Attachment 1 - Evaluating Sample Design Issues in the NCS

Section on Survey Research Methods – JSM 2010

Evaluating Sample Design Issues
in the National Compensation Survey October 2010
Gwyn R. Ferguson1, Chester Ponikowski2, and Joan Coleman3
1

U.S. Bureau of Labor Statistics, 2 Massachusetts Ave., NE, Room 3160,
Washington, DC 20212
2
U.S. Bureau of Labor Statistics, 2 Massachusetts Ave., NE, Room 3160,
Washington, DC 20212
3
U.S. Bureau of Labor Statistics, 2 Massachusetts Ave., NE, Room 3160,
Washington, DC 20212

Abstract

The National Compensation Survey is conducted by the Bureau of Labor Statistics to
measure employment cost levels and trends, incidence of employer-provided benefits,
benefit plan provisions, occupational earnings by geographic area, and occupational pay
comparisons between areas. The current survey design uses a three stage sample design
to select samples of areas, establishments, and jobs for which wage and benefit data are
collected periodically over a five-year rotation. In recent years, several potential changes
to this design have been explored due to budget cuts, known issues with the current
design, and an on-going effort to make the survey more efficient. This paper discusses the
issues and alternative approaches to the current design being explored and presents some
recommended changes to the general survey design.

Key Words: survey design, dependent sampling, respondent burden, sample rotation
1. Introduction
The National Compensation Survey (NCS) is an establishment-based survey conducted
by the U.S. Bureau of Labor Statistics (BLS). Over the last several decades, the NCS has
undergone many changes leading up to the survey design currently in operation which
has been used by NCS since the mid-1990’s. In recent years, several potential changes to
this design have been explored due to budget cuts, known issues with the current design,
and an on-going effort to make the survey more efficient. This paper presents an
overview of the NCS that includes its scope and major product lines in Section 2. Section
3 provides an overview of the current sample design and rotation strategy. Section 4
outlines several issues with the design while Section 5 describes the objectives for any
design changes. Some challenges and constraints on design changes and research are
described in Section 6. Section 7 describes the efforts currently underway to evaluate
potential changes to the sample design and rotation strategy and future plans for
additional research to evaluate the issues. The paper concludes with a summary of the
status of the analysis in Section 8.

Section on Survey Research Methods – JSM 2010

2. Survey Background
The NCS provides comprehensive measures of occupational earnings, employer costs of
employee compensation, compensation trends, wages in one geographic area relative to
other geographic areas, and the incidence and provisions of employer-provided benefits
(BLS Handbook of Methods, Chapter 8). The Employment Cost Index (ECI)—a
Principal Federal Economic Indicator—is estimated from data collected by the NCS.
The NCS produces several types of data with varying degrees of frequency as
summarized below.
Employment Cost Index (ECI) data are released quarterly
Employer Costs for Employee Compensation (ECEC) data are released quarterly
Incidence and Provisions of Employer Provided Benefits data are released
annually
Detailed Provisions for employer provided health insurance, defined benefit
retirement plans, and defined contribution retirement plans are released once a
year with a focus on one of these benefit areas each year
Occupational earnings data for the nation, each Census Division, and selected
geographic areas are released once a year
Occupational pay comparisons for approximately 80 geographic areas are
released once a year
The NCS covers workers in private industry establishments, and in State and local
government, in the 50 States and the District of Columbia. For the NCS, the term
“civilian workers” denotes workers in private industry and workers in State and local
government. Establishments with one or more workers are included in the survey.
Excluded from the survey are workers in the Federal Government and quasi-Federal
agencies, military personnel, agricultural workers, workers in private households, the
self-employed, volunteers, unpaid workers, individuals receiving long-term disability
compensation, individuals working overseas, individuals who set their own pay (for
example, proprietors, owners, major stockholders, and partners in unincorporated firms),
and those paid token wages.

3. NCS Current Survey Design
The BLS Quarterly Census of Employment and Wages (QCEW) serves as the sampling
frame for the NCS survey. The QCEW is created from State Unemployment Insurance
(UI) files of establishments, which are obtained through the cooperation of the individual
state agencies (BLS Handbook of Methods, Chapter 5).
The NCS sample consists of five rotating replacement sample panels for private industry
establishments, an additional sample panel for State and local government entities, and an
additional panel for private industry firms in the aircraft manufacturing industry. Each of
the sample panels is in the sample for at least five years before it is replaced by a new
sample panel selected annually from the most current frame.
The NCS sample is selected using a three stage stratified design with probability
proportionate to employment size (PPS) sampling at each stage. The first stage of sample
selection is a probability sample of areas; the second stage is a probability sample of

Section on Survey Research Methods – JSM 2010

establishments within sampled areas; and the third stage is a probability sample of
occupations within sampled areas and establishments.
The first stage of the NCS sample occurs at the national level across geographic areas.
These Primary Sampling Units (PSUs) are based on the 2003 Office of Management and
Budget (OMB) area definitions. Under the OMB definition there are three types of
statistical areas. These area types are defined as Metropolitan, Micropolitan, and
Combined Statistical Areas. Combined Statistical Areas (CSAs) are defined as a
combination of adjacent Metropolitan and Micropolitan areas that meet certain conditions
set by OMB. Outside of these areas exists a number of counties. These counties are
referred to as Outside Core Based Statistical Areas (CBSA). For selection purposes,
PSUs in these outside CBSA’s consist of one or more adjacent counties. Where possible
the counties are organized into clusters to create heterogeneous primary sampling units.
In 2004, a new area sample was selected for the NCS. This sample contains 152 areas. In
this sample 57 areas were selected with certainty, where certainty areas are defined as
having employment greater than 80 percent of the final sampling interval, which is
obtained through an iterative process. The remaining areas consisted of 60 non-certainty
metropolitan areas, 22 non-certainty micropolitan areas, and 13 non-certainty outside
CBSA county clusters. All establishment samples selected since 2005 have been selected
from the 2004 area sample while establishment samples selected in 2004 and earlier were
chosen from the previous sample of areas. Until all establishment samples selected before
2005 rotate out of the survey in 2012, data collected from samples selected from the
previous area sample and from the 2004 area sample will be included in the various NCS
published estimates.
The second stage of this design occurs at the establishment level within each selected
area. Establishments in the sampling frame are stratified by ownership and industry.
Industries for the NCS are defined using the North American Industry Classification
System (NAICS). Within each of the ownership by industry strata, NCS employs PPS
systematic sampling with frame employment as the measure of size (MOS). To ensure
that no unit has a probability of selection greater than one, we identify all units that
would be selected with certainty before the sampling process, designate them as part of
the sample, and set their sampling weights to one. These certainty units with a weight of
one are identified once every five years and are included in each yearly sample until we
identify a new set of certainty units. These units are referred to as multi-year certainties.
By including them in every annual sample, we ensure that each sample represents the
target population while making it easier operationally to process data for the various NCS
outputs. During the selection process, approximately one-half of the establishments, the
index portion, are sub-sampled and flagged to support the ECI, ECEC, and NCS Benefits
products as well as the NCS wage products. The remaining establishments, the wage-only
portion, are flagged to support the NCS wage products only. After the sample of
establishments is selected, it is used for the third stage of the sampling process.
The third stage of this design occurs at the occupational level within each selected
establishment. A sample of jobs is drawn from each of these establishments using PPS
systematic sampling where the number of employees in the job is the measure of size. To
ensure consistency across all establishments, the Standard Occupational Classification
(SOC) manual is used to classify the selected jobs into occupations based upon the
assigned duties. After this selection and classification we create our smallest aggregate
unit known as a quote, which is a distinct combination of time or incentive pay, work

Section on Survey Research Methods – JSM 2010

level, collective bargaining status, full-time or part-time status, and establishment defined
occupation.
Establishments in each sample are initiated over a one-year time period. During the
initiation process, respondents are identified, jobs are selected, and respondents provide
BLS with initial information about each selected job quote. All establishments are asked
to provide BLS with employer provided wages and salaries for all workers in each
selected job quote. Establishments in the index portion of the sample are also asked to
provide the cost of each employer provided benefit, a description of each benefit offered
to the employees in each selected occupation, and benefit access and provisions data such
as the number of employees who are offered the benefit, the number who partake of the
benefit, and detailed descriptions of the benefit.
Respondents are asked to provide periodic updates for the initiated occupations for the
next five years. Index respondents are asked to provide quarterly updates while wage
respondents are asked to update their data annually. At the end of the five year update
period, NCS thanks the respondents for supporting our survey and ceases to ask for
updated data unless the respondent has been selected in a subsequent sample. The chart
below shows the current rotation strategy. Each row in this chart is an independent
establishment sample. The yellow rows indicate the introduction of a new area sample
while the vertical red lines indicate the transition period when estimates include
establishments selected from two area samples.
Sample Group

2005

2006

2007

2008

2009

2010

2011

2012

2013

2014

2015

2016

2017

NCS 90 - Governments
NCS 68 - Aircraft Manufacturing
NCS 05 - Private Industry
NCS 101 - Private Industry
NCS 102 - Private Industry
NCS 103 - Private Industry
NCS 104 - Private Industry
NCS 901 - Governments
NCS 105 - Aircraft Mfg
NCS 106 - Private Industry
NCS 107 - Private Industry
NCS 108 - Private Industry
NCS 109 - Private Industry
NCS 110 - Private Industry
NCS 111 - Private Industry
NCS 112 - Private Industry
NCS 113 - Private Industry
NCS 114 - Private Industry
NCS 115 - Private Industry
NCS 902 -Governments
NCS 116 - Aircraft Mfg
NCS 117 - Private Industry

<-------- 2004 Area Sample Transition --------->
Area Sample Rotation
Sample Selection & Refinement

Initiation

Update/Estimation

Figure 1: NCS Sample Rotation Chart

4. Issues with the Current Design
Since the introduction of the current sample design, NCS has used the same basic design
and rotation strategy. Initially, the design worked well and there were no major problems
with its implementation. But as we used the design for a longer period of time, several
things changed leading the survey to conclude that we had some potential issues with the
design.

Section on Survey Research Methods – JSM 2010

In the last several years, NCS has completed several studies of the current sample design.
In 2009, NCS completed an internal review of the sample design and issues related to
it from the NCS employee’s perspective. This review identified several concerns with
the sample design that prompted us to begin an evaluation of several major
components of the design.
From 2006 to 2009, the BLS compared the National Compensation Survey to the
Occupational Employment Statistics (OES) survey in terms of concepts,
measurement objectives, and sample design. The OES is an establishment based
survey that provides employment counts and earnings by occupation (BLS Handbook
of Methods, Chapter 3). Based on this comparison, we identified several areas where
the two survey designs were different and could potentially be changed to be more
consistent, although we agreed that additional research was needed before making
any changes.
From fiscal year 2004 through fiscal year 2009, the NCS budget was cut four
different times, resulting in reduced sample sizes each time. While the budget
reductions were implemented and the sample sizes were reduced, the process for
implementing the reductions became more difficult due to the design of the sample
allocation process which assumes fixed sample sizes for five years at a time. Most
recently, the President’s proposed fiscal year 2011 budget calls for the elimination of
the Locality Pay Survey component of the NCS with a transition to a national design
and a reduction in the overall sample size for the remaining portions of the survey.
Although no final budget for fiscal year 2011 has been approved as of the writing of
this paper, NCS will need to make several changes to the sample design if the
President’s proposed budget is implemented.
Using the results of each of these studies and budget cuts, NCS staff compiled a
comprehensive list of issues with the sample design as described below.
1) As a result of independent PPS sampling and identification of multi-year certainty
establishments, many units are selected year after year in the NCS. This frequent
sampling and the continuous need to obtain current wage and benefits data from
employers has the potential of overburdening some respondents by requesting data
for multiple samples simultaneously.
2) Under the current five-year rotation, it takes several years to make changes to our
scope, definitions, sampled areas, etc. In order to meet publishability criteria, we
need to initiate several years of samples under new definitions, areas, and/or
processes before fully moving to the new coverage.
3) The five-year rotation scheme is a concern, especially for private industry
establishments which remain in the survey for six or more years, up to one year for
sample initiation and at least five years for sample updates. This also creates an
attrition problem -- the rate of attrition in private industry is approximately 1 percent
per quarter; thus, sample groups approaching replacement may be more prone to bias
due to larger nonresponse.
4) The current allocation methodology assumes a constant budget and fixed sample size
levels over a five year period. Under this methodology, the allocation process should
be executed at the beginning of each five year period, multi-year certainty
establishments are to be identified and held constant for the five year period, and
annual samples should be selected from a new frame that excludes the multi-year
certainties using the previously identified sample sizes. The recent budget climate has
not supported this continuity. In fact, NCS has seen budget reductions in four of the

Section on Survey Research Methods – JSM 2010

5)

6)

7)

8)
9)

last seven budget years, forcing the survey to implement sample size changes in the
middle of the five year period. These budget changes have resulted in uneven sample
sizes from one year to the next and a set of multi-year certainty establishments that is
larger than would be selected given the current sample size.
The President’s proposed budget for fiscal year 2011 calls for the elimination of the
locality pay portion of the NCS and a change to a national sample design instead of
the current area-based design. Under this proposal, NCS will cease to produce
occupational earnings estimates as direct computations from sample data, eliminating
the requirement to sample geographic areas in support of the area-based earnings
data. If Congress approves this budget proposal, several changes will need to be
made to the sample design including eliminating the first stage of sampling to select
areas, preparing a sample frame from establishments across the nation instead of the
sampled areas, eliminating the process of identifying establishments that will support
only the wage products, and updating the weighting modules to reflect these changes.
Including establishments that appear on the frame with zero employment during the
frame reference month may be causing NCS to sample establishments that are no
longer in business, also leading to sample inefficiencies due to a smaller useable
sample. NCS currently looks at a single month of employment data on the frame. In
seasonal establishments, this month’s employment may be zero even though the
establishment has many employees in other months of the year. Thus, NCS decided
to include all establishments on the frame and set the employment to one when it is
equal to zero on the frame. However, this may lead to inclusion of establishments
that have gone out of business during the quarter from which the frame was extracted
or giving establishments very small employment values when they typically have
much larger employment.
Complexity is also an issue with the current sample, but it is inherent. The NCS
sample data are used to produce many outputs within a limited budget constraint and
different subsets of the NCS data are used for the various outputs. For example, the
ECI is produced using all samples that have completed the initiation process and have
not yet rotated out of the survey. However the detailed provisions products only
include data from the most recently initiated sample. Thus, synchronization of
changes to the methods and concepts is difficult especially since the sample affects
every aspect of the surveys, from data collection to review, estimation, and
publications.
Response rates are lower than desired and require nonresponse bias studies. NCS has
begun conducting nonresponse bias studies on the wage outputs produced by the
survey as documented in Ponikowski et al (2006) and Crockett et al (2008).
New establishments on the sample frame, sometimes called birth establishments, are
not included in every establishment sample included in the NCS estimates. Since
most of the NCS estimates include data from multiple samples, this can lead to
underrepresentation of the current frame due to exclusion of these newly formed
businesses in samples that have been in the survey for two or more years.

5. Design Objectives
As NCS has begun an evaluation of each of the sample design and rotation issues, we
have agreed that all design changes must meet several objectives. First, the new design
alternatives must meet the NCS measurement objectives related to the ECI, ECEC, and
Benefits products. The critical measurement objectives for the NCS are that the not
seasonally adjusted Total Compensation and Wages series of the ECI have a standard

Section on Survey Research Methods – JSM 2010

error of change that does not exceed 0.3 for the 3-month changes and 0.5 for 12-month
changes at least 75 percent of the time. In addition, the survey should support production
of the various outputs for the Civilian, Private, and State and local government sectors of
the economy at various specified levels of detail by ownership, industry, and geographic
area. Second, the design shall allow quicker implementation of survey changes,
especially mandated changes such as area definitions, NAICS, and SOC. Third, any
design changes must address respondent burden issues related to frequent sampling of
some employers. Fourth, the new design should provide the ability to adjust sample sizes
as necessary with fluctuating budgets. Fifth, the new design should address response rate
concerns. And, finally, the new design changes should decrease the complexity of the
survey design, where possible.

6. Challenges and Constraints
As with all survey design research, there are several challenges and constraints on this
work, some imposed by external forces and some imposed internally. For the NCS, the
primary external challenges deal with the Federal budget cycle. If the fiscal year 2011
budget as proposed by the President is approved by Congress, we will need to begin
implementing sample design changes quickly. However, the Bureau of Labor Statistics
and NCS in particular have not received any extra funds or resources to develop or
evaluate changes to our design. In addition, due to the budget cuts in recent years, NCS
has been unable to replace staff as they have left the Bureau resulting in fewer available
research staff than we had for earlier design efforts. Thus, all research must be conducted
using current staff and resources by implementing production efficiencies and delaying
other planned improvements to the survey processes. The primary internal constraint
facing us as we evaluate potential design changes is the time needed to implement any
approved changes. Since NCS selects a single sample each year, design changes must be
approved far enough in advance of the selection process to ensure adequate time to
develop and test all changes to our sampling systems before the next sample selection
process begins. For example, the design changes need to be known and approved in the
fall of 2010 for the sample scheduled to begin collection in the spring of 2012.

7. Research and Testing
7.1 Previous Research
Several different research efforts to explore the issues and identify potential alternatives
for resolving them have already been completed. As mentioned in Section 4, NCS
completed an internal review of the sample design from the NCS employee’s perspective
in 2009. This review identified several issues with the design but did not propose any
specific ways to resolve those issues.
Also in 2009, BLS compared the NCS survey design and measurement objectives to
those for the Occupational Employment Statistics survey. This comparison identified
several areas where the two BLS surveys were different and made four recommendations
for further research. The first recommendation was that the two surveys should adopt
common concepts, definitions, and coverage. One of the specific areas mentioned under
this recommendation was that the two surveys should look at the definition of an
establishment and the establishment’s measure of size used for sample selection purposes
since their current approaches were different. The second recommendation was that the

Section on Survey Research Methods – JSM 2010

BLS should use econometric modeling techniques to present a single set of wage
estimates to users. The third recommendation was to streamline and gradually expand
coordinated collection and coding of establishments which appear in both of these
surveys. The fourth recommendation from this comparison was to implement fully
overlapping private industry sampling in 2014 or later. This fourth recommendation
included some specific areas for further study between the two surveys to find ways to
better coordinate their survey designs. Several of the detailed recommendations included
in this report are being researched at this time as described in Section 7.2.
In early 2010, NCS completed a detailed analysis of the effort required to collect NCS
data. This analysis included a detailed study of the amount of time needed to initiate an
establishment into the survey as well as the amount of time needed to update the data for
each establishment on a periodic basis. Using this data, NCS was able to evaluate the
staffing levels needed for both a five-year rotation and a three-year rotation. This analysis
showed that the NCS did not have sufficient staff levels to implement a three-year
rotation under the current design. However, if the proposal to eliminate the locality wage
portion of the survey is approved in the fiscal year 2011 budget, the analysis showed that
a three-year rotation would be feasible from a staffing viewpoint and would help balance
the workload over the course of each year. Since a three-year rotation would permit NCS
to implement change faster and helps balance data collection workload, NCS decided to
focus our remaining sample design research on a three-year rotation. Where feasible, we
will compare the proposed three-year rotation design changes to the results that would be
obtained under the current sample design and five-year rotation.

7.2 Current Research

Using the previous research as a base and the design objectives as a goal, NCS identified
several areas for further research. By looking at the sample life-cycle, we identified and
have begun researching several potential changes related to national design stratification,
sample allocation, sample selection of establishments and alternatives for large
establishments, the method of assigning an employment measure of size to establishment
on the sample frame, and an analysis of the frame coverage.
The current NCS establishment sample design stratifies the frame by geographic area and
industry as defined by NAICS. The geographic areas are defined to be the areas which
were selected during the area selection phase of the design. If the President’s proposed
fiscal year 2011 budget is implemented, NCS will move to a national design and cease
the selection of geographic areas. In evaluating the stratification needs under a national
design, NCS considered the critical estimates desired for the ECI, ECEC, and benefits
product lines. The NCS currently produces ECI and ECEC estimates for each of the
fifteen largest metropolitan statistical areas and plans to continue these series. In addition,
NCS publishes several data products by Census Division and wishes to continue
publication of these series. Based on these needs, we identified 24 geographic strata
covering the 50 states and the District of Columbia – one for each of the fifteen largest
metropolitan areas and one for the remaining states and counties in each of the nine
Census Divisions. In addition, NCS publishes many outputs based on industry. Currently
the NCS publishes data based on 24 detailed industry groupings for private
establishments. Twenty-three of these detailed industry groupings are included in each of
the private industry establishment samples while the last detailed private industry

Section on Survey Research Methods – JSM 2010

grouping, aircraft manufacturing, is included in a separate sample panel1. For
stratification purposes, we will continue to use these detailed industry groupings in the
sample design. However, all current research efforts will focus on the 23 detailed
industries typically included in each private industry establishment sample.
One of the major challenges with the current NCS sample design is that it stratifies the
private industry establishment samples by 152 geographic areas and 23 detailed industry
groupings, resulting in 3,496 potential sampling cells for private industry establishments.
With an overall sample size of approximately 13,400 private industry establishments to
be selected over a five year period, there are not enough establishments to fully allocate
the sample size to every potential sampling cell. So the NCS implemented a controlled
rounding approach to distribute the allocation across the sampling cells and the five years
of establishment samples. This process is complicated and not designed for frequent
changes in sample size. With the move to a National design, NCS would only have 24
geographic cells. If we keep the 23 detailed private industry cells (excluding aircraft
manufacturing) we would have a total of 552 sampling cells. But with a new private
industry sample size of slightly less than 10,000 establishments over a full three-year
rotation, it would still be difficult to ensure coverage of each sampling cell every year.
So, we are researching the impact of moving to a design that has five aggregate industry
strata within each geographic area for allocation and creation of independent sampling
cells while using the 23 detailed industries for implicit stratification. In this application,
implicit stratification is done by sorting the establishments by detailed industry within
each aggregate industry sampling cell prior to selecting a systematic sample to ensure
that selections are made in each of the detailed industries.
Using the 120 geographic area by aggregate industry cells (i.e. 24 areas by 5 industries),
we have begun evaluating the effectiveness of using fewer sampling cells and replacing
the controlled rounding approach with a systematic/distributed rounding approach to
obtain allocated annual sample sizes. Under the proposed systematic/distributed rounding
approach, each area by aggregate industry cell will be assigned a three year sample
allocation. This allocation will be divided into three integer allocations by dividing the
total allocation by three and then adding any remainder to that number of individual
yearly allocations. The individual years to receive the extra allocation will be assigned
systematically so that the total sample size within an aggregate industry is held constant
each year. The process will also work to ensure that the total sample size across all
industries is constant from one year to the next by beginning the systematic assignment of
the remainder allocations randomly. The table below shows an example of the resulting
yearly rounded allocations with 4 areas (A1, A2, A3, and A4) within one industry stratum
where the first random year start is 2.

1

Aircraft manufacturing is selected separately due to the size of the industry and through a
contract with the Aerospace Industries Association under which BLS provides more detailed
data for this industry than would otherwise be feasible. A new sample of establishments is
selected approximately once every ten years.

Section on Survey Research Methods – JSM 2010

Table 1: Example of NCS Systematic/Distributed Allocation Rounding
Area

3-year allocation

A1
A2
A3
A4

10
12
5
7

Year 1
Rounded
Allocation
3
4
2
2

Year 2
Rounded
Allocation
4
4
1
3

Year 3
Rounded
Allocation
3
4
2
2

Our analysis shows that this approach yields approximately the same number of selected
establishments in each of the more detailed industries as our current approach that
includes controlled rounding. By using a recent frame and the detailed industry sample
sizes, we created new sampling cells based solely on the aggregate industry groupings.
We then selected 100 simulated samples within each of the aggregate industry sampling
cells and computed the mean sample size in each of the geographic area and detailed
industry cells across the 100 samples. The mean sample sizes from this simulation were
close to the original allocated sizes, showing that we can get comparable sample sizes for
the 23 current industries with implicit allocation as we obtain under our current explicit
sample allocation process. We also learned that it is acceptable to use the proposed
systematic/distributed rounding approach to distribute the sample size across geographic
areas within each industry because that is what implicit allocation accomplishes by
design.
We then continued an analysis of a new allocation process for a three-year rotation and a
national sample design. To conduct this analysis, we obtained a full frame of
establishment data from the Bureau’s QCEW database and assigned a geographic area,
detailed industry grouping, and aggregate industry grouping to every establishment on the
frame. Under this approach, we assigned a fixed proportion of the total size to each
industry using prior response and variance information and distributed the industry size
across each of the 24 national geographic areas in proportion to the total employment in
the area. As with the current design, we identified multi-year certainties in each
geographic area by aggregate industry cell and set the probability of selection for these
units equal to one. The remaining non-certainty sample size for each sampling cell was
distributed across the three years using the systematic/distributed rounding approach
described above. We then selected fifty simulated samples from the full frame for each of
the yearly allocations, i.e. Year 1, Year 2, and Year 3 allocations. We computed the mean
sample size across all 150 simulated samples by aggregate industry, detailed industry,
geographic area, and BLS collection region. Along with each mean sample size, we also
computed the standard error of the mean sample size as well as the minimum, maximum,
mode, and median sample sizes across the samples. We are currently using this data to
help us better understand the impact of using this design and allocation approach.
We are also evaluating some sample design alternatives to deal with establishments that
appear in multiple samples. The main alternative being investigated at the current time is
dependent sampling. Under the current sample design, all establishments on the sample
frame (except for the multi-year certainties) are eligible for inclusion in the sample.
Under dependent sampling, we would exclude all establishments in the two most recent

Section on Survey Research Methods – JSM 2010

samples from the frame used to select the current sample. This research effort has been
conducted using the current area-based sample and five-year rotation design and is
described in the paper by Ojo and Ponikowski (2010). We plan to repeat this research
using the three-year rotation and national sample design before making any changes from
our current independent sample design with multi-year certainties.
The NCS sample frame is drawn from the BLS QCEW database. This database contains
one record for each quarter that an establishment reports to its state unemployment
insurance file. Each record contains the establishment’s monthly employment for each
month in the quarter and the total wages paid to employees for the entire quarter. On the
BLS QCEW database, it is possible for an establishment to exist on the database with an
employment value of zero for one or more months of the quarter. This can occur for
many reasons including newly forming establishments (sometimes called births),
seasonal employment, and establishments that are in the process of going out of business.
When the NCS encounters establishments on the frame with an employment value of
zero, we currently set their employment to one so that the establishment has a chance of
being selected in the sample. But this assumes that all of these establishments are
seasonal or birth establishments and ignores the potential that some of them will not be in
business by the time the sample is selected and data collection begins. So we are
currently evaluating the nature of these establishments on the frame by tallying their
frequency and employment trends over a multi-year time period. We are also evaluating
several different approaches to decide when to use these establishments in the frame and
how to set their frame employment values. For each option, we are setting the assigned
employment value, selecting 100 simulated samples, computing the mean monthly wages
across the simulated samples, and comparing the mean monthly wages from the samples
to those computed using the full frame. Until this analysis is completed, we will not
change this portion of the sample design.
NCS is also evaluating the frame for a National design. On the QCEW database, each
establishment is assigned numeric codes representing the state and county in which the
company conducts business. Under the current area based design, only establishments
with a state and county code located in one of the sampled areas are included in the NCS
sample frame. Under the proposed National design, all establishments in each county will
be included on the frame as long as they belong to the in scope industries. But some of
these establishments do not have a distinct county code on the QCEW as they represent
firms that conduct business state-wide, outside the given state, or in foreign locations.
NCS is currently evaluating the magnitude and employment of firms with these county
designations and gathering information about how other BLS surveys treat these
establishments in order to decide whether to include some or all of these unusual
reporters in the NCS sample frame.

7.3 Future Research Plans

As we continue to evaluate the above mentioned research topics, we have also identified
several other questions for future research. We expect to begin research on these new
topics in the near future and plan to report on the results of all research efforts through
published papers.
The first new area we would like to study is what impact the proposed sample design
changes would have on our survey response rates. Since we are not changing the data
being collected or the collection methods or forms, we expect to be able to achieve our
current response rates for establishments of similar size and in the same industries. So we

Section on Survey Research Methods – JSM 2010

will compute our historical response rates by establishment size class and detailed
industry across all samples currently supporting the ECI and ECEC product lines. We
will then compute the mean number of establishments in the same establishment size
classes and detailed industry groupings from our 150 simulated samples. By applying the
historical response rates to the simulated sample results, we will compute an aggregate
expected response rate across all size classes and industries.
Next, we want to evaluate the impact of the new design on our published outputs. To do
this, we are identifying all critical and high priority estimate publication categories. We
will then tabulate the number of establishments in the sample for each of these categories
and compare that to the minimums needed to support the NCS published outputs.
There are several other issues that we believe need to be addressed as we transition from
our current design to a new design. For example, we would like to conduct additional
research to evaluate the potential impact on the survey accuracy due to the design
changes by evaluating the theoretical aspects of the survey design as well as results of
related types of changes to the design in the past. We will also need to evaluate the
current strategy of selecting establishment samples for the State and local government
establishments and the aircraft manufacturing industry approximately once every ten
years to determine if we should continue with this approach of separate samples or if we
should include these establishments in every sample. In addition, we need to determine if
we will have sufficient data collection staff to continue to update all the establishments in
the current private industry samples until we finish transitioning to the new design. If not,
we will need to figure how to reduce the size of the samples in update mode 2, figure out
how to update them more efficiently, and/or identify additional short-term resources to
conduct the update process.
Each of the samples in update mode reflects the establishments in the frame at the time
the sample was drawn. The sample weights are updated to reflect losses due to
establishments that have gone out of business since the sample was drawn. But these
samples (up to 80% of the overall private industry sample) do not include any
establishments that have begun business since the frame for that sample was prepared.
These are called the birth establishments. We expect to conduct future research to
determine if we should supplement the older samples with birth establishments.

8. Conclusion
As described in section 7.3, NCS is currently conducting several pieces of research on
potential changes to the sample design. We plan to implement survey design changes that
are based on research results with positive outcome, that is, ones that lead to survey
design improvements and that are not very costly to implement starting with the selection
of our sample in late summer 2011. The transition from the current to a new design will
need to be completed one establishment sample at a time over a multi-year time frame –
three years of initiation efforts for private industry followed at some point by a new State
2

Since it could be several years before we reselect the State and local government sample, NCS
plans to reduce the size of the current sample for these establishments by subsampling the
existing sample and continuing to update the data only for those establishments selected
during this process. However, no decision has been made about whether or not to reduce the
size of the private industry samples at this time.

Section on Survey Research Methods – JSM 2010

and local government sample and a new private aircraft manufacturing sample. Outputs
during the transition will be computed using a combination of data selected under the old
and new designs.
Although there is still a lot of work to do before a new sample design is completed for the
NCS, there are many positive results that we have learned from both prior knowledge and
experience as well as from current research project results. The research will continue
and even after the new sample design has been implemented, the NCS staff will continue
to look for ways to improve our sample design to provide our users with the best possible
outputs.

References
Cochran, W. G. (1963), Sampling Techniques, New York: John Wiley & Sons, Inc.
Crockett, J., Ponikowski, C.H., and McNulty, E.E. (2008), ―Upd
ate on Use of
Administrative Data to Explore Effect of Establishment Nonresponse Adjustment on
the National Compensation Survey Estimates,‖ 2008 Proceedings of the Section on
Survey Research Methods, Alexandria, VA: American Statistical Association.
Ernst, L.R., Guciardo, C., Ponikowski, C.H., and Tehonica, J. (2002), ―Sa
mple
Allocation and Selection for the National Compensation Survey,‖ 2002 Proceedings
of the Section on Survey Research Methods, Alexandria, VA: American Statistical
Association.
Izsak, Y., Ernst, L. R., Paben, S. P., Ponikowski, C.H. and Tehonica, J. (2003). ‖
Redesign of the National Compensation Survey.‖ 2003 Proceedings of the Section on
Survey Research Methods, [CD-ROM], Alexandria, VA: American Statistical
Association.
Izsak Y., Ernst, L. R., McNulty E., Paben, S. P., Ponikowski, C. H., Springer G., and
Tehonica, J. (2005). ―
Update on the Redesign of the National Compensation Survey.‖
2005 Proceedings of the Section on Survey Research Methods, [CD-ROM],
Alexandria, VA: American Statistical Association
Ojo, O. E. and Ponikowski, C. H. (2010), ―
Evaluating the Effect of Dependent Sampling
on the National Compensation Survey Earnings Estimates‖, 2010 Proceedings of the
Section on Survey Research Methods, [CD-ROM], Alexandria, VA: American
Statistical Association.
Ponikowski, C. H., and McNulty, E. E. (2006), ―Us
e of Administrative Data to Explore
Effect of Establishment Nonresponse Adjustment on the National Compensation
Survey Estimates,‖ 2006 Proceedings of the Section on Survey Research Methods,
Alexandria, VA: American Statistical Association.
U.S. Bureau of Labor Statistics (1997) BLS Handbook of Methods, Employment and
Wages Covered by Unemployment Insurance, Chapter 5. http://www.bls.gov/
opub/hom/homch5.htm
U.S. Bureau of Labor Statistics (2008) BLS Handbook of Methods, National
Compensation Measures, Chapter 8. http://www.bls.gov/opub/hom/homch8_a.htm
U.S. Bureau of Labor Statistics (2008) BLS Handbook of Methods, Occupational
Employment Statistics, Chapter 3. http://www.bls.gov/opub/hom/pdf/homch3.pdf
Any opinions expressed in this paper are those of the authors and do not constitute policy
of the Bureau of Labor Statistics.