Skip to main content
PMC home page
NIHPA Author Manuscripts logoLink to NIHPA Author Manuscripts
. Author manuscript; available in PMC: 2008 Jun 17.
Published in final edited form as: Psychol Addict Behav. 2007 Dec;21(4):498–507. doi: 10.1037/0893-164X.21.4.498

The Theory of Planned Behavior as a Model of Heavy Episodic Drinking Among College Students

Susan E Collins 1, Kate B Carey 1
PMCID: PMC2430391  NIHMSID: NIHMS53832  PMID: 18072832

Abstract

This study provided a simultaneous, confirmatory test of the theory of planned behavior (TPB) in predicting heavy episodic drinking (HED) among college students. It was hypothesized that past HED, drinking attitudes, subjective norms and drinking refusal self-efficacy would predict intention, which would in turn predict future HED. Participants consisted of 131 college drinkers (63% female) who reported having engaged in HED in the previous two weeks. Participants were recruited and completed questionnaires within the context of a larger intervention study (see Collins & Carey, 2005). Latent factor structural equation modeling was used to test the ability of the TPB to predict HED. Chi-square tests and fit indices indicated good fit for the final structural models. Self-efficacy and attitudes but not subjective norms significantly predicted baseline intention, and intention and past HED predicted future HED. Contrary to hypotheses, however, a structural model excluding past HED provided a better fit than a model including it. Although further studies must be conducted before a definitive conclusion is reached, a TPB model excluding past behavior, which is arguably more parsimonious and theory driven, may provide better prediction of HED among college drinkers than a model including past behavior.

Keywords: college drinking, alcohol use, theory of planned behavior, heavy episodic drinking, college students

Ajzen’s (1988) theory of planned behavior (TPB) has been used widely to describe health behaviors over which one has volitional control (Godin & Kok, 1996). Research over the past decade has applied this theory to describe the psychological sequelae involved in college drinking. Most studies to date, however, have focused on contributions of individual model components instead of evaluating simultaneous model fit, and the results of some of these studies have not been entirely consistent with the predicted model (e.g., Wall, Hinson, & McKee, 1998). Moreover, the role of past behavior in the TPB has not been adequately defined and remains somewhat controversial (Ajzen, 2002c). Finally, although the theory has been applied to college drinking (Conner, Warren, Close, & Sparks, 1999; O’Callaghan, Chant, Callan, & Baglioni, 1997), it has been used less often to describe the occurrence of heavy episodic drinking (HED) among college students (see Johnston & White, 2003; Norman, Bennett, & Lewis, 1998; Wall et al., 1998). Because HED can be particularly harmful to college students (Clements, 1999; Hingson, Heeren, Zakocs, Kopstein, & Wechsler, 2002), it deserves further examination. To address these issues, this study used longitudinal, simultaneous and confirmatory model testing to examine the ability of the TPB to predict intention and HED among college drinkers.

Theory of Planned Behavior

According to the TPB, a certain set of motivational factors leads to intention to act in a certain way (Ajzen, 1991). Given the right opportunity, people will translate this intention into behavior. The role of intention in this model is believed to be two-fold: attitudes, subjective norms and perceived behavioral control should predict intention, and intention, in turn, should predict behavior. Several studies focusing on college drinkers have established the prediction of behavior by intention (Armitage, Norman, & Conner, 2002; Conner et al., 1999; O’Callaghan et al., 1997).

Attitudes refer to people’s evaluation of their own behavior. Attitudes concerning alcohol use appear to be important in concurrently predicting drinking quantity and frequency (Leigh, 1989). Further, longitudinal evidence has shown that positive attitudes towards alcohol use are positively correlated with future alcohol use among college drinkers (Stacy, Bentler, & Flay, 1994).

The subjective norm is a target person’s perception of others’ evaluations of the target person performing a behavior (Ajzen, 1991). This construct may be broken down into two components: (a) perception of others’ evaluations (also referred to as normative beliefs) and (b) the importance of the others’ opinions to the target person (representing motivation to comply with perceived norms). In other words, a person is more likely to perform a behavior if others, whose opinions the person values, approve of the person performing the behavior. A conceptualization of HED among college students that includes subjective norms is important because adolescents and college students report being influenced by their peers more than adults (Beck & Treiman, 1996; Perkins, 2002). Injunctive norms, as commonly conceptualized in the college drinking literature, resemble the evaluative component of subjective norms in the TPB, and ample evidence has linked not only descriptive norms (perceptions of how others drink) but also injunctive norms (perceived peer approval of drinking) to college drinking (Larimer, Turner, Mallett, & Geisner, 2004; Perkins, 2003).

Perceived behavioral control, or the perceived ease or difficulty of performing a behavior, is believed to both indirectly (through its association with intentions) and directly influence behavior (Ajzen, 1988, 1991). Both self-efficacy, which has been defined as one’s perceived control over a certain behavior in a specific situation (Bandura, 1977; Marlatt & Gordon, 1985), and controllability beliefs, or beliefs that the performance of a behavior depends on the individual alone, make up the concept of perceived behavioral control (Ajzen, 2002b). Perceived behavioral control and drinking refusal self-efficacy have been supported in the literature as strong predictors of college drinking, such that higher drinking refusal self-efficacy and perceived behavioral control predict lower drinking quantity, frequency and experience of problems (Johnston & White, 2003; Young, Connor, Ricciardelli, & Saunders, 2006).

TPB and Past Drinking Behavior

Including past behavior in the TPB model may improve the prediction of future behavior, but this modification has engendered controversy. The debate regarding the role of past behavior in TPB models typically centers on challenges defining the past behavior construct and the relative importance of including a significant yet perhaps atheoretical predictor of future behavior. Contrasting arguments for and against the inclusion of past behavior in TPB models are presented below.

First, it has been asserted that consistent performance of a behavior represents habit, which may be a relevant predictor of various types of behavior (Ouellette & Wood, 1998). The results of a meta-analysis of 64 studies, which evaluated past behavior prediction of future behavior, indicated that past behavior was a stronger predictor of behaviors repeated on a weekly basis in stable, conducive contexts than intention (r = .45 and r = .27, ps < .001; Ouellette & Wood, 1998). For behaviors that were not performed regularly, such as voting or getting a flu shot, past behavior was a weak predictor of future behavior relative to intention (r = .12 and r = .82, ps < .001). Another study found that past or habitual behavior significantly predicted future behavior even after controlling for a range of other cognitive components, including attitude accessibility, self-concept, intent, attitudes, subjective norms and perceived behavioral control (Ouellette, 1996). These authors have argued that, taken together, this evidence supports past behavior as a habit construct that may be helpful in understanding longitudinal behavior patterns.

However, it has been argued that the habitual nature of particular behaviors may not be assumed simply because they are regularly performed or because other variables do not account for the variance in future behavior (Ajzen, 2002c). Instead, it has been proposed that a direct measure of habit should be used to establish its role in behavior prediction (Ajzen, 2002c; Verplanken, Aarts, van Knippenberg, & van Knippenberg, 1994). Measures of habit, however, have been difficult to develop because it is unclear whether people can accurately judge the habituality of their own behavior (Ouellette & Wood, 1998). Difficulties defining and directly testing past behavior as a measure of habit have thus complicated the interpretation of past behavior as a predictor in TPB models.

Second, proponents of the inclusion of past behavior have asserted that it is a consistently strong predictor of future behavior that should be accounted for (e.g., Stacy et al., 1994). In the interest of correct model specification, or the inclusion of all variables known to be important in predicting a dependent variable, it can be argued that past behavior should be included in models of TPB. On the other hand, it has been asserted that past behavior’s predictive strength may be due to a methodological artifact: common method variance, or the correlation between identical indicators repeatedly assessed using the same method (Ajzen, 2002c). To test the effect of common method variance on the prediction of past on future behavior, one study employed both self-report and observation methods in the alcohol-use assessment (Conner et al., 1999). Contrary to hypotheses, past drinking explained variance above and beyond the TPB variables even when common method variance was eliminated. This study, however, did not use a simultaneous confirmatory test of the TPB and did not focus on HED specifically. Considering these differing perspectives and findings, an empirical comparison of competing theoretical models (inclusion vs. exclusion of past HED) would be a helpful addition to the literature.

TPB and Heavy Episodic Drinking

Three studies to date have examined the influence of attitudes, subjective norms and perceived behavioral control on intention to engage in HED (also, to “drink too much” and “binge drinking” in Johnston & White, 2003; Norman et al., 1998; Wall et al., 1998). Wall et al. (1998) used stepwise or statistical regressions (conducted separately by gender) to test the ability of concurrently assessed TPB variables to predict intention and HED. For women, attitudes, perceived behavioral control and sociability expectancies predicted intention, whereas attitudes, perceived behavioral control and subjective norms predicted intentions in men. Intentions, perceived behavioral control, and assertiveness expectancies predicted HED for women, whereas intentions and expectancies of higher sexual functioning predicted HED among men.

Another study used partial correlations and regressions to test the concurrent associations among attitudes, perceived behavioral control, subjective norms, beliefs (i.e., behavioral, normative and control beliefs) and HED (Norman et al., 1998). In this study, gender explained 9% and TPB variables an additional 29% of the variance in HED. Of the TPB variables, only perceived behavioral control and positive control beliefs reached significance. Thus, in spite of its high R2, this study provided only partial support for the concurrent prediction of HED by TPB variables.

Finally, a more recent study used a longitudinal model to test the prediction of intentions and HED by TPB variables (Johnston & White, 2003). The results of the multiple regression indicated that attitudes, subjective norms and self-efficacy significantly predicted 69% of the variance in intention to engage in HED. Additionally, intention predicted HED assessed two weeks later and accounted for 51% of the variance.

In exploring the ability of TPB variables to predict HED among college students, the above studies evinced similar methodological limitations. Perhaps the most striking limitation in the first two studies was the concurrent assessment of the TPB variables and HED. Because the theory posits that TPB variables at time 1 predict time 2 behavior, a simple prospective design would provide a better test of model prediction. Further, all three studies used regression models to test the contributions of TPB variables to the prediction of intentions and HED but did not simultaneously test the TPB model as a whole. Because the TPB is clearly defined and well-researched (e.g., Godin & Kok, 1996), a confirmatory and simultaneous testing method (e.g., latent factor structural equation modeling) would provide a more appropriate test of the model. Finally, despite the fact that research supports a strong correlation between past and future drinking behavior (Conner et al., 1999; McMillan & Conner, 2003; O’Callaghan et al., 1997), none of the studies mentioned above accounted for this effect.

Proposed Study

The TPB is a potentially useful conceptual framework for understanding drinking among college students. The studies that have tested the prediction of HED by TPB model components have not provided methodologically optimal tests of the theory. Our goal in this study was therefore to improve on the methodology used in previous studies by using a longitudinal design, a simultaneous and confirmatory rather than exploratory testing method, and error-free latent factors (i.e., latent factors from which measurement error is estimated and removed). Further, in this study, we compared models including and excluding the potential contribution of past HED to the prediction of future HED in the context of the TPB. In doing so, this study aimed to provide a better test of whether past HED, attitudes, self-efficacy and subjective norms predict behavioral intention and future HED among college students.1

Hypotheses

The first set of hypothesized models included a past HED factor (time 1 HED). As shown in Figure 1, it was predicted that time 1 HED, subjective norms, drinking refusal self-efficacy (referred to from this point as “self-efficacy”) and attitudes would predict time 1 intention. In turn, time 1 HED and intention would predict future HED (time 2 HED). In order to test Ajzen’s (1991) assertion that self-efficacy both indirectly (via intention) and directly predicts behavior at time 2, additional, nested models tested whether direct effects of self-efficacy on time 2 HED would improve model fit over a reduced model in which self-efficacy served as an indirect explanatory variable via intention (represented by dotted line in Figure 1). The second set of models excluded time 1 HED. It was hypothesized that self-efficacy, attitudes and subjective norms would predict intention, which would, in turn, predict time 2 HED (see Figure 2). It was further predicted that the model containing time 1 HED would be a better-fitting model than the model excluding time 1 HED.

Figure 1.

Figure 1

Open in a new tab

Top panel: Hypothesized model including time 1 drinking. + = hypothesized positive association. − = hypothesized negative association. D = disturbance (unexplained variance in endogenous latent factors). T1 HED = time 1 heavy episodic drinking. ATT = time 1 attitudes. NORMS = time 1 subjective norms. SE = time 1 self-efficacy. T2 HED = time 2 heavy episodic drinking. Double-headed arrows imply nondirectional covariance between the variables. The dotted line represents the nested model including a direct effect for SE on T2 HED. For ease of interpretation, measured indicators are not shown. Bottom panel: Final model including time 1 drinking. * = significance at the .05 level.

Figure 2.

Figure 2

Open in a new tab

Top panel: Hypothesized model excluding time 1 drinking. + = hypothesized positive association. − = hypothesized negative association. D = disturbance (unexplained variance in endogenous latent factors). ATT = time 1 attitudes. NORMS = time 1 subjective norms. SE = time 1 self-efficacy. T2 HED = time 2 heavy episodic drinking. Double-headed arrows imply nondirectional covariance between the variables. The dotted line represents the nested model including a direct effect for SE on T2 HED. For ease of interpretation, measured indicators are not shown. Bottom panel: Final model excluding time 1 drinking. * = significance at the .05 level.

Method

Participants

Participants (N = 131) consisted of students (63% female, n = 83) enrolled in introductory psychology classes at Syracuse University. All participants were at least 18 years of age (M = 18.95, SD = 2.67). Most participants were freshmen or sophomores in college (93%, n = 122) and lived on campus (90%, n = 118). Hispanic/Latino students represented 2%, White students 92%, Asian students 5%, and self-described multiracial students 1% of the sample.

Measures

Demographic information

The Personal Information Questionnaire was used to assess participants’ age, gender, year in college, ethnicity, residence, and membership in an on-campus Greek organization.

Intention factor

Intentions were measured by the Behavioral Intentions Questionnaire (BIQ; Neal & Carey, 2004). The BIQ consists of seven items asking participants to rate their intentions of increasing and decreasing their drinking quantity/frequency, peak drinking and heavy, episodic drinking in the next two weeks. Responses were scored on a 1 to 6 scale, where 1 = definitely will not do and 6 = definitely will do. Questions 2, 4, and 6, which assessed the probability of decreased alcohol use, were reverse scored. The BIQ was supplemented with an item that asked participants to use the 6-point scale described above to rate the probability of their “drinking to get drunk” in the next two weeks.

At the suggestion of the scale’s author (D. J. Neal, personal communication, November 7, 2005), the seventh BIQ item was used separately from its original scale. An exploratory principal components analysis was conducted on the modified, six-item scale and indicated a one-factor solution (Eigenvalue = 3.82), which accounted for 64% of the variance and evinced acceptable internal consistency (α = .88). In the analyses, three indicators represented time 1 intention: a) the sum of the six BIQ items, b) participants’ intention to engage in a heavy-drinking episode (as defined by Wechsler et al., 2002) in the next 2 weeks, and c) participants’ intention to “drink to get drunk” in the next 2 weeks. The internal consistency of the three indicators was acceptable (α = .83).

Attitudes factor

The two attitudes indicators in the current study were selected from the Global Attitude Scale (AS; Simons and Carey, 1998). The AS has previously evinced good reliability (α ≥.91) and discriminant validity (Simons & Carey, 1998). Participants rated their “overall opinions” about “drinking to get drunk” along an unnumbered, nine-point scale framed by two opposing word pairs on either end. The two global attitude word pairs used in the current study were like/dislike and desirable/undesirable. In this study, higher attitudes scores represented more positive attitudes towards drinking. These items evinced good internal consistency (α = .87).

Subjective norms factor

Two items from the Subjective Norms Questionnaire (labeled SNQ; based on guidelines from Ajzen, 2002a) served as indicators of the subjective norms latent factor. Participants reported how much “an average American college student” and their “closest friend” would approve or disapprove of their “drinking to get drunk” on a five-point Likert scale, where 1 = highly disapprove and 5 = highly approve. Next, participants rated the importance of these groups’ opinions to them on a scale ranging from 1 (highly unimportant) to 5 (highly important). Subjective norm indicators were ascertained for each group by multiplying the approval of the target group by the participants’ report of the importance of the target group’s opinion (Ajzen, 1991, 2002a). The “average American college student” and “closest friend” indictors evinced good reliability (α = .91).

Self-efficacy factor

Total scores for the social pressure and emotional relief scales of the Drinking Refusal Self-efficacy Questionnaire (DRSEQ; adapted from Young, Tian, & Crook, 1991) served as two of the three indicators of the self-efficacy factor. On a six-point Likert scale, participants indicated their confidence they could resist “drinking to get drunk” in each of the hypothetical situations presented. Good internal consistency (α = .87 –.94) and concurrent and discriminant validity have been established for this measure (Baldwin, Tian, & Young, 1993; Young et al., 1991). In this study, summary scores for the social pressure and emotional relief scales were formed, such that higher scores indicated higher self-efficacy to resist “drinking to get drunk” (α = 88 and .93, respectively).

For the third indicator of this factor, participants answered the question, “How confident are you that, if you wanted to, you could cut down on your drinking?” Responses were made on a 7-point Likert scale, where 1 =not at all confident and 7 =very confident. Pilot testing indicated that higher scores on this item were significantly associated with lower one-week quantity and fewer drinks consumed on the heaviest drinking day (r =−.33 and −.34, respectively, ps < .05). The alpha coefficient for all three indicators indicated an internal consistency of α = .65.

HED factors

Two, open-ended quantity and frequency items measured by the F-Q questionnaire (adapted from Collins, Carey, & Sliwinski, 2002) served as the indicators for time 1 and 2 HED: number of drinks consumed during peak drinking occasion in the past 2 weeks and number of heavy-drinking episodes. Heavy-drinking episodes were assessed using two gender-specific items and were defined as having consumed 5 or more drinks on one occasion for men and 4 or more for women (Wechsler et al., 2002). The internal consistencies of the time 1 and 2 HED indicators were good (α = .78 and .82, respectively).

Procedure

College students (N = 234), who signed up to participate in an alcohol-use study, gave written, informed consent and filled out all questionnaires listed in the Measures section as part of a larger questionnaire battery (see Collins & Carey, 2005). Participants who reported having experienced a heavy drinking episode in the past 2 weeks (N = 131) were randomized to receive no treatment or a minimal intervention involving a single session decisional balance exercise (see Collins & Carey, 2005). Approximately 2 weeks following the interventions, all participants attended a posttest questionnaire session. Participants were given course credit for their participation.

Data Analysis Plan

The main hypotheses in this study were tested with a series of latent factor structural equation models (SEMs) using the EQS 6.1 program (Bentler, 2004). This type of analysis allows for simultaneous, confirmatory model testing, which was deemed appropriate for evaluating an established model like the TPB. All models except the MIMIC model described below made use of the full information maximum likelihood method to estimate parameters and standard errors for the entire data set. Two types of model fit assessment were used in the current study: a) descriptive goodness-of-fit indices (CFI ≥ .95 and RMSEA ≤ .06 were interpreted as indications of close model fit; Hu & Bentler, 1999) and b) exact model testing using the model chi-square. Traditionally, “exact fit” is determined when the null hypothesis is accepted using the model chi-square test (Hayduk & Glaser, 2000). Although exact fit is the recognized term for meeting this criterion, it does not imply “perfect” model fit and is not interpreted as such in this manuscript.

The SEM analyses were conducted according to the two-step modeling approach (Anderson & Gerbing, 1988; Pentz & Chou, 1994). In the first step, saturated measurement models, which are essentially confirmatory factor analyses, tested the relationships of the measured variables to their hypothesized latent constructs. After the appropriateness of the measurement model was established, hypothesized relationships among the latent factors were tested in the second step.

Results

Exploratory Data Analysis

At the baseline assessment, participants reported having consumed an average of 28.24 (SD = 25.39) drinks and having experienced 2.62 (SD = 2.23) heavy-drinking episodes in the past 2 weeks. On their peak drinking occasion, participants reported having consumed 8.95 drinks (SD = 5.77).

The HED variables (i.e., peak drinks and heavy-drinking episodes) exhibited positively skewed distributions and were therefore transformed using a square-root transformation for the main analyses. Other variables (i.e., the subjective norm variables, the emotional drinking scale of the DRSEQ and two intentions items) did not exhibit normal distributions but could not be successfully transformed. For this reason, only robust statistics (i.e., Santorra-Bentler T and other tests using robust standard errors) are reported (Bentler, 2004).

As shown in Table 1, bivariate Spearman correlations between the measured variables provided further support for the construct validity of the proposed factors. Although not reported here, all preliminary independence SEMs, which tested the hypothesis that the variables were uncorrelated, were easily rejected (all ps < .001).

Table 1.

Bivariate Spearman’s Rho Correlations Between Model Indicators

Measure 1 2 3 4 5 6 7 8 9 10 11 12 13
1. Int: Heavy- drinking episode --
2. Int: drinking to get drunk .72b --
3. Int: overall .45b .49b --
4. Attitude: like/dislike .30b .41b .22a --
5. Attitude: desirable/undesirable .32b .46b .28b .76b --
6. SE: social −.33b −.44b −.24b −.40b −.38b --
7. SE: emotional −.24b −.32b −.15 −.22a −.29b .65b --
8. SE: general −.04 −.11 −.07 −.03 −.11 .18a .24b --
9. SN: college student .06 .24b .01 .27b .23b −.33b −.21a .003 --
10. SN: closest friend .23b .36b .09 .37 .35 −.27b −.13 .02 .48b --
11. Time1: Peak drinks .42b .30b .04 .15 .16 −.11 −.04 −16 −04 −.02 --
12. Time 1: Heavy-drinking episode .35b .22a .09 .22a .32b −.24b −.18a −.11 .06 .11 .49b --
13. Time 2: Peak drinks .41b .35b .22a .23b .23a −.15 −.20a −.17 .04 .12 .50b .27b --
14. Time 2: heavy-drinking episode .33b .38b .20a .38b .39b −.29b −.23b −.18a .10 .17 .38b .37b .6b
Open in a new tab

Notes. Int=intention. SE=self-efficacy. SN=subjective norms.

a

correlation significant at the .05 level.

b

correlation significant at the .01 level.

Models Including Time 1 HED

Evaluation of the measurement models

The baseline measurement model included the following 5 latent factors (measured indicators in parentheses): attitudes (AS-dislike/like, AS-desirable/undesirable), subjective norms (SNQ-college student, SNQ-closest friend), self-efficacy (DRSEQ-social, DRSEQ-emotional, general self-efficacy), time 1 HED (peak drinks, frequency of heavy-drinking episodes), and intentions (BIQ, intention to “drink to get drunk”, intention to engage in a heavy-drinking episode). All latent factors were initially allowed to covary. To ensure model identification, equality constraints were placed on the error variances of the indicators of subjective norms, attitudes and time 1 HED, respectively.

The first model tested was the baseline measurement model. Although fit indices indicated good fit of the hypothesized model, CFI = .97, RMSEA = .05, the robust scaled chi-square test did not indicate exact fit, T(N = 131, df = 47) = 65.38, p = .04. To improve model fit, the multivariate Wald test was examined, and accordingly, a nonsignificant covariance path, ϕF1F3 (time 1 drinking, subjective norms), was dropped from the model. The resulting hypothesized model evinced adequate fit, T(N = 131, df = 48) = 65.68, p = .05, CFI = .97, RMSEA = .05.

The next step confirmed the appropriateness of the longitudinal measurement model. Except for the addition of the time 2 HED factor, the measured indicators and latent factors were identical to those in the baseline model. When using a longitudinal model, tests of measurement invariance are necessary to ensure that factors represent the same constructs across time and may be consistently interpreted (Grouzet, Otis, & Pelletier, 2006; Little, 1997). In accordance with these standards, the final measurement model fulfilled the criteria for configural, metric and partial scalar invariance (see Thompson & Green, 2006) and exhibited good model fit, T(N = 131, df = 66) = 81.03, p = .10, CFI = .98, RMSEA = .04.

MIMIC model

Because the second data collection period involved a “treated” sample, a preliminary model was run to test the appropriateness of collapsing across treatment conditions. It was hypothesized that no group effects would be found on time 2 HED, as observed in a previous study (Collins & Carey, 2005). To test this hypothesis, a multiple indicators and multiple causes (MIMIC) model, which is the SEM equivalent of multiple regression with dichotomous predictors (Hancock, 2004), was conducted. MIMIC models allow for the inclusion of one or more dummy-coded variables to test population differences on a specific dependent variable within an SEM model. This procedure was deemed appropriate because a) the sample size in the current study was too small to conduct a structured means analysis, b) any potential group differences would occur late in the recursive model and would thus not influence previous model factors, and c) there was no indication that the groups had different covariance structures (Hancock, 2001).

The MIMIC model examined the structural relationships among the time 1 latent factors (attitudes, subjective norms, time 1 HED, self-efficacy and intention), the 2 dummy-coded group assignment variables (intervention effects) and the time 2 HED factor. In the case that group was not a significant predictor of time 2 HED, the groups would be collapsed for the longitudinal models. Although this analysis would not rule out group x TPB variable interactions, it would rule out the more plausible and critical intervention effects on time 2 HED. The covariance structure was analyzed using the maximum likelihood estimation method. No means were estimated for this preliminary model.

The hypothesized model indicated close but not exact fit, T(N = 120, df = 94) = 122.07, p = .03, CFI = .95, RMSEA = .05. The dummy-coded group variables did not predict time 2 HED (written motivational intervention vs. no-contact control, γ = −.05; in-person motivational intervention vs. no-contact control, γ = −.02; ps > .05). Due to the lack of association between the groups and time 2 HED, this model was rejected, and the groups were collapsed for the main outcome analyses.

Evaluation of the structural models

The first structural model tested the effects of attitudes, self-efficacy, time 1 HED and subjective norms on intention, and intention and time 1 HED on time 2 HED (see Figure 1 for hypothesized model). The hypothesized model test indicated exact fit, T(N = 131, df = 70) = 86.73, p = .09, CFI = .98, RMSEA = .04. As shown in the bottom panel of Figure 1, significant direct effects on intention were found, including attitudes (γF5F2 = .23) and self-efficacy (γF5F4 = −.32). These variables accounted for 47% of the variance in intention. Further, as predicted, time 1 intention and time 1 HED significantly predicted time 2 HED (βF6F5 =.26 and γF6F1 = .38, respectively) and accounted for 29% of the variance.

A further nested model was tested to examine the hypothesis that self-efficacy may also have had a direct effect on time 2 HED (represented by the dotted line in Figure 1). The hypothesized model met the criterion for exact fit, T(N = 131, df = 69) = 84.66, p = .10, CFI = .98, RMSEA = .04. However, the direct effect of self-efficacy on time 2 HED was nonsignificant (γF6F4 = −.17, p > .05). Further, a chi-square difference test Td, which was corrected for use with the scaled robust chi-square statistic (Satorra & Bentler, 2001), indicated that the model including the direct effect did not significantly improve model fit, Td (1) = 1.95, p > .10. Finally, the multivariate Lagrange Multiplier (LM) test, which is a modifier index indicating possibilities for improving model fit, indicated a nonsignificant improvement in fit with the addition of the direct self-efficacy effect, χ2(N = 131, df = 70) = 2.34, p = .13. Based on these tests, this model was rejected.

Models Excluding Time 1 HED

The next set of models tested the TPB excluding time 1 HED (see Figure 2 for the hypothesized model). Tests of the saturated measurement model indicated exact fit, T(N = 131, df = 46) = 47.93, p = .39, CFI = 1.00, RMSEA = .01. No post hoc modifications were conducted. Structural model tests indicated that the hypothesized model evinced exact fit, T(N = 131, df = 49) = 54.91, p = .26, CFI =.99, RMSEA = .03. As shown in the bottom panel of Figure 2, the path coefficients were similar to those found in the previous model including time 1 HED. Lower self-efficacy and more positive attitudes regarding HED significantly predicted higher intention and accounted for 45% of the variance. Intention, in turn, predicted greater levels of time 2 HED and accounted for 21% of the variance.

A further nested model tested the addition of a direct effect of self-efficacy on time 2 HED. The hypothesized model evinced good fit indices, CFI =.99, RMSEA = .03, and met the criterion for exact fit, T(N = 131, df = 48) = 54.17, p = .25. However, the direct effect of self-efficacy on time 2 HED was nonsignificant (γ F6F4 = −.14, p > .05). Further, the scaled chi-square difference test indicated that the larger model did not significantly improve model fit, Td(1) = .80, p > .25. The LM test for the original model also indicated a nonsignificant improvement in fit with the addition of the direct effect, χ2(N = 131, df = 49) = .98, p = .32. Based on these tests, the model including the direct self-efficacy effect was rejected.

Model Comparison: Inclusion and Exclusion of Time 1 HED

The Akaike Information Criterion (AIC; Akaike, 1974) and Bayesian Information Criterion (BIC; Schwarz, 1978) were used to compare the structural models including and excluding time 1 HED. AIC and BIC values may be used to compare the goodness-of-fit of nonnested models and take both parsimony and descriptive accuracy into account (Wagenmakers & Farrell, 2004). Because AIC and BIC values contain arbitrary constants and are difficult to interpret alone, the raw AICc (AIC corrected for small sample sizes; Burnam & Anderson, 2004) and BIC scores were converted into weighted scores (i.e., wAICc and wBIC; Wagenmakers & Farrell, 2004) and compared (see Table 2). The weighting procedure normalizes the values and results in an estimate of the probability that a particular model provides the “best fit” (Wagenmakers & Farrell, 2004). Both wAICc and wBIC values indicated probabilities of at least .99 that the model excluding time 1 HED provided a better fit than the model including time 1 HED.

Table 2.

Nonnested Model Comparisons

Model AICc ΔAICc wAICc BIC ΔBIC wBIC
1 25.926 25.017 ≈0 60.879 13.07 .001
2 0.909 0 ≈1 47.809 0 .999
Open in a new tab

Notes. Model 1= structural model including time 1 HED. Model 2= Stuctural model excluding time 1 HED. AICc= Akaike Information Criterion corrected for smaller sample sizes (Burnham and Anderson, 2004). ΔAICc = AICci - AICcminimum or the difference between the AICc. of the ith model and the lower AICc. wAICc = rounded weighted AICc or the probability that the ith model is the better fitting model. BIC= Bayes Information Criterion. ΔBIC = BICi –BICminimum. wBIC = rounded weighted BIC or the probability that the ith model is the better fitting model.

Discussion

In the current study, we explored the ability of the TPB to predict intention and HED among college students. In all models, results indicated that attitudes and drinking refusal self-efficacy predicted intention. Further, intention and time 1 HED predicted time 2 HED. This study therefore provided some support for the TPB in predicting intention and HED among college drinkers.

Despite its significant correlation with intent in the baseline and longitudinal measurement models (rs = .48 and .45, respectively, ps < .05), the subjective norm factor did not significantly predict intention in the structural models and was thus the only aspect of the TPB that was not confirmed in this study. This finding was somewhat surprising in light of previous research that has found subjective norms to be a predictor of HED among college students (e.g., Johnston & White, 2003). It is possible that the significant moderate correlation of the subjective norms factor with attitudes and self-efficacy suppressed the subjective norm effects. On the other hand, this pattern of findings, in which all variables except subjective norms contribute to the model, has been observed in previous TPB research (Ajzen, 1991). It may be that more proximal, intrapersonally generated components, such as attitudes and self-efficacy, have a more direct and salient affect on intention to engage in HED than more distal and interpersonally dependent components, such as subjective norms.

The current study also addressed a controversial issue in the TPB literature: the relative merits of including and excluding past behavior in the prediction of future behavior. Initial findings supported the inclusion of past HED by evincing good overall model fit and a significant path from past to future HED. Additionally, analyses indicated a larger effect for past HED than for intention on future HED, which replicated findings reported in a previous meta-analysis (Ouellette & Wood, 1998). According to Ouellette and Wood (1998), this pattern of findings would suggest that HED represented habitual behavior, particularly since HED in this sample was reportedly repeated on a weekly basis (Ouellette & Wood, 1998). Another potential reason for the weaker intention effect is that intention may have simply represented participants’ prediction of their future behavior based on their previous behavior (Ouellette & Wood, 1998).

On the other hand, intention was a significant predictor of future HED, and cognitive factors—not past behavior—predicted intention. Specifically, stronger intentions to engage in HED were predicted by more positive attitudes and lower levels of self-efficacy regarding HED. Thus, as a group, heavier drinking students acknowledged feeling relatively positively about HED, yet indicated that they were not fully in control of their drinking. The fact that this discrepancy could be identified and verbalized indicates a role for cognitive factors in predicting HED. Further, 95% of college students included in the current study were not of legal drinking age, which presents an external barrier to the habitual use of alcohol. To minimize this barrier to drinking, students must, to a greater or lesser extent, plan their HED. For example, students have to frequent the “right” parties, bars, stores, and associate with the “right” peers to ensure access to alcohol. Taken together, the findings for cognitive factors in this study indicate that more than automatic habitual behavior is at play in predicting HED among college students.

This point may also help explain the next finding, which was contrary to hypotheses and to initial model findings: the model excluding past behavior provided comparatively better statistical fit to the data than the model including past behavior. This finding corroborates Ajzen’s (2002c) assertions by suggesting that the more parsimonious and theory-driven model better predicted future behavior. Comparing the R2 statistics from the two models, the addition of past behavior added only 2% explained variance to the baseline model and 8% to the longitudinal model. Further, it is difficult to attribute the effect of past drinking on intention and future drinking to habitual behavior without having included a direct measure of habit (Ajzen, 2002c). Taken together, these findings indicate that the model including past behavior was, in this study, neither statistically nor theoretically superior to a model comprised solely of cognitive TPB predictors.

Considering the preliminary and somewhat conflicting findings of this study, however, these results should be replicated before a definitive conclusion is reached as to the potential role for past behavior in the TPB model of college drinking. Further, the current study sample consisted of college drinkers endorsing relatively high levels of heavy episodic drinking, and thus these findings may not be applicable to college drinkers in general. In order to determine the generalizability of these results, this type of study could be replicated in college drinkers with more variability in drinking habits.

Some limitations of the current study deserve mention. First, perceived behavioral control was represented by a unitary self-efficacy factor. Controllability beliefs, which are hypothesized to be an additional component of perceived behavioral control, were not directly measured in this study. It may therefore be argued that the self-efficacy factor included in its place did not fully represent participants’ perceived behavioral control. However, the self-efficacy factor was highly predictive of intention, which is consistent with the findings of other studies involving substance misuse (Godin & Kok, 1996). Further, a number of other studies including both components of perceived behavioral control found self-efficacy to be a significant predictor of intention, whereas controllability had no significant effect (Armitage & Conner, 1999a, 1999b; Terry & O’Leary, 1995). Future studies should, however, assess both self-efficacy and beliefs about controllability to further test their relative contributions to the model.

Next, the small sample size (N = 131) did not ensure adequate power to optimally test the model. In fact, power analyses based on RMSEA estimates indicated that the structural models including and excluding time 1 HED would have required 200 and 248 participants, respectively, to achieve a power of .80 (Kim, 2005). The fact that the TPB models tested provided so-called “exact fit” to the current data may be interpreted in two ways. First, it indicated evidence of model robustness in predicting HED in a small sample of college students. On the other hand, exact chi-square testing can “benefit” from smaller sample sizes because the associated lower power is more likely to produce nonsignificant effects. Thus, despite these encouraging findings, future studies using SEM to replicate these results should clearly aim for larger sample sizes.

Finally, the fact that this study was conducted using data from a randomized study complicates the prediction of time 2 HED. Specifically, participants underwent a brief intervention for at-risk drinking between times 1 and 2. The preliminary MIMIC model confirmed the Collins and Carey (2005) findings that there were no significant group differences on participants’ time 2 drinking. However, this analysis could not rule out group x TPB variable interactions or undetected differences in covariance structures.

Despite its limitations, this study indicated overall support for the TPB in predicting HED among college drinkers. It also provided the first simultaneous, confirmatory test of the TPB in predicting HED among college students and empirically compared TPB models including and excluding the influence of past behavior. Future studies may replicate these findings using a larger sample size to ensure greater power and may further probe the importance of subjective norms and past HED in the TPB model as applied to college drinkers.

Acknowledgments

This research was supported by a Creative Research Grant from the College of Arts and Sciences, Syracuse University to Susan E. Collins and by NIAAA grants R01-AA12518 and K02-AA15574 to Kate B. Carey.

Biographies

Author Note Susan E. Collins, Center for Health and Behavior, Syracuse University; Kate B. Carey, Center for Health and Behavior, Syracuse University.

Susan E. Collins is now at the Smoking Cessation Research Group, University Hospital Tübingen.

Footnotes

1

In the current study, self-efficacy, but not beliefs about controllability, was directly assessed. Thus, perceived behavioral control was operationalized as a unitary self-efficacy factor in this study and will be referred to as such from this point forward.

References

  1. Ajzen I. Attitudes, personality, and behavior. Chicago: Dorsey Press; 1988. From intentions to actions; pp. 112–145. [Google Scholar]
  2. Ajzen I. The theory of planned behavior. Organizational Behavior and Human Decision Processes. 1991;50:179–211. [Google Scholar]
  3. Ajzen I. Constructing a TpB questionnaire: Conceptual and methodological considerations. 2002a Retrieved July 20, 2006, from http://www.people.umass.edu/aizen/pdf/tpb.measurement.pdf.
  4. Ajzen I. Perceived behavioral control, self-efficacy, locus of control and the theory of planned behavior. Journal of Applied Social Psychology. 2002b;32:1–20. [Google Scholar]
  5. Ajzen I. Residual effects of past on later behavior: Habituation and reasoned action perspectives. Personality and Social Psychology Review. 2002c;6:107–122. [Google Scholar]
  6. Akaike H. A new look at the statistical model identification. IEEE Transactions on Automatic Control, AC- 1974;19:716–723. [Google Scholar]
  7. Anderson JC, Gerbing DW. Structural equation modeling in practice: A review and recommended two-step approach. Psychological Bulletin. 1988;103:411–423. [Google Scholar]
  8. Armitage CJ, Conner M. Distinguishing perceptions of control from self-efficacy: Predicting consumption of a low-fat diet using the theory of planned behavior. Journal of Applied Social Psychology. 1999a;29:72–90. [Google Scholar]
  9. Armitage CJ, Conner M. The theory of planned behaviour: Assessment of predictive validity and “perceived control”. British Journal of Social Psychology. 1999b;38:35–54. [Google Scholar]
  10. Armitage CJ, Norman P, Conner M. Can the theory of planned behavior mediate the effects of age, gender and multidimensional health locus of control? British Journal of Health Psychology. 2002;7:299–316. doi: 10.1348/135910702760213698. [DOI] [PubMed] [Google Scholar]
  11. Baldwin AR, Tian OPS, Young R. To drink or not to drink: The differential role of alcohol expectancies and drinking refusal self-efficacy in quantity and frequency of alcohol consumption. Cognitive Therapy and Research. 1993;17:511–530. [Google Scholar]
  12. Bandura A. Self-efficacy: Toward a unifying theory of behavior change. Psychological Review. 1977;84:191–215. doi: 10.1037//0033-295x.84.2.191. [DOI] [PubMed] [Google Scholar]
  13. Beck KH, Treiman KA. The relationship of social context of drinking: Perceived social norms, and parental influence to various drinking patterns of adolescents. Addictive Behaviors. 1996;21:633–644. doi: 10.1016/0306-4603(95)00087-9. [DOI] [PubMed] [Google Scholar]
  14. Bentler PM. EQS 6 structural equations program manual. Encino, CA: Multivariate Software, Inc; 2004. [Google Scholar]
  15. Burnam KP, Anderson DR. Multimodal inference: Understanding AIC and BIC in model selection. Sociological Methods and Research. 2004;33:261–304. [Google Scholar]
  16. Clements R. Prevalence of alcohol-use disorders and alcohol-related problems in a college student sample. Journal of American College Health. 1999;48:111–118. doi: 10.1080/07448489909595682. [DOI] [PubMed] [Google Scholar]
  17. Collins SE, Carey KB. Lack of effect for decisional balance as a brief motivational intervention for at-risk college drinkers. Addictive Behaviors. 2005;30:1425–1430. doi: 10.1016/j.addbeh.2005.01.004. [DOI] [PMC free article] [PubMed] [Google Scholar]
  18. Collins SE, Carey KB, Sliwinski MJ. Mailed personalized normative feedback as a brief intervention for at-risk college drinkers. Journal of Studies on Alcohol. 2002;63:559–567. doi: 10.15288/jsa.2002.63.559. [DOI] [PubMed] [Google Scholar]
  19. Conner M, Warren R, Close S, Sparks P. Alcohol consumption and the theory of planned behavior: An examination of the cognitive mediation of past behavior. Journal of Applied Social Psychology. 1999;29:1676–1704. [Google Scholar]
  20. Godin G, Kok G. The theory of planned behavior: A review of its applications to health-related behaviors. American Journal of Health Promotion. 1996;11:87–98. doi: 10.4278/0890-1171-11.2.87. [DOI] [PubMed] [Google Scholar]
  21. Grouzet FME, Otis N, Pelletier LG. Longitudinal cross-gender factorial invariance of the Academic Motivation Scale. Structural Equation Modeling. 2006;13:73–98. [Google Scholar]
  22. Hancock GR. Effect size, power, and sample size determination for structured means modeling and MIMC approaches to between-groups hypothesis testing of means on a single latent construct. Psychometrika. 2001;66:373–388. [Google Scholar]
  23. Hancock GR. Experimental, quasi-experimental, and nonexperimental design and analysis with latent variables. In: Kaplan D, editor. The SAGE handbook of quantitative methodology for the social sciences. Thousand Oaks, CA: SAGE Publications; 2004. [Google Scholar]
  24. Hayduk LA, Glaser DN. Jiving the four-step, waltzing around factor analysis, and other serious fun. Structural Equation Modeling. 2000;7:1–35. [Google Scholar]
  25. Hingson RW, Heeren T, Zakocs RC, Kopstein A, Wechsler H. Magnitude of alcohol-related mortality and morbidity among US college students ages 18–24. Journal of Studies on Alcohol. 2002;63:136–144. doi: 10.15288/jsa.2002.63.136. [DOI] [PubMed] [Google Scholar]
  26. Hu L-t, Bentler PM. Cutoff criteria for fit indexes in covariance structure analysis: Conventional criteria versus new alternatives. Structural Equation Modeling. 1999;6:1–55. [Google Scholar]
  27. Johnston KL, White KM. Binge-drinking: A test of the role of group norms in the theory of planned behaviour. Psychology and Health. 2003;18:63–77. [Google Scholar]
  28. Kim KH. The relation among fit indexes, power, and sample size in structural equation modeling. Structural Equation Modeling. 2005;12:368–390. [Google Scholar]
  29. Larimer ME, Turner AP, Mallett KA, Geisner IM. Predicting drinking behavior and alcohol-related problems among fraternity and sorority members: Examining the role of descriptive and injunctive norms. Psychology of Addictive Behaviors. 2004;18:203–212. doi: 10.1037/0893-164X.18.3.203. [DOI] [PMC free article] [PubMed] [Google Scholar]
  30. Leigh BC. Attitudes and expectancies as predictors of drinking habits: A comparison of three scales. Journal of Studies on Alcohol. 1989;50:432–440. doi: 10.15288/jsa.1989.50.432. [DOI] [PubMed] [Google Scholar]
  31. Little TD. Mean and covariance structures (MACS) analyses of cross-cultural data: Practical and theoretical issues. Multivariate Behavioral Research. 1997;32:53–76. doi: 10.1207/s15327906mbr3201_3. [DOI] [PubMed] [Google Scholar]
  32. Marlatt GA, Gordon JR, editors. Relapse prevention: Maintenance strategies in the treatment of addictive behaviors. New York: Guilford Press; 1985. [Google Scholar]
  33. McMillan B, Conner M. Using the theory of planned behaviour to understand alcohol and tobacco use in students. Psychology, Health & Medicine. 2003;8:317–327. [Google Scholar]
  34. Neal DJ, Carey KB. Developing discrepancy within self-regulation theory: Use of personalized normative feedback and personal strivings with heavy-drinking college students. Addictive Behaviors. 2004;29:281–297. doi: 10.1016/j.addbeh.2003.08.004. [DOI] [PubMed] [Google Scholar]
  35. Norman P, Bennett P, Lewis H. Understanding binge drinking among young people: An application of the theory of planned behavior. Health Education Research. 1998;13:163–169. doi: 10.1093/her/13.2.163-a. [DOI] [PubMed] [Google Scholar]
  36. O’Callaghan FV, Chant DC, Callan VJ, Baglioni A. Models of alcohol use by young adults: An examination of various attitude-behavior theories. Journal of Studies on Alcohol. 1997;58:502–507. doi: 10.15288/jsa.1997.58.502. [DOI] [PubMed] [Google Scholar]
  37. Ouellette JA. How to measure habit? Subjective experience and past behavior (Doctoral dissertation, Texas A&M University, 1996) Dissertation Abstracts International. 1996;57:4093. [Google Scholar]
  38. Ouellette JA, Wood W. Habit and intention in everyday life: The multiple processes by which past behavior predicts future behavior. Psychological Bulletin. 1998;124:54–74. [Google Scholar]
  39. Pentz MA, Chou CP. Measurement invariance in longitudinal clinical research assuming change from development and intervention. Journal of Consulting and Clinical Psychology. 1994;62:450–462. doi: 10.1037//0022-006x.62.3.450. [DOI] [PubMed] [Google Scholar]
  40. Perkins HW. Social norms and the prevention of alcohol misuse in collegiate contexts. Journal of Studies on Alcohol. 2002;(Suppl 14):164–172. doi: 10.15288/jsas.2002.s14.164. [DOI] [PubMed] [Google Scholar]
  41. Perkins HW, editor. The social norms approach to preventing school and college age substance abuse. San Francisco: Jossey-Bass; 2003. [Google Scholar]
  42. Satorra A, Bentler PM. A scaled difference chi-square test statistic for moment structure analysis. Psychometrika. 2001;66:507–514. doi: 10.1007/s11336-009-9135-y. [DOI] [PMC free article] [PubMed] [Google Scholar]
  43. Schwarz G. Estimating the dimension of a model. Annals of Statistics. 1978;6:461–464. [Google Scholar]
  44. Simons J, Carey KB. A stuctural analysis of attitudes toward alcohol and marijuana use. Personality and Social Psychology. 1998;24:727–735. [Google Scholar]
  45. Stacy AW, Bentler PM, Flay BR. Attitudes and health behavior in diverse populations: Drunk driving, alcohol use, binge eating, marijuana use, and cigarette use. Health Psychology. 1994;13:73–85. doi: 10.1037//0278-6133.13.1.73. [DOI] [PubMed] [Google Scholar]
  46. Terry DJ, O’Leary JE. The theory of planned behavior: The effects of perceived behavioural control and self-efficacy. British Journal of Social Psychology. 1995;34:199–220. doi: 10.1111/j.2044-8309.1995.tb01058.x. [DOI] [PubMed] [Google Scholar]
  47. Thompson MS, Green SB. Evaluating between-group differences in latent variable means. In: Hancock GR, Mueller RO, editors. Structural equation modeling: A second course. Greenwich, CT: Information Age Publishing, Inc; 2006. pp. 119–169. [Google Scholar]
  48. Verplanken B, Aarts H, van Knippenberg A, van Knippenberg C. Attitude versus general habit: Antecedents of travel mode choice. Journal of Applied Social Psychology. 1994;24:285–300. [Google Scholar]
  49. Wagenmakers E, Farrell S. AIC model selection using Akaike weights. Psychonomic Bulletin & Review. 2004;11:192–196. doi: 10.3758/bf03206482. [DOI] [PubMed] [Google Scholar]
  50. Wall AM, Hinson RE, McKee SA. Alcohol outcome expectancies, attitudes toward drinking and the theory of planned behavior. Journal of Studies on Alcohol. 1998;59:409–419. doi: 10.15288/jsa.1998.59.409. [DOI] [PubMed] [Google Scholar]
  51. Wechsler H, Lee JE, Kuo M, Seibring M, Nelson TF, Lee H. Trends in college binge drinking during a period of increased prevention efforts. Journal of American College Health. 2002;50:203–217. doi: 10.1080/07448480209595713. [DOI] [PubMed] [Google Scholar]
  52. Young RM, Connor JP, Ricciardelli LA, Saunders JB. The role of alcohol expectancy and drinking refusal self-efficacy beliefs in university student drinking. Alcohol & Alcoholism. 2006;41:70–75. doi: 10.1093/alcalc/agh237. [DOI] [PubMed] [Google Scholar]
  53. Young RM, Tian OPS, Crook GM. Development of a drinking self-efficacy questionnaire. Journal of Psychopathology and Behavioral Assessment. 1991;13:1–15. [Google Scholar]

RESOURCES