In Friday’s post covering the March report on Arkansas employment and unemployment, we noted an anomalous implication of the data used to calculate the unemployment rate: With the number of unemployed going up by 18,526 and the number of employed rising slightly (+530), the labor force expanded by over 19,000. This change is far larger than any historical precedent, runs counter to the labor force change at the national level, and makes little economic sense.
Several factors appear to have contributed to this result, presenting a number of questions. What does it mean for interpreting the unemployment rate, measured by dividing the number unemployed by the labor force? And for that matter, how should the underlying statistics on the number employed and unemployed by interpreted and understood?
The Arkansas Division of Workforce services has been very helpful in our examination of the data for March. Their primary explanation for the anomaly is that the BEA methodology led to some double-counting. According to spokesperson Zoë Calkins, “March was truly a unique situation, were workers were counted as employed (because they were drawing a paycheck) even if they were not on the job. Many of those same workers were also eligible to and did apply for Unemployment Insurance benefits. So under BLS definition, they were counted as both employed and unemployed.”
There are several factors that might have contributed to the the double counting, but the first matter to consider is how much it matters. Suppose that all of the newly unemployed were double-counted. Subtracting 18,526 from the reported number of employed implies an increase in the labor force is only 513. The direction is unexpected, but well-within the range of past magnitudes of change. More importantly, the impact on the unemployment rate is insignificant. In fact, rounded to the nearest tenth, the adjustment does not change the conclusion that the Arkansas unemployment rate was 4.8% in March. Moreover, confidence intervals around monthly unemployment rate estimates for Arkansas are generally in the magnitude of +/-0.5% in a normal month, and there is clearly more uncertainty for this unusual month.
So at least on one level, the answer to the question posed in the title of this article is: 4.8%.
Nevertheless, the processes used to estimate U.S. and Arkansas unemployment rates for March (and the problems encountered in the process) make for an interesting story that helps shed light on the interpretation of the March statistics. Our analysis shows how rapidly and severely the COVID-19 crisis has affected the economy, both nationally and here in Arkansas.
The U.S. Unemployment Rate
At the national level, the unemployment rate is estimated using the Current Population Survey (CPS), a monthly survey of households conducted by the Bureau of Census for the Bureau of Labor Statistics. It asks survey respondents to describe their employment status during a “survey reference week” which was, for this particular report, the week of March 8 -14.
The BLS news release for the U.S. unemployment statistics noted that the reference week survey “predated many coronavirus-related business and school closures that occurred in the second half of the month.” In that respect, the estimate of a nationwide increase in the unemployment rate from 3.5% to 4.4% was somewhat surprising in the suddenness of its impact.
But there were problems with the March CPS survey. The BLS news release noted that “in-person interviews were suspended, and that survey response rates were down 10 percentage points from recent months. There were also problems with the classification those affected by COVID-19 related closures: Some workers who should have been classified as ‘unemployed on temporary layoff’ were instead categorized as ’employed but absent from work.’”
The mis-classification in the survey would tend to underestimate the unemployment rate. In a special FAQ release accompanying the national employment report, the BLS considered one scenario in which 1.4 million workers considered not at work for “other reasons” (the number exceeding the average for recent March observations) were hypothetically reclassified as unemployed instead of employed. In this calculation the U.S. unemployment rate would be 5.3% instead of 4.4%. This may be the relevant benchmark for comparing Arkansas’ unemployment rate to the national average.
Despite the recognized problems with the CPS, the BLS maintained a policy of methodological consistency: “[T]he data from the household survey are accepted as recorded. To maintain data integrity, no ad hoc actions are taken to reassign survey responses.”
The Arkansas Unemployment Rate
State-level estimates of the unemployment rate begin with the national CPS survey, but are subject to the methodology of the Local Area Unemployment Statistics (LAUS) program.
In particular, additional information is used to refine state-level estimates, to compensate for uncertainty associated with smaller sample sizes. “Payroll employment estimates from the (CES) survey of establishments and unemployment insurance (UI) claims counts from the state workforce agencies are also used as model inputs to help mitigate volatility in the monthly state-level CPS estimates.”
In the news release announcing the Arkansas unemployment results, the Arkansas Division of Workforce services noted that “The increase in the number of unemployed Arkansans is, in large part, a reflection of the number of Unemployment Insurance claims filed during the week of March 12th.” (The survey reference week is typically the week including the 12th of the month.)
This suggests that the mis-classification at the national level might have been particularly disruptive to the application of methodologies to measure unemployment in Arkansas. And in fact, the news release went on to note “The small gain in employment is based largely on a monthly survey of Arkansas households, which was significantly impacted by the outbreak.”
With regard to the number unemployed, if the mis-classification of unemployment status was particularly prevalent in the Arkansas sub-sample, LAUS procedures to incorporate additional information–in particular, data from state unemployment insurance (UI) claims–might well have been influential in refining the estimate for Arkansas unemployment.
The problem with this simple explanation is the timing of unemployment insurance claims. During the survey reference week (the week ending March 14th, there was no perceptible change in initial claims or continued claims (insured unemployment). In fact, both measures declined slightly.
But the LAUS methodology specifies that unemployment insurance (UI) claims data are used as “as model inputs to help mitigate volatility in the monthly state-level CPS estimates.” It does not specify that UI data for the reference week alone are used. In fact, for the purpose of helping to “mitigate volatility” in the survey data, it would make sense to use data for the full month, or at least a centered moving average around the reference week, which would incorporate data from later weeks in the month. Under ordinary circumstances, this distinction would make little difference and would help refine survey estimates, but could very well be of consequence when it comes to measuring unemployment in the rapidly-changing situation we had in March.
On the employment side: The estimate for the number of employed Arkansans was reported to have increased slightly, in contrast to a sharp decline in payroll employment. This is another indication of a tainted Arkansas survey. If data from the payroll survey was used to “mitigate volatility” in the CPS survey data, the raw survey data must have shown an even larger increase in employment.
So, the unemployment numbers were “a reflection of the number of Unemployment Insurance claims” and the employment totals were “based largely on a monthly survey … significantly impacted by the outbreak.” These observations suggest that the problems identified in the national CPS survey were particularly problematic for Arkansas. The established procedures for estimating state-level estimates ended up placing routine data-refinement techniques center-stage in the estimation of state-level unemployment rate calculations.
Throwing out the CPS March Survey
If you are the type of reader who might enjoy wandering even deeper into the weeds than we’ve already gone, let’s consider the possibility that the CPS (household) survey for March is too tainted to be of use at the state level (at least for Arkansas). Using contemporaneous information from unemployment insurance claims and nonfarm payroll employment, what might we estimate the unemployment rate to be?
In order to answer that question, we estimated monthly models for one-month-ahead forecasts of unemployment and employment, using UI claims and nonfarm payroll employment data. The models were estimated for the period from January 1986 through February 2020, then used to forecast values for March 2020. The exercise here is not to forecast actual employment and unemployment, but to predict March values for the series reported by the BLS.
We derived unemployment estimates from a time-series model relating monthly (log) changes in LAUS unemployment to changes in monthly averages of continuing unemployment claims (insured unemployment). The estimation model also included a constant, time-trend, and lagged dependent variable. The March forecast from this model produces a not-seasonally-adjusted estimate indicating an increase of 10,974 in the number unemployed. This is remarkably close to the reported value for not-seasonally-adjusted data, 11,303.
The BLS and DWS noted that the usual seasonal adjustment procedures were modified to treat the March 2020 figures as “level-shift outliers.” The level shifts “preserved movements in published estimates that the models otherwise would have discounted, without requiring changes to how the models create estimates at other points in the time series.”
In an effort to replicate the seasonal adjustment process, we implemented a Census X-13 procedure, treating March 2020 as an outlier. The resulting, seasonally adjusted forecast value for March indicated an increase of 16,933 in the number unemployed. This is a close, but imperfect approximation of the BLS seasonal adjustment procedures which use a combination of X-11 seasonal adjustment techniques along with a special smoothing algorithm. If we simply take the implicit seasonal factors calculated by taking the ratios of seasonally-adjusted to not-seasonally-adjusted values in the published data, the adjustment to our March forecast value implies an increase of 18,190 in the number of unemployed — remarkably similar to the published estimate of 18,526.
A similar approach was used to estimate a monthly change in the number of employed. A model relating monthly (log) changes in household employment to (log) changes in nonfarm payroll employment was estimated through February 2020, then used to forecast a value for March. Using not-seasonally-adjusted data, this yields a forecasted employment decline of 1,431. Using the implied seasonal factors from the BLS to seasonally adjust the estimates, we get a seasonally-adjusted decline of 7,375.
With these estimates, we can construct an alternative measure of the unemployment rate that uses no information from the household survey. The model-generated estimates yield an unemployment rate of 4.8%.
Although the forecast-model version of the unemployment rate is nearly identical to the published value, the information used to construct the estimates was not limited to the survey reference week. Monthly averages of UI claims data were used to generate the unemployment estimates. The employment estimates are derived using nonfarm payroll employment from the CES survey of establishments. While the reference period is the same as the household survey, the reporting is on a pay-period basis, with many employers reporting bi-weekly or monthly figures in their reports. Hence, the alternative figures presented here are not representative of the early-March situation, which “predated many coronavirus-related business and school closures in the second half of the month.”
When it comes to estimating a figure for March unemployment, however, there is nothing magical about the survey-reference period. What made the situation in March unique was the suddenness with which a wave of unemployment spread across the economy, as evidenced by the UI Claims data. It is clear that the unemployment rate was rising during the month, and the estimates, official or alternative, attempt to capture point-in-time estimates of the March situation.
On might note that, just like the published estimates, the forecast-generated values for employment and unemployment imply an unusually large increase in the labor force, +10,815. This isn’t quite so large as the change implied by the published numbers, but it is still uncharacteristically large and moves opposite to what one might expect. The reason for this anomaly appearing again is the same as it was for the published estimates: double-counting. In this case, estimates of employment using payroll data and estimates of unemployment using UI insurance claims counted some people twice. In particular, those who were on active payrolls early in the month and were subsequently unemployed later in the month. The double-counting is an inescapable implication of trying to estimate fixed values for rapidly-changing numbers.
Ultimately, March will be seen as a transition month, wedged between February, when no impact of COVID-19 was evident, and April, when the effects of the virus and efforts to control it will be fully reflected in the employment data.
For now, as best we can tell, the Arkansas unemployment rate in March was about 4.8%.
