ABS Sampling

Attachment 9. Address-based Sampling Frame
RTI has conducted a significant body of research on the use of address-based frames for
household surveys, particularly regarding issues of coverage compared with traditional “counting
and listing” approaches. (Iannacchione, Staab, and Redden, 2003; Iannacchione et al, 2007;
McMichael, Ridenhour, and Shook-Sa, 2008; Iannacchione et al, 2010; Shook-Sa et al, 2010;
Staab and Iannacchione, 2003). Our research has shown that the coverage of the household
population in North Carolina using mailing addresses is comparable to the coverage of counting
and listing in urban areas, but is somewhat lower in rural areas (Iannacchione et al, 2007).
Currently, the sampling frame for the NCVS is maintained by the U.S. Census Bureau. As such,
it is subject to Title XIII restrictions which do not allow it to be shared with research contractors.
In contrast, mailing addresses are offered to the public by the U.S. Postal Service (USPS)
through a nonexclusive license agreement with qualified private companies. One such company
is Valassis Direct Mail, Inc. In July 2010, the Valassis Lists product accounted for all but 35,000
of the more than 137 million residential mailing addresses on the U.S. Postal Service
Computerized Delivery Sequence (CDS) File. The CDS File contains all postal delivery points
serviced by the USPS. In addition to the CDS file, the USPS makes available the No-Stat file, a
file of mailing addresses that supplement the CDS file with both active and vacant addresses that
are excluded from the CDS file.
Selecting the Sample. The sampling frame will be comprised of addresses from the CDS and NoStat files. We will select a two-stage sample of addresses. In the first stage, we will select a
probability proportional to size (PPS) sample of 64 primary sampling units (PSUs). Primary
sampling units (PSUs) are five-digit ZIP codes and the size measure is the number of addresses
in the PSU. After selecting PSUs, we will randomly assign them to one of the four mode and
incentive conditions. In the second stage of sample selection we will select a sample of addresses
from each sampled PSU and append phone numbers to as many sampled addresses as possible.
All English-speaking adults residing at the selected addresses will be eligible to complete the
interview.
Although it is not unreasonable to assume that virtually every household in the United States has
a mailing address, not all mailing addresses are suitable for in-person household surveys because
interviewers must be able to locate a mailing address “on the ground.” Households with citystyle1 mailing addresses are considered locatable for in-person household surveys and constitute
the vast majority of elements on the CDS file. Households with mailing addresses that are not
locatable include those with simplified rural addresses2 and households that only receive mail
through residential Post Office (P.O.) Boxes. In addition, the CDS file contains some addresses
that are incomplete. Drop points are addresses where mail is delivered to a single location for
multiple units. The CDS file contains the drop point address and the number of drop units but
does not include drop unit descriptors. The No-Stat file contains drop-unit descriptors for a
1

A city-style mailing address contains a street name and number as well as city, state, and ZIP Code.
A simplified rural address does not have a street address. Mail delivery is based on the resident’s name, city, state,
and ZIP Code. Typically, simplified rural addresses are assigned to all households on a rural carrier route.
2

portion of drop units on the CDS. Thus inclusion of drop points that are not contained on the NoStat file would require in-field sample selection procedures.
The sampling frame for this study consists of 16,567,614 active, complete, locatable residential
mailing addresses within NC, OH, PA, and VA. These addresses were derived from Valassis’
May, 2011 CDS and No-Stat Files3. We did not include vacant and seasonal addresses on the
frame because the target population is limited to occupied households. We excluded the 170,7034
active drop units for which we could not identify complete drop unit addresses from the frame
because including them would require additional field selection procedures.
In order to maintain an EPSEM (equal-probability-of-selection-method) sample, the same
number of addresses will be selected from each PSU. To achieve this, each PSU must have at
least n / nc addresses, where n = the total sample size, and nc = the number of clusters. For the
SCV, n=3,840 and nc=64, so each PSU must contain at least 60 addresses. However, in addition
to selecting 60 primary addresses from each PSU, we will select 30 additional addresses as a
hold sample. This hold, or supplemental sample, will be released if response rates are lower than
expected and additional completed interviews are required. To allow for the selection of a hold
sample, and to prevent every address within a PSU from being selected, we set the minimum size
as 120 addresses. The 256 zip codes (containing 15,657 addresses) with fewer than 120
addresses were excluded from the frame.
The total number of active locatable addresses excluded from the frame (i.e. drop points without
drop unit designators and addresses in zip codes below the minimum size criterion) represents
1.1% of active locatable addresses in the four states.
The number of mailing addresses on the SCV frame compares favorably with the estimated
number of occupied housing units in these four states which was 16,423,5525 in 2010. Table 9-1
compares the number of addresses on the SCV frame to the 2010 Census estimate for the number
of occupied housing units for each of the four states in the SCV. It also provides the estimated
English-speaking adult household population and the number of zip codes on the frame for each
state.
Matching Telephone Numbers to Addresses. Selected addresses will be matched to land line and
cell telephone number databases maintained by commercial vendors. Matching addresses to cell
number databases will enable us to increase telephone coverage of cell-phone-only households6.
A two-tiered approach will be used to attain the highest rate of correct telephone number
matches. First, the sample will be batched through a telephone number and name appending
process. By appending name and telephone number, an additional identifier will be developed to
3

The two types of active, locatable supplemental addresses contained on the No-Stat file are locatable city-style
addresses for P.O. Box throwbacks on rural and highway contract carrier routes and locatable city-style addresses
including unit type and number for approximately 20 percent of the units within drop points.
4
This assumes that all drop units on the No-Stat file correspond to drop units on the CDS file.
5
Source: 2010 Census.
6
Blumberg and Luke (2011) report that persons living in over 45 percent of U.S. households have little or no chance
of being contacted for a landline telephone survey: http://www.cdc.gov/nchs/nhis.htm.

ensure that the end results are active telephone lines. This batch process will match the sampled
address with the occupants currently believed to be associated with that address.

Table 9-1. SCV State Summary
State
North Carolina
Ohio
Pennsylvania
Virginia

English-Speaking
Household
Population 18+1
6,372,180
8,180,115
8,886,405
5,362,377

Occupied
Housing Units2
3,745,155
4,603,435
5,018,904
3,056,058

Total
Addresses on
Frame3
3,801,620
4,694,717
4,976,512
3,094,765

Number of 5-Digit
ZIP codes on Frame
711
988
1,277
761

Total
28,801,077
16,423,552
16,567,614
3,737
2009 American Community Survey 1-Year Estimates. (B16004, B15001, B07013, B09001)
2
2010 Census (H1)
3
Addresses consist of active, complete, locatable mailing addresses in zip codes above the minimum size criterion
1

The second tier of the matching process will use the names, telephone numbers, and addresses to
conduct a comprehensive telephone search. We will obtain up to three verified telephone
numbers for the provided address (including cell phone numbers) as well as the date of the most
recent association of that number and person with the sampled address. Using these data, we can
determine the most likely current occupants.
Based on previous research, we expect to associate a name and telephone number with between
50 and 60 percent of the sampled addresses. As part of the experiment, match rates will be
monitored as well as the proportion of incorrect or nonworking numbers.