Technical Information: Estimation Methods for Business Births and Deaths

Background

The Current Employment Statistics (CES) program, also known as the payroll survey, produces nonfarm employment, hours, and earnings series each month based on a monthly sample of nearly 400,000 business establishments nationwide. The CES is designed as a high volume, quick turnaround survey, collecting a few readily available data items from each sampled firm's payroll. The CES is a simple random sample, stratified by geography, industry and employment size. The entire sample is redrawn annually, and a supplemental sample of recent business births is selected midway through the year. About one-fourth of the sample is rotated out each year and replaced with newly drawn units. The current design has been fully implemented since 2003 and follows modern design principles for an establishment survey. Firms from all sizes, industries, and States are included in the sample. For more information on the sample design see www.bls.gov/web/empsit/cestn2.htm.

The over-the-month change in nonfarm payroll employment is the primary measure of interest for most CES data users. BLS estimates the level of payroll employment with the CES survey; the over-the-month change is derived simply as the difference between the current and previous month's employment levels.

Why CES uses non-sample methods to account for business births and deaths - Although the CES sample is very large and follows standard design principles, it alone is not sufficient for estimating the total employment level because each month new firms generate employment that cannot be captured through the sample. There is an unavoidable lag between a firm opening for business and its appearance on the CES sample frame. The sample frame is built from Unemployment Insurance (UI) quarterly tax records. These records cover virtually all U.S. employers and include business births, but they only become available for updating the CES sampling frame 7-9 months after the reference month. After the births appear on the frame, there is also time required for sampling, contacting and soliciting cooperation from the firm, and verifying the initial data provided. In general, the CES can not sample and begin to collect data from new firms until they are at least a year old.

There is a parallel though somewhat different issue in capturing employment loss from business deaths through monthly sample collection. Businesses that have closed are less likely to respond to the survey and data collectors may not be able to ascertain until after the monthly collection period that firms have in fact gone out of business. As with business births, hard information on business deaths eventually becomes available from the lagged UI tax records.

Difficulty in capturing information from business birth and death units is not unique to the CES; virtually all current business surveys face these limitations. Unlike many surveys, CES adjusts for these limitations explicitly, using a statistical modeling technique. Other surveys that do not explicitly adjust for business births and deaths are implicitly using the continuing sample units to represent birth and death units. This approach is viable when the primary characteristic of interest is an average measure of some type. However, because the goal of the CES program is to estimate an employment total each month and business births and deaths are important components contributing to these totals, CES uses a model-based adjustment in conjunction with the sample. Without the birth/death model-based adjustment, the CES nonfarm payroll employment estimates would be considerably less accurate.

Rationale for the CES birth/death modeling technique - Prior to BLS adopting the current birth/death modeling technique, research using historical information indicated that the business birth and death portions of total employment were substantial, but the net contribution of, or the difference between, the two components was relatively small and stable. The research was done using the nearly complete counts of employment developed from the UI tax records that are tabulated under the BLS Quarterly Census of Employment and Wages (QCEW) (www.bls.gov/ore/pdf/st020090.pdf). These QCEW tabulations also form the basis for both the sample frame and annual benchmark for the CES program.

Beyond the research cited above, the Business Dynamics (BED) series published quarterly by BLS, also illustrate how business birth and death employment substantially offset each other. The BED series are also derived from the QCEW. The BED series demonstrate that most of the net employment change each quarter is generated by the expansions and contractions in employment of the continuing businesses and a relatively smaller piece from business openings and closings (which CES refers to as business births and deaths). As shown in the chart below, continuing businesses which are adding employees (expansions) or subtracting employees (contractions) over the quarter comprise the vast majority of total change; these movements are measured by the CES sample. Employment change contributions from openings (or births) and closings (or deaths) are much smaller and more stable, and the two series offset each other to a large degree. It is these underlying relationships among the components of net employment change that allow the CES to produce accurate estimates using a current monthly sample of continuing businesses and a model-based approach for the residual of net business births and deaths.

Business Employment Dynamics series, seasonally adjusted, 1997-2007 Total Private Employment in thousands

 

Description of the CES methodology for capturing net employment change from business deaths and births

The CES methodology has two steps.

Step One - Employment losses from business deaths are excluded from the sample in order to offset the missing employment gains from new business births. Because employment increases from births nearly offset employment decreases from deaths in most months (as illustrated above by the BED data), this step accounts for most of the net of business birth and death employment.

Operationally this is accomplished in the following manner each month. Business deaths that are non-respondents to the survey are automatically excluded because they have no current month data. Death establishments that report zero employment to the survey for the current month are treated the same as non-respondents and also excluded. As a result, the over-the-month change calculation from the sample is based solely on continuing businesses.

For the months subsequent to a business death, the deaths are "kept alive" in the CES estimation process; the growth rate of the continuing units in the sample is applied to them each month. This estimates for the growth of the new business births in the months after their birth but before they can be brought into the sample.

This step accounts for most of the net birth/death employment but not all of it. The residual net employment that is not captured by this step is estimated through an econometric model, described below as Step 2.

Step Two - Modeling for the residual of net/birth death employment change. In this step, the CES adjusts its sample-based estimates for the residual net birth/death employment that step 1 misses. This adjustment is derived from an econometric technique known as Auto Regressive Integrated Moving Average (ARIMA) modeling. ARIMA is a standard econometric modeling technique that is often used to estimate relatively stable series. CES refits the ARIMA models each year, for each basic estimation cell, as part of its annual benchmarking process.

The inputs to the ARIMA model are historical observations of the residual net birth/death employment that is not captured by either the sample or the step 1 imputation described above. These historical observations are derived empirically, from the most recent five years of QCEW historical data. From the QCEW universe employment series, CES classifies each establishment each month as a continuing unit, a birth, or a death. Then sample-based estimates are simulated using the month-to-month change of the continuing units, and using the deaths-to-impute-for-births technique described above in step 1. The difference between these simulated estimates and the actual total employment measured by the QCEW each month, is the residual net birth/death employment.

Five years of monthly observations of these net birth/death residuals are calculated for each estimation cell; they are then input to the individual cells' ARIMA models to produce a net birth/death residual for each cell that is used in the current monthly estimates.

The table below shows the actual residual net birth/death employment (column 3) for 2000-2007, as calculated from the QCEW universe in the manner described above. Comparing this residual to the overall net change in employment (column 2, also as measured by the QCEW) shows that it does not correlate closely with underlying employment growth, but is relatively stable.

Over the year total nonfarm and residual net birth/death employment from the QCEW

CES benchmark year	Total nonfarm employment over-the-year change in the QCEW, not seasonally adjusted (in thousands)	Actual residual net birth/death employment over-the-year change derived from the QCEW, not seasonally adjusted (in thousands), total private employment*
Mar 00-01	1163	735
Mar 01-02	-2017	607
Mar 02-03	-524	700
Mar 03-04	871	718
Mar 04-05	2019	726
Mar 05-06	2830	1122
Mar 06-07	1665	792

* The model is not used for government series.

 

Because the residual net birth/death employment component is relatively stable, the ratio of it to total employment change can vary substantially from year to year. In slower growth years (for example, March 03-March 04), the ratio is much different than in stronger growth years (for example March 04-March 05). The table also shows than even in a year where total nonfarm employment declines, the residual net birth/death employment component is positive (for example March 01-02). Put another way, the residual net birth death amount itself is relatively stable but its relationship to overall net employment change varies, depending on the magnitude of the overall change, almost by definition.

How Effective are the CES Methods for Measuring Net Business Birth/Death Employment?

Benchmark Revisions - On an annual basis BLS recalculates nearly two years of CES estimates in a process known as benchmarking. The benchmark process re-anchors the CES estimates to a nearly complete count of employment based on the Unemployment Insurance tax records tabulated through the QCEW. During the benchmark process the March CES estimate for a given year is replaced by the employment counts derived from the QCEW.

The benchmark process helps to correct for sampling and modeling error in the CES estimates. It provides a method of both validating and improving the CES employment series. If the birth/death estimator or any other aspect of the CES estimation process has sustained large statistical error over the course of a year, it will be corrected by the benchmarking process. In most years, the benchmark error, measured as the difference between the CES estimate for March and the final QCEW-based March employment level, is relatively small, indicating that the CES estimation process is producing accurate employment estimates. The benchmark error is generally used as a proxy for total CES estimation error although this interpretation is not entirely accurate, because there is statistical error in the QCEW as well as in the CES. Both data series are subject to non-response, imputation, reporting, and processing errors, which are common to all surveys and administrative records tabulations. However, because the QCEW is not subject to sampling error and provides a reliable source for business birth/death employment, the benchmarking process improves the CES employment series.

The table below gives a recent history of benchmark revisions or benchmark errors. They have ranged from 0.1 percent to 0.6 percent of the total nonfarm payroll employment level; the average is two-tenths of one percent since 2000, the year CES began phasing in a new sample design along with the birth/death modeling technique. Beginning with 2003, all industries were estimated using the new sample design and birth/death model.

CES total nonfarm benchmark revisions, recent years*, numbers in thousands

Benchmark Year	Benchmark revision	Percent benchmark revision
March 2000	468	0.4
March 2001	-123	-0.1
March 2002	-313	-0.2
March 2003	-122	-0.1
March 2004	203	0.2
March 2005	-158	-0.1
March 2006	752	0.6
March 2007	-293	-0.2
Average	52	0.1
Average Absolute	304	0.2

* CES began phasing in use of the birth/death model by industry beginning in March 2000; by March 2003 the model was used for all industry series.

 

How the net birth/death model reduces benchmark error - The table below shows that the CES birth/death model adjustment effectively reduces error in CES estimates. The table compares actual benchmark revisions to revisions which would have resulted if CES had not adjusted sample-based estimates with the residual birth/death model, for the March 2003 benchmark year forward. The March 2003 benchmark is the first in which all industries were estimated using the net birth/death model. As an example, for March 2003-2004, if there were no model-based adjustment, a benchmark revision of 838,000 would have occurred for the year; the incorporation of the modeled residual (635,000) reduced the error to 203,000. In every year, the birth/death adjustment reduced the error in the CES estimate of over-the-year change.

Simulated CES benchmark revisions if net birth/death adjustments not made; Numbers in thousands

Benchmark Year	Birth/death model amount	Actual benchmark revision	Simulated benchmark revision if birth/death adjustments not made
Mar 02-03	458	-122	366
Mar 03-04	635	203	838
Mar 04-05	830	-158	672
Mar 05-06	880	752	1632
Mar 06-07	1073	-293	780

 

Limitations of the residual net birth/death model

The current modeling technique consistently reduces error in the estimate of nonfarm payroll employment, as compared to making no adjustment, however it has limitations. The primary limitation stems from the fact that the model is, of necessity, based on historical data. If at some future time, there is a substantial departure from historical patterns of employment changes associated with the residual of net business births and deaths, the model's contribution to error reduction could erode.

Because there is no current monthly information available on business births, and because only incomplete sample data is available on business deaths, estimation of this component will always be potentially more problematic than estimation of change from continuing businesses.

Interpretation of the birth/death model adjustment relative to overall monthly change in payroll employment

The birth/death model component is added to the sampled-based component to form the not seasonally adjusted, employment estimate for each month, as described above. These employment estimates are subsequently seasonally adjusted. Seasonal adjustment smoothes the employment series by removing normal seasonal variations due to factors such as weather and holidays; therefore the seasonally adjusted over the month employment changes are generally much smaller than the unadjusted changes.

Users who wish to compare the model's contribution to overall employment change reported for a month need to compare against the unadjusted estimates, not the seasonally adjusted series. Comparing the model amounts to seasonally adjusted estimates generally results in an overstatement of the model-based component's contribution to over-the-month employment change.

The birth/death model component generally shows the same overall seasonal patterns as the sample-based component. For example, total nonfarm employment shows a large seasonal increase in employment each April; the model also shows a relatively large net addition to employment each April. Similarly total nonfarm employment records a large drop in employment each January and the model estimates a substantial drop in net birth/death employment each January. An example of the net birth/death model components versus overall net employment change for March 2006 to March 2007 (prior to the March 2007 benchmark implementation) is shown below. The April model amount of 207,000 should be viewed as a component of the 934,000 not seasonally adjusted employment change, rather than as a component of the 144,000 seasonally adjusted change.

Birth/death model adjustment and over the month change in total nonfarm employment, in thousands, April 2006-March 2007

	Apr	May	Jun	Jul	Aug	Sep	Oct	Nov	Dec	Jan	Feb	Mar
Model amount	207	192	176	-71	127	50	57	22	63	-192	116	133
Not seasonally adjusted total change	934	827	516	-1139	225	675	737	409	-93	-2770	715	922
Seasonally adjusted total change	144	103	124	222	186	198	109	196	226	162	90	175

 

Comparing the net birth/death adjustment to the previously-used bias adjustment

The CES program has always included a model-based adjustment to adjust its sample-based results because the lag between new businesses opening and their appearing on the sample frame is unavoidable and intractable.

Prior to the year 2000, CES used "bias adjustment factors" to adjust its sample-based estimates each month. BLS began using the birth/death model in 2000 for one industry (wholesale trade) and gradually expanded its use to all industries, concurrent with the phase-in of a new sample design and estimation techniques over the 2000-2003 time period. (www.bls.gov/opub/mlr/2006/05/art4full.pdf)

The redesign replaced an outmoded quota sample design with a more modern and technically sound probability-based design. It also introduced net business birth/death modeling as a replacement for the less precise bias adjustment factors, as the technique for estimating employment change in components not measurable from the monthly sample.

There are major conceptual and methodological differences between bias adjustment and net business birth/death modeling. Although one primary purpose of bias adjustment was to account for employment from business births, it also attempted to adjust for other elements of both sampling and non-sampling error in the estimates. This is the case because the major input to the model was historical total estimation error, as measured by the difference between purely sample-based estimates and UI universe employment counts (benchmarks). In contrast, the net birth/death model estimates are more targeted. They adjust only for the residual component not measurable from the sample; the birth/death model does not attempt to adjust for other sampling or non-sampling errors. Because of the replacement of the old quota sample design with the probability design, the potential for substantial error from these other sources was greatly diminished, and no general "bias adjustment" is applied.

Summary

The net birth/death model is used by the CES program in order to produce a comprehensive estimate of total payroll employment on a very timely basis each month. The model estimates a residual net business birth/death employment contribution that is not measurable by the sample, due to the unavoidable lag between a firm's opening for business and its appearing on the BLS sampling frame.

BLS uses a relatively simple modeling technique to estimate the net contribution of business births and deaths because the historical data on birth/death employment contributions, as measured by the UI tax records, indicates that this method is appropriate. The modeling improves the accuracy of the CES estimates by reducing the annual benchmark revision.

BLS continues research on business birth/death estimation for possible enhancements to the current modeling technique. However, there are no changes to the current modeling technique scheduled at this time.

 

Last Modified Date: May 15, 2009