Forget the Phantoms, the Real Districts are Behind

Much attention was paid last week the errant phantom districts created by the federal administrators of the stimulus bill. Online policy watchers at Watchdog.org have done a good job of counting the 30,000 jobs “created or saved” in 440 phantom districts created by coding errors.

Since then the top level information on the Recovery.gov website has been changed so that districts are now compiled as “unassigned congressional district.” The errors remain in the downloadable data files.

The phantom districts don’t really matter, except that they undermine the data. These simple code errors reveal a lack of oversight. As Ed Pound, director of communications for the Recovery Accountability and Transparency board, told Watchdog, “Our job is data integrity, not data quality.”

It’s this lack of attention that allows for massive over-reporting from Georgia to Colorado, to name a few. These repeated errors cast doubt upon the entire dataset.

Most economists will tell you that Congressional district is a terrible breakdown for job analysis. The choice to mandate such reporting the stimulus bill was likely more to garner bragging rights about local job generation. While low in economic meaning, those numbers could be powerful talking points next November. Despite potential inflation, most districts have not made surprisingly large job gains.

When the stimulus bill was being debated, district by district job creation estimates were created based on a wider analysis by CEA Chairwoman Christina Romer and Jared Bernstein. Those estimates were for the full two years after passage and most districts are far from reaching the predicted levels. Yet 16 districts have already surpassed their two year creation expectations. Of the 16, 13 are represented by Democrats, most notably; California’s 5^th district in Sacramento and represented by Representative Doris Matsui (D) has already produced 12 times as many jobs as predicted. Even without those outliers, Republican districts report on average 521 jobs compared to the average 679 reported in Democratic districts. Of course, it could be a phantom difference.

Be prepared to work more

The Federal Reserve Bank of Philadelphia released its forth quarter suvery of forecasters today. The survey predicts that economic growth will be stronger than expected in coming years and that unemployment will be higher compared to last quarter's survey. Those two factors combined are likely to mean more work for those Americans who are working.

The basic conception of national output (Y) is usually given as Y=f(K, AL). This means that output is a function of capital (K) and labor (L) with a multiplier for technological advancement (A). The new Philadelphia numbers indicate that L is going to continue to decline for some time. The capital stock, K, seems unlikely to rise anytime soon. As the figure below shows non-defense capital goods orders have declined drastically this year. While the figure appears to be leveling off, it will likely take some time to regain lost levels.

This leaves A to drive growth. Increases in A are by no means synonymous with more work. During the tech bubble A increases were the result of things like e-mail and internet access. In the present recovery, the growth is likely to come from one place: hours 1. Of course hourly increases are not going to be distributed evenly. The average workweek for nonsupervisory production works is 33 hours a week, its lowest level since WWII. That of course is linked to the lower panel shown above, with lots of unfilled orders businesses just are not producing right now.

The situation may not be all bad news. The New Yorker ran a story in March that those who are working are seeing steady wage increases. Plus research out this year indicates that it’s the un- and under-employed who are the most likely to experience depression. Happy work hours.

1. We could of course think of more hours as an increase in L. I’ve eschewed that here so as to separate out the impact of number of people employed and work per person employed.

More Marathon analysis

Several weeks ago I posted upon the results of a half marathon that I recently ran. Since then, the marathon season has run it's biggest races of the fall, including Chicago, New York, and DC's Marine Corps. I haven't had the pleasure of running any of these trials (I'm hoping for a debut next fall). I've been encouraged to see that more analysis has come from watching those races.

One of the most past along was this post by Paul Kedrosky of Infectious Greed asking why the wining finish times at New York have been less volatile overtime. The post provides a great question with little answer. I've discussed this disparity with several runners (including one who finished in the top 400 males runners at Boston this year) and several hypothesis have struck us.

I. Weather. The New York and Boston marathons are run at different times of the year and weather maybe in more flux in the Boston spring than the New York fall.

II. Competing Races. The New York marathon competes with Chicago, DC, and Detroit for runners. Boston competes with London, in 2010 the two are a week apart. As a result New York may attract a wider international field and Boston may have a more American appeal.

III. Prize money disparity. I've checked and this explanation seems unlikely. Boston paid out$150,000 to winning runners and New York paid out $130,000 to winners. Unless the evolution of the two races' prizes has been different, I'd expect that Boston with the higher prize would have less volatile times as it should consistently attract top talent.

Analysis of this type is a test, evaluate, and reject kind. There are lots of possible reasons but probably only a few that hold up to rigorous analysis.

Speaking of rigor, I've done a bit more work on the half-marathon I ran and owe readers an update. I originally provided a simple linear regression on gender and age. I tested other specifications and did not get any more significant explanitory power. Well, at least not statistically. Here's a case where common sense needs to play an important role. The model that I originally reported implies that someone of the lowest age should run fastest, for example that a 8-year old boy should best a 21-year old man. This makes little sense. Instead the model should and now does include a variable for the square of age.

The new regression is shown below.

This regression is only for men. The linear line is the original formulation. The upward-sloping curve is the squared regression. This new model has two important implications. First, it estimates that a man should run his fastest time at 23.4 years of age, eliminating the problem of superfast pre-teens. Second, at older ages, beginning around age 55, the curve increases much faster than the line. The squared model indicates that at each age above 23 each additional year of age adds more time to finish than the last year. As a result the move from age 60 to 61 is much more significant than 24 to 25.

Also plotted on the figure is my finish time. In my modest defense, I'd like to say that the models are best fit. The full data shows that I'm still within a dense cloud of finish times. I'm taking all this as an incentive for improvement rather than a reason to retire.

Depression Fatigue?

I wrote over the summer about the manic consumer mood during recessions. My argument then was that by calling a recession, policymakers could send consumer confidence into a tailspin.

With the current recession creeping toward it's second full year (even if NBER calls June the end of the downturn, as some predict, that will make the recession 18 months), this is the longest contraction since the Great Depression. Calling a recessions is academic for this instance. Today, I want to know what happens in long recessions. The figure below charts the unemployment rate (the BLS' headline U-3) and the ICS (the University of Michigan and Reuters confidence index that I used before.The model that I provide is hyperbolic, allowing it to take the curved shape in the blue regression line. The model, which accurately predicts the current ICS from the September unemployment rate, indicates that small changes in the unemployment rate drag on confidence more than at high levels.


The model is far from perfect because unemployment is one of the last indicators to recover form a recession. Yet unemployment tends to peak at the end, or just after the end of a recession. So we can use high unemployment levels as as a proxy for the length of a recession.

People appear to be most respond emotionally to the beginning of a contraction. We're seeing this now. Chairman Bernanke called the recession "officially over" recently and the Fed said in it's FOMC statement yesterday that consumer spending is "expanding."


Oh for the wonks, the model specificaiton is:
ICS=187 (1/unemployment rate)+54
R2=.25
*The model outperforms ln, linear, or quadradic specifications
Data from 1978 to present