The results of the study are analyzed in two steps: main component analysis, confirmatory factor analysis.
3.1. Main Component Analysis
The initial data exploration reveals that some students did not want to participate by answering the questionnaire, so the valid sample consisted of 319 students, as shown in
Table 6. Forty-one percent of the respondents were male, while 59% were female. Regarding the type of institution, 49% came from Universities, 13% from Professional Institutes, and 38% from Higher Level Technical Training Centers.
The responses were analyzed according to a Likert scale from 1 to 7, where 1 was strongly disagree and 7 was strongly agree (
Table 7).
The collinearity of the variables was investigated using a Pearson correlation matrix. When using the matrix determinant, a
p-value of 0.01 was achieved. This indicates that the set of variables in the correlation matrix has a high level of collinearity. Bartlett’s test of sphericity was employed for multicollinearity assumptions, yielding a
p-value of 0.0001; and the Kaiser–Meyer–Olkin (KMO) sample adequacy measure provided a value of 0.93. A main component analysis was carried out based on these findings. Two criteria were analyzed to extract the components: the parallel analysis, which suggested three components, and the variance percentage criterion, which took the variance explained by the three components suggested in the previous analysis. The results of the first oblique rotation are summarized in
Table 8.
Then, after three rotations, the items with comparable scores in more than one dimension were removed, affecting the total variance explained.
Table 9 shows the final findings of the rotated components, which exclude items 1 and 3.
The final version of the instrument was created based on the results collected. Two questions (Q01 and Q03), as well as two categories (information organization and teamwork), were removed. The Planning category was moved to the KCT dimension. The last 11 topics were divided into three categories: Component 1 (Open Learning), which included six questions; Component 2 (Information Management), which included two questions; and Component 3 (Knowledge Creation and Transfer), which included three questions.
The reliability analysis of the instrument yielded a Cronbach’s alpha coefficient = 0.92, which is considered very good according to Landero and González [
38]. The analysis for each component was as follows: IM component α = 0.87; KCT component α = 0.79; and OL component α = 0.90. The final version of the questionnaire contains 14 questions, 11 of which correspond to a seven-level scale, and one to the scale: Yes/No (but I know it)/I do not know it; one corresponds to the frequency: Every day, two hours or more/Every day less than two hours/A few times a week/A few times a month/Never; and one question has answers of the type I learned: alone (a)/with my teachers/with my family/with my friends/I do not know how to do this.
The final structure of the B-PLE instrument with three validated components, which explain 72% of the total variance, is presented in
Table 10.
In terms of component correlations, a weak positive correlation was found between IM and KCT (rs = 0.48, p < 0.001) and a moderate positive correlation between IM and OL (rs = 0.57, p < 0.001) and KCT and OL (rs = 0.60, p < 0.001).
3.2. Confirmatory Factor Analysis
A Confirmatory Factor Analysis (CFA) was performed to validate the results of the psychometric properties obtained from the principal components analysis. This technique is considered a type of structural equation modeling (SEM). Its purpose is to analyze the structure and conformation of items and indicators of a previously hypothesized latent variable [
39] and to represent their relationships with observed or indicator variables [
33,
34], to confirm that all questions fit the proposed model [
27,
35].
In the CFA, subjects’ scores on each item are due to an unobserved variable (or latent factor), which explains the variability of scores on each item. It is likely that the latent factor does not fully explain the variability of responses on each item; this unexplained situation is known as measurement error [
40]. The CFA was carried out in three stages: (i) proposal of the measurement model; (ii) analysis of indices and goodness-of-fit criteria of the model; and (iii) re-specifying the model.
Stage 1. Measurement model proposal
According to the results obtained in the principal component analysis, the proposed model is composed of 11 observed variables, grouped into 3 latent variables: IM, with two questions; KCT, with three questions; and OL, with six questions (see
Table 10).
Stage 2. Analysis of the indexes and goodness-of-fit criteria of the model
In this stage, the structural equation modeling methodology was used, implemented in the lavaan package of the R Studio statistical software. This package is used to analyze multivariate statistical models, CFA, structural equation models, and growth curve models [
41].
The procedure consists of analyzing the covariance structure to confirm that the measurement model matches the proposed model structure by comparing two covariance structures: the covariance matrix was obtained from the observed variables and the covariance matrix that the model reproduces [
42].
To load the model, the notation presented in
Table 11 should be considered:
The proposed model can be represented in the statistical software R Studio with the following notation:
Model1 <-’IM =~ Q03 + Q04
KCT =~ Q01 + Q02 + Q07
OL =~ Q05 + Q06 + Q08 + Q09 + Q09 + Q10 + Q11’.
By complementing the lavaan package with the semPlot package, it is possible to graphically represent the proposed conceptual structure, as shown in
Figure 2. The ovals represent the three latent variables (or constructs), and the squares represent the 11 observed variables, which together make up the dimensionality of the instrument. The bidirectional arrows represent covariances between the latent variables. The latent variables’ influence on observable variables is represented by the unidirectional arrows.
To analyze the model fit, there are different indices and criteria. The first goodness-of-fit index is the χ2 (Chi-square) statistic, which is proposed to represent the point of best fit or the minimum discrepancy between the compared matrices [
43]. With this index, the overall fit of the measurement model is evaluated. The model has an acceptable fit if the Chi-square/gl values are 2 to 3 and with limits up to 5 [
35]. This statistic is very sensitive when used in samples with more than two hundred subjects. In these cases, it is suggested to omit the χ2 interpretation and evaluate the model fit with other goodness-of-fit indices [
34,
42,
44,
45], which are explained below.
The Comparative Fit Index (CFI), developed by Bentler [
46], compares the χ2 discrepancies between an independent model that maintains that there is no relationship between the variables in the model and the model proposed by the researcher [
40]. CFI values range from 0 to 1. If the CFI value is greater than 0.90, it is considered a reasonable fit between the model and the data [
42]. The unnormalized fit index (NNFI), also known as Tucker–Lewis index (TLI), considers the complexity of the model by comparing the degrees of freedom of the independent model with those of the proposed model. TLI values vary between 0 and 1. Values greater than 0.9 are considered as a good model fit. Values greater than 1 indicate an overparameterization of the model [
34]. The root mean squared error of approximation (RMSEA) refers to the amount of variance not explained by the model per degree of freedom [
40]. It is interpreted as the absolute measure of the difference between the relationship structures between the proposed model and the covariances in the measured population [
47]. When the RMSEA presents values less than 0.10, it is considered a reasonable fit between the measurement model and the structure of the data. If the RMSEA is less than 0.05, the fit is superior; if the RMSEA value is less than 0.01, the fit is outstanding [
42]. Another index used is the RMSR, which analyzes the absolute measure of the differences between the elements of the covariance matrix reproduced by the model and the covariance matrix derived from the observed sample data. It is interpreted in its standardized version, known as SRMR. If this value is less than 0.05, it is considered a reasonable model fit [
42]. The summary of the indicators analyzed is presented in
Table 12.
In addition to the indicators analyzed, Herrero [
40] suggests that it is necessary to consider the standardized variances and covariances, to corroborate that there is no negative variance (Heywood cases) and that none of the correlations between the variables of the model exceeds ±1. If this happens, the results of the model are invalidated, because it offers a solution that is not possible (negative variance or correlation exceeding ±1).
The inspection of the goodness-of-fit indicators associated with the proposed model is presented in
Table 13.
As can be seen, the proposed model complies with the minimum parameters suggested in each goodness-of-fit index, except for the χ2 indicator, which, in the case of samples with more than 200 subjects, it is suggested to omit.
Stage 3. Re-specifying the model
Although the indicators are acceptable, it is possible to analyze the relationships of the original proposed model and evaluate whether it is possible to improve them by reviewing the modification indexes option.
Table 14 presents some of the values of the output obtained with the modindices function, available in the lavaan package. Each row indicates the amount by which the χ2 value would decrease if any of the proposed relationships were included. For example, the first row (Q10~~Q11) suggests that there would be a 42-point change in the χ2 value (the current value is 172.893). The two questions pertain to the same latent variable. The difference between the two is that the context changes. Q10 refers to activities with ICT use in class period, while Q11 refers to activities with ICT use outside class time. This could be one of the reasons why this correlation in measurement errors between the two variables is suggested.
The new evaluation model would be as follows:
Model2 <-’IM =~ Q03 + Q04
KCT =~ Q01 + Q02 + Q07
OL =~ Q05 + Q06 + Q08 + Q09 + Q10 + Q11
# Error correlation
Q10 ~~ Q11’
With the re-specification of the model, a modification index was obtained that reduced the χ2 value by 42 points, improving the rest of the global goodness-of-fit indexes of the empirical model, complying (as with model 1) with four of the five proposed indicators. The comparison of the goodness-of-fit indicators associated with the two models is presented in
Table 15.
In
Figure 3, the graphical representation of model 2 is presented. Some changes in the values of the regression coefficients are observed, compared to model 1 (see
Figure 2); and a covariance relationship of 0.40 emerges between the error terms of variables Q10 and Q11. This implies that both observed variables share an alternative source of variance, different from the latent variables of model 2 [
42].
The regression coefficients are understood as follows. The first two values from left to right (0.60 and 0.84) are interpreted as the influence of the latent variable IM on the observed variables Q03 and Q04. This implies that when the latent variable increases by one unit, Q03 (Digital identity) and Q04 (Information seeking) increase by proportions of 0.6 and 0.84, respectively. For the rest of the latent variables the same analysis is followed. When the latent variable KCT increases by one unit, the observed variables Q01 (Knowledge construction), Q02 (Information processing), and Q07 (Planning) increase in proportions of 0.73, 0.82, and 0.70, respectively. Finally, when the latent variable OL increases by one unit, the observed variables Q08, Q10, and Q11 (linked to the category Open-ended Learning) increase by 0.71, 0.77, and 0.78, respectively, while the observed variables Q06 (Knowledge transfer), Q05, and Q09 (linked to the Motivation category) increase by 0.75, 0.80, and 0.82 respectively, when the latent variable OL increases by one unit.
To confirm the accuracy of the results, the coefficient omega, a reliability estimator used in multidimensional factorial models [
48,
49,
50], was calculated [
51]. This indicator is suggested because of its higher sensitivity compared to other estimators [
52], its resistance to sampling in heterogeneous populations, and the reduced risk of overestimation of reliability [
50,
53].
Using the MBESS library and the ci.reliability function of the statistical software R Studio, the result shown in
Table 16 is obtained.
According to the results obtained for the coefficient, the instrument has an internal consistency of 0.91; additionally, according to the level of confidence, there is a 95 percent chance that the true value of omega is found in the interval (0.893, 0.934).
To be considered an acceptable level of confidence, the omega coefficient must be between 0.70 and 0.90 [
51,
54,
55].