Redesign Of The Medicare Current Beneficiary Survey Sample

REDESIGN OF THE MEDICARE CURRENT BENEFICIARY SURVEY SAMPLE
Annie Lo, Adam Chu, and Richard Apodaca, Westat
Annie Lo, Westat, 1650 Research Boulevard, Rockville, Maryland 20850
Key Words:

Multistage
probability
sample,
resampling methods, Ernst method for
overlap
maximization,
longitudinal
sample, rotating panels, Medicare
population

1.

Introduction
The Medicare Current Beneficiary Survey (MCBS)
conducted by the Centers for Medicare and Medicaid
Services (CMS), Department of Health and Human
Services, is a continuous sample survey of Medicare
beneficiaries residing in the United States and Puerto Rico.
The MCBS collects data on access to health care, health
status, source of care, health care utilization and costs,
satisfaction with health care, and other health-related topics
(e.g., see Sharma, Chan, Liu, and Ginsberg, 2001). A
representative sample of Medicare beneficiaries (referred to
as a “panel”) is selected for the MCBS each year using a
stratified multistage probability sample design. The sample
of first-stage or primary sampling units (PSUs), which
includes MSAs (metropolitan statistical areas) and groups
of rural (nonMSA) counties, was designed and selected in
1991. Although new beneficiary samples are selected each
year to supplement the original sample, the new samples
are always selected from the same PSUs. Over time, the
continued use of the original PSU sample has resulted in
losses in both sampling precision and operational
efficiency. In 2000, based on an evaluation of the existing
PSU sample, a decision was made to reselect the PSUs.
This paper summarizes some of the analyses leading to that
decision and describes the procedures used to update and
select the new MCBS PSU sample.
2.

The MCBS Sample Design
The MCBS employs a stratified multis tage
probability sample with three stages of selection. The first
stage involved the selection of PSUs consisting of MSAs
and groups of rural counties. The PSUs were selected with
probabilities proportionate to 1980 population within strata
defined by Census region, metropolitan status, and selected
PSU-level socio-economic characteristics. Two PSUs were
selected per stratum. The second sampling stage consisted
of the selection of ZIP Code areas within each sampled
PSU. To facilitate linking with available county-level data,
the second-stage sampling unit was defined to be the part
of the ZIP Code area that was physically contained within a
given county. In other words, ZIP Code areas that crossed
county borders were subdivided by county into separate
units called “ZIP fragments.” For sampling purposes, small

ZIP fragments were combined into clusters where
necessary to ensure that each ZIP cluster would provide a
reasonable workload for interviewers if selected for the
sample. At the third and final stage of selection,
beneficiaries within the sampled ZIP clusters were
stratified by age and subsampled at rates designed to yield
self-weighting (equal probability) samples of beneficiaries
within each of seven age groups. Additional details about
the original MCBS sample design are provided in Apodaca,
Judkins, Lo, and Skellan (1992).
The MCBS was originally intended to be a true
longitudinal survey in which sampled Medicare
beneficiaries would be interviewed three times a year
throughout the remainder of their lives. However, after two
years of data collection, it became clear that this would be
impractical. Thus, a decision was made to switch from a
fixed panel design to a rotating panel design in which
roughly one-third of the existing sample (i.e., the oldest
panel) is retired each year, and a new panel is selected to
replace it. Under this design, beneficiaries in each newly
selected panel are interviewed three times a year for a
maximum of four years. Table 1 illustrates the basic
features of the rotating panel design developed for the
MCBS. Additional details are given in Westat (2001).
Table 1. Panel rotation scheme for the MCBS*
Panel year
1994
1995
1996
1997
1998
1999
2000
2001

1998
A4
B3
C2
D1
––
––
––
––

Data collection year
1999
2000
2001
––
––
––
B4
––
––
C3
C4
––
D2
D3
D4
E1
E2
E3
––
F1
F2
––
––
G1
––
––
–

2002
––
––
––
––
E4
F3
G2
H1

*Panel year refers to the year of the fall round in which the panel is
introduced into the study. Data collection year refers to subsequent data
collection rounds. The letters A, B, C, etc. are used to designate a
particular panel. The numeric values indicate the data collection year.
For example, C4 refers to the fourth year of data collection for the 1996
panel.

Over 15,000 beneficiaries were selected for the
initial round of the MCBS. In each of the following two
years, supplemental samples of about 2,400 beneficiaries
per year were added to the original sample to compensate
for sample attrition and to give coverage to newly enrolled
Medicare beneficiaries. With the implementation of the
rotating panel design in 1994, the number of beneficiaries

selected for each annual supplement (i.e., nationally
representative panel) has been between 6,300 to 6,400
beneficiaries per year.
3.

Considerations for Updating MCBS PSUs
The main reasons for updating and reselecting an
existing PSU sample are: (a) to improve sampling precision
through the use of more up-to-date sampling measures of
size and stratification schemes, and (b) to maintain more
balanced sample workloads across PSUs. Kish (1965, page
482) also cites the need to avoid “inertia in continuing
operations” that can hinder improvements of outdated and
inefficient procedures. The Current Population Survey
(CPS), for example, has traditionally updated its PSU
samp le at 10-year intervals using data from the most recent
decennial Census to redesign the sample (U.S. Department
of Labor, 2000).
For the MCBS, there was evidence that design
effects (defined to be ratio of the variance of an estimate
derived from the MCBS to the corresponding variance
based on a simple random sample of the same size) were
increasing over time. For a number of statistics examined
in Westat (2000), the median design effect for the total
beneficiary sample increased by an average of four percent
annually between 1992 and 1995. By age group, the
average increase in median design effects varied from less
than two percent for older beneficiaries (75 years or older)
to around three to six percent annually for younger
beneficiaries (74 years or younger). Although the estimated
design effects fluctuated widely from year to year, the
overall patterns did suggest that design effects had
generally increased between 1992 and 1995. By 1996,
however, there was a noticeable drop in design effects,
probably due to the fact that over two-thirds of the original
MCBS panel had been phased out of the study by this time.
While design effects after 1996 did not increase as greatly
as in previous years, there did appear to be modest
increases for some subgroups. The results shown in Table 2
for selected health related variables illustrate the magnitude
and variability of the change in design effects over time.
Table 2. MCBS design effects, 1992-1999

Survey
year
1992
1993
1994
1995
1996
1997
1998
1999

None
2.00
1.78
1.91
1.75
1.30
1.33
1.30
1.84

Characteristic
Functional limitation*
IADL
1 or 2
3 to 5
only
ADLs
ADLs
1.29
1.28
1.44
1.46
1.45
1.07
1.34
1.52
0.55
1.38
1.04
1.06
1.38
1.29
1.36
1.29
1.17
1.45
1.41
1.10
1.08
1.35
1.97
1.00

Hypertension
1.55
1.52
1.61
1.50
1.71
1.82
1.86
NA

*Instrumental activities of daily living (IADLs). Activities of daily living
(ADLs).

It should be noted that the design effects
summarized in Table 2 reflect the increase in variance

arising from a variety of sources. The continued use of an
increasingly inefficient PSU sampling measure of size can
lead to both increased clustering effects as well as
increased variation in sampling weights. In Section 3.1, the
variance of an estimate based on the MCBS design is
decomposed to show how various design features affect
overall sampling precision.
In addition to increased variances, the use of an old
PSU sample can lead to a less efficient distribution of
workload across PSUs. This occurs because the measure of
size used to select the original PSU sample may no longer
adequately control PSU workloads (sample sizes). For the
MCBS, the original PSU measure of size was 1980
population. Not only was the source of data 10 years older
than the PSU sample, the measure of size assigned to PSUs
reflected total U.S. population rather than the Medicare
population. Updating the PSU measure of size with current
counts of Medicare beneficiaries, therefore, was expected
to ameliorate the worsening imbalance in the PSU
workloads. Implications of the aging PSU sample on
survey operations and costs are discussed further in
Section 3.2.
3.1

Precision of Estimates
As is the case with virtually all sample surveys,
approximately unbiased estimates of totals derived from the
MCBS are weighted sums of the form:

yˆ =

P np

∑ ∑ w pi y pi ,

(1)

p =1 i =1

where y pi is the observed value of the characteristic being
estimated for the i-th sample beneficiary in panel p, and
wpi is the corresponding sampling weight. Annual cost
and use estimates are typically based on three complete
(continuing) panels, while access-to-care estimates are
based on four panels. The panel sample sizes, np , vary
slightly with newer panels being somewhat larger than
older ones due to attrition. The sampling weights, wpi ,
reflect the beneficiaries’ overall probabilities of selection,
and include adjustments for nonresponse and
undercoverage. Additional details about the weighting
procedures employed in the MCBS are given in Judkins
and Lo (1993).
To bring out important features of the sample design
employed for the MCBS, it is useful to express the
estimated total given by equation (1) in the following
alternative form:

yˆ =

G

G

P

g =1

g =1 p =1

∑ yˆg = ∑ ∑ a gp yˆgp ,

(2)

where
ngp

NR
yˆ gp = ∑ wgpi
y gpi
i =1

(3)

is the estimated total for the g-th “combination group” in
panel p. A combination group is a subset of the beneficiary
sample within which the individual panel-specific estimates
are “composited” to form an overall combined estimate.
NR
Note that the weights, wgpi
, in equation (3) are panel-

specific nonresponse-adjusted weights that inflate the
results for panel p to population levels; thus,

yˆ g =

P

∑ agp yˆgp is a composite estimate of the population
p =1

total for the g-th combination group based on P panels. The
combination groups used to construct the MCBS estimates
are defined in terms of age group and initial year of
Medicare eligibility (also referred to as “accretion” status).
For combination group g, the a gp ' s in equation (2) are
generally proportional to the panel sample sizes, ngp and
are subject to the condition that a g1 + ag 2 + K + agP = 1 .
From equation (2), the variance of the estimated
total can be written as:

var ( yˆ ) =

G

P

∑ ∑ a2gp var ( yˆgp ) + D ,

(4)

g =1 p =1

to a redistribution of the ZIP cluster sample sizes within
PSUs.
As mentioned earlier, the design effect provides a
rough measure of the relative precision of the MCBS
samp le design with respect to a simple random sample of
the same size. For example, the design effect (DEFF) for
an estimated total is defined as
2
DEFF = var ( yˆ ) σ SRS
,

(6)

2
where var ( yˆ ) is given by equation (4) and σ SRS
is the
hypothetical variance that would have been obtained from a
simple random sample of the same size. In general, it is
difficult to disentangle the various sources of variance
contributing to the design effect. In particular, MCBS
estimates are subject to both clustering and unequal
weighting design effects. For estimates of means and
proportions, an approximation that is useful for separating
out the different effects is given by:

{

DEFF = 1 +  cv ( wi ) 

2

}{1 + ( n −1) ρ}
*

,

(7)

where ρ is the intraclass correlation between beneficiaries
within PSUs, cv ( wi ) is the coefficient of variation of the

{

}

where D represents the total covariance between pairs of
panel estimates, yˆ gp and yˆ g ′p′ . Although D cannot be

sampling weights, n * = n 1 +  cv ( ni ) 

assumed to be zero, it is expected to account for a relatively
small part of the total variance. Moreover, the variance of
the panel estimates can be written as:

PSU sample size adjusted for varying cluster sizes, n is

( )

var yˆ gp =

2
M 2 Bgp

m NC

+

2
N 2Wgp

mn

,

(5)

where mNC = the number of noncertainty PSUs in the
sample, m = the total number of PSUs in the sample,
M = the number of PSUs in the population, n = the average
number of sample beneficiaries per sample PSU, and N =
2
the number of beneficiaries in the population. Bgp
and
2
Wgp
are unit variances associated with the different stages
2
of selection; Bgp
is the “between PSU” unit variance and

is a function of the PSU selection probabilities, while

2
Wgp

is an average “within PSU” variance that reflects all stages
of selection within the PSU (Hansen, Hurwitz, and Madow,
1953).
Although equation (5) is an oversimplification, it
does serve to point out that both between-PSU and withinPSU components can change as the PSU sample ages.
2
Since Bgp
is a function of the original PSU selection

probabilities (e.g., see Hansen, et al., 1953, page 397), it
can increase if the distribution of Medicare beneficiaries
within PSUs changes dramatically over time. Similarly,
2
these same changes can lead to inflated values of Wgp
due

2

is the average

the average PSU sample size, and cv ( ni ) is the coefficient
of variation of the PSU sample sizes (United Nations,
1993). Note that formula (7) can be written as

DEFF = Dw Dc where Dw = 1 +  cv ( wi ) 

(

2

is the unequal

)

weighting design effect and Dc = 1 + n* − 1 ρ

is the

clustering effect.
Design effects were computed for selected
characteristics derived from the 1997 Access to Care data
file (see Sharma, et al., 2001). Fay’s modification of the
balanced repeated replication (BRR) technique was used to
compute the requisite standard errors (Judkins, 1990).
Using equation (7) with n = 149 and cv ( ni ) = 0.29 , the
calculated design effects were then used to estimate the
intraclass correlation. As shown in Table 3, the intraclass
correlations range from less than 0.005 to 0.02 for the items
considered. The unequal weighting design effect for the
statistics in Table 3 was estimated to be Dw = 1.22 . Thus,
on average, the ratio of the clustering design effect to the
unequal weighting design effect ranged from 1 to over 3.
Finally, the speculated gains in precision that could
be achieved with a new PSU sample are summarized in
Table 4. The design effects were calculated using equation
(7) for a range of values of cv ( ni ) . A value of cv ( ni ) = 0
corresponds to the situation where the new PSU measure of
size has controlled the PSU workloads perfectly. While this
is highly unlikely in view of the panel rotation employed in

the MCBS, it does provide lower bounds on the design
effects that can be achieved. On the other hand, values of
cv ( ni ) in the range of 0.10 to 0.20 are more realistic. In
this case, the reduction in DEFFs (as compared with the
DEFFs in Table 3) would range from two to five percent.
Although the reductions are modest, the introduction of
new PSUs is expected to improve sampling precision.
Table 3. Intraclass correlations and design effects
Characteristic
Poor health status
Hypertension
Difficulty bathing
Difficulty walking
Limited activity
Medicaid
Risk HMO
High school graduate
Married
Income <$25,000

ρ
0.005
0.002
0.007
0.013
0.003
0.010
0.022
0.008
0.003
0.007

DEFF
2.13
1.52
2.53
3.80
1.87
3.09
5.45
2.87
1.76
2.52

Dc
1.74
1.24
2.08
3.12
1.54
2.53
4.47
2.35
1.44
2.07

Dc Dw
1.43
1.02
1.70
2.55
1.26
2.07
3.66
1.93
1.18
1.70

Table 4. Speculated design effects with new PSU sample
cv ( ni )

Characteristic
Poor health status
Hypertension
Difficulty bathing
Difficulty walking
Limited activity
Medicaid
Risk HMO
High school graduate
Married
Income <$25,000

3.2

0.00
2.06
1.49
2.43
3.60
1.82
2.94
5.12
2.74
1.72
2.42

0.10
2.06
1.50
2.44
3.62
1.83
2.96
5.16
2.76
1.72
2.43

0.20
2.09
1.51
2.48
3.70
1.85
3.01
5.28
2.81
1.74
2.47

maximize efficiency. Over time, the cost of adjusting to the
relative change in workloads can be absorbed into the
yearly hiring and training process. The absolute costs
associated with these sample disbursement changes are
difficult to measure because they are offset by the overall
efficiency of the data collection system. It is easy to see,
however, that as new ZIP Codes are added to the sample,
interviewers must travel longer distances to reach
unclustered areas. Moreover, the workloads in existing ZIP
fragments can become uneven. The longer the PSU sample
remains in place, the more dispersed and inefficient the
sample becomes. While moving to a new set of PSUs will
not totally solve the dispersion problem, it will serve to
lessen its impact and help maintain desired levels of
operational efficiency.
4.

Selection of the New PSU Sample
Based on considerations summarized in Section 3, a
decision was made to reselect the sample of PSUs. In order
to retain as much of the existing field operations as
possible, the PSUs were selected using procedures
designed to maximize overlap with the existing MCBS
PSUs.
4.1

0.30
2.13
1.52
2.54
3.82
1.88
3.10
5.47
2.88
1.76
2.53

Cost and Operational Implications
There are two factors leading to increased costs
associated with remaining in the original PSUs: (a) greater
dispersion of the sample within PSUs, and (b) increasingly
unequal workloads across PSUs. The selection of new ZIP
fragments each year to represent newly-created ZIP Codes
tends to disperse the sample within PSUs, thereby
increasing travel costs and data collection time. The
supplemental sample (i.e., new panel) selected each year
also changes the relative sample sizes between the PSUs,
thereby changing the individual PSU workloads. This
results in additional travel and hiring costs to accommodate
the changing workloads within the PSUs.
A comparison of the 1999 panel with the initial
MCBS sample showed that if the 1999 panel were
expanded to the size of the initial sample, the sample sizes
for each PSU would fall between 89 percent and 192
percent of the original sample. Since the 1999 sample was
selected using the original MCBS PSU measure of size, it
provides a good indication of how the PSU workloads can
fluctuate over time. Such changes would necessitate a very
different staffing configuration than the original sample to

Definition of PSUs
Experience has shown that the types of PSUs
defined for the MCBS and many other national in-person
surveys (i.e., PSUs consisting of metropolitan areas or
groups of rural counties) are generally robust and efficient
for the purpose of maximizing sampling precision and
minimizing survey costs. For this reason, the same PSU
definitions developed for the original MCBS sample were
maintained whenever possible.
In the nonMSA areas, a PSU was defined to be a
single county unless it was too small to provide an
adequate workload for an interviewer. In such cases, the
county was combined with an adjacent county or counties
to form the PSU. Each nonMSA PSU was designed to have
a minimum measure of size of roughly 3,100 Medicare
beneficiaries.
4.2

Certainty PSUs
For the redesign, those PSUs with at least 224,000
Medicare beneficiaries were included in the sample with
certainty. (For cost reasons, Alaska and Hawaii, which
together account for 0.6 percent of all Medicare
beneficiaries, were not included in the sampling process.)
The cutoff of 224,000 corresponds roughly to a probability
of selection of 75 percent under a probabilityproportionate-to-size (PPS) sample design. The use of the
specified cutoff resulted in designating the 28 largest PSUs
in the United States as certainties. Of these, 27 were also
certainties in the original MCBS design. In addition, the
largest MSA in Puerto Rico was included in the sample
with certainty.

4.3

Noncertainty PSUs
The remaining (noncertainty) PSUs were grouped by
Census region and MSA status (where Puerto Rico was
treated as a separate “region” for sampling purposes).
Within these major groups of PSUs, detailed sampling
strata were formed by sorting PSUs by the percent of
Medicare beneficiaries enrolled in HMO plans (and in
some cases also by the percentage of minority
beneficiaries), and then forming strata of roughly equal size
from this sorted list. The measure of size (MOS) assigned
to a PSU was the weighted sum of the number of Medicare
beneficiaries in the PSU in seven age groups, where the
“weights” used to calculate the MOS were proportional to
the corresponding overall target sampling rates. The use of
the weighted measure of size was designed to obtain selfweighting samples of beneficiaries within each of the seven
age groups, while at the same time maintaining a roughly
constant PSU sample size (workload) across all
noncertainty PSUs (e.g., see Folsom, Potter, and Williams,
1987). Thirty-eight noncertainty strata were formed within
the continental United States and one was formed in Puerto
Rico. Two PSUs were then selected with probabilities
proportionate to size from each stratum using procedures
designed to maximize the overlap with the existing MCBS
sample.

Fh were determined by maximizing the unconditional
expected overlap defined by
L Rh A

∑ ∑ ∑ chij xhij ,
subject to the constraints that
L Rh

∑∑ xhij = π j ,

∑ xhij = phi yh, h = 1,K , L and i = 1,K , Rh ,
j=1
L

∑ yh = 1,
h=1

where chij is the conditional expected sample overlap
and yh is the probability of selecting the h-th original
MCBS stratum.
Step 3—One of the original MCBS strata was selected
with probability yh determined as part of the solution

n

of the LP problem.
Step 4—Finally, a new sample of PSUs was selected
from the given intersection using conditional
probabilities derived from the LP procedure.

4.4

n

( F1 , F2, K , FL )

Step 1—All intersections

between a

given “new” stratum and the “old” MCBS strata were
identified and labeled. In each Fh, all possible old
samples were listed
selection probabilities

(s

o
hi , i

( phi )

= 1,2, K , Rh

)

and the

samples were listed

(

s nj ,

j = 1,2,K , A

corresponding selection probabilities
n

4.5

Results and Workload Implications
Overall, 63 of the 107 original MCBS PSUs were
retained for the new sample. The achieved overlap of 59
percent was consistent with preliminary estimates (Westat,
2000). Table 5 summarizes the results of the PSU sampling
process. Also shown are estimates of the expected relative
workload per PSU in the four years after the selection of
the new PSU sample. The workload estimates in the table
are “typical” workloads reflecting the PSU workload
associated with four active panels. Due to sample attrition,
the panels are not equal in size. Older panels are generally
smaller in size than newer ones. The percentages shown are
rough estimates intended to reflect the different sample size
losses in the component panels over time.
Table 5. Distribution of MCBS PSU sample by selection
status and approximate workload

were computed for each

o
shi
. Similarly, within a new stratum, all possible new

)

and the

(π j )

were

computed.
Step 2—Next, the optimal joint selection probabilities

( xhij ) of each combination of new and old samples in

j = 1,2, K , A,

h=1 i =1
A

n

The Ernst Algorithm
To maximize overlap with the existing PSUs, the
method developed by Ernst (1986) was used to select the
new sample of noncertainty PSUs. In the Ernst approach,
each stratum in the new design is treated as a separate
linear programming (LP) problem where the objective is to
maximize the unconditional overlap subject to certain
constraints involving relevant selection probabilities. The
results of the optimization process are then used to select
the new sample. A very superficial summary of the Ernst
algorithm is given below. Readers interested in the
mathematical details are referred to the excellent paper by
Ernst (1986).

(8)

h=1 i=1 j =1

1.
2.
3.
4.
5.
6.

No.
Relative workload (%)
Status of PSU *
PSUs Year 1 Year 2 Year 3 Year 4
C in both samples
28
100
100
100
100
C in new sample, NC in
1
100
100
100
100
old sample
NC in new sample C in
1
100
100
100
100
old sample
NC in both samples
33
100
100
100
100
In new sample but not
44
30
57
80
100
in old sample
In old sample but not in 44
70
43
20
0
new sample

*C = “certainty”; NC = “noncertainty”.

Under the rotating panel design, the MCBS will be
operational in 151 PSUs until the beneficiary samples in
the old PSUs are completely phased out of the study. For
the 44 original PSUs that are not included in the new
sample, there will be no new (supplemental) samples of
beneficiaries. However, beneficiaries in the three most
recent panels in these PSUs will continue to be interviewed
for up to three years. The workload in these PSUs will start
out at roughly 70 percent of the desired workload because
there will be no supplemental sample to replace the oldest
panel. In the ensuing two years, the workload will dwindle
further to roughly 43 percent and 20 percent of the
maximum workloads, respectively, as the older panels are
released from the study.
At the same time, the workload in the 44 newly
selected PSUs will start out at a reduced level of
approximately 30 percent since they will include only the
newest panel. However, with the introduction of new
panels in each of the following three years, the workload
will increase to 57 percent in the second year, 80 percent in
the third year, and eventually to full capacity in the fourth
year of operation. For the 63 PSUs that are included in both
the new and original samples, the workload will be
maintained at the desired 100 percent level since the annual
supplement will replace the panel that is scheduled to be
released under the rotating panel design.
5.

Summary
The MCBS PSU sample underwent a redesign 10
years after it was first introduced in 1991 for a number of
reasons. Between 1996 and 1999, design effects for the
total beneficiary sample increased by an average of three
percent annually. It was anticipated that if the same PSUs
remained in place, further deterioration of sampling
precision would occur. Continued sampling from the
existing PSUs also led to unbalanced PSU workloads and a
more dispersed sample within PSUs.
In order to maximize overlap with the existing
MCBS PSUs, the Ernst optimization algorithm was used to
select the new PSU sample. An overall 59 percent overlap
between the old and new MCBS PSUs was attained. The
overlap of 46 percent among the noncertainty PSUs
compared favorably with the 25 percent overlap expected
with independent sampling. With the redesign, the PSU
measure of size was updated with current counts of
Medicare beneficiaries. Stratification of the noncertainty
PSUs was also enhanced through the use of relevant
information on the Medicare population. The redesign will
have little or no impact on existing weighting and
imputation procedures. Comparing the response rates for
the initial fall interview for the 2000 panel in old PSUs
with those for the 2001 panel in new PSUs, the overall
response rate was slightly higher in the new PSUs (87.7%)
than the old PSUs (86.7%).

6.
References
Apodaca, R., Judkins, D., Lo, A., Skellan, K. (1992).
Sampling from HCFA lists. Proceedings of the Section
on Survey Research Methods of the American Statistical
Association, pp. 250-255.
Ernst, L. (1986). Maximizing the overlap between surveys
when information is incomplete. European Journal of
Operational Research, 27, pp. 192-200.
Folsom, R., Potter, F., Williams, S. (1987). Notes on a
composite size measure for self-weighting samples in
multiple domains. Proceedings of the Section on Survey
Research Methods of the American Statistical
Association, pp. 792-796.
Hansen, M., Hurwitz, W., Madow, W. (1953). Sample
Survey Methods and Theory, New York: John Wiley &
Sons.
Judkins, D. (1990). Fay’s method for variance estimation.
Journal of Official Statistics, 3, pp. 223-240.
Judkins, D. and Lo, A. (1993). Components of variance and
nonresponse adjustment procedures for the Medicare
Current Beneficiary Survey. Proceedings of the Section
on Survey Research Methods of the American Statistical
Association, pp. 820-825.
Kish, L. (1965). Survey Sampling, New York: John Wiley
& Sons.
Sharma, R., Chan, S., Liu, H., Ginsberg, C. (2001). Health
and Health Care of the Medicare Population. Data from
the 1997 Medicare Current Beneficiary Survey.
Rockville, MD. Westat.
United Nations (1993). Sampling Errors in Household
Surveys. Department for Economic and Social
Information and Policy Analysis. Statistical Division,
National Household Survey Capability Programme.
U.S. Department of Labor, Bureau of Labor Statistics and
U.S. Department of Commerce, Economics and Statistics
Administration, U.S. Bureau of Census (2000).
Technical Paper 63, Current Population Survey, Design
and Methodology, Washington, DC.
Westat (2000). Medicare Current Beneficiary Survey:
Evaluation of Alternative Measures of Size for Primary
Sampling Unit Selection. Technical Report prepared for
Health Care Financing Administration, Rockville, MD.
Westat (2001). Medicare Current Beneficiary Survey:
Design and Selection of the 2001 PSU Sample. Technical
Report prepared for Health Care Financing
Administration, Rockville, MD.