Appendix L_sample design

Appendix L
HINTS 2007 Sample Design

APPENDIX L
HINTS 2007 SAMPLE DESIGN

The sample design for HINTS 2007 consists of two samples with each sample being selected
from a separate sampling frame. One sample will be a list-assisted random digit dialing (RDD) sample
selected from all telephone exchanges in the United States, following the design of HINTS 2003 and
HINTS 2005. This will result in a nationally representative sample of telephone households. The second
sample is new and comprises addresses selected from a list based on United States Postal Service (USPS)
administrative records. Questionnaires will be mailed to the address sample, and telephone interviewers
will follow up with mail nonrespondents.
During the household screener for the RDD sample, one adult will be sampled within each
household and recruited for the extended interview. For the address sample, all adults in the household
will be asked to complete a survey. The sample design will yield approximately 7,000 completed
interviews: 3,500 from the RDD sample and 3,500 from the address sample.

L.1

RDD sample
A list-assisted RDD sample is a random sample of telephone numbers from all working

banks in U.S. telephone exchanges (see for example Tucker, Casady, and Lepkowski, 1993). A working
bank is a set of 100 telephone numbers (e.g., telephone numbers with area code 301 and first five digits
294-44) with at least one listed residential number.1

L.1.1

Size of RDD Sample and Expected Yields
Table L-1 presents expected sample sizes for the RDD sample. A total of 59,020 telephone

numbers are to be sampled, with an expected yield of 3,500 completed interviews. A reserve sample of
29,510 telephone numbers will also be sampled and set aside to be used in case expectations are not met

1

Note that all numbers whether listed as residential or not are part of the sampling frame, as long as they are in working banks.

L-1

(i.e., a total of 88,530 telephone numbers will be initially sampled, with 29,510 then set aside as the
reserve).
Table L-1.

RDD sample’s expected completed screeners and completed extended interviews
RDD sample
59,020
91.7%
35.9%
19,422
35.0%
85.8%
5,833
60.0%
3,500

Sampled telephone numbers
Reduction rate due to nonmailable subsampling
Residency rate
Residential numbers
Screener response rate
Reduction rate due to refusal subsampling
Completed screeners
Extended interview response rate
Yield of extended interviews

NOTE: All figures in the table are rounded, leading to arithmetic inconsistencies (a*b equals c, but rounded a * rounded b is not equal to
rounded c).

Our sample design for HINTS 2007 will subsample numbers of two kinds. First, we will
sample out 13.2 percent of the nonmailable numbers (numbers for which we have no address
information), as discussed in Section L.1.2. This subsampling will deselect a total of 4,917 nonmailable
numbers (8.3% of the entire sampled numbers) leaving 54,102 numbers. Then we will further subsample,
from each stratum by mailable status, 49.7 percent of the initial screener refusals and noncontacts
(numbers for which we obtain a certain form of human contact with nonhostile refusal or noncontact at
the first round of screening calls). This subsampling will exclude 4,288 such numbers allowing us to
focus only on 50.3 percent of such numbers for conversion. The first type of subsampling was conducted
for HINTS 2005 but the second type of subsampling was not. Sections L.1.2 and L.1.3 discuss these two
types of subsampling in detail.

L.1.2

Stratification by Mailable Status
We will use stratification by mailable status (see for example Brick, et al., 2002). Numbers

that are mailable are those for which we have an address. In HINTS 2005,2 36.9 percent of the RDD
sample was mailable, with 63.1 percent nonmailable. We expect similar percentages in HINTS 2007
(though there may be some change). The mailable numbers have a much higher percentage of residential

2

We computed the HINTS 2005 mailable rate and response rates from all of the Wave 1 data and a portion of Wave 2 data. We believe this
provides the most reliable estimates of various HINTS 2005 rates for purposes of planning the HINTS 2007 RDD sample.

L-2

numbers. Additionally, we have seen in HINTS 2005 that we obtain a higher screener response rate
among the mailable numbers, especially when we send incentives with an advance letter to the mailable
numbers. Computations using HINTS 2005 results show that the cost for each completed extended
interview among the nonmailables can be expected to be 1.33 times larger than that for a completed
extended interview among the mailables, thereby justifying an explicit stratification by mailable status.
The optimal rate for the nonmailable stratum is 86.8 percent of that of the mailable stratum. Table L-2
presents the sample design with this subsampling rate.
The residency rates and mailable percentages are from HINTS 2005. The differential
screener response rates (36.0% for the mailable stratum and 28.8% for the nonmailable stratum) reflect
the difference in response rates (7.2 percentage points) that we saw in HINTS 2005 and other recent
Westat RDD studies between the nonmailable stratum and the mailable stratum with $2 incentive.
The same extended-interview response rates for the mailable and nonmailable strata reflects
our HINTS 2005 observation that there was no difference in extended interview response rates between
the mailable stratum with the $2 advance letter incentive and the nonmailable stratum.

L.1.3

Subsampling of Refusals
We will also apply refusal subsampling (e.g., Brick, et al., 2005). Numbers that are screener

refusal or noncontact are those for which we obtain either human contact with nonhostile refusal or
nonhuman contacts during the first round of the screener calling attempt. Based on HINTS 2005 data, we
anticipate 25 percent and 44.4 percent of residential numbers will be cooperative and refusal-noncontact
numbers, respectively, in the first round of the screening calls. From about 8,618 screener refusal and
noncontact numbers, we will select 50.3 percent of them (4,342 numbers) for carrying out refusal
conversion. Table L-2 shows the sample design with this subsampling rate. We anticipate a total of 5,833
screener completes with 4,856 initial screener completes and 978 additional screener completes that are
either initially refused or noncontact.

L-3

Table L-2.

Proposed mailable stratification and refusal subsampling sample design

Original numbers

Mailable
21,766

Subsampling rate

100.0%

Sampled telephone numbers

21,766

Residency rate

77.2%

Residential numbers

16,783

Screener response rate
Initial response rate
Initial refusal rate
Refusal conversion rate
Refusal subsampling rate

36.0%
26.0%
44.7%
22.9%
50.3%

Completes with initial cooperation
Completes with initial refusal
Screener completes
Extended interview response rate
Overall response rate
Yield of extended interviews
Effective sample size*

Mailable
percentage
of total
36.9%

Nonmailable
37,254

Nonmailable
percentage
of total
63.1%

Total
59,020

59.8%

54,102

86.8%
40.2%

32,337
8.1%

86.5%

2,619

35.9%
13.5%

28.8%
18.6%
42.6%
19.9%
50.3%

4,369
866
5,235

90.0%
88.5%
89.7%

60.0%
21.6%

486
112
5980

35.0%
25.0%
44.4%
22.5%
50.3%
10.0%
11.5%
10.3%

60.0%
17.3%

3,141

89.7%

359

19,422

4,856
978
5,833
60.0%
21.0%

10.3%

3,500
2,692

* Effective sample size was figured with assuming design effect of 1.3.

L.1.4

Effective Sample Size for Domains of Interest
In HINTS 2005, the number of completed RDD extended interviews was 5,493.3 For HINTS

2007 the expected number of completed RDD extended interviews is 3,500. The HINTS 2007 RDD
sample is smaller than previously because there will also be a HINTS 2007 address sample. Table M-3
presents Current Population Survey (March 2005 supplement) estimates of adults within the domains of
interest, with expected sample sizes proportional to these estimates. The effective sample sizes (the
sample size of a simple random sample with the same precision) are smaller by a factor of 1.3: we expect

3

In HINTS 2005, we have 5,586 extended interview completes, including 93 web-extended interview completes.

L-4

a design effect4 of 1.3, which allows for adult selection within households (generating variable weights
for adults for differing size households) that generally has a design effect of 1.2, mailable-nonmailable
subsampling (see Section L.1.2), screener refusal subsampling (see Section L.1.3), and nonresponse
weighting adjustments.
Table L-3.

RDD sample’s expected completed extended interviews by race

Hispanic
Non-Hispanic black
Non-Hispanic white and other race
Total

L.2

Adults in U.S.
population
(in 1,000s)
27,509
24,916
164,910

Percentage
of adults
12.66
11.46
75.88

217,334

100.00

Expected
completes
399
407
2,694
3,500

Expected
effective
sample size
341
309
2,043
2,692

Address Sample
The address sample will be a stratified sample selected from a list of addresses.

Questionnaires will be mailed to the address sample, and all adults at each sampled address will be asked
to complete a separate questionnaire. Telephone interviewers will follow up those mail nonrespondents
for which it is possible to reverse match telephone numbers to addresses. Telephone followup interviews
will use the RDD calling protocol and interview instruments, in which only one adult will be sampled and
interviewed (see Appendixes I and J).

L.2.1

Sampling Frame for Address Sample
The sampling frame for the address sample will be a database of addresses used by

Marketing Systems Group (MSG) to provide random samples of addresses. Our decision to use this
database as a sampling frame is the result of an evaluation study conducted by Link and colleagues.
(2005). This study compared five address vendors in terms of the coverage of their lists for a six-state
area. Three vendors had high levels of undercoverage in one or more of the six states. Of the remaining
two vendors, only MSG could provide sampling services for a single-stage sample of addresses. The use
4

Design effect is defined as the ratio of the actual sample variance to the variance of a simple random sample with the same sample size. See for
example Kish (1965, p. 162).

L-5

of the other vendor would have required two stages of sampling—first the sampling of carrier routes and
then the sampling of individual addresses. Compared with a single-stage design, a two-stage design for
selecting addresses is more costly and provides less precision for a given sample size.
The MSG address database is updated bimonthly from the USPS’s Computerized Delivery
Sequence (CDS) File. Licensed by the U.S. Postal Service to qualified address vendors, the CDS is an
electronic data product that provides and updates addresses by carrier route (USPS, 2006). Address
vendors must initially qualify for the CDS information for a given five-digit ZIP Code area by having at
least 90 percent but not more than 110 percent of the all the addresses in the ZIP Code area. Once a
vendor has qualified for a five-digit ZIP Code area, CDS information is made available bimonthly via
electronic media.
The CDS contains current information on all mailing addresses serviced by the U.S. Postal
Service, with the exception of general delivery. CDS information is available for the following types of
addresses:


Address Type 1. Addresses that currently receive or have received mail delivery.



Address Type 2. Addresses on city routes to which carriers do not deliver because of
alternative delivery arrangements, e.g. to post office boxes. (Referred to as
throwbacks, these addresses can be included in or excluded from MSG-provided
samples of addresses.)



Address Type 3. Addresses on city routes vacant longer than 90 days and that are
likely to be long-term vacancies, that are not considered seasonal. (Referred to as
vacants, these addresses can also be included in or excluded from MSG-provided
samples of addresses.)



Address Type 4. Addresses delivered seasonally. (No CDS information is available,
however, on the dates of the mailing season. Referred to as seasonals, these addresses
can also be included in or excluded from MSG-provided samples of addresses.)

The availability of bimonthly CDS data allows address vendors to frequently update their
address lists. Another address vendor that uses CDS data for updating is Advo, Inc. Staab and
Iannachione (2003) evaluated the coverage for the Advo mailing list in 2002. They compared the number
of addresses on the Advo mailing list (excluding vacant and seasonal addresses) to census household
projections generated by Claritas, Inc., for 29,000 local areas. They found that for local areas containing
10,000 or more households, the totals number of residential city-style addresses exceeded the number of
households. For areas containing fewer than 10,000 households, the total number of residential city-style

L-6

addresses was only 86.3 percent of the total number of households. However, if post office boxes and
rural route addresses were also included, the total number of addresses exceeded the total number of
households for local areas containing fewer than 10,000 households.
Link and colleagues (2005) evaluated the coverage of the MSG address list for the six states
of California, Illinois, New Jersey, North Carolina, Texas, and Washington. For each of the counties in
this six-state study area, they compared the number of addresses on the MSG list as of April 1, 2005, to
the U.S. Census Bureau’s estimated number of households for July 1, 2003. They tabulated the number of
counties in which there was a high level of undercoverage that they defined as the number of addresses on
the MSG list for the county as less than the number of households in the county by at least 10 percent.
They found that in counties where less than 25 percent of the population lives in an urban area that nearly
90 percent of the counties had a high level of undercoverage; whereas in counties where 75 percent or
more of the population lives in an urban area, only 4.3 percent of the counties had a high level of
undercoverage.
We plan to include a question on the mail questionnaire about the different ways respondents
receive mail. The responses to this question will be used in the calculation of sampling weights to adjust
for the duplication of households in the sampling frame. We also plan to investigate weighting
adjustments to reduce the effects on survey estimates of the sampling frame’s undercoverage of rural
areas.

L.2.2

Selection of Main-Survey Address Sample
The sampling unit for the address sample will be an individual address. We plan to subject to

sampling all residential addresses on the MSG database, including post office boxes, throwbacks, vacant
addresses, and seasonal addresses. Following the selection of the address sample, we will use the
AUTOMATCH computer program to compare the address sample with the addresses of telephone
numbers assigned to the mailable stratum of the RDD sample. Addresses in both the address sample and
the RDD sample will not be contacted by RDD telephone interviewers. (Telephone numbers assigned to
the nonmailable stratum will not have addresses, so they cannot be tested for membership in the address
sample.) Envelopes containing mailed questionnaires will be marked “Do Not Forward” so that address
changes will not provide multiple opportunities for a household to be selected for the address sample.

L-7

The address sample will have two strata: one containing a high concentration of minority
adults and the other containing a low concentration. Each address on the MSG database will be assigned
to one of the two minority strata by an algorithm based on linking each address to a geographic
assignment area for which MSG has demographic data by race and ethnicity. One possibility for the
stratification algorithm is to use ZIP+4 Codes to link addresses to census block groups, for which MSG
has Claritas-provided demographic data. Another possibility is to use ZIP Codes to link addresses to
telephone exchanges, for which MSG also has associated Claritas-provided demographic data. The
development of the stratification algorithm will be done in consultation with MSG data-product experts.
For planning purposes, we are using telephone exchanges as assignment areas and defining the highminority stratum as those addresses in ZIP Codes linked to telephone exchanges in which the black or
Hispanic population proportion is 16 percent or greater.
An equal-probability sample of addresses will be selected from each sampling stratum. The
high-minority stratum will be oversampled by 50 percent to increase the yield of blacks and Hispanics.
For example, if 50 percent of all the addresses in the sampling frame were assigned to the high-minority
stratum, the oversampling of the high-minority stratum would assign to the high-minority stratum
75 percent of the sample, rather than the population percentage of 50 percent.
Unlike the RDD sample, all adults in the household at a sampled address will be asked to
complete a questionnaire. Hence, the mail sample is a stratified cluster sample, in which the household is
the cluster. Our decision to not subsample the adults in sampled households is the result of an evaluation
study conducted by Battaglia and colleagues (2005). This study compared three respondent-selection
methods for household mail surveys: (1) any adult in the household, (2) the adult in the household having
the next birthday, and (3) all adults in the household. The study found that the next birthday and all-adults
methods yielded household-level completion rates that were comparable with the any-adult method, the
method that the researchers assumed to have the least respondent burden. Another finding from this study
was that differences in response rates by gender and age were less for the all-adults methods than for the
next birthday and all-adults method.
Because we will be using the all-adults method to select respondents, two types of
household-level nonresponse must be considered. One type of nonresponding household will be withinhousehold nonrespondent—that is, a household in which some but not all adults in the household,
complete the mail questionnaire. To handle this situation the mail questionnaire will include a question
about the number of adults living in the household. The responses to this question, plus the number of

L-8

adults in the household that do respond, will be used to calculate a within-household sampling weight to
be applied to the data provided by the household’s respondents. The other type of nonresponding
household will be an entire-household nonrespondent. A sample of entire-household nonrespondents will
receive telephone followup using the RDD data collection protocol. Only one adult will be interviewed in
the cooperating address sample of households assigned to telephone followup.
The target number of completed mail questionnaires is 3,500, and the target number of
completed RDD questionnaires resulting from telephone followup is 457. Table L-4 contains the number
of sampled addresses needed to obtain these targets and the planning assumptions we used to determine
these results.
Table L-4.

Address sample expected completions and telephone followup calls

Total
6,944

Highminority
stratum
5,208

Lowminority
stratum
1,736

25%

25%

25%

1,736

1,302

434

Average number of adults per household

2.52

2.52

2.52

Within-household response rate

80%

80%

80%

Number of completed mail questionnaires

3,500

2,625

875

Number of households not responding to mail survey

5,208

3,906

1,302

70%

70%

70%

3,657

2,743

914

Screener response rate for telephone followup

25%

25%

25%

Extended response rate for telephone followup

50%

50%

50%

457

343

114

3,957

2,968

989

1.6

1.20

1.20

2,473

2,473

824

Number of sampled addresses
All-mailings household response rate
Number of households responding to mail survey

Subsampling rate for telephone followup
Number of households assigned to telephone followup

Number of completed telephone followup questionnaires
Number of completed mail + followup questionnaires
Design effect
Effective sample size

L.2.3

Effective Sample Sizes by Domain of Interest
For HINTS 2007, the expected number of completed mail questionnaires is 3,500 and the

expected number of completed telephone followup interviews is 457. The third to last row of Table L-4

L-9

contains the expected number of completed mailed questionnaires and telephone followup interviews by
stratum (using telephone exchanges as assignment areas). The effective sample sizes in the last row of
Table L-4 are smaller by a factor of 1.2. We expect an approximate design effect of 1.2 for the completed
mailed questionnaires (due to within-household correlation and weighting adjustments for withinhousehold nonresponse) and also for telephone followup (due to the selection of one adult within
households that generates variable weights for adults for differing size households). Table L-5 contains
estimates of the address sample’s effective sample sizes by strata and analysis domains of interest.
Table L-5.

Address-sample’s effective sample sizes by stratum and analysis domains of interest
Proportion of
population
(%)
11.9
12.2
75.9
100.0

Proportion of
stratum
(%)

Expected
completes
651
665
2,641
3,957

Effective
sample size
520
528
1,425
2,473

Stratum
Total

Analysis domain
Hispanic
Non-Hispanic black
White and other race
All

High minority

Hispanic
Non-Hispanic black
White and other race
All

10.5
10.6
28.7
49.8

21.0
21.4
57.6
100.0

624
634
1,711
2,968

520
528
1,425
2,473

Low minority

Hispanic
Non-Hispanic black
White and other race
All

1.4
1.6
47.2
50.2

2.8
3.2
94.0
100.0

27
32
930
989

23
26
775
824

L.3

Calculation of Weights for Composite Estimates
Domains are subsets of samples defined by respondent-provided data. Domain statistics

estimate parameters for corresponding subpopulations. For example, data from respondents that indicate
they are Hispanic are used to estimate parameters for the Hispanic subpopulation. We plan to include on
the mail questionnaire a question about whether or not the respondent’s household has one or more landline telephone numbers that answers to conduct household telephone calls. The responses to this question
and the mode of data collection define the following three estimation domains:


Noncallable Domain. Address-sample respondents who complete the mail
questionnaire and indicate their household does not have one or more land-line
telephone numbers.

L-10



Address-Sample Overlap Domain. Address-sample respondents who (1) complete a
mail questionnaire and indicate that their household does have one or more land-line
telephone numbers or (2) provide data via telephone followup.



RDD-Sample Domain. RDD-sample respondents.

We will use the estimation domains to calculate weights for composite estimates—i.e.,
estimates based on data from both the RDD sample and the address sample. A composite estimate for a
population totals combines estimated totals from the three estimation domains as follows:
( address )
( address )
( RDD )
Tˆcomposite  TˆNC
  Tˆoverlap
 (1   )TˆRDD
,

where

Tˆcomposite  the composite estimate of a population total,
( address )
TˆNC
 the estimated total for the noncallable domain, calculated from address-sample

data,
( address )
 the estimated total for the address-sample overlap domain, calculated from
Tˆoverlap

address-sample data,
( RDD )
TˆRDD


the estimated total for the RDD domain, calculated from RDD-sample data, and

 , called the mixing parameter, satisfies 0 <  < 1 and is chosen to minimize the variance
of resulting composite estimates.
We will calculate composite weights so that composite estimates of totals can be calculated
as weighted sums of the data from both the RDD sample and the address sample and composite estimates
of means and proportions can be calculated as weighted totals divided by sums of composite weights. A
four-step procedure will be used to calculate weights, with the results of each step denoted as follows:
ws

=

sample-specific base weights

ws’

=

sample-specific nonresponse-adjusted weights

wc’

=

composite nonresponse-adjusted weights

wc” =

composite calibration weights

L-11

The sample-specific base weight, ws, for a respondent is the reciprocal of its probability of
being included in a particular sample (i.e., RDD sample or address sample). The RDD-sample base
weight for a respondent will be adjusted for multiple selection opportunities if the respondent’s household
has two or more land-line telephone numbers that it answers. Similarly, the address-sample base weight
for a respondent will be adjusted for multiple selection opportunities if the respondent’s household
receives mail at two or more addresses, such as home delivery to a street address and a post office box.
The sample-specific, nonresponse-adjusted weight, ws’, will be calculated by multiplying ws by samplespecific, nonresponse adjustment factors calculated from counts of sampled units and responding units
within nonresponse adjustment cells. The composite nonresponse-adjusted weight, wc’, will be calculated
as follows:





In the noncallable domain, wc’ = ws’ (where ws’ is for the address sample).
In the address-sample overlap domain, wc’ =  ws’ (where 0<  < 1 and ws’ is for the
address sample).
In the RDD-sample domain, wc’ = (1-  ) ws’ (where 0<  < 1 and ws’ is for the RDD
sample).

If we find that it is possible to assign RDD-sample respondents to their corresponding
address-sample strata (e.g., if ZIP Code areas are assignment areas, and it is possible to link telephone
exchanges to ZIP Code areas), then the value of used for the high-minority stratum can differ from the
value used for the low-minority stratum. If this is not possible, however, then the same value of  will be
used for all respondents. The composite calibration weights, wc”, will be calculated by modifying the
nonresponse-adjusted calibration weights so they aggregate to control totals computed from the Current
Population Survey, which has a much larger sample size than HINTS 2007.
Table L-6 contains the effective sample sizes by stratum for RDD-sample estimates, addresssample estimates, and composite estimates. For the composite estimates, Table L-7 contains the
maximum standard errors of estimated proportions in the race/ethnicity domains of interest and the halfwidths of the associated 95-percent confidence intervals.

L-12

Table L-6.

Effective sample sizes for composite estimates by stratum

Proportion of adults with landline telephones*
Telephone-sample effective sample size
Address-sample effective sample size
Effective sample size for composite estimates for various mixing
parameters:
0.40
0.45
0.50
0.55
0.60
0.65
0.70

High-minority
stratum
93%
1,346
2,473

Low-minority
stratum
91%
1,346
824

3,127
3,346
3,534
3,674
3,752
3,761
3,698

1,971
1,927
1,855
1,761
1,652
1,536
1,417

* Based on Blumberg, S.J., Luke, J.V., and Cynamon, M.L. (2006). Telephone coverage and health survey estimates: Evaluating the need for
concern about wireless substitution. American Journal of Public Health, 96, 926-931.

Table L-7.

Maximum standard errors and half-widths of associated 95-percent confidence intervals for
composite estimates of proportions in race/ethnicity domains of interest

Hispanics
(%)

Non-Hispanic
blacks
(%)

Non-Hispanic
whites and
other races
(%)

All adults
(%)

Maximum standard error of estimated domain
proportions for various mixing parameters*:
0.40
0.45
0.50
0.55
0.60
0.65
0.70

1.90
1.85
1.82
1.80
1.80
1.81
1.84

1.87
1.83
1.80
1.78
1.78
1.80
1.83

0.85
0.85
0.85
0.87
0.89
0.91
0.95

0.89
0.86
0.84
0.82
0.82
0.82
0.82

Half-width of 95-percent of confidence
intervals about estimated domain proportions
for various mixing parameters:
0.40
0.45
0.50
0.55
0.60
0.65
0.70

3.72
3.62
3.56
3.52
3.52
3.55
3.61

3.67
3.58
3.52
3.49
3.49
3.52
3.58

1.66
1.66
1.67
1.70
1.74
1.79
1.85

1.75
1.69
1.65
1.62
1.60
1.60
1.61

* Standard error when estimating stratum proportions equal to 50 percent.

L-13

REFERENCES

Battaglia, M.P., Link, M.W., Frankel, M.R., and Mokhad, A.H. (2005). An evaluation of respondent
selection methods for household mail surveys. Proceedings of the Section on Survey Research
Methods (pp. 2727-2731), American Statistical Association.
Blumberg, S.J., Luke, J.V., and Cynamon, M.L. (2006). Telephone coverage and health survey estimates:
Evaluating the need for concern about wireless substitution. American Journal of Public Health,
96, 926-931.
Brick, J.M., Judkins, D., Montaquila, J., and Morganstein, D. (2002). Two-phase list-assisted RDD
sampling, Journal of Official Statistics, 18, 203-215.
Brick, J.M., Montaquila, J., Hagedorn, M.C., Roth, S.B., and Chapman, C. (2005). Implications for RDD
design from an incentive experiment. Journal of Official Statistics, 21.
Kish, L. (1965). Survey sampling. New York: John Wiley & Sons.
Link, M.W., Battaglia, M.P., Giambo, P., Frankel, M.R., Mokhad, A.H., and Rao, S.R. (2005, May).
Assessment of address frame replacements for RDD sampling frames. Paper presented at the 2005
American Association for Public Opinion Research Annual Conference, Miami, FL.
Staab, J.M., and Innacchione, V.G. (2003). Evaluating the use of residential mailing addresses in a
national household survey. Proceedings of the section on survey research methods (pp. 40284033), American Statistical Association,
Tucker, C., Casady, R., and Lepkowski, J. (1993). A hierarchy of list-assisted stratified telephone sample
design options. Paper presented at the annual conference of the American Association for Public
Opinion Research, St. Charles, IL.
U.S.

Postal Service. (2006). CDS User
http://ribbs.usps.gov/files/CDS/CDSug.pdf.

Guide.

L-14

Retrieved

April

20,

2006,

from