Data were collected through both Google form and printed questionnaires.
Raw data from Google forms were exported into an Excel file. Data from printed responses were then manually input into that same file. Coding for each item was the same as in the phase of piloting. Since data in this phase were collected from different faculties, the faculties were categorized as History (coded as 1), Literature (coded as 2), Physics (coded as 3), Chemistry (coded as 4), and Information Technology (coded as 5). History and Literature were then grouped into a category called social (coded as 1) whereas the remaining was categorized into science (coded as 2). The whole data set were then imported into SPSS 22 for statistical data analysis.
To begin with, data screening was conducted to eliminate univariate outliers.
Then, internal consistency tests were performed using acceptable values of Cronbach’s Alpha being greater than .7 and Corrected Item-Total Correlation being greater than .3 (Hair et al., 2018; Nunnally, 1978). Next, mean, standard deviations, skewness, and kurtosis coefficients of all items were preformed to determine if data met the assumption of normality (Hair et al., 2018; Warner, 2008).
Cut-off values for skewness and kurtosis criteria were ranging from -2 and +2, and from -7 and +7, respectively (Hair et al., 2018). Next, Pearson correlations were run on all dimensions of reading self-efficacy and all types of metacognitive reading strategies. Specially, the absolute correlation coefficient values among independent variables, namely, the four dimensions of reading self-efficacy, were examined to check if multicollinearity occurred among them. If those values are smaller than .8, multicollinearity is not likely to occur, which allowed the researcher to move to the next step (Senaviratna & Cooray, 2019). To further confirm the non-existence of multicollinearity, Tolerance and Variance Inflation Factor values were checked in SPSS to see if their values satisfied the criteria of being less greater .2 and smaller than 10, respectively (Field, 2009).
After multicollinearity was confirmed not to be a big issue, factor analyses, including both exploratory and confirmatory types, were run to confirm whether
49
the classified items converged in the pre-determined dimensions of reading self- efficacy and the types of metacognitive reading strategies. Regarding exploratory factor analysis (EFA), since the present study specified the roles of reading self- efficacy as an independent variable and metacognitive reading strategies as a dependent variable, items accepted in the reliability tests were imported separately for EFA. This is because Hair et al. (2010, p.99) stated that "mixing dependent and independent variables in a single factor analysis and then using the derived factors to support dependence relationships is inappropriate.". Hair et al. (2015, p.411) further emphasized that EFA “can be used to factor analyze either independent or dependent variables considered separately".
The principal components method with varimax rotation were applied in EFA.
Factor loading for EFA was set at .5 to ensure observed variables had good statistical significance (Hair et al., 2018). The KMO value ranging between .5 and 1.0 (Kaiser, 1974) and Bartlett’s test of sphericity being smaller than .05 were set as cut-off values (Hair et al., 2018) to indicate the appropriateness of factor analysis.
Furthermore, eigenvalues being greater than 1 and the cumulative variance being greater than 50% indicate a good structure (Hair et al., 2018). Importantly, all the items should converge correctly in the pre-determined dimensions of reading self- efficacy and types of metacognitive reading strategies.
EFA results were cross-validated using confirmatory factor analysis (CFA).
This step was performed on Analysis of Moment Structures version 24 (AMOS 24). It includes the examination of model fit indices and other criteria of construct validity. Before examining model fit indices, it is necessary to ensure the standardized regression weight coefficients of the items being greater than .5 (Tabachnick & Fidell, 2001). The following fit indices were used in the present study: the chi-square divided by the degrees of freedom (χ2/df), Tucker-Lewis fit index (TLI), incremental fit index (IFI), comparative fit index (CFI), standardized root mean square residual (SRMR), root mean square error of approximation (RMSEA). Cutoff values were determined as χ2/df < 2 (Schermelleh-Engel et al., 2003), TLI and IFI and CFI > .90 (Tabachnick & Fidell, 2001), SRMR < .08, RMSEA< .06 (Hu & Bentler, 1999). Regarding construct validity, it is indicated
50
by convergent and discriminant validity. Convergent validity is indicated by construct reliabilities being greater than .7 (Bagozzi & Yi, 1988) and average variance extracted being greater than .4 (Brunelle & Lapierre, 2007; Menguc &
Auh, 2006; Zhou et al., 2005). Eventually, all constructs are considered discriminant if their average variance extracted values are greater than maximum shared variance and if their square root of average variance extracted exceed their correlations with other constructs (Fornell & Larcker, 1981).
All items that were kept in CFA phase were assessed in a complete structural equation modeling (SEM) which is defined as “a comprehensive statistical approach to testing hypotheses about relations among observed and latent variables”
(Hoyle, 1995, p.1). SEM can provide more information regarding the fit of the framework with measurement errors under control and the use of parameters to determine interdependencies (Bosak et al., 2013). This approach not only allowed the researcher to “understand patterns of covariances among a set of observed variables” but it can also “explain as much of their variance as possible” within the hypothesized model (Kline, 1998, p.14). In this study, covariance-based structural equation modelling (CB-SEM) was run on AMOS 24. Model fit indices were examined again using the same cut-off values as listed in the CFA phase to guarantee that the model had sufficient goodness of fit prior to the assessment of relations in the model. Probability values should be smaller than .05 to demonstrate associations between dimensions of reading self-efficacy and types of metacognitive reading strategies. To further confirm CB-SEM results, bootstrap sampling with 500 replications were conducted (Efron & Tibshirani, 1993;
Streukens & Leroi-werelds, 2016). If 95% confidence interval values in bootstrap contain zero, there are no associations between the variables (Xu, 2016).
Additionally, supplementary analysis was provided using one-way analysis of variance (one-way ANOVA) on SPSS 22 to examine differences in reading self- efficacy and metacognitive reading strategies with regard to the control variables, namely, genders and fields of study (Armstrong et al., 2000). In detail, if the probability of Levene’s test is greater than .05, there is no need to transform data because they are homogeneous. Then, if the probability value of F-ratio is smaller
51
than .05, the control variables can make impact on the main variables. On the contrary, if the probability value of F-ratio is greater than .05, the control variables do not make any influences on the main variables. To specify how students from each faculty are different from their counterparts in other faculties, post-hoc tests were performed. Criteria to confirm differences are the probability of Levene test being greater than .05, then the probability of F-ratio being smaller than .05, and lastly the probability in multiple comparison being smaller than .05 (Armstrong et al., 2000).
52