The Fallacy of the Ecological Fallacy: The Potential Misuse of a Concept and the Consequences
Abstract
Ecological studies have been evaluated in epidemiological contexts in terms of the "ecological fallacy." Although the empirical evidence for a lack of comparability between correlations derived from ecological- and individual-level analyses is compelling, the conceptual meaning of the ecological fallacy remains problematic. This paper argues that issues in cross-level inference can be usefully conceptualized as validity problems, problems not peculiar to ecological-level analyses. Such an approach increases the recognition of both potential inference problems in individual-level studies and the unique contributions of ecological variables. This, in turn, expands the terrain for the location of causes for disease and interventions to improve the public's health.
The
Fallacy
of
the
Ecological
Fallacy:
The
Potential
Misuse
of
a
Concept
and
the
Consequences
Sharon
Schwartz,
PhD
A
Proposed
Validity
Scheme
Epidemiology
texts
offer
a
consistent
appraisal
of
ecological
studies:
they
are
crude
attempts
to
ascertain
individual-
level
correlations.
The
flaws
in
such
studies
limit
their
usefulness
to
"hypoth-
esis
generation,"
leaving
the
more
es-
teemed
process
of
"hypothesis
testing"
to
individual-level
data.
The
problems
are
generally
attributed
to
the
"ecological
fallacy,"'14
a
logical
fallacy
inherent
in
making
causal
inferences
from
group
data
to
individual
behaviors.9"10
The
consequences
of
this
ecological
fallacy
are
well
documented.
More
than
40
years
ago,
Robinson
demonstrated
that
the
correlation
coefficient
between
two
individual-level
variables
is
generally
not
the
same
as
that
between
those
same
variables
for
aggregates
into
which
the
individuals
are
grouped.'1"2
Many
papers
have
examined
this
problem
statistically,
confirmed
Robinson's
findings,
and
sug-
gested
methods
for
making
ecological
and
individual
correlations
more
compa-
rable.1-20
Epidemiology
texts
have
used
these
analyses
to
support
their
evaluation
of
ecological
studies.
The
use
of
the
ecological
fallacy
to
explain
the
discrepancy
between
indi-
vidual
and
ecological
correlations
may
have
unintended
consequences.
Examin-
ing
this
issue
from
a
different
perspec-
tive-as
a
general
validity
problem-will
show
that
the
ecological
fallacy,
as
often
used,
encourages
three
interrelated,
falla-
cious
notions:
(1)
that
individual-level
models
are
more
perfectly
specified
than
ecological-level
models,
(2)
that
ecologi-
cal
correlations
are
always
substitutes
for
individual-level
correlations,
and
(3)
that
group-level
variables
do
not
cause
dis-
ease.
We
begin
with
a
description
of
the
validity
framework
and
the
definition
of
key
terms.
Cook
and
Campbell
developed
an
analytic
scheme
to
assess
the
validity
of
causal
relationships.2'
Two
questions
they
pose
are
of
salience
here:
(1)
Given
a
statistically
significant
correlation
be-
tween
two
variables,
is
it
valid
to
assert
a
causal
relationship
between
these
two
variables
as
measured?
(2)
Given
a
plausi-
bly
causal
relationship
between
two
vari-
ables
as
measured,
what
are
the
causal
concepts
involved
in
the
relationship?
The
first
question
deals
with
internal
validity
and
the
second
with
construct
validity.
Internal
Validity
The
essence
of
internal
validity
is
accounting
for
third-variable
alternative
interpretations
of
presumed
A-B
relation-
ships
in
which
A
represents
the
indepen-
dent
variable
and
B
the
dependent
vari-
able.
It
is
precisely
here
that
a
source
of
noncomparability
between
an
individual
and
an
aggregate
correlation
of
the
same
variables
may
arise.
"In
shifting
from
one
unit
of
analysis
to
another,
we
are
very
likely
to
affect
the
manner
in
which
outside
and
possibly
disturbing
influences
are
operating
on
the
dependent
and
independent
variables."22(P97)
As
a
result
of
the
grouping
operation,
one
may
have
The
author
is
with
the
Division
of
Epidemiol-
ogy,
School
of
Public
Health,
Columbia
Univer-
sity,
New
York,
NY.
Requests
for
reprints
should
be
sent
to
Sharon
Schwartz,
PhD,
Columbia
University,
School
of
Public
Health,
600
W
168th
St,
7th
Floor,
PH18-332:
Epidemiology,
New
York,
NY
10032.
This
paper
was
accepted
January
14,
1994.
Editor's
Note.
See
related
articles
by
Susser
(p
825
and
p
830)
and
Koopman
and
Longini
(p
836)
and
editorial
by
Poole
(p
715)
in
this
issue.
American
Journal
of
Public
Health
819
Introduction
controlled
for
the
effects
of
other
vari-
ables,
making
the
ecological
estimate
less
biased
than
the
individual
estimate,23
or
one
may
have
included
various
confound-
ing
variables,
making
the
ecological-level
correlation
more
biased.24
If
a
difference
occurs
between
ecological-
and
individual-
level
correlations,
the
problem
may
be
due
to
a
failure
to
specify
the
correct
model
and
not
to
an
inherent
logical
fallacy
in
moving
from
individual
to
group
correlations.
Construct
Validity*
However,
discrepancies
between
indi-
vidual
and
ecological
correlations
often
remain
after
controlling
for
confounding
variables.2
To
some
extent,
this
may
be
due
to
further
misspecifications,
where
other
confounding
variables
are
not
taken
into
account.
But
there
may
be
another
problem
as
well.
"The
demystification
of
cross-level
bias
begins
with
the
recogni-
tion
that
an
aggregate
variable
often
measures
a
different
construct
than
its
name-sake
at
the
individual
level."25(P560)
The
construct
referenced
on
the
ecologi-
cal
level
may
be
the
context
or
social
environment
in
which
individuals
live,
distinct
from
the
attributes
of
those
individuals.26-30
Thus,
poverty
as
an
indi-
vidual
characteristic
and
poverty
as
a
neighborhood
characteristic
may
exert
different,
independent
effects
on
health.
Consequently,
individual
and
aggregate
correlations
of
this
variable
will
be
discrep-
ant.
Internal
validity
and
construct
valid-
ity
can
therefore
be
used
to
explain
disparities
in
correlations
between
indi-
vidual-
and
ecological-level
variables.
We
examine
the
benefits
of
doing
so
through
a
discussion
of
three
assumptions
associ-
ated
with
the
ecological
fallacy.
Evaluations
of
Ecological
Studies
Based
on
the
Ecolgical
Fally:
Hidden
Assumptions
and
Their
Conseqences
Assunption
1:
Individual-Level
ModelsAre
More
Perfectly
Specified
Than
Ecolocal-Level
Models
The
problem
of
internal
validity,
confounding,
is
considered
a
particularly
egregious
fault
in
ecological
studies.l.49
Indeed,
the
ecological
fallacy
is
often
defined
as
a
problem
of
confounding.
For
example,
Lillienfeld
and
Lillienfeld
con-
tend
that
ecological
correlations
"may
suffer
from
an
'ecological
fallacy',
that
is,
the
two
communities
differ
in
many
other
factors
and
one
or
more
of
those
may
be
the
underlying
reason
for
differences
in
their
observed
mortality
or
morbidity."3(p$8)
But
to
conclude
that
differences
in
rel-
evant
third-variable
effects
at
the
ecologi-
cal-
and
individual-levels
of
the
same
variable
constitute
an
ecological
fallacy,
a
weakness
in
ecological
studies,
requires
one
to
assume
that
individual-level
mod-
els
are
more
accurately
specified
than
ecological-level
models.
This
is
often,
but
not
inherently,
true.
If
individual-
and
ecological-level
analyses
are
both
based
on
historical
records,
information
neces-
sary
for
including
confounding
variables
may
be
extant
for
the
aggregate
but
not
for
the
individual
level,
allowing
better
specification
of
the
ecological
study.
For
example,
employee
records
may
have
less
information
on
smoking
and
dietary
hab-
its
than
sales
records
for
company
towns.
Similarly,
certain
confounding
variables
intrinsic
to
survey
research,
such
as
response
bias,
recall
bias,
and
naysaying,
may
be
avoided
in
ecological
studies.
In
particular,
when
the
variables
of
interest
probe
sensitive
issues,
ecological-level
data
may
be
more
accurate.
(E.g.,
sales
of
alcoholic
beverages
or
rates
of
abortions
may
be
more
useful
than
statements
of
alcohol
use
or
an
individual's
abortion
experience.)
Additionally,
the
grouping
process
itself
may
control
for
some
con-
founding
variables
not
controlled
for
in
an
individual-level
model.9'31'32
In
practice,
it
may
be
that
confound-
ing
usually
poses
a
more
intractable
problem
for
ecological-
than
for
individual-
level
studies.
But
this
is
due
to
the
greater
reliance
on
secondary
data
and
proxy
measures
in
ecological
studies,
not
to
any
problem
inherent
in
ecological
studies.
An
inability
to
control
confounding
vari-
ables
occurs
under
these
conditions,
no
matter
what
the
unit
of
analysis.
However,
the
view
that
ecological
studies
may
be
used
for
only
hypothesis
generation
or
evaluation
of
interventions,
while
not
generally
supportable,
is
valid
under
cer-
tain
conditions.
If
it
is
suspected,
in
a
specific
instance,
that
an
ecological
corre-
lation
will
yield
a
biased
estimate
of
an
individual
correlation
that
is
perfectly
specified,
due
solely
to
problems
of
internal
validity
(i.e.,
problems
of
con-
founding
variables),
the
solution
to
the
problem
would
be
a
careful
respecifica-
tion
of
the
model.
If
the
potentially
confounding
factors
are
unknown
or
unmeasured,
the
ecological
correlation
is
useless.
In
such
a
case,
the
ecological
correlation
is
merely
a
poor
substitute
for
an
individual-level
correlation.
Although
it
may
yield
some
hypotheses
for
explora-
tion,
it
will
be
of
little
real
help
because
confounding
may
alter
not
only
the
magni-
tude
of
the
correlation
coefficient
(or
other
measure
of
association)
but
the
direction
of
the
effect
as
well.
Seeing
this
confounding
problem
in
terms
of
intemal
validity
raises
a
number
of
questions
to
be
evaluated
on
a
case-by-
case
basis.
Are
internal
validity
problems
the
only
possible
source
of
discrepancy
between
a
particular
ecological
correla-
tion
and
a
correlation
of
these
same
variables
at
the
individual
level?
If
so,
what
are
the
sources
of
confounding
that
are
likely
to
be
problematic
at
each
level?
At
what
level
of
analysis
can
these
confounding
problems
best
be
controlled?
The
answers
will
not
always
favor
the
individual-level
study.
Assumption
2:
Ecological
ModelsAre
Substitutes
for
Individual-Level
Models
As
we
have
seen,
ecological
and
individual
correlations
may
be
discrepant
not
only
because
of
internal
validity
problems
but
also
because
of
construct
validity
problems.
That
is,
the
aggregated
variable
may
measure
a
different
con-
struct
than
its
namesake
on
the
individual
level.
This
source
of
discrepancy
was
not
mentioned
in
any
of
the
epidemiological
textbooks
reviewedl-5,l0**
despite
consid-
erable
discussion
in
other
fields.253233
The
reason
for
this
lies,
perhaps,
with
an
assumption
Robinson
makes:
that
re-
searchers
undertake
ecological
studies
only
when
individual-level
data
are
un-
available
and
that
the
individual-level
analysis
is
their
real
concern."1
This
assumption
is
accepted
in
the
main
epide-
miological
texts
and
is
implicit
in
discus-
sions
of
the
ecological
fallacy.2-10
For
example,
Morgenstern
writes:
The
key
feature
of
ecological
data
relative
to
cohort
data
is
the
lack
of
information
about
the
joint
distribution
of
the
study
factor
and
the
disease
within
each
group....
In
ecologic
analy-
*This
use
of
construct
validity
is
an
expansion
of
the
concept
as
developed
by
Cook
and
Campbell,21
who
explicitly
argue
that
it
refers
only
to
constructs
at
the
same
level
of
reduc-
tion.
An
analysis
of
Cook
and
Campbell's
position
and
the
development
of
the
reinterpre-
tation
used
here
are
available
from
the
author.
**While
none
of
the
epidemiological
texts
refer
to
construct
validity,
Morgenstern's
article
evaluating
ecological
studies
in
epidemiology
does.9
However,
Morgenstern
views
macroso-
cial
or
contextual
effects
only
as
confounding
variables
and
not
as
causal
variables
in
their
own
right.
820
American
Journal
of
Public
Health
May
1994,
Vol.
84,
No.
5
Ecolical
Falacy
sis
the
independent
variable
(X)
is
the
proportion
of
exposed
subjects
within
the
group
and
the
dependent
variable
(Y)
is
the
rate
(or
risk)
of
disease.9(P'337)
This
is
sometimes
the
case
but
only
if
the
ecological
variable
is
an
aggregate
vari-
able
rather
than
a
characteristic
of
a
group
and
if
there
are
no
contextual
effects-that
is,
only
when
the
ecological-
and
individual-level
variables
measure
the
same
construct.
But
when
the
ecological-
level
variable
measures
some
group
prop-
erty,
it
is
no
longer
the
proportion
of
exposed
subjects
that
is
the
independent
variable.2630
Rather,
the
proportion
of
subjects
with
a
certain
factor
of
interest
is
itself
the
exposure.
In
this
case,
the
ecological
study
is
not a
substitute
for
an
individual-level
study
but
an
examination
of
unique
variables
not
measurable
on
the
individual
level.3435
The
neglect
of
this
possibility
leads
to
two
interrelated
prob-
lems:
(1)
a
failure
to
recognize
the
ecological
fallacy
in
individual-level
stud-
ies,
and
(2)
a
failure
to
recognize
the
full
range
of
cross-level
studies.
Ecological
fallacies
in
individual-level
studies.
Epidemiological
discussions
frame
the
issue
of
ecological
inference
problems
in
terms
of
a
lack
of
consistency
between
the
measure
of
association
for
the
inde-
pendent
(A)
and
dependent
(B)
variables
at
the
individual
level
and
the
measure
of
association
for
the
independent
(A')
and
dependent
(B')
variables
at
the
ecological
level.
Therefore,
the
measure
of
associa-
tion
is
the
focus
of
analysis.
But
when
the
issue
is
framed
in
terms
of
construct
validity,
it
becomes
apparent
that
there
are
two
other
points
of
potential
disagree-
ment
between
ecological-
and
individual-
level
correlations.
A
may
not
equal
A'
and
B
may
not
equal
B'.
This
conceptualiza-
tion
helps
to
clarify
the
logical
fallacy
involved
in
cross-level
inference
and
al-
lows
one
to
think
more
fully
about
levels
of
analysis
and
their
relationships.
As
a
logical
fallacy,
the
ecological
fallacy
is
a
problem
of
construct
validity
and
not
of
a
measure
of
association.
Aristotle
refers
to
it
as
"the
fallacy
of
division."36
It
is
a
problem
of
confusing
the-group
with
the
members
of
that
group,
of
assuming
that
because
a
group
has
a
certain
characteristic
the
members
of
that
group
also
have
that
characteristic.
Zito37
provides
an
illuminative
example.
A
hung
jury
is
a
jury
that
is
indecisive,
it
cannot
decide
whether
the
accused
is
guilty
or
innocent.
However,
to
deduce
that
the
members
of
such
a
jury
are
indecisive
would
be
absurd.
Members
of
a
hung
jury
are
very
decisive,
so
much
so
that
they
can
not
be
persuaded
to
change
their
mind.
Attributing
to
the
members
of
this
group
the
characteristic
of
that
group
(indecisive-
ness)
is
thus
a
case
of
the
ecological
fallacy.
A
construct
validity
approach
raises
the
awareness
that
the
ecological
fallacy
is
a
ubiquitous
problem
and
may
occur
with
individual
as
well
as
ecological-
level
data.
Note,
for
example,
the
following
comment
in
one
epidemiological
text:
"In
most
epidemiologic
contexts
as
opposed
to
sociologic
or
anthropologic
contexts,
one
is
interested
in
drawing
inferences
about
disease
etiology
in
individual
per-
sons
[emphasis
addedJ."2(P4
)
But
epidemi-
ology
is
not
concerned
with
disease
etiology
in
individual
persons.
As
defined
by
Susser
et
al.,
"Epidemiology
('epi'
upon,
'demos'
the
people)
is
the
science
concerned
with
the
health
of
populations
or
communities
[emphasis
added]."'39(Pl6)
Indeed,
the
empirical
analysis
of
sample
data,
whether
it
is
aggregate
or
individual,
cannot
be
used
to
study
the
behavior
of
individuals.
The
objective
of
most
em-
pirical
analyses
is
to
determine
the
independent
effects,
in
a
probabilstic
way,
of
some
households
or
individuals
possessing
that
characteristic
[emphasis
added].24(P")
For
example,
if
an
experimental
vaccine
trial
provided
evidence
that
20%
of
the
vaccinated
and
50%
of
the
unvaccinated
people
contracted
the
disease,
one
would
conclude
that
there
is
an
association
between
getting
the
vaccine
and
not
getting
the
disease.
In
fact,
one
would
conclude
that
there
is
probably
a
causal
relationship
between
avoiding
the
disease
and
being
vaccinated.
Yet
for
any
particu-
lar
vaccinated
person
it
would
be
a
logical
fallacy-indeed,
an
ecological
fallacy-to
suggest
from
these
data
alone
that
he
or
she
did
not
contract
the
disease
because
of
the
vaccination.39
This
logical
fallacy
is
ubiquitous
when
proxy
measures
are
used.
For
example,
in
an
individual
study
of
the
relationship
between
an
exposure
and
disease,
controlling
for
diet
may
be
desir-
able.
However,
data
on
dietary
habits
may
be
difficult
to
obtain,
and
another
variable
collected
for
each
individual-perhaps
educational
level-may
be
used
as
a
proxy
measure
for
diet.
Doing
so
involves
the
ecological
fallacy,
however,
because
it
implicitly
assumes
that
since,
as
a
group,
people
with
different
educational
levels
exhibit
different
dietary
habits,
an
indi-
vidual
within
a
specific
educational
group
will
exhibit
the
dietary
pattern
of
that
group.
This
is
particularly
problematic
because
it
leads
to
significant
measure-
ment
error
and
therefore
to
an
underad-
justment
for
this
control
variable.
Thus,
one
may
erroneously
conclude
that
the
exposure
is
associated
with
the
disease
controlling
for
diet
when,
in
fact,
diet
has
not
been
controlled.
Examining
the
eco-
logical
fallacy
in
terms
of
construct
valid-
ity
has
the
advantage
of
increased
vigi-
lance
in
the
search
for
greater
validity
in
all
studies.
The
ftdl
range
of
cross-level
studies.
Viewing
ecological
studies
as
substitutes
for
individual-level
studies
leads
to
an-
other
consequence.
There
is
a
tendency
to
dichotomize
studies
as
either
ecological,
in
which
case
the
independent
and
depen-
dent
variables
are
aggregated
individual-
level
variables,
or
nonecological,
in
which
case
the
independent
and
dependent
variables
are
individual
level.
For
ex-
ample,
the
usefulness
of
ecological
studies
has
been
limited
as
follows:
"If
broad
social
or
cultural
processes
are
of
interest
then
the
individual
may
not
be
the
most
appropriate
unit
of
analysis,
since
infer-
ences
are
to
be
drawn
about
whole
societies
rather
than
about
individ-
uals."2(Pm)
Thus,
there
is
a
general
conclu-
sion
that
ecological
studies
cannot
be
used
to
make
inferences
about
individual
phe-
nomena
or
behaviors.Z10
But
social
and
cultural
factors
and
processes
do
not
have
effects
solely
on
whole
societies
but
on
individuals
as
well.
By
analyzing
ecologi-
cal
studies
in
terms
of
an
ecological
fallacy-a
problem
in
measures
of
associa-
tion
rather
than
in
terms
of
the
construct
validity
of
component
variables-the
full
range
of
potential
cross-level
relation-
ships
is
attenuated.
For
example,
both
the
independent
and
dependent
variables
in
a
study
can
be
group
characteristics
that
cannot
be
measured
by
the
aggregation
of
individual
behaviors
(e.g.,
the
relationship
between
level
of
industrialization
and
number
of
hospitals
per
capita).
A
second
possibility
is
a
study
of
contextual
effects
in
which
the
focus
of
interest
is
the
relationship
between
an
individual's
be-
havior
and
the
group
context
in
which
that
behavior
exists.
In
this
instance,
the
independent
variable
may
be
-a
group-
level
variable
and
the
dependent
variable
may
be
an
individual-level
one
(e.g.,
the
effects
of
living
in
a
minority
neighbor-
hood
on
infant
mortality).
There
can
also
be
structural
analyses
that
focus
on
the
group
but
make
reference
to
differenti-
ated
roles
of
individuals
that
interrelate
to
form
a
group's
internal
structure
(e.g.,
an
examination
of
the
social
network
pat-
American
Journal
of
Public
Health
821
May
1994,
Vol.
84,
No.
5
Schwartz
tems
of
immune
and
vulnerable
individu-
als
that
potentiate
herd
immunity).33A40
In
this
case,
the
independent
variable
is
individual
level
and
the
dependent
vari-
able
is
group
level.
Thus,
there
are
many
study
designs
that
are
neither
purely
ecological
nor
purely
individual
level.
One
example
of
the
consequence
of
viewing
studies
as
acting
on
only
one
level
is
the
analysis
of
Durkheim's
Suicide4l
in
epidemiological
contexts.
We
examine
this
example
closely
because
it
is
often
cited
as
the
exemplar
of
the
ecological
fallacy.9'10
One
analysis
in
an
epidemiol-
ogy
textbook
is
as
follows:
He
[Durkheim]
found,
on
the
average,
[that]
provinces
with
greater
propor-
tions
of
Protestants
had
higher
suicide
rates
and
that
provinces
with
greater
proportions
of
Catholics
had
lower
suicide
rates.
Durkheim
concluded
from
these
data
that
Protestants
are
more
likely
to
commit
suicide
than
are
Catho-
lics.
While
the
conclusion
may
be
true,
the
causal
inference
is
not
logically
correct,
because
it
may
have
been
Catholics
in
predominantly
Protestant
provinces
who
were
taking
their
own
lives.
This
logical
flaw,
called
the
ecologi-
cal
fallacy
(Selvin,
1958),
results
from
making
a
causal
inference
about
an
individual
phenomenon
or
process
(e.g.,
suicide)
on
the
basis
of
observations
of
groups.10(P9)
In
faimess
to
Durkheim,
it
should
be
noted
that
he
based
his
conclusions
on
ecological-level
correlations
in
tandem
with
an
examination of
suicide
rates
among
Catholic
and
Protestant
individu-
als
within
provinces.4142*
However,
it
is
worthwhile
to
examine
this
ecological
fallacy
with
the
assumption
that
these
facts
were
correct.
In
assessing
plausible
alternative
ex-
planations,
Morgenstern
suggests
that
minority
status
may
be
related
to
a
propensity
for
suicide.9
It
may
be
that
the
higher
suicide
rates
in
Protestant
coun-
tries
are
accounted
for
by
the
suicides
of
the
Catholics
who
have
minority-group
status
in
such
places.
Regardless
of
the
merits
of
this
hypothesis,
it
could
not
be
*The
authors
quote
the
Selvin
article
on
Durkheim
as
the
source
for
the
term
ecological
fallaCy.41
Indeed,
this
does
appear
to
be
the
first
use
of
the
term
in
the
literature,
although
MenzelM
and
Thorndike43
both
referred
to
the
ecological
correlation
problem
as
a
fallacy
prior
to
this
and
many
sources
erroneously
cite
Robinson
as
coining
that
expression.5,27,28
How-
ever,
while
Selvin
does
say
that
Durkheim
is
at
points
guilty
of
the
ecological
fallacy,
he
explicitly
states
that
Durkheim
recognized
this
problem
and
solved
it
by
looking
at
individual-
level
data
when
he
could-for
example,
in
the
religion
issue.41(P6O8)
tested
by
an
individual-level
study.
A
comparison
of
suicide
rates
among
indi-
viduals
of
different
religious
persuasions
could
reveal
only
a
higher
or
lower
rate
for
Catholics
versus
Protestants.
Only
in
conjunction
with
the
aggregate
variable
of
"proportion
Catholic"
and
"proportion
Protestant"-that
is,
only
in
conjunction
with
contextual
analysis-could
this
hy-
pothesis
be
tested.
Furthermore,
accord-
ing
to
this
definition,
Morgenstem's
alter-
native
hypothesis
would
also
constitute
an
ecological
fallacy.9
Causal
inferences
about
an
individual
process,
suicide,
would
be
made
from
observations
of
groups.
Thus,
this
alternative
hypothesis
also
suggests
that
a
group
variable
influences
behaviors
carried
out
by
individuals.
Indeed,
Durkheim
contends
that
sui-
cide
is
a
social
rather
than
an
individual
phenomenon,44
for
although
it
is
an
act
committed
by
an
individual
with
idiosyn-
cratic
reasons
for
its
commission,
varia-
tions
in
suicide
rates
are
caused
by
social
factors.
In
this
case,
Durkheim
posits
the
effect
of
living
in
a
Protestant
area
as
a
sociological
phenomenon,
related
to
the
rules
goveming
attitudes
and
behaviors
that
influence
the
propensity
to
commit
suicide.
Examining
Durkheim's
study
in
terms
of
construct
validity
makes
it
clear
that
an
ecological
fallacy
would
exist
if
it
were
assumed
that
the
variable
mea-
sured-living
in
a
Protestant
country-
was
equivalent
to
the
individual
variable-
being
Protestant.
There
is
no
ecological
fallacy
in
relating
observations
of
groups
to
behaviors
performed
by
individuals.
Thus,
while
neither
studies
of
groups
of
groups
nor
studies
of
groups
of
individuals
can
explain
the
behavior
of
a
particular
individual,
they
can
both
help
to
explain
behaviors
performed
by
individuals.
Assumption
3:
Only
Characteristics
of
Individuals
Cause
Disease
The
use
of
the
ecological
fallacy
in
epidemiology
also
fosters
a
dismissal
of
social
variables
as
causal
factors
in
dis-
ease.
First,
as
seen
above,
it
leads
to
a
consignment
of
sociological
and
anthropo-
logical
studies
to
examining
impacts
on
whole
societies,2
and
it
denies
that
ecologi-
cal
variables
can
affect
individual
pro-
cesses.28
Second,
it
reinforces
an
assump-
tion
that
aggregated
variables
are
substitutes
for
individual-level
variables.
Under
such
an
assumption,
the
potential
etiological
influence
of
aggregate-level
variables,
distinct
from
the
effects
of
the
same
measures
on
an
individual
level,
would
not
be
considered.
Usually,
this
assumption
is
implicit
in
statements
about
not
making
causal
inferences
about
indi-
vidual
phenomena
on
the
basis
of
observa-
tions
of
groups.
Sometimes,
however,
this
denial
is
made
explicit,
as
in
Rothman's
statement
that
"social
class
...
itself
is
presumably
causally
related
to
few
if
any
diseases
but
is
a
correlate
of
many
causes
of
disease.4(P90)
This
evaluation
is
perplexing
because
the
concept
of
cause
in
epidemiology
does
not
preclude
and
often
times
explicitly
includes
non-individual-level
variables.
As
Susser
suggests,
"A
determinant
can
be
any
factor
...
[that]
brings
about
change
for
better
or
worse
in
a
health
condition."45(P3)
Such
social
factors
as
socioeconomic
status
and
social
disorgani-
zation
surely
lie
within
the
purview
of
this
definition.
It
may
be
that
the
prominence
of
the
"germ
theory"
paradigm
has
reinforced
a
focus
on
individual
factors.
While
this
model
is
clearly
useful,
it
has
limitations
because
"some
health
problems
may
be
more
parsimoniously
understood
and
more
efficiently
controlled
by
viewing
them
as
products
of
community
dynamics."35(P11)
For
example,
decreas-
ing
economic
disparity
may
decrease
the
rates
of
a
wide
range
of
physical
and
psychiatric
disorders.
The
perception
of
a
longer
and
more
indirect
chain
of
causation
for
social,
ecological-level
variables
may
also
prompt
a
neglect
of
such
factors.45
But
the
length
and
complexity
of
the
causal
chain
does
not
determine
the
importance
of
the
cause.
For
all
variables,
"behind
the
'intimate'
cause
of
disease-disease
agent,
stands
the
ultimate
causal
factors
of
the
social
and
physical
environment
providing
the
linkages
between
agent
and
host.46(P11)
The
"intimate"
and
"ultimate"
causes
each
deserve
attention,
and
neither
ne-
gate
the
validity
of
the
other.47
Indeed,
"the
idea
of
cause
has
become
meaning-
less
other
than
as
a
convenient
designa-
tion
for
the
point
in
the
chain
of
event
sequences
at
which
intervention
is
most
practical.""48(Pl8l)
No
matter
what
the
variable
of
interest
or
the
level
of
analysis,
unless
cause
is
viewed
in
terms
of
a
particular
purpose,
the
problem
of
infinite
regress
ensues.
As
Zito
writes:
When
we
begin
to
supply
intervening
and
antecedent
variables
to
a
model
...
there
is
no
end
to
such
a
series
of
questions.
We
can
continuously
de-
scend
to
lower
and
lower
orders
of
questioning
and
higher
and
higher
levels
of
abstraction.37(P143)
By
viewing
discrepancies
between
ecological
and
individual
studies
as
valid-
822
American
Journal
of
Public
Health
May
1994,
Vol.
84,
No.
5
ity
issues,
it
becomes
apparent
that
all
causes
are
indirect
and
that
all
variables
can
be
viewed
as
either
ecological
or
individual
level,
depending
on
one's
per-
spective.
Robinson
contended
that,
unlike
social
groups,
individual
persons
consti-
tute
an
indivisible
entity."
But
this
is
true
only
if
one
views
the
individual
as
the
level
of
analytic
interest.
From
another
view-
point,
the
individual
is
an
ecological-level
variable,
an
aggregated
measure
of
body
parts
that
become
diseased.
For
example,
smoking,
an
activity
performed
by
individu-
als,
is
generally
viewed
as
a
cause
of
lung
cancer
although
the
chain
of
causation
is
long
and
indirect.
However,
a
molecular
biologist
concerned
with
a
lower
order
of
pathogenesis
would
not
consider
smoking
a
cause
of
interest,
for
the
smoking
behavior
of
the
individual
is
at
a
level
of
analysis
too
remote
from
his
or
her
concern.
According
to
Susser
et
al.:
Investigators
conceptualize
variables
and
abstract
them
from
a
given
ecological
setting
within
a
limited
frame
of
refer-
ence....
The
choice
is
the
outcome
of
the
needs
and
consciousness
of
an
investigator
in
a
particular
situation,
but
on
logical
grounds
it
is
an
arbitrary
procedure.38(p43)
Thus,
discussions
of
cause
in
epidemiol-
ogy
include
both
the
social,
ecological
level
and
the
individual
level
as
valid
arenas
of
causal
inquiry.
However,
the
manner
in
which
Robinson's
observations
have
been
adopted
tends
to
hinder
a
serious
consideration
of
social
factors
in
disease
etiology.
Conclusions
In
1979,
Kasl
suggested
that
epidemi-
ologists
need
to
develop
guidelines
for
comparing
ecological
analyses
with
stud-
ies
of
individuals.49
This
paper
posits
that
the
concept
of
the
ecological
fallacy,
the
framework
used
to
juxtapose
and
contrast
ecological
and
individual
studies,
cannot
fully
address
these
issues.
A
validity
approach,
examining
all
studies
in
terms
of
internal
and
particularly
construct
validity
problems,
may
prove
a
useful
addition
to
understanding
cross-level
infer-
ence.
This
perspective
would
suggest,
in
agreement
with
the
ecological
fallacy
perspective,
that
ecological
studies
cannot
usually
be
used
as
substitutes
for
indi-
vidual
correlational
studies.
However,
it
does
not
indicate
that
ecological
studies
are
etiologically
useless,
for
they
are
not
viewed
as
crude
estimates
of
individual-
level
studies.
Rather,
ecological
variables
are
necessary
to
examine
structural,
con-
textual,
and
sociological
effects
on
human
behavior
and
disease
development.
One
example
of
a
contextual
effect
that
has
important
public
health
conse-
quences
is
the
relationship
between
job
characteristics
and
myocardial
infarction.
Karasek
and
colleagues
found
that
indi-
viduals
in
jobs
characterized
by
high
levels
of
psychological
demands
coupled
with
low
decision
latitude
are
at
increased
risk
of
myocardial
infarction.50
These
job
characteristics,
while
clearly
operating
through
mechanisms
that
influence
indi-
viduals,
are
not
reducible
to
individual
characteristics.
They
are
variables
describ-
ing
the
psychosocial
work
environment
that
has
an
influence
on
the
workers'
health.
The
most
effective
intervention
to
reduce
this
risk
would
be
to
change
the
organization
of
these
occupations-an
intervention
at
the
ecological
rather
than
the
individual
level.
The
question
of
disease
etiology
is
complex.
It
is
likely
that
a
multitude
of
causes
is
involved
in
the
development
of
any
particular
disease.
Where
in
the
causal
chain,
among
the
myriad
of
vari-
ables,
one
chooses
to
examine
and
ascer-
tain
causation
is
often
a
question
of
where
intervention
is
most
efficacious.
That,
in
turn,
is
often
a
political
and
not
a
scientific
issue.
An
examination
of
the
full
range
of
variables
potentially
involved
in
disease
etiology,
with
a
synthesis
of
findings
from
all
levels
of
analysis,
provides
the
best
opportunity
for
a
full
understanding
of
disease
etiology.
O
Acknowledgments
This
work
was
supported
by
National
Institute
of
Mental
Health
grant
T32MH13043.
I
would
like
to
thank
Janice
Husted,
Jennifer
Kelsey,
Bruce
Link,
Steve
Ng,
Karen
Raphael,
and
Elmer
Struening
for
their
helpful
comments
on
earlier
drafts
of
this
manuscript.
References
1.
Mausner
JS,
Kramer
S.
Epidemiology-An
Introductory
Text.
Philadelphia,
Pa:
WB
Saunders;
1985.
2.
Kelsey
J,
Thompson
DW,
Evans
AS.
Methods
in
Observational
Epidemiology.
New
York,
NY:
Oxford
University
Press;
1986.
3.
Lillienfeld
AM,
Lillienfeld
DE.
Founda-
tions
of
Epidemiology.
New
York,
NY:
Oxford
University
Press;
1980.
4.
Rothman
KJ.
Modem
Epidemiology.
New
York,
NY:
Little
Brown;
1986.
5.
Friedman
GD.
Primer
of
Epidemiology.
New
York,
NY:
McGraw-Hill;
1974.
6.
Fletcher
RH,
Fletcher
SW,
Wagner
EH.
Clinical
Epidemiology:
The
Essentials.
Balti-
more,
Md:
Williams
&
Wilkins;
1988.
7.
Abramson
JH.
Making
Sense
of
Data.
New
York,
NY:
Oxford
University
Press;
1988.
8.
Selvin
S.
Statistical
Analsis
of
Epidemio-
Ecological
Fallacy
logic
Data.
New
York,
NY:
Oxford
Univer-
sity
Press;
1991.
9.
Morgenstern
H.
Uses
of
ecologic
analysis
in
epidemiologic
research.
Am
J
Public
Health.
1982;72:1336-1344.
10.
Kleinbaum
DG,
Kupper
LL,
Morgenstern
H.
Epidemiologic
Research:
Principles
and
Quantitative
Methods.
London,
England:
Lifetime
Learning;
1982.
11.
Robinson
WS.
Ecological
correlations
and
the
behavior
of
individuals.
Am
Sociol
Rev.
1950;15:351-357.
12.
Hammond
J.
Two
sources
of
error
in
ecological
correlations.
Am
Sociol
Rev.
1973;38:764-775.
13.
Meckstroth
TW.
Some
problems
in
cross-
level
inference.
Am
JPolitical
Sci.
1974;18:
45-66.
14.
Goodman
LA.
Ecological
regressions
and
the
behavior
of
individuals.
Am
Sociol
Rev.
1953;18:663-664.
15.
Goodman
LA.
Some
alternatives
to
ecologi-
cal
correlation.
Am
J
SocioL
1959;64:610-
625.
16.
Irwin
L,
Lichtman
AJ.
Across
the
great
divide:
inferring
individual
level
behavior
from
aggregate
data.
Pol
Methodology.
1976;3:411-439.
17.
Iversen
G.
Recovering
individual
data
in
the
presence
of
groups
and
individual
effects.
AmJSociol.
1974;79:421-434.
18.
Scheuch
EK.
Social
context
and
individual
behavior.
In:
Dogan
M,
Rokkan
S,
eds.
Social
Ecology.
Cambridge,
Mass:
Massa-
chusetts
Institute
of
Technology;
1969:133-
155.
19.
Marx
JH,
Suchman
EA,
Heller
DB.
A
systematic
framework
for
relating
the
be-
havioral
sciences
to
the
health
care
fields.
Inquiry.
1967;4:48-58.
20.
Hannan
M,
Burstein
L.
Estimation
from
grouped
observations.
Am
Sociol
Rev.
1974;39:374-393.
21.
Cook
T,
Campbell
D.
Quasi-Experimenta-
tion:
Design
and
Analysis
Issues
for
Field
Settings.
Boston,
Mass:
Houghton
Mifflin;
1979.
22.
Blalock
H.
Causal
Inferences
in
Nonexperi-
mental
Research.
Chapel
Hill,
NC:
Univer-
sity
of
North
Carolina
Press;
1964.
23.
Gove
WR,
Hughes
M.
Reexamining
the
ecological
fallacy:
a
study
in
which
aggre-
gate
data
are
critical
in
investigating
the
pathological
effects
of
living
alone.
Soc
Forces.
1980;58:1157-1177.
24.
Hanuschek
E,
Jackson
JE,
Kain
JF.
Model
specification,
use
of
aggregate
data
and
the
ecological
correlation
fallacy.
PolMethodol-
ogy.
1974;1:89-107.
25.
Firebaugh
G.
A
rule
for
inferring
individual-
level
relationships
from
aggregate
data.
Am
Sociol
Rev.
1978;43:557-572.
26.
Blau
P.
Structural
effects.
Am
Sociol
Rev.
1960;25:178-193.
27.
Blau
J,
Blau
P.
The
cost
of
inequality:
metropolitan
structure
and
violent
crime.
Am
Sociol
Rev.
1982;47:114-129.
28.
Davis
JA,
Spaeth
JL,
Huson
C.
A
tech-
nique
for
analyzing
the
effects
of
group
composition.
Am
Sociol
Rev.
1961;26:215-
225.
29.
Farkas
G.
Specification,
residuals
and
contextual
effects.
Sociol
Methods
Res.
1974;2:333-363.
30.
Meltzer
L.
Comparing
relationships
of
individual
and
average
variables
to
indi-
May
1994,
Vol.
84,
No.
5
American
Journal
of
Public
Health
823
Schwartz
vidual
response.
Am
Sociol
Rev.
1963;28:
117-123.
31.
Lichtman
A.
Correlation,
regression
and
the
ecological
fallacy:
a
critique.
JInterdisci-
plinary
Hist.
1974;4:417-433.
32.
Langbein
LI,
Lichtman
AJ.
Ecological
Inference.
Beverly
Hills,
Calif:
Sage;
1978.
33.
Riley
MW.
Sociological
Research:
A
Case
Approach.
New
York,
NY:
Harcourt,
Brace
and
Work;
1963.
34.
Menzel
H.
Comment
on
Robinson's
"eco-
logical
correlations
and
the
behavior
of
individuals."Am
Sociol
Rev.
1950;15:674.
35.
Catalano
R.
Health,
Behavior
and
the
Community:An
Ecological
Perspective.
New
York,
NY:
Pergamon
Press;
1979.
36.
Wolf
A.
Textbook
of
Logic.
New
York,
NY:
Collier;
1938.
37.
Zito
G.
Methodology
and
Meanings:
Variet-
ies
of
Sociological
Inquiry.
New
York,
NY:
Praeger;
1975.
38.
Susser
M,
Watson
W,
Hopper
K.
Sociology
in
Medicine.
New
York,
NY:
Oxford
Univer-
sity
Press;
1985.
39.
MacMahon
B,
Pugh
TF.
Epidemiology:
Principles
and
Methods.
Boston,
Mass:
Little
Brown;
1970.
40.
Lazarsfeld
PF,
Menzel
H.
On
the
relation
between
individual
and
collective
proper-
ties.
In:
Etzioni
A,
ed.
A
Sociological
Reader
on
Complex
Organizations.
New
York,
NY:
Holt,
Rinehart
and
Winston;
1969:499-516.
41.
Durkheim
E.
Suicide.
New
York,
NY:
Free
Press;
1951.
42.
Selvin
H.
Durkheim's
suicide
and
prob-
lems
of
empirical
research.
Am
J
Sociol.
1958;63:607-619.
43.
Thomdike
EL.
On
the
fallacy
of
imputing
the
correlations
found
for
groups
to
the
individuals
or
smaller
groups
composing
them.Am
JPsychol.
1939;52:122-124.
44.
Durkheim
E.
Rules
of
the
Sociological
Method.
New
York,
NY:
Free
Press;
1959.
45.
Susser
M.
Causal
Thinking
in
the
Health
Sciences.
New
York,
NY:
Oxford
Univer-
sity
Press;
1973.
46.
Coe
R.
Sociology
of
Medicine.
New
York,
NY:
McGraw
Hill;
1978.
47.
Bahnson
CB.
Epistemological
perspectives
of
physical
disease
from
the
psychody-
namic
point
of
view.
Am
J
Public
Health.
1974;64:1034-1039.
48.
Helman
C.
Culture,
Health
and
Illness.
Bristol,
England:
Wright;
1984.
49.
Kasl
SV.
Mortality
and
the
business
cycle:
some
questions
about
research
strategies
when
utilizing
macro-social
and
ecological
data.
Am
JPublic
Health.
1979;69:784-788.
50.
Karasek
RA,
Theorell
T,
Schwartz
JE,
Schnall
PL,
Peiper
CF,
Michela
JL.
Job
characteristics
in
relation
to
the
prevalence
of
myocardial
infarction
in
the
US
Health
Examination
Survey
(HES)
and
the
Health
and
Nutrition
Examination
Survey
(HANES).
Am
J
Public
Health.
1978:910-
919.
824
American
Journal
of
Public
Health
May
1994,
Vol.
84,
No.
5
- CitationsCitations459
- ReferencesReferences52
- "Ecological analyses are used when data are not accessible on an individual level (e.g., for data privacy regulations), and may give indications for risk factors on the basis of aggregated disease counts. A general limitation of these analyses is the potential for ecological fallacy [29]. Furthermore, many of the socio-economic variables in our study were given as percentages, which bears the risk of spurious correlations. "
[Show abstract] [Hide abstract] ABSTRACT: In 2011, a large outbreak of entero-hemorrhagic E. coli (EHEC) and hemolytic uremic syndrome (HUS) occurred in Germany. The City of Hamburg was the first focus of the epidemic and had the highest incidences among all 16 Federal States of Germany. In this article, we present epidemiological characteristics of the Hamburg notification data. Evaluating the epicurves retrospectively, we found that the first epidemiological signal of the outbreak, which was in form of a HUS case cluster, was received by local health authorities when already 99 EHEC and 48 HUS patients had experienced their first symptoms. However, only two EHEC and seven HUS patients had been notified. Middle-aged women had the highest risk for contracting the infection in Hamburg. Furthermore, we studied timeliness of case notification in the course of the outbreak. To analyze the spatial distribution of EHEC/HUS incidences in 100 districts of Hamburg, we mapped cases' residential addresses using geographic information software. We then conducted an ecological study in order to find a statistical model identifying associations between local socio-economic factors and EHEC/HUS incidences in the epidemic. We employed a Bayesian Poisson model with covariates characterizing the Hamburg districts as well as incorporating structured and unstructured spatial effects. The Deviance Information Criterion was used for stepwise variable selection. We applied different modeling approaches by using primary data, transformed data, and preselected subsets of transformed data in order to identify socio-economic factors characterizing districts where EHEC/HUS outbreak cases had their residence.- "Contextual effects are related to a broader political, cultural and/or institutional context, for example the presence of infrastructure such as available health services [9, 52] , but also ecological or environmental influences such as air pollution, noise pollution and temperature [40, 53]. In order to understand place effects on health, research should indeed consider higher-level variables too [41] . The impact of such effects seems to vary considerably in literature [1, 8, 18, 26, 43]. "
[Show abstract] [Hide abstract] ABSTRACT: Background Country averages for health outcomes hide important within-country variations. This paper probes into the geographic Belgian pattern of all-cause mortality and wishes to investigate the contribution of individual and area socio-economic characteristics to geographic mortality differences in men aged 45-64 during the period 2001-2011. Methods Data originate from a linkage between the Belgian census of 2001 and register data on mortality and emigration during the period 2001-2011. Mortality rate ratios (MRRs) are estimated for districts and sub-districts using Poisson regression modelling. Individual socio-economic position (SEP) indicators are added to examine the impact of these characteristics on the observed geographic pattern. In order to scrutinize the contribution of area-level socio-economic characteristics, random intercepts Poisson modelling is performed with predictors at the individual and the sub-district level. Random intercepts and slopes models are fitted to explore variability of individual-level SEP effects. Results All-cause MRRs for middle-aged Belgian men are higher in the geographic areas of the Walloon region and the Brussels-Capital Region (BCR) compared to those in the Flemish region. The highest MRRs are observed in the inner city of the BCR and in several Walloon cities. Their disadvantage can partially be explained by the lower individual SEP of men living in these areas. Similarly, the relatively low MRRs observed in the districts of Halle-Vilvoorde, Arlon and Virton can be related to the higher individual SEP. Among the area-level characteristics, both the percentage of men employed and the percentage of labourers in a sub-district have a protective effect on the individual MRR, regardless of individual SEP. Variability in individual-level SEP effects is limited. Conclusions Individual SEP partly explains the observed mortality gap in Belgium for some areas. The percentage of men employed and the percentage of labourers in a sub-district have an additional effect on individual MRR aside from that of individual SEP. However, these socio-economic factors cannot explain all observed differences. Other mechanisms such as public health policy, cultural habits and environmental influences contribute to the observed geographic pattern in all-cause mortality among middle-aged men. Keywords Belgium, Mortality, Geographic distribution, Socio-economic position, Poisson regression, Multilevel- "Future work on how online searches relate to health behavior or outcomes might require detailed analysis of socio-demographic, educational, or even psychological features of internet users. Speculations about online search behavior conducted using the methodology used here inevitably carries a risk of ecological inference fallacy; i.e., inferring individual behavior from the group to which those individuals belong [11,12] Anyone performing or interpreting this kind of study need to be aware of this risk, which is intrinsic to infodemiology. Despite these limitations, infodemiological studies remain a good way to shed light on internet search behavior at the population level. "
[Show abstract] [Hide abstract] ABSTRACT: Most patients with cancers and cancer survivors use the internet to obtain health information and support each other. Our aim was evaluate whether relationships exist between the information prevalences and search volumes of terms related to various cancers and their actual incidence and mortality figures in the USA and the UK.
Data provided are for informational purposes only. Although carefully collected, accuracy cannot be guaranteed. Publisher conditions are provided by RoMEO. Differing provisions from the publisher's actual policy or licence agreement may be applicable.
This publication is from a journal that may support self archiving.
Learn more