We are unable to display your institutional affiliation without JavaScript turned on.
Browse Book and Journal Content on Project MUSE

Find using OpenURL

Comments and Discussion

From: Brookings Papers on Economic Activity
Fall 2012
pp. 299-315 | 10.1353/eca.2012.0022

In lieu of an abstract, here is a brief excerpt of the content:

Comment by David Laibson

It's very impressive to have called the housing bubble a few years before it popped. It's even more impressive to have conducted a perfectly timed series of housing surveys that anticipated the bubble. These authors are prophetic.

This paper by Karl Case, Robert Shiller, and Anne Thompson offers key insights into the thinking of participants in the housing market. The survey findings are required reading for anyone trying to explain the extraordinary up-and-down price movements that have whipsawed homeowners, homebuilders, banks, and by extension, most of the world economy.

There is much to praise in the remarkable survey studied in this paper, but I will focus on a survey design problem that leads me to reinterpret a few of its findings. Stated briefly, the 10-year forecasting data need far more trimming than the 10 percent trim (discarding the top and bottom 5 percent of the data) that the authors adopt.

In most survey data a 10 percent trim would be adequate—often more than adequate—to remove outliers. The survey data used here, however, suffer from an unusual bias that is not corrected with a 10 percent trim. Some fraction of the respondents—let's say a quarter for now—appear to be confusing the concepts of annualized returns and total returns. For the 1-year forecast data, these two concepts are the same, so that the bias that I am highlighting does not apply. But for the 10-year forecast data there is an enormous difference between an annualized return (what the survey question asks for) and a total return (what perhaps a quarter or more of the subjects are thinking).

To see how this bias works, consider the following simplified example. Suppose that all respondents believe that housing will appreciate 3 percent per year for the next 10 years. Assume as well that three quarters of the subjects respond correctly to the question about 10-year returns, giving an answer of 3 percent, and one quarter give an answer of 30 percent (the total 10-year return, ignoring compounding to keep things simple). Then the mean response is an "annualized return" of 9.75 percent per year, more than triple the subjects' true belief. In this simple example, the researchers would need to trim 50 percent of the data to unbias the mean (since the bias is not symmetric).

Four related empirical facts about the survey results point to the existence of this bias:

  • —   Some subjects give an answer for the 10-year annualized return forecast that is exactly 10 times the answer they give for the 1-year forecast (see the authors' appendix).

  • —   A substantial fraction of the answers to the 10-year forecast question are so high that they are far more likely to be total returns than annualized returns. For example, my table 1 shows that at least 10 percent of the 2004 survey respondents say (if you take their answers literally) that housing prices will appreciate 50 percent per year for the next 10 years, implying a total price appreciation of 5,670 percent. Note that the authors' 10 percent trim removes only half of these respondents from the calculation of the mean.

  • —   The 10-year annualized forecasts are far more right skewed than the 1-year annualized forecasts (see table 1 and the authors' appendix figures A.1 and A.2).

  • —   The mean 10-year forecast exceeds the mean 1-year forecast (both using the 10 percent trim) in all study years. In most years this gap is considerable (appendix tables A.1a and A.1b).

It is hard to know what to do about this bias. For some of the reasons discussed above, I believe that the authors' 10 percent trim is not adequate. Indeed, I think that a quarter or more of the data may be corrupted by respondent confusion about annualized versus total returns.

Consider the following suggestive evidence. The 1-year forecast mean is unaffected by trimming. Whatever the trimming parameter used (10, 15, 20, 50, or 100 percent), the 1-year forecast "mean" barely budges (see appendix table A.1a). But for the 10-year forecast mean, each additional...

You must be logged in through an institution that subscribes to this journal or book to access the full text.


Shibboleth authentication is only available to registered institutions.

Project MUSE

For subscribing associations only.