The summation term then introduces cross-product correlation in tastes. The term
β
ikt
is consumer i’s marginal utility for characteristic k, often modeled as a function of
the observed demographics and unobserved “tastes” of the consumer:
β
ikt
¼
X
k
x
kjt
υ
ik
σ
k
+ γ
k
z
i
½: (3.4)
The observed demographics are the vector z
i
, and the utility interaction between product
characteristic k and the consumer demographics is parameterized by γ
k
z
i
. In empirical
practice, the number of such interactions generally has to be limited. Unobserved tastes
for characteristic k are captured by the scalar υ
ik
, often assumed to be normally distributed.
The variance of unobserved tastes for characteristic k is increasing in the parameter σ
k
.
The combination of unobserved product characteristics (as in
3.2) with a normally dis-
tributed random coefficients logit (as in
3.4) is introduced in Berry et al. (1995), referred
to as “BLP.”
The distribution of marginal tastes for characteristics, β
ikt
, drives the cross-product
substitution patterns implied by the model. A higher variance of marginal tastes for prod-
uct characteristic k implies, ceteris paribus, that consumers will more closely substitute
between products with similar values of x
kjt
. For example, in radio markets a larger σ
k
on the unobserved taste for country music will imply that consumers will be more
“loyal” to the country music format as market conditions change. The simple reason
is that when σ
k
is large, those consuming country music are disproportionately those with
a systematic taste for country music, whereas when σ
k
is small many consumers listening
to a given country music station are doing so simply because of a large idiosyncratic match
value ε
ijt
.
The idiosyncratic match value is almost always present in a classic discrete-choice
demand model, in part because it “smooths out” market shares and avoids kinks in the
demand function (
Berry and Pakes, 2007). However, the economics of the idiosyncratic
match value is sometimes questioned, as it implies that each new product is “born” with
some fresh component that brings value to some subset of consumers. This assumption
seems more reasonable in media markets than in some other markets. Further, the i.i.d.
match value implies that the share of media consumers increases in the number of product
choices, an assumption that matches our stylized fact. The degree to which the market
expands is governed by the relative importance of the systematic β
ikt
as opposed to the
idiosyncratic ε
ijt
and this relative importance is determined by the parameters σ
k
and γ
k
.
There are several variations and classic special cases. The nested logit (
Cardell, 1991;
McFadden, 1978
) assumes that the only random coefficients are those on a set of “nested”
product group dummies. With one level of product nests, the unobserved random taste
component of utility for a product in nest k is given by
2
2
Arbitrarily rich observed interactions with consumer demographics are easily added to the nested logit
model; they are omitted here only for exposition.
97
Empirical Modeling for Economics of the Media