Data were collected from 508 younger adolescents (48.8% male and 51.2% female) with a mean age of 12.56 years (DS=0.50, range: 12-13 years).

Exploratory factor analysis was conducted on a first random subsample of 254 participants, 48% men and 52% women, with a mean age of 12.54 years (SD=0.50, range: 12-13 years).

Confirmatory factor analysis was conducted on a second randomly selected 254 participant subsample, 49.6% men and 50.4% women, with a mean age of 12.56 years (SD=0.50, range: 12-13 years). These two subsamples of participants were compared with regard to demographic characteristics. Chi square test and Student's independent samples t-test were used to evaluate differences in the distribution of gender and age between the two groups. The two sets of data are comparable in terms of both gender (χ²=0.283, p=0.594) and age (t=-0.53, df=506, p=0.593).

The instrument was administered to students in southern Italy. All participants provided written consent.

TAS-2012,¹³ is a self-report measure of alexithymia. It consists of three subscales: Difficulty identifying feeling (7 items; e.g., "I am often confused about what emotion I am feeling"); Difficulty describing feelings (5 items; e.g., "It is difficult for me to find the right words for my feelings"); and Externally oriented-thinking (8 items; e.g., "I prefer talking to people about their daily activities rather than their feelings"). Cut-off scores are as follow: ≤50=no alexithymia, 51-60=borderline alexithymia, and ≥61=alexithymia. The Italian reliability, construct, and criterion validity of scores on the TAS-20 have been well established in various samples of adults.¹⁹²⁰

A preliminary inspection of the item distribution was conducted to assess the extent to which TAS-20 items could be factor analyzed using normal-theory estimation procedures. The normality of data was checked through Kolmogorov-Smirnov and Shapiro-Wilk tests.

Psychometric evaluation of the TAS-20 was initiated with examination of the distributional properties and response frequencies.

Exploratory factor analysis was performed to determine the underlying dimensions of the questionnaire, using various criteria of item selection according to the number of selected factors and item factor loadings. Prior to EFA, data were inspected to ensure items were significantly correlated, using Bartlett's Test of Sphericity, and shared sufficient variance, using Kaiser-Meyer-Olkin (KMO) Measure of Sampling Adequacy to evaluate whether items share sufficient variance to justify factor extraction.

Sampling adequacy values that are less than 0.50 are considered unacceptable, values that are between 0.50 and 0.60 are considered marginally acceptable, and values greater than 0.80 and 0.90 are considered excellent.²¹ Principal axis factoring was selected as the method of factor extraction. An oblique rotation method (promax criterion) was selected to obtain a simple structure since there was no theoretical assumption suggesting that the factors were independent from each other.

The number of factors to extract was determined performing random data parallel analyses.²² The eigenvalues derived from the actual data were compared to the eigenvalues derived from the random data. Factors were retained as long as the ith eigenvalue from the actual data was greater than the ith eigenvalue from the random data.²³ Both Kaiser's²⁴ criterion and the scree test²⁵ were checked for agreement. Salience was detected applying the following three criteria: 1) a factor loading of at least 0.3 on the primary factor, ensuring a high degree of association between the item and the factor, 2) when an item was loading simultaneously on two factors, a difference of 0.3 between loading on the primary factor and loading on other factors, ensuring that each item could be considered salient to one factor, 3) a minimum of three items for each factor, so ensuring meaningful interpretation of stable factors.²⁶

To investigate to what extent the factor scores were intercorrelated the standard Pearson product moment correlation coefficient was used.

The reliability of the scale, in terms of internal consistency, was computed by coefficient alpha. Corrected item-scale correlations were examined for each of the revealed subscales. For item selection it was decided that adjusted item-total correlations for each item of a scale should exceed 0.30, recommended as the standard for supporting item-internal consistency.²⁷

A confirmatory factor analysis, using ML Maximum Likelihood robust estimation procedures, was performed using the EQS Structural Equation Program Version 6.1.²⁸ To statistically evaluate the closeness of the hypothetical model to the empirical data, multiple goodness-of-fit indexes were used, including the ratio of the chi-square to degrees of freedom (χ²/df), the Non-Normed Fit Index (NNFI), the Comparative Fit Index (CFI), the Standardized Root Mean Square Residual (SRMR), and the Root Mean Square Error of Approximation (RMSEA). NNFI and CFI values of 0.95 or greater and SRMR and RMSEA values of 0.05 or less are interpreted as evidence of models that fit well.²⁹

Table 1 gives item characteristics for both examined subsamples. Included are mean, standard deviation, skewness and kurtosis, Kolmogorov-Smirnov and Shapiro-Wilk tests of normality, and response frequency of all the 20 items.

The mean values of the TAS-20 items ranged from 1.81 to 3.37 for the first random subsample and from 1.78 to 3.57 for the second random subsample.

The distributional properties of each item were examined by inspecting the skewness and kurtosis and the pattern of response frequency. The univariate skewness values ranged from -0.36 to 1.38 for the first random subsample and from -0.48 to 1.25 for the second random subsample, and the univariate kurtosis values ranged from -1.63 to 0.60 for the first random subsample and from -1.54 to 0.50 for the second random subsample, thus suggesting that some items deviated from a normal distribution.

The statistical significance of both Kolmogorov-Smirnov and Shapiro-Wilk tests of normality revealed that each item had a distribution that was significantly different from normal and, as a result, suggested that estimation procedures that assume a normal distribution may not be appropriate for examining the underlying factor structure of the TAS-20.³⁰³¹ Based on these findings, the principal axis factoring method was chosen for exploratory factor analyses, and Maximum Likelihood robust estimation procedures were applied for confirmatory factor analyses.

Data from the first random subsample of respondents were subjected to exploratory factor analysis to identify the likely factor structure of the questionnaire. With our 20-item scale, we were able to satisfy the minimum ten participants-per-item ratio that is usually recommended for factor analysis.³² A sample of 12.7 subjects per item ensured that reliable factors would emerge from the factor analysis.

Bartlett's Test of Sphericity (chi-square=645.55; df=190) was significant (p<0.001), indicating that the correlation matrix is factorable based on a suitable level of variable interrelations, and the Kaiser-Meyer-Olkin Measure of Sampling Adequacy was 0.73, demonstrating a sufficient proportion of common variance in our variables.³³ Both results, thus, suggest that items were appropriate for proceeding with factor analysis.

Parallel analysis determined six factors to be extracted (Table 2, Figure 1). The resulting number of factors is evidently over-defined, with several factors comprised by only two indicators, some items with loadings less than 0.30 on all factors, and a number of items loading simultaneously on two factors, without a difference of at least 0.30 between loading on the primary factor and loading on other factors.

To determine the number of factors to extract, the scree plot and eigenvalues were also examined. The Kaiser-Guttman's criterion is known to potentially inflate the number of factors to be extracted, because it is sensitive to the number of variables in the analysis. Hence, Cattell's²⁵ scree test is considered a more reliable indicator of the number of factors to be extracted because it draws on the relative values of the eigenvalues and so is not sensitive to the number of variables in the analysis.³⁴ Cattell²⁵ recommended that the number of factors to be extracted is the number of eigenvalues that lie well above the scree slope and is a more reliable test for the number of factors in cases where there is a clear and easily interpretable scree slope. The eigenvalue greater than one criterion suggested extracting seven factors, accounting for 35.63% of the total variance. Inspection of the scree plot suggested a solution of up to four factors.

Based on the examination of the pattern of loadings and accepting a minimum of three items for each factor, we retained four factors explaining 27.54% of the total variance. Based on the resultant pattern matrix, item 20 "Looking for hidden meanings in movies or plays distracts from their enjoyment" that failed to load on any of the four factors was not retained (this item loaded: on F1 at -0.252, on F2 at -0.131, on F3 at <0.10, and on F4 at 0.168).

Factor 1, with an eigenvalue of 3.40 and responsible for 14.50% of the total variance for the questionnaire, included 7 items which loaded above 0.32. The items with the highest loading are items 7 and 13.

Factor 2, with 4 items loading above 0.32, had an eigenvalue of 1.71, and accounted for 5.64% of the total variance explained. The items with the highest loading are items 4 and 2.

Factor 3, with 4 items loadings above 0.35, had an eigenvalue of 1.52 and accounted for 3.85% of the total variance. The items 19 and 10 have the highest loading on this factor.

Factor 4, with 4 items loading above 0.35, had an eigenvalue of 1.41, and accounted for 3.55% of the total variance. Items 17 and 15 have the highest loadings on this factor.

Items and factor loadings are presented in Table 3.

The revealed dimensions correlated significantly (p<0.01, two-tailed tests) but moderately, with each other (-0.22≤r≤0.21, p<0.01), except for the non-significant correlation between factor 2 (Difficulty describing feelings) and factor 3 (Externally oriented thinking) (r=-0.02) and between factor 3 (Externally oriented thinking) and factor 4 (Lack of subjective sig-nificance or importance of emotions) (r=-0.02). The correlation between F2 and F4, although still significant (p<0.05, two-tailed tests) was very weak (r=0.159) (Table 4).

The first factor "Difficulty identifying feelings", consisting of seven variables, had a Cronbach's alpha of 0.69, which delineates good internal consistency of this subscale. If one or more items of the first factor are eliminated, Cronbach's alpha drops off. The second factor "Difficulty describing feelings", involving four items and had a Cronbach's alpha of 0.52, which can be considered to be undesirable. One item was removed from this factor scale due to corrected item-total correlation below 0.30 (item 5 "I prefer to analyze problems rather than just describe them"). Thus, alpha for the second factor, now consisting of three variables, was 0.60. The third factor "Externally oriented thinking", consisting of four items, had a Cronbach's alpha of 0.40. There were no items that would have increased the scale reliability if they were deleted. The fourth factor "Lack of subjective significance or importance of emotions", comprising four items, had a Cronbach's alpha of 0.40. The deletion of any of the eight items could not increase the internal consistency of this dimension. These values are considered unacceptable and indicate poor internal homogeneity.

The confirmatory factor analysis performed on the second random subsample showed reasonable goodness-of-fit for the oblique bi-factorial model: [chi]² (32, n=254)=54.22; p=0.008; χ²/df=1.69; NNFI=0.92; CFI=0.95; SRMR=0.05; RMSEA= 0.05; 90% confidence interval=0.027-0.078 (Table 5). All manifest variables loaded significantly (p<0.05) on their hypothesized latent factors. Figure 2 presents the standardized parameter estimates.