consumer (e.g., stronger side effects) to exploit monopoly power among the con- sumers who do not experience the strong side effects. The results of the hedonic price regression are discussed in the results section of the paper.

3.2

# Quality Adjusted Price Index

We compute a second set of price indices that explicitly calculate the changes in welfare from the introduction of new products. These indices are derived from the estimation of an equilibrium model of colon cancer drug pricing. Trajtenberg (1990), who introduced the idea, proposed a two-stage method to construct a quality-adjusted price index. In the first stage, the welfare gains from product innovation are obtained, and in the second stage the price index is built upon those welfare gains.

In this paper the first step is performed by estimating a logit model, where t h e p r e f e r e n c e s o f t h e p h y s i c i a n i o v e r r e g i m e n s j ∈ { 0 , . . . , J t } a t t i m e t a r represented by the following indirect utility function e

u_{ijt }

=

p_{jt }

+

βx

_{jt }

+ ξ j

+

ξ j

+

ε

_{ijt }

where p_{jt }is the price of regimen j at time t, x_{jt }are the observable attributes

of

the

regimen,

ξ j

is

the

mean

of

the

unobserved

characteristics,

and

ξ j t

is

a

time-specific deviation from this mean. ε_{ijt}, which is an idiosyncratic shock to preferences for regimen j, is assumed to follow a Type I Extreme Value distri- bution.^{6 }The outside option (j = 0) in this paper includes off-label colon cancer treatments and regimens with very small market shares for which a complete set

of is not observed. ^{6}In this model all the individual-specific heterogeneity is contained in the idiosyncratic shock to pref- erences, and therefore, it su ers from the well-known independence of irrelevant alternatives criticism (see Nevo (2000) for a complete discussion of the limitations of this approach). In addition, Petrin (2002) points out that the welfare calculations based on these models depend heavily on the error term. We are cur- rently estimating more flexible specifications such as the ones proposed by Berry et al. (1995), and the pure characteristics approach proposed in Berry and Pakes (2007)

10