Download as pdf or txt
Download as pdf or txt
You are on page 1of 9

Original

Self-reported
Michael

The

compared
nationally
from

error

with measured
representative

the

Survey

second

and height12

of 1976-1980.

reporting
status-bias

Health

small

weight
and

population

to a persons

increased

with

a predictor

were

their

reported,

weight

and

Errors

Errors

mobile

and

age after

ofreporting

height

unreliability

error

self-reported
NHANES

Body

weight,
II, regression

weight,

body

status

reporting
nature

age range,

were
was

Am J Clin NuIr

self-reported
analysis,

height,

height,
end-digit

body

mass

population
preference

68%

report

describes

report

their

the

and

with the accuracy

ofreported

since

study

the

seminal

in The American

These

accuracy

heights

weight

Journal ofClinical
reflect the importance

studies

either

in their

representativeness
earlier

study

based

the

bias

on self-reported

defines
reported

the distribution
height
and

general
National

population
Health

II).

Knowledge

and
weight

and

NHANES

II

NHANES
Statistics

(1 1 864)

This

height

weights

paper
The

the

direction
of reporting
in a representative

and

an

that

are

error
sample

subjects

in the

sample

Excluded

from

direction

of reporting
in which

designed

l990;52:

by the

probability
I 125-33.

Printed

National
sample

to adults

adolescents

national

and

adult

of 17 390

weights
were
during
the ex-

of reported

weights

subjects

no evidence

be of

In the

sample

Heights
measured

aged

analysis

may

in adults.

nonexamined

indicated

the

among

of a nonresponse

were

2 1 8 people

99 pregnant

women;

and

in reported
and measured
the mean difference
(outliers).
proxies,
and subjects
alter the conclusions

sample

national

with

missing

mea-

or height;
1 1 3 subjects
whose
response
else (proxies);
92 people
of unknown
58 subjects
body mass
Reanalysis

with unknown
of the study.

consisted

for whom
index were
including

respondent

of 1 1 284 people,

sta-

65% of

sample.

study

weight

Weight

was

reported

About

how

much

About
information

Center

and

height
you

are

you

without

was

also

gathered,

in response

including

interviews.

clothes

in answer

shoes?

question,
or shoes?

to the question,

Other
age,

as

obtained

to the

without

inches

acting

Statistics

in house-to-house

weigh

in feet and

tall

of the Census

for Health

in pounds
do

was reported
how

the US Bureau

National

self-reported

Height

for selfof the

from

for the

demographic

sex,

race,

and

ed-

ucation.

error
height
I From
the Division of Health Examination
Statistics, National Center
for Health Statistics,
Hyattsville,
MD.
2 Address
reprint
requests
to ML Rowland,
Department
of Health

Center
ofthe

in USA.

in

analyzed.

An interviewer

methods

II was

original

and

the study

status;

an agent

States from the second


Survey
(NHANES

and interpreting
studies
rather
than measured.

history
constructed

Variables

or their

present

during

y. People

from

of the distribution

examined

The final analytic

extends

original

ofthe

heights

the outliers,
tus did not

conducted

indices
(8).

ofthe

and
error

Comparison

(7) appeared

and

ofadults
in the United
and Nutrition
Examination

to be a national

Am J C/in Nutr

and

of the degree

which
dealing

variables
in epistudies
have been

in weight-for-height

weight

will be useful in planning


and weight
are reported

Subjects

US population.

ofthe
on

ofheights

reporting

and

the differences
3 SDs from

Nutrition
(AJCN)
in 1981.
of weight and weight-for-

and as exposure
these accuracy

distribution

with
studies

Albaum

for children

from

sured or reported
weight
was provided
by someone

have been

and

error

excluded

amination.

survey,

reliability
Several

(1-6)

by Stunkard

height
indices
in themselves
demiologic
research.
However,
limited

and

weights.

place

specially

was limited

20 y were

<

interviewed
and examined.
during
the interview
and

respondent

This

took
6 mo-74

a medical
into

adults were
self-reported

bias

index,

Introduction

people

administered
brought

for this study

aged

all interviewed

WORDS

later

sample

y. People

a different

1990;52:l 125-33.
KEY

aged

sample

because

in self-reporting

in height.

survey

people

for examination.

The analytic
20-74

weight

age 45 y. Overweight

were

were

(9). The

included

sample

and

centers

Analytic

in self-

in self-reported

in self-reporting

for the
home

and

height

than in overweight
males.
were ancillary
predictors
of

Errors

US population

1976-1980

selected

related
to a persons
overweight
in self-report
increased
directly

age-bias

directly

the years

Examination
Statistics,
National
Center for Health
1661, Hyattsville,
MD 20788.
Received
January
26, 1990.
Accepted
for publication
March 21, 1990.

for Health
civilian,

non-

1990 American

Society

for Clinical

Nutrition

Statistics,

P0

Box

1125

Downloaded from ajcn.nutrition.org by guest on July 8, 2014

related

in weight.

height

subgroups.

ofoverweight.

error

and

institutionalized

height

Examination

self-reported

were greater
in overweight
females
Race, age, and end-digit
preference
reporting

and

evaluated
in a
aged 20-74
y

Nutrition

weight

were directly
unreliability

the magnitude

and

errors,

in important

weight

and height
was
of 1 1 284 adults

Although

with

are unreliable

in self-reported

weight
sample

National

on the average,

also

Communications-surveys

L Row/and

ABSTRACT

with

weight

Research

1 126

ROWLAND

Technicians

obtained

medical

examinations

aminees

were

and

weighed

foam-rubber

the

examinees

0. 1 cm
age

record
head.

and

to the

nearest

in

OH).

Height
level

both

to the

education

in these
years),

12

or 6-9).
Body

mass

included

was

regression

index
used

wt/ht2

(BMI;
both

analyses

the bivariate

and

BMI

in the consensus
ference

studies,

statement
was

here

19.2-27.2,
32.3 kg/rn2.

Limitations

ofthe

Both
lected

used

NIH

overweight

the

20.7,

as 27.3-32.2,

and

Con-

as 20.8-27.7,
as
3 1 . 1 kg!
19. 1 , midweight
severe

and

standardization

interview

overweight

and

processes

were

quality-control

weight.
be time

or loss ofseveral
kilograms.
in the sample
may have had
Many

A small percentage
some slight growth

examinees

to be weighed,

may

leading

have

to more

during
the time between
interview
occurred,
leading
to overreporting
in a face-to-face

questionnaire,

interview

possibly
were

in the

of this
influence.

to reporting

than

reporting

were

Dieting

accuracy.

may

times

have
were

influenced

were

This

to reporting

self-reported

weight

to estimate
and

height.

The

validity

and

variables

having

variables

still

stage

measures

of validity

was

(15).

at the
the

into

elsewhere(l6,

17, 19).

The

independent
the control
analyses

variable
used

affected

study (2) and follow from the classic test-theory


measured
height
and weight
are considered
values
[because
the test-retest
reliabilities

analysis
is not
Unweighted

unchanged

indication

and

two-way
entered

the dependent

the analyses,

variable

and

a logarithmic

(1 5). Because

with

the

results

transformation,

reported
samples

of the

the

adequacy

here.
are also
of the

the
reported
analytic

In the

in the

first-

design,

interaction
into

ie,

as described
terms

in-

multiple-regresTo assess

whether

of the dependent
data

the

of
all

by use of pro-

distribution

normalize

set
with

level.

sample

variable.

in the

the

(R2),

groups,

were

to stabilize

pendent

complex

variables

normality

for performing
selects

screening

of selected

in a

a standard

chosen

variables

against
from

the

were

values)

confidence

set ofvariables

account

oversampling

the

related

used

power

to further

and

measured
stage

procedure

95%

The

account
models

explanatory

subjected

take

clustering

reliability
(and the underlying
assumptions)
used in this study
are the same as those used by Stewart
in her excellent
reliability
model.
Explicitly,
the true (criterion)
were
1 .00 for both

significant

evaluated

significantly

(1 8). MAXR

largest

ofscreening,

that

and

the

the
mul-

was calculated
were

approach

were

SAS

was

When

t value

minus

MAXR

the

variable

significant.

first

used
signifi-

finding

Interactions

The

regression

was

were

dependent
the

conrelated

variable

for the

(14).

this

strongly

of statistical

5% that

(self-reported

program,

grams

of

ofthis
Tests

the critical

process.

errors

for

age was

multiple-regression

multiple

Mea-

reliability

standard

to control

variables

stepwise,

volving

the

error

in con-

(ie, reporting

and

(ANCOVA)

computer

deviations
seeks

made,

which

screening

of the

weight.

units

variable

as statistically

Weighted

re-

within those strata. Because


to reporting
error of weight,

of

approach

to determine

model

re-

to metric

used

of covariance

Multivariate.

the
of

preference).

programs
used in the analysis
took into
sample
design
of NHANES
II ( 16, 1 7).

stage

analysis
analysis

an analysis

inches.

because

end-digit

means

values

A probability

the Bonferroni

and

predictors

because

to contrast

colData

potential

categories

comparisons

po-

finally,

were

units

confounding.

tiple

this,

and,

measurements.

(ie,

and

were

ofheight,

was taken

After

in pounds

the dependent

ofchance

per-

the uniand

standards.

Similarly,

used

strata.

were

in these

converted

strata,

result

sion

Data

were

final

these measureUse of the word

have

done

done

the data

for possible

two-stage

in a self-administered

II at various

questions

off their

they

were

variable

error

to control

and height

in this study

and

for weight

into

weight

controlled

data

publication

divided

founding

used

are used

analyses

variables.
were

considered

show

AJCNs

categories

using

values

approach
was to screen
between
reporting-error

analysis

In general,

was

computer
complex

people
during

reporting.

of the year; therefore,


seasonal
variations.

interview

to round

rather

in NHANES

day and different


seasons
ments
reflect
diurnal
and

that

for both

ofweights

particularly

Screening.

with

self-report
weight
gain

and examination
may
of weight.
Questions

encouraging

made

known

accurate

sub-

between
contribself-re-

ofyoung
in stature

error

SD-of-

undertaken.

system

data

tables

cance

procedures,

This time lapse between


enough
for an actual

measured

of reporting
error were determined
BMI (wt/ht2)
was strongly
related

(1 1, 12). For men,

midweight

were

were

with

error)

The

correlation

gender-specific

was necessarily

error,

formity

data

ported
and measured
and weighing
could

spondents

themselves

to divide

some factors
may have affected
the data. The time lapse
interview
and examination
was 2-6 wk and may have
uted to the overall
variability
in the difference
between

about

units

between

to rigorous

surements

in the

Summary

Development

ofObesity

as

the measurement

period.

analysis

ie,

and

and examination

and

Data

index,

measured

bias.

product-moment

influences

in the English

porting

minus

estimates

body-measurement

analyses

and

weight

women,
analytical
relationships

and

interview

in

ofadjusting

was

1985

Implications

defined

ofihe

Although

asked

multivariate

the

ofreporting
and

confounding

variable

as Quetelets

method

overweight
as 27.8-3
1.0, and severe
overweight
m2. For females,
underweight
was defined
as

going

patterns

The general
and bivariate

in
1-4

four classes
of measured
weight.
The ratiowas an adaptation
ofcriteria
recommended

on the Health

underweight

variable

stratification

used

in epidemiological

weight

predictor

(defined

a commonly

measured

and

self-reported

for men

y),

value

and

of reliability.

demographic

12

weight
ending
in

(self-reported

transformation

distribution
of the

of the
analysis

untransformed
in order
sample.

to provide

dewere
data
an

Downloaded from ajcn.nutrition.org by guest on July 8, 2014

height

from

as a categorical

analyses.

in kg/rn2),

computed

as a continuous

analytic
sample
into
nale for classification

this

differed

tential

y,

error

of validity

between

preparation

height)

ject

coefficient

formed.
variate

(dichotomized-

(dichotomized-<

ofthe

measure
estimates

Because
as

nearest

analyses

race

mean

is the

as estimates

same

preference
(dichotomized-self-reported
ending
in 0 or 5, self-reported
weight

as

that

directly

0. 1 in.
considered

The

value)

the-difference

at the

end-digit
pounds

for

scale

Toledo,

measured

Ex-

uniforms

pound)

secured

(13)].

during
(10).

examination

Scale,

was

variable

black),

height
site

of a self-balancing

camera

Height

and

(to one-quarter

(Toledo

variables

(continuous

white,

by use

by a Polaroid

Independent

paper

the weight

a permanent
recorded

weight
examination

in disposable

printed

was

measured

slippers

mechanically
onto

the

at a mobile

SELF-REPORTED

WEIGHT

Results

Because

age

case,

age was

entered

The distribution
overall

ofreporting

tendency

average

was

for

men

of 0.4 kg, whereas

by an

average

of

error.

to overreport

women

1.0 kg. The

As shown

men

of 2.3 kg whereas

variability

overreported

severely

their

weight

underreported

their

in reporting

self-reported
minus measured
weight)
women:
the SD of the reporting
error
Reporting
error varied considerably
Underweight

in Table

groups.

by an

young

(eg,

BMI

weight

category):

kg for the age group

and

55-74
porting

men

underreported

their

error)

was

accompanied

in a skewing

in the direction

of reporting

error

for

End-digit

report

ofunderreporting

percent

18. 1% for the

of men

overreporting

underweight

to 4.5%

by
for the

severely

Conversely,
ranged from

the percent
of men underreporting
0.5% for the underweight
to 13.9%

overweight.

Relatively

few

women

overreported

greater

overweight.

by > 4.5 kg
for the severely

was

those

not

study

(ie, age, race,

-4.

COVA

was performed
to evaluate possible important
categories
of BMI in explaining
differences

end-digit

preference),

for men.

was

found

age group

association

For

women,

among

the

graduates
-4.2

the

severely

underreported
vs -2.6
kg,

in Table

shown

by chance

2, a strong

in self-reported
alone

ending

weight

people

that

in each

ending

with

extent

For men,

than

category (ie, 1-3%). The percent of women


underreporting
by > 4.5 kg ranged
from 1.0% for the underweight to 30.9% of the severely overweight.
For each ofthe other independent
variables
considered
in this
and

more

y). The average


men was -2.1

significant

education

observed

weights

greater

4.5 kg in any weight

education,

with

weight.

10% ofthe

of the digits

re-

digit

It would

sample

0 through

in 0 or 5. A comparison

error for the two groups


vs no end-digit
preference)

from

by more

y underreported

would

9. In

the analytic
sample
of 1 1 284, 6773 (60%) reported
weights
ending
in 0 or 5. Thus,
---40% of the sample
erroneously

and

4.5 kg ranged

>

weight

people,

y.

where
high school
than did nongraduates,

be expected

variability

(indicating

four

1.2 kg for the

55-74

preference-As

(indicating

in location

the

spectively.
preference

by a shift

20-34

y and

difference

overweight,
much
more

for

overweight

was no statistically
error

in each

overweight
women,
the average
rekg for the age group
20-34
y and

age group

significant

reported

The

with

kg for the

men and women were greater


midweight
men and women;
greater

20-34

Education-There
only

by age

severely

ages

severely
was -4.3

of reporting

bias).

>

y. For
error

-2.3

by an average

error,

men and women


(55-74
among
severely
overweight

(ie,

weight

differed

among

women

reporting

variable.

end-digit

than

did

there

overweight
1 kg for those

those

women

rounding

the groups

indicating

was a pronounced
category

who

with

(-2.0

end-digit

respond

to a

preference.

was small

that

rounding

as rounding

difference
kg for

no

and
was
down.

in the se-

preference

preference),
with

of 0 or 5
including

underreported

no end-digit

up was as common

For women,

interactions
in reporting

between

significant,

random-rounding

an AN-

with

of reporting

preference
the group

preference

those

the difference

statistically

verely

(ie, end-digit
shows
that

vs

indicating

that

a 5 or 0 preference

are

down.

TABLE

1
Mean, SD, and percent distribution
(United
States 1976-1980)

of error

for weight

estimation

(ie, self-reported

Deviation
Sex and measured-body-weight
category

Males
Total (n = 5396)
Underweight
(n
51 1)
Midweight
(n = 3552)
Overweight
(n = 892)
Severely overweight
(n = 441)
Females
Total (n
5888)
Underweight
(n
410)
Midweight
(n = 3702)
Overweight
(n = 1093)
Severely overweight
(n = 683)

The corresponding

measured)

weight

from

in kilograms,

measu red weight

by sex and BMI category

in kilograms5

2.3-4.4
kg
(too high)

2.2 kg
(correct)

-2.3--4.4
kg
(too low)

7.0
18.1
6.3
5.3
4.5

16.0
24.4
17.2
9.8
8.8

62.2
53.8
65.3
58.3
53.9

10.7
3.3
9.0
18.0
19.0

4.0
0.5
2.2
8.7
13.9

0.41
2.25
0.59
-0.51
-1.35

1.4
1.1
1.2
1.4
3.2

5.2
10.3
4.7
6.0
3.1

70.0
82.0
76.4

15.0
5.5
14.0
20.0
20.6

8.4
1.0
3.6
17.2
30.9

-1.03
0.36
-0.63
-1.88
-3.41

>4.5 kg
(too high)

categories

of self-reported

minus

in pounds:

>

10.0, 5.1-10.0,

55.5

42.2
-5.0-5.0,

-l0.0--5.l,

and

<

-10.0.

-4.5 kg
(too low)

<

SEM

SD
kg

0.04
0.17
0.05
0.14
0.19

3.02
2.85
2.70
3.30
3.65

0.04
0.10
0.04
0.12
0.22

3.02
1.96
2.12
3.61
5.24

Downloaded from ajcn.nutrition.org by guest on July 8, 2014

in reporting

and
older
error

and

the SDs

overweight
and severely overweight
than those for underweight
and
the difference
in size of the SDs

men

error

weight by an average of 1.4 kg. Underweight


women overreported
their weight by an average of0.4 kg whereas severely overweight
women
underreported
their weight by an average of 3.4 kg.
The distribution
of reporting
error shown in Table 1 and Figures 1 and 2 also varied systematically
with measured
weight
status

error

In particular,

than did
reporting

with

as a control

Age-Reporting

I, the
weight

is associated

was similar for men


was 3.0 kg for both.
with weight status
their

overweight

1127

HEIGHT

error.

Weight

sex.

AND

1 128

ROWLAND

E
.2!
.E

0.

Midwoight

Overweight

Sever.
overweight

FIG 1. Box graph of the distribution


of reporting
error for weight according
to measured
men aged 20-74 y, 1976-1980.
0, SOth-75th
percentile
ofthe distribution
ofreporting
error,

weight category
for US
0, 25th-SOth
percentile.

E
0.
0!

.E
2!

to
0.
0

Underweight

Midweight

FIG 2. Box graph of the distribution


of reporting
error
women aged 20-74 y, 1976-1980.
G, SOth-75th
percentile
percentile.

Overweight

Severe
overweight

for weight according


to measured
weight category
for US
range of the distribution
of reporting
error; 0, 25th-SOth

Downloaded from ajcn.nutrition.org by guest on July 8, 2014

Underweight

SELF-REPORTED
TABLE 2
Mean error of weight
States l976-l980)

estimation

(kg) and sample

WEIGHT

size for people

aged 20-74

AND

1 129

HEIGHT

y by body-mass-index

End-digit

category,

sex, and end-digit

1 (underweight)
2 (midweight)
3 (overweight)
4 (severely

overweight)

Analysis

No

Yes

2.34
(194)
0.77
(1271)
-0.26
(294)
-0.62
(131)

2.21
(317)
0.50
(2281)
-0.63
(598)
-1.66
(310)

Women
Combined
2.25
(511)
0.59
(3552)
-0.51
(892)
-1.35
(441)

of covariance

No

Yes

0.31
(228)
-0.38
(1770)
-1.13
(401)
-2.00
(222)

0.43
(182)
-0.85
(1932)
-2.29
(692)
-4.14
(461)

Combined
0.36
(410)
-0.63
(3702)
-1.88
(1093)
-3.41
(683)

<0.001

End-digit
Interaction
Covariate

<0.553
<0.139
<0.544

preference)

<0.00 1
<0.001
<0.001
<0.001

in parentheses.

t End-digit preference:
0, 1 (last digit other than 0 or 5, last digit 0 or 5).
1:Age was treated as a continuous
variable expressed
in years of age.

Race-Reporting
bias did not differ significantly
(P < 0.05)
between
whites and blacks of either sex for the four weight
groups when age was controlled
for. However,
there was
an observed
difference
between
black and white men at the
extremes
ofthe weight distribution:
underweight
black men
overreported
more than did white men (3.3 vs 2. 1 kg), and

severely
did

overweight

black

men

white

men

1 .4 vs -0.9

(-

underreported

more

than

kg).

ference

from
between

term

was

the

identity

the slopes

significantly

line.
for men

less

There

was

and

women,

for women.

no

significant

dif-

but the constant

A graphical

comparison

of the sex-specific
regressions
(not shown)
indicates
that men
overreport
at the lower end of the distribution
and underreport
at the upper end of the distribution,
whereas
women
are fairly
accurate
at the lower end ofthe weight distribution
but increasingly

underreport

relations

for men and


A stepwise

variables
reported
measured
variables
end-digit

and weight.
with

BMI

at the

between
0.98

end

ofthe

and

distribution.

measured

weight

The

cor-

were

0.97

for women.

multiple
regression
was performed
to explore which
would predict the discrepancy
between
measured
and
weight (the dependent
variable
is self-reported
minus
weight).
Sex-specific
models
were fit. Independent
entered
into the analyses
were age, race, education,
preference,

Two-way
(derived

and

the race-by-BMI

BMI

interaction
from

entered
into the regression.
For men, the significant
and

upper

self-reported

derived

terms

measured

predictors

interaction;

error
for

BMI,

age, age-by-BMI,

indicate

that

accounted
women.

actual

from

measured

for each ofthese

weight

(P

<

for women,

and

height)

0.05)

were
significant

height

and

weight

ported weight for men,


weight on self-reported
presented
here. Note,
prediction
prediction
in correcting
dom

States collect
the prediction
(proportion

R2 values

were

weight.
actual
weight

Because

weight,
the
on self-re-

here

been

more
error

with
most

nor
than

multiple-regression
marginally

resulting

self-reported

from

and

ran-

(see discussion

values

epidemiological

successful
bias

studies

in the

United

reported
weight and height in pounds
and inches,
equations
are expressed
in those units.) The R2
of variance

dictor
variables)
respectively.
For men, the

Measured

self-reported
to estimate
of measured

presented
have

associated

in ref 8). (Because

partial

0.16

weight,

respectively.

from

misclassification

error

and

and

and the prediction


equation
of measured
weight and age in years for women
are
however,
that neither
the simple
linear

equations
equations

height

predictor:
women,

The

in reporting

is 0. 12 for men

measured

may want
equation

preference.

of variation

variables
from

important

0. 1 3 for men

Estimating

and end-digit

proportion

derived

was by far the most


0. 1 1 and

the

for by these
BMI,

some
investigators
regression
prediction

Predicting
reporting
error. The simple linear regression
of selfreported
weight on measured
weight for men and women
separately is shown in Table 3. Both regression
lines differed
significantly

tors were
R2 values

were

simple

linear

weight

weight

0.95

and

explained

0.96

regression

for men

by the
and

pre-

women,

was

(pounds)

= -4.1259
For women,

in actual

values

+ 1.0185

the simple

(self-reported

linear

weight

regression

in pounds)

was

variables
were
BMI,

also
race,

predic-

Measured

= -3.

weight
1974

(pounds)

+ 1 .0438

(self-reported

weight
-

in pounds)

0.0175

(age

in years)

Downloaded from ajcn.nutrition.org by guest on July 8, 2014

BMI category
preferencet
(BMI-digit
ages

(United

preference

Men
BMI category

preference

1 130

ROWLAND

TABLE 3
Distribution

and regression

parameter

Sex

estimates

Percentage
overweight

Agef
y

Weight (kg)
Male (n = 5396)
Female (n
5888)
Height (cm)
Male (n = 5396)
Female (n = 5888)

42
43

42
43

for measured

and self-reported

height

and weight

Measured
variable (X)t

Reported
variable (Y)t

78.52
65.45

78.93
64.41

for NHANES

II (United

95% Confide

nce interval

Intercept

States

l976-l980)

Slope

Sy.x

Ty

15
16

24
27

15
16

24
27

0.7
0.7
0.7
0.7

13.34
14.24

175.5
161.8

7.1
6.6

177.0
162.3

12.75
13.38

5.782
4.345

7.4
6.6

5.415
12.517

0.648
0.497

0.932
0.918

2.824
2.283

0.977
0.926

0.008
0.008

0.974
0.978

2.88 1
2.783

0.016
0.014

0.938
0.910

2.581
2.767

S Measured
weight is denoted
by X, reported
weight by Y. The 95% confidence
intervals
for both slopes and intercepts
are for the regressions
on X. rxy refers to the correlation
coefficient
of X and Y, whereas
S1. x is the root mean square error of the regression
of Y on X.

t I

SD.
SEM.

Height

overreported

by an average

2.4 cm

in the

65-74-y

age group.

The distribution
of reporting
error shown in Table 4 varied
across age groups: although
the SDs varied within a narrow range
across age groups (2.3-3.3
cm), indicating
little change in the
spread of the error distribution
with age, with increasing
age,
the location

ofthe

distribution

shifted

in the direction

of overre-

porting.

Most subjects
reported
heights were within 2.54 cm of their
measured
height, though the percent within this range decreases
with age as the distribution
of reporting
error skews towards
overreporting
cm increased
Similarly,

height.
The
from 22.7%
the

percent

percent
at age

of men overreporting
>
2.54
25-34
y to 46.3%
at 65-74
y.

of women

overreporting

by

2.54

increased
from 9.7% at 20-24 y to 43.6% at 65-74 y.
For each ofthe other independent
variables
considered
study (ie, race, education,
measured
height, and measured
ANCOVA

interactions
reporting

was

performed

to evaluate

with categories
of age
error.
(For the ANCOVA,

categories:

20-34,

35-54,

and

in this
BMI,

important

in explaining
differences
in
age was grouped
into three

55-74

Education-Reporting

school graduates
and non-high-school
graduates
of either
sex for the three age groups.
Race-Reporting
bias did not differ between
whites and
blacks of either sex for the three age groups.
Measured
height-For
men aged 55-74 y and women aged
35-74
y, people
below median
0.5 cm more than did those

BMI-For
age group

reported

height

not

differ

between

high-

height overreported
by nearly
above
median
height.

both men and women, overweight


people in each
overreported
their heights by ---0.5 cm over their

nonoverweight

Predicting

did

y.)

bias

possible

cm

counterparts.

reporting

error. The

on measured

simple
linear regression
of selfheight (Table
3) differed
from unity

for men and women,


demonstrating
a slight tendency
to overreport height at the lower end ofthe height distribution.
The correlations
between
measured
and reported
height were 0.94 for
men and 0.9 1 for women.
Systematic
differences
in reporting
error were evaluated
by
using

a stepwise

multiple-regression

approach

for each

sex (the

dependent
variable
was self-reported
minus measured
height).
Independent
variables
included
in the analysis
were age + age2
(continuous),
race (dichotomous),
education
(dichotomous),
measured
height (continuous),
and BMI derived from measured
height
each

and

weight

ofthese

(continuous).

variables

with

Two-way
age were

also

interaction
entered

terms

into

for

the regres-

sion.

For men, the significant


predictors
were age, age-squared,
and
for women,
significant
predictors
were age, age-squared,
BMI, and the race-by-age
interaction.
The main effect, race, was
not statistically
significant,
but was included
in the final model
because
of its interaction
with age. The R2 values (proportion
of variability
in reporting
error for height explained
by the
BMI;

regression

equation)

were

small:

0.04

in men

and

0. 10 in women.

Age and age2 were the most important


predictors:
partial R2
values were 0.03 and 0.09 for men and women,
respectively.
Estimating
actual height from self-reported
height. Genderspecific

prediction

variable)
here.

The

equations

on self-reported
proportions

for

height

measured

ofvariance

= +7.

height
1987

(self-reported

+ 0.0222

Measured

= +7.4583

the linear

height

height

were

0.89

are given
accounted

and

0.85

for

(inches)

+ 0.8865

For women,

(dependent

variable)

in measured

for by self-reported
height (ie, R2 values
men and women,
respectively.
For men the linear equation
was
Measured

height

(independent

equation

height

in inches)

(age in years)

0.0004

(age2)

was

(inches)

+ 0.8745

(self-reported
+

0.0424

height

(age in years)

in inches)
-

0.0007

(age2)

Downloaded from ajcn.nutrition.org by guest on July 8, 2014

The distribution
of reporting
error. On average,
men and
women overreported
their height by 1.4 and 0.6 cm, respectively.
The distribution
of reporting
error was similar
for men and
women:
the SDs of differences
were 2.59 cm for men and 2.81
cm for women.
As seen in Table 4, adults aged 45-74 y overreported more than did adults aged 20-44 y. Both men and women

and

of

SELF-REPORTED

TABLE 4
Mean, SD, and percent
(United

States

distribution

oferror

WEIGHT

for height estimation

(ie, self-reported

>5.1 cm
(too high)

=
=

=
=
=
=
=

minus

measured)

in centimeters,

by sex and age category

height

from

measured

height

in centimeters

2.6-5.1
cm
(too high)

-2.5-2.5
cm
(correct)

-2.6--5.l
cm
(too low)

5396)
655)
1039)
723)
665)
1 166)
I 148)

5.7
5.6
2.7
4.3
5.4
7.8
13.3

22.9
17.9
20.0
20.5
23.9
27.5
33.0

67.0
69.6
73.2
70.6
67.2
60.7
49.7

3.4
5.8
3.5
3.7
2.3
2.8
2.1

1.1
1.1
0.6
1.0
1.2
1.2
1.9

1.44
1.23
1.15
1.18
1.48
1.78
2.43

5888)
678)
1088)
803)
724)
1261)
1334)

4.9
1.1
2.7
1.7
3.8
6.5
18.0

13.0
8.6
8.2
8.7
12.1
20.9
25.6

74.4
80.8
80.3
81.0
77.3
66.4
51.6

6.2
8.2
7.6
7.5
5.4
4.0
3.5

1.4
1.3
1.2
1.2
1.5
2.2
1.2

0.57
-0.03
0.06
-0.01
0.47
1.28
2.42

Sex and age in years

(n
(n
(n
(n
(n
(n
(n

1131

HEIGHT

1976-1980)
Deviat ion of self-reported

Males
20-74
20-24
25-34
35-44
45-54
55-64
65-74

AND

<

-5.1 cm
(too low)

SEM

SD
kg

0.05
0.16
0.07
0.09
0.1 1
0.08
0.10

2.59
2.94
2.24
2.33
2.46
2.68
3.04

0.06
0.09
0.07
0.1 1
0.10
0.1 1
0.13

2.81
2.19
2.47
2.47
2.68
3.12
3.30

Females

(n
(n
(n
(n
(n
(n
(n

=
=
=
=
=
=
=

The corresponding

categories

in inches

are as follows:

2.0,

>

1.1-2.0,

1.0-1.0,

Discussion
These

findings

extend

those

for the same NHANES


that when self-reported
expressions

of relative

ofan

study

weight

(Metropolitan

index), overreporting
act together
to make

combined

effect

Weight

in the

numerator

and

with

4.5 kg.
Severely
overweight

women

tend

more

the

and

rounding

error

have

on

reporting

error.

liability)

ulation
S

in self-reported

values

were

found

in important

pop-

subgroups:

overweight
average

and relative
weight were predictors
in weight.
Heavier
people underreport

weight
more
Overweight

than do lighter
people.
and severely
overweight

their weight
Proportionately,

more

weight
women
vs 14%).

of retheir

women

underreport

as men

underreport

by

4.5

kg

class.
over(31%

their

weight

a high-school

by

women

(4.3

aged

kg) than

20-34

On

Approximately

their

height

by

the average,
heights

overweight

slightly

do

aged

more

more

55-74

than

y (2.3

do younger

aged 65-74

y overre-

cm).

than

did

tended

to overreport

midweight

and

under-

people.

A strong
digit
preference
weight.
Overweight
women

to round
and

y underreport

people

more

kg) than

category.

do women

1 in (2.4

>

education

(4.2

weight

45% ofpeople

down

women,

was observed
in particular

to the nearest
rounding

self-reported

a reduced

it more

difficult

5 or

resulted

weight,

expect

with

power

10 lbs.

For

in greater

the

result

that

men
of

researchers

tests,

associations

both

unreliability

that

in statistical

to find

in self-reported
showed
a tendency

thereby

can

making

actually

exist.

As other authors
have conjectured,
the misreporting
of weight
appears
to be in the direction
of cultural
ideals-overweight
men

and

school
port

women

underreport
being

education;
their

physiological

than do men in the same


weight
more
than twice
as many
severely

the average

people.

weight
S

with
on

(2.6 kg) in the same

underreporting

Both absolute
porting
error

overreport

group.
their height

their

Al-

though
height and weight were reported,
on the average,
with
small errors, larger systematic
differences
in the average reporting
error (or bias) and in the distribution
of reporting
error (or re-

men

kg) in the same weight


Older people overreport
port

(8).

status,

to underreport

on the

By considering
interactions
between
predictors
of reporting
error in self-reported
height and weight, the present study further
clarifies
the roles that sex, age, education,
measured
height,
weight

1 8% ofunderweight

e Severely

overestimate
in the denominator
for relative weight leads to distortions
in the distribution
ofrelative
weight status across weight
groups, which in turn leads to misclassification
ofweight
status
as well as overestimation
of underweight
and underestimation
ofoverweight.
The biological
significance
ofthis misclassification
to epidemiologists
who are concerned
with measures
of association such as relative
risk and attributable
risk has been discussed

-2.

<

Nearly

nongraduates

of height as well as misreporting


relative weight inaccurate.
The

underestimate

>

(8) by the author

Relative

and

II data set. The earlier study showed


height and weight are combined
into

Quetelets
of weight

ofthe

earlier

-2.0--l.l,

weight.

their

greatest

on the other
The

basis

weight,

among

hand,

misreporting

due

to the

the magnitude

women

with

underweight
of height

effects

of age

men

appears
and

of the

at least

a high-

overreto have

gravity:

older

people possibly report their heights as they remember


them from
an earlier age before the shrinkage
caused by osteoporosis
occurs
(20,

2 1). In a longitudinal

ministration

a decrease

Normative

in height

analysis
Aging

Study,

based

on

Borkan

of 2.6 cm between

the

Veterans

Ad-

et al (2 1) estimated

ages 45 and

75 y in a

Downloaded from ajcn.nutrition.org by guest on July 8, 2014

20-74
20-24
25-34
35-44
45-54
55-64
65-74

1132

ROWLAND

TABLE 5
Characteristics

Authors

of eight

published

(publication

studies

of self-reported

Percent
sample

date)

and
size

40t

0.99
NA
NA

-l.201
NA
NA

3.10
NA
NA

100(3226)
46
54

15-61

0.99
0.98
0.98

-1.09
-0.73
-1.41

2.49
2.42
2.57

100(3407)
47
53

20-59

NA
0.96
0.97

-1.23
-0.54
1.86

NA
3.24
3.03

100(1337)
63
37

30-69

NA
NA
NA

NA
-1.50
-2.40

NA
3.20
4.00

100(11284)
48
52

20-74

0.98
0.97
0.98

-0.34
0.41
1.03

3.11
3.02
3.02

100(1519)
64
36

35-65

0.98
NA
NA

-0.58t
NA
NA

NA
NA
NA

in US health

insurance

study

Face-to-face
population

interview
study

100(752)
79

Women

16-66

NA
0.95-0.96
0.91-0.97

21

et al (1981)

Jalkanen
et al (1987)
Total
Men
Women

NA
NA
NA

program

interview

population

measured

NA
NA
NA

-0.50
-0.40
-0.60

interview

in New
study

Questionnaire

in Danish

applicants

for medical

NA
NA
NA

insurance

Questionnaire
in Danish
applicants
for medical
insurance
Questionnaire
in Eastern
population
study

[6]
30-64

in US

study

population

Zealand

[5]

100(11880)
49
51

in urban

Face-to-face
interview
in
hypertension
screening

Face-to-face

Women
Stunkard
and Albaum
(1981) [7]
Total
Men

minus

in US medical

NA
3.00
2.00

Finland

weight.

t Mean.
No differencebetween sex groups.

Not available.
white

male

of2.4

cm

sample.

NHANES

reported

height

Seven

weight

sample

extend

tributions.

reported

compares
and

women

II studies
and

representative

States
where

men

with

an

aged

65-74

overreporting

bias

which

also

(2-5,

y in the present

of validity

[The

of the United

height

other
weight

from
less
have

reliability

are the first to be done

to populations

estimated
(10)].

and

with
and

weight

NHANES

representative
appeared

States.

similar

The

height

distributions

II are

results
and

of this

weight

of the

dis-

United

else-

studies

of self-

literature

(1-7),

five

high

of

studies

correlations

sured

in a nationally

documented

large-sample
in the

of self-

included

an

evaluation

of

self-reported

height

7).

These

study.
The

study

This

for both

height

in estimating

relative

are consistent

with

of self-reported

and weight.
the

magnitude

the present

height

However,
ofbias

one

in finding

and

weight

with

they differ

among

themselves

and

unreliability

and

meain the

importance
of such predictor
variables
as age, education,
measured
height, measured
weight, and relative
weight (computed from measured
weight and height).
As shown in Table 5, among
the US studies,
reporting
bias
for weight
was generally
smaller
in the NHANES
II than in
other studies. This indicates
that the present study may be con-

Downloaded from ajcn.nutrition.org by guest on July 8, 2014

100(550)
19
81

[1]

Total
Men

and sample

kg

Face-to-face

Men
Women
Foreign studies
Stewart et al (1987)

Reported

of method

Questionnaire

Women

Description

SD

[2]

Pirie et al (1981) [3]


Total
Men
Women
Palta et al (1982) [4]
Total
Men
Women
Rowland
(1989) [8]
Total

Schlichting

Mean
difference5

Questionnaire
patients

Men
Women
Stewart (1982)
Total
Men

Correlation
coefficient
of reported
and
measured
weight

Age group

%(n)
US studies
Stunkard
and Albaum
(1981) 171
Total

weight

SELF-REPORTED
sidered

as a lower

in studies

relying

bound

for estimating

on self-reported

The implications
ofthese
studies
relative weight based on self-reported
used as continuous
variables,
given
measured

height

and

weight.

WEIGHT

the error that may appear

weight.

are that height, weight, and


height and weight are better
their high correlations
with

As categorical

variables,

self-reports

are subject
to misclassification,
a problem
compounded
by
subgroup
differences.
The present study presented
the distribution
ofreporting
error
in detail and used more than one approach
to enable researchers
to evaluate
for themselves
the adequacy
of self-reported
height
and weight relative to their own study designs and goals and to
anticipate
potential
areas of bias.
N
I wish
Schoenborn

to thank

Katherine

Regal,
Dedun
Ingram,
and Charlotte
Center for Health Statistics;
Willis Foster
ofHealth; and David Williamson
ofthe Centers

of the National

ofthe National Institutes


for Disease Control for their reviews of the manuscript.

References

reported height and weight with controlled


women and men. Int I Obes 198 l;5:67-76.
6. Jalkanen
reported

height

and weight

in

L, Tuomilehto
I, Tanskanen
A, Puska P. Accuracy
of selfbody weight compared
to measured
body weight. Scand I
SocMed
1987;15:191-8.
7. Stunkard
Al, Albaum
JM. The accuracy
of self-reported
weights.
Am J Gin Nutr 198 1;34: 1593-9.
8. Rowland
M. Reporting
bias in height and weight data. Stat Bull
Metropol Life Ins Co. I 989;70(2):2-1
1.

1133

HEIGHT

9. McDowell
A, Engel A, Massey IT, Maurer
K. Plan and operation
ofthe Second National
Health and Nutrition
Examination
Survey,
1976-80.
Hyattsville,
MD: National
Center for Health Statistics,
1981. (Vital and health statistics
series 1, #15 [DHHS
publication
(PHS) 81-13 17].)
10. Najjar MF, Rowland
M. Anthropometric
reference
data and prey-

alence ofoverweight,
United States, 1976-80. Hyattsville, MD: National Center for Health Statistics, 1987. (Vital and health statistics
series 1 1, number
238 [DHHS publication
(PHS) 87-1688].)
1 1. Rowland ML. A nomogram
for computing
body mass index. Dietetic
Currents
1989;16(2):
1-8.
12. Consensus
Development
Panel. Health implications
ofobesity:
National Institutes
of Health
Consensus
Development
Conference

statement.

Ann Intern

Med 1985;l03:

1073-7.

13. Marks cx::, Habicht


J-P, Mueller
WH. Reliability,
dependability,
and precision ofanthropometric
measurements:
the second National
Health and Nutrition
Examination
Survey,
1976-80.
Am I Epidemiol l989;l 30:578-87.
14. NeterJ, Wasserman
W. Applied linearstatistical
models Homewood,
IL: Richard D Irwin, Inc, 1974.
15. Kleinbaum
DO, Kupper
LL. Applied regression
analysis and other
multivariable
methods.
North Scituate,
MA: Duxbury
Press, 1978.
16. Shah BY. SESUDAAN:
standard
errors program
for computing of
standardized
rates from sample survey data. Research
Triangle Park,
NC: Research
Triangle Institute, 1981.
17. Holt MM, revised by Shah BY. SURREOR:
standard error of
regression
coefficients
from sample survey data. Research
Triangle

Park, NC: Research

Triangle

Institute,

1982.

18. SAS Institute


Inc. SAS users guide: statistics.
5th ed. Cary, NC: SAS
Institute
Inc. I 982.
19. Landis JR, Lepkowski
JM, Eklund SA, Stehouwer
SA. A statistical
methodology
for analyzing
data from a complex
survey: the first
National
Health and Nutrition
Examination
Survey. Hyattsville,
MD: National
Center for Health Statistics,
1982. (Vital and health
statistics series 2, number
92 [DHHS publication
(PHS) 82-1366].)
20. Freidlaender
IS, Costa PT Jr, Bosse R, Ellis E, Rhoads JO, Stoudt
HW. Longitudinal
physique
changes among healthy white veterans

at Boston.
21.

Hum Biol l977;49:541-58.

Borkan GA, Hults DE, Olynn


secular trend in age differences
l983;55:629-41.

Ri. Role oflongitudinal


in male body dimensions.

change
Hum

and
Biol

Downloaded from ajcn.nutrition.org by guest on July 8, 2014

1. Stewart AW, Jackson


RT, Ford MA, Beaglehole R. Underestimation
of relative weight by use of self-reported
height and weight. Am I
Epidemiol
l987;l25:
122-6.
2. Stewart AL. The reliability
and validity ofself-reported
weight and
height. J Chron Dis 1982;35:295-309.
3. Pirie P, Jacobs D, Jeffery R, Hannan
P. Distortion
in self-reported
height and weight data. J Am Diet Assoc 198 l;78:60l-6.
4. Palta M, Prineas
RI, Berman
R, Hannan
P. Comparison
of selfreported and measured
height and weight. Am I Epidemiol l982;l 15:
223-30.
5. Schlichting
P, Hoilund-Carlsen
PF, Quaade F. Comparison
of self-

AND

You might also like