Causes of higher levels of stress among students in higher education who used eLearning platforms during the COVID-19 pandemic

2.1 Sampling procedure

In this study a convenient sampling procedure was used to select the research sample. A total of 1,000 email IDs of students were collected from the teachers of the 12 higher education institutions where they were enrolled, in India. Bearing in mind ethical issues and confidentiality the names of students were not disclosed. Using random number method, a total of 800 email IDs were selected. The Google form was used to construct the questionnaire which was later sent via email. From the total questionnaires sent by email, 602 questionnaires were received. All entries were selected since the Google form does not allow any question to be left unanswered. After detection and removal of outliers, the final sample size was restricted to 538 responses for the current study, which represents an excellent sample rate of 65.25%.

To evaluate the causes of student stress levels deriving from the COVID-19 pandemic emergency, the Partial Least Square Structural Modeling (PLS-SEM) was used, and a conceptual model designed. The PLS-SEM method is nonparametric in nature, which means that it does not need any assumption concerning the distribution of data. The PLS-SEM presents an opportunity to determine multifaceted procedure of associations and causal relationships that are otherwise hard to uncover (Hair et al. 2012). PLS-SEM processes data for the assessment of path coefficients. This study uses the PLS-SEM to better understand the factors that most influence student stress level because of COVID-19. The conceptual model explains the relationships between the latent variables and their related manifest variables. By using the PLS-SEM technique, a model was developed and a total of 43 factors were also identified as the observed variables, which were followed and clustered into six groups such as Academic environment (Busari 2012; Kotter et al., 2017), Family (Byron et al., 2008; Chernomas & Shapiro 2013; Goff 2011), Social (Bennett & Holloway 2014; King et al., 2014), Personal (Deb et al., 2015; Verma et al., 2002), Health (Dallman et al., 1993; Deb et al., 2015; King et al., 2014) and Cognitive (Shinto 1998; Liu 2015; Liu and Lu 2011).

2.2 Statistical evaluation

The primary data was analyzed using some statistical tools, such as partial least squares structural equation modeling and line charts, Cluster analysis, Mean and Error Bar plot. In this research, the statistical package R Language 3.6.3 was used with other packages, such as Hmisc, plssem, ggplot2, ggdendro and dendextend and MS-Excel 2007.

2.3 Cluster Analysis

In order to identify the similarity and dissimilarity among constructs, the cluster analysis was conducted. The variables were clustered based on the Wald minimum variance measures, using different colors to identify the six constructs.

Mean and Error Bar Plot: For every construct, mean and standard deviation were computed on the basis of the average measure of its variables.

Partial Least Square-Structural Equation Modeling (PLS-SEM): PLS-SEM is one of the non-parametric statistical tools that is generally used to develop the theoretical model (Hair et al., 2012; Henseler et al., 2009).

Measurement model: This model was assessed through Unidimensionality, Loadings and Communalities, Discriminant validity and Cross loading. One needs to evaluate the outer model to estimate the inner model settings.

3.1 Respondent profile and its influence on stress variables

Table 1 shows a brief description of the profile of respondents in terms of gender, nature of degree, graduation and years of study. Out of the 538 respondents, men accounted for 40.9%, while women amounted to59.1 % of total respondents. The majority of the students were from the Commerce department (39.6%), followed by the Science department (38.7%) and Arts department (16.0%), while the rest were from the Education and Management departments. With regards to Education, undergraduate students account for 70.8 %, while postgraduate students were 29.2 %. With regards to the years of study, the largest group was third year under-graduate UG (29.2%) and second year UG (27%) students. First-year UG students accounted for 17.1 %, while the second year PG students accounted for 13.0%. In case of first year post-graduate PG students, the value was 11.3 %. The smallest number of respondents was from the Education department i.e., 2.5 %. These details are tabulated below (Table 1).

Fig. 1(a) shows attending the undergraduate final year students are under a lot of stress from the academic, family, social, personal, health and cognitive environment in comparison with first- and second-year undergraduate students. Final year undergraduate students also suffered from stress during the lockdown period. In line with this, the stress from academic, family, social, personal and cognitive environment was higher for final-year postgraduate students than the first year PG students. Concerning stress from health, levels were almost the same for first- and second-year students. At the end, first- and second-year undergraduate students showed contrasting results. While first-year Education students are under a lot of stress compared to second-year students. The observed stress score for first-year education students was in the range of 2.87 to 3.93, whereas final-year students’ stress levels were in the range of 2.40 to 3.20. This shows a significant difference between first-year education students and second-year students.

Close

Fig. 1

(a): Stress levels of academic, family, social, personal, health and cognitive environment during year of study; (b) Stress levels by male and female students during year of study.

Export to PPT

Stress levels were found to be higher among male students than female students according to the Fig. 1(b). In academic environment, the stress level was slightly higher among women than men. Personal stress remained same for both male and female students. Family, social, health and cognitive types of stress were higher in male students than female students.

3.2 Cluster analysis

In order to identify the similarity and dissimilarity among constructs, the cluster analysis was conducted. The variables were clustered based on the Wald minimum variance measures, using different colors to identify the six constructs. This provides explanatory insight about variables that are linked with constructs. Due to mismatch between one another, few variables are dropped to arrive at the conceptual idea shows the final set of variables in Fig. 2 (a).

Close

Fig. 2

(a): Dendrogram showing similarities and dissimilarities among MVs; (b) Bar chart with mean and error of family, social, personal, health and cognitive environment.

Export to PPT

For every construct, mean and standard deviation were computed on the basis of the average measure of its variables. Fig. 2 (b) shows the error bar chat that was drawn using R language with ggplot2 package. The bars illustrate the mean of every construct, while the head of the bars upside and downside show the plus or minus standard deviation. The constructs, such as academic environment, cognitive, family, personal and social mean values ranged between 3 and 5. This infers the intention of participants’ response towards their stress level. Meanwhile, the health mean value was 2.5, which is less than 3, thus denoting that stress related to health is less than other constructs.

Looking at Fig. 2 (b) it is possible to identify academic environment as the main factor responsible for stress level, followed by family, and personal is the third top factor. We would like also to highlight the weak influence that the health factor has on the level of stress.

3.3 Partial least Square-Structural Equation Modeling (PLS-SEM)

PLS-SEM is one of the non-parametric statistical tools that is generally used to develop the theoretical model. According to some researchers, it exhibits extraordinary statistical outcomes despite the small sample size, since it has no assumption about the distribution of data. However, it should not be far away from the normal so that the model significance can be assessed. The ordinary least square regression-based method is the usual procedure for estimating the parameters in PLS-SEM. PLS-SEM has two types of model constructions, namely the inner model and outer model. The relationship between latent constructs is explained in the inner model (structural model), which is shown in Fig. 3. The relationship between manifest variables and latent construct is explained in the outer model (measurement model). Few constructs, such as academic environment, family, social and personal, were exogenous latent constructs, while the endogenous latent constructs were health and cognitive. According to Fig. 3, the stronger relationships that can be identified are between the constructs” Personal” “Health”,”Personal” “Cognitive” and “Academic environment” “Cognitive”. Also, the weak relationship is between “Academic environment” and “Health”.

Close

Fig. 3

PLS-SEM diagram with regression coefficients of family, social, personal, health and cognitive environment of the students;

Export to PPT

3.4 Measurement model:

This model was assessed through Unidimensionality, Loadings and Communalities, Discriminant validity and Cross loading. One needs to evaluate the outer model to estimate the inner model settings.

3.5 Unidimensional:

Unidimensionality represents the correlation between manifest variables and associated latent variables. When there is a change in the latent variable position (either decrease or increase), then the indicators associated with it also tend to change in the same direction. C. alpha, DG.rho and Eigen values are used to measure the unidimensionality and are illustrated in Table 2.

3.6 Factor loadings and communalities

The correlation between a latent variable and its manifest variables is explained by the loadings of construct. Communality is no more than the square of loadings of a construct. Fig. 4(a) shows the loadings, whereas Fig. 4 (b) shows the communalities. The rule of thumb is that each manifest variable loading should be more than 0.7, which is satisfied in all the manifest variables under specified construct, while communality values above 0.49 is accepted.

Close

Fig. 4

(a) Loadings of each Construct of family, social, personal, health and cognitive environment; (b) Communalities of each construct and its threshold value of family, social, personal, health and cognitive environment.

Export to PPT

The communalities represent the amount of variability explained by a latent variable. When the communality value is greater than 0.5, it denotes that more than 50% of the variability in an indicator is captured by its latent construct. Fig. 4 shows that the current study meets the assumption made herewith.

3.7 Measurement model validity

The measurement model needs to be validated using Average Variance Explained (AVE), Construct Reliability (CR) and Discriminant Validity (DV), expressed as follows:

The AVE values of academic environment, family, social, personal, health and cognitive were 0.603, 0.582, 0.596, 0.632, 0.595 and 0.714, respectively. These observed AVE values were higher than 0.5 (Hair et al., 2017) and Construct Reliability values were higher than 0.7 (Cohen, 1988). Both AVE and CR met the requirement for model performance in the current study. Table 3 shows the Construct Reliability (CR) and Average Variance Extracted (AVE) values.

One more DV form was developed with the help of cross loading of manifest variables. Table 4 shows the cross loading of each manifest variable. There are five manifest variables (MV) which measure highly loaded academic environments ranging from 0.725 to 0.829. Similarly, the four Manifest Variables (MVs) are highly loaded to family and range from 0.700 to 0.787. The seven MVs are highly loaded to social, ranging from 0.747 to 0.826, five MVs are highly loaded to personal and range from 0.738 to 0.847, seven MVs are highly loaded to health and range from 0.704 to 0.818 and five manifest variables are highly loaded to cognitive and range from 0.795 to 0.881. The cross loading of each manifest variable supports that these manifest variables are highly loaded to their underlying construct compared to other constructs. Therefore, these outcomes provide the adequate level of DV of constructs.

3.8 Assessment of Structural model

The Structural Model is explained in Eqs. (4) and (5). The coefficient of each construct was estimated using PLS-SEM procedures. The estimated coefficients are given in Table 5.

Where:

The proposed nine hypotheses must be examined based on the estimated regression coefficients and their significance (Table 5). The significant value of academic environment ( ${\beta }_1^{\ast }$  = 0.0025, t = 0.0641) and social ( ${\beta }_3^{\ast }$  = 0.0817, t = 1.8800) towards health was more than 0.05, yet insignificant at 5% alpha level. This result infers that both academic environment and social factors do impact the health of the respondents. The significant values of family ( ${\beta }_2^{\ast }$  = 0.1960, t = 5.1500^***) and personal ( ${\beta }_4^{\ast }$  = 0.5390, t = 12.400^***) towards health were less than 0.001, yet significant at 0.1 % alpha level. This phenomenon infers that both family and personal impact the health of respondents. Out of these two constructs, personal seems to influence the health of the respondents significantly. The significant values of academic environment ( ${\beta }_1$  = 0.2640, t = 8.2400^***), social ( ${\beta }_3$  = 0.1320, t = 3.7200^***), personal ( ${\beta }_4$  = 0.3630, t = 9.0500^***) and health ( ${\beta }_5$  = 0.1510, t = 4.3100^***) towards cognitive stress were less than 0.001, but highly significant at 0.1 % alpha level. However, the significant value of family ( ${\beta }_2$  = 0.0913, t = 2.8800^**) was less than 0.01, but significant at 1% alpha level. The regression coefficients of structural model were found to be significant at alpha level. However, there is still a requirement to evaluate the overall fitting model using fit indices. To quantify model efficacy, the coefficients of determination and Goodness of Fit (GoF) are the two fit indices generally used. The coefficient of determination is the amount of variance which is shown in the latent endogenous construct by the latent exogenous constructs. The GoF index helps assess the performance of both the outer (measurement) model and the inner model (Structural Model).

The estimated R-square value was 0.527, which underscores the relationship between health and academic environment, family, social and personal. Health is explained through the amount of variance, i.e., 52.7 %,in the constructs academic environment, family, social and personal. Similarly, the estimated R-square value was 0.690, which proves the relationship that exists between cognitive stress and academic environment, family, social, personal, health. The cognitive stress explains the amount of variance, i.e., 69.0%,inconstructs such as academic environment, family, social, personal and health. The GoF index is computed as the geometric mean of the average communality and the average R-square value. The GoF index value 0.6134 was observed for the developed model. The prediction power of the model was good at 61.34 % and it can be concluded that the derived model possesses high explanatory power in terms of performance of both outer and inner models.