On the Dynamic Specification of Money Demand in Kenya1

Christopher Adam, University of Oxford

Recent developments in the theory of dynamic specification are applied

to the estimation of the demand for money in post-independence Kenya.
The models use a broad specification of the demand function, allowing
for parallel market currency substitution effects, from which robust
error-correction models of money demand are estimated for a range of
standard and Divisia monetary aggregates and are shown to
encompass existing studies.

1. Introduction

The demand for money is probably one of the most heavily researched aspects
of applied econometrics, both in industrialized and, increasingly, in developing
countries. Nevertheless in Africa there remains such a wide range of results,
particularly with respect to the behaviour of the demand function in the short-
run when agents are in a process of adjustment towards equilibrium, that the
issue remains relevant. This is particularly so since at the present time since
many of the policy objectives currently being pursued by sub-Saharan African
governments, including the Kenyan government, concerning financial sector
liberalization and the development of open-market techniques for monetary
control require the accurate calibration of the impact of relative price effects
in both official and parallel markets on the demand function. Closely linked
to this is the question of the implications of liberalization on the magnitude
and sustainability of the fiscal effects of seigniorage and inflation tax which

1 I am grateful to Paul Collier, David Hendry and John Muellbauer and an anonymous
referee for comments on earlier versions of this paper which is taken from my D. Phil
thesis. I am also indebted to Francis Mwega of the University of Nairobi for providing the
data used in his model (Mwega (1990)) discussed in Section 5 of this paper.
234 JOURNAL OF AFRICAN ECONOMIES, VOLUME 1, NUMBER 2

depend not only on the long-run inflation elasticity of the demand for money
but also on the dynamics of private sector adjustment towards its long-run
equilibrium. Again the failure of existing work to provide a comprehensive
analysis of both the long-run and dynamic considerations ensures that further
analysis of the demand for money remains necessary.
There are two reasons why now is a particularly good time to re-visit the
relationship in Kenya. The first is that the historical span of data is now
sufficient to allow for a comprehensive analysis of dynamic effects in the
demand for money while simultaneously allowing for the underlying long-run
equilibrium in the data to be identified. Hitherto, such dynamic analysis has
not been possible (see for example the papers by Pathak (1981), Darrat (1985)
and Mwega (1990) which are discussed in detail later in this paper). From a
technical perspective, the availability of data also allows us to use recent
techniques in the analysis of cointegrating time series relationships developed
by, inter alia Johansen (1988), the results of which suggest that such methods
may be profitably applied more widely on African macroeconomic data.
Second, the evidence from the Kenyan economy in the 1970s and 1980s,
particularly on the characteristics of the demand for money, may now provide
some indication of the likely evolution of monetary aggregates in countries,
such as Tanzania, who are currently undergoing economic liberalization
programmes. For many sub-Saharan countries past economic conditions make
it difficult, if not impossible, to infer much about the likely long-run
behaviour of the demand for money post-liberalization, and hence the long-run
behaviour of the demand function in Kenya may provide some valuable
archetypical evidence. Though the Kenyan economy has been subject to a
number of significant positive external shocks, principally the 1976/7 coffee
boom, it has also been prone to periods of macroeconomic imbalance in
which fiscal and monetary control was weak. Nonetheless, the economy did
not experience the severe economic crises common elsewhere in sub-Saharan
Africa: domestic inflation, though volatile, never exceeded 25 per cent per
annum and averaged approximately 15 per cent; domestic interest rates moved
gradually upwards over the period such that by the time domestic asset
markets were fully liberalized in mid-1991 only very marginal adjustments in
nominal interest rates ensued; while the relative lack of asset market
distortions has ensured that parallel markets, particularly in foreign currency,
have remained relatively small.
The remainder of the paper is structured as follows. Section 2 provides a
brief overview of die literature on the demand for money in developing
 Christopher Adam 235

countries, including Kenya, and discusses the nature of money demand for a
small open economy where the.non-bank private sector faces regulated official
domestic asset markets, capital controls and informal asset markets. In
addition, this section also considers the use of Divisia monetary aggregates as
an alternative pleasure of the demand for money. Section 3 discusses the data
problems in estimation and presents the results of the preliminary time-series
analysis of the data. Section 4 reports and discusses the results of the
cointegration analysis, while Section 5 extends the model to examine its short-
run dynamics and compares the results with existing work. Section 6
concludes by drawing together the rriain policy implications.

2. The demand for money in small open economies with capital

constraints and parallel markets

The empirical analyses of the demand for money in developing economies has
been dominated by research on Latin America and Asia (see Adekunle (1968),
and Coates and Kathkhate (1980)). Work on sub-Saharan Africa has been
somewhat more limited, although exceptions include the studies on Kenya
discussed below and a study of the Sudan by Domowitz and Elbadawi
(1987).2
The generic long run model which underpins this body of work, and which
nests the various microeconomic models of money demand,3 is of the form

(1) M" = kPya(l +nf(l+r)\

where P is the price level, y the level of real income, 71 the rate of inflation,
r the "interest rate", k a constant, and a, P, and y are long-run (semi-)
elasticities of demand. Generally, long-run price homogeneity is assumed such
that real money balances are expected to be increasing in real income and

2 Domowitz and Elbadawi's (1987) dynamic error-correction model for the Sudan from
1956 to 1982 is perhaps the most robust of the studies on Africa. Using annual data, an
etrpr correction model is estimated aro.un'd (he long-run proportionality between money
and income, where in the short-run domestic inflation and the US dollar exchange rate are
significant explanatory variables.
3, See, for example, Laidler (1985) for a discussion of how individual transactions,
portfolio and speculative motives for holding money can be aggregated into this general
form.
236 JOURNAL OF AFRICAN ECONOMIES, VOLUME 1, NUMBER 2

decreasing in inflation, while the elasticity with respect to the interest rate
may be positive or negative depending on the composition of the monetary
aggregate. The transition from this general specification to the analysis of the
demand for money in Africa raises a variety of issues. First, money held with
the commercial banking system is the principal financial asset in most sub-
Saharan economies. Alternative financial assets are scarce while equity and
bond markets are thin. Opportunities for substitution between money and other
financial assets are therefore more limited than in developed economies.
However there remain important substitution effects between money and real
assets which, in addition to their own intrinsic income flows, may substitute
for money holdings for speculative and precautionary purposes. Second,
borrowing and deposit interest rates have traditionally been administered by
the monetary authorities and, for much of the period since the 1960s,
domestic interest rates have been maintained at low or negative levels in real
terms. Third, however, in many cases there are thriving parallel markets in
foreign exchange and the demand function will involve currency substitution
considerations (see, for example, Blejer (1978) and Calvo and Rodriguez
(1978)). Failure to incorporate these effects may induce omitted variable
biases by overstating the substitution from money into domestic real assets.
The nature of currency substitution in the demand for money will depend
inter alia on the nature of the macroeconomic control regime. In a number of
economies the parallel market is large and foreign currency holdings
constitute a significant proportion of private sector wealth (for example the
Sudan, Uganda and Zambia), while in others, notably Kenya, the parallel
market in foreign exchange is illegal and more limited in scale.
Three published studies estimating demand for money functions for the
post-independence Kenyan economy can be cited. These are by Pathak (1981),
Darrat (1985) and Mwega (1990). Darrat and Mwega estimate the quarterly
demand for money4 for the period 1969 to 1978, and 1973 to 1988
respectively, while Pathak uses annual data for the period 1968 to 1978.5
Measures of income (Pathak and Mwega use GDP, Darrat GNP) and inflation
are included in all the models, but the specification of the asset market
variables differs across the models. For example, though Mwega and Pathak

4 Darrat estimates models for Ml and M2, Mwega for Ml, M2, and M3.
5 Consequently, the reported regressions show only 5 degrees of freedom which
invalidates any serious inference.
 Christopher Adam 237

follow standard practice by including the short-term Treasury bill rate to

proxy the return on financial assets, Darrat argues that since domestic asset
markets are thin, the domestic interest variable should be substituted by a
foreign interest rate variable. Similar differences exist with respect to the
specification of the dynamic structure of the model. Mwega estimates the
relationship using a partial-adjustment structure for the lagged dependent
variable, while Darrat admits no short-term dynamic effect of money itself,
although the explanatory variables are all estimated on the basis of a third
order Almon lag. Pathak's study estimates only a static regression.
Despite overlapping sample periods and a relatively congruent set of
explanatory variables, the models report widely differing elasticities for the
main variables (see below), and a wide range in their statistical significance.
These three models of the demand for money thus provide the benchmark for
this study, although in practical terms the Mwega (1990) model is the
dominant specification and it is this work to which we shall return in Sections
4 and 5 when we evaluate the results from our own estimation. It is clear,
however, that inconsistencies between these results can only be partly
explained by the small differences in the sample size: more important is the
composition of the vector of explanatory variables, which we consider in the
next section, and the appropriate econometric specification of the processes
of dynamic adjustment in the demand for money.
The origin of the debate on dynamic specification in time-series models
derives from the view that though equation (1) characterises the long-run, or
static, behaviour of the demand function it does not provide any insight into
the behaviour of agents out of equilibrium. The appropriate specification of
the relationship between the long-run theory and short-run dynamics has
dominated much of the time series econometric research in the 1980s and
represents the principal response to the collapse of many aggregate macro-
economic relationships in the 1970s (inter alia Davidson etal. (1978), Hendry
(1980, 1985), Hendry and Richard (1982)). This analysis, and particularly the
development of the ideas of cointegration and error-correction, have led to
important revisions in the modelling of macroeconomic relationships in recent
years, not least with respect to the demand for money. (See, for example,
Granger (1986), Engle and Granger (1987), Hendry (1986), Banerjee et al.
(1986) and (1992), Johansen and Juselius (1990).)
Research on cointegration in time-series data has suggested the relatively
widespread presence of cointegrating relationships between macroeconomic
238 JOURNAL OF AFRICAN ECONOMIES, VOLUME 1, NUMBER 2

variables,6 which can be efficiently modelled within an error-correction model

of the general (logarithmic) form7

(2) A(L)Am, = B(L)Az, - Tfm-kz),., + e,

where A(L) and B(L) are lag polynomials, z is a vector of explanatory

variables, and the second term on the right-hand side of (2) is the error
correction term which is the stationary linear (cointegrating) combination of
the non-stationary levels of the variables m and z, where it is a scalar.8
Early error-correction models of money demand, including that by
Domowitz and Elbadawi (1987), tended to be based around bivariate
cointegrating relationships between money and income or money and financial
wealth (see also Hendry (1985)). More recently, the analysis of cointegrating
relationships suggests that multivariate cointegrating vectors spanning a
broader number of variables including wealth, inflation and interest rates,
provide a fuller characterization of the long-run determinants of money
demand (see Johansen and Juselius (1990), Hendry and Mizon (1990) and
Adam (1991)). The identification of such multiple cointegrating vectors
between non-stationary variables employs procedures developed by Johansen
(1988) which nest the original Engle-Granger (1987) method for identifying
cointegration as a special case.9

6 A series is said to be integrated of order d if it becomes stationary after differencing d

times, but not d-l times. Such a series is denoted as x, ~ l{d). Using this terminology, a
stationary series is an 1(0) series, and a non-stationary series with a single unit root is 1(1).
Many (logarithmic) macroeconomic variables are 1(1) (Engle and Granger (1987)), but
some non-stationary series are of order 2 or higher, in which case the first difference or
growth rate of the series will be 1(1). In general, linear combinations of 1(1) series are
themselves non-stationary, but an exception is in the case of cointegration. Formally, if
y, - l(d) and x, ~ f(b) and the linear combination of the two, namely, z, = v, - kx, - l(d-
b) then x, and y, are cointegrated of order (d-b). The statistical properties of cointegrating
combinations of 1(1) variables allow long-run information (i.e. on the relationship between
the levels of the variables) to be combined with short-run information on their dynamic
behaviour within a stationary error-correction framework.
7 Engle and Granger establish the isomorphism between the presence of a cointegrating
relationship between variables and the error correction dynamic specification. TTius if data
are cointegrated then the error correction specification will be the dominant dynamic
parameterization of the model.
8 See Nickell (1985) and Domowitz and Hakkio (1985) for discussion of how this
functional form relates to standard models of adjustment costs.
9 An extended discussion of the techniques used in this subsection is available on request
from the author.
 Christopher Adam 239

The Johansen method commences from a standard vector autoregression

(VAR) of the form

(3) X, = FI,*,., + ... + F l ^ +/i + 0 D , + £,

where X is a p x 1 vector of the 1(1) variables of interest, u is a vector of

constants, D is a vector of centred seasonal dummies, and e, ....e T are IN(0,
A) error terms. Using the difference operator, A, (3) can be reparameterized
as:

(4) AX, = F;AX,.; + ... + ruAx,.t., +nx,.k + ^ + eD, + el

where F, = - (I - U, - ... - UJ (i= 1 k-1),

and n = - (I - Tl, - ... - n j

Equation (4) is a stationary error-correction first-differences VAR where the

term UXhk contains information on the long-run (levels) relationship between
the variables in the VAR. By extension from the bivariate cointegration case,
this term will be consistent with the stationary VAR only if it is also
stationary, and this will occur if the elements of X are cointegrated. The
number of cointegrating vectors, v, between the elements of X will therefore
determine the rank of the vector Fl. There are three alternative situations
depending on the rank of FI (i.e. the number of independent linear
combinations spanning the 1(1) vector space). If n is of full rank p, then the
matrix is stationary, and the data in vector X are not, in fact, 1(1). If the rank
is zero, then the variables are all individually 1(1) but not cointegrated, in
which case the error-correction isomorphism cannot be employed. Finally, if
the rank, v, is greater than zero but less than p, this means that there are v
cointegrated vectors which can be identified and can be embedded in an error-
correction model.10 In the latter case, n can be factorised as FI = -a(3' where
a and P are v x p, with (3 containing the v cointegrating vectors, and a their
corresponding adjustment, or feedback, parameters. The matrix P' contains the
parameters of economic interest in the VAR, and its vectors constitute the

10 The rank of the matrix n is established using the eigenvalues (i,, derived from the
minimization of the product matrix of the concentrated likelihood of (3), in the form of
tests based on the value of -Tlog(l-/j;), where T is the sample size. The critical values of
the statistic are computed in Johansen and Juselius ((1990) Appendix A3).
240 JOURNAL OF AFRICAN ECONOMIES, VOLUME 1, NUMBER 2

long-run (levels) relationships between the variables. The Johansen procedure

thus defines cointegrating relationships between the endogenous variables in
the VAR which represent cross-equation restrictions on the joint long-run
evolution of the variables. One implication of this result is that the presence
of the same cointegrating vectors across the different equations in the VAR
violates the implicit weak exogeneity assumptions employed when the
relationships are estimated individually. In such circumstances the individual
relationships may need to be estimated within a simultaneous equation model.
In Section 5 we examine whether such cross equation restrictions prevail in
the case of the demand for money function for Kenya.

3. Data definitions and characteristics

On the basis of this review of the theoretical literature we define the vector
of variables of interest in determining the demand for money in Kenya, X, as

(5) X = ( (m-p), (y-p), r, r, n),

where m-p is real money balances," y is some measure of nominal income,

r is the domestic rate of interest, r* a measure of the rate of return on holdings
of illegal foreign currency, and n the rate of inflation.12 Note that this
specification is sufficiently general to nest all the previous models. Empirical
representation of these variables are as follows, starting with the definitions
of money.
We estimate models for the money aggregates reported by the Central Bank
of Kenya, namely: MO (defined as notes and coin in circulation outside the
banking system); Ml (MO plus demand deposits with commercial banks); and
M2 (Ml plus time and savings deposits). In addition we examine the demand
for the broader measure of money, M3, which is defined as M2 plus the

11 The demand for real balances is assumed to be homogeneous of degree 1 in prices.

This neutrality restriction has been pre-tested and is accepted by the data for Kenya.
However we do not assume that real balances are neutral in terms of inflation.
12 The data cover the period 1973 to 1990 and are derived from official sources,
principally the Economic Survey (Central Bureau of Statistics), the Quarterly Economic
Review (Central Bank of Kenya), International Financial Statistics (IMF) and
International Currency Yearbook (Picks).
 Christopher Adam 241

deposit liabilities of the non-bank financial institutions (NBFls), but for this
latter aggregate we examine both its simple-sum and its Divisia
representations. Generally monetary aggregates such as Ml, M2 and M3 are
constructed as a simple sum of their components. However, for such
aggregates to have economically valid micro-foundations then it follows that
the components of the aggregate must be perfect substitutes, so that the utility
from holding a monetary aggregate is independent of its composition. In
reality, however, perfect substitutability does not hold (cash, demand deposits,
savings deposits and deposits with non-bank financial institutions, for
example, differ in terms of their rate of return, liquidity and risk), and thus the
simple sum aggregate is non-neutral with respect to the composition of the
aggregate. Work undertaken in the early 1980s has attempted to address this
deficiency by constructing monetary aggregates in which the weights reflect
the relative marginal liquidity of the component assets. Barnett ((1980),
(1982)) and others have suggested the Divisia Index as an appropriate
aggregate for discrete-time analysis.11 The Divisia monetary index, Q, is
defined

(6) logQ, - logQ,., = Is',, (logm,, - logmihl)

where the share weights s'jt = 'Ms-,, + sjhl) and slt = ^lpxt/lX^kfnkl. m,, i = l...k
are the asset components and 03, the user cost of each asset, is defined as

(7) Ot = p ( « , - a / + / ? J

where p* is the general price level, r,, the (tax-adjusted) current period yield
on the component m,, and R the "maximum available holding yield in the
economy for that period" (Barnett (1982) p 689). The Divisia thus uses the
interest rate differential between assets to define their relative liquidity, the
intuition being that higher interest rates are required to compensate the holder
for lower liquidity (or lower "moneyness") of a component asset. In an
economy where relative interest rates are market-determined this method is

13 Strictly, the Divisia index is an exact index (i.e. it is always equal to the underlying
implicit aggregator function when the data are generated by microeconomic maximizing
behaviourfonly in continuous time. There are no known exact indices in discrete time,
although Diewert's (1976) superlative indices provide second-order approximations to any
weakly separable aggregator function. The Divisia index belongs to this class of index.
242 JOURNAL OF AFRICAN ECONOMIES, VOLUME l, NUMBER 2

likely to produce a reasonable approximation to the true aggregator function

(see Barnett (1980) on US aggregates, Cockerline and Murray (1981) on
Canada, and Mills (1983) on the UK, all of whom find significant deviations
between simple-sum and Divisia representations of broad money aggregates).
In less monetized economies where interest rates are administered by fiat, the
share-weighting metric of the Divisia may constitute only a poor
approximation to the true relative liquidity of the component assets. Moreover,
in economies where there are sizeable parallel markets in foreign currency, or
where unrecorded trade represents a significant proportion of economic
activity, the utility of cash will be higher relative to equivalent non-interest
bearing demand deposits than would be indicated by the Divisia index based
on official interest rate differentials. Moreover, in the case where demand
deposits do not attract interest, the Divisia index will indicate a perfect
substitutability between cash and demand deposits since both are non-interest
bearing. Clements and Nguyen (1980) therefore suggest as an alternative that
the elasticity of substitution between cash and demand deposits may be
derived from auxiliary regression estimates of the elasticity of substitution and
then replaced in the Divisia index, although this is rarely done and most
indices ignore this specific form of substitution.
While Kenya does exhibit a number of these characteristics which may
render the Divisia imprecise, their gravity is less severe than elsewhere in sub-
Saharan Africa. For example, as noted, the parallel market is relatively small,
and although money market rates were administered until July 1991, a
discernible yield curve was maintained between different monetary
instruments. Divisia indices have therefore been computed for the three
monetary aggregates, Ml, M2, and M3. The Ml and M2 aggregates, as
expected, deviate very little from their simple-sum equivalents, but there is a
significant deviation on the part of the M3 and M3 Divisia aggregates with
the M3 growing significantly faster than the Divisia. The reason for the
deviation is that though in absolute terms there has been a dramatic rise in the
volume of NBFI deposits over the period, their relatively illiquidity vis a vis
other forms of deposit ensures that this volume increase is offset by their
relatively low user-cost share weight in the Divisia index.14 We therefore

14 Non-bank financial institutions are unable to offer checking facilities and withdrawal
of deposits is by passbook only, in many cases requiring advance notice. See Killick and
Mwega (1990).
 Christopher Adam 243

consider only the Divisia M3 aggregate (denoted M3d) as well as the simple
sum aggregates MO, Ml, M2, and M3.
The second issue concerns the choice of income series. Traditionally some
measure of GDP is employed as a proxy for income in models of the demand
for money. However in the case of an open economy such as Kenya the
volatility in the terms of trade means that taking GDP as a measure of income
will fail accurately to reflect the transactions demand for money. We therefore
use a gross national income (GNY) measure which is defined as Gross
National Product adjusted for changes in the terms of trade. The series
(denoted v) is reported only annually and an interpolation exercise is required
to create suitable quarterly data.'516
One final caveat which needs to be mentioned here is whether some
measure of wealth may be a more appropriate conditioning variable for broad
money aggregates. A growing body of literature, stimulated by the publication
of data on financial wealth for the UK, suggests that wealth serves as a more
appropriate conditioning variable than income for broad money aggregates
(see Adam (1991), and Barr and Cuthbertson (1991)). No such official data
however yet exist for Kenya on which to assess this hypothesis.17
The third issue concerns the price series. Most models of the demand for
money in developed economies use the GDP or GNP deflator. In an open
economy such as Kenya, however, the GDP deflator is not appropriate since
it is constructed as a value-added deflator which includes exports (which
domestic residents do not purchase), but excludes imports (which they do).
The CPI deflator, on the other hand, avoids this particular problem since it
includes imports and excludes exports, but being a consumption deflator it
excludes expenditure on investment goods. Since, however, the majority of

15 Details on the construction of this scries are available on request from the author.
16 The method adopted to create the quarterly income series ensures, by design, that the
interpolated series is consistent with known annual observations for GNY. However a
concern remains that the interpolated quarterly observations exhibit an artificial pattern of
seasonality which may be inconsistent with their true (but unknown) seasonally. Though
the seasonality of the series may be non-stationary, (see Osborn et al. (1988), and
Hylleberg et al. (1990)), we find that for all of the data used in this model the seasonality
is stochastic around the non-seasonal series, and can be efficiently modelled using a
dummy-variable regression model.
17 Evidence from a demand-share model in which asset demands are conditioned on a
constructed wealth stock variable does however support the claim that wealth rather than
income is the appropriate variable for modelling the demand for broad money (Adam
(1992)).
244 JOURNAL OF AFRICAN ECONOMIES, VOLUME 1, NUMBER 2

total expenditure is on consumption the CPI provides a reasonable first-order

approximation to the true price deflator.
Fourth, concerning interest rates we note that there are very few interest-
bearing assets held by the private sector and throughout most of the period
under examination domestic interest rates have been controlled by the
authorities. However, all interest rates have generally been adjusted in a
consistent manner over the period such that despite the absence of an active
market mechanism through which interest rate changes are transmitted, all the
main rates of interest have tended to move together over time, and we
therefore use the Treasury Bill rate (denoted rtb) as a proxy for the rate of
return on all (capital certain) financial assets. We note also that in an
economy such as Kenya where there is only limited substitution between
money and other financial assets, this rate serves principally as a measure of
the own rate of return on money.
Finally, in measuring the effect of currency substitution, we face a number
of difficult empirical issues. The expected return to holding foreign currency
is defined as a function of the rate of interest on foreign currency holdings
plus the expected depreciation of the home currency. In an open economy
with no capital controls and assuming rationality in the foreign exchange
market the return on currency substitution would equal expected depreciation
of the exchange rate plus, say, the dollar interest rate in the US or on the
Eurocurrency markets. However when considering illegal foreign currency
holdings, it is likely that interest rate considerations are less important and
that currency substitution is determined principally by the devaluation effect.
Thus in an economy such as Kenya where domestic agents are prohibited
from holding foreign currency balances the appropriate rate would be the
expected depreciation of the parallel market rate of exchange.18 In the absence
of the observability of this forward looking rate we take the depreciation of
the ex post actual parallel market exchange rate (denoted exb) to proxy the ex
ante expected rate of depreciation. The parallel market rate has remained
relatively close to the official exchange rate over the period, with the premium
averaging little more than 15 per cent over the official rate. Only in 1982 at

18 This assumes that if the parallel market exchange rate is freely determined it will fully
reflect expected inflation differential between the domestic and world economy (Blejer
(1978)).
 Christopher Adam 245

the height of the political and economic crisis did the premium exceed 25 per
cent.

Time series characteristics of the data

In this section we investigate the time-series characteristics of the data to

assess the possibility of cointegration in the data and to ensure consistency in
subsequent stationary econometric modelling. The Dickey-Fuller and Sargan-
Bhargava unit root tests are the standard tests of the null that p = 1 in the
regression

(8) y, = Oo + py,., + u, ; u, - (0, a2); y0 = 0.

Failure to reject p = 1 indicates that y, is non-stationary and will need to be

differenced at least once to render it stationary. The test results are reported
in Table 1, but because of known weaknesses in these tests (see for example
Banerjee et al. (1992)), we complement the full-sample tests by examining
their behaviour at each point in the sample using the recursive equivalent of
the tests.19

19 The actual Dickey-Fuller test is against the equivalent null H,,: P, = (1-p) = 0 in

(i) Ay, = Po + p;y,.; + e r

The distribution of the test statistic is a non-standard 't' for which Fuller (1976) has
tabulated the distribution. One of the drawbacks of the Dickey-Fuller test, however, is that
it necessarily assumes the DGP is an AR(1) process under the null. If it is not, then
autocorrelation in the error term in (i) will bias the test. In order to overcome this problem
the Augmented Dickey-Fuller (ADF) test can be used. The ADF is identical to the
standard DF test but is embedded within a regression model of the form

(ii) Ay, = p 0 + P,y,.; + XyyYy,., + u,

where the lag lengthy is set so as to ensure that any autocorrelation in Ay, is absorbed, and
the error term is distributed as white noise. The Sargan-Bhargava Durbin-Watson test is
based on the standard Durbin Watson statistic, but the test is applied not to the residuals
of the regression as usual, but on the level of each individual series. The test statistic is
defined as

(iii) DW(y) = Kyry,.,)2/Z(y,-yf.

246 JOURNAL OF AFRICAN ECONOMIES, VOLUME 1, NUMBER 2

Table 1:
Unit root tests on quarterly data

Variable DW DF ADF Longest Order of

lag integration

(mO-p) 0.324 -1.98 -1.32 4

(ml-p) 0.254 -1.92 -2.42 8
(m2-p) 0.139 -1.49 -2.91 8
(m3-p) 0.055 -1.28 -2.07 8
(m3d-p) 0.182 -1.27 -1.75 8
(y-p) 0.366 -2.19 -1.95 4
7i 0.285 -2.19 -1.89 4
rtb 0.061 -0.81 -0.83
exb 0.034 0.03 0.83 - 1

Note: With a sample of approximately 6 0 observations, values less than -2.93 represent
rejection of the null that p = l in favour of p<l (i.e. stationarity) at the 5 per cent critical
value for the D F and A D F tests. Value less than 0.65 represent rejection of the null than
p < l in favour of the alternative that p=l (i.e. non-stationarity) at the 5 per cent
significance level under the D W test. All series are in natural logs indicated by lower case
letters, with the interest rate expressed as r = Ln(I+rq) where rq is the rate of interest per
quarter.

The results are typical of many problems in time-series econometrics,

namely that although the DW statistics strongly reject stationarity, the DF and
ADF test statistics are, in many cases, close to their critical values. Though
not reported, the recursive test plots (which recalculate the test at every point
in the sample) echo these results by revealing a very high variance in the
value of the statistics through the sample. However, with the exception of a
small number of points in the sample space, it was not possible to
comprehensively reject the null of non-stationarity for any of the series. Thus
we can interpret these results as indicating that the five money aggregates,
income, inflation, and all the interest rates are indeed 1(1), or non-stationary.

Unlike the Dickey-Fuller tests, the test is against the null that the series is 1(0), in which
case the value or the DW statistic will tend towards a value of 2. If the statistic is low
then this is evidence of an 1(1) series.
 Christopher Adam 247

We therefore proceed to test for the presence of cointegration between the

variables.

4. Cointegration analysis and the long-run demand for money

The VAR corresponding to (3) above for Kenya consists of the five variables
in (5), and is estimated for each of the five money aggregates, without a trend
term, but with a constant and seasonal dummies.20 Seven lags were included
for each variable to ensure that the error term e, is distributed white noise.
The sample period extends from 1973(1) to 1989(2). Though containing 66
observations this sample is relatively short and, as is characteristic of high
frequency data estimated over short periods, is subject to a low signal-to-noise
ratio, thus requiring careful interpretation of the statistics. Table 2 reports the
maximal eigenvalue statistics, -7(7-/7,) derived from the minimization of the
product matrix of the concentrated likelihood of the VAR.2'
In each case the 1(1) vector space is spanned by at least one cointegrating
vector (the maximal eigenvalues for v>l are all greater than the 5 per cent
critical value), although there is evidence that there may also be two
cointegrating combinations between the variables: with the exception of the
(m2-p) aggregate the maximal eigenvalue statistic for v^2 is approximately
equal to the critical value. Since the sample is relatively short, and subject to
a relatively high level of noise, it is advisable to maintain the hypothesis that
there are at least two significant cointegrating relationships in each vector.
The dramatic fall in the eigenvalue statistics for v>3 suggests that it is not
necessary to search for a third vector between the variables. We therefore
accept v=2. Examination of the two significant rows of the p' matrix of
eigenvectors, and the corresponding loadings, a,22 indicates that for all the
aggregates except MO, the first cointegrating vector is dominated by the
relationship between the interest rate (rib) and inflation (7t) variables, whose
linear combination is virtually stationary relative to the other variables, and
seems to be capturing the rise in administered interest rates in absolute terms

20 All estimation is carried out using the PC-GIVE and PC-FIML routines within PC-
GIVE v6.01 (Hendry (1989)).
21 See Johansen and Juselius (1990) pp 174-180.
22 Full data and results can be supplied by the author on request.
248 JOURNAL OF AFRICAN ECONOMIES, VOLUME 1, NUMBER 2

(relative to the falling rate of inflation over the period). The second vector,
however, is more clearly interpretable in terms of a money demand function,
consistent with theoretical priors that the demand for money is a positive
function of income and the own rate of interest, and a negative function of the
rate of inflation and the expected devaluation of the parallel market exchange
rate.

Table 2:
Maximal eigenvalues l-T(l-ft)] and critical values for five-variable VAR

Rank of n (v) (mO-p) (ml-p) (m2-p) (m3-p) (m3d-p) 5%cv

5 2.37 2.68 0.89 0.30 2.45 9.09

4 5.58 5.79 5.55 5.37 6.54 15.72
3 12.34 10.56 10.51 10.03 8.26 21.89
2 36.50 30.39 19.13 26.38 26.20 28.16
46.28 46.75 50.15 57.91 49.91 34.39

In the case of the narrow money measure it was found that the order of the
significance of the two vectors was reversed, with the money demand
relationship being identified in the first vector and the interest rate relationship
in the second. The interest rate relationship is virtually identical across all the
aggregates, while the long run solutions to the demand for money functions
(i.e. the second row of the P' vectors) are all intuitively plausible. Because
the P' matrix is not identified we are obliged to impose identifying
restrictions on the matrix so as to define unique cointegrating relationships
(which may include linear combinations of the vectors). In this case
identification is facilitated by the fact that the interest rate and inflation
relationship can be approximated as

(9) (m-p) (y-p) n rtb exp

0.0 0.0 0.4 1.0 0.0

which reflects the convergence towards a constant positive real domestic

interest rate. Taking this approximation, which can be tested formally using
a standard likelihood ratio test, we can then derive a linear combination of the
first and second relationships to define economically interpretable demand for
money vectors for the five aggregates. This allows us to uniquely identify two
 Christopher Adam 249

cointegrating vectors for each aggregate which exactly span the vector space
of the P' matrix, along with their feedback vector a. These are reported in
Table 3.

Table 3:
Final cointegrating money demand vectors and loadings

(mO-p) Aggregate
Variable (mO-p) (y-p) 71
B'Eigenvectors 1.00 -1.01 6.15
a Coefficients -0.19 0.08 -0.03

(ml-p) Aggregate
Variable (ml-p) (y-p) 71 rtb exb
B'Eigenvectors 1.00 -0.89 5.46 -0.52 0.16
a Coefficients -0.36 0.07 -0.04 0.20 -0.04

(m2-p) Aggregate
Variable (m2-p) (y-p) 71 rtb exb
B'Eigenvectors 1.00 -0.84 6.73 -2.25 0.11
a Coefficients -0.23 0.03 -0.06 0.33 -0.03

(m3-p) Aggregate
Variable (m3-p) (y-p) 71 rtb exb
B'Eigenvectors 1.00 -1.10 6.19 -18.14 0.09
a Coefficients -0.09 -0.06 -0.06 0.34 -0.02

(m3d-p) Aggregate
Variable (m3d-p) (y-p) n exb
B'Eigenvectors 1.00 -0.84 5.51 0.07
a Coefficients -0.07 0.22 -0.10 -0.02

Note: Each eigenvector represents a stationary linear combination of the 1(1) variables such
that, for example, ((mO-p) - (y-p) + 6.15 n) = ecmO ~ 1(0).

Two features need to be noted. First, it was found that for the model in M0
only by excluding the parallel market exchange rate term ex ante could an
interpretable money demand cointegrating vector be determined. The
interpretation of this exclusion is discussed below. Moreover, since narrow
250 JOURNAL OF AFRICAN ECONOMIES, VOLUME 1, NUMBER 2

money is, by definition, outside money, it was found that the linear
combination which eliminated the rtb term between the two vectors provided
the most intuitive interpretation of the data.23 Second, a similar sort of
argument holds for the Divisia M3 aggregate. Since the own rate of interest
had already been captured within the construction of the aggregate, it was
found that the linear combination of the vectors which yielded the most
plausible relationships was that which netted out the own (internal) rate of
return.

Interpreting the long-run solutions

These cointegrating vectors therefore define the long-run demand for money
functions.24 It is instructive to discuss the results in detail, and to compare
them with those from the three other studies on Kenya.
First, the long-run demand has an income elasticity equal to unity in the
case of cash balances, but as the aggregate widens the income elasticity of
demand falls slightly to approximately 0.8 (excluding the M3 measure where
the elasticity unaccountably rises to over unity). This fall reflects the relative
shift from transactions to portfolio considerations in holdings of money
balances, and is correspondingly reflected in the stronger portfolio
considerations, as proxied by the inflation, interest rate and currency
substitution effects in the wider aggregates. These income elasticities are
significantly lower than those reported by Darrat (whose income elasticities
of demand of 1.8 for Ml and 1.9 for M2 seem implausibly high and
inhomogeneous). Mwega's results on the other hand are quite variable across
the aggregates (0.25 for Ml, 0.78 for M2, and 0.56 for M3) and, with the
exception of the M2 aggregate, are significantly lower than those emerging
from our cointegration analysis.
Much of the divergence between the results presented here and those from
Mwega's and Darrat's studies can also be explained by the large differences
in the other price effects in the model. Consider first the inflation elasticity,
which ranges around -5.0 to -6.0 across all the aggregates in the cointegration

23 Note that we could have equally taken a linear combination of the vectors to eliminate
to the inflation term, n. In either case the demand for money function in which either
interest rates and inflation appear is recoverable by substitution from the first cointegrating
vector.
24 Note that the constant and seasonal effects have not been reported in Table 3.
 Christopher Adam 251

analysis. This quarterly elasticity is equivalent to an annual inflation elasticity

of approximately -1.5 which lies well within any theoretical priors, and is also
very much in line with Hendry and Mizon's (1990) cointegrating vector
analysis for the UK.25 Mwega however finds both that the inflation elasticity
is much higher in absolute terms, and that it almost doubles between
aggregates (-8.0 for the Ml aggregate, and -14.7 for M2). A possible
explanation for these differences may lie with the specification of the
domestic interest rate, which, from the cointegration analysis is positively
related to the money aggregates as expected for the broader aggregates, while
in Mwega's study, the relationship is negative.
Finally, we note that the currency substitution effect, though relatively weak,
is consistently significant and correctly signed in the Ml, M2 and M3d
aggregates: a depreciation in the black-market exchange rate (represented by
a rise in exb) leads to a shift out of domestic money balances. The
determination of the price effects for the simple-sum M3 aggregate, however,
seem less clear.
Turning to the a vectors, which indicate the rate of feedback of deviations
from the long-run money demand relationship to the dynamic behaviour of the
endogenous variables of the VAR, we note that the main effect of the money
demand cointegrating vector is, as expected, through their feedback onto (m-
p), i.e. real money holdings, and to a lesser extent onto inflation. Moreover
the direction and magnitude of the feedback on all the equations is consistent
with a stable, error-correcting model of the demand for money. Since the data
are in logarithmic terms, the coefficients in the a vectors can be interpreted
as the percentage feedback per period, which therefore indicate a feedback of
approximately 20 per cent per period for (mO-p) and (m2-p), a higher
feedback for (ml-p), and just under 10 per cent per quarter for the broadest
aggregates (m3-p) and (m3d-p).
Of particular note is the comparison of the simple-sum and Divisia
aggregates for M3. Though the eigenvalues indicate that the Divisia aggregate
is marginally less significant, and while both measures reveal intuitive
inflation elasticities, the M3 aggregate reports an improbably high own
interest rate elasticity of demand, an unexpectedly high income elasticity, and,
more notably, a perverse (i.e. positive) dynamic feedback effect on income.
Note also that lower income elasticity for the Divisia-based M3 is consistent

25 They report an income elasticity of 0.97 and an inflation elasticity of -6.9.

252 JOURNAL OF AFRICAN ECONOMIES, VOLUME 1, NUMBER 2

with Barnett's (1980) assertion that the simple-sum aggregate will tend to
overstate the income elasticity of demand.
In general, then, the long-run cointegrating vectors, with the exception of
the simple-sum M3 aggregate, are consistent with the data and provide the
basis for a coherent economic interpretation of the underlying relationships
determining the demand for money. They are consistent with the limited
comparable evidence from other applied studies using this methodology, and
at the same time are somewhat more plausible than the results emerging from
other applied work on Kenya. Consequently this enables us to accept the
Engle-Granger Representation Theorem (1987) and proceed directly to an
error-correction specification in which the cointegrating vectors, defined in
Table 3, represent the long-run demand for money relationships.

5. Dynamic demand for money functions for the Kenyan economy

Adopting a general-to-specific procedure we estimate a series of dynamic

error correction models for each money aggregate based around the
cointegrating vectors reported in the preceding table. Each model is of the
general form

(10) A(LWm-p), = Oo + B(L)A(y-p), + C(L)An, + D(L)Artb, + E(L)Aexb, +

F(L)Ecmhl + Ly£, + e,

where A(L)...F(L) are polynomials of the form A(L) = YxxJJ in which L is the
lag operator such that Wx, = xhr Ecm is the error correction term, and 5, are
seasonal dummies. Throughout the reduction sequence various
parameterizations of the variables were considered. In particular alternative
specifications of the currency substitution effect were examined (using the
actual and anticipated26 depreciation of the parallel market rate, the actual and
anticipated official rate of depreciation, and also the premium on the official
rate). The results of these alternative specifications are not reported here, but
it was found that for all aggregates a model specification in terms of the
actual parallel market exchange rate encompassed (i.e. variance dominated)
alternative specifications.

26 As proxied by the lead of the ex post actual depreciation of the exchange rate.
 Christopher Adam 253

It may also be noted that (10) is initially estimated with the two error
correction terms for each monetary aggregate, the relevant money demand
relationship and the interest rate-inflation vector. In each case, however, it was
found that the latter did not enter significantly into the model.27
The process of reduction has thus resulted in a series of parsimonious
models which uniformly have increased efficiency over their nesting model
(10), and allow for a clearer specification of the dynamic processes. The steps
required to achieve these final models are not reported, although in general
it was found that the short run dynamic effects were felt within a relatively
short period (in no case were there significant lagged effects beyond four
quarters). However, as shall be seen, the dynamic structures which most
efficiently captured the data were often complex, but nonetheless interpretable.
The final full sample equations spanning the period from 1974 to 1989 are
reported in the Appendix, but to facilitate an economic analysis of the results
a stylised summary of the results is presented in Table 4, where for
convenience we do not report the seasonality of the models and we suppress
the details of their individual short-run dynamics. Thus, for example, while
in the Appendix we note that for the MO and Ml models the short run
dynamics of the dependent variable are best captured by a four-period (0 to
3) moving average process, while those for the M2 model are captured by a
two period (0 to 1) process, in Table 4 we report both these under the single
line lA(m-p).

Analysis of results

The statistical characteristics of the dynamic equations are good and all
broadly accept the null that the error term is normally and homoscedastically
distributed and exhibits no significant first or higher-order autocorrelation.
In the case of Ml, M2, and M3, however, the null of error normality is
rejected by the Jarque-Bera test. This is due entirely to the presence of a
significant outlier in 1977 Ql: for all aggregates the model significantly

27 This can be demonstrated using a simple F-test on the marginal significance of the
cointegrating vector in each money-demand function. In each case the null that the
marginal significance of the vector was zero could not be rejected at any level of
significance. This implication of this is that the model may, ceteris paribus, be estimated
independently of the other dynamic relationships in the VAR.
254 JOURNAL OF AFRICAN ECONOMIES, VOLUME 1, NUMBER 2

Table 4:
Summary demand for money models: OLS estimates 11974(4)-1989(2)]

Monetary aggregate
Explanatory
variables A(mO-p) A(ml-p) A(m2-p) A(m3-p) A(m3d-|

ZA(m-p) 0.54 0.57 0.40 -0.23 -0.30

(2.98) (2.11) (1.79) (2.61) (2.83)
IA(y-p) - -0.17 -0.11 - 0.42
(2.24) (1.95) (2.29)
A7t -1.20 -0.52 -0.93 -0.93 -0.74
(6.61) (2.23) (3.42) (3.87) (4.09)
lArtb -3.01 -6.77 -7.81 -5.46 -11.06
(2.56) (1.78) (2.68) (2.17) (4.13)
Aexb 0.19 -0.17 0.11 -0.09 -0.06
(2.58) (2.78) (2.02) (1.26) (0.99)
Ecm -0.19 -0.32 -0.20 -0.07 -0.09
(3.21) (4.15) (4.70) (2.63) (2.16)

R- 0.853 0.594 0.599 0.560 0.643

s.d 0.078 0.061 0.047 0.041 0.055
d 0.032 0.042 0.032 0.029 0.032
DW 1.946 2.016 2.083 2.024 1.981
C, (6) 0.68 0.86 1.07 1.16 0.83
0.79 0.44 0.25 0.27 1.08
0.77 1.17 1.24 0.94 0.96
1.54 7.18* 4.84* 10.85* 1.54
C 0.63 0.43 0.53 0.69 1.41

• indicates rejection of null.

Notes: (1) t statistics are reported in parenthesis; s.d is the standard error of the dependent
variable; o the equation standard error; DW is the first-order error autocorrelation Durbin-
Watson statistic.
(2) £, (6) is the sixth order Lagrange Multiplier test for autocorrelation; C,2 (6) the
sixth order Lagrange Multiplier test for autoregressive conditional heteroscedasticity; £3 is
White's (1980) test for unconditional heteroscedasticity; £, the Jarque-Bera (1983) test for the
normality of the error distribution; and £, a measure of 16-period (1985 Q3 to 1989 Q2)
forecast stability (see Hendry (1989)). Tests £,, £,, and £_, are reported as sample-adjusted
F-Tests (see Kiviet (1985)): £, and ($ are distributed as x2 with two degrees of freedom.
 Christopher Adam 255

under-predicted the rise in money holdings in this quarter. This was due to
the effect of the rapid increase in unsterilized income from coffee exports.
The compulsory surrender of coffee earnings (which accrued directly to the
non-bank private sector and which were not sterilized by the government)
combined with the lags in the implementation of investment projects, forced
the private sector off its demand function. This disequilibrium was also
reflected in the rapid rise in the liabilities of the banking sector, which
experienced severe excess liquidity at the period. It was this rise in excess
liquidity (liquid assets were equivalent to almost 40 per cent of eligible
liabilities) which prompted the abolition in the minimum lending rate in 1977.
In the case of MO, however, the strong seasonality in that aggregate seems to
have captured the first quarter surge, while the built-in elasticity of
substitution effects in the Divisia-based aggregate meant that the cash-surge
in that quarter carried less weight in the Divisia aggregate than in the simple
sum aggregates M2 and M3. Fitting a dummy variable for this quarter ex post
to the Ml, M2, and M3 equations improves their overall fit quite substantially
while leaving the other equation coefficients unchanged. However in order to
facilitate direct comparison between the aggregates we have chosen not to
include any ex post dummy variables of this kind.
A clear pattern emerges in terms of the overall goodness-of-fit of the
aggregates. Due to the strong seasonal component in the MO model, R2 is high
at 0.85, but this falls off sharply to around 0.60 when the aggregates are
widened to include inside-money deposits with the financial sector. Noting
that the dependent variable in each case is in terms of a quarterly first
difference, this level of fit is relatively high, and compares well with
comparable money demand studies elsewhere (e.g. Hendry (1985, 1988),
Adam (1991)). It may also be noted that the Divisia-based aggregate has a
slightly higher R2 statistic than the simple-sum M3.28
Over the sample as a whole three distinct phases in the overall fit of the
aggregates can be identified (the plots are not reported though). The first is
the period of the coffee boom in 1976 and 1977 where, because the boom was

28 Since the two models have a different number of parameters and the dependent
variables have differing standard deviations, the a and R2 statistics are not strictly
comparable. However it can be noted that while the M3 equation standard error is
approximately 70 per cent of the standard deviation of the dependent variable, that for the
Divisia aggregate is less than 60 per cent of the standard deviation, indicating a slightly
greater degree of efficiency in the latter model.
256 JOURNAL OF AFRICAN ECONOMIES, VOLUME 1, NUMBER 2

known to be transitory, movements in cash balances and in the corresponding

aggregates were large and transitory, as money fulfilled a buffer stock
function in the presence of lagged portfolio adjustment effects. However more
interesting is the second phase from late 1979 to mid 1982, where the
predictive power of the models seems to be at its weakest, and where it may
be possible to argue that the stabilization and trade liberalization efforts
during that period served to alter agents' short-term money demand function.
The explanation for this has some interesting bearing on the implicit super-
exogeneity conditions in the model (see Hendry and Engle (1986) and Hendry
and Ericcson (1991)). In particular, this period was characterised by a level
and volatility in domestic inflation which were much higher than previously
experienced in Kenya, while the increased political uncertainty and loss of
government credibility following the attempted coup d'etat in 1982 served to
increase the parallel market premium. Finally, in the third phase, as
stabilization in Kenya was gradually re-established from mid-1983 onwards,
the fit of the money demand functions improved markedly although (as
occurred in 1977) a similar failure to capture the cash surge accompanying the
1986 coffee boom can again be noted.
Within-sample stability of the model and its coefficients can be gauged
using recursive estimates of the individual parameters of the equations, and
from within-sample tests of forecast accuracy. Though we do not plot the
recursive least squares coefficient plots, they confirm that the equation
residuals for all aggregates are relatively stationary, independent and
homoscedastic over time, and that for all aggregates the error correction and
short-run inflation effects are well defined and relatively stable throughout the
sample. However, the plots suggest that the short-run currency substitution
effect (the Aexb term) is only reasonably well defined for the MO, Ml and the
M3d aggregates. The full-sample values for the M2 and M3 aggregates need
therefore to be interpreted with extreme caution. Despite this, the recursive
least square plots do quite clearly confirm the sign switch in this effect
between the MO and wider aggregates as was noted in Table 4. We discuss
this effect in the next section.
The direct tests of forecast stability of the model (see Appendix and the C,5
statistics in Table 4) confirm that forecast stability cannot be rejected for the
16 quarter forecast although there is a systematic forecast error during the
period of the 1986 coffee boom. Finally, we note that for the full-sample
model all five aggregates perform satisfactorily against the standard battery
of diagnostic tests, and in no case can any violation of the conditions of error
 Christopher Adam 257

homoscedasticity, or zero autocorrelation be identified. Combined with the

within-sample forecast stability and the evidence from the recursive analysis
of the model coefficients we can therefore conclude that from amongst the
equation we can obtain a tentatively adequate characterization of the data. On
the basis of this we now briefly assess the encompassing implications of the
models before considering the economic interpretation of the dynamic demand
for money models.

Encompassing existing money demand studies in Kenya

One of the minimum conditions for acceptance of new empirical work is that
it encompasses existing studies in the field. Though the literature is not
widespread it is possible to examine the encompassing issue with reference
to Mwega (1990). His estimated demand functions are reported in equations
(11) and (12), where the same nomenclature as above is used, 't' statistics are
reported in parentheses

(11) (ml-p),= 0.94 + 0.04(y-p), -1.28 K, -0.04 rtb, + 0.84 (ml-p),.,

(1.5) (0.7) (3.0) (3.1) (12.7)
R2 = 0.82 DW = 2.03

(12) (m2-p)= 0.20 + 0.07(y-p),-1.32 n,-0.02 rtb, + 0.91 (m2-p),_,

(0.5) (1.7) (5.1) (2.7) (19.7)
R2 = 0.92 DW = 2.00

In both equations, the current values for income,inflation and the interest
rate are derived from Almon lag representations of the variables. To evaluate
these rival models we first consider their performance relative to standards
goodness of fit indicators (estimated over the same data and sample period)
and then relative to each other using non-nested encompassing tests. In the
latter case, the Mwega model is transformed into a first difference growth
model. The results are presented in Table 5, which tells an unambiguous
story.
While the standard performance criteria (equation standard error and
Schwarz) indicate that our model marginally dominates Mwega's, the forecast
performance indicators suggest a much higher level of within-sample forecast
stability. These conclusions are borne out by the non-nested encompassing
258 JOURNAL OF AFRICAN ECONOMIES, VOLUME 1, NUMBER 2

tests where not only do our models (i.e. the error-correction dynamic model)
significantly encompass the MWega model but also easily encompass the joint
nesting model.29

Table 5:
Competing model performance

Performance criteria
Equation Schwarz 16 Forecast

Ml Aggregate
std error criterion x2
Model I (per Table 4) 0.043 -5.81 0.43
Model 2 (Mwega eq (11)) 0.059 -5.15 2.83*
M2 Aggregate
Model 1 (per Table 4) 0.032 -6.32 0.53
Model 2 (Mwega eq (12)) 0.046 -6.05 1.43

Encompassing tests
Model vs Model 2 Model 2 vs Model 1
Cox Joint Cox Joint

Ml Aggregate 0.36 0.37 -28.37* 8.32*

M2 Aggregate 0.48 0.25 -24.76* 7.52*

Notes:

1 • indicates significant at 5 per cent

2 The Schwarz criteria is defined as SC = Ina + klnT/T where a is the degrees-of-freedom
adjusted equation standard error, k the number of parameters, and T the sample size. It is
increasing in both the equation standard error and the number of regressors: a fall in the
Schwarz Criterion is an indication of model parsimony. See Hendry (1989).
3 Under the null that M, encompasses M2 the Cox test has a standard normal distribution. A
significant value for the statistic indicates a failure to encompass. If both models fail to
encompass each other, then the Cox statistic is significant for both M, vsM, and M2 vs M,
(note that for two models to fail to encompass each other is valid and is evidence of two
weak models). The F test is a test of each model against the joint nesting model.

29 The two models are strictly non-nested since we have retained Mwega's original
Almon lag transformation for the independent variables.
 Christopher Adam 259

The ability of the error-correction models to encompass the form specified

by Mwega is not surprising given the evidence on the long-run cointegration
between the variables of the model, such that the addition of long-run
information to the dynamic model will clearly improve the explanatory power
of the model, but this exercise does serve to underline both the significance
of the cointegrating relationships in the dynamic demand relationships and
also the importance of currency substitution effects in of the demand for
money.

Economic interpretation of the models

In discussing the economic interpretation of these demand for money

functions it needs to be borne in mind that in each case we are considering
the dynamic behaviour of the equations around the long-run demand function
characterised by the cointegrating vectors discussed in the previous section.
In this respect the most important coefficient of the dynamic model is the
feedback on the error correction term. Comparing Tables 3 and 4 we note that
the feedbacks reported in the dynamic single equation models (estimated
unrestrictedly, and reported as Ecm) are virtually identical to the
corresponding money demand feedbacks reported in the a loadings matrix
under the Johansen procedure. This closeness supports the validity of the
structure of the cointegrating vectors used in the models. More generally,
these results suggest a relatively intuitive picture in which the narrow
aggregates MO, Ml and M2 enjoy a much faster adjustment than broader
aggregates where possibly higher transactions costs may preclude rapid
adjustment. Thus in the case of the narrower money balances approximately
20 per cent of the previous disequilibrium from the long run demand is
corrected per period, while for the broader M3 aggregates the figure is closer
to 10 per cent per period. This speed of adjustment is again comparable to
similar work for industrialised economies (for example Hendry and Mizon
(1990) report a 10 per cent feedback effect for UK Ml), but is almost twice
as rapid as the adjustment speeds implied by Mwega's (1990) partial
adjustment models which, in view of the autoregressive nature of the money
aggregates, imply relatively slow adjustment to equilibrium.
In terms of the other short-run dynamic effects in the model a number of
features warrant attention. The first is the relatively consistent short-run
inflation effects for all aggregates. In each case the growth rate of inflation
260 JOURNAL OF AFRICAN ECONOMIES, VOLUME 1, NUMBER 2

(i.e. the acceleration in the price level) has an immediate and significant effect
on real money holdings (aside from other considerations of price
homogeneity).30 In this economy then, while the price level may be neutral,
its growth rate is not (as shown by the cointegrating vector), and its
acceleration is similarly non-neutral in the short-run. This result is integral to
the analysis of inflation tax, not least to the extent to which as a result of this
relatively high, and rapid, substitutability between money and real assets
and/or consumption, the scope for raising inflation tax from the private sector
is likely to be limited. The lower short-run interest elasticity for the broader
aggregates probably reflects the lower liquidity of broader aggregates and also
the greater complementarity of broader money aggregates with respect to non-
monetary assets.
A second feature of note concerns the short-run interest elasticity in the
model. We noted during the analysis of the long-run cointegration
relationships between the variables of the model that there existed a strong
and stable positive relationship between the rate of inflation and the market
interest rate as proxied by the Treasury Bill discount rate. Table 4 indicates
not only that in the short run there is a sign-switch indicating substitution
effects, but that as the money aggregates widen the short-run elasticity of
demand with respect to the growth of the interest rate increases, indicating
broader portfolio considerations in determining money holdings. While this
result is standard, one that may require some interpretation is the large
difference in the interest elasticity between the two M3 aggregates. This can
be explained as follows: changes in the interest rate will induce two sets of
effects, the first being out of money into other financial assets, but the second
is the switch between non-interest bearing and interest bearing components of
money. As we have discussed above, the principle behind the Divisia index
is that the simple-sum aggregate overstates the substitutability between
components of money, and consequently for any given rise in the interest rate
the apparent elasticity of the simple-sum aggregate to external substitution
effects will be dissipated by the internal substitution effects brought about by
switches from non-interest bearing to interest bearing components of money.
In contrast, however, the Divisia index by definition is neutral to changes in
the composition of the money aggregate (in other words it has already

30 Allowing for the possible endogeneity of the inflation effect and re-estimating the
model using a FIML estimator does, however, reduce this short-run elasticity slightly.
 Christopher Adam 261

captured the internal substitution effects) and as a result the interest elasticity
represents only the external interest elasticity of demand (see also Barnett
(1980) p. 41).
Finally we consider the interpretation of the currency substitution effects. In
the cointegrating vectors for the Ml, M2, M3 and M3d aggregates we
identified a significant, but small, negative elasticity with respect to non-
interest bearing (and illegal) foreign currency holdings as proxied by the
parallel market exchange rate. There was however no identifiable long-run
levels relationship between narrow money holdings and the parallel foreign
exchange rate. This could be rationalised quite simply in terms of a economy
where narrow money balances are held in the long run exclusively for
domestic currency denominated transactions while currency substitution
entered as a portfolio/precautionary consideration in the determination of the
level of broader money holdings. In the short-run, however, the picture is
altered as we observe an economy in which there remains a currency
substitution motive but in which the narrow money holdings act as a buffer
stock through which foreign currency transactions are effected. Thus, while
(with the exception of the M2 aggregate) depreciation of the parallel market
exchange rate reduces short run holdings of broader aggregates, it in fact
increases holdings of cash in the short run. Since currency substitution is
illegal in Kenya, transactions are of necessity conducted in cash and thus
agents switch from deposits into cash in anticipation of currency substitution
opportunities. Having said this, however, it may be noted that the currency
substitution effect in the demand for money seems to be only weakly
identified in these models, and the dynamics not particularly well understood
(see our discussion of the recursive stability of the models above). In part this
reflects not only that the parallel market is illegal, but, because the
government has avoided the overvaluation of the official exchange rate, it is
also a relatively small and specialised market, certainly in comparison to
many other economies.

6. Conclusions and policy implications

In this paper we have successfully established a series of single-equation

demand for money functions for the Kenyan economy from 1973 to 1989. In
general, though estimated using an interpolated series for income, the
equations embody long-run solutions which are fully consistent with
262 JOURNAL OF AFRICAN ECONOMIES, VOLUME 1, NUMBER 2

theoretical priors from the literature on the demand for money. Moreover the
paper has demonstrated that it is also possible to develop dynamic
specifications for the demand for money which efficiently track actual
movements in holdings of real money balances around the long-run demand
functions throughout the period. In doing so we have successfully
encompassed existing studies of Kenyan demand for money and have, in
addition, established a number of important results.
First, the successful identification of a stable and well defined demand
function for narrow money balances, MO, has allowed for the detailed analysis
of the fiscal implications of inflation, both by identifying both the steady-state
inflation tax capacity of the private sector, and also by quantifying the short-
run dynamics of money holdings in response to changes in the rate of
inflation. Following Friedman (1971), standard analyses of inflation tax derive
the revenue maximizing rate of inflation as

(12) n = 1/p - (agy)

where P is (minus) the inflation elasticity of the demand for money, a is the
income elasticity of the demand for money, and g, is the average rate of
growth of income. From the long-run demand function for narrow money, we
derive an average revenue-maximizing rate of inflation of 13.5 per cent
(l/P=16.5%-3%, where P=-6.15, a= 1.01 and g=3% per annum). Above this
rate of inflation, the private sector systematically reduces its real holdings of
base money more rapidly than the rate of inflation, so that the real inflation
tax revenue declines. It is interesting to note that this average rate is not only
somewhat lower than may have been expected, but will also be a decreasing
function of the degree of domestic and foreign asset market liberalization. The
greater the range of substitutes for domestic base money, and the more rapidly
the private sector can adjust their real holdings, the lower the revenue
maximizing level of inflation will be. Moreover, the coefficient on the error-
correction term is a direct measure of this speed of adjustment: according to
these results, the Kenya private sector adjusts at a rate of approximately 20
per cent per quarter to any disequilibrium (i.e. increase in real base money).
Again, the further liberalization proceeds, the more rapidly the private sector
can adjust to disequilibria and the less will be the short-run inflation tax
revenue.
Second, we have identified a limited currency-substitution effect which
operates, through a well-defined cash-based buffer-stock model. Since the
 Christopher Adapt 263

parallel market premium has remained relatively small this has resulted in
currency substitution not being a major determinant of the demand for money.
However, in other countries where the rnarket is widespread and the premium
larger, such considerations may be expected to play an important role.
Finally, we have demonstrated that, while not dramatic, the use of a Divisia-
based aggregation for money stock aggregates which include the deposit
liabilities of the NBFIs results in a more data coherent specification of broad
money demand than is possible using simple-sum aggregates.

APPENDIX
FINAL OLS MONEY DEMAND EQUATION AND FORECAST ANALYSIS

Modelling A(mO-p) by QLS Sample is I974( 3) to 1989( 2)

Variable Coefficient Std Error H.C.S.E. t-Value Partial r

Constant .0818376 .00946 .01186 8.65368 .5949

Ql -.1309355 .01284 .01434 -10.19459 .6708
Q2 -.1391293 .012% .01421 -10.73314 .6931
Q3 -.0619953 .01225 .01411 -5.05899 .3341
A2rtb -3.0139988 1.17588 1.36665 -2.56319 .1141
Z03A(mO-p) 1 .5425604 .18183 .14587 2.98381 .I486
An -1.2007735 .18152 .22492 -6.61498 .4618
EcmO 1 -.1932344 .06017 .05954 -3.21123 .1682
Acxb2 .186,9660 .07240 .07145 2.58232 .1156

R2 = .8531734 a = .0320912 F( 8, 51) = 37.04 | .0000] DW = 1.946

LM Autocorrelation LM Autoregressive Conditional i

F( 6, 45) = .68 | .6693) F( 6, 39) == .79|.58I3|

F( 5, 46) = .64 | .6697] F( 5,41) == .88 | .50581
F( 4,47)= .68 | .6116] F( 4, 43) == .86 | .4932]
F( 3, 48) = .69 | .56061 F( 3, 45) == .25 | .8598]
F( 2, 49) = .25 | .7830] F( 2, 47) == .21 | .8113]
F( 1, 50)= .01 [ .9268] F( 1,49) == .00 | .9980]

Jarque - Bera test for error normality x1(2)= 1.538

White's heteroscedastic error test F(I8, 28) == .7710 [ .6838]
Tests of parameter constancy over 1985(3)- 1989( 2)
Forecast %3( 16)/16 = .63
Chow test F(I6, 34) = .50
Zero forecast innovation t(15) = -.36
264 JOURNAL OF AFRICAN ECONOMIES, VOLUME 1, NUMBER 2

Modelling N,ml-p) by OLS Sample is I974( 3) 10 I989( 2)

Variable Coefficient Std Error H.C.S.E. t-Value Partial r

Constant .0608067 .01229 .01256 4.94741 .3287

Qi -.0740473 .01653 .01990 -4.48008 .2864
Q2 -.1110592 .01847 .01625 -6.01436 .4198
Q3 -.0494678 .01812 .01873 -2.73070 .1298
Ecml 1 -.3252310 .07830 .08922 -4.15363 .2565
H)2Artb 6.7767617 3.79990 3.46558 -1.78340 .0598
Arc -.5190740 .23266 .27635 -2.23100 .0905
A3Aexb2 -.1869746 .06753 .06170 -2.76860 .1329
£03A<ml-p)l .5757866 .27242 .25592 2.11361 .0820
A3A(y-p)l -.1736261 .07735 .06680 -2.24481 .0916

R : = .5938998 a = .0425577 F( 9, 50) = 8. 12 1.00001 DW = 2.016

LM Autocorrelation LM Autoregressive Conditional He

F( 6,44) = .86 | .5319] F( 6, 38) = .44 [ .84501

F( 5, 45) = .31 | .9028) F( 5,40) = .42 [ .83391
F( 4, 46) = .33 | .85971 F( 4, 42) = .42 | .79601
F( 3, 47) = .29 [ .83571 F( 3,44) = .45 | .72081
F( 2, 48) = .30 | .7457] F( 2,46) = .41 | .6671]
F( 1,49) = .02 | .9004| ' F( 1, 48) = .08 [ .7727]

Jarque - Bera test for error normality x2(2) = 7 .189

White's heteroscedastic error test F(18, 28) == 1.1706 | .33881

Tests of parameter constancy over I985( 3 ) - 989( 2)

Forecast jf( 1 6 )/l6 = 43

Chow test F(I6, 34) = 41

Zero forecast innovation t(l5) _ 05

 Christopher Adam 265

Modelling Mm2-p) by OLS Sample is 1975( 2) to I989i 2)

Variable Coefficient Std Error H.C.S.E. t-Value Partial r

Alt -.9253064 .27015 .22358 -3.42513 .1997

I02Artb -7.8143049 2.91120 2.68445 -2.68422 .1329
Ecm2 1 -.2015763 .04281 0.03849 -4.70899 .3206
A3A(y-p)l -.1157329 .05916 .05248 -1.95615 .0753
Constant .0364967 .00937 .00744 3.89478 .2440
Ql -.0324361 .01305 .01589 -2.48510 .1161
Q2 -.0721695 .01471 .01234 -4.90619 .3387
Q3 -.0237910 .01448 .01249 -1.64314 .0543
I03A(m-p)l .4002921 .22272 .18555 1.79732 .0643
AexbO5 .1123071 .05537 .05177 2.02831 .0805

R; = .5998514 o = .0326962 F ( 9, 47) = 7.83 | .0000) DW = 2.083

LM Autocorrelation LM Autoregressive Conditional He

F( 6, 41)= 1.07 | .39761 F( 6, 35) = .25 | .9551]

F( 5, 42) = .09 | .9926) F( 5, 37) = .30 1 .9122]
F( 4, 43)= .12 | .9759) F( 4, 39) = .18 | .9481]
F( 3,44)= .16 | .9232) F(3, 41) = .13 | .9413]
F( 2, 45) = .05 ( .9523] F( 2, 43) = .13 1 .8803]
F( 1,46)= .IO| .7551) F( 1,45) = .05 | .8198]

Jarque - Bera test for error normality X: (2) = 4.845

White's heteroscedastic error test F(12, 36) = 1.2452 | .2928]

Tests of parameter constancy over: 1985( 3) - 1989( 2)

Forecast X2( 16V16 = .53

Chow test F(16, 34) = .49

Zero forecast innovation (15) _ .23

266 JOURNAL OF AFRICAN ECONOMIES, VOLUME 1, NUMBER 2

Modelling t^mi-p) by OLS Sample is 1975< 2) to I989< 2)

Variable ' Coefficient Std Error H.C.S.E. t-Value Partial r

Alt -.9331610 .24081 .22912 -3.87516 .2346

Ecm3 3 -.0726517 .02761 .04075 -2.63125 .1238
Constant -.0289555 .00609 .00503 4.75577 .3158
Ql -.0321299 .00998 .01299 -3.22009 .1747
Q2 -.0304054 .01053 .01059 -2.88746 .1454
Aexb2 -.0892926 .07040 .07420 -1.26834 .0318
AA(m3-p)3 -.2354050 .09002 .10731 -2.61497 .1225
I02Artb -5.4656929 2.51074 3.34212 -2.17692 .0882

R- = . 5621595 a = .0289741 F( 7, 49) = 8.99 [ .0000) DW = 2.024

LM Autocorrelation LM Autorcgressive Conditional Heteroscedasticity

F( 6,43) = 1.16 [ .3448] F( 6, 37) = .27 .94571

F( 5. 44) = .40 | .8458] F( 5, 39) = .32 .8971]
F( 4, 45) = .15 [ .9610] F( 4,41) = .31 .8690)
F( 3, 46) = .03 | .99321 F( 3, 43) = .32 .8127]
F( 2, 47) = .03 1 .96681 F( 2, 45) = .47 .6270]
F( 1,48)= .01 [ .9203] F( 1,47) = .20 .6597]

Jarque - Bera test for error normality r (2) = 0.085

White's heteroscedastic error test F(12, 36) = .9368 | .52281

Tests of parameter constancy over: I985( 3) - 1989( 2 )

Forecast X3(I6)/16 = .69

Chow test F(16, 34) = .60

Zero forecast innovation t(15) _ -.36

 Christopher Adam 267

Modelling &(m3d-p) by OLS Sample is I975( 2) to I989( 2)

Variable Coefficient Std Error H.C.S.E. t-Value Partial r

A(m3d-p)3 -.3013017 .10621 .09989 -2.83683 .1462

ZO2A(y-p)3 .4231503 .18445 .17907 2.29406 .1007
An' ' -.7432367 .18129 .16539 -4.09981 .2634
EOpArtb -11.0640513 2.67007 2.53958 -4.14373 .2676
Gojistant .0322678 .00997 .00798 3.23810 .1824
Ql" -.0188569 .01435 .01374 -1.31431 .0355
Q2 -.0616722 .01547 .01450 -3.98665 .2527
Q3 -.0115919 .01348 .01382 -.85988 .0155
Ecm3d4 -.0933928 .04323 .02551 -2.16053 .0903
A2Aexb.' -.0609182 .06139 .06516 -.99237 .0205

= .6430255 a = .0329316 F( 9, 47) = 9.41 | .0000) DW = 1.981

1 Autocorrelation Lfyl Autorcgressive Conditional He

F( 6.41) = .83 I .55561 F( 6, 35) = .08 | .2470]

5,42) = .38 ( .86041 F( 5, 37) = .40 1 .3257]
F( 4,43) = .43 [ .78771 F( 4, 39) = .20 | .3257]
F( 3,44) = .42 | .73811 F( 3. 41) = .14 | .3548)
F( 2,45) = .01 | .9879) F( 2, 43) = .79 | .17941
F( 1,46) = .02 | .8816] F( 1,45) = .!.15 | .0827)

Jarque - Bera Test for Error Normality X: (2) = 1.544

White's Heteroscedastic Error Test F(12, 36) = .9620 1 -51351

Tests of Parameter Constancy over 1985(3)- I989( 2)

Forecast x2(16)/16 1.41

Chow Test F(16, 34) 1.28

Zero Forecast Innovation t(!5) 1.07

268 JOURNAL OF AFRICAN ECONOMIES, VOLUME 1, NUMBER 2

References

Adam, C.S. (1991) "Financial Innovation and the Demand for £M3 in the
UK 1975-86", Oxford Bulletin of Economics and Statistics, 53, 401-424.
- (1992) The Demand for Money, Asset Substitution and the Inflation Tax in
a Liberalizing Economy: an Econometric Analysis for Kenya, D.Phil Thesis,
University of Oxford.
Adekunle, J.O. (1968) "The Demand for Money: Evidence from
Developed and Less Developed Economies", IMF Staff Papers, 15.
Banerjee, A., J.J. Dolado, G.W. Galbraith and D.F. Hendry (1992)
Equilibrium, Error Correction and Cointegration in Econometrics, Oxford:
Oxford University Press.
Barnett, W.M. (1980) "Economic Monetary Aggregates: An Application
of Index Number and Aggregation Theory", Journal of Econometrics, 14,
11-48.
- (1982) "The Optimal Level of Monetary Aggregation", Journal of Money
Credit and Banking, 14, 688-710.
Barr, D.G. and K. Cuthbertson (1991) "Neoclassical Consumer Demand
Theory and the Demand for Money", Economic Journal, 101, 855-876.
Blejer, M.I. (1978) "Black Market Exchange Rate Expectations and the
Domestic Demand for Money: Some Empirical Results", Journal of
Monetary Economics, 4, 161-1 A.
Calvo, G.A. and C. Rodriguez (1977) "A Model of Exchange Rate
Determination under Currency Substitution and Rational Expectations",
Journal of Political Economy, 85.
Central Bank of Kenya (various years) Quarterly Economic Review,
Nairobi.
Centra] Bureau of Statistics Business Expectations Enquiry, Nairobi
(unpublished).
- (various years) Statistical Abstract, Nairobi.
Clements, K.W. and P. Nguyen (1980) "Economic Monetary Aggregates -
A Comment", Journal of Econometrics, 14, 49-53.
Coates, W.L. and D.R. Khathkhate (1980) "Money and Monetary Policy in
Less Developed Countries", Oxford: Pergammon Press.
Cockerline, J. and J. Murray (1981) "A Comparison of Alternative
Methods of Monetary Aggregation: Some Preliminary Evidence", Bank of
Canada Technical Report 28.
 Christopher Adam 269

Darrat, A. (1985) "The Demand for Money in a Developing Economy:

The Case of Kenya", World Development, 13, 1163-70.
Davidson, J., D.F. Hendry, F. Srba, and S. Yeo (1978) "Econometric
Modelling of the Aggregate Time-Series Relationship between Consumers'
Expenditure and Income in the United Kingdom", Economic Journal, 88,
661-92.
Diewert, W.E. (1976) "Exact and Superlative Index Numbers", Journal of
Econometrics, 4, 115-45.
Domowitz, I. and I. Elbadawi (1987) "An Error-Correction Approach to
the Money Demand: The Case of Sudan", Journal of Economic
Development, 26, 257-275.
Domowitz, I. and C. Hakkio (1987) "Error Correction, Forward Looking
Behaviour, and Dynamic International Money Demand", Journal of Applied
Econometrics.
Engle, R.F. and C.W.J. Granger (1987) "Co-integration and Error
Correction: Representation, Estimation and Testing", Econometrica, 55, No.
2.
Friedman, M. (1971) "Government Revenue From Inflation", Journal of
Political Economy, 79, 846-56.
Fuller, W.A. (1976) Introduction to Statistical Time Series, New York: John
Wiley.
Granger, C.W.J. (1986) "Developments in the Study of Cointegrated
Variables", Oxford Bulletin of Economics and Statistics, 48, 213-28.
Hendry,D.F. (1980) "Econometrics - Alchemy or Science", Economica, 41,
387-406.
- (1985) "Monetary Economic Myth and Econometric Reality", Oxford
Review of Economic Policy, 1, No. 1, 72-83.
- (1986) "Econometric Modelling with Cointegrated Variables: An Overview",
Oxford Bulletin of Economics and Statistics, 48, 201-12.
- (1988) "The Encompassing Implications of Feedback and Feedforward
Mechanisms in Econometrics", Oxford Economic Papers, 40, 132-49.
- (1989) PC-GIVE: An Interactive Econometric Modelling System, Institute
of Economics and Statistics, Oxford.
Hendry, D.F. and R. Engle (1986) "Testing Superexogeneity and
Invariance" (mimeo).
Hendry, D.F and N. Ericcson (1991) "Modelling the Demand for Narrow
Money in the UK and US", European Economic Review, 35, 833-81.
270 JOURNAL OF AFRICAN ECONOMIES, VOLUME 1, NUMBER 2

Hendry, D.F. and G.E. Mizon (1990) "Evaluating Dynamic Econometric

Models by Encompassing the VAR" (unpublished).
Hendry, D.F. and J-F. Richard (1982) "On the Formulation of Empirical
Models in Dynamic Econometrics", Journal of Econometrics, 20, 3-33.
Hylleberg, S., R.F. Engle, C.W.J. Granger, and B.S. Yoo (1990) "Seasonal
Integration and Cointegration", Journal of Econometrics, 44, 215-38.
International Monetary Fund (various years) International Financial
Statistics, Washington, D.C.
Jarque, CM. and A.K. Bera (1980) "Efficient Tests for Normality,
Homoscedasticity and Serial Independence of Regression Residuals",
Economic Letters, 6, 255-9.
Johansen, S. (1988) "Statistical Analysis of Cointegrating Vectors", Journal
of Economic Dynamics and Control, 12, 231-54.
Johansen, S. and K. Juselius (1990) "Maximum Likelihood Estimation and
Inference on Cointegration - with Applications to the Demand for Money",
Oxford Bulletin of Economics and Statistics, 52, 169-210.
Killick, T. and F.M. Mwega (1990) "Monetary Policy in Kenya, 1967-88",
ODI Working Paper, 39.
Kiviet, J.F (1985) "Model Selection Test Procedures in a Single Linear
Equation of a Dynamic Simultaneous System and their Defects in Small
Samples", Journal of Econometrics, 28, 327-62.
Laidler, D.E.W. (1985) The Demand for Money, London: Harper Row.
Mill, T.C. (1983) "Composite Monetary Indicators for the United
Kingdom: Construction and Empirical Analysis", Bank of England
Discussion Paper 3.
Mwega, F.M. (1990) "An Econometric Study of Selected Monetary Policy
Issues in Kenya" (unpublished).
Nickell, S.J. (1985) "Error Correction, Partial Adjustment and All That: an
Expository Note", Oxford Bulletin of Economics and Statistics, 41, No. 2,
119-29.
Osborn, D.R., A.P.L. Chui, J.P. Smith, and C.R. Birchenhall (1988)
"Seasonality and the Order of Integration for Consumption", Oxford Bulletin
of Economics and Statistics, 50, 361-377.
Pathak, D.S. (1981) Demand for Money in a Developing Country: An
Econometric Study, Indian Economic Journal, 29 10-16.
Picks (1990) International Currency Yearbook, New York.