Sie sind auf Seite 1von 7

Sampling Methods for Random Digit Dialing

JOSEPH WAKSBERG*

A method of sample selection foi household telephone interviewing


via random digit dialing is developed which significantly reduces the
cost of such surveys as compared to dialing numbers completely a t
random. The sampling is carried out through a two-stage design and
has the unusual feature that although all units have the same probability of selection, it is not necessary to know the probabilities of
selection of the first-stage or the second-stage units. Simple random
sampling of possible telephone numbers, within existing telephone
exchanges, is inefficient because only about 20 percent of these
numbers are actually telephone numbers assigned to households. The
method of selection proposed reduces the proportion of unused
numbers sharply.

KEY WORDS : Random digit dialing; Telephone surveys; Sample


selection for telephone surveys ; Clustering in random digit dialing.

1. INTRODUCTION

1975). Most of these numbers are unused; others are assigned to businesses, institutions, government bodies, etc.
The effort in designating these numbers, dialing them,
and then ascertaining whether or not they are working
numbers located in housing units adds considerably .to
the cost of a survey. Chilton Research Services has stated
that a t considerable expense, it has eliminated large
blocks of nonworking numbers from its sampling frame
and has been able to increase the proportion of household
numbers to 40 percent in the frame (Chilton 1967). The
reduced list is not available to the general public, and
most researchers cannot afford to duplicate the Chilton
effort.
Sudman (1973) has described a method of sampling
using banks of numbers as clusters, which reduces the
proportion of nonhousehold numbers to be dialed.
Sudmans procedure involves selecting numbers from
telephone directories, having each number identify a
bank of numbers, and dialing random numbers within
each bank until a predetermined number of listed telephone numbers has been reached. Although this system
reduces the workload dra.matically, it contains several
awkward operational features. First, for each dialed
household it involves ascertaining whether or not the
number has been listed. Where directories arranged in
numerical order do not exist, it is necessary to obtain the
information from the households interviewed raising
problems in cases of refusals or no-answers, as well as
some uncertainty regarding accuracy of response. Second,
for a national survey, it requires working with a large set
of telephone directories and unduplicating overlapping
sections of the directories. Finally, the procedure produces
inequality in the number of households per cluster, which
usually adds to the variance of the results as well as
creates uncertainty about the number of clusters needed
to achieve a specific sample size.
A different method of selecting the sample clusters
(using Sudmans definition of clusters) avoids the problems cited in the preceding paragraph and also achieves
sizeable reductions in t,he workload, although probably
not as much as Sudmans method does. This new method
was initially developed by Warren Mitofsky of CBS

Recently there has been an increasing interest in telephone surveys in lieu of personal interview surveys. The
main motivation, of course, is to avoid the high cost of
personal interviews. There appear to be other advantages
to telephone surveys: They can be completed more
quickly; the way the questions are actually asked and the
practices of the interviewers can be observed and controlled more effectively ; and interviewers willing to
work in inner-city areas can be recruited without difficulty. As a result of this rising interest, a number of
studies have been instituted on the quality of responses in
telephone surveys. The evidence seems to indicate that
generally there is little difference in accuracy of reporting
between telephone and personal interviews (Rogers 1976).
There are essentially two kinds of sampling frames used
for telephone surveys. One is the list of names and numbers in telephone directories. The other is the set of all
possible four-digit numbers within existing telephone exchanges. The use of the latter is referred to as random
digit dialing. Random digit dialing is generally preferred
because telephone directories do not contain unlisted
numbers and t,hose issued since the publication dates of
the most recent directories, and thus are subject to
potentially serious biases (Fletcher and Thompson 1974 ;
Glasser and Metzger 1975; Roslow and Roslow 1972).
Several recent papers describe techniques for sampling
and interviewing with random digit dialing (Glasser and
Metzger 1972; Hauck and Cox 1976; Sudman 1973).
Unless there is a considerable amount of advance work,
the use of unrestricted random scmpling for random
digit dialing turns out to be quite inefficient because a
Another procedure that is sometimes used involves selecting
very high percentage of the numbers in the sampling numbers from directories and replacing the last digit (or two) by the
addition of a fixed number. This does not result in known probabilities
frame-approximately 80 percent-are not assigned to of
selection.
households (Chilton 1967; Glasser and Metzger 1972,
Q

* Joseph Wahberg is Vice President, Westat, Inc., RockviUe, MD


20852. Research was supported by CBS News and Westat, Inc.

Journal of the American Staqstical Assoclation


March 1978, Volume 73, Number 361
Applications Section

41

Waksberg: Random Digit Dialing Sampling

News who used it in a number of telephone surveys prior


to the authors work.2 Mitofsky proposed this technique
to the author for further analysis.
Sudman has defined his clusters as blocks of 1,000
numbers, i.e., sets of telephone numbers with identical
first seven digits (including area code and prefix number)
but has pointed out that the methodology is equally applicable t o clusters of 100 or 10. Clusters can be similarly
defined for the alternative procedure described in this
article. However, the greatest practical advantage comes
with the use of fairly small clusters, and we shall define
the clusters as the sets of 100 numbers with the identical
first eight digits.
The proposed sample-selection method for random
digit dialing is as follows: Obtain from ,4T&T a recent
list of all telephone area codes and existing prefix numbers
within the areas. T o these add all possible choices for the
next two digits, and thus prepare a list of all possible
first eight digits of the ten digits in telephone numbers.
These eight-digit numbers are treated as Primary Sampling Units (PSU). A random selection is then made of an
eight-digit number, and also of the next two digits. The
number is then dialed. If the dialed number is a t a residential address, the PSU is retained in the sample. Additional last two digits are selected at random and dialed
within the same eight-digit group, until a set number, k,
of residential telephones is reached. Interviews are attempted both a t the initia.1 number and the additional k
numbers. If the original number called was not residential,
the PSU is rejected. This process is repeated until a predesignated number of PSUs, m, is chosen. The total
sample size is, therefore, m(k
1). The values of m and
k are chosen t o satisfy criteria for an optimum sample
design.
An improvement (over use of unrestricted random
sampling of digits) is possible because a fairly high proportion of blocks of numbers contains no residential
numbers. The savings depend on the proportion of such
blocks and the number of times a survey will be performed. The latter is an important factor because the
improved procedure essentially amounts to first identifying and selecting a sample of blocks of numbers and
dialing random numbers within the blocks. The first step
has t o be done only once (with updating every few years
or when the telephone numbers within the blocks are
exhausted, if surveys are performed frequently), and thus
its cost can be amortized over subsequent surveys.
The theory developed in this article does not require
the PSUs t o be sets of 100 numbers. They could conceivably be defined as sets of 1,000 numbers, or of other
sequences. However, the efficiency of the scheme depends
on the proportion of PSUs that have no residential numbers. Defining a PSU as a set of 100 numbers appears to
be particularly useful for this purpose.
The sample design has several interesting properties.

2 Mitofsky, Warren (1970),Sampling of Telephone Households,


unpublished CBS memorandum.

First, it does not require knowledge of either the first- or


second-stage selection probabilities, but it does produce
an equa.1 probability sample of telephone numbers.
Second, the PSUs are selected with probability proportionate to size (pps), the measures of size being the actual
numbers of residential numbers within the PSUs rather
than estimates based on past data, which is the more
common situation with pps selection. Third, because the
mea.sures of size are the exact numbers, the cluster sizes
within the PSUs are exactly equal, instead of being only
approximately so. This, of course, is a highly desirable
feature of a two-stage sample. The method of selecting
PSUs is somewhat similar to the Lahiri procedure for
pps selection (Lahiri 1951) but does not require advance
estimates of the measure of size.

2. NOTATION
Let the number of telephone prefix areas (PSUs) be M .
For this analysis, we assume that the prefix area includes
the three-digit area code, the three-digit prefix code, and
the first two digits of the suffix. (The development which
follows does not really depend on this assumption and
others may be considered.)
The following notation will be used :

M = number of prefix areas in the population,


n = total sample size (of residential numbers),
m = number of sample prefix areas desired,
k + l = cluster size in sample,
n = m(k
11,
P = intraclass correlation within prefix areas,
Pi = number of residential numbers in ith prefix
area,
7 r = C Pi/100M, i.e., the proportion of residential numbers in the population. (Note: If
the prefix area is defined differently, e.g., to
include only the three-digit area and threedigit prefix code, the denominator of A will
change to 10,000; all other aspects of the
notation will be unchanged.)
lOO7rM = number of residential units in the
population,
t = proportion of prefix areas with no residential
numbers,
uz2 = sampling
variance of statistic being
estimated,
= population variance ,for statistic,
s = number of surveys to be carried out with the
sample design,
c, = cost of an unproductive call, i.e., to a nonresidential number,
c, = cost of a productive call; includes both interviewing and processing costs.

Where a survey involves screening a large number of


households to identify a subset eligible for interview, it
is possible to define the sample size as the number screened
or the number interviewed. The values of C , and C ,
should, of course, relate to the definition of the units.

Journal of the American Statistical Association, March 1978

42

The probability that the ith area is selected in a single


To make decisions on which values of m and k to choose
for the sample, estimates of some of the parameters given draw is 1/M; the probability, that once it is selected it
in the preceding list are needed. The most crucial one ap- will be retained in the sample, is P,/lOO. Its overall
pears to be the value of t. Groves (1977) provides data probability of entering the sample in a draw is thus
for a national study implying a value of t of about .65. Pi/lOOM. The probability that nothing is selected in a
There do not appear to be any data which can provide single draw is 1 - T . The probability that the ith area is
estimates of t for individual states, metropolitan areas, or the first unit selected is
cities. It seems reasonable to assume that in the more
densely populated urban areas the value of t will be much
lower than .65 and in rural areas it will be higher. The
efficiency of the sample design, relative to simple random
Since the sampling is with replacement, the same probasampling, will, therefore, be lower in the larger metropolibilities apply to each of the m selections and the probatan areas and better in areas with large rural populations.
bility of area i entering the sample with m units selected
The only restriction necessary on the definition of a
is mP,/lOOrM.
block is that it should be large enough so that each
To obtain a self-weighting sample, the sampling within
Pi 2 s(k
1 ) . A two-digit block should generally satisfy
the prefix area must be inversely proportionate to these
this criterion. The value of p will generally be fairly small.
numbers. Since the overall sampling rate should be
Groves (1977) reports values of p of about .02 or .03 for
m ( k 1)/100rM, the within-area rate has to be
attitude questions and about .05 for economic or social
(k l ) / P , . If this is applied to the P , residential numstatistics. The value of IT seems to be about .2 for national
bers within an area, the sample size within the area will
studies but is probably larger within the larger metropolibe exactly k
1 . This is similar to the usual situation
tan areas. The ratio of C,/C, will depend on the survey
for a self-weighting two-stage sample, with pps sampling
plans.
a t the first stage, except that generally the cluster sizes
average
out to k
1 ; here they are all equal to k
1.
3. CLUSTER SAMPLING W I T H EQUAL PROBABILITIES
Let
us
examine
the
total
number
of
telephone
calls
OF SELECTION OF PREFIX AREAS
required to implement this system. Two groups of teleTo derive a sample size of m ( k 1 ) cases, the overall
phone
calls are necessary. The first is to establish the set
sampling rate is
of m sample prefix areas. The calls that establish the
m(k
1)/1001~M= n/100?rM .
(3.1) sample areas also provide one residential number within
With equal probability of selection of m prefix areas, the each area. The second group is necessary to locate an
probability of selecting an area is m / M and the probabil- additional k residential numbers within each area. The
number of calls is a random variable, and the expected
ity of selection within areas must be k
1/100s.
value
is derived in the following paragraphs.
Since the overall probability of selection is n/100nM,
In
a
single draw, the probability that a prefix area will
and there are 1OOM possible numbers in the population,
be
selected
is
the expected number of telephone calls that will be made
is n / r .
1 b4 Pi
The expected number of calls to be made is thus in-X--=7r.
(4.2)
M 1 100
dependent of k. Since the total cost is a function of the
number of telephone calls, it is independent of the cluster The probability of not making a selection is 1 - I T . The
size. Consequently, with this sampling scheme, the most probability that k calls are needed to retain one PSU is
efficient choice is to use no clustering; i.e., for each tele- ( 1 - I T ) ~ - ~ I T . Hence the expected number of calls t o
phone call a new ten-digit number is to be selected a t retain one PSU is
random.
d
1
4. CLUSTER SAMPLING W I T H PROBABILITY PROPORC k ( l - I T ) ~ - ' T= -T- C ( 1 - T ) =~ -. (4.3)
k=O
dr
?r
TIONATE TO SIZE A N D W I T H I N CLUSTER RATE
SET T O PRODUCE A SELF-WEIGHTING SAMPLE
To retain m PSU's, the expected number of calls is,

00

Assume that each P , 2 k


1. (This assumption will
be partly eliminated in the next section.) Universe data
on the values of Pi do not exist, and, therefore, the usual
methods of pps selection cannot be applied. The following
procedure will be used t o create the required probabilities
of selection :Select a prefix area at random. Dial a number
within it a t random. If the dialed number is a residential
address, accept the prefix area. If it is not, reject the
area. Continue to select prefix areas until m accepted
areas have been reached. Use sampling with replacement.

therefore, m/T assuming sampling with replacement.


In order to get k additional cases in a prefix area that
contains Pi residential numbers, the expected number of
calls is 100k/Pi. (We assume that the original number
can be reselected in order to keep the probabilities and
expected values simple. I n practice one would prefer t o
exclude the original units. This would tend .to increase
the number of expected calls slightly.)
The expected number of calls is the sum of probabilities
of getting the ith area multiplied by the expected number

Waksberg: Random Digit Dialing Sampling

43

of calls in the area.


E(cal1s)

mPi lOOk

C -__
1

__ =

lOOTM Pi

mk

-.

(4.4)

The expected value of the total calls is the sum of the


two preceding expected values.
.
T

7r

The expected number of calls is the same with this


sa.mpling scheme as with the preceding one (prefix areas
selected with equal probability). As in the preceding case,
the total number of calls is independent of the value of k
or m. An optimum design, therefore, calls for making
k = 0, i.e., to have no clustering and to select n different
prefix areas.
5. CLUSTER SAMPLING W I T H PPS A N D W I T H SOME
PREFIX AREAS HAVING NO RESIDENTIAL
TELEPHONE NUMBERS
We now assume that P , = 0 in some prefix areas: t is
the proportion of such prefix areas. We also assume that
in the remaining ( 1 - t ) M areas, P, 2 k
1. (To
1, it is necessary
eliminate the assumption that P , 2 k
to give up self-weighting samples or to arrange to group
prefix areas so that each group has P , 2 k
1. It does
not seem worthwhile to complicate the theory in this
way.)
The expected number of calls to select a sample of m
prefix areas, as derived in Section 4, does not involve k ,
or the assumption that P , > 0. Therefore, it applies to
the assumption in this section and equals m/n.
The calls within the selected prefix areas are obviously
restricted t o the subset of (1 - t)M areas in which
P, > 0 since the t M cases have zero chance of selection.
The probabilities of selection of areas are the same as in
Section 4, and the expected number of calls within the
selected areas is thus

C
1

mPi lOOk

mk

(1 - t ) -.

--=

100TM Pi

optimum ( k

(5.1)

+ 1)
1-T-t

=-

LP,CU

Optimum Value of k

+ (1 - t ) k ] .

To determine the optimum values of m and kJ we note


that
=

m(k

+ l)Cp + [:- (1 + (1 - t ) k )
- m(k

-cu
7r

+ 1 for a Single Survey

.I

.3
7r

.6

.4

.2

.6

.4

.4

.2

4
3
2

5
3

6
4

3
2
1

3
2
1

4
3
2

3
2
1

3
2
1

4
2
2

2
2
1

3
2
1

2
1
1

2
1
1

.2

CJC, = 2

.02

.05
.10

2
1
1

2
1
1

2
2
1

3
2
1

4
2
2

c./c,. = 5
.02
.10

1
1
1

1
1
1

2
1
1

2
1
1

.02
.05
.10

1
1
1

1
1
1

1
1
1

2
1
1

.05

3
2
1

CX..=7

.10

1
1
1

1
1
1

1
1
1

2
1
1

.02
.05
.10

1
1
1

1
1
1

1
1
1

1
1
1

2
1
1

2
1
1

1
1
1

= 10

2
2
1
= 20

2
1
1

We shall now examine the sample sizes when a group


of surveys is to be carried out. We assume that the conditions described in Section 5 apply; i.e., that P , = 0 in
tM of the prefix areas. There will be s surveys and, in each
1 households are taken in each prefix area.
survey, k
The expected number of calls can be obtained by using
data from Section 5. The number of calls necessary to

.5

6. OPTIMUM SAMPLE SIZES W I T H REPEATED SURVEYS

+ 111C,
mt

CJC,

(5.2)

. (5.5)

.05

expected cost

- t)

7r

A table of optimum values of k


1, for a range of
values of the parameters, is shown in the table. (Note
that T must be less than 1 - t . ) Only slight gains as compared to simple random sampling are possible unless t is
fairly high; i.e., equal to .5 or greater.

.02

- [l

+ (1 -

___

c,/c.

(31

7r

+ (1 - t)mk

cu

Therefore,
E(tota1 calls)

tCu

(l--OM

where p is the intraclass correlation within clusters. The


optimum value of the cluster size then follows immediately (see Hansen, Hurwitz, and Madow 1953, formula
16.2, Ch. 6). I t is

(5.3)

+.

Journal of the American Statistical Association, March 1978

44

identify the m prefix areas is the same as for Section -5,


i.e., m/n. The number of calls within the prefix area is
m(1 - t ) [ s ( k

+ 1) - l]/n

(6.1)

so that
expected total calls

mrs(k

+ 1)(1 - t ) + t ] / n .

variance is

m(k

+ 1)/(1+ kp)[Cp + (1 - n)C,/nl

(7.2)

The cost of s such surveys is s times as large.


For the clustered sa.mple, the cost is

Of these, s m ( k
1) will cost C, per call and the balance
will cost C, per call. The variance of each survey will be
the same as (5.4).
The cost function can be expressed as :

sm(k
1)C,
m ( s - st

+ t ) / n { ( l- n s / ( s - st + t ) )

+ k ( l - ( t + n s ) / ( s - st + t)))CU .

This can be put into the same form as (5.3) by the


following transformation :

+ t) ,
t / ( s - st + t ) ,

and the ratio of the cost of the clustered sample to the


unclustered i s thus

[CP/C

+:

+ s ( k + 1) c,/cu + n

(1 - t

T)]

(1

+ kP)
. (7.4)

1-n

Carrying out some illustrative calculations, and using


values o f t = .5, n = .2 or .4, s = 1, 6, or 12, C,/C, = 2,
t =
and p = .02, the reductions in cost run between 20 and
C P = s c ,
40 percent. Higher gains can occur only if s is greater
than 12 or if t is greater than .5. Larger gains are possible
C, = s c , ,
if C,/C, is smaller than 2. This is likely to occur when the
C,/C, = cp/c,.
survey involves a large screening operation to identify a
small population of interest, and the screening cost is not
As a result, the optimum value of k
1 is
much higher than the cost of determining whether the
telephone number represents a household. TeIephone
(6.2) surveys are particularly efficient for such situations, and
L i C
(1
t - A - t)
the advantage of the cluster-sample approach is also
Since t < t and A > n, the optimum values of k
1 greatest under these conditions.
in this scheme will be less than the values in Section 5.
If n t and A
t are fairly large, say 2.7, the effect
8. WEIGHT1 NG, STRATIFICATION, A N D
of the term 1 - n - t will be negligible. The optimum
POSTSTRATI F ICAT10 N
will be close t o
If information on the number of households in the area
to be surveyed is not available in advance, the probabili(6.3) ties of selection are not known in advance of the survey.
However, the sample units are randomly selected and all
units
have the same chance of selection. There seems to
and it calls for k
1 for a single survey to be smaller
be
some
ambiguity in the literature as to whether the
than in the Section 5 method by a factor of about l / d s .
term
probability
sampling requires the exact probOf course, there are bigger savings in the long run, since
abilities
to
be
known.
We see no problem in including
the survey is repeated s times, and the total number of
samples
of
this
type
in
the
class of probability samples. It
units selected in each prefix area is therefore ( k
1)ds.
is
clear
that
unbiased
estimates
of proportions or other
It should be noted that the s surveys all use the same
characteristics
of
the
population
being surveyed can be
sampling frame of prefix areas and blocks. Over time,
prepared.
A
problem
does
arise
in
attempting to estimate
new prefix areas will be added, the measures of size of
totals
for
the
population.
Without
an independent estithe blocks will change, and there will be an increased bias
mate
on
the
total
number
of
households,
such estimates
arising from the uses of an out-dated frame. The s surveys
require
weighting
the
sample
numbers
by
the reciprocals
should, therefore, be thought of as conducted over a
of
the
probabilities
of
selection.
relatively short period of time, possibly a year or two.
With random digit dialing, the sampling units are the
possible
numbers in the sampling frame, which are known.
7. REDUCTIONS IN COST
If records are kept on the number dialed, the proportion
With an optimum sample design, the variance for a th a t are households will be known, permitting estimates
I
given survey will be
of the total households in the survey area, and almost
unbiased weights. (Because of the use of a stop-rule, there
may be a small bias to the estimate, but this should be
The cost of an unclustered sample having the same trivial if a reasonably large sample is used.) Unfor7

= TS/(S

- st

(+)I4.

45

Waksberg: Random Digit Dialing Sampling

tunately, the estimates of total households will be subject to a fairly high sampling error. If a moderately accurate independent estimate of the number of households
is available, it is probably preferable to use the independent estimates.
This uncertainty about the sampling rates also affects
the use of stratification. The information available for
stratification on the AT&T tapes of existing telephone
prefixes is the geographic location. Stratification by
region or state is thus possible and can be performed fairly
easily. However, stratification in the sense that the term
is usually used cannot really be effectively applied. With
finite populations, stratified sampling is usually performed by sampling a t fixed rates within strata. With
random digit dialing, because of the variability on sample
size it seems preferable to work with fixed sample sizes
rather than with predetermined sampling rates. The proposed sampling method follows this principle.
Under these conditions, stratification can only be applied by assigning fixed quotas to the strata. The quotas
are necessarily based on information availabe prior to
the start of the survey. If the information is not very
accurate, the procedure will introduce biases in the survey
results. Normally, with stratified sampling, poor prior
information will reduce the efficiency of a sample but will
not introduce any bias.
Assigning quotas to strata involves obtaining independent estimates of the number of households in each
state or region and using these estimates to establish the
quotas. Alternatively, the same data can be used to
establish poststratified estimates (or ratio estimates).
With fairly large samples, there seem to be no special
advantages to stratification as contrasted with ratio
estimates. For small samples, it probably is worth the
effort of stratification. I n either case, it is important to
superimpose on the weighting system some form of ratio
estimate, to reduce the biases in the survey arising from
the exclusion of nontelephone households and to adjust
for nonresponse. The items used to produce ratio estimates generally depend on the subject of the survey and
the kinds of information available on a 100 percent basis.
However, the major source of bias arises from the exclusion of telephone households and, obviously, availability of telephones is closely related to family income.
For example, according to the 1970 Census of Population
and Housing, only 76 percent of the households with
family incomes under $5,000 had telephones available
whereas about 95 percent of those with incomes of $25,0tx)
or more had telephones. The close relationship of telephones with income is reflected in items that correlate
with income. Thus the 1970 Census reported that 89
percent of white households had telephones but only 70
percent of black households.
For most surveys, ratio estimates by income and race
should be effective in sharply reducing potential biases.
We have found it useful to replace income in the ratio
estimates by items that correlate highly with income.
Census Bureau statistics on income are produced an-

nually, but the detailed distributions appear fairly late


in the year. Also, most surveys probably cannot produce
accurate data on income. It, therefore, seems preferable
to replace income by a correlated item that is usually reported relatively completely and accurately and also
changes slowly over time. Thus even if data were not
available for the current year, data for another period
could be used. The use of educational attainment has been
suggested by Warren Mitofsky of CBS News. Westat,
Inc. and CBS News have used it for a number of telephone surveys.

9. OTHER STATISTICAL ISSUES W I T H


TELEPHONE SURVEYS
There are several other statistical issues that arise
under any system of random digit dialing. The major
ones are summarized in the following paragraphs.
(1) A telephone sample comprises a sample of households, not persons. If one person is interviewed in the
household, a weight should be superimposed on the response; the weight is the number of persons in the household. If the interviews are to be performed for only certain classes of the population (e.g., all persons over 25
years of age) the weight is the number of such persons.
To retain the features of a probability sample, the person
in the household should be selected a t random and not
necessarily be the person who happened to answer the
phone. (Of course, if the survey is interested in household
characteristics rather than personal items, this problem
does not arise.)
(2) Households with more than one telephone number
will have multiple chances of selection. To have an unbiased system, it is necessary t o ask households if they
have more than one telephone number. If they have two,
a weight of one-half is needed, etc.
(3) I n any telephone survey, follow-up calls are necessary for calls in which there are no answers. Some noanswers represent nonworking numbers. It is sometimes
necessary to call the telephone company to check this,
after three or four calls have been made with no replies.
An efficient procedure is to make the first or second call
to a number during the daytime when business firms will
answer, so that business numbers can be eliminated
quickly.
(4) Random digit dialing to achieve a fixed sample
size is, in a sense, a sequential procedure. Theoretically,
beyond a certain point it is necessary to clear up the
classification of all uncertain rings, busy signals, etc.
before proceeding further. This creates some awkwardness
in the operations of the survey. A good record-keeping
system and careful supervision are necessary to keep the
process under control. The two-stage sample design proposed in this article is somewhat more complex than
simple random sampling and thus increases the need for
tight control. However, there are some elements that tend
to simplify the operations. I n using the system, both
Westat and the University of Michigan Survey Research

Journal of the American Statistical Association, March 1978

46

Center have found it convenient to complete most of the


work on the first stage of the sample before proceeding
with the second. For the second stage, each block can be
treated as a separate entity within which k additional
units are to be selected. The sequential aspects a t this
point are contained within small groups and are kept
track of more easily.
(5) Finally, not all surveys can be performed as telephone interviews. Personal interviewing is usually necessary if the interview is very long, flash cards are necessary,
the respondent needs to refer to records, personal observation by the interviewer is required, there is a special need
for rapport between the interviewer and respondent, or
identification of the interviewer as a bona fide employee
of the research organization is necessary. I n addition]
since telephone surveys start off with a bias (exclusion of
nontelephone households) and it is frequently not possible to attain as high a response rate as with personal
interviews. telephoning may not be advisable when the
survey focuses particularly on the low-income population
or when a very high degree of accuracy is required. These
considerations do not necessarily apply to the use of
telephones for follow-up, when initial information was
obtained in a personal interview.
IReceived December 1976. Revised July 1977.3

REFERENCES
Chilton Research Services (1967), Continuous Tracking Studies
via WATS Lines and Personal Interviewing, report presented to
the Advertising Research Foundation, 13th Annual Conference,
Chilton Research Services, Phila., Pa.
Fletcher, James E., and Thompson, Harry B. (1974), Telephone
Directory Samples and Random Telephone Number Generation,
Journal of Broadcasting, 18.2, 187-191.
Hansen, Morris H., Hurwitr, William N., and Madow, William G.
(1953), Sample Survey Methods and Theory, 1, New York: John
Wiley & Sons.
Glasser, Gerald J., and Metrger, Gale D. (1972), Random Digit
Dialing as a Method of Telephone Sampling, Journal of Marketing
Research, 9, 5964.
, and Metzger, Gale D. (1975), National Estimates of Nonlisted Telephone Households and Their Characteristics, Journal
of Marketing Research, 12, 35S361.
Groves, Robert M. (1977), An Empirical Comparison of Two
Telephone Sample Designs, unpublished report of the Survey
Research Center of the University of Michigan, Ann Arbor, Mich.
Hauck, Mathew, and Cox, hlichael (1976), Locating a Sample by
Random Digit Dialing, Public Opinion Quarterly, 40, 253-260.
Lahiri, D.B. (19.51), A Method of Sample Selection Providing Unbiased Ratio Estimates, Bulletin International Statistical Institute,
33, 133-140.
Rogers, Theresa F. (1976), Interviews by Telephone and in Person:
Quality of Responses and Field Performance, Public Opinion
&uarterly, 40,51-65.
Roslow, Sydney, and Roslow, Laurence (1972), Unlisted Phone
Subscribers Are Different, Journal of Advertising Research, 12,
3S38.
Sudman, Seymour (1973), The Uses of Telephone Directories for
Survey Sampling, Joumal of Marketing Research, 10, 204-207.

Das könnte Ihnen auch gefallen