For many classes of goods—and perhaps especially services—it can be extremely difficult to identify which characteristics are actually associated with price. Despite the early success of hedonics to move quality adjustment in the CPI toward a statistically based approach, considerable judgment by researchers is still required. For instance, early introduction of video memory as an explanatory characteristic in regressions for PCs yielded “implausibly high coefficient values,” so the variable was left out of initial specifications. Later the values “settled down and behaved much more reasonably” and the feature was included in the specification (Fixler et al., 1999:11). Likewise, in the hedonic model for TVs developed by Moulton et al. (1998:10) the “stereo sound” indicator was dropped because it predicted a negative (though insignificant) impact on price. Given the short history of this type of research at BLS, it is not clear what the benchmark should be for assessing what is or is not reasonable. Strange-looking variable coefficients could be indicative of larger problems—including omission of key value indicators, characteristic mismeasurement, and functional form issues.
Whether for a standard comparability decision or for hedonic modeling, identifying and quantifying relevant characteristics is tricky when quality is tied to consumer perceptions that may not be constant over time. Also, it is next to impossible to collect full and timely information on certain types of product characteristics. Quality of fabric in clothing, for example, is determined by a complicated combination of characteristics—not simply by material type, but also by threads per inch, type of weaving, quality of dye, etc. Given the changing nature of fashion, a characteristic may be viewed as a negative quality at one time and as a positive quality at another. For instance, the original move from cotton to synthetic shirts was considered a quality improvement—but so too was the move back to cotton.
Once identified, it is not necessarily any simpler to measure the characteristics thought likely to affect price: consider stylishness in clothing or handling in cars. Even for the best candidates, such as computers, attribute measurement can be problematic. For instance, how does one quantify the user friendliness of hardware or software? For most products, certain elements that contribute to its value will always be difficult to measure consistently.
Theory provides little guidance to help determine the appropriate functional form for hedonic equations. Experience suggests that characteristics often interact in complex ways to affect value. When characteristics work in combination, nonlinear functional forms, perhaps involving interaction variables, must be used to produce reasonably robust results.35 Furthermore, when one product works as
Curry et al. (2001) summarize some of the advantages of flexible functional forms (and even neural networks) in the context of hedonic modeling applied to consumer goods. The authors use detailed scanner data to estimate and test hedonic models with interaction effects for the U.K. television market.