Psychometric properties of the Internet Addiction Test in a sample of Malaysian undergraduate students

A total of 104 undergraduate students from Univerisiti Teknologi Malaysia (UTM) answered the questionnaire. As shown in Table 2 , the sample consisted of 46 students from Arts, Humanity, and Social Science, 27 from Science and 27 from Engineering. There were 50 males and 54 females.

The pencil-paper questionnaire used in this study was comprised by two parts: first, basic information of undergraduate students including gender, major field, time spent online per day, and years of Internet use; second, the Internet Addiction Test (IAT) —a 20-item self-report instrument used to measure an individual's Internet use from the perspective of psychological symptoms and behaviours, such as psychological dependence, compulsive use, withdrawal, problems of school, sleep, family, and time management. It was developed based on Young's YDQ ( Young & Nabuco de Abreu, 2011; Young, 1996 ). The original English version of IAT was translated into Malay using translation and back translation procedures. Both English and Malay were shown in the questionnaire in this study. In Young and Nabuco de Abreu's latest book, Internet Addiction: A Handbook and Guide to Evaluation and Treatment , the items are rated on a six-point scale regarding participants’ experience of their Internet use: 0 = not applicable, 1 = rarely, 2 = occasionally, 3 = frequently, 4 = often, 5 = always . The score range is 0 to 100, and the higher score, the greater level of PIU. An individual who gets a total score between 0 and 30 is deemed as normal Internet user, between 31 and 49 mild Internet user, between 50 and 79 moderate PIU, and between 80 and 100 he is supposed to suffer from severe PIU. In this study, the individual who got 80 or above on IAT was categorized as PIU, the remaining were non-PIU.

First, this study exmined the items of IAT using Winsteps, version 3.75.0, which is a Rasch Model analysis software. The Rasch Model theory is a kind of item response theory (IRT) which intends to measure item responses rather than total scores ( Thissen, 2001 ). There are some critical concepts used in this study under Rasch Model analysis. Person fit in the Rasch model is an index of individual's response to items. People may be considered as “misfit” when they respond in an inconsistent manner because of feeling bored and inattentive to the task, confused, or an item evokes an unusually salient response from an individual ( Linacre, 2012). Linacre (2012) suggested that the value of INFIT and OUTFIT MNSQ should be in the range of 0.6 and 1.4 for rating scales. Separation coefficient is the signal-to-noise ratio, the ratio of “true” variance to error variance. Person separation is used to classify people. Low person separation implies that the instrument may not be sensitive enough to distinguish person with high and low performance. More items may be needed. Item separation is used to verify the item hierarchy. Low item separation implies that the person sample is not large enough to confirm the item difficulty hierarchy of the instrument. This is analogous to the Fisher Discriminant Ratio. Reliability (separation index) is separation reliability. The person reliability is equivalent to KR-20, Cronbach's alpha coefficient. And the item reliability is equivalent to construct validity ( Linacre, 2012).

Second, exploratory factor analysis (EFA) was conducted by Mplus, version 6. The EFA in Mplus could provide the goodness of fit statistics as CFA. This study took the following model fit index to evaluate the EFA model. First, there is the chi-square and degrees of freedom, which suggested that a model can be considered to fit well if ? 2/df ratio is below 2. Second, the Root Mean Square Error of Approximation (RMSEA) suggested that the value between 0 and .05 indicated a good fit and between .05 and .08 indicated an acceptable one. Third, the Standardized Root Mean Square Residual (SRMR) was suggested to be in the range of .05 and .10 as acceptable, between 0 and .05 as good fit ( Schermelleh-Engel & Moosbrugger, 2003 ). The fourth index is the Comparative Fit Index (CFI), which was suggested to be greater than .95 as good fit, and above .90 acceptable ( Hu & Bentler, 1999 ). Last, there is the Tucker-Lewis Index (TLI), also known as the Non-normed Fit Index, (NNFI), whose value was recommended to be greater than .95 as good fit ( Hu & Bentler, 1999).

To examine the rating scale with six categories, the result of category structure for IAT is shown in Table 3 . The observed average measure increases with the category score (-1.75, -0.77, -0.24, 0.31, 0.70 and 1.03 for categories 1, 2, 3, 4, and 5 respectively) and is close to sample expected value. The value of structure calibration also increases with the category value, which indicated that there was no disordered category. The value of INFIT and OUTFIT is close to 1 on categories 1, 2, 3, 4, and 5 (from 0.88 to 1.15).

Table 4 is the Rasch analysis result of item fit statistics in misfit order, which showed that all the point-measure correlations (CORR.) are positive and high, range from .43 to .78, and all are close to the expected correlation (EXP.). It implied that all the items are aligned with the abilities of person. The infit and outfit mean-square (MNSQ) values in Table 4 showed that all the items fit the Rasch model very well with mean infit and outfit of 0.99 and 1.02 respectively, except IAT 7 and IAT 12 with higher infit and outfit MNSQ (> 1.40). Further examination on a person's performance of items found that the misfit on IAT 7 and IAT 12 was due to the abnormal response from five students. The infit and outfit MNSQ of IAT 7 and IAT 12 ped below the cut-off value of 1.4 (IAT 7: 1.39 and 1.35; IAT 12: 1.35 and 1.38) once the response of these five students was removed from the dataset. Therefore, IAT 7 and IAT 12 could be kept, as the misfit was caused by the unexpected responses of five students.

The result of Rasch principal component analysis (PCA) in Table 5 indicated that the raw variance in observations of IAT was 54%, with 23.5 eigenvalue units. The unexplained variance in the first contrast was 7.4%, with 3.2 0.500 units; the second contrast was 6.5%, with 2.8 0.500 units; and the third contrast was 5.2%, with 2.3 eigenvalue units. The 0.500 units of first, second, and third contrast are bigger than 2.0, which implied that IAT may be multidimensional with items measuring different constructs. To further test the unidimensionality of IAT, the items were segmented into subtests according to the cluster numbers to perform the disattenuated correlation on person measures, which got significant high positive value of disattenuated correlation, ranging from .6604 to 1.00 ( Table 6 ). The high positive disattenuated correlation implied that the person measures on the different clusters of items are statistically the same, which implied that the three clusters of items measure the same thing. Based on the result of PCA and attenuated correlation, all the items of IAT measure the same construct with four subdimensions, which suggested to identify the sub-construct of IAT.

The overall property of IAT showed high person and item separation (3.52 and 4.61 respectively) corresponding to person reliability of .93 and item reliability of .95. The high person separation indicated the students were separated into more than three groups by IAT, while the high item reliability meant that the item ability was widely spread, and could distinguish approximately five different levels of Internet addiction.

Exploratory factor analysis (IAT) was run in Mplus v. 6. to identify the underlying sub-construct of IAT using weight least square with mean and variance (WLSMV) estimation. As previous research have found one- to six-factor solutions, this research identified the one- to six-factor models respectively using oblique direct quartimin rotation. The goodness-of-fit of the six EFA models are listed in Table 7 , which indicates that a 6-factor model is fit better and acceptable (? 2/df < 2, RMSEA = .075, SRMR = .029, CFI = .986, TLI = .969).

The factor loading and correlations of the 6-factor model are shown in Table 8 . All items had strong primary loadings on their corresponding factors, ranging from .344 to .786. There were four salient cross-loading items that were IAT1, IAT13, IAT15, and IAT18. All the six factors were correlated weakly to moderately ( r = .233 -.517).

Factor 1, named as tolerance, time management problems, and neglect of duty, comprised four items (IAT2, IAT 3, IAT6, IAT8). Factor 2, termed neglect of social life, consisted of three items (IAT4, IAT5, IAT7). Factor 3, named problematic use and reality substitute, contained three items (IAT 10, IAT11, IAT14). Factor 4 was termed as withdraw and emotional conflict, and contained five items (IAT 12, IAT13, IAT15, IAT19, IAT20). Factor 5, namely, intentionally concealing behavior, had two items (IAT 9, IAT18). Factor 6, termed lack of control, comprised three items (IAT1, IAT16, IAT17).

The overall mean IAT score was 49.567 ± 19.323. Result of two- way ANOVA without interaction showed that there were no significant main effect of gender F (1, 100) = 3.838, p = .053 and major field F (2, 100) = 0.554, p = .576 on overall IAT mean score. There were only four students reported the overall IAT score above 80, who were categorized as PIU. As shown in Table 9 , the mean item scores ranged from 1.600 ± 1.310 to 3.220 ± 1.397 for the non-PIU students and from 2.750 ± 1.258 to 5.000 ± 0.000 for PIU students.

Table 10 provided the correlation of six IAT factors, time spent online per day and years of Internet experience, which showed that time spent online per day was significantly related to each IAT factors, while years of Internet experience was significantly correlated to two IAT factors (F2: neglect of social life; F5: intentionally concealing behavior).