IEEE Technology and Society Magazine - December 2015 - 76

method for estimating the value of accountability services.
One way of conducting a stated preference analysis is a
choice-based conjoint (CBC) experiment. This type of
experiment is popular in order to understand consumers' preferences for attributes of different goods, in this
case the attributes of an accountability tool. In other
words, our CBC experiment will enable us to research
the influence of different considerations in the decisionmaking about whether to make use of, and pay for,
accountability services or not.
We want to establish the relative importance individuals attribute to transparency offered, by whom the
transparency is offered, who has access to the information, and the possible costs associated with such
information. Our four hypotheses underlying the four
attributes in our experiment are:
H1a: Costs will decrease utility.
H1b: T ra nspa rency infor mation offered w ill
increase utility.
H1c: Independent party involved will increase utility.
H1d: Public availability will decrease utility.
We will explore what kind of transparency information needs to be offered to counterbalance the likely
negative utility associated with having to pay for a cloud
service. We will therefore analyze the market share of
a paid cloud service that provides better utility, in a
market full off free cloud services that offer little to no
transparency. We will attempt to put a price tag on certain transparency tools that may be offered in the cloud.
This way we will contribute to existing knowledge by
providing insight in what consumers would be willing to
pay for better transparency in the cloud.

A key inhibitor for movement to
software-as-a-service cloud models
is lack of consumer trust.

Method
The CBC experiment presents respondents with a number of choice sets. Each choice set will offer a number
of alternatives defined as a combination of different
attributes and per attribute different levels. Respondents
are asked to determine their preference for one of the
alternatives in each set of choices. Based on a series
of these types of choices, the influence of each attribute and level of the attribute can be determined. The
underlying idea is that people base their choice on the

characteristics of the different accountability services
they can choose from, and that the choices are founded
by utility maximization (the best alternative). Because
the variables that are relevant for the decision-making
process are observed (attributes), we can establish the
relative importance of each attribute. We do so by modeling the chance that a product (the accountability service) is chosen, is related to the score of each attribute.
In our experiment the individual users' preferences
are determined via an electronic survey specifically
designed to measure each attribute's effect. Respondents are presented with a series of choices of scenarios constructed out of the different attributes for which
we want to measure influence. The attributes vary
in a predetermined range of possible values that are
perceived as realistic by cloud users for that particular
attribute (the so called levels). Scenarios are thus constructed of attributes and levels (see Table 1).

Sample
Respondents were recruited through Amazon Mechanical Turk. This service allows requesters to put tasks
online, in this case a survey, which will then be completed by so-called workers. Previous studies have
shown that Mechanical Turk is viable for scientific
research with no more downsides than other methods
of obtaining a sample [9]. The Amazon Mechanical
Turk respondents were offered 2 dollars as a reward
for completing our 15-min survey. We gathered the
responses of 254 respondents. Only respondents that
completed at least 6 of the 12 real decision tasks were
kept for analysis. This led to the removal of 43 respondents (16.9%), leaving 211 respondents. In this sense,
workers from Mechanical Turk proved more reliable
compared to a sample we gathered for the pilot where
61.4% of respondents had not completed over half of
the survey.
In addition to 12 real decision tasks, 2 control decision tasks were included in which one of the options
was clearly preferable over the others, which respondents were expected to choose. Respondents who
"failed" one such task were assumed to have done so
through random error, whereas those who "failed" both
tasks were dropped from analyses. This resulted in 52
respondents being removed (24.6%). This is considerably more than during our pilot, where only 6.8% of
respondents had to be removed in this step. Finally, this
left 159 respondents for our analysis.
Respondents were 18- to 67-years-old (M = 36.13,
SD = 10.83), and 45.3% were male and 53.5% female
(1.3% were unknown). The majority of respondents were
American (N = 68, 42.8%) or from India (N = 35, 22.0%)
as is typical for a Mechanical Turk sample [10], [11].
The remainder of the respondents were from various

IEEE Technology and Society Magazine

∕

December 2015

Table of Contents for the Digital Edition of IEEE Technology and Society Magazine - December 2015