Quality Of The Dmi File As A Business Sampling Frame

QUALITY OF U.S. BUSINESS E S T A B L I S H M E N T FRAMES: DISCUSSION
Brenda G. Cox, Mathematica Policy Research, Inc.
600 Maryland Avenue, SW, Suite 550, Washington, DC 20024-2512
Keywords:

business frames,
business surveys.

business

register,

•

Creating and maintaining sampling frames for
business surveys is difficult and expensive. Many
problems relate to the unit being measured. A business
is not a natural unit like a person or household that can
be neatly or crisply defined. Businesses add employees
and locations, lose employees and locations, and in
some cases have no employees or locations. The rate
of change in business organizations is high and they
begin operations, change ownership, and go out of
business at surprising rates.

The Standard Statistical Establishment List
(SSEL) maintained by the U.S. Bureau of the
Census.
The Dun's Market Identifiers File (DMI)
maintained by Dun & Bradstreet.

With some exceptions, use of the BEL and SSEL is limited
to the maintenance agencies. Other agencies of the Federal
government tend to use a private vendor such as Dun &
Bradstreet to obtain their business frames.
The Business Establishment List

In particular, small businesses are particularly
difficult to maintain in a business register. The point at
which they can be said to have started business can be
imprecise, with new businesses being incubated out of
homes with the only "employee" the owners who may
not consider themselves to be employees. Capturing
small businesses that die can be equally difficult. The
end of a small business can be no more dramatic than
turning off the lights and disconnecting the telephone.

Searson and Farmer (1997) present an excellent
description of the methodology underlying the Business
Establishment List or BEL. The BEL is an establishmentlevel business register compiled from administrative reports
submitted as a part of the Unemployment Insurance (UI)
program. As such, it demonstrates the complexities and
coordination problems associated with administrative
databases as well as the advantages. The complexities are
associated with the UI laws which vary across states and
result in variations in terms of exactly what businesses (and
employees) are covered. The advantages relate to having an
efficient mechanism for identifying (1) businesses just
starting operations, (2) changes in existing businesses, and
(3) businesses that have ceased operations.

On the other extreme, large businesses have
complex ownership and management structures that
make it difficult to define the characteristics of the
business entity. Large enterprises buy independent
companies, sell component companies, and merge
operations with other organizations. What constitutes
the "enterprise" can become murky, especially for the
large, complexly organized entities.

Searson and Farmer's discussion of coverage focused
on percent of employees covered. I would have also liked
to have heard about the BEL's coverage of businesses. The
BEL covers 98 percent of non-farm employment, for
instance. What fraction of non-farm employers does it
cover? Business registers tend to have more difficulty
capturing small businesses which constitute a large portion
of the nation's total business but not nearly the same
proportion of the nation's total employees. Because
inclusion equates to UI taxes in this instance, small
businesses may also have incentives to neglect to report
their existence.

These examples illustrate why building and
maintaining a high quality business register resembles
forging through a swamp riddled with quicksand. Each
step must be carefully thought out and the end
objectives must be kept in mind. The authors in this
session discuss these three business registers which
most Federal agencies use as sampling frames for
business surveys:

I fred it curious that BEL classifies contract employees
working off-site at another business as a separate
establishment. This approach seems atypical of the standard
definition which links employees who work out of a
particular establishment back to that establishment rather
than defining a new establishment for their temporary j o b

The Business Establishment List (BEL)
maintained by the U.S. Bureau of Labor
Statistics (BLS).

31

business registers (the BEL and SSEL) at the level of
individual statistical units would, no doubt, strengthen both
registers and the Federal statistical system in general;
unfortunately, confidentiality restrictions presently do not
allow such effort." I concur with this observation that both
registers would be enhanced by cross comparisons. The
Federal statistical community needs to work to see the
removal of such wasteful, artificial barriers to data quality
and productivity.

site. With this approach, a business could have as
many establishments as they have employees, yet not
have a physical space except perhaps at the central
office. The locations at which their staff works as onsite contract labor might tend to be volatile over time.
Coordination is another issue that arises for
administrative databases such as the BEL. The Bureau
of Labor Statistics does not control the legislation
under which these data are collected, nor do they have
direct control over the State Employment Security
Agencies who collect much of BEL data. The quality
issues raised by this shared data collection and the
extent to which state-to-state differences occur should
be investigated.

Dun's Market Identifiers File

Marker and Edwards (1997) discuss Dun's Market
Identifiers (DMI) File, which is used to construct the
sampling frame for private surveys as well as for federal
surveys with sponsors other than BLS and the Census
Bureau. 1 The DMI file is constructed from lists derived
from many sources, the principal ones being the Dun &
Bradstreet business ratings and telephone listings. The DMI
file is maintained for sale to users interested in direct
marketing as well as business surveys. Over the last decade
or so, considerable effort has been expended in improving
the coverage of the DMI file and that effort has bourne fruit.
That DMI file coverage is "near 98 or 99 percent" as Marker
and Edwards assert was not quite proven to my satisfaction,
however. Deadwood and duplicate records make their
direct comparisons of frame counts problematic to interpret.

Continuing business-based data collection can
achieve gains in efficiency and data accuracy by using
automated electronic methods of data collection. BLS
is moving in this direction, and in doing so is creating
or encouraging approaches that reduce respondent
burden.
The Standard Statistical Establishment List

Walker (1997) describes the second contender for
the title of the "Federal Government's Best and Most
Complete Business Register." The U.S. Bureau of the
Census uses tax return data to build the Standard
Statistical Establishment List augmented with (1)
"birth" data from the Social Security Administration
[based upon applications for Employer Identification
Numbers (EIN)] and (2) SIC codes for unidentified
businesses from the BLS. SSEL is maintained through
a register proving survey, the Company Organization
Survey, and information derived from the Bureau's
economic censuses and surveys. Use of these multiple
data sources adds complexity to the operations and has
an associated time delay attached to them. However,
maintaining an accurate business register would not be
possible without their use. Even with multiple data
sources used, noticeable deteriorization in data quality
occurs in the five years between the Company
Organization Surveys. It was unclear how or if the
Bureau uses sample-derived data to update the SSEL.
Use of sample-derived data can bias future samples
selected from the register, depending upon the
circumstances.

Out-of-business operations are not an uncommon
occurrence in the DMI file. In part, such deadwood appears
in the file because Dun & Bradstreet does not have access
to good mechanisms for identifying business deaths.
Rather, an absence of confirming evidence of life is
observed by DMI operations. Once a firm goes out of
business, the telephone listing may cease to exist, for
instance.
Quite rightly, in the absence of positive
information as to death, DMI tends to retain the business
record.
With known deadwood in the DMI file, comparisons
of DMI frame counts to BEL or SSEL derived counts can be
misleading. Such comparisons might be better made using
weighted estimates derived once the sample has identified
out-of-business operations. To the extent possible, the
counts derived for comparison purposes should also be
confined to types of businesses represented in both files.

1The Bureau of Economic Analysis gets access to
SSEL through the 1990 Foreign Direct Investment and
International Financial Data Improvements Act. Federal
agencies can use the BEL after obtaining permission from
the individual states.

An evaluation of SSEL data quality is now being
planned, and I strongly encourage such action as many
important economic surveys are derived from the
SSEL. Walker also notes that, "Reconciliation of these

32

Comparisons should be made at the enterprise and at
the establishment level when possible and ideally
should include breakdowns by business activity
(manufacturing,
transportation,
service,
etc.),
ownership type (corporation, partnership, sole
proprietorship), and size (number of employees).
The mechanisms used to build the DMI file also
affect the quality of the individual DMI data. The
number of employees variable is not reliable enough to
use to exclude ineligible businesses. DMI-derived SIC
codes show more correlation to survey-derived SIC
codes but are often in disagreement, although the
company usually reports a related industry type to that
of the DMI. I recommend caution in using DMI data
values to eliminate out-of-scope businesses. Incorrect
entries can lead to sample undercoverage when
database items are used to automatically exclude
businesses from the sample. It may be preferable to
include potential out-of-scope businesses in an initial
screening interview which can also update location
information and current operating status.
As did Searson and Farmer for BEL, Marker and
Edwards compare coverage in terms of employment
counts. This makes sense for BEL as BLS is interested
in employment data. This comparison may also make
sense for the National Employer Health Insurance
Survey that Marker and Edwards mention, because
here employee results are clearly of interest. However,
larger businesses account for a substantial percentage
of total employment, so coverage results for number of
employees can be expected to be higher than for the
number of businesses covered.
Employee-based
coverage rates can be deceiving for statistics unrelated
to business size or surveys that collect such statistics.
Examples might be a statistic such as the percent of
businesses offering health insurance benefits or
similarly a survey interested in the distributional
attributes of businesses.

Some businesses may never be added to the DMI
database. It is not uncommon for small businesses to be
born and to die within a year or so. Marker and Edwards
fail to factor in this consideration in projecting how long it
takes before new businesses enter the DMI database.
My previous experience with the DMI file led me to
resolve not to use the establishment-level data for sampling
purposes.
I was pleased that Marker and Edwards
suggested that coverage at the establishment level has
improved. I would still recommend that surveys sample at
the enterprise level whenever possible, as DMI coverage of
enterprises is more complete, almost by definition, than its
coverage of the individual establishments each enterprise
contains.

Concluding Remarks
It is sad that we have three different business registers
being used for Federal business surveys. A friend of mine,
Michael Colledge at the Australian Bureau of Statistics,
calls this lack of data sharing between Federal statistical
agencies--"that distinctly American problem." The Federal
statistical community is making some progress in promoting
data sharing, but the pressure should continue to remove
these unnatural impediments to data quality and efficiency.

References
Marker, David A., and W. Sherm Edwards (1987). "Quality
of the DMI File as a Business Sampling Frame." Invited
paper, Joint Statistical Meetings, Anaheim, CA.
Searson, Michael A., and Tracey E. Farmer (1997).
"Quality of the Bureau of Labor Statistics' Business
Establishment List as a Sampling Frame." Invited paper,
Joint Statistical Meetings, Anaheim, CA.
Walker, Ed (1997). "The Census Bureau's Business
Register: Basic Features and Quality Issues." Invited paper,
Joint Statistical Meetings, Anaheim, CA.

33