study. For a specific behaviour and population one or more antecedents may indeed
not be predictive, without disproving the theory. For example, social approval may
be crucial to some health behaviours but not to others. Thus finding that a particular
cognition is not relevant to a particular behaviour does not disconfirm the theory.
However, finding that none of the cognitions specified by the theory predicted useful
proportions of the variance in intention or behaviour across behaviours would indeed
disconfirm the theory. In fact, available evidence suggests that the theory is very useful,
explaining on average 40–50 per cent of the variance in intention and 21–36 per cent
of the variance in behaviour across studies (Conner and Sparks, 2005; McEachan
et al., 2011).
Ogden also claimed that the theories contain only analytic truths (as opposed to
synthetic or empirical truths that are based on evidence) because the correlations
observed between measured cognitions are likely to be attributable to overlap in the
way the constructs are measured. She claimed that this argument extends to measures
of behaviour because these are often based on self-report. This interpretation of the
literature has been disputed for two main reasons. First, it is not at all apparent that this
explanation would account for the observed patterns of correlations among cognitions
that are commonly reported in the literature. Second, high levels of prediction of
behaviour are also found with objective measures of behaviour that do not rely on self-
report and thus cannot be biased in the way Ogden describes. For example, Armitage
and Conner (2001) in their meta-analysis of the TPB showed that intention and
perceived behavioural control still accounted for an impressive 21 per cent of variance
in behaviour when behaviour was objectively measured across a number of studies (see
also McEachan et al., 2011 who reviewed prospective tests of the TPB to a range of
health behaviours using either self-reported or objectively measured behaviour).
This examination of the percentage of variance explained by SCMs has discouraged
some health psychologists. For example, Mielewczyk and Willig (2007: 818–819) in
reviewing this evidence conclude that:
the TPB is therefore unable to account for around 60 per cent of the variance in
intentions and for up to almost 80 per cent of that in behaviour. Since the SCM
approach is directed purely at providing explanations of variance in outcomes, the
extent of that unexplained across such a large body of literature is highly damning.This pessimism is unfounded for a number of reasons (see Abraham, Sheeran and
Orbell, 1998, for a useful discussion). First, as Sutton (1997) notes, the percentage
of variance explained by any model, including physiological models of symptom
appearance, is directly related to the reliability of the measures employed. The
maximum variance that can be accounted for will always be the square root of the
product of the two reliabilities. Thus there are inherent measurement limitations on
the percentage of variance that any model can explain. Second, when cognition and
behaviour measures are not ‘compatible’ (i.e. do not refer to the identical action, target,
time and context), the R^2 will be reduced (Ajzen and Fishbein, 1980). Similarly, if the
number of response options used to measure cognitions and behaviour is not equal
(e.g. if attitude is measured on a 7-point scale and behaviour on a 3-point scale), this
will also reduce R^2. Finally, when sampling biases lead to a restricted range in either
the independent (cognition) or the dependent (behaviour) variable compared to the
HEALTH COGNITIONS AND BEHAVIOURS 157