Participants were employees of a technology company in Silicon
Valley and voluntarily participated in the experiment. They all paid
for their own electricity bill. More than 50% of the participants
reported income levels higher than $150,000. However, it is worth
mentioning that the mix of households in our study (i.e., welleducated,
upper and middle class families who are also early
adopters of new technologies such as home energy monitoring
systems) is also more likely to respond to energy efficiency programs
by investing in energy-efficient products [34]. Hence, the
results of our analysis can be particularly helpful to energy efficiency
program managers and policy makers to develop programs
specifically targeted towards the households represented by our
sample.
After collecting the data, 952 households for which reliable
smart meter and survey data were available were selected for the
analysis. Less than 3% of survey responses were inconsistent or
missing, for which we imputed data using iterative model-based
imputation techniques ([35,36]). After the initial screening, we
did not observe any outliers in the data set. Selected households
were located in 419 different Zip Codes, 140 different counties, 26
different states, and were spread across all six climate zones
defined by the Department of Energy [1]. California had the largest
representation (53% of households) of all states in the data set.
During the data collection process, the weather conditions in most
areas where participant households resided were similar to the 30-
year average climatic conditions; however, some areas, especially
in the north east of the U.S., experienced slightly higher-thannormal
temperatures [37]. Average electricity consumption in our
sample lied between California and US averages. Some structural
determinants such as household size, square footage of the house,
and the proportion of single family detached units in our sample
were close to US population averages [33].
We analyzed the consumption data at hourly level to ensure that
the fluctuations in electricity consumption are considered, but not
obscured by sudden spikes in the consumption. This also makes the
results of our analysis comparable with those of previous studies on
smart meter data and electricity market analysis [38].
We transformed some variables to better reflect the technical
characteristics of buildings. For example, we transformed the
construction year to a categorical variable that indicated the residential
building code that was effective at the time of the construction
(i.e., different revisions of ASHRAE 90.2 [1]). We also
included a categorical variable for House Size to capture the effects
of the floor area that are not completely explained by square
footage. For example, when a building’s floor area passes a certain
threshold, the type of structural and architectural material that is
used in the building often changes significantly. Since we do not
have a separate variable for floor area and are not dividing the
electricity consumption of the dwelling by its floor area, introducing
the House Size variable does not create a multicollinearity
problem (Table 1).