Transcript of the talk, from the video's captions. Auto-generated: speaker names in particular are unreliable. ======================================================================== # Harmonized Population and Labor Force Statistics Authors: John Coglianese, Seth Murray, Christopher J. Nekarda Discussant: Thomas F. Crossley Video: https://www.youtube.com/watch?v=aXOe_NS3iVo&t=17426s ## Talk (04:50:26 – 05:11:00) [04:50:26] we're at our our last session and we're going to end this one promptly at 2:45, so we want to keep everything on schedule for people who need to catch flights and whatnot. First up is John Kanezi from uh the Federal Reserve Board [04:50:40] talking about harmonized population and labor force statistics. [04:50:45] >> Excellent. Well, thank you very much. Thank you to the organizers for um putting this paper on the program. Um thanks to all of you for sticking around for the final session of today. Um this is a a project joint with two of my [04:50:59] colleagues at the Federal Reserve Board, Seth Murray and Chris Narta. And the usual disclaimer applies that these are our views uh and not the views of the Fed board. [04:51:11] Um many of you in this room are probably very familiar with the current population survey. Uh for those of you who are not, this is uh one of the main work uh workforce surveys uh within the US. Every month, the Bureau of Labor [04:51:25] Statistics collects a survey of 60,000 households uh with a variety of different information on labor force. Um this is where the headline unemployment rate each month comes from. Uh and a variety of other statistics like the [04:51:40] labor force participation rate and so on. Um in contrast to a lot of the other data sets we've been talking about at this conference, this is one that is focused on people. When we talk about employment here, these are going to be [04:51:53] people who are employed, not just counts of jobs. [04:51:57] Um, once these responses are collected, a key step in the processing is that uh the responses are weighted to match uh demographic population totals at a demographic group cell level. They weight all the responses within that [04:52:12] cell to ensure it adds up to the total. These totals are taken from the latest available estimates of population at the demographic group level from the Census Bureau. [04:52:23] And the key aspect here is that these population estimates get updated annually and new ones are released but previous year's data in the CPS is with few exceptions essentially never [04:52:36] revised. Um, and the result is that for the current year, CPS statistics reflect current population estimates, but last year's CPS statistics reflect last year's population estimates, and the [04:52:50] year before reflects that year's population estimates, and so on. So, each previous year reflects uh estimates that are now out of date. Let me show you exactly what this looks like. Um, [04:53:03] here I've taken uh a specific group. This is the civilian non-institutionalized population for ages 20 to 24. And I'm showing you uh the black dashed line. This is the [04:53:16] population uh released by BLS constructed from the current population survey. And you can see it has these echoes of different vintages of Census Bureau estimates which are each of the [04:53:29] colored lines. Um so some years it jumps up, some years it jumps down, but each year it tracks the latest uh estimate for that year. But this is not just level shifts. So if you pay attention to [04:53:43] 2025 uh here at the very end of the chart, the yellow line shows the uh Census Bureau estimates that were available when the CPS was being published for that year. and it it population ticks up [04:53:58] slightly over the course of the year. But now we have a new set of population estimates, the V2025 that are shown in blue. And uh not only is the level different, but the slope is actually different. It shows that over the course [04:54:11] of the year uh population in this group, our best estimate is now that it contracted uh by a fair a fairly uh substantial amount. So, you know, you [04:54:23] see these uh shifts in population uh and that's the clearest place to see it. But these have other implications beyond just population. Uh individuals ages 20 to 24 tend to have higher unemployment [04:54:37] rates. So, shifts in population here affect the unemployment rate. Uh shifts in the population of other groups can affect the labor force participation rate. Um and that's where this paper is going to come in. [04:54:50] So, our goal, what we're going to do is try and replicate as closely as possible uh what BLS would have estimated, what's the CPS uh statistics they would have put together had they known then in each [04:55:04] month what we know now. We're going to take the latest population estimates and use that to construct harmonized population estimates stretching back all the way through the CPS period uh at the [04:55:17] demographic group level. Then with those demographic population totals, we're going to rewe the CPS responses to match those harmonized uh population totals. [04:55:29] And then with the new weights, we're going to reestimate all sorts of labor force statistics uh from the CPS. [04:55:37] And the results we're going to I'll show you uh we're going to get smooth series with no sharp breaks due to population changes. [04:55:45] And importantly for the labor force participation rate, there's going to be some different patterns around the pandemic um that I'll highlight uh when we get there. Now um we've found this a very beneficial method for our own uh [04:56:00] policy work at the Fed, but we want this to be available as a public good for the community. So we have put on the Fed website uh harmonized time series stretching all the way back to 1948. Um [04:56:11] and uh micro data weights. So uh if you're doing some analysis or regression analysis or something where you need uh uh updated weights, you can merge these weights into your NBR or IPAM's extracts [04:56:26] and uh and rerun your analysis uh and get the the value of these um uh smooth population series. [04:56:36] So that's the overview of what I'm going to talk about here today. Let me take you through the details and I'm going to start with uh in in one slide I've tried to fit the essential aspects of BLS's methodology. This is how they go from [04:56:50] raw collected survey responses for the CPS to finished estimates like the headline unemployment rate. They're going to do this in four steps. The first step is waiting based on geography [04:57:03] alone. Uh this is what they call the first stage. This accounts for differences in the probability of sampling in different areas. This is all very standard survey methodology. [04:57:14] The second stage is the second stage is where demographics come in. That's where they take those first stage weights and then rewe within each demographic group uh the the [04:57:27] total uh weight adds up to uh population estimates uh set externally. Um they do this in a in a a variety of ways. [04:57:38] There's an initial national and state coverage step and then they do this sort of triple raking procedure to hit three sets of targets simultaneously. Um we go through there's a lot of details in the BLS documentation and in the paper uh [04:57:52] but I won't take you through that here. Just know that this is where demographics come into play. [04:57:58] Once you have weights for the CPS you can add up. You can get an estimate of the unemploy of unemployment in levels and in differences using the longitudinal structure and then you can combine those. This is the standard [04:58:12] practice of composite estimation to improve stability of the estimates and then that gives you the NSA estimates. [04:58:19] So the last step is seasonal adjustment and further back in history there are some additional time series adjustments for known breaks. [04:58:26] So those are the key four steps. That's how we get an unemployment rate. Um what I will say is the first step we're going to take as given. We're not going to make any changes. We're going to take [04:58:38] that uh entirely uh uh as given. Where we're going to come in is in the second step where we're going to use our own population totals uh to rewe match. And [04:58:52] then once we have those weights, we're going to need to repeat steps three and four, which we're going to try and do as closely as possible following BLS's methodology to be consistent. So really the key change is two, but we're going to have to do steps three and four to [04:59:05] get the the final estimates. So let me talk first about how we get harmonized population estimates. [04:59:15] We're going to use two different sources of population numbers from the Census Bureau. uh one is that every 10 years there's a dennial census. So in that month we have an exact count of population within demographic groups and [04:59:30] then in between there are these postsal estimates. These use information on uh births and deaths and migration to imputee uh changes in population at the demographic group level. And these cover the period in between but with some [04:59:45] error. We're going to smooth that error back over a decade. And I'll show you what I mean by that. Um, and then I'm not going to talk about this much today. [04:59:53] We have a lot of detail in the paper. Uh, the further back you go in history, there are some, uh, fairly significant breaks we had to sm spend a lot of time smoothing through. If you're interested in that, there's more detail in the paper. [05:00:06] Let me show you what this looks like in practice. So, here I've picked a single demographic group cell. This is for 16-year-old non-Hispanic black men. And I'm showing you uh the different uh [05:00:19] postsensil estimates. Those are the the colored lines. Uh the dennial census counts. Those are the the solid circles. [05:00:27] You can see every 10 years. And then our harmonized series is the black line. I find this is easiest to think about starting from right and going to the left. So uh with the most recent decade from 2020 forward, we're making no [05:00:42] adjustments. We take the latest postal estimates uh exactly as given and that's why you can see the yellow line and black lines are on top of each other over that period. [05:00:53] Moving one decade uh earlier for the 2010s, you can see the pink line shows the um last postsensil estimate for that decade. And by the end of that decade, [05:01:06] by April 2020, there was actually a bit of an error. they they um underestimated uh the number of people in this group and when we got the 2020 census it came in somewhat higher. We take that [05:01:18] difference between the uh last value of the pink line and that yellow dot and we smooth it back linearly over the previous decade. We don't know exactly when that error occurred. So um uh we [05:01:31] just smooth it back linearly. We then do the same for the 2000s and the same for the 90s and the same for the 80s. There are a few issues. You'll note that in in 2003 we actually have a break in our harmonized series and [05:01:44] that's because uh CPS switches from single race to multi-race. We there's a lot more detail on that in the paper if you're curious. Um and there are some special details for the for the 1980s where we have a slightly different set of population targets. [05:01:59] Um but this is the essential idea. We're going to combine data from the dennial censuses, these uh points where we have an exact count with the estimates that we have in the intervening period to get [05:02:12] a single population series covering this whole time period at the demographic group level. [05:02:20] Once we have this, we can then rewe the CPS to match those targets. Um we're going to match three sets of targets. [05:02:28] We're going to do an initial national coverage step and then we'll do two sets of second stage targets that we match iteratively in a raking procedure. Um we're going to use the same demographic group cell definitions that that BLS [05:02:42] uses essentially. Um but I want to highlight two caveats. We're going to try and get as close as we can reasonably. Uh but there are two reasons why we can't get exactly the same uh uh [05:02:56] as what BLS would have produced. Um the first is that we're actually only going to have national targets. Um BLS also does a state coverage step and has state targets in the second stage waiting process. [05:03:09] um uh the Census Bureau estimates that we have uh that are publicly available at the state level are only once a year and uh don't have the right uh [05:03:20] population base. Um so if if anybody's here from BLS and knows exactly how uh the CPS group gets to monthly CNIP estimates at the state level, we would [05:03:33] be very interested to hear. Um so please come talk to us afterwards if you know exactly how that works. Um the second caveat is that starting in 2011 uh the age information in the public [05:03:47] microdata files that we're using as our starting point is perturbed slightly. [05:03:52] That means that occasionally we will assign a person to a demographic group cell when really they should be in the adjacent one. Um we don't think this has a big effect on our estimates but wanted to flag these two uh in the interest of [05:04:05] disclosure. Once we have new weights then we can repeat the composite estimation steps BLS uses. We can uh we're not going to do our own seasonal adjustment. We're just going to take the published [05:04:18] seasonal factors. Um we're going to adjust for the 94 redesign in a standard way. And uh we're also going to we we spent a lot of time extending this far back before the micro data period. Uh it only ends up being one ball on the [05:04:32] slides, but there's a lot more in the paper if you're curious. [05:04:36] Let me show you what this looks like. So here I'm showing the uh labor force participation rate for ages 16 plus. [05:04:45] That's the share of the population that's either working or looking for work. [05:04:50] In blue is our harmonized estimate and in orange is the published estimate. And this bottom panel shows you the difference between those two series. [05:04:59] You can see that before 1994 our estimate was actually a little bit above the published estimate, but since then it's tended to be below the published estimate. particularly here uh [05:05:12] in the latest years uh more than 4/10en below that's more than a million people lower in terms of the labor force level. [05:05:20] Um let me zoom in uh on the pandemic period so you can see that a little bit uh more clearly. This is the same data just zoomed into uh the pandemic period. [05:05:33] Even before the pandemic, we had a slight gap between the two measures, but that's really expanded uh since then, particularly in late 2021 and early 2022. [05:05:46] Um, this was a period where we were recovering and there was a lot of commentary and uh debate about the speed of recovery. If you looked at the published LFPR, you'd see this rapid run [05:05:59] up over the course of 2021 that then looked like it almost stalled out a little bit. Um, but in fact, our harmonized series shows a more gradual improvement over this period. And we think this is important. This was a time [05:06:14] period where policy makers and lots of observers were trying to figure out how quickly the economy was recovering. [05:06:21] And it's valuable to have data series that don't have this contamination from population revisions so you can get a clean read from just the trajectory of the underlying data. [05:06:34] So we think this is one area where um our harmonized statistics really show a difference that uh that can be helpful in real time. [05:06:44] One area where it matters a little bit less is for the unemployment rate. So here I'm showing the same thing. are harmonized in blue, published in orange and the difference on the bottom. Uh you can see that before 1994 there are some [05:06:58] differences but since then these two measures have been have been quite quite close. Um we think this is actually somewhat uh significant in that uh there are a lot of reasons to think that maybe [05:07:10] with population control revisions in the last few years adjusting for the immigration surge among other things you know there were there were some wondering whether this could affect the uh headline unemployment rate. It turns [05:07:24] out that um uh not to be affecting this which we think is helpful. This is one of the best labor market indicators we have. So, it's helpful to know that it's not being affected by these population [05:07:36] control revisions. So, those are our main results. [05:07:42] Uh, let me say just a few words about one exercise we do in the paper. Um, in everything I've shown you so far, we've used official Census Bureau estimates of the population as uh the the foundation [05:07:57] for our uh harmonized population. But in principle, if you have any set of population targets, you could use this methodology. And one example where you might want to do this, uh, back in 2023, [05:08:11] there was a lot of discussion about an immigration surge happening that we weren't yet seeing in the official population data. [05:08:19] Um, and questions about whether that would be affecting the unemployment rate, the LFPR, etc. [05:08:27] So we do an exercise where we take the at that time the estimates that did not incorporate the immigration surge and we add on top of that our own uh estimate. [05:08:38] We take CBO's estimate of the total immigration surge and then we aortion that to demographic groups based on the demographics of recent immigrants in the ACS with those new targets. Then we pass [05:08:51] that through our machinery. We rewe the CPS. We estimate new labor force statistics uh and then apply seasonal factors and get new labor force statistics accounting for the immigration surge. [05:09:06] The result are um series that show large revisions to the levels but relatively little effect on the rates. Um, and that was reassuring because then when the official estimates came out that incorporated the immigration surge, [05:09:20] those also showed large revisions to the levels but little effect on the rates. [05:09:26] So this is one example where if you want to be thinking about population in real time before official estimates are available, you could use this methodology. or if there are any other uh set of population targets that you [05:09:40] want to that you think uh are more correct, you can use those two. [05:09:45] Let me quickly sum up. So I've shown you here today that uh the annual update of population estimates creates these breaks within the CPS that matters and is very clear for population but then [05:09:58] also filters into uh employment and labor force and potentially unemployment as well. [05:10:05] We uh undertake an effort to harmonize population and construct new CPS weights to match. [05:10:12] And what we get out of this are labor force series that are smoother in the sense that they don't feature these breaks from population revisions. And that's particularly important. I gave you an example of the LFPR during the pandemic. that can be particularly [05:10:27] important in real time to uh to know that what you're seeing is what's happening in the data and not an artifact of changing population. Um and we want this to be a benefit to the whole community. So we've put on the Fed [05:10:40] website um all sorts of time series, unemployment rates, LFPRS, EPOPS, etc. [05:10:47] Uh and also micro data weights if you want something that's uh even beyond the time series. So with that, let me stop there. Thanks >> team uh their paper on US unemployment. ## Discussion — Thomas F. Crossley (05:34:49 – 05:47:14) *Shared discussion for Session 6: The CPS and Employment Statistics; the discussant covers every paper in the session.* [05:34:49] [applause] Okay. Thank you, Bruce. And Tom Crossley is the discussant for this session. [05:35:03] So uh let me start by saying well I enjoy reading these two papers very much and so thank you to the organizers for inviting me to the conference and giving opportunity to comment. [05:35:23] Okay. I'm sorry. [05:35:36] So, I'm going to start with John's paper. Um, let me say it's a very nice paper. Um, I think the methods, as far as I can tell, are sound. Um, there's not just a nice paper here, but there's also a a data product for the research [05:35:50] community. So I think they're very much to be commended for both writing the paper and sharing uh that product the the new weights. Um so I think that's great. Um I'm going to just make a couple of comments. I'm going to say something about uh you know [05:36:05] international comparison. Um I want to say something conceptually about you know why they're doing this and maybe why the BLS shouldn't do this. So how we think about revision or not revision uh official statistics. And then I just because I'm the discussant I have to [05:36:19] maybe think about whether there are any downsides to what they're doing. [05:36:25] Okay. I'm not getting this to move. [05:36:40] >> I need to get the cursor on the slide. >> Okay. [05:36:51] Okay, hold on. Oh, I see. Maybe this way. [05:36:55] >> I've got it. >> Yeah, sorry. Okay, so um you know, so I just wanted to offer a comparison of what the BLS or my understanding of what the BLS does uh with a couple of other jurisdictions and this paper. So we have [05:37:10] these uh population data which are being updated in the in the US annually uh every January and then when and then you know we reerate to the new population totals but we don't go backwards so we only rewe going forward so there's no [05:37:24] you know even though we have new updated population information about you know month prior months we don't update and this leads as John described to a sharp uh December January step which could be up or down every year but this is not [05:37:38] unique to the BLS. So, you know, this h exactly the same thing goes on in the UK um where the population well in fact actually it's a little bit worse. Um so there are mid-year [05:37:52] population estimates um and up until the pandemic what would happen is roughly annually those would come up and there'd be a forward for forward only revision to the labor force uh to the population numbers to which the labor force [05:38:06] statistics were related. So exactly like here with a sharp step at each rebasing. [05:38:11] Um the difference was you know in the US the my the um the revisions are you know lock step Janu January um always the same whereas in the UK they moved around a little bit. So here you know there's this everyone knew what was going to [05:38:25] happen when it was going to happen. That wasn't quite the case in the UK. And then um since the pandemic uh there's been a whole bunch of revisions to LFS weights to those of you who follow the UK case and things happened like [05:38:39] Scotland had a brave heart moment and did their census in a different year and so it's been a bit of a mess. Um so things here were already maybe a little bit better than some jurisdictions. Then there's these pesky people who live up [05:38:52] north, right? So, Statistics Canada, interestingly, does something quite like what I think this paper pro proposes. [05:38:59] So, there the population numbers are rolled forward from each census using administrative data, vital statistics and administrative data on immigration um until they get to the next census in which they're revised and then they do [05:39:13] revise backwards unlike the BLS or the ONS, but only back to the last census. [05:39:19] Um, and in Canada, censuses are every 5 years. Um, what John and his co-authors are proposing is they're going to roll backwards at each update, which is going to be more frequent than the Canadian case, and they're going to go back as far as they need to. So, 10 years if [05:39:34] there's only changes back to the last sentences, but farther, if I understand correctly, when necessary. U So, there was something at least in Canada that looked a little bit like this already, which I thought was an interesting thing. [05:39:47] The other thing I wanted to um think about a little bit is you know when you think about economic statistics broadly sometimes we revise and sometimes we don't and and the way that I think about that is sort of this 2x2 table you know there are cases and and you know [05:40:01] national output or GDP is an obvious case where we need to publish before we have most of the data so there's sort of a strong incentive to revise and there's not much cost to revision because there aren't things you know actions or legal ties to the data the contrasting case in [05:40:16] the upper left hand corner is the CPI right where there are tips and colas and if we revise the data that would cause big problems but at least the way the CPI is currently done most of the data is there in time we have the prices [05:40:29] pretty quickly we're using lagged quantities um of course we could change that and we could move the inflation measure into the right top right corner if we use tornquist indices or we could move GDP to the top right corner if we [05:40:43] had GDP linked bonds like I guess the Greeks do, but we don't at the moment. [05:40:47] Um, and so the point I wanted to make is the case that's in the top right hand corner where you both have a strong incentive to revise and a strong incentive not to revise is labor force statistics because they are tied to some things like triggering extended UI [05:41:02] benefits. Um, and so you want to keep them frozen. But on the other hand, as John described, you do get the updates to the population information, right? [05:41:10] And so the way I think about this paper is, you know, and the Canadian example shows that different statistical agencies might make different choices in this quadrant. Um, and so what I think about John's paper is doing and his co-author's paper is doing is res [05:41:24] resolving this tension for us by giving us two separate series, a research series and an official statistics series. [05:41:30] Um, so are there any downsides to having a par parallel harmonized series? I don't think there are very big ones. [05:41:38] There's obviously a resource cost. I guess uh they've decided to bear that cost. Um you know we no longer have a single canonical record of what the employment rate was in June 1990 for [05:41:51] example. Um we have we'll have multiple unemployment rates and you know there is a sort of proliferating series thing here which is um you know they'll there'll be every time they update there's a new backward series. And so we [05:42:05] sort of ch tra traded jumps for a need to really do do version control and and release stamping. Um you know I served for a long time on Canadian and and UK advisory groups to national statistical [05:42:18] agencies and and my colleagues in the agencies really dislike proliferating series because they worry that it'll undermine confidence in the uh [snorts] official statistics. But I you know I was never fully persuaded by that argument. So the bottom line for me is [05:42:33] these issues are manageable. the benefits outweigh the costs and this is a really welcome contribution. Okay, so now let me talk about uh Bruce's paper with his co-authors. So this is another really important paper uh from Bruce and his co-authors. I find it very [05:42:48] convincing. I I'll try to make a few caveats because it's my job is to discuss it. It's a worrying finding for sure, but what I like about the paper is it's constructive, right? They're not just try documenting a problem, but trying to show how to do a correction [05:43:01] and how to overcome the problem. Okay, so here's the things that you might want to think about. [05:43:08] Is it possible to legitimately not be unemployed and have you unemployment compensation on your U 1099G? Okay, so I spent the last few days trying to learn about 1099gs because I didn't know anything about it. And it turns out [05:43:22] there are a lot of reasons that could happen. All right, for one thing, they're on a cash basis. So, you know, you could be unemployed at the end of the year, um, file, not get paid until the next year. You might be back at work [05:43:35] and you would have income in the second year even though you're never unemployed in the second year. So, that sort of cash versus acrruel basis could cause some problems. Um, there some states have short time compensation. So, they play unemployment benefits people who [05:43:50] have lost hours rather than employment entirely. Um, as as the paper notes, there could be other payments in this box. You know, there could be um you you know, disability as UI substitute and so on. And for exactly that reason, Bruce [05:44:04] and his co-authors dropped California from the analysis. There could be claimment fraud um so people could receive benefits when they're not actually unemployed. There could be identity theft. So apparently there was a surge in identity theft targeting [05:44:18] unemployment insurance in the United States in uh 20 and 21 um where you know people were never unemployed never filed and then got a 1099 saying they'd received uh um unemployment insurance compensation. [05:44:32] And another thing is that when um when when unemployment benefits are repaid if they were inappropriate if there's an overpayment um that's often not corrected in the 1099. So that would be another reason where somebody could have [05:44:46] a 1099 that said that they had payments when they in fact were not did not and they weren't unemployed. All of these are small uh Bruce uh and his co-authors look at a number of them and try to do something about it. I think the question [05:44:59] I would ask is do a lot of small things add up to something that's not quite as small and in particular many of these things change over time like the you know there was a spike in identity theft [05:45:10] fraud was moves over time um the the cash basis problem was particularly acute during the pandemic because states got really behind on processing so for [snorts] all of these reasons I don't think none of these thing even in total these things won't made the main result [05:45:25] go away but they could change the sort of cyclical pattern And I think what we'll see over time is that Bruce and his co-authors will you know refine their estimates as they as they document these things. Um [snorts] a second thing you might be concerned [05:45:39] with is you know this um extrapolation from the ASC to the monthly CPS. And you know this sort of rests on two arguments. One is that the SIP which collects unemployment data in a very [05:45:52] different way from ASC has a very similar bias and that the two series are close together and I understand the argument. Nevertheless, the ASC and the monthly CPS are so different in the way that they collect unemployment that it's [05:46:05] just I'm still getting my head around this. um you know the the ASAC in particular has these funny gate questions where you you know if you you sort of say um did you do any full-time work at all during the year and if you [05:46:20] say no and then you say um you spend any time trying to find work or find work or on a layoff. So if you say no to if you get that gate wrong for the last year then you never get the how many weeks question right and so I think there's [05:46:34] work in the past not looking so at the ASC but it's at its predecessor the work experience supplement showing that these rates are pretty different um so that's a worry if I can say one more thing um there the other I think big assumption [05:46:47] here is that the um the UI recipients and non-reients uh were imputing from one to the other but of course that goes exactly the other way Right. Um probably that means that Bruce's and his co-author's numbers [05:47:01] are too low. Um so I you know I think there's convincing evidence of of an important problem. I would and this is a brand new paper as as Bruce said and I would imagine that the magnitudes will continue to be refined. Uh but I really ## Q&A (05:47:14 – 05:57:05) *Shared floor discussion for Session 6: The CPS and Employment Statistics.* [05:47:14] enjoyed both papers. Thanks very much. [applause] >> Okay. Thank you Tom. We have a few minutes for audience questions and [05:47:26] comments. We have one Katherine. >> Um just a a a couple of quick comments. [05:47:32] One on the Tom's comment about the the first paper and the revisions and the the CPI being in the corner where it's you high significance and not revised. [05:47:45] There is actually a revised CPI that's now used for important purposes specifically for indexing the tax code and we we found a way to do it. I remember when that measure was first introduced having back and forth with [05:47:59] Janet Yellen who at the time was at the CEA about um how you could use a an a revisable CPI for all of these purposes. [05:48:10] And we were convinced then that it was doable. And I I think it it is, but um it it makes life a little more complicated. The the main thing I wanted to say is that you I I'm a huge admirer [05:48:23] of all the work that Bruce has done on the you know issues with reporting in in surveys and I think it's very clear that the annual ASEC numbers you under miss a lot of people who were unemployed at [05:48:38] some point during the year. But where I really have trouble making the leap is from that to saying that the monthly unemployment rate is similarly biased. [05:48:47] And Tom pointed out, you know, one of the the the re the the sequence of questions is completely different for identifying whether you're un whether you were unemployed at any point during the last year and whether you were [05:49:00] unemployed this month. asking people whether they were unemployed at any point during the last year. It's a it's a very different memory exercise than asking about what [05:49:12] you did last week. The evidence in support of the idea that these problems carry over is that the you know the unemployment the annual unemployment rate and the monthly unemployment rate [05:49:25] kind of track each other. But they that sort of assumes that the reported durations are right. [05:49:36] We know what happens when pe you monthtomonth in the CPS when people go from being out of the labor force say to being unemployed. [05:49:45] They should report zero weeks of unemployment or you know zero to four weeks of unemployment when they come in. [05:49:51] But they often say things like, "Oh, I was unemployed 24 weeks or whatever." And so if you if you have understatement of the incidents and overstatement of the duration, you could have [05:50:06] unemployment rate based on the annual data that looked a lot like the unemployment rate based on the monthly data. But that could be because the monthly unemployment rate is right and the annual unemployment rate is affect [05:50:20] reflecting two offsetting errors. So I I I could be convinced but it at this point I I am not. [05:50:35] Um uh this is for you Bruce. There's two things I want to think about this are related and related to what Katherine was just saying. One is I wonder whether or not you can use the longitudinal sorry my name is Seth Murray from the Federal Reserve. Sorry. Um whether or not you can use the longitudinal nature [05:50:48] of the CPS to look at individuals who are in the fifth to eighth wave when they're reporting in the ASAC and to look at what they were responding in the previous year in terms of their unemployment status and see whether or not those for those individuals who reported as being unemployed whether or [05:51:02] not you're seeing uh a failure to report being unemployed along the lines of what you're what you're expecting. Another is is you could then also couple that with my second comment which is uh I wonder how much this might be driven by temp layoffs where people might think of themselves as being employed [05:51:17] continuously last year but actually they were on temporary layoff and so they're going to show up as unemployed in the basic um but then not be reporting as as unemployed when you're asking them uh about their status a year prior and in particular it could actually be accounting for the very large magnitude [05:51:32] you're finding because I'd expect that people who are on temp layoff probably have shorter durations of temp layoff And so that's going to be even though there in any given month they're only 12% of the unemployed in that month, they're probably actually pretty large number of individuals when you look over [05:51:46] the course of the year. Um, and so that might be more than 12% of of individuals when you look at who is unemployed over the last year. Um, and so those are my two suggestions. [05:52:06] Uh I'm Hijan from the Federal Reserve Board and uh so I have one question to Bruce and another question to John. So I think Bruce research makes sense to me and is largely consistent with the research that I've been doing. Um but I [05:52:21] have one questions about the causes of the understatement of the unemployment rate. So it seems like your presentation suggests that you know the causes of the mclassification um I mean the cause of the understatement of the unemployment rate [05:52:35] is likely to be due to mclassification and I always wonder whether the mclassification is between employment unemployment or unemployment nonparticipation and also you know then the the second question maybe more important one is [05:52:50] that what are the causes of the mclassification is it because of the shame of reporting active job search but you know turned out to be unsuccessful or is it because of the ambiguity of the concept of activity job search active [05:53:03] job search the reason I'm asking this is you know depending on the cause you know the solution can be different and the question to John is that you know I really think that you know your research is very useful um and this can be an aggressive statement but the [05:53:18] overstatement of the labor force participation rate you know seems to be you know mclassification due to population you know weights Um but I was wondering whether the workers in the gray area you know the level of those who are in the part-time for economic [05:53:32] reasons or marginal attached workers you know how they are affected by this you know population rates. So I'm asking this because you know if the main source you know of the bias is you know heavily affected by the people who are in the [05:53:45] gray area then the implications for you know assessing the economic slack can be different. [05:53:52] >> Okay great. Uh so I think we'll give one minute each if we can to we'll do first response for for John. [05:54:03] >> Thanks. Um I'll just thank Tom for great comments. Uh really appreciated that. Um uh Hij on your question. I think uh a lot of what we find is the differences [05:54:15] uh in the LFPR is driven by uh people at older ages versus population at younger ages. sort of very broad differences. [05:54:24] And so I would be a little less worried about the the gray area driving some of that. But that's something we can have an exact answer to um uh to speak to. [05:54:35] Did you want to say anything? >> I had one other followup. Thanks again, Tom. Uh you mentioned the real time the sort of why might we not want to do this? uh sort of effects on real-time indicators and one of the things we showed is the effect on the unemployment [05:54:50] rate which is where most of those are based on was was dimminimous essentially. So I think you'd be unlikely to draw different conclusions in real time. Um [clears throat] and then the proliferation of series problem [05:55:03] is is a real one and we don't have a great solution to that other than by noting there still is an official series and we would point you to that for official statistics. [05:55:15] And Bruce, I see you have a microphone over by you. It's it's it's right there. [05:55:20] Oh, okay. Uh, I'd like to thank Tom Crossley for his comments and I'd like to thank Abra Katherine Abraham for her kind words and [05:55:34] her engagement in uh my work. I want to respond to two of her comments. Um, first it's a tricky game to try and infer [05:55:48] what's going to happen from differences in wording and and survey design. I can remember the 93 CPS redesign and how everyone thought it would make the [05:56:01] unemployment rate go up, but unemployment rate went down. Um and um so I think really the proof is in the two measures lining up having a [05:56:15] correlation of.99 and and almost a fixed difference over time. And uh Katherine's comment was that assumes that the durations are right. No, it's a [05:56:30] check on them being right or at least the bias in them being constant over time because the fact that they line up says that [05:56:43] either the durations were right or differed by a constant such that it it uh doesn't uh uh it leads to a constant [05:56:54] difference in them. So, Okay, let's thank our presenters and discuss them.