Transcript of the talk, from the video's captions. Auto-generated: speaker names in particular are unreliable. ======================================================================== # New U.S. Business Establishments: Surging or Stalling? Authors: Dan Cao, Henry R. Hyatt, Toshihiko Mukoyama, Erick Sager Discussant: Benjamin Pugsley Video: https://www.youtube.com/watch?v=aXOe_NS3iVo&t=7510s ## Talk (02:05:10 – 02:27:03) [02:05:10] [applause] And the next speaker is Henry Hyatt from the Census Bureau. [02:05:50] I've got until 11:25. >> Okay. [clears throat] Uh, good morning. Uh I'd like to thank the organizers for including this paper on the program. Uh this is joint work with Dan Cow and Eric Sager who are here [02:06:03] today as well as Toshi Mukoyama. Uh please note the disclaimer. [02:06:09] So I'm going to tell you about a puzzle that I first noticed uh when I was preparing some discussant comments back in 2018. I was discussing a paper by Dan Toshi and Eric who are co-authors in what I'm presenting right now. and their [02:06:22] paper said that the number of employer establishments or workplaces in their BLS micro data was growing. And to assess their findings, I made an early version of the chart you'll see on the next slide. [02:06:35] So here I'm showing you the total number of employer establishments, also called work sites, workplaces, or plants from two data sources. Uh these are the business employment dynamics uh [02:06:48] published by BLS. that's the solid red line. Uh and the business dynamic statistics published by the census bureau. Uh that's the blue dashed line. [02:06:58] Uh both data sources provide estimates of the total number of employer establishments operating in March of a given year. And as you can see the total number of employer establishments in the business employment dynamics has been [02:07:10] increasing by more uh than the number in the business dynamic statistics. Uh in 2023, the most recent year for which data is available for both uh sources, there were just over 9 million employer establishments in the business [02:07:25] employment dynamics and just under 7.5 million in the business dynamic statistics. Uh a difference of more than 1.5 million establishments. And the business employment dynamics has published estimates for more recent [02:07:38] years. uh and in March of 2025 the total number of employer establishments was in excess of 9.3 million. Now when I uh considered these totals in my discussant comments uh almost a decade ago, the [02:07:51] latest published aggregates from the business dynamic statistics uh were for 2014 uh which means I could only see the data series until about here uh when the the difference was on the order of half a million. Uh but even then [clears throat] I could see that the [02:08:04] establishment count uh published by BLS was growing more rapidly than the analog published by the Census Bureau. [02:08:12] And here's the year-over-year change uh in the number of establishments. The first difference in the number of establishments uh in the late 1990s uh and into the as both data sources show growth on the order of 100,000 [02:08:27] establishments per year, but there's a fairly consistent gap between these two series. And if you take the average is about 35,000 establishments. [02:08:36] Uh after the downturn associated with the 2008 financial crisis, the business employment dynamics continues to grow uh by more than 100,000 per year. But the growth in the business dynamic statistics hovers in the tens of [02:08:50] thousands. Overall uh this gap for the 2010s uh is about 70,000. So it roughly doubled in the 2010s compared to the preceding decades. And after the 2020 [02:09:03] pandemic recession, this gap widened again. And uh just look at the change from 2021 uh to 2022 when the number of establishments in the business employment dynamics increased by 461,000 [02:09:16] but only by 155,000 in the business dynamic statistics. Uh there's some speculation that this difference is related to the proliferation of remote work after 2020. uh as Julie mentioned in the panel discussion earlier today uh [02:09:30] because BLS microdata collection is done by individual US states if businesses are reporting employment and payroll to more states this could increase the number of measured establishments in the BLS data so I won't have much to say [02:09:43] about these more recent years uh in my presentation today as the linked micro data that we explore ends in 2016 but I did at least want to acknowledge that as background before I move on uh and this gap continued to grow in the most recent [02:09:56] year of available data from 2022 to 2023 the number of establishments increased by 181,000 in the business employment dynamics uh but only by 86,000 in the business dynamic statistics. [02:10:09] So there's a gap of more than 1 and a.5 million between the latest establishment counts from the business employment dynamics and the business dynamic statistics. Uh why should we care? So establishment level microdata is the basis of how economists make inferences [02:10:23] about productivity enhancing reallocation and dynamism uh innovation especially in services the growth of mega firms and associated changes in market concentration which in turn can affect the prices that businesses can charge and how much of these rents are shared with workers in the form of [02:10:37] higher wages. So I'll show results today from our own microdata analysis soon. Uh but first I want to show you some more published aggregates uh that I think are helpful to keep in mind in interpreting the findings that you'll see. These are [02:10:50] the establishment counts by industry. These are published estimates for 2023. [02:10:55] Uh again because that's the most recent year for which we have data from both sources. Uh the red bars on top are estimates from the business employment dynamics published by BLS and the blue bars on the bottom are from the business dynamic statistics published by census. [02:11:10] uh two industry groups stand out. Uh the largest difference in published establishment counts is found for the education and health services super sector. Uh and the business employment dynamics uh reports uh only about 1 [02:11:22] million. Uh but the businesses the business dynamic statistics reports only about 1 million and the business employment dynamics uh reports 1.7 million. Uh that's a difference of roughly 700,000. [02:11:35] And then for professional and business services, we find the second greatest difference here. The business employment dynamics from BLS uh publishes a count that's in excess of 1.8 million and the business dynamic statistics uh reports about 1 and a4 million uh a difference [02:11:49] of more than half a million. So in the other groups uh the differences are smaller and in all but one leisure and hospitality the business employment dynamics reports more establishments than the business dynamic statistics. [02:12:01] So we'll be able to make some progress on why these published aggregates differ by linking the establishment level micro data that underlies these estimates. [02:12:09] Specifically, we use the quarterly census of employment and wages or QCW micro data which is the data source that underlies the BLS business employment dynamics data product. This QCEW microdata is collected by US states and [02:12:23] these states have also provided these data to the longitudinal employer household dynamics or LEHD program at the census bureau who include it in the employer characteristics file or ECF. [02:12:35] Different states have provided different amounts of historical data and we have complete QCW micro data for all 50 states in DC for 2004 to 2016. uh we link this micro data that underlies the BLS estimates with the longitudinal [02:12:49] business database or LBD uh the microata from which the business dynamic statistics are derived. This link was done at the employer identification level uh employer identification number or EIN level and these EIN's are grouped [02:13:02] together at the census firm ID or alpha level. All the micro data estimates that I'll show you today will be for the most recent year for which we have complete data 2016. [02:13:14] So I'll show you the results we obtained from this linked data set and I'm going to spend uh a little bit of time on this slide uh because most of the remaining results will follow this format. So we linked the employer characteristics file or ECF with the longitudinal business [02:13:27] database or LBD and aggregated this data set to the firm level. The ECF is shown in red at the top and the LBD is shown in blue at the bottom. We're comparing whether a particular firm is single unit [02:13:39] or multi-unit in the ECF and the LBD. If a firm only operates one establishment with positive employment in March of 2016, we count that as a single unit firm. If a firm operates more than one establishment with positive employment [02:13:52] in March of 2016, we count it as a multi-unit firm. A firm can be of one of five groups, four of which are shown on this slide. First of all, a firm can be single unit in both data sources. If a firm is single unit in both data sources, by definition, it doesn't [02:14:07] contribute to the disagreement in the number of establishments. So we omit that category from uh our bar charts. [02:14:13] Although you can find the omitted number in the notes at the bottom of the slide. [02:14:17] Uh overall for 2016 4.2 million firms are recorded as single establishment in both data sources. So to include that group we'd need to more than double the range of the horizontal axis. [02:14:28] Uh so for the remaining groups starting at the top a firm can be single unit in the ECF but multi-unit in the LBD. That means that the firm should contribute only one establishment to the business employment dynamics published by BLS uh [02:14:42] but more than one establishment to the census business dynamic statistics. Note that these firms do not contribute to the greater number of establishments in the ECF. In fact, they have an offsetting effect. There were 52,000 [02:14:55] such firms who operated 164,000 establishments in the LBD. So, this offsets the difference by more than 110,000. Next, consider firms that are single unit in the LBD but multi-unit in [02:15:08] the ECF. There were 167,000 such firms who operated 621,000 uh establishments in the ECF. These firms account for more than half of the disagreement between the number of [02:15:22] establishments in the ECF and the LBD. Uh next, consider firms that were multi-unit in both data sources. There were around 1.8 8 million establishments associated with such firms in each data source, but about 75,000 more [02:15:35] establishments in the ECF than the LBD. Finally, we have firms that are only present in one data source. There were 1.2 million ECF establishments that did not match the the LBD, but only 800,000 establishments that were LBD only. This [02:15:50] figure suggests that there are two drivers of the disagreement between the number of establishments in the BLS and the census uh sources. coverage which is shown here and the establishment detail especially among this singleunit versus [02:16:04] multi-unit um disagreement. Now I'll go through particular industries. Earlier I showed that the greatest disagreement uh between the business employment dynamics and the business dynamic statistics appeared in education and health services. So it turns out that most of the disagreement [02:16:18] in this industry group is due to differential coverage in the BLS and census data sources. Uh in 2016 there were about 600,000 establishments that only appeared in the ECF but not the LBD and 100,000 establishments that were LBD [02:16:32] only. Otherwise the total number of establishments largely agrees. For all the remaining charts from our linked micro data analysis I'm going to keep the range of the horizontal axis at 650,000. Uh this will allow us to easily compare the magnitude of how particular [02:16:46] industries contribute to the overall disagreement in the number of establishments. So there's one detailed industry where this uh coverage difference is uh concentrated and that's services to the elderly and persons with [02:16:59] disabilities. There are about 540,000 such establishments in the ECF and these are not in the LBD. This is a clear difference in coverage between these two data sources and that is because many employees in this industry are employed [02:17:12] by private households. Employment by private households is out of scope for the Census Bureau's establishment frame. [02:17:18] Uh, and this is related to how households report their employment to the IRS. Specifically, while most employers report their employment using IRS form 941, uh, which is a main input to the Census Bureau's business register, private households instead use [02:17:32] IRS form 1040, schedule H. Disagreement in most other industry groups looks different. Uh, next we'll consider professional business services, which is the second largest source of disagreement between the published aggregates of the BEED and the BDS. [02:17:46] Disagreement in this industry group is driven by firms that appear as single unit in the LBD but multi-unit in the ECF. [02:17:54] There are almost 60,000 such firms who operate more than 230,000 establishments with positive employment in the ECF. [02:18:03] This contributes a difference of about 175,000 establishments between these two data sources. A secondary source of disagreement occurs among firms with multiple establishments in each data source. uh there are about 280,000 establishments associated with such [02:18:16] firms in the ECF uh but less than 180,000 establishments in the LBD uh adding to this agreement by more than 100,000 and in this uh industry differential coverage actually has an offsetting effect of about 40,000 [02:18:30] establishments uh and note that disagreement in professional and business services is diffuse across many detailed industries if we zoom in on the six-digit NAKES codes that contribute the most to the dis disagreement uh custom computer programming services you [02:18:44] know the the contribution is small. Uh if we look at computer system design services uh the contribution again uh is small and those two industries uh contribute less than a third of the overall disagreement in professional and business services. I'll quickly go [02:18:58] through some other industries. The third largest difference is is found for trade transportation and utilities. Uh the largest contribution comes from the 35,000 firms that appear as single unit in the LBD uh but in the ECF operate 150 [02:19:10] thou 115,000 establishments yielding a difference of about 80,000 establishments. Uh in manufacturing most of the disagreement is found among firms that are multi-unit in both data sources but where the ECF reports uh an [02:19:24] additional 65,000 establishments. Uh the single multi- disagreement uh only contributes about 40,000 in manufacturing. And uh in information there are 7,000 firms uh that operate a total of 21,000 establishments in this [02:19:38] uh single multi- disagreement. So what have we learned so far? In 2016, the QCEW had on the order of a million more establishments than the LBD. Now, of course, that difference has grown. There are two main reasons for this. The first [02:19:51] is differential coverage. When a private household employs someone to provide health care services, uh they are counted by the QCEW, but they don't enter into the census business register. [02:20:01] The second is disagreement among matched establishments, especially those that appear as single unit in the LBD, but report multiple establishments to the QCW. [02:20:09] So, I started with published aggregates. Then I used a unique micro data set to understand why the number of establishments published by the BLS and Census Bureau are different. And I'll discuss the economic implications of these results uh in the context of the well-known Hopenheind and Rogerson [02:20:23] model. But first, I'd like to discuss published aggregates some more. [02:20:26] Specifically, how the results I've shown you can help us understand the differences between the published establishment size distributions from the BLS and the Census Bureau. Uh this the establishment size distribution provides moments that the uh economic model that we calibrate targets. [02:20:40] Uh so here's the number of establishments by their number of employees in the business employment dynamics in the BDS for 2023. Both data sources are skewed toward large uh towards small establishments. Uh the BLS data show that there are over 7 million [02:20:53] establishments with 1 to 19 employees while the Census Bureau shows uh closer to 5 1.5 million. Uh the absolute differences among the other establishment size categories are of course small. Um and if we convert this to logarithms, uh you can see that there [02:21:08] are similar proportionate differences at uh at the extremes. Uh differences of roughly one quarter. there's a about a 25% difference between the number of establishments in the BLS and the Census Bureau among these small establishments. [02:21:22] You know, 25% being the proportionate difference between um 7 and 5 1/2. And then there's also a 24% difference among establishments with uh 500 or more employees. The business employment dynamics uh publishes a count of 7 or [02:21:36] 17,000 while the business dynamic statistics has 22 thousand. And this matters for how employment is allocated across establishments of different sizes. Smaller establishments by definition employ fewer workers. So even though the number of such establishments [02:21:50] may differ by 1.5 million, the BLS reports nearly 35 million employees compared to the Census Bureau's 33 million. Uh but at the opposite extreme, the uh uh difference in employment is is much larger. Uh the Census Bureau reports 31 million employees while the [02:22:04] BLS uh reports less than 23 million. Uh okay. And uh and if you sum these numbers, you'll note that there's also a um uh there's a a difference in the [02:22:19] total amount of employment uh in these two data sources. So the the Census Bureau publishes a a total employment count uh that's about uh 10 million higher than the um uh that that of uh BLS. Uh we explore that a little bit in the context of this project and I can [02:22:34] tell you that uh it's not due to differential coverage. uh the greatest source of this discrepancy uh occurs among uh uh employers firms that appear as multi-unit uh in both data sources. [02:22:46] Okay. Um so I've shown you a lot of numbers. [02:22:50] Hopefully I've convinced you that the BLS and census data frames differ in how employment is allocated across establishments. Uh how much does this matter? So I'll incl explore this in the context of a well-known model. [02:23:00] Specifically, we're using the Hopenheind and Roger model. I'll get through all the working parts of this model on one slide. So these are all the equations that I'll show you uh for the economists in the room. So uh a representative consumer provides uh labor services for [02:23:14] which they're compensated receives its share of business profits as well as their share of uh government receipts collected from a firing tax. Uh the firms in Hopenheind and Roger are more interesting than the consumers. They have a decision to make regarding how many employees to hire and they also [02:23:29] make a decision whether or not to exit. Uh to save space, I'm not showing you the value of exit, but it involves setting the number of employees to zero. [02:23:36] Uh and it gets to avoid paying the fixed cost, but it still has to pay the firing cost. Uh a decreasing returns to scale production function transforms a single uh input labor into output. Uh there's a um a returnsto-scale parameter theta as [02:23:50] well as a productivity shifter s. Productivity shifter is determined by an AR1 uh process. And the firm begins the period. observes its productivity, its fixed cost and the number of employees that it had in the previous period. So why does the number of employees it had [02:24:04] in the previous period matter? Uh because if it downsizes, it has to pay a firing cost to as a share of the change in employment. And uh that's the model setup. So some of the standard parameters I put in the appendix slides, but the interesting ones uh are the ones that are determined by the establishment [02:24:18] size distribution. Now before I show you the parameter estimates I want to mention the working parts of the model are going to yield some uh some inferences uh as we match the establishment size distribution in the census and BLS frames. The census frame contains more large establishments than [02:24:33] the BLS frame and the BLS has more small establishments. So large establishments are rare. So the question is how you can thicken the tail of the establishment size distribution. In Hopenheind and Rogerson there's a tight link between an establishment size and its productivity. [02:24:47] uh being a rare big establishment uh is easier when shocks are more persistent. [02:24:52] Uh and this slide puts numbers to that intuition. So here we're matching the establishment size distribution uh in each data source. Uh matching the BLS uh uh impi or me uh matching the census uh [02:25:04] BDS uh implies a larger cost of entry and it also uh implies uh lower uh less volatile uh innovations which are in turn more persistent. So this persistence parameter uh is higher when [02:25:17] we match the uh the census data. So this framework links the stability or volatility of the economic environment and the establishment size distribution and our calibration illustrates how a more dynamic economic environment uh makes frictions more costly which you [02:25:32] know we we calibrate on the on the next slide. So following Hopenheind and Roger we consider the incidence of a firing tax. Uh in this case we make the firing tax equal to one year's wages. Uh the 0.2 2 is because it's calibrated at 5-year intervals. Uh when we target the [02:25:46] BLS establishment size distribution, uh this friction implies a greater rise in the price of consumer goods as well as greater declines in output uh labor and labor productivity. Uh and these are [02:26:00] smaller when we uh target the the census establishment size distribution. So if productivity moves around a lot and there's a friction that keeps employers from operating at their optimal size, this is going to matter more when firms are more likely to have to reoptimize [02:26:14] due to shocks. So to conclude, the difference between the business employment dynamics and business dynamic statistics has grown over time and is now more than a half a million and a half. Uh there are two reasons for this. [02:26:27] The first is differential coverage specifically employees of private households who provide healthcare services. The second is the disagreement in the way that firms appear in the micro data that underlies the business employment dynamics and business dynamic statistics especially firms that are [02:26:41] single single unit in the census micro data but multi-unit in that of BLS and in the context of an example that we work through if there are more large establishments that suggests a more stable less dynamic economic environment [02:26:54] uh in which the costs of frictions are smaller and uh I look forward to your questions and comments as well as those from our discussant. Thank you. ## Discussion — Benjamin Pugsley (02:27:03 – 02:37:48) *Shared discussion for Session 5: Measuring Employment Using Tax Administration Data; the discussant covers every paper in the session.* [02:27:03] >> [applause] >> Ben Pugsley will be the discusser. [02:27:50] Terrific. Thank you very much to the organizers for inviting me to uh uh to read these papers and uh and participate in in this discussion. So right off the bat, these are two excellent [02:28:04] presentations. Um I and it may the second one in particular it it made me think back because I I like Henry many years ago I had looked at the establishment um entry rate of [02:28:17] establishments in in the BLS data and entry rate of the establishments in in the the census data and noticed some differences and I couldn't really and I' I've talked to many of you in the in the room about some of those differences years ago and I couldn't really get to the bottom of it but the the census data [02:28:31] were a bit easier to use and it was a a little bit harder to get to get to the to the data on the on the BLS side and I and I kind of stuck with the the census data and of course over time these these differences got much bigger and and so I was really excited to have the chance to [02:28:46] read the paper by Henry and his co-authors to to learn uh learn more about the source of those divergences. [02:28:52] Um Henry concluded with the discussion about the Hopenheim and Rogerson model. [02:28:56] I love the Hopenhe and Roger model. I I don't have a I don't have a lot to say about the Hope and Hiden Rogers model in this particular paper because I think it's just an instantiation of uh a something that matters what that depends on the uh that depends on the [02:29:10] establishment size distribution. So we got to get the establishment size distribution right. Okay. So in the spirit of in the spirit of the conference and thinking about this frames program and thinking about uh integrating the the BLS and and the census data, I want to talk a little bit [02:29:24] about the challenges and the payoffs to integration. Um we have learned a lot and many of us already know about the two different sort of data streams that are going into the QCW program and going into the um and going into the census business register. Right? So the QCW is [02:29:39] built up from state level UI reports. Um Jason made an a joke yesterday about hurting 51 sheep. It's actually I mean in some ways that ends up being a little bit easier than than hurting one big [02:29:53] sheep in particular. uh but it's an extraordinary example of of uh federal state cooperation and coordination. [02:29:59] Okay, the census business register is coming from different streams. So primarily from federal tax reports and supplemented by the economic census company organization survey and other data streams. What I didn't appreciate until I read Henry and his co-author's paper is that the census led of which [02:30:14] I'm a more recent user uh links the QCW and the business register in an incredibly helpful way. Uh the problem is is that the QCW frame and the business register frame can disagree in really meaningful ways. And so my comments are uh about the paper and the [02:30:28] frames project and and the source of this disagreement. So I'm going to I'm going to start to by by saying yes, there's some differences in coverage. Um if you looked at the lines that that Henry presented, I would say that the [02:30:41] measures of total total in in quotation marks private payroll employment largely agree. Okay, so there yes, there's some differences, but they're not diverging. [02:30:49] the what's diverging is the number of establishments. So what does that mean? [02:30:52] That QCW shows far more establishments in the business register. That means that the um they disagree over the allocation of of this employment across places and and that's really really important. It's important for the Openheim Rogers model and it's important [02:31:05] for a lot of questions. So the QCW feeds into the business employment dynamics, the business register feeds into the BIS county business patterns and and and the BDS. um those narrow the gaps a little bit because some things about zero employment observations and and household employees and so on, but [02:31:20] there's still a really big large gap. Okay. Um so which one's right? I hate it when students do this. Yes, there there there's there there's both uh they're both right to some extent. Okay. So I want to narrow things down a little bit [02:31:34] and I want to say that if you're looking at larger establishments, I wouldn't sweat it. Okay. the the uh on the orange line here. So this is just the uh splitting the the uh the BDS and splitting the business employment [02:31:47] dynamics by by size. Um and this is so this is much smaller than than Henry did. So this is one to f one to four employees and and five plus employees. [02:31:57] And so the uh the orange and lines the solid one is the BDS and the broken one is is the business employment dynamics. [02:32:04] And they're they're pretty close. Of course, those are big establishments and so that that is a difference. But even in logs, it's it's it's relatively uh small compared to looking at the BDS, which is the solid green line and looking at the BED, which is the broken [02:32:17] green line. And that's there's a huge difference. Um and so small, very small establishments are growing really quickly in the the QCW frame and they're not growing uh as quickly in the the um [02:32:31] on the census side. So there's some differences in coverage. Uh the home health aids are on the on the BLS side. [02:32:37] You can quibble about whether they should be counted or not. But there also is some differences in conceptual differences in in what counts as an establishment. Um for example, the different ways of feeding into um different different ways of of taking [02:32:51] reporting units and converting it into uh converting it into uh work sites or or uh locations of of business activity. [02:32:59] Okay. So the surging and stalling paper I view it primarily as understanding the source of divergence especially for these really small establishments and I would say kind of despite the title the title's a little bit of a red herring. [02:33:12] It it doesn't read as a critique on on declining dynamism. What it tells us is there's more employment in these small establishments potentially and that could be good or bad for for efficiency. [02:33:23] The problem is is that there's real discrepancies in these publicly facing tabulations that are designed to measure similar things. And so the rest of my presentation is is a is a comment. It's kind of an easy one to make. Uh but it's a comment and a wish. Okay. And both of [02:33:37] these things apply to the frames project as well. So what's the comment? Uh the comment is we have to pay attention to the rising incident of remote work. And the QCW frame is particularly uh sensitive to this. You already heard about this this morning, but as you [02:33:49] know, UI tax is collected typically by the state where the work is performed. [02:33:54] And so if you're an employer where you have fully remote or even hybrid employees, they're going to pay their state unemployment tax to the state in which the employee performs the work. [02:34:02] Now, sometimes this can get quite complicated. Um but there's a whole uh framework for localization of work rules that provide guidance for these complex cases. And at the end of the day, that um employee and that payroll goes to a particular state and that state might [02:34:16] not be where the where the headquarters or where the actual business itself is located. So multiple work site reporting actually does try to take this into account. So they try to not count the the remote employees. But the problem is is when those remote employees are in [02:34:30] another state, it generates another reporting unit. It can be hard to sync these things back up to realize that this is not an actual establishment in another state. It's just somebody who is a coder and and working in Michigan and [02:34:42] and um or let's say working in um New Jersey, but actually living in uh living in Michigan. Okay. And we know remote work is important. So this is taken from a a census paper and you can see that [02:34:55] remote work triples uh since since 2020. Okay. [02:35:01] Okay. So this is rising incident of remote work. Um the Henry and his co-authors showed that the 54 and 56 the professional business services that's a large part of this this gap. Now their data only goes back to 2016 where they have the the joint distribution but even [02:35:16] then if we look at those those industries these are these are ones that were early adopters of of uh of remote work. Okay. So the census and the the business register and the county business patterns the the way in which they ingest that information on on [02:35:30] locations wouldn't be contaminated if if we think it's actually contamination by these employee work arrangements. And so one big question for for this paper is, you know, how much of this divergence of small establishments is is explained by [02:35:43] the changes over time in in work arrangements. Okay. Um I got about a minute left. Uh I that was my my comment. Oh, two minutes left. Okay. Uh here here's here's my wish and and this [02:35:56] is this is aspirational. Um, I wanted to say that the so I can just use the BDS, I can just use the LBD because it's not necessarily sensitive to to these problems. But the as we know on the the census side there's there's [02:36:10] shortcomings as well. Uh, timeliness is an issue and and timing of an establishment birth and death is is a is a really significant issue. Even if we have it um even at years that end in two and seven, we we get a fairly precise [02:36:23] measure. So there are uh shortcomings to the the the census side and there are many advantages to to the QCW frame. So for the surging and the the the stalling paper um this is sort of a a challenge [02:36:37] to the the authors. Um I think everybody in this audience is probably already convinced. Um, I'm I'm less interested in the the economic uh sort of the outcome of a Hopenheind and Rogerson model that's matched to a particular [02:36:51] establishment size distribution than I am in in a user guide for how to harmonize the the data that we have these two we have these two data sources that are both providing signal uh not not just not just one adding noise and [02:37:03] and so I'm more interested in in how how do I use the QCW data to to enhance the um the business dynamic statistics and for the frames project. Uh Henry alluded [02:37:15] to this as well. Um, the QCW obviously feeds into the jobs frame, but I think what we're what we're learning and and this is a great example is that the QCW has great potential uh not just for for [02:37:28] jobs obviously, but but for the higher frequency flow of UI data to help us more precisely measure business locations in the geospatial frame uh as as well as the boundaries of of the business in the the um what is referred [02:37:42] to as the businesses frame. So I will uh I I will end on on on that. Thank you. ## Q&A (02:37:48 – 02:50:51) *Shared floor discussion for Session 5: Measuring Employment Using Tax Administration Data.* [02:37:48] [applause] >> So we we have a little time for uh questions and comments and I will take advantage of having the mic in my hand [02:38:02] to ask a question. um which is I I mean I make a comment and and ask a question which is I I'm not terribly surprised that there's more establishments in the [02:38:13] BLS frame. um the it it was described as being based primarily on unemployment insurance reports, but that's not exactly right because for a long time the BLS has also done uh uh created the [02:38:27] multiple work site report, which is an effort to go out to businesses that have a single unemployment insurance account number to ask them about where the locations are of all the people who are working from them and identify the [02:38:41] individual establishments. So, it's not really surprising to me that there's more establishments in the BLS frame. [02:38:47] Um, but I I I guess there's also a question about when you really care about these establishments and when you care about the firm, if you will. So if you're if you're concerned about, you [02:39:01] know, productivity and um the responses to shocks to the the the e economic environment, it may be that what you really care about is the decision-making unit, which would be [02:39:15] the firm rather than the establishment. If you're a local, you know, community development person, you may care about the, you know, where the people are actually working because, you know, that's going to affect, you know, who's paying property taxes and and all of [02:39:30] that kind of thing. So, I this is maybe more a comment than a question, but I I don't know. I'd be interested in Henry's thoughts about how how I should think about that and which is relevant when and may John may have a comment that's [02:39:46] related to to this. >> Why why don't we actually why don't we collect a few comments and then >> so I've been watching both of these projects which are are great. It's it's [02:40:00] the the frames project's fantastic. I kind of question for Dedra that might be hard to answer which is you know this is amazing breaking down of the silos and integrating all this data. I don't know whether you can give us a timeline of of [02:40:13] when we're going to start seeing uh of I'll say new products. You might say these two papers we just saw now and and whatever are are the new products but but what are the timelines? and and I I guess I'm uh particularly interested in [02:40:27] will this speed up the timing of the business register and then various data products like the CBP and uh and the BDS. So for for Henry I I I've been talking to that team for the last couple [02:40:40] years about this paper. uh ju just so these authors already know this this comment but I think there's another puzzle out there that's kind of interesting even within the BLS data which is if you if you take the beed [02:40:54] data and you compute the number of establishments using a longitudinally consistent method you start in a particular year and then you take establishments and you add add all the openings and you subtract all the closings and you and you track that [02:41:08] measure it's actually much closer to the BDS than then than than uh actually the total number of establishments who reported the QCW and BED. So it doesn't look like this is coming from at least [02:41:21] measured openings and closings. And so that's I think a puzzle that we need to figure out. Um the other comment I really echoing what uh uh Katherine said it when actually look at the firm level [02:41:36] data, the data sets aren't so different. The firm counts are actually pretty similar. And I agree with Katherine. I mean, one one for those of us working in firm dynamics and establishment dynamics, we always kind of confront the problem that Katherine's already identified. [02:41:54] >> There's a mic. >> Um Henry, um if you wouldn't mind identifying yourself. [02:42:01] >> Oh, Andrew Reamer, George Washington University. Um great work. uh within sectors to what extent are you uh do you think that there's differences in classification by census and BLS of the same [02:42:16] establishment so that some of the differences within uh nakes codes particularly at the four and five six digit level is that the establishment appears in another another nakes code and context for that is in in [02:42:30] the development of the AIES annual integrated economic survey census told OM I'm quoting nakes classifications can be unnatural or challenging for some businesses. So, uh there's a challenge for census to get the NAKES code right [02:42:44] because businesses are self-reporting. Okay. Uh Ryan Decker from Federal Reserve Board. This is such awesome stuff. Um but I have a comment sort of picking up on Katherine's comment about [02:42:58] uh like the boundary of the firm. Uh if you drill into these 624 establishments where I think about half the excess is uh and if you go look at the QCW what you find is that in 2013 [02:43:11] um they were they became 624 and had previously been 814 private households as Henry was was talking about. So that was a decision that was made for the QCW. Um it wasn't sent back into history. QCW is not a time series. They [02:43:24] will tell you if you ask uh them about that. Um but my first sort of idea is uh should we have more scope for coordination across the agencies for deciding where to put some of these different establishments is kind of related to Andrew's comment. Um but then [02:43:38] the broader thing is these look to be like basically individuals who have some sort of uh employer uh business records but in some ways are acting a little bit more like non-emp employers. And so I think the bigger P and and indeed if you go drill down it looks like they [02:43:52] probably have one employee per establishment and these are some very small number kind of like what Ben showed. So what I'm wondering is so we're treating these folks as establishments employer establishments. [02:44:03] Are we consistent across all the other sort of quasi employer uh uh arrangements that could be out there like temp help workers construction dispersed location workers. There's some documentation on that in the BLS uh stuff. contractors, the remote work [02:44:17] thing has come up. Um, when we're going to go calibrate these models to the firm size distribution, you know, saying, okay, well, this one decision QCW made, we should probably let that count, but what about all these other, you know, uh, decisions? What should we do about [02:44:31] them? And then the real elephant in the room, which is the non-employer universe and this very clear line between the employer and non-employer universe that exists in all of the data that we're going to hear more about later today. [02:44:41] So, I'm it's not a question, it's a comment that um I think we have a real challenge here and I'd be very hesitant to say, "Okay, let's go with QCW for this firm size stuff or or establishment size." I think we have a much bigger mess on our hands. Uh as non-employer [02:44:56] universe has grown tremendously, uh as folks here have shown, um I I I think we've got a lot of work to do here. [02:45:05] respond. >> Um, let me just go in the order of the papers and ask Dear Dre to respond and then and then you thanks John. Thank you for your [02:45:23] question. I think it's uh important to remind everyone that the work that's happening within the frames program is primarily concentrated on the foundational elements that are used to [02:45:36] create these new products. And so when we talk about enhancing the geospatial frame to include both residential and non-residential addresses and then also [02:45:49] attribute mixeduse structures and ensure that those flags are represented to show where we have one of those addresses in the geospatial frame, the business frame, the demographic frame. Those are all foundational pieces that are [02:46:04] strengthening what we release next. Okay. [02:46:08] A good example of a new product coming out the door is the self-employment statistics that we talked about that's going to come out in December. And then another piece of what's making the timing better is that we are moving from [02:46:23] off on prem servers to the cloud and we're moving away from SAS and Access DB and SQL code to Python and R. And John and I were speaking a week ago or so and [02:46:37] I talked to him about a briefing that I had just received at the Census Bureau about two products related to justice and criminal justice. And I said it was amazing to hear these two products that [02:46:50] had moved off of Access DB to um R or Python and are now maintaining all of their documentation and code in GitLab repositories. The one product was able [02:47:04] to go from and I wrote this down three weeks in production time to four days. [02:47:10] The other was able to go from one month to one week. So you see how this is kind of all coming together. The advancement of technology, the strengthening of the foundation, it kind of fosters creativity, innovation along the way. I hope that [02:47:25] helps. And then to the other commenter about having more conversations across agencies, I definitely agree, especially after reading Henry's paper and sitting here with all of you. We need to do [02:47:39] better at that and I think the leadership is starting to be in place to enable that. Um, but we also need to talk more within agency. So, we'll work on that too. Thank you. [02:47:52] So, so I I'll try to go through um the comments in roughly roughly reverse order. So, uh I'd like to I'd like to take issue with some of the numbers that we've been throwing around, including by myself. I was saying like census [02:48:05] publishes 7 12 million employer establishments. Dearra said 8 million and if you go to the the county business patterns there there are some some definitional differences and you can get a 8 million number. I am going to say the Census Bureau doesn't publish 7.5 [02:48:18] million establishments. It doesn't publish 8 million. It publishes 32 million establishments uh operating in in 2023 because if you add the number of employer and non-employer establishments and when it comes to non-employers, Census Bureau does call them [02:48:33] establishments. So, so yeah, it it's not it's not 7 12, it's not 8, it's 32 million. So, uh it it's true like uh how do we define an establishment? [02:48:42] Non-employer establishments, it's generally just a sole proprietor. It's one person uh and we count that as an establishment. uh for home health aids, it's one person going to one like location. They're employed by a particular household who's reporting the the services of the person they employ, [02:48:56] right? It looks very similar to a non-employer. So, I'm I'm I'm very uh similar uh to Andrew Reamer's comment. [02:49:01] Uh so, uh I am certainly uh sympathetic to the idea that we really need to be releasing uh data that's relevant for policy analysis. Uh I can't tell you how many times when I present this, I'm asked about uh particular associate application to the chips act. uh and uh [02:49:16] really in order to do these sort of industry level classifications comparisons uh we need to be working on uh not just looking at things at the firm level which is what I'm doing uh but more like uh linking the establishments to each other which is which is a challenging exercise uh [02:49:31] Martha Stinson and her team uh have been looking into this some of my colleagues in LHD have been looking into this uh but you know we we still really need to to be uh pushing on that so in my opinion uh industry largely agrees and [02:49:43] it agrees except when it when uh doesn't uh and then um uh John you have your your your comment is you know it's going to be a paper so the um the the reason that you can't recover the total number of establishments and the business [02:49:57] employment dynamics uh is because of a very straightforward reason when um a business previously existed and it you know it responds to the the um multiple work site report for the first time or [02:50:12] the annual refiling survey. uh it that those uh transitions are suppressed, right? It's it's really that uh and the states actually provide the data that we need uh in LEHD to to to quantify that. [02:50:23] But that that's really that's where the bodies are buried. Uh and then uh yes, Katherine, I I am really uh uh sympathetic to the the question that economists have been asking for centuries or or uh what are the boundaries of a firm? Why do farms matter? Uh what are the boundaries of an [02:50:36] establishment? And uh yeah uh having those sort of uh conceptual issues really pinned down uh will really help us make progress in in measuring things in a in the most relevant way. Thanks.