Concurrent Seasonal Adjustment for Industry Employment Statistics
The Current Employment Statistics (CES) Survey, conducted monthly by the
Bureau of Labor Statistics, obtains payroll employment, hours, and earnings from
business establishments and produces industry-based estimates. Accurate seasonal
adjustment is an important component in the usefulness of these monthly data.
The CES program will convert to the North American Industry Classification System
(NAICS) with the publication of May 2003 first preliminary estimates, published
in June 2003. Simultaneous with this conversion CES will replace the current
projected-factor seasonal adjustment methodology with concurrent seasonal
adjustment. This paper compares the two seasonal adjustment methodologies,
examines results from recent research evaluating the two methods, and discusses
the implications of a conversion to concurrent seasonal adjustment.
Introduction: The CES is a monthly survey of over 300,000 business
establishments. The national CES estimates of employment, hours, and earnings
are some of the most timely and sensitive economic indicators published by the
federal government. They are widely viewed as a key measure of the health of the
economy and are closely tracked by both public and private policy makers alike.
Most CES data users are interested in the seasonally adjusted over-the-month
employment changes as a primary measure of overall national economic trends.
While seasonally adjusted series go through several monthly revisions and an
annual benchmark revision before they are finalized, the first published
estimates are the most widely anticipated and analyzed. Thus it is important to
use the most efficient and reliable methods for seasonal adjustment of current
months' data. Currently, CES uses seasonal adjustment methodology that applies
forecasted seasonal factors to the employment estimate. Twice a year seasonal
factors are forecasted for 6 months into the future and applied to the not
seasonally adjusted estimates during the subsequent 6 months. Beginning in
2003, CES will convert to concurrent seasonal adjustment in which new seasonal
factors are calculated each month using all relevant data, up to and including
the current month.
Background on CES Estimates: One of the benefits of CES is the
timeliness of the estimates. CES estimates are published each month after only 2
½ weeks of data collection. The primary deadline for data receipts, referred to
as "first closing," is the last Friday of the reference month, and
preliminary estimates are published on the first Friday following the reference
month. In order to incorporate additional sample received after the primary
deadline, each estimate undergoes two monthly revisions before being finalized.
The secondary cut-off, or "second closing," is usually 3 weeks
after the primary deadline, and the third deadline, or "third closing,"
is 3 weeks after the second. Therefore, for any given reference month,
second closing estimates are published the following month, and third closing
estimates are published 2 months subsequent.
CES estimates also undergo annual revisions called "benchmarks."
Each year, the sample-based estimates for the previous year are adjusted to
universe employment counts derived from State unemployment insurance tax
records. This constitutes the final estimate for all reference months in the
benchmark period.
To seasonally adjust the estimates, CES uses X-12 ARIMA software developed by
the US Census Bureau. Seasonal adjustment factors are recalculated
semi-annually, in April and November, and projected factors are published in
advance for the next 6 months. Currently, new seasonal factors are published in
June and December of each year. The June CES publication incorporates annual
benchmark revisions that include recalculation of seasonally adjusted data for
the most recent 5 years. After 5 years of seasonal adjustment revisions, figures
are frozen. For example, the March 2001 benchmark revision, published in June
2002, provided revised seasonally adjusted data for 1997 through the first
quarter of 2002.
Research approach: During the past several years, BLS has been
researching the impact that a change in seasonal adjustment methodology would
have on both the CES data itself and on users. Each month, parallel to the
monthly production of CES seasonally adjusted data using projected-factor
methodology, CES runs concurrent seasonal adjustment for research purposes. The
parallel tests are structured in a way to measure only the effect of
incorporating additional months of data into the seasonal adjustment process.
Current CES standard practice requires 10 years of historical data to be
used as input to the X-12 ARIMA model. The same historical data set was used for
the experimental concurrent run. Therefore, any prior adjustments originally
made to the data during production, such as to account for strikes or editing
and screening, are included in the concurrent simulations as well. The only
difference in inputs between the two runs is that concurrent adjustment also
incorporates up to 5 months of additional estimates when calculating the
seasonally adjusted data.
In the parallel series, incorporation of revised seasonal factors is handled
within the normal CES monthly revisions procedures. With the calculation of
first closing estimates for a current month, the second closing and third
closing estimates for the prior 2 months are revised on an unadjusted basis to
incorporate further sample receipts. Likewise, the concurrent seasonally
adjusted data are recalculated using revised second closing and third closing
estimates, mirroring the current production process. Finally, all published data
types are seasonally adjusted under both methods; however, because the all
employee series is the most closely watched series published by CES, it is the
focus of this paper.
Results: The two methods are compared in terms of (1) mean absolute
revisions to the over-the-month changes evident from first preliminary estimate
to the benchmarked series, (2) the variation between monthly revisions, and
(3) the smoothness of the seasonally adjusted series. Looking first at the
smoothness of the series, Figure 1 compares the third closing over-the-month changes of the
seasonally adjusted employment figures for Total Nonfarm from January 2001 to
June 2002 for the two methodologies. The dashed line shows the published
over-the-month changes for third closing, while the solid line shows the third
closing over-the-month change for the experimental series (i.e., what the
over-the-month change would have been if CES had been using concurrent seasonal
adjustment). As the graph illustrates, concurrent adjustment produces a slightly
smoother seasonally adjusted series with less variability in the over-the-month
changes.
Figure 1. Third Closing Over-the-Month Change, Total
Nonfarm, January 2001 through June 2002
Table 1 underscores the smoothness of the concurrent seasonally adjusted
series for Total Nonfarm plus all nine industry divisions. The smoothness ratio
shown in column B of Table 1 is a comparison measure of variability in the third
closing over-the-month change of the seasonally adjusted estimate. The
calculation compares the sum of the squared over-the month changes in the
concurrently adjusted series to the sum of the squared over-the-month changes in
the projected-factor adjusted series. A smoothness ratio below 1 indicates that
concurrent seasonal adjustment has less variability in the over-the-month
changes than does a series adjusted using projected seasonal factors. As Table 1
illustrates, concurrent adjustment produces a smoother seasonally adjusted
series for Total Nonfarm plus all nine industry divisions. Taken with the results from
Figure 1, this indicates that CES will benefit from a switch to concurrent seasonal adjustment
by producing employment series with less variability in the over-the-month changes.
Table 1. Smoothness Ratio, January 2001
through June 2002
(A) Group |
(B) Smoothness Ratio (Third Closing) |
Total Nonfarm |
.67 |
Mining |
.77 |
Construction |
.47 |
Manufacturing |
.87 |
Transportation and Public Utilities |
.78 |
Wholesale Trade |
.88 |
Retail Trade |
.56 |
Finance, Insurance, and Real Estate |
.68 |
Services |
.58 |
Government |
.67 |
Results to this point focused solely on estimates of seasonally adjusted
over-the-month changes in employment. Also of interest is the revision to the
estimate of the seasonally adjusted over-the-month change, both from first
closing to the final benchmarked series, and between monthly closings.
Table 2 illustrates the size of the mean absolute revision to the
over-the-month change from first preliminary to the final benchmarked series for
all nine major industry divisions and their topside aggregate, total nonfarm.
Column B shows the mean absolute revision in the over-the-month change for the
6-month projected method for March 1998 through March 2001, while column C shows
the same for the concurrent adjustment method. Column D shows the difference
between the two methodologies (concurrent minus 6-month projected). As the table
illustrates, CES employment estimates seasonally adjusted under the concurrent
method have a smaller revision from first closing estimates to final benchmarked
series in eight of the nine industry divisions plus total nonfarm. In Wholesale Trade,
the revision statistic was larger for concurrent adjustment, but only
by 0.2 percent.
Table 2. Mean Absolute Revision in Over-the-Month Changes,
March 1998 through March 2001
(A)
Group |
(B) CES Published Series (6-month Projected) |
(C) Experimental Series (Concurrent) |
(D)
Difference |
Total Nonfarm |
77,973 |
64,973 |
-13,000 |
Mining |
1,892 |
1,865 |
-27 |
Construction |
22,892 |
17,838 |
-5,054 |
Manufacturing |
13,757 |
12,487 |
-1,270 |
Transportation and Public Utilities |
7,892 |
6,568 |
-1,324 |
Wholesale Trade |
11,135 |
11,162 |
27 |
Retail Trade |
32,162 |
21,946 |
-10,216 |
Finance, Insurance, and Real Estate |
6,919 |
5,703 |
-1,216 |
Services |
38,784 |
29,703 |
-9,081 |
Government |
23,135 |
17,432 |
-5,703 |
In addition to a smaller revision between first closing and the final
benchmarked series, revisions in the over-the-month changes between closings
are of concern as well. In particular, there is the potential for these monthly
revisions between closings to increase under concurrent adjustment because the
seasonal factors can change with each iteration of the monthly adjustment
process. However, results indicate that, in addition to a smaller revision
between first closing and the final benchmarked series, concurrent seasonal
adjustment leads to equal or even less variability in the over-the-month changes
between closings.
Figure 2 shows the revision to the over-the-month change between seasonally
adjusted first closing and second closing total nonfarm estimates under both
methods. The graph illustrates that, in general, the concurrently adjusted
series shows slightly less variability in the seasonally adjusted over-the-month
changes between revisions. Results were very similar for revisions between first
closing and third closing.
Figure 2. Over-the-month changes between Revisions, 1st
Closing to 2nd Closing Seasonally Adjusted Total Nonfarm All Employees Series,
March 1998 - March 2002
Likewise, Table 3 illustrates a comparison of revisions between closings for
the currently published CES series and the same series adjusted concurrently. As
the table shows, the mean revision and mean absolute revision in the
over-the-month change does not differ between first closing and second closing
across the two methods. However, from first closing to third closing, both the
mean revision and mean absolute revision are lower in the concurrently adjusted
series. These results, when combined with results shown in Figure 2, suggest that
concurrent seasonal adjustment will not increase revisions between closings.
Table 3. Mean and Mean Absolute Revisions in Over-the-Month
Changes, Total Nonfarm All Employees Series, March 1998 through March 2002
Type |
CES Published Series (6-month Projected) |
Experimental Series (Concurrent) |
Difference |
First Closing to Second Closing |
Mean Revision |
-4 |
-7 |
3 |
Mean Absolute Revision |
37 |
34 |
-3 |
First Closing to Third Closing |
Mean Revision |
19 |
4 |
-15 |
Mean Absolute Revision |
48 |
36 |
-12 |
Summary of Advantages and Disadvantages of Concurrent Seasonal Adjustment
Advantages
More accurate seasonal factors - Concurrent seasonal adjustment is
technically superior to the 6-month forecasted factors because it takes into
account the timeliest information available. Empirical results from this
analysis illustrate that seasonally adjusted CES data are closer to the final
benchmarked series under concurrent adjustment, leading to smaller revisions
between first primary estimates and the final benchmark series. Furthermore,
monthly revisions between first closing and third closing are slightly lower for
concurrent adjustment.
Conversion to NAICS - Using concurrent seasonal adjustment will be
especially advantageous during the first few years following the CES conversion
to NAICS because most of the NAICS historical data will be reconstructed from
the SIC-based sample. Only 2 years of NAICS history from a NAICS-based sample
will be available. Therefore, under the projected-factor method, in the first
year of the NAICS conversion, there would be only two historical NAICS-based
estimates per month used to calculate projected seasonal factors, while with
concurrent adjustment three actual NAICS-based estimates would be used (the
previous 2 years of NAICS-based estimates plus the current one). The
additional observations will be valuable because X-12 weights the most recent
years more heavily than the past when calculating seasonal factors.
Familiarity with revisions - As discussed earlier, CES already revises
2 prior months of estimates with each month's release. As part of the current
monthly production process, not-seasonally adjusted estimates are revised for
the previous 2 months, and projected seasonal factors are applied to the
revised estimate to calculate the new seasonally adjusted figures. No additional
revisions would occur under concurrent seasonal adjustment. Under concurrent
adjustment, the not-seasonally adjusted estimate for the previous 2 months
would still be revised as before, and the seasonally adjusted data for these
months will be based on these revisions.
Potential Disadvantage
Factors will not be available ahead of time - As discussed earlier,
CES traditionally calculates seasonal factors twice a year and projected factors
are published in advance for the next 6 months. Under concurrent seasonal
adjustment, CES will not publish factors in advance because the new seasonal
factors are calculated each month. However, it is possible to make available
beforehand the ARIMA model specifications used by BLS so that the seasonal
adjustment run can be replicated if desired.
Summary and Implementation Plans: The research done with the National
CES employment series indicates that the CES will benefit from conversion to
concurrent adjustment through smaller revisions to the over-the-month changes
from the first closing estimates to the final benchmarked estimate. Furthermore,
it shows that concurrent adjustment would not increase revisions between
closings, and would actually reduce revisions from first closing to third
closing. Based on these results, simultaneous with the program's conversion to
NAICS industry coding in June 2003, CES will switch to concurrent seasonal
adjustment methodology. At that time, the practice of publishing forecasted
seasonal factors will be discontinued.
Last Modified Date: October 30, 2002