What pairs of questions test respondent's consistency?
They don't exist ;-)
The SUS wasn't designed with pairs of questions in mind. The SUS wasn't designed to test particular factors (although there has been some interesting post-hoc factor analysis of the SUS).
To quote from the original SUS paper:
SUS is a Likert scale. It is often assumed that a Likert scale is
simply one based on forced-choice questions, where a statement is made
and the respondent then indicates the degree of agreement or
disagreement with the statement on a 5 (or 7) point scale. However,
the construction of a Likert scale is somewhat more subtle than this.
Whilst Likert scales are presented in this form, the statements with
which the respondent indicates agreement and disagreement have to be
The technique used for selecting items for a Likert scale is to
identify examples of things which lead to extreme expressions of the
attitude being captured. For instance, if one was interested in
attitudes to crimes and misdemeanours, one might use serial murder and
parking offences as examples of the extreme ends of the spectrum. When
these examples have been selected, then a sample of respondents is
asked to give ratings to these examples across a wide pool of
potential questionnaire items. For instance, respondents might be
asked to respond to statements such as “hanging’s too good for them”,
or “I can imagine myself doing something like this”.
Given a large pool of such statements, there will generally be some
where there is a lot of agreement between respondents. In addition,
some of these will be ones where the statements provoke extreme
statements of agreement or disagreement among all respondents. It is
these latter statements which one tries to identify for inclusion in a
Likert scale, since, we would hope that, if we have selected suitable
examples, there would be general agreement of extreme attitudes to
them. Items where there is ambiguity are not good discriminators of
attitudes. For instance, while one hopes that there would be a
general, extreme disagreement that “hanging’s too good” for those who
perpetrate parking offences, there may well be less agreement about
applying this statement to serial killers, since opinions differ
widely about the ethics and efficacy of capital punishment.
SUS was constructed using this technique. A pool of 50 potential
questionnaire items was assembled. Two examples of software systems
were then selected (one a linguistic tool aimed at end users, the
other a tool for systems programmers) on the basis of general
agreement that one was “really easy to use” and one was almost
impossible to use, even for highly technically skilled users. 20
people from the office systems engineering group, with occupations
ranging from secretary through to systems programmer then rated both
systems against all 50 potential questionnaire items on a 5 point
scale ranging from “strongly agree” to “strongly disagree”.
The items leading to the most extreme responses from the original pool
were then selected. There were very close intercorrelations between
all of the selected items (± 0.7 to ± 0.9). In addition, items were
selected so that the common response to half of them was strong
agreement, and to the other half, strong disagreement. This was done
in order to prevent response biases caused by respondents not having
to think about each statement; by alternating positive and negative
items, the respondent has to read each statement and make an effort to
think whether they agree or disagree with it.