Revisions to Employment and Retail Sales and What They Mean
What's more exciting than the NCAA Final Four?
Data revisions! (well, not really, but for those of you who follow economic indicators closely, data revisions are important).
From July 2010 to December 2010, the estimates of private payroll creation have been revised up an average of 56,000. On the one hand, these revisions are small relative to a base of about 130 million workers on payrolls. On the other hand, these revisions are large relative to the initially reported private sector job gains (the number that received a lot of attention) which averaged 87,000 during these 6 months. So although the headlines emphasized the numbers that made up the 87,000 average, subsequent revisions suggest that private payrolls increased an average of 143,000 per month.
As an example, let’s go back a month in time. The headline from last month’s jobs report was that nonfarm employment (private and public) rebounded from a snow-covered January, rising by 192,000 jobs in February. However, below the headlines, we also learned that December and January data were revised upward by a combined 58,000 jobs—continuing a string of positive revisions to payroll employment. Adding together the 192,000 new jobs in February PLUS the 58,000 “new” jobs that appeared in the previous two months due to revisions, we actually found that there were a quarter million more jobs in the economy than we had previously thought. Was this figure part of the jobs report news cycle? No (well, I know of one reporter who caught this point). But revisions should be a part of the discussion, whether we’re talking about the jobs report or any other economic indicator. So let’s spend a moment on them.
Before jumping to conclusions about revisions, let me stress that a lot of analysis has been conducted on the issue of how data is originally reported (referred to as “real-time data” in the biz) and how it is subsequently revised. The story about revisions to economic indicators is complicated, however. The Bureau of Labor Statistics (BLS), like the Census Bureau and the Bureau of Economic Analysis, go to great lengths to balance data timeliness, accuracy and limited resources. Timeliness generally translates into publishing initial economic estimates using an incomplete sample, and as additional data reports come in and the sample becomes more complete, estimates of jobs, retail sales, GDP, and other indicators change and become more accurate. Revisions are unpredictable. If they were predictable, the statistical agencies would change their data crunching methodology to capture those predictable elements. A whole field of economic literature on “real-time data analysis,” much of which takes place at the Philadelphia and St. Louis Federal Reserve Banks, explores these data revisions at great lengths. For folks interested in a great survey of the academic literature on the topic, I recommend Dean Croushore’s article in the latest Journal of Economic Literature. An earlier working paper is also available. What experience does tell us is that the statistical agencies have struck the right balance between timeliness and accuracy. Revisions will change the numbers, but rarely do they change the underlying narrative.
Getting back to the jobs data, a quick glance at Figure 1 shows that revisions do appear to be all over the place. Figure 1 charts revisions not just to nonfarm employment published by BLS, but also to retail sales published by the U.S. Census Bureau here at the Commerce Department. It is important to note that there are many flavors of revisions, and statistical agencies use various (and not necessarily consistent) nomenclature to identify which data are “preliminary,” “advance,” “revised,” “benchmarked,” etc. Here we are just focusing on initial revisions to the data. How much did the headline figures on nonfarm jobs and retail sales change between when they were first published and when they were revised over the following two months? To understand what the figure shows us, let’s focus on the last point on each line, which corresponds to December 2010.
When the December jobs data were first released on January 7, BLS estimated that nonfarm employment had risen by 103,000 jobs. The following month, BLS published a revised estimate of 121,000, and last month, BLS further revised the figure up to 152,000 jobs. Thus, the revision from the first to third published figure was 152,000 - 103,000 = 49,000 jobs. (To its credit, BLS makes it easy to track these revisions by publishing them directly on its website.) For retail sales, the first estimate in December showed 0.6-percent growth, which was subsequently revised down to 0.5 percent and then back up to 0.6 percent. Thus, there was essentially no revision from the first to third published figure for December.
On average, both series show hardly any revisions at all. From 2008 through 2010, the retail sales revision rounds to zero, and the employment revision was just 7,000 jobs downward. That’s just 7,000 out of a base of more than 130,000,000 jobs, or 0.005 percent. This examination supports the idea that revisions are generally a) small, in relative terms; b) unpredictable, and c) random. However, a longer look at Figure 1 hints that there may be some underlying pattern to these revisions. Notice that revisions tended to be negative in 2008, when the economy was deep in recession. They were mixed in 2009 as the economy began to emerge from the recession. And, they were mostly positive in 2010 when the recovery began to strengthen.
You can see this pattern more clearly in Figure 2, which shows the entire last decade — including two recessions — and smoothes the data’s considerable month-to-month volatility by using 12-month averages of the revisions. This information strongly suggests that revisions to both employment and retail sales have patterns that follow the business cycle. Furthermore, revisions to these series appear to have been especially sensitive to the last recession.
As Dean Croushore’s article makes clear, the how's and why's of revisions are complex. They can involve, among other things, the blog-unfriendly topic of concurrent seasonal adjustment. Suffice it to say that the latest recession and recovery have put our economic statistics under a strong stress test and given our statistical agencies evidence of how well they capture data in real time. Overall, statistics have performed admirably well, and agencies have already begun to apply lessons learned from the past few years.
On Friday, we all hope that forecasters’ expectations of strong growth in nonfarm jobs in March are met, if not exceeded. But I will also be looking at any revisions to previous months in hopes that the news about jobs and the economy is even better than we originally thought.
~Mark Doms, Chief Economist, U.S. Department of Commerce
March 30, 2011
