But wait, there's more.
The "birth/death model", the BLS's statistical technique for guesstimating the number of jobs they think were created by businesses they don't know exist yet, added 182,000 jobs to this month's (unseasonally adjusted) total. (See earlier post for an explanation of how the BLS imputes the existence of new jobs from new businesses by looking at the number of jobs destroyed by recently deceased businesses.) When you do all the math involved, it turns out that the BLS actually detected a job LOSS of about 68,000 jobs through it's direct sample measurements, and that only by virtue of applying the birth/death statistical fiction could any positive number be discussed at all.
But Even That's Not All.
In that recent blog entry "Most of Those New Jobs Reported Are Imaginary", I calculated that 85% of the new jobs touted as created since March 2003 are the imaginary product of statistical fiction. Updating the charts and numbers with the just-announced June numbers and the April and May downward revisions, it's even more drastic. At this point only 97,000 of the 1,380,000 jobs purported to have been created since March 2003 were actually measured by the BLS sample surveys. The other 93% were "imputed" by the birth/death model. Here's the updated graph:
So What's It Mean?
It means that job growth is far less robust than what's been advertised. That carries with it all kinds of implications for stock markets, interest rates and business decisions. So Caveat Emptor. Here's what it doesn't mean: It doesn't mean the BLS are a bunch of lying scoundrels. The birth/death model is a serious attempt to measure something real, albeit undetectable: the number of jobs created by new enterprises that haven't checked in with the unemployment insurance offices that the BLS samples to collect it's establishment job data. The BLS has been completely straightforward in it's admissions of the shortcomings of the statistical estimation techniques it uses to guess at the undetectable. The BLS has also said that birth/death modeling stinks at detecting changing trends. (I'm liberally paraphrasing.) It also doesn't necessarily mean that all the jobs imagined by the birth/death model don't exist. SOME of them do, since there has to be some new businesses out there that DID create some jobs that haven't reported in to their state bureaucracies yet.
However, when looking at other employment measures, it doesn't look too good for too many of those imputed jobs to be real. We've been stuck at 5.6% unemployment for months now. The number of unemployed is up 84,000 from April to June. There is very little evidence that we have anything more than the same "jobless recovery" we've been experiencing for months. If most of these imputed jobs have not come to pass in reality, then we should see that in ongoing lackluster jobs reports, as the unrealistic projections of the birth/death model continue to be reconciled with reality and month after month of data ends up being revised downward gradually. June's BLS release at least follows this pattern.
Footnotes
There's been an ongoing discussion about my methodology in the comments section of my previous post. Thanks to a handy and brand new FAQ posted at the BLS site, many of the issues in question are now put to rest. Some comments about all of this are posted to the comments page of this post.
8 comments:
Notes on the Jobs Report: The Bureau of Labor Statistics is ListeningThat much is clear. Posted on the BLS site is a new document dated July 2nd, a FAQ on the Birth/Death Model directly answering many of the questions posed in the comments discussion here generated by my post "Most of Those New Jobs Reported Are Imaginary", as well as in similar discussions taking place throughout the blogosphere. I'm impressed. These guys clearly take their work seriously.
The FAQ appear to confirm that I had measured the effect of the birth/death model on total reported jobs correctly. There's been some discussion as to whether the birth/death numbers should be seasonally adjusted separately. The BLS says, no, the sample based numbers and the birth/death numbers are all added up first and then the total estimate is seasonally adjusted, (which is the methodology I used.) The BLS also confirmed FTM's original criticism of my first attempt at this, pointing out that the birth/death numbers were not seasonally adjusted and therefore should not be directly compared to seasonally adjusted numbers.
Beyond that, there is one more point here to address. FTM's objection to my numbers center around his point that when looking at UNseasonally adjusted total jobs, the contribution from the imputed jobs added by the birth/death model only add up to about 40%, not the 85% I had calculated. If you stop at that series, he is right. However, we shouldn't stop at the unseasonally adjusted series. By accident of timing, the unseasonally adjusted numbers from March 2003 through June 2004 show an increase of over 3.2 million jobs. This is apparently an accident of seasonal variation, as the BLS arrives at a much smaller number of seasonally adjusted job growth of 1.38 million jobs over the same period. In other words, the effect of seasonal adjustment has been to lower the number of new jobs by 57%.
1.3 million jobs have been added to the total nonfarm payroll estimates by the statistical fiction of birth/death modeling since March 2003. It SEEMS logical that if the effect of seasonal adjustment reduces new jobs by 57%, then applying the same seasonal adjustment to the birth/death jobs would decrease them by 57% as well, leaving FTM's 40% total impact intact. However, that's an arithmetic mistake to do so. Seasonal adjustment is performed not on new jobs, but on the entire total estimate. So, seasonally adjusting the total estimate has the effect of changing the seasonally adjusted total from 99.2% to 101.5% of the raw numbers month after month. That wobble from 99 to 101 is enough to account for the big difference between unseasonally adjusted and seasonally adjusted job growth. Since the BLS adds the birth/death model numbers to the total smaple based estimate and applies seasonal adjustment to the whole estimate, the effect of applying the seasonal adjustment to the birth/death model numbers reveals a difference of + or - 1.5%, NOT 57%.
Great work! I do believe that Paul Krugman is reading you: http://nytimes.com/2004/07/06/opinion/06KRUG.html?hp
John Eaton
http://beclear.blogspot.com/
It’s good to see the BLS is trying to clear up the questions raised by its birth/death model. They’ve still got a ways to go before their methodology is well described.
I haven’t got the time to really respond to the new issues you’ve raised. After all the discussion and the BLS clarification, it’s disappointing that you are still overestimating the impact of the Birth/death estimates. At this stage it is unclear to me whether you actually believe your analysis or whether you are just sticking with it for polemic reasons.
Most of your claims are conditioned on your method of seasonally adjusting the B/D numbers. Assuming you are using the method you described in an earlier post, that method is equivalent to assuming the B/D numbers have no seasonal component. The end result of your calculations is that you are doing exactly what BLS says not to do – treating the B/D numbers as if they have no need for seasonal adjustment. Play around with a simple example and you’ll see your method of seasonal adjustment just returns the original unadjusted numbers plus or minus rounding errors.
In fact if you really want to convince readers there is something to your B/D critique, post a three-period example that shows how your method of seasonal adjustment results in adjusted numbers that are actually different then the unadjusted numbers you start with.
Absent a convincing example, I’m forced to conclude you are just offering up polemics. I’m very sympathetic to the position that employment growth has been far weaker than it would have been had we had rational fiscal policy but continuing to misinterpret the impact of the Birth/Death model is a disservice to the debate. You need to consider the future possible implications of your position. With significant probability the July B/D number will be negative because July is seasonally an extremely weak month for B/D jobs. Looking ahead with your analysis, you may be placed in the position of arguing the July non-farm number is too low due to the B/D adjustment. And if employment really starts heading downhill, you may be placed in that position repeatedly. The job numbers are weak enough -- particularly if you look at wages and hours -- to support a case that Bush has severely mismanaged fiscal policy without ginning up the B/D issue. The case has to be made on solid ground. Otherwise, things could easily turn around and bite you from behind.
Cheers,
FTM
FTM: It's good to see the BLS is trying to clear up the questions raised by its birth/death model. They've still got a ways to go before their methodology is well described.
I haven't got the time to really respond to the new issues you've raised. After all the discussion and the BLS clarification, it's disappointing that you are still overestimating the impact of the Birth/death estimates. At this stage it is unclear to me whether you actually believe your analysis or whether you are just sticking with it for polemic reasons.
RT: If you didn't have the time to go through the issues I addressed perhaps you should have refrained from making this charge until such time when you did.
FTM: Most of your claims are conditioned on your method of seasonally adjusting the B/D numbers. Assuming you are using the method you described in an earlier post, that method is equivalent to assuming the B/D numbers have no seasonal component. The end result of your calculations is that you are doing exactly what BLS says not to do – treating the B/D numbers as if they have no need for seasonal adjustment. Play around with a simple example and you'll see your method of seasonal adjustment just returns the original unadjusted numbers plus or minus rounding errors.
RT: Nonsense. Seasonal adjustment, as explained by the BLS, takes place on the whole estimate, including the jobs added by the birth/death model. There is no seasonal adjustment that effects that month's sample more than plus or minus 1.6%. Since the net birth/death numbers are folded into the whole estimate before seasonal adjustment is performed, then the total effect on those numbers as well is plus or minus 1.6%, which is why they do not appear to change much to you. Your mistake is conceptual. Seasonal adjustment is NOT performed on the delta of jobs from month to month, but on the whole estimate from month to month.
FTM: In fact if you really want to convince readers there is something to your B/D critique, post a three-period example that shows how your method of seasonal adjustment results in adjusted numbers that are actually different then the unadjusted numbers you start with.
Sure thing. First example:
April 2003. Here are the numbers reported: Unseasonally adjusted total is 129,781. Seasonally adjusted total is 129,901. Jobs added by net birth/death model:128.
Applying the seasonal adjustment to the unseasonally adjusted numbers reveals that the seasonal adjustment altered the former by +0.0925%. (129901/129781).
We know from the BLS that they add the net birth/death model numbers to the sample derived numbers before seasonal adjustment, at the cellular level:
Cell1 estimate=sample number plus net birth/death model derived number
Cell2 estimate=sample number plus net birth/death model derived number
Cell3 estimate=sample number plus net birth/death model derived number
so therefore
Sum of all cells= Sum of all sample numbers plus sum of net birth/death numbers
Sum of all cells = unseasonally adjusted published estimate or 129,781 for April 2003
Sum of net birth/death numbers for April 2003 is 128.
and
The unseasonally adjusted estimate = the sample derived estimate plus the net birth/death model derived estimate. From this we know that the sample derived estimate equals 129,653. (129781 - 128).
Now then, we know from the BLS that they conduct seasonal adjustment not on the cellular level, but on the 3-digit NAICS level (each of which contains many cells). And we know that the aggregate effect of the seasonal adjustment on the entire estimate for April 2003 was to increase the estimate by 0.0942%. Arithmetically, it looks like this:
(Sum of all unseasonally adjusted sample based estimates + Sum of all birth/death model based estimates) X 1.000942 = Seasonally adjusted estimate.
which resolves to
(Sum of all unseasonally adjusted sample based estimates X 1.000925) + (Sum of all birth/death model based estimates X 1.000925) = Seasonally adjusted estimate.
using April 2003 data, this becomes
(129,653 X 1.000925) + (128 X 1.000925) = 129,901, which becomes 129,772.929 + 128.118 approximately equalling 129901.
128.118 does not equal 128.
QED.
One more example of the same process: We'll use January 2004, which shows the effect of a large negative number on the estimates.
Here are the numbers reported: Unseasonally adjusted total is 128190. Seasonally adjusted total is 130194. Jobs added by net birth/death model:-321.
Applying the seasonal adjustment to the unseasonally adjusted numbers reveals that the seasonal adjustment altered the former by +1.5633%. (130194/128190).
The unseasonally adjusted estimate = the sample derived estimate plus the net birth/death model derived estimate. From this we know that the sample derived estimate equals 128,511 (128190 - (-321)).
So January 2004 looks like
(Sum of all unseasonally adjusted sample based estimates X 1.015633) + (Sum of all birth/death model based estimates X 1.015633) = Seasonally adjusted estimate.
using January 2004 data, this becomes
(128511 X 1.015633) + ((-321) X 1.015633) = 130194, which becomes 130,520 + (-326) which equals 130194.
-326 does not equal -321.
This shows that it is NOT the case that the birth death numbers aren't changing. They are simply changing as much as the entire estimate changes via seasonal adjustment, no more no less.
The other part of this that makes the use of the birth/death model so problematic is in one step of the math that I omitted above for simplicity. That is, the sample derived numbers are just that: a sample. What is really happening is that the total sample is compared to last months sample and a delta computed. Then that delta is multiplied to last month's total estimate. Last month's total estimate also included birth/death model numbers as well, so you have to factor that in as well. This is why I started at the month which began the 100% use of the birth/death model, March 2003.
FTM: Absent a convincing example, I'm forced to conclude you are just offering up polemics. I'm very sympathetic to the position that employment growth has been far weaker than it would have been had we had rational fiscal policy but continuing to misinterpret the impact of the Birth/Death model is a disservice to the debate. You need to consider the future possible implications of your position. With significant probability the July B/D number will be negative because July is seasonally an extremely weak month for B/D jobs. Looking ahead with your analysis, you may be placed in the position of arguing the July non-farm number is too low due to the B/D adjustment. And if employment really starts heading downhill, you may be placed in that position repeatedly. The job numbers are weak enough -- particularly if you look at wages and hours -- to support a case that Bush has severely mismanaged fiscal policy without ginning up the B/D issue. The case has to be made on solid ground. Otherwise, things could easily turn around and bite you from behind.
RT: I think the only thing we can do to continue this discussion is for you to show in as much detail as I've given you the math you employ to arrive at the conclusion that my analysis is just "offering up polemics." Until you can show me on paper where I'm doing it wrong, I'm at the point where I have to conclude you are arguing for arguing's sake.
Kind Regards,
Richard
Thanks for the examples. They really help to clarify what’s at issue.
First, you’ve gotten caught up in how the seasonal adjustments are applied rather than considering what creates the need for seasonal adjustments in the first place. It’s undeniable that monthly seasonal adjustments have a cumulative impact on total seasonally adjusted employment. But the magnitude of the adjustment is independent of the level of total employment because the vast majority of employment is non-seasonal. Seasonal adjustments arise due to seasonal employment which is but a fraction of total employment. Second, not all seasonal employment results in seasonal adjustments. Some seasonal jobs are offsetting. The job that starts in January and ends in June perfectly offsets the job that starts in July and ends in December. The part of seasonal employment that drives monthly seasonal adjustments is the monthly net change in seasonal jobs. If there were never any net change in seasonal jobs, there would be no need for seasonal adjustments.
Your examples want to disaggregate total employment into new bd jobs and all other jobs and then apply the same overall adjustment factor to each component. This is a conceptual error because the adjustment factors are the product of a time series model where bd jobs ( the high monthly variability they contribute ) are included. If you took bd jobs out of the employment time series, then BLS would supply very different adjustment factors. And the factor for total employment less bd jobs would be much smaller than the factor for new bd jobs. The adjustment factors are conditional on the inclusion of BD jobs in the total series.
Just think about it statistically. You have one series with very low percent change variation (total employment less new bd jobs) and one with very high percent change variance (new bd jobs) applying the same set of multiplicative seasonal adjustment factors simply can’t work to smooth both series. Again you are correct about how BLS constructs the numbers, but you didn’t account for the fact that adjustment factors always must be conditioned on a particular time series structure. You can’t disaggregate the series and apply the same factors to each piece. The assumption that the same adjustment factors apply to either component series is where your analysis goes awry.
The best way of allocating monthly seasonal adjustments between net jobs created by new businesses and net jobs created by existing businesses requires knowledge of the proportion of net change in seasonal jobs to the net change in total jobs for both new and existing businesses. Unfortunately these figures are not observeable. But I know of no evidence that the share of seasonal jobs differs between existing and new businesses. If anything one would guess the share of seasonal jobs to be higher for new businesses than existing businesses because a larger share of new businesses are likely temporary seasonal businesses. ( Note: Your method has been implicitly assuming the proportion of seasonal jobs in net new business jobs is near zero ) I’ll stick to the conservative assumption that if x% of the new jobs are seasonal for existing businesses then x% of the new jobs are seasonal for new businesses.
Since approximately 40% of total new jobs are estimated to come from new businesses; assuming the same proportion of new jobs are seasonal for both new and existing businesses, then it’s reasonable to attribute 40% of the monthly seasonal adjustments to jobs from new businesses.
I offered a numerical example using this assumption a few posts back which is the best way I can think of do a seasonal adjustment on such a short data set.
For anyone who doesn’t want to wade through this post, the message to take away is that I see no evidence there is anything out of the ordinary with the recent B/D numbers.
Cheers
FTM
RT: I do not believe that your point produces much of a problem. The X-12 ARIMA software seasonally adjusts using two types of data sets in mind: historical patterns of total employment going back 10 years, as well as known quirks of the calendar, moving holidays, number of weekends and such. In the former case, at no point do the bd numbers amount to more than 0.25% of the monthly total, suggesting that stripping them out of the historical experience from which X-12 ARIMA examines should not amount to a significant difference. Much more importantly however, is the basic fact that the bd numbers are not treated in any way different or separate from the total estimate by virtue of the fact that seasonal adjustment is ONLY performed on the total, not of each component. BLS is on record saying to split these up to handle separately is not mathematically significant. Finally, the other factors of seasonal adjustment, like moving holidays and such are irrelevant to the discussion.
Perhaps you are correct in that seasonal adjustment should be fine tuned in the manner you describe. It's beyond the scope of my analysis to critique this aspect of BLS work. I simply want to isolate the net contribution of bd model numbers to what is being advertised. And that is what I have accomplished to at least some degree of precision.
There is an additional step to fine tune these numbers further which I am not likely to spend time on. That would be to replicate the BLS work down to the cellular or at least the 3 digit NAICS level. There are surely slight differences between seasonal adjustments and contributions of bd numbers from industry group to industry group. If I wanted to publish a formal paper, I would do that. Since my goal in this analysis was to answer the basic question of the disparity between the advertised positive employment numbers and other less positive economic statistics, I probably won't spend the time. Let an ambitious graduate student take up the challenge.
FTM: For anyone who doesn't want to wade through this post, the message to take away is that I see no evidence there is anything out of the ordinary with the recent B/D numbers.
RT: I agree there is no reason to doubt that the product of the statistical model reported by the BLS is anything but a faithful representation of that model. What I question is the basic logical premise of that model and it's applicability to changing circumstances. The outsourcing of domestic jobs overseas is a HUGE economic trend much larger than what had been taking place when this model was developed. That fact alone is enough to question the very premise that you can impute new domestic jobs based on the jobs lost by business deaths. Remember, statistically speaking you can accurately predict the number of saloons in a census tract based on the number of churches in the same tract. Change any of a million socioeconomic factors at work though, and the statistical model will fail utterly. The BLS recognizes this and says:
"The most significant potential drawback to this or any model-based approach is that time series modeling assumes a predictable continuation of historical patterns and relationships and therefore is likely to have some difficulty producing reliable estimates at economic turning points or during periods when there are sudden changes in trend. BLS will continue researching alternative model-based techniques for the net birth/death component; it is likely to remain as the most problematic part of the estimation process."
My hypothesis is that the outsourcing revolution, as well as the incredible rise in worker productivity seen in the last couple of years are just such economic turning points. If I am wrong in this hypothesis, then the sample based estimates of payroll employment will 'catch up' to the imputed numbers of the net birth/death model. For the last sixteen months, it is clear that they haven't.
Thanks for a thoughtful and stimulating discussion.
Best,
Richard
no, if you buy the logic of my critique of your seasonal analysis, then your analysis is completely off the mark. Your graph in your post which shows the B/D jobs piling up and the other jobs flatlining would instead show two lines moving together slowly upward.
I've been repeating in post after post that your method of seasonal adjustment is leading you to seriously astray but I don't sense it's getting through.
You may be right about offshoring changing the dynamics of the labor market. But there is no evidence that The B/d model would not pick up these effects as they have been going on in certain sectors for many years.
I originally started to examine this issue because people were linking to your site and I was curious whether there was anything to your critique. But we've clearly reached the end of the road.
Good Luck
FTM
FTM: no, if you buy the logic of my critique of your seasonal analysis, then your analysis is completely off the mark.
RT: Quite frankly, I don't see the logic of your critique at all relative to what I am trying to measure. I don't care if your conception of seasonally adjusting B/D numbers is superior to the way the BLS in fact performs their seasonal adjustment. I've simply tried to duplicate what they have in fact done with their seasonal adjustment. They come right out and spell out the fact that they do not perform the seasonal adjustment on the sample based numbers and the B/D numbers differently -- they lump all together and perform on seasonal adjustment of the whole. If you think that is the incorrect way to do it, take it up with them, not me.
FTM:Your graph in your post which shows the B/D jobs piling up and the other jobs flatlining would instead show two lines moving together slowly upward.
RT: Not if the number of jobs in the real world were not in fact increasing.
FTM: I've been repeating in post after post that your method of seasonal adjustment is leading you to seriously astray but I don't sense it's getting through.
RT: Yes, it's true that you haven't gotten that point across. I'm not the only one who calculates the seasonal adjustment this way, Check out this TheStreet.com article quoting other analysts looking at the same thing. From the article:
"The BLS says it is unfair to compare the birth/death estimates with total nonfarm payrolls because total payrolls are seasonally adjusted and the birth/death figures are not. But Doug Henwood, co-editor of the Liscio Report, said the results are practically unchanged when compared on an apples-to-apples basis."
FTM: You may be right about offshoring changing the dynamics of the labor market. But there is no evidence that The B/d model would not pick up these effects as they have been going on in certain sectors for many years.
RT: Well I'd say the evidence is staring us in the face. There is no way the total estimate and the estimate stripped of the B/D numbers would show the dramatic divurgence the graph shows if the B/D model was imputing jobs that actually did in fact exist.
FTM: I originally started to examine this issue because people were linking to your site and I was curious whether there was anything to your critique. But we've clearly reached the end of the road.
Good Luck
FTM
RT: Thanks. You too. That I think will be the last word, seeing as how it's my blog ;>).
Post a Comment