The design of the instrument is adopted from previous results in the literature with appropriate modifications for LINE users (e.g. terminologies). The scale of perceived enjoyment was revised from Moon and Kim [7] , which includes three items. The scale of perceived critical mass was revised from Ellison et al. [8] , which includes three items. The scale of self-efficacy was revised from George [9] , which includes three items. The scales of behavioral attitude and subjective norms were revised from Taylor and Todd [10] , which includes two items and three items respectively. The scale of usage intention was revised from Moon and Kim [7] , which includes three items. All of the items were measured on a 7-point Likert-type scale, where possible answers ranged from strongly disagree (1) to strongly agree (7). The instrument of the study is shown in Table 1.

The sample of this study is from Taiwan. The main reason is that LINE is up to a total of 230 million registered users in August 2013 [11] . In terms of the use proportion of population, Chinese Taipei obviously is the biggest country to use LINE in the world. Thus, it has representative to investigate the LINE users in Taiwan.

This study conducted an online survey. The questionnaire was translated by a language professional to ensure that the wording used in the Chinese and English versions were consistent. In the announcement, we declared the purpose of the study and the qualifications for participating in the survey. The respondents were informed that their responses are anonymous and only for academic purposes. The questionnaire was posted in the Google online survey system, which links to PTT (ptt.cc) website, the largest bulletin board system in Taiwan. The respondents who have the experience of LINE were invited to participate in the survey. To increase the sample return rate, this study offered gifts to the respondents for increasing the participation to the survey.

The questionnaires were collected for one month. A total of 470 respondents were received, of which 12 copies were deleted due to regular or incomplete data. The valid respondents totaled 458. The demography of the respondents is shown in Table 2. Females (52.8%) surpass males (47.2%). The largest age group is 40 - 49 years (32.3%), and the largest education group is undergraduate (62.2%). Regarding the frequency of using LINE in the past three months, most of the respondents use LINE every day or almost every day (36.5%).

In addition, the measurement of frequent users in the study was revised from Johnson [12] , which classified the frequent user group as the users who use the Internet a few times a week or every day or almost every day, and the infrequent user group as the users who never or rarely use the Internet. As thus, frequent users in this study totaled 270 (58.9%), whereas infrequent users totaled 113 (24.7%).

This study employed Cronbach’s alpha (a) for examining the internal consistency of the constructs [13] [14] . The a in Table 3 indicates the reliability of the measurement constructs: perceived enjoyment is 0.92, perceived critical mass is 0.91, behavioral attitude is 0.90, subjective norms are 0.83, perceived behavioral control is 0.84, and usage intention is 0.95. These numbers satisfy the general requirements in the field (e.g. [13] ) suggest a reliability coefficient above 0.7, and Robert and Wortzel [14] want the number to be between 0.70 and 0.98. Therefore, we content that this study carried good reliability.

Note: valid samples = 458.

Confirmation factor analysis (CFA) was performed for scale validity assessment [15] . Convergent validity was measured by average variance extracted (AVE) in each construct. The criterion of AVE should be greater than 0.5 [16] . The evidence showed in Table 3 reveals that all constructs were satisfied. Thus, this study possessed adequate convergent validity.

Discriminate validity was also tested. The result shows that the AVE square root of each research variable is larger than the related coefficients of the variables, as shown in Table 4. This is a clear case of positive proof [16] . Thus, this study had adequate discriminate validity.

A structural equation modeling using AMOS 20.0 was conducted to test the postulated hypotheses. Figure 2 presented the estimation results. From the model fitness indexes, χ² = 289.560, DF = 114, χ²/DF = 2.54, GFI = 0.938, AGFI = 0.913, CFI = 0.962, RMSEA = 0.067, showing the collected data fits the postulated model. The estimated structural coefficients were used to test each hypothesis. Results in Table 6 showed that the model explained 57.9%, 42.6% of the variance in behavioral attitude and perceived behavioral control, and 67.5% of

Note: Diagonal elements in boldface represent the square root of AVE.

Note: Estimate is unstandardized; *** means significant at the level of 0.001.

the variance in usage intention. All paths in the research model were statistically significant at the level of 0.001. Perceived enjoyment and perceived critical mass are positively associated with behavioral attitude (Estimate = 0.302, SE = 0.038, CR = 7.857, p < 0.001; Estimate = 0.535, SE = 0.047, CR = 11.312, p < 0.001). Self-efficacy is positively associated with perceived behavioral control (Estimate = 0.436, SE = 0.037, CR = 9.472, p < 0.001). Moreover, behavioral attitude, subjective norms, and perceived behavioral control are positively associated with usage intention (Estimate = 0.590, SE = 0.040, CR = 14.650, p < 0.001; Estimate = 0.138, SE = 0.036, CR = 3.881, p < 0.001; Estimate = 0.226, SE = 0.047, CR = 4.810, p < 0.001). Thus, H1, H2, H3, H4, H5, and H6 are supported.

Multiple group structural equation modeling was conducted to test the differences of the causal relationships across two groups. Yang and Lee [19] suggested that significant differences in two groups were determined by using a χ² difference test. As thus, Table 7 showed that full constrained model with limited the same path coefficients between frequent users and infrequent users is significant (Δχ²/6 = 12.262, p = 0.031). In addition, the path between subjective norms and usage intention, and the path between perceived behavioral control and usage intention all are significant (Δχ²/1 = 5.889, p = 0.015; Δχ²/1 = 5.694, p = 0.017), but other paths are insignificant across frequent users and infrequent users. For more details, Figure 3 and Figure 4 showed the results of estimated structural modeling analysis for frequent users and infrequent users respectively. Moreover, the path between subjective norms and usage intention and the path between perceived behavioral control and usage intention both are significant for frequent users (Estimate = 0.123, SE = 0.037, CR = 3.366, p < 0.001; Estimate

Note: *means significant at the level of 0.05.

= 0.360, SE = 0.062, CR = 5.829, p < 0.001), but insignificant for infrequent users (Estimate = 0.088, SE = 0.101, CR = 0.869, p = 0.385; Estimate = 0.055, SE = 0.085, CR = 0.648, p = 0.517), as shown in Table 8. Thus, H7 is supported.

The study has yielded several important findings. First, it affirms that behavior attitude, subjective norms, and perceived behavioral control are positively associated with usage intention. This result is in line with the study of Ajzen [2] , which asserted that behavioral intention is a function of attitude, subjective norms, and perceived behavioral control. This study also finds that perceived enjoyment and perceived critical mass are two important antecedents influencing behavior attitude for the users. Obviously, three main forces, including behavior attitude, subjective norms and perceived behavioral control, influence LINE users to use the service. Thus, the TPB model has been confirmed in the context of LINE service. This result implies that perceived enjoyment and perceived critical mass are two important antecedents influencing behavior attitude, which in turn increases usage intention for LINE users. It is believed that users can get more fun and communicate with more people in LINE service. Moreover, subjective norms also affect usage intention for LINE users. Indeed, peer pressure can encourage users to participate the service for communicating with friends in the mobile community. Furthermore,

Note: Estimate is unstandardized; **means significant at the level of 0.01; ***means significant at the level of 0.001.

self-efficacy is another antecedent to influence usage intention in the LINE. Maddux [20] argued that if a person has higher self-efficacy in a particular job, this will make him/her more active participation in efforts to complete the task or the job. Thus, a person’s self-efficacy in the use of LINE can influence perceived behavioral control to the service, which in turn increases usage intention.

Second, this study also finds that the effect of behavioral attitude on usage intention (Estimate = 0.590) is stronger than the effects of subjective norms and perceived behavioral control (Estimate = 0.138; Estimate = 0.226). This finding implies that behavioral attitude is a major determinant influencing usage intention for LINE users. The possible reason is that users may likely enjoy the communication with many friends to enhance usage intention in LINE service, followed by the effects of perceived behavioral control and subjective norms.

Third, interestingly, this study found that the causal relationship is varied between frequent and infrequent users. In particular, the intention is only influenced by behavioral attitude for infrequent users, but for frequent users, usage intention is affected by behavioral attitude, subjective norms and perceived behavioral control. The possible reason is that users who are infrequent use of LINE service may lack of interactive targets and the ability of self-efficacy to user the services. Thus, they may passively use the services when some people “line” them.