An Experimental Component Index for the CPI: From Annual Computer Data to Monthly Data on Other Goods. (Preliminary and Incomplete)

Until recently the Consumer Price Index consisted solely of “matched model” component indexes. The latter are constructed by BLS personnel who visit stores and compare prices of goods with the same set of characteristics over successive periods. This procedure is subject to a selection bias. Goods that were not on the shelves in the second period, and hence whose price comparisons were discarded, were disproportionately goods which were obsoleted over the period, and consequently represented goods whose prices were falling. Pakes (2003) provided an analytic framework for analyzing this selection effect and showed that it could be partially corrected using a particular hedonic technique. Using personal computer data he showed that the hedonic correction could be substantial. The BLS staff has recently increased the rate at which they incorporate techniques to correct for selection effects in their component indexes. However their work and the work of other researchers shows very little difference between hedonic and matched model indices for other (non computer) components of the CPI. This paper explores why. We look carefully at the data on the component index for TV’s and show that differences between the TV and computer markets, together with the fact that the BLS data are high frequency, imply that to obtain an effective selection correction we need to use a more general hedonic procedure than has been used to date. The computer market is special in having both well defined cardinal measures of the major product characteristics, and exiting goods with relatively low values for them. In markets where such measures are absent and where turnover can be at the high quality end, we need to allow for selection on unmeasured, as well as measured, characteristics. We develop a hedonic selection correction that accounts for these phenomena and show that when applied to TVs it yields much larger selection corrections. In particular we find that matched model techniques underestimate the rate of price decline by over 20%. Moreover the BLS staff’s recent successful push to modernize their data gathering procedures has made it possible to compute our index within the BLS’s time constraints, making it a “real time” alternative to current procedures. ∗We would like to thank Teague Ruder and Aylin Kumcu for excellent research assistance, Paul Liegey for his patience in explaining the BLS’s algorithm for constructing the TV component index to us, Dale Jorgenson and Jack Triplett for helpful comments, and the BLS NSF and Tolouse Network for Information Technology for the support that allowed us to undertake this project. All errors and omissions are those of the authors and not of any of these institutions. In particular Table 8 contains empty entries because we have not yet finished all the calculations.