Você está na página 1de 12

Volume 139

Number 10
May 15, 1994
American Journal of
EPIDEMIOLOGY
Copyright 1994 by The Johns Hopkins University
School of Hygiene and Public Health
Sponsored by the Society for Epidemiologic Research
REVIEWS AND COMMENTARY
Telephone Sampling in Epidemiologic Research: To Reap
the Benefits, Avoid the Pitfalls
Richard F. Potthoff
Recently, the use of telephone sampling methods in epidemiology has been sharply
increasing. Properly applied, these methods provide powerful tools. Improperly applied,
they may produce invalid results. This review covers many points to which the inves-
tigator should be alert. An underlying theme is that bias in studies that use telephone
sampling can potentially spring from many sources and should be avoided wherever
feasible. In epidemiology, there are two main uses of telephone samplingin general
surveys (cross-sectional studies) and in case-control studies. For the former, the prin-
ciples differ little from those for general surveys in other fields. For the latter, most of the
same principles apply, but case-control studies also have their own unique aspects. In
this review, several topics receive detailed treatment. Valid combinations of area code
and prefix can be found through careful processing of a file that is available commercially.
Three options that can be used singly or in any combination provide broadened adapt-
ability for the Mitofsky-Waksberg method of random digit dialing. Bias can be thwarted
by certain steps in the interviewing and by weighting. For population-based and then
center-based case-control studies, a scheme that samples controls from household
censuses and avoids usual problems is offered. Am J Epidemiol 1994;139:967-78.
bias (epidemiology); case-control studies; cross-sectional studies; data collection; epi-
demiologic methods; interviews; sampling studies; telephone
In epidemiologic and other health-related
research, the use of telephone sampling (1,
2)generally based on some form of ran-
Received for publication June 8, 1993, and in final
form February 9, 1994.
Abbreviations: PSU, primary sampling unit; RDD,
random digit dialing.
From the Center for Demographic Studies, Duke Uni-
versity, Durham, NC.
Reprint requests to Dr. Richard F. Potthoff, Center for
Demographic Studies, Duke University, 2117 Campus
Drive, Box 90408, Durham, NC 27708-0408.
dom digit dialing (RDD)has grown rap-
idly in recent years. Telephone sampling is
done for both large and small studies and by
investigators both experienced and inexpe-
rienced in its use. Many pitfalls await the
unwary. Most of them will be indicated, im-
plicitly if not explicitly, in what follows. The
main purpose of this review is to describe
suitable telephone sampling practices de-
signed to avoid pitfalls that could lead to
damaging bias. To obtain full benefit and
forestall potential criticism, one should
967

a
t

C
o
m
m
i
s
s
a
r
i
a
t

A

L
'
E
n
e
r
g
i
e

A
t
o
m
i
q
u
e

o
n

S
e
p
t
e
m
b
e
r

2
9
,

2
0
1
4
h
t
t
p
:
/
/
a
j
e
.
o
x
f
o
r
d
j
o
u
r
n
a
l
s
.
o
r
g
/
D
o
w
n
l
o
a
d
e
d

f
r
o
m

968 Potthoff
strive to do telephone sampling in the best
way possible and to remove readily correct-
able flaws, even if a telephone method with
such flaws is already better than nontele-
phone alternatives.
This review will suggest rough criteria for
telephone sampling in thorough, high-qual-
ity research. Hard-and-fast rules will not be
proposed because individual situations may
vary and judgment may be needed: In par-
ticular, one faces trade-offs between cost
and accuracy, between two costs, or be-
tween two aspects of accuracy.
Over the past decade, the article by Waks-
berg (3) on the Mitofsky-Waksberg RDD
method has been cited in more than 200 re-
ports of health-related studies, most of them
epidemiologic. These and other reports have
also noted the use of other RDD methods. In
many cases, it appears that the telephone
sampling was properly carried out. In other
cases, though, the description of telephone
sampling is vague or too brief, reveals a
flaw, indicates that a variation of the
Mitofsky-Waksberg method was used but
fails to explain the nature of the variation, or
implies that the Mitofsky-Waksberg method
was used but leaves one to wonder if it really
was. Thus, in a number of the published
reports there may not be enough detail for
the reader to assess whether the telephone
sampling was done properly.
There are two main types of applications
of telephone sampling in epidemiologic re-
search. Each will be treated in turn. First,
and less complex, is the general surveythe
cross-sectional or prevalence study (4), in
which one surveys the general population or
some subset of it. The principles of tele-
phone sampling for such a survey are es-
sentially the same as for market-research,
social, and political surveys. Second is the
case-control study. Certain issues concern-
ing telephone sampling for case-control
studies are unique to those studies.
For general surveys, this review will dis-
cuss ways, both good and questionable, of
obtaining a sample; how to set up the RDD
sampling frame (i.e., the conceptual list of
population telephone numbers from which
the sample is to be drawn); the Mitofsky-
Waksberg method and several ways to ex-
tend it to make it more flexible; sampling
issues that arise during interviewing; and
weighting. For case-control studies, the em-
phasis will be on those issues in telephone
sampling that do not arise in general sur-
veys. Population-based and then hospital- or
center-based case-control studies will be
covered, with particular attention to im-
provement of sampling schemes for con-
trols.
This review is not intended to deal with
statistical analysis, except for a few matters
of weighting that are related to sampling.
Special aspects of statistical design and
analysis that are peculiar to case-control
studies are covered elsewhere (4-9).
Telephone numbers in the United States
have 10 digits. The first three comprise the
area code. The next three will be called the
prefix; although the term exchange is some-
times used instead, that term is best avoided
because it also refers to a set of one or more
prefixes that serve the same territory. The
last four digits may be called the suffix.
GENERAL SURVEYS
Numerous reports of epidemiologic and
other health-related RDD general surveys
have appeared in the recent literature. Al-
though sample selection for these surveys
was by RDD, interviewing and other fol-
low-up could be either by telephone or in
person. The surveys have pertained to such
subjects as diet, tobacco, alcohol, caffeine,
injuries, disability, depression, headaches,
mammograms, prescription drugs, contact
lenses, radon in homes, health practices,
health knowledge, and health care. One
large study (10-12) dealt with a number of
behavioral risk factors.
Ways to get a sample
Different ways have been used to get
samples of residential telephone numbers.
Several will be dealt with here, but further
detail appears elsewhere (1315). Many of
the ways have major drawbacks.

a
t

C
o
m
m
i
s
s
a
r
i
a
t

A

L
'
E
n
e
r
g
i
e

A
t
o
m
i
q
u
e

o
n

S
e
p
t
e
m
b
e
r

2
9
,

2
0
1
4
h
t
t
p
:
/
/
a
j
e
.
o
x
f
o
r
d
j
o
u
r
n
a
l
s
.
o
r
g
/
D
o
w
n
l
o
a
d
e
d

f
r
o
m

Telephone Sampling in Epidemiologic Research 969
One method is to draw numbers from tele-
phone directories. This method is not RDD
and is generally avoided, mainly because the
omission of numbers not in directories may
lead to serious bias (3, 13, 14), but also be-
cause of overlap among directories plus
clerical difficulties.
A less objectionable method again in-
volves drawing numbers from directories,
but one replaces the last one or two digits of
each phone number with digit(s) randomly
chosen. This method has received much use
in some fields. It produces a high yield (a
high ratio of residential numbers to numbers
dialed), so dialing time and cost are low.
Unfortunately, though, different numbers
have unequal (and unknown) selection prob-
abilities. As a result (14, 15), the method
does not remove all of the bias against num-
bers that are not in directories, and it is bi-
ased as well against numbers in blocks that
have fewer residential numbers (where a
block, also called a bank, refers here to a
group of 100 or 10 numbers with the same
first eight or nine digits). The method is not
recommended for use in careful research.
Another way to get a sample is through
commercial means. This avenue may be
ruled out because of cost. If not, one should
proceed cautiously. Consultation with a
competent statistician familiar with sam-
pling theory is no less advisable with this
approach than with others, nor is one re-
lieved of the obligation to report enough de-
tail about the sampling for readers to judge
whether the sampling was properly done. In
some cases, a sample obtainable through
commercial means is more or less a com-
puterized equivalent of drawing from direc-
tories and replacing the last two digits. In
these cases, the pros and cons are just as
before: The yield of residential numbers is
higher than with proper RDD methods (a
strong selling point), but the sample is bi-
ased.
Various sampling methods (13, 15, 16)
are based wholly or partly on "working
bank" datacommercially prepared infor-
mation that is supposed to show how many
listed residential numbers there are in every
group (bank) of 100 phone numbers with the
same first eight digits. High cost of the data
and of associated processing may render the
methods uneconomical except in organiza-
tions that do a heavy volume of telephone
sampling. A second problem is that the data
are not fully up to date, fail to reflect num-
bers not in directories, and may be imperfect
because of clerical error. A recent method
that uses the data just for stratification pur-
poses (16) neutralizes the second problem,
but not the first.
The method that is cleanest statistically
and simplest conceptually consists of simple
random sampling from the frame of tele-
phone numbers. This technique is occasion-
ally used (10, 17). Its drawback is its low
yield of residential numbers (less than 25
percent in a nationwide sample). Low yield
increases costs because more interviewer
time is required and also can indirectly cause
lower accuracy by adversely affecting in-
terviewer morale.
Finally, the Mitofsky-Waksberg method
has virtually no bias and gives a reasonable
yield of residential numbers. This method
and its extensions are covered in detail later.
The frame for RDD surveys
Before sampling can start under either the
Mitofsky-Waksberg method or simple ran-
dom sampling, one has to set up the frame
of telephone numbers from which the
sample is to be drawn. Generally, the frame
should consist of the 10,000 phone numbers
in each combination of area code and prefix
whose territory lies partially or wholly
within the region covered by the survey.
This region can be the entire nation, an en-
tire state or a group of states, a county or a
group of counties, or something else. If a
region includes only part of a state, deter-
mining the frame can be complex, but for an
entire state one just uses all the prefixes in
the area code(s) for the state. Where the ter-
ritory for a prefix is partly inside the survey
region and partly outside, a call to a house-
hold number with that prefix requires
screening question(s) to find out whether the
residence is within the survey region.

a
t

C
o
m
m
i
s
s
a
r
i
a
t

A

L
'
E
n
e
r
g
i
e

A
t
o
m
i
q
u
e

o
n

S
e
p
t
e
m
b
e
r

2
9
,

2
0
1
4
h
t
t
p
:
/
/
a
j
e
.
o
x
f
o
r
d
j
o
u
r
n
a
l
s
.
o
r
g
/
D
o
w
n
l
o
a
d
e
d

f
r
o
m

970 Potthoff
With a small enough region, it may be
feasible to obtain the required list of prefixes
from telephone directories or possibly from
the telephone company. If one relies on di-
rectories and nothing else, however, one
may miss prefixes that were added recently,
an omission that could result in bias.
It is possible to forgo the task of identi-
fying all active combinations of area code
and prefix in the survey region by simply
using a list based on all 1,000 possible pre-
fixes, with exclusion of ineligible ones
(mainly those below 200, which are not in
use). However, this approach is usually in-
efficient because it substantially increases
the number of nonworking numbers dialed
and reduces the yield of residential numbers.
Except when the survey region is small,
ordinarily the best means for creating the
frame is through use of data prepared by
Bellcore (Bell Communications Research,
Livingston, New Jersey). Bellcore sells sev-
eral products that could be used to get a
frame, but the best of these for most situa-
tions is one called "NPA/NXX V&H Co-
ordinates" (priced at $350, for either disk or
tape, in early 1993). This product is reissued
each month. It includes seven files, but the
main file of interest to RDD samplers is the
second one, which hereafter will be referred
to simply as the "V&H file." One should use
an up-to-date file to avoid missing new pre-
fixes. Familiarity with Bellcore's V&H file
may not be widespread, although the file or
its AT&T predecessor has been covered in
standard references (3, 13, 18).
The V&H file has an 80-position record
for each combination of area code and prefix
(which are in positions 32-34 and 35-37,
respectively). Although the record has a
number of fields, only certain ones relevant
to RDD sampling are mentioned here. The
file appears to be well maintained, espe-
cially so given its size and complexity. It
serves well the needs of RDD samplers even
though it was not created for their purposes.
However, three steps need to be taken before
the file is used for a frame; some details are
provided here because the documentation
that comes with the file is rather limited for
purposes of RDD sampling.
First, certain data on a record may indicate
that its combination of area code and prefix
is not supposed to have any dialable resi-
dential numbers. Such records should be de-
leted. This can be accomplished by dropping
a record if position 25 (nondialable indica-
tor) is not 0 and/or if positions 38-39 (type
of NXX) are other than 00,50,51,52, or 54.
Second, any record whose territory is
wholly outside of the survey region should
be dropped. Doing this may not be simple if
the region covers only part of a state, but is
easy otherwise. Even for a nationwide sur-
vey, many records have to be removed be-
cause the V&H file covers most of North
America. For a national survey, one deletes
any record for which position 80 (point iden-
tification) is other than 3 (Alaska), 6 (Ha-
waii), or 0 (the rest of the United States).
Alaska and Hawaii can also be dropped if
desired.
The April 1993 Bellcore V&H file con-
tained 60,734 records to start. Application of
the first two steps (with retention of the en-
tire United States, including Alaska and Ha-
waii) resulted in a reduction to 44,723
records.
The most important reason for doing the
third step is to avoid bias when an area code
is split or rearranged (which is now occur-
ring with increasing frequency). When this
happens, generally there is first a period of
"permissive dialing" for several months,
during which the affected prefixes can be
reached through either the old or the new
area code. After this, these prefixes can be
reached only through the new area code. For
some time before, during, and after the pe-
riod of permissive dialing, the V&H file will
temporarily have two records for each af-
fected prefix, one for the old area code and
one for the new. All such pairs of duplicate
records should be consolidated into a single,
record. If this is not done, sampling will be
biased during the period of permissive di-
aling (because phone numbers with the af-
fected prefixes will have a double chance of
selection) and will be inefficient at other
times.
For the April 1993 V&H file, the third
step was applicable to splits of area codes in

a
t

C
o
m
m
i
s
s
a
r
i
a
t

A

L
'
E
n
e
r
g
i
e

A
t
o
m
i
q
u
e

o
n

S
e
p
t
e
m
b
e
r

2
9
,

2
0
1
4
h
t
t
p
:
/
/
a
j
e
.
o
x
f
o
r
d
j
o
u
r
n
a
l
s
.
o
r
g
/
D
o
w
n
l
o
a
d
e
d

f
r
o
m

Telephone Sampling in Epidemiologic Research 971
three states (California, Georgia, and Texas)
and to a transfer of the Bronx, New York,
from area code 212 to 718 in New York.
Pairs of duplicate records for these four, situ-
ations had to be found by matching on V-
coordinate (positions 1-5), H-coordinate
(positions 6-10), and prefix (positions 35-
37). In addition, for New York, a match was
allowed only if place name (positions 60-
69) was "Bronx" on both records; without
this constraint, many nonduplicates (e.g:, in
Manhattan and Brooklyn, New York) would
have been incorrectly consolidated. Across
all four situations, 745 pairs of records were
consolidated, which reduced the 44,723
records to 43,978.
The V- (vertical) and H- (horizontal) co-
ordinates are somewhat like longitude and
latitude. They can be used in a scheme
to order the records geographically and
roughly by city size to allow for systematic
sampling of records in a way that can reduce
sampling variability (18). They also have
another use: If two places not too far apart
have V- and H-coordinates (V, H) and ( V,
H') and one wants the distance in miles be-
tween them, it is approximately the square
root of (1/10) [(V - V' )
2
+ (H - H' )
2
] .
Other useful data on the V&H file include
time zone and a daylight savings indicator.
The Mitofsky-Waksberg method and
some extensions
Briefly, the Mitofsky-Waksberg method
(3) works as follows. There are two stages
of sampling. In the first stage, one draws
telephone numbers from the frame and dials
each one. If a number is residential, then its
associated primary sampling unit (PSU), de-
fined to be the set of 100 numbers with the
same first eight digits as the number dialed,
is retained for trie second stage. If a number
is not residential, its PSU is discarded. In the
second stage, from each retained PSU one
draws and dials numbers until k residential
numbers are found (where is a chosen in-
teger). .
It is proper to do interviews either at the
second-stage numbers only or at the num-
bers in both stages. Factors such as sched-
uling may prevent interviewing at the first-
stage numbers in some cases.
The same set of retained PSUs can be used
for more than one survey in the same survey
region. Bias problems may result, however,
if too much time elapses after the first stage
(19).
The value of k need not be chosen in ad-
vance. Whatever k is, though, there should
be sample-management procedures to pre-
vent one from ending up with more than k
second-stage residential numbers in any
PSU or with fewer than k numbers unless not
enough exist.
The Mitofsky-Waksberg method has cer-
tain drawbacks (13, 20), which may have
more impact in some, situations than in oth-
ers. First, determining whether a number is
residential may be impossible or at least may
require many callbacks, such as when the
number repeatedly rings without answer.
Second, the sequential nature of the second
stage (stemming from uncertainty about
how many numbers will need to be dialed to
find k residential numbers in a PSU) causes
delays and administrative complexity.
Third, PSUs with fewer than (Jk + 1) resi-
dential numbers will be "exhausted" if cho-
sen and will bring about an exception to the
property that the method gives all residential
numbers the same chance of selection.
Three options are available that extend or
generalize the Mitofsky-Waksberg method
to eliminate or alleviate these drawbacks.
Although these options have their own
drawbacks, often the trade-off is favorable.
Any one, any two, or all three of the options
may be used.
First, instead of requiring from each PSU
in the second stage a fixed number of resi-
dential numbers, one can sample exactly k
numbers regardless of how many are found
to be residential (21, p. 91; 22). The prob-
lems of sequential sampling are thus
avoided. However, there are now differing
probabilities of selection, so weights must
be applied to compensate. If the k second-
stage numbers in the gth PSU yield x resi-
dential numbers, the relative weight for an
interview in that PSU can be taken as k/x
g
if
interviewing is done in the second stage

a
t

C
o
m
m
i
s
s
a
r
i
a
t

A

L
'
E
n
e
r
g
i
e

A
t
o
m
i
q
u
e

o
n

S
e
p
t
e
m
b
e
r

2
9
,

2
0
1
4
h
t
t
p
:
/
/
a
j
e
.
o
x
f
o
r
d
j
o
u
r
n
a
l
s
.
o
r
g
/
D
o
w
n
l
o
a
d
e
d

f
r
o
m

972 Potthoff
only, or as (k + l)/(x
g
+ 1)for an inter-
view in either stageif interviewing is done
in the first stage as well.
Second, suppose that, in the foregoing
(except for part of the paragraph describing
drawbacks), "residential" is replaced by
"auspicious" wherever the former appears,
with the understanding that the choice of
how to define "auspicious" is made by the
user, but under the restriction that auspi-
cious numbers must at least include all resi-
dential numbers (20). Then all that was
stated before remains valid under the gen-
eralization from residential to auspicious
numbers. An auspicious number might be
defined simply as one that rings when not
busy, except where the ringing precedes a
message that indicates the number is not in
service. Other definitions are possible, such
as one that embraces those numbers just de-
scribed plus nonworking numbers with cer-
tain messages suggesting that the number
was recently in service (15,20). With a suit-
able definition, one largely avoids the kinds
of delays, costs, and error possibilities as-
sociated with having to find out if numbers
are residential. There is a drawback, how-
ever: Some PSUs with auspicious numbers
but no residential numbers will be retained.
Third, the Mitofsky-Waksberg method
can be further generalized to a scheme
where c (^1) numbers from each PSU are
drawn and dialed in the first stage (20). As
c increases, first-stage costs also increase,
but the sequentiality problems are dimin-
ished. Briefly, the scheme works as follows.
APSU is retained and called a Type II PSU
if two or more of its c first-stage numbers are
auspicious, is retained and called a Type I
PSU if exactly one is auspicious, or is dis-
carded if none are auspicious. For a Type II
PSU, the second-stage sample is totally non-
sequential and consists of exactly kc tele-
phone numbers (where, as before, k is cho-
sen by the user). For a Type I PSU, the
second-stage sample has both a nonsequen-
tial segment, consisting of exactly k(c - 1)
numbers, and a sequential segment, consist-
ing of as many numbers as are needed to find
k auspicious numbers.
The third option provides equal probabil-
ity of selection of all residential telephone
numbers (except for complications from ex-
hausted PSUs) if it is used without the first
option, but not if the first option is also ap-
plied. To combine the first and third options,
one makes the second-stage sample in a
Type I PSU completely nonsequential by
giving it exactly kc numbers. Weights are
needed, however. If the kc second-stage
numbers in the gth Type I PSU yield x
g
aus-
picious numbers, the relative weight for an
interview in that PSU can be taken as 1 +
{kc - x
g
)/(cx
g
) if interviewing is done in
the second stage only, or as 1 + (kc - x
g
)l
[c(x
g
+ 1)]for an interview in either
stageif interviewing is done in the first
stage as well. The first formula was derived
earlier (15, p. 618, third complete para-
graph); the second is derived similarly. For
any interview from a Type II PSU, the rela-
tive weight is 1. A benefit of using both the
first and third options rather than the first
option alone is that the disparity among
weights diminishes as c rises.
If c = 1 and auspicious numbers are de-
fined as residential, this is the Mitofsky-
Waksberg method, and the two weight for-
mulas just given agree with the two given
previously. Note also that for x
g
0 the
weight formulas reflect an imperfection if
no first-stage interviewing occurs but not if
it does occur.
Finally, research reports that indicate use
of the Mitofsky-Waksberg method or some
variant of it will better inform the reader if
these reports include such details as the
number of retained PSUs, the value of k,
whether first-stage interviewing was done,
and any options or modifications that were
used.
Sampling issues encountered during
interviewing
During the interviewing, certain steps are
needed to avoid bias related to sampling:
1. Sometimes a sample number has been
changed to a new number, and one may
learn the latter upon dialing the former. Gen-
erally, to avoid double chance of selection

a
t

C
o
m
m
i
s
s
a
r
i
a
t

A

L
'
E
n
e
r
g
i
e

A
t
o
m
i
q
u
e

o
n

S
e
p
t
e
m
b
e
r

2
9
,

2
0
1
4
h
t
t
p
:
/
/
a
j
e
.
o
x
f
o
r
d
j
o
u
r
n
a
l
s
.
o
r
g
/
D
o
w
n
l
o
a
d
e
d

f
r
o
m

Telephone Sampling in Epidemiologic Research 973
(once through the old number and once
through the new), one should then terminate,
and thus not use the new number.
2. One should check that the number
reached is the number that was to be dialed.
This check not only uncovers misdialing; it
also detects call forwarding and "fall-
throughs," either of which, if undetected,
would distort the chance of selection. Fall-
throughs, largely confined to rural areas, oc-
cur when the dialing of a nonworking num-
ber results in reaching another number (18,
pp. 47-48).
3. The use of answering machines (23-25)
has grown rapidly in recent years. Espe-
cially under the Mitofsky-Waksberg
method, if an answering machine is reached,
one should infer nothing definite about
whether the number is residential because of
the possibility of call forwarding or a fall-
through.
4. One should find out if the number
reached is for a home, a business, or both,
and continue only if it is for a home or both.
Interviewing at a business number answered
by "Hello" would give extra chance of se-
lection.
5. For any sample number that might not
be inside the survey region, one should find
out whether it is and then terminate if it is
outside.
6. Some surveys require no selection
within a household, such as in situations in
which all eligible household members are
chosen or those in which household-related
rather than individual-related information is
sought. In surveys that pick one member,
however, each eligible member should have
the same chance of selection. Interviewing
whoever answers the phone is not proper,
nor is a rule that chooses a male if one is
home or a female otherwise. Accepted se-
lection techniques include some older ones
(26), as well as a newer one based on birth-
days (27, 28) that consists of choosing the
eligible member with the next birthday or
the most recent birthday. Among suitable
techniques, a birthday technique was found
to be most widely used in a recent survey of
political surveys (14).
7. Let E (>1) denote, for a household, the
number of eligible members from whom one
member is picked (in surveys in which this
is done). Let F (^1) denote the number of
residential telephone numbers at which the
household can be reached. Chance of selec-
tion decreases with higher E and increases
with higher F. E and F have to be obtained
in each interview, for later use in weighting,
if estimation is to be free of bias.
Weighting
Estimates from surveys have several po-
tential sources of bias that one can try to
offset through weighting. Several weight-
ings can be applied together, in the same
analysis (29).
Let R denote the relative weight whose
different formulas were given above in
discussion of the first option under the
Mitofsky-Waksberg method, or a weight
(29) that applies instead if the first option is
not used and an exhausted PSU is encoun-
tered. Define W = RE/F. Relatively weight-
ing each interview by its associated value of
W will correct for bias corresponding to R,
E, and F. Many reports of epidemiologic
surveys ignore some or all of these three,
leading one to wonder whether any correc-
tion was made. Mitofsky (14) considered
weighting based on both E and F to be "fun-
damental," but he found that few election
polls used E and even fewer used F.
Calling mainly during productive evening
and weekend hours (30) and making suffi-
cient callbacks will hold down the number
of sample members not reached. Nonavail-
ability bias that remains can be corrected for
by giving more weight to interviews that re-
quired more callbacks (31) or by poststrati-
fication. One uses poststratification as well
to try to remove refusal bias and bias re-
sulting from households without telephones.
Poststratification weights are typically as-
signed by cells, with a cell weight based on
population size for the cell divided by num-
ber sampled (adjusted for any earlier
weighting) from the cell. With too many
cells, these weights can be unstable, but an

a
t

C
o
m
m
i
s
s
a
r
i
a
t

A

L
'
E
n
e
r
g
i
e

A
t
o
m
i
q
u
e

o
n

S
e
p
t
e
m
b
e
r

2
9
,

2
0
1
4
h
t
t
p
:
/
/
a
j
e
.
o
x
f
o
r
d
j
o
u
r
n
a
l
s
.
o
r
g
/
D
o
w
n
l
o
a
d
e
d

f
r
o
m

974 Potthoff
alternative poststratification scheme that ac-
commodates more cells (32) is available.
CASE-CONTROL STUDIES
Much of the preceding discussion of tele-
phone sampling for general surveys also ap-
plies fully to case-control studies. What fol-
lows will cover the exceptions and other
points that involve something special or
unique to case-control studies. Schemes that
sample controls so as to avoid common
problems will be described. Population-
based studies are covered before center-
based studies because the former are less
complex.
Population-based studies
In a population-based case-control study,
the cases are all of the eligible ones that oc-
cur in a given geographic region in a given
period. RDD controls are sampled from the
same region, typically by strata (categories)
based on age, sex, race, and/or geography.
Everyone within a stratum is supposed to
have the same chance of selection, but
sample sizes for the strata are chosen so that
the controls are frequency matched (4) to the
cases. (Individual matching instead of fre-
quency matching could be done but seems to
be uncommon for population-based studies
with RDD controls.) The many population-
based studies that have used RDD controls
include two large ones, one on bladder can-
cer (33) and the other on breast, ovarian, and
endometrial cancers (34).
A partial RDD approach is possible. For
example, investigators with access to Health
Care Financing Administration files (35)
might use those files to sample for controls
aged 65 years and over, but use RDD to get
younger controls.
Standard odds ratio analysis is not strictly
applicable in case-control studies that use a
cluster sampling plan such as the Mitofsky-
Waksberg method. Modified analysis that
handles the clustering ensures applicability,
however (36).
Selection and enlistment of RDD controls
to frequency match them to cases, plus any
telephone interviewing, can all be done to-
gether using either of two one-step proce-
dures, but neither can work well unless the
strata are few (37). Instead, usually one
takes a household census on an initial call,
draws the sample for each stratum from the
assemblage of censuses, and then calls again
to enlist anyone selected and to either ar-
range for a visit or do an interview by tele-
phone.
In the sampling, one faces the same
sources of potential bias that were covered
before. The usual method of dealing with
them is not the same as before, however,
because avoidance of weighting is generally
desired.
Thus, the only ways used to reduce non-
availability bias and refusal bias may be to
make many callbacks and to try to minimize
refusals. Bias stemming from households
without telephones can be dealt with by ex-
cluding cases without telephones (38), but
one wonders how often this is done because,
outside of some reports of studies in Wash-
ington State, one rarely sees it mentioned.
The Mitofsky-Waksberg method in its se-
quential form avoids the weighting by R, ex-
cept where a PSU has fewer than (k + 1)
residential numbers. Households reachable
at more than one telephone number can be
handled by discarding them with probability
(F - 1)/F (2,38-40) instead of by weighting
them by 1/F.
From the eligible persons in the censuses,
one can draw the sample for each stratum
separately. However, the same household
can yield more than one control, a condition
often deemed undesirable. From such a
household, one can randomly pick one per-
son to retain, but bias against higher E then
results. If controls include both males and
females, this bias is of especial concern but
can be largely removed by randomly pre-
designating each household as "male" or
"female," and excluding anyone not of the
designated sex (2).
Averting the various potential biases is
important. This is illustrated by a question
that was raised (41) about why a certain
population-based study with RDD controls

a
t

C
o
m
m
i
s
s
a
r
i
a
t

A

L
'
E
n
e
r
g
i
e

A
t
o
m
i
q
u
e

o
n

S
e
p
t
e
m
b
e
r

2
9
,

2
0
1
4
h
t
t
p
:
/
/
a
j
e
.
o
x
f
o
r
d
j
o
u
r
n
a
l
s
.
o
r
g
/
D
o
w
n
l
o
a
d
e
d

f
r
o
m

Telephone Sampling in Epidemiologic Research 975
(34), unlike some other studies, showed no
positive relation between breast cancer and
alcohol. Because alcohol use is associated
positively with socioeconomic status, any
biases that brought about overly upscale
controls would give rise to some doubts.
Improvement over current schemes for
selecting RDD controls from the assem-
blage of censuses is possible. The following
scheme has not previously been proposed,
as far as I know, but it pieces together certain
standard concepts. Its aim is to simulta-
neously avoid any bias, avoid picking more
than one member from a household, and
avoid need for a weighted analysis.
The scheme has two stages of selection
from the assemblage of household censuses.
The first stage consists simply of choosing
one member from each household, in such
a way that each eligible member has the
same chance (1/E) of selection.
The second stage operates separately for
each stratum. Although it essentially equal-
izes probabilities of selection for everyone
within a stratum, mechanically it is just like
systematic sampling with probability pro-
portional to size (42, table 1). To begin, sort
in random order the first-stage selectees in
each stratum. Let the value W be as before,
except that it could be modified as desired
to incorporate poststratification weights or
other weights in addition to R, E, and F. For
a stratum, let W
+
denote the sum of the Ws
across all first-stage selectees and let W
3
*
denote the maximum W. The number of sec-
ond-stage selectees for a stratum, to be de-
noted by n, will be the number of controls
for the stratum and generally will be set
equal, or otherwise proportional, to the num-
ber of cases for the stratum. Some bias will
exist unless (for each stratum) n < WVW"".
Define / = W
+
/n, which is the sampling in-
terval for the stratum, to be used as in sys-
tematic sampling. Obtain a random point u
between 0 and 1, drawn from a uniform (rec-
tangular) distribution. Form the cumulative
sums of the Ws (42, table 1). Finally, for
each integer h(h = 1,2,... ,ri) choose the
first first-stage selectee for whom the cu-
mulative sum of the Ws is ^ I(u + h - 1).
The resulting n persons are the n controls for
the stratum.
Center-based studies
In a center-based case-control study, the
cases are all of the eligible ones that arise in
specified hospitals or other centers during a
given period. Selection of controls generally
poses more problems for the center-based
study than for the population-based study. A
center-based study, however, is often the
only feasible choice.
Telephone sampling methods that have
been used to select controls for center-based
studies have their imperfections (which ap-
pear to be partially correctable), but may
nevertheless be better than any alternative
methods. The telephone methods normally
are based on the telephone numbers of the
cases.
Under one approach, telephone numbers
that agree with the number of the case except
in the last two digits are chosen randomly
and are dialed until enough control(s) who
match the case in designated ways are found
(38). If no more than one eligible person per
household is chosen, this approach is biased
against persons in households with higher E,
as Greenberg (38) pointed out. Although he
was concerned mainly with studies of chil-
dren, in an adult study the bias also exists,
albeit to a lesser degree if the adult study
covers only one sex. Greenberg argued that
(for child studies) the bias of the approach
results in underrepresenting children of
lower socioeconomic status because of
closer spacing of the births of such children.
The approach can avoid a bias favoring
households with more than one telephone
number by discarding them with probability
(F - 1)/F, but one rarely reads of this being
done. Again, the bias produces too many up-
scale controls.
When the approach is applied, some cases
may be difficult to match, so high costs or
unmatched cases may result. Overmatching
may also be a concern (38).
Under a second approach to obtain con-
trols, one dials numbers with the same area
code and prefix as the case's number. The

a
t

C
o
m
m
i
s
s
a
r
i
a
t

A

L
'
E
n
e
r
g
i
e

A
t
o
m
i
q
u
e

o
n

S
e
p
t
e
m
b
e
r

2
9
,

2
0
1
4
h
t
t
p
:
/
/
a
j
e
.
o
x
f
o
r
d
j
o
u
r
n
a
l
s
.
o
r
g
/
D
o
w
n
l
o
a
d
e
d

f
r
o
m

976 Potthoff
last four digits are varied. Reports of the use
of the approach are mainly vague about how
these four digits are selected, so there may
be several variants; no assessment is at-
tempted here.
Early in the development of telephone
methods for finding controls, a third ap-
proach, based on directories, was used (43).
One can obtain control(s) matched to each
case by drawing from the telephone direc-
tory that covers the locality where the case
lives. However, an extra pitfall arises: Be-
cause no controls can come from numbers
not in directories, the approach is open to
potential bias as long as one retains cases
whose numbers are not in directories.
Some ideas for better telephone sampling
methods for selection of controls in center-
based studies will now be sketched. The best
route to improvement appears to lie in the
direction of methods more like those used in
population-based studies. Thus, instead of
matching one case at a time, one could, at the
least, group all cases whose prefixes serve
the same territory. Dialing to seek controls
would then be by territory rather than by
individual case. Household censuses could
be taken, and something like frequency
matching could be done. It appears that the
prefix grouping by territory, or coarser
grouping as desired, could be obtained
through the use of Bellcore products.
The scheme described earlier for popula-
tion-based studies might be used for center-
based studies as wellif the word "stratum"
in the earlier description is changed to
"cell," where the cells correspond to all the
possible combinations of stratum and terri-
tory. Unless the number of cells is small,
though, the scheme may often fail because
not enough controls may be obtainable from
some of the cells. In the most extreme situ-
ation, a cell could have case(s) but not have
any first-stage selectees who could serve as
controls.
If necessary, however, one generally can
modify the earlier scheme to make it work
for center-based studies. The modification
involves only the way that the n's are cho-
sen; otherwise the scheme is the same as
before, and so it retains the properties of
avoiding biases, avoiding more than one
control from a household, and avoiding
weights. The n' s no longer have the same
constraints by cell as before, but their mar-
ginal totals, both by stratum and by territory,
are still constrained (except for rounding) to
bear a fixed ratio to the marginal totals of
cases. To determine the n's, one uses itera-
tive proportional fitting and then controlled
rounding, as follows.
Let Njj be the number of cases in the cell
for stratum i and territory;. Define the mar-
ginal case proportions P
(
. = N
L
IN and
P,j - Nj/N , where a dot indicates summa-
tion over the index that it replaces.
Let n' be the total number of controls to
be drawn, and let n
tj
(yet to be determined)
be the number of controls for cell (i,j). Let
W? be the sum of the Ws across the first-
stage selectees in cell (/, ;).
Using iterative proportional fitting (44,
pp. 115-118,45, p. 97 ff.), one can generally
find values n
lj0
= a^a.jW^ satisfying the
constraints n
L0
= P,.n', n
j0
= P.j/i', where
the dots indicate summation as before.
Were n
ij0
an integer, it would be the num-
ber of controls to draw for cell (/ , ; ). The
constraints are the desired ones that re-
quire the marginal control proportions
(for both margins) to be equal to the cor-
responding marginal case proportions.
The territory factors a.j should not vary
greatly from each other if RDD sample
sizes are chosen so that numbers of house-
hold censuses by territory will turn out
roughly proportional to theP,
;
' s. The stra-
tum factors a
jt
could vary sharply, how-
ever. Thus, if the strata involve age
groups, a,, could increase with rising age.
Finally, using unbiased zero-restricted
controlled rounding (46), one can obtain the
n
tj
's (integers) by rounding the n
lj0
's (usually
not integers). With this rounding technique,
each n,y, n
L
, or w
;
differs from the corre-
sponding n
ij0
, n
L0
, or n
j0
by less than 1, and
the rounding entails no bias (46). One draws
n
tj
controls from cell (i, j).
Let W?
3
* denote the maximum W
in cell (i, j). Let {n
ij0
} denote the smallest
{} p f
j
integer > n
ij0
. Unless {n
ij0
} <
for

a
t

C
o
m
m
i
s
s
a
r
i
a
t

A

L
'
E
n
e
r
g
i
e

A
t
o
m
i
q
u
e

o
n

S
e
p
t
e
m
b
e
r

2
9
,

2
0
1
4
h
t
t
p
:
/
/
a
j
e
.
o
x
f
o
r
d
j
o
u
r
n
a
l
s
.
o
r
g
/
D
o
w
n
l
o
a
d
e
d

f
r
o
m

Telephone Sampling in Epidemiologic Research 977
each cell, some bias will exist (just as with
population-based studies). The problem can
be avoided by using a smaller n' or by hav-
ing more household censuses available.
CONCLUDING REMARKS
The recent rapid growth in the use of RDD
sampling in epidemiologic and other health-
related research reflects the benefits that this
type of sampling can provide. As with any
new methodology, increasing experience
can lead to better practices. One can expect
that the current wide use of RDD sampling
will continue. Various changes are likely,
however. New techniques may be devel-
oped. Concern has been expressed over de-
clining response rates for surveys in general
(47), although refusals may not be as hard to
prevent in health-related research as in some
other types of research. New types of infor-
mation related to telephone sampling frames
could become available. Telephone technol-
ogy changes will continue, with unpredict-
able effects on RDD sampling.
For current conditions, this review has
suggested suitable and efficient telephone
sampling strategies that are aimed at avoid-
ing bias. The user of RDD sampling should
be alert, however, to future changes and
their implications.
ACKNOWLEDGMENTS
Supported by National Institute on Aging
grants 5R37AG07198 and 5R37AG07025.
The author thanks Fred Ederer for his encour-
agement and help.
REFERENCES
1. Marcus AC, Crane LA. Telephone surveys in
public health research. Med Care 1986;24:
97-112.
2. Hartge P, Brinton LA, Rosenthal JF, et al. Ran-
dom digit dialing in selecting a population-based
control group. Am J Epidemiol 1984; 120:
825-33.
3. Waksberg J. Sampling methods for random digit
dialing. J Am Stat Assoc 1978;73:40-6.
4. Kelsey JL, Thompson WD, Evans AS. Methods
in observational epidemiology. New York, NY:
Oxford University Press, 1986.
5. Ibrahim MA, ed. The case-control study: consen-
sus and controversy. J Chronic Dis 1979;32:
1-144, i-viii.
6. Schlesselman JJ. Case-control studies: design,
conduct, analysis. New York, NY: Oxford Uni-
versity Press, 1982.
7. Wacholder S, McLaughlin JK, Silverman DT, et
al. Selection of controls in case-control studies. I.
Principles. Am J Epidemiol 1992;135:1019-28.
8. Wacholder S, Silverman DT, McLaughlin JK, et
al. Selection of controls in case-control studies.
II. Types of controls. Am J Epidemiol 1992;135:
1029-41.
9. Wacholder S, Silverman DT, McLaughlin JK, et
al. Selection of controls in case-control studies.
III. Design options. Am J Epidemiol 1992; 135:
1042-50.
10. Marks JS, Hogelin GC, Gentry EM, et al. The
behavioral risk factor surveys. I. State-specific
prevalence estimates of behavioral risk factors.
Am J Prev Med 1985;1 (6):l-8.
11. Gentry EM, Kalsbeek WD, Hogelin GC, et al.
The behavioral risk factor surveys. II. Design,
methods, and estimates from combined state
data. Am J Prev Med 1985;1 (6):9-14.
12. Remington PL, Smith MY, Williamson DF, et al.
Design, characteristics, and usefulness of state-
based behavioral risk factor surveillance:
1981-87. Public Health Rep 1988;103:366-75.
13. Lepkowski JM. Telephone sampling methods in
the United States. In: Groves RM, Biemer PP,
Lyberg LE, et al., eds. Telephone survey meth-
odology. New York, NY: John Wiley & Sons,
1988:73-98.
14. Mitofsky WJ. The state of state election polls.
Chance 1993;6 (1):9-16.
15. Potthoff RF. Generalizations of the Mitofsky-
Waksberg technique for random digit dialing:
some added topics. In: American Statistical As-
sociation. 1987 proceedings of the section on
survey research methods. Alexandria, VA:
American Statistical Association, 1987:615-20.
16. Casady RJ, Lepkowski JM. Stratified telephone
survey designs. Surv Methodol 1993;19:103-13.
17. Klecka WR, Tuchfarber AJ. Random digit
dialing: a comparison to personal surveys. Public
Opinion Q 1978;42:105-14.
18. Groves RM, Kahn RL. Surveys by telephone: a
national comparison with personal interviews.
New York, NY: Academic Press, Inc. 1979.
19. Voigt LF, Davis S, Heuser L. Random digit
dialing: the potential effect on sample character-
istics of the conversion of nonresidential tele-
phone numbers. Am J Epidemiol 1992;136:
1393-9.
20. Potthoff RF. Some generalizations of the
Mitofsky-Waksberg technique for random digit
dialing. J Am Stat Assoc 1987;82:409-18.
21. Waksberg J. Discussion. In: US Bureau of the
Census. First annual research conference pro-
ceedings. Washington, DC: US Department of
Commerce, Bureau of the Census, 1985:87-92.
(Document no. C3.2:R31/9/985).
22. Brick JM, Waksberg J. Avoiding sequential sam-
pling with random digit dialing. Surv Methodol
1991;17:27-41.

a
t

C
o
m
m
i
s
s
a
r
i
a
t

A

L
'
E
n
e
r
g
i
e

A
t
o
m
i
q
u
e

o
n

S
e
p
t
e
m
b
e
r

2
9
,

2
0
1
4
h
t
t
p
:
/
/
a
j
e
.
o
x
f
o
r
d
j
o
u
r
n
a
l
s
.
o
r
g
/
D
o
w
n
l
o
a
d
e
d

f
r
o
m

978 Potthoff
23. Tuckel PS, Feinberg BM. The answering ma-
chine poses many questions for telephone survey
researchers. Public Opinion Q 1991;55:200-17.
24. Piazza T. Meeting the challenge of answering
machines. Public Opinion Q 1993;57:219-31.
25. Xu M, Bates BJ, Schweitzer JC. The impact of
messages on survey participation in answering
machine households. Public Opinion Q 1993;57:
232-7.
26. Czaja R, Blair J, Sebestik JP. Respondent selec-
tion in a telephone survey: a comparison of three
techniques. J Marketing Res 1982; 19:381-5.
27. Salmon CT, Nichols JS. The next-birthday
method of respondent selection. Public Opinion
Q 1983;47:270-6.
28. O'Rourke D, Blair J. Improving random respon-
dent selection in telephone surveys. J Marketing
Res 1983;20:428-32.
29. Massey JT, Botman SL. Weighting adjustments
for random digit dialed surveys. In: Groves RM,
Biemer PP, Lyberg LE, et al , eds. Telephone
survey methodology. New York, NY: John
Wiley & Sons, 1988:143-60.
30. Weeks MF, Jones BL, Folsom RE Jr, et al. Op-
timal times to contact sample households. Public
Opinion Q 1980;44:101-14.
31. Potthoff RF, Manton KG, Woodbury MA. Cor-
recting for nonavailability bias in surveys by
weighting based on number of callbacks. J Am
Stat Assoc 1993;88:1197-207.
32. Potthoff RF, Woodbury MA, Manton KG.
"Equivalent sample size" and "equivalent de-
grees of freedom" refinements for inference
using survey weights under superpopulation
models. J Am Stat Assoc 1992;87:383-96.
33. Hartge P, Cahill JI, West D, et al. Design and
methods in a multi-center case-control interview
study. Am J Public Health 1984;74:52-6.
34. Wingo PA, Ory HW, Layde PM, et al. The
evaluation of the data collection process for a
multicenter, population-based, case-control de-
sign. Am J Epidemiol 1988;128:206-17.
35. Psaty BM, Cheadle A, Curry S, et al. Sampling
elderly in the community: a comparison of com-
mercial telemarketing lists and random digit di-
aling techniques for assessing health behaviors
and health status. Am J Epidemiol 1991;134:
96-106.
36. Graubard BI, Fears TR, Gail MH. Effects of
cluster sampling on epidemiologic analysis in
population-based case-control studies. Biomet-
rics 1989;45:1053-71.
37. Harlow BL, Davis S. Two one-step methods for
household screening and interviewing using ran-
dom digit dialing. Am J Epidemiol 1988;127:
857-63.
38. Greenberg ER. Random digit dialing for control
selection: a review and a caution on its use in
studies of childhood cancer. Am J Epidemiol
1990;131:l-5.
39. Strader CH, Weiss NS, Daling JR. Vasectomy
and the incidence of testicular cancer. Am J Epi-
demiol 1988; 128:56-63.
40. Olson SH, Kelsey JL, Pearson TA, et al. Evalu-
ation of random digit dialing as a method of
control selection in case-control studies. Am J
Epidemiol 1992;135:210-22.
41. Longnecker MP. Re: "The evaluation of the data
collection process for a multicenter, population-
based case-control design." (Letter). Am J Epi-
demiol 1989;129:1311.
42. Hartley HO, Rao JNK. Sampling with unequal
probabilities and without replacement. Ann Math
Stat 1962;33:350--74.
43. Gaziano JM, Buring JE, Breslow JL, et al. Mod-
erate alcohol intake, increased levels of high-
density lipoprotein and its subfractions, and de-
creased risk of myocardial infarction. N Engl J
Med 1993;329:1829-34.
44. Deming WE. Statistical adjustment of data. New
York, NY: John Wiley & Sons, 1943.
45. Bishop YMM, Fienberg SE, Holland PW. Dis-
crete multivariate analysis. Cambridge, MA:
MIT Press, 1975.
46. Cox LH. A constructive procedure for unbiased
controlled rounding. J Am Stat Assoc 1987;82:
520-4.
47. Groves RM, Lyberg LE. An overview of nonre-
sponse issues in telephone surveys. In: Groves
RM, Biemer PP, Lyberg LE, et al., eds. Tele-
phone survey methodology. New York, NY:
John Wiley & Sons, 1988:191-211.

a
t

C
o
m
m
i
s
s
a
r
i
a
t

A

L
'
E
n
e
r
g
i
e

A
t
o
m
i
q
u
e

o
n

S
e
p
t
e
m
b
e
r

2
9
,

2
0
1
4
h
t
t
p
:
/
/
a
j
e
.
o
x
f
o
r
d
j
o
u
r
n
a
l
s
.
o
r
g
/
D
o
w
n
l
o
a
d
e
d

f
r
o
m

Você também pode gostar