Participants
A convenience sample of 662 voluntary Spanish first-year HE students (age range = 17–23 years, Mdn = 18), mostly composed of women (60.0%, n = 397), was used for this study. Participants were selected in two consecutive academic years: sample 2014/15 (n = 338, age range = 17–23 years, Mdn = 18) and sample 2015/16 (n = 324, age range = 17–23 years, Mdn = 18), who attended HE for the first time. As a function of the study area, most of the students (66.2%) were enrolled in degrees in the juridical-social field, and the rest were in the scientific-technological field. No association was found between study area and academic years, χ2(1, N = 662) = 1.10, p = .295, or between gender and academic years, χ2(1, N = 662) = 3.28, p = .073.
Materials and procedure
Students’ planning was operationalized through the “Inventario sobre Estrategias Metacognitivas” (IEM; Metacognitive Strategies Inventory; Martínez-Fernández, 2004, 2007), a Spanish adaptation of the Reduced Revised State Metacognitive Inventory (RRSMI; O’Neil & Abedi, 1996). Participants were asked what they do or think when they face a learning activity or problem.
The structure of the 20 IEM items, rated on a Likert-type scale ranging from one (never) to five (always), was examined with a sample of college students through maximum likelihood (ML) exploratory factor analysis (oblique rotation), showing a bifactorial model. These factors, each with 10 items, were named planning (e.g., item 9: “You are aware of the need to plan your course of action”) and self-checking (e.g., item 2: “You check your work while you are doing it”) and presented, respectively, alpha values of .80 and .82. However, the factors presented a very high correlation (.71), and the author did not test an alternative unifactorial model.
Students’ AEs were operationalized using the final version of the “Cuestionario de Percepciones Académicas (CPA): Versión Expectativas” (Deaño et al., 2015), named the English Academic Perceptions Questionnaire (APQ; Almeida et al., 2018). The APQ has seven factors, each with six items rated on a Likert-type scale ranging from one (strongly disagree) to six (strongly agree): training for employment (TE), referring to the conditions of obtaining training to obtain better jobs or enter into the working world (e.g., item 15: “Obtain training to obtain a good job”); personal and social development (PSD), which includes autonomy, self-confidence, critical thinking, and personal improvement through new experiences of academic life (e.g., item 16: “Use academic opportunities to improve my identity, autonomy and self-confidence”); student mobility (SM), linked to the attitude of carrying out part of the studies in international mobility programs, internships or jobs abroad (e.g., item 24: “Obtain international-quality training”); political engagement and citizenship (PEC), which reflects the desire to engage in the political, social, and economic life of the country, to understand how to help improve it, and to participate in specific associative or volunteer activities (e.g., item 25: “Participate in volunteer activities”); social pressure (SP), which includes the items referring to the desire to respond to parents' expectations or to please significant others (e.g., item 5: “Meet my family’s expectations”); quality of education (QE), linked to feeling challenged to deepen one's knowledge and having the personal and material means to stimulate it (e.g., item 13: “Deepen my knowledge of specific subjects”); and social interaction (SI), which includes the will to enjoy some moments of conviviality and fun, dedicating a scheduled weekly time to these activities, different from the study time, which may entail a relationship with classmates (e.g., item 28: “Attend university student parties”).
Regarding the APQ’s psychometric properties, Almeida et al. (2018, Table 2) showed that factors’ convergent validity (CV) and discriminant validity (DV) and its composite reliability (CR) (Fornell & Larcker, 1981) were suitable across countries and genders.
Procedure
Data collection
Students (initial sample, N = 669) were selected to ensure the heterogeneity of the major subjects. Data were collected at the beginning of the first semester in the classroom after obtaining teachers’ permission and students’ informed consent. Students who attended HE in previous academic years left the classrooms. The instruments were administered in a counterbalanced way. Seven students were excluded from the sample due to incomplete data (gender variable = 4; IEM protocol = 3).
Data analysis
IBM SPSS Statistics for Windows (version 21.0) was used for descriptive data analysis, and LISREL 8.80 (Jöreskog & Sörbom, 2006) was used for model estimation and testing.
Given the ordinal categorical nature of the data, analyses were performed using the underlying bivariate normal approach (Jöreskog, 2005). PRELIS 2 (Jöreskog & Sörbom, 1996) produces the polychoric covariance matrix of the underlying latent continuous and normal counterparts of items’ observed responses, the respective asymptotic covariance matrix, and the vector means of the latent responses. They were taken as input for model estimation and testing with the robust Satorra-Bentler (SB) scaled correction (Satorra & Bentler, 1994) in LISREL 8.80 (Jöreskog & Sörbom, 2006) using the SIMPLIS command language (Jöreskog & Sörbom, 1993). Factor measurement units were assigned by fixing the path of one of their items to one.
Models’ fit to the data were examined through the following practical goodness-of-fit (GOF) indices and recommended cutoff values (Hu & Bentler, 1998): the root mean square error of approximation (RMSEA; values close to or below .06), the standardized root mean square residual (SRMR; values close to or below .08), and the comparative fit index (CFI; values close to or above .95). The expected cross validation index (ECVI; Browne & Cudeck, 1993) was also used for the comparison of the two alternative mediation models in Fig. 1; the model with lower ECVI should be chosen.
Following a two-step modeling approach (Jöreskog & Sörbom, 1993), the mediation models presented in Fig. 1 were only tested after the assessment of the structural validity of the IEM model.
First, a confirmatory factor analysis (CFA) was performed with the 2014/15 sample. The obtained ML completely standardized estimates allowed for the examination of IEM factors’ CV, DV, and CR (Fornell & Larcker, 1981). CV was examined through items’ average variances extracted (AVEs), which should be at least .50. DV was assessed by comparing factors’ shared variance (φ2 = squared disattenuated correlation) and AVE of each compared factor: DV should be lower than AVE. A factor reliability of .80 is recommended for group comparisons (Nunnally & Bernstein, 1994).
Second, the resulting factorial solution was tested through a multigroup cross-sectional measurement invariance analysis, using both the 2014/2015 and the 2015/2016 samples. It typically starts with the testing of the form-invariant model, where all parameters are freely estimated across groups, followed by the testing of more stringent equality conditions, specifically weak, strong, and strict invariance (the latter compares with the former) (Meredith, 1993). Under weak invariance, factor loadings are equal across groups. Under strong invariance, factor loadings and intercepts (item values corresponding to the zero value of the factor) are equal across groups. To ensure construct comparability across samples, strong invariance is a sufficient criterion. However, to complete the study of measurement invariance, strict invariance was also examined. In strict invariance, factor loadings, intercepts, and residual (item-specific factor plus random error) variances are equal across groups. Finally, a model can also be partially invariant, indicating differential item functioning (Byrne, Shavelson, & Muthén, 1989).
Comparisons between baseline models (with parameters unconstrained for all groups; smaller df) and restricted models (with specific parameters constrained to be equal across groups; larger df) were based on the difference (Δ) between models’ CFI and, subsidiarily, on GOF statistics. The cutoff value of .01 was used for the ΔCFI results (Cheung & Rensvold, 2002).
Finally, the two alternative mediation models with latent variables (see Fig. 1) were tested using the full sample. After the selection of the model with better ECVI, the following expression, based on the difference between its total and direct effects (unstandardized), was applied: Δz = total effect − direct effect/root square [(SE2(total effect) + SE2(direct effect))/2]. If the value of the test statistic Δz was higher than 1.96, p < .05, then it means that the contribution of the indirect effect to the total effect was significant.