Intervention Analysis with Applications to Economic and Environmental Problems

Author(s): G. E. P. Box and G. C. Tiao

Source: Journal of the American Statistical Association , Mar., 1975, Vol. 70, No. 349
(Mar., 1975), pp. 70-79
Published by: Taylor & Francis, Ltd. on behalf of the American Statistical Association

Stable URL: https://www.jstor.org/stable/2285379

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

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

Taylor & Francis, Ltd. and American Statistical Association are collaborating with JSTOR to
digitize, preserve and extend access to Journal of the American Statistical Association

This content downloaded from

138.38.44.178 on Fri, 14 Jun 2024 13:23:02 +00:00
All use subject to https://about.jstor.org/terms
 Intervention Analysis with Applications to
Economic and Environmental Problems
G. E. P. BOX and G. C. TIAO*

This article discusses the effect of interventions on a given response A. Monthly Average of Hourly Readings of 03 (pphm)
variable in the presence of dependent noise structure. Difference in Downtown Los Angeles (1955-1972)a
equation models are employed to represent the possible dynamic
WEIGHTS
characteristics of both the interventions and the noise. Some proper-
ties of the maximum likelihood estimators of parameters measuring
level changes are discussed. Two applications, one dealing with the
photochemical smog data in Los Angeles and the other with changes pphm

in the consumer price index, are presented. 9.0

8.0
Intervening events
7.0
1. INTRODUCTION
6.0-

Data of potential value in the formulation of public 5.0-

and private policy frequently occur in the form of time 4.0

series. Questions of the following kind often arise: "Given 3.0

a known intervention,' is there evidence that change in 2.0

the series of the kind expected actually occurred, and, if 1.0

so, what can be said of the nature and magnitude of the

58 56 57 58 59 60 61 62 63 64 65 66 67 68 69 70 71 72
change?" TIME

For example, in early 1960 two events occurred, here

a With the weight function for estimating the effect of intervening events in 1960.
referred to jointly as the intervention, which might have
been expected to reduce the oxidant (denoted by 03)
pollution level in downtown Los Angeles. These events However, the ordinary t test would be valid only if the
were the diversion of traffic by the opening of the Golden observations before and after the event of interest varied
State Freeway and the coming into effect of a new law about means gu and ,2 not only normally and with
(Rule 63) which reduced the allowable proportion of constant variance but independently. In the examples
reactive hydrocarbons in the gasoline sold locally. The quoted, however, the data are in the form of time series
expected effect of this intervention would be to produce in which successive observations are usually serially
a more or less immediate reduction (i.e., a step change) dependent and often nonstationary, and there may be
in the oxidant level in early 1960. Figure A shows the strong seasonal effects. Thus the ordinary parametric or
monthly averages of oxidant concentration level from nonparametric statistical procedures which rely on in-
1955-72 in downtown Los Angeles [6]. Using this highly dependence or special symmetry in the distribution
variable and seasonal time series, is there evidence for a function are not available nor are the blessings endowed
change in level and, if so, what is its magnitude? by randomization.
Many other problems of this kind have come to our An approach we initiated earlier [2] was to build a
attention in recent years. These have included the stochastic model which included the possibility of change
possible effect of the opening of a nuclear power station of the form expected. Such model building is necessarily
on measurements made on river samples, the possible iterative and, as discussed, e.g., in [3], involves inferences
effect of the Nixon Administration's Phases I and II on from a tentatively entertained model alternating with
an economic indicator, and the possible effect of pro- criticism of the appropriate tentative analysis. The
motions, advertising campaigns and price changes on the process proceeds [1] by successive use of Identification
sale of a product. (tentative specification of the model form), Fitting, and
Available procedures such as Student's t test for Diagnostic Checking. Using these ideas in the present
estimating and testing for a change in mean have played context, we come to the following general strategy:
an important role in statistics for a very long time.
1. Frame a model for change which describes what is expected
to occur given knowledge of the known intervention;
* G.E.P. Box is R. A. Fisher professor and G. C. Tiao is professor and chairman,
Department of Statistics, University of Wisconsin, Madison, Wis. 53706. This
? Journal of the American Statistical Association
research was supported in part by a grant from the American Petroleum Institute
March 1975, Volume 70, Number 349
andin part by U.S. Army Research Office under Grant DA-ARO-D-31-124-72-G162.
1 A term introduced in [5], based on our earlier work E2]. Invited Paper, Theory and Methods Section

70

This content downloaded from

138.38.44.178 on Fri, 14 Jun 2024 13:23:02 +00:00
All use subject to https://about.jstor.org/terms
 Intervention Analysis with Applications 71

2. Work out the appropriate data analysis based on that model; where the polynomials X1(B), 42(BS), 01(B), 02(BS) are
3. If diagnostic checks show no inadequacy in the model, make degrees pt, P2., ql, q2, respectively.
appropriate inferences; if serious deficiencies are uncovered,
make appropriate model modification, repeat the analysis, etc.
3. A DYNAMIC MODEL FOR INTERVENTION
Suppose the data . . . Yte1, Yt, Yt?, . . . are avail-
Frequently the effects of exogenous variables t can be
able as a series obtained at equal time intervals. Follow-
represented by a dynamic model of the form
ing, e.g., [Ell we will employ models of the general form
k k

yt = f(v., , t) + Ne (1.1) A6, t,t) = E Et3= E {fw(B)/bj(B)}Itj (3.1)

i=1 j=1
where:
where:
yt = F(Yt) is some appropriate transformation of Yt, say
1. The Wtj represent the dynamic transfer from tit;
log Yt, (Yt)1 or Yt itself;
2. The parameters Kc previously lumped together are now
f(v., t, t) can allow for deterministic effects of time, t, the effects
denoted by B and (;
of exogenous variables, j, and in particular, interventions;
3. The polynomials in B
Nt represents stochastic background variation or noise;
vc is a set of unknown parameters. 5j(B) = 1 - 51jB- *.. -r1jBri and

In Section 2 we discuss a general integrated mixed cj (B) = woj - w jB- w -,,jBi

autoregressive moving average model for representing are of degrees rj and sj, respectively;
the noise Nt. A class of general dynamic models capable 4. We shall normally assume that coj(B) has roots outside, and
of representing the effect of interventions is given in 5j(B), outside or on, the unit circle.
Section 3. The associated parameter estimation pro- In general, the individual ttj could be exogenous time
cedures are given in Section 4. In Section 5 two illustra- series whose influence needs to be taken into account.
tive examples of intervention analysis are presented. The For the present purpose, however, some or all of them
first concerns the Los Angeles oxidant data, and the will be indicator variables taking the values 0 and 1 to
second considers possible effects on the consumer price denote the nonoccurrence and occurrence of intervention.
index of recent government actions. Finally, in Section 6, For illustration, suppose for a single exogenous variable
the nature of the maximum likelihood estimators for (k = 1) the model is
some specific level-change parameters is discussed in
some detail. Yt = 't + Nt = (co(B)/&(B))St + (O(B),/p(B))at ; (3.2)

then the transfer 'jt to the output from it is generated

2. A STOCHASTIC MODEL FOR THE NOISE the linear difference equation
We suppose that the noise Nt = yt - f(k, j, t) may be
b(B)'yt =co(B%S
modeled by a mixed autoregressive moving average
process Figures B(a), B(b) and B(c) show the response St
ep(B)Nt = O(B)at (2.1) transmitted to the output for various simple dynamic
systems by an indicator variable representing a step.
where: We can denote such an indicator by it = St(T) where
1. B is the backshift operator such that Bye = y-1i;
Sta)
2. ... a-,, at, at+,, ... is a sequence of independently = 1,t<>T
O,t
distri-> T * (3.3)
buted normal variables having mean zero and variance (a,.)2
which for brevity we refer to as "white" noise;
3. @(B) = 1 -01B-02B2 _qBq, ep(B) = 1 - wpiB -
Similarly, we use P(T) for a pulse indicator where
.2.B2. -PBP are "moving average" and "autoregres-
sive" polynomials in B of degrees q and p, respectively; Pt(T) {I t-T- (3.4)
4. the roots of 0(B) lie outside,.and those of (B) lie on or outside
the unit circle.
Referring to the figure for the case w
For the representation of certain kinds of homogeneous for the Los Angeles 1960 intervention,
nonstationary series, the operator (p(B) is factored so that
that the change could be modelled as in Figure B(a), so
that immediately following the known step change in the
p (B) = (1 - B)d4(B) (2.2)
input, an output step change of unknown magnitude
where the roots of + (B) all lie outside the unit circle. This would be produced according to
corresponds to the use of a stationary model in the dth
Wt = cBS jT)
difference. Also, for seasonal data with period s (e.g.,
Sometimes a step change would not be expected to pro-
monthly data with s = 12), it is often helpful to write
ep (B) = p1 (B) p2(B8) and 0(B) = 01 (B)02(B8) with duce an immediate response but rather a "first order"
(P2(B8)
- (1 - B)D2(Bs) to allow for seasonal nonstationarity. dynamic response like that in Figure B (b). The ap-
Finally, we entertain a class of noise model of the form propriate transfer function model is then

fl(B+2(s(- B)d(1 -BS)DNe = 0(B)02(B8)ae (2.3) Wt= {coB/(1 - eB)} SiT)

This content downloaded from

138.38.44.178 on Fri, 14 Jun 2024 13:23:02 +00:00
All use subject to https://about.jstor.org/terms
 72 Journal of the American Statistical Association, March 1975

(8 < 1). It is readily shown that the time constant of If it were believed that the full impact of intervention
this system is estimated by { -log,b }-' and the steady might not be felt until the second month, after which
state gain is w/ (1 - 8). When 6 approaches the value there would be a decay and possibly a residual effect as
unity, we have the transfer function model in the previous case, the model

t {=wB/(1 - B)I}StT) jt = {cooB + (coiB2/(l - SB)) + (co2B2/(1 -B))IPt )

in which a step change in the input produces a "ramp"'
might be appropriate. This would insert a preliminary
response in the output (Figure B (c)).
value wc into the output (which in the preceding context
Note that since
would usually be less than wi). The same form of model,
(1 - B)St(T) -p t(T) (3.5) shifted forward and with some sign changes in the
parameters, could be useful to represent the effect of
any of these transfer functions could equally well be price changes. In the application shown in Figure B(f),
discussed in terms of the unit pulse Pt(T) and some- coo would be positive and would represent an immediate
times matters are best thought of directly in terms of rush of buying when a prospective price change was
Pt(T). Thus, suppose we have monthly sales data and announced. The reduction in buying immediately after
wish to represent the effect of a promotion or advertising
the change occurred would be represented by wl + '2
campaign lasting less than a month. The simple first
and the final effect of the change would be represented by
order model
C2 which is shown as negative but, of course, could have
a zero or positive value.
=t { wB/(1 - B)}Pt(T)
Obviously, these difference equation models may be
might do this (Figure B(d)) with cow indicating the initial readily extended to represent many situations of potential
increase in sales immediately following the intervention
interest.
and 8 representing the rate of decay of this increase. The following points are worthy of note:
This particular model implies that no lasting effect will
occur as a result of the intervention. When this might (i) The function yjt represents the additional effect of the
not be so, the model B(e) intervention over the noise. In particular, when Nt is non-
stationary, large changes could occur in the output even with

St = {I (B/(1 (- B)) + (W2B/(1 - B))}Pt(T) no intervention. Fitting the model can make it possible to
distinguish between what can and what cannot be explained by
could be used in which the possibility is entertained that the noise.
a residual gain (or loss) in sales W2 persists. (ii) Intervention extending over several time intervals can be
represented by a series of pulses. A three month advertising
campaign might be represented, for example, by three pulses
B. Responses to a Step and a Pulse Input a whose magnitude might represent expenditure in the three
months.
Input

stT) I,,4 @ @ , I _ _ L _ _ PT
4. CALCULATIONS BASED ON THE LIKELIHOOD
STEP PULSE

Output Suppose we entertain a model of the form

w(B) S(T) s(B) (T) k
8(B) t 8(B) t
yt = E 'gtj + Nt (4.1)
3=1

where Ek=j Wtj is the transfer function given in

aaBs(T)t wB (T)
associated with known interventions, Nt assumes the
form in (2.3), and a time series is available of length
(0) (d) n + d + sD. Then the likelihood may be obtained in
terms of an n dimensional vector w whose tth element is
Wt = (1 - B)d(l - B8)D(yt - E= tj). The corre-
sponding model for wt,
iB (T) - .P
Wt = {O,(B)O2(B8)/41(B)02(B8)}at , (4.2)
(b) (e)
is stationary. Thus, following the argument given, e.g.,
in [1, p. 273], and with the vector 5 having for its g
elements the stochastic and dynamic parameters in the
wB (T) 0B w B (T)
model, the likelihood function may be written
(c) (f) L(T/ (-a)2 1y) = (2- (-a)2Y-(n/2) I M' 12
* exp {-S (I)/2GTa)2 } (4.3)
8 (a), (b), (c) show the respon
tion models; (d), (e), (f) show the response to a pulse for some models of interest. where M' (~a)2 is the covariance mnatrix of the vector

This content downloaded from

138.38.44.178 on Fri, 14 Jun 2024 13:23:02 +00:00
All use subject to https://about.jstor.org/terms
 Intervention Analysis with Applications 73

w and States consumer price index, to determine the effect of

n
known interventions.
S(0) = w'Mw = E [at I y, 5]2 (4.4)
t =-x0
5.1 Example 1: The Los Angeles Oxidant Data
with [at I y, 5] as the expected value of at conditional on
1M[onthly averages of the oxidant (03) level in Down-
1 and y.
town Los Angeles from January 1955 to December 1972
If none of the roots in (4.2) is close to the unit circle,
are shown in Figure A.
then for moderate and large n, the likelihood is dominated
Identification (Specification) of the Model. The periods
by the exponent. The values of the elements of 5 minimiz-1955-60 and 1960-65 were regarded as containing no
ing (4.4), which we shall call the least squares values, are
major intervention which would affect the 03 level. The
to a close approximation also the maximum likelihood
series themselves and the sample autocorrelation func-
values. Alternatively, if we introduce a prior distribution
tions within these periods suggest nonstationary and
such that in the neighborhood where the likelihood is
highly seasonal behavior. The autocorrelation functions
nonnegligible p(5, a) a: p(5)(0Ta)Y', we obtain the pos- of such differences (1 - B'2)yt taken twelve months
terior distribution
apart show significant correlations only at lags 1 and 12.
p( y)cc p([ [ 12 {S0 -(n/2)* (45 This suggests the following model for the noise Nt:
Again for moderate or large samples and for a non- (1 - B'2)Nt = (1 - 01B)(I - 02B12)at . (5.1)
informative distribution p(5), the term involving S(0)
Interventions I, and I2 of potential major importance
dominates and approximately
are:

p(jy) a{S(y)}-(n/2) (4.6) I,: In 1960 the opening of the Golden State Freeway and the
coming into effect of a new law (Rule 63) reducing the allow-
so that the least square estimates correspond with the
able proportion of reactive hydrocarbons in locally sold
point of maximum posterior density. gasoline.
Now if, over the region where the density is ap- I2: From 1966 onwards regulations required engine design
preciable, S(5) is approximately quadratic (and in anychanges in new cars which would be expected to reduce the
given case it is easy to check this numerically), then the production of 03.

posterior distribution is approximately a multivariate t. As already argued, I, might be expected to produce a

Then, step change in the 03 level at the beginning of 1960. The
effect of I2 might be most accurately represented if we
P(1 y) c { 1 + (E Sij(3i - Ai)(j -,j)
knew, for example, the proportion of new cars having
ij (n -) (8a)2) }Jn2) (4.7) specified engine changes which were in the pool of all
where
cars driven at any point in time. Unfortunately, such
Sij = 21a2 S(5) } /0,49-#jj#,=-
data are not available to us presently. We have, therefore,
and (Sa)2 = S(')/(n - g). Thus, for moderate orthe
represented possible effect of intervention as a con-
large
stant intervention
n, 5 is approximately distributed as multivariate normal change from year to year reflecting
with mean 0 and covariance matrix the increased proportion of "new design vehicles" in the
car population. As explained more fully in [6], the engine
V(0) = (Sa)2{Sij}l changes would be expected to slow down the photo-
The square roots of the diagonal elements of V(0) will chemical reactions which produce 03 and, because of the
be referred to as standard errors (S.E.). summer-winter atmospheric temperature inversion differ-
In practice we may obtain ', V(0) and (sa)2 using a ential and the difference in the intensity of sunlight, the
standard nonlinear least squares computer program for net effect would be different in winter when oxidant
pollution is low from that in summer when it is high.
the numerical minimization of S(5). To do this we need
A model form was, therefore, tentatively entertained
only to be able to compute the quantities [atly, 5] for
any 5 and we may proceed as follows. Since the modelfor all the available monthly 03 data from January 1955
to December 1972, which may be conveniently written as:
for wt is stationary, [at I y, 5] will be negligible for values
t <- Q where Q is some suitably chosen positive num- 4t2 t0
ber. We, therefore, replace S(5) by the finite sum Yt = Woi4:tl + '.02 l + W03

,t=-Q [aty, I ]2. It is shown in [1] that the initial 1B 12 1-B 12

values [ao], [a-1], *.., [a-Q] may often be obtained + (1 -01B)(1 - 02B 12) at
conveniently by a process of "back forecasting" which + (1 - B12) a 52
also indicates an appropriate value for Q. where

_ J0, t < January, 1960

5. TWO ILLUSTRATIVE EXAMPLES tl 1 t > January, 1960

The theory developed here is illustrated in this section rl,={i "summer" months June-October beginning 1966
by two examples, one employing the Los Angeles oxidant
data and the other, the rate of change in the United (t=l0,
f ,otherwise.
"winter" months November-May beginning 1966

This content downloaded from

138.38.44.178 on Fri, 14 Jun 2024 13:23:02 +00:00
All use subject to https://about.jstor.org/terms
 74 Journal of the American Statistical Association, March 1975

This allows for a step change in the level of 03 be- controls in August 1971. As indicated in the figure, in the
ginning in 1960 of size woo associated with I, and for three months beginning September 1971, Phase I control
progressive yearly increments in the 03 level beginning was applied; and after that to the end of the recorded
1966 of (002 and C003 units, respectively, for the summer
period, Phase II was in effect.
and the winter months. This representation is admittedly Inspection of the autocorrelation functions of the first
somewhat crude, and we hope to improve on it as more 218 observations and their differences prior to Phase I
data become available. suggests a noise model of the form
Estimation Results. The maximum likelihood estimates
(1 - B)Nt = (1 - OB)at . (5.3)
and the associated standard errors are as follows:

Parameter MLE S.E.

The maximum likelihood values for the parameters are:

Parameter MLE S.E.

Woi -1.09 .13
W02 -0.25 .07 O 0.84 .04
W03 -0.07 .06 xta 0.0019
al -0.24 .03
02 0.55 .04 Inspection of the residuals and their autocorrelations
reveals no obvious inadequacies of this model, so we
Since examination of residuals At fails to show any
adopt it.
obvious inadequacies in the model, we interpret the
results as follows. The marginal distributions a posteriori We now ask the question, "What are the possible
of col, w02 and O003 are very nearly normal and centered effects of Phases I and II?" To answer, we suppose:
at the maximum likelihood estimate values with the (i) that Phases I and II can be expected to produce changes in
approximate standard deviations shown. level of the rate of change of the CPI,

Thus, there is evidence that (ii) that the form of the noise model remains essentially the
same.
(i) associated with I, is a step change of approximately L30
= -1.09 units in the level of 03; On these assumptions, the approximate model (ignoring
(ii) associated with I2 there is a progressive reduction in 03. estimation errors in the noise structure) is
Over the period studied, there is a yearly increment of
approximately -O2 = -.25 in the summer months, but the Yt =01itl + W02it2 + {(1 - .84B)/(1 - B)}at (5.4)
increment (if any) in the winter is slight.
where
5.2 Example 2: The Rate of Change in the U.S. Consumer
11, t = September, October and November 1971
Price Index O0, otherwise
1, t > December 1971
A second example supplies further intuitive apprecia- ~t2 y 7
=0, otherwise
tion for the kind of calculations being performed.
Figure C shows the latter part of a record of the which may be written
monthly rate of change in the consumer price index
Zt= Coo1xt1 + C002Xt2 + at * (5.5)
(CPI) given more completely in [4]. The complete (July
1953 to December 1972) data include 234 successive The sequences {ztj, {xtI}, {xt2} may be readily calcu-
values, 218 of which occurred prior to the institution of lated from the equations

(1- .84B)zt = (1 - B)yt

C. Monthly Rate of Inflation of the U.S. Consumer
Price Index: January 1964-December 1972 (1- .84B)xt1 = (1 -B%
INFLATION (1 -.84B)xt2 = (1 -B) 2
RATE .3

9.0 _ x Phase I in effect using, e.g., the initial approximation zi =X11 = X12 = 0.
o Phase 11 in effect
8.0 _
Also, since

7.0
(1 - B)/(1 - OB)
= 1-B(1-0)(1 + OB + 02B2 +*),

5.0 6 we have
4.0 T
Zt = ye - yt-i, Xtl = 61- it-1 l, Xt2 = t2- (t-1,2
3.0
where gt-e, kt-i,l and et-1i2 are exponentially
2.0
moving averages of values prior to time t, e.g.,
1.0

ytl= (1 - )(ye-i + Oet-2 + O2lIe-3 +--

We see that (5.5) is very much like the regression
equations we are all familiar with in which the deviation

This content downloaded from

138.38.44.178 on Fri, 14 Jun 2024 13:23:02 +00:00
All use subject to https://about.jstor.org/terms
 Intervention Analysis with Applications 75

of yt from its average is related to the deviations of iti so that the maximum likelihood estimator of 3 is
and it2 from their averages. Notice, however, that the n n

model copes with nonstationarity by using not the usual = E Ztxt/E (Xt)2
arithmetic averages, but local exponentially weighted t=1 t=1

with (6.4)
averages which change as the series progresses.
Using (5.5), the constants cool and CO02 may now be Var (A) = (a) 2(E (xt) 2)1
estimated by ordinary linear least squares as t=1

Parameter MLE SE For large n, we apply the results (A.6) and (A.7) in
the appendix to obtain
cool -0.0022 0.0010
002 -0.0007 0.0009 00 00 00

E ztxt = E Q(B)ytR(B)%t = E {tR(F)Q(B)yt

Alternatively, a nonlinear least squares program may be t=1 t=1 t=1 = R(F)Q(B)Ctv(O)
employed to estimate wol, W02 and 0 simultaneously from
where F = B-1 and
the complete set of 234 data values yielding the
n 00
estimates (essentially as before):
E (Xt)- E R(B%R(B% = R(F)R(B)Ct(O)
Parameter MLE SE t=1 t=1

0 0.85 .05 where

cool -0.0022 0.0010
Ca,(k) = E f3tat-k , k = 0, 411 42 ?1, 2
002 -0.0008 0.0009 t=1

The analysis suggests that a real drop in the rate of and for a given k
increase of the CPI is associated with Phase I, but the
effect of Phase II is less certain.
B'Ca,(k) = Ca,(k - 1) , 1 = 0, 41, 42,
Thus,

6. NATURE OF THE MAXIMUM LIKELIHOOD = R (F) Q (B) C V (0)/R (F)R (B) C (0) (6.5)
ESTIMATORS FOR SOME LEVEL and
CHANGE PARAMETERS Var (A) = (cTa)2/R(F)R(B)Ctt(O)
The maximum likelihood estimnators of parameters Making use of (A.10) in the appendix, we can write
such as col, C02 and WO3 in (5.2) and (5.4) which measure R (B)R (F) as
level changes are functions of the data. It is instructive
to consider the nature of these functions. Several results R(B)R(F) = ro + E ri(Bl + Ft) . (6.6)
1=1
in the summation of series useful in the following discus-
sion are given in the appendix. Suppose that it = Pt(T) is a pulse at time T, and a
large number of observations are available before and
6.1 One Parameter "Linear" Dynamic Model after T. In this case

Consider first the dynamic model in (3.2). Formally, rI, k = 0

it can be written
Ctt(k) = %, k o and Cy(k) = YT-k (6.7)

Q(B)yt = ( p(B)/0(B))(w(B)/6(B))it + at (6.1) so that

where Q(B) = qp(B)/0(B), even though in practice the A = (ro)-'R(F)Q(B)yT and Var (A) = (a)2(ro)-1 (6.8)
Yt are only available for t = 1, , n. Since the roots
where it is understood that B is operating on T.
of @(B) all lie outside the unit circle, Q(B) can be ex-
Now, nonstationarity in time series data can often be
pressed as a power series in B which converges for
removed by differencing. In what follows we suppose that
BI = 1.
the polynomial (p(B) in (6.1) is divisible by (1 - B). We
Here we discuss the situation where
consider two special cases of interest.
( (B)10 (B)) (o(B)6f(B)) = lR(B) (6.2)
Case (i). cw(B)/6(B) = fiB , (6.9)
and investigate the nature of the maximum likelihood
that is, the pulse input Pt (T) gives rise to a response at
estimator of f, assuming that (i) the coefficients in Q(B)
time (T + 1) measured by : which dissipates completely
and R (B) are known and (ii) the power series R (B)
after the (T + 1)th period. It should be noted that with
converges for IB = 1. any number of periods of pure delay, the response will
Letting
follow the same pattern but be appropriately shifted. In
Zt = Q(B)yt and Xt = R(B)%t, this case, Q(B) = R(B)F so that, from (6.6) and (6.8),
we can write (6.1) in the form of the usual linear model
f3 = YT?1 - a E2 XI (yT?1?l + YT?1l-) , (6.10)
zt = f3Xt + at (6.3) 1=1

This content downloaded from

138.38.44.178 on Fri, 14 Jun 2024 13:23:02 +00:00
All use subject to https://about.jstor.org/terms
 76 Journal of the American Statistical Association, March 1975

where Xi = - 2ri/ro. Also, since q(B) is assumed di- interventions after 1966 and discuss the step change
visible by (1 - B), rO + 2E,1l ri =0 , and hence
(fB/(1 - B))Pt(T) = wo,i I T = December 1959
1 -=1 x\1 = 1.
As an example, consider the integrated moving average in the level of the series due to the intervening events
model of order one for the noise term N, for which around that time. In this case, the noise model is such
that
q'(B) = 1-B and 0(B) = 1-B . (6.11)
(B) = (1 - B12)
Since
and

R (B)R(F)= (1 -B) (1 -F) 0(B) = (1 - 1B)(1 -02B12)

(1- OB)(1- OF)
Thus,

= (1 + O)-1 * [2 - (1- 0) E O1-'(B1 + Fl)]

1=1
(E Bj)(E Fj)
we find that R(B)R(F) j== ;=o
Al = (1 - 0)0-1 (6.12) (1 - 61B)(1 -02B12)(1 -06F)(1 - 02F12)
00 00

Thus, A represents a comparison between YT+i and the

=(E 7rjBi)(E rjFj) (6.17)
mean of two exponentially weighted averages, one of the
observations before time (T + 1) and the other after, so that from (A.10), j=O j=O
00
with the magnitude of the weights (1 - 0)01-1 mono-
ri E 7rj7rj+l
tonically decreasing as 1 increases. j=o
This formulation is applicable to situations where the
response to the pulse input is expected to be short-lived, The -rj can be obtained from the relationship
e.g., the effect on the demand for electricity during a 00 11

sudden heat wave in the summer or the sale of beer in (1 - 61B)(1 6-02B12) E rjBi = E B.
j=O j_o
Wisconsin should the Packers win the Super Bowl.
Essentially, we are comparing the observation YT+1 with By writing urj = 12n + m, we
the neighboring ones to determine if YT+1 is an "aberrant"
or "outlying" observation. The results in (6.10) and lrl2n+m - (1 - 01) 1( - 02)-l[(01)m+l (1 -

(6.12) are appealing since, in forming the comparison, - ( 1 - 02) (02) } n+ (4 - 02) (02)n]

more weight is given to observations close to the interven- m = O, ., 11; n = O, ,oo (6.18)
ing event and less and less weight to observations remote
where c = (01)12.
from the time of the event.
From (6.18) and after some algebraic reduction, we
Case (ii). c(B)/6(B) = B/(1 - B) (6.13) obtain, on setting 1 = 12k + s,

Here, the response to the pulse P,(T) is a step change rl2k+s (1 - 1)-2(1 - (02)2))1
in the level of the observations measured by A. Thus
*[12 -s(1 -02) + l (16;):
Q(B) = (1 - B)R(B)F (6.14)
and, from (6.6), (6.8) and (A.11), we have that *( _ - ] ( 2)k + (1 _ 01)-2

= (ro)-'R(B)R(F)(1 -B)YT+l 1 - 002 ?>02

00 00
* ( - 02)-1(1 ( - (6i)2)-1
= alYT+1+1 - E alYT-1 (6.15)
1=0 1=0 - (14_+)2(61)8+14k ,(6.19)
s = O,*, 11; k =O, *,oo
where al = (ro)-1(ri - rl+i) so
The quantity A is,
The resulting therefore,
weight function for the Los Angeles data
weighted averages, one of observations before the inter- is shown in Figure A above the observations.
vening pulse P,(T) and the other afterward, where the
weights are symmetrical. 6.2 The General "Linear" Dynamic Model
As a first example, consider again the integrated
The result in (6.5) can be readily extended to the
moving average model in (6.11). We find
case of more than one parameter. In the general dynamic
model with k inputs in (4.1), letting
00 00

: = (1 - 0) E O1YT+1+l - (1 - 0) E O1YT-1 (6.16)

1=0 1=0
(qp(B)/6(B))(wj(B)/6j(B)) = IjRj(B) (6.20)
as obtained in [2]. we can write
As a second example, we return to the model in (5.2)
for the monthly averages of ozone in downtown Los
Q(B)yt = E fl,Rj(B)itj + at, t = 1, * **, n (6.21)
Angeles. For illustration, we shall ignore the effect of j=1

This content downloaded from

138.38.44.178 on Fri, 14 Jun 2024 13:23:02 +00:00
All use subject to https://about.jstor.org/terms
 Intervention Analysis with Applications 77

where, as before in (6.1), Q(B) = 4p(B)/0(B). Assuming Some Properties of 3 and 02*
that all the coefficients in Q(B) and Rj(B) are known and
(i) Both bi and b2 are linear functions of the observa-
these k + 1 power series converge for B = 1, the
tions yt. By setting B = F = 1, the sum of the co-
model is then linear in the k parameters 5 = (f1, , Ok)'.
efficients associated with yt is zero for both of these
It readily follows that, for large n, the maximum likeli-
functions. Thus, A and A2 are linear contrasts in yt.
hood estimator satisfies the normal equations
(ii) The estimator A2 can be expressed in the form
A = b (6.22) 00 00

32 =E allYT+l+l - E a2lYT-l
where A is a k X k matrix and b a k X 1 vector such that
where ?=0 1=0 (6.25)
00 00

A = [a]ij , aij = Ri(F)Rj(B)C j j(O)

E al= E a2j = 1
b = (bl, .., bk)' 1=0 1=0

with i.e., a contrast between two weighted averages, one of

observations on or before the pulse input and the other
bj = Ri(F)Q(B)Cejy(O); i, j = 1, , Ik
afterward. To see this, since A2 is a linear contrast, it
In what follows, we investigate the special case having
suffices to show that El%o ali = 1.
two parameters, From the expression for b2 in (6.24), letting

yt = {J1O-(B)B + 02(1 - B)-'B}Pt(T) G(B) = R (B)R (F) , H (B) = 1 - B

+ (6(B)/qn(B))at . (6.23) and
00

In this model, d317q(B)BPt(T), where r(B) is assumed b2= E2 dlYT+1-1

to converge for |Bj = 1, measures the transient effect, l=-Co

and 132 represents the eventual change in the level of the it follows from (A.11) that EO= _ di = a22.
observations induced by the pulse input Pt(T) (see Further, making use of (A.12) and (A.13), we see that
Figure B(e) for the special case r1(B) = (1 -B)-1). a12 in (6.24) is also the coefficient of BO in R(B)R(F)
When fl = 0, the model reduces to that considered in * (1 - F)r(F). If we now set
(6.13). It is, therefore, of particular interest to know to
what extent the nature and precision of the estimator of G1(B) = R(B)R(F)(1 - F)>(F) , H1(B) = 1 - B
12 is affected by the presence of 01. We again suppose and
that the noise term is nonstationary so that s (B) is
b= E
l=00 lyrll
divisible by (1 - B).
To facilitate comparison with the model (6.13) we
again define a quantity R(B) such that we then have E0= _ di* = a12. The desired result
follows since
Q(B) = (1 - B)R(B)F, 00 0 0

so that in (6.22) E, ali A A-{anl E di - a2 E, di*} 1.

1=0 l=-oo 1=-oo

R1 (B) = Q (B) >i(B)B = R (B) >q (B) (1 -B) This property is similar to that of A in (6.15) for the
and model (6.13), except that the weight functions are no
R2(B) = R (B) longer symmetrical. From least squares theory, we have

It follows that, provided IA # 0, 32 = - (al2/jAj)(bl - a12) (6.26)

and the second term on the right side measures the effect
A1 = jA-1j{a22bj - al2b2} (6.24)
of the presence of the term /31q(B)BPt(T) in the model.
32 =I A|-1{aijb2 -al2bl (iii) One would expect that addition of the parameter
where /1 to the model would reduce the precision with which
02 could be estimated. A useful measure of the loss of
= AI alla22 - (a12)2- information is the variance ratio Var (p2)/Var ( A) where
b = R (B)R (F) (1 -F) F(F) (1 - B)YT+? it is understood that the denominator corresponds to the
model in (6.13). Now
b2= R(B)R(F)(1 -B)YT+l ,
Var (A2)/Var ( A) = (1 p2)-l
and aii, a12 and a22 are, respectively, the coefficients of
where (6.27)
BO in the power series
p = al2/(alla22)2

R(B)R(F>ri(B)r,(F)(1 -B) (1 -F) We illustrate these results in terms of a specific

R (B)R (F)r7 (B) (1 -B) example. Consider the case of (6.23) in which

R (B)R (F) rB()= (1 - 6B)1, (p(B) = 1 -B and 0(B) = 1 - B.

This content downloaded from

138.38.44.178 on Fri, 14 Jun 2024 13:23:02 +00:00
All use subject to https://about.jstor.org/terms
 78 Journal of the American Statistical Association, March 1975

We find and the noise parts of the model have tended to be

ignored. The application of time series methods can
A32 A (1-0)(1 + 6) X
: 2 : ( E [(I - 6 ) amend this situation. This is illustrated in this article in
the particular case where the object is to study the
- (1 - 0)01]YT+1+z, (6.28) possible effect of interventions in the presence of de-
pendent noise structure.
where A is given in (6.16). In this case only the
weights associated with the observations after the inter-
APPENDIX
vening pulse Pt (T) are affected by the presence of
-1(l -B)-1BPt(T) in the model. The weight function We here state sorne useful results in the sumrnation of series.
is shown in Figure D for 0 = .5 and 6 = .25.
Lemma 1: Let lVk Io' be a sequence of numbers and let {xt I-.
and Jyt _}' be two sequences of numbers such that xt-=yt0--
for t < 0. If one of the following three double sums is absolutely
D. Comparison of Weights Associated with YT +1+ convergent,
42 and for (o= .05, 8=.25, I= 0, 1, 2, ...) 00 OD 00 X0

Si = E E XtVkYt-k , S2 E E YuVkXu+k, (A.1)

IMPULSE t=1 k=0 u-i k=O
RESPONSE
o: oo

S3 = E: E VkYuXu+k,
k=O ui1

weights for, the other two are absolutely convergent and

- ----- weights for SI = S2 = S3 .

Proof of the lemma can be found in any standard text on infinite

series.
It is convenient to express Si, S2 and S3 in terms of the backshift
operator B and its reciprocal, the forward shift operator F = B-1.
Letting

0' I 2 3 4 5 6 7 LAG 00 00

V(B) - E VkBk and V(F) = E vkFk (A.2)

k - kcO

we can then write

Also, for this model the variance ratio is
00 00

SI E xtV(B)yt and S2 = L ytV(F)xt . (A.3)

V = Var (A 2)/Var (4) t-i t-i

= 1 + ((1- )(1 + 6)/(1 + )(1 - a)) . (6.29) Further, suppose we define

The value of this ratio for various values of 0 and 6 is 00 00

CZV(k) = E YtXt-k, C,z(k) =E XtYt-k , k 0, 41, ?2,

shown in the following tabulation: t-i t-i

a so that

O -.6 -.26 0 .26 .6 C.y (k) = Cy.(-k) * (A.4)

-.5 2.00 2.80 4.00 6.00 10.00 The quantity S3 in (6.1) can be expressed as
-.25 1.56 2.00 2.67 3.78 6.00
0 1.33 1.60 2.00 2.67 4.00 S3 = E VkCz1y(-k)
k -0
.25 1.20 1.36 1.60 2.00 2.80
.5 1.11 1.20 1.33 1.56 2.00 and, by letting C,( (-k) = BkCzv (0), we have

Thus, the presence of f1 in S3 = V(B)Crv(O)

the model *can (A.5) cau
increases in the variance of 42, compared with 4, when 0 It follows that when the conditions of
is negative and a is positive.
L_ xtV(B)yt = L ytV(F)xe = V(B)Czv(0) * (A.6)
tl1 t=l
7. CONCLUDING REMARKS
This result can be readily
In the past, much attention has been given to sta-
tistical analysis linking phenomena which are coinci- Lemma 2: Suppose W(B) = Vl(B) + V2(F) where V1(B) and
dental in time. In practice, it is perhaps more often the V2(F) are two power series in B and F, respectively, such that the
sum _tI l xtW(B)yt is absolu-tely convergent. Then
case that a response at a given point of time depends on
events, both known and unknown, which have occurred
00

EI xW (B)yT = W (B)CI (?) . (A.7)

not necessarily coincidentally but over the recent past. t-i

Statistical methods have, in a word, "lacked memory." Lemma 3: Let GJ(B) = 2? q, gBi and H(B) = ,^??_o0 hk?Bk
The dynamic characteristics of both the transfer function be two power series in B and converge for lB l 1, and let D (B)

This content downloaded from

138.38.44.178 on Fri, 14 Jun 2024 13:23:02 +00:00
All use subject to https://about.jstor.org/terms
 Intervention Analysis with Applications 79
= G(B)H(B). Then
(v) if g, = gjandhi = 0 j > 1, then
0

D(B) = I d1B1 (A.8)

do = E hjgj . (A. 13)

where di = x: gjhl11 [Received October 1973. Revised August

In particular

(i) if gj = g9j and hk = h-k, then REFERENCES

[1] Box, G.E.P. and Jenkins, G.M., Time Series

u0 u casting and Control, San Francisco: Holden-Day
(A.9)
[2] - and Tiao, G.C., "A Change in Level of a N
(ii) if gj = 0, j < -1 and H(B) = G(F), then Time Series," Biometrika, 52 (June 1965); 181-9
[3] - and Tiao, G.C., Bayesian Inference in Stat
Reading, Mass.: Addison-Wesley Publishing Co
di = d-i = - g j9j+1, I = O,*, oo (A.10)
ijo [4) Feige, E. and Pearce, D.K., "Inflation and In
(iii) if H(B) = 1 - B, then Application of Time Series Models," Report
Systerns Research Institute, University of Wisc
di = gi gi-, 1 0, 4-1, *-4-, vo (A. 11) 1973.

so that x-:*=I di = -go and ?=_-o di = go; [5] Glass, G.V., "Estimating the Effects of Int
Nonstationary Time Series," American Educat
(iv) if gi = g- and hj = Oj- <-1, then Journal, 9, No. 3 (1972), 463-77.
[6] Tiao, G.C., Box, G.E.P. and Hamming, W.J., "Analysis of Los
do= F hjgj; (A.12) Angeles Photochemical Smog Data: A Statistical Overview,"
Journal of Air Pollution Control Association, 25 (March 1975).

This content downloaded from

138.38.44.178 on Fri, 14 Jun 2024 13:23:02 +00:00
All use subject to https://about.jstor.org/terms