Information Technology Reference
In-Depth Information
coding procedure achieve higher accuracy than
ordinal scales, since the former do not require
human judgment and interpretation, thus render-
ing the use of a second coder unnecessary. As a
check on intracoder reliability and consistency
(Bauer, 2000; Stempel & Wesley, 1981), all texts
were recoded in August 2003 to eliminate the
possibility of error. The agreement between the
two codings was 98.71%.
Content analysis examines a text only in light
of the questions included in the coding frame.
However, since content analysis should not only
be concerned with what a text is about but also
with its vocabulary and, in particular, lexical
patterns in the text corpus (Bauer, 2000), a word
frequency list of the total text corpus was created
to identify words of potential interest and to un-
cover patterns that otherwise may not be obvious
(Wolfe et al., 1993). The most frequent words in
a corpus were a clear indicator of those concepts
and ideas that received special attention in the
text (Krippendorff, 1980). The word frequency
list was calculated with WordSmith Tools, a
software tool for textual analysis. The initial list
of 4,016 different words (representing 108,570
running words) was lemmatized automatically
to remove inflectional suffixes (e.g., plural end-
ings,
-ing
forms) and to separate contracted verb
forms. In addition, all words of less than three
letters and all words appearing only once were
eliminated, as they were considered insignificant
to the content of a text (Lebart et al., 1998). For
the same reason, all grammatical words (e.g., the
auxiliary verbs
be
and
have
, articles, conjunc-
tions, prepositions), numbers, names (e.g., for
cities, states, companies, months), and Internet
domain extensions were excluded. The elimina-
tion of all non-content words produced a list of
1,637 content words, representing 52.15% of the
running words.
This list was then used for the computer-as-
sisted textual analysis. Typically, such frequency
counts are subject to distortions. For one, words
may have multiple meanings (polysemy), and for
another, authors deliberately may vary the terms
they use when they refer to the same concept
(synonymy) (Weber, 1985). Therefore, frequency
counts require human interpretation, which can
be facilitated by keyword-in-context (KWIC)
lists. These lists show all instances of a certain
word in their immediate contexts, thus helping to
identify lexical patterns (Stubbs, 2001).
results oF the Content
analysis
general Characteristics
On average, a privacy policy consists of 1.12 Web
pages and 2,157 words. Ninety percent of the pri-
vacy policies are accessible with just one click on
a link on the bottom of the Web site's home page.
Table 2 gives an overview of additional charac-
teristics of these documents, including document
name, contact information, and last update. Most
commonly, the documents are referred to as
poli-
cies
or
statements
. Other documents bear titles
like
notices
,
policy statements
,
information
, or
have no titles at all. Almost all of them provide
some form of contact information. A significant
portion of the companies (44%) do not provide any
indication of when they last updated their privacy
policies. While 86% of the companies point out
that their privacy policies are subject to change,
only 20% promise to post a notice on the Web
site prior to the change, and 14% notify registered
users of significant changes by e-mail.
The majority of the companies (62%) do not
display or refer to any privacy seal; 30% have one
seal, and 8% have two seals. The seals displayed
include TRUSTe (n = 11), BBB Online (n = 8), and
AOL (n = 4). Only 12% of the companies have
their privacy practices and the enforcement of
their policies audited either by internal review-
ers or as part of external reviews by experts or
consulting firms.
Search WWH ::
Custom Search