Methodology Report #26:
|
||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||
|
_^top Abstract
A series of public use files (PUFs) are released from the Medical Expenditure
Panel Survey (MEPS) Household Component every year. The most commonly used files
are at the person level (with household and family identifiers) but some files
are at lower levels such as medical conditions and medical events files. Eight
event-level files are published from MEPS each year. For each person in the MEPS
sample, these event files may include no record, a single record, or multiple
records depending on the number of events the person had during the year.
Therefore, an event file only includes records related to a subset of persons in
the full year person file. Since the number of persons in an event file is not
known before conducting the survey, any analysis of estimates from event files
should be treated as a domain analysis which requires the entire sample to take
all variability into account in estimating the variance of domain estimates. That
is, the analysis should ideally include all persons with and without events in
the file. However in practice, since it is convenient to deal with the subset of
cases with events only, users generally compute event level estimates without
merging persons with no event from the full person-level file. The impact of not
doing a domain analysis is usually negligible if the subset (the domain) is large compared to the full file. This report looks into the issue of variance estimation from the MEPS event files and evaluates the impact of not doing a proper domain analysis on variance estimates. Suggested Citation Chowdhury, S., Machlin, S. Variance Estimation from MEPS Event Files, Methodology Report No. 26. September 2011. Agency for Healthcare Research and Quality, Rockville, MD. http://www.meps.ahrq.gov/mepsweb/data_files/publications/mr26/mr26.pdf
* * * The estimates in this report are based on the most recent data available at the time the report was written. However, selected elements of MEPS data may be revised on the basis of additional analyses, which could result in slightly different estimates from those shown here. Please check the MEPS Web site for the most current file releases. Center for Financing, Access, and Cost Trends _^top Background
The Medical Expenditure Panel Survey (MEPS) Household Component The Medical Expenditure Panel Survey (MEPS) provides nationally representative
estimates of health care use, expenditures, sources of payment, and health
insurance coverage for the U.S. civilian noninstitutionalized population. The
MEPS Household Component (HC) also provides estimates of respondents' health
status, demographic and socio-economic characteristics, employment, access to
care, and satisfaction with health care. Estimates can be produced for
individuals, families, and selected population subgroups. The panel design of
the survey, which includes five rounds of interviews covering two full calendar
years, provides data for examining person level changes in selected variables
such as expenditures, health insurance coverage, and health status. Using
computer assisted personal interviewing (CAPI) technology, information about
each household member is collected, and the survey builds on this information
from interview to interview. All data for a sampled household are reported by a
single household respondent. The MEPS-HC was initiated in 1996. Each year a new panel of sample households is
selected. Because the data collected are comparable to those from earlier
medical expenditure surveys conducted in 1977 and 1987, it is possible to
analyze long-term trends. Each annual MEPS-HC sample size is about 15,000
households. Data can be analyzed at either the person or event level. Data must
be weighted to produce national estimates. The set of households selected for each panel of the MEPS-HC is a subsample of
households participating in the previous year's National Health Interview
Survey (NHIS) conducted by the National Center for Health Statistics of the
Centers for Disease Control and Prevention. The NHIS sampling frame provides a
nationally representative sample of the U.S. civilian noninstitutionalized
population and reflects an oversample of blacks, Hispanics and, starting in
2006, Asians. MEPS oversamples additional policy relevant subgroups such as low
income households. The linkage of the MEPS to the previous year's NHIS provides
additional data for longitudinal analytic purposes. Medical Provider Component Upon completion of the household CAPI interview and obtaining permission from the household survey respondents, a
sample of medical providers are contacted by telephone to obtain information
that household respondents cannot accurately provide. This part of the MEPS is
called the Medical Provider Component (MPC) and information is collected on
dates of visit, diagnosis and procedure codes, charges, and payments. The
Pharmacy Component (PC), a subcomponent of the MPC, does not collect charges or
diagnosis and procedure codes but does collect drug detail information,
including National Drug Code (NDC) and medicine name, as well as date(s) prescriptions
are filled, sources and amounts of payment. The MPC is not designed to yield
national estimates. It is primarily used as an imputation source to supplement
and/or replace household reported expenditure information. Survey Management MEPS-HC and MPC data are
collected under the authority of the Public Health Service Act. Data are
collected under contract with Westat. Data sets and summary statistics are
edited and published in accordance with the confidentiality provisions of the
Public Health Service Act and the Privacy Act. The National Center for Health Statistics (NCHS) of the Centers for Disease Control and Prevention provides
consultation and technical assistance related to the selection of the MEPS
household sample. As soon as data collection and
editing are completed, the MEPS survey data are released to the public in
staged releases of summary reports, micro data files, and tables via the MEPS
Web site: http://www.meps.ahrq.gov. Selected data can be analyzed through
MEPSnet, an on-line interactive tool designed to give data users the capability
to statistically analyze MEPS data in a menu-driven environment. Additional information on MEPS is available from the MEPS
project manager or the MEPS public use data manager at the Center for Financing,
Access, and Cost Trends, Agency for Healthcare Research and Quality, 540 Gaither Road, Rockville, MD 20850; 301-427-1406, or e-mail mepspd@ahrq.gov. Be sure to specify the AHRQ number of the document or CD-ROM
you are requesting. Selected electronic files are available through the
Internet on the MEPS Web site: http://www.meps.ahrq.gov/ _^top Introduction
The Medical Expenditure Panel Survey (MEPS) is a set of
large-scale surveys of families and individuals, their medical providers
(doctors, hospitals, pharmacies, etc.), and employers across the United States.
MEPS provides estimates of specific health services use by the U.S. civilian
noninstitutionalized population, the payments for these services, sources of
payment, and the cost and scope of health insurance of U.S. workers. MEPS has
three components: the Household Component (HC), Medical Provider Component
(MPC) and the Insurance Component (IC). The Household Component collects data
from individual households and their members in selected communities across the
United States, drawn from a nationally representative subsample of households
that participated in the prior year's National Health Interview Survey
(conducted by the National Center for Health Statistics). The data collected
from households are supplemented by data from their medical providers collected
in the MPC. The Insurance Component is a separate survey of employers that
provides data on employer-based health insurance. The MEPS Household Component (which will be generally
referred to as MEPS hereafter) collects detailed information for each person in
the household on demographic characteristics, health conditions, health status,
use of medical services, charges and source of payments, access to care,
satisfaction with health care, health insurance coverage, income, and
employment. The panel design of the survey, which features five rounds of
interviewing covering two full calendar years, makes it possible to determine
how changes in individuals’ health status, income, employment, eligibility for
public and private insurance coverage, use of services, and payment for care
are related. MEPS data are available on the MEPS Web site in data tables,
downloadable data files (person, job, event, or condition level) and
interactive data tools, as well as in publications using HC data. The main
public use files (PUFs) released from MEPS are the full-year (FY) consolidated
file and related medical conditions and medical event level files. The FY file
includes records at the person level with family and dwelling unit (DU)
identifiers and provides individual level information on health status,
socio-demographic characteristics, employment, insurance, access to care, and
various other related data items. The conditions file provides detailed
information about each heath condition reported by the household respondent. It
includes records at the person/condition level. Each event file consists of a
specific type of event for persons in the corresponding full-year consolidated
file. In the conditions file or an event file, some persons may have multiple
conditions/events within the year and thus have multiple records and other
individuals may have no medical conditions or events and have no records on the
particular file. The conditions file or an event file includes identifiers to
link each event to the individual in the person-level file who had the
condition or the event. In analyzing MEPS event (or conditions) files, estimates are
often produced from the event file without merging all person records from the
full file (i.e., only using the subset of persons with events). For point
estimates, since the persons with no events will have no contribution to event
related estimates, there is no need to keep those persons in the file. However,
for variance estimation, since the persons with events (or conditions) are a
subset of all persons, the subset total in the population is not known and the
sample size is random, the theoretically correct approach is to include all
persons (with or without an event) in the file and variance estimates are
produced by treating the persons with and without events as separate domains.
This approach to analysis is called domain analysis. Domain analysis takes the
variability into account by using the entire sample in estimating the variance
of subgroup estimates. For more information about domain analysis, see Lohr
(1999), Cochran (1977), and Fuller et al. (1989). The ‘domain’ statement in SAS
survey procedures, ‘subgroup’ statement in SUDAAN, and similar features in
other variance computation software are designed to correctly estimate
variances in such situations. If the estimates are produced using the records
in the event file only i.e., using the subset of persons with events from the
FY file, the variances may not be estimated correctly. The extent of deviation
of the estimated variance from the correct estimate for this technically
improper analysis depends on the size of the subset compared to the full file.
If the size of the subset is large and the number of variance strata with
singleton primary sampling units (PSUs) is small or zero, the impact on the
variance estimates will generally be negligible. This study examines this issue
of variance estimation from the MEPS event files, with particular focus on the
impact on variance estimates of analyzing the subset of cases with events only
without doing a domain analysis that incorporates persons with no reported
medical events. _^top MEPS Event Files
Eight public use event files from MEPS are published each
year: prescription medicines, dental visits, other medical expenses, hospital
inpatient stays, emergency room visits, outpatient department visits,
office-based medical provider visits, and home care. Each of these files is an
event-level file consisting of specific type of events for persons in the
corresponding full-year consolidated file. Table 1 summarizes the sizes of 2008
MEPS event files and brief descriptions of different event files are provided
below. In a prescription medicines event file, each record
represents a unique prescribed medicine event that is reported by the household
respondent as being purchased or otherwise obtained for a household member. The
file contains an identifier for each unique prescribed medicine and information
on the detailed characteristics associated with the event; selected Multum
Lexicon variables; conditions, if any, associated with the medicine; the date
on which the person first used the medicine; total expenditures and sources of
payments; and types of pharmacies that filled the household’s prescriptions. The dental visits event file contains variables
pertaining to household reported dental visits. The file includes the date of
the dental event, type of provider seen, if the visit was due to an accident,
reason for the dental event, whether or not medicines were prescribed,
expenditures, and sources of payment. The other medical supplies event file contains
information on the purchase of and expenditures for medical equipment,
supplies, glasses and other medical items purchased, and sources of payment.
Each record in this file contains information for the whole calendar year for
all items except glasses for which each record contains information for a data
collection round. The hospital inpatient stays event file contains
characteristics associated with the hospital inpatient stay event, such as the
date of the hospital inpatient stay, reason for the stay, types of services
received, condition(s) and procedure(s) associated with the hospital inpatient
stay, whether or not medicines were prescribed, expenditures, and sources of
payment. The emergency room visits event file contains
characteristics associated with the emergency room visit, such as the date of
the visit, types of care and services received, types of medicine prescribed
during the visit, condition codes, expenditures, and sources of payment. The outpatient visits event file contains
characteristics associated with the outpatient visit data, such as the date of
the visit, type of provider seen, type of care received, type of services provided,
expenditures, and sources of payment. The office-based medical provider visits event file
contains characteristics associated with the office-based visit, such as date
of the visit, type of provider seen, time spent with the provider, types of
treatment and services received, types of medicine prescribed, condition codes,
expenditures, and sources of payment. The home health event file can be used to make
estimates of the utilization and expenditures associated with home health care.
The file contains monthly information on expenditures for home health visits,
types of providers, types of services received, lengths of visits, reasons for
the visits, expenditures, and sources of payment. Each record in this file
represents a month of care. Moreover, the above files include various ID variables that
can be used to link events to individuals in the FY person-level file or to
other events or conditions in those files. _^top Domain Variance Estimation
Estimates from sample surveys like MEPS are often produced
for different subgroups or subpopulations into which the population can be
divided. For example, estimates may be required for groups with different types
of health insurance, or persons with and without a health condition or events,
or for a particular ethnic group. These subgroups are called domains or
subdomains of study. The interest may concentrate on a particular domain in
which the persons have certain characteristics or events e.g., in the analysis
of a particular type of event in MEPS. In that case, point estimates can be
produced from the domain but the full sample is required for variance
estimation. This is called domain analysis. If the file is subset to the domain of interest only there
will be no problem in producing the point estimates such as mean, percentage,
or total. However, the variances or standard errors of these point estimates
may be computed incorrectly from the subsetted file because the subset may not
contain the full sample design information or the share of the domain to the
full population to compute the variance correctly. This is not a problem if
the sample is selected separately from each domain and the domain size in the
population is known and the weighting adjustment is made independently within
each domain. When the sample is not selected independently within each domain
and the size is not fixed, the sample size becomes random in repeated draws.
Also, if the population total of the domain is not known and not benchmarked at
the domain level, the variance of the estimate of a total of a variable (say
total expense) not only depends on the variance of the mean expense per person,
but also the variance of the estimate of the total number of persons in the
domain. The full file with all domains is required to compute the variance of
the total or the proportion that belong to the domain. The estimate of the mean
in this case becomes the case of an estimate of a ratio because both the
numerator and the denominator of the mean are estimates and the variance of the
mean needs to be correctly estimated by treating it as a ratio. For means, this
complication can be avoided by assuming that the sample size for the domain is
fixed over repeated draws of the sample of same overall size. The problem is
more complex for computing the variance of an estimate of total. If a simple random sample of
size
and the population total,
if
This shows for the point estimation of the mean and total
for the domain, only the cases within the domain are required irrespective of
whether the domain total
where,
variance is calculated from the cases in the domain only, it
will not reflect the full variance of the estimate. In addition to the theoretical reasons, there are practical
reasons for keeping the full file and using the domain option for estimating
the variance of a domain estimate. When a complex cluster sample design is
used, the variance of a survey estimate is often computed based on variance
strata and clusters (PSUs) using the Taylor series approximation. This approach
needs at least two PSUs within each variance stratum to compute the variance by
accounting for the variance contributions from all strata. If the domain of
interest is small or clustered in certain areas so that some PSUs do not
include any case from the domain then some variance strata appear to have no
PSU or only one PSU when the file is subset to the domain. The situation of one
PSU within a stratum is known as the singleton PSU problem. In this case, it is
not possible for variance computation software to correctly compute the
variance from that stratum unless the full file is provided and a domain
analysis is requested. In the absence of the full file, different software
packages deal with a singleton PSU differently. SAS complex survey procedures
exclude the singleton PSUs from variance calculation, SUDAAN imputes a value
equal to the overall mean of all other PSUs for the missing PSU if the MISSUNIT
option is used, and STATA offers different options including the approach SAS
and SUDAAN use. Therefore, variance estimates from different software may not
be identical. When there are more than two PSUs in a stratum and the domain has
cases in at least two PSUs but not in all PSUs, none of the software can
account for the missing PSUs when the file is subset to the domain. This can
also lead to an under estimation of variances. Calculations of degrees of
freedom (df), design effect, hypothesis testing, etc. are also affected
by this. For example, to compute df, SUDAAN counts the number of PSUs
and strata with at least one observation from the domain. In contrast, SAS 9.1
Survey procedures compute df as the number of clusters (PSUs) in the
non-empty strata minus the number of non-empty strata after excluding the
singleton PSUs. When the df is not correctly computed it may affect the
confidence interval or hypothesis testing and the resulting inference,
particularly when the available df is small. Theoretically, the variance is generally
underestimated if the full file is not used. But the impact of a singleton PSU
may be positive or negative depending on the situation and the software.
However, overall the impact of using the subset and not the full file depends
on the size and clustering of the domain compared to the full population. As
the domain size gets larger, the impact becomes smaller and smaller both from
theoretical grounds and because of the smaller number of singleton PSUs. For
further information about variance estimation for domain estimates, see Lohr
(1999), Cochran (1977), Fuller et al. (1989), and user manuals for SAS survey
procedures (SAS, 2004) and SUDAAN (Shah et al., 1997). _^top Comparison of Variance
Estimates from Event Files
Table 1 presents a comparison of
MEPS event files for 2008 in terms of the two factors that may affect the
variance estimation from an event file—the file size and the number of variance
strata with singleton PSUs. Three files with one or more singleton PSUs are
Home Health, Hospital Inpatient Stays, and Outpatient Visits. These files also
have the smallest number of persons. To investigate the impact on variance
estimates when using the file subset to persons with events only, we compared
the variances for selected estimates from these three files. Since these files
have the smallest number of persons and have one or more singleton PSUs, any
impact on variance from not doing a proper domain analysis should be more
pronounced on the estimates from these files. _^top
Event file Total number of event
records Corresponding number of
persons and percentage of the full person file Number of variance
strata with single PSU Number Percentage A. Prescribed Medicine 293,379 17,969 57.5 0 B. Dental Visits 26,253 11,639 37.2 0 C. Other Medical Expense 6,787 5,251 16.8 0 D. Hospital Inpatient Stays 2,821 2,113 6.8 4 E. Emergency Room Visits 6,115 4,165 13.3 0 F. Outpatient Visits 11,173 3,967 12.7 1 G. Office-Based Medical Provider Visits 136,460 21,208 67.8 0 H. Home Health 4,372 692 2.2 50 For the purpose of the analysis, each of these event files
is merged with the 2008 FY person file by dwelling unit-person identifier
(DUPERSID) and all records with necessary variables from both files are kept on
the merged file. The merged file becomes an expanded event level file with one
record for each person with no event (with missing values for event related variables)
but one or more records for persons with events depending on the number of
events. An indicator variable (say, event indicator) is created to indicate if
the record came from the event file or not. Estimates and standard errors (SEs) are then produced using
SAS survey procedures in two different ways: 1) by subsetting only the records
with events i.e., using a ‘by’ statement in SAS and 2) by performing a domain
analysis using ‘event indicator’ as the domain. Using the ‘by’ statement only
the persons with events are included in the analysis, which is equivalent to
using the event file. In contrast, the expanded file with all person records
with and without an event is used and estimates are produced when using the
domain statement in SAS. Tables 2 to 4 present comparative results under the two
approaches for a selection of estimates of percentages, means, and totals from
the three selected event files. As explained previously, the point estimates
are the same under both approaches and the differences are only in standard
errors (SEs). The differences in SEs are more pronounced for the estimates from
the Home Health file which has 692 person records and 50 singleton PSUs,
negligible for the estimates from the Hospital Inpatient Stays file which has 2,113
person records and only four singleton PSUs, and also negligible for the
Outpatient Department visits file which has 3,967 person records and only one
singleton PSU. For example, the SE of the estimate of mean expense per month
for home health is 82.05 when produced from the event file and 85.38 when
produced from the full file with the domain statement. For hospital inpatient
stays, the SE for the estimate of mean expense per stay is 22.09 when produced
from the event file and 22.15 when produced from the full file. For outpatient
department visits, the SE for the estimate of mean expense per visit is 51.75
when produced from the event file and 51.66 when produced from the full file. A
similar analysis was done using the Emergency Room Visits file, which has 4,165
person records and no singleton PSU, and not surprisingly found no difference
in SEs of almost all estimates. It appears that the difference in SEs decreases
or disappears as the number of persons increases and the number of singleton
PSUs decreases in the event file. When there is a difference, SEs are generally higher when
estimated from the full file with domain analysis than when estimated from the
event file. The differences are slightly more pronounced for SEs of totals than
for SEs of means and percentages. There is a big difference in degrees of
freedom (df) for estimating SEs from the event file and the full file.
This is because in the event file some PSUs have no records but in the full
file all PSUs have some records with or without events. However, since df
is large in both cases, this difference in df will not have any impact
on the inference here. If the df were small (say, less than 30) in one
or both cases, the inference in terms of statistical testing or forming
confidence intervals would be more precise from the full file. As mentioned earlier, the treatment of singleton PSUs are
different in SAS and SUDAAN. In SAS, singleton PSUs are excluded from the
estimation of variances but in SUDAAN, when the MISSUNIT option is used, the
overall mean of PSUs is used for the missing PSU to compute variances from a
stratum with a singleton PSU. Since the number of stratum with singleton PSUs
is small for the Inpatient and Outpatient files, there was no noticeable
difference between the SE estimates from SAS in tables 3–4 and those from
SUDAAN (not shown in any table). Table 5 presents a comparison when the estimates are
produced for subgroups within the Inpatient and Outpatient event files to see
if there is any higher difference in SEs at that level. For this comparison,
SEs are produced using three approaches: 1) subsetting the event file to the
records in the subgroup of interest (i.e., using a ‘by’ statement in SAS), 2)
using the event file and specifying subgroups as a domain, and 3) using the
full file (with and without events) and specifying subgroups as a domain. The
table shows that when a domain analysis is done either using the event file or
the full file, the differences in SEs are small and negligible. However, if the
analysis is done by subsetting the file to the subgroup of interest or by using
a ‘by’ statement in SAS, the differences in SEs are substantial. For example,
the SE of the estimate of mean expense for hospital inpatient stays for
Hispanics is $1,295.6 when the analysis is done using the ‘by’ statement,
$1,363.8 when the domain statement is used in the event file, and $1,360.6 when
the domain statement is used in the full file. There are substantial
differences in df. The df is substantially smaller when the
estimates are produced by subsetting to the subgroup than when the estimates
are produced using the domain statement either from the full event file or from
the full file. The df available for variance estimation is large enough
in either the full event file or the full file that the difference can be
ignored. That means, if the domain statement is used for subgroup analysis,
either the event file or the full file can be used without worrying about a
substantial impact on the estimates of SEs or df for all event files
except for the Home Health file. A ‘by’ statement or further subsetting of file
to the subgroup of interest should never be used for analyzing any subgroup
within an event file. In this comparison, the Home Health file is not included
as the SEs are showing differences even at the overall level. Therefore, a
domain analysis with the full Home Health file should always be used for
estimation either at the overall or at the subgroup level. _^top
Variable Estimate SE of estimate Degrees of freedom Event file Full file Event file Full file Insurance status Private 29.76% 2.70 2.86 106 205 Public 69.49% 2.72 2.88 Uninsured 0.75% 0.18 0.20 Race/ethnicity
Hispanic 9.93% 1.70 1.78 106 205 NH-black 16.83% 1.49 1.68 Provider
work for agency, hospital, nursing home? Yes 74.64% 2.42 2.52 106 205 Any
care due to hospitalization? Yes 35.33% 2.50 2.59 106 205 Expense
($) Mean/visit $1,366 82.05 85.38 106 205 Total $48.87B 4.72B 4.99B OOP
expense ($) Mean/visit $153.7 37.17 37.62 106 205 Total $5.49B 1.54B 1.55B _^top Variable Estimate SE of estimate Degrees of freedom Event file Full file Event file Full file Insurance status Private 54.58% 1.65 1.66 179 205 Public 40.27% 1.60 1.61 Uninsured 5.15% 0.60 0.60 Had
surgery 39.20% 1.34 1.35 179 205 Race/ethnicity
Hispanic 10.58% 1.03 1.03 179 205 NH-black 12.79% 1.01 1.02 Expense
Mean/visit $11,349 427.11 424.59 179 205 Total $329.9B 17.43B 17.73B OOP
expense ($) Mean/visit $312.6 22.09 22.15 179 205 Total $9.09B 0.669B 0.675B Number
of nights Mean/visit 5.22 0.21 0.21 179 205 Total 151.8M 8.46M 8.58M _^top Variable Estimate SE of estimate Degrees of freedom Event file Full file Event file Full file Insurance status Private 66.07 2.82 2.82 191 205 Public 30.48 2.90 2.90 Uninsured 3.45 0.51 0.51 Any
surgery? Yes 12.24 0.78 0.78 191 205 Race/ethnicity
Hispanic 6.08 0.77 0.77 191 205 NH-black 10.48 1.45 1.45 Expense
($) Mean/visit $792.67 51.75 51.66 191 205 Total $97.76B 7.23B 7.25B OOP
expense ($) Mean/visit $66.60 5.28 5.28 191 205 Total $8.21B 593M 595M _^top SE of estimate Degrees of freedom Estimate Event file Full file1 Event file Full file1 Subset Domain2 Domain2 Subset Domain2 Domain2 Expense: Hospital
Inpatient Stay Mean for Hispanics $12,081 1,295.6 1,363.8 1,360.6 57 179 205 Mean for blacks $10,557 887.0 1,011.5 1,003.4 49 Total for Hispanics $37.2B 5.73B 6.14B 6.16B 57 Total for blacks $39.3B 3.18B 4.37B 4.46B 49 Expense: Outpatient Department Visits Mean for Hispanics $694 76.73 79.83 79.87 79 191 205 Mean for blacks $734 63.19 72.15 72.20 74 Total for Hispanics $5.24B 627M 665M 666M 79 Total for blacks $9.06B 1,025M 1,310M 1,311M 74 1 Merging of the event file and full person file _^top Conclusion
In
estimating variances of estimates from MEPS event files, theoretically the
event file should be merged with the FY person file and then a domain analysis
should be used. This is required to account for the extra variance due to the
fact that the number of persons in an event file (i.e., the sample size) is
random and the corresponding population size is unknown. However, the impact of
not doing a domain analysis on variance estimates is usually small when the
subgroup is large and the number of singleton PSUs is small. To assess the
impact on variances of producing estimates from MEPS event files without
merging with the FY person file, an analysis is performed using the four
smallest event files and the SEs of some estimates are compared. The
analysis shows that SEs are somewhat distorted (about 5 percent) for the Home
Health file if the estimates are not produced from the full file with the
domain option but this is not a notable problem for the other event files.
Generally, the differences in variances between full and subsetted files are
slightly higher for the estimates of totals than for means and proportions.
There are differences in df available for estimating variances but the
difference is ignorable since the available df under both approaches is
sufficient. For estimating variances of estimates of subgroups within an event
file, the differences in variances are negligible whether the estimates are
produced from the full event file or the full file as long as the subgroup is
treated as a domain. However, if the domain analysis is not done, the estimates
of variance can be considerably biased and the available number of df
can be less than sufficient. In summary, for analysis of estimates from the Home Health file, the estimates of
variances should always be computed by merging the event file with the full
file with domain option. For all other event files, the analysis can be done
using the event file only (i.e., without merging with the full file) without
having any noticeable impact on the estimates of variances or df. For
analyzing subgroups within the event file, a domain option should always be
used and the file should never be further subset to subgroup of interest or the
‘by’ statement should never be used. However,
the above conclusion is based on the sizes of 2008 event files and this
conclusion will be valid as long as the MEPS sample size (and hence the sizes
of event files) remain stable from year to year. If there is a substantial
decrease in the overall sample size, this conclusion may not be applicable and
the caveats described above may need to be extended to other event files than
just Home Health. Finally, the reasons and the need
for domain analysis discussed in this report are also applicable for analyzing
any subset of a full person file. Generally, a domain analysis should be used
for analyzing subgroup estimates using the full person file unless the impact
of subsetting the file is assessed. This is particularly important when the
subgroup size is not large and may be clustered geographically. _^top References
Botman S. L., Moore T. F.,
Moriarity C. L. (2000). Parsons V. L. Design and Estimation for the National
Health Interview Survey, 1995–2004. National Center for Health Statistics.
Vital Health Stat 2(130). Cochran
W. G. (1977). Sampling Techniques. New York, John Wiley & Sons, Inc. Ezzati-Rice,
T. M., Rohde, F., Greenblatt, J. (2008). Sample Design of the Medical
Expenditure Panel Survey Household Component, 1998–2007. Methodology Report
No. 22. March 2008. Agency for Healthcare Research and Quality, Rockville, MD. http://www.meps.ahrq.gov/mepsweb/data_files/publications/mr22/mr22.pdf Machlin S. R., Chowdhury S. R., Ezzati-Rice
T., DiGaetano R., Goksel H., Wun L.-M., Yu W., Kashihara D. Estimation
Procedures for the Medical Expenditure Panel Survey Household Component. Methodology
Report #24. September 2010. Agency for Healthcare Research and Quality,
Rockville, MD. http://www.meps.ahrq.gov/mepsweb/data_files/publications/mr24/mr24.pdf | |||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||
is selected from a population of size 
and the
sample randomly includes
units
from total
units in domain
. Then
is the
sampling weight (
.
, for domain 
can be estimated as
