Walfogel-style instruments, interacted with exogenous rival characteristics (BLP instru-
ments). In her case, she uses consumer demographics in the other counties served by a
rival newspaper. For example, if a paper sold in a sparsely populated suburban county is
headquartered in a populous central metropolitan county, it is likely to have high levels of
quality and variety. The headquarters county is taken as exogenous so that the demo-
graphic levels in that county are available as instruments that shift the choice set in
the suburban county. More specifically, the population demographics of “other
counties” where the newspaper is sold are excluded from the mean utility of the paper
in a given county. Therefore, “other county demographics” are available as instruments
for endogenous characteristics and prices in the “own county” mean utility equation. In
Fan (2013) , the Waldfogel-style instruments include other counties’ education level,
median income, median age, and urbanization.
Fan (2013) considers extending the model of (3.9) to include random coefficients on
some of the characteristics. As noted, random coefficients allow for richer substitution
patterns than simple logit models. For example, if some consumers have a larger than
average taste for local news, they will likely substitute from one locally focused paper
to another one, whereas the pure logit generates substitution patterns that vary only with
demographic-specific market shares. It is not clear whether Fan has adequate instruments
for this purpose, and so she models only one random coefficient (on local news). Once
again, we consider Fan’s supply side below.
Continuing with a discussion of random coefficients,
Sweeting (2013) and Jeziorski
(2014)
consider listener demand models for radio that include random coefficients. In
addition to demographically varying tastes for formats, both Sweeting and Jeziorski have
normally distributed random coefficients on format, which in practice is very similar to
the nested logit model. Though the model with normal coefficients lacks a closed form
solution, it also lacks any a priori constraints on the value of the substitution parameters
(which in the nested logit must fall between zero and one). Jeziorski also adds a random
coefficient on the quantity (in minutes per unit time) of advertising, which introduces a
new dimension of substitution beyond the nested logit. Both of these authors consider
dynamic models on the supply side, focusing either on joint ownership (Jeziorksi) or
format choice (Sweeting).
Both Sweeting and Jeziorski want to model market structure endogenously, and so do
not want to assume that the numbers and formats of competing products are uncorrelated
with the demand unobservable ξ
jt
and so available as potential instruments. Instead, both
use a combination of panel data and IV assumptions. For example, Jeziorski uses a “quasi-
difference” in the unobserved product characteristic for station j at time t:
ξ
jt
¼ρξ
j,t1
+ ν
jt
(3.10)
and then assumes that the innovation ν
jt
is uncorrelated with past market and product
characteristics. These are the BLP instruments, but with the exogeneity (uncorrelated)
103Empirical Modeling for Economics of the Media