terns by layering the error components on top of the random-parameter logit
model.
2.2 Specification of Utility
Recognizing the preceding discussion about heterogeneous adoption behavior, we
assume that the household’s utility Vij is negatively effected by the investment
cost Cij, and positively effected by the decline of the building’s annual primary
energy demand ∆Qij, measured in megawatt hours (MWh), both of which are
associated with a specific retrofitting alternative j . We control for the economic
background of the households by including annual disposable income into the
analysis.3 Further, we expect that the level of the household’s energy consump-
tion influences the decision of whether to renovate, either positively because a
household with a high energy consumption level is more inclined to lower its
energy cost, or negatively because a high level reflects low energy awareness.
Moreover, because there is a quality differential between the building stocks in
western and eastern Germany, a binary variable indicates whether the household
lives in the eastern part of Germany. Finally, we include a measure of the ac-
cessibility of information on home retrofits within the immediate vicinity of the
household. This variable is intended to proxy for the transaction costs of infor-
mation acquisition, and is defined as the relative availability of certified home
auditors within a 20 kilometer radius of the household’s location.4
3As is typical for survey data, information on income is missing for a large share of the
households - roughly 20%. To impute these missing values, we employ the expectation-
maximization algorithm recommended by King et al. (2001). The employed algorithm can
be implemented using a program compatible with the statistical software R, and is download-
able from http://gking.harvard.edu.
4To derive this measure we drew upon a list of certified home auditors and their addresses
published by the German government. We read the data as a map-layer into a Geographical
Information System and overlaid this with a layer of household locations. We then created a
circular buffer around each household having a radius of 20 kilometers and generated a count
11