Content uploaded by Paula Katherine Hagedorn Diehr
Author content
All content in this area was uploaded by Paula Katherine Hagedorn Diehr
Content may be subject to copyright.
The
Validity
of
Self-Reported
Smoking:
A
Review
and
Meta-Analysis
Donald
L.
Patrick
PhD,
MSPH,
Allen
Cheadle,
PhD,
Diane
C.
Thompson,
MS,
Paula
Diehr,
PhD,
Thomas
Koepsell,
MD,
MPH,
and
Susan
Kinne,
PhD
Introduction
Smoking
continues
to
be
the
largest
single
preventable
cause of
premature
mortality
and
morbidity
in
the
United
States,
yet
29%
of
American
adults
continue
to
smoke.'
Efforts
to
promote
cessation
of
smoking
include
interven-
tions
conducted
with
patients
in
clinical
practices,
in
group
environments
such
as
schools
and
work
sites,
and
in
entire
communities.
Self-reports
of
smoking
be-
havior are
often
assessed
to
determine
the
efficacy
of
these
interventions.
Observa-
tional
studies
and
epidemiological
studies
of
risk
also
incorporate
measures
of
smoking
behavior.
Smoking
is
assessed
in
these
studies
to
discriminate
between
smokers
and
nonsmokers,
to
measure
change
in
smoking
status,
or
to
calculate
pack-years
of
exposure
retrospectively
for
risk
assessment.
The
validity
of
self-reported
smoking
is
often
questioned
because
of
the
wide-
spread
belief
that
smokers
are
inclined
to
underestimate
the
amount
smoked2'3
or
to
deny
smoking
at
all.45
As
more
attention
is
paid
to
smoking
in
the
media
and
in
public
places,
work
sites,
and
clinical
practice,
individuals
become
sensitized
to
socially
desirable
forms
of
behavior.
Thus,
smokers
may
be
more
likely
to
exaggerate
the
extent
to
which
their
behavior
con-
forms
to
the
perceived
social
norm
of
"not
smoking."
Bias
may
be
more
common
wherever
social
desirability
is
greater,
such
as
in
community-based
studies
in
which
intervention
programs
often
seek
explicitly
to
change
community
norms
about
the
social
acceptability
of
smoking.
Biochemical
assessments
of
smoking
by-products
in
body
substances
are
often
made
to
validate
self-reports
of
smoking.
Biochemical
assessments
can
be
viewed
primarily
as
measures
of
the
point
preva-
lence
of
current
smoking.6
Because
they
are
believed
to
be
more
objective
and
less
susceptible
to
bias,
biochemical
measures
are
most
often
considered
the
"gold
standard"
in
validation
studies
(i.e.,
they
are
considered
more
accurate
than
self-
reports
of
smoking).
Cotinine
(in
plasma,
saliva,
or
urine),
thiocyanate
(in
plasma
or
saliva),
and
carbon
monoxide
(in
expired
air)
are
the
most
commonly
used
bio-
chemical
assessments.
Participants
are
either
told
in
advance
that
such
assess-
ments
will
be
made
or
asked
for
informed
consent
and
specimens
"on
the
spot."
Sometimes
the
bogus
pipeline
procedure
has
been
used,
wherein
subjects
are
informed
that
their
self-reports
can
or
will
be
objectively
verified
by
the
researchers
by
means
of
a
biochemical
test.
In
actuality,
no
verification
takes
place,
although
speci-
mens
are
collected
and
left
unanalyzed.4
Despite
their
believed
objectivity,
biochemical
measures
do
not
provide
a
gold
standard,
nor
are
they
perfect
mea-
sures
of
accuracy
for
use
in
assessing
criterion
validity.
Carbon
monoxide
and
thiocyanate
can
be
elevated
in
those
who
do
not
use
tobacco,
and
cotinine,
although
a
specific
metabolite
of
nicotine,
can
be
elevated
in
users
of
snuff
and
chewing
tobacco.
When
biochemical
tests
are
Donald
L.
Patrick,
Allen
Cheadle,
Paula
Diehr,
and
Thomas
Koepsell
are
with
the
Department
of
Health
Services,
University
of
Washington,
Seattle.
Donald
L.
Patrick,
Tho-
mas
Koepsell,
Diane
C.
Thompson,
and
Susan
Kinne
are
with
the
Fred
Hutchinson
Cancer
Center,
Seattle.
Paula
Diehr
is
also
with
the
Department
of
Biostatistics,
and
Thomas
Koep-
sell
with
the
Department
of
Epidemiology,
University
of
Washington.
Requests
for
reprints
should
be sent
to
Donald
L.
Patrick,
PhD,
MSPH,
Department
of
Health
Services,
SC-37,
University
of
Wash-
ington,
Seattle,
WA
98195.
This
paper
was
accepted
October
27,
1993.
July
1994,
Volume
84,
No.
7
Self-Reported
Smolng
repeated, the
results
may
be
different
even
when
smoking
status
has
not
changed.
Biochemical
measures
also
have
practical
drawbacks.
Although
nonreactive,
these
measures
are
obtrusive:
blood,
saliva,
or
breath
samples
need
to
be
collected
from
the
individual.
Collection
of
samples
involves
more
contact
with
respondents
than
usual
in
conducting
large-scale
field
studies3
and
may
result
in
increased
refusals.6
Because
of
the
short
half-life
of
smoking
by-products
in
the
body,
bio-
chemical
assessment
validates
only
smok-
ing
status
near
the
time
of
specimen
collection.7
Costs
can
be
considerable,
ranging
from
less
than
$1
per
sample
for
carbon
monoxide
to
$20
per
sample
for
cotinine
analysis.8
The
cost
of
collecting,
handling,
and
arranging
for
frozen
storage
of
the
specimens
can
add
significantly
to
these
estimates.
In
contrast,
self-reported
smoking
is
assessed
easily
by
using
self-administered
questionnaires
in
person
or
by
mail
or
by
using
interviewer-administered
question-
naires
in
person
or
on
the
phone.
Ques-
tionnaires
are
noninvasive
and
inexpen-
sive,
and
assurances
of
confidentiality
of
information
can
reduce
refusals
to
partici-
pate.
Self-reported
information
can
be
used
to
measure
behavioral
change,
to
calculate
exposure
risk,
or
to
study
path-
ways
to
smoking
cessation
or
continua-
tion.
The
meta-analyses
reported
in
this
paper
combine
findings
from
a
number
of
studies
that
validated
self-reported
smok-
ing
with
biochemical
measures,
making
it
possible
to
examine
the
importance
of
different
aspects
of
the
studies,
the
popu-
lations,
and
the
validation
process.
This
paper
addresses
four
major
questions.
First,
what
evidence
exists
to
document
the
validity
of
self-reported
measures
of
smoking
behavior?
Second,
under
what
circumstances
is
it
most
important
for
investigators
to
consider
biochemical
as-
sessment
in
studies
of
smoking
behavior?
Third,
how
do
results
change
when
using
different
statistical
approaches
for
analyz-
ing
variation
in
the
measures
of
accuracy
and
for
pooling
information
across
stud-
ies?
Finally,
how
does
this
literature
review
inform
the
conduct
of
future
validation
studies,
reports
of
smoking
behavior,
and
the
publication
of
results?
Methods
Descnption
of
Meta-Analysis
Procedures
Meta-analyses
are
becoming
com-
mon
in
both
clinical
and
social
science
research.11
Meta-analytic
techniques
were
applied
in
this
study
to
observations
of
the
association
between
biochemical
measures
and
self-reported
smoking,
simi-
lar
to
meta-analyses
of
diagnostic
tests.
Similar
applications,
such
as
the
accuracy
of
the
exercise
electrocardiogram12
and
the
human
immunodeficiency
virus
anti-
body
test,13
have
appeared
in
the
litera-
ture.
This
application,
like
meta-analyses
of
randomized
clinical
trials,
is
an
observa-
tional
study
of
previously
published
studies.
Standard
procedures
were
followed
in
accumulating
and
evaluating
research
studies
for
the
meta-analysis.-l'
We
de-
fined
the
problem
as
accuracy
of
self-
reported
smoking,
with
biochemical
assess-
ment
as
the
criterion
or
concordance
measure
for
evaluating
validity.
Biblio-
graphic
searches
were
conducted
on
all
articles
published
between
1982
and
1991.
Initially
the
MEDLINE
database
was
used,
with
"smoking"
as
the
subject
head
and
the
keywords
"intervention
studies,"
"evaluation,"
"community-based
pro-
grams,"
and
"education"
as
subheads.
The
Science
Citation
Index
was
used
to
trace
articles
referenced
in
studies
previ-
ously
identified
in
the
bibliographic
search.
The
Current
Contents
database
was
scanned
for
more
recent
articles
through
mid-1991,
and
references
in
these
articles
were
also
evaluated.
Investigators
familiar
with
smoking
research
at
the
Fred
Hutchinson
Cancer
Research
Center
were
also
asked
to
identify
appropriate
studies
from
their
literature
files.
Thirty
studies
containing
compari-
sons
between
self-reported
smoking
and
biochemical
assessments
were
identified.
Studies
confined
to
pregnant
women
were
excluded.
All
studies
were
reviewed
for
information
on
the
following
characteris-
tics:
method
of
administration
(self-
administered
vs
interviewer
adminis-
tered),
biochemical
measures
(cotinine,
thiocyanate,
or
carbon
monoxide),
type
of
sample
(air,
blood,
saliva),
cutoff
value
used
for
biochemical
definition
of
smok-
ing,
population
(student
vs
general
popu-
lation),
study
design
(intervention
vs
observational),
sample
size,
ability
to
classify
participants
according
to
a
2
x
2
table
based
on
self-report
of
smoking
(yes/no)
and
the
gold-standard
definition
of
exceeding
the
defined
cutoff
level
on
the
biochemical
measure,
and
the
smok-
ing
rate
(i.e.,
prevalence
of
smoking),
defined
by
the
gold
standard
and
self-
report
measures.
The
2
x
2
table
for
calculating
the
accuracy
of
reports
is
shown
in
Figure
1.
All
the
required
data
were
available
in
26
of
the
30
articles
identified
as
validity
studies
of
self-reported
smoking.4-5'1437
Four
studies
did
not
contain
sufficient
information
to
calculate
accuracy
mea-
sures,
and
these
studies
were
eliminated
from
further
analysis.
Three
members
of
the
study
team
(Donald
L.
Patrick,
Diane
C.
Thompson,
and
Susan
Kinne)
ab-
stracted
data
independently
to
ensure
quality
control
of
the
data
used
in
the
analyses.
Discrepancies
among
the
three
abstractors
were
investigated
and
re-
solved
after
discussion.
The
26
studies
contained
32
comparisons
based
on
inde-
pendent
samples
and
51
comparisons
wherein
2
or
more
comparisons
were
made
on
partial
or
total
analyses
of
the
same
individuals.
Table
1
contains
the
essential
data
abstracted
from
the
studies
included
in
the
meta-analysis.
Measures
ofAccuracy
Data
abstracted
from
the
26
studies
were
used
to
calculate
two
measures
of
self-report
accuracy.
For
the
purposes
of
this
study,
in
which
biochemical
measures
were
considered
the
criterion
measure,
sensitivity
was
defined
as
a/(a
+
c)
in
Figure
1,
or
the
proportion
of
respondents
with
a
positive
level
on
the
biochemical
measure
that
reported
smoking.
Specificity
was
defined
as
dI(b
+
d),
or
the
propor-
tion
of
respondents
with
a
negative
level
on
the
biochemical
measure
that
reported
absence
of
smoking.
Analytic
Models
and
Procedures
Cutoff
levels
were
standardized
into
comparable
units
of
measurement
across
studies
for
each
type
of
biochemical
measure
and
biological
specimen.
Studies
using
thiocyanate-saliva
samples
were
eliminated
from
our
analyses
because
of
their
outlying
values
for
sensitivity
and
specificity
after
such
standardization.
We
thus
analyzed
47
of
the
51
comparisons
(allowing
more
than
1
comparison
per
American
Journal
of
Public
Health
1087
July
1994,
Volume
84,
No.
7
Patrick
et
al.
00
)z
--L
a
T)CV)
Cl)
M N
C)Cf
4CV
MT-))CM
N
)V
OU
O--
000
C)
0Mc
WMCO
DCI0
CDN
W
I
C\
O
(/)
C/)
o~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~C
0
~~~~~~~~~~~~~~~~~~~~~~~~~~~C
lt
C)V
1,
f-
q
LO
_
0
LO
CjC)L
D0
-0C
DC
"
)T-
\
-
LO
\j
0\0O0
C0
-
VNN
\0C\
)400M
-00
CV)CVT-N-Cl
Cl
l
CV)
CV)
CM
%CJm'-t
I
M
V
,
l-lt
C
0
-C
)
0
tC
D
CD
-CO
-LO
'a
05
~ ~ ~ ~ ~ ~ ~ ~ ~
~ ~ ~ ~
75
~~~~~~~~~~~~~~~~~~~~~~~~~~~E
0
~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~0)
cm0
c
NNmLor
C
)
or00
*
l
OC
)L
0
Y Y
-L)(
i)00
CY)CDL
COO[lV0)
00
O
0
i0o
CY)
~
0
0
)
0
0
00D
0
N
W
N 0
0
U ) ' c
t 0
W0
WI
O0
)
_
Cl)
0~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~N
.0
CD
g~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~&
LO
C)
coi
0~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
>CD
Cu
~~~~~000000000000000000
0---
0--
0
C
u)
oA
0C
0
0
0
~
0
0
0
-
0
.CM
CM
.M
.C
cm
0
If
0.rl
:
a)~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~>
8
0
0
0
>
>>
o>C
co.22
000000000000000000000000000000000000000000000000000000
C
0
0
Cu000000000
0000000000000000
<
<
CMC
~ ~ ~ ~ ~ ~ ~ ~ ~ ~ ~ ~ ~ ~ ~ ~ ~ ~ ~ ~ ~ ~ ~ ~ ~ ~ ~ ~
0o
C
Cl))o
E
c~~~~~~~~~~~~~~~~N
00
0)
v-
0'-
*W
'Co
Co
'-)
-'-'co-0
00'
(D
O).Q
0)0
C
M
M.
Id"-'
)0)0.
CCDCuCu
.CM
E
Q
(o
0
CD
0
C03
.)
xx'
0wI0C
00I0
XvX
CMJ
Ct0
0
0
0
t
.=
Z75
~~~~~c
c
c~~
om
cuc
oW.
c
V.
~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~c
1088
American
Journal
of
Public
Health
1
--
July
1994,
Volume
84,
No.
7
Self-Reported
Smoking
study)
and
30
of
the
32
strictly
indepen-
dent
comparisons
(allowing
only
1
com-
parison
per
study,
chosen
at
random).
Since
there
is
uncertainty
about
how
best
to
combine
proportions
across
stud-
ies,40
several
analytic
methods
were
used
to
obtain
point
estimates
and
confidence
intervals
of
average
sensitivity
and
specific-
ity.
The
fixed-effects
approach
assumes
that
all
studies
are
homogeneous
and
that
the
"true"
sensitivity
and
specificity
are
the
same
for
each
study.
Under
this
model,
estimates
from
each
study
differ
from
this
assumed
true
value
only
because
of
sampling
variation,
and
the
estimates
from
small
studies
tend
to
be
worse
than
those
from
larger
studies.
In
the
fixed-
effects
model,
the
most
efficient
point
estimate
for
sensitivity,
for
example,
is
the
mean
of
the
sensitivities
weighted
by
sample
size.
This
is
equivalent
to
putting
all
of
the
people
from
all
studies
into
one
2
x
2
table
and
estimating
sensitivity
and
specificity
from
it.
Confidence
intervals
are
calculated
from
the
weighted
stan-
dard
error.
Regression
analyses
to
deter-
mine
correlates
of
accuracy
can
be
achieved
by
using
weighted
ordinary
least
squares,
with
the
study
as
the
unit
of
analysis,
or
by
using
logistic
regression.
We
used
logistic
regression
to
conduct
fixed-effects
analyses
of
the
correlates
of
accuracy.
In
actuality,
however,
studies
are
not
homogeneous:
they
involve
different
study
populations
and
different
data
collection
methods.
The
"true"
sensitivity
and
speci-
ficity
values
being
estimated
probably
do
vary
among
studies
because
of
this
hetero-
geneity.
Under
a
random-effects
model,
variation
among
the
studies
is
assumed
to
be
due,
in
part,
to
the
variation
among
the
"true"
values,
as
well
as
to
sampling
variation
within
each
study.
The
correct
weight
in
this
case
depends
on
both
sources
of
variation.
If
the
between-
studies
heterogeneity
is
large,
this
results
in
approximately
equal
weights
for
each
study;
if
it
is
small,
then
the
weights
are
approximately
proportional
to
the
sample
size,
as
in
the
fixed-effects
model.
Laird
and
Mosteller38
suggested
a
moment-
based
estimate
for
the
mean,
with
its
associated
variance,
that
can
be
used
to
obtain
confidence
intervals
for
the
ran-
dom-effects
method.
This
method
com-
putes
appropriate
weights
for
each
study;
these
weights
are
a
combination
of the
variance
within
studies
and
the
variances
among
studies.
If
there
is
almost
no
variation
among
studies,
then
each
study
is
weighted
by
its
sample
size.
If
there
is
great
variance
among
studies,
then
each
study
will
receive
approximately
equal
weight.
We
computed
the
weights
and
found
that
the
latter
case
was
true
for
our
data.
The
sensitivity
and
specificity
esti-
mates
for
our
studies
were
so
variable
that
the
random-effects
analysis
essentially
gave
all
studies
equal
weight.
For
this
reason,
to
examine
the
effects
of
covari-
ates
under
a
random-effects
assumption,
we
used
simple
ordinary
least
squares
regression
with
the
study
as
the
(un-
weighted)
unit
of
analysis
and
sensitivity
and
specificity
as
the
dependent
variables.
The
sensitivity
and
specificity
mea-
sures
were
negatively
skewed.
To
test
the
sensitivity
of
the
findings
to
this
deviation
from
normality,
we
analyzed
the
loga-
rithm
of
100
minus
sensitivity
plus
1
and
100
minus
specificity
plus
1,
which
did
have
a
reasonably
normal
distribution.
The
results
of
analyses
using
the
logarith-
mically
transformed
measures
were
sub-
stantially
the
same
as
analyses
with
un-
transformed
data;
hence,
for
ease
of
interpretation,
we
report
the
original
analyses.
We
compared
fixed-effects
and
ran-
dom-effects
models.
For
the
purposes
of
this
paper,
we
present
unadjusted,
bivari-
ate
results
and
multivariate
analyses
using
the
ordinary
least
squares
random-effects
regression
(most
conservative)
and
the
fixed-effects
logistic
regression
(least
con-
servative).
Independent
variables
in
the
regression
analyses
were
method
of
admin-
istration
(self-administered
vs
interviewer
administered),
study
design
(observa-
tional
vs
intervention),
population
type
(student
vs
general
population),
and
type
of
biochemical
measure
and
specimen
(cotinine-plasma
as
the
reference
vs
cotinine-saliva,
thiocyanate-plasma,
or
carbon
monoxide
[air
and
blood]).
Two-
way
interactions
among
the
study
charac-
teristics
were
examined
in
the
random-
effects
ordinary
least
squares
analyses
of
47
comparisons;
one
interaction
was
en-
tered
at
a
time.
Results
A
total
of
36
830
respondents
were
included
in
the
26
studies
and
51
compari-
sons.
Of
the comparisons
reported
in
these
studies,
37.7%
were
obtained
from
self-administered
questionnaires
and
62.3%
were
obtained
by
interviewers.
Students
represented
22.8%
of
respon-
dents
for
the
comparisons;
77.2%
of
respondents
were
drawn
from
the
general
population.
Only
5.8%
were
respondents
in
intervention
studies
in
which
biochemi-
cal
assessments
were
used.
Carbon
monox-
ide,
thiocyanate,
and
cotinine
were
used
as
the
biochemical
measure
in
54.8%,
18.2%,
and
27%
of
the
comparisons,
respectively.
As
shown
in
Table
1,
sensitivity
values
ranged
from
6
to
100
in
the
51
comparisons;
specificities
ranged
from
33
to
100.
The
average
sensitivity
for
these
studies,
unweighted
by
sample
size,
was
87.5;
the
average
specificity
was
89.2.
Tables
2
and
3
show
results
for
sensitivity
and
specificity
for
all
47
comparisons
and
for
the
30
strictly
independent
compari-
sons,
respectively.
For
each
sample,
re-
sults
are
shown
for
the
ordinary
least
squares
random-effects
and
logistic
fixed-
effects
models.
The
bivariate
analyses
show
the
effect
on
sensitivity
or
specificity
of
each
study
characteristic
by
itself.
The
multivariate
analysis
results
come
from
regression
models
that
included
as
predic-
tors
all
of
the
study
characteristics
listed.
Thus,
they
show
the
effect
of
each
predictor
on
sensitivity
or
specificity
after
adjusting
statistically
for
all
of
the
other
predictors.
For
the
fixed-effects
analysis,
the
logistic
regression
coefficients
were
used
to
calculate
the
effect,
in
percentage
points,
of
the
study
characteristic
of
interest
when
other
characteristics
(if
any)
in
the
model
were
held
constant
at
their
mean
values.
Effects
for
the
ordinary
least
squares
random-effects
regression
were
the
coefficients
themselves.
For
example,
per
the
random-effects
ordinary
least
squares
regression
treating
comparisons
as
independent
(Table
2),
interviewer-
administered
survey
self-reports
had
a
sensitivity
that
was
5.2
percentage
points
higher
than
that
of
self-administered
survey
reports,
but
this
difference
was
not
significantly
different
from
zero
(P
=
.135).
For
sensitivity,
the
fixed-effects
result
for
interviewer-administered
studies
was
a
1.1
percentage
point
increase,
which
was
significant
at
the
.05
level.
After
control-
ling
for
all
other
study
characteristics,
the
effect
was
a
4.0
percentage
point
increase
in
self-report
sensitivity
for
the
random-
effects
model
(not
significant)
and
a
1.2
percentage
point
increase
in
the
fixed-
effects
analysis
(marginally
significant
at
P
=
.066).
None
of
the
study
characteris-
tics
showed
significantly
different
esti-
mates
for
sensitivity
in
either
the
bivariate
or
multivariate
random-effects
analyses.
For
the
fixed-effects
approach,
student
populations
and
the
type
of
sample
were
significantly
related
to
sensitivity.
Student
populations
yielded
significantly
lower
sensitivity
than
general
population
stud-
ies.
All
biochemical
samples
yielded
higher
American
Journal
of
Public
Health
1089
July
1994,
Volume
84,
No.
7
Patricket
al.
sensitivity
than
cotinine-plasma,
and
inter-
viewer-administered
studies
provided
mar-
ginally
greater
sensitivity
(P
=
.047).
For
specificity,
the
random-effects
model
showed
statistically
significant
lower
estimates
for
thiocyanate-plasma,
saliva
cotinine,
and
carbon
monoxide
(both
air
and
plasma)
in
comparison
with
plasma
cotinine.
When
all
study
characteristics
July
1994,
Volume
84,
No.
7
1090
American
Journal
of
Public
Health
TABLE
2-Estimated
Effects
of
Study
Characteristics
on
Sensitivity
and
Speclflcity
of
Self-Reported
Smoking:
All
Comparisons
(n
=
47)
Sensitivity
Specificity
Ordinary
Least
Ordinary
Least
Squares
Logistic:
Fixed
Squares
Logistic:
Fixed
Random
Effects Effects
Random
Effects
Effects
Study
Characteristic
Effecta
P
Effecta
P
Effecta
P
Effecta
p
Bivariate
analyses
Interviewer
administered
5.2
.135
1.1
.047
2.5
.313
6.3
<.001
Student
population
-4.6
.227
-3.7
<.001
1.8
.510
-3.4
<.001
Observational
study
-0.3
.937
-1.0
.135
4.2
.124
4.8
<.001
Cotinine-salivab
-0.4
.944
2.3
.017
-4.9
.212
-5.5
<.001
Thiocyanate-plasmab
2.1
.750
3.0
.001
-13.5
.001
-7.6
<.001
Cob
-4.5
.441
2.6
.003
-8.5
.023
-6.8
<.001
COHbb
-6.2
.382
6.3
<.001
-10.6
.019
-5.0
<.001
Multivariate
analyses
Interviewer
administered
4.0
.402
1.2
.066
3.8
.204
8.1
<.001
Student
population
-5.2
.301
-3.3
.001
0.1
.983
-0.3
.551
Observational
study
3.7
.480
0.8
.310
3.5
.286
5.2
<.001
Cotinine-salivab
1.4
.845
2.9
<.001
-5.4
.222
-8.6
<.001
Thiocyanate-plasmab
2.5
.716
2.6
<.001
-12.3
.007
-10.7
<.001
COb
-1.5
.809
2.9
<.001
-6.9
.085
-3.8
<.001
COHbb
-4.5
.556
5.6
<.001
-8.2
.097
-9.2
<.001
Note.
Four
studies
using
thiocyanate-saliva
samples
were
omitted
from
analysis.
CO
=
carbon
monoxide;
COHb
=
carboxyhemogobin,
measured
in
percent.
aEstimated
percentage
point
difference
in
sensitivity
or
specificity
due
to
the
study
characteristic.
bCotinine-plasma
is
the
reference.
TABLE
3-Estimated
Effects
of
Study
Characteristics
on
Sensitivity
and
Specificity
of
Self-Reported
Smoking:
Independent
Comparisons
Only
(n
=
30)
Sensitivity
Specificity
Ordinary
Least
Ordinary
Least
Squares
Logistic:
Fixed
Squares
Logistic:
Fixed
Random
Effects
Effects
Random
Effects
Effects
Study
Characteristic
Effecta
P
Effecta
P
Effecta
P
Effecta
p
Bivariate
analyses
Interviewer
administered
4.9
.326
2.4
.001
4.2
.056
6.2
<.001
Student
population
-4.7
.351
-4.6
<.001
-1.5
.508
-4.7
<.001
Observational
study
5.3
.389
2.1
.033
0.9
.748
1.3
.239
Cotinine-salivab
0.1
.993
3.3
.010
-8.1
.026
-7.1
<.001
Thiocyanate-plasmab
2.5
.804
3.6
.017
-9.0
.031
-9.2
<.001
COb
-4.5
.546
1.6
.128
-8.4
.007
-7.0
<.001
COHbb
-5.9
.511
8.0
<.001
-7.0
.051
-4.3
<.001
Multivariate
analyses
Interviewer
administered
4.0
.554
2.9
.002
4.6
.088
6.7
<.001
Student
population
-7.1
.364
-3.7
.003
2.3
.454
0.4
.469
Observational
study
9.3
.272
8.7
<.001
1.3
.688
0.8
.406
Cotinine-salivab
4.4
.667
3.9
<.001
-9.5
.022
-10.7
<.001
Thiocyanate-plasmab
5.3
.639
3.9
<.001
-6.9
.117
-8.9
<.001
Cob
1.2
.888
3.9
<.001
-7.4
.030
-5.1
<.001
COHbb
-0.9
.931
-8.5
<.001
-4.9
.245
-7.5
<.001
Note.
Two
studies
using
thiocyanate-saliva
samples
were
omitted
from
analysis.
CO
=
carbon
monoxide;
COHb
=
carboxyhemoglobin,
measured
in
percent.
aEstimated
percentage
point
difference
in
sensitivity
or
specificity
due
to
the
study
characteristic.
bCotinine-plasma
is
the
reference.
Self-Reported
Smoking
were
entered
into
the
regression
equa-
tion,
significant
beneficial
effects
for
speci-
ficity
remained
only
for
cotinine-plasma
in
comparison
with
plasma
thiocyanate
samples.
Estimates
of
self-report
accuracy
us-
ing
the
fixed-effects
analyses
were
much
more
likely
to
show
statistically
significant
differences
due
to
study
and
population
characteristics
than
were
the
results
of
random-effects
analyses.
The
ordinary
least
squares
random-effects
results in
Table
2
(columns
2
and
4)
illustrate
the
most
conservative
approach;
the
logistic
fixed-effects
results
(columns
3
and
5)
are
less
conservative
under
the
assumption
that
studies
are
homogeneous.
The
ordi-
nary
least
squares
random-effects
analy-
ses,
both
bivariate
and
multivariate,
showed
no
significant
effects
for
sensitiv-
ity
and
significantly
lower
estimates
of
specificity
for
plasma
thiocyanate
samples
using
statistical
significance
as
the
crite-
rion.
Table
3
shows
the
results
of
even
more
conservative
analyses
using
only
30
independent
comparisons,
with
1
compari-
son
chosen
at
random
from
each
study
eligible
for
analyses.
Results
were
similar
for
sensitivity
using
the
random-effects
or
the
fixed-effects
approach
in
both
bivari-
ate
and
multivariate
analyses.
No
study
characteristics
produced
significantly
higher
or
lower
estimates
of
self-report
accuracy.
Fixed-effects
analyses
of
sensitiv-
ity
were
also
similar
for
bivariate
and
multivariate
analyses,
indicating
that
all
study
characteristics
produced
the
same
higher
or
lower
estimates.
Results
of
studies
with
student
populations
yielded
lower
estimates
of
sensitivity.
For
specific-
ity
and
the
random
effects
analyses,
significantly
lower
estimates
were
ob-
tained
for
the
different
samples
in
com-
parison
with
plasma
cotinine.
All
study
characteristics
were
significantly
different
in
the
bivariate,
fixed-effects
analysis
of
specificity,
except
for
observational
stud-
ies
(in
comparison
with
intervention
stud-
ies).
This
same
pattern
emerged
with
multivariate
analyses
of
specificity
using
the
fixed-effects
approach,
although
the
differences
obtained
from
student
popula-
tions
no
longer
were
evident.
An
analysis
of
two-way
interaction
effects
using
the
random-effects
model
and
47
comparisons
yielded
only
one
significant
interaction
each
for
sensitivity
and
specificity
(interviewer-administered
questionnaires
in
observational
studies).
Our
power
to
detect
interaction
effects,
however,
was
small
given
the
small
num-
ber
of
studies
included
in
the
analysis.
Discussion
This
meta-analysis
of
published
stud-
ies
comparing
self-reported
smoking
sta-
tus
with
results
of
biochemical
validation
suggests
generally
high
levels
of
sensitivity
and
specificity
for
self-report.
Across
all
studies,
the
sensitivity
of
self-report
was
87%,
and
the
specificity
was
89%.
None-
theless,
both
measures
of accuracy
proved
quite
variable
among
studies,
as
shown
in
Table
1,
suggesting
that
specific
aspects
of
the
setting,
study
population,
measure-
ment
methods,
and
study
purpose
are
important
to
the
accuracy of
smoking
self-reports.
Our
search
for
systematic
patterns
of
variation
in
sensitivity
and
specificity
across
studies
was
only
partially
success-
ful.
Two
different
methods
of
meta-
analysis
yielded
generally
similar
results
on
the
sign
and
magnitude
of
the
effect
of
each
study
characteristic
on
sensitivity
or
specificity,
but
they
often
produced
widely
divergent
verdicts
on
the
statistical
signifi-
cance
of
those
effects.
The
fact
that
the
random-effects
model
produced
P
values
that
were
usually
much
higher
than
those
from
the
fixed-effects
model
indicates
a
large
amount
of
between-study
variability
that
could
not
be
accounted
for
by
the
study
characteristics
measured.
This
con-
clusion,
in
turn,
suggests
that
other
unmea-
sured
study
characteristics
may
confound
our
results
about
the
effects
of
method
of
administration,
study
population,
study
type,
and
biochemical
test
and
specimen.
One
of
the
most
important
unmea-
sured
study
characteristics
is
the
specific
wording
of
questions
on
smoking
status.
Very
few
studies
reported
this
critical
information,
despite
considerable
evi-
dence
from
survey
research
that
re-
sponses
are
heavily
influenced
by
how
a
question
is
phrased
and
the
order
in
which
questions
are
asked.
Because
of
these
limitations,
the
observed
patterns
of
association
between
accuracy
of
self-report
and
study
charac-
teristics
must
be
interpreted
with
caution.
The
results
suggest
that
interviewer-
administered
questionnaires
yielded
higher
estimates
of
sensitivity
and
specific-
ity
than
did
self-administered
question-
naires.
Interviews
identified
more
of
the
smokers
correctly
and
classified
nonsmok-
ers
more
accurately.
This
may
reflect
smokers'
awareness
of
sensory
cues
about
their
smoking
(visible
cigarettes,
nicotine
stains
on
teeth
or
hands,
smoke
odor)
that
would
be
obvious
to
an
interviewer.
More
respondents
may
attempt
to
hide
smoking
behavior
in
self-administered
question-
naires,
even
when
biochemical
validation
is
known.
Even
in
the
most
conservative
analy-
ses,
student
self-reports
had
lower
sensitiv-
ity
than
studies
using
reports
from
sub-
jects
in
the
general
population;
however,
the
results
were
not
always
statistically
significant.
That
is,
students
appear
more
likely
to
deny
smoking,
even
when
bio-
chemical
measures
classify
them
as
smok-
ers.
This
is
not
surprising,
since
smoking
by
minors
is
illegal
in
most
states
and
many
young
tobacco
users
have
not
yet
defined
themselves
as
smokers.
Both
of
these
conditions
would
contribute
to
a
tendency,
whether
conscious
(fear
of
being
found
out)
or
unconscious
(self-
definition
inconsistent
with
behavior),
to
underreport.
The
different
analyses
do
not
suggest,
however,
that
student
self-
reports
have
higher
specificity
than
do
those of
subjects
from
the
general
popula-
tion.
Unfortunately,
all
studies
of
student
populations
reported
here
were
observa-
tional
in
nature;
no
intervention
studies
with
students
that
reported
biochemical
results
were
found
among
published
stud-
ies,
although
the
bogus
pipeline
proce-
dures
have
been
used
with
students.
Reports
of
accuracy
from
intervention
studies
with
student
populations
might
have
yielded
lower
estimates
of
sensitivity
given
the
results
from
observational
studies.
In
the
most
conservative
analyses,
using
only
one
comparison
from
each
study
and
the
random-effects
approach,
observational
studies
had
higher
levels
of
sensitivity
than
intervention
studies,
a
conclusion
supported
by
a
qualitative
review
of
studies
in
the
1990
surgeon
general's
report.3
Self-reports
from
sub-
jects
in
intervention
studies,
in
which
there
is
an
expectation
of
cessation
of
smoking,
are
more
likely
to
involve
under-
reporting
of
actual
smoking.
Self-reports
of
participants
in
inter-
vention
programs
also
have
lower
specific-
ity,
meaning
that
more
biochemically
validated
nonsmokers
reported
smoking.
Biochemical
tests
have
limited
ability
to
detect
the
very
low
levels
of
smoking
that
would
be
expected
from
recent
quitters
who
"slip"
and
smoke
an
occasional
cigarette.3m
Discussions
of
quitters'
reac-
tions
to
such
slips
(the
abstinence
viola-
tion
effect)
suggest
that
the
would-be
quitter
is
likely
to
exaggerate
the
impor-
tance
of
a
few
cigarettes
under
these
conditions.39
This
can
produce
a
report
of
"smoking,"
although
its
magnitude
is
smaller
than
the
test
can
detect.
In
observational
studies,
with
little
focus
on
American
Journal
of
Public
Health
1091
July
1994,
Volume
84,
No.
7
Patrick
et
al.
cessation,
this
reaction
may
not
be
trig-
gered.
Consistent
with
previous
reports,40
self-reports
of
smoking
validated
by
means
of
cotinine-plasma
biochemical
measures
appear
to
have
higher
specificity
than
those
reports
validated
by
other
biochemi-
cal
tests
and
specimens.
The
P
values
for
tests
comparing
cotinine-plasma
and
other
biochemical
measures
were
not
always
significant
in
the
more
conservative
ran-
dom
effects
model,
but
the
signs
and
magnitudes
of
the
coefficients
were
mostly
consistent
across
analytic
models
and
between
Tables
2
and
3.
The
observed
pattern
shows
that
more
false-positive
self-reports
(i.e.,
respondents
report
that
they
are
smokers
when
the
biochemical
test
does
not
confirm
such
a
report)
tend
to
be
observed
in
studies
that
use
methods
other
than
cotinine-plasma.
This
finding
may
reflect
variation
in
the
accuracy
of
the
other
biochemical
tests
in
relation
to
self-reported
smoking.
These
so-called
false-positives
may
actually
be
smokers,
but
the
poorer
biochemical
tests
were
too
insensitive
to
detect
relatively
low
levels
of
biochemical
abnormality.
There
is
no
substantive
reason
to
expect
differences
in
self-report
behavior
due
to
the
form
of
biochemical
validation.
The
fixed-effects
and
random-effects
analyses
yielded
both
different
point
esti-
mates
of
the
effects
and
different
signifi-
cance
levels.
Between-study
variation
was
highly
significant
in
all
analyses,
calling
into
question
a
key
premise
underlying
the
fixed-effects
model.
The
results
ob-
tained
from
using
multiple
nonindepen-
dent
comparisons
in
the
different
studies
(Table
2)
and
from
using
only
indepen-
dent
measures
(Table
3)
were
substan-
tially
similar.
Thus,
the
lack
of
indepen-
dence
in
samples
did
not
prove
to
be
a
serious
problem
for
this
meta-analysis.
In
summary,
our
results
suggest
that
biochemical
validation
may
be
more
im-
portant
in
intervention
studies,
in
studies
with
student
populations,
and
in
studies
using
self-administered
rather
than
inter-
viewer-administered
questionnaires.
For
greatest
accuracy,
self-administered
ques-
tionnaires
given
to
students
might
benefit
from
biochemical
validation,
given
the
lower
estimates
obtained
from
these
groups.
Cotinine-plasma
may
be
the
biochemical
test
of
choice
if
adequate
resources
are
available
for
collection
and
analysis.
The
decision
to
use
biochemical
validation
is
not
as
straightforward
as
it
might
appear.
Biochemical
validation
is
costly
and
sometimes
difficult
to
obtain
1092
American
Journal of
Public
Health
for
all
participants.6
The
bogus
pipeline
procedure
and
use
of
biochemical
assess-
ments
with
random
subsamples
of
the
target
population
are
alternative
strate-
gies.4
The
conclusions
from
this
meta-
analysis
are
subject
to
three
major
caution-
ary
notes
that,
in
turn,
indicate
needed
improvements
in
the
conduct
and
report-
ing
of
future
studies
of
smoking
behavior.
First,
it
is
known
that
the
form
and
content
of
self-report
questions
about
smoking
influence
the
responses
given
and,
hence,
the
categorization
of
respon-
dents
as
smokers.25
This
observation
ar-
gues
that
studies
asking
about
smoking
should
report
or
reference
the
exact
questions
used
so
that
this
source
of
misclassification
can
be
controlled.
Sec-
ond,
authors
of
published
studies
need
to
make
clear
how
biochemical
validation
was
presented
to
study
participants.
In
studies
in
which
participants
know
that
biochemical
assessment
will
occurr,
such
as
in
the
bogus
pipeline
procedure,
self-reported
smoking
rates
may
be
differ-
ent
from
those
in
studies
in
which
bio-
chemical
assessment
is
presented
only
at
the
point
of
collection.
The
exact
proce-
dures
used
in
studies
should
be
identified
in
the
methods
sections
of
articles
to
permit
evaluation
of
the
results
according
to
the
potential
bias
introduced
by
precol-
lection
announcement
of
biochemical
vali-
dation.
Finally,
in
any
meta-analysis,
publica-
tion
bias
and
the
"file
drawer
phenom-
enon,"
the
failure
to
submit
for
publica-
tion
studies
that
do
not
produce
effects,
have
an
impact
on
the
data
available
for
analysis
and
can
bias
the
outcome.4142
Although
this
bias
is
likely
to
be
reduced
in
smoking
studies
that
do
not
seek
to
test
a
specific
hypothesis
about
validity,
the
universe
of
potential
data
sets
on
the
validity
of
self-reports
is
still
influenced
by
investigator
analysis
and
submission
of
data
for
publication.
This
meta-analysis
also
excluded
studies
of
pregnant
women,
and,
thus,
the
generalizability
of
the
results
must
be
considered
in
comparison
with
the
population
under
investigation.
The
failure
to
include
sufficient
data
on
biochemical
assessment
resulted
in
the
exclusion
of
several
studies
that
reported
validations
of
self-report.
This
suggests
that
in
future
studies
with
smoking
valida-
tion,
sufficient
data
should
be
published
for
investigators
to
confirm
and
evaluate
the
accuracy
of
self-reports.
O
Acknowledgments
This
work
was
supported
by
grant
CA
34847
from
the
National
Cancer
Institute
awarded
to
the
Fred
Hutchinson
Cancer
Center
Public
Health
Sciences
Division,
Cancer
Prevention
Research
Unit,
for
work
on
the
methodology
of
community-based
studies.
An
earlier
version
of
this
paper
was
presented
at
the
International
Conference
on
Measurement
Errors
in
Survey,
Tucson,
Ariz,
November
1990.
References
1.
Reducing
the
Health
Consequences
of
Smok-
ing:
25
Years
of
Progress.
A
Report
of
the
Surgeon
GeneraL
Washington,
DC:
US
Dept
of
Health
and
Human
Services;
1989.
DHHS
publication
CDC
89-8411.
2.
Haley
NJ,
Hoffmann
D.
Analysis
for
nicotine
and
cotinine
in
hair
to
determine
cigarette
smoker
status.
Clin
Chem.
1985;31:
1598-1600.
3.
The
Health
Benefits
of
Smoking
Cessation.
Washington,
DC:
US
Dept
of
Health
and
Human
Services;
1990.
DHHS
publication
CDC
90-8416.
4.
Murray
DM,
O'Connell
CM,
Schmidt
LA,
Perry
CL.
The
validity
of
smoking
self-
reports
by
adolescents:
a
reexamination
of
the
bogus
pipeline
procedure.
Addict
Be-
hav.
1987;12:7-15.
5.
Luepker
RV,
Pallonen
UE,
Murray
DM,
Pirie
PL.
Validity
of
telephone
surveys
in
assessing
cigarette
smoking
in
young
adults.
Am
JPublic
Health.
1989;79:202-204.
6.
Velicer
WF,
Prochaska
JO,
Rossi
JS,
et
al.
Assessing
outcome
in
smoking
cessation
studies.
Psychol
Bull.
1992;1 11:23-47.
7.
Benowitz
NL.
The
use
of
biologic
fluid
samples
in
assessing
tobacco
smoke
con-
sumption.
In:
Grubowski
J,
Bell
CS,
eds.
Measurement
in
the
Analysis
and
Treatment
of
Smoking
Behavior.
Washington,
DC:
US
Dept
of
Health
and
Human
Services;
1983;6-26.
NIDA
research
monograph
48.
8.
Windsor
T,
Morris
J,
Cutter
G,
et
al.
Sensitivity,
specificity
and
predictive
value
of
saliva
thiocyanate
among
pregnant
women.
AddictBehav.
1989;14:447-452.
9.
Light
RJ,
Pillemer
DB.
Summing
Up:
The
Science
of
Reviewing
Research.
Cambridge,
Mass:
Harvard
University
Press;
1984.
10.
Rosenthal
R.
Meta-Analytic
Procedures
for
Social
Research.
Newbury
Park,
Calif:
Sage;
1984.
Applied
Social
Research
Methods
Series
Volume
6.
11.
Hunter
JE,
Schmidt
FL.
Methods
of
Meta-
Analysis.
Newbury
Park,
Calif:
Sage;
1990.
12.
Gianrossi
R,
Detrano
R,
Mulvihill
D,
et
al.
Exercise-induced
ST
depression
in
the
diagnosis
of
coronary
artery
disease.
A
meta-analysis.
Circulation.
1989;80:87-98.
13.
Phillips
KA.
The
use
of
meta-analysis
in
technology
assessment:
a
meta-analysis
of
the
enzyme
immunosorbent
assay
human
immunodeficiency
virus
antibody
test.
J
Clin
Epidemiol.
1991;44:925-931.
14.
Bauman
KE,
Dent
CW.
Influence
of
an
objective
measure
on
self-reports
of
behav-
ior.
JAppl
Psychol.
1982;67:623-628.
15.
Bauman
KE,
Koch
GG,
Bryan
E.
Validity
of
self-reports
of
adolescent
cigarette
smok-
ing.
!ntJAddict.
1982;17:1131-1136.
16.
Bauman
KE,
Koch
GG.
Validity
of
self-
July
1994,
Volume
84,
No.
7
reports
and
descriptive
and
analytical
con-
clusions:
the
case
of
cigarette
smoking
by
adolescents
and
their
mothers.
Am
J
Epidemiol.
1983;1
18:90-98.
17.
Cohen
JD,
Bartsch
GE.
A
comparison
between
carboxyhemoglobin
and
serum
thiocyanate
determinations
as
indicators
of
cigarette
smoking.
Am
J
Public
Health.
1980;70:284-286.
18.
Fortmann
SP,
Rogers
T,
Vranizan
K,
et
al.
Indirect
measures
of
cigarette
use:
expired-
air
carbon
monoxide
versus
plasma
thiocya-
nate.
Prev
Med.
1984;13:127-135.
19.
Gillies
PA,
Wilcox
B,
Coates
C,
et
al.
Use
of
objective
measurement
in
the
validation
of
self-reported
smoking
in
children
aged
10
and
11
years:
saliva
thiocyanate.
J
Epidemiol
Community
Health.
1982;36:205-
208.
20.
Haddow
JE,
Palomakii
GE,
Knight
GJ.
Use
of
serum
cotinine
to
assess
the
accuracy
of
self-reported
non-smoking.
BMJ.
1986;293:1306.
Letter.
21.
Luepker
RV,
Pechacek
TF,
Murray
DM,
et
al.
Saliva
thiocyanate:
a
chemical
indica-
tor
of
cigarette
smoking
in
adolescents.
Am
JPublic
Health.
1981;71:1320-1324.
22.
McNeill
AD,
Jarvis
MJ,
West
R,
et
al.
Saliva
cotinine
as
an
indicator
of
cigarette
smoking
in
adolescents.
BrJAddict.
1987;82:
1355-1360.
23.
Noland
MP,
Kryscio
RJ,
Riggs
RS,
et
al.
Saliva
cotinine
and
thiocyanate:
chemical
indicators
of
smokeless
tobacco
and
ciga-
rette
use
in
adolescents.
J
Behav
Med.
1988;11
:423-433.
24.
Pechacek
TF,
Murray
DM,
Luepker
RV,
et
al.
Measurement
of
adolescent
smoking
behavior:
rationale
and
methods.
J
Behav
Med.
1984;7:123-130.
25.
Petitti
DB,
Friedman
GD,
Kahn
W.
Accu-
racy
of
information
on
smoking
habits
provided
on
self-administered
research
questionnaires.
Am
J
Public
Health.
1981;
71:308-311.
26.
Pierce
JP,
Dwyer
T,
DiGiusto
E,
et
al.
Cotinine
validation
of
self-reported
smok-
ing
in
commercially
run
community
sur-
veys.
J
Chronic
Dis.
1987;40:689-695.
27.
Pojer
R,
Whitefield
JB,
Poulos
V,
et
al.
Carboxyhemoglobin,
cotinine,
and
thiocya-
nate
assay
compared
for
distinguishing
smokers
from
non-smokers.
Clin
Chem.
1984;30:1377-1380.
28.
Prignot
J.
Quantification
and
chemical
markers
of
tobacco-exposure.
Eur
J
Respir
Dis.
1987;70:1-7.
29.
Saloojee
Y,
Vesey
CJ,
Cole
PV.
Carboxyhe-
moglobin
and
plasma
thiocyanate:
comple-
mentary
indicators
of
smoking
behavior?
Thorax
1982;37:521-525.
30.
Sillett
RW,
Wilson
MB,
Malcolm
RE,
et
al.
Deception
among
smokers.
BMJ.
1978;2:
1185-1186.
31.
Slattery
ML,
Hunt
SC,
French
TK,
et
al.
Validity
of
cigarette
smoking
habits
in
three
epidemiologic
studies
in
Utah.
Prev
Med.
1989;18:11-19.
32.
Stookey
GK,
Katz
BP,
Olson
BI,
et
al.
Evaluation of
biochemical
validation
mea-
sures
in
determination
of
smoking
status.
J
Dent
Res.
1987;66:1597-1601.
33.
Van
Vunakis
H,
Tashkin
DP,
Rigas
B,
et
al.
Relative
sensitivity
and
specificity
of
salivary
and
serum
cotinine
in
identifying
tobacco-smoking
status
of
self-reported
nonsmokers
and
smokers
of
tobacco
and/or
marijuana.
Arch
Environ
Health.
1989;44:
53-58.
34.
Vogt
TM,
Selvin
S,
Widdowson
G,
et
al.
Expired
air
carbon
monoxide
and
serum
Self-Reported
Smoldng
thiocyanate
as
objective
measures
of
ciga-
rette
exposure.AmJPublic
Health.
1977;67:
545-549.
35.
Wagenknecht
LE,
Cutter
GR,
Haley
NJ,
et
al.
Racial
differences
in
serum
cotinine
levels
among
smokers
in
the
coronary
artery
risk
development
in
(young)
adults
study.
Am
J
Public
Health.
1990;80:1053-
1056.
36.
Wald
NJ,
Idle
M,
Boreham
J,
et
al.
Carbon
monoxide
in
breath
in
relation
to
smoking
and
carboxyhemoglobin
levels.
Thorax.
1981;36:366-369.
37.
Williams
CL,
Eng
A,
Botwin
GJ,
et
al.
Validation
of
students'
self-reported
ciga-
rette
smoking
levels
with
plasma
cotinine
levels.
Am
J
Public
Health.
1979;69:1272-
1274.
38.
Laird
NM,
Mosteller
F.
Some
statistical
methods
for
combining
experimental
re-
sults.
Int
J
Technol
Assess
Health
Care.
1990;6:5-30.
39.
Curry
SG,
Marlatt
GA,
Gordon
JR.
Absti-
nence
violation
effect:
validation
of
an
attributional
construct
with
smoking
cessa-
tion.
J
Consult
Clin
Psychol.
1987;55:145-
149.
40.
Murray
RP,
Connett
JE,
Lauger
GG,
Voelker
HT.
Error
in
smoking
measures:
effects
of
intervention
on
relations
of
cotinine
and
CO
to
self-reported
smoking.
AmJPublic
Health.
1993;83:1251-1257.
41.
Jenicek
M.
Meta-analysis
in
medicine:
where
we
are
and
where
we
want
to
go.
J
Clin
EpidemioL
1989;42:35-44.
42.
Fleiss
JL,
Gross
AJ.
Meta-analysis
in
epidemiology,
with
special
reference
to
studies
of
the
association
between
expo-
sure
to
environmental
tobacco
smoke
and
lung
cancer:
a
critique.
J
Clin
Epidemiol.
1991;44:127-139.
July
1994,
Volume
84,
No.
7
American
Journal
of
Public
Health
1093