Factor analysis of the data partially supported the previous findings regarding the distinct motivations for playing online games. While some factors were not cor distinct within one construct(with somewhat high cross-loadings), we decided to refine and retain them. The logic was that first, our pilot test sample was somewhat cond, our scales were translated to Chinese and hence items may have slightly shifted from their original meaning, even though a back-and-forth translation procedr employed. As a result, the items adopted were preserved and the translation was rechecked. We modified some ambiguous expressions. The final English version be found in Appendix A Data Collection Participants were selected from middle schools in a large Chinese city. This was done for two reasons. First, from a convenience and practical stand point-it wa to encompass many online game players, which where potentially reachable by one of the researchers. Second, the penetration of Internet in this city is the second hig China( CNNiC, 2007), and presumably, adolescents in this city are susceptible and exposed to online games more than others in the nation. Two data collection approaches were taken: in school-and on street-collection surveys. First, 600 copies of paper-based que middle schools in this city over a period of two months. The surveys were administrated by a research assistant with the supervision and help of the teachers of the rel asses. Survey completion was voluntary, and was encouraged with small monetary incentives(less than S1). Additionally, 200 copies of the survey were handed out ocations on the streets. Potential adolescents were approached in person at locations such as McDonalds and Intemet Cafes. Potential participants were asked if they online games before given a copy of the survey. In total, 800 surveys were distributed and 682(85%)were returned, out of which 623(78%)were valid. A Multivari lied to the data showed that the source of data( school vs. street) had no significant omnibus effect(Pillais Trace =0. 16, p<0. 13), implyin were no significant differences between the datasets. Thus, subsequent analyses were performed on the whole dataset. Participants' ages ranged from 12 to 18 with an of about 15 years. The modal age was 14 with 36% of our sample representing this age group. The sample was slightly male dominant(56%). ANALYSIS AND RESULTS Since our research model contains both reflective and formative components, PLS (Parcial Least Square)was chosen for data analysis. PLS can easily support odels with no identification issues, as demonstrated in past MIS research( Chin Gopal, 1995; Turel et al., 2007). The hypothesis testing was conducted ma version 2.0(Ringle et al., 2005) following the two-step approach for model estimation(Anderson Gerbing, 1988). The measurement model We first examined factor loadings. Almost all were above 0.7, but the loading of Addiction item I was. 52. In addition, the average variance extracted (Ave)of ded threshold of 0.5. Hence, we deleted the problematic item. As a result, the Ave of Addiction with 6 it was acceptable(59). The same procedure was applied to the perceived cost construct. The loading of Cost5 was low(31), and the Ave wa able(57). Neve deleted the item. As a result, all loadings were over 0.7, and the AvE was 72. Consequently, reliability coefficients were above. 70 and all AVE scores were over .50 Appendix B). This indicated that the measurement scales were reliable and that the latent variables account for more than 50 percent of the variance in the items. As s Appendix B, the loadings are in an acceptable range and the t-values indicate that they are significant at least at the 01 level. The results in Appendix B further sugge discriminant validity, because the square root of the AVE is greater than all of the related inter-construct correlations( Chin, 1998). In order to further assess validity, a oadings table(Appendix C)was constructed. It can be seen that each item loading is much higher on its assigned construct than on the other constructs, supporting ac convergent and discriminant validity To further evaluate the formative composite variables(Attention Switching, Parental Monitoring and Resource Restrictions), we followed the guidelines provide Cenfetelli and Bassellier(2009). With the first guideline, we checked multicolinearity among the indicators with Variance Inflation Factor (VIF)scores. The highest V alculated was 1. 461 (Table 3)and was thus below the recommended upper border( Diamantopoulos Siguaw, 2006) Table 3. vif, factor weights, p-value and factor loadings for the formative measurement Factor Weights Attention switching 1.333 Parental Monitor PM Passive .246 73 RESI 043 RES2 0.05 131 108 15 <000Factor analysis of the data partially supported the previous findings regarding the distinct motivations for playing online games. While some factors were not com distinct within one construct (with somewhat high cross-loadings), we decided to refine and retain them. The logic was that first, our pilot test sample was somewhat second, our scales were translated to Chinese and hence items may have slightly shifted from their original meaning, even though a back-and-forth translation procedu employed. As a result, the items adopted were preserved and the translation was rechecked. We modified some ambiguous expressions. The final English version of t can be found in Appendix A. Data Collection Participants were selected from middle schools in a large Chinese city. This was done for two reasons. First, from a convenience and practical stand point - it wa to encompass many online game players; which where potentially reachable by one of the researchers. Second, the penetration of Internet in this city is the second hig China (CNNIC, 2007), and presumably, adolescents in this city are susceptible and exposed to online games more than others in the nation. Two data collection approaches were taken: in school- and on street-collection surveys. First, 600 copies of paper-based questionnaires were randomly distribute middle schools in this city over a period of two months. The surveys were administrated by a research assistant with the supervision and help of the teachers of the rel classes. Survey completion was voluntary, and was encouraged with small monetary incentives (less than $1). Additionally, 200 copies of the survey were handed out locations on the streets. Potential adolescents were approached in person at locations such as McDonalds and Internet Cafes. Potential participants were asked if they online games before given a copy of the survey. In total, 800 surveys were distributed and 682 (85%) were returned, out of which 623 (78%) were valid. A Multivari of Variance procedure applied to the data showed that the source of data (school vs. street) had no significant omnibus effect (Pillai’s Trace = 0.16, p < 0.13), implyin were no significant differences between the datasets. Thus, subsequent analyses were performed on the whole dataset. Participants' ages ranged from 12 to 18 with an of about 15 years. The modal age was 14 with 36% of our sample representing this age group. The sample was slightly male dominant (56%). ANALYSIS AND RESULTS Since our research model contains both reflective and formative components, PLS (Parcial Least Square) was chosen for data analysis. PLS can easily support models with no identification issues, as demonstrated in past MIS research (Chin & Gopal, 1995; Turel et al., 2007). The hypothesis testing was conducted using Sma version 2.0 (Ringle et al., 2005) following the two-step approach for model estimation (Anderson & Gerbing, 1988). The Measurement Model We first examined factor loadings. Almost all were above 0.7, but the loading of Addiction item 1 was .52. In addition, the average variance extracted (AVE) of A with all items was .49 which is slightly below the recommended threshold of 0.5. Hence, we deleted the problematic item. As a result, the AVE of Addiction with 6 it was acceptable (.59). The same procedure was applied to the perceived cost construct. The loading of COST5 was low (.31), and the AVE was acceptable (.57). Neve deleted the item. As a result, all loadings were over 0.7, and the AVE was .72. Consequently, reliability coefficients were above .70 and all AVE scores were over .50 Appendix B). This indicated that the measurement scales were reliable and that the latent variables account for more than 50 percent of the variance in the items. As s Appendix B, the loadings are in an acceptable range and the t-values indicate that they are significant at least at the .01 level. The results in Appendix B further sugge discriminant validity, because the square root of the AVE is greater than all of the related inter-construct correlations (Chin, 1998). In order to further assess validity, a loadings table (Appendix C) was constructed. It can be seen that each item loading is much higher on its assigned construct than on the other constructs, supporting ad convergent and discriminant validity. To further evaluate the formative composite variables (Attention Switching, Parental Monitoring and Resource Restrictions), we followed the guidelines provide Cenfetelli and Bassellier (2009). With the first guideline, we checked multicolinearity among the indicators with Variance Inflation Factor (VIF) scores. The highest V calculated was 1.461 (Table 3) and was thus below the recommended upper border (Diamantopoulos & Siguaw, 2006). Table 3. VIF, factor weights, p-value and factor loadings for the formative measurement VIF Factor Weights p-value Attention switching InnerAS 1.333 .651 < 0.001 ExternalAS 1.333 .500 < 0.001 Parental Monitoring PM_Passive 1.246 .604 < 0.001 PM_Active 1.246 .573 < 0.001 Resource Restrictions RES1 1.253 .043 < 0.5 RES2 1.461 .104 < 0.05 RES3 1.343 .131 < 0.01 RES4 1.085 .153 < 0.001