This thread over at metafilter might be of interest.

Snip:

I spent four years as an economist at the BLS, in the office responsible for researching methodologies for calculating the CPI. I worked on the article's principle target: hedonic adjustments to the CPI. The CPI uses code I wrote to make hedonic adjustments. ...

Hedonic adjustments attempt to account for changes in product quality over time. The CPI (basically) measures the cost of purchasing a representative "basket" of goods, and the amount by which the cost of the basket changes over time is reported as "inflation". But some goods in the basket disappear as they are discontinued and replaced by newer models. Think audio equipment: in the '80s, the basket might have included your favorite cassette player. When CDs gained market share, your cassette player might have been discontinued. This poses a problem for calculating inflation - the old basket of goods cannot be obtained at any price if the good is no longer available for sale, and inflation is technically infinite. So the CPI generally just replaced the old good with a newer model - say, your new favorite CD player. Problem: the CD player costs more (indicating positive inflation), but it also provides characteristics preferred by most consumers. Because most CPI users (mistakenly) think of the CPI as a cost of living index, which should measure the cost of obtaining a certain level of "happiness" or "satisfaction" or "utility" for a representative consumer, the CPI was actually overstating the increase in the cost of living due to its failure to account for the increase in quality of the market basket. A pure cost of goods index is rendered useless by the disappearance of one of its goods; a cost of living index is overstated if the change in goods does not account for a change in quality of goods.

The principal reason the CPI did not adequately account for changes in quality of goods is that the data required to estimate those changes was unavailable until retail scanner data became economically obtainable. You need a huge amount of price and characteristic data to estimate hedonic values. The BLS simply could not obtain such data much before the '90s, and could not have processed the data in a timely fashion, either. And each product type requires detailed studies on which characteristics are important, measurable, and prevalent enough to gather product pricing data for estimating an hedonic model (not to mention extensive modeling exercises to develop the best - or even an acceptable - hedonic model).