Is the “Institutions” series an overestimation of NFP employment?

This proposal put forward Kevin Drum in Jabberwocking. Verbatim:

So: have we really created 2.5 million new jobs since March? I’m not sure I believe it.

The chart he posted briefly summarizes his argument:

Source: Drums (2022).

The stylized fact is that the row of households adds little to the information contained in the rows of establishments about cyclical behavior or employment, as pointed out by practitioners (see fig. Furman, 2016P CEA, 2017). However, the possibility of incorrect measurement, especially at this time, remains. Some suggestion of mis-measurement gives the behavior of non-seasonally adjusted non-farm payroll (NFP) employment and total employment covered by the Quarterly Census of Employment and Wages (QCEW). For a number of reasons related to employee coverage, the numbers vary, but general trends should give some idea.

Figure 1: Log ratio of NSA NFP to QCEW (blue line) and 2001-2019 average (red dotted line). Peak-to-trough dates as determined by the NBER are in grey. Source: BLS via FRED. BLSNBER and author’s calculations.

In 2021 and the first half of 2022, the NFP was below what the historical relationship with QCEW might suggest.

While the Establishment Survey is just such a survey, the QCEW is a census that should in principle capture all covered employment through tax reports, etc. These data are used, among other things, to benchmark a number of establishments. The last preliminary benchmark was March 2022 (but not yet included in the official series – I do it in this mail). Note that in Figure 2, the ratio dropped sharply in early 2022 and rebounded again in June. This tells me that as of June the NSA NFP series was just around the corner. QCEW is now significantly behind (June data was released Nov. 22), so we cannot use QCEW to judge data from July to November.

With this in mind, what does net employment growth, measured in different ways, mean since March 2022 (when was the last reference month)? In addition to the official NFP BLS series by establishment, Figure 2 shows the QCEW series (seasonally adjusted) and household series adjusted for the NFP concept.

Figure 2: Employment change from NFP data (black), QCEW seasonally adjusted using author’s Census X-11 (dark blue), QCEW using author’s multiplicative moving average differences (sky blue) and using household employment series adjusted from considering the NFP concept (red), all in 1000, relative to 2022M03. Source: BLS via FRED. BLS, BLSand the author’s calculations.

The Household Employment Survey Series adapted to the NFP concept is not an official series and is described here. A word of caution: Given that the household employment series is based on a relatively small sample, the variability associated with it—and therefore, more broadly, its derived series—will be relatively large.

During June, it is not clear whether the NFP underestimates or overestimates the number of QCEW, given the different results obtained during the seasonal adjustment process. The adjusted series of households shows that only about 800 thousand 2700 thousand of the official NFP series were created for November.

One way to get around seasonal adjustment issues (which have been exacerbated by the pandemic-related sharp drop) is to look at 12-month changes. This obviously belittles recent events, but at least we can look at trends.

Figure 3: 12-month change in employment compared to non-seasonally adjusted NFP (black), QCEW (dark blue) and NFP concept-adjusted household employment series (red), all in thousands since 2022M03. Source: BLS via FRED. BLS, BLSand the author’s calculations.

All three series match up pretty well until 2022M06. Looking up to 2022M11, both NFP and adjusted household series show a decline. Which one to believe? As noted earlier, Furman, 2016 as well as CEA, 2017 give almost full weight to a number of establishments. A more formal analysis showing near total weight in the top ranks of establishments for real-time analysis is presented Goto, Jacobs, Sinclair & van Norder (2021). (Furthermore, if I consider the final NFP data and the adjusted household series, the statistically adjusted household series appears to respond to gaps between the two series, while the NFP series does not; the formally adjusted series corresponds to the error correction time.)

So, for now, while I’m open about what (NFP) employment growth is, I think that (1) it’s probably positive and (2) it’s probably slowing down.