Selasa, 02 Februari 2010

Efficacy of the Theory of Planned Behaviour



Efficacy of the Theory of Planned Behaviour:
A meta-analytic review

Armitage, C. J., & Conner, M. (2001). Efficacy of the theory of planned behaviour: A metaanalytic review. British journal of social psychology, 40(4), 471-499.

The Theory of Planned Behaviour (TPB) has received considerable attention in the literature. The present study is a quantitative integration and review of that research. From a database of 185 independent studies published up to the end of 1997, the TPB accounted for 27% and 39% of the variance in behavior and intention, respectively. The perceived behavioural control (PBC) construct accounted for signi. cant amounts of variance in intention and behaviour, independent of theory of reasoned action variables. When behaviour measures were self-reports, the TPB accounted for 11% more of the variance in behavior than when behaviour measures were objective or observed (R2s = .31 and .21, respectively). Attitude, subjective norm and PBC account for signi. cantly more of the variance in individuals’ desires than intentions or self-predictions, but intentions and self-predictions were better predictors of behaviour. The subjective norm construct is generally found to be a weak predictor of intentions. This is partly attributable to a combination of poor measurement and the need for expansion of the normative component. The discussion focuses on ways in which current TPB research can be taken forward in the light of the present review.


Since Wicker’s (1969) review of research examining the relationship between attitudes and behaviour, and his conclusion that attitudes probably do not predict behaviour, social psychologists have sought to improve the predictive power of attitudes. In recent years, the main approach within this area has been to develop integrated models of behaviour, including additional determinants of behavior such as social norms or intentions (Olson & Zanna, 1993). Arguably the most widely researched of these models are the Theories of Reasoned Action (Ajzen & Fishbein, 1980; Fishbein & Ajzen, 1975) and Planned Behaviour (Ajzen, 1988, 1991). The Theory of Planned Behaviour (TPB) is essentially an extension of the Theory of Reasoned Action (TRA) that includes measures of control belief and perceived behavioural control (see Fig. 1).

*Request for reprints should be addressed to Chris Armitage, Centre for Research in Social Attitudes, Department
of Psychology, University of SheYeld, Western Bank, SheYeld S10 2TP, UK.

Ajzen (e.g. 1991) extended the TRA to include a measure of perceived behavioural control—a variable that had received a great deal of attention in social cognition models designed to predict health behaviours (e.g. health belief model, protection motivation theory; see Armitage & Conner, 2000; Conner & Norman, 1996a). Perceived behavioural control (PBC) is held to in uence both intention and behaviour (see Fig. 1). The rationale behind the addition of PBC was that it would allow prediction of behaviours that were not under complete volitional control.
Thus, while the TRA could adequately predict behaviours that were relatively straightforward (i.e. under volitional control), under circumstances where there were constraints on action, the mere formation of an intention was insuYcient to predict behaviour. The inclusion of PBC provides information about the potential constraints on action as perceived by the actor, and is held to explain why intentions do not always predict behaviour.1
With respect to the in uence of PBC on intention, Ajzen (1991, p. 188) states that: ‘The relative importance of attitude, subjective norm, and perceived behavioral control in the prediction of intention is expected to vary across behaviors and situations’. That is, in situations where (for example) attitudes are strong, or where normative in uences are powerful, PBC may be less predictive of intentions. Thus, Ajzen (1991) argues that the magnitude of the PBC–intention relationship is dependent upon the type of behaviour and the nature of the situation. Indirect evidence for this claim has been demonstrated in studies that have shown that measures of attitude strength (e.g. Sparks, Hedderley, & Shepherd, 1992) and individual diVerences in sociability (e.g. Tra.mow & Finlay, 1996) increase the relative predictive power of attitudes and subjective norms, respectively.
In general, individuals are more disposed (i.e. intend) to engage in behaviours that are believed to be achievable (cf. Bandura, 1997).
PBC is also held to exert both direct and interactive (with behavioural intentions) eVects on behaviour. This is based on the following rationale: that however strongly held, the implementation of an intention into action is at least partially determined by personal and environmental barriers, thus: ‘The addition of perceived behavioural control should become increasingly useful as volitional control over behavior decreases’ (Ajzen, 1991, p. 185). Therefore, in situations where prediction of behaviour from intention is likely to be hindered by the level of actual (i.e.volitional) control, PBC should (1) facilitate the  mplementation of behavioural intentions into action, and (2) predict behaviour directly.


1The authors wish to thank Russell Spears for his helpful comments with respect to the following section.

Figure 1. The theory of planned behaviour.


In the prediction of social behaviours, there are no absolutes. However, it is instructive to consider Ajzen’s (1991) predictions by examining the impact of PBC on behaviour under both optimal (i.e. complete volitional control) and suboptimal (i.e. problems of volitional control) conditions.2
In conditions of complete volitional control, the intention–behaviour relationship should be optimal, and PBC should not exert any in uence on this relationship. In contrast, where the behaviour is not under complete volitional control, PBC should moderate (see Baron & Kenny, 1986) the relationship between intention and behaviour. Under such conditions, greater PBC should be associated with stronger intention–behaviour relationships. In earlier versions of the TPB, Ajzen (e.g. 1985) emphasized the fact that the interaction between behavioural intention and PBC should be independently predictive of behaviour. That is, under conditions where volitional control is relatively low (i.e. where intention is only weakly related to behaviour), increased PBC should facilitate the implementation of intentions into action. However, in his meta-analysis of the TPB, Ajzen (1991) reported that only one study had found the (marginally) signi. cant (p<.10) intention–PBC interaction that would support this moderator hypothesis. Ajzen (1991) argued that this . nding re ected the fact that linear models account well for psychological data—even if interaction terms are known to be present. Yet, several, more recent studies (e.g. Terry & O’Leary, 1995) have found signi. cant PBC–intention interactions, and the present meta-analysis examines these to test this moderator hypothesis.
Following the lack of evidence for the interactive eVects of PBC on the intention–behaviour relationship, Ajzen (1991) argued for a direct relationship between PBC and behaviour which more closely . tted the available data. Thus, Ajzen argues that under conditions where behavioural intention alone would account for only small amounts of the variance in behaviour (i.e. where there are problems of volitional control), PBC should be independently predictive of behaviour. This is based on the rationale that increased feelings of control will increase the extent to which individuals are willing to exert additional eVort in order successfully to perform a particular behaviour. In contrast, under conditions of very high volitional control, behavioural intention should be the only predictor of behaviour. This ceiling eVect occurs because where the behaviour is relatively straightforward, exerting additional eVort to engage in the behaviour will not impact on the actual performance of the behaviour, over and above the eVects of intention.

2Note that by ‘problems of volitional control’, we are referring to environmental and personal constraints on
behaviour. For example, consider cigarette smoking: here, an environmental barrier might be that everyone at work
smokes; a personal barrier might be the level of craving for cigarettes (for further discussion of these issues see
Armitage & Conner, 1999a, 1999b).

However, predictions concerning the eVects of PBC on behaviour are clouded by the explicit assumption that PBC is an accurate representation of actual (volitional) control. Indeed, Aizen3 states that ‘When PBC is inaccurate all kinds of possibilities open up’ (I. Aizen, personal communication, 8 November 1999). Thus, where PBC and actual control are discrepant, the eVect of PBC on behaviour is more problematic. Given the extant literature on ‘illusions of control’ (e.g. Langer, 1975; Lerner, 1977), it seems likely that PBC will rarely re ect actual control in a very accurate way. In short, adequate tests of predictions concerning the eVects of PBC on behaviour would either (1) experimentally manipulate individuals’ levels of perceived control, or (2) obtain independent measures of volition (actual control). These are matters for future research, and cannot be adequately addressed using meta-analysis.4
As we have already noted, within the TPB, PBC is held to aVect both intentions and behaviour. There are two further antecedents of intention: subjective norm and attitude toward the behaviour, which are retained from the earlier TRA. Subjective norm refers to the individual’s perceptions of general social pressure to perform (or not to perform) the behaviour. If an individual perceives that signi. cant others endorse (or disapprove of) the behaviour, they are more (or less) likely to intend to perform it. Attitude towards the behaviour re ects the individual’s global positive or negative evaluations of performing a particular behaviour. In general, the more favourable the attitude towards the behaviour, the stronger should be the individual’s intention to perform it.
The antecedents of attitude, subjective norm and PBC are corresponding beliefs, reflecting the underlying cognitive structure. Each behavioural belief links a given behaviour to a certain outcome, or to some other attribute, such as the cost incurred in performing the behaviour. The attitude towards the behaviour is determined by the strength of these associations, and by the beliefs that are salient at the time. This works on the principle of Fishbein and Ajzen’s (1975) Expectancy-value Model: the subjective value of a given outcome aVects the attitude in direct proportion to the strength of the belief. Subjective norm is considered to be a function of salient normative beliefs. While subjective norm relates to perceptions of general social pressure, the underlying normative beliefs are concerned with the likelihood that speci. c individuals or groups (referents) with whom the individual is motivated to comply will approve or disapprove of the behaviour. According to Ajzen (1991), control beliefs are the antecedents of PBC, and are concerned with the perceived power of speci. c factors to facilitate or inhibit performance of the behaviour. Like the other beliefs, the equation takes account of the relevance of the belief to the individual, in this case by taking a measure of the frequency of occurrence of the promoting (or inhibitory) factor.

3Note that ‘Ajzen’ recently changed his name to ‘Aizen’.
4A number of previous meta-analyses have suggested that the TPB adds very little explained variance beyond that bnwhich is explained by the TRA (e.g. Sutton, 1998). One possibility is that as volitional control decreases, the in uence of PBC on intention and behaviour increases, although even studies designed to directly test this hypothesis have not produced clear-cut . ndings (e.g. Madden, Ellen, & Ajzen, 1992). We therefore attempted to code studies for ‘level of volitional control’. On 3-point scales, raters were asked to judge whether the behaviour in question was under volitional control, not under volitional control, or whether it was unclear. Initial analysis of coding reliability indicated 68% agreement. Following discussion, this increased to 79%, leaving over 20% of cases unresolved. Analyses of the categories revealed no substantive diVerences between groups, and no decrement in between-study variance. An alternative procedure—suggested by one of our anonymous reviewers—was therefore adopted, which is set out in the Appendix.

PBC will therefore be increased by salient beliefs concerning adequate resources and opportunities and fewer anticipated obstacles or impediments. Reviews have provided support for the TPB (e.g. Blue, 1995; Conner & Sparks, 1996; Godin, 1993; Jonas & Doll, 1996; Manstead & Parker, 1995; Sparks, 1994), as have four previous meta-analyses (Ajzen, 1991; Godin & Kok, 1996; Hausenblas, Carron, & Mack, 1997; Van den Putte, 1991). However, these meta-analyses have been limited in scope and sampling. For example, although Van den Putte (1991) reported that PBC explained an additional 14% of the variance in intention and 4% in behaviour (over and above attitude and subjective norm), discussion of issues surrounding this . nding was limited because the focus of his study was the TRA.5 Ajzen’s (1991) meta-analysis of the TPB found an average multiple correlation of attitude, subjective norm and PBC, with intention of R = .71 (19 correlations), and an average multiple correlation of R = .51 (17 correlations) for prediction of behaviour from intention and PBC. However, these analyses considered only the direct antecedents of intention and behaviour, and were based upon a limited data set, including studies that have never been published. Godin and Kok’s (1996) meta-analysis found that PBC contributed a mean additional 13% of variance to the prediction of intentions and 12% to the prediction of behaviour. However, Godin and Kok considered only health behaviours, and reported values that were derived only from studies that reported the relevant data. The tendency for authors to report only signi. cant . ndings may have in ated the reported values (cf. Rosenthal, 1979). Finally, Hausenblas et al. (1997) report a meta-analysis on applications of the TRA and TPB to exercise behaviour. They conclude that the TPB is more useful than the TRA, but base this conclusion solely on the magnitude of correlations between PBC, intention and behaviour.
More generally, previous meta-analyses of the TRA/TPB have tended to analyse data from participants more than once, have failed to report reliability statistics, and treated all studies as equivalent, with no attempt to weight their data in favour of studies with more participants. However, in spite of these weaknesses, evidence from narrative and meta-analytic reviews suggests that the TPB is a useful model for predicting a wide range of behaviours and behavioural intentions. The present meta-analysis aims to overcome some of the methodological weaknesses of previous meta-analyses and to focus on several of the issues in current TPB research.

Issues surrounding the TPB
Self-report
Behavioural decision-making models such as the TRA and TPB have tended to rely on self-reports, despite evidence to suggest the vulnerability of such data to self-presentational biases (e.g. Gaes, Kalle, & Tedeschi, 1978). To a great extent, this has been ignored in the literature pertaining to the TRA/TPB, in spite of the threat to the validity and reliability of the models. Beck and Ajzen (1991) provided an exception, applying the TPB and a Marlowe–Crowne Social Desirability Scale (SDS; Crowne & Marlowe, 1964) to predicting dishonest intentions and actions (cheating, shoplifting and lying). SDS scores were entered into a regression equation and accounted for 5% of the variance in intentions, providing some evidence to suggest that individuals may provide socially desirable answers in terms of their attitudes and intentions.6 Six months later, TPB variables were able to account for between 12% and 55% of the variance in self-reported behaviour. In contrast, however, Armitage and Conner (1999c) reported few eVects of social desirability on relationships between TPB components.
More closely related to the concerns of the present study, Hessing, ElVers, and Weigel (1988) examined the TRA in relation to tax evasion, and contrasted self-reports with oYcial documentation. Findings indicated that attitudes and subjective norms signi. cantly correlated with self-reported behaviour, but did not correlate with documentary evidence, in spite of considerable eVort to maintain the anonymity of respondents. The implication was that self-reports of behaviour were unreliable, compared with more objective behaviour measures (see also Armitage & Conner, 1999a, 1999b; Norwich & Rovoli, 1993; Pellino, 1997). In terms of the present study, we expected TPB variables (i.e. intention and PBC) to predict self-reported and observed behaviour, but that prediction of objective behavior would be less accurate.

Control

It has already been noted that the diVerence between the TRA and TPB lies in the control component of the TPB. Ajzen (1991) argues that the PBC and self-efficacy constructs are interchangeable. However, several authors (e.g. Terry, 1993) have suggested that self-efficacy and PBC are not entirely synonymous. For example, Bandura (1986, 1992) has argued that control and self-efficacy are quite diVerent concepts. Self-efficacy is more concerned with cognitive perceptions of control based on internal control factors, whereas PBC also re ects more general, external factors. Researchers such as de Vries, Dijkstra, and Kuhlman (1988) have advocated the use of measures of self-efficacy, as opposed to PBC in the prediction of intentions and behaviour. Further, Dzewaltowski, Noble, and Shaw (1990), in a comparison of the theories of reasoned action, planned behaviour and social cognitive theory, found that self-efficacy, rather than PBC, had a direct impact on behaviour.
Terry and colleagues have closely examined the distinction between PBC and self-efficacy. For safer sex behaviours, White, Terry, and Hogg (1994) reported that PBC only had an eVect on a behavioural measure of discussing the use of condoms with any new partner, while self-efficacy had a strong eVect on intentions to discuss and intentions to use condoms. Consonant with White et al. (1994), Terry and O’Leary (1995) found that self-efficacy only predicted intentions, while PBC predicted exercise behaviour. These studies therefore provide support for a distinction between self-efficacy and PBC (see also Manstead & van Eekelen, 1998).

5Indeed, this analysis was presented in the introduction to his meta-analysis of the TRA.

6However, this . nding must be interpreted with extreme caution, as SDS measures were not taken at the same time
as the six-month self-report of behaviour. Perhaps more importantly, SDS also rely on self-report.

More recently, Sparks, Guthrie, and Shepherd (1997) have proposed a distinction between ‘perceived diYculty’ and ‘perceived control’ (see also Chan & Fishbein, 1993). These authors argue that items which tap ‘perceived diYculty’ are both more meaningful to participants and are closer to Ajzen’s (1991) original conceptualization of PBC. Sparks et al. (1997) report two studies to support their position. In their study 1, although they found diVerences in the pattern of intercorrelations, neither ‘perceived diYculty’ nor ‘perceived control’ predicted intention. In study 2, ‘perceived diYculty’ independently predicted intention but ‘perceived control’ did not. These . ndings were interpreted as evidence to support the use of ‘perceived diYculty’ over ‘perceived control’. Armitage and Conner (1999a, 1999b) have critiqued this approach, arguing that asking individuals about the ‘ease’ or ‘diYculty’ of performing a particular behaviour does not allow discrimination between ease or diYculty in relation to external (e.g. ‘availability’) and internal (e.g. ‘con. dence’) factors. Moreover, the Sparks et al. study employed a cross-sectional design, with no data to test the eVects on subsequent behaviour, which formed the basis of Terry and colleagues’ distinction. Armitage and Conner (1999a, 1999b) also provide evidence to support a distinction between self-efficacy and ‘perceived control over behaviour’, utilizing measures that do not rely on perceived ease or diYculty. The present study sought meta-analytic evidence to support this position.

Behavioural intentions

The intention construct is central to both the TRA and TPB. Intentions are assumed to capture the motivational factors that in uence a behaviour and to indicate how hard people are willing to try or how much eVort they would exert to perform the behaviour (Ajzen, 1991, p. 181). In applications of the TRA/TPB, researchers have not always employed measures that clearly tap the intention construct. For example, Sheppard, Hartwick, and Warshaw’s (1988) review of the TRA argued for the need to consider both behavioural intentions and selfpredictions when predicting behaviour. Warshaw and Davis (1985) noted a number of diVerent ways in which intentions had been measured, and distinguished measures of behavioural intentions (e.g. ‘I intend to perform behaviour x’) from measures of self-predictions (e.g. ‘How likely is it that you will perform behavior x?’). Sheppard et al. (1988) went on to argue that self-predictions should provide better predictions of behaviour as they are likely to include a consideration of those factors which may facilitate or inhibit performance of a behaviour, as well as a consideration of the likely choice of other competing behaviours. Sheppard et al.’s meta-analysis supported this view: measures of self-predictions were found to have stronger relationships with behaviour (mean r = .57) than did behavioural intentions (mean r = .49), although attitude and subjective norm accounted for more of the variance in intentions (mean R = .73) than self-predictions (mean R = .61).
In the TPB, the PBC construct should tap perceptions of the factors that may facilitate or inhibit performance of behaviour. One might therefore expect little diVerence in the predictive power of intentions vs. self-predictions once PBC is taken into account. More speci. cally, the relationship between PBC and behavior should be stronger when intention (as opposed to self-prediction) measures are used, because intention measures do not take facilitating/inhibiting factors into account.
Beyond this, Bagozzi (1992) has suggested that attitudes may . rst be translated into desires (e.g. ‘I want to perform behaviour x’), which then develop into intentions to act, which direct action. From this perspective, one might expect that desires would inform intentions, upon which behavioural self-predictions are partly based. Given that desires take no account of facilitating/inhibiting factors, PBC should contribute more unique variance to the prediction of behaviour when measures of desires are employed than self-predictions. Congruent with the view that desires do not take account of facilitating/inhibiting in uences on behaviour, PBC should be more closely associated with self-predictions than with desires. On the other hand, intentions are held to mediate the relationship between desires and self-predictions, suggesting that eVects associated with intentions will fall between the desire and self-prediction . ndings. The present meta-analysis considers the role of intentions, desires and self-predictions in the context of the TPB.

Subjective norms

The normative component was the last addition to the TRA (Fishbein & Ajzen, 1975), and several authors have argued that it is the weakest component. For example, Sheppard et al. (1988) and Van den Putte’s (1991) meta-analyses of the TRA found that the subjective norm component was the weakest predictor of intentions (see also Godin & Kok, 1996). As a result, several authors have deliberately removed subjective norms from analysis (e.g. Sparks, Shepherd, Wieringa, & Zimmermanns, 1995). While these . ndings could merely re ect the lesser importance of normative factors as determinants of intentions in the behaviours studied, Tra.mow and Finlay (1996) suggest that this is unlikely. Across 30 behaviours, they found evidence to suggest a distinction between individuals whose actions are driven primarily by attitudes, and those whose actions are driven primarily by subjective norms.7 In addition, across several diVerent types of behaviour, variables thought to tap diVerent facets of normative conduct (e.g. descriptive and moral norms) have been found to be independently predictive of intentions (e.g. Beck & Ajzen, 1991; Conner, Martin, Silverdale, & Grogan, 1996; for a review see Conner & Armitage, 1998). The most likely explanation for poor performance of the subjective norm component lies in its measurement: many authors use single item measures, as opposed to more reliable multi-item scales (e.g. Nunnally, 1978). The present meta-analysis therefore considered type of measurement
as a moderator of subjective norm–intention correlations.

7See also Prislin and Kovrlija (1992) for an application of the TPB to high- and low self-monitors.

Aims

The aims of the present meta-analysis were . vefold: (1) to test the overall efficacy of the TPB; (2) to assess the predictive validity of the TPB in relation to observed 478 Christopher J. Armitage and Mark Conner
or self-reported behaviour; (3) to consider diVerences in the conceptualization of intentions, and to assess the evidence for discriminant validity between the constructs; (4) to examine the role of PBC as opposed to self-efficacy or ‘perceived control over behaviour’, and consider the proposed intention–PBC interaction; and
(5) to consider measurement adequacy as a moderator of the subjective norm–intention relationship, given that this construct has been found to be the weakest predictor in both the TRA and TPB.

Method
Selection of studies

Mullen (1989) presents several useful strategies for the retrieval of studies for meta-analysis. In terms
of the present study, the main approaches used were: ancestry and descendancy; abstracting services;
on-line computer searches; the ‘invisible college’; and browsing. In total, references to 161 journal
articles and book chapters testing the TPB (up to the end of 1997) were found.
The decision to include only published articles renders the present meta-analysis susceptible to
publication bias. The publication bias refers to the assumption that studies with signi. cant . ndings
are more likely to be submitted for publication. Several studies have examined this phenomenon (e.g.
Greenwald, 1975; Rosenthal, 1984; White, 1982), although . ndings are inconsistent. Related to this,
the ‘. le drawer problem’ (Rosenthal, 1979) refers to the possibility that all published articles are the
result of Type I errors, whereas all non-published (i.e. . le drawer) studies are the remaining 95%.
Reliability of the data included in the present study is assessed using Rosenthal’s (1984) fail-safe
number (the number of studies required to nullify the present . ndings). All relationships in the
present study exceed this tolerance level, unless otherwise stated.
Tables reporting meta-analytic data also include 2 values. These allow assessment of betweenstudy
variance (i.e. the variability of (in this case) eVect sizes around the mean presented). All 2
values in the present study indicate considerable variability around the mean, indicating that even
moderator analysis failed to reduce between-study variance to non-signi. cance. Used in conjunction
with Rosenthal’s fail-safe number, it is possible to assert that the present . ndings are robust (i.e.
require large numbers of additional studies to overturn them), but are subject to variability around the mean.

Characteristics of studies
The 161 articles contained 185 independent empirical tests of the TPB. Of these, 44 contained
prospective self-reported behaviour measures and 19 prospective measures of behaviour that were
independently rated or were objective (e.g. taken from records).8
Self-efficacy, PBC and ‘perceived control over behaviour’. The present meta-analysis distinguishes
between three types of PBC measure: self-efficacy, PBC and ‘perceived control over behaviour’.
Congruent with Armitage and Conner (1999a, 1999b), self-efficacy was de. ned as ‘con. dence in one’s
own ability to carry out a particular behaviour’; perceived control over behaviour was de. ned as
‘perceived controllability of behaviour’; and PBC was de. ned as the perceived ease or diYculty of
performing behaviour (Ajzen, 1991), and also included studies that utilized measures of both
self-efficacy and perceived control over behaviour in multiple-item scales.

8Note that only prospective measures of behaviour were included in the present meta-analysis. This is because a measure of behaviour taken contemporaneously with intention is actually a measure of past behaviour. Measures of past behaviour have been shown to contribute unique variance to the prediction of future behaviour, over and above TPB variables, introducing a possible confound (for reviews, see Conner & Armitage, 1998; Sutton, 1994).


Studies were coded as measuring self-efficacy if they included items such as: ‘I believe I have the
ability to . . .’; ‘To what extent do you see yourself as being capable of . . .’; ‘How con. dent are you
that you will be able to . . . ’; and ‘If it were entirely up to me, I am con. dent that I would be able
to . . .’ (N of studies = 28). ‘Perceived control over behaviour’ was coded when items such as
‘Whether or not I do x is entirely up to me’, ‘How much personal control do you feel you have over
. . .’, and ‘How much do you feel that whether you do x is beyond your control?’ were employed
(N = 7). Items that assessed perceived ease or diYculty were not included when coding for
self-efficacy or ‘perceived control over behaviour’. Where studies employed mixed measures (i.e. any
combination of the above or ‘easy–diYcult’ items) these were coded as ‘PBC’ (N = 101).
Desires, intentions and self-pred ictions. Desires, intentions and self-predictions were coded
according to the criteria discussed in Bagozzi (1992), Fishbein and Stasson (1990), Norman and Smith
(1995), Sheppard et al. (1988) and Warshaw and Davis (1985). Brie y, ‘desire’ was coded if studies
employed items such as ‘I want to perform behaviour x’; ‘self-prediction’ was coded for measures such
as ‘I will perform behaviour x’ or ‘How likely is it that you will perform behaviour x?’; and ‘intention’
was coded for studies that employed only measures such as ‘I intend to perform behaviour x’. Where
studies employed some combination of the above, these were coded as ‘mixed’ measures. We were
able to locate 88 studies that used mixed measures of behavioural intentions, 20 with measures of
intention, 40 with measures of self-prediction, and six of desire.
Subjective norms. Studies were also coded for measurements of the subjective norm component.
These fell into six categories: multiple-item scale (N = 32), single item (N = 52), general social
pressure multiplied by motivation to comply (N = 14), normative beliefs9 as direct predictors of
intention (N = 26), social support (N = 1) and unspeci. ed (N = 12).

Analyses
Analyses are based on bivariate correlations: where the appropriate statistics were not reported in the published article, the authors were contacted, and several have generously supplied copies of their correlation matrices. This allowed us to run additional analyses.
For the purpose of analysis, rs were converted to Fisher z scores, weighted by sample size (N 3), before a mean Fisher z was calculated (see Hedges & Olkin, 1985). The weighted mean Fisher zs were then converted back to rs, for the purpose of reporting the results. R2 change values (for eVects of PBC controlling for TRA variables) were converted to r, and combined in the same way as bivariate correlations.
Comparisons between correlation coeYcients were conducted using Cohen’s (1977) qs statistic, which evaluates diVerences in the magnitude of Fisher z. For samples with unequal Ns, a harmonic mean (n ) was used. Note, however, that this only provides an estimate of diVerences between magnitude of correlation coeYcients because the technique ignores dependencies between variables.
In the case of the TPB, most of the interesting comparisons involve diVerences between correlations that hold intention in common.
It has been noted that the meta-analyses of Godin and Kok (1996) and Sheppard et al. (1988) can be criticized for analysing groups of participants more than once. For example, in the Sheppard et al.
(1988) meta-analysis, two groups of participants from Warshaw and Davis (1985) were treated as independent tests of the TRA and included 18 times in the analysis. Clearly, this threatens the validity of meta-analysis. In order to avoid this, where studies examined more than one behaviour with one group of participants (e.g. Madden, Ellen, & Ajzen, 1992), the rs were converted to Fisher zs and meta-analysed in their own right before inclusion in the main data set.
Finally, because meta-analytic data tend to be based on large sample sizes, even the smallest correlations are likely to reach statistical signi. cance. Cohen (1992) presents a useful guide to interpreting the magnitude of eVect sizes: medium eVect sizes are de. ned as those that approximate the average eVect sizes across a variety of . elds of research. Small eVect sizes are ‘noticeably smaller than medium, but not so small as to be trivial’; large eVect sizes are ‘the same distance above medium as small was below it’ (Cohen, 1992, p. 156). These eVect size categories equate with correlations of .10 (small), .30 (medium), and .50 (large), and this provides a useful heuristic that acts as a standard of comparison.
 
9By ‘normative beliefs’, we are referring to a summed scale derived from the product referent beliefs and
motivations to comply.
 =====


Discussion
The present meta-analysis provides evidence supporting the use of the TPB for predicting intention and behaviour, although the prediction of self-reported behaviour is superior to observed behaviour. Moreover, there is some evidence for discriminant validity between desire, intention and self-prediction, and for a distinction between self-efficacy and perceived control over behaviour. Finally, subjective norm shows a reasonably strong relationship with intention when appropriately measured with multiple-item scales.

Overall findings
The present meta-analysis of the TPB compares favourably with previous metaanalyses.
The present study found R = .52 (R2 = .27) for the multiple correlation of intention and PBC with behaviour; previous meta-analyses have reported similar . ndings (range of R = .46–.58). Further, congruent with these studies, PBC was found to contribute uniquely to the prediction of behaviour, demonstrating the efficacy of the PBC construct. Similarly, the intention–behaviour correlation from the present meta-analysis is comparable with those of recent meta-analyses devoted to intention–behaviour relations. The intention–behaviour correlation in the present meta-analysis is r = .47. Randall and WolV (1994) report a corresponding relationship of .45 (98 studies), while Sheeran and Orbell (1998) reported a mean correlation of .44 (28 studies of condom use).



Table 5. Meta-analysis of single vs. multiple-item measures of subjective norms in regression with intentions

Measure
N of tests
ra
Fail-safe
N 2
Multiple items
32
.38
11,403
237***
Single item

52 .28 17,936

270***
Subjective norm Motivation to comply
14
.30
1195
45***
Social support
1
.25
[–]
 [–]
Normative beliefs
26
.38
9987
213***
Unspeci. ed
12
.45
4095
188***
***p<.001.
Note. aWeighted by sample size

Further support for the efficacy of the TPB over the TRA is provided by the multiple correlation of attitude, subjective norm and PBC with intention. The . ndings of the present meta-analysis are comparable with those of previous studies (Rs = .64–.71). More importantly, from the present meta-analysis, PBC adds—on average—6% to the prediction of intention, over and above attitude and subjective norm. Therefore, it would appear that PBC in uences behaviour directly and indirectly, independent of TRA variables, and therefore represents a useful addition to the TRA.

Self-report vs. objective behaviour
It is clear that many TPB studies do not employ prospective designs or measure behaviour. Where behaviour is measured, it is typically through self-report.
Congruent with Hessing et al. (1988), intention and PBC were better predictors of self-reported behaviour than observed behaviour. Clearly this is not a problem speci. c to the TRA/TPB, but provides indication of the wider debate within social psychology.
Potentially, however, this may simply reflect the fact that measurement correspondence
is typically maximized where subjective measures of behaviour are used (cf. Fishbein, 1980). For example, in a study of low-fat diet consumption, Armitage and Conner (1999a) reported a comparable discrepancy between self-reported ehaviour (e.g. ‘I ate a low-fat diet’) and a more objective assessment of behavior (validated measure of percentage of calories derived from fat). While the divergence between the two may represent a subjective–objective distinction, it may also re ect the fact that the subjective measure of behaviour directly mapped onto the prior measure of intention, whereas the objective measure could not. Sutton (1998) has suggested that showing participants the measure of behaviour on which they will later be assessed is one way of circumventing such problems. Researchers should be cognizant of the problems of self-report data and, wherever possible, take accurate multiple measures of actual  behaviour.

Desire, intention and self-pred iction
The present study provides some support for work proposing a distinction between intention, desire and self-prediction (e.g. Bagozzi, 1992). TPB variables were most closely associated with desires, although PBC contributed relatively little additional variance. In turn, desires were the weakest predictors of behaviour, with PBC contributing the most additional variance. PBC contributed most additional variance to the explanation of intention and self-prediction, and contributed least to prediction of behaviour when intention and self-prediction were statistically controlled. These . ndings can be accounted for by the fact that PBC takes account of factors that may facilitate or inhibit behaviour; under such circumstances, one would expect little diVerence between intention and self-prediction.
Overall, the present . ndings provide some support for Bagozzi’s (1992) position: intentions and self-predictions were superior predictors of behaviour than desires; attitudes, subjective norm and perceived control (i.e. self-efficacy, PBC or perceived control over behaviour) were the best predictors of desires. Thus, individuals may . rst translate their attitudes into desires, taking perceptions of social pressure and (to a lesser extent) control into account. However, these desires are weak direct predictors of behaviour, but may instead be mediated by intentions or selfpredictions,
or may co-determine behaviour with perceived control. Further evidence for this is reported in Bagozzi and Kimmel (1995), who showed that the impact of attitudes on intention was almost entirely mediated by desires. Future work is required to test the proposed causal relationships between these variables.

Perceptions of control
The present meta-analysis found diVerences between measures of PBC, self-efficacy and perceived control over behaviour. Self-efficacy and PBC were significantly more strongly correlated with both intention and behaviour than was perceived control over behaviour. Congruent with this, analysis of the proportion of additional variance explained indicated that the . ndings for perceived control over behaviour were both weak and unreliable. In general, self-efficacy accounted for the most additional variance in intention, and both PBC and self-efficacy accounted for equivalent proportions of variance in behaviour. The implication is that individuals form intentions that they are con. dent they can enact (i.e. those they perceive self-efficacy over), and that translation of intention into action may be facilitated both by self-efficacy and an assessment of more external factors tapped by PBC.
The analyses concerning the proportion of additional explained variance contributed by perceived control over behaviour were shown to be unreliable: more studies are required that more fully investigate this construct. Indeed, the possibility exists that the predicted diVerential eVects of self-efficacy and perceived control over behaviour may vary as a function of behaviour studied (see Armitage & Conner, 1999a, 1999b; Manstead & van Eekelen, 1998). However, where the data were reliable, perceived control over behaviour was signi. cantly more weakly related to intention and behaviour. The . ndings also suggest that self-efficacy and PBC are both useful predictors of intention and behaviour. While there is no clear evidence for which is to be preferred, self-efficacy is more clearly de. Ned and operationalized than is PBC (cf. Bandura, 1997), which consists of ‘mixed measures’ (see Method). Moreover, while self-efficacy and PBC account for equivalent proportions of the variance in behaviour, self-efficacy explains somewhat more of the variance in intention than does PBC. The implication is that self-efficacy should be the preferred measure of ‘perceived control’ within the TPB, but further research is required that more fully evaluates the impact of diVerent operationalizations of perceived control on intention and behaviour.
In addition, researchers have paid relatively little attention to precisely what PBC is tapping: up to the end of 1997, only 18 published studies reported control belief–PBC relationships. Given the role of PBC as a powerful determinant of both intention and behaviour, further exploration of both the nature and antecedents of the PBC construct is clearly required. For example, a recent study by Armitage and Conner (1999a) provides some evidence to suggest that control beliefs (as conceptualized by Ajzen, 1991) are the antecedents of self-efficacy, but correlate only weakly with perceived control over behaviour.

Subjective norms
Several researchers have argued that the subjective norm component of the TPB is inadequate and rarely predicts intention, and so have removed it from analysis (e.g. Sparks, Shepherd, Wieringa, & immermanns, 1995). The present meta-analysis provides some support for this view: subjective norm was the TPB component most weakly related to intention. However, when type of measure was used as a moderator, the poor performance of the subjective norm component was shown to be a function of measurement. Clearly, this component requires further empirical attention, and the present study points to measurement as its principal weakness, given that the majority of TPB studies have used single-item measures. Beyond this, a number of authors have argued that the way in which norms are conceptualized within the TRA/TPB framework fails to tap important facets of social in uence (e.g. Conner & Armitage, 1998; Terry, Hogg, & White, 1999).
Some researchers have suggested a reconceptualization of the mechanism by which normative pressure is exerted. Tra.mow and Finlay (1996) have argued that the weakness in the subjective norm component stems from a minority of individuals whose actions are driven primarily by perceived social pressure.
Although they provide some evidence to support this view (see also DeBono & Snyder, 1995), it seems unlikely that the majority of people’s behaviour is unaVected by social pressure. There is also evidence to suggest that alternative conceptualizations of norms exert independent eVects on intentions, controlling for subjective norms.
Subjective norm is operationalized as a global perception of social pressure either to comply with the wishes of others or not (Ajzen, 1991). However, social pressure is rarely so direct or explicit, leading a number of researchers to suggest alternative conceptualizations. For example, Terry and colleagues (e.g. Terry & Hogg, 1996; Terry et al., 1999; Terry, Hogg, & White, 2000; White et al., 1994) have drawn on Self-categorization and Social Identity Theories (see Hogg & Abrams, 1988; Turner, 1985). Speci. cally, Terry and colleagues have shown that identi. cation with a behaviourally relevant group moderates the eVects of group norm on intention (Terry & Hogg, 1996). In addition, they present some evidence to support a distinction between group- and subjective-norms (Terry et al., 1999; cf. Deutsch and Gerard’s (1995) distinction between informational and normative in uence).
Related work has investigated a number of diVerent types of norm. Cialdini, Kallgren, and Reno (1991) distinguish between personal, descriptive and injunctive norms. Injunctive norms map onto subjective norms. Personal norms have been operationalized as either self-identity or moral norms (see Conner & Armitage, 1998). A number of studies have shown that self-identity explains additional proportions of the variance in intentions over and above TPB variables (e.g. Armitage & Conner, 1999a, 1999b; Sparks & Shepherd, 1992; for a review, see Conner & Armitage, 1998). There is also evidence to support the inclusion of moral (e.g. Beck & Ajzen, 1991; Conner & Armitage, 1998) and descriptive (e.g. Conner et al., 1996) norms within the TPB. Further research is required to test the suficiency of such additional variables by testing them against adequate measures of subjective norm.

Conclusions
The present meta-analysis provides support for the efficacy of the TPB as a predictor of intentions and behaviour. Although prediction is superior for self-reported than observed behaviour, the TPB is still capable of explaining 20% of the variance in prospective measures of actual behaviour (i.e. a medium to large eVect size). The present . ndings therefore corroborate those of previous TPB meta-analyses, as well as expanding on some of the theoretical debate surrounding the model. The present study showed that PBC independently predicted intentions and behaviour in a wide number of domains. There was also evidence to suggest that measures of intention, self-prediction and desires possess discriminant validity, although only relatively weak evidence for the proposed self-efficacy–perceived control over behaviour distinction. Finally, work on additional normative variables (e.g. moral or descriptive norms) may increase the predictive power of the normative component of the model.
 




Tidak ada komentar:

Poskan Komentar