EPA 560/3-76-002
PILOT STUDY OF THE ASSOCIATION BETWEEN
CANCER MORTALITY AND INDUSTRIAL
CONCENTRATION IN 200 U.S. COUNTIES
Environmental Protection Agency
Office of Toxic Substances
Washington, D.C. 20460
December 1976
-------
ABSTRACT
The Office of Toxic Substances is undertaking studies which will attempt
to identify potentially regulable sources of health hazards in the
environment.
These studies involve statistical analyses of relationships'
between industrial concentrations and mortality rates from various
diseases in all 3,069 u.S. counties.
A pilot study, involving 200
counties, was undertaken to anticipate and resolve as many methodological
problems as possible before encountering them in the full-scale effort.
This report describes the results of that pilot study.
-------
TABLE OF CONTENTS
Page
Introduction and Purpose
1
. Analysis of the 200 County Data
3
Methodological Problems
25
Methodological Analyses
34.'>
-: "I
-------
LIST OF FIGURES
1
Plot of Respiratory Cancer Residual by
Concentration of a Group of Correlated Industries
2
Plot of Respiratory Cancer Residual by
Concentration of Women's Dress Manufacture
.3
Plot of Digestive Cancer Residual by
Concentration of Fertilizer Manufacture
Page
2.3
41
42
-------
LIST OF TABLES
1
Thirty Industries Selected for Study
2
Cancer Mortality Categories Selected for Study
3
Industries Selected by Stepwise Regression as Significantly
Related to Residual Cancer Mortality
4
Correlation Coefficients for All Industries Significantly
Correlated with Respiratory Cancer
5
Industries Selected by Stepwise Regression as Significantly
Related to Residual Cancer Mortality in Each of Two Random
100-County Groups
6
The 30 Selected Industries Combined into Six Groups
7
200-County Ranking of Age-Adjusted Mortality Rates for
Respiratory Cancer, for Four Race-Sex Categories, 1950-
1969. Top 15, Ordered by White Male Rankings
8
200-County Ranking of Age-Adjusted Mortality Rates for
Respiratory Cancer, for Four Race-Sex Categories, 1950-
1969. Top 15, Ordered by Nonwhite Female Rankings
9
Comparison of Stepwise Regressions on Industry for Counties
by Percent of Workforce Commuting
10
Percent Distribution of Deaths by Cause, 1970: U.S.,
Counties Contributing to Negative Stepwise Regression,
and Their States
iv
Page
5
6
13
15
16
21
36
37
42
44
-------
PILOT STUDY OF THE ASSOCIATION
BETWEEN CANCER MORTALITY AND INDUSTRIAL CONCENTRATION IN
200 U.S. COUNTIES
INTRODUCTION AND PURPOSE:
As part of an attempt to identify health hazards posed by potential
regulatory targets~ the Office of Toxic Substances (OTS) is undertaking
a computerized statistical analysis of the relationship between indus-
trial concentrations and mortality rates from various diseases in all
3,069 U.~. counties.. Since this involves compiling and correlating
several massive data sets heretofore not considered conjunctively, many
unforeseen methodological problems can arise which~ once the full-scale
project is under way, would be expensive or impossible to correct.
To anticipate and resolve as many such problems as possible, OTS implemen-
ted a smaller, more manageable~ 200-county pilot study.
The experience
gained in this pilot study, carried out by the same OTS staff members
who developed the methodology for the full-scale project, was incorporated
directly into that larger effort.
This report describes the methodological
problems discovered during the pilot study, and discusses how they are
to be handled in the full-scale project.
It is in four sections: .'..
1.
Introduction and Purpose
2.
Analysis of ZOO. County Data
Describes the difficulties encountered in defining the relationship
between four broad categories of cancer and 30 selecte~ industries
in the 200 counties.
J
-------
3.
Methodological Problems
Describes the methodological problems discovered in the pilot study
and.:the steps taken to avoid them in the full-scale project.
4.
Methodological Analyses
Describes some additional analyses done to shed light on methodological
issues.
-------
ANALYSIS OF THE 200 COUNTY DATA:
This pilot study was limited to an analysis of the relationship between
30 industries and four broad categories of cancer mortality among the
200 most populous counties in the U.S.
The mortality rates were adjusted
to neutralize the effect of certain known non~industrial correlates of
mortality.
As shown by the following summary, the pilot study was far
more modest in scope than the full-scale project will be.
Pilot Study
Full-Scale Project
200 most populous counties
in the continental U.S.
All counties in the continental
U.S.
30 selected industries in 1959.
Over 600 manufacturing, extrac-
tive and transportation industries
in 1959, 1967, 1973.
Four broad categories of cancer.
All causes of death divided into
56 categories.
Crude mortality rates for total
county population.
Age-adjusted mortality rates by
race and sex for the age band 35-74.
10 demographic control variables.
Approximately 40 non-regulable
control variables.
A.
Description of data sets
1.
Industrial
The index of industrial concentration for each county and
specific industry was the percent of the total county workforce engaged
in the specific industry in 1959 expressed as a multiple of the percent
of the total U.S. workforce in that industry in the same year.
Thus,
Index = 100~NSI/CWF
uSSr/USWF
Where:
cwsr = Number of county workers in the specific industry.
CWF
= Total county workforce.
ussr = Number of U.S. workers in the specific industry.
USWF = Total U.S. workforce.
-------
Thus, if the index for a specific industry in a specific county is 800, the
percent of that county workforce employed by that industry is eight times
the percent of the U.S. workforce employed by the industry.
For most of the statistical analyses discussed here, it makes no difference
that the county percentage (CWSI/CWF) is, divided by the U.S. percentage
(USSI/USWF) for the same industry; the index is equivalent to the county
, percentage alone.
However, when index values for several industries are
added together to produce a group index, division by the U.S. percentage does
make a difference.
This procedure will be discussed in Section F.
Social Security employment figures published by the Census Bureau were
used in computing the index (1).
Although exact employment figures were
withheld in some instances, they were estimated from other data available
in the same publication.
Thirty specific industries, listed in Table 1, were selected for analysis
because of their relatively high concentration of workers in counties
with high age-adjusted cancer mortality rates among white males during
the period 1950-1969.
Table 2 lists the cancer mortality categories
that were selected for study.
~1)
U.S. Bureau of the Census and U.S. Bureau of Old-Age and Survivors'
Insurance; Cooperative Report, County Business Patterns, First
Quarter 1959, Part 1, U.S. Su~ry. Washington, D.C., 1961.
-------
Abbreviation
-
IRONMIN
ANTHMIN
PETROMIN.
OILSERV
SULFMIN
BEER
CIGARS
NAROFAB
BROADFAB
BLOUSES
DRESSES
COALTAR
DYES
ORGANICS
PLASTICS
ELASTICS
DRUGS
MEDCHEM
FRTLIZER
PETROREF
ASPHALT
RUBRPROD
GYPSUM
STLMILLS
STLFOUND
COP SMELT
COPRLNG.
WIRE
DOCKS
STEVEDOR
TABLE I
Thirty Industries Selected for Study
S. 1. C. Number
Industry
101
1111
131
138
1477
2082
212
224
Iron ore mining
Anthracite mining
Crude petroleum and natural gas mining
Oil and gas,field services
Sulfur mining
Malt liquors
Cigars
Narrow fabrics and other small wares
mills: cotton, wool, silk and man-
made fibers .
Finishers of broad woven fabrics of
man-made fiber and silk
Women's, misses', and juniors' blouses,
waists, and shirts -
Women's, misses', and juniors' dresses
Cyclic (coal tar) crudes
Dyes, dye (cyclic) intermediates,
and organic pigments (lakes and toners)
Industrial organic chemicals, not
elsewhere classified
Plastics materials, ,synthetic resins
and non-vulcanizable elastomers
Synthetic rubber (vulcanizable elastomers)
Drugs
Medicinal. chemicals and botanical products
Fertilizers
Petroleum refining
Asphalt felts and coatings
Fabricated rubber products, not elsewhere
classified
Gypsum products
Blast furnaces (including coke ovens),
steel works and rolling mills
Steel foundries
Primary smelting and refining of copper
Rolling, drawing, and extruding of copper
Drawing and insulating of nonferrous wire
Piers and docks
Stevedoring
2262
2331
2335
2814
2815
2818
2821
2822
283
2833
2871
291
2952
306
3275
3312
3323.
3331
3351
3357
4462
4463
Source: U.S. Bureau of the Budget, Standard Industrial Classification Manual,
prepared by the Technical Committee on Industrial Classification, Office of
Statistical Standards, Washington, D.C., 1957.
5
-------
Abbreviation
TOTALCAN
DIGSTCAN
.' RESPCAN
URINCAN
TABLE 2
Cancer Mortality Categories Selected for Study
I.C.D. Numbers
140-209
150-159
160-163
188-189
Disease Catego~
Malignant neoplasms, including neoplasms of
lymphatic and hematopoietic tissues
Malignant neoplasms of digestive organs
and peritoneum
Malignant neoplasms of respiratory system
Malignant neoplasms of urinary organs
Source: World Health Organization, Manual of the International Statistical
Classification of Diseases, Injuries, and Causes of Death, 8th Revision,
Vol. I, Geneva, 1967.
-------
(2)
(3)
(4)
2.
Mortality
Mortality rates per 100,000 population by county for total
cancer, respiratory cancer, digestive cancer, and urinary
cancer, were calculated from published mo~tality (2) and
population (3) data.
Crude rates were computed because deaths
by age are not published at the county level.
3.
Non-industrial correlates
,The following non-industrial correlates of mortality were
used as control variables:
population, median age, percent
aged 65 and over, .percent non-white, population per square mile,
percent urban, median family income, median last school year
completed, physicians per 100,000 and hospital beds per 100,000
(3,4).
A number of counties and cities in Virginia and the five
boroughs comprising New York City were combined into single
observations to compensate for inconsistencies between data
sources.
The methods used to produce aggregate values for
these observations are described in a computer printout available
from OTS.
u.s. Public Health Service. Vital Statistics of the United
States 1970, Vol. 2, Mortality, Part B, Tables 7-9. Rockville,
Md., 1974.
U.S. Bureau of the
Characteristics of
Section 1, Chapter
D.C., 1973.
Census. Census of Population 1970, Vol. 1,
the Population, Part 1; U.S. Summary,
A, Number of Inhabitants, Table 24. Washington,
American Medical Association. Distribution of Physicians in the
United States 1970, Regional State, County, Metropolitan Areas,
Table 12, Medical Practice Data by State and County. Chicago;
'Illinois, 1971. .
7
-------
B.
Stepwise re~ression
Stepwise regression was used to remove the effects of non-industrial
correlates from the mortality data and to select industries significantly
related to the resulting residual mortality.
Regression analysis is a procedure for determining the mathematical
relationship between a dependent variable and a set of independent'
variables when that relationship is not exact (i.e., when there is random
error).
There are many possible techniques for determining that.
relationship; the standard technique used in the pilot study is known
as linear least squares.
Linear,means that the relationship between the
independent and dependent variables is assumed to be linear.
If there
is only one ,independent variable, the r~lationship is depicted by a '
straight line on a graph.
For three or more variables the relationship
is still conceptually a straight line, although it can no longer be
depicted on a single graph.
Least squares is a method for selecting the
most suitable straight line.
Stepwise regression is a method for selecting the "best" regression.
equation by including only those independent variables which improve the
degree to which the mathematical equation fits the observed data by more
than some predetermined amount.
It is called stepwise because it selects
variables one at a time in a series of steps.
The first variable
selected is that which has the highest significant correlation with the
-------
dependent variable.
Subsequent variables are selected on the basis
of a partial correlation between an independent variable and the dependent
variable after the effects of all previously selected independent variables
have been removed.
The stepwise regression used in the pilot study retained
only those variables whose correlations were significant at probability =
.10 level.
The set of independent variables selected by this procedure, and the
estimated slopes which describe their mathematical relationship to the
dependent variable, are known collectively as the independent variable
model.
This language is used whether or not model building is the
primary.purpose of the analysis.
Thus this report will refer to the
"demographic model" or the "industrial model".
An important characteristic of stepwise regression is that it selects
only those independent variables which have a significant relationship to
the dependent variable over and above the effect of all other independent
variables in the model.
Thus a variable may fail to be selected even
though, taken alone, it has a high correlation with the dependent variable.
For example, both median age and percent aged 65 and over are highly correlatec
with cancer mortality but one or the other alone may sufficiently account
for the relationship of both in the demographic model.
c.
Removal of the effects of demographic and health care variables
~- . ."
Purpose and procedure:
in order to isolate and identify the role of
industry, it was first necessary to control for the non-industrial factors
-------
correlated with mortality.
Because it is capable of handling a
large number of variables simultaneously, regression analysis was
used to remove the effects of the ten demographic and health care
variables from the mortality data before the industrial analysis
was performed.
It is worth noting that some recent studies of the.
relationship between industry and cancer lack such safeguards.
The
stepwise procedure was used to make the computation more efficient
by eliminating non-essential variables, not because of any interest
as to which non-industrial variables fell into the model.
The
following procedure was used to remove the effects of the non-
industrial variables from each of the four cancer mortality variables:
a.
Determination of the non-industrial model by
stepwise regression
b.
Computation of the predicted mortality for each
county based upon the model determined in (a).
c.
Computation of the residual mortality for each
county by subtracting the predicted from the
observed mortality.
Thus the residual mortality for a particular county may be positive
or negative depending upon whether that county's observed mortality
is higher or lower than the non-industrial model would predict.
A
residual of 12, for example, indicates that there were 12 more deaths
per 100,000 population than would have been predicted on the basis
of median age, population per square mile, etc.
After a statistician
joined the study, another step was included:
d.
Examination of scatter plots of the residuals
-------
to insure elimination of all observable effects of
non-industrial variables.
This step turned out to be important; it was discovered that
a potentially biasing curvilinear relationship remained between
residual mortality and several non-industrial correlates. To
remove these effects, additional non-industrial variables were
created from the cross-products of the old ones (e.g., median
age squared, population per square mile multiplied by income).
These variables have little meaning in themselves but provide a
way of depicting a curvilinear relationship with an essentially
linear model. (True non-linear regression is a much more complex
process.)
Steps (a) through (d) were repeated with the newly created
variables added to the set.
This process removed most of the
curvilinear effect, but some irregularities remained in the median
age and percent aged 65 and over variables.
They appeared to
be due to a single observation (Pinellas, Florida) with a
very high median age.
Since these irregularities could not be
removed without the undesirable effect of excluding nearly five
percent of the 200 counties, the Pine11as data remained.
Development of the industrial model
D.
The relationships between the residual mortalities computed in
Section C and the 30 specific industries were determined by stepwise
11
-------
regression.
At this stage the model itself is of primary interest,
particularly the set of industries falling into it.
As shown in Table 3, there were 12 instances of significant
relationships between an industry and cancer mortality.
Two
industries were associated with both total and digestive cancer.
Two relationships were negative, indicating lower than expected
cancer mortality; eight industries had significant positive
relationships to one or more of the four categories of cancer.
These results will not be detailed for reasons which will
become apparent in the next section.
However, two points need to
be made:
1.
Several of the significant relationships were due entirely
to one or two counties because of the extremely skewed
distribution of the industrial variables.
Each industry was
contained in only a small minority of the counties (ranging
from a low of one percent for sulfur mining to a high of 35
percent for rubber products).
However, even among these few
counties the distribution was skewed, most having fairly low
index values while one or two were extremely high.
1f the .few
counties with high index values also had high residual.
mortalities, there would be a significant positive relationship
regardless of mortality rates in the other counties.
Similarly, .
if they had low residual mortalities, there would be a
signific~nt negative relationship.
One or two counties were
responsible for both negative relationships and two of the
positive relationships shown in Table 3.
1?
-------
. _._~-~._~--~- --~--
Table 3
Industries Selected by Stepwise Regression as Significantly
Related to Residual Cancer Mortality
Type of Cancer
Positively
Related
Negatively
Related.
Total cancer
Iron ore mining
Dye and pigment.
manufacturing
Organic chemical
manufacturing
Respiratory
Sulfur mining
Dress manufacturing
Steel works & .mills
Stevedoring
Digestive
Iron ore mining
Fertilizer manufacturing
Dye and pigment
manufacturing
Gypsum products
Wire drawing
,
11
-------
2.
As pointed out in Section B, an industry could be
significantly related to residual. mortality but fail to be
selected by the model because of its correlation (i.e., concen-
tration in the same counties) with another industry.
Comparison
of Table 3 and Table 4 shows that fewer than half the
industries significantly correlated with residual respiratory
cancer were selected by the model.
E.
'Testing of the industrial model
In order to avoid spurious significance and instability
produced by intercorrelations'among industries, the relationship
between the industries and residual cancer mortality was more
stringently tested. The 200 counties were split into two groups
of 100 counties each by random assignment and an industrial model
was developed independently for each. The test was whether an
industry was selected into both models.
The traditional method for
testing in independent groups is, to deyelop the model with one
group and apply the resulting mathematical equation to the other
to determine how well, it predicts the dependent variable values of
the second group.
But because this_pilot study was more concerned
with the industries chosen than with the estimated slopes of the
equation, the procedure of developing two independent models was
used.
The two independent analyses produced entirely different
sets of industries (Table 5).
Three factors wer~ responsible for
this result.
:t/~,
-------
Table 1+
Correlation Coefficients for All Industries Si~nificantly
Correlated with Residual Respiratory Cancer
Industry (abbreviation)
Correlation Coefficient
SULFMIN
.205
STLMILLS
.117
STEVEDOR
.159
DRESSES
-.210
PETROREF
.117
ORGANICS
.148
ANTHMIN
-.201
DOCKS
.146
OILSERV
.130
. ,.:r5
-------
'rable 5
Induetr1ea Selected by Stepwise Regre8810n 88 S1Rn1f1c8ntl~
ftelated to Residual Mortality in Each &f Two Random lOO-County aroup~.
'.
Cancer'type
(Res 1dua1)
"Group 0"
"Croup I"
Positively correlated
. industries
(abbreviation)
Negatively correlated
, industries
(abbreviation)
Positively correlated
industries
(abbreviation)
Negatively correlated
industries
(abbreviation)
Total
----------
Respirat01'1: .
...
'"
Digestiva
Urinary
IRONHIN
ORGANICS
. PLASTICS
DYES
STUIILLS
DOCKS
IItONHIN
PLASTICS'
COALTAR
DYES
ANTHMIN
BEER
PETROREF
STEVlmOR
GYI.~UH
WlltE
..----....---
. MEDCHEM
STEVEDO!
_rTF"" .......
, .
.GYPSUM
. BLOUSES
WIRE
DRESSES
ELASTICS
CIGARS
FRTLtZE!
CIGARS
MEDCP.~
-------
1.
This was intentionally a stringent test equiva1e~t to
using the .01 level of significance. Reducing the number
of observations from 200 to 100 further decreased the
probability of detecting a true difference. This problem
will be mitigated somewhat in the full-scale study when
there will be over 1500 observations in each group.
2.
As previously observed, there were very few counties
with high concentrations of each industry;.. there were
fewer still when the number of counties was cut in half.
Estimates of residual mortality at the high industrial
concentrations, being based on few observations, were
unstable, and therefore produced unstable estimates of
the slope.
3.
Results of additional analyses of the industrial variables
indicated that multicolinearity (correlation of industries
with each other) was a serious problem in this data set,
with some correlations greater than .80.
The high mu1tico-
linearity of a set of independent variables can produce
estimates so unstable that a small difference in the data
set can make a very large difference in the estimates and
hence in the variables selected by stepwise regression.
This is almost certain to be a problem in the full-scale
study because of the many independent variables, even though
the correlations are expected to be much lower for industries
not in this set.
-------
F.
Combining industries to reduce mu1tico1inearity
Three general approaches to dealing. with the problem of mu1ti-
co1inearity were considered:
1.
Although use of recently developed sophisticated
statistical techniques (e.g. ridge regressions) would produce
better regressio~ estimates for mathematical relationships
. they would be less useful for the variable selection in this
study.
Because these techniques are new, reliable computer
programs are not widely available to compute calcula-
tions and the techniques themselves require more
explanation than the more familiar tests used here.
2.
Individual regressions for each industry would avoid
computational problems and would be easy to understand.
However, with the number of variables involved in the
full-scale study, it would be very expensive.
Also,
there would be the possibility that the significant
correlation of one industry with another would be ignored.
An analysis should underline this effect rather than ignore
it.
3.
The industrial data set could be reduced by combining
correlated industries.
Industries too highly correlated
for their effects to be separated would be treated as
an entity.
This procedure was used in the pilot study
and will probably be used in the full-scale ~tudy.
-------
Although it is easy to describe and does not require new
computer programs, there are several disadvantages.
One
disadvantage is the necessity of combining industries
with nothing in common except location, thus producing
a new variable which is difficult to refer to except by
a listing of the component industries.
Another is the possi-
bility that some relationships between individual industries and
mortality would be obscured. . Additional analyses are under
way. to resolve the problem.
The procedure used in this study was to combine highly correlated industries
while also keeping similar segments together as much as possible.
The
Standard Industrial Code (5) was useda~\a guide in combining..industries
as follows:
1.
Correlated industries within the same three-digit
code were combined by addition of index values;
correlations were computed for the new set of industry
groups.
2.
Correlated industries within the same two-digit code
were combined; correlations were computed for the new set.
(5)
U.S. Bureau of the Budget. Standard Industrial Classification Manual.
Prepared by the Technical Committee on Industrial Classification,
Office of Statistical Standards, Washington, D.C., 1957.
-------
3.
Industries were combined on the basis of size of
correlations until a set of essentially uncorre1ated
industries was produced.
(A mathematical characteristic
of the correlation matrix determined when the set was
essentially uncorre1ated.)
This procedure produced
a set of six uncorre1ated industry groups, defined
in Table 6.
The reduction in the number of industries is
not expected to be as drastic in the full-scale study
where correlations will be lower on the average.
Note
that the sums were weighted to protect against domination
of the combined variable by large industries.
This followed
naturally from the way the original index was computed.
G.'
Relationships between combined industries and residual mortality
The relationship between the residual mortalities computed in Section
C and the newly defined set of six industry groups was determined
independently by stepwise regression for each of the randomly
selected subgroups of 100 counties.
Results are shown in Table 6.
The "Combo 1" industry group was selected by both of the independently
determined models as significantly related to residual respiratory
cancer mortality.
This was an indication that the failure to find
any overlap in the 30-industry test was largely due to mu1tico1inearity;
the probability of finding overlap by chance alone would be much
less with six independent variables than with 30.
-------
- -
Table 6
The 30 Selected Industries Combined into Six Groups
Group (abbreviation)
Industry (abbreviation)
ANCICLTH
ANTHMIN
BLOUSES
CIGARS.
DRESSES
BEER
BEER
cmmo 1
DOCKS
ELASTICS
FRTLIZER
GYPSUM
OILSERV
ORGANICS
PETROREF
PETROMIN
STEVEDOR
SULFMIN
cmmo 2
ASFHALT
,
COAL TAR
DYES
STLFOUND
STLMILLS
COMBO 3
BROADFAB
COPRLNG
COP SMELT
*DRUGS
NAROFAB
PLASTICS
I RUBRPROD
WIRE
IRONMIN
IROl\"MIN
, -
*Since MEDtHEM (S.I.C. 2833) is a subdivision of DRUGS (S.I.C. 283), it is
not listed separately. .
-------
The regression line is shown in Figure 1 fOr each of the
200 counties according to their values on combined index and
residual respiratory mortality.
The five counties with highest
values on the combined variable are given along with their
ranking among the 200 counties for each industry contained in
Combo 1.
Figure 1 illustrates how the correlation among
industries necessitated combination.
The two counties which
contributed most to this regression, Jefferson, Texas, and Jefferson,
Louisiana, contained 7 and 8, respectively, of the industries
included in Combo 1; together they contained all 10.
It would
be misleading .to report a relationship between respiratory cancer
and the Combo 1 industries without also indicating that the same
I
few counties were responsible for all of these correlations.
There
is no way to determine which industries if any were responsible for
increased respiratory cancer mortality without fo11owup studies.
Studies relating industry to cancer mortality by county which
deal with only one, or very few, industries, selected a priori,
can be misleading.
For example, a recent study by Thomas Mason (6)
reported an increase of lung cancer among males and females in four
counties with synthetic rubber manufacturing due largely to the
high concentration of synthetic rubber manufacturing in Jefferson,
Texas.
However, as Figure 1 shows, although Jefferson, Texas,
ranked high in concentration of synthetic rubber manufacturing
it ranked even higher in petroleum refining and equally high in sulfur
(6)
Mason, Thomas J. Cancer Mortality. in U.S. Counties with Plastics
and Related Industries. Environmental Health Perspectives,
2:79-84, June 1975.
-------
...
!:
3.
.,
tI
~ t~
'"
tI
U
C
c:
u
>.
'"
'"
..
..,
'"
<04
Co
'"
tI
aI
12.00000000
6.00000000
.....
~
CI
AI
U
"
e-
x
!oJ
I
'0
CI
>
'"
"
'"
.D
o
......
0.0,0000000
-~.oooooooo
-12.01)000000
.8
IA
I
I
IA
IAA
I A
18
IBAA
19
oDAAA A A
IC A
If'
18
18
IC A
I BAAA A
ICBCA
IKq
1(9
A
I CEUA A
If"
If' AA
I GAAAA A
IDA AA A
IE 11 A
IBAA
I
lEI) A
oOC
IA A
ItJ A A
I
IA
I
IA
IA
I
I
oA
I
I
IA
I
I
I
I
Figure 1
PLOT OF RESPIRATORY CANCF.R RESIDUAL BY CONCENTRATION'
OF CROUP OF CORRELATED INDUSTRIES.
A
A
A
A
Jefferson, Louisiana ~
Petroleum mining (13)
Oil & gas field services (1)
Sulfur mining (1)
Organic chemicals (5)
Fertilizers (5) .
Petroleum refining (24)
~ast Baton Rouge. Louisiana
Organic chemicals (3)
Synthetic rubber (1)
Petroleum refining (10)
~0
effers~n. Texi!S
Petroleum clnlng (7)' .
Oil A gas field services (11)
Sulfur mining (2)
Organic chemicals (6)
Synthetic rubb~r (2)
Petroleum refining (1)
Stevedoring (3)
1---------.-------------------.-------------------.-------------------.-------------------.--------------.----.--;--
~~anawha, West virlfnia
Petroleum mining 8)
Organic chemicals (1)
Synthetic rubber (3)
Petroleum refining (28)
Gypsum products (3)
Piers & docks (1)
A
A
A A
A
A
A
A
A
.
A
A A
A
"
A
2350.00000.
" 7350.00000
12350.00000
Inrlll"t.rhi f'onr:('nt.fnt,lon Inlll'!!.
17350.00000
Lfrr~o~ ~ ~ 1 OA~
I) II , OB!7
rn:
22350.00000
27350.00000
-------
mining.
H.
Summary.
In the full-scale study all sectors of a wide range of
industries will be screened and correlated industries will be .
combined to reduce the likelihood of drawing misleading inferences.
As an integral part of the analysis of the pilot study data, several
methodological issues were identified, addressed and resolved:
.3.
1.
The control procedure for non-industrial correlates was expanded
to include non-linear relationships.
2.
A method was developed and used to reduce spurious- res~lts py
independent testing.
The problem of multicolinearity was studied and a viable
solution adopted.
4.
Excess respiratory cancer mortality rates were found-in counties
containing a group of correlated industries.
This finding
brings into question the results of some other recent studies.
',1.\'
-------
METHODOLOGICAL PROBLEMS
The major purpose of the pilot study was to anticipate and resolve problems
prior to implementing the full-scale project.
In this respectt it was a
success; many methodological problems and deficiencies were discovered.
In terms of resolution they fall into three categories:
A.
Those which could be and were resolved in both studies;
B.
Those which were resolved for the full-scale project but
could not be resolved in the pilot study;
C.
Those for which we have not yet been able to find a solution and
which may be insoluble.
The problems in each of the three categories are discussed below in greater
detail in other sections of this report and in the Proposed Methodology
Descriptiont available from OTSt for the full-scale study.
A.
Problems resolved in both studies
1.
Intercorrelations'among industries
Some of the industries in the pilot study were highly correlated
with one anothert i.e't concentrated in the same counties.
These
correlations created difficulties in the analysis by multiple re-
gression techniques because they reduced the efficiency of estimates
and significance tests.
They also increased the possibility of
wrongly attributing increased mortality to a particular industry.
A similar problem is expected in the full-scale study.
In the
previous section a method for resolving this problem is described.
The intercorrelation of ~ndustries is a problem in any geographic
study of mortality and industrial concentration whether or not this
25
-------
is recognized or acknowledged by the authors.
As noted earlier,
some recent country mortality studies attributed increased cancer
mortality to a particular industry when it could have been equally
attributed to anyone of several other industries.
2.
Oversimplification
Statisti~al summaries of numerous complex relationships give an
incomplete picture and sometimes distort very dissimilar relation
ships, causing them to appear similar.
Plots of significant
regressions were used to flush out the summary statistics in the
pilot study and this same procedure will be used for the full scale
study.
3.
Nonlinear relationships between control variables and mortality
rates
The regression techniques used in the pilot study (and also to be
used in the full scale study) to control for nonindustrial variables
with a known or suspected relationship to mortality are convenient
for handling a large number of variables but present difficulties
with outliers and nonlinear relationships.
During the course of this study it was discovered that some of the
control variables, most notably the two age variables, were curvi-
linearly related to cancer mortality.
The problem was largely
corrected, as previously described, although a small margin of error
could not be corrected in the pilot study without arbitrarily excluding
many counties.
However, two major improvements were made for the
26
-------
B.
fullscale project as a result of this discovery:
The regression
analyses of the control variables will allow for curvilinear
relationships, and age-adjusted rates will be used.
Problems resolved in the full scale study but not in the pilot study
1.
Uncontrolled variables
Many variables frequently shown to be correlated with cancer mortality
were not considered in this pilot study although they will be included
in the full scale project.
Some uncontrolled variables include elevation,
latitude, mean precipitation and hours of sunlight, water hardness,
alcohol consumption and gasoline sales.
Other important variables were not considered because the data were
unavailable.
Cigarette smoking is one such important variable for
which data are unavailable at the county level.
Many studies which relate certain industries to cancer mortality,
including some recently reported, have no adequate control over
related variables.
In this respect, this pilot study is superior.
2.
Migration of industries
The migration of industries between counties could lead
to distortion of mortality resulting from industrial exposure.
Industrial concentration data for several points in time, beginning
in 1959, will therefore be examined for the full-scale study.
3.
Latency period
The latency period between exposure and mortality was assumed in the
pilot study to be 11 years (1959 to 1970).
27
This was the longest
-------
i
I
latency period allowable with the available data.
It is usually
assumed that the latency period for cancer is longer than this and
it doubtless varies according to the type of disease.
The full-scale study will use industrial distributions for 1959,
1967, and 1973, thus permitting analysis of three latency periods.
The best latency period can then be empirically determined for each
disease.
However, the longest latency period studied will still be
a few years short of that generally assumed for cancer because
employment data prior to 1959 are not available in the level of
detail needed for thi~ analysis.
4.
Commuting
Intercounty commuting could obscure the effects of industries on
/
mortality rates since deaths are assigned to the county of residence.
Thus deaths resulting from daytime exposures, whether occupational
or ambient, could be attributed to the wrong county when the worker
commutes across the county line.
Since these deaths are not related
to industries within the county, the relationships between mortality
and industrial concentration would be obscured.
There are two types
. of commuters for any given county:
the residents who commute outside
the county to work, and the non-residents who commute into the county
to work.
There was a much stronger relationship between industry
and cancer in those counties with a small proportion of workers
commuting outside the county than in all 200 counties considered
together.
It is anticipated that information on commuting will also
be used in the full-scale study to reveal such relationships and to
help differentiate between occupational and industrial exposure.
28
-------
5.
Age bias
The mortality data were neither age-specific nor age-adjusted.
An
attempt was made to correct this deficiency by using median age as
a control variable, but it was not adequate.
The full-scale project
will use age-adjusted rates for the age band 35-74.
6.
Race/sex aggregation
The mortality data were presented for the total population and were
not broken down by race or sex.
Since mortality rates for specific cancer
sites differ .sharply by race and sex, this seriously' clouded many
relationships.
These problems are discussed further in the next section.
the full:-scale project will use race and sex specific rates.
I
7.
Small numbers
There is a possibility that the mortality rates in the pilot study were
unstable since they were based on data for a single year in small
geographic units, and therefore sometimes based on only a small number
of deaths.
The full-scale project will use five years of data and
will delineate causes of death in greater detail.
8.
Limited disease categories
Cancer mortalities were the only health measures used.
For reasons
discussed in a report of mortality trends;. in prepar<;1t~ori.by' Orr:S,
cancer is probably not the most promising indicator of health effects due
to industry.
The full-scale project will cover all mortalities, divided
into 56 specific etiologic categories.
The need for expansion of cause-
of-death categories is further discussed in the next section.
29
-------
9.
Migration of county residents
The relationship between mortality and industrial concentration
may be obscured by migration between counties when workers are
exposed to the industry in one county but die in another county
not containing the industry.
High rates of migration into retirement
areas or out of Appalachian coal fields represent obvious examples of
the regression inaccurately reflecting'mortality as a result of
exposure to locally generated industrial pollution.
At least two
measures of migration will be used for the full-scale study to partially
correct for this factor:
net migration rate from 1960 to 1970 and per~
cent of units occupied during 1965-1970.
10.
Small population counties
The statistical methods used in the pilot study weighed each county
equally, regardless of its population.
However, because counties
with larger .populations have more stable mortality rates, they
should be given more weight to produce a more accurate estimate of
the relationship between mortality rates and other variables.
The
full-scale study will use regression analysis to weigh each county
according to the population used in computing the mortality rate.
11.
Selected counties
Since only the 200 most populous counties were studied, the results -
cannot be generalized to the entire U.S.
The full-scale project will
deal with all 3,069 U.S. counties.
12.
Selected industries
The pilot study included 30 industries which were selected by a
30
-------
method which produced a highly intercorrelated set.
Of these, only
about six industry groups were sufficiently independent so that the
effects of one could be separated from the effects of the others.
The full-scale study will screen all manufacturing, mining, trans-
portation, agricultural, forestry, and fishery industries.
They
will not be pre-selected and are expected to have lower correlations
than the industries in the pilot study.
13.
Community vs. occupational exposure
Although the pilot study was most concerned with community rather than
occuvational exposure, in practice it was difficult to distinguish the
results of one from the other.
Counties with occupationally related effects can be isolated by the
use of race and sex specific rates, the addition of other variables
such as percent commuting outside the county to work, and knowledge
from other sources about the effects of exposure in certain industries.
While these approaches were beyond the scope of the pilot study, most
have been built into the full-scale project, and all will be pursued in
that study.
C; . .Problemsnot resolved in either study
1.
Biased industrial index
The industrial index value for a specific industry in a given county
is derived by dividing the number of workers in that industry employed
in that county by the county's total workforce.
The purpose of the
ratio is to permit valid county-by-county comparisons of the relative
31
-------
significance of the size of an industry workforce, after controlling
for county size.
The denominator in the ratio is designed to avoid exaggerating the
significance of the size of an industry workforce in large, densely
populated or high employment counties or underestimating its signifi-
cance in small, sparsely populated or low employment counties.
Thus,
the index values stress those industries that are located in counties
with small workforces or are heavily concentrated in a few areas.
This bias is purposely built into the methodology to avoid the large
county bias noted above, but the present index may overcorrect for
county size.
Alternative indices were considered for the full-scale study.
The one
selected, suggested by Christoper Gordon of System Science, Inc. (551),
divides the square of the industry workforce by the total county
workforce, thus giving more weight to counties with large workforces
than the index used in the pilot study.
It will be recognized that the problem is inherent in the methodological
. approach.
In particular, the adverse health effects of those industries
that are distributed in general acccordance with the population
distribution (e.g., gasoline service stations) cannot be detected
by any index.
2.
Surrogate measures
The industrial index is obviously not a direct measure of an industry's
environmental pollution.
The requirement for an index with interindustry
32
-------
comparability, and our general ignorance regarding the true adverse
health effect potential of specific pollutants, both make such a sur~
rogate index essential.
3.
Chance findings
The probability levels reported in the pilot study applied to individual
regressions, or to multiple regressions after individual regressions
were scanned and selected to provide a strong multiple relationship.
Since so many regressions were run or scanned, one would expect to
find some "significant" at the .05 or .10 level by chance alone.
Therefore the probability levels should not be taken literally.
This
is true of all large-scale hypothesis-formulating studies.
The
probability of spurious results will be reduced in the full-scale
study but cannot be eliminated.
4.
Community exposure across county lines
People living and working in one county may be exposed to industrial
emissions from another county if, for example, the industry is near
county line or if the water supply is polluted.
Such exposures would tend to obscure the effects of industrial
exposure in the approach used in the pilot study and in the full-
scale study.
This is an unavoidable limitation of a large-scale
correlational analysis.
33
-------
'.
METHODOLOGICAL ANALYSES:
Many small studies have been done by OTS to elucidate methodological
problems.
Only a few are reported here.
Others are too technical for
general interest or are still in progress.
One (mortality trends analysis)
is the subject of a special report.
A.
Importance of race and sex specific rates
Data for race and sex specific mortalities by county were not in the
published tables upon which the pilot study was based.
These data will
be available for the full-scale study.
. There is a great difference in rankings of county mortality by sex
and by color.
Tables 7 and 8 show the 200 county ranks (200 were
ranked, although only 15 appear in each table) on the age-adjusted
respiratory cancer mortality rates in 1950-1969 for the specific race-sex
group indicated.
Table 7 shows the top 15 counties ranked on rates
for white males.
Counties similarly ranked for each of six cancer types
were used to select the 30 industries examined in the pilot study.
Table 8 contains the top 15 countie~ ranked on rates among non-white
females.
Only one county, Hudson, N.J., occurred in both ratings.
Had industries been selected from counties with high rates among
non-white females instead of white males, an entirely different set
34
-------
of counties would have been included, which may have led to a completely
different set of industries.
This problem will not occur in the
full-scale project, of course, since all industries are being screened.
The averagerankings in Table 7 are low for white males, of course,
(indicating high rates) and higher for the other three race-sex
groups (indicating lower rates).
They differ much more by race
than by sex.
Similarly the average rankings in Table 8 are low
for non-white females and also differ more by race than by sex.
This has significance for the full-scale project because it suggests
an approach for dealing with the occupational versus community
exposure problem.
Counties with high rates for both sexes of the
one or the other race may reflect differential housing patterns
and hence community exposure.
Th~y may also reflect SOIDe other
living pattern aspect such as diet.
High rates for males of both
races may be due to occupational exposure.
Charleston, South Carolina (Table 7) had the highest mortality
rates from respiratory cancer for white males and white females but.
close to the lowest for non-white groups; Northampton, Pennsylvania
. (Table 8) had just the opposite situation.
Because of the racial
difference in rates, if an industry associated with respiratory
cancer mortality were present in one of these counties it would
suggest that residential, rather than occupational, patterns may be
involved.
This information is lost when only one race-sex group
is used or when all are combined in a total rate, as was the case
in the pilot study.
These data illustrate the need for race-sex
specific rates.
35
-------
Table 7
200-County Ranking of Age Adjusted Mortality Rates
for Respiratory Cancer, 1950-1969
Top 15, Ranked by White Male
RANKING
County and State WHITE NONWHITE
Male Female Male Female
Charleston, SC 1 1 181 151
Orleans, LA 2 30.5 23.5 51
Jefferson, LA 3 26 11 143
Baltimore City,MD 4 43.5 21 87
Norfolk City, VA 5 6 108 95.5
Mobile, AL 6 26 128 155
Hudson, NJ 7. 32 9 7
Duval, FL 8 20.5 99.5 151
Middlesex, NJ 9 107 5 25
Chatham, GA 10 20.5 160.5 118
Caddo, LA 11 35.5 164.5 168.5
Harris, TX 12 4.5 85 99.5
Jefferson, TX 13 16.5 77.5 38
Prince Georges, MD 14 16.5 155 102
Hillsborough, FL 15 26 124 107
AVERAGE:
8
27.5
90.2
99.9
36
-------
Table 8
200-County Ranking of Age-Adjusted Mortality Rates for
Respiratory Cancer, For Four Race-Sex Categories 1950-1969
Top 15, Ordered by Nonwhite Female Rankings
RANKING
WHITE NONWHITE
Male Female Male Female
County and State
Northampton, PA 153.5 157.5 2 1
New London, CT 113 146.5 116 2
Winnebago, IL 190.5 136.5 36 3
Lancaster, PA 194 141 83.5 4
Cumberland, ME 61.5 107 191 5
York, PA 192.5 191. 5 87 6
Hudson, NJ 7 32 9 7
Polk, IA 76 168 12 8
Washington, PA 135 168 110 9
Will, IL 150 182 168 10
Albany, NY 23.5 89.5 23.5 11.5
Westmoreland, PA 181. 5 182 106 11.5
Sonoma, CA 155.5 68.5 167 14
Lucas, OH 47.5 107 35 14
Nueces, TX 66.5 12 70.5 14
AVERAGE:
116.5
125.9
81.1
8.0
37
-------
B.
Effects of commuting
Occupational exposure of workers who commute outside the county of
residence could obscure the effect of any industrial exposure
inside the county.
To determine the extent to which this factor
affected relationships in the pilot study analysis, the percent of
county workforce commuting outside the county was added to our data
set.
Table 9 shows results of the stepwise analysis for two groups of
counties, divided according to the percent working outside the
county.
A ten percent level was used since it divided the 200
counties nearly down the middle.
As the table shows, more industries
were significantly related to cancer mortality residuals and the
relationships for all four cancer types were generally stronger
among those counties with fewer than 10 percent of the workers
commuting.
The individual plots show that almost all the industries
related to a type of cancer for the counties with more than 10
percent commuting were also related for the counties with less than
10 percent.
They sometimes failed to be selected by stepwise
regression, evidently because of their high correlation with an
industry which was selected.
In most cases the counties with more
than 10 percent commuting had fewer and weaker significant regressions
than the 200 counties combined.
An exception was respiratory
cancer, where both the counties above and below 10 percent showed
more positive regressions on industry than did all 200 counties.
Further study is required to determine the kind of relationship
38
-------
Table 9
Comparison of Stepwise Regressions on Industry
for Counties by Percent ~f Workforce CommutinR
Less than 10% 10% or'More
Commuting Commu't ing All
Type of Industry (8ase-91 (Base~109 200 I
Cancer (Abbrevia tion) Counties) Counties) Count1e~ I
R R 1t
Total IRONMIN .33 .19
Cancer . DYES .18 .14
ORGANICS .29 .16
ASPHALT ~.24
DRUG'S .21
STLFOUND .17
All significant
industries .53 .18 .29
Respiratory .~. ~
Cancer SULFMIN .28 .17
STLMILL S .15 .13
STEVEDOR .22 .12
DOCKS .24
DRUGS .22
PETRO REF .20
PLASTICS .17
DRESSES -.23 -:-19
COPSMELT .11 I
,CIGARS -.30
FRTLIZER -.17
All significant
industries .48 .49 .33
Positive R only .43 .32 .27
Digestive
Cancer IllONMIN .23 .25
GYPSUM .17
DYES .17 .13
WIRE .12
ASPHALT .40
STLMILLS .23
COALTAR. -.24 .16
lRTLIZER -.23 -.14
BROADFAS .25
STEVEDOR .15
COPSMELT .21
All significant
industries .68 .31 .36
Positive R on1y .52 .31 .33
Urinary
Cancer . ANTHMIN .35 -.19
GYPSUM .23.
BLOUSES -.22
BEER .20
SULFMIN .16
. PETROMIN -.17
All significant
industries .46 '-.25 .00
Positive R only .35 .00 . .
39
-------
being summarized in these statistics.
However, for the other-
cancer types, there was greater statistical association between
industry concentration and cancer mortality on a county basis
when more than 90 percent of the workers were employed in their
county of residence.
It is anticipated that this variable will be
used in the full-scale project to separate industrial from community
exposure.
c.
Problem of negative relationships
The initial 200 county stepwise regression of residual cancer
mortality on the 30 industries produced two significant negative
regressions: respiratory cancer residuals on dress manufacturing,
and digestive cancer residuals on fertilizer manufacturing.
Scatter plots of these two relationships are shown in Figures 2 and
3.
In Figure 2, all counties appearing in the lowest 10 percent of the
200 counties on respiratory residuals and in the top 10 percent on
dress manufacturing are indicated.
Only two of these, Lackawanna
and Luzerne, Pennsylvania, accounted for the significant negative
relationship; without them, there would be no significant relationship
for the two variables.
In Figure 3, Polk, Florida, similarly
accounts for the negative relationship.
It is interesting to note that the actual cancer ~ortality rates in
Lackawanna and Luzerne, Pennsylvania, which are coal mining as well
40
-------
. --~- 0-
. --0._0_0 -. .., o__.-'-'O__.._"'O~- ..
I
I .
Figure Z
PLOT OF RESPIRATORY CANCER RESIDUAL BY CONCENTRATION OF WOKEN'S DRESS MANUFACTURE
.'
12.1)000"110/' . \
I ' .
, - A
Q . I
I t,
I :. /I
I 1\
I to II
..... I ~ A
"0 I Q 1\
u
.. !t.OOOOoH)OO . r. ~ II A
\J
~ "
;c f\ AliA II
...
"
'":J r
u ~ II
>
.. r. 1\
-------
20.01)00'/0011 .
I
I
I
I'
I
I
, I
I
I
..-
...
u
..
u
g, 12.0(001)00'1 .'
I(
...
. '
":J
:.0
>
..
:.0
..
..:I
.8.
!:
-0
4.0/)00/10/)1\ .
I
I
I,
I
I
I
I
I
I
..
:.0
.-:-..::
N..
'J
U
C
t)
U
>-
..
a
~
..
..
~ -4.0/)000001) .
o
'"
,-12.00001101)1) .
'..'. . . '.". "
~ "--~ --.-
"--....-----'--'--'-' .... .
o
Pol k, Florida
1---------.-------------------.-----------------------~---------------.--~----------------._----------------.-.-----
Figure 3
PLOT OF OICESTIVE CANCER RESIDUAL BY CONCEUTRATION OF FERTILIZER HANUF~CTURE
4.
II
A' .
a 0
q
A
1\
o
a II
k II
o
C
L
OAA
. !(
M
L
.J A
A
4
A
4
,..
II
A
.J
\.14
G
.J
.14 I.
J
r,A
C
A
A
R
A
4
A
o
Hlllsborough, Florida
..
o
Chesapea~e, Virginia
A
A
o
Ingham, Mich.
o
~ullford, N. Carol.
-4/)1). (;13(1""
! ,1"0. ""'100'
. ;'>"'lfl.I)Onnn
4 "I),). oon"o
1'000,')'.00000
. LEr.r~JI): II = I 11.'<;
. '"' = ;. nit" .
. 'r..
Concentration In~e~:
Fertilizer Manufacture
Hllo.oonoo
-------
as dress manufacturing counties, were so far below the predicted
level.
Coal mining also had a negative relationship with respiratory
cancer, although it was not selected by the stepwise procedure; its
negative relationship was also accounted for by these two counties.
Competing causes could account for this.
Table 10 shows the percent of all mortalities attributed to certain
causes for those counties in Figures 2 and 3; for the States in
which they are located; for all metropolitan counties in the U.S.;
and for the U.S. as a whole.
A high percentage of deaths in Lackawanna and Luzerne were in the
residual disease category.
This heterogeneous category contains,
among other causes, anthracosilicosis (lCD 515.1) and anthracosis
(a subcategory of lCD 516); many of the deaths in this category
were probably due to these two diseases.
Unfortunately a more
specific category is not available by county.
The residual respiratory
category is close, but it is available only by state.
Pennsylvania
had a rate nearly three times that of the U.S. for this category
although its rate for the broader disease category was only slightly
higher than for the U.S.
These data support the hypothesis that the low respiratory cancer
mortality rates in the counties which account for the negative
regression between respiratory cancer and dress manufacturing, were
due in part to a competing respiratory disease.
43
-------
Table 10
Percent Distribution of Deaths by Cause, 19701
U.8., -Counties Contributln~ to Nep,at!ve Stepwise Regression. and Their States
Area Hal1gnant Dinbetes Major Selected Selected Residual Diseases
Neoplasms Hcllitus Cardiovasf;ular Respiratory Digestive of 34 Cnuses
Diseases Diseases and Urinary Respiratory
Diseases AU:
~40-209 250 390-448 470-474 571 501-508
480-486 580-584 512
490-493 514-516
519
United States 17.2 2.0 52.5 4.9 2.1 0.6 7.2
Metropolitan Counties 18.0 2.0 51.7 4.8 2.4 7.3
Lackavonna, PA 16.2 2.6 56.8 3.2 2.0 10.8
Luzcrne, PA 15.8 2.8 52.2 2.8 2.0 17.2
Cambria, PA 15.8 2.1 57.2 4.3 1.5 10.2
York, PA 17.8 2.2 54.5 4.7 1.3 8.7
Pennsylvania 1-7.3 ' 2.2 54.9 4.0 1.9 1.4 8.0
Richlond, SC 14.7 3.2 50.6 5.8 1.9 --- 6.9
South Carolina 13.7 . 2.3 50.7 4.8 1.8 0.5 6.9
~
~ 46.7
Polk. FL 17.1 -1.5 6.1 1.4 7.5
Florida 18.3 1.7 50.7 5.2 2.0 0.5 7.2
.. ..
Ingham, MI 17.5 3.4 48.5 5.4 1.4 7.3
~Iichlgan 17.6 2.9 51.5 4.6 2.4 0.3 6.7
Chesapeake, VA 16.1 1.3 49.5 5.0 2.5 8.0
, Virginia 16.2 1.5 51.0 5.2 1.9 0.6 '.8
Guilford, NC 16.5 2.1 .48.4 5.1 2.3 8.5
North Carolina 14.6 1.9 . 50.9 5.2 1.9 0.5 7.6
Source: Vital Statistics of the United States,' 1970 . Respiratory Rcsidual column: Table 1-27 (U.S. and
States onlY)inll other columns: Table 7-9
-------
No such pattern was found for Polk, Florida.
However, Polk had an
unusual age distribution.
,
It ranked 43rd from the highest median
age of the 200 counties but 15th in percent aged 65 and over.
As
noted earlier, controlling for differences in age distribution by
regression analysis is much less effective than using age-adjusted
rates.
Both negative relationships may disappear in the full-scale
study.
These data illustrate the need to use age-adjusted rates and to
include diseases other than cancer in the full-scale study.
45 .
------- |