AMERICAN JOURNAL or EPIDEMIOLOGY VoL 131, No.

1
Copyright <© 1990 by The Johns Hopkins University School of Hygiene and Public Health Printed m U S A.
All nghtt reserved

MORTALITY AND AIR POLLUTION IN LONDON: A TIME SERIES

ANALYSIS
JOEL SCHWARTZ 1 AND ALLAN MARCUS 1

Schwartz, J. (Environmental Protection Agency, Washington, DC 20460), and

A. Marcus. Mortality and air pollution in London: a time series analysis. Am J
Epidemiol 1990; 131:185-94.
The relation between air pollution and mortality in London was examined for
the winters of 1958-1972. The data exhibited a high degree of autocorrelation,
requiring analyses using autoregressive models. There was a highly significant
relation between mortality and either particulate matter or sulfur dioxide (after
controlling for temperature and humidity), both overall and in each individual
year. Graphic analysis revealed a nonlinear relation with no threshold, and a
steeper exposure-response curve at lower air pollution levels. In models with
both pollutants, particulate matter remained a significant predictor with about a
10% reduction in its estimated coefficients, while sulfur dioxide was insignificant,
with a large drop in its estimated coefficient The authors conclude that particu-
lates are strongly associated with mortality rates in London, and the relation is
likely causal.

air pollution; longitudinal studies; mortality

Particulate matter in the air has long
been believed to be a health hazard. Simple
clinical observation during modern epi-
sodes of comparatively high particulate air
pollution such as in Donora, Pennsylvania,
Received for publication October 17, 1988, and in
final form May 4, 1989.
1
Environmental Protection Agency, Washington,
DC.
1
Batelle Memorial Laboratories, Research Trian-
gle Park, NC.
Reprint requests to Dr. Joel Schwartz, Environ-
mental Protection Agency, Office of Policy Analysis
PM 221, 401 M Street SW, Washington, DC 20460.
This work was supported in part by Contract
RP1001 from the Electric Power Research Institute
to the Harvard School of Public Health, Boston, MA.
Much of this work was done while Dr. Joel
Schwartz was a Visiting Scientist at the Department
of Biostatisbcs, Harvard School of Public Health,
Boston, MA.
The authors are indebted to Dr. John Moore of the
California Air Resources Board for his computational
assistance and constructive advice. They are grateful
to the California Air Resource Board, to the British
Medical Research Council and the Office of Popula-
tion Census and Statistics for providing the data, and
to Dr. David Wypij at the Department of Biostatistics,
Harvard School of Public Health, for his advice on the
two-stage multivariate modeL
in 1948 or London in 1952 suggested that
particulates were a cause of premature mor-
tality (1). Effects were strongest among
populations that are especially vulnerable
to such stress, including the chronically ill
and the elderly.
Progress in air pollution control has
greatly reduced particulate and sulfur diox-
ide levels in the United States, Great Brit-
ain, and other countries so that such major
episodes resulting in excess mortality are
less likely. The present challenge is to de-
termine whether smaller degrees of excess
mortality are likely in contemporary pop-
ulations exposed to much lower levels of
particulates and sulfur dioxide. This ques-
tion has been a major issue in the Environ-
mental Protection Agency's recent regula-
tion of particulates and requires much more
powerful methods of data analysis since the
effects in exposed populations are more
subtle. Detection of such subtle effects is
easier when important confounding factors
are controlled or when covariates are in-
cluded in the analysis.

186
186 SCHWARTZ AND MARCUS

Weather variables are important stress nent of the air and levels of British Smoke,
factors explaining part of the variations in an optical measure of particulate matter,
daily mortality. Other factors such as influ- were not substantially greater than current
enza epidemics that may occur simultane- levels. The dramatic and evidently causal
ously with elevated air pollution levels are role of elevated air pollution in excess
harder to measure, especially for historical deaths among vulnerable populations (5)
data. Some investigators have speculated lends credibility to statistical analyses.
that even within a single season, collinear- While several analyses of these data have
ity with omitted weather variables that already been performed (3, 4, 6, 7), they
vary slowly over the season (long wave- have not led to a consensus. In our view,
length) may artificially enhance the corre- the methodology used in those studies was
lation. A number of approaches have been inadequate to resolve some central issues.
used to eliminate such longer-term effects Specifically:
from mortality data so as to better assess 1) Is the observed relation between air
other suspected causes of excess daily mor- pollution and mortality an artifact of the
tality. These include subtraction of sinu- earlier use of ordinary least-squares regres-
soidal (seasonal) trends (2) and subtraction sion techniques on highly autocorrelated
of 15-day moving averages of daily mortal- data?
ity (3, 4). None of these have dealt with the 2) Is the relation a spurious one due to
high degree of short-term autocorrelation coincidence of long-wavelength (seasonal)
in the data and the subsequent lack of
oscillations in the mortality and air pollu-
independence of the regression residuals.
tion series?
In this paper, we will demonstrate some
3) Is the relation between daily particu-
advantages of autoregressive regression
lates and death counts linear, and does it
methods currently available.
manifest a threshold at 750 Mg/m3 as
The data set consisting of daily measure- Lawther et al. (8) suggested?
ments of air pollution and mortality in 4) Does controlling for weather factors
London during the winters of 1958-1959 to remove the correlation?
1971-1972 has played a central role in dis- 5) Can the effects of particulates and
cussions of the possible relation between sulfur dioxide be separated despite their
particulates, sulfur dioxide, and mortality. high degree of collinearity?
These data are the only available long-term We have attempted to resolve these is-
data set with reliable daily measurements sues using modern time series and graphic
of particulates and verified mortality rec- techniques.
ords. The use of every-sixth-day monitor-
ing in the United States has prevented the
DATA AND METHODS
construction of a similar data set. In addi-
tion, the time series data for daily particu- Data
lates and sulfur dioxide levels are averages The data consist of daily measurements
from seven stations and contain no missing of British Smoke, sulfur dioxide, tempera-
values; this is unusual in air pollution stud- ture, humidity, and death counts from all
ies and allows more detailed statistical ex- nontraumatic causes for the winters (112
ploration than with other data sets. days each year) of 1958-1972. The 24-hour
The data also remain of interest because collection period for pollution measure-
there is a clear spectrum of effect from the ments was lagged 15 hours behind the mor-
earlier years (especially 1958 and 1962), tality period. Deaths from respiratory
when air pollution was unmistakably a fac- causes and deaths from cardiovascular
tor in mortality, to the later years when causes were also available. We do not ana-
coal smoke was no longer a major compo- lyze them here in the interest of brevity,
 MORTALITY AND AIR POLLUTION IN LONDON
and because earlier authors (6) have shown
that the relation between all three mortal-
ity measures and pollution were similar.
The British Smoke method measures
light reflectance from dark particles col-
lected on filter paper. The method is mod-
erately efficient for collecting small parti-
cles (over 90 percent efficient for particles
below 1 nm), falling to 50 percent at 4.5 ^iin
and continuing to fall for large particles.
British Smoke and total suspended partic-
ulates were fairly well correlated at the end
of the period (1970) (9). Site-specific gravi-
metric calibrations for London were avail-
able through the early 1960s. The relation
between British Smoke and total sus-
pended participates appears to be nonlin-
ear, with British Smoke representing a
larger fraction of total suspended particu-
lates on highly polluted days than on less-
polluted days (9-11).
The British Clean Air Act was passed in
1956, banning the open-air combustion of
coal for home heating, and by 1965 and
onward, pollution was lower than in the
past and also had a different composition.
Both the level and the nature of the pollu-
tion in the last seven years were more typ-
ical of what is currently found in both
Britain and the United States. Table 1
187
shows the mean levels of British Smoke, by
year, in London, as well as the fraction of
days on which British Smoke levels were
less than 500 Mg/m3 and 200 Mg/m3. The
much lower fraction of high-pollution days
in the latter years is evident.
Analytic approach
We first plotted the data to exhibit the
shape of any dose-response relation, with
particular attention to low levels of pollu-
tion. We summarized the data in the plots
by sorting the observations in order of in-
creasing pollutant and taking the means of
adjacent observations. Similar plots were
constructed for temperature and humidity.
We then estimated multiple regressions of
daily mortality on temperature, humidity,
and air pollution. If indicated by the plote,
independent variable transforms were ex-
amined in the regressions. British Smoke
and sulfur dioxide were examined sepa-
rately because of the high collinearity be-
tween them. The residuals of these regres-
sions were examined for autocorrelation,
and autoregressive models were then spec-
ified. These models have been discussed
extensively in the time series literature,
such as the article by Box and Jenkins (12).

TABLE 1
Mean levels of British Smoke (BS) in London and the percentage of days on which BS levels were less than 500
fig/m* and less than 200 ng/m3, 1958-1971

MeanBS % of dayi on which BS level* were:

Year level 3200 Ms/m1 >200 xi/m 1
1958 536 62 38 6 94
1959 354 87 13 17 83
1960 257 92 8 38 62
1961 197 94 6 65 35
1962 208 93 • 7 69 31
1963 193 96 4 67 33
1964 132 99 1 86 14
1965 109 100 0 90 10
1966 100 100 0 94 6
1967 86 100 0 97 3
1968 68 100 0 99 1
1969 76 100 0 98 2
1970 67 100 0 98 2
1971 59 100 0 99 1
 188 SCHWARTZ AND MARCUS
These models were of the form:
+ ... + e,, (1)
where Y, is the number of deaths on day i,
X, is the vector of independent variables on
day i, £p is the vector of regression coeffi-
cients for lagp, ru r2,... are the autoregres-
sion parameters at lags of 1, 2, etc., days,
and e, is normally and independently dis-
tributed. The residuals of these models
were then reexamined to assure that no
autocorrelation remained, and the models
were reestimated with additional autore-
gressive terms if necessary. Regressions
were done in PROC AUTOREG of the
Statistical Analysis System, using maxi-
mum likelihood (13, 14). If it was indicated
by the ploto, the data were transformed
before the regressions were performed.
To control for year-to-year differences,
separate models were estimated for each of
the 14 years. The resulting regression coef-
ficients were then combined over all years
using a two-stage random effects model to
account for uncontrolled sources of varia-
tion between years. The model estimates
the regression coefficient for pollution as
the mean of the individual year coefficients,
weighted by the inverse of the estimated
covariance matrix. The random effects
model assumes that year-to-year variability
in the regression coefficients may exceed
that predicted by the within-year variances
because of omitted, possibly collinear, fac-
tors such as influenza epidemics, which
vary widely between years. This is de-
scribed in Appendix 1. We believe this ap-
proach is responsive to our first issue, es-
tablishing whether a relation still exists
when the regressions incorporate autocor-
relation in the residuals. The autoregres-
sive models we describe below also deal
with the long-wavelength issue in series of
only 112 days.
We also examined models using both
British Smoke and sulfur dioxide as co-
variates to attempt to resolve whether one
pollutant was more likely to be the princi-
pal cause of any association with mortality.
To test the robustness of our results to the
analytic approach, we filtered the depen-
dent and independent variables to remove
oscillations with periods longer than 15
days before performing our regressions.
. This tests the hypothesis that the observed
correlation is due to collinearity with omit-
ted long-wavelength seasonal factors. An
alternative approach is to determine
whether the expected value of daily mor-
tality conditional on its past and the past
of the independent variables has any fur-
ther association with the independent vari-
ables (15).' Hence, all of the series would be
prefiltered to white noise, and correlations
among the innovations of the series would
be examined. We have not used this as our
primary approach because we believe that
serial correlation in mortality rates is, in
fact, due to serial correlation in causative
factors, with no true biologic dependence
on past rates. However, we have used this
more conservative approach to test the ro-
bustness of any associations found using
the autoregressive models. These ap-
proaches are described in more detail
below.
RESULTS
Graphic analysis
Figure 1 shows the scatter plot of daily
mortality versus British Smoke, where each
point represents the mean of 20 consecutive
observations in increasing order of British
Smoke. Starting at the lowest observed lev-
els, 20 Mg/ni3, there is a relation between
British Smoke level and mortality. In ad-
dition, the slope of the relation decreases
at the higher levels. This is consistent with
previously published results. For example,
Mazumdar et al. (4) found higher regression
coefficients in the later years, when pollu-
tion was low, than in the early years, when
the average pollution level was much'
higher. Ostro (7) also reported a higher
coefficient below 150 /ig/m3 than above 150
3
in his spline analysis of the data.
 MORTALITY AND AIR POLLUTION IN LONDON 189
400

350 -
2

<
8 300

250
500 1000 1500

BRITISH SMOKE
FIGURE 1. Daily mortality in London, by British Smoke level, for the winters of 1958-1972. Each point
represents the mean total mortality and mean British Smoke (in tig/m3) for 20 adjacent values of British
Smoke, after sorting by British Smoke.

However, both of these analyses ignored 330

autocorrelation in the residuals.

The curvilinear slope is not a function of
some change that occurs over time, since it
occurs within individual years, as is shown
in figure 2, which plots daily mortality ver-
sus British Smoke in the winter of 1963-
1964. The graphic analysis allows us to
reject the threshold hypothesis. It also sug-
gests that a natural logarithm transforma-
tion or fractional power transformation of
220
the British Smoke data would be appropri- 100 200 300 400 500
ate. A similar curvilinear relation (not BRITISH SMOKE
shown) was obtained for sulfur dioxide. Fig- FIGURE 2. Daily mortality in London, by British
ure 3 shows the relation for temperature. Smoke level, for the winter of 1963-1964. Each point
In this case, the relation appears quite lin- • represents the mean of total mortality and British
ear, with lower temperature associated with Smoke (in fig/m3) for 10 adjacent values of British
Smoke, after sorting by British Smoke.
higher mortality. A similar plot (not
shown) indicated a linear relation with hu-
midity as well, and these variables were not of the British Smoke (or sulfur dioxide)
transformed for the regression analysis. data substantially linearized figure 1. To
Regression results choose between the natural log and frac-
Plots (not shown) showed that both log- tional powers, we estimated autoregressive
arithmic and square-root transformations models using hi (British Smoke) and Brit-
190 SCHWARTZ AND MARCUS

400 the greatest incremental R2 (the 0.6 power

was also superior to the natural log trans-
formation) and was used in our subsequent
analysis.
350 - On the basis of these results, we esti-
mated autoregressive regressions of mor-
tality on temperature, humidity, and
square-root British Smoke (or sulfur diox-
300 • ide) for each year. Cross-correlation func-
tions showed only a prompt effect of pol-
lution (and weather), and lags of one or two
days were likewise insignificant in the
250 regression analysis. We therefore report
-10 only results with contemporaneous expo-
sure measures.
<catsus) For daily mortality, British Smoke was
FIGURE 3. Daily mortality in London, by temper-
significant for 13 out of 14 years (table 2).
ature (Celsius), for the winters of 1958-1972. Each The random-effects model gave an overall
point represents the mean of total mortality and tem- coefficient for transformed British Smoke
perature for 20 adjacent values of temperature, after of 0 = 2.31 (£=14.45).
sorting by temperature. A similar pattern was obtained when sul-
fur dioxide was used as the pollutant, with
TABLE 2
sulfur dioxide significant for 12 out of 14
Daily mortality and British Smoke* in London,
regressions^ by year, 1958-1971 years. The random-effects model gave a
coefficient of (3 = 2.15 (t = 9.41). In almost
Year Coefficient t statistic
every year, sulfur dioxide was less signifi-
1958 2.81 7.23 cant than British Smoke, and we will report
1959 1.16 2.45 details only for British Smoke.
1960 2.00 3.79
1961 3.20 5.28
1962 3.08 5.21 Sensitivity analysis
1963 2.06 3.85 Our first sensitivity check was to deter-
1964 2.22 3.83 mine whether exclusion of the long-
1965 2.37 3.48
1966 2.52 3.19 wavelength variations in mortality and
1967 3.11 3.24 British Smoke would seriously alter the
1968 2.87 3.66 relation, which might suggest that weather
1969 3.05 2.86 or some other seasonal factor was inade-
1970 1.08 1.34 quately controlled. All series were filtered
1971 2.73 2.65
using deviations from 15-day moving aver-
Random-effects model for all years: ages. (A less traditional approach, using
fi = 2.31, t'- 14.45 Butterworth filters, gave essentially iden-
• Square-root transformation (/ig/m1).
tical regression results.)
t Controlling for temperature and humidity. The use of the filter reduced levels of
autocorrelation. We used an autoregressive
ish Smoke to the powers 0.3, 0.4, 0.5, 0.6, model to eliminate significant autocorrela-
0.7, and 0.8. All models controlled for tem- tion remaining in the regression residuals.
perature and humidity and included For filtered daily mortality, British Smoke
dummy variables for each year. The square was significant in all 14 years (table 3),
root of British Smoke minimized the with an overall coefficient of j9 = 2.53 (t =
Akaike information criteria (16) and had 13.23). Sulfur dioxide was insignificant in
 MORTALITY AND AIR POLLUTION IN LONDON 191
TABLE 3 with British Smoke, temperature, humid-
Daily mortality and British Smoke* m London, ity, and autocorrelation terms. In the two-
regressions^ by year, using filtered data, 1958-1971 stage model, British Smoke was a signifi-
Year Coefficient t 1UUMIC cant predictor of mortality, with a slope of
1958 2.74 722 /3 = 2.03 (t = 5.77), even when sulfur dioxide
1959 1.23 2J34 was included in the model. Thus, the Brit-
1960 1.98 2.84 ish Smoke slope was reduced by only 10
1961 2.79 4.91 percent when sulfur dioxide was also in-
1962 3.68 6.15
cluded as a predictor. On the other hand,
1963 1.62 3.27
1964 2.23 3.67 the mean sulfur dioxide slope was statisti-
1965 2.65 4.08 cally insignificant (0 = 0.39, t = 0.96) when
1966 2.61 3.75 British Smoke was also included in the
1967 3.86 3.63 model. Thus, the inclusion of sulfur dioxide
1968 2.73 3.73
3.15
as a covariate does not greatly reduce the
1969 3.06
1970 2.14 2.90 magnitude or significance of the British
1971 3.61 3.48 Smoke effect, yet sulfur dioxide consist-
ently fails to show up as significant in
Random-effects model for all years: models including British Smoke.
0 » 2.53, t =- 13.23
The same pattern was observed in the
* Square-root transformation filtered data. When both pollutants were in
t Controlling for temperature and humidity.
the model, the coefficient of British Smoke
(over all years) was 1.83 (t = 4.14). For
one year (not shown) but again highly sig- sulfur dioxide, the coefficient was 0.75 {t =
nificant overall. When innovations in daily 1.34).
mortality were regressed on innovations in
temperature, humidity, and air pollution, DISCUSSION
British Smoke 03 = 2.54, t = 13.81) was The overall conclusion from our analysis
again significant. is that there is a strong relation between
air pollution levels and daily mortality in
Separating British Smoke from sulfur London in these data which holds both in
dioxide the whole data set and in individual years.
The high degree of collinearity between Moreover, controlling for temperature and
British Smoke and sulfur dioxide makes it humidity increased the significance of Brit-
difficult to distinguish between them. No ish Smoke compared with models that
statistical analyses can completely control omitted those factors, and inclusion of the
for the intrinsic daily correlation of sulfur autoregressive terms to control for serial
dioxide and British Smoke levels within correlation also increased the significance
each year. This can induce a significant of British Smoke compared with ordinary
negative correlation between estimated least-squares regressions (not shown).
slopes for British Smoke and sulfur dioxide Thus, the hypothesis that the correlation
effects such that overstating the effect of between pollution and mortality is due to
one pollutant must, on the average, lead to ignoring weather variables or to serial cor-
understating the effect of the other. While relation can be rejected. The similar strong
this makes conclusions based on one year relation found in the data after removing
difficult, we have 14 replicate years. This long-term cycles, by deviations from 15-day
may give us enough information to separate moving averages or by prefiltering all the
out the effects. series to white noise, also indicates that
We fitted our autoregressive regression seasonal trends are not responsible for
models as above using sulfur dioxide along these correlations.
 192 SCHWARTZ AND MARCUS
Despite the high degree of collinearity
between British Smoke and sulfur dioxide
(correlation coefficients exceed 0.8 in all
years), in joint regressions British Smoke
remained a highly significant predictor of
mortality (p < 0.0001). Moreover, the over-
all coefficient of British Smoke when sulfur
dioxide was included dropped only by about
10 percent from its value in the identical
model without sulfur dioxide. We believe
that considerable confidence can be given
to the conclusion that there is a significant
correlation between British Smoke and
daily mortality in these data, independent
of any effect of sulfur dioxide.
The relation between sulfur dioxide and
mortality is much less stable. While the
mean coefficient of sulfur dioxide was
highly significant in the absence of British
Smoke, the inclusion of British Smoke sub-
stantially changed this. The coefficient was
not significant, and its magnitude was sub-
stantially reduced from the model which
did not include British Smoke. Overall, we
believe that the evidence for an indepen-
dent relation between sulfur dioxide and
mortality is weak, but the data do not allow
us to definitively preclude the possibility.
Lawther et al. (8) suggested that excess
mortality occurred only when British
Smoke was greater than 750 Mg/m3 and
sulfur dioxide was greater than 715 ng/mz.
Our analysis for years after 1965 demon-
strates a strong correlation in a period
where these levels were never exceeded. We
find that the apparent effect of British
Smoke on mortality is significant even at
these lower levels.
The curvilinear relation evident in figure
1 and the regression results show a larger
slope at lower pollution levels than at
higher levels. The most obvious explana-
tion is that there is a linear relation with
suspended particulates, and figure 1 reflects
the curvilinear relation between British
Smoke and suspended particulate levels.
Lee et al. (9) and Bailey and Clayton (10)
have presented data suggesting that British
Smoke was 80 percent of suspended partic-
ulates when air pollution levels were quite
high, but only 40 percent of suspended par-
ticulates when pollution levels were low.
Waller (11) shows a nonlinear relation be-
tween British Smoke and total suspended
particulates based on five years of parallel
measurements (1973-1979). Therefore, a
linear relation between mortality and par-
ticulates implies a curvilinear relation be-
tween mortality and British Smoke, with
the slope of the relation doubling at low
levels. This is at least qualitatively consis-
tent with the relation between British
Smoke and mortality seen in figure 1.
There are other possible explanations. Pol-
lution itself is highly autocorrelated. Since
very-high-air-pollution days usually follow
high-pollution days, the population of re-
sponders may have been depleted so that it
cannot respond proportionately to the high
levels on the following day. The essentially
identical results in regressions where each
series was prefiltered on its past makes this
explanation less likely. In addition, the
highest pollution levels were accompanied
by extremely low visibility, and both logic
and anecdotal reports suggest avertive be-
havior on those days.
The obvious question is whether the re-
lation between air pollution and acute mor-
tality observed here is causal. The relation
could fail to be causal either because some
omitted factor was correlated with both
particulate levels and mortality and in-
duced a correlation between the two of
them or because some omitted factor plays
an etiologic role in both, and therefore in-
duces a correlation. While no one can ever
prove there is no omitted covariate problem
in an analysis, our approach offers some
information on the issue. First, the relation
is significant in 13 out of 14 years when
analyzed separately in regressions with ap-
proximately 100 degrees of freedom. This
makes the hypothesis of a random factor
unlikely. An omitted variable bias would
have to occur in each of the 14 years and
most likely would have to be a systematic
factor. The most obvious one is weather,
 MORTALITY AND AIR POLLUTION IN LONDON
but, as noted above, inclusion of weather
terms in fact increases the strength of the
relation between particulates and mortal-
ity. Moreover, the similar relation in the
filtered data set indicates that the omitted
factor cannot be one described by long-term
cycles (such as bad weather increasing to a
peak in February and then falling), because
all oscillations of over 14 days duration
have been removed. Even non-weather-
related candidates for omitted variable bias
would have to be ones whose short-term
fluctuations were similarly correlated with
both particulates and mortality.
Indoor air pollution levels as well as out-
door levels may be associated with mortal-
ity. However, there was no significant dif-
ference between the slope coefficients esti-
mated for outdoor pollution during the
early years, when home combustion of coal
would have led to high indoor air pollution
levels, compared with the last seven years,
when such coal combustion was banned.
Thus, trends in indoor pollution are an
unlikely source of bias. The nature of the
relation observed does not readily suggest
that omitted variable bias is present.
The object of our analyses has been to
quantify the relation between British
Smoke and mortality at the lower levels of
British Smoke of current environmental
concern. These are, by design, levels at
which there will not be dramatic evidence
of excess mortality; hence, it is necessary
to assume a mathematic relation that pro-
vides a continuous extrapolation of effect
(mortality) from higher to lower concentra-
tions. This does not prove that British
Smoke caused excess mortality in individ-
uals exposed to low concentrations in Lon-
don in 1958-1972. Additional evidence for.
the hypothesis that particulates are asso-
ciated with mortality is given in recent
reanalyses of New York City mortality data
(1) in which the coefficient of haze and the
visibility range were found to be statisti-
cally significant predictors of mortality,
more so than sulfur dioxide.
A number of possible biologic mecha-
193
nisms exist, which suggests that the rela-
tion between British Smoke and mortality
is causal. These include chemical and me-
chanical irritation or stimulation of recep-
tors, reduced mucociliary clearance, insults
to macrophages and other defense mecha-
nisms, and other forms of tissue damage.
When particulate levels were very high in
London in the early 1950s, the excess daily
mortality was so high as to leave no ques-
tion of a causal relation (1). Our analysis
suggests that even at today's lower levels,
a 10 percent reduction in particulate matter
in London would result in several hundred
fewer early deaths per year. The possible
connection between British Smoke and
London mortality has been discussed for
over 25 years, with no more plausible ex-
planation advanced thus far. We believe
that it is prudent to continue to assume
that control of British Smoke levels
will control the direct and/or indirect
relation(s) of British Smoke to excess
mortality.
REFERENCES
1 Environmental Protection Agency. Air quality cri-
teria for particulate matter and sulfur oxides. Vol
3. Research Triangle Park, NC: Environmental
Protection Agency, 1982. (EPA publication no.
600/8-82-029c).
2. Ozkaynak H, Spengler J, Garzd A, et al. Assess-
ment of population health risks resulting from
exposure to airborne particles. In; Lee SD, ed.
Aerosols: research, nsk assessment, and control
strategies. Chelsea, MI: Lewis Publishers, 1986.
3. Martin AE, Bradley WH. Mortality, fog, and at-
mospheric pollution: an investigation during the
winter of 1958-59. Monthly Bull Min Health Pub-
lic Health Lab Serv 1960:19:56-72.
4. Mazumdar S, Schimmel H, Higgins ITT. Relation
of daily mortality to air pollution: an analysis of
14 London winters, 1958/59-1971/72. Arch En-
viron Health 1982;37 213-20.
5. Pnndle J. Notes made during the London smog
in December 1962. J Air Pollut Control Assoc
1963;7:493-6.
6. Shumway RH, Tai RY, Tai LP, et al. Statistical
analysis of daily London mortality and associated
weather and pollution effects. Sacramento, CA:
California Air Resources Board, 1983. (Contract
no. AX-1M-33).
7. Ostro B. A search for a threshold in the relation-
ship of air pollution to mortality- a reanalysis of
data on London winters. Environ Health Perspect
1984^:397-9.
8. Lawther PJ, Waller RE, Henderson M. Air pol-
194 SCHWARTZ AND MARCUS

lution exacerbations of bronchitis. Thorax 1970; the independent variables. However, we assume that
25:525-9. the true slopes (S,) may actually differ from year to
9. Lee RE Jr, Caldwell JS, Morgan GB The evalu- year from their mean value B because of random
ation of methods for measuring suspended partic- factors. In particular, if
ulates in air. Atmos Environ 1972;6:593-622.
10. Bailey DLR, Clayton P. The measurement of sus- B.|S, D-N{B, D) (1)
pended particle and total carbon concentrations
in the atmosphere using standard smokeshade represent the distribution of the Bi and
methods. Atmos Environ 1982;16.2683-90.
11. Waller RE Assessing effects of suspended par- 6,\Bt ~ N(Bi, Vt) (2)
ticulate matter on health. European Commu- are the estimated slopes (6<) and covanances (t/,) from
nity Workshop RTVM-BUthoven, European Eco-
the first stage, then
nomic Community, 1984.
12. Box GEP, Jenkins GM. Time series analysis: fore-
6. |B, D ~ N(B, V, + £». (3)
casting and control. Revised ed. San Francisco:
Holden-Day, Inc. 1976. Berkey and Laird (17) showed, using the EM algo-
13. Kay A, ed. Statistical Analysis System (SAS) rithm, that the joint multivanate maximum likelihood
Cary, NC: SAS Institute, Inc, 1985. estimates of B and D can be obtained by iteratively
14. Harvey AC, Phillips GDA. Maximum likeli- solving
hood estimation of regression models with
autoregressive-moving average disturbance. Bio- (4)
metnka 1979;66.49-58.
15. Haugh LD, Box GEP. Identification of dynamic and
regression (distributed lag) models connecting two
tune series. J Am Stat Assoc 1977;72:121-30.
16. Akaike H. Information theory and an extension
of the mBTimiim likelihood principle. In: Petrov (5)
BM, Csaki F, eds. 2nd International Symposium Initial estimates of D were obtained using the method
on Information Theory, Budapest: Akademiki of moments. The asymptotic covanance of B is esti-
Kiado, 1973:267-81. mated by
17. Berkey CS, Laird NM. Nonlinear growth curve
analysis: estimating the population parameters. (6)
Ann Hum Biol 1986;13:lll-28.
For highly collinear models, such as those containing
APPENDIX both British Smoke and sulfur dioxide, the off-
diagonal elements of V, can be quite substantial, and
Two-stage random effects model for slopes D is likely to have significant off-diagonal elements as
We assume a first-stage linear autoregression esti- welL These reflect the high negative correlation be-
mation of the coefficients 6t and the covariance fy of tween the coefficients.