Journal of the Air Pollution Control Association

ISSN: 0002-2470 (Print) (Online) Journal homepage: http://www.tandfonline.com/loi/uawm16

Analysis of Los Angeles Photochemical Smog Data:

A Statistical Overview

G. C. Tiao , G. E. P. Box & W. J. Hamming

To cite this article: G. C. Tiao , G. E. P. Box & W. J. Hamming (1975) Analysis of Los Angeles
Photochemical Smog Data: A Statistical Overview, Journal of the Air Pollution Control Association,
25:3, 260-268, DOI: 10.1080/00022470.1975.10470082

To link to this article: http://dx.doi.org/10.1080/00022470.1975.10470082

Published online: 13 Mar 2012.

Submit your article to this journal

Article views: 1946

View related articles

Citing articles: 7 View citing articles

Full Terms & Conditions of access and use can be found at

http://www.tandfonline.com/action/journalInformation?journalCode=uawm16

Download by: [181.232.23.178] Date: 02 March 2017, At: 21:15

 Analysis of Los Angeles Photochemical Smog
Data: A Statistical Overview

G. C. Tiao and G. E. P. Box

The University of Wisconsin, Madison, Wisconsin

W. J. Hamming
Los Angeles County Air Pollution Control District, Los Angeles, California

A research project has been under way to investigate air pollution A research project has been underway to perform statistical
problems in Los Angeles County with the help of the data supplied analysis of aerometric data from January 1955 to December
by the Los Angeles County Air Pollution Control District. These 1972 assembled by the Los Angeles Air Pollution Control
District. The principal pollutant source data are hourly
data consist of measurements of primary pollutants such as nitric
readings on such primary contaminants as nitric oxides,
oxide, hydrocarbons, carbon monoxide, sulfur dioxide and particu- hydrocarbons, carbon monoxide, sulfur dioxide and partic-
lates, and secondary pollutants such as ozone and nitrogen diox- ulates, and secondary pollutants such as ozone* and nitro-
ide, recorded hourly at a number of different stations in Los An- gen dioxide, recorded at seven appropriately distributed
geles County over the past seventeen years. This present discus- locations in the Los Angeles Basin. They are: Downtown
Los Angeles, West Los Angeles, Burbank, Pasadena, Azusa,
sion deals in a preliminary way with a particular aspect of this
Long Beach, and Lennox. Atmospheric variables such as
analysis, namely, the occurrence of photochemical smog in Los mixing height, wind speed and direction, and base inver-
Angeles. The paper is divided Into two main sections. The first is sion height are considered. Additional exogenous factors,
intended to provide a brief survey of the problem of photochemical such as the chronology of control measures, vehicle popula-
smog in Los Angeles as presently understood in relation to the tion, and traffic patterns will also be included in the analy-
sis.
available field data and also in relation to chamber experiments
The eventual objective of this study is to utilize informa-
which have been run in various laboratories. The second part of
tion from the data to build mathematical models which will
the paper discusses a class of intervention problems that arise in adequately represent the chronological and spatial move-
studying the data. It is noted that parallel problems occur in the ments of the pollutants. It will then be possible to produce
study of other ecological material and elsewhere. Statistical meth- efficient forecasts of current and future trends of the pollu-
ods for dealing with this class of problems are illustrated with tants and to assess the effectiveness of control measures.
This paper is the first of a series reporting on our re-
some of the Los Angeles data.
search findings. It deals in a preliminary way with only one
particular aspect of the problem, namely, the movement of
photochemical smog in Log Angeles. The paper is divided
into two main sections. The first is intended to provide a
brief survey of the problem of photochemical smog as pres-
Dr. Tiao is Professor and Chairman of the Department of ently understood. Using O3 as the basic indicator of smog,
Statistics and Dr. Box is R. A. Fisher Professor of Statistics, various graphical devices are employed to illustrate the ef-
University of WisconsinMadison, 1210 West Dayton fect of atmospheric variables and the historical develop-
Street, Madison, WI 53706. Mr. Hamming has recently re- ment of the pollutant over the seven locations. In particu-
tired from the Los Angeles County Air Pollution Control lar, it will be seen that considerable differences exist in the
District. This is the revised version of Paper No. 73-79, pre-
sented at the 66th Annual Meeting of APCA at Chicago in
1973.
' Strictly, other oxidants as weir as O3 are measured by the chemical test.

260 Journal of the Air Pollution Control Association

historical variations of O3 over these locations. In places umn represents the average behavior of O3 during the day
such as Downtown Los Angeles and West Los Angeles, the for a particular month, and each row indicates the chrono-
level of concentration was lower in the 1960's compared logical variation of the pollutant concentration for a partic-
with the 1950's and there seems to have been some progres- ular hour (Pacific Standard Time) of the day. To facilitate
sive improvement over the last several years. Correspond- comparison, the entries are classified into four ranges by
ing changes are, however, not evident in the more easterly shading as labelled. The solid line in the figure indicates
regions such as Pasadena and Azusa. the time of occurrence of the peak level of O3.
The second part of the paper discusses a number of sta- By reading Figure 1 vertically, we see that in Downtown
tistical problems that arise in studying the data. As pollu- Los Angeles the level of O3 begins to rise about 7 A.M.,
tant data are expected to be correlated through time, ap- peaks around the noon hour and dissipates in the after-
propriate time series models which can adequately repre- noon. Most of the accumulation of O3 during the daytime
sent the data and allow for projection of current and future hours is the result of photochemical reaction of contami-
trends are considered. A new and widely applicable tech- nants from automobile emissions and other sources.
nique, called "intervention analysis," has been further de-
veloped and is here illustrated in assessing the effect of ex-
ogenous factors such as control legislation, opening of new The Seasonal Effect
freeways and other environmental changes.
From June to October which we shall loosely call "the
summer," Los Angeles County provides an ideal natural re-
The Photochemical Smog Problem in Los Angeles actor in which these chemical changes can take place. Dur-
ing this period occur (i) strong and persistent night and
Los Angeles has what may be called general and special daytime inversions; (ii) weak and, for several hours each
air pollution problems. The nature of these problems has day, stagnant winds; (iii) bright sunshine.
been reasonably well understood for some time.1"3 The Because of (i) and (ii) the primary pollutants from heavy
general pollution problem, which is currently found to morning automobile traffic and other sources are not dis-
some extent in any large city, is attributable directly to the persed and occur in rather high concentrations while (iii)
properties of the primary contaminants themselves. For ex- allows the photochemical reaction to proceed. During the
ample, NOX, SO2, and particulates are of themselves unde- remainder of the year November-May, which we shall call
sirable and can cause lung damage if inhaled for sufficient- "the winter," the inversion, although initially lower, usually
ly long periods in high concentrations. rises higher by noon each day than in the summer. Also, the
sunlight is of shorter duration and is less intense. In the
The Product of Photochemical Smog winter, therefore, there is much less time and less available
energy for the chemical reactions to take place and the
The special pollution problem of Los Angeles, and the problem is less serious.
most serious one, comes not from the primary pollutants The seasonal difference between the summer and the
themselves. It comes from substances produced by chemi- winter can be seen by studying Figure 1 horizontally which
cal reactions among certain of these primary pollutants. shows that the daytime concentration of O3 is much lower
The products of these chemical reactions are responsible in the winter time than in the summer. Also, the peak hour
for the famous Los Angeles Smog. They include eye and in DOL A shifts from 10 to 11 A.M. in the summer to 1-2
lung irritants, and other agents, which in high enough con- P.M. in the winter.
centrations can produce damage to living things both ani-
mal and vegetable. In addition, these products are also as- The figure also seems to suggest that the level of concen-
sociated with visibility-reducing haze and high levels of tration in Downtown Los Angeles was lower in the 1960's
ozone. The principal contaminants taking part in the reac- compared with the 1950's and that there has been some
tion are: (1) oxides of nitrogen, denoted by NOX, whose progressive improvement over the last few years. This will
main constituent is nitric oxide, denoted by NO, and (2) re- be discussed in further detail later in the paper.
active hydrocarbons, denoted by HC. A measured product
which is indicative of the degree of photochemical pollution Effect of Inversion Mixing Height and Wind Speed
is ozone, denoted by O3.
The chemistry is very complicated and is still not com- It is of interest to know to what extent changes in atmo-
pletely understood, but very roughly we can say that under spheric conditions within a season may affect the concen-
certain very special conditions NOX and HC take part in a tration level of the pollutant. The maximum mixing height
photochemical reaction which eventually forms sufficient affects the volume available for diffusion of the pollutants,
O3 and other associated substances to produce undesirable while the wind speed affects their dilution and transport.
effects. Thus, in several steps, This is illustrated in Table I showing means of daily av-
erages of O3 for various wind speeds and for various maxi-
sunlight
+ NO + HC O3; and other products mum mixing heights in Downtown Los Angeles during the
four summers (1967-1970). As one would expect, on the
The very special conditions needed are: (1) that the main days on which the mixing height and wind speed are lower,
reactants (NO and HC) be present; and (2) that sunlight be the concentration of O3 is increased.
present of sufficient intensity and duration to initiate and
sustain the reaction in the air for several hours. The reac-
tion is rather slow and it is typically three to five hours
after its initiation before significant quantities of O3 build Table I. Means of daily averages of O3 (pphm) for various wind
up. speeds and maximum mixing heights in Downtown Los Angeles:
June-October 1967-1970.

Daily Movement of Ozone Maximum Mixing Ht. (feet)

Wind Speed
The behavior of O3 during the day is illustrated in Figure (MPH)6-12AM 0-2000 2100-2500 2600-3500 3600 and up
1. The chart is based on a two way table of O3 measured in 0 -- 4 5.7 5.6 4.3 3 .8
Downtown Los Angeles(DOLA). The entries are the 4.1 -- 5 4.8 4.5 4.1 3 .1
monthly averages (in pphm, parts per hundred million) of 5.1 i& up 4.3 3.9 3.0 2 .2
each hour of the day from Jan. 1955 to Dec. 1972. Each col-

March 1975 Volume 25, No. 3 261

8p.m.~

4 p.m.

Noon

8 a.m. -

4 a.m

1955 1956 1957 1958 1959 1960 1961 1962 1963 1964

1 10-14
15-19
20&up

8 p.m.

4 p.m.

Noon

8 a.m.

4 a.m.

1964 1965 1966 1967 1968 1969 1970 1971 1972 1973

Figure 1. Monthly average of hourly readings of O3 (pphm) at DOLA 1955-1972,

262 Journal of the Air Pollution Control Association

 Geographical Location

In the above, we have confined our discussion to one

locationDowntown Los Angeles. Charts and tables simi-
lar to Figure 1 and Table I have been developed for the
other six locations.4
Preliminary study indicates that appreciable differences
exist in the level of O3 concentration over these stations.
Some clue as to why this is so may be found in Figure 2
which shows the geographical location of the seven record-
ing stations. At each location, the first two numbers are, re-
spectively, the gross seventeen-year (1956-1972) summer
averages for O3 and the primary pollutant NO, and the
third number is the average vehicle miles traveled per acre
within a three-mile radius of the air monitoring station, es- Figure 2. Geographical location of recording stations, with
timated from actual counts collected over the last three or the gross seventeen-year (1956-1972) summer averages for
four years. O3 and NO, and the estimated vehicle miles traveled per acre.
The arrows in the figure roughly indicate the direction of Azusa (A), Burbank (B), Downtown Los Angeles (DOLA), Len-
nox (L), Long Beach (LB), Pasadena (P), West Los Angeles
the prevailing winds and show the "pipe reactor effect." (WLA).
Thus, although the heaviest concentration of traffic is in
the region around Downtown Los Angeles and southwest of
it, the highest concentration of O3 is at Azusa at the end of
the "reactor pipe line." This is so even though Azusa itself
has a traffic density which is much lower than that in
Downtown Los Angeles. Similarly the concentration of NO Intervention Problems
is in general reduced as we move through the pipe line.
This may be partly because it is used up in the reaction and A class of problems likely to be important in many dif-
partly because the local emissions are in general less in the ferent kinds of ecological study concerns the question "Did
Pasadena and Azusa region. a change of a certain kind made on a certain date make any
difference?" More specifically "Did it produce the kind of
An Overall Look at the Past and Present
difference to be expected?"
For example, in 1960 the Golden State Freeway which
A natural question to ask is "Are things getting better, may have had an effect on the traffic pattern in Downtown
worse, or staying about the same?" Some preliminary an- Los Angeles was opened. Also, at about this time Rule 63
swer may be found (Figure 3) in the relative frequency di- was introduced which reduced the proportion of reactive
agrams of daily maximum hourly concentrations of O3 in hydrocarbons in the gasoline sold in Los Angeles County.
four ranges for the seven locations over the last eighteen Thus about 1960 two events occurred either or both of
years. The range 8 pphm or below is of particular impor- which, could have, produced a change in level of ozone con-
tance since it is the Federal Air Quality Standard for oxi- centration in Downtown Los Angeles. Is there evidence
dant. that such a change did occur at this time? Figures 1 and 3
The diagrams illustrate how the situation has developed seem to indicate that it did. However, appearances can be
differently at different locations. Thus, from 1966 to 1972 deceiving and more sophisticated statistical methods are
the proportional frequency with which the Federal Air needed in considering questions of this kind.
Quality Standard was met, increased in Downtown Los An- Available procedures, such as Student's t test for esti-
geles from 0.35 to 0.62 and in West Los Angeles from 0.51 mating and testing for a change in level, have played an im-
to 0.82. However, in Pasadena and Azusa only minor portant part in Statistics for a very long time. Unfortunate-
changes occurred from 1966 to 1970, although an apparent ly, however, the employment of most such procedures re-
improvement of about 0.10 did occur in 1971 for both loca- quires assumptions that are invalid in the present context.
tions. It should be noted that the sudden change in fre- The ordinary t test for comparison of two means (i) would
quency occurring at Long Beach in 1963 is not necessarily be valid if the observations before and after the event of in-
an indication of improvement, since there was a change in terest varied about means HI and /*2, with errors normally
the location of the monitoring station in that year. and independently distributed with constant variance, or
(ii) would be approximately valid if randomization could be
Some Statistical Problems introduced into the conduct of the data gathering process.
In fact, as we see for the monthly averages of O3 in Down-
What are some of the problems that must be considered town Los Angeles in Figure 4 (obtained by averaging the
in the analysis of data of this kind? Of course, a major entries of each of the columns of Figure 1 from January
problem is that of coping with so many numbers. There are 1955 to December 1965), our data are time series in which
in these records collected over 18 years more than twenty not only are successive observations autocorrelated but also
million separate pieces of data. Considerable effort has had strong seasonal effects occur. Thus the ordinary statistical
to be expended in the organization of the data and numer- procedures (parametric or non-parametric) which rely on
ous problems have been encountered, which we will not independence or special symmetry in the distribution func-
discuss here. At a more technical level analysis of these tion are not available to us nor are the blessings endowed
data has raised some interesting problems in statistical by randomization.
methodology. In the remaining part of this paper, we dis- One way to proceed which we initiated earlier5 is to build
cuss one such problem which has resulted in new proce- a stochastic model for the data which takes cognizance of a
dures for time series intervention analysis. possible change of the form hypothesized.

March 1975 Volume 25, No. 3 263

 AZUSA LONG BEACH

70
68
66

64

62
/m/m/mm 60
58 '/mmm/mm. 58
56 '/mm

.8 .6 .4 .6 .4 .2
PASADENA LENNOX

64

58 58

56 56

.6 ,4 .8 .6 .4 .2
BURBANK DOWNTOWN LOS ANGELES

.2 .6 .4 .2
WEST LOS ANGELES

PPHM
18 & below
|9-19
20-49
50 & up

60

58

56
Figure 3. " Relative frequency of daily maxi-
mum hourly readings of O 3 at various levels
for seven locations (1956-1972). .6 .4

264 Journal of the Air Pollution Control Association

 WEIGHTS Jan. 60-

illinium.
""""""IIIIIIIIIH 1

Dec. 59 Intervening
9.0 events
occur
8.0

7.0

6.0

5.0

4.0

3.0

2.0

1.0

55 56 57 58 59 60 61 62 63 64 65 66
Figure 4. Monthly averages of O3 in Downtown Los Angeles (January 1955 to December 1965) and weight func-
tion for determining the change in 1960.

Time Series Models where the zeroes of <t>(B) all lie outside the unit circle. This
corresponds to the use of a stationary model in the dth dif-
Let us denote the time series obtained at equal intervals ference. Thus if we let
of time by . . . yt-i,yt,yt+it and use B for the back shift
operator such that Bzt = zt-\. Also let . . . at-\,at,at+\, ... (1 - B)% = wt (4)
be a sequence of independently distributed normal vari-
ables having mean zero and variance <ra2 which for brevity the model may be written
we refer to as "white" noise. Then a useful model6 for such
a series may be written 4>(B)wt = 9(B)at, (5)
a stationary autoregressive moving average process.
Y; = f (K, 0 + (1) In representing seasonal models with period s (for
monthly data s = 12) it is often convenient to write <p(B) =
<Pi(B) (&&*) and 6(B) = di(B) 62(BS). Also, non-stationarity
where Yt = F(yt) is some appropriate transformation of yt can frequently be eliminated by seasonal differencing so
(say log yt or yt1^2 or perhaps yt itself), f(x, t) is some func- that we come finally to seasonal models of the form
tion of time containing parameters K,
- B)d(l-B")Dzt = (6)
6(B) = \-BxB- 02B2.. .-eQBq,
cp(B) = l-cpiB-(p2B2...-<ppBp where the polynomials <f>i(B), 02(5S), 0\(B), d2(BB) are of de-
grees pi, P2, q\, 92, respectively.
The process of model building is necessarily iterative
are "moving average" and "autoregressive" polynomials in and, as discussed for example in Ref. 6, entails the succes-
B, and we shall require that the zeroes of 6(B) lie outside sive use of Identification, Fitting, and Diagnostic Checking
and those of <p(B) lie on or outside the unit circle. Then if zt in establishing the model. Applied in the present context
= Yt f(x,) the model may be written our strategy is as follows: (i) first frame a model for change,
which describes what is expected to occur given knowledge
(p(B)zt = 9(B)at. (2) of the actions which are known to have been taken; (ii)
work out the appropriate data analysis based on that
For the representation of certain kinds of stable non-sta- model; (iii) if diagnostic checks show no inadequacy in the
tionary series the operator (p(B) can be factored so that model form, conclusions as to the effects of the change may
be drawn directly from the analysis; but (iv) if the model
(3) appears to be inadequate, then one tries to learn how and

March 1975 Volume 25, No. 3 265

why and to check up on new possibilities which suggest as illustrated in Figure 5(a).
themselves to those having understanding of the problem. Some events might not be expected to produce an imme-
Thus we have a strategy for learning which we shall illus- diate response but rather a "first order" dynamic response
trate with a few examples. like that in Figure 5(b) for which the transfer function is
wo/(l-5iB). With this model it is readily shown that the
A Model for Ozone
time constant of the system is estimated by T = {
loge^i}"1 and the steady state gain is a?o/(l-5i).
For monthly ozone averages denoted below by zt, prelim- It is to be noted that the model for Eq. (8) has great flex-
inary identification studies show that, apart from a possi- ibility, since need not merely be an indicator variable. It
ble deterministic component to be discussed later, a model could be any exogenous variable such as atmospheric con-
of the simple form ditions, traffic patterns, vehicle population, etc. whose ef-
fect on the response needed to be modelled dynamically.
Given this model the likelihood function associated with
- Bi2)zt = (1 - 0 t B)(l - B2Bn)at (7) any given set of data can be written down. This can be used
to produce maximum likelihood estimates of the parame-
seems adequate to describe the stochastic time-dependent ters of both the stochastic and dynamic parts of the model
relationship of the data for all the seven locations. In par- and the approximate standard errors. Alternatively we may
ticular, the factors 1-B12 and I-62B12 jointly take account of adopt a Bayesian viewpoint and analyze the estimation re-
the pronounced but somewhat drifting seasonal pattern sults in terms of posterior distributions of the parameters
shown for example in Figure 4 for Downtown Los Angeles. on the basis of noninformative prior distributions.8
The nature of the two factors can best be understood as fol-
lows. Suppose 6\ were zero, then the forecast of, say, next Some Examples of Time Series and Intervention Analysis
January would be a weighted average of the data on all past Applied to the Pollution Data
Januarys, with weights (l-02), 02(l-02), h20--h), de-
creasing exponentially into the remote past. The factor We now apply the above time series methods and inter-
l-0iB represents the "non-seasonal" relationship between vention analysis techniques to three selected but important
observations in successive months. Using the model, fore- cases.
casts follow a stable seasonal pattern, which is appropriate-
ly updated as new data become available. The model in Eq. Possible Change of O 3 Level in Downtown Los Angeles in 1960
(7) is of a similar nature to those obtained in Ref. 7 where,
however, biweekly averages of daily maximum hourly con- For this study we employ the series of monthly averages
centrations were considered. of O3 in Downtown Los Angeles from January 1955 to De-
cember 1965, denoted by zt, shown earlier in Figure 4.
A General Model for Intervention Analysis
Combining the model in Eq. (7) together with the transfer
function in Eq. (9), we have
In the above, we have argued that variation in, say the
monthly result can be represented by a time series model
which is a stochastic difference equation. That is, a differ- - 62Bi2)
ence equation (p(B)zt = 0(B)at driven by "white" noise at. - 1 _ R12 f (10)
The effects of environmental changes are often dynamic in
character with transfer functions which may also be repre- where
sented by a suitable difference equation model. * i1,0, beginning
prior to January, 1960
January 1960.
Let & be a variable (which, for example, could take the
value zero before 1960 and unity after that time thus repre-
senting a step change) and consider the model
(B)

0(B)
(8)
(a)
r
In general, 8(B) = l-5aB- . .. -5rB and a(B) = <a0 -
. . . -cosBs are polynomials in B of degrees r and s, respec-
tively. We shall normally require that co(B) has zeroes out-
side and 5(B) zeroes outside or on the unit circle. Figure 5
shows the output response to a step change for various
simple transfer functions.
For the Freeway and/or Rule 63 change of 1960, an ap-
propriate model will then be of the form of (8) with 1-B

, , . , . , ui * ( 0 , p r i o r to 1960
the indicator variable , = < '*.. ., __ O O o O O

11, after 1960

and
Input ooooo-

the transfer function (9) Figure 5. Response to a step change for vari-
, , = u>0 ous transfer functions w(B)/5(B).

266 Journal of the Air Pollution Control Association

The maximum likelihood estimates (MLE) of the parame- sion differential, the net effect would be different in the
ters together with their estimated standard errors are given winter from that in the summer. To examine these possibil-
below. ities the following model was postulated for all the avail-
able monthly averages of O3 in Downtown Los Angeles
MLE S.D. from January 1955 to December 1972,
>o -1.10 0.10
i -0.24 0.07
2 0.54 0.06

Examination of residuals fails to show any obvious inade- (11)

quacies in the model so we proceed to discuss the findings. where
From a Bayesian viewpoint, we may interpret the result l, " s u m m e r " months June-October
concerning coo as saying that, on the basis that little is
known initially about these parameters, the posterior dis- beginning 1966
tribution of OJ0 is nearly a normal distribution centered at 0, otherwise
o>o = 1.10 with standard deviation 0.10. The evidence is
thus very strong that a reduction in the level of O3 in 1, " w i n t e r " months November-May
Downtown Los Angeles occurred in 1960, the best estimate kt" \ beginning 1966
of the reduction being 1.10 pphm.
It is illuminating to consider the form of the estimate o>o. 0, otherwise
Let the month January 1960, when the effect took place be The parameters X\ and X2 represent, respectively, for the
denoted by T+l. Suppose for the moment that the parame- summer and winter months, progressive changes in the lev-
ters di and 02 are known. Then it is shown9 that a>o can be els of the observations from one year to the next beginning
written as in 1966 and are, therefore, slopes of two deterministic year-
ly trends for these two "seasons." As before, the parameter
a>o is introduced to allow for the change in 1960.
Estimation results are as follows:
f2(z T,
ZT_2,
MLE S.D
Xi 0.25 .07
where /1 is a weighted average of the observations x2 -0.07 .06
ZT+2, after the event occurred and fi is a weighted aver- 030 -1.09 .13
age of the observations ZT, ZT-I, - - prior to the event. In 0i -0.24 .03
practice, 0\ and 02 would be unknown, but for large sam- 02 0.55 .04
ples, as in the present example, they can be approximated
by the corresponding maximum likelihood estimates. Using The estimate of Xi provides very strong evidence that
the values $1 = 0.24 and 82 = 0.54, the weights applied to there has been a steady improvement for the summer sea-
the observations zt in forming o>o are shown above the se- sons, the yearly reduction being estimated at 0.25 pphm.
ries in Figure 4. On the other hand, the improvement, if any, in the winter
The weight function gives large weight to observations in is very slight. In particular, the parameter value X2 = 0,
the years immediately before and after the event took place representing no improvement, is well inside the approxi-
and less and less weight to observations remote from it. mate 95% highest posterior density interval -0.07 0.12,
This is an intuitively pleasing result, as one would certainly and hence is not contradicted by the data.
expect that, for data of this kind, observations farther re-
moved from the time the event took place should be less A Useful General Indicator of Change
and less relevant in determining its effect.
A device which is sometimes useful, at least as a prelimi-
nary identification tool, is to compare the forecast made
A Test of the Progressive Reduction of O3 at Downtown Los from some point at which change is known to have been in-
Angeles since 1966
stituted with what has actually occurred. This can be done
This second example illustrates a situation where the graphically. But it is easy to deceive the eye and more for-
model for change must take a more elaborate form. From mal statistical procedures are needed. Suppose we have n
1966 to 1970 regulations for the reduction of hydrocarbon observations (zu ..., zj and it is suspected that changes
emissions resulted in engine design changes in new cars might have occurred in the pattern level after time T(T <
which had the effect of decreasing HC but sometimes at n). The nature of the changes may be studied by comparing
the expense of increasing NOX.3 Also, legislation since 1970 forecasts .Ml), M 2 ) , . . . , zT(n-T) of zT+u . . . , made at
progressively reduced the emissions of both HC and NO* in time T with the actual observations themselves, from
the new cars. It has been suggested that these changes which OT+I, . . . , an can be readily calculated.1" If in fact
could slow down or reduce the photochemical reactions and there were no change, the a^'s would be "white noise" and
hence could lead to a reduction in the level of O3 near a pri- hence
mary pollutant center such as Downtown Los Angeles. Vi-
sual inspection of Figures 1 and 3 indicate this may very
well be the case. In this example the change in O3 should
show, not as a step function, but as a trend reflecting an in-
Z
t =T
creasing proportion of new design vehicles in the car popu- t Notice that, for example, SIT+I is not the forecast error ZT+I 2T0) but may be comput-
lation. Furthermore, because of the summer-winter inver- ed from the complete set of forecast errors.

March 1975 Volume 25, No. 3 267

would be distributed as o-a2x2 with (n-T) degrees of free- PPHM
dom. This x 2 test could thus provide a useful general indi- 8
cator of various changes in pattern* from that expected.
For illustration we consider the recent ozone results for
Azusa. 6
Has there been improvement in Azusa since 1971? We
now employ monthly averages of O3 in Azusa to investigate 4
whether changes have actually occurred since 1971, as a re-
sult of new auto emissions standards introduced at that
time. First, the data from January 1956 through December 2
1970 was used to estimate the parameters in the basic
model (7). The results are 1
1 1 1 1 1 1 1 1 1
2 4 6 8 10 12 ' 2 4 6 8 10 12
1971 1972
MLE S.D. 2\-
-0.15 0.07
0.91 0.04
1.00
x Forecast
Based on the fitted model, forecasts of the 24 months of o Actual
1971 and 1972 made at time T, December 1970, are shown The errors a.
in Figure 6(a) together with the actual observations. The Figure 6. Comparison of forecasts with actual observa-
errors a t 's for these months are plotted against time in Fig- tions: monthly averages of O3 in Azusa January 1971-De-
ure 6(b). The sum of squares value cember 1972.

K2)
2\-l = 35
t = T
is larger than the x2 value with 24 degrees of freedom at the primary and the secondary pollutants, (ii) the phase rela-
10% level, suggesting that the hypothesis of no-change is tions between the locations (iii) the influence of exogenous
untenable. Indeed, Figures 6(a) and 6(b) suggest that (i) atmospheric variables and other factors such as vehicle
there has been improvements in these two years, (ii) the population and traffic density and (iv) the effect of control
improvement is much greater in 1972, and (iii) the magni- measures. New statistical tools in modelling multivariate
tude of the reduction is appreciably larger in the summer time series which are necessary to achieve these goals are
months. If desired, one can proceed to introduce dynamic currently being investigated.
models of the type in (11) to estimate the magnitude of the
reductions as well as to project the level of O3 for future Acknowledgments
years.
This research has been carried out under a joint project
between the Department of Statistics, the University of
Wisconsin, Madison and the Los Angeles County Air Pollu-
tion Control District, supported by a grant from the Ameri-
Conclusions and Further Work can Petroleum Institute. The authors wish to thank W. S.
Wei, J. Schaap, and S. Grover for computing assistance and
In the preceding sections, a preliminary statistical analy- for preparing the figures in this paper.
sis has been presented for the ozone data over the seven
locations in the Los Angeles Basin. The level of ozone con-
centration is subject to strong seasonal fluctuations and, References
within the summer season, to the influence of changes in 1. B. Dimitriades, "Effect of hydrocarbon and nitrogen oxides on
atmospheric conditions. In Downtown Los Angeles and photo-chemical smog formation, "Environ. Sci. Technoi, 6:
West Los Angeles, the level was lower in the 1960's com- 253 (1972).
pared with that in the 1950's and there have been progres- 2. A. J. Haagen-Smit, "Chemistry and physiology of Los Angeles
sive improvements beginning around 1966. Corresponding smog,"/nd Eng Chem, 44:1342 (1952).
3. W. J. Hamming and J. E. Dickinson, "Contrl of photochemical
improvements were, however, not evident at the end of the smog by alteration of initial reactant ratio," J. Air Poll. Con-
pipe line in Pasadena and Azusa until 1971-1972. trol Assoc, 16: 317 (1966).
Various graphical devices have been presented to bring 4. G. C. Tiao, G. E. P. Box, M. Grupe, S. T. Liu, S. Hillmer, W. S.
out important features of the data. We believe that such Wei, and W. J. Hamming, "Los Angeles Aerometric Ozone
Data 1955-1972." Technical Report #346, Department of Sta-
presentations provided a powerful means to promote dis- tistics, University of Wisconsin, Madison, Oct. 1973.
cussion about the nature of the pollution problem and thus 5. G. E. P. Box and G. C. Tiao, "A change in level of a non-sta-
play an indispensable role in the preliminary stage of tionary time series," Biometrika, 52:181 (1965).
model building. At the same time, analysis of the data gives 6. G. E. P. Box and G. M. Jenkins, Time Series Analysis, Fore-
casting and Control, Holden-Day, San Francisco 1970.
rise to the need for new statistical methods. In particular, 7. P. H. Merz, L. M. Painter, and P. R. Ryason, "Aerometric data
the intervention analysis technique has been developed analysistime series analysis and forecast and an atmospheric
and is illustrated using the ozone data. smog diagram," Atmos. Environ., 6: 319 (1972).
Using similar statistical tools, much preliminary work 8. G. E. P. Box. and G. C. Tiao, Bayesian Inference in Statistical
Analysis, Addison Wesley, Reading 1973.
has already been done on other pollutants including nitric 9. G. E. P. Box and G. C. Tiao, "Intervention Analysis with Ap-
oxide, carbon monoxide, sulphur dioxide and particulates, plications to Economic and Environmental Problems," Tech-
and the findings will be reported in the near future. Work nical Report #335, Department of Statistics, University of
has begun to explore (i) the interdependence between the Wisconsin, Madison (June 1973). (To appear in the Journal of
American Statistical Association.)
10. G. E. P. Box and G. M. Jenkins, "Models for Prediction and
Control VI. Diagnostic Checking," Technical Report #99, De-
* There is evidence10'11 that examination of the aj's is not sensitive to all kinds of model partment of Statistics, University of Wisconsin, Madison Dec,
change. Especially it tends to be insensitive to changes in the memory parameters of 1966.
6(B). 11. R. A. Johnson and M. Bagshaw, Private communication.

268 Journal of the Air Pollution Control Association