Employment Estimating Methods Evolving
By Arthur Famiglietti, Research Analyst

The Way We Were

When current estimates of the economy's employment levels became a topic of more than casual interest, a method was devised to develop timely estimates in a cost effective manner. The Current Employment Statistics program, first authorized by a Congressional Act in 1930 as a federal-state cooperative survey, became the source for current estimates of the nation's nonfarm employment, hours and earnings based on a quota sample methodology, the standard in its time. Part of today's difficulties with that methodology is that it was developed when the structure of America and its economy were different. Another is that the existing CES survey is a quota sample whose inception predated the introduction of probability sampling as the internationally recognized standard for survey sampling.

The Way I Always Heard It Should Be

Over time, criticism has been made regarding the CES time series estimates, not the least of which centered on accuracy and revisions. The quota sample design is known to be at risk for potentially significant biases. Additionally, the production of sampling errors and confidence intervals, standard survey accuracy measures, are not possible under the current design. Accuracy assessment came only after the fact via the quantity of change experienced for the primary estimate of interest, the total nonfarm employment level when benchmarked to employment data gathered for the unemployment insurance (UI) program.

Here Comes the Judge

At the request of the Bureau of Labor Statistics, in the fall of 1992 the American Statistical Society began to examine the CES program. In November 1993 the panel members came to agreement on five important problems to which they recommended BLS pay early attention. They were: needs of users, probability sampling, births and deaths of establishments, quality measurement and improvement of the ES-202, and methodological documentation and dissemination of research. On the topic of probability sampling the ASA stated "The BLS should adopt immediately the goal of moving to probability samples both for national estimates and for the state data. Probability sampling should assist in reducing bias in the employment estimates and should assist in reducing bias in the hours and earnings estimates, which are not benchmarked. Research will be required on some aspects of probability sampling and on how best to introduce the probability estimates into the current series." The progression toward that end will form the basis of the following discussion.

In June of 1995, the BLS began a comprehensive sample redesign of its monthly payroll survey. Three goals of the redesign are: development of a probability-based sample design to replace the current quota sample along with improved estimators; improved techniques for estimating employment change from business births and deaths; and improved sample solicitation procedures designed to achieve higher response rates than historically experienced. The initial research phase for the CES sample redesign was completed in 1997, and the Bureau launched a production test of the new sample design at that time.

The implementation phase began in June 2000 when the first national estimates from the new design for the wholesale trade industry were published with the 1999 benchmark. The remaining industry divisions will be phased in with subsequent years' benchmark releases, between 2001 and 2003. (See table below.) In Connecticut the first estimates based on the new sample design will commence in February 2001. Estimates will be produced for wholesale trade at that time on a statewide basis, but not all areas in Connecticut will have adequate sample coverage under the new probability design to provide estimates using that method.

The new CES Sample Design is a stratified, simple random sample, where the strata, or sub-populations, are defined by state, industry, and employment size. The sampling rates for each stratum are determined through a method known as optimum allocation, which distributes a fixed number of sample units across a set of strata to minimize the overall variance, or sampling error, on the primary estimate of interest. The total nonfarm employment level is the primary estimate of interest, and the new design gives top priority to measuring it as precisely as possible; in other words, minimizing the statistical error around the statewide total nonfarm employment estimates.

Total sample sizes for the states are to remain at current levels, i.e., not to change in quantity of respondents from the old quota sample to the new. Therefore optimum allocation is achieved for a set sample size using a program which seeks to optimally select a sample by drawing it across a state's industry and employment size makeup. This achieves an optimum allocation for the sample size that has been chosen, not the optimum level that could be sampled to achieve a desired variance level. Only time will tell how well the new sample method will estimate, and while standard survey accuracy measurements will be produced, overall performance evaluation will continue to be gauged by the level of benchmark revision.

Another change of significance being introduced with the probability sample involves establishment of birth and death modeling. This effort has been undertaken to address the issue of sample bias resulting from the non-capture of employment attributable to establishment openings. The most timely UI universe files available always will be a minimum of six months out of date. Exploratory research indicated that while both the business birth and death portions of total employment are generally significant, their net contribution is relatively small and stable. To account for this net birth/death portion of total employment, BLS is implementing an estimation procedure with two components. The first component uses business deaths to impute employment for business births. The second component is an ARIMA (AutoRegressive Integrated Moving Average) time series model designed to estimate the residual net birth/death employment not accounted for by imputation.

Half Breed

Starting in 2001 data users of the industrial nonfarm employment estimates will be faced with a changing data series. The new probability method will be phased in resulting in total nonfarm employment estimates based on a combination of old and new methods. The CES method of producing total nonfarm employment is a bottom up system of aggregation. Each industry for which an estimate is made is independently estimated and then aggregated to division and summary levels. This will be the first time that a mixed sample method will be used in the production of estimates.

The sample under the probability methodology will not consist of entirely new reporters. Establishments with employment of 100 or greater will be retained as grandfathered units in the sample. The new probability sample will assign weights to reporting units that are equal to the inverse of their probability of being selected. This will result in small size reporters having a much greater impact on the algorithm calculated to produce the monthly estimate. The sample will continue to operate on a matched sample concept. That means that in order for a report to be used in the current monthly estimate its previous monthly report must have been received. Under this method grandfathered units will have a weight of 1.000, thus representing only themselves in the estimate.

Under redesign, industries in MSAs that do not have 20 sampled unemployment insurance accounts and 10 responding unemployment insurance accounts will be excluded from the probability sample estimation. The criteria exclude UI accounts grandfathered into the probability sample, i.e. current reporters not selected under the probability sample but retained because of their size.

Presently, the methods employed to produce estimates are the same for all sub-state areas and the statewide estimate in Connecticut. This will change with the 2000 benchmark as the Danbury, New London-Norwich, and Waterbury Metropolitan Statistical Areas (MSAs) do not have sufficient probability sample to implement wholesale trade redesign estimates. The very small areas of Lower River and Danielson will continue to use the current method.

A source of question surrounding CES estimates involves the non-additivity of disaggregated geographies. Some data users are confused by the fact that summing the fifty states will not produce the national estimate. Similarly the addition of MSAs does not yield a statewide estimate. The new probability design will not eliminate this issue. The estimates produced for each geographic grouping will continue to be independently estimated and benchmarked. A number of factors bear upon this issue, including the method by which UI covered employment is assigned to an area.

Clearly, extensive research over the past two to three years has proved that, in some cases, proposed changes in CES methodology would not produce improved results. The methods which will be used to select the CES sample and produce the estimates are evolving, yet some elements of the existing methodology will remain. Standard survey measurement statistics will be produced, but the ultimate gauge of accuracy will continue to be based on the extent of annual benchmark revisions. A new sample will be drawn using a random sampling method, while self-representing, non-probability units will remain in the system. The goal of additivity across geographies will not be achieved. All areas will not move to the new sample method and benchmark employment issues will continue to exist. In the end, this incremental advance will be a work in progress for some time to come. To our data users I would say, expect improvement, not perfection. After all, evolution is a lengthy process and a journey we will make together.

Return to Top

Industry Clusters
State Acclaimed

Three recent indicators demonstrate Connecticut continues to stay in the forefront nationally. The highest average annual pay in the nation, improved income distribution, and strongly competitive existing businesses earned Connecticut "Honor Roll" status in the 2000 Development Report Card for the States by the non-profit Corporation for Enterprise Development (CFED). In one of the most broad-based ratings of the 50 states, with more than 70 indicators, Connecticut made "Honor Roll" (scoring an A or B in each index) by earning an "A" in Performance, a "B" in Business Vitality, and an "A" in Development Capacity.

The Milken Institute, a Santa Monica based research center, ranked Connecticut the third best state in the nation for growth opportunities in the "new economy." Massachusetts, California, Connecticut, Colorado and Washington were the five top-ranking states based upon 12 criteria critical to future high-tech growth, including research and development dollars, advanced degrees, patents, venture capital investment, business starts, and initial public offerings.

Connecticut also received the 2000 National Alliance of Business Distinguished Performance Award for "State of the Year." Governor Rowland accepted the award at a ceremony in Washington, D.C. October 31st. Connecticut was recognized for its innovative efforts to raise student achievement and improve workforce quality.

Return to Top

Housing Update
New Permits Up 3.9% From A Year Ago

Commissioner James F. Abromaitis of the Connecticut Department of Economic and Community Development announced that Connecticut communities authorized 776 new housing units in October 2000, a 3.9 percent increase compared to October of 1999 when 747 units were authorized.

The Department further indicated that the 776 units permitted in October 2000 represent an increase of 3.3 percent from the 751 units permitted in September 2000. The year-to-date permits are down 14.0 percent, from 9,123 through October 1999, to 7,850 through October 2000.

Fairfield County documented the largest number of new, authorized units in October with 165. New Haven County followed with 162 units and Hartford County had 145 units. Danbury led all Connecticut communities with 50 units, followed by Wallingford with 23 and Berlin and Tolland tied with 18.

Return to Top

Leading Index Recovers: Coincident Index Reaches Prior Peak

Last month, we reported that the leading index, a barometer of future employment activity, had declined during four of the previous five months, suggesting that some uncertainty surrounded the continued expansion over the next year or so. The release of (preliminary) September data added to the uncertainty as the leading index rose significantly. In addition, the coincident index, a gauge of current employment activity, returned to its previous peak in June 2000 with the September release. As noted last month, the current expansion continues to roll along.

So what's it all mean? At the national level, the current 10-year expansion, the longest on record in the United States, shows little evidence of ending. Analysts argue about whether Chairman Alan Greenspan and the Federal Reserve might again raise interest rates to head off an outbreak of the dreaded inflation disease. Those concerns have diminished as recent reports on the economy suggest a slowdown from its recent rapid growth. The third quarter growth in real GDP slowed to only 2.7 percent - still a healthy rate by the standards of the last 20 years or so, but dramatically below the second quarter growth of 5.6 percent. Moreover, consumer confidence fell recently and retail sales saw slower growth. Thus, the best current thinking thinks that the Federal Reserve will not raise interest rates for the rest of this year.

Other analysts of the national economy see the third quarter slowdown as only a temporary blip on the economic radar that will disappear (reverse) in future quarters. We have already seen quarterly growth rates slow on several occasions only to about face in subsequent quarters. The big question: What is the sustainable level of non-inflationary real growth? New and old economy pundits disagree. For economic historians, it is no big deal. The passage of time will answer the question. For most of the rest of us, we need to make decisions based on our outlook. We cannot wait for the real-time data to become historical information.

In summary, the coincident employment index rose from 97.6 in September 1999 to 103.5 in September 2000. All four components of the index point in a positive direction on a year-overyear basis with higher nonfarm employment, higher total employment, a lower total unemployment rate, and a lower insured unemployment rate.

The leading employment index rose from 88.2 in September 1999 to 89.2 in September 2000. Three index components sent positive signals on a year-over-year basis with a lower short-duration (less than 15 weeks) unemployment rate, lower initial claims for unemployment insurance, and a higher average workweek of manufacturing production workers. The remaining two components sent negative signals on a year-overyear basis with lower total housing permits and lower Hartford help wanted advertising.

SOURCE: Connecticut Center for or Economic Analysis, University of Connecticut. Developed by Pami Dua [Economic Cycle Research Institute; NY,NY] and Stephen M. Miller [(860) 486-3853, Storrs Campus]. Stan McMillen and Jingqui Zhu [(860) 486- 3022, Storrs Campus] provided research support.

Return to Top