to product characteristics and time, and then use the results to obtain estimates of the price changes keeping the characteristics constant. For our application, we estimate the following hedonic price regression

05:3

ln p_{it }= βx_{it }+

X

γ t d t + ε i t

(1)

t=93:2

where p_{it }is the price of regimen i in quarter t, x_{it }contains the attributes of each regimen, including effectiveness attributes such as survival rate, time-to- progression, and response rate, a second line treatment indicator, the interactions between the effectiveness attributes above and the second line indicator, grade 3 or 4 side effects such as diarrhea, nausea, abdominal pain, vomiting and neutropenia, and an indicator of whether the regiment contains a tablet or not. The time dummies denote quarters and the base quarter is the first quarter of 1993.

# The change in prices is obtained from the parameters of the time dummies as

P I t

P I_{t }_{1 }

P I_{t }_{1 }

= e x p ( γ t

γ t 1 )

1

and therefore,

P I t

= e x p ( γ t

γ_{t 1 }

)P I_{t }_{1 }

The parameters β in equation (1) have been referred to as “implicit prices”, however, the theoretical foundations of the hedonic surface are not clear. Pakes (2005) argues that in oligopolistic markets, price is a combination of marginal costs and markups, and therefore the hedonic price regression constitutes a re- duced form or projection of these markups on the characteristic space. In the in- dustry we study, given the high costs of R&D relative to marginal cost, markups should be large and Pakes’s insights are applicable. The parameters β do not nec- essarily follow any restriction based on economic theory. That is, the β parameters may have the “wrong” sign. For example, if there is more entry of products into a particular region of the attribute space that has more desirable characteristics for the average consumer (e.g., survival), this will reduce the markups of these products, and allow the product with less desirable characteristics for the average

9