The LAUS program uses time-series models to estimate the true values of household employment and unemployment from the state Current Population Survey (CPS) series, which contain sampling error. The estimates consist of trend and seasonal components that are summed to produce the not seasonally adjusted data series.
Each trend model consists of the trend of a complementary series as a regression variable with a fixed coefficient, as well as a residual that accounts for variation in the CPS trend that is unexplained by the regression variable. The trend models for unemployment estimation use counts of continued claimants from the regular state unemployment insurance program as covariates, while the trend models for household employment estimation use estimates of total nonfarm payroll jobs from the Current Employment Statistics (CES) survey as covariates.
The model-based estimates are benchmarked using a two-stage hierarchical scheme where an internal constraint is imposed at each stage on the estimation process. In the first stage, model estimates for the nine census divisions are benchmarked to sum to the monthly not seasonally adjusted national employment and unemployment estimates from the CPS. In the second stage, the state model estimates within each census division are constrained to their corresponding first-stage census division benchmarks. These two steps taken together ensure that the sum of the model-based estimates across all states equals the corresponding national CPS estimate. For current-year estimates, this process, called real-time hierarchical benchmarking, makes the models more robust to sudden shocks to the overall economy as they occur.
To produce seasonally adjusted estimates, the LAUS program applies an X-11 type of seasonal adjustment filter to the benchmarked not seasonally adjusted estimates to reduce distortions to seasonality after benchmarking and then further smooths out variability introduced by the benchmarking process with a trend-type filter. Note that use of the smoothing filter had been suspended effective with revised estimation for April 2020, in order to avoid artificial dampening of outlier effects attributable to the coronavirus (COVID-19) pandemic.
Because of the unusually large magnitude of pandemic-induced breaks in the Current Population Survey (CPS) and covariate series, it was necessary to perform real-time identification and estimation of outlier effects by adding level-shift dummy variables to the models to prevent serious distortions to their trend and seasonal components.
Accounting for outlier effects in the models is complicated by the fact that there are two sources of level shifts in each model. The first is the level-shift regression variable added directly to the model, and the second is a level shift in the covariate that is transmitted to the CPS model via the covariate-trend regression component. In the past, the latter was not an important source of level shifts, but it has become the dominant source with the onset of the pandemic. Unfortunately, the covariate trend relationship has become unstable, too.
This is illustrated with the unemployment model. Historically, unemployment insurance (UI) continued claims counts have been substantially lower than the total unemployment estimates from the CPS, because UI programs covered primarily experienced workers on layoff who satisfied relatively stringent eligibility requirements for benefits, such as having qualifying earnings in recent quarters (i.e., having paid into the system) and actively seeking work if there is no expectation of recall. This category of unemployed workers receiving UI benefits historically accounted for one-third to no more than half of total unemployment. As a result, the UI regression coefficients historically were greater than one. Prior to the pandemic, level shifts in UI continued claims series tended to be small in magnitude relative to total unemployment and had little effect on the unemployment models.
Things changed dramatically in April 2020, when the full impact of COVID-19 hit state labor markets. Unemployment insurance continued claims rose dramatically above their previous historical levels and, in many states, even exceeded the CPS measure of total unemployment. During this time, many state workforce agencies relaxed their requirements for receipt of benefits, which may have resulted in some people receiving benefits who would not have been classified as unemployed using the activity-based criteria of the CPS.
Because level shifts in the continued claims series are directly transmitted to the unemployment model, there is a potential for overestimating the COVID-19 effect at the start of the pandemic. Later in 2020, workers began exhausting their UI benefits while still unemployed. At that point, the bias in the regression component of the trend tended to switch to underestimating COVID-19 effects.
In the past, sudden breaks in the Current Population Survey (CPS) series were infrequent and easily could be controlled using level-shift variables specific to the CPS. Moreover, level shifts in the covariates were small in magnitude and largely independent of the CPS. While outlier detection and estimation are normal parts of estimation, COVID-19 presents an unprecedented challenge due to its magnitude and scope. The regression coefficients showed some flexibility in real time between the CPS and covariate inputs, but not enough to prevent bias due to the influence of past data. Other sources of model flexibility exist: the residual trend tends to offset the bias in the regression component of the trend, and the addition of new level shifts specific to the CPS also helps to reduce bias.
The current way the models adapt to the ongoing effects of the coronavirus pandemic provides short-run, ad hoc means for countering bias generated by the model specification of the regression component of the trend. What is needed is a re-specification of the trend to provide a more flexible relationship between the CPS and covariate trends and allows for a more direct and simple approach to accounting for structural breaks.
Specifically, the trend relationship will be re-specified in bivariate form where the two trends are jointly related through their corresponding disturbance terms. Each trend depends upon its previous value, plus a random disturbance term that is independent of its own past values. At a given point in time, the values of these trend disturbances are mutually correlated. The correlation coefficient varies between zero and one. A value of zero indicates no relation between the two trend disturbances; a value of one indicates a perfect relationship, which is a special case of a fixed regression coefficient where the two trends are proportional to each other. Within these two extremes lies a wide range of possibilities. Structural breaks in either series are modeled as independent external effects estimated by suitably specified outlier regression variables. In this way, the bivariate trend model allows for a wide range of possible relationships between the two trends and simplifies the modeling of structural breaks.
The LAUS program also will simplify the real-time benchmarking procedure, which constrains the sum of states’ not seasonally adjusted model estimates of employment or unemployment for a given month to equal the national Current Population Survey (CPS) estimate of the same labor force component. Currently, benchmarking is done internally in the estimation process, which has a number of theoretical advantages in variance estimation but requires combining all state models within a census division into a single, large multivariate model in which a stochastic constraint is imposed. Because of the complexity of incorporating bivariate models into these large multivariate models, a simpler external ratio adjustment was tested as an alternative. It was decided to switch to the ratio-adjustment approach, since there was little empirical difference between the two approaches.
No additional substantive changes to the LAUS modeling procedures are planned. However, data users may want to note that effective with revised estimation for April 2020, use of the smoothing filter in conjunction with seasonal adjustment had been suspended. Use of the smoothing filter will resume for the seasonally-adjusted series upon implementation of the modeling changes. Information on the seasonal adjustment and smoothing procedures for model-based estimates is available at https://www.bls.gov/lau/ssachanges2018.htm.
The LAUS program will implement the new estimation procedures for model-based areas in early 2021, effective with the Regional and State Unemployment 2020 Annual Averages news release scheduled for issuance on March 3. Data using the new procedures will replace all published historical data. These methodological changes, therefore, will not create discontinuities in any time series.
Last Modified Date: March 3, 2021