Structural Equation Modelling applied to proposed Statistics Attitudes-Outcomes Model: A case of a University in South Africa

The purpose of the study is to investigate the structural relationships among constructs of the statistics attitudes-outcomes model (SA-OM) using exploratory structural equation modelling (ESEM) methodology. The sample consists of 583 first-year undergraduate students enrolled for statistics courses at the university in South Africa. ESEM reveal that all but two of the nine constructs have well to excellent reliability. To enhance the model, we deleted the eight variables. All other indicators have a significant loading into a construct. Congruency of the SA-OM and expectancy value model (EVM) is noted. The SRMR for all modified models are less than 0.10 suggesting that all these models have acceptable fit. Moreover, all the modified models have RMSE values within the ranges of adequate fit. On the contrary, all the models have unacceptable fit according to PCF, CFI, AGFI and PGFI statistics, i.e. according to all parsimony fit indices except the RMSE. The results also reveal that all incremental fit indices but the BBNFI approve the modified models as acceptable since most of these indices are almost equal to a cut-off point of 0.9. However, BBNNI disapprove the ML3 and ML5 models as being acceptable. A host of inconsistencies in fit indices are noted.


Introduction
Students' achievement in statistics depends heavily on how they learn, understand and apply the course content in their careers of choice. More important is how they perceive the course, the effort they put in, and their ability to deal with its cognitive demands. Generally, one could tell if students have negative attitude towards a course or not. Studies by Carnell (2008), Dempster and McCorry (2009), and Wiberg (2009) suggest timely revision of statistics courses to motivate students. The degeneration of positive to negative perceptions escalate to negative attitudes which inhibit any learning of the course, as well as apprehension and application of the course content. Large bodies of research have focused on these individual constructs' relationship to statistics achievement rather than their interactive and mediating effect on it. Prevention of positive perception, a possibly snowballing into adverse attitudes towards students' statistics achievement becomes imperative. Negative perception often precedes a poor performance in statistics as highlighted by Galli et al. (2010). Very few studies in the area have attempted to uncover causal relationships and covariance among measured and latent variables. This also applies to how development of a positive outlook on statistics can help generate interest, relevance, motivation, effort, and the worth of the course. Most of the studies have focused primarily on relationships between attitudes and achievement, but have not investigated the underlying complex structural relationships (Dempster and McCorry, 2009).
In recent research, factor structure and correlations within structure models have been studied, leaving room for the studies to identify varying variables and constructs. It is for these reasons that an interaction between attitudinal affect and perception constructs, and relevance are viewed as concomitants of statistics achievement. The interaction between these constructs is of interest and there is a need to explore them with a view of suggesting remedial intervention strategies to be implemented at the beginning of every academic year or consistently and concurrently throughout the semesters. Structural Equation Modelling has the capability to do both single-level and multi-level analysis simultaneously, the duty which other first and second generation models failed to do. The method can perform analyses of comparisons of multi-groups using different multivariate techniques such as, multiple regression, multivariate analysis of covariance, confirmatory factor analysis (CFA) to mention a few.

Structural Equation Modelling
(SEM) has the ability to calculate the prediction model and the associated power of analysis, and do path model analysis as well. Thus, SEM in this study helps to remedy deficiencies of past researches on students' attitudes towards statistics by looking at the significance of causal relationships and possible confounding factors that may arise from the model. Specifically, the study aims to confirm underlying perceptions and attitudes constructs derived from ideas gathered from statistics students at a university in South Africa. The primary objective of the study is to determine the effect of students' perceptions and attitudes on statistics achievement or statistics outcome. We hope to be able to analyse causal links between manifest variables and constructs, and further among constructs. Furthermore, the study looks at the relationship between of students' attitudes on their self-efficacy. With SEM, one is able to simultaneously determine parsimony and spuriousness of the final prediction model. Though CFA has been proven to perform similar duties as the SEM (Suhr and Shay, 2009), its limitations have been well documented (see for e.g., Marsh et al., 2004;Hopwood and Donnellan, 2010). This study is expected to contribute to literature by suggesting ways to avert negative attitudes and promote positive perceptions. The results of the study would give an indication of basic constructs that need the instructors' attention. These constructs may be developed in students or used to improve instruction of the statistics subject to ensure students' achievement in related statistics courses. The study will serve as a guide to redesign programmes for statistics for non-majors and for the faculty at this university to relook the entrance requirements for degree programmes offered, paying more attention to how students learn statistics first rather than focus on the pedagogical content of the subject. The study will finally make a significant contribution to the advancement of SEM and its possible application in other disciplines.

Literature Review
Theoretical Framework: Bad experiences are often a precursor to debilitating statistics anxiety effects on academic achievement. Anxiety emanates not from a lack of proper instruction or training, and insufficient skills (Pan and Tang, 2004). Statistics courses are not only a terrible experience to the majority of non-majors, they also pose a threat to completion on time of their degrees (Onwuegbuzie and Wilson, 2003). Students' misperceptions and little or no proper mathematical background is contributory if not intermediary in students' achievement in statistics (Hulsizer and Woolf, 2009;Pan and Tang, 2004). Anxiety inducing factors are classified into three categories according to Baloğlu (2003). These are dispositional, course-related or situational, and person-related factors. Dispositional factors are emotional and psychological traits of students which include their perceptions, attitudes, and mathematical self-concepts according to Baloğlu (2003) and (Dykeman, 2011). Situational factors according to Onwuegbuzie and Wilson (2003), Pan and Tang (2004;2005) and Dykeman (2011) are dependent on whether the course is mandatory or an elective, prior knowledge of statistical course and mathematical content. There is a growing body of research describing the relationship between students' attitudes towards statistics and statistics achievement (Emmioğlu and Çapa-Aydın, 2011;Sorge and Schau, 2002). Most studies focus on one aspect at a time and do not explore covariation or causality. The current study attempts to address this gap.
A meta-analysis study of students' attitudes toward statistics by (Emmioğlu and Çapa-Aydın, 2011) has cited positive relationships between affect and perceived cognitive competence and course grade. Reported by this study is that course value has a small but positive effect on perceived difficulty. It is evident that with a high correlation between Mathematics and Statistics (Onwuegbuzie and Wilson (2003); Onwuegbuzie, 2003), statistics anxiety may possibly be born from existing and prevalent Mathematics anxiety oftentimes accompanied by negative expectations. Many scholars and educators believe that negative perceptions and attitudes towards Statistics are important in a student's academic life. It is for this particular reason that Schau et al. (1995) developed SATS-28 (with four constructs) and later SATS-36 (with six constructs). Dauphinee et al. (1997) further researched the factor structure of the SATS-28 survey instrument. The instruments' constructs are congruent with Eccles' Expectancy-Value Model's (EVM) theoretical framework Eccles et al., 2005). Of interest to note is how Eccles' EVM is consistent with SATS-36 six components and other variables from the MPSP presented in the appendices, emphasizing the multidimensionality of perceptions, attitudes and motivation. The framework aids instructors and researchers alike to determine the factors that directly or indirectly affect the perception and attitudes of students toward statistics, their achievement related choices and the relationships among them. According to Ramirez et al. (2010), SATS-36 complements Eccles' EVM by demonstrating that some of the constructs are relevant to university students' statistics course. The EVM further allows for researchers to determine the interrelation between attitudinal and motivational factors, suggests (Ramirez et al., 2010). This framework additionally gives room for an extension into Statistics and Mathematics domain or other academic domains.
The EVM has been acknowledged as an appropriate instrument and theoretical framework-with its implications in pedagogy, evaluation and research, for investigating the complexity of students' perceptions and attitudes toward statistics subject. There is consistency between the EVM and SATS-36 and selected variables from the MPSP. Ramirez et al. (2010) suggest that when selecting an instrument, instructors and researchers aiming to measure students' perceptions and attitudes towards a statistics subject should consider SATS-36. This is due to its consistency with the EVM and its psychometric properties (Hilton et al., 2004 andTempelaar et al., 2007). Another model of interest in the Statistics Attitudes-Outcome Model (SA-OM) is cited in Ramirez et al. (2012), Emmioğlu and Çapa-Aydın (2012) and Arumugan (2014). The latter applied the PLS algorithm to confirm the SA-OM. The SA-OM is also said to be congruent with many learning theories as it assumes that affective factors, besides cognitive factors play a pivotal role in the students' statistics outcomes attainment. The model is based on the Self-Efficacy theory (Wigfield et al., 2006), Self-Determination Theory (Deci and Ryan, 2002;Wigfield et al., 2006), and The Theory of Planned Behaviour (Ajzen, 2005). SEM tests, simultaneously, the validity of measures in a model and plausibility of theory (Chin, 1998;Gefen, Straub, and Boudreau, 2000). SEM has its foundations deeply rooted in the classical, path analysis (Wright, 1918) and CFA (Jöreskog, 1966). This text follows a two-step approach first described by Anderson and Gerbing (1988) and later Anderson and Gerbing (1992) and highlighted also by Chin (1998), for performing analysis of covariance structures.
The initial step involves developing an acceptable measurement model using CFA, based on theory and a priori specified causal hypotheses (Mueller and Hancock, 2008). The items' statements and sub-scales mentioned above are shown in appendices. Correlations, multiple regression, etc., cannot test for instrument convergent, discriminant, and nomological validity simultaneously, thus researchers have to follow a 'twostep approach' (Gefen et al., 2000). It is for this reason that the SEM method is followed for this study, as it also offers a solution to the two-step approach (Chin, 1998), with its capability to test convergent validity and discriminant validity simultaneously (Chin and Todd, 1995), reduce the likelihood of false negatives (Type 2 errors) (Chin, 1998) and better tests moderators (Lowry and Gaskin, 2014.).

Empirical Evidence:
There is growing evidence on literature that points to the positive effect of perceptions on positive attitudes, with achievement in Statistics course. An SEM model including measures of mathematical aptitude, statistics anxiety and attitudes, motivation to learn statistics, and effort was tested by (Lalonde and Gardner, 1993) in predicting Statistics performance. Their study found that there is a direct positive relationship between aptitude and performance, and the relationship is negative in statistics anxiety, which in turn appeared to be positively related to both motivation and performance. The study also found that the path from Statistics anxiety to performance was statistically non-significant. In a replicated study by (Tremblay, Gardner, and Heipel, 2000), contrary to the expected outcome, there was a significant and negative relationship between statistics anxiety to performance, and a negative path from attitudes to anxiety. Misperception of statistics may give rise to consistent and increased avoidance of the subject. Statistics-related stress develops this avoidance, or poor performance on the subject. Cherney and Cooney (2005) in their study revealed that the lower the Mathematics and Statistics perceptions, the lower the final grade. Students' misperceptions about both Statistics and their Mathematical skill (or lack of it thereof) are due to anxiety and do not necessarily emanate from their limited skills or bad instruction received (Pan and Tang, 2004;Onwuegbuzie and Wilson, 2003). Another relevant work is the Anxiety-Expectation Mediation (AEM) model (Onwuegbuzie, 2003) where both statistics anxiety and achievement's expectation were expected to mediate the relationship between cognitive, personality, and person's characteristics, and performance.
Statistics anxiety and achievement are reported to play a pivotal and significant role in mediating the relationship between performance and anxiety, study behaviour, course load, and the number of statistics courses taken in an academic year (see for e.g. Chiesi and Primi, 2009;Chiesi et al., 2011). (Nasser, 2004) study obtained a high positive effect of mathematical aptitude and a lower, but significant, positive effect of attitudes on performance with a SEM approach. Anxiety was also found to be directly and negatively linked to attitudes and the path between anxiety and performance was non-significant, consistent with Lalonde and Gardner (1993). These findings were in contrast with Onwuegbuzie's (2003) and Tremblay et al. (2000) studies. Chiesi and Primi (2010) proposed an SEM model where mathematical background affects both mathematical knowledge and attitudes toward Statistics. These two variables influenced statistics anxiety, which in turn was directly related to attitudes and performance. The results showed that both post-test attitudes and mathematical knowledge were directly and positively related to performance, but anxiety only indirectly affected performance through attitudes. The SEM approach showed attitude as the stronger direct predictor of performance, and played a full mediating role in determining the relationship between statistics anxiety and performance. Mathematics background also appeared as a negative predictor of anxiety. Finally, test anxiety was a positively direct predictor of statistics anxiety.

Methodology
This section discusses the proposed methods by the study and the related pertinent issues governing them. The section also reports and discusses the results.

Data description and assumptions:
Used in this study is the altered SATS-36 and MPSP self-administered questionnaires as instruments for data collection from the university first-year undergraduate class who availed themselves for Statistics lectures on that particular day. A proportionate stratified random sampling method was used to select respondents from a population of about 1000 students during the 2015 academic year. Respondents were randomly selected from their respective classes. A random sample was selected within each stratum to make up the final sample of statistics students. The instrument was chosen due to the fact that it had recently been used to assess students' perceptions, attitudes and achievements towards statistics (Emmioğlu, 2011). The psychometric properties of SATS-36 are well documented (e.g. Chiesi and Primi, 2009). The subscales are based on Eccles and Wigfield's (1995) EVM and Sorge and Schau's (2002) statistics attitudes-achievement structural model. The hypothesised model ( Figure 1) in the next sections consists of nine latent constructs, as opposed to seven in previous studies.

Sample size:
Structural equation modelling is based on large sample theory (Lehmann, 1999), and minimum sample requirement is asserted to reliably conduct an SEM study. The reader is reminded that the initial step involves finding an acceptable measurement model using confirmatory factor analysis (CFA) (Mueller and Hancock, 2008). Some authors recommend a sample of minimum 200 observations while others say a threshold of 300 is more appropriate. According to Hair et al. (2010), a sample size of 50 observations suffices under ideal conditions using maximum likelihood estimation (MLE). However, Hair et al. highlighted that a sample size of 200 provide a sound basis for estimation. Researchers vary in their elucidations, in terms of the number of cases required. Ratios of 5:1 (about 200 cases) and 2:1 (about 100 cases) respectively of subjects to variable are proposed by Kline (2011). The sample size for the present study is 583 from four first year Statistics classes. This sample size is in agreement with Hair et al. sample size requirements and fulfils the minimum sample and ratio of subjects per variable requirements (Kline, 2011). The observed KMO 0.884 reported in Table 1 further confirmed that the sample used in this study is adequate and suitable for the proposed methods. Pett et al. (2003) recommended a sample adequacy measure of at least 0.70 to less than 0.80 as middle. Values in excess of 0.80 are meritorious. Multicollinearity and Factorability: One other assumption concerns the extent of correlations between the variables. Though a certain degree of collinearity between the variables is allowed, it is of importance to protect the assumption of multicollinearity. This assumption may not be valid as far as social science, psychometric or psychological data is concerned. This study analyses a psychometric or psychological data and did not check the violation of this assumption as it is highly expected that some or most of the variables are highly correlated. A correlation matrix is not invertible if its determinant is equal to zero, that is, it is not factorable as cited in (Field, 2013). Shown in Table 1 also is that the determinant of the correlation matrix is not zero, but very close to zero. This suggests a hint of multicollinearity but at the same time confirms that the data is factorable.
Normality Assumption: The univariate analysis of data was run, to determine the skewness and kurtosis, this includes their normality. All variables showed no issues with regard to skewness (< 2) and kurtosis < 3, although most variables have a negative kurtosis (platykurtic), that is, the responses are highly dispersed or most responses share about the same amount of frequency. This kind of distribution is less peaked than the mesokurtic (normal) distribution. Table 2 shows the mean scaled univariate kurtosis as -0.2807, smaller than lower bound of -0.0392. This is mainly due to leptokurtic cases where most respondents selected four for "Neither agree nor disagree" for most of the items, and sometimes sparseness in responses throughout the questionnaire. Multivariate kurtosis given by Mardia's Multivariate Kurtosis equals 371.7651 in Table 1 also reveals that the data were multivariate kurtose, that is, they presented a multivariable 226 deviation from normality, as this value is expected to be as close to zero as possible. These are no surprising result given the type of data from the behavioural sciences, which is psychometric or 226 psychological in nature.   (Schau, 2003a) that each construct has three or more variables per construct (identified constructs). The exception to this condition is mentioned in Hair et al. (2010), that is, an unidentified construct maybe be seen in analysis involving large samples with many constructs. O'Rourke and Hatcher (2013) outlines the rules to be followed in specifying the model (see Saris et al., 2009).

Model Identification:
A topic believed to be most often confusing and difficult in SEM is identification. According to (Suhr, 2006), a model is said to be identified if there exist numerical solutions for every parameter in the model. Estimation is only possible with identified models. A hypothesized model where it is possible to express parameters as functions of variance and covariance of the observed or manifest variables is said to be identified. To attain a solution, number of known parameters in the sample covariance matrix must be greater than or equal to the number of free parameters q. This is denoted as * ( +1) 2 where p is the number of measured variables as per sample covariance matrix, ≤ * . A model is just-identified if = * , that is, a parameter can be estimated through only a single manipulation of the data. If the number of estimated parameters is greater than the unique variance and covariance, where > * , the model is underidentified. This implies, one or more unique estimations of free parameters are possible from the data. When a model is over-identified, < * , it implies that free parameters can be estimated in multiple ways from the sample data.

Parameter estimation:
The sole aim of parameter estimation is to obtain numerical values for free (unknown) parameters through an estimation technique, as discussed in (Bollen, 1989) for a technical approach, and Ullman (2006) for an applied viewpoint. In SEM, MLE method is widely used. This method is adopted in this study as the sample parameter estimates are approximate in value to the population parameters and robust. The discrepancy function criterion for parameter estimation aims to optimize the difference between the population as estimated sample covariance estimates and the covariance matrix derived from the hypothesised model. The error variance of each latent variable and the parameter value associated with one of the paths from the latent variable to the indicator variable should be fixed at 1.0, suggests (O'Rourke and Hatcher, 2013). The disturbance is on the endogenous latent variable. According to Suhr (2006), the significance of the variable is set at the 0.05, and the value should exceed t ≥ 1.96 or 0.10 level of significance if t ≥ 2.56. Effect size or weight is the value of the path coefficient. Standardised coefficients ≤ 0.10 may indicate a small effect, values ± 0.30 a medium effect, and ≥ 0.50 a large effect.

Measurement Model Validity:
The first stage of SEM is CFA, where a measurement model is estimated through the parameter estimation model and evaluated via the Cronbach's alpha threshold of α ≥ 0.70. This includes convergence, discriminant, construct and nomological validity. Cronbach's alpha ≥ 0.70 are employed by the study to assess convergent reliability. Discriminant validity is defined as the extent to which constructs are distinct from one another (Hair et al., 2006); this includes a simple factor structure with virtually no cross-loadings. This is assessed through inter-factor correlations. Nomological validity is the degree to which a construct correlates or covariance with other construct with the structural model (Bagozzi, 1980). It is established when the correlations or covariance within the structural model makes contextual sense (Hair et al., 2006). If the initial or hypothesized model (CFA) fails, i.e. no clear factor structure, EFA is performed followed by iterative exploratory structural equation modelling (ESEM until the best model is selected using the model fit indices. Model fit assessment: A model of best fit is recognized by looking at individual fit indices for each model during analysis. Several types of these indices have been distinguished by many authors in the last three decades (see Mulaik et al., 1989;Bentler, 1990;Tanaka, 1993;Byrne, 1998), among others. The ability of the hypothesized or a priori model to reproduce the covariance structure of variables of interest is assessed by absolute fit indices (McDonald and Ho, 2002). The other batch of indices concerned with the ability of the hypothesized model to account for the sample data relative to a restricted model which is less complex are called comparative (Miles and Shevlin, 2007) or incremental fit indices. The final batch selects or identifies better fit of a model by an increment of the number of estimated parameters. These are termed parsimonious fit indices. Steiger and Lind (1980) proposed the root mean squared error approximation (RMSEA), which measures the discrepancy between the fitted model and the inferenced covariance matrix of the population. RMSEA value falls within a certain specified confidence value. The model fit indices that are widely used presented in Appendix B, adapted from Hair et al. (2010) show the goodness-of-fit across different model situations. These were based on simulation research that considers model complexity, degrees of error in model specification for different sample sizes (Marsh et al., 2004).

Results
We run Confirmatory factor analysis on the proposed statistics attitudes-outcomes model (SA-OM) and SATS-36 structure. The results summarised in Table 4 are for the hypothesised and modified models. The usual cutoff criterion is used, and is set at absolute 0.3 for factor loadings as cited in (Howell et al., 2012). SAS 9.3 PROC FACTOR with PROMAX rotation resulted in PCF < 0.001; implying that the null hypothesis that the hypothesised model fits the sample data be rejected in favour of the alternative hypothesis that it does not. The null hypothesis asserts acceptable to good fit if > 0.80 1.00, respectively, but bad to unacceptable fit if < 0.80. The results discussed here are with regard to the estimated model as discussed in previous sections.

Hypothesized model:
The researcher used MLE to obtain all models in Table 4. Table 4 also presents model assessment indices obtained from the SAS 9.3. According to the output, the hypothesized model indices fall way below the acceptable levels. A χ 2 value < 0.0001 (significant) implies bad model fit. The model has an SRMR that is > 0.08, RMSEA is > 0.05 but less than 0.08 suggesting adequate fit, the probability of close fit (PCF) is < 0.0001 that is a sign of no fit at all. The other fit indices such as the GFI = 0.7040, AGFI = 0.6701, CFI = 0.6783 and BBNFI = 0.6145 are below the suggested acceptance value of 0.90. Owing to the discrepancies reported by the results with only one of the indices (RMSEA) suggesting good fit whilst the rest suggest otherwise. The study concluded that the hypothesised model does not fit the sample data. The next step was to run an exploratory factor analysis (EFA) to determine the factor structure of the sample data. A simple sample factor matrix calculated has eight of the forty-nine redundant loadings, that is, with loadings less than the cut-off value of |0, 3|. The researcher deleted all such loadings as a result. The study used significant constructs of EFA to fit a modified ESEM model. The results of the modified constructs were found to be acceptable (coefficients more than 0.7) except the "Difficulty' and "Value" constructs which were reported to be mediocre according to Cronbach and Shavelson (2004). Modified models: This step is a result of under-fitting or over-fitting the hypothesised structural model. The step involves either truncating or re-specifying the model, deleting cases or deleting variables that were marked during the preliminary data analysis process. Due to the unacceptable fit of the hypothesized model. We performed EFA to determine the factor structure of the sample data. We made improvements to the hypothesised model as reported in Table 4. All modified models (ESEM) (ML1-ML7) have a significant chisquare value < 0.0001, implying bad fit. The SRMR < 0.08 for acceptable to good fit, whilst the RMSEA ranges between 0.05 and 0.08 for good fit. The result 0.60 ≤ < 0.0001 with a small probability implies bad to no fit, whereas = 0.6684 or 66.84% suggests a better fit than the rest of the models. The fit indices range from 0.8400 to 0.8806, obtained from deletion of observations that contributed immensely to the overall negative impact of the kurtosis on data. From the results in Table 4, SRMR for all modified models are less than 0.10 suggesting that all these models have acceptable fit (Hu and Bentler, 1995). Moreover all the modified models have RMSE values within the ranges of adequate fit. On the contrary, all the models have unacceptable fit according to PCF, CFI, AGFI and PGFI statistics, i.e. according to all parsimony fit indices except the RMSE. The results also reveal that all incremental fit indexes but the BBNFI approve the modified models as acceptable since most of these indexes are almost equal to a cut-off point of 0.9. However, BBNNI disapprove the ML3 and ML5 models as being acceptable. The overall best model is not clear from the results due to inconsistent modification indices across all competing models. Since all standardized linear equations for ML6 (see Appendix C) are statistically significant with 6.7275 ≤ ≤ 25.2993, this means all the indicators of this model are a real measure of the constructs in which they load. The standard errors of the paths equations range between 0.0164 ≤ standard error ≤ 0.0526 for ML6. The standardized covariance among exogenous (Factors or Latent) variables represent covariance among constructs in ML6. These showed only two statistically insignificant of the model. The results (in Appendix D) suggest that there is no significant covariance ( = −0.37519) between Factor 1 (Affect) and Factor 5 (Difficulty) and ( = 0.30096) between Factor 4 (Effort) and Factor 7 (Anxiety). This makes ML6 the best candidate model due to fewer statistically insignificant linear equations between indicator and latent variables. These results suggest that when one's affect deteriorates, then statistics subject content becomes difficult. It also means that the more effort students put into dealing with the statistics content, the lesser the anxiety they experience.

Discussion of Findings:
This section discusses the findings in relation to the objectives and the hypotheses. Primarily, this study applied the SEM to investigate the effect of students' perceptions and attitudes on their self-efficacy and statistics outcome. Exploratory and confirmatory factor analyses were applied to the sample data collected during the second semester of the 2015 academic year. Respondents were 538 students registered for first year Statistics modules at the university in South Africa. The findings of this study are intended to give guidance to instructors, policy makers and researchers in identifying variables (including somewhat redundant variables) that need to be developed further, monitored and require student-instructor attention. Moreover, the study has identified shortcomings of large data volumes with many observed variables.

Research Question 1: Is there is a relationship between the perceptual and attitudinal constructs?
To answer this question, the study considered the SA-OM by Ramirez et al. (2012) which has been deemed congruent to EVM and The Theory of Planned behaviour for selecting variables for use in the study. The study identified the relationships between students' attitudes, their statistics self-efficacy and statistics outcome. The construct validity results have shown the difference in factor structures between the SA-OM and SATS-36. From the twenty-three hypotheses tested, only fourteen were supported and were found to be statistically significant owing to the t-values obtained. Construct and convergent validity and reliability of the measures were also examined. According to theory and literature, all the constructs are related. These findings bring a new dimension to the students' attitudes research area, as far as South Africa is concerned.

Research Question 2: Is there a relationship between students' attitudes and their self-efficacy?
The results showed statistical significant relationships between all students' attitudinal indicators measuring their self-efficacy construct. The results confirmed relationship between current statistics self-efficacy and statistics outcome measures and this is consistent with Bandura's (1996) claim as cited in (Finney and Schraw, 2003). This study highlights the fact that students with high self-efficacy may see the relevance of the subject in the degree programme of their choice.

Index
Hypothesised model

Research Question 3: Does Statistics Anxiety have a negative effect on a students' Statistics Outcome?
Regards to this question, it has been established according to the findings that there is virtually no effect of anxiety on statistics outcomes. This is in contradiction to the findings by Onwuegbuzie (2000;, Pan and Tang (2005), Keeley et al. (2008), and Hamid and Sulaiman (2014) that statistics anxiety is a strong predictor of academic performance. It is important to note that a decrease in anxiety throughout a student's academic life has an impact in terms of the heightened perceived usefulness and statistics achievement.
Recommendations: Based on the findings of this study, the following recommendations should be considered.

Implications for Further Research:
The current study was based on undergraduate students enrolled in different sections of statistics courses in a public university in South Africa. Therefore, the results do not make general conclusion to the target population of the study outside the country and possibly to the world or other universities in the country. It is suggested that further studies should examine structural relationships in a nation-wide context; so that, the hypothesised relationships can be further generalised by extending the current study to different student populations in developing countries. In addition, the authors suggested that further studies should make cross-cultural comparisons as also suggested by Mvududu's (2003). This may generate new insight into statistical pedagogy. The current study also suggests further studies to use direct measures, such as Mathematics and Statistics tests and quizzes, for assessing students' self-efficacy and statistics outcome. This is consistent with Arumugam's (2014) findings that mathematics achievement and self-efficacy both have a positive significant relationship with statistics outcomes.
Based on the findings we propose to the university to monitor students' attitudes continuously. The proposed relationships generated by the results were static rather than longitudinal. It is suggested that further research should expand on the current study by using a longitudinal design in which the data are collected prior to, during, and after taking statistics courses. For example, Eccles and her colleagues Wigfield and Eccles, 2000; suggested that individuals' performances and achievement choices influence their previous achievement-related experiences across time. We suggest further studies that will test their proposal to extend the findings of the current study. Although, the current study revealed that the hypothesized "Statistics Attitudes-Outcomes Model" did not fit well to the data, it does not mean that this model is the only best possible model. We further suggest methodological research to determine the effect of sample size, number of constructs, under-identification of constructs, free parameters and observed variables on the fit of the measurement and structural models. Further studies may also evaluate redundant variables of a valid construct; construct interactions and moderations to determine the gravity of their effect on the overall SEM performance.
We suggest that further research should investigate alternative models. The current study based the hypothesised and tested "Statistics Attitudes-Outcomes Model" on Eccles and colleagues' application of expectancy value theory to mathematics education (Eccles and Wigfield, 1995;Wigfield and Eccles, 2000). The study also highlighted imminent congruence between the Eccles' Expectancy-Value Framework and the SATS-36 and some MPSP constructs. The data used in this paper explained the adaptation of Eccles' Model to the statistics education context very well. Further studies could expand on the present findings by adapting "Statistics Attitudes-Outcomes Model" to different subject domains. Due to the lack-of-good fit of the model to the data, the study suggest further research be carried out to determine the effect of sample size, validity and reliability on the overall global model fit. This paper used ML and suggests that further research be done using real life data and different estimation methods such as, the weighted least squares (WLS), full information maximum likelihood (FIML), and/or asymptotic distribution free method (ADF).

Implications Practice of Statistics Education:
This study has demonstrated that affective domain is important for explaining students' statistics outcomes. As a result, the university should give students' attitudes a high priority when designing and implementing statistics curricula. It is highly important to suggest that students' positive attitudes toward statistics should be among the main goals of the statistics education; and accordingly, a statistics curriculum should involve various instructional practices, which enhance students' positive attitudes toward statistics. The authors also suggest the evaluation of effectiveness of the statistics curriculum by assessing students' attitudes toward Statistics as well as by assessing short term and long-term outcomes.
Statistics instructors in universities should enrol for training that will inform them about the importance of their students' attitudes. This also includes how to implement and evaluate the instruction in a way aimed at enhance students' positive attitudes. The study formulates suggestions for instructors to give statistics activities that are interesting, enjoyable and fun for students to participate. This would help students to have more interest and positive affect toward statistics (Lesser and Pearl, 2008). These may range from possible aptitude checks prior to students' registration, continuous remedial mathematics skills classes, and development of students' interest, worth or value, and appreciation of statistics by giving immediate feedback and have them provide solutions to real-life problems statistically in groups and give feedback as suggested by numerous researchers. I plan to attend every statistics class session Self-Efficacy SE_1 Can identify the scale of measurement for a variable SE_2 Can identify if a distribution is skewed. SE_3 Can select the correct statistical procedure to be used to answer a question SE_4 Can communicate statistical results without any problem SE_5 Can read a value from any statistical table. SE_6 I am confident that I have mastered introductory statistics material up to this point in the present academic year SO_1 If I could, I would choose to take another statistics module SO_2 In the field in which I hope to be employed when I finish school, I will use statistics. SO_3 As I complete the remainder of my degree program, I will often use statistics * Added items