Is February employment growth overstated due to statistical problems?
From “February Employment Report: Has the Economy Seen Its Shadow?,” by James Sherk and Salim Furth, a discounting of the recent positive employment figures:
Another concern is the pattern of “spring boom, summer swoon” in BLS reports since 2010. In each of the past four years, the BLS employment report has shown strong growth in the late winter and early spring, followed by anemic growth in the summer. The winter and spring reports have led to hope for summer growth that then fizzles.
This pattern may be a result of the economic collapse of late 2008 and early 2009 interfering with the BLS seasonal adjustment process. The BLS’s algorithms may “expect” large job losses in the winter and spring that do not materialize, making their seasonally adjusted job reports appear artificially strong.[1] This month’s good numbers could represent a similar statistical mirage, similar to the growth in early 2010 that led the Administration to predict the “recovery summer”[2] that never arrived.[3]
On this topic, the Economic Report of the President, 2013 reports in “Data Watch 2-1: Seasonal Adjustment in Light of the Great Recession” (pages 47-48):
…detailed studies of a wide range of principal economic indicators suggest that the seasonal adjustment techniques that had already been employed by the Bureau of Labor Statistics (BLS) adequately accounted for the effects of the Great Recession. BLS analysts calculated alternative seasonal factors for total nonfarm payroll employment after manually excluding the sharp declines that were recorded during the downturn (Kropf and Hudson 2012). This counterfactual experiment failed to generate meaningful revisions to the actual published estimates of total nonfarm payroll employment since January 2010. In fact, the BLS analysts concluded that the implementation of these counterfactual seasonal factors would have revised total nonfarm payroll employment upward by a mere 24,000 jobs over the second and third quarters of 2011 (in other words, an average of 4,000 jobs a month) and downward by just 19,000 jobs over the fourth quarter of 2011 and the first quarter of 2012 (or an average of roughly 3,000 jobs a month). BLS analysts also thoroughly investigated the seasonal adjustment of the Current Population Survey data over the course of the recovery (Evans and Tiller 2012). This inquiry showed that alternative assumptions regarding seasonal adjustment did not meaningfully affect estimates of the unemployment rate since 2007.
While Sherk and Furth’s conjecture predates the release of the ERP, the referenced Kropf and Hudson article predates the Sherk and Furth issue brief by several months. The Kropf and Hudson (2012) article is excellent reading, and I am glad the Heritage Foundation’s random musings drew me to the paper (it saved me a lot of work trying to tease out the seasonals myself!). Chart 2 is particularly illuminating:
Source: Kropf, Jurgen, and Nicole Hudson. 2012. “Current Employment Statistics Seasonal Adjustment and the 2007–2009 Recession.” Monthly Labor Review 135, no. 10: 42–53.
The straightforward (and I believe correct) interpretation of the graph is that excluding the Great Recession from the sample period over which the seasonals are estimated does not have a large impact that is consistent with the Sherk and Furth conjecture.
None of the foregoing means that employment growth won’t decrease (I expect it very well might with the sequester going into effect [1]). However, if employment growth decreases, it won’t likely be attributable to a statistical artifact associated with the seasonal adjustment process.
I’m curious. The first quote speculates about the BLS. Didn’t they ask? My first inclination would be to call and ask. Given a mild anomaly that has only existed for a few years, I’d be curious rather than judgmental; by writing speculation, the implication is the BLS are idiots who can’t do seasonal adjustments. And that ignores how they revise both reports and forecasts. I don’t understand that kind of thinking.
In a related vein, aren’t the issues underlying this – other than any politics – questions of model integrity? I see this play out often, including here, with arguments about the validity of models. I see models questioned for their statistical work, which is one part of the coin – and which is my first paragraph above. I guess that’s natural, especially since so many models are based on lies. But assuming models work properly internally, I’m reminded of an old metaphor, which I’ve seen resurface lately, for larger validity, that of flipping a coin to decide whether it’s 2 sided. The model of flipping may be fine but the method is a lousy test related to the final question.
Jonathan: In answer to your first question, I don’t know the answer. It could be they would prefer to be uninformed in a way that is conducive to being negative in their assessment; it could be they do not have access to ARIMA X-12, which is built into many statistical software packages, but not for instance in Excel.
Menzie Well, X-12 ARIMA is free (note…the Census just switched to X-13 ARIMA), as is Eurostat’s TRAMO-SEATS. And Eurostat just came out with a new and freely available version of their Demetra program that automates the diagnostics for batch seasonality adjustments across hundreds of different variables.
2slugbaits: Well, perhaps the Heritage Foundation’s budget could not handle the cost…
Menzie,
I posted too early in another thread. Here is the Heritage Foundation’s analysis of the Democrats budget. Perhaps you could tell us why it is so much better than the Republican budget.
Seasonal patterns vary over time for a variety of reasons which are often just viewed as background noise. But if there are known influences that suddenly shift the pattern, then your choices are to adjust the data for the shift or to let the data ride for several cycles and let the software adjust the seasonal patterns.
The first choice is subject to error in estimating the extent of the shift versus the variations in normal change. The second choice works, but provides less than stellar seasonally adjusted numbers until the new pattern is recognized. I’d go with the first method if the change were known to be a one-time event and the input value could be calculated fairly well. Otherwise, I’d let the data ride and after a few years, the corrected pattern will emerge.
has anyone looked at seasonally adjusted housing data? zero hedge has been running comparisons on starts and new home sales & they sure looked distorted to me (though i dont buy their theories)…
here’s an example: http://www.zerohedge.com/news/2013-02-20/chart-day-housing-starts-adjusted-vs-unadjusted-redux
since new starts & sales were 5 times greater during the boom than the bust, would that be enough to throw recovery the adjustment off?
Are Democrats really liars?
Last night Sen. Jeff Sessions proposed the following amendment:
Mr. Sessions moves to commit S. Con. Res. 8 back to the Committee on the Budget with instructions to report back no later than March 22, 2013 with such changes as may be necessary to achieve unified budget balance by fiscal year 2023.
Now we may debate whether balancing the budget will actually do anything but the Democrats have been very clear in their support for a balanced budget. Here is a list of Democrat senators and dates they voiced their support for a balanced budget.
SEN. SHERROD BROWN on 11/01/06
SEN. DEBBIE STABENOW on 10/22/00
SEN. MARK BEGICH on 04/15/11
SEN. BILL NELSON on 03/29/11
SEN. BEN NELSON on 03/04/11
SEN. MARK UDALL on 02/01/11
SEN. MICHAEL BENNET on 03/06/11
SEN. CLAIRE McCASKILL on 06/29/11
SEN. KIRSTEN GILLIBRAND on 06/04/07
SEN. TOM CARPER on 06/14/04
SEN. HARRY REID on 02/12/97
SEN. MARY LANDRIEU on 02/25/97
SEN. DIANNE FEINSTEIN on 02/26/97
SEN. TOM HARKIN on 02/10/95
SEN. TIM JOHNSON on 10/26/95
SEN. MAX BAUCUS on 02/10/95
SEN. DICK DURBIN on 01/11/95
SEN. JON TESTER on 06/25/06
SEN. BOB CASEY on 09/03/06
They all voted no on Sen. Sessions amendment. We couldn’t be so cynical as to say that they lied to get elected and now are voting their true positions could we? Perhaps they have just grown in their jobs.
Oh, I forgot to mention Sen. Joe Manchin. He also supports a balanced budget, but he is not a liar. He voted for the Sessions amendment. I guess he is still growing.
Eternal problem of correlations and causations, but still with improving employment, the Federal surplus/ Deficit as a percentage of GDP is a difficult gauge. When attempting to corroborate an increasing tax receipts and a better employment, this graph is not validating a strong trend.
Federal Surplus or Deficit [-] as Percent of Gross Domestic Product (FYFSDFYGDP)
2011: -8.68795 Percent of GDP
Still few economists are seeing improvements and discarding the effect of public debts on GDP growth.
Bloomberg “Economists See No Crisis With U.S. Debt as Economy Gains”
As often debated through Econbrowser posts and comments the zoological cycles and the geological ones do not share the same time.
Perhaps they finally figured out that balancing the budget on an arbitrary date is a really stupid idea. Everybody starts out ignorant, some people learn, some people, like Sessions, stay ignorant.
BLS states on their website they use ARIMA X-12 so there’s that.
Bruce Hall Seasonal patterns vary over time for a variety of reasons which are often just viewed as background noise.
Be careful that seasonal patterns aren’t wandering about because of a seasonal unit root.
Menzie You forgot the stupidest part of the Heritage Foundation report:
Policymakers should not worry that sequestration will disrupt growth. Economic research shows that spending-based deficit-reduction plans, unlike tax-based plans, do not damage the economy.
A little off topic, but only a little. I’ve never been comfortable combining seasonally adjusted data across different variables; e.g., in a VAR or OLS regression. Whenever practical I prefer to use NSA data and adjust for seasonality in the model. I have a couple of concerns. The first is that many time series are aggregates of seasonally adjusted components, and a seasonally adjusted aggregate of seasonal components is not guaranteed to capture the same information as a a seasonally adjusted aggregate of NSA components. Another problem is that the filtering and smoothing process imposes different trend factors, which might be hard to explain if the model you’re developing assumes a common trend across all of the variables. Finally, if you’re doing something like an error correction model what you’re mainly interested in are the short-run dynamics of the deviations. I’ve always worried that the filtered and smoothed nature of seasonally adjusted data purges the very deviations that are of greatest interest. Any thoughts? Are my concerns misplaced?
2slugbaits: I don’t think your concerns are misplaced; I was taught that it was best to estimate using NSA data, and include seasonal terms to account for seasonality. However, this is often impractical because NSA GDP is not reported, for instance. In my experience, the impact has not been too large, when I could estimate using both SA and NSA data. However, this has mostly been in the context of exchange rate modeling, where the volatility of the LHS variable is very high (and does not typically exhibit seasonality).