Content uploaded by Eyal Peer
Author content
All content in this area was uploaded by Eyal Peer on Jan 01, 2016
Content may be subject to copyright.
Scaling the Security Wall
Developing a Security Behavior Intentions Scale (SeBIS)
Serge Egelman
International Computer Science Institute
Berkeley, CA, USA
egelman@cs.berkeley.edu
Eyal Peer
Bar-Ilan University
Ramat Gan, Israel
eyal.peer@biu.ac.il
ABSTRACT
Despite the plethora of security advice and online education
materials offered to end-users, there exists no standard mea-
surement tool for end-user security behaviors. We present the
creation of such a tool. We surveyed the most common com-
puter security advice that experts offer to end-users in order
to construct a set of Likert scale questions to probe the ex-
tent to which respondents claim to follow this advice. Using
these questions, we iteratively surveyed a pool of 3,619 com-
puter users to refine our question set such that each question
was applicable to a large percentage of the population, ex-
hibited adequate variance between respondents, and had high
reliability (i.e., desirable psychometric properties). After per-
forming both exploratory and confirmatory factor analysis,
we identified a 16-item scale consisting of four sub-scales
that measures attitudes towards choosing passwords, device
securement, staying up-to-date, and proactive awareness.
Author Keywords
Psychometrics; Individual differences; Security behavior
ACM Classification Keywords
H.1.2. Models and Principles: User/Machine Systems; K.6.5.
Management of Computing and Information Systems: Secu-
rity and Protection; J.4. Social and Behavioral Sciences: Psy-
chology
INTRODUCTION
Each year, billions of dollars are spent on computer secu-
rity education programs. Many employers require employ-
ees to undergo training regimens, prior to granting network
or computer access, or on an ongoing basis. The U.S. Na-
tional Institute of Standards and Technology (NIST) offers
guidelines for these programs [46], and even the U.S. Depart-
ment of Homeland Security has a “National Cyber Security
Awareness Month” [42]. The NIST guidelines state that “for-
mal evaluation and feedback mechanisms are critical com-
ponents of any security awareness, training, and education
program” [46], yet provide no evaluation metrics. Measuring
attack rates is a suboptimal strategy: fewer successful attacks
Permission to make digital or hard copies of all or part of this work for personal or
classroom use is granted without fee provided that copies are not made or distributed
for profit or commercial advantage and that copies bear this notice and the full cita-
tion on the first page. Copyrights for components of this work owned by others than
ACM must be honored. Abstracting with credit is permitted. To copy otherwise, or re-
publish, to post on servers or to redistribute to lists, requires prior specific permission
and/or a fee. Request permissions from permissions@acm.org.
CHI 2015, April 18 - 23 2015, Seoul, Republic of Korea.
Copyright is held by the owner/author(s). Publication rights licensed to ACM.
ACM 978-1-4503-3145-6/15/04...$15.00.
http://dx.doi.org/10.1145/2702123.2702249
may simply be due to fewer attempts; increased attack suc-
cess may be due to new vulnerabilities, increased sophistica-
tion, or other factors that are independent of user behavior.
In addition to training and education programs, a plethora of
organizations offer tips and advice to end-users on how they
can stay safe online. However, due to the wide range in spe-
cific actionable items offered, most end-users are likely to be
confused about which advice to follow and which to ignore.
For example, the U.S. Computer Emergency Readiness Team
(US-CERT) offers over 500 actionable items [41], whereas
Verizon, a large U.S. cellular phone provider and ISP, offers
only 4 items [43]. If there are a core set of general behaviors
underlying most of this computer security advice, the action-
able items offered to end-users can be drastically reduced to
be made more memorable. Similarly, in certain cases, reli-
able self-reporting of these behaviors could replace the need
to perform costly and time-consuming observational studies.
When studying human behavior, researchers across many dif-
ferent fields will often use scales as proxies for observation.
Scales “are intended to measure elusive phenomena that can-
not be observed directly” [11]. Scales are therefore highly
useful when behavioral or observational experiments are ei-
ther too costly, complex, or simply not possible. A scale can
also measure how attitudes or behaviors change over time.
For instance, the Westin privacy index is used to segment
the population based on attitudes towards online privacy, and
while there is substantial debate about whether it is predictive
of actual privacy behaviors [48], it is still a useful tool that
researchers use to examine how privacy attitudes evolve over
time [22]. While carefully constructed scales are a core tool
when studying human behavior in other domains, they have
yet to be applied to the study of computer security behaviors.
We developed a new scale for assessing the computer se-
curity behaviors of end-users: the Security Behavior Inten-
tions Scale (SeBIS). In short, our scale measures users’ self-
reported adherence to computer security advice. We em-
ployed the 4-step scale development procedure offered by
Netemeyer et al. [29], wherein we used common security ad-
vice offered to end-users as our content domain, generated
and refined the individual questions, performed Exploratory
Factor Analysis (EFA) to explore underlying constructs and
develop sub-scales, and then performed Confirmatory Factor
Analysis (CFA) to finalize our scale and measure its reliabil-
ity. Our resulting security behaviors scale was evaluated on
3,619 respondents, and consists of 16 items that measure 4
underlying constructs: device securement, password genera-
tion, proactive awareness, and updating behaviors.
RELATED WORK
Several models have been created to analyze how humans
make security decisions (e.g., [47, 9, 30]), which researchers
have used to offer recommendations. For instance, Egel-
man et al. showed that when less-frequent high-risk warn-
ings appear similar to frequent low-risk security warnings,
users ignore both [12]. Felt et al. showed that smartphone
permission warnings are overlooked because they occur too
frequently and with poor timing [16]. In this vein, usable
security research has greatly improved the security interven-
tions to which users are exposed. We posit that further gains
can be made by differentiating users according to individ-
ual traits and offering different intervention designs based on
those traits. However, for this to be possible, we first need a
reliable tool for measuring users’ traits, such as their adher-
ence to common computer security practices.
Scales are most often used in psychology, as a means of mea-
suring underlying psychological constructs. One scale that
might be relevant to our work is the Domain-Specific Risk-
Taking (DoSpeRT) scale, which measures a person’s self-
reported propensity to engage in risky behaviors across four
dimensions: ethical, financial, health/safety, recreational, and
social [4]. We hypothesize that an individual’s propensity to
take safety risks may predict their computer security behav-
iors. Other psychometrics may also apply: for instance, high
impulsivity (e.g., [31]) may predict an individual’s willing-
ness to visit questionable URLs, low concern for future con-
sequences (e.g., [21]) may predict an individual’s disinclina-
tion for running security software, and so forth.
While we are not aware of any studies to develop a scale to
measure end-user security behaviors, researchers have raised
questions about the psychological underpinnings of computer
crime for over three decades. Copes et al. created a sur-
vey instrument to examine correlations between victimiza-
tion and demographic factors [8]. Sherizen suggested that
there may exist certain “markers” that could be used to de-
tect the propensity for committing online crimes [38]. Only
recently have others started to explore this concept. For in-
stance, self-control is a psychological construct with corre-
sponding scales that can be used to measure its prevalence in
people. Researchers have administered these scales to see if
they correlate with either committing online crimes [27], or
being the victim of online crimes [5, 26].
Measuring more general types of computer security behaviors
has been limited to within organizations. Stanton et al. cre-
ated a taxonomy of workplace computer security behaviors,
and showed that they could be grouped by two dimensions:
intentions and expertise [39]. Ng et al. [30] used factor anal-
ysis to create and evaluate a model of end-user security be-
havior. Their goal was to use this model to predict email se-
curity behaviors, whereas our goal is the development of a
composite measurement scale.
Privacy is related to security and has several of its own scales
that have been developed over the years and are actively used
by researchers to better understand people’s attitudes and in-
tentions. The most well known privacy scale is the Westin
index, which is used to classify consumers into three groups
based on their privacy attitudes: fundamentalists, pragmatists,
and the unconcerned [22]. While the Westin index is well-
cited and frequently used by privacy researchers, it has been
shown to be a poor predictor of privacy-preserving behavior
(e.g., [48, 16]). The scale is therefore used to illustrate the
gap between stated preferences (as measured by the Westin
index) and observed behaviors (as measured through empiri-
cal studies)—the so-called “privacy paradox.” Thus, the self-
reported data may not be “wrong,” but instead there may be
many other factors at play that influence behavior [1, 2].
The Westin index is unidimensional and therefore only mea-
sures one aspect of privacy, whereas other scales measure
privacy attitudes and behaviors across multiple dimensions.
Malhotra et al. developed the Internet Users’ Information Pri-
vacy Concerns (IUIPC) scale by performing careful factor
analysis, and showed that privacy concerns can be measured
across three dimensions: control over information, awareness
of privacy practices, and attitudes about information collec-
tion [23]. Buchanan et al. developed three different unidi-
mensional privacy scales, which together measure concerns,
general behaviors, and use of technical solutions [6].
The Security Behavior Intentions Scale (SeBIS) is a new scale
to measure end-users’ intentions to comply with computer se-
curity advice. We followed the scale development procedure
outlined by Netemeyer et al. [29]:
1. Construct Definition and Content Domain: Clearly
defining the construct that the scale intends to measure.
2. Generating and Judging Measurement Items: Creating
a pool of candidate questions and then evaluating the ques-
tions to remove ones that are invalid.
3. Designing and Conducting Studies to Develop and Re-
fine Scale: Performing EFA to reduce the set of questions
and explore latent constructs to build a model.
4. Finalizing the Scale: Performing CFA to confirm that the
scale fits the intended model, as well as reliability analysis.
In the remainder of this paper, we describe how we applied
each of these steps and their results. We then describe how we
used our scale to explore correlations between self-reported
security behaviors and existing psychometrics.
CONSTRUCT DEFINITION AND INITIAL ITEMS
The goal of our scale is to measure end-users’ intentions to
comply with common security advice. This created two con-
straints: limiting our metric to security behaviors applicable
to most end-users (vis-`
a-vis advice for corporate users, etc.),
and only consisting of “widely accepted” security behaviors.
We examined the advice offered to end-users by nationwide
U.S. ISPs, US-CERT, and industry consortia. Based on this
advice, we enumerated a preliminary set of computer security
behaviors, omitting advice that was too technical.
Our initial list consisted of 26 behaviors. Next, we met with a
group of six computer security experts who offered additional
items, increasing our set to 30 behaviors. We transformed
these 30 behaviors into personal statements that could be an-
swered based on degree of agreement (i.e., “strongly agree” to
“strongly disagree”). To minimize acquiescence bias (i.e., the
# Question N/A µ σ
A1 I apply software updates as soon as my computer prompts me. 5 (1.0%) 3.20 1.221
A2 I am happy to use an older version of a program, as long as it meets my needs. 5 (1.0%) r1.99 1.000
A3 Whenever I step away from my computer, I lock the screen. 5 (1.0%) 2.50 1.306
A4 Others can access my smartphone or tablet without needing a PIN or passcode. 21 (4.4%) r3.34 1.545
A5 When I discover a computer security problem at work, I’m likely to promptly report it to my employer. 64 (13.4%) 4.08 0.995
A6 It’s important to use a WiFi password to prevent unauthorized people from using my home network. 11 (2.3%) 4.66 0.690
A7 I frequently click links in email messages to see what they are, regardless of who sent the message. 5 (1.0%) r4.51 0.922
A8 It’s important to run anti-virus software on my computer. 7 (1.5%) 4.35 0.941
A9 When browsing websites, I frequently mouseover links to see where they go, before clicking them. 4 (0.8%) 4.13 0.977
A10 When using public WiFi, I visit the same websites that I would visit when using the Internet at home. 20 (4.2%) r2.93 1.266
A11 I usually do not pay attention to where I’m downloading software from. 2 (0.4%) r4.38 0.900
A12 I frequently backup my computer. 5 (1.0%) 3.07 1.165
A13 I frequently visit websites even when my web browser warns me against it. 8 (1.7%) r3.98 1.028
A14 I circumvent my employer’s computer usage policies when they prevent me from completing a task. 86 (18.0%) r3.54 1.184
A15 I am careful to never share confidential documents stored on my home or work computers. 15 (3.1%) 4.36 0.757
A16 Frequently checking the access control settings on social networking websites isn’t worth the time it takes. 18 (3.8%) r3.56 1.165
A17 I always write down my passwords to help me remember them. 6 (1.3%) r3.60 1.313
A18 Creating strong passwords is not usually worth the effort. 6 (1.3%) r4.05 1.047
A19 I frequently check my financial accounts for fraudulent charges. 10 (2.1%) 4.11 0.914
A20 If I receive a suspicious email from a company that I do business with, I’ll phone the company 22 (4.8%) 3.52 1.236
to make sure the email is accurate.
A21 I never give out passwords over the phone. 7 (1.5%) 4.53 0.787
A22 I frequently purchase things that I see advertised in unsolicited emails. 4 (8.8%) r4.51 0.840
A23 I tend to ignore computer security stories in the news because they don’t impact me. 4 (8.8%) r3.83 1.050
A24 I use encryption software to secure files or email messages. 10 (2.1%) 2.74 1.225
A25 Once I create a password, I tend to never change it. 5 (1.0%) r3.30 1.182
A26 I try to create a unique password for every account I have. 5 (1.0%) 3.21 1.284
A27 Rather than logging out of websites, I usually just navigate elsewhere or close the window when I’m done. 7 (1.5%) r3.06 1.299
A28 I always make sure that I’m at a secure website (e.g., SSL, “https://”, a lock icon) 4 (0.8%) 3.80 1.173
when transmitting information online.
A29 I frequently use privacy software, “private browsing” or “incognito” mode when I’m online. 6 (1.3%) 3.17 1.247
A30 I frequently let others use my computing devices (e.g., smartphone, tablet, laptop). 3 (0.7%) r3.79 1.172
Table 1. Initial set of security questions evaluated on a 5-point Likert scale (from “strongly disagree” to “strongly agree”) by 479 participants. Depicted
are the questions, the rate of “N/A” responses, and the average responses and standard deviations after recoding negatively-phrased questions (r).
propensity for respondents to agree with every question) [34],
we reworded half the questions to be negatively phrased (i.e.,
agreement with the statement indicates not following com-
mon security advice). Thus, our initial question pool con-
sisted of the 30 questions listed in Table 1.
Method and Demographics
In August of 2014, we performed an initial evaluation of our
questions. We recruited participants from Amazon’s Mechan-
ical Turk, restricting the survey to participants based in the
U.S. who had an approval rate of 95% or greater. We paid
each participant $1.00 to complete our survey, which con-
sisted of the 30 aforementioned statements evaluated on a
Likert scale (“strongly disagree (1),” “disagree (2),” “neither
agree nor disagree (3),” “agree (4),” and “strongly agree (5)”).
We gave participants a “N/A” option and the ability to simply
not answer questions so that we could determine whether any
questions were not applicable to our audience.
We were concerned that social desirability bias may have an
impact on participants’ responses [10]. That is, some par-
ticipants may misreport their behaviors in order to make it
appear as though they engage in more socially acceptable be-
haviors (e.g., participants may be ashamed to admit engaging
in ill-advised behaviors). We took two steps to mitigate this
concern. First, we advertised the experiment as a “computer
behavior survey,” so as to not specifically mention security
during recruitment (thereby minimizing a potential selection
bias). Second, we asked all participants to complete the 10-
item Strahan-Gerbasi version of the Marlowe-Crowne Social
Desirability Scale [40], which we then correlated with partic-
ipants’ responses to the security behavior questions.
As with all surveys, we were concerned with being able to
deter and detect careless responses. Based on Meade and
Craig’s recommendation [25], we included two separate “in-
structed response items” in our survey. On the first page of
the survey after the consent form, we included the following:
This study requires you to voice your opinion using
the scales below. It is important that you take the time
to read all instructions and that you read questions
carefully before you answer them. Previous research
on preferences has found that some people do not
take the time to read everything that is displayed in
the questionnaire. The questions below serve to test
whether you actually take the time to do so. Therefore,
if you read this, please answer ‘three’ on the first
question, add three to that number and use the result
as the answer on the second question. Thank you for
participating and taking the time to read all instructions.
I would prefer to live in a large city rather than a
small city. [Strongly disagree (1), (2), (3), (4), (5), (6),
Strongly agree (7)]
I would prefer to live in a city with many cultural oppor-
tunities, even if the cost of living was higher. [Strongly
disagree (1), (2), (3), (4), (5), (6), Strongly agree (7)]
If participants failed to select 3 and 6, respectively, they were
given a second opportunity featuring bolded instructions. If
they failed a second time, they were prevented from com-
pleting the survey. In addition to this attention check task,
we included a second question hidden amongst the true/false
Strahan-Gerbasi social desirability statements: I do not read
the questions in surveys. We excluded participants who an-
swered “true” for this question from our analysis.
We received a total of 503 responses, though removed 24
(4.8%) who did not answer the second attention check ques-
tion correctly. Our resulting sample of 479 participants con-
sisted of 273 males (57.0%) and ages ranged from 18 to 69
(µ= 33.2,σ= 10.5). Thirty-six percent of our partici-
pants (173 of 479) held bachelor’s degrees, while an addi-
tional 12.9% had completed some graduate school, and me-
dian household income was between $35,000 and $50,000.
We therefore consider our sample to be representative of the
U.S. online population, albeit with a slight male skew.
Results
We observed that the internal reliability of our initial question
set was quite high (Cronbach’s α= 0.83). We performed
Pearson correlations between the Strahan-Gerbasi social de-
sirability scale and each question, and only observed one to
be statistically significant: A28 (r= 0.160,p < 0.0005).
However, given the small effect size (i.e., 2.6% of variance is
shared between this question and the social desirability scale),
we chose to retain the question. This suggests that partici-
pants answered truthfully and consistently. At the same time,
we discovered several issues with our items, which high-
lighted the need for question refinement: non-applicability,
poor item-total correlation, and ceiling effects.
Non-Applicability
We measured how many participants skipped items or re-
sponded with a “N/A” option. Two questions stood out, A5
and A14, which both had to do with workplace computer us-
age. Participants did not respond to these questions 13.4%
and 18.0% of the time, respectively. In hindsight, possible
confounds are obvious: for instance, not having a job involv-
ing computer use. Thus, we decided to remove the question
about circumventing policies (A14) and reworded A5 to focus
on promptly fixing/reporting security problems in general.
Additionally, questions A22 and A23 were not answered by
8.8% of participants. The former concerned buying items
advertised via unsolicited emails, which we decided is not
a security concern per se. The latter had to do with ignor-
ing computer security stories in the news, which we felt was
ambiguous: it did not describe a specific behavior, and a neg-
ative response could indicate either not following the news or
disagreeing with the impact of computer security stories. In-
stead, we reworded it to be about proactively changing pass-
words after publicized data breaches.
We also realized that file sharing (A15) and social networking
(A16) might not be widely applicable. We removed A17 be-
cause there is no longer a clear consensus that writing down
passwords is a poor practice (especially if it leads to creat-
ing stronger passwords). Similarly, we removed A30, which
concerned sharing devices with others, because research sug-
gests that this behavior is quite nuanced [13], and therefore
this question may have been ambiguous.
Poor Item-Total Correlation
Item-total correlation measures how well the responses to one
item correlate with the other items in the scale (i.e., homo-
geneity). Questions are deemed to not represent the rest of
the scale when their item-total correlations are below 0.2 [14].
This occurred in four questions: A2 (0.117), A3 (0.147), A17
(0.113), and A29 (0.196). We observed that these questions
all used qualifiers (e.g., “I frequently...” or “I always...”).
Given that the purpose of the scale was for participants to re-
port how frequently they engaged in each activity, we realized
that we needed to remove the qualifiers from each statement
and change the response format so that, instead of measuring
agreement with a certain behavior, our scale would measure
the frequency of engaging in that behavior using the options
of, “always,” “often,” “sometimes,” “rarely,” and “never.”
Ceiling Effects
Because a scale is designed to quantify one respondent’s re-
sponse relative to others, it is important that individual ques-
tions exhibit adequate variance; if everyone responds in the
same manner, the scale is of little utility. Ceiling and floor
effects occur when questions fail to capture sufficient vari-
ance; responses lie at one end of the spectrum because the
provided responses did not adequately capture the true range
of opinions. We observed that this was the case with ques-
tions A6,A7,A8,A11,A15,A21, and A22. The average re-
sponses to these questions were close to the maximum (i.e.,
4.0<µ<5.0), and had relatively low standard deviations
(i.e., σ < 1.0). As a result, we explored alternate wordings.
Additional Testing
We evaluated the impact of wording changes on responses,
both with regard to variance and applicability. These itera-
tions involved 2,184 additional participants, who we recruited
in the same manner as our initial sample. Additionally, we
also screened out participants who had completed a previ-
ous iteration of the survey. We produced a refined set of 24
questions that exhibited adequate variance, were applicable
to almost all participants, were free from ceiling effects, and
showed high item-total correlations.
REFINING THE SCALE
We recruited an additional cohort of participants so that we
could perform Exploratory Factor Analysis (EFA) on the re-
sulting 24 items. We describe our methodology and results.
Methodology
We recruited 500 new participants to respond to a set of 24
security questions (Table 2). Additionally, we randomly as-
signed a fifth of our sample to complete the 16-item Pri-
vacy Concerns Scale [6]. We did this to assess the discrim-
inant validity of our scale: whether it was measuring a new
construct or something already captured by existing privacy
scales (PCS correlates significantly with other established pri-
vacy scales, the Westin Index [22], and the IUIPC [23]). (We
piloted established psychometric tests on the participants who
did not receive PCS, which we describe later in this paper.)
# Question N/A µ σ
B1 When I’m prompted about a software update, I install it right away. 1 (0.2%) 3.23 0.993
B2 When my computer wants me to reboot after applying an update or installing software, I put it off. 0 (0.0%) r3.08 1.116
B3 I try to make sure that the programs I use are up-to-date. 1 (0.2%) 3.89 0.914
B4 I manually lock my computer screen when I step away from it. 5 (1.1%) 2.58 1.386
B5 I set my computer screen to automatically lock if I don’t use it for a prolonged period of time. 3 (0.7%) 3.47 1.552
B6 I log out of my computer, turn it off, put it to sleep, or lock the screen when I’m done using it. 1 (0.2%) 3.87 1.283
B7 I use a PIN or passcode to unlock my mobile phone. 18 (3.9%) 3.27 1.798
B8 I use a password/passcode to unlock my laptop or tablet. 22 (4.8%) 3.99 1.525
B9 If I discover a security problem, I continue what I was doing because I assume someone else will fix it. 15 (3.3%) r3.98 0.972
B10 When someone sends me a link, I open it without first verifying where it goes. 0 (0.0%) r3.85 1.068
B11 I verify that my anti-virus software has been regularly updating itself. 9 (2.0%) 3.59 1.215
B12 When browsing websites, I mouseover links to see where they go, before clicking them. 0 (0.0%) 3.48 1.115
B13 I know what website I’m visiting based on its look and feel, rather than by looking at the URL bar. 3 (0.7%) r2.92 1.107
B14 I backup my computer. 2 (0.4%) 2.97 1.216
B15 When I hear about websites that have been hacked, I wait to change my passwords 16 (3.5%) r3.41 1.232
until I have been personally notified.
B16 I use some kind of encryption software to secure sensitive files or personal information. 20 (4.4%) 2.56 1.401
B17 I do not change my passwords, unless I have to. 0 (0.0%) r2.40 1.103
B18 I use different passwords for different accounts that I have. 1 (0.2%) 3.59 1.084
B19 I do not include special characters in my password if it’s not required. 2 (0.4%) r2.95 1.349
B20 When I create a new online account, I try to use a password that goes beyond the site’s minimum requirements. 2 (0.4%) 3.37 1.193
B21 When I’m done using a website that I’m logged into, I manually log out of it. 0 (0.0%) 3.64 1.125
B22 I submit information to websites without first verifying that it will be sent securely 1 (0.2%) r3.65 1.103
(e.g., SSL, “https://”, a lock icon).
B23 I use privacy software, “private browsing,” or “incognito” mode when I’m browsing online. 3 (0.7%) 2.55 1.119
B24 I clear my web browsing history. 0 (0.0%) 3.31 1.114
Table 2. Revised set of security questions evaluated on a 5-point Likert scale (from “never” to “always”) by 456 participants. Depicted are the questions,
the rate of “N/A” responses, and the average responses and standard deviations after recoding negatively-phrased questions (r).
510 15 20
123456
Determining Number of Factors
Components
Eigenvalues
Eigenvalues (>mean = 6 )
Parallel Analysis (n = 4 )
Optimal Coordinates (n = 4 )
Acceleration Factor (n = 1 )
(OC)
(AF)
Figure 1. Scree plot showing eigenvalues from EFA, acceleration factor,
optimal coordinates, and parallel analysis. We extracted 4 components.
Since we found that responses to the security questions did
not correlate with social desirability, we removed the Strahan-
Gerbasi scale, which also removed our second attention check
(i.e., I do not read the questions in surveys). To compensate
for this, we followed Meade and Craig’s recommendation to
screen out careless responses based on completion times [25].
Because reading speed is so varied and because there was no
obvious threshold in our data, we removed participants whose
completion times were in the top 10% (i.e., those finishing the
survey in under 227s), which left us with 456 valid responses.
Results
We examined individual questions in terms of means, stan-
dard deviations, and applicability. Table 2 shows that the ceil-
ing effects in our initial experiments disappeared: all means
were below 4.0 and standard deviations exceeded 0.9. Simi-
larly, each question was applicable to over 95% of our partic-
ipants. Thus, we proceeded to factor analysis and measured
reliability. While we allowed participants to omit specific
questions so that we could measure applicability, factor anal-
ysis and reliability metrics (e.g., Cronbach’s alpha) require
all scale items. As a result, we removed 80 participants who
omitted responses to any questions, leaving us with a sample
size of 376 for the remainder of our analysis. This sample
size still exceeds the recommendation of Hair et al. to use a
minimum of 5 participants per item [18].
Factor Analysis
Bartlett’s test of sphericity (χ2(276) = 2037.7,p < 0.0005)
and the Measure of Sampling Adequacy (0.869) showed
that our data were correlated and measured common factors.
We performed an exploratory Principal Component Analysis
(PCA) and extracted six components based on the Kaiser cri-
terion (i.e., retaining all components with eigenvalues >1.0).
Since the Kaiser criterion is considered to be overly inclu-
sive [15], we also determined the optimal coordinates and
performed parallel analysis (Figure 1). Determining optimal
coordinates involves linear regression to fit a line to the small-
est eigenvalues and then determining the point at which an
eigenvalue diverges [33]. As can be seen in Figure 1, this test
indicated that we should retain four components.
Passwords Securement Awareness Updating
α0.764 0.728 0.668 0.719
IIC 0.449 0.407 0.288 0.469 ITC
B18 .742 0.512
B20 .742 0.629
B19 .735 0.580
B17 .692 0.543
B5 .760 0.526
B8 .724 0.513
B7 .720 0.502
B4 .714 0.547
B22 .727 0.496
B10 .697 0.488
B9 .628 0.402
B13 .552 0.326
B12 .540 0.403
B1 .797 0.474
B3 .773 0.589
B11 .726 0.580
Table 3. Factor loadings (PCA with Varimax rotation; loadings <0.25
removed) from EFA. The first two rows depict reliability measures:
Cronbach’s αand average inter-item correlation. The last column de-
picts each item-total correlations.
Parallel analysis involves extracting eigenvalues from random
data [19]. Any retained eigenvalues should be greater than
those extracted due entirely to random sampling error. Thus,
the eigenvalues generated by parallel analysis are compared
to those generated by EFA, and when the ith eigenvalue of
the former is greater than the ith eigenvalue of the latter, it is
concluded that there are i−1underlying components. This
corroborated our decision to extract four components.
We applied a Varimax rotation and considered an item to
be loaded on a factor if its loading exceeded 0.5. We also
applied Saucier’s criterion of only considering an item to
be loaded on a component if its loading on that component
was more than twice as high as its loading on other compo-
nents [36]. Based on these requirements, we removed the
following seven items: B6,B14,B15,B16,B21,B23, and
B24. Finally, we observed that removing B2 increased Cron-
bach’s alpha of its associated sub-scale from 0.707 to 0.719.
Thus, we retained the remaining 16 items and reran PCA. The
rotated factor loadings are depicted in Table 3.
The four components that we extracted predicted over 55.6%
of variance. Based on each component’s items, four distinct
themes emerged: password generation (e.g., creating strong
passwords, changing passwords), device securement (e.g., us-
ing a PIN on a smartphone, locking a desktop screen when
stepping away), proactive awareness (e.g., checking links be-
fore clicking them), and updating (e.g., applying software
updates in a timely manner). Each of these components ac-
counted for 26.7%, 11.6%, 9.0%, and 8.3% of variance, re-
spectively. Our scale thus consists of four sub-scales.
Reliability
We examined scale reliability in three parts. First we com-
puted Cronbach’s alpha for the full scale (α= 0.801) and
its sub-scales (see Table 3). We observed that our data com-
plied with the metric used by McKinley et al. [24]: a multi-
component scale is reliable if α > 0.6for all sub-scales and
α > 0.7for a majority of sub-scales.
Next, we calculated the item-total correlation of each item,
which is the Pearson correlation between the item and the av-
erage of all other items in its sub-scale. Everitt recommends
a threshold of 0.2 [14], which our items all exceeded. Finally,
we calculated the average inter-item correlation; Robinson
et al. classify average inter-item correlations between 0.20
and 0.29 as “extensive” and above 0.30 as “exemplary” [35].
The lowest average inter-item correlation was found in the
“awareness” component, though it was still 0.288 (i.e., “ex-
tensive”), whereas the reliability of the other three compo-
nents was consistently above 0.4. Thus, by all three measures,
we concluded that our sub-scales each had high reliability.
Discriminant Validity: Correlation with Privacy Concerns
We randomly selected one fifth of our participants1to com-
plete an additional question set: the 16-item Privacy Concern
Scale developed by Buchanan et al. [6]. The reason for us-
ing this scale in particular was that it is unidimensional and
has high internal reliability; across the 68 participants (18%
of 376) who completed it, α= 0.935. We performed Pear-
son correlations between participants’ average privacy con-
cern score, and the average scores for each of our new scale’s
four sub-scales: passwords (r= 0.151,p < 0.219), se-
curement (r= 0.153,p < 0.213), awareness (r= 0.251,
p < 0.039),2and updating (r= 0.164,p < 0.183). The lack
of significant correlations suggests that security behaviors are
orthogonal to privacy concerns, and therefore our new scale
is measuring something different.
FINALIZING THE SCALE
Based on the previous EFA, we created a final survey instru-
ment that included only the 16 final questions (Table 4), in
random order, with the N/A option removed (i.e., all ques-
tions were now required). We included additional questions
from several well-studied psychometric tests in an attempt to
correlate the self-reported security behaviors measured by our
instrument with existing psychometrics. In this section we
describe the Confirmatory Factor Analysis (CFA) that we per-
formed and how it showed reasonable goodness of fit between
our questions and the hypothesized latent factors. We show
how the scale also has good reliability and how its underlying
constructs relate to well-established psychometrics.
Factor Analysis
We recruited an additional cohort of 500 participants from
Mechanical Turk to participate in a final experiment. We fol-
lowed the recommendation of Peer et al. [32], who showed
that restricting tasks to workers who have completed over 500
previous tasks and at least 95% approval rates negates the
need for attention check questions. In addition to these two
requirements, we restricted participation to those 18 or over
in the U.S. who had not been exposed to one of our previ-
ous question sets. We included the initial instructed response
question, but did not do any post hoc screening.
1As we noted earlier, the remaining participants completed other
psychometrics, which we piloted prior to our CFA experiment, dis-
cussed in the next section.
2Upon applying the Bonferroni correction to account for multiple
testing, we do not consider this to be statistically significant.
#Device Securement (28.47% of variance explained; λ= 4.555)µ σ
F4 I set my computer screen to automatically lock if I don’t use it for a prolonged period of time. 3.20 1.559
F6 I use a password/passcode to unlock my laptop or tablet. 3.78 1.525
F3 I manually lock my computer screen when I step away from it. 2.63 1.343
F5 I use a PIN or passcode to unlock my mobile phone. 3.21 1.733
#Password Generation (12.95% of variance explained; λ= 2.071)µ σ
F12 I do not change my passwords, unless I have to.r2.65 1.091
F13 I use different passwords for different accounts that I have. 3.75 1.037
F15 When I create a new online account, I try to use a password that goes beyond the site’s minimum requirements. 3.31 1.096
F14 I do not include special characters in my password if it’s not required.r3.30 1.292
#Proactive Awareness (8.36% of variance explained; λ= 1.337)µ σ
F8 When someone sends me a link, I open it without first verifying where it goes.r4.01 1.014
F11 I know what website I’m visiting based on its look and feel, rather than by looking at the URL bar.r3.17 1.077
F16 I submit information to websites without first verifying that it will be sent securely (e.g., SSL, “https://”, a lock icon).r3.69 1.102
F10 When browsing websites, I mouseover links to see where they go, before clicking them. 3.69 1.027
F7 If I discover a security problem, I continue what I was doing because I assume someone else will fix it.r4.08 0.976
#Updating (6.77% of variance explained; λ= 1.082)µ σ
F1 When I’m prompted about a software update, I install it right away. 3.07 1.035
F2 I try to make sure that the programs I use are up-to-date. 3.78 0.890
F9 I verify that my anti-virus software has been regularly updating itself. 3.55 1.228
Table 4. The final questions for the Security Behavior Intentions Scale and associated sub-scales. The means of reverse-scored questions (denoted by
r) have been recoded. The final question numbers have been enumerated from F1-F16 in order to differentiate them from prior exploratory questions.
Responses were reported on the following scale: Never (1),Rarely (2),Sometimes (3),Often (4), and Always (5).
Participants’ genders were split fairly evenly: 46.8% were fe-
male, 52.8% were male, and two declined to specify. Partici-
pants’ ages ranged from 19 to 71 (µ= 34.3,σ= 10.78), me-
dian income was between $35,000 and $50,000, and 37.8% of
participants held bachelor’s degrees (an additional 14% had
completed some or more graduate school). We therefore be-
lieve our sample is comparable to our previous samples and
is also representative of the U.S. online population.
We performed PCA using a Varimax rotation, extracted four
components, and observed that all items were loaded on the
same unique component that they were in the previous exper-
iment. These four components predicted 56.6% of variance,
and are shown in Table 4 along with the mean responses.
Next, we performed CFA by examining the goodness-of-fit of
our data to the latent variable model. Multiple metrics showed
that our data adhered to the model. We did not rely on the
chi-square goodness-of-fit test, because of our large sample
size [44]. Instead, we examined the relative chi-square [28],
the Root Mean Square Error of Approximation (RMSEA), the
Standardized Root Mean Square Residual (SRMR), the Com-
parative Fit Index (CFI), and the Tucker-Lewis Index (TLI).
Our relative chi-square statistic, χ2/df, was approximately
2.71. While there is no firm consensus on the exact cutoff
value to indicate a “good” fit, our statistic is below 3, which
is the more stringent requirement recommended by Hair et
al. [18]. Our analysis yielded a RMSEA of 0.058 and a
SRMR of 0.050, which are both below recommended cut-
off points; Hu and Bentler recommend a cutoff of 0.06 for
RMSEA and 0.08 for SRMR [20]. Finally, our CFI and TLI
were 0.920 and 0.902, respectively, which are above the 0.90
cutoff recommended by Netemeyer et al. [29].
Reliability
We calculated the composite reliability of the four sub-scales
(Table 5) [17]. Since these values are all above the threshold
of 0.6 recommended by Bagozzi and Yi [3]. Additionally,
Securement Passwords Awareness Updating
CR 0.78 0.78 0.64 0.69
ICC 0.89 0.86 0.81 0.81
Table 5. Each sub-scale’s composite reliability (CR) and average mea-
sures intraclass correlation coefficient (ICC) during test-retest.
we conducted a test-retest procedure to make sure responses
were stable over time. Eleven days after our 500 participants
had completed our survey, we invited them to participate in
a followup survey for an additional $1 payment. Of the 500
invited, 354 participated. We calculated the average measures
intraclass correlation coefficient between their sub-scale av-
erage scores (Table 5), and found that all of them were above
the 0.6 threshold recommended by Weir [45].
Relation to Psychological Constructs
We explored whether certain psychological constructs may
be correlated with an individual’s security behaviors. Thus,
participants completed several additional scales:
•Domain-Specific Risk-Taking Scale (DoSpeRT) [4],
which measures propensity for risk-taking across five di-
mensions: ethical (1), financial (2), health/safety (3), recre-
ational (4), and social (5).
•General Decision-Making Style (GDMS) [37], which
measures how people approach decision-making with re-
gard to five dimensions: rational (1), avoidant (2), depen-
dent (3), intuitive (4), and spontaneous (5).
•Need for Cognition (NFC) [7], which is a unidimensional
scale that measures the propensity to engage in “cognitive
endeavors.”
•Barratt Impulsiveness Scale (BIS) [31], which measures
impulsivity across three dimensions: attention (1), motor
(2), and non-planning (3).
•Consideration for Future Consequences (CFC) [21],
which is a unidimensional scale measuring how much at-
tention people pay to long-term consequences.
Securement Passwords Awareness Updating
DoSpeRT1
∗∗-.201 ∗∗ -.226 ∗∗ -.201
DoSpeRT2
DoSpeRT3
∗∗-.204 ∗∗ -.164
DoSpeRT4
DoSpeRT5
∗.141
GDMS1
∗∗.145 ∗∗ .224 ∗∗ .229
GDMS2
∗-.133 ∗∗-.220 ∗∗ -.230 ∗∗ -.247
GDMS3
∗∗-.157
GDMS4
GDMS5
∗-.129
NFC ∗∗.164 ∗∗ .290 ∗∗ .231 ∗∗ .253
BIS1
∗∗-.243 ∗-.140 ∗∗ -.218
BIS2
∗∗-.147 ∗∗ -.145
BIS3
∗∗-.235 ∗∗ -.171 ∗∗ -.247
CFC ∗∗.184 ∗∗ .317 ∗∗ .307 ∗∗ .303
Table 6. Correlations between sub-scales and various psychometrics.
∗p < 0.005,∗∗p < 0.001.
Because this work is exploratory, we performed Pearson cor-
relations rather than building a regression model (Table 6). To
counteract effects from repeated testing, we only report cor-
relations significant at the p < 0.005 level. Across all four
sub-scales, participants who were more inquisitive (as deter-
mined by the NFC scale) were more likely to report engaging
in better security practices. We also noted that the highest
correlation across all four sub-scales was with consideration
for future consequences (CFC), suggesting that good security
behaviors are tied to long-term thinking.
Our initial hypothesis was that willingness to take risks in-
volving health or safety would be inversely correlated with
computer security behaviors; we observed this was true, but
only for behaviors involving proactive awareness and keep-
ing software updated. Similarly, across all four sub-scales,
people who engaged in “better” security behaviors were less
likely to procrastinate (as determined by the GDMS avoidant
sub-scale). People scoring low on GDMS dependence scored
high on SeBIS awareness; being proactive about computer se-
curity means not relying on other people. Many of the secu-
rity behaviors also correlated inversely with impulsivity: this
suggests that adhering to security advice involves foresight.
DISCUSSION
The final Security Behavior Intentions Scale (SeBIS) consists
of 16 items, which map uniquely onto four factors: device se-
curement (locking devices using passwords, PINs, etc.), pass-
word generation (creating and using passwords), proactive
awareness (noticing and taking into account environmental
security cues), and updating (ensuring software is up-to-date).
The scale showed high psychometric properties: all sub-
scales were internally consistent, all items loaded highly on
their respective component and did not load highly on other
components, correlations between sub-scales were low, two
administrations of the same test to the same sample showed
high correlations, and some of the sub-scales were correlated
with some other, well-established, psychological constructs.
Both researchers and practitioners can use this new scale for
various ends. Researchers may use it to measure intended se-
curity behaviors and examine how these change between dif-
ferent populations, over time, or following educational inter-
ventions. They may also use the scale to determine users’ in-
volvement in security-related situations, which could mediate
how users may respond to different security mitigations. For
example, researchers who are interested in compliance with
security warnings may use the scale to explore whether some
types of warnings are more effective for one type of user or
another. The scale may also be used to explore its predictive
ability of real world behaviors (e.g., falling for phishing at-
tacks) and estimate the effectiveness of various interventions
(e.g., security education) by measuring changes in the scale
following the interventions, and over time.
Practitioners may use this scale in various organizational set-
tings. First, they may use it to estimate the security behaviors
among their employees and to identify prevalent weaknesses
which could then be remedied using targeted interventions.
This may help reduce the cost, and increase the effectiveness,
of organizational training aimed at improving end-user secu-
rity behavior. Second, they may use the scale to examine the
common behaviors among their employees in order to try to
match the organization’s security policy with employees’ nat-
ural behavior. For example, if an organization realizes that
many of its employees do not lock their computers when step-
ping away, the organization may instigate a policy of forcing
all computers to be locked after a short period of idle time.
Lastly, managers may use the scale to identify “extreme” in-
dividuals (both on the high and low ends of the scale) and
take targeted actions toward these individuals (e.g., reward
the highly secure ones and educate the lowly secure ones).
Another potential implication of this scale could be for public
policy. As described in the beginning, current computer se-
curity recommendations targeted at end-users are highly scat-
tered, inconclusive and inconsistent. This could be very con-
fusing to end-users who are not highly educated on this topic.
Instead, a different type of policy, or set of policies, could be
formed if policy makers knew the current prevalent attitudes
and behaviors of the general public. The scale we developed
can help assess these attitudes and behaviors on a large scale,
and inform policy makers on the specific domains and situ-
ations that pose a potentially higher risk due to users’ lower
adherence to security standards.
Future Work
While we established the reliability and content validity of
SeBIS in this paper, there is still much work to be done. We
showed that it correlates with several well-studied psycho-
metrics. However, it is unclear how they interact with each
other and how they could be used to predict security behav-
iors. In future work, we expect to perform studies to examine
other psychometrics and build a predictive model. One of
our ultimate goals is to be able to use existing psychometrics
to infer a user’s computer security attitudes and behaviors so
that adequate mitigations can be tailored to their needs and
abilities based on observations of other behaviors.
Some may be concerned that SeBIS targets both personal and
business contexts at the same time. While the final scale con-
tains questions that are likely to be impacted by workplace
policies (e.g., screen locking, software updates), we believe
this is a minor concern: like the vast majority of psychomet-
ric tests, responses are self-reported, and self-reports are im-
pacted by exogenous factors (i.e., not just personal factors).
We do not presume to explain why people report behaving
the way they do, just how secure they report behaving. In
fact, if our measure is impacted by policies, it actually makes
it more relevant for organizational experiments: it can show
how a change in workplace policy affected (or failed to affect)
employee behaviors. Still, we agree that this issue does need
to be addressed empirically, which is why studies are needed
to compare SeBIS results between corporate and home users.
Similarly, while SeBIS is reliable and stable over time, it is
not clear how well it correlates with actual security behaviors,
since it only measures intentions. We plan to perform ex-
periments to examine how each of its component sub-scales
correlate with behaviors, as well as how other psychometrics
could be used to predict computer security behaviors. Given
its demonstrated reliability and discriminant validity, our re-
sults show that SeBIS is still a useful tool for measuring inten-
tions and how they might change over time or vary between
individuals. However, if construct validity is also strong, then
in many cases SeBIS may be a reasonable substitute for per-
forming costly and time-consuming laboratory experiments.
ACKNOWLEDGMENTS
This work was supported by NSF under award CNS-1343433.
We would also like to thank Alessandro Acquisti, Chris
Hoofnagle, Diogo Marques, David Wagner, members of the
Berkeley Laboratory for Usable and Experimental Security
(BLUES), and Refjoh¨
urs Lykkewe.
REFERENCES
1. Acquisti, A., and Grossklags, J. Losses, gains, and
hyperbolic discounting: An experimental approach to
information security attitudes and behavior. In
Proceedings of The 2nd Annual Workshop on Economics
and Information Security (WEIS ’03) (2003).
2. Acquisti, A., and Grossklags, J. Privacy and rationality
in individual decision making. IEEE Security & Privacy
(January/February 2005), 24–30.
3. Bagozzi, R. P., and Yi, Y. On the evaluation of structural
equation models. Journal of the academy of marketing
science 16, 1 (1988), 74–94.
4. Blais, A.-R., and Weber, E. U. A domain-specific
risk-taking (dospert) scale for adult populations.
Judgment and Decision Making 1, 1 (2006), 33–47.
5. Bossler, A. M., and Holt, T. J. The effect of self-control
on victimization in the cyberworld. Journal of Criminal
Justice 38, 3 (2010), 227–236.
6. Buchanan, T., Paine, C., Joinson, A. N., and Reips,
U.-D. Development of measures of online privacy
concern and protection for use on the internet. Journal
of the American Society for Information Science and
Technology 58, 2 (2007), 157–165.
7. Cacioppo, J. T., Petty, R. E., and Feng Kao, C. The
efficient assessment of need for cognition. Journal of
personality assessment 48, 3 (1984), 306–307.
8. Copes, H., Kerley, K. R., Huff, R., and Kane, J.
Differentiating identity theft: An exploratory study of
victims using a national victimization survey. Journal of
Criminal Justice 38, 5 (2010), 1045–1052.
9. Cranor, L. F. A framework for reasoning about the
human in the loop. In Proceedings of the 1st Conference
on Usability, Psychology, and Security, USENIX
Association (Berkeley, CA, 2008).
10. Crowne, D. P., and Marlowe, D. A new scale of social
desirability independent of psychopathology. Journal of
consulting psychology 24, 4 (1960), 349.
11. DeVellis, R. F. Scale Development: Theory and
Applications, 2nd ed., vol. 26 of Applied Social
Research Methods Series. Sage Publications, 2003.
12. Egelman, S., Cranor, L. F., and Hong, J. You’ve been
warned: An empirical study of the effectiveness of web
browser phishing warnings. In Proceeding of The 26th
SIGCHI Conference on Human Factors in Computing
Systems, CHI ’08, ACM (New York, NY, USA, 2008),
1065–1074.
13. Egelman, S., Jain, S., Portnoff, R. S., Liao, K.,
Consolvo, S., and Wagner, D. Are you ready to lock?
understanding user motivations for smartphone locking
behaviors. In Proceedings of the 2014 ACM SIGSAC
Conference on Computer & Communications Security,
CCS ’14, ACM (New York, NY, USA, 2014).
14. Everitt, B. S. The Cambridge Dictionary of Statistics.
Cambridge University Press, Cambridge, UK, 2002.
15. Fabrigar, L. R., Wegener, D. T., MacCallum, R. C., and
Strahan, E. J. Evaluating the use of exploratory factor
analysis in psychological research. Psychological
methods 4, 3 (1999), 272.
16. Felt, A. P., Ha, E., Egelman, S., Haney, A., Chin, E., and
Wagner, D. Android permissions: user attention,
comprehension, and behavior. In Proceedings of the
Eighth Symposium on Usable Privacy and Security,
SOUPS ’12, ACM (New York, NY, USA, 2012).
17. Fornell, C., and Larcker, D. F. Evaluating structural
equation models with unobservable variables and
measurement error. Journal of Marketing Research 18, 1
(1981), pp. 39–50.
18. Hair, J. F., Tatham, R. L., Anderson, R. E., and Black, W.
Multivariate Data Analysis, 6 ed. Prentice Hall, 2006.
19. Horn, J. L. A rationale and test for the number of factors
in factor analysis. Psychometrika 30, 2 (1965), 179–185.
20. Hu, L.-t., and Bentler, P. M. Cutoff criteria for fit indexes
in covariance structure analysis: Conventional criteria
versus new alternatives. Structural Equation Modeling:
A Multidisciplinary Journal 6, 1 (1999), 1–55.
21. Joireman, J., Shaffer, M. J., Balliet, D., and Strathman,
A. Promotion orientation explains why future-oriented
people exercise and eat healthy evidence from the
two-factor consideration of future consequences-14
scale. Personality and Social Psychology Bulletin 38, 10
(2012), 1272–1287.
22. Kumaraguru, P., and Cranor, L. F. Privacy Indexes: A
Survey of Westin’s Studies. Tech. Rep.
CMU-ISRI-5-138, Carnegie Mellon University,
December, 2005. http://reports-archive.adm.cs.
cmu.edu/anon/isri2005/abstracts/05-138.html.
23. Malhotra, N. K., Kim, S. S., and Agarwal, J. Internet
users’ information privacy concerns (iuipc): The
construct, the scale, and a causal model. Information
Systems Research 15, 4 (December 2004), 336–355.
24. McKinley, R. K., Manku-Scott, T., Hastings, A. M.,
French, D. P., and Baker, R. Reliability and validity of a
new measure of patient satisfaction with out of hours
primary medical care in the united kingdom:
development of a patient questionnaire. British Medical
Journal 314, 7075 (1997), 193.
25. Meade, A. W., and Craig, S. B. Identifying careless
responses in survey data. Psychological methods 17, 3
(2012), 437.
26. Modic, D., and Lea, S. E. G. How neurotic are scam
victims, really? the big five and internet scams.
http://ssrn.com/abstract=2448130, September 2012.
27. Moon, B., McCluskey, J. D., and McCluskey, C. P. A
general theory of crime and computer crime: An
empirical test. Journal of Criminal Justice 38, 4 (2010),
767–772.
28. Mueller, R. O. Basic principles of structural equation
modeling: An introduction to LISREL and EQS.
Springer, 1996.
29. Netemeyer, R. G., Bearden, W. O., and Sharma, S.
Scaling Procedures: Issues and Applications. SAGE
Publications, 2003.
30. Ng, B.-Y., Kankanhalli, A., and Xu, Y. C. Studying
users’ computer security behavior: A health belief
perspective. Dec. Sup. Systems 46, 4 (2009), 815–825.
31. Patton, J. H., Stanford, M. S., et al. Factor structure of
the barratt impulsiveness scale. Journal of clinical
psychology 51, 6 (1995), 768–774.
32. Peer, E., Vosgerau, J., and Acquisti, A. Reputation as a
sufficient condition for data quality on amazon
mechanical turk. Behavior Research Methods 45, 4
(December 2013).
33. Raˆ
ıche, G., Walls, T. A., Magis, D., Riopel, M., and
Blais, J.-G. Non-graphical solutions for cattell’s scree
test. Methodology: European Journal of Research
Methods for the Behavioral and Social Sciences 9, 1
(2013), 23.
34. Ray, J. J. Reviving the problem of acquiescent response
bias. The J. of Social Psychology 121, 1 (1983), 81–96.
35. Robinson, J. P., Shaver, P. R., and Wrightsman, L. S.
Criteria for scale selection and evaluation. In Measures
of personality and social psychological attitudes.
Academic Press, 1991, ch. 1, 1–16.
36. Saucier, G. Mini-markers: A brief version of goldberg’s
unipolar big-five markers. Journal of Personality
Assessment 63, 3 (1994), 506–516.
37. Scott, S. G., and Bruce, R. A. Decision-making style:
The development and assessment of a new measure.
Educational and psychological measurement 55, 5
(1995), 818–831.
38. Sherizen, S. Criminological concepts and research
findings relevant for improving computer crime control.
Computers & Security 9, 3 (1990), 215–222.
39. Stanton, J. M., Stam, K. R., Mastrangelo, P., and Jolton,
J. Analysis of end user security behaviors. Computers &
Security 24, 2 (2005), 124–133.
40. Strahan, R., and Gerbasi, K. C. Short, homogeneous
versions of the marlowe-crowne social desirability scale.
Journal of clinical psychology (1972).
41. United States Computer Emergency Readiness Team.
Tips. https://www.us-cert.gov/ncas/tips. Accessed:
September 12, 2014.
42. United States Department of Homeland Security.
National Cyber Security Awareness Month 2014.
http://www.dhs.gov/
national-cyber- security-awareness- month-2014,
September 8 2014. Accessed: September 12, 2014.
43. Verizon. Security. http://www.verizon.com/Support/
Residential/Internet/FiosInternet/General+
Support/Security/Security.htm, 2014. Accessed:
September 12, 2014.
44. Wang, L., Fan, X., and Willson, V. L. Effects of
nonnormal data on parameter estimates and fit indices
for a model with latent and manifest variables: An
empirical study. Structural Equation Modeling: A
Multidisciplinary Journal 3, 3 (1996), 228–247.
45. Weir, J. P. Quantifying test-retest reliability using the
intraclass correlation coefficient and the sem. The
Journal of Strength & Conditioning Research 19, 1
(2005), 231–240.
46. Wilson, M., and Hash, J. Building an Information
Technology Security Awareness and Training Program.
Special Publication 800-50, National Institute of
Standards and Technology, Gaithersburg, MD, US,
October 2003. http://csrc.nist.gov/publications/
nistpubs/800-50/NIST- SP800-50.pdf.
47. Wogalter, M. S. Communication-Human Information
Processing (C-HIP) Model. In Handbook of Warnings,
M. S. Wogalter, Ed. Lawrence Erlbaum Associates,
2006, 51–61.
48. Woodruff, A., Pihur, V., Consolvo, S., Brandimarte, L.,
and Acquisti, A. Would a privacy fundamentalist sell
their dna for $1000...if nothing bad happened as a
result? the westin categories, behavioral intentions, and
consequences. In Proceedings of the 2014 Symposium
on Usable Privacy and Security, USENIX Association
(2014), 1–18.
