Thursday, July 17, 2014

Thoughts on the Fed's New Labor Market Conditions Index

I'm usually one to get excited about new data series and economic indicators. I am really excited, for example, about the Fed's new Survey of Consumer Expectations, and have already incorporated it into my own research. However, reading about the Fed's new Labor Market Conditions Index (LMCI), which made its debut in the July 15 Monetary Policy Report, I was slightly underwhelmed, and I'll try to explain why.

David Wessel introduces the index as follows:
Once upon a time, when the Federal Reserve talked about the labor market, it was almost always talking about the unemployment rate or the change in the number of jobs. But the world has grown more complicated, and Fed Chairwoman Janet Yellen has pointed to a host of other labor-market measures. 
But these different indicators often point in different directions, which can make it hard to tell if the labor market is getting better or getting worse. So four Fed staff economists have come to the rescue with a new “labor markets conditions index” that uses a statistical model to summarize monthly changes in 19 labor-market into a single handy gauge.
The Fed economists employ a widely-used statistical model called a dynamic factor model. As they describe:
A factor model is a statistical tool intended to extract a small number of unobserved factors that summarize the comovement among a larger set of correlated time series. In our model, these factors are assumed to summarize overall labor market conditions. What we call the LMCI is the primary source of common variation among 19 labor market indicators. One essential feature of our factor model is that its inference about labor market conditions places greater weight on indicators whose movements are highly correlated with each other. And, when indicators provide disparate signals, the model's assessment of overall labor market conditions reflects primarily those indicators that are in broad agreement.
The 19 labor market indicators that are summarized by the LMCI include measures of unemployment, underemployment, employment, weekly work hours, wages, vacancies, hiring, layoffs, quits, and sentiment in consumer and business surveys. The data is monthly and seasonally adjusted, and the index begins in 1976.

A minor quibble with the index is its inclusion of wages in the list of indicators. This introduces endogeneity that makes it unsuitable for use in Phillips Curve-type estimations of the relationship between labor market conditions and wages or inflation. In other words, we can't attempt to estimate how wages depend on labor market tightness if our measure of labor market tightness already depends on wages by construction.

The main reason I'm not too excited about the LMCI is that its correlation coefficient with the unemployment rate is -0.96. They are almost perfectly negatively correlated--and when you consider measurement error you can't even reject that they are perfectly negatively correlated-- so the LMCI doesn't tell you anything that the unemployment rate wouldn't already tell you. Given the choice, I'd rather just use the unemployment rate since it is simpler, intuitive, and already widely-used.

In the Monetary Policy Report, it is hard to see the value added by the LMCI. The report shows a graph of the three-month moving average of the change in LMCI since 2002 (below). Values above zero are interpreted as an improving labor market and below zero a deteriorating labor market. Below the graph, I placed a graph of the change in the unemployment rate since 2002. They are qualitatively the same. When unemployment is rising, the index indicates that labor market conditions are deteriorating, and when unemployment is falling, the index indicates that labor market conditions are improving.


The index takes 19 indicators that tell us different things about the labor market and distills the information down to one indicator based on common movements in the indicators. What they have in common happens to be summarized by the unemployment rate. That is perfectly fine. If we need a single summary statistic of the labor market, we can use the unemployment rate or the LMCI.

The thing is that we don't really need or even want a single summary statistic of the labor market to be used for policymaking. The Fed does not practice rule-based monetary policy that requires it to make policy decisions based on a small number of measures. A benefit of discretionary policy is that policymakers can look at what many different indicators are telling them. As Wessel wrote, "the world has grown more complicated, and Fed Chairwoman Janet Yellen has pointed to a host of other labor-market measures." Yellen noted, for example, that the median duration of unemployment and proportion of workers employed part time because they are unable to find full-time work remain above their long-run average. This tells us something different than what the unemployment rate tells us, but that's OK; the FOMC has the discretion to take multiple considerations into account.

The construction of the LMCI is a nice statistical exercise, and the fact that it is so highly correlated with the unemployment rate is an interesting result that would be worth investigating further; maybe this will be discussed in the forthcoming FEDS working paper that will describe the LMCI in more detail. I just want to stress the Fed economists' wise point that "A single model is...no substitute for judicious consideration of the various indicators," and recommend that policymakers and journalists not neglect the valuable information contained in various labor market indicators now that we have a "single handy gauge."

2 comments:

  1. I can see the usefulness of the index (sans wages for the reasons you cite) as a sort of defense from those who are always saying, "yes but the REAL unemployment rate is XX."

    ReplyDelete
  2. great post! factor models are attractive but their usefulness depends partly on what they're used for. one of the differences in factor and component extraction is the source of common variance (communalities/uniqueness). it would be nice to see the extraction details as well as the loadings/rotations used to better interpret the signals that each variable send out through the index.

    ReplyDelete

Comments appreciated!