The Quantitative Analysis of Reasoned Action Theory

The structural and measurement components of the IM

The IM is primarily a theory of intention formation, with a focus on the roles of attitudes, norms, and efficacy relative to one another in explaining variation in intention. The “direct measures” model of the IM (Figure 1A) includes both intention and behavior and shows the structural relationships between the three predictors of intention. Figure 1B shows the complete IM from a measurement model perspective, with each of the latent constructs (except for the direct measure of self-efficacy/control) having multiple indicators that reflect the causal influence of the construct.

OPEN IN VIEWER

Figure 1

The IM as an Analytic Object. (A) As a Structural Model with Direct Measures as Predictors of Intention. (B) As a Measurement Model Showing Beliefs, Direct Measures, and Intention.

One implication of the structural and measurement features of the IM is that the causal model is already “prespecified.” ³ When using the reasoned action approach, there is no need for “theory trimming.” Nonsignificant effects in an IM analysis are important exactly because nonsignificant effects are important findings, not parameters to discard. The reasoned action approach does not require that all determinants of intention are relevant for all behaviors; some intentions (and, therefore, some behaviors) may be best predicted by subsets of the three theoretical mediators. In fact, an important question about every behavior of interest is always “Which combination of determinants, or which single theoretical determinant, are the most important in predicting intention?”

Another analytic implication of the IM is that its measurement complexity does not easily allow for analyzing the complete IM simultaneously (i.e., including both the underlying beliefs and the primary determinants in the same analysis). For example, the data matrix of the hypothetical IM in Figure 1B includes 435 nonredundant variances and covariances.

Because of this complexity, the analysis is typically broken down into three steps. The first step is to estimate a path model with the direct measures predicting intention and, if available, extending the analysis to prospective behavior. The direct measures may be treated as latent constructs or parceled (Little et al. 2002; Matsunaga 2008). The second step is to have separate equations for each arm of the IM that include the underlying beliefs in the analysis. In the third step, segmentation analysis is used to identify the relevant beliefs. This last step is useful for intervention design and when the identification of specific beliefs is relevant to the study goals. We elaborate on each step below and also discuss the correct treatment of background factors/precursor variables in each case.

Step 1: Analyzing the direct determinants of intention using path analysis

As shown in Figure 1A, the direct measure model of the three determinants of intention can be analyzed using path analysis (Wolfle 2003), where the determinants are either observed (as in Figure 1A) or latent. When the constructs are observed (represented in standard SEM notation by boxes), the determinants are represented as scales that are created using the multiple indicators for each construct. This process (“parceling”) is very common and appropriate in instances when the contribution of each indicator to the overall construct is not of interest. The exception is self-efficacy/control, which in most cases is measured using only one item (but see Yzer [this volume] for measurement suggestions for the perceived behavioral control construct).

One point to consider is the sequencing of measures used in the path analysis. Intention to perform a behavior is always prospective. That is, the intention measure always refers to performing the target behavior in a specific future context. Longitudinal data in which data on intention precedes data collected on behavior is ideal (i.e., intention is measured at time 1 and behavior is measured at time 2 or later). Data on the determinants of intention can be collected during the same time period as the intention measures, which makes cross-sectional data suitable for an IM analysis as well. However, as is the case with the use of cross-sectional data in general, one’s ability to draw causal inferences is limited if (prospective) intention and (retrospective) behavior are simultaneously analyzed.

The analytic approach changes when precursor variables are included in the analysis of the direct measures. As mentioned, the effects of precursor variables are completely mediated by the determinants of intention (Bryan, Schmiege, and Broaddus 2007). Thus, for example, with one or more precursor variables, the IM analysis model for the direct measures will have five equations (three equations with the theoretical mediators as the dependent variable, one equation with intention as dependent, and one equation with behavior as dependent).

The reasoned action approach is sometimes criticized because there is not an explicit causal model at the level of the three theoretical mediators when the direct measure model is analyzed with precursor variables. Reasoned action theory presents no assertions about how the three IM mediators are mutually determined. For example, do attitudes cause normative pressure? Does self- efficacy cause normative pressure and normative pressure then cause attitude? In other words, reasoned action theory is a causal theory of intention, not a theory of the causal relationships between the proximal determinants of intention.

However, this does not imply that the three IM mediators are statistically independent. Note that in the model with precursors, the mediators become endogenous variables. Figure 2 shows this situation. Although there are no causal arrows between any of the three IM mediators, all three mediators are correlated. The principle of “path tracing rules” (Maruyama 1998) demonstrates that the implied correlation between attitude and normative pressure is the A*B product when the coefficients are completely standardized. Similarly, the implied correlation between self-efficacy/control and attitude is A*C, and the implied correlation between normative pressure and self-efficacy/control is B*C. None of these implied correlations is zero unless the precursor variables are unrelated to the IM mediators.

OPEN IN VIEWER

Figure 2

How Precursor Variables Create Associations between the Direct Measure Mediators

That said, to ignore other sources of IM mediator association besides the common cause represented by the precursors is a mistake. Doing so will produce model miss-fit and lead to poor SEM goodness-of-fit statistics, such as the root mean square error of approximation (RMSEA) or the Tucker-Lewis index (Hu and Bentler 1995). The solution to this issue is to correlate the error terms of the three mediators. The resulting model fit is exactly the same as imposing a causal model between them and is an example of the common paradox of “equivalent models” in SEM (MacCallum et al. 1993). Correlated errors of the direct measures are a necessary component of any direct measure analysis when a precursor variable is included in the analysis (Hennessy et al. 2010).

Step 2: Analyzing the underlying beliefs with an arm-by-arm approach

When the underlying beliefs are included in the analysis, each theoretical arm of the IM is analyzed separately. As an example, here we discuss the underlying behavioral beliefs of the attitude component of the IM, because in our experience it is typically the outcome expectancies of attitude that are the most difficult to model. Figure 3A displays the relationships for the attitudinal arm of the IM from the precursor variable, to the scale of underlying beliefs (based on behavioral belief items B₁ to B_k), to the direct measure of attitude (based on semantic-differential items SD₁ to SD_k), and to intention. While Figure 3 depicts the arm-by-arm approach using latent variables for the direct measure of attitude and intention, parceled variables may be used as an alternative for the direct measure of attitude and intention. However, the specification in the figure for the underlying beliefs must be maintained regardless of whether the other constructs are latent or observed because it is the specific belief item performance that is of interest.

OPEN IN VIEWER

Figure 3

The IM with Precursor Variables, Underlying Beliefs, Direct Measures, and Intention. (A) With Underlying Beliefs Defining a Scale. (B) With Underlying Beliefs Defining an Index.

The correlations between the underlying beliefs and the precursor variable are parameter set A, the factor loadings of the underlying beliefs scale are parameter set B, the correlation between the underlying beliefs latent factor and the direct measure attitude factor is parameter C, the factor loadings of the semantic-differential attitude items defining the direct measure attitude scale are parameter set D, and the correlation between attitude and intention is parameter E. ⁴

One of the most common issues with the underlying behavioral beliefs measures is that they are often extremely heterogeneous and often do not structure well into a single factor using Cronbach’s α (Streiner 2003a), ordered difficulty scaling (Hennessy et al. 2008), or confirmatory factor analysis (Brown 2006). Each belief is often quite substantively different from every other belief, and such item heterogeneity does not generate the conceptual or statistical consistency that should be present in a scale. However, the underlying beliefs do not have to scale into a common factor. One solution is to create an index of the belief items. When observed measures are used, the process for creating an index is essentially the same as creating a scale except that the α coefficient or other statistics of internal consistency are not reported. In general, one can expect that in most instances an index is appropriate for the underlying behavioral beliefs and that scales can be created for the normative pressure beliefs and the self-efficacy beliefs.

The use of “causal indicators” is the SEM equivalent of creating an index of the observed variables (Diamantopoulos and Winklhofer 2001). While measurement models for scales consist of effect indicators that have causal arrows going from the latent construct to the observed variables (i.e., the latent construct explains the covariation between the items), measurement models for causal indicators have causal arrows going from the observed variables to the latent construct. That is, the indicators or belief items, in this case, define the latent construct, and the regression coefficients attached to these arrows show the change in the latent construct due to a one-unit change in the indicators (Bollen and Lennox 1991; Streiner 2003b). While Figure 3A shows effect indicators in a scale of the underlying beliefs, behavioral beliefs are treated as causal indicators in Figure 3B. Note the difference of the arrow direction in Figure 3B: the arrows are going from each belief item to the underlying beliefs index. For a fuller discussion of causal indicators and the IM, see Hennessy, Bleakley, and Fishbein (this volume).

Step 3: Segmentation analysis for intervention design purposes

In this section, we discuss ways of analyzing a reasoned action dataset with the purpose of designing effective behavior change messages. There are two analytic preconditions that need to be established before effective messages can be designed. First, the target behavior (and the associated intention measure) needs to be associated with its relevant theoretical determinants because mismatches of message type with target behavior determinant (e.g., attempting to manipulate underlying attitudinal beliefs when the major theoretical predictor of intention is actually normative pressure) will not lead to behavior change. Second, the constructed messages need to address the relevant underlying beliefs of the major theoretical predictor. The correlations between the message features (which would be the “precursor variable” in this case) and the underlying beliefs are the A parameters in Figure 3. Effective message content needs to reinforce beliefs that are positively associated with intention to perform the target behavior and counterargue beliefs that are negatively associated with intention to perform the target behavior (see Jordan et al., this volume). This dual process is necessary because the reasoned action mechanism of behavior change depends on initial identification and then subsequent manipulation of salient underlying beliefs relevant to the determinants of intention (Kim and Hunter 1993).

The analysis of the direct measures in step 1 represents a causal model of intention formation that identifies which (or possibly all) of the three direct determinants is best in predicting intention for the target behavior. The arm-by-arm analysis in step 2 identifies which underlying beliefs are most relevant in defining the belief scale or index (parameter B is the relevant set in this case). Once those influential underlying beliefs are identified, the next step is to determine which beliefs differ by who intends to perform the behavior and who does not (i.e., intender status). This step is called a “segmentation analysis.”

We typically classify the respondents into one of three groups based on their responses to the reasoned action intention items and retrospective reports of performing (or not performing) the target behavior. Performers are respondents who report performing the target behavior in the survey and intend to perform the behavior in the future. Intenders are respondents who report high intention to perform the behavior but are not performing the behavior currently. Nonintenders are respondents who do not report high intention to perform the behavior and are not currently performing the behavior. To measure intender status, intention to perform the behavior is dichotomized into intending to perform the behavior or not intending to perform the behavior. How one defines intending to perform the behavior is discretionary. On a 7-point scale, for example, any value above the neutral position (5–7) could be considered as having an intention, while values 1 to 4 would be recategorized into no intention. The tests of association to perform next depend on the measurement level of the belief item. Most commonly, for example, if the belief measure is on a 7-point scale, t-tests would compare the intender and nonintender means of a belief item. Performers are included with respondents who are intending to perform the behavior and should not be broken out as a separate group unless maintaining the behavior is of interest to the researcher.

The proportions of the total analysis sample that fall into each of the segmentation categories are important because respondent status suggests different kinds of relevant messages. Performers, for example, require messages designed to maintain their performance of the target behavior, and these kinds of messages are rarely researched (Rothman 2000). Intenders, on the other hand, require messages that emphasize increase in skills and abilities and changes in the environmental context to enable performance of the behavior and movement into the performer category. Nonintenders require messages that change intention via attitudes, norms, or control belief changes to move these respondents to the positive intention intender category. ⁵

For each behavioral belief, a segmentation analysis should look at the correlation between the individual underlying belief and intention, the average value of the specific belief for each intender and nonintender (and performer, if applicable), and the percentage who endorse the specific belief by respondent segment (Fishbein and Cappella 2006). If the intenders are more likely to hold the belief, and the belief is positively related to the behavior, the belief should be emphasized in any future interventions. Alternatively, if nonintenders are more likely to hold a belief and the belief is negatively associated with intention, this belief should be counterargued. Once again, this analysis is only performed if specific beliefs are of interest for the purposes of intervention development and are only conducted on the underlying beliefs of a determinant or determinants that are significantly related to one’s intention to perform a behavior.

Should past behavior be included in the model?

In our opinion, the inclusion of past behavior in an application of the IM is often a tautological exercise. It is evident that one’s past behavior will predict one’s future behavior, but when included in an analysis, it often accounts for a majority of the variance without adding any substantive explanatory value. As Fishbein and Ajzen (2010, 286) state, “To argue that we behave the way we do now because we performed the behavior in the past begs the question as to why we previously behaved that way.”

The effect of a background variable on intention and on behavior is not completely mediated

Occasionally a mediation analysis will demonstrate that the effects of a background variable on intention are not completely mediated by attitudes, normative pressure, and efficacy or (much more rarely) that there is a direct effect of a (nonpast behavior) background variable on behavior. While modeling such an effect is discouraged because it is not compatible with the specification of the IM, how does one explain when this happens? Is it a failure of the IM?

We assert that in such a case the direct effects can probably be explained by poor measurement. Either the appropriate underlying beliefs were not elicited or the direct measures of the determinants of intention were incomplete. (For more on accurate elicitation of beliefs, see Middlestadt [this volume] and Jaccard [this volume]). Although the measurement standardization of reasoned action theory is an important practical benefit to researchers, such standardization cannot guarantee adequately performing measures, especially when the target behavior is not well researched. Therefore, consistent with conventional practice in SEM (Anderson and Gerbing 1988), we always begin with an examination of the measurement models of the IM constructs before we perform an analysis of the direct measures or the underlying beliefs.

Should injunctive and descriptive norms be combined into a single normative pressure construct?

The direct measure items for normative pressure are typically combined into one normative pressure construct and usually scale quite well. With the underlying beliefs, it is sometimes useful to separate the injunctive norms from the descriptive. The distinction between the two with regard to a particular behavior may be meaningful depending on the empirical questions of interest. Parsimony is a priority with the IM, so whenever possible combining the different types of normative items is preferred, especially for the normative pressure construct and less so for the underlying beliefs. The same position can be taken in relation to the measures of autonomy and capacity that underlie perceived behavioral control. Again, we prefer to combine these items into a single direct measure if possible. For more on the autonomy and capacity measures, see Yzer (this volume).

Is there a proper way to test for moderation in the context of the IM?

One of the strengths of the IM is the notion that the roles of specific underlying beliefs as well as attitudes, normative pressure, and efficacy in intention formation and the performance of a behavior may vary by population, context, and behavior. For example, does the importance of a particular determinant of intention vary by racial or ethnic group, gender, media exposure levels, or intervention group? One way to test for moderation of the IM is to conduct a multiple group analysis. This approach is essentially a stratification of the model by the classification of interest followed by tests for statistically significant differences of the regression coefficients of precursor, theoretical mediators, and the intention–behavior association across groups. This method is preferred to including interaction terms because the stratification keeps the interrelationships between the model constructs intact, reduces multicollinearity, and allows for all model parameters to vary across the groups. In addition, the use of interaction terms usually requires the dichotomization of continuous variables, which is undesirable (Streiner 2002; Royston, Altman, and Sauerbrei 2006; MacCallum et al. 2002).