Title: The Theory of Planned Behavior (TPB) and Texting While Driving Behavior in College Students MS # Manuscript ID GCPI-2015-02298 Appendix 1 Role of TPB in changing other behaviors TPB has been applied to explain a variety of health behaviors, including exercise behavior 17, smoking 18, drug use 19, STD/HIV prevention behaviors 20 and driving behaviors. 11,21-23 For example, among Arab Americans adults in Houston, attitude, ative beliefs, and motivation to comply were significant predictors of the intention to quit water pipe smoking, after adjusting for age, gender, income, marital status, and education. 18 In a prospective study of pregnant women, the authors found that intention significantly predicted exercise behavior, and attitude followed by perceived behavioral control, and subjective was the strongest determinant of exercise intention. 17 The TPB model also has been widely used to predict individuals Zhou et al reported that TPB explained the over 40% variance in intention to use a hands-free mobile phone. An earlier meta-analysis of 185 studies concluded that nearly 40% of the variance in intention to perform a behavior was explained by variables in attitude, subjective, and perceived behavioral control. Study measures descriptions In addition to gender and past behaviors, we examined the following TPB-based predictors. Intention to send and read TWD: Intention was measured through three statements with 7- point Likert scale responses ranging from strongly disagree = 1 to strongly agree = 7: 18 I plan to send/read TWD in the next week; It is likely that I will send/read TWD in the next week; and I intend to send/read TWD in the next week. We created a composite scale wherein higher scores indicated higher levels of intention for each behavior (send, α =.90; read, α =.89). Attitude towards texting while driving: This construct was assessed using three semantic differential items: For me to use a cell phone while driving would be (bad = 1 to good = 7); For me to read TWD would be (worthless = 1 to valuable = 7); For me to send TWD would
be (unwise = 1 to wise = 7). 15 We created a composite scale with higher scores representing positive attitude (send, α =.75; read, α =.79). Subjective : This construct was measured through 7-point Likert scale responses to three statements: 18 Those people who are important to me would approve of me sending/reading TWD; Those people who are important to me would want me to send/read TWD to/from them, and Those people who are important to me think I should send/read TWD. Responses ranged from strongly disagree = 1 to strongly agree = 7. The mean of these three items yielded a composite scale, with higher scores indicating higher agreement for each behavior (send alpha=0.72, read alpha=0.79). Perceived behavioral control (PBC): We assessed PBC through two statements with 7-point Likert scale responses, ranging from strongly disagree = 1 to strongly agree = 7: 18 I have complete control over whether I send/read TWD in the next week, and It is mostly up to me whether I send/read TWD in the next week. The mean of these two items produced a composite scale, with a higher score meaning higher perceived control for each behavior (Pearson's correlation send r(218)=0.54, p<0.001; read(222)=0.54, p<0.001) Group : We assessed this through 7-point Likert scale responses ranging from none = 1 to all = 7 to two statements: Think about your friends and peers. How many of them do you think would send/read TWD? How many of your friends and peers would think that sending/reading TWD is a good thing to do? The mean of these two items produced a composite scale for each behavior (Pearson's correlation send r(212)=0.51 p<0.001, read r(212)=0.52 p<0.001). Moral : We measured this through 7-point Likert scale responses, ranging from strongly disagree = 1 to strongly agree = 7, to three statements: 18 I would feel guilty if I read/sent TWD; I personally think that reading/sending TWD is wrong, and Reading/sending TWD goes against my principles. The mean of these three items produced a composite scale for each behavior (send, α =.76; read, α =.78). Statistical Analysis
All analyses were conducted using IBM Statistical Package for the Social Sciences (SPSS) software, version 20, except for the mediation analysis, which was conducted using Mplus 7.4. 32 We used frequency (count, percentage, means, standard deviation) to depict the overall characteristics of the sample for the categorical variables (gender, ethnicity, marital status, year in college, history of driving and accident/injury). We used Cronbach s alpha and Pearson correlation reliability coefficients to identify and assess any correlations between the main study predictors, and reading and sending TWD. Prior to the hierarchical regression, we examined the relationship between independent variables for collinearity. In the first model, we tested the ability of the TPB to predict students intention to TWD, using intention as the dependent variable in the regression. In the second model, we used willingness to read and send TWD as the dependent variable. We entered background variables in both regression models, and included TPB variables. Statistical significance was established at p 0.05. In addition, we used hierarchical multiple regressions (HMR) to determine the effect of intention and perceived behavioral control in predicting willingness to send and read TWD. Finally, the indirect relationship between perceived behavioral control on willingness to read and send TWD was assessed using a latent mediation analysis. Bivariate Correlations and Reliability Coefficients for Sending and Reading TWD The last column in Table 1 reveals strong correlations between the variables of interest and willingness to send TWD. The strongest correlation is between intention and willingness (0.50), the strongest negative correlation is between moral and willingness (-0.33). The negative correlation indicates that higher scores on moral are associated with lower scores on willingness to send TWD. This suggests participants who believe that sending and reading TWD is wrong are less likely to do it. Table 2 reveals similar results in regards to reading TWD. The last column indicates a strong correlation between most TPB constructs and willingness to read TWD. The strongest positive
correlation is seen between intention and reading TWD (0.42). The strongest negative correlation exists between moral and reading TWD (-0.37). As with sending text, respondents who believe it is wrong to read TWD are less likely to do so. Table 1. Bivariate correlations for sending texts while driving Variable 1.Attitude 2.Subjective 3.PBC 4. Intention 5.Group 6.Moral 7.Gender 8.Past 9.Willingness behavior 1.Attitude (0.95) 0.40** 0.03 0.50** 0.21* -0.25** 0.01 0.08 0.27** 2.Subjective (0.72) 0.08 0.49** 0.32** -0.14 0.00 0.11 0.30** 3. PBC (0.69) 0.24** 0.12 0.26** -0.01 0.05 0.14 4. Intention (0.85) 0.36** -0.23** -0.09 0.29** 0.50** 5. Group (0.67) -0.14 0.01 0.09 0.21* 6.Moral (0.64) 0.05-0.13-0.33** 7.Gender - 0.02-0.03 8.Past behavior - 0.28** 9.Willingness (0.73) Note: PBC=perceived behavioral control *p < 0.05 *p < 0.01 **p < 0.001 Table 2. Bivariate correlations coefficients for reading texts while driving Variable 1.Attitude 2.Subjective 3. PBC 4. Intention 5.Group 6.Moral 7.Gender 8.Past 9.Willingness behavior 1.Attitude (0.95) 0.44** 0.09 0.42** 0.28* -0.30** -0.04 0.14 0.36** 2.Subjective (0.79) 0.04 0.40** 0.34** -0.19* -0.01 0.11 0.26** 3. PBC (0.69) 0.58** 0.12 0.24** -0.04 0.09 0.15 4. Intention (0.84) 0.33** -0.05-0.12 0.26** 0.42** 5. Group (0.68) -0.21* -0.03 0.14 0.24** 6.Moral (0.63) 0.05-0.17-0.37** 7.Gender - -0.04-0.02 8.Past behavior - 0.28** 9.Willingness (0.79) Note: PBC=perceived behavioral control *p < 0.05 *p < 0.01 **p < 0.001 Mediation Analysis of Willingness to Text on Perceived Behavioral Control To assess the indirect relationship of perceived behavioral control on willingness to text, a latent mediation analysis of intention as the mediator of this relationship was tested 32,33. All path coefficients were standardized using the variances of the continuous latent variables, as well as the variances of the outcome variables (i.e., STDYX standardization in Mplus). 34
To assess overall model fit, Hu and 35 recommend using a combination of goodness of fit indices. Their simulation study findings suggest cutoff values close to.95 for Comparative Fit Index (CFI) resulted in the least Type I and II error rates and Root Mean Square Error of Approximation (RMSEA) > 0.05 resulted in acceptable Type II error rates. Given these criteria, overall model fit indicates that (714) = 2756.04, Comparative Fit Index (CFI) = 0.819, and Root Mean Square Error of Approximation (RMSEA) = 0.111. The indices suggest that although the fit is reasonable, alternative models can be tested to improve the model fit. However, since the latent mediation model is of most relevant theoretical interest, the model was chosen despite non-optimal fit. For model identification, indicator paths for the first factor loading of each latent factor was set to 1.0, and the variances of each latent factor estimated as a free parameter. Due to poor initial model fit, residual covariances between send and read items were specified for all factors. For example, for the latent factor Attitude towards texting while driving, the residual covariance between two items were estimated: For me to read messages while driving would be (worthless = 1 to valuable = 7); and For me to send messages while driving would be (unwise = 1 to wise = 7). The residual covariances of items with similar send/read pairings were similarly specified; however due to a problem with convergence, the residual covariance of the following items were constrained to be zero: Within the past week, how often did you use your cell phone to read/send text messages while driving? ; and How many of your friends and peers do you think would read/send an SMS message while driving during the next week. The covariances of all exogenous factors as well as the covariance between the residual were set to zero. The measurement model was specified as eight latent factors each measured with manifest variables as specified in the Study Questionnaire and Measurements and Outcome Variable sections of this paper, with the exception that individual item responses were used in lieu of construct scales. In general, most items loaded highly on each factor, with STDYX loadings ranging from 0.454 to 0.981. The structural model includes the regression paths of intention on attitude, group, moral, past behavior, and subjective ; as well as the indirect path of willingness to text on perceived behavioral control, mediated by intention. See Figure 1.
Based on STDYX standardization, the total effect c of willingness to text on perceived behavioral control was 0.177 and was significant at alpha = 0.05, p = 0.035. The indirect effect of willingness on perceived behavioral control is the product of the a and b paths, 0.292*0.684 = 0.200. Since the direct effect c of willingness to text on perceived behavioral control was -0.023 which was not significantly different from zero at the alpha = 0.05 level, then accounting for the effect of willingness on intention, the effect of willingness on perceived behavioral control was reduced. To assess the statistical significance of the indirect path, the bias-corrected bootstrap using 10,000 replications was conducted. The 95% confidence interval of the indirect path was (0.096, 0.369), see Table 3. Since the interval does not contain zero, the indirect effect is significant at the alpha = 0.05 level. The results suggest that intention mediates the relationship between perceived behavioral control and willingness to text. Table 3. Bias-corrected bootstrap confidence interval of total, direct and indirect paths using STDYX standardization Effects Path Estimate 95% CI Total (c path) PBC WIL 0.177 (0.002, 0.348) Direct (c path) PBC WIL -0.023 (-.0.224,0.171) Indirect (a*b path) PBC INT WIL.0.200 (0.096,0.369) Note: WIL= Willingness to Text, INT = Intention to Text, PBC=perceived behavioral control