data is assumed to fall on neoclassical
demand functions (rather than to be unrestricted independent variables, as
assumed by Frisch), then the economic quantity and dual price indices
discussed above ‘do meet the spirit of all of Fisher’s criteria in the only case
in which a single index number of the price of cost of living makes economic
sense’. Furthermore, economic aggregates can be constructed at multiple
levels of monetary aggregation in such a manner that the indices arc nested.
As a result, internal contradictions cannot arise at varying levels of aggre-
gation. We shall use only such economic quantity and dual price indices (or
statistical approximations to them), and we shall accept the assumptions
necessary and sufficient for the existence of a hierarchy of nested aggregates.
The economic quantity index cannot be known exactly without knowledge
of the representative consumer’s utility function. since the economic quantity
index depends upon and is defined in terms of consumer preferences.”
Conventional accounting practices generate economic indices only to the
degree that those indices imply plausible preference orderings over com-
ponents. Yet, all current monetary quantity indices are constructed from
simple addition of components. [f’those indices measure economic variables.
then it follows that the variables have been generated by a utility function for
financial assets possessing the same simple unweighted summation form used
in constructing the index. But such a utility function requires the goods over
which it is defined to be perfect substitutes in identical ratios. In other words,

(lY74), Diewert (1974). Parkm. Cooper, Henderson. and Danes (1975~ C’lements (1976).
Donovan (1977), Phlips (1978). Offenbacher (197’)). and Clements and Nguyen (lY7Y). Early
advocates of the utility approach include Friedman and Patlnkin.
The utility approach is based upon impllcit modehng rather than the explicit modehng used in
the transactions demand or portfolio analysis approaches. In an economic (rather than
empirical) sense, the utility approach is a reduced form approach which mod&, restricts, and
characterizes the results of the consumer’s decisions without the need to consider the explicit
structure of the decison. The resultin g approach has the merit of unifying the modellnp of the
demand for all money market instruments within a single framework wlthout the need to
explore the detailed and different structures of the comumer‘s decisions wlthin each sector of the
money market.
It would be preferable to aggregate within each relevant sector (firms. wealthy households.
etc.) separately, and then over the sector aggregates. But adequate data currently ih not available
by sector.
‘In this paper we postulate the existence of a representative consumer, although we argued
against that practice in Barnett (lY79b). WC use a commumty utility function here because of its
usefulness as a means of imposing functional regularity on demand systems. rather than out of
any conviction that such a community utlhty function actually need exist. However, there ilr
some empirical and theoretIcal evidence that, under some conditions, the beha\mr of aggregate
consumption data may be approximated by a consistent and transitive prefcrencc ordering. See
Dixon (1975), Maks (1978). and DonoLan (1979).
 WA. Brrmrtt, Economic monetary aggregutes 15

the components of the quantity index must be indistinguishable to the

consumer.’
We frequently seek information about the ‘price’ of money. In various
studies, the price of money has been viewed as an interest rate, an index of
interest rates, the rate of change of prices, the price level, or an index of some
subset of those subindices. We derive the Jorgensonian user cost (equivalent
rental price) of money. Furthermore, when we aggregate over monetary
assets, we can aggregate over those user costs (in parallel with aggregation
over quantities) by using the theory of economic price indices. Once the
consumer’s preference structure has been estimated, the ‘price’ and the
quantity of the aggregate are simultaneously implied as duals.’
In section 3 we define the consumer’s intertemporal decision problem,
derive the user cost of monetary assets, and derive the consumer’s current
period monetary asset allocation problem. In section 4 we block the
consumer’s current period monetary asset utility function, and we present the
consumer’s resulting multistage budgeting procedure along with the implied
functional quantity and price aggregates. In section 5 we present our
recursive estimation and recursive functional aggregation procedure, which
works its way up the utility tree. In sections 6, 7, and 8 we present our results
with passbook accounts and transaction balances and reject aggregation by
summation. In section 9 we present a procedure for empirical selection of
a more extensive blocking of preferences, and we discuss some preliminary
results. In section 10 we define parumeter-free statistical index numbers that
can approximate the functional aggregates, and we construct an example; we
plot the velocity of the resulting quantity index. We conclude by advocating
use of the Divisia parameter-free statistical index number to measure the
quantity of money. We reject further use of the obsolete simple sum index.

3. The consumer’s decision

3.1. lntrrtemporal ullocation

In this section we shall derive the Jorgensonian user cost (equivalent rental
price) of monetary assets from a rigorous Fisherine intertemporal con-
sumption expenditure allocation model. Since the model is formulated in

‘In the existing simple sum M,, for example, the consumption characteristics of one dollar of
currency must be identical to those of one dollar of long-term small time deposits. The violation
of aggregation theory increases as the level of aggregation increases, since the higher the level of
aggregation the less substitutable the components of the aggregates becomes.
‘These indices satisfy the accounting identity of equality between expenditure and the product
of quantity and price, and the consumer can be shown to behave in a rational manner relative
to the aggregate good whose price and quantity have been defined. This rationality obtains both
relative to the aggregates and relative to their components. and consumers’ decisions at all levels
of aggregation are consistent with a single joint rational choice criterion.
16 WA. Bnrnett, Economic monetuq irggregutes

discrete time, a structure of assumptions is required regarding the timing of

interest rate and price changes and of portfolio transactions.’
We define time period t to be the time interval [r, t + l), closed on the left
and open on the right. Hence the instant of time t is included in interval t,
but the instant t + 1 is not. We assume that consumption of goods can
proceed continuously throughout any time interval, although our model will
use only the total (integral) of that consumption for any time period. Stocks
of monetary assets and bonds are constant during each period, and can
change only at the end of an interval. Hence during period t, any changes in
holdings occurring at instant t+ 1 are not seen until the initial instant of
interval t + 1. In short, all portfolio transactions take place at the boundaries
between intervals.
Interest on bonds and on monetary assets is paid at the end of each
period. Since the end (right hand boundary) of period t is included in period
t+ 1, but not in period t, interest paid for asset holdings during period t
cannot be consumed until period t + 1. Interest rates. prices, and wage rates
remain constant within the interior of each period, but can change discretely
at the boundaries of periods. Hence capital gains (or losses) resulting from
changes in market bond yields can take place only at the boundaries of
periods.
We treat labor supply as exogenously determined, and we assume that
labor supplies, L,, ...,
L,,T, during all periods of the consumer’s planning
horizon are blockwise weakly separable from all other arguments of his
utility function, so that we can use the subutility function defined only over
the other arguments.
Let t be the current period (or equivalently the instant of time at the start
of the period).” Let T be the length of the planning horizon, so that the
consumer currently plans through all periods, s. in (s: t 2 s 5 t + TJ . We now
define our variables:’ ’

XS =vector of per capita (planned) consumption of goods and services

(including those of durables) during period s.
P, =vector of goods and services expected prices and of durable goods
expected rental prices during period s.
mi, =planned per capita real balances of monetary asset i during period s
(i=l,..., n).
'i6 =the expected nominal holding period (including capital gains and
losses) yield on monetary asset i udring period s (i = 1,. ., 11).

‘Although we shall not fix the time interval, it could be set at one day, since interest on
savings deposits rarely is paid more frequently than daily. For our purposes a daily discrete time
model could be viewed as approximating a continuous time model. since our average quarterly
data corresponds to a substantially longer period.
“In effect the instant at the start of each period indexes the period.
’ 1We require planned to equal actual values only during the current period I.
A, =planned per capita real bond holdings during period s.
R, = the expected (one-period holding) yield on bonds during period s.12
LS =per capita labor supply during period s.
U’S = the wage rate during period s.

We let u, be the representative consumer’s current intertemporal, T-period,

utility function. We assume that U, is weakly separable in each period’s
consumption of goods and monetary assets, so that U, can be written in the
form

%=U,(m,,...,m,+,;x ,,..., X,+r;A,+T)

=~t(~(m,),c,+,(m,+,)....,c,+~(m,+,);
%G), V+,(x,+,h..., I/;+~(x,+T-);A~+T-X (3.1)

for some monotonically increasing, linearly homogeneous, strictly quasicon-

cave functions, u,, , , . . ,, z),+ T’ V; I: + 1,. . ., V,, T.l 3 The function, v, is monotoni-
cally increasing and strictly quasiconcave, but not necessarily linearly homo-
geneous. l4 The functiov U, also is monotonically. increasing.
Dual to the functions, V and V, (s = t + 1,. . ., t + T), there exist current and
planned true cost of living indices, p,? =p(p,) and pt=pf(p,) (.s=t + 1,
. ., t+ T).15 Those indices will be used to deflate all nominal quantities to
real quantities, as in the definitions of mi, and A, above.
Assuming replanning at each t, we write the consumer’s decision problem
during each period s (15s 5 t + T) within his planning horizon as to choose

“As will be seen in the formulation of the consumer’s decision problem below, R, is the
expected one-period holding (including realized or unrealized capital gains or losses) yield
(during period s) on assets accumulated to transfer wealth between multi-period planning
horizons rather than to yield liquidity or other services during the current period. As a result, A,
will enter the consumer’s utility function only during period s = t + 7; and A, need not necessarily
be bond holdings. We use the word ‘bonds’ (also sometimes referred to as the benchmark asset
in this context) to simplify exposition. The benchmark asset’s one-period holding yield during
period s is defined to contain all premiums available in the market for foregoing the services
provided by monetary assets. Note that the holding period used in defining R, must be the same
as that of ris, which is a short rate.
“Observe that c, = L’and I: = V independently of t for the current period f. Also observe that U,
can be acqmred as a ‘derived’ utility function from the ‘true’ utility function, which does not
depend upon monetary assets. The derivation postulates the existence of an arbitrary time-using
transactions technology. See Arrow and Hahn (1971, p. 350). Quirk and Saposnik (1968, p. 97)
further argue that if such a derived utility function does not exist, then no money will be held in
equilibrium.
14We later shall assume that 1: is linearly homogeneous in supernumerary quantities.
“For a discussion of the relevant duality theory, see the appendix. The true cost of living
index for a weakly separable block of goods equals expenditure on those goods divided by the
(category) indirect utility function for those goods.
subject to (3.2)

The real value of assets carried over (endowed) from the prior planning
period is

and the real value of the consumer’s provisions for later planning periods is

Let

p,=L s= t
r-l

= n (liR,), t+1ssst+7:
U=C

Then ps is the discount factor for discounting period s transactions. Observe

that p,# fliE1 (1 + R,), since R, is not paid during [s, s + l), but rather at the
start of [s + 1, s + 2). In problem (3.2), ( m,,x,) is actual consumption of goods
and monetary assets during period t, while (m,, 1,. . ., m,+T; x,, ,, .,x,+ T) is
planned consumption of goods and monetary assets.”
Solve (3.2) for A, and write the resulting equation for each s between t and
t+ 7: Then back substitute for A,7 starting from A,,, and working down to
A,, always substituting the lower subscripted equation into the next higher

“Since we assume replanning each period and permit u, to vary over time. the consumer’s
behavior is bound only by his decislons regarding current period consumption. Actual
consumption patterns need not evolve in agreement with prior plans. However. further
restrictions (stationary preferences. intertemporal strong separability. and constant rate of time
preference) could be imposed upon U, to assure that the sequence of current consumption
quantities evolves over time in agreement with plans whenever correct expectations exist for all
variables that are not under the consumer’s control. Agreement between actual and planned
consumption paths is not necessary to estimation of our model.
 WA. Barnrtl, Economic monrtar? aggregatvs 19

one. Completion of the sequence of back-substitutions results in the single

wealth constraint

(3.3)

The consumer now can be viewed as maximizing utility subject to the single
wealth constraint, (3.3). which is interpreted in Barnett (1978).

3.2. The user cost of‘ monetary assets

From (3.3) we see immediately that the user cost (equivalent rental price) of
illis is

,_Pi+ P,h(l+ris) (3.4)

’ P, PS+l

Finally the current period user cost, nil, of m, reduces to

nit =
PF(Rr-rit)
7x-m -
17

Correcting the formula for taxation, we get

pr*(R,-ri,)(l -T,) ,8
(3.5)

“It can be shown that z,, is the monetary asset analog of the well-known Jorgensonian user
cost (rental price) of durable consumer goods. See Donovan (1979).
‘“User costs c’ommonly are viewed as the prices of the services of durables rather than of their
stocks. See Donovan (1978). In that interpretation, services are assumed to be proportional to
stocks. and units of quantities and prices are assumed to have been chosen such that the
proportionality constants are one. Hence user-cost evaluated stocks (stocks multiplied by
corresponding user costs) are expenditures on the sercicrsof the stocks.
Observe that our model is not a disequilibrium stock adjustment model. Since we assume
continuous optimal adjustment, consumers optimally select quantities consumed for their
services.
where T, is the marginal income tax rate. Observe that financial asset i is a
free good if yil =R,, and observe that the current period user costs of
financial assets are independent of expectations. We shall use formula (3.5) to
compute the user costs of financial assets.
It is interesting to observe that although (3.5) does not depend directly
upon inflation rates, the nominal interest rates within the formula can be
expected to respond to expected inflation rates. Furthermore, since the well-
known user cost formula for non-monetary durables services does depend
inversely upon the expected inflation rate, it follows that the user cost of
monetary assets relative ro durables increases as the expected inflation rate
increases. Hence consumers will respond to increased inflationary expec-
tations by substituting consumer durables for monetary assets.

3.3. Supwnunzerury quantities

We have not assumed linear homogeneity of r, since that assumption

would be unnecessarily strong for our purposes. However in this section we
assume a form of marginal homogeneity that will be required for
aggregation.
We assume that 1’depends upon m,- 1 as well as upon m,. This assumption
introduces no complications into the earlier sections. since the consumer
selected m,- I during the prior planning horizon and hence m,- 1 is given and
fixed during the current horizon. We further assume that there exists
constants, 6 = (6 ],. ., ii,)‘, and a linearly homogeneous function, II. such that
o(m,;m,_,)=u(y,), where J~~=(Jx,~,..., Y,~)’ and Jo, = nzit -iip~,.,_~. In short, we
assume the existence of proportional habit formation in current (but not
future planned) consumption.” In the language of the habit formation
literature, y, is supernumerary consumption of monetary assets and Simi,,_ 1
is the quantity of monetary asset i consumed out of habit (independently of
current interest rates or income) during period t.
From (3.1), (3.3), (3.4). and (3.5), we see that the consumer’s intertemporal
decision problem can be rewritten as to choose (y,, m,, ],. ., m, +.,.;
X,,...,-q.T> .A,+,.)~0 to

max U,(u(y,),L’t+l(m,+,),...,c,+,.(m,+.);

(3.6)

subject to the single wealth constraint

“The theoretical implications of habit formation have been considered by Pollak (1976).
 WA. Burrwtt. Economic monrtary aggregutrc 21

1+7

= 2 (W,y/Ps)L,+ i C(l +‘.i.r-1)Pt*-I-6i~i,1mi.,-,

+ (~+R,-~M-,P,*_,.
(3.7)

We now have established the model and assumption structure needed to

apply aggregation theory to monetary aggregation. If we must use ag-
gregates, a case can be made for accepting whatever assumptions are
required to render economic aggregates meaningful. If we cannot accept the
assumptions, we have no economic aggregates at all. As Samuelson and
Swamy (1975, p. 592) conclude: ‘One must not expect to be able to make the
naive measurements that untutored common sense always longs for; we must
accept the sad facts of life, and be grateful for the more complicated
procedures economic theory devises.’

3.4. Conditional current period allocation

Our assumptions on the homogeneous blockwise weakly separable struc-

ture of the intertemporal utility function, (3.6), are sufficient for consistent
two-stage budgeting. Hence by Green’s (1964) Theorem 4 it follows that the
consumer can maximize utility, (3.6), subject to the wealth constraint, (3.7) in
two stages. In the first stage, the consumer selects aggregate monetary asset
expenditure (supernumerary expenditure for the current period), aggregate
consumer goods expenditure for each period within his planning horizon,
and his terminal bond holdings, A,+T.20 In the second stage, he allocates
current aggregate monetary asset expenditure and current aggregate con-
sumer goods expenditure over individual current period monetary assets and
consumer goods.
The second-stage allocation decision over individual current period super-
numerary monetary assets is to select y, to

maxU(Y, 1
subject to (3.8)

*‘The chosen bond holdings are to be carried forward to the start of his next planning
horizon.
22 WA. Burnett, Economic monetary aggregutes

where ni’f=nJp: is the real current period user cost of monetary asset i, n:
= (n,“, . . .,T$)‘, and M: is the real value of aggregate supernumerary mo-
netary asset holdings allocated to the current period in the consumer’s first-
stage decision. Observe the nz =(R,-vi,)(l -s,)/[l +R,(l -z,)] independently
of PT.21
We model the conditional current period monetary asset allocation
decision, (3.8), in sections 4 through 9 of this paper, and we explore its
implications for aggregation.22

4. Preference structure over financial assets

4.1. Blocking of thr utility function

Suppose that yr contains only total transaction balances and passbook

savings deposits at three institution types, and we seek to aggregate
passbook savings deposits over institution types and to nest that aggregate
within an aggregate of all of the components of yr. We partition the vector,
yt, such that Y~=(F~~, Y;,)‘, where 4’lt is per capita real supernumerary
transaction balances and y,, is a vector of per capita real supernumerary
passbook account deposits. We correspondingly partition $ and 6 such that
7~: = (TTY,,&‘)’ and 6 = (6,, 6;)‘.
We assume that the utility function, u(y,), can be written in the blockwise
weakly separable form

(4.1)

with the function u2 being linearly homogeneous. As discussed below, these

conditions are both necessary and sufficient for the existence of the economic
aggregates we seek.23
Backsubstituting (3.7) into (3.6), observe the way in which we have nested
weakly separable blocks within weakly separable blocks. We have established
a fully nested utility tree. As a result we can acquire a rational multi-stage
budgeting procedure, in which the structured utility function itself defines the

*‘The choice between the real values, KC and M:, and the corresponding nominal values. z,,
and M,, is arbitrary. since p,? can be canceled out of each side of the budget constraint in the
nominal case. This observation IS just a restatement of the well-known homogeneity of demand.
We further could multiply the budget constraint through by [I + R,( I -T,)]/( I -7,) in order to
use R,-r,, as prices. The simplified formulation then would correspond with that of Klein
(1974) and Offenbacher (1979).
‘*We treat MT as exogenous, although MT actually may be endogenous. Although nearly all
of the demand systems literature estimates such conditional current period demand, we
nevertheless should recognize the possibility of simultaneous bias in the estimates.
2’This conclusion, based upon Green’s (1964) Theorem 4, assumes that yr is held exclusively
by consumers. For firms, the analogous conditions would be applied to the production
functions.
 WA. Burn&t, Economic monerury uggrrgutrs 23

relevant theoretical quantity index at each stage, and duality theory defines
the corresponding functional price index.24
In the next section we elaborate on the multi-stage budgeting properties of
decision (3.6) and the implications for quantity and price aggregation.

4.2. Multi-stuge budgeting

Our assumptions on the properties of u are sufficient for a two-stage

solution of the decision problem (3.8).2” We define that two-stage decision in
this section.
Let 77E = K7,(nz,) be a function of the user costs nz,. The first stage of the
two-stage decision is to select ylr and Y,, to solve

max ~(4’~~~
Gt),
(Y,,,Y,,)

subject to (4.2 I

From the solution to problem (4.2), the consumer determines aggregate

supernumerary consumption of real passbook account services, l7,*,Y,,.
In the second stage, the consumer allocates m&Y,, over consumption of
the services of passbook accounts at individual institution types. He does so
by solving the decision problem:

max u2 (y2, ),

subject to (4.3)

It follows from Green’s (1964) Theorem 4 that there exists some function,
II?, such that the solution for y, to problem (3.8) is the same as the solution
for y, acquired from the two stage decison, (4.2) and (4.3). for any
theoretically admissible values of MT and 7~:. It furthermore can be shown
that if we use that function, I7,, in (4.2), then Yzr=zd2(yZr) at the solution
values for Y,, and y2, to the two-stage decision. We shall say that Y,,
“Other financial assets (repurchase agreements, money market mutual funds, Treasury bills,
commercial paper, etc.) could be included in the analysis by increasing the dimension of y,.
partitioning it into more than two subsectors, and blocking u into multiple blocks accordingly.
“Recall that decismn (3.8) itself was defined as the second stage of a two-stage decision.
Hence we now are acquiring multi-stage budgeting rather than just two-stage budgeting. Our
separability conditions are also sufficient for modeling structural change through the use of the
household production function approach. That approach introduces production functions
which model the production of monetary services from monetary asset portfolios. See Barnett
(1977b).
 =nz(y2,) is the economic (or functional) quantity aggregate (or index)
corresponding (or dual) to the economic (or functional) user cost aggregate
(or index), L’$=n,(nT,). We shall call ~1~the quantity aggregator function,
and we shall call 17, the user-cost (or price) aggregator function.
In general, the quantity aggregator function is the corresponding (ca-
tegory) utility function. We show in the appendix that the corresponding
price (user cost) index is equal to expenditure, flF,Y,,, divided by the
(category) indirect utility function (induced by the direct utility function, IJ~).
This two-stage decision process is two-stage budgeting, and can be
extended to n-stage budgeting simply by nesting weakly separable blocks
within weakly separable blocks, etc., in the analogous manner. The result
that follows from such nesting is purely mathematical and need not be
related to actual multistage decision processes. We need only observe that
the consumer acts ‘as if’ he were making his decision in stages. if his
preferences are nested.
The price index, l7;,, and the quantity index, Y,,, are economic price and
quantity indices. As can be seen from problem (4.2), those indices have all of
the properties of quantities and prices of actual goods (whether or not
aggregates).”

5. Recursive estimation approach

The consumer is viewed as making his budgeting decisions from the top of
the tree down, as he decentralizes his budgeting to lower levels of aggre-
gation; but we can estimate the entire implied model recursively from the
bottom up. We begin at the bottom of the tree and estimate the most
disaggregated demand decisions. We compute the implied price (user cost)
and quantity indices, based upon the utility functions we have estimated, and
we then move up to estimate the next level using the just-computed price
aggregates as instrumental variables. This approach to recursive estimation
of utility trees has been developed by Barnett (1977a), Fuss (1977), and
Anderson (1979). Our data consists of quarterly average values from the first
quarter of 1970 to the first quarter of 1978. The data sources are described in
Barnett (1980, ch. 7).
Recall that the current period monetary asset allocation problem, (3.8), is
defined conditionally upon the consumer price index, p:, which is dual to
(and therefore derivable from) the consumer goods current period utility
function, V. Hence to apply this instrumental variables approach most fully,
we should estimate the function, V, defined over the consumer goods sector,

Z61n particular, observe that the consumer acts as if actual aggregate goods existed. Also
observe that quantity indices depend exclusively upon quantities, and that price indices depend
exclusively upon prlccs. Furthermore, the budget constraint of problem (4.3) shows that the
product of a dual price index and its corresponding quantity index always equals actual
expenditure on the goods wlthin the aggregate.
prior to estimating U, defined over the monetary asset sector. But aside from
p:, we seek no other information from the consumption sector. Hence the
cost of strict adherence to the recursive instrumental variables approach is
excessive in the case of computation of p,*.
As a result, we use a statistical index rather than a functional index for p,*.
Statistical price indices can depend upon quantities as well as prices, but
cannot depend upon unknown parameters.2’ We assume that I/(x,)
==(x;Bx,)” locally for some square matrix, i?, of unknown parameters. That
specification can provide a quadratic approximation to any aggregator
function. Diewert (1976) has shown that if a representative consumer exists,
then the Fisher Ideal statistical price index (geometric mean of the Laspeyres
and Paasche indices) is always equal to the true value of the functional
index, p:, regardless of the values of the parameters in the matrix. B. We
shall use the Fisher Ideal price index for p:.”
Having computed p:, we begin our empirical ascent up the utility tree.
Recalling the form of equation (4.1) we begin by estimating u2. Then u2(~y2,)
becomes the economic quantity index used with ~‘i, in the next (higher) stage.
We compute the implied price index dual to u2 and estimate the demand
system generated by p. The procedure could be carried to any level of
aggregation, but will be terminated at ,u.

6. Passbook savings

6.1. C.E.S. specification

In the current subsection we present our specification for passbook savings

conditional demand, which is the solution to decision (4.3). Since m,, is a
vector, we implicitly have segmented passbook deposits into categories. We
let m2t=(m21t, m221,m23L)‘, where mzll= real per capita holdings of com-
mercial bank passbook accounts, m221= real per capita holdings of savings
and loan passbook accounts, and mz3,= real per capita holding of mutual
savings bank passbook accounts. 29 We then write tth period supernumerary
real per capita holdings in passbook account category i as y2it =m2il
-62im2i.t- 1’

To cldrify our notation, we replace the subscript 2 with p (for passbook).

“Statistical Indices are introduced more rigorously in section IO.

*aIn computing the Fisher Ideal index, we use the Bureau of Labor Statistics’ CPI as the
Laspeyres index and the Commerce Department’s Implicit Price Deflator as the Paasche Index.
Some approximation error exists m the use of the CPI as the Laspeyres index. although the
error is small. See Triplett (1976).
29Lack of appropriate data prevented us from usmg credit union passbook deposits.
 Then u,(y,,)=u,(y,,), etc. The C.E.S. specification for up is

where a= (ccl, cxz,ct3)’ and /J are parameters satisfying fi < 1 and a 20.30 In
decision (4.3) we let Ezr = n,*, Y&, which is total user-cost-evaluated expendi-
ture allocated to passbook account services, determined from the prior
allocation stage (one level higher in the utility tree).
The solution to (4.3) is the demand system

(6.1)

where

with

2=(~?,,5,,2,)‘>0 and fl<l.

The vector of parameters a is not jointly identified, since the demand

system is homogeneous of degree zero in a. Hence we impose the identifying
restriction xi ii = 1.31
We seek to estimate (6.1) in a form that will impose all theoretical
restrictions. We do so by transforming the parameters into other parameters
that are free of inequality restrictions. We then impose our restrictions by
substitution. We can acquire the maximum likelihood estimates (M.L.E.) of
the transformed parameters and then acquire the unique M.L.E.‘s of the

30While more flexible utility functions exist than the C.E.S., they did not appear to be
appropriate to our objectives. Our approach estimates a demand system that ts mtegrable to a
marginally homothetic utility function and has known closed form representations both for the
demand system and for the utility function. The model also is a generalization of the sample sum
utility function which provides the conventional quantity Indices. The C.E.S. satisfies all of those
objectives and is a very substantial generalization of the simple sum function. Since the simple
sum aggregate is widely used, it could be impractical (at this stage of research) to consider a
quantity index more general than the C.E.S. Furthermore, the use of a common elasticity of
substitution appears reasonable with our passbook savings data. At higher levels of ag-
gregation, a more flexible functional form would be required.
“‘We do so by estimating (6.1) with the normalizatron cl= 1. and then renormalizing the
resulting estimates to get 1, 3Li= 1. The choice of normalization is arbitrary: we can renormalize
at will.
 W,4. Barnrtt. Economic monrtcrry aggregates 27

original restricted parameters by using the invariance property of the M.L.E.

In particular, we substitute the transformation Ej=yT (j= 1,2,3) to impose cj
20, and we estimate the unrestricted parameters y=(y1,1.‘2,y3)‘. Since Q< 1
defines an open set, that restriction (or any other such strict inequality
restriction) cannot be imposed. We replace /!I< 1 with the approximation fl
50.9. We then substitute the transformation fl= 1.9 -cash 0 into (6.1) and
estimate the unrestricted parameter 0.
Since y,,>O, it follows that for any i, we must have mpir >S,,mpi,,_ 1 for all
t. Since passbook deposits never changed by more than 207; between
quarters in our data, a sufficient condition for that inequality would be bpi
SO.8 for all i=l, 2, 3. We shall impose that sufficient condition. In addition,
we require that 6, z0.32 We jointly impose all of these restrictions on 6, by
substituting the transformations 6,, =0.4( 1 +sin $J for i = 1,2,3, and estimat-
ing the unrestricted vector 4 = (4,, 42, $3)‘.
Multiplying (6.1) by 7$,/E;, to acquire desired expenditure shares, wii,
= ~p*itmpitlE& and making all of the parameter substitutions described above,
we acquire our model for the consumer’s desired expenditure shares. Since
adjustment costs may exist, we permit actual expenditure shares, wpit, to
differ from desired expenditure shares, w&, in accordance with the partial
adjustment scheme: wgit = Lwzit + (1 - A)w,~.L 1, where 0 5 A $1. We use the
same adjustment rate, i. for each institution type to assure that the budget
constraint will be satisfied in actual expenditure shares as well as in desired
budget shares. In addition, equality of adjustment rates appears plausible
for passbook accounts at different institution types. Performing all of these
transformations on (6.1) we have our passbook deposits allocation model.
We take xpir, i= 1, 2, 3, as endogenous and Ep*I and n,*,,, i = 1, 2, 3,
as exogenous. We adopt a conventional additive error structure without
serial correlation.33

6.2. Theoretical index number properties

We now consider the properties of the functional price and quantity index
numbers for passbook savings, when aggregation over institution types is to
be consistent with the C.E.S. consumer preferences specified in the previous
section.

32Although theory does not require this restriction, the logic of the multi-stage budgeting
process becomes more difftcult to mterpret when 6, contains negative elements. In addition, our
prior views on 6, impute low probability to negative elements of 6,, and Barn&t (1977) has
found that negative estimates of 6, tend to have low precision and hence to be statistically
indistinguishable from zero at conventional levels of significance.
33Serially correlated disturbances did not appear to be a potential problem, since our
specificatron contains lagged values both of quantity demanded (through habit formation) and of
expenditure shares (through partial adjustment).
 The functional quantity index is the utility level itself. Normalizing the
index to equal 1.0 at the first observation, we acquire the normalized
functional quantity index Q,,(Y,,) = u,,(Y,,,)/u~(~~, ).34 The nominal functional
price index that is dual to our C.E.S. specification of up is rcp(np,)
= (I?=, @g,p, where (en) are as defined in the previous section. The
corresponding normalized nominal user-cost price index is P,(n,,)
= 77,(n,,)/77,(n,, ).35
We seek to consider the limiting case in which x1 =x2 =a3 and fl= 1. In
that case the functional quantity index equals the simple sum of its
components. Since the elasticity of substitution, 0, equals l/(1 -/3), we see
that G+% as /?- 1. Hence the special case we are considering is that of
three ‘goods’ (or, more appropriately, services) that are perfect substitutes in
equal proportions, i.e. indistinguishable goods. When p= 1 (but the ri’s are
not necessarily equal), the functional quantity index acquires the form of a
Laspeyres-type (fixed weight linear) quality index. The functional price index
that is dual to the Laspeyres quantity index is the Leontief price index,
17,(lr,,)=minj7ipir/3i: i= I, 2, 3).“” Hence if the monetary quantity index is
the usual simple sum index (so that CI, =ixZ =x3), then the corresponding
price index is just the minimum user cost.

6.3. Results with prmhook savings

The parameter estimates for eq. (6.1) using passbook data and joint
maximum likelihood (FIML) estimation are displayed in table 1 with
standard errors in parentheses and with ;s3 normalized to equal one.37 The
estimates of 4i and (+z,yl,) imply boundary solutions for b,, and (ci,,,(S,,)
at their lower and upper bounds, respectively. Transforming back to the
original parameters of y,,(rpr), we find that the implied joint maximum
likelihood estimates are p=O.62 and r=(0.55, 0.26, 0.20)‘, where a has been
renormalized such that c’= , xi = 1.
Precisions (t-ratios) are generally high. The implied elasticity of sub-

34A functional quantrty index must be lmearly homogeneous in its arguments. Whtle u,, is
linearly homogeneous in J’~,, up, is not homogeneous in m,, unless 6,, =0 for all i. Hence up
cannot strictly be viewed as an aggregator function for m,, when some ii,,, is non-zero, although
u,(y,,) is always the functional quantity aggregate for the supernumerary quantities, T,,,.
“The corresponding real price indices are n,,(nz,) and P,(xz,). If we were to reqmre an index
of toto/ (rather than per capita) supernumerary nominal balances, we could compute Q,,(y,,)
using total passbook deposit data in place of the fvzercaptta real balances. mpf, in the defmuion
of _v,,, The result would be identical to computing Q,,(J,,) with populatton and p: fixed at index
year levels, since those fixed index year levels would be cancelled out of the numerator and
denominator of Q,(J),,,),
‘%ee Samuelson and Swamy (1974. p. 574).
j’The standard errors were computed from Theorem 4 of Barnett (1976). The data is
described in Barnett’s (1981) Chapter- 7.
 WA. Burnett, Economic monetary uggrrgutrs 24

stitution, 0, equals 2.66, which is very high.j” We can see just how high that
value is by observing that CTis monotonically increasing in [j, and /I must lie
between - cr3 and 1. Clearly /?=0.62 is very close to the upper bound of 1, at
which the utility function (and hence the functional quantity index) is linear
and demand functions become set valued correspondences.39
Thus we see that passbook accounts at different institution types are
highly substitutable, and a simple linear quantity index may be a reasonable
approximation to the theoretical quantity index. However the simple sum
index requires equal weights in the linear index, and d, differs substantially
from oi,, which does approximately equal 3i3.” The tail area of the
asymptotic likelihood ratio test of equal ai’s is less than 0.00001. Since that
tail area is well below 0.05, we reject the hypothesis of equal xj’s4r

Table I
Parameter esttmates.’

0, 4, 42 43 ‘r’l 72 L

1.94 - Jr/‘2 rl;2 rr,i2 3.83 1.43 0.206

(0.24) (1.31) (0.35) (0.27) (0.18) (0.06) (0.06)

“Standard errors in parentheses.

A functional quantity index measures the quantity of a properly aggre-

gated economic ‘good’. Since ~xr clearly exceeds a2 or x3, we see that
commercial bank passbook accounts contribute more heavily to that mean-
ingful economic ‘good’ than mutual savings bank or savings and loan
passbook accounts. An explanation may lie in the fact that commercial
bank passbook accounts possess all of the basic consumption characteristics
of the other two types, but greater liquidity through the ‘one-stop-banking’
property made available during routine trips to the bank to deposit funds

j”This elasticity is the short run elasticrty of substitution, as is relevant to the aggregator
function and hence to aggregation and index number theory. Regarding the long-run utility
function, see Pollak (1976).
390bserve from ,? that the estimated quarterly adjustment rate from desired to actual shares is
about 21”‘./
?f E, ‘ZE,=~, with /j’= 1. then u(y,,) is a linear function of the usual simple sum index,
I;‘=, mpl,. But with unequal E,‘s, our economic quantity index is a linear function of c:=, aimDi,.
not of the simple unweighted sum.
4’To test the hypothesis of a simple sum aggregate, we should test the hypothesis that /j = 1
jointly with the hypothesis of equal intensity parameters (x~s). However the likelihood function
is not uniquely defined when B= 1, since demand functions become set valued in that case,
Hence a likelihood ratio test is not applicable. We could construct an approximate test by
testing the hypothesis that p =0.999, at which demand remains a point valued function.
 into checking accounts.‘2 If funds were transferred from savings and loan
passbook accounts to commercial bank passbook accounts, our functional
quantity index would increase, evidently to reflect the economy’s increased
liquidity. The usual sum index would not change.5”
There appears to be information contained in the fact that 6,, is at its
lower bound, while dp2 and a,, are large. Recall that 6PimPi,r_ 1 is a vector of
quantities consumed out of habit (or for ‘subsistence’) regardless of the
variations in user costs or in total consumption expenditure within the
sample period. Evidently commercial bank passbook accounts contain ac-
tively managed primary balances, while mutual savings bank and savings
and loan passbook accounts contain a greater percentage of less actively
managed secondary balances and saved consumer reserve funds.“’

7. Transaction balances

We now progress to the next level of the utility tree in (4.1) to estimate ,u.
We again use a C.E.S. utility function.4s We specify ~1 to be C.E.S. in two
goods: real per capita supernumerary transactions balances, xlt, and the
economic real per capita supernumerary passbook savings aggregate, upt

“Aggregation theory does not attach a name (such as ‘moneyness’ or ‘liquidity’) to the
functional quantity index. However our use of user costs does dictate that the quantity index is
the quantity of services provided by the components of the aggregate. Hence it may not be
unreasonable to deduce that commercial bank passbook accounts appear to provide greater
‘monetary services’ than passbook accounts at the other two institution types.
‘“We also observed that computed values of the normaltzed functional quantity Index. Q~(J-~,),
and the normahred user-cost price index. Pp(xpl). tended to move in opposite directions. as
would be expected from movement along a demand curve. This result is not surprismg since
Regulation Q cannot decrease the user cost of passbook account deposits to below the
equilibrium price, although the regulation can raise the user cost to above the equilibrium level.
Hence an excess supply but not an excess demand can exist in the passbook account market.
We therefore can expect the data always to lie on the demand function, even when the market IS
out of equilibrium. In addition, governmental rate setting tends to mimmize simultaneous bras
rn estimators that conditmn upon cxogcnous user costs.
“‘When integrability conditions are imposed, as we have done, it is common for some of them
to be binding. Hence the extstence of bmding regularity conditions is not surprising.
Nevertheless, it 15 also possible that the boundary solutions on the habit formation parameters
may have resulted from the joint use of the habit formation dynamics and partial adjustment
dynamics. Despite the fact that all of the model’s parameters are identified. the data may not
contain sufficient informatron to permit distinguishing adequately between the two sources of
dynamic consumer behavror.
“‘At this level of aggregation, It no longer would be reasonable to assume that elasticities of
substitution are constant between all monetary assets. But we now have only two ‘goods’ and
hence only one elasticity of substitution. The flexrbility of the C.E.S. specification therefore still
remains satisfactory for our purposes. Furthermore a constant lintte elasticity of substitution.
even between all monetary assets, would be more reasonable than the uniformly infinite
etastictties of substitution implied by the usual simple sum indices.
 =Up(Ypt).46 We introduce no additional habit formation at this level of
aggregation (in yIr and the aggregate ppt), since habit formation already is
built into u,(y,,) through the specification of ypt, and since we expect short-
run Engel curves in ylr to pass through the origin.“7
We impute to m,, the user cost price, (3.5), with the own rate set equal to
zero. We impute to the supernumerary passbook aggregate, up(yp,), the dual
user-cost functional price index, 77Pl= 77&n,,,). We do not introduce adjust-
ment dynamics at this level of aggregation. Since transaction balances
turnover rates are high, we believe that adjustment to the desired transaction
balances share in monetary asset consumption is rapid.“”
The utility function is of the C.E.S. form

where (x1, CQ, p) are parameters satisfying p < 1 and (cz,, r,)>O.
The conditional decision problem at this level of aggregation is to choose
(?n It, upr 1 to

maxyh,,, up,L
subject to (7.1)

where E, is user-cost-evaluated expenditure allocated to the services of real

transaction balances and of real supernumerary passbook savings deposits
during the current period.
We define the expenditure share of transaction balances in E: to be H’,,
=m,,$JE,*. The share of supernumerary passbook deposits then is M’~,=
1 = N’,,. After employing parameter transformations analogous to those in

“hTransaction balances are measured as the sum of M, plus NOW accounts tat all institution
types) plus share drafts at credit unions plus demand depostts at mutual savings banks.
Offenbacher’s (1979) results suggest that currency and demand deposits do not satisfy the
conditions for aggregation by summation, however separate treatment of those two components
requires imputation of separate own rates to each. In this paper we avoid such ambiguous and
controversial imputations. Hence we condition upon summed transaction balances as an
elementary good.
470bserve therefore that y,, = M,, and that ACIS hornothetic in real per capita transaction
balances and in aggregate real per capita supernumerary (not total) passbook savings deposits.
%ombining both stages of the decision over transaction balances and passbook savings
depostts, we find that consumers are viewed as allocating expenditure over transaction balances
and passbook savings deposits (either jointly or through the equivalent two-stage decisron) by
utility maximizatron (with habit formation in passbook savings preferences) to acquire desired
consumption levels. The desired level of transaction balances then 1s purchased without lags. In
addition the desired level of current total user-cost-evaluated expenditure on passbook savings
deposits services is actually consumed, but its distribution over institution types differs from the
desired allocation in accordance with the linear partial adjustment mechanism used in section
6.1.
32 WA. Barn&t, Economic monetary aggregates

section 6.1, we find that the solution to (7.1) can be written in the form

(7.2)

and

W pf= 1 -WI,,

where I7;, = II,(nzZ).

Let i’gr be the value of fI,(lrzr) with the parameters of 111,replaced by their
estimates acquired in section 6.3. We replace flzt with l?;,, normalize y2 to
equal 1.0, and estimate (7.2) with an additive disturbance term.49
Letting E, (t= 1,. . ., 7’) be the additive error in equation (7.2), we introduce
first-order autocorrelation by specifying that (Ed,. . .,+) is a sample from a
stationary scalar autoregressive stochastic process satisfying the stochastic
difference equation Ed= PEG_ , -I u,, where the sequence (IA,: t=2,..., T) con-
sists of independently and identically distributed normal random variables
with mean zero.5o The parameter p is subject to the constraint - 1 sp 5 1.
To impose that restriction, we let p =sin I/I. We eliminate that equality by
substitution and estimate the unconstrained parameter, I/I.“’

7.2. Estimates

The resulting maximum likelihood estimates of (y,,8, ti) are presented in

table 2. Transforming back to the original parameters of 1.4,we find that /I=
-2.53, ,?=0.96, and (&,,G,)= (0.77, 0.23), where (~,,a,) have been re-
normalized to sum to one.52 The implied elasticity of substitution is l/(1 -/?)
=0.28. Substitutability between transaction balances and passbook savings

4’F~ss (1977) has considered the properties of such nested estimation procedures.
“The same value, p, is used in defining the error structure for each of the two demand
equations derived from (7.1). That procedure follows from Berndt and Savin (1975) when no
serial correlation of disturbances exists across equations.
“To estimate (7.2) with the additive autoregressive drsturbance, E,. we use the following
transformation. Let the right-hand side of (7.2) be written as,f(lr:,. f7:,; :‘r, 0). so that

w,,=pw,.r~I-cS(n:,.n*. p,.jll,o)--l!f‘(~[:.,~,,rr,*.,-,;~,,O)l.
If we add I-:,to the right-hand stde of (7.2), then tt follows that the disturbance to be added to
the right-hand side of the transformed equation is a,-ppc,_ r =I+.. So we can estimate the
transformed equation using maximumhkelihood estimation with a conventional disturbance, u,.
sZOur estimate of the intensity parameter, o(r, is more than three times our estimate of rt.
Hence we might deduce that transaction balances, RI,,, contribute to our monetary asset
economic quantity aggregate more heavily than our nested passbook deposits aggregate, upl.
However one should be cautious about viewing the intensity parameters as simple weights m
this case, since p is a nonlinear function rather than a linear weighted average.
 W,4. Barnett, Economic monetary aggregates 33

deposits is far lower than between passbook accounts at different institutions

types. The elasticity of substitution of 0.28 is too low and the precision of its
estimator is too high to justify a linear approximation (requiring infinite
elasticity of substitution) to p.

Table 2
Parameter estimakxa

ll ;’ i

0.597 1.20 1.29

(0.22) (0.17) (0.17)

“Standard errors in parentheses.

7.3. Functional index numbers

In the present section, our highest level aggregator function is p. Hence

our highest level economic quantity aggregate is u, =p(mir,u,(y,,)). The
nominal dual user cost aggregate is

where
~.=,?‘C1-p)
I I and fl=fi/(fi- 1).

In summary, we have acquired the following nested pair of quantity and

nominal dual user cost indices, with all indices normalized to equal 1.0 in the
first quarter. For passbook accounts we have the maximum likelihood
estimate of the normalized functional quantity index, Q,(y,), and its nominal
dual user cost index, P,(np,). For our higher level (M,-type) monetary asset
aggregate we have the maximum likelihood estimate of the normalized
functional quantity index,

and its nominal dual user cost index

P($, npl) = 77(x,,, 77,(75Jl77(71,‘,‘, 77&J,)).

8. Implications of estimates

While passbook accounts at different institutions are excellent substitutes.

we find no evidence to support equal weighting of the accounts across
institutions. Although a simple linear (Laspeyres-type) index of passbook
deposits may be useful. the conventional unweighted sum index should be
understood to be based upon accounting practice rather than upon any
economically meaningful index number construct. If one sought no more
than total dollar deposits in passbook accounts in all institution types, the
use of simple summation would be dictated tautologically by an accounting
identity.
The simple sum index in economics corresponds to the degenerate limiting
special case of preferences having linear indifference curves at 45 degree
angles, and the corresponding dual price index is the poorly behaved
Leontief fixed coefficients index. In our case, consumers would use passbook
accounts in only one institution type, unless all institutions paid the exact
same interest rate. If all institutions did pay the exact same interest rate, then
the budget constraint would lie on top of a linear indifference’curve, and
consumers would not care how they allocated funds over institution types.
No unique solution would exist. But in fact commercial banks pay lower
interest rates than the other two institution types. yet acquire stable non-zero
deposits. Since passbook accounts across institution types do provide very
similar services, we should expect to find even poorer support for the simple
sum index at higher levels of aggregation within the money market, and that
conclusion generally is supported by our results with transaction balances at
the next aggregation level.
When we pass to a higher level of aggregation to incorporate transaction
balances into our monetary aggregate, the possibility of a useful linear
approximation, even with unequal coefficients, disappears. Transaction bal-
ances and passbook savings are not perfect substitutes and possess an
elasticity of substitution of only 0.28. The usual simple sum monetary
quantity index is rejected. The current M, aggregate provides useful account-
ing information on commercial bank liability structure, but is badly designed
as an economic monetary quantity index. The use of simple sum monetary
quantity aggregates as economic indices of the quantity of monetary services
should be discontinued.

9. Empirical selection of blocking

9.1. Conditions 011elristic’ities cf suhtitution

In section 4.1, we we selected our homothetic weakly separable blocking of
the current period conditional utility function, u, on a priori grounds. That
blocking then dictated the components of each subindex and index at all
levels of aggregation within our hierarchy of aggregates. Conditionally upon
that blocking, we have determined in sections 6, 7, and 8 that the form of the
aggregator function over the components of each index precludes use of
aggregation by simple summation. In the current section we briefly consider
the possibility of formally testing for the blocking itself, rather than solely for
the form of the preblocked utility (aggregator) function.
We begin with the current period monetary asset utility function, I, for
the vector of real supernumerary per capita holdings, yr, of all monetary
assets in the economy. We seek a partitioning, yr = (y’,,, . ., yh,)‘, such that I(
can be written in the blockwise weakly separable form

with uk linearly homogeneous for all k= 1,. ., M. The existence of such a

homogeneous weakly separable blocking is necessary and sufficient for the
existence of consistent quantity aggregation (to the functional quantity
aggregates. u, (y,,), ~~(y~,), . ., ~~(y(~,~)).~~ Clearly our earlier a priori block-
ing, (4.1) was a special case of (9.1) with one dimensional yi, and with
M=2.
Necessary and sufficient conditions for that homogeneous weakly separ-
able blocking are that the elasticity of substitution between any component
of yk, (for fixed k= 1,. . ., M) and any (supernumerary) monetary asset not in
ykt be independent of the element of ykt selected. We shall refer to those
conditions on elasticities of substitution as the Aggregation Conditions.
Systematic testing for those conditions with monetary assets has not yet been
undertaken and is a promising area for future research.54 However Barnett’s
(1981) Chapter 7 contains elasticity of substitution estimates (without formal
separability hypothesis tests) between many categories of monetary assets.
The conclusions suggested (at unknown statistical significance levels) by
comparisons of those elasticity of substitution estimates follow.

9.2. Empirical evidence

Referring to Barnett’s (1981) Chapter 7, we observe the following. Over the

past decade substitutability among passbook accounts at the three institution
types (commercial banks, S&Ls. and MSBs) has risen substantially and to
high level (&=2.66 jointly). In addition substitutability is high between small

53The conditions could be substantially weakened by dropping the homogeneity condition, if

we permit Fisher’s factor reversal test to be violated.
54A testing approach potentially applicable to that problem is contained in Denny and Fuss
(1977).
time deposits at S&Ls and MSB (c?= 12.82). However substitutability is low
between time deposits at commercial banks and at either of the two thrift
institutions.55 In general, substitutability within the many diverse groups of
financial assets considered in Barnett (1981) has tended to rise over the past
decade. However, with the exception of the two cases just described,
substitutability between financial assets has remained cer!’ 10w.~‘j
We now consider the implications of those elasticity of substitution
estimates for the selection of the components of aggregates. From Barnett
(1981), we find that the elasticities of substitution between passbook accounts
at different institution types are far higher than the elasticities of substitution
between passbook accounts at any one of those institution types and any
other financial asset. Hence any aggregate (such as the old M-2 index) which
contained passbook accounts at some but not at all institution types would
violate the Aggregation Conditions. Similarly we find that any aggregate
containing small time deposits at S&Ls must also contain small time deposits
at MSBs. In short, the empirical evidence in Barnett (1981) tends to support
aggregation of like-assets over institution types, as proposed in Barnett, Beck,
Ettin, Kalchbrenner, Lindsey, Porter, Simpson, and Tinsley (1979).
In sections 6, 7, and 8, we considered the separate question of whether
aggregation over given components can be accomplished by simple sum-
mation. Aggregation by summation is a special case of linear aggregation.
The necessary and sufficient conditions on elasticities of substitution for
linear aggregation are infinite elasticities of substitution between all components
bt,ithin the uggregutr. We call those conditions the LineuritJ Conditions. The
frequently very low elasticities of substitution found in Barnett (1981) further
strengthen our rejection of the Linearity Conditions in sections 6, 7, and 8.
It should. however, again be observed that our inferences drawn from
Barnett (1981), without formal statistical testing, are highly tentative. Our
conclusions in this section should be viewed as suggestive of areas for future
research through systematic hypothesis testing with models specifically
designed for that purpose.

55Those individuals who purchase small time deposits at commercial banks perceive them to
possess properties that are, in some ways, significantly different from those of small time deposits
at S&Ls or MSBs. This result is not surprising since those individuals who purchase small time
deposits at commercial banks generally are locked into the lower yields paid by the commercial
banks, as a result of the penalty structure imposed on early redemption. In fact it would be
difficult to understand why anyone would hold commercial bank small time deposits if he
considered them to be close substitutes for small time deposits at thrift institutions.
5”Earlier published studies of substitutability between monetary assets have all indicated very
low substitutability between monetary assets. Hence our results are in general agreement with
the earlier findings, and our finding of current high substitutability between passbook accounts
at the three institution types and between small time deposits at thrift institutions are thereby
strengthened by contrast.
10. Statistical index numbers

10.1. Definition

In the prior sections, we have been using aggregation theory. In aggre-

gation theory, aggregator functions are utility functions for consumers and
production functions for firms. Aggregator functions provide the foundations
of aggregation theory, and hence their existence and properties are important
in understanding aggregation. By estimating aggregator functions in the prior
sections, we have acquired information regarding the components of con-
sistent aggregates. and we have determined that aggregator functions defined
over financial assets cannot be adequately approximated by simple sum-
mation. Aggregation theory itself then would leave us with the alternative of
using the actual nonlinear aggregator function in aggregating over monetary
assets.
However, as we have seen, functional quantity aggregators depend upon
the quantities of the component goods and upon unknown parameters.
Estimates of the unknown parameters depend upon the specified model, the
data, and the estimator. Hence aggregator functions, although important in
theory and in hypothesis testing, are not generally useful in constructing
index numbers which are publishable as data by governmental agencies. For
precisely that purpose, the theory of statistical index numbers has been
developed. We introduce and then use that highly practical theory in this
section.
A functional quantity aggregator depends only upon component quantities
and unknown parameters. Functional quantity aggregators cannot depend
upon prices, and the definition of a functional quantity aggregator does not
depend upon maximizing behavior by economic agents. On the other hand,
statistical index numbers do not depend upon any unknown parameters, but
quantity index numbers can depend upon component prices as well as upon
component quantities, and the definition of exact statistical index numbers
does depend upon the maximizing behavior of economic agents. In brief, the
introduction of prices (and maximizing behavior in the exact case) into index
number theory permits us to dispense with the unknown parameters that
exist in the aggregator functions. The merits of the resulting index numbers
are not dependent upon any specialized properties of the aggregator function
(such as linearity of the function).
A quantity index between periods t - 1 and t, Q(q l,n,; M,_ 1, mt), is a
function of the vectors of prices (user costs) in periods t - 1 and t, rc_ , >O
and nL,>O, and the corresponding quantity vectors, m,_, >O and m, >o.57
Diewert defines such an index to be exact for a given aggregator function, J

57Given a quantity index the correspinding price mdex can be computed from Fisher’s weak
factor reversal test. See Diekert (1976, p. 115).
 if Q(n,_,.?I,;m,~,.m,)=,J‘(m,)~~(m,~,) whenever m,>O is the value of m>O
which maximizes f(m) subject to $ mzn;m,. In other words, an index
number is exact if it exactly equals the aggregator function whenever the
data is consistent with microeconomic maximizing behavior.“” Since the
aggregator function depends only upon quantities. the index number is a
quantity index number despite the existence of prices in its formula.
Two particularly noteworthy contributions exist in the recent literature on
index numbers. Hulten (1973) has proved that in continuous time the Divisia
index is always exact for LII~J consistent (blockwise homothetically weakly
separable) aggregator function.“” Hence no index number can be better than
the Divisia in continuous time. Although no always-exact index numbers are
known in the discrete time case, Diewert (1976) has constructed an elegant
theory of superlative index numbers in discrete time. Diewert defines an
index number to be ‘superlative’ if it is exact for some aggregator function,J,,
which can provide a second-order approximation to any linearly homo-
geneous aggregator function. We call such an index number Diewert-
.superlative.
Fisher (1922) advocated the following quantity index number, called the
Fisher Ideal index:

Tornquist (1936). and subsequently Theil (1967), advocated the following

quantity index number. called the TiirnquisttTheil Divisia index:

where sit = 7ci,miJ~~= , nkpxkt. Taking logarithms of each side, observe that

log QT - log Q,‘_ , = 1 s; (log mit - log nr,, t , ), (10.1)

i=l

581n this paper we do not consider the sophtsttcated issue of aggregating over economtc
agents. Relevant references are Muellbauer (1976). Dixon (1975). Barnett (1979a, 1979b), and
Maks (197X). The form of the index numbers does not depend upon whether the a’ggregator
function is a utility function or a production function. If dtstributional data were available on
shares held by firms (LS. households) or by different categories of wealth holders. that information
could be incorporated directly into the tndcx number. SW Thcil (1967. ch. 5) for an Information
theoretic interpretation of the resulting index numbers.
“The Divisia index is the line integral defined by the differential

d log Q = 1 s, d log q, where .A,= p,z,Jp’x

11
where ,sZ=i(si,+si,,_ ,). The same index numbers result, regardless of whe-
ther the aggregator functions are utility functions or production functions.
Diewert (1976) has proved that both the Fisher Ideal and Tornquist-Theil
Divisia indices are Diewert-superlative. In addition, as can be seen from
(lO.l), the Tiirnquist-Theil Divisia index provides a discrete time approxi-
mation to the optimal continuous time Divisia index. In fact the Tiirnquist-
Theil Divisia index can be derived by numerical integration of the Divisia
line integral. The TiirnquisttTheil Divisia index and the Fisher Ideal index
are highly reputable throughout all segments of the current literature on
index numbers, both for their statistical and economic properties.
As a quantity index, the Tiirnquist-Theil Divisia index is more widely used
than the Fisher Ideal index, since eq. (10.1) permits a natural interpretation
of the index. Observe that the growth rate of the index is a weighted average
of the growth rates of the components. The weights are the share contri-
butions of each component to the total value of the services of all
components. Because of the ,availability of that transparently clear in-
terpretation, we advocate use of the Tiirnquist-Theil Divisia index to
measure the quantity of money at all levels of aggregation (higher than Mi ).

In this section we consider the case of an aggregate having the following

components: transaction balances, passbook savings at the three institution
types and at credit unions, small time deposits at the three institution types,
and negotiable and nonnegotiable large C.D.‘s at commercial banks. The
components were selected on the basis of ready availability of the data rather
than as a proposal.hO The collection of components will be called M,. Table
3 displays the GNP velocity of the Tiirnquist-Theil Divisia index, of the
Fisher Ideal index, and of the simple sum index for seasonally adjusted data.
Velocity is normalized to be one in the first quarter. Observe that the
velocities of the Fisher Ideal and TiirnquisttTheil Divisia indices are iu’entical
to three decimal places, so that the choice between those two indices is of no
importance.61
However the ordinary simple sum index differs substantially from the two
Diewert-superlative indices. In addition the range of values of the velocity of
the sum index (0.201) is more than twice that of the superlative indices
(0.089). The velocity of the simple sum index (labeled ‘M, simple sum’) and

“The proper procedure for selecting components is described in section 9, but we seek only
an example in the current section.
“This phenomenon resulted from the fact that each is a Diewert-superlative index number.
Hence if an aggregator function exists and maximizing behavior obtains, then the two indices
can differ only by a third-order remainder term. In addition, each of the two indices should
agree with the unknown aggregator function equally as well as they agree with each other, since
the remainder term is of the same order in either case.
40 NA. Burnett, Economic monetary aggregates

of a Diewert-superlative (labeled ‘M, Diewert-sup’) index are plotted in fig. 1.

The Diewert-superlative indices are too close to be plotted separately.
The velocity of the simple sum index continues declining secularly from
1972(3), while the velocity of the Diewert-superlative index rises. Our
aggregate does not include many money market instruments such as RP’s,

Table 3
GNP velocities (seasonally adjusted data).

Fisher TGrnquistkTheil Simple

Quarter Ideal Divisia sum

1968(l) 1.0000 1.oOoO 1.oooo

1968(2) 1.0141 1.0141 1.0141
1968(3) 1.0131 1.0131 1.0094
1968(4) 1.0088 1.0088 1.0027
1969(l) 1.0174 1.0174 1.0188
1969(2) 1.0311 1.0312 1.0385
1969(3) 1.0524 1.0525 1.0713
1969(4) 1.0570 1.0571 1.0793
1970(l) 1.0626 1.0626 1.0825
1970(2) 1.0574 1.0574 1.0671
1970(3) 1.0473 1.0472 1.0412
1970(4) 1.0231 1.0229 1.0072
1971(l) 1.0219 1.0217 0.9975
1971(2) 1.0123 1.0121 0.9846
1971(3) 1.0041 1.0037 0.9726
1971(4) 0.9997 0.9993 0.9617
1972(l) 1.0031 1.0027 0.9611
1972(2) 1.0013 1.0009 0.9553
1972(3) 0.9919 0.9916 0.9419
1972(4) 0.9942 0.9939 0.9407
1973(l) 0.9998 0.9996 0.9365
1973(2) 0.9977 0.9976 0.9241
1973(3) 1.0061 1.0060 0.9205
1973(4) 1.0172 1.0171 0.9248
1974(l) 1.0104 1.0103 0.9065
1974’(2) 1.0234 1.0233 0.9019
1974(3) 1.0348 1.0347 0.9043
1974(4) 1.0339 1.0338 0.8997
1975(l) 1.0170 1.0169 0.8823
1975(2) 1.0263 1.0262 0.8895
1975(3) 1.0517 1.0516 0.9116
1975(4) 1.0555 1.0554 0.9127
1976(l) 1.0670 1.0668 0.9209
1976(2) 1.0681 1.0680 0.9206
1976(3) 1.0661 1.0660 0.9 I63
1976(4) 1.0608 1.0607 0.9097
1977(l) 1.0709 1.0707 0.9 164
1977(2) 1.0815 1.0814 0.9237
1977(3) 1.0801 1.0800 0.9197
1977(4) 1.0766 1.0765 0.9128
1978(l) 1.0733 1.0732 0.9057
 WA. Barnett, Economic monetary aggregates 41

treasury bills, commercial paper, money market funds, etc., while our
aggregate includes many assets subject to governmental rate regulation.
Hence we should expect substitution (disintermediation) to occur out of our
aggregate and into such substitutes during periods of rising interest rates and
high inflation, if our M, index approximates an economic monetary good. In
such cases velocity should rise. Clearly the declining velocity of the simple
sum index is very misleading.

Logarithm of normalized velocity

11.105

- ,980

- ,896

1968 1970 1972 1974 1976 1978

auarter

Fig. 1. Seasonally adjusted velocity (normalized).

Comparing fig. 1 with the ten-year government bond rate in fig. 2, we see
that variations in the velocity of the Diewert-superlative index make econ-
omic sense; the interest elasticity of money demand has the right sign.
Internalizing further money market substitution by aggregating over further
money market instruments can be expected to further stabilize the velocity of
the superlative index. The substitution effect (defined to hold utility constant)
of a change in the relative prices of components within an aggregate cannot
change the value of an economic quantity aggregate (utility level)!
In contrast, the trend in velocity of the simple sum index would suggest
that, in response to rising interest rates and rising inflationary expectations,
monetary asset holders have increased the fraction of GNP allocated to
consumption of the services of the lowest yielding (largely rate controlled)
sector of the market.62 Disintermediation thereby would appear (mislead-
ingly) to have proceeded within the money market in the wrong direction. It
is not surprising that simple sum aggregates frequently provide conflicting
information.

Percent per annum

r 19.03

j 6.49

i 593

j 5.42

LLLuLLdu.d LLLLLLLI L - 4.95

1968 1970 1972 1974 1976 1978
Quarter

Fig. 2. Ten-year government bond rate

The computational reason for the divergence between the Diewert-

superlative and sum indices can be seen from eq. ( 10.1).h3 The TGrnquist&
Theil Divisia index (or therefore, approximately, any Diewert-superlative
index) weights transaction balances more heavily than any of the other
components of the aggregates, since transaction balances provide the largest
share of monetary services, sir. An economic reason for the heavy weighting
of transaction balances is that their liquidity contributes heavily to monetary
services. But the velocity of transaction balances has been rising rapidly in
recent years. Hence the inadequate weighting of transaction balances in the

“‘Since GNP does not include the user-cost-e\.aluatetl services of durables or of monetary
;ISVAS. OLII-conclusion k based upon the UK of GNP a\ an appr’~~\lmation to the corrc.\ponding
theoretical national product concept.
“It is tempting to conclude that the reason the velocity of the Diewert-superlative index
tracks the gobcrnment bond rate is the fact that the Dleuert-superlatl\e index depends upon
Interest rates. However the index is constructed to approximate the aggregator function. which
depends only upon quantities and therefore not upon R,.
simple sum M, has permitted velocity to be drawn down by the substitution
effect of the increasing relative price (user cost) of transaction balances
relative to less liquid monetary substitutes.
To further verify our interpretation, we now incorporate elements of the
unregulated money market into n/l, to create M;. We incorporate dealer
and directly placed commercial paper, repurchase agreements (R P’s) of
commercial banks with the non-bank public, bankers’ acceptances, and
negotiable Treasury securities with less than one year remaining to maturity.
In fig.1 we plot the velocity of M.T. with Ml computed as a simple sum
index (labeled ‘Ml simple sum’), as a Diewert-superlative index (labeled ‘Ml
Diewert-sup’), and as a chained Laspeyres index (labeled ‘111.: Laspeyres’).
We continue to normalize all velocities to equal 1.0 in the first quarter.
Clearly internalizing those additional segments of the money market has
further stabilized the velocity of the Diewert-superlative index. The velocity
of the simple sum index continues to trend in the wrong direction. The
Laspeyres index is seen to provide a far better approximation than the
simple sum index, despite the fact that the Laspeyres index provides only a
first-order approximation to the value of the aggregator function.‘” The
slight variations remaining in the velocity of the Diewert-superlative index
continue to correlate with the ten-year bond rate and to reflect the fact that
some elements of the unregulated money market remain outside of the
aggregate.”
Further information regarding the comparison between the simple sum
and the Tiirnquist Theil Divisia index of money supply has been provided
by Barnett and Spindt (1979. 1980). They used information theory to
compute the information gained about various final targets (GNP, the
inflation rate. etc.) from knowledge of the rate of growth of a measure of the
money supply. They found that, regardless of the targets used or of the
components included in the aggregate, the TBrnquist~-Theil Divisia index
dominated the sum index computed using the SU~JZEmonetary components.
In addition, the gains in information acquired by going from the sum to the
Tiirnquist- Theil Divisia index frequently were extremely large.
The simple sum index is a Laspeyres quantity index with the weights
erroneously set to be equal. Clearly the arbitrary weighting destroys the
index’s critical independence of substitution effects (within the aggregate),
and hence the simple sum index c~~n~ot approximate the economic aggregate.

“‘The Diewert-superlative indlces provide a second-order approximation.

h5An entirely rigorous conclusion would be based upon the observation that the velocity of
the Diewert-superlative index reveals (to the second-order) movements along the aggregator
function and therefore movements of the underlying economic aggregate. Hence fig. 1 indicates
that the velocity of the simple sum Index has been moving rn the wrong direction, in the sense of
moving in the direction opposne to that of the economic aggregate.
11. Conclusion

In computing monetary quantity indices, the simple sum index number

formula is not satisfactory. The aggregates should be computed as
Tiirnquist-Theil Divisia indices. The components of the aggregates should be
selected to satisfy the conditions for consistent aggregation described in
section 9. These conclusions apply so long as the indices are to be used as
quantity indices of monetary services. as required in economics. Simple
summation would provide valid indices of the stock of nominal monetary
wealth, as required in national accounting, or indices of bank liability
structure, as required in bank accounting, but MN valid structural economic
variables.
The discussion in this paper has related to the economic theories of
aggregation and index numbers. However, there is also a statistical theory of
index numbers, which does not depend upon economic theory for its
foundations. Statistical index number theory considers the ability of index
numbers to pass certain classicial tests, such as factor reversal and circu-
latory tests. During the past decade, results from both approaches have
converged on the Tiirnquist Theil Divisia and Fisher Ideal indices as being
clearly among the best, and advocates of both the economic and statistical
approaches view the simple sum index as being among the very worst index
numbers ever devised.hh
According to Fisher, the two worst statistical properties that an index
number can possess are called ‘bias’ and ‘freakishness’. Regarding the simple
sum (or equivalently the arithmetic average) index, which Fisher called his
Formula 1, Fisher (1922, p. 363) observed: ‘There are two objections to
Formula 1, the simple arithmetic, l-i-_.: (1) that it is ‘simple’, and (2) that it is
arithmetic! -~ that it is at once freakish and biased. In the case of
Sauerbeck’s index number, for instance, the bias alone reaches 36 percent!’
In our case we found that the simple sum index dismally failed to
internalize the long-run substitution effects that have occurred within the
money markets during the past decade, and Barnett and Spindt (1979, 1980)
found that much component information is unnecessarily lost when
monetary components are aggregated by simple summation.h7 These results
are consistent with our empirical rejection of aggregator function constraints
that are necessary for economic aggregation by simple summation. As Fisher
(1922, p. 361) long ago deduced correctly: ‘The simple arithmetic (Formula
1) should not be used under any circumstances.’

‘*As we have observed, Hulten’s and Diewert’s work strongly supports the economic
foundations of the Tiirnquist Theil Divisia and Fisher Ideal indices. In addition Fisher (1922)
and Theil (1967) strongly support those same Indices on the basis of their statistical index
number properties.
“-I-‘or an overview of additional ongoing research in this area, see Barnett. Offenbacher and
SDindt (1981).
 We conclude with the following quotation from Fisher’s (1922, p. 29)
classical book, written over half a century ago:
‘The simple arithmetic average is put first merely because it naturally
comes first to the reader’s mind, being the most common form of
average. In fields other than index numbers it is often the best form of
average to use. But we shall see that the simple arithmetic average
produces one of the very worst of index numbers, and if this book has
no other effect than to lead to the total abandonment of the simple
arithmetic type of index number, it will have served a useful purpose.’

Appendix: Duality

A quantity aggregator function and its corresponding price aggregator

function are duals. The mathematics of function duals is not the subject of
this paper and will not be discussed in detail below. Nevertheless, those
familiar with classical duality relationships will recognize the foundations for
the following observation. We begin with the two stage decision of section
4.2.
Dual to the (quantity aggregator) function ~~(y,,,) exists the function
(L’,(Q) such that the identity u,(y,,)f7,(71,,)=yb,71,, will hold whenever y,,, is
the solution to the problem

min YbtzPt.
yp,
subject to
$(YpI)= k,,

where k, is a positive constant.

This duality relationship demonstrates that knowledge of the function u is
sufficient for determination of the function f7,. Hence we need only estimate
the conditional demand system solving (4.3) to estimate 77, and therefore to
compute estimates of the passbook real user cost index, 17zt=77,J$). We
thereby can acquire 17;* without estimating the higher level utility function,
~1, of eq. (4.1). Hence in (7.1) we were thereby justified in treating 17zt as
given and recursively estimating the utility tree from the bottom up. In fact it
can be shown that I7,(lr$) is just real expenditure on passbook account
services, Y;&, divided by the indirect utility function corresponding to
uP(Y,,). Since preferences are assumed to be homogeneous of degree one, it
follows that the resulting function, LIP, depends only upon R:~ (and is
independent of expenditure on passbook account services).
The function J7, is homogeneous of degree 1. Hence IZ,(n~~)=L’,(~,,)/p:.
As a result, we can compute the real value of the user-cost price aggregate,
I~,(R,,)/P:, by using real user costs, zz,. as arguments for II,,. Thus our earlier
observation that our estimates do not depend upon the use of p: is verified.
It is interesting to obser\Je that this nesting process immediately can be
carried to a higher level to acquire a user cost index for the economic
aggregate over passbook accounts and transaction balances taken jointly.
Since economic quantity aggregates always are utility functions, the quantity
aggregate immediately is seen to be ~~,=/~(~~z~,.~~,(r,,))=~(m,,,~r,,). We define
the dual user-cost index by observing that dual to the function (quantity
index) ~(rn II, IL,,,) exists a function (price index) 17(71,,. 17,,)=11(7r,,. Il,(n,,,))
such that the identity

will hold whenever (M,,, u,,) is the solution to the problem

where k, is a positive constant.

By Fisher’s factor reversal test (equality of expenditure to the product of
the price and quantity index). the price (user-cost) index dual to a functional
quantity index must equal total expenditure on the aggregated assets divided
by the indirect category (conditional) utility function defined on those assets.
Because of our linear homogeneity assumption on category utility functions.
total expenditure cancels out of the quotient leaving a functional price index
depending solely upon prices.

References

Anderson, Ronald W., 1979, Perfect price aggregation and empirical demand analysis,
Econometrica 47, 1209 1230.
Arrow, Kenneth J. and F.H. Hahn, 1971. General competitive analyw (Holden-Day, San
Francisco, CA).
Barn&t, William A., 1976, Maximum likelihood and iterated Altken estimatmn of nonlinear
systems of equations, Journal of the American Statistical Association 71, 354 360.
Barnett, Wilham A.. 1977a. Recursive subaggregation and a generalized hypocycloidal demand
model, Econometrica 45, I 117-l 136.
Barn&t, William A., 1977b, Pollak and Wachter on the household production function
approach, Journal of Political Economy 85, 1073 1082.
Barnett, William A., 1978. The user cost of money. Economics Letters 1. 145- 14Y.
Barnett, William A.. 1979a. Theoretical foundations for the Rotterdam model. Review of
Economic Studies 46. 109 130.
Barnett, William A., 1979b, The joint allocation of leisure and goods expenditure, Economctrica
45, 1117~-1136.
 WA. Bartwrt, Economic monetary aggregates 41

Barnett, William A., 1980, Divisia Indices, in: Norman Johnson and Samuel Katz, eds.,
Encyclopedia of stattstical sciences 1 (Wiley, New York) forthcoming.
Barnett, William A., 1981, Consumer demand and labor supply: Goods, monetary assets, and
time (North-Holland, Amsterdam) forthcoming.
Barnett. William A. and Paul A. Spindt, 1979. The velocity behavior and information content of
Divisia monetary aggregates, Economics Letters 4. 51- 57.
Barnett. William A. and Paul A. Spindt. 1980. The informatton content of Divisia monetary
quantity indices. Mimeo. (Federal Reserve Board, Washington, DC).
Bartlett. William A.. Edward K. Offenbacher and Paul A. Spindt. 1981, New concepts of
aggregated money. Journal of Finance, forthcoming (Papers and Proceedings of the 39th
Annual Meeting of the American Finance Association).
Barnett, William A., D. Beck, E. Ettin. J. Kalchbrenner, D. Lindsey, R. Porter, T. Simpson and
P. Tinsley, 1979, A proposal for redefining the monetary aggregates, Federal Reserve Bulletin
65, 13 42.
Berndt, E.R. and N.E. Savin, 1975, Esttmation and hypothesis testing in singular equation
systems with autoregressive disturbances. Econometrica 43, 937-957.
Bisignano, J., 1974, Real money st!bstitutea. International Economic Review, forthcoming.
Blackorby, D., D. Primont and R.R. Russell, 1978, Duality, separability, and functional structure
(North-Holland, Amsterdam).
Chetty, V.K., 1969, On measuring the nearness of near-moneys, American Economic Review 59,
27@281.
Clements. Kenneth W., 1976, A linear allocation of spending-power system: A consumer demand
and portfolio model. Economic Record 52. 182-198.
Clements, Kenneth W. and Phuong Nguyen. 1979, Money demand, consumer demand and
relative prices in Australia, Mimeo. (Reserve Bank of Australia, Sydney).
Denny, M. and M. Fuss, 1977, The use of approximation analysis to test for separability and the
existence of consistent aggregates, American Economic Review 67, 404418.
Diewert, W.E., 1974, Intertemporal consumer theory and demand for durables, Journal of
Econometrics 42. 4977516.
Diewert, W.E.. 1976. Exact and superlative index numbers, Journal of Econometrics 4, 115-146.
Dixon, Peter B., 1975, The theory of joint maximization (North-Holland, Amsterdam).
Donovan, Donal J., 1978, Modelling the demand for liquid assets: An application to Canada,
IMF Staff Papers 25, 676-704.
Enzler. J., L. Johnson and J. Paulus, 1976. Some problems of money demand, Brookings Papers
on Economic Activity 1, 261 280.
Fisher, I., 1922. The making of index numbers (Houghton Mifflin, New York).
Frisch. R. 1930. Necessary and sufficient conditions regarding the form of an index number
which shall meet certain of Fisher’s tests. Journal of the American Statistical Association 25,
297406.
Fuss, Melvyn A., 1977, The demand for energy in Canadian manufacturing, Journal of
Econometrics 5. 89 116.
Galbraith, John Kenneth, 1979, On inflation. Consumer Reports 44, 87-98.
Goldfeld, S.M.. 1976. The case of the missmg money. Brookings Papers on Economic Activity 3,
683 730.
Green, H.A. John. 1964, Aggregation in economic analysis (Princeton University Press,
Princeton, NJ).
H&en, C.R., 1973, Divjisia index numbers, Econometrica 63, 1017- 1026.
Klein, Benjamin, 1974, Competitive interest payments on bank deposits and the long-run
demand for money, American Economic Review 64, 931-949.
Maks, J.A.H., 1978, Consistency and consumer behaviour in the Netherlands, 1921-1962,
European Economic Review I 1. 343- 362.
Muellbauer. J., 1976. Community preferences and the representative consumer, Econometrica 44,
979 -999.
Offenbacher, Edward K., 1979, The substitutability of monetary assets, Mimeo. (Board of
Governors of the Federal Reserve System, Washington, DC).
Parkin, J.M.. R.J. Cooper, J.R. Henderson and M.K. Danes, 1975. An integrated model of
consumption. investment and portfolio decisions. in: Papers in monetary economics II
(Reserve Bank of Australia, Sydney).
48 WA. Burnett, Economic monetary aggregate\

Philips, L., 1978, The demand for leisure and money, Econometrica 46, 102551044.
Pollak, R.A., 1976, Habit formation and the long-run utility functions, Journal of Economic
Theory 13, 272-297.
Quirk, James and Rubin Saposnik, 1968, Introduction to general equilibrium theory and welfare
economics (McGraw-Hill, New York).
Samuelson, P.A. and S. Swamy, 1974, Invariant economic index numbers and canonical duality:
Survey and synthesis, American Economic Review 64, 566- 593.
Theil, H., 1967, Economics and information theory (North-Holland. Amsterdam).
Tornquist, L., 1936. The Bank of Finland’s consumptton price index. Bank of Fmland Bulletin
10, I-8.
Triplett, Jack E., 1976. The measurement of intlatton: A survey of research on the accuracy of
price indices, in: Paul H. Earl. ed., Analysis of inflatton (D.C. Heath, Lexington, MA) 19-82.