We’re well into a severe housing downturn by every measure except for the number of people working in residential construction.
The figure above is similar to one that has often been highlighted by one of our favorite bloggers, Calculated Risk. The graph shows the relation between the number of new housing units completed each month and the number of people employed in residential construction. The latter is calculated as the sum of the BLS series for those working in residential building (currently a little under a million people) and residential specialty trade contractors (over two million). It is puzzling that the decline in the number of people working in residential construction has so far been quite modest relative to the decline in the number of units completed.
One plausible explanation that is sometimes offered is that many of those laid off in the current episode may be undocumented immigrants. Although there must be some truth to that claim, it is not correct to assume that undocumented workers do not show up in the BLS series. According to the Wall Street Journal, the IRS has issued 11 million individual taxpayer identification numbers to people without a social security number. In practice, the penalties for hiring an undocumented worker appear to be lighter than the penalties for failing to withhold taxes on your workers. As a result, government counts of the number of people working may be more accurate than one might have supposed.
Another possibility is that some of those firms that had previously been classified by the BLS as involved primarily in residential construction are perhaps now being sustained at least in part by commercial construction. Nominal spending on nonresidential construction was up 15% in 2007:Q1 relative to 2006:Q1 (BEA NIPA Table 1.1.5).
Or, some may conclude that residential construction firms are simply postponing the inevitable layoffs, although it is unclear what these firms’ calculations and motivations would be.
I wanted to do some statistical analysis of this relation, but unfortunately the BLS does not provide the values for specialty trade residential construction employment prior to 2001, and 7 years is not enough data for the questions I’d like to ask. However, the series for residential building employment, though only 1/3 of total residential construction employment, goes back to 1985 and seems to have some similar properties. Certainly one is again struck by the failure of this series to fall as sharply as completions over the last six months:
Let yt denote the residential building employment (excluding specialty trade contractors) in month t and let ct denote the number of completions in month t. An OLS regression over 1985:1 to 2007:3 yields the following estimates (t-statistics in parentheses):
This appears, both to the eye and to the regression, to be a very significant relation. But one feature that can fool both the eye and the regression is the fact that employment yt has a tremendous amount of persistence– its level last month is an extremely important predictor of where it is likely to be today. The same is true of completions ct. When one regresses one such variable on another (or just plots them together in a graph, inviting your eye to do the regression for you) it is very easy to mistake the common swings in the two series for a reliable predictive relationship.
One way to investigate the magnitude of these concerns is to examine the serial correlation of the residuals of the above regression, which we can do with a second regression of these estimated residuals on their own lagged values:
These coefficients sum to 0.96, meaning that when the level of employment deviates from the value predicted by the relation, it takes a very long time to return. In fact, one accepts the null hypothesis that these coefficients sum to 1.0, in which case a discrepancy between employment and completions in the original regression would never be completely eliminated. To get a sense of what this could mean, the blue line on the graph below plots the actual residuals through March 2007. At that time there were several hundred thousand more people working in residential building than one would have predicted on the basis of the number of homes completed. The green dashed line indicates what those residuals would be if from here they follow the dynamic pattern of the above equation. Even two years from now, we might still be seeing 100,000 more workers than predicted by the historical regression, and must accept the hypothesis that the true discrepancy could end up much bigger than that.
How could there be a drop in the number of homes completed with residential construction employment never following it down? In addition to the issues about how we measure the number of people working in residential construction employment noted in the beginning of this post, there are changes each month in the types of units that are built and the means by which they are constructed. All of these changes would cause discrepancies between the numbers for residential construction employment and houses built in a given month. The statistical evidence is that these discrepancies can take a long time to be eliminated, and indeed may never be.
So is residential construction employment about to follow completions off the cliff? Maybe. Or maybe not.