ArticlePDF Available

Prospective Prediction of Health-related Behaviours with the Theory of Planned Behaviour: A Meta-analysis

Taylor & Francis
Health Psychology Review
Authors:

Abstract and Figures

This meta-analysis explored the efficacy of the Theory of Planned Behaviour (TPB) dependent on behaviour and methodological moderators. A lack of hierarchical analysis in previous reviews risks confounding these moderators. Here moderating roles of behaviour type, length of follow-up, sample age and behavioural measure are explored hierarchically amongst prospective tests of the TPB, controlling for past behaviour where possible. Searching identified 237 prospective tests from 206 articles. Random-effects meta-analytic procedures were used to correcting correlations for sampling and measurement error. Behaviour type moderated the model; physical activity and diet behaviours were better predicted (23.9% and 21.2% variance explained, respectively) whilst risk, detection, safer sex and abstinence from drugs were poorly predicted (between 13.8 and 15.3% variance explained). Methodological moderators were also apparent: age of sample moderated relations with student samples better predicted for physical activity, and adolescent samples better predicted for abstinence behaviours. Behaviours assessed in the shorter term, and those assessed with self-reports (compared with objective measures) were also better predicted. Both behavioural and methodological characteristics moderated relations amongst model components. The results can aid selection of appropriate targets upon which to base interventions.
Content may be subject to copyright.
Prospective prediction of health-related behaviours with the Theory of
Planned Behaviour: a meta-analysis
Rosemary Robin Charlotte McEachan
a,b
*, Mark Conner
b
, Natalie Jayne Taylor
b
and Rebecca Jane Lawton
b
a
Bradford Institute for Health Research, Bradford Teaching Hospitals NHS Foundation Trust,
Bradford Royal Infirmary, Duckworth Lane, Bradford BD9 6RJ, UK;
b
Department of
Psychology, Institute of Psychological Sciences, University of Leeds, Leeds LS2 9JT, UK
(Received 7 December 2009; final version received 3 September 2010)
This meta-analysis explored the efficacy of the Theory of Planned Behaviour
(TPB) dependent on behaviour and methodological moderators. A lack of
hierarchical analysis in previous reviews risks confounding these moderators.
Here moderating roles of behaviour type, length of follow-up, sample age and
behavioural measure are explored hierarchically amongst prospective tests of
the TPB, controlling for past behaviour where possible. Searching identified
237 prospective tests from 206 articles. Random-effects meta-analytic procedures
were used to correcting correlations for sampling and measurement error.
Behaviour type moderated the model; physical activity and diet behaviours were
better predicted (23.9% and 21.2% variance explained, respectively) whilst risk,
detection, safer sex and abstinence from drugs were poorly predicted (between
13.8 and 15.3% variance explained). Methodological moderators were also
apparent: age of sample moderated relations with student samples better
predicted for physical activity, and adolescent samples better predicted for
abstinence behaviours. Behaviours assessed in the shorter term, and those
assessed with self-reports (compared with objective measures) were also better
predicted. Both behavioural and methodological characteristics moderated
relations amongst model components. The results can aid selection of appropriate
targets upon which to base interventions.
Keywords: theory of planned behaviour; meta-analysis; random effects; health
behaviour
The ability to effectively predict and explain health-related behaviour is important to
a range of researchers and professionals concerned with developing interventions to
change those behaviours. A variety of social cognition models exist which purport
to delineate the key determinants of behaviour (see Conner & Norman, 2005). The
Theory of Planned Behaviour (TPB: Ajzen, 1988, 1991; an extension of the earlier
Theory of Reasoned Action, TRA: Ajzen & Fishbein, 1980) has been one of the most
widely tested models of the factors influencing health-related behaviour. The TPB is
a parsimonious model that has been applied to a wide range of behaviours (for
reviews see Albarracin, Johnson, Fishbein, & Muellerleile, 2001; Albarracin,
Kumkale, & Johnson, 2004; Armitage & Conner, 2001; Cooke & French, 2008;
*Corresponding author. Email: r.mceachan@leeds.ac.uk; rosie.mceachan@bradfordhospitals.
nhs.uk
Health Psychology Review
2011, 148, iFirst
ISSN 1743-7199 print/ISSN 1743-7202 online
# 2011 Taylor & Francis
DOI: 10.1080/17437199.2010.521684
http://www.informaworld.com
Downloaded By: [Conner, Mark T] At: 17:41 15 February 2011
Godin & Kok, 1996; Hagger, Chatzisarantis, & Biddle, 2002; Hausenblas, Carron,
& Mack, 1997; Sheeran & Taylor, 1999). Its popularity is due in part to the fact that
it is clearly operationalised with guidelines concerning how to measure (Ajzen, 2006;
Conner & Sparks, 2005; French & Hankins, 2003), analyse (Hankins, French,
& Horne, 2000) and develop interventions (Sutton, 2002) using the theory. Although
the TPB is criticised for not providing specific guidance on change techniques, its
parsimony makes the model particularly useful in applied settings. The present
manuscript reports a meta-analysis of prospective tests of the TPB with health-
related behaviours controlling for the effects of past behaviour. In addition to
reporting the overall power of the model to predict behaviour, we also assess the
moderating impacts of behaviour type, sample (age group) and methodological
factors (length of follow-up and nature of behavioural measure) and address some of
the weaknesses in previous meta-analyses of the TPB.
Focus of the meta-analysis
The TPB states that behaviour is determined by an individual’s behavioural intention
(INT) and perceived behavioural control (PBC). Intention is held to be the motiva-
tional component that spurs an individual to engage in a particular behaviour. PBC
captures the extent to which people have control over engaging in the behaviour.
Intentions, in turn, are determined by an individual’s attitude towards the behaviour
(ATT: e.g., whether engaging in the behaviour is evaluated to be positive or negative),
subjective norms (SN: e.g., perceptions of whether others think one should engage in
a behaviour) and PBC. These are referred to as direct predictors. Different sets of
beliefs underlie ATT, SN and PBC, and are referred to as indirect predictors.
According to the model attitudes are comprised of beliefs about the likelihood of
salient outcomes of the behaviour weighted by the evaluation of each outcome.
Subjective (injunctive) norms are comprised of beliefs about whether salient referents
think that one should perform the behaviour weighted by the motivation to comply
with that referent. Finally, PBC is comprised of beliefs about the frequency of
occurrence of facilitating or inhibiting factors towards engaging in a behaviour
weighted by the perceived power of each factor to impact on engagement with the
behaviour (Conner & Sparks, 2005). Importantly in relation to measurement, Ajzen
(2006) recommends that measures of behaviour and TPB components follow the
principle of compatibility, that is they are matched on their target (e.g., engaging in
physical activity), action (e.g., of at least moderate intensity), and also context (e.g.,
on at least 5 days of the week) and time frame (e.g., over the next 3 months) in order
to gain the strongest relationships between model components.
Overall, the TPB has been shown to be an adequate predictor of intention and
behaviour explaining 4049% of the variance in intention and 2636% of the
variance in behaviour (Ajzen, 1991; Armitage & Conner, 2001; Godin & Kok, 1996;
Hagger et al., 2002; Schulze & Whittmann, 2003; Trafimow, Sheeran, Conner, &
Finlay, 2002). However, accounting for the effects of past behaviour has proved
a challenge for the TPB. Within the context of physical activity, Hagger et al. (2002)
found that past behaviour contributed an extra 19% variance to the prediction of
physical activity controlling for other TPB variables. These authors also found that
past behaviour significantly attenuated the relationships between attitude and
intention and between intention and behaviour (see also Hagger, Chatzisarantis,
2 R.R.C. McEachan et al.
Downloaded By: [Conner, Mark T] At: 17:41 15 February 2011
Biddle, & Orbell, 2001). In extreme cases, past behaviour has been found to be the
only significant predictor of prospective behaviour (e.g., Norman & Smith, 1995).
Thus, there is no doubt that inclusion of past behaviour improves the prediction of
future behaviour. However, as Sutton (1998) notes, prediction has limited value in an
applied setting, and it is perhaps more useful to be able to explain behaviour so that
interventions can be developed and behaviour changed. Testing the impact of TPB
variables on intention and behaviour controlling for the effects of past behaviour was
the first focus of the present meta-analysis and one not addressed by previous general
reviews of the TPB.
A second focus of the present review was to assess the moderating impact of
behaviour type on the effectiveness of the TPB. The efficacy of the model varies
depending on behaviour type. For example, Godin and Kok (1996) found that
clinical and screening behaviours were predicted poorly by the model (with 15.6%
explained variance), whereas HIV- and AIDS-related behaviours were relatively well
predicted (42.1% explained variance). The prediction of intention was less hetero-
geneous, but values still ranged from 39.6% (automobile behaviours) to 46.8% (oral
hygiene). Other reviews of the TRA/TPB have also reported variation in predictive
power across behaviours. For example, Van den Putte (1993) found that the
prediction of intention from attitude and SN was higher for procreation behaviours
(R
2
0.80) compared with health behaviours (R
2
0.61) and other behaviours
(R
2
0.71). There is also evidence of the differential predictive utility of the TPB
constructs for different behaviours. For example, the SN intention relationship
appears to be stronger for condom use (r
0.390.42; Albarracin et al., 2001, 2004;
Sheeran & Taylor, 1999) and screening behaviour (r
0.41; Cooke & French, 2008)
than for physical activity (r
0.250.27; Hagger et al., 2002; Hausenblas et al.,
1997). In one meta-analysis, Randall and Wolff (1994) classified behaviours into
seven types (encompassing domains such as food- and beverage-consumption
activities, sexual reproductive behaviour and political/voting behaviour) and reported
that behaviour type explained over 19% of the variance in the strength of
the intention behaviour relationship. Taken as a whole, these variations are not
necessarily problematic for the model. Ajzen and Fishbein (1980) are clear that
variations in the size of relationships between constructs are likely for different
behaviours and also populations. The present review provides an examination of
these systematic variations that should be useful to those developing interventions
and wanting to identify the most appropriate target for intervention.
The evidence cited above suggests that the nature of the behaviour may be an
important factor that has implications for the predictive utility of the TPB as well as
its potential for use in the development of effective interventions. Research has
suggested that health behaviours can be distinguished along dimensions based on
individual perceptions (McEachan, Lawton, & Conner, 2010) or theoretical
constructs (e.g., health promoting vs. health risk, frequency of performance,
Ouellette & Wood, 1998; habits vs. discrete behaviours, see Borland, 2010; initiation
vs. maintenance, Van Stralen, De Vries, Mudde, Bolman, & Lechner, 2009; emotions
and hedonic behaviours, see Cameron, 2010; also Lawton, Conner, & McEachan,
2009). Authors have suggested that the nature of these similarities or differences
amongst behaviours may be important moderators of the ease with which we can
predict or change them (e.g., Johnston & Dixon, 2008; McEachan et al., 2010;
Michie et al., 2005). The grouping of behaviours used here was based on existing
Health Psychology Review 3
Downloaded By: [Conner, Mark T] At: 17:41 15 February 2011
classifications that propose both a hierarchical (Roysamb, Rise, & Kraft, 1997) and
functional approach (Rothman & Salovey, 1997) to the classification of health
behaviours. Behaviours were therefore divided into those that improve health (health
promoting) and those that reduce health (health risk). Consistent with Rothman and
Salovey (1997), the behaviours were then further divided into three groups:
preventive, detection and curative behaviours. The numbers of studies available led
to a number of these groups being dropped or further sub-divided into additional
groups based on descriptive characteristics (see Method section). This analysis
allowed us to explore differences in prediction (for constructs and the model as
a whole) between behaviours in order to identify for which types of health
behaviours the TPB is most or least effective. We were also able to explore broad
differences based on the theoretical dimensions of health-promoting vs. health-risk
behaviours (Conner & Norman, 2005) and frequent vs. infrequent health behaviours
(Ouellette & Wood, 1998).
A third focus of the present review was to assess the moderating impact of sample
characteristics on the effectiveness of the TPB. In addition to differences in
prediction due to behaviour type, the TPB appears to be moderated by sample
moderators. Few sample characteristics have been explicitly explored as moderators
in the context of the TPB. In addition, published studies are not generally stratified
by meaningful sample characteristics (e.g., social class) allowing the meta-analytic
examination of such differences. One exception is the age of the sample. For example,
both Hagger et al. (2002) and Sheeran and Orbell (1998) report the intention
behaviour relationship to be significantly stronger for adult samples (r
0.57
and 0.50, respectively) compared with younger samples (r
0.48 and 0.25,
respectively). In the present meta-analysis, we distinguished between adult, student
and adolescent samples. Differences between such groups are of interest for
theoretical and practical reasons. For example, it has been speculated that
adolescents compared to adults are driven less by rational considerations and
more by affective associations, impulsivity and direct social pressure (e.g., Gibbons,
Houlihan, & Gerrard, 2009; Hofmann, Friese, & Wiers, 2008) perhaps because of a
less well-developed executive control function (Hall, Fong, Epp, & Elias, 2008). As
such one might expect attitudes to be less important and SN more important
predictors in adolescent samples. The TPB has also been widely applied in student
samples. Such samples are known to be younger, better educated, from higher socio-
economic status groups and more motivated to accurately respond than the general
adult population (Hooghe, Stolle, Maheo, & Visser, 2010; Sears, 1986). It is therefore
of interest to know whether such differences translate into a better fit for the TPB in
student compared to adult samples.
A fourth focus of this meta-analysis was to assess the extent to which
methodological factors moderate the effectiveness of the TPB. In terms of
methodological factors, two issues are particularly pertinent in relation to the
application of the TPB to prospective studies of health-related behaviours: length of
follow-up and objective vs. self-report behaviour measures. The length of follow-up
between measurement of TPB variables and subsequent measurement of behaviour is
held to be a limiting condition of the TPB (i.e., TPB variables should only predict
behaviour to the extent that they remain stable between the point at which they are
measured and the point at which the behaviour occurs and this should be less likely
as the time interval increases; Ajzen, 1985, 1991). However, empirical findings on this
4 R.R.C. McEachan et al.
Downloaded By: [Conner, Mark T] At: 17:41 15 February 2011
issue are mixed. In one meta-analysis on this issue, Randall and Wolff (1994) found
no difference in the ability of intention to predict behaviour for time periods as long
as 15 years. In relation to physical activity, Hausenblas et al. (1997) also reported no
difference in the ability of intention to predict proximal vs. distal measures in
behaviour amongst a small number of studies (k 4). In contrast, for condom use,
relations between intention and behaviour are typically weaker over longer time
intervals (Albarracin et al., 2004; Sheeran & Orbell, 1998). In relation to measures of
behaviour, current evidence exists to suggest that self-report measures of behaviour
are better explained than objective measures of behaviour (Armitage & Conner,
2001). This may be attributable to self-report measures being more open to
consistency biases or being more readily assessed with measures strictly following
the principle of compatibility (Armitage & Conner, 2001). The present review took
the opportunity to confirm this pattern of effects in a larger sample of prospective
applications of the TPB to health-related behaviours.
An important issue in considering these various moderators of the efficacy of the
TPB in predicting health-related behaviours is the fact that they are not necessarily
independent. Previous meta-analyses of the TPB have not satisfactorily addressed
this issue. For example, screening studies typically report objective measurements of
behaviour (e.g., actual attendance at appointment). If screening behaviours are
posited to be relatively poorly predicted (cf. Cooke & French, 2008; Godin & Kok,
1996), we have no way of determining whether this is due to specific characteristics of
the behaviour or the type of behavioural measure. One way to address this problem is
to report the degree of confounding of moderators. A second is to examine the
influence of different moderators in a hierarchical fashion. In the present review, we
report the degree of confounding and where possible use a hierarchical approach to
examine the impact of different moderators. In particular, in examining the impacts
of age of sample and length of follow-up, analyses are performed within each
identified group of health behaviours.
Limitation of previous meta-analyses
Previous meta-analyses and reviews are also subject to limitations that the present
meta-analysis attempts to address. First, many of them include both cross-sectional
and prospective measures of behaviour (Ajzen, 1991; Godin & Kok, 1996; Hagger
et al., 2002; Hausenblas et al., 1997; Notani, 1998; Sheeran, Abraham, & Orbell,
1999; Sheppard, Hartwick, & Warshaw, 1988; Trafimow et al., 2002; van den Putte,
1993). Measuring behaviour cross-sectionally means that studies are actually
providing measures of past or current behaviour rather than future behaviour, and
are therefore less appropriate measures to use as a test of the sufficiency of the model
than future behaviour (see Weinstein, 2006). Both Albarracin et al. (2001, 2004) and
Manning (2009) found that relations between intention and behaviour were stronger
for studies measuring concurrent behaviour than those measuring behaviour
prospectively. Given the problems interpreting cross-sectional TPB studies the
current meta-analysis therefore only includes studies reporting prospective measure-
ment of behaviour.
Second, previous meta-analyses have varied according to whether they have used
fixed-effects (FE) or random-effects (RE) meta-analytic procedures. It has been
argued that RE meta-analytic procedures are more appropriate models on which to
Health Psychology Review 5
Downloaded By: [Conner, Mark T] At: 17:41 15 February 2011
base meta-analyses of this type as they allow for the possibility that effect sizes may
vary randomly as if they were sampled from a super-population rather than
assuming a FE (Hedges, 1992; Hunter & Schmidt, 2000). By correcting for both error
accrued from sampling studies from a population of studies, and error accrued from
the sampling of populations from a super-population, the RE meta-analytic
procedures employed here provide a more accurate and conservative method of
summarising information from a range of studies and more appropriate reflections of
real-world data (Field, 2003; Hunter & Schmidt, 2000). In the current context, this
would mean allowing for the possibility that the magnitude of relations vary by
behaviour type (RE) vs. assuming a fixed magnitude regardless of behaviour type
(FE).
Third, all meta-analyses routinely correct for sampling error within their studies
by calculating frequency-weighted correlations, but not for other sources of error.
There are statistical artefacts in addition to sampling error, which can serve to
attenuate the true strength of correlations between model components (Hunter
& Schmidt, 1990). One of these artefacts is measurement error incurred by less than
perfect reliability of measures used to assess the variables under consideration.
The result is that the true correlation is lowered by a factor correspondent to the
square root of the estimate of the reliability. No general meta-analyses of the TPB
have corrected for measurement error in addition to sampling error (but see Hagger
et al., 2002, in relation to physical activity; Manning, 2009, in relation to norms). By
correcting for such artefacts, the present review provides better estimates of the true
relationship between TPB model components.
Summary
In summary, the present meta-analysis explored the role of behavioural, sample and
methodological moderators within the TPB, whilst controlling for past behaviour.
We also examined a number of moderator variables. Most importantly we examined
whether type of behaviour moderated relations amongst model components. In
addition, we examined the moderators of age of sample, length of follow-up and
type of behaviour assessment (self-report or objective). Where possible, analyses were
conducted hierarchically to reduce confounding of moderators. Finally, to overcome
some of the limitations of previous meta-analyses, the current research included only
prospective tests of the TPB, applied the RE meta-analytic procedures (Hunter
& Schmidt, 1990, 2004) and corrected correlations individually for both sampling
and measurement error.
Method
Literature review and inclusion/exclusion criteria
To ensure that all relevant studies were included in the review, a variety of search
strategies were used. Electronic databases (ISI Web of Science ISI WOS;
Cumulative Index to Nursing & Allied Health Literature CINAHL; MEDLINE;
and PsycINFO) were searched to 10 May 2010 using the following search strings
1
:
(1) attitud* and norm* and control and intention*; (2) theory of planned behavi*;
and (3) planned behavi* and Ajzen.
6 R.R.C. McEachan et al.
Downloaded By: [Conner, Mark T] At: 17:41 15 February 2011
Second, citation searches were performed within ISI WOS on three key papers
(Ajzen, 1991; Armitage & Conner, 2001; Godin & Kok, 1996) within the TPB
literature. Third, reference lists of all included articles were manually searched.
Fourth, the contents pages of several key journals were hand searched: British
Journal of Health Psychology, 19962010; Health Education Research, 19862010;
Health Psychology, 19822010; Journal of Applied Social Psychology, 19932010; and
Psychology and Health, 19892010. Finally, key authors were contacted to identify
any further published work eligible for inclusion. In total 5802 potential papers were
identified. These were subjected to the following inclusion/exclusion criteria:
1. Studies had to report prospective tests of health- or health-risk-related
behaviours. Thus, studies with cross-sectional or retrospective designs or
those which did not provide a measure of behaviour at follow-up were
excluded. Measurement of behaviour could range from self-report to
objective; however, studies providing only a stage of change algorithm of
behaviour were excluded (e.g., Courneya, Estabrooks, & Nigg, 1997).
2. For the purposes of the current study, health behaviours were defined as
behaviours which impact or have the potential to impact upon the health of an
individual in a positive or negative way. This definition encompassed such
behaviours as physical activity, safer sex, drug use and screening, etc. Studies
measuring help-seeking behaviour (e.g., Hunter, Grunfeld, & Ramirez, 2003),
physician behaviour (e.g., Millstein, 1996), those measuring dieting or
weight control behaviour amongst normal samples (e.g., Nejad, Wertheim,
& Greenwood, 2004; because dieting is not necessarily a healthy behaviour for
people of normal weight to engage in), and those involving professional
athletes (e.g., Palmer, Burwitz, Dyer, & Spray, 2005) were excluded.
3. Studies had to explicitly test the TPB, and measure all components (intention,
attitude, SN and PBC) to be eligible. Studies measuring the first three variables
but only including a general control construct (e.g., locus of control) were
excluded (e.g., Terry, Galligan, & Conway, 1993).
4. Studies were required to report all items used to measure TPB constructs or
provide at least enough information for them to be reliably coded. Measures
not matching definitions and studies which could not provide this information
were excluded (e.g., Bozionelos & Bennett, 1999).
5. Studies needed to report at minimum, bivariate correlations between intention
and behaviour, and PBC and behaviour, in line with the model tenets.
6. Papers from meeting abstracts or unpublished research were not included.
2
Of the papers identified, 5051 were excluded on the basis of their title or abstract.
The remaining 749 papers were obtained (two papers were requested via an inter-
library loan system but did not arrive in time to be included in the review), and 416
excluded on the basis of content within the paper. A further 44 papers were excluded
as they reported analyses based on the same sample as another included paper.
Searches identified a total of 287 independent papers reporting 320 tests of the
model. Where the required information was not available authors were contacted to
provide the required information. A total of 206 papers (237 tests) provided the
necessary information to be included in the analysis. Eighty-nine of these tests
provided information on at least one correlation with past behaviour.
Health Psychology Review 7
Downloaded By: [Conner, Mark T] At: 17:41 15 February 2011
Coding
As noted in the introduction, the grouping of health behaviours was based on
existing classifications that propose both a hierarchical (Roysamb et al., 1997) and
functional approach (Rothman & Salovey, 1997) to classification. A second
consideration that determined the grouping of behaviours was the need to ensure
adequate numbers in each group for the behaviour-type moderator analysis.
Behaviours were divided into those that are generally carried out to promote health
(health promoting) and those that risk health (health risk). The health-promoting
behaviours were then divided into three groups according to Rothman and Salovey
(1997): preventive, detection and curative behaviours (only one paper fitted into the
latter group and so this group was not included in analysis). The numbers of tests
within the preventive group (80.6% of all studies) allowed us to further sub-divide
this category into four descriptive categories: abstaining from/quitting drugs (e.g.,
reducing binge drinking and quitting smoking); physical activity, safer sex and
dietary behaviours. The health-risk behaviour category included studies that focused
on engaging in risk-taking behaviours (e.g., speeding, drinking alcohol, smoking,
using drugs). However, the relatively modest number of studies in this category
(k29) did not allow us to further sub-divide this grouping.
Of the 237 valid tests, 207 were successfully coded into these groups
3
(risk,
k29; detection, k17; physical activity, k103; dietary, k 30; safer sex, k 15;
and abstinence, k13 this latter category contained 12 tests related to abstinence
from smoking, and one test related to abstinence from binge drinking). Types of
behaviours which were not coded were those classed as preventive but not fitting into
any of the above categories and were diverse (e.g., breast feeding, general health
protection). These studies (k37) were excluded from all analyses by behaviour type.
Of the included papers, four reported tests for physical activity and dietary
behaviours, and one reported tests for physical activity and binge drinking (risk
behaviour). For these papers the relevant tests were included in both categories. For
this and the following analyses there were insufficient studies to populate groups for
belief-based measures as the vast majority of studies used direct measures of TPB
constructs. Belief-based and direct measures of attitudes, SN and PBC were thus
analysed in the same group. Where a study reported both direct and belief-based
measures, the direct measure was used, in line with theory tenets.
For the methodological moderator analyses, length of follow-up and type of
behavioural measure were split into dichotomous groups. Length of follow-up was
split at the median (5 weeks) into a group with shorter follow-up periods (55 weeks,
k126) and those with longer follow-up periods (5 weeks, k111). Type of
behavioural measure was split into self-report (k204) or objective measurement
(k33, four studies provided both an objective and a self-report measure so were
used twice). To explore age of sample, studies were split into three groups according
to whether they were adolescent or school-age samples (e.g., those recruited directly
from schools or youth clubs and who were 17 years or under, k51), student samples
(e.g., undergraduate student samples recruited from university settings, k70) or
adult samples (e.g., excluding latter two groups and recruited from community
setting, k108). Six tests were based on samples that combined two or more of these
groups (e.g., a mixed adolescent and student sample); these were excluded from
analyses by age of sample. Studies were originally coded into these groups by the first
8 R.R.C. McEachan et al.
Downloaded By: [Conner, Mark T] At: 17:41 15 February 2011
author, and blind-double coded by the fourth author. Agreement between the two
raters was 97.0%. Disagreements were discussed and resolved.
Analysis
The RE meta-analytic procedures of Hunter and Schmidt (2004) were used.
Correlations were corrected for both sampling and measurement error. A variety
of statistics are reported to aid interpretation of the meta-analysis. Within the tables,
the number of tests the meta-analysis is based on is denoted by the letter k. N refers
to the total sample size across all the included studies. To allow comparison with
previous meta-analyses, the average correlation corrected only for sampling error (r
c
)
is reported; however, this correlation is not interpreted in the text. Interpretation of
results uses the mean true score correlation corrected for sampling and measurement
error, which is denoted by mean r. The standard deviation of the correlation
corrected for measurement and sampling error (SD r) is presented along with
80% credibility intervals (an indication of the values 80% of studies fall between,
calculated using the corrected standard deviation). Finally, the percentage variance
of variation amongst correlations attributable to statistical artefacts (e.g., sampling
and measurement error) is reported. As a rule of thumb, Hunter and Schmidt (1990)
state that if 75% or more of the variance in study correlations is due to statistical
artefacts then the studies can be considered homogeneous.
In addition to the above statistics, further results are reported when examining
moderator variables. First, the RE standard error of the mean true score correlation
is reported (SE mean r), which is calculated using the following formula:
SE mean q ¼ SDr
c
=
ffiffi
k
p
(Hunter & Schmidt, 2004, p. 206).
The SE mean r is then used to construct 95% confidence intervals around the
mean r. Confidence intervals are used to aid in the interpretation of the significance
of the difference between two groups. If confidence intervals do not overlap (or only
partially overlap) then one can be more confident that groups are significantly
different to one another. These are a conservative test of significance as they are
constructed using the standard error of the number of studies (k) rather than
the total sample size (N). Within the sample, significant differences (where
confidence intervals do not overlap) are indicated either by letters (capital if
confidence intervals do not overlap, or lowercase if intervals overlap by less than
0.2 points) in the differences column (where there are three or more groups
compared) or by an asterisk (where two groups are compared). Due to the low
sample sizes involved in these calculations, intervals with an overlap of 0.2 or less are
interpreted as if they are significant. In addition, if the percentage artefact variance
explained is higher in the grouped vs. the ungrouped data one can be more confident
a moderator is explaining significant variance across the studies.
As a large number of model relationships are reported, to simplify interpretation
and explore how TPB variables react in combination with one another we performed
a series of multiple regressions to predict behaviour and intention, using correlation
matrices comprising the mean r. For the purposes of the analysis, the harmonic
Health Psychology Review 9
Downloaded By: [Conner, Mark T] At: 17:41 15 February 2011
mean N was used to specify sample size. The percentage variance explained is
reported (R
2
) and the relative contribution of each variable to the final equation are
reported by way of beta weights (B). Significance levels are not reported as they are
not meaningful due to the high numbers of participants the regressions are based on.
Calculations for the meta-analysis were conducted using Microsoft Office Excel
(2003). Regressions were conducting using SPSS (v14).
Hierarchical moderator analysis
A general test of the model is presented first. Studies are then grouped according to
behaviour type, then finally sample and methodological moderators explored. Un-
fortunately there were not enough studies to populate a full hierarchical analysis (e.g.,
behaviour typeage of samplelength of follow-uptype of behavioural measure),
therefore each sample and methodological variable was assessed hierarchically together
with behaviour type. Groups with three studies or less were excluded, as were groups
where substantial confounding of moderators occurred (described in more detail below).
Results
Full details of included studies are provided in the supplementary information
(Supplementary Table 1). Included studies are marked by an asterisk within the
References section.
General test of the model
In order to allow comparison with previous meta-analyses, a general test of the
model is presented first. The magnitude of the mean correlations corrected for both
sampling and measurement error (mean r), the standard deviation (SD r), total
number of studies (k) and the percent variation accounted for by statistical artefacts
are presented in Table 1. A full table including total sample size, mean correlation
corrected only for sampling error and 80% credibility intervals can be found in the
supplemental information (Supplementary Table 2). For this analysis we separated
tests between direct constructs (e.g., attitude) from those with belief-based constructs
(e.g., behavioural beliefs or behavioural beliefsoutcome evaluations).
In line with TPB tenets, intention showed the strongest relationship with
prospective behaviour (mean r0.43). This represents a mediumlarge effect
according to Cohens (1992) classification of effect sizes. Direct measures of attitudes
and PBC also showed medium-sized relationships with behaviour (both mean
r0.31). In relation to predicting intention, direct attitude exhibited the strongest
correlation with a mean r of 0.57 (indirect mean r0.42), followed by PBC (direct
construct mean r0.54 and indirect mean r0.44) and finally SN (direct mean
r0.40 and indirect mean r0.37). Past behaviour exhibited mediumlarge
correlations with behaviour (mean r0.50), intention (mean r0.47) and medium
correlations with attitude (mean r0.32) and PBC (mean r0.33), and
a smallmedium relationship with SN (mean r0.22). Table 1 also indicates that
the relationship between direct and indirect predictors was moderate at best (mean
r0.330.53). In general, these values are smaller than those reported in previous
general meta-analyses of the TPB (e.g., Armitage & Conner, 2001).
10 R.R.C. McEachan et al.
Downloaded By: [Conner, Mark T] At: 17:41 15 February 2011
The mean true score correlations were used to construct a correlation matrix and
were subjected to regression analyses to determine the unique contribution of each
variable to the prediction of behaviour and intention, controlling for past behaviour.
In line with theory tenets, only direct variables were used in this analysis. These were
entered together in the first step, with past behaviour entered in the second step.
Table 1. General test of model relationships. Mean r, standard deviation (in brackets),
number of studies (k) and percentage variance accounted for by statistical artefacts.
INT ATT
BB/
BBOE SN
NB/
NBMC PBC
CB/
CBP PastBR
a
BR 0.43
(0.19)
k 237
6.77%
0.31
(0.16)
k 209
11.52%
0.23
(0.13)
k 54
13.93%
0.21
(0.16)
k 196
12.97%
0.22
(0.16)
k 66
11.07%
0.31
(0.18)
k 219
9.24%
0.27
(0.13)
k 42
11.38%
0.50
(0.21)
k86
5.29%
INT 0.57
(0.18)
k 212
6.59%
0.42
(0.15)
k 55
7.86%
0.40
(0.21)
k 199
6.26%
0.37
(0.15)
k 68
9.30%
0.54
(0.22)
k 217
4.58%
0.44
(0.17)
k 44
5.73%
0.47
(0.23)
k87
4.21%
ATT 0.43
(0.20)
k 35
6.79%
0.44
(0.17)
k 191
8.52%
0.36
(0.19)
k 52
7.46%
0.45
(0.21)
k 198
6.13%
0.47
(0.20)
k 37
4.92%
0.32
(0.20)
k86
7.41%
BB/BBOE 0.25
(0.17)
k 39
9.41%
0.29
(0.15)
k 50
9.80%
0.29
(0.17)
k 44
8.30%
0.30
(0.13)
k 34
10.92%
SN 0.53
(0.21)
k 35
4.62%
0.29
(0.22)
k 186
6.40%
0.27
(0.21)
k 35
5.10%
0.22
(0.22)
k85
6.56%
NB/NBMC 0.25
(0.16)
k 54
9.95%
0.23
(0.16)
k 37
8.28%
PBC 0.41
(0.19)
k 27
5.15%
0.33
(0.33)
k 86
5.37%
CB/CBP
a
There were few studies exploring relations between indirect measures and past behaviour (BB/BBOE
PastBR: 7; NB/NB MCPastBr: 12; CB/CBPPastBR: 3), therefore for analyses with past behaviour
belief-based and direct constructs were combined.
Note: BR, behaviour; INT, intention; ATT, direct attitude; BB/BBOE, indirect attitude either
behavioural beliefs or behavioural beliefs outcome evaluation; SN, subjective norm; NB/NB MC,
indirect subjective norm either normative beliefs or normative beliefs motivation to comply; PBC,
perceived behavioural control; CB/CB P, indirect perceived behavioural control either control beliefs
or control beliefspower; PastBR, past behaviour.
Health Psychology Review 11
Downloaded By: [Conner, Mark T] At: 17:41 15 February 2011
Together, behavioural intention and PBC accounted for 19.3% of the variance in
behaviour. Intention was the main predictor of behaviour contributing three times
more to the final equation (B0.37) than PBC (B0.11). The percentage variance
explained in the prediction of behaviour was smaller than that of previous reviews
(2634%: Armitage & Conner, 2001; Hagger et al., 2002; Rivis & Sheeran, 2003a;
Schulze & Whittmann, 2003; Sheeran & Taylor, 1999; Trafimow et al., 2002). In the
present data, the addition of past behaviour as a predictor of behaviour at Step 2 of
the equation added an additional 10.9% variance (B0.38) and greatly attenuated
the effects of both intention (B 0.22) and PBC (B0.07), although both remained
significant.
In relation to predictions of intention, attitude, SN and PBC together accounted
for 44.3% of the variance. Attitude was the strongest predictor (B0.35), followed
by PBC (B0.34) and SN (B0.15). The percentage variance explained in the
prediction of intention was similar to that of previous reviews (Armitage & Conner,
2001; Hagger et al., 2002; Rivis & Sheeran, 2003a; Schulze & Whittmann, 2003;
Sheeran & Taylor, 1999; Trafimow et al., 2002). The addition of past behaviour at
Step 2 added an additional 5% variance and was a significant predictor (B0.25).
Inclusion of past behaviour slightly attenuated the impacts of attitude (B0.31), SN
(B0.13) and PBC (B0.28), although attitude remained the strongest predictor of
intention and all predictors remained significant.
It is worth noting that all the overall mean correlations reported in Table 1 were
subject to substantial variability. The mean percentage artefact variation within the
correlations explained by sampling and measurement error was 7.81%, with values
ranging from 4.21% (past behaviourintention) to 13.93% (indirect attitudeinten-
intention). These values do not approach the 75% rule of thumb proposed by Hunter
and Schmidt (1990, 2004) for homogeneity of study correlations. Therefore a search
for additional moderators is warranted.
Behaviour type as moderator
The impact of behaviour type as a moderator was explored next. Table 2 contains
relationships between TPB variables and behaviour or intention. A full list of model
relationships can be found in Supplementary Table 3. Examination of the mean rs
for each relationship identifies a number of differences, and the final column in Table 2
highlights which relationships are significantly different across behaviours. Modera-
tion by behaviour type is clearest for the PBCBR and SNINT relationships. For
example, PBC is a significantly stronger predictor of physical activity (mean
r0.34) and dietary (mean r0.35) behaviours compared with detection and
safer sex behaviours (both mean r0.22). The magnitude of the SNINT
relationship is significantly stronger for safer sex (mean r0.56) compared with
detection, physical activity and abstinence (mean rs 0.340.38) behaviours. In
contrast, the correlations between SN and behaviour are strongest for risk
behaviours. Other substantial differences are also apparent. Intentions appear to
be relatively more important in the prediction of physical activity (mean r0.48)
compared with safer sex and abstinence (both mean r0.37) behaviours. There were
sufficient studies to populate comparisons of relations with past behaviour for risk
(k11), physical activity (k44), dietary (k6) and abstinence (k9) behaviours.
The smaller number of studies leads to lower power to detect differences between
12 R.R.C. McEachan et al.
Downloaded By: [Conner, Mark T] At: 17:41 15 February 2011
Table 2. Relationships between TPB variables with prospective behaviour and intention: split by behaviour type.
Test N k r
c
Mean
r
SD
r
SE
r
10%
CV
90%
CV CI low CI high
Variance
(%) Diffs
INTBR A. Risk 13,710 29 0.37 0.38 0.24 0.04 0.08 0.69 0.30 0.47 2.82
B. Detection 8370 17 0.37 0.38 0.18 0.04 0.15 0.61 0.29 0.47 4.62
C. Physical
activity
23,376 103 0.45 0.48 0.18 0.02 0.25 0.70 0.44 0.51 8.57 E, f
D. Dietary 9047 30 0.38 0.44 0.16 0.03 0.23 0.64 0.38 0.50 9.81
E. Safe sex 2605 15 0.34 0.37 0.11 0.03 0.22 0.51 0.30 0.43 27.45 C
F. Abstinence 4406 13 0.35 0.37 0.12 0.04 0.21 0.52 0.29 0.44 13.83 c
PBCBR A. Risk 13,713 29 0.22 0.24 0.20 0.04 0.01 0.50 0.17 0.32 4.88 c
B. Detection 8370 17 0.20 0.22 0.10 0.03 0.09 0.35 0.17 0.27 15.87 C, D
C. Physical
activity
23,385 103 0.31 0.34 0.15 0.02 0.15 0.54 0.31 0.38 14.29 a, B, E
D. Dietary 9047 30 0.30 0.35 0.15 0.03 0.16 0.55 0.29 0.41 11.69 B, e
E. Safe sex 3674 19 0.21 0.22 0.13 0.03 0.06 0.39 0.16 0.29 23.24 C, d
F. Abstinence 4406 13 0.26 0.28 0.17 0.05 0.05 0.50 0.18 0.38 7.95
ATTBR A. Risk 13,713 29 0.27 0.29 0.16 0.03 0.08 0.49 0.23 0.35 7.01
B. Detection 8370 17 0.22 0.25 0.21 0.05 0.02 0.51 0.15 0.35 4.63
C. Physical
activity
23,141 101 0.30 0.34 0.15 0.02 0.14 0.53 0.30 0.37 15.81 E
D. Dietary 9046 30 0.29 0.33 0.18 0.03 0.11 0.56 0.26 0.40 9.26
E. Safe sex 2234 14 0.23 0.26 0.07 0.03 0.16 0.35 0.20 0.31 53.92 C
F. Abstinence 4406 13 0.26 0.29 0.15 0.04 0.10 0.48 0.20 0.37 11.55
SNBR A. Risk 13,189 29 0.26 0.29 0.18 0.04 0.05 0.53 0.22 0.36 5.22 c, d
B. Detection 8370 17 0.19 0.20 0.17 0.04 0.02 0.43 0.12 0.29 5.83
C. Physical
activity
22,849 100 0.18 0.21 0.14 0.02 0.03 0.39 0.18 0.24 19.33 a
D. Dietary 9049 30 0.15 0.17 0.15 0.03 0.02 0.37 0.11 0.23 14.35 a
Health Psychology Review 13
Downloaded By: [Conner, Mark T] At: 17:41 15 February 2011
Table 2 (Continued )
Test N k r
c
Mean
r
SD
r
SE
r
10%
CV
90%
CV CI low CI high
Variance
(%) Diffs
E. Safe sex 2235 14 0.21 0.23 0.15 0.05 0.03 0.43 0.14 0.32 21.97
F. Abstinence 4406 13 0.21 0.21 0.18 0.05 0.02 0.44 0.11 0.32 8.05
PBBR A. Risk 2637 11 0.60 0.62 0.20 0.06 0.37 0.88 0.50 0.74 4.31 D
C. Physical
activity
10,014 44 0.53 0.54 0.19 0.03 0.30 0.79 0.49 0.60 6.23 D
D. Dietary 442 6 0.34 0.36 0
a
0.04 0.36 0.36 0.28 0.44 100.00 A, C
F. Abstinence 2032 9 0.38 0.42 0.25 0.09 0.09 0.74 0.25 0.59 5.13
ATT INT A. Risk 14,673 29 0.46 0.52 0.21 0.04 0.25 0.78 0.44 0.59 3.03 c, d
B. Detection 8370 17 0.45 0.52 0.20 0.05 0.26 0.77 0.42 0.61 3.59
C. Physical
activity
23,905 101 0.51 0.60 0.14 0.02 0.42 0.79 0.57 0.63 11.92 a
D. Dietary 9823 30 0.52 0.62 0.14 0.03 0.45 0.80 0.57 0.67 9.12 a
E. Safer sex 4958 15 0.51 0.59 0.11 0.03 0.44 0.73 0.53 0.65 12.59
F. Abstinence 6351 13 0.47 0.52 0.16 0.05 0.31 0.72 0.43 0.61 5.04
SNINT A. Risk 14,673 29 0.40 0.45 0.20 0.04 0.19 0.71 0.37 0.52 3.52
B. Detection 8370 17 0.33 0.35 0.21 0.05 0.08 0.63 0.25 0.46 3.52 E
C. Physical
activity
23,499 100 0.32 0.38 0.20 0.02 0.12 0.64 0.34 0.42 9.01 E
D. Dietary 9823 30 0.35 0.43 0.14 0.03 0.25 0.61 0.38 0.49 12.54
E. Safer sex 4958 15 0.45 0.56 0.17 0.05 0.34 0.78 0.47 0.65 8.16 B, C, F
F. Abstinence 6351 13 0.33 0.34 0.15 0.04 0.14 0.53 0.25 0.42 6.63 E
PBCINT A. Risk 14,673 29 0.43 0.49 0.22 0.04 0.21 0.77 0.41 0.57 2.94
B. Detection 8370 17 0.45 0.51 0.17 0.04 0.30 0.72 0.43 0.59 4.85
C. Physical
activity
23,996 102 0.47 0.55 0.20 0.02 0.29 0.80 0.51 0.59 6.98
D. Dietary 9823 30 0.44 0.53 0.18 0.03 0.31 0.75 0.46 0.59 7.11
14 R.R.C. McEachan et al.
Downloaded By: [Conner, Mark T] At: 17:41 15 February 2011
Table 2 (Continued )
Test N k r
c
Mean
r
SD
r
SE
r
10%
CV
90%
CV CI low CI high
Variance
(%) Diffs
E. Safer sex 4958 15 0.44 0.49 0.25 0.07 0.17 0.81 0.36 0.62 3.37
F. Abstinence 6351 13 0.43 0.45 0.20 0.06 0.20 0.70 0.34 0.56 3.36
PBINT A. Risk 2905 11 0.58 0.63 0.17 0.05 0.41 0.85 0.52 0.73 5.92 F
C. Physical
activity
10,332 43 0.53 0.56 0.17 0.03 0.35 0.78 0.51 0.61 7.51 F
D. Dietary 514 5 0.34 0.39 0.24 0.11 0.09 0.70 0.17 0.62 13.94
F. Abstinence 5586 10 0.23 0.23 0.14 0.05 0.05 0.41 0.14 0.32 7.62 A, C
a
Calculations returned a negative standard deviation, this is interpreted as 0 hence no variation around the mean r (cf. Hunter & Schmidt, 2004).
Note: BR, behaviour; INT, intention; PBC, perceived behavioural control; ATT, attitude; SN, subjective norm; PB, past behaviour; r
c
, average correlation corrected for
sampling error; mean r, true score correlation (corrected for sampling and measurement error); SD r, standard deviation of true score correlation; SE mean r, standard
error of mean true score correlation; 10% CV, 10% credibility interval; 90% CV, 90% credibility interval; CI low, lower 95% confidence interval; CI high, upper 95%
confidence interval; Variance, percentage variance in true score correlation attributable to statistical artefacts; Diffs, letters in capital indicate groups for which confidence
intervals are not overlapping, and letters in lowercase indicate confidence intervals that overlap by less than 0.02 points.
Health Psychology Review 15
Downloaded By: [Conner, Mark T] At: 17:41 15 February 2011
behaviour types. Nevertheless, the magnitude of the PBBR relationship was
significantly weaker for dietary (mean r0.36) compared with risk and physical
activity (mean r0.62 and 0.54, respectively) behaviours. With regard to intention,
past behaviour was a significantly weaker predictor for abstinence (mean r0.23)
compared with risk (mean r0.63) and physical activity (mean r0.56)
behaviours.
In order to further analyse these findings and ascertain which behaviours are
most efficiently predicted by the TPB as a whole, a series of multiple regressions were
computed for each behaviour type. TPB variables were added in the first step, and for
risk, physical activity, dietary and abstinence from drug behaviours, past behaviour
was added at the second step. These results are reported in Table 3 (top panel).
Looking first at the analyses without past behaviour, it can be seen from the top
panel of Table 3 that the TPB is most effective at predicting physical activity and
dietary behaviours (respectively, 23.9% and 21.2% variance predicted in behaviour)
and least effective at predicting risk, detection, safer sex and abstinence behaviours
(13.815.3% explained variance). Intention is the strongest predictor of all behaviour
types and is particularly important for physical activity behaviours (B0.42). PBC
appears to have a negligible impact on the prediction of detection (B 0.04) and
safer sex behaviours (B0.05) but has a more substantial impact for other
behaviours. It was possible to add past behaviour to the regression equations to
predict behaviour for risk, physical activity, diet and abstinence behaviours. This
variable added 25.3% variance to the prediction of risk behaviours, 11.4% variance to
abstinence behaviours, 10.3% to physical activity behaviours, but only 3.4% to
dietary behaviours. When added past behaviour becomes the only significant variable
predictive of engaging in risk behaviours. Importantly, although past behaviour is the
most important predictor for physical activity and dietary behaviours, intentions
remain significant predictors (with intentions remaining the strongest predictor for
dietary behaviours).
The pattern of results was not the same for intention (see Table 4, top panel). In
this case, dietary and safer sex intentions are relatively better predicted (50.3% and
51.3% variation explained, respectively) when compared to other behaviours.
Abstinence intentions show the lowest level of prediction by the model (36.6%
variation explained). Attitudes are the strongest predictor of intentions for all
behaviours except detection behaviours (where PBC is most important) and risk
behaviours (where attitude and PBC make similar contributions). SNs are most
predictive of safer sex intentions (where they make a greater contribution than PBC;
B0.32 vs. 0.26) and are relatively more important for risk (B0.23) and dietary
behaviours (B0.23). Past behaviour added 8.2% to the prediction of risk and
physical activity intentions (and was the strongest individual predictor), but only
around 2% to the prediction of diet and abstinence behaviours (2.1% and 1.6%,
respectively).
In order to examine whether behaviour type was a valid moderator in the meta-
analysis, the percentage variance in statistical artefacts accounted for was examined
(Table 2). If the moderator is valid, one would expect a greater amount of variation
to be explained by statistical artefacts after accounting for the moderator, as one is,
in principle, removing the variation caused by the moderator. The mean percentage
artefact variance accounted for across all reported model relationships (excluding
past behaviour correlations) was 17.1% for safer sex, 12.1% for physical activity,
16 R.R.C. McEachan et al.
Downloaded By: [Conner, Mark T] At: 17:41 15 February 2011
Table 3. Multivariate prediction of behaviour: split by all moderators. Past behaviour included where available.
Step 1 Step 2 (addition of past behaviour)
Group B INT B PBC
Total
R
2
B INT B PBC
B
PastBR
R
2
change Total R
2
Behaviour A. Risk 0.345 0.071 0.148 0.019 0.160 0.698 0.253 0.401
B. Detection 0.362 0.035 0.145
C. Physical activity 0.420 0.109 0.239 0.222 0.074 0.388 0.103 0.340
D. Dietary 0.354 0.162 0.212 0.293 0.128 0.204 0.034 0.247
E. Safer sex 0.345 0.051 0.138
F. Abstinence 0.306 0.142 0.153 0.235 0.122 0.348 0.114 0.267
Behaviour age of sample A. Adult: physical activity 0.390 0.113 0.212 0.239 0.056 0.403 0.127 0.339
B. Students: physical activity 0.481 0.097 0.297
C. Adolescent: physical
activity
0.400 0.120 0.222 0.402 0.120 0.004 0.057 0.222
D. Adult: dietary 0.265 0.314 0.267
E. Adolescent: dietary 0.270 0.085 0.096
F. Students: abstinence 0.381 0.055 0.155
G. Adolescent: abstinence 0.177 0.315 0.209
Behaviour length of follow-up A. Long : detection 0.260 0.069 0.093
B. Short : detection 0.550 0.027 0.314
C. Long : physical activity 0.319 0.127 0.157 0.179 0.066 0.377 0.111 0.267
D. Short : physical activity 0.500 0.100 0.320 0.252 0.096 0.391 0.090 0.410
E. Long : dietary 0.219 0.265 0.184
F. Short : dietary 0.387 0.142 0.227
Behaviour by type of
measurement
A. Self-report: physical
activity
0.439 0.105 0.257
B. Objective: physical activity 0.315 0.079 0.121
Note: INT, intention; PBC, perceived behavioural control; ATT, attitude; SN, subjective norm; PBC, past behaviour; Long, longer follow-up period; Short, shorter
follow-up period.
Health Psychology Review 17
Downloaded By: [Conner, Mark T] At: 17:41 15 February 2011
Table 4. Multivariate prediction of intention: split by behaviour and age of sample.
Step 1 Step 2 (inclusion of past behaviour)
Group B ATT B SN B PBC Total R
2
B ATT B SN B PBC B PastBR R
2
change Total R
2
Behaviour A. Risk 0.298 0.227 0.297 0.403 0.205 0.147 0.148 0.380 0.082 0.484
B. Detection 0.337 0.126 0.345 0.395
C. Physical activity 0.399 0.119 0.325 0.463 0.316 0.102 0.251 0.320 0.082 0.545
D. Dietary 0.392 0.230 0.304 0.503 0.369 0.217 0.266 0.156 0.021 0.524
E. Safer sex 0.355 0.315 0.258 0.513
F. Abstinence 0.358 0.119 0.309 0.366 0.336 0.130 0.295 0.131 0.016 0.382
Behaviourage of
sample
A. Adult: physical activity 0.366 0.137 0.329 0.437 0.320 0.119 0.282 0.214 0.038 0.474
B. Students: physical activity 0.396 0.099 0.383 0.521 0.272
C. Adolescent: physical
activity
0.488 0.127 0.203 0.463 0.394 0.098 0.156 0.332 0.091 0.562
D. Adult: dietary 0.401 0.176 0.292 0.496
E. Adolescent: dietary 0.329 0.385 0.200 0.445
F. Students: abstinence 0.552 0.147 0.223 0.436
G. Adolescent: abstinence 0.258 0.186 0.450 0.591
Note: INT, intention; PBC, perceived behavioural control; ATT, attitude; SN, subjective norm; PBC, perceived behavioural control; PB, past behaviour.
18 R.R.C. McEachan et al.
Downloaded By: [Conner, Mark T] At: 17:41 15 February 2011
10.2% for dietary, 7.1% for abstinence, 5.6% for detection and 4.1% for risk
behaviours. This indicates that behaviour type is a valid moderator, but that the
groupings for risk and detection behaviours are less optimal. This might be expected
considering the former three groupings contain single behavioural types, whilst the
latter two are more general behaviour domains. Despite the fact that behaviour type
appears to be a valid moderator of the TPB a substantial proportion of variation
remains indicating there may be other moderators at work.
Sample and methodological moderators
The next step in the analysis was to explore the impact of age of sample, length of
follow-up and type of behavioural measure. Where possible this was performed
hierarchically within behaviour type. Following Lipsey (2003) and Patall, Cooper,
and Robinson (2008) these moderators were first assessed for potential confounding.
Table 5 displays frequencies for each level of the behaviour, age of sample and length
of follow-up moderators (there were too few studies measuring objective behaviour
to warrant inclusion in this analysis). Due to small cell sizes, the FreemanHalton
extension to the Fisher exact test for 23 contingency tables (Freeman & Halton,
1951) was used to explore whether significant confounding was apparent (calcula-
tions were performed using the Vasser Stats website: http://faculty.vassar.edu/lowry/
VassarStats.html, accessed 26 June 2010, # R Lowry). Calculations were not
performed for detection behaviours as there was no variation on age of sample. It can
be seen from Table 5 that potential confounding was apparent for physical activity
(pB0.001) and risk behaviours (pB0.0001). In the former case, it can be seen that
adolescent and student samples tend to report shorter follow-ups compared with
adult samples. In the latter, adolescent samples tend to report longer follow-ups and
Table 5. Exploring of moderator confounding: number of tests as each moderator level by
behaviour.
Behaviour Adolescent Student Adult Total Fishers exact
Risk Shorter 1 12 0 29 p B0.0001
Longer 11 3 2
Detection Shorter 0 0 6 17 n/a
Longer 0 0 11
Physical activity
a
Shorter 19 30 16 100 p B0.0001
Longer 6 5 24
Dietary
b
Shorter 2 12 11 29 p0.17 n/s
Longer 1 0 3
Safer sex
b
Shorter 0 1 0 14 p0.16 n/s
Longer 3 1 8
Abstinence Shorter 1 1 0 13 p0.68 n/s
Longer 4 2 5
a
Three studies not reported in this table: two contained a mixed samples(e.g., mixture of students and
adolescents), one could not be coded according to sample type.
b
Two tests missing as they reported a mixed sample.
Health Psychology Review 19
Downloaded By: [Conner, Mark T] At: 17:41 15 February 2011
student samples tend to report shorter follow-ups. This indicates that caution must
be exercised when drawing conclusions about the impact of each moderator as it is
difficult to tease apart their independent influences.
For the purposes of the current analysis, it was not deemed appropriate to split
studies by these two moderators simultaneously due to the resultant small cell sizes.
Instead, we separately split behaviour type with first, age of sample and second,
length of follow-up. In order to minimise problems of confounding, we did not
analyse groupings that showed high levels of confound (i.e., where 90% or more of
studies were in one group for the moderator not directly being assessed). For
example, taking the behaviourage of sample analysis, the adolescent risk group
was excluded as 11 out of 12 studies reported longer follow-ups. Within the
behaviourlength of follow-up analysis we excluded the shorter follow-up risk
group as within this group 12 out of 13 studies were conducted with student samples.
For clarity we have indicated which groups have been excluded and why in the
appropriate footnotes.
Age of sample
For the age of sample moderator there were sufficient studies to compare results for
physical activity (adult k40, students k35 and adolescent k25), dietary (adult
k14 and adolescent k3) and abstinence (students k3 and adolescent k5)
4
behaviours. For relations with past behaviour it was only possible to explore age of
sample within physical activity studies for adult (k12) and adolescent (k16)
groups. Table 6 shows the individual model relationships. For the sake of brevity,
only relationships with behaviour and intention are displayed. The full data can be
found in the supplemental information (Supplementary Table 4).
Comparisons between all three different age groups could only be made within
physical activity behaviours. It can be seen that although the magnitude of the mean
r does vary, none of these differences reach significance. Comparison across other
behaviours is difficult as dietary and abstinence behaviours do not report the same
groups. Taking dietary behaviours it can be seen that adult samples display stronger
relationships with behaviour for intention, PBC and attitude (mean r IN-
TBR0.45 vs. 0.30; PBCBR0.47 vs. 0.18; ATTBR 0.45 vs. 0.18, respectively)
compared with adolescent samples. However, the opposite trend is apparent for the
SN intention relationship. Here adolescent samples exhibit stronger relationships
compared with adult samples (mean r0.53 vs. 0.36). With regard to abstinence
behaviours, the PBC behaviour relationship is significantly stronger for adolescent
samples compared with student samples (mean r0.44 vs. 0.12). The SNbehaviour
relationship is stronger for adolescent samples compared with older samples for both
physical activity (mean r0.27 adolescent vs. 0.17 adult) and dietary behaviours
(mean r0.30 adolescent vs. 0.24 students), although significant only in the former.
Thus it may be indicative of a tendency for social norms to be more predictive of the
behaviour of adolescent samples.
Table 3 (second panel from top for both) and 4 (bottom panel) display the results
of the multivariate regressions to predict behaviour and intention. It can be seen that,
amongst physical activity behaviours, intentions and behaviour in student samples
appear to be better predicted (29.7% behaviour and 52.1% intention) than in either
adult or adolescent samples (21.222.2% behaviour and 43.749.6% intention).
20 R.R.C. McEachan et al.
Downloaded By: [Conner, Mark T] At: 17:41 15 February 2011
Table 6. Model relationships: behaviour typeage of sample: intention and behaviour.
Nkr
c
Mean r SD r SE r 10% CV 90% CV CI low CI high Variance (%) Diffs
INTBR A. Adult: PA 8761 40 0.42 0.45 0.18 0.03 0.21 0.68 0.39 0.51 8.93
B. Students: PA 7152 35 0.51 0.54 0.16 0.03 0.34 0.74 0.48 0.59 10.34
C. Adolescent: PA 6654 25 0.42 0.46 0.14 0.03 0.27 0.64 0.40 0.52 11.86
D. Adult: dietary 3333 14 0.43 0.45 0.13 0.04 0.28 0.62 0.37 0.52 14.16 E
E. Adolescent: dietary 3011 3 0.24 0.30 0.04 0.03 0.25 0.34 0.24 0.35 45.96 D
F. Students: abstain 908 3 0.33 0.39 0.06 0.05 0.32 0.47 0.30 0.49 49.57
G. Adolescent: abstain 1402 5 0.37 0.40 0.10 0.05 0.27 0.53 0.30 0.50 22.20
PBCBR A. Adult: PA 8761 40 0.31 0.32 0.18 0.03 0.10 0.55 0.27 0.38 10.95
B. Students: PA 7152 35 0.34 0.39 0.14 0.03 0.22 0.57 0.34 0.44 19.21
C. Adolescent: PA 6654 25 0.27 0.32 0.12 0.03 0.16 0.48 0.27 0.38 19.55
D. Adult: dietary 3333 14 0.42 0.47 0.13 0.04 0.31 0.64 0.40 0.55 15.69 E
E. Adolescent: dietary 3011 3 0.17 0.18 0.10 0.06 0.05 0.31 0.06 0.30 8.59 D
F. Students: abstain 908 3 0.10 0.12 0.14 0.09 0.06 0.31 0.05 0.30 16.91 G
G. Adolescent: abstain 1402 5 0.41 0.44 0.10 0.05 0.31 0.56 0.34 0.53 19.85 F
ATTBR A. Adult: PA 8524 38 0.26 0.30 0.16 0.03 0.09 0.51 0.24 0.35 14.28
B. Students: PA 7152 35 0.30 0.34 0.14 0.03 0.17 0.52 0.29 0.40 20.12
C. Adolescent: PA 6647 25 0.31 0.36 0.10 0.02 0.23 0.48 0.31 0.40 26.65
D. Adult: dietary 3333 14 0.40 0.45 0.10 0.03 0.32 0.58 0.39 0.51 23.52 E
E. Adolescent: dietary 3011 3 0.14 0.16 0.04 0.03 0.11 0.21 0.11 0.22 43.60 D
F. Students: abstain 908 3 0.26 0.31 0.05 0.05 0.24 0.38 0.22 0.40 54.52
G. Adolescent: abstain 1402 5 0.36 0.38 0.14 0.07 0.20 0.57 0.25 0.52 11.87
Health Psychology Review 21
Downloaded By: [Conner, Mark T] At: 17:41 15 February 2011
Table 6 (Continued )
Nkr
c
Mean r SD r SE r 10% CV 90% CV CI low CI high Variance (%) Diffs
SNBR A. Adult: PA 8159 37 0.15 0.17 0.12 0.02 0.01 0.32 0.12 0.21 25.17 c
B. Students: PA 7056 35 0.18 0.21 0.10 0.02 0.08 0.33 0.17 0.25 36.15
C. Adolescent: PA 6654 25 0.23 0.27 0.18 0.04 0.04 0.51 0.20 0.35 10.51 a
D. Adult: dietary 3333 14 0.15 0.17 0.16 0.05 0.03 0.37 0.08 0.26 16.01
E. Adolescent: dietary 3011 3 0.14 0.17 0.10 0.06 0.03 0.30 0.04 0.29 9.51
F. Students: abstain 908 3 0.19 0.24 0
a
0.02 0.24 0.24 0.20 0.29 100.00
G. Adolescent: abstain 1402 5 0.28 0.30 0.26 0.12 0.03 0.63 0.07 0.53 4.95
PBBR A. Adult: PA 3026 12 0.53 0.53 0.14 0.04 0.35 0.72 0.45 0.62 9.22
C. Adolescent: PA 3800 16 0.45 0.47 0.21 0.05 0.21 0.74 0.37 0.58 6.35
ATTINT A. Adult: PA 8939 39 0.48 0.57 0.15 0.03 0.38 0.76 0.52 0.62 11.87
B. Students: PA 7604 35 0.52 0.63 0.16 0.03 0.43 0.83 0.58 0.69 10.78
C. Adolescent: PA 6544 24 0.54 0.64 0.11 0.02 0.50 0.77 0.59 0.68 16.80
D. Adult: dietary 3557 14 0.56 0.65 0.07 0.02 0.57 0.74 0.61 0.70 33.11
E. Adolescent: dietary 3491 3 0.44 0.52 0.14 0.08 0.35 0.70 0.37 0.68 3.35
F. Students: abstain 908 3 0.55 0.61 0.02 0.03 0.58 0.64 0.56 0.66 74.05
G. Adolescent: abstain 1523 5 0.57 0.65 0.16 0.07 0.45 0.85 0.51 0.79 6.46
SNINT A. Adult: PA 8533 38 0.34 0.39 0.18 0.03 0.16 0.62 0.33 0.45 11.60
B. Students: PA 7604 35 0.32 0.38 0.13 0.02 0.21 0.54 0.33 0.42 21.46
C. Adolescent: PA 6544 24 0.31 0.39 0.29 0.06 0.02 0.77 0.27 0.51 4.05
D. Adult: dietary 3557 14 0.31 0.36 0.13 0.04 0.19 0.53 0.28 0.44 17.86 e
E. Adolescent: dietary 3491 3 0.41 0.53 0.08 0.05 0.43 0.64 0.43 0.63 10.16 d
F. Students: abstain 908 3 0.36 0.42 0
a
0.01 0.42 0.42 0.39 0.45 100.00 g
G. Adolescent: abstain 1523 5 0.46 0.56 0.14 0.06 0.38 0.73 0.43 0.68 12.42 f
22 R.R.C. McEachan et al.
Downloaded By: [Conner, Mark T] At: 17:41 15 February 2011
Table 6 (Continued )
Nkr
c
Mean r SD r SE r 10% CV 90% CV CI low CI high Variance (%) Diffs
PBCINT A. Adult: PA 9030 40 0.46 0.53 0.20 0.03 0.26 0.79 0.46 0.59 6.67
B. Students: PA 7402 34 0.52 0.61 0.18 0.03 0.38 0.84 0.55 0.67 8.40
C. Adolescent: PA 6544 24 0.42 0.50 0.20 0.04 0.25 0.75 0.42 0.58 7.02
D. Adult: dietary 3557 14 0.53 0.59 0.14 0.04 0.42 0.77 0.52 0.67 11.19
E. Adolescent: dietary 3491 3 0.32 0.35 0.14 0.08 0.18 0.53 0.20 0.51 4.10
F. Students: abstain 908 3 0.16 0.17 0.06 0.05 0.09 0.25 0.08 0.27 48.26
G. Adolescent: abstain 1523 5 0.64 0.71 0.17 0.08 0.50 0.93 0.56 0.87 4.20
PBINT A. Adult: PA 2883 11 0.45 0.45 0.10 0.03 0.33 0.58 0.39 0.52 17.76
C. Adolescent: PA 3800 16 0.50 0.56 0.19 0.05 0.31 0.80 0.46 0.65 6.56
a
Calculations returned a negative standard deviation, this is interpreted as 0 hence no variation around the mean r (cf. Hunter & Schmidt, 2004).
Note: BR, behaviour; INT, intention; PBC, perceived behavioural control; ATT, attitude; SN, subjective norm; PB, past behaviour; r
c
, average correlation corrected for
sampling error; mean r, true score correlation (corrected for sampling and measurement error); SD r, standard deviation of true score correlation; SE mean r, standard
error of mean true score correlation; 10% CV, 10% credibility interval; 90% CV, 90% credibility interval; CI low, lower 95% confidence interval; CI high, upper 95%
confidence interval; Variance, percentage variance in true score correlation attributable to statistical artefacts; Diffs, letters in capital indicate groups for which
confidence intervals are not overlapping, and letters in lowercase indicate confidence intervals that overlap by less than 0.02 points, comparisons are only made between
individual behaviour types; PA, physical activity; abstain, abstinence from drug related behaviours.
Health Psychology Review 23
Downloaded By: [Conner, Mark T] At: 17:41 15 February 2011
Intentions are the strongest predictor of physical activity for all samples. The
addition of past behaviour is important in the prediction of behaviours for adult
samples (an additional 12.7% variance explained), and becomes the most important
predictor of behaviour, but does not explain any additional variance for adolescent
samples. For predicting intentions, attitudes are the strongest predictor, although
PBC appears to be relatively more important for adult and student samples.
Including past behaviour adds the most variance for adolescent samples (9.1%), but
attitudes remain the strongest predictor of intention.
The pattern is different for abstinence behaviours (Table 3, second panel from
top) with adolescent samples (20.9% behaviour and 59.1% intention) better predicted
than student samples (15.5% behaviour and 43.6% intention). It is also interesting to
note that for students intention is the strongest predictor of abstaining behaviours
(B0.38), whilst for adolescent samples PBC is the strongest predictor of abstaining
(B0.32). PBC is also particularly important with regard to the prediction of
abstinence intentions for adolescent samples, with the intention for student samples
being more under attitudinal control (see Table 4, second panel from top).
Finally, dietary behaviours (Table 3, second panel from top) amongst adolescent
samples appear to be very poorly predicted (9.6% variance) compared with adult
samples (26.7%). PBC is the most important predictor of adults dietary behaviours
(B0.31), whilst intentions are the most important predictor of adolescent
behaviours (B0.27). Although the percentage variance in intention accounted
for is similar across the two groups, SNs are the most important predictor of
adolescent intentions whilst attitudes are the most important predictor of adult
dietary intentions (see Table 4, second panel from top).
The mean percent artefact variation explained for each relationship generally
increased with the inclusion of the age of sample moderator indicating its validity
when used in conjunction with behaviour type. The most artefact variation was
explained for abstinence behaviours amongst student samples (59.8%), and for the
remaining groups ranged between 9.9 and 18.3%. In fact, for the SNbehaviour
relationship 100% of artefact variation was explained by sampling and measurement
error (although caution should be exercised here due to the low number of studies
within this group, k3). These findings point to the validity of age of sample as
a moderator of TPB model relations.
5
Length of follow-up
Upon inspection of levels of the different moderators the following groups could be
explored with reference to length of follow-up: physical activity behaviours (longer
k36 and shorter k67), dietary behaviours (longer k4 and shorter k26).
Detection behaviours were all conducted with adult samples, and analyses are
conducted with 11 longer follow-up tests and six shorter follow-up tests.
6
For
correlations with past behaviour, comparisons could only be made for physical
activity behaviours (longer k8 and shorter k30). For this moderator only
analyses with behaviour are reported as length of follow-up is not a relevant variable
for cross-sectionally measured relationships between intention, attitude, PBC and
SN.
The results of this analysis can be found in Table 7. It can be clearly seen that
length of follow-up moderates relations with a number of variables. For example,
24 R.R.C. McEachan et al.
Downloaded By: [Conner, Mark T] At: 17:41 15 February 2011
Table 7. Model relationships: behaviour typelength of follow-up.
Nkr
c
Mean r SD r SE r 10% CV 90% CV CI low CI high Variance (%) Diffs
INTBR A. Longer : detection 5884 11 0.29 0.30 0.09 0.03 0.18 0.41 0.24 0.36 16.81 B
B. Shorter : detection 2486 6 0.55 0.56 0.21 0.09 0.30 0.83 0.40 0.73 2.63 A
C. Longer : physical activity 9971 36 0.35 0.38 0.11 0.02 0.24 0.52 0.34 0.42 19.00 D
D. Shorter : physical activity 13,404 67 0.52 0.56 0.17 0.02 0.35 0.77 0.52 0.60 9.38 C
E. Longer : dietary 2391 4 0.30 0.37 0.17 0.09 0.15 0.59 0.20 0.55 5.84
F. Shorter : dietary 6656 26 0.41 0.46 0.15 0.03 0.26 0.65 0.39 0.52 11.34
PBCBR A. Longer : detection 5884 11 0.20 0.22 0.12 0.04 0.07 0.37 0.14 0.29 11.46
B. Shorter : detection 2486 6 0.20 0.22 0.07 0.03 0.13 0.31 0.15 0.29 35.19
C. Longer : physical activity 9971 36 0.26 0.28 0.13 0.02 0.11 0.44 0.23 0.32 16.01 D
D. Shorter : physical activity 13,414 67 0.35 0.40 0.15 0.02 0.21 0.59 0.36 0.44 16.69 C
E. Longer : dietary 2391 4 0.34 0.39 0.15 0.08 0.19 0.58 0.23 0.54 5.90
F. Shorter : dietary 6656 26 0.29 0.34 0.15 0.03 0.14 0.53 0.28 0.40 13.97
ATTBR A. Longer : detection 5884 11 0.16 0.17 0.13 0.04 0.00 0.35 0.09 0.26 10.29
B. Shorter : detection 2486 6 0.37 0.39 0.23 0.10 0.09 0.69 0.20 0.58 3.16
C. Longer : physical activity 9863 35 0.23 0.26 0.12 0.02 0.12 0.41 0.22 0.31 21.57 D
D. Shorter : physical activity 13,278 66 0.34 0.39 0.15 0.02 0.20 0.58 0.35 0.43 16.52 C
E. Longer : dietary 2391 4 0.25 0.29 0.23 0.11 0.00 0.58 0.06 0.51 3.19
F. Shorter : dietary 6655 26 0.30 0.35 0.15 0.03 0.15 0.54 0.29 0.41 13.80
Health Psychology Review 25
Downloaded By: [Conner, Mark T] At: 17:41 15 February 2011
Table 7 (Continued )
Nkr
c
Mean r SD r SE r 10% CV 90% CV CI low CI high Variance (%) Diffs
SNBR A. Longer : detection 5884 11 0.10 0.11 0.06 0.02 0.03 0.18 0.06 0.15 33.75 B
B. Shorter : detection 2486 6 0.35 0.36 0.18 0.07 0.13 0.58 0.21 0.50 5.48 A
C. Longer : physical activity 9475 34 0.18 0.20 0.10 0.02 0.07 0.33 0.16 0.24 27.28
D. Shorter : physical activity 13,212 66 0.19 0.21 0.16 0.02 0.00 0.42 0.17 0.25 16.77
E. Longer : dietary 2391 4 0.10 0.11 0.17 0.09 0.11 0.33 0.07 0.28 6.17
F. Shorter : dietary 6658 26 0.17 0.19 0.14 0.03 0.02 0.37 0.13 0.25 19.40
PBBR C. Longer : physical activity 4157 14 0.46 0.48 0.09 0.03 0.36 0.60 0.43 0.53 20.18 d
D. Shorter : physical activity 5857 30 0.57 0.59 0.22 0.04 0.30 0.88 0.51 0.67 4.56 c
Note: BR, behaviour; INT, intention; PBC, perceived behavioural control; ATT, attitude; SN, subjective norm; PB, past behaviour; r
c
, average correlation corrected for
sampling error; mean r, true score correlation (corrected for sampling and measurement error); SD r, standard deviation of true score correlation; SE mean r, standard
error of mean true score correlation; 10% CV, 10% credibility interval; 90% CV, 90% credibility interval; CI low, lower 95% confidence interval; CI high, upper 95%
confidence interval; Variance, percentage variance in true score correlation attributable to statistical artefacts; Diffs, letters in capital indicate groups for which
confidence intervals are not overlapping, and letters in lowercase indicate confidence intervals that overlap by less than 0.02 points, comparisons are made within each
behaviour type only; Longer longer follow-up periods: 5 weeks; Shorter shorter follow-up periods: 55 weeks.
26 R.R.C. McEachan et al.
Downloaded By: [Conner, Mark T] At: 17:41 15 February 2011
intention is a better predictor of behaviour measured in the shorter compared to
longer term for both detection (mean r0.56 vs. 0.30) and physical activity (mean
r0.56 vs. 0.38) behaviours. Although the same trend is evident for dietary
behaviours (mean r0.46 vs. 0.37), this difference is not significant. With regards to
PBCbehaviour and attitudebehaviour relations, there again appears to be an
advantage for shorter compared to longer follow-up periods, although only
differences with physical activity behaviours are significant (mean r0.40 vs. 0.28
for PBC and 0.39 vs. 0.26 for attitude). For SNbehaviour relationships there
appears to be a clear advantage for shorter compared to longer time periods for
detection behaviours only (mean r0.36 vs. 0.11). Past behaviour was marginally
more predictive of behaviour across shorter follow-up periods compared with longer
follow-up periods (see Table 7).
7
Multiple regressions revealed that for all three behaviours where these analyses
were possible the overall efficacy of the model is better with shorter follow-ups
compared with longer follow-ups (see Table 3, third panel from top). The most
pronounced difference is for detection behaviours with just 9.3% variance in
behaviour explained over longer-term follow-ups compared with 31.4% variance
explained over shorter-term follow-ups. With the exception of dietary behaviours
across the longer term, intention is the most important predictor of behaviour. For
dietary behaviours measured in the longer terms SN is the most important predictor,
although caution must be noted as this analysis is based on only four tests. Amongst
physical activity behaviours, past behaviour seems to be of similar importance in
predicting behaviours across both the shorter and longer terms explaining between
9.0 and 11.1% additional variance. Across both time frames it becomes the most
important predictor of behaviour.
The mean percent of variation in statistical artefacts (Table 7) across all tests with
behaviour (e.g., INTBR, PBCBR, ATTBR, SNBR) was examined for each
group (excluding past behaviour). The most artefact variation was explained across
physical activity behaviours (mean artefact variance 20.96% longer term and 14.84%
shorter term), closely followed by detection behaviours (mean artefact variance
18.08% longer term and 11.62% shorter term). For dietary behaviours, relatively
more artefact variance was accounted for across studies with shorter follow-ups
(mean 14.63%) compared with longer follow-ups (mean 5.27%). This indicates that
there is still considerable study variation in this set of correlations.
Type of behavioural measure
Finally the impact of type of behavioural measure was explored. Unfortunately there
were only enough studies to populate physical activity behaviour studies (objective
k14 and self-report k 91).
8
Again, as these analyses were concerned with
measurement of prospective behaviour, only relationships with behaviour are
presented (see Table 8). It can be seen that prediction of objective behaviour is
more modest than that of self-reported behaviour for intention (mean r0.34 vs.
0.50), PBC (mean r0.18 vs. 0.36) and attitude (mean r0.17 vs. 0.35). It can be
seen from the multivariate regression (Table 3, bottom panel) that only 12.1% of
variance in objective physical activity is predicted compared with 25.7% variance in
self-reported physical activity. The mean variation in statistical artefacts was 26.4%
for objective measures of behaviour and 14.1% for self-report behaviour.
9
Health Psychology Review 27
Downloaded By: [Conner, Mark T] At: 17:41 15 February 2011
Table 8. Self-report vs. objective measures of physical activity.
Nkr
c
Mean r SD r SE r 10% CV 90% CV CI low CI high Variance (%) Diffs
INTBR PA: self-report 20,578 91 0.47 0.50 0.17 0.02 0.28 0.71 0.46 0.53 9.32 *
PA: Objective 2723 14 0.30 0.34 0.13 0.04 0.18 0.50 0.26 0.41 23.92
PBCBR PA: self-report 20,587 91 0.32 0.36 0.14 0.02 0.18 0.54 0.33 0.39 17.02 *
PA: Objective 2734 14 0.18 0.18 0.11 0.04 0.04 0.33 0.11 0.25 28.36
ATTBR PA: self-report 20,343 89 0.31 0.35 0.14 0.02 0.18 0.53 0.32 0.38 17.58 *
PA: Objective 2734 14 0.14 0.17 0.10 0.03 0.04 0.29 0.10 0.23 35.87
SNBR PA: self-report 20,284 89 0.19 0.21 0.14 0.02 0.03 0.39 0.18 0.24 18.62
PA: Objective 2734 14 0.16 0.18 0.12 0.04 0.03 0.33 0.10 0.25 29.41
Diffs: *denotes pair significantly different.
Note: PA, Physical activity; BR, behaviour; INT, intention; PBC, perceived behavioural control; ATT, attitude; SN, subjective norm; PB, past behaviour; r
c
, average
correlation corrected for sampling error; mean r, true score correlation (corrected for sampling and measurement error); SD r, standard deviation of true score
correlation; SE mean r, standard error of mean true score correlation; 10% CV, 10% credibility interval; 90% CV, 90% credibility interval; CI low, lower 95% confidence
interval; CI high, upper 95% confidence interval; Variance, percentage variance in true score correlation attributable to statistical artefacts. Note for confounding: Self-
report measure: 28 longer, 63 shorter; 31 adult, 34 students, 24 adolescent. Objective measure: 9 longer, 5 shorter; 10 adult, 1 student, 3 younger.
28 R.R.C. McEachan et al.
Downloaded By: [Conner, Mark T] At: 17:41 15 February 2011
Summary of results
Overall the TPB explains 19.3% of the variance in behaviour and 44.3% of the
variation in intention. However, the efficacy of the TPB varies depending on
behaviour type with physical activity and dietary behaviour best predicted, and safer
sex, detection, risk and abstinence behaviours relatively poorly predicted. With
regard to intentions, safer sex and diet behaviours are best predicted, with abstinence
behaviours poorly predicted. We next compared results hierarchically with behaviour
type for adult, student and adolescent samples. Comparisons with all three age
groups could only be carried out for physical activity behaviours, where it appears
that the physical activity behaviours and intentions of students are better predicted
that adult or adolescent samples. With regards to other behaviours it could be
seen that adolescent dietary behaviours were particularly poorly predicted. In
addition, with the exception of physical activity behaviours, SNs appear to have more
of a role in the prediction of intentions for adolescent samples. Methodological
characteristics were also of importance. Length of follow-up moderated model
relationships with behaviour better predicted in the shorter term. Self-report
behaviour measures were better predicted than objective behaviour measures. Where
possible we controlled for past behaviour in our analyses. Overall, past behaviour
adds 10.9% variance to the prediction of behaviour and an additional 5% to the
variance in intention. Past behaviour was the most important predictor of behaviour
but not intention.
Discussion
The current study aimed to conduct a meta-analysis of the TPB when applied to
health behaviours which addressed the limitations of previous reviews by including
only prospective tests of behaviour, applying RE meta-analytic procedures, correcting
correlations for sampling and measurement error, and hierarchically analysing the
effect of behaviour type and sample and methodological moderators. Some 237 tests
were identified which examined relations amongst model components. Overall the
analysis indicated that the TPB could explain 19.3% of the variance in behaviour and
44.3% of the variance in intention across studies. This level of prediction of behaviour
is slightly lower than that of previous meta-analytic reviews which have found between
27% (Armitage & Conner, 2001; Hagger et al., 2002) and 36% (Trafimow et al., 2002)
of the variance in behaviour to be explained by intention and PBC. Both the
intentionbehaviour correlation (0.43 vs. 0.47) and the PBCbehaviour correlation
(0.31 vs. 0.37) were weaker in the present study when compared to Armitage and
Conner (2001). However, the prediction of intention from TPB variables was similar
to levels identified by past reviews of between 42% and 45% variance explained
(Armitage & Conner, 2001; Hagger et al., 2002; Rivis & Sheeran, 2003a; Schulze
& Whittmann, 2003; Sheeran & Taylor, 1999; Trafimow et al., 2002). Indeed, the
attitudeintention correlation (0.57 vs. 0.49), the SNintention (0.40 vs. 0.34) and
the PBCintention correlation (0.54 vs. 0.43) were stronger in the present study
compared to Armitage and Conner (2001). The variations in findings are of course
a function of the studies included. They may also, in part, be attributable to
differences such as the focus on prospective studies, the use of RE meta-analyses and
the controlling for both sampling and measurement error in the present analyses.
Health Psychology Review 29
Downloaded By: [Conner, Mark T] At: 17:41 15 February 2011
Based on the sub-sample of studies that provided the relevant correlations it
was found that past behaviour exhibited the strongest association with behaviour
(mean r0.50). However, intention was also a strong predictor of behaviour (mean
r0.43). Attitude and PBC showed similar levels of prediction (both mean rs0.31).
Similar to previous meta-analyses attitudes emerged as the strongest predictor of
intentions (mean r0.57), and remained so after controlling for past behaviour. The
medium to large effect sizes of these latter relationships suggest that the TPB identifies
important targets for interventions to change health behaviours. Whilst from
a predictive perspective it is useful to take past behaviour into account, from an
intervention perspective, past behaviour is not so readily changed as traditional TPB
variables and so is of limited use to those tasked with changing behaviour. Therefore, it
is encouraging to note that controlling for past behaviour, intentions still emerged as
important predictors of behaviour (albeit only equivalent to a small effect size), and
attitudes remained the most important predictor of intention.
Moderation by behaviour type
As expected the overall analysis exhibited substantial heterogeneity and thus a search
for further moderators was warranted. Behaviour type was found to be a moderator
of TPB relationships. Generally, splitting studies according to behaviour type
captured more variance in statistical artefacts than grouping studies together,
indicating behaviour type to be a valid moderator of the TPB (the evidence was
weaker for detection and risk behaviours, groupings that contain a more diverse set
of health behaviours). Examination of the overall levels of prediction and relation-
ships between components of the TPB revealed a number of differences. At the
overall level (Tables 3 and 4), physical activity (23.9%) and dietary (21.2%)
behaviours were generally better explained, with intentions being the key predictor
of subsequent engagement in behaviour. In contrast safer sex, detection, risk and
abstinence behaviours were relatively poorly predicted by the model, with on average
8% less variance explained. Although the explained variances might suggest the value
of examining addition predictors of behaviour within the TPB (Ajzen, 1991),
the present data would suggest this to be particularly the case in relation to these
latter behaviours. For example, affective influences including affective attitudes
(Conner & Sparks, 2005) and anticipated regret (Sandberg & Conner, 2008) have
been suggested as useful possible additions to the TPB (see Conner & Armitage,
1998, for a broader review of additional variables in the TPB).
At the level of relationships between components of the TPB a large number of
differences emerged across behaviours (Table 2). Notably the PBCbehaviour, the
SNintention, the past behaviourbehaviour and the past behaviourintention
relationships showed variations across behaviours. In relation to the PBCbehaviour
relationship, this was stronger for physical activity and dietary behaviours compared
to detection and safer sex behaviours. Previous work has noted that SN are the
weakest predictor in the TPB (e.g., Armitage & Conner, 2001; Godin, Savard, Kok,
Fortin, & Boyer, 1996; Hagger et al., 2002; Hausenblas et al., 1997). The present
analyses indicate that the role of SN varies dependent on behaviour type. The
SNintention relationship was stronger for safer sex (as observed in other reviews;
Albarracin et al., 2001, 2004) compared to detection, physical activity and abstinence
behaviours. Finally, the past behaviourbehaviour and past behaviourintention
30 R.R.C. McEachan et al.
Downloaded By: [Conner, Mark T] At: 17:41 15 February 2011
relationships were stronger for risk and physical activity behaviours compared to diet
and abstinence behaviours. These findings may provide useful guidance as to which
variables to target when attempting to change different health behaviours. Analysing
the differences in predictors by behaviour type rather than by a higher-level
theoretical categorisation not only translates more easily into practical advice for
those attempting to change behaviour, but also appears to reveal more differences.
For example, classifying our studies simply into health-promoting vs. health-risking
behaviours (mean r0.20 vs. 0.29 for SNBR) or frequent vs. infrequent behaviours
(mean r0.34 vs. 0.22 for PBCBR) each revealed only one significant difference
(although an imbalance in groups means limited power to detect significant effects).
Sample and methodological moderators
In addition to behavioural moderators, other factors also served to moderate
relationships in the model. We examined these moderators hierarchically (i.e., within
behavioural categories) and excluded groups where moderators were confounded.
Full comparisons within behaviours across the three age groups were difficult,
because it was not possible to populate all cells. Nevertheless, we were able to make
comparisons for physical activity, diet and abstinence behaviours and a number of
differences emerged. Whilst the differences were relatively modest for physical
activity behaviour (slightly stronger predictions of behaviour and intentions for
students compared to the other two groups; no significant differences for individual
correlations; cf. Peterson, 2001), larger differences emerged for dietary and
abstinence behaviours. In relation to dietary behaviours, the TPB showed consider-
ably better predictions of behaviour in the adult compared to the adolescent samples
(26.7% vs. 9.6%; Table 3) with both intentions and PBC showing significantly
stronger correlations with behaviour in the adult compared to the student samples
(Table 6). This may reflect both the less reasoned nature of adolescent dietary
behaviour (Gibbons et al., 2009; Hall et al., 2008) and perhaps a lower degree of
control over this behaviour in this group. At a practical level it suggests a weakness of
the TPB in relation to explaining dietary behaviours in adolescents. In relation to
abstinence behaviours, the level of prediction was slightly higher for adolescents
compared to students (Table 3). However, this masked larger differences in relation
to individual correlations. In particular, PBC was a much stronger predictor of both
behaviour and intentions for the adolescent compared to student samples. This
difference might represent a useful focus for future research to explore.
Length of follow-up of behavioural assessment also moderated the efficacy of the
model such that behaviours measured over the shorter term (less than 5 weeks) were
better predicted that those over the longer term (more than 5 weeks). This effect was
strongest for detection (9.3% vs. 31.4% explained variance) and physical activity
(15.7% vs. 32.0%), and less so for dietary behaviours (18.4% vs. 22.7%; Table 3).
Examination of the individual correlations indicated that it was particularly the
intentionbehaviour relationship that was attenuated as the length of follow-up
increased (Table 7). This finding replicates previous results (Albarracin et al., 2004;
Manning, 2009; Sheeran & Orbell, 1998), and is likely to be attributable to the influence
of other factors being more likely to change intentions between the time point at which
they are measured and the time point at which they might influence behaviour when the
follow-up period is longer (cf. Ajzen, 1991). Various studies have demonstrated that
Health Psychology Review 31
Downloaded By: [Conner, Mark T] At: 17:41 15 February 2011
temporally stable intentions are more predictive of behaviour (for a review see Conner
& Godin, 2007) and interventions to promote strong and stable intentions may
be particularly important for distal behaviours. It was also the case that the
PBCbehaviour, attitudebehaviour and past behaviourbehaviour correlations for
physical activity and the SNbehaviour correlation for detection behaviours were
similarly attenuated. Manning (2009) reported an advantage of SN (measured as
descriptive norms) in predicting behaviour over longer periods. The present
hierarchical analysis did not support this prediction within behavioural categories
(indeed SN was significantly weaker for detection behaviours).
Objective vs. self-reported behavioural measure was the last moderator examined
(this was only possible for physical activity due to the limited numbers of studies).
Similar to previous reviews (e.g., Armitage & Conner, 2001; see also Abraham &
Graham-Rowe, 2009, in relation to workplace physical activity interventions) we
found (Table 3) that objective behaviour (12.1% explained variance) was less well
predicted than self-report behaviour (25.7% explained variance). The reduction in
predictive power for objective compared to self-report behaviours was consistent for
intention, PBC and attitudes as predictors (the effect sizes were generally in the small
to medium range for objective behaviour measures compared to the medium to large
range for self-report behaviour; Table 8). In general, one might expect objective
measures of health behaviours to be more strongly associated with health outcomes.
As such these results are disappointing and suggest that the power of the TPB to
predict behaviours with strong relationships to health outcomes is more modest than
the overall relationships might suggest. However, it would be important to confirm
the size of these differences for behaviours other than physical activity.
Strengths and weaknesses
The current meta-analysis has a number of strengths. First, by using RE meta-
analytic procedures the results of this meta-analysis can justifiably be generalised to
studies out with those included in the review. Second, by correcting for sampling and
measurement error a better indication of the true correlations between model
components can be ascertained. Third, the meta-analysis hierarchically examined
moderating variables to get a better understanding of the impact of each factor
without fear of confounding different levels of the moderators. Finally, this meta-
analysis has been the first to systematically explore differences in prediction
according to type of behaviour using meta-analytic techniques. The identification
of significant differences amongst model components at different levels of the
moderators is especially impressive given the comparatively low power of the
comparisons to detect differences.
There were, however, limitations to the present meta-analysis. Although the
behavioural moderator was examined hierarchically with each methodological
moderator, unfortunately there were insufficient studies to warrant a full hierarchical
analysis including sample, methodological and behaviour-type moderators simulta-
neously. Nonetheless, a number of interesting findings were apparent which might
not have been found had studies simply been grouped together in a traditional
manner. In addition, despite the moderators considered, there generally remained
large amounts of heterogeneity amongst study correlations. This could be due to
statistical artefacts, which were not corrected for in the current study, or the presence
32 R.R.C. McEachan et al.
Downloaded By: [Conner, Mark T] At: 17:41 15 February 2011
of additional moderator variables not considered here. Finally, the current meta-
analysis included only published studies meaning results may be inflated by
publication bias (if studies with significant effects are more likely to be published).
However, it is noted that a meta-analytic review of unpublished studies by Schulze
and Whittmann (2003) found similar levels of prediction to those containing only
published studies.
In revealing gaps in the literature in relation to the application of the TPB to
specific behaviours with specific populations and follow-up periods (Table 5), the
present review also identifies useful directions for future research with the TPB. On
this basis, further studies in relation to particular behaviours (e.g., abstinence),
samples (e.g., adults) and follow-up periods (e.g., less than 5 weeks) appear more
informative than other combinations (e.g., physical activity in students over shorter
follow-up periods). Studies that include objective measures of both behaviour and
past behaviour in testing the TPB are particularly lacking in this area and represent a
useful area for future research.
Conclusions
The current study demonstrated that prospective tests of the TPB provide strong
predictions of intention and behaviour across a range of health behaviours. It was
found that behaviour type moderated relationships amongst the model components
and that physical activity and dietary behaviours were better predicted by the model.
Age of sample moderated relationships for specific relationships and specific
behaviours. Length of follow-up and type of behavioural measure also moderated
the TPB such that shorter follow-up time periods, and self-report measures of
behaviour were better predicted. The present findings can usefully inform the design
of interventions to change different health behaviours in different populations.
Acknowledgements
We would like to thank the following individuals for providing further information regarding
their studies: Charles Abraham, Madelynne Arden, Chris Armitage, Anne Nordrehaug
Astrom, Desiree Backman, Vasilis Barkoukis, Blair Beadnell, Angela Bryan, Jill Budden,
Mary Bursey, Nikos Chatzisarantis, Marilia Corne´lio, John De Wit, Ron Dishman, Joan
Dodgson, Constance Drossaert, Mark Elliott, Chris Fife-Schaw, Rebecca Ellis Gardner,
Sylvmarie Gatt, Melanie Giles, Gaston Godin, Daniel Gredig, Martin Hagger, Kyra
Hamilton, Mary Hanson, George Higgins, Melissa Hyde, Cath Jackson, Lee Jones, Ian
Kellar, Matthew Kerner, Bethany Kwan, Andrew Lac, Amy Latimer, Deirbhile Lavin, Lilian
Lechner, Winnifred Louis, Anil Mather, Brian McMillan, Ilse Mesters, Eric Nehl, Inger
Synnove Moan, Diane Morrison, Paul Norman, Roman Pawlak, Jost Reinecke, Juliette
Richetin, Ryan Rhodes, Derek Rutter, Paschal Sheeran, Falko Sniehotta, Liz Steadman,
Yannis Theodorakis, Stewart Trost, Monique van de Ven, Bas van den Putte, Rolf van Hulten,
Liesbeth van Osch, An Victoir and Karen Wambach.
Notes
1. Please note that asterisk refers to truncated term.
2. A decision was made to only include published research in the current meta-analysis. It was
not felt this would pose a threat to the validity of the meta-analysis for two reasons. First,
a recent review of unpublished TPB research (Schulze & Whittmann, 2003) found similar
levels of prediction to meta-analyses of published studies. Second, previous meta-analyses
Health Psychology Review 33
Downloaded By: [Conner, Mark T] At: 17:41 15 February 2011
have indicated that the fail safe N (the number of studies with null findings needed to make
results non-significant) to number in the tens of thousands for relations amongst model
components (e.g., Armitage & Conner, 2001).
3. Six tests were excluded at this point as the required data were not available for the relevant
behaviour (four for physical activity and two for abstinence).
4. Risk had k2 adult, k15 student and k12 adolescent. The latter group was
confounded with age of follow-up as 11 school-age tests reported longer follow-ups.
Detection behaviours were all with adult samples. Dietary student samples k12 were
excluded as all reported a shorter length of follow-up. Safer sex contained k3 adolescent,
k2 students and k8 adult. The former and latter group were confounded as all reported
longer follow-ups. Abstinence behaviours reported k5 adolescent, k3 students and
k5 adult. The latter group was confounded as all reported longer follow-ups.
5. Analysis of this moderator was repeated not split by behaviour type for correlations with
intention and behaviour. This analysis revealed that INTBR, PBCBR, ATTBR and
ATTINT relationships were significantly moderated such that student samples were better
predicted than either school or adult samples (all student vs. school comparisons
significant, comparisons between student and adult samples significant for INTBR and
ATTINT and marginally significant for PBCBR and ATTBR). Where SN was involved
adult samples had weaker relationships with either student or adolescent samples. Less of
the artefact variance was explained in the non-hierarchical analysis (range from mean
6.06% adolescent to mean 14.50% students). Further details of this analysis can be
requested from the first author.
6. For risk behaviours there were k16 longer follow-up and k13 shorter follow-up,
although of the latter 12 were conducted with student samples. For safer sex behaviours
there were k12 longer follow-up and k2 shorter follow-up. For abstinence behaviours
there were k11 longer follow-up and k2 shorter follow-up.
7. Non-hierarchical analysis of this moderator variable found that all relations with behaviour
including SNs were significantly moderated with shorter follow-ups predicted better than
longer follow-ups (SNs were marginally significant). Mean percent artefact variation
explained ranged from 8.81 to 13.01%, again, lower than the hierarchical analysis. Full
details can be requested from the first author.
8. This included three studies which measured both objective and self-reported physical
activity (Armitage, 2005; Kwan & Bryan, 2010 additional objective measure provided on
request; Maddison et al., 2009). One study was excluded as it used both objective and self-
report elements to make an index (Theodorakis, 1994). Risk, diet, safer sex and abstinence
studies all had self-report measures of behaviour, detection studies had k 15 objective
measures and k2 self-report measures.
9. The non-hierarchical analysis found self-report measures were significantly better predicted
by all variables including past behaviour. Mean percent artefact variation was 10.47% for
self-report measures and 10.67% for objective measures.
References
*Abraham, C., Clift, S., & Grabowski, P. (1999). Cognitive predictors of adherence to malaria
prophylaxis regimens on return from a malarious region: A prospective study. Social Science
& Medicine, 48, 16411654.
Abraham, C., & Graham-Rowe, E. (2009). Are worksite interventions effective in increasing
physical activity? A systematic review and meta-analysis. Health Psychology Review, 3,
108144.
*Abraham, C., Henderson, M., & Der, G. (2004). Cognitive impact of a research-based school
sex education programme. Psychology & Health, 19, 689703.
*Abraham, C., & Sheeran, P. (2003). Acting on intentions: The role of anticipated regret.
British Journal of Social Psychology, 42, 495511.
Ajzen, I. (1985). From intention to actions: A theory of planned behaviour. In
J. Kuhl & J. Beckmann (Eds.), Action control: From cognitions to behaviour (pp. 1139).
New York: Springer-Verlag.
34 R.R.C. McEachan et al.
Downloaded By: [Conner, Mark T] At: 17:41 15 February 2011
Ajzen, I. (1988). Attitudes, personality and behaviour. Milton Keynes, UK: Open University
Press.
Ajzen, I. (1991). The theory of planned behaviour. Organizational Behaviour and Human
Decision Processes, 50, 179211.
Ajzen, I. (2006). Constructing a TPB questionnaire: Conceptual and methodological considera-
tions. Retrieved from http://www.people.umass.edu/aizen/pdf/tpb.measurement.pdf
Ajzen, I., & Fishbein, M. (1980). Understanding attitudes and predicting social behaviour.
Englewood Cliffs, NJ: Prentice-Hall.
Albarracin, D., Johnson, B.T., Fishbein, M., & Muellerleile, P.A. (2001). Theories of reasoned
action and planned behaviour as models of condom use: A meta-analysis. Psychological
Bulletin, 127, 142161.
Albarracin, D., Kumkale, G.T., & Johnson, B.T. (2004). Inuences of social power and
normative support on condom use decisions: A research synthesis. AIDS Care, 16, 700723.
*Arden, M.A., & Armitage, C.J. (2008). Predicting and explaining transtheoretical model
stage transitions in relation to condom-carrying behaviour. British Journal of Health
Psychology, 13, 719735.
*Armitage, C.J. (2004). Evidence that implementation intentions reduce dietary fat intake:
A randomized trial. Health Psychology, 23, 319323.
*Armitage, C.J. (2005). Can the theory of planned behaviour predict the maintenance of
physical activity? Health Psychology, 24, 235245.
*Armitage, C.J. (2007a). Effects of an implementation intention-based intervention on fruit
consumption. Psychology & Health, 22, 917928.
*Armitage, C.J. (2007b). Efcacy of a brief worksite intervention to reduce smoking: The roles
of behavioral and implementation intentions. Journal of Occupational Health Psychology,
12, 376390.
*Armitage, C.J., & Arden, M.A. (2008). How useful are the stages of change for targeting
interventions? Randomized test of a brief intervention to reduce smoking. Health
Psychology, 27, 789798.
*Armitage, C.J., & Conner, M. (1999a). Distinguishing perceptions of control from self-
efcacy: Predicting consumption of a low-fat diet using the theory of planned behaviour.
Journal of Applied Social Psychology, 29,7290.
*Armitage, C.J., & Conner, M. (1999b). The theory of planned behaviour: Assessment of
predictive validity and perceived control. British Journal of Social Psychology, 38,3554.
Armitage, C.J., & Conner, M. (2001). Efcacy of the theory of planned behaviour: A meta-
analytic review. British Journal of Social Psychology, 40, 471499.
*Armitage, C.J., Conner, M., Loach, J., & Willetts, D. (1999). Different perceptions of control:
Applying an extended theory of planned behaviour to legal and illegal drug use. Basic and
Applied Social Psychology, 21, 301316.
*Armitage, C.J., Norman, P., & Conner, M. (2002). Can the theory of planned behaviour
mediate the effects of age, gender and multidimensional health locus of control? British
Journal of Health Psychology, 7, 299316.
*Astrom, A.N. (2004). Validity of cognitive predictors of adolescent sugar snack consumption.
American Journal of Health Behaviour, 28, 112121.
*Backman, D.R., Haddad, E.H., Lee, J.W., Johnston, P.K., & Hodgkin, G.E. (2002).
Psychosocial predictors of healthful dietary behaviour in adolescents. Journal of Nutrition
Education and Behaviour, 34, 184193.
*Bai, Y., Middlestadt, S.E., Peng, C.Y.J., & Fly, A.D. (2010). Predictors of continuation of
exclusive breastfeeding for the rst six months of life. Journal of Human Lactation, 26,
2634.
*Barkoukis, V., & Hagger, M.S. (2009). Test of the trans-contextual model of motivation in