0% found this document useful (0 votes)
54 views34 pagesDynamic Model
Distributed-lag models incorporate lagged values of exogenous and/or dependent variables to explain economic relationships, reflecting how past experiences influence current behavior. Various lag structures, such as declining, rectangular, and inverted V patterns, can be employed to model these relationships, with researchers often experimenting to find the best statistical fit. The estimation of these models can be complex due to issues like multicollinearity and the need for meaningful reduction of lagged variables, leading to methods like the Almon scheme for polynomial lag estimation.
Uploaded by
dinkyaneja1Copyright
© © All Rights Reserved
We take content rights seriously. If you suspect this is your content, claim it here.
Available Formats
Download as PDF or read online on Scribd
0% found this document useful (0 votes)
54 views34 pagesDynamic Model
Distributed-lag models incorporate lagged values of exogenous and/or dependent variables to explain economic relationships, reflecting how past experiences influence current behavior. Various lag structures, such as declining, rectangular, and inverted V patterns, can be employed to model these relationships, with researchers often experimenting to find the best statistical fit. The estimation of these models can be complex due to issues like multicollinearity and the need for meaningful reduction of lagged variables, leading to methods like the Almon scheme for polynomial lag estimation.
Uploaded by
dinkyaneja1Copyright
© © All Rights Reserved
We take content rights seriously. If you suspect this is your content, claim it here.
Available Formats
Download as PDF or read online on Scribd
/ 34
13. Lagged Variables
Distributed-lag Models
Distributed-lag models are models which include lagged values of the exogenous
variables and/or lagged values of the dependent variables among the set of
explanatory variables.
The general form of a distributed-lag model with only lagged exogenous
variables is
Y= at boX; + By Xp_4 + boXra +... tb Xpaig to. te
They are called distributed lag models because the influence of the explanatory
variable on the dependent variable is distributed over a number of past values of
X. The number of lags, s, may be either finite or infinite. However, in order to
avoid explosive values of ¥;, we assume that the b’s have a finite sum
‘
Eb <0
0
The average lag is defined as the weighted average of all the lags involved,
with weights being the relative size of the respective b coefficients
s
Din,
Average lag ae =z i
=o
Lo | Db
i=0 i=0
Lagged values of the variables are important explanatory variables in most
economic relationships, because economic behaviour in any one period is toa
great extent determined by past experience and past patterns of behaviour.
Some examples will illustrate the importance of lagged variables.
Example 1. The consumption function
Recent versions of the consumption function postulate that the current level
of consumption depends on past levels of consumption, due to ‘habit persistence’,
on current income and past levels of income and other factors
Ce = LGa as Con as es Yor Yew as Yrmas- +e Xan Xae- +)
Example 2. The demand for durables
The demand for durables (Dg) depends, among others, on present income (Y;).
past levels of income (Y;-,) which determine the amount saved for the acquisi-
294 Lagged Variables 295
tion of the durables, on the stocks of durables or, equivalently, on past
acquisitions of durables (Sa, +1), prices (P,), and so on
Da = SY Kaas Sa, tn Pe)
Example 3. The demand for inventory investment
The firms usually define their inventories on the basis of the actual sales of
the three past periods (and other factors)
= f(Xr-1s Xe-2, Xr-3 --)
Example 4. The investment function
Investment projects depend on past outputs, on expectations about future
profits, on capital stock and other factors
Tee Xe Xe Xena Me Kea tna)
where X= level of output
l= profit
K = capital stock
r = interest rate
Example 5, Demand for nondurables (e.g. tobacco)
The habit-persistence is a characteristic of human behavioup? The quantity
demanded for food, tobacco and other nondurables depends, among others,
on the past levels of consumption of these commodities
Or = S(Qe—-a. Yo Pros)
Indeed lags are involved in all economic behaviour. We live in a dynamic world
of continuous adjustment. Clearly an adjustment process takes time, the length
of the time period depending on the nature of the Particular phenomenon. “Yet
Ww] erent ~~
equilibria are compared. adjustments are assumed instantaneous. The proces of
adjustment and the lags involved are scarcely discussed. Yet these lags are o}
paramount importance for decision making. It is ccuciat forthe goverment
official to know how fast, after how many time periods will the econamic units
react to changes of various policy variables (instruments), How-fast will
Consumers react to the imposition of a sales tax or a credit squeeze? How fast
will firms react to tax concessions and other incentives for investment and
innovation? How fast will be the effects of a devaluation? How fast will
‘investors react to changes in the interest rates?
Similarly the lags involved in the demand for consumers goods following a
change in the ‘policy instruments’ (price, quality, style, advertising) of a firm are
of crucial importance for managerial decisions. Again the microeconomic theory
of the firm is mostly static. The lag patterns in the adjustment of buyers to
changes in various variables are scarcely mentioned. 296 Econometric Problems: Second-order Tests
In view of the nature of economic behaviour any realistic formulation of
economic models should involve some lagged variables among the set of
explanatory tarables-Lagged saablr sane. for ‘ekg ito seeoua he
length of time in the adjustment processes of economic beh:
mamic. Lagged models have become
increasingly popular in applied econometric research. Their refinement has
made possible the Handling of expectation3about future events, if only in a
rigid and not very satisfactory way. In short, lagged models offer much flexibility
to the formulation of models of economic behaviour.
Distributed lag models involve a high degree of empiricism, which is due to
the unsatisfactory state of economic theory with regard to the length of the
adjustment processes of econ phenomena-Eeonemic-theary, even where
it recognises the importance of time lags, never suggests the precise number of
igs that shoul. be motoded ina function: Rather the pattern of lags is explored
and determined from the available sample observations, by adopting an experi-
mental approach involving various lag patterns and then choosing among them
the one that gives the best statistical fit. The researcher experiments with models
including different lag patterns geometric lags, arbitrary lags, polynomial lags,
compound geometric lagsPand chooses among them the one that gives the most
satisfactory fit on the basis of statistical (mainly) criteria.
We shall start with the examination of models including lagged values of
exogenous variables only. We shall next extend the analysis by examining
models including lagged values of the endogenous variable (Y;-1, ¥¢—2 etc.)
among the set of explanatory variables. In a last section we shall discuss methods
of estimation of the parameters of distributed lag models.
13.1, EXOGENOUS LAGGED VARIABLES
Assume that ¥ depends on the values of X over s periods. This it.a finite lag
structure of the form
¥, =a + boX, +byXp-1 +... +b, Xr—g tu,
13.1.1, ESTIMATION OF LAGS BY APPLYING OLS
If the model includes only lagged values of the exogenous variable(s) in the
set of explanatory variables we can, in principle, make the usual assumptions
about the error term w
u~N(O, 02)
Euyy; Gt)
E(u;X;) = 0 G=1,2...4)
and proceed with the application of OLS to the sample observations,
However, two difficulties are almost certain to arise in attempting to apply
OLS to this type of model. FirstIK{f the number of lags is large and the sample
is small (which is usually the case with time series data), we may be unable to Lagged Variables 297
estimate the parameters, because there will be no adequate degrees of freedom.
“To carry out the tradittonat statistical tests of significance, Secondly, it is almost
certain that we will havémulticollinearity problems, since one might expect a
_strong correlation between successive values of the samé-variable. With strong
collinearity the values of the estimates will be imprecise and their standard
errors will be large so that we may be led to mis-specification of the model by
in fact be important.
To avoid these difficulties various methods have been suggested, all of which
have as their basic aim a ‘meaningful’ reduction of the number of lagged variables.
This is achieved by imposing restrictions on the b’s and constructing new
variables (we shall call them MW) from-atinear-combination of the lagged
‘variables, The methods differ in the weights which are used in constructing these
new variables, i.e. in the way the restrictions on the b’s are imposed.
13.1.2, ESTIMATION BY ASSIGNING ARBITRARY VALUES TO THE WEIGHTS OF
THE LAGGED VARIABLES
Our concern is mainly to find ways foCGeonomisig degrees of freedom phat
is reduce meaningfully the number of parameters to be directly estimate
the given sample. The simplest method is to construct new variables (W’s) from
a linear combination of the lagged X’s, assigning arbitrary numerical values as
weights for the individual lagged variables.
There are various versions in this approach.
(a) The declining lag-scheme
In this scheme it is assumed that the weights are declining, ie. the more
recent values of X have a greater influence on Y than more remote values. For
example assume s = 4, Then
Wye = WoXs twrXi-a + WoXe—-2 + Wasa t WAX 4
where
Wo > W, >We > Ws >W4
For example
Wye = EX A EXe1 + Xt-a gM a tg Xena
(b) The rectangular lag-scheme
In this scheme all lagged values are given equal weights, i.e. it is assumed
that each past value of X has the same influence on Y.
For example
Woe =H Xe + EXi-1 + EXt—-2 + EXt-3 + HA 4
(0) The ‘inverted V’ lag pattern
Here it is assumed that the weights are initially increasing, and subsequently
declining. 298 Econometric Problems: Second-order Tests
For example
War = fo Xe +4 Xe—1 + EXt-2 tea th Kea
any applications economic considerations may suggest which is the most
probable pattern of lag structure. For example in a consumption function the
warts shapes of weights are shown in figure 13.1.
w, w ,
$ 5 %
3 g g
etitit'si_
01234 t 01234567 t '23456789t
(time tog) (time |ag) (time tog)
Geometically declining weights Rectangular Inverted V
Figure 13,1
declining weights for past levels of income seem more plausible, while in an
investment Tunction, where the Tagged X"s are ‘ast capital appropuialions, it may
be more reasonable to assume an ‘inverted V” lag scheme.
In general, the method of arbi ights is an extremely arbitrary method
ot edn ant age Te ours apecies aot ony the genta form of
the weights (declining, rectangular, ‘inverted V°), but he also assigns actual
values to the weights (w's) of the lagged variables. Having constructed various
W;’s the researcher experiments with functions including each W; in tum asa
single explanatory variable. For example one may apply OLS to the models
CD) Y= eo tea Wie tue
2) Ye = ag + ay Woy + Uy
QB) Yp=do + dy Wy, + uy
() p= bo +b X, + bo, ty,
where
Web Xin tEX 2 tha ty Mina
and so on. The choice among these alternative models is based mainly on
statistical criteria, that is on the goodness of fit as measured by R? and on the
size of the standard errors of the estimates. Sometimes rationalisation of the
choice is attempted on @ priori economic considerations as well. Lagged Variables 299
A good example of the arbitrary-weights lag scheme is provided by De Leeuw’s
study of investment, based on quarterly data.’ He experimented with the follow-
ing three types of distributed lag functions:
CI) fy = br + AX aa tN Xan te tN) tue SACI
This is a standard distributed lag function with declining weights. (The weights
are geometrically declining; this form has been suggested by Koyck and will be
discussed in detail in section 13.2 below.)
(2) Te= baw, #WXran HWXea tt WK ena) te
This is a lag scheme with constant weights (rectangular distribution of weights),
in which all output levels are assigned an equal weight of w for all s past periods.
3) ty = bg [Xp + 2Xpa + 3X te. t AX ea +R Xin t
+ (K-2)Xe—n—1 + (K-3)Xp ea t+ Xan ead tu
This is a function with an ‘inverted V’ lag pattern, where the weights increase
initially and then fall. In particular DeLeeuw assumed that for a total lag of s
periods the first half of the weights are proportional to the increasing series
1, 2,3...,8/2 (for even values of s), and the other half of the weights are
proportional to the declining series (s/2) — 1, (s/2)-2,.. ., 3, 2, 1. DeLeeuw,
after numerous experimentations with various lag periods, concluded that the
best fit was obtained by a function of the third type (with an ‘inverted V" lag
scheme) and with a twelve-quarters lag period (i.e. s = 12 quarters). His preferred
function, chosen on the basis of highest R?, was of the form
Ty = bg(Xp + 2Xp-1 + 3Xp-2 + 4Xp—g + Xp 4 + OXp 5 + 7Xp—g + OXt-7 +
+ 9X pg #4X,—9 + 3X:—10 + Xena + Xe)
13.1.3. THE ALMON SCHEME OF POLYNOMIAL LAG
(S. Almon, ‘The Distributed Lag Between Capital Appropriations and
Expenditures’, Econometrica, vol. 33, 1965, 178-96.)
‘Almon proposed the following method for estimating the parameters of the
lagged exogenous variables.
The lagged model is finite and includes only exogenous lagged variables
Y, = boXe + Bi Xp-1 +... + Xen te (3.1)
Instead of attempting to estimate directly all the b’s (which are s + 1 in number)
+ F, DeLeeuw, “The Demand for Capital goods by Manufacturers: A Study of Quarterly
Time Series’, Econometrica, vol. 30, July 1962, pp. 407-23. A similar arbitrary-weight
approach has been applied by P. J. Lund and K. Holden, ‘An Econometric Study of Private
Sector Gross Fixed Capital Formation in the United Kingdom, 1923-1938", Oxford
Economic Papers, vol. 20, 1968, pp. 56-73. 300 Econometric Problems: Second-order Tests
by applying OLS to the above function, we may obtain estimates of all the b’s
indirectly as follows.
‘Assume that the 6s in the lagged model can be approximated by some
function 5 ~ f(z). The function f(z) is unknown if we do not make any prior
assumptions about its form, The usual assumption is that the function f(z) may
be approximated by a polynomial in z of the rth degree
S$) * ag +a,2 +4327 ta32? +... 40,2"
This is called ‘the approximation polynomial’, because of its role. That is f(z)
yields the values of the b's (approximately) if we know the a’s and the degree
(©) of the polynomial. Almon develops a general method for estimating f(z)
which is highly complex and has several disadvantages.’ In most cases the
following simpler approach is adopted.
Step 1. We specify the degree of the approximating polynomial, r, and the
number of lags, s. Usually the degree of the polynomial is assumed low (r= 3
or 4). If the degree of the polynomial is high we do not achieve the intended
reduction in the number of the parameters which we will estimate from our
sample. (In a later stage we will examine how one may increase the degree r and
test whether the higher-degree polynomial yields a better fit. See below.)
Step 2. We express the b’s in terms of the a’s of the approximation poly-
nomial, by assigning to z the successive integer values z = 0, z= 1,
2=2,...42 =5. Thus we obtain the following system.
bo = f(0) = Go)
by =fl)=@ota, ta, + a3 +...+ a,
by =f(Q)= (ao + 2a, + 20q + Pay t...+ Ya, (13.2)
bs = f(3) = (do + 3a, + 37a, + 3305 +. a,
b, =fls) =(@o tsa, +87a, + say +...4 Sa,
In this system the b’s are expressed as linear functions of the a’s. We shall call
this system “the b-system”. If we knew the a’s, we could obtain the b’s by
substitution in the above system.
Before we show how to obtain the estimates of the a’s it is useful to rewrite
the coefficients of the ‘b-system” in a table, so as to show clearly the numerical
pattern of these coefficients. This is done in table 13.1.
" Note that the general Almon scheme is computationally cumbersome and has several
disadvantages in applied research (see J. Johnston, Econometric Methods, 2nd Edition,
McGraw-Hill, 1972, pp. 296-7). The method presented in this section is a simplified version
of the general Almon scheme, which is most often used in practice because of its simplicity
and its computational advantages, Lagged Variables 301
Table 13.1 Numerical Coefficients of the *b-system’
bj 4 G=0,1,2,... 07)
G=01, 2.08) a, a a, Oy... say
be 1 0 0 0.
>, 1 1 1 Leese
b, 1 2 ? De. 2
b 1 5 2 2
In general
b= Ey § GHO 1a)
jm
Step 3. We obtain estimates of the a’s by applying OLS to the transformed
model
Y,=GoWo +a,W, +a,W, +... +a,W, + Uy (13.3)
where the W’s are linear combinations of the lagged X’s, with the weights shown
in table 13.2.
Table 13.2, Weights in an Almon Lag Scheme
x
ww x Xena Xena Me-a Xtna Xt wee Xas
W, 1 1 1 1 1 1 1
W, 0 1 2 3 4 5 s
W, 0 1 2 3 e 3 2
W, 0 1 QP » ” » 2
wy, 0 1 2 cy a s st
w, 0 1 a x a So y
Looking at the pattern of the weights we may derive the following generalisations,
1. There will be as many constructed W's as the arbitrarily chosen degree of
the polynomial plus one (that is the number of the W’s isr + 1).
2. The W’s are linear combinations of all the X values (current and lagged).
(2) The weights used in the first constructed variable (Wo) are all equal to
unity
Wo= Xp Xe tXea +... + Xee 302 Econometric Problems: Second-order Tests
(b) The weights of W; will be the simple increasing series of integers
0,1,2,3,...,8
that is
Wy = Xeon + 2Xpoa + 3Xp-3 +
(c) The weights of the third constructed variable, Wy, will be the squares
of the weights of W,, that is
Wy =Xpoy + PX pg + 7X pg tet PX pag
(d) The weights used in the construction of Ws are the weights of W;
raised to the third power, that is
Wy = Xpoy + PXpa + HPXp-9 +
(e) In general the weights of W, are the weights used in constructing W;
raised to the rth power
We = Xp UX pg 3X to tT Xs
Lit SXeg
tS Xey
Formal Derivation of the W's
The procedure for the derivation of the W's is the following. We substitute the b's for the
a's in the original function 13.1, and obtain
Ye=anX~ + (p40, +a, +...44)Xt,
+ (ay + 2a, + 2a, +... + Zapp,
+ (a, + 3a, + 3a, +... + 3 a,)Xpy 3.4)
+ (ag +80, +540, +...4 80) Neg tue
Rearranging we obtain
Yee a(Xe+ Xt + Xen + Xt tee Xt)
aXe, + Xe. + BXey +... 4 SX pg)
taXe + PXp, + PAs tt PX) (13.5)
+a(Xy, + Xe, + Xe, +FXy ty
Clearly the expressions in brackets are the W's appearing in 13.3.
Step 4. We substitute the OLS estimates of the a’s obtained from 13.3 into
the “b-system” (13.2) and we find the estimates of the parameters of the lagged
model, bo, 5: ,52,..., Bs.
Example
1, Assume that the number of lags is 7 (ie. = 1, 2,..., 7), and the degree of the poly-
nomial is 3 (that is, r = 3). Thus our original model is
Vp = by Xp +b, Xp.4 +b, X pg te. tbs X py Hay
2. We express the b's
toz the values z = 0,2 = 1,
terms of the a's of the approximation polynomial, by assigning
soy2 = 7. The “b-system’ is Lagged Variables 303
ta, +a, ta,
9 + 2a, + 22a, + 2a,
9 + 3a, + 370, +3?
otha, + 42a, + 4?
9 + Sa, + 52a, + 52a,
2, + 6a, + a, +60,
+ 7a, + 7a, +70,
3. The required number of W's is 4(= r + 1). They are
We=Xp +X, +Xtg tee
W, =X + Keg t 3X +.
Xp HK tN toe
Wy = Xp + VX te PX, te.
Xter
Mey
PX,
Px,
‘Thus the transformed model is
Y¥,=a9W, +a,W, +0,W, +2,W, +uy
Applying OLS we obtain the estimates d,,4,, 4,45.
4. We substitue the OLS estimates, a's, in the d-ysiem” and we obtain the estimates
of the B'sof the original model, By With the application of the Almon
scheme we managed to obtain four estimates vay +4,,2,) from our sample, instead of
the eight (by, b, ,...,5,) which we would have to compute if we applied OLS to the
original model directly.
In summary, the steps involved in the simplified Almon polynomial lag
scheme are the following.
1. We choose an arbitrary degree for the approximation polynomial. Usually
alow degree, say r= 2, orr = 3.
2. We assign the values 0, 1, 2,.. .,$ to the 2’s of the approximation poly-
nomial and we find the system of b’s in terms of the a’s of this polynomial. We
call this the ‘b-system’. It is of the form (13.2).
3, The coefficients of the ‘b-system’ are used to construct r + 1 composite W
variables. The weights in the W’s are given in table 13.2.
4. We estimate the a’s by applying OLS to the equation
Y, =aoWo + a,W, +a,W2 +a,W;+...+a,W, tu,
5, We use the OLS @ to substitute into the ‘b-system’ and obtain the B's.
This simplified scheme has the following advantages over the general Almon
scheme. (See J. Johnston, Econometric Methods, 2nd edition, p. 297.)
Firstly, it is computationally simpler.
Secondly, it yields a direct test of the degree of the approximation polynomial.
In our example a third degree polynomial was specified and ay denotes the coef-
ficient of the third degree term; and it is clear that a test of significance for as is
provided by the standard error of a3 from OLS in step 4.
Thirdly, changing the degree of the polynomial involves adding extra W
explanatory variables, leaving the previous variables unchanged. 304 Econometric Problems: Second-order Tests
The above simplified Almon scheme can easily be extended to many
explanatory lagged variables, each having a different length of lag. For example
assume that we have k variables with their lagged values
Xue Ma-y Xie-2
Xo Xaa-1 X2e-2)
Xne Xece-1y Xece-2y---Xece-eyy
Clearly the lags s,, 52 ... $, need not be the same. To estimate such a model
‘we must specify the degree of the polynomial r and the length of the lag period s
for each explanatory variable and repeat the Almon scheme for all X’s.
Tinsley has extended the model of lagged variables to allow for changes in the
lag coefficients (b's) over time. (P. A. Tinsley, ‘An Application of Variable
Weight Distributed Lags’, J. Am. Statist. Assoc., vol. 62, 1967, pp. 1277-89.)
It is assumed that the effect of X,_; on Y, depends not only on the length
of the time period of the lag and on the value of X;_,, but also on the value of
some other variable, y/, which causes the value of the lag coefficient b; to
change. Thus we may write
be=catrh-1 =0,1,2,.
Substituting in the lagged model
Vy = oXy + B:Xp a +. +B Xray te
+3)
we obtain
Ye = CoXe tcp Xe-a +--+ CeXe—e + YO(WXe) + 11 (Wea Kea) +.
+ Ya(We-aXe—s) + Ur
This form includes lagged values of two explanatoty variables, X and WX, and
takes into account changes in the lag coefficients of the original model. For
the estimation of this transformed model we need a twofold application of the
Almon scheme.
13.2, ENDOGENOUS LAGGED VARIABLES
Up to now we have examined lagged models including lagged values of
exogenous variables only. In this section we will discuss models with lagged
values of endogenous variables among the set of regressors.
13.2.1, KOYCK’S GEOMETRIC LAG SCHEME
s one of the most popular distributed lag models in applied research.
‘See L. M. Koyck, Distributed Lags and Investment Analysis, North-Holland, 1954. Lagged Variables 305
Koyck’s distributed lag model assumes that the weights (lag coefficients) are
declining continuously following the pattern of a geometric progression.
The original model includes only exogenous lagged variables
Y, = a9 + boX, + by Xi-1 + baXy-a to. ty
where
u~N(O, 07)
Etuiw)=0 Ft
EUX)=0 G=1,2,...k)
Koyck’s geometric lag-scheme implies that more recent values of X exert a
greater influence on Y than remoter values of X. In particular the lag coefficients
of this model decline in the form of a geometric progression
by = Abo
ba = No
and in general
bj=Nbo = O 0 and all
As a consequence of the violation of Assumption 6 of the linear regression
model the OLS estimates will be biased in small samples.
Thirdly, autocorrelation of v,, superimposed on values of Y;— 1, which are
interdependent with the error term v, renders the OLS estimates not only biased,
but also inconsistent in large samples. The OLS estimates are asymptotically
biased, that is the Bias in small sampies, due to £(¥; 1, ¥;) # 0, does not
vanish as n + ©, hence the estimates are inconsistes.t.
Proof: We will prove these results for the simple model
Y= bY, + (13.9)
vee Ate te
where e, satisfies the usual stochastic assumptions,
(a) The smalt sample bias in b
Applying OLS to (13.9) we obtain
Lagging (13.9) by one period and multiplying by \ we find
AY p= = ADY pg + Ay (13.10)
Subtracting (13.10) from (13.9) we find
Yy = (B+ A) ¥y-a — DAY; 2 + 3.11)
Multiplying through by ¥;_,, and summing over all sample observations we have
LY, Yy—g = (b+ A) LY -y - PALY: Nea + Leen (13.12)
Divide by DY", and obtain
z YY
eet “py Sher Ye-2
ST TN = RS (13.13),
5 Ly
Bea 308 Econometric Problems: Second-order Tests
By rearranging we obtain
Taking expected values we have
= (LY a Mee
rireo[ fetete Benes
‘so that
(bias) = £(8) -b #0
(b) The asymptotic bias in b
We established in (13.13) that
Be @+ny— pacientes , Zea
Dy BY
Taking probability limits
(plim by = (b + a) - ba (plim 5)
EM-1¥e-2 5)
(since, as n + ©, D¥;_1¥;_-2 ~ LY;¥;—1, and hence asymptotically
Lye
Rearranging we find
and
[asymptotic bias] = (plim 5) — b (13.14)
Fourthly, the combined violation of two assumptions of the OLS method
impairs the power of the d statistic in detecting autocorrelation.
We established that the asymptotic bias is equal to
(1-5?)
Tox
It is obvious that the asymptotic bias depends on A (the autocorrelation coef-
ficient). In the usual case of positive autocorrelation (A > 0) the bias will be
positive. In other words contemporaneous violation of Assumptions 5 and 6
leads to overestimation of the b, and the bias does not vanish as n increases.
Monte Carlo type studies’ have provided evidence that the bias can be very large
especially for low b’s and large }’s. Thus in the Koyck type distributed lag models
where lagged Y's appear in the set of explanatory variables detection of auto-
correlation is crucial. Nerlove and Wallis found? that the classical Durbin—Watson
} See Z. Griliches, *A note on the Serial Correlation Bias in Estimates of Distributed Lags’,
Econometrica, vol. 29, 1961, pp. 65-73.
? See M. Nerlove and K, Wallis, ‘Use of the Durbin— Watson Statistic in Inappropriate
Situations’, Econometrica, vol. 34, 1966, pp. 235~8. Lagged Variables 309
d statistic is biased towards 2 (its asymptotic value in the absence of auto-
correlation) if ¥;-1 appears as an explanatory variable in the right-hand side of
the equation. This finding has alarmed econometricians unduly, because the bias
of d (towards 2) is serious for models containing only Y;-1. Malinvaud! has
shown that the bias in d tends to decrease if apart from Y;-, there are exo-
genous X’s in the model. Taylor and Wilson” have explored the power of d i
detecting autocorrelation in various models which included various values of X
(i.e. A), various values of R?, and various autoregressive schemes, namely first
order and second order schemes, stable and explosive. They adopted the
‘amended’ form of the Durbin—Watson test (see Chapter 10), namely they
adopted the procedure of rejecting the null hypothesis (Ho: zero autocorrelation)
if d d,,. They found that the d test performs well
(a) the larger the size of the sample, (b) the larger R?, (c) the larger the absolute
value of X (or A), (d) in stable second order autoregressive schemes. But d was
found to perform badly in small samples, in models with low R?, when IX! (or
|p|) was small, and in cases of unstable explosive first order and second order
schemes.
In view of the defects of the d statistic, Durbin? has suggested the following
test for models involving lagged Ys and for large samples (n > 30). Firstly, apply
OLS to the original model and estimate the residuals, e's. Secondly, regress e, on
€¢-1 and all the other variables of the original model and conduct the classical
tests of significance for the coefficient of e,-1. For example assume the model
Y, = bo + by Ye-1 + baXe + bgXr-1 +... tue
Applying OLS we obtain the residuals ¢,'s. We next apply OLS to the function
ep = Co +0, 1 +C2%e—-1 Hea Ket caXea +...
If the standard error of @, is small (s@) <é/2) we accept that there is auto-
correlation in the function. Of course this is a large sample test and nothing
is as yet known about its power in small samples.
In summary, if the model is
Y, = bo +d Yy-1 + baXp +... ty
where 0; = uy — Aue-a O
