Why Do College Students Cheat ? A Structural Equation Modeling Validation of the Theory of Planned Behavior

Cheating on tests is a serious problem in education. The purpose of this study was to test the efficacy of a modified form of the theory of planned behavior (TPB) to predict cheating behavior among a sample of Saudi university students. This study also sought to test the influence of cheating in high school on cheating in college within the framework of the TPB. Analyses were conducted on a sample of 322 undergraduate students using structural equation modeling. The results were consistent with the TPB model’s predictions. The TPB model explained a modest variance in cheating in college. When cheating in high school added to the model, the proportion of explained variance increased and cheating in high school was the best predictor of cheating in college. Although not hypothesized by the TPB, subjective norm had a direct effect on attitude.

The theory of planned behavior (TPB) (Ajzen, 1991) is used to predict a variety of social behaviors.Several studies have investigated the efficacy of the TPB in predicting student cheating behavior (e. g., Beck & Ajzen, 1991;Harding, Mayhew, Finelli, & Carpenter, 2007;Passow, Mayhew, Finelli, Harding, & Carpenter, 2006).Each of these studies found strong support for the TPB's explanation of student cheating behavior.However, no research to date has included all of the constructs in the model nor has their influence been estimated in a causal sequence.Only three studies (Mayhew, Hubbard, Finelli, Harding, & Carpenter, 2009;Stone, Jawahar, & Kisamore, 2009, 2010) provide a comprehensive and specific test of the model in causal sequence using structural equation modeling.However, these studies did not test the modified form of the TPB, with the exception of Mayhew, Hubbard, Finelli, Harding, and Carpenter (2009).While extensive research on cheating has been conducted on a western context, very little is known about the cheating behavior of Arab studentsspecifically of Saudi students.Saudi Arabia is considered both conservative and collectivist society and great importance is placed on the needs, attitudes, and goals of the family rather than on the individual.Islam is the religion of Saudi Arabia and cheating is forbidden in Islam and against Islamic values.Thus, a major goal of the present study is to determine the explanatory power of the modified form of the TPB model in explaining cheating behavior of Saudi students using structural equation modeling.A second purpose of the study is to extend the model by considering an additional construct, high school cheating.It seeks to test the influence of cheating in high school on cheating in college within the framework of the TPB.
Cheating on tests is a serious problem in higher education because of its implications.Cheating violates institutional regulations and damages reputations.It negatively affects the accuracy of the evaluation by adding more sources of errors which decreases the validity of the measures of student learning.In addition, the

Data Analysis
The data were analysed using Structural Equation Modelling (SEM) with Amos 18 software.SEM uses various types of models to depict relationships among variables.Various theoretical models can be tested in SEM to understand how sets of variables define factors and how these factors are related to each other (Schumacker & Lomax, 2004).
Before running the SEM, the data were prepared for the analysis.The negatively worded items were reverse scored.In addition, the data were checked for missing values, outliers, and normality distributions according to the guidelines provided by Tabachnick and Fidell (2001) with SPSS 14 software.
As suggested by Jöreskog (1993) and Anderson and Gerbing (1988), a two-step structural equation modelling procedure was used in estimating parameters: a measurement model followed by a structural model.The measurement model, which is a confirmatory factor analysis, specified the relationships between variables and factors.It provided an assessment of reliability and validity of variables for each factor.The structural model specified the relationships among factors (Schumacker & Lomax, 2004).

Prevalence of Cheating on Tests during College and High School
Percentages of male and female students admitting to cheating on tests during college and high school are shown in Table 1.Nearly half (46.5%) of college students reported never cheating on tests.More than one quarter (29.3%) reported cheating on tests few of the time.The remainder was students who reported cheating on tests about half of the time (14.9%),almost every time (4.5%), every time they had an opportunity (2%), or did not report cheating behavior (2.8%).There is also a significant gender difference (p<.05) in reporting never cheating on tests: 41% of males reporting never cheating on tests compared to 58.6% of females.
Similar patterns were seen in high school.Less than half (44.5%) reported never cheating on tests, and more than one quarter (25.9%) reported cheating on tests few of the time.The remaining 29.6% of students reported cheating on tests half of the time (10.4%),almost every time (7%), every time they had an opportunity (6.5%), or did not report cheating behavior (5.6%).In terms of gender differences, 49.5% of female students reported never cheating on tests compared to 42.2% of male students.This difference is not statistically significant.

Data Screening
In preparation of data for the analysis, data were screened for missing values, outliers, and normality distributions.There were some missing values.Missing values were evaluated with respect to both cases (Table 2) and variables (Table 3).293 cases (82.54%) had valid, non-missing values and 62 cases (17.46%) had missing values.One variable had no missing values (SN1).The two variables with the highest proportion of missing values were cheating in college and cheating in high school with 5.6% and 2.81% of missing cases, respectively.Little's MCAR test was used to assess the pattern of missing values.If the p-value for Little's MCAR test is not significant, then the data can be assumed to be MCAR.Little's MCAR test showed that the missing values can be assumed to be MCAR (χ²= 529.95, df= 559, p= 0.806).There were 17 students did not report their cheating behaviour either during college or high school.Because cheating behaviour variables are important in this study, it was decided to delete them.In addition, it was decided to delete cases with more than 10% of missing values.The number of these cases was 11.Therefore, the remaining data contained of 327 cases.Univariate and multivariate Outliers were detected.To assess univariate outliers, all variables were converted to z scores.Tabachnick and Fidell (2001) recommend considering cases with Z scores higher than 3.29 (p<.001, two-tailed test) to be outliers.All cases were less than 3.28.Multivariate outliers were identified by computing each case's Mahalanobis distance and a case is considered as a multivariate outlier if the probability associated with its D² is 0.001 or less (Tabachnick & Fidell, 2001).Five multivariate outliers were identified and deleted.
After deleting these cases, the remaining data contained of 322 cases.
Normalitiy distribution was assessed using skewness and kurtosis.Tabachnick and Fidell (2001) suggest that skewness and kurtosis values should be within the range of -2 to +2 when the variables are normally distributed.The values ranged between -.06 to 1.51 for skewness and between -1.32 and 1.75 for kurtosis.This indicated that the data is normality distributed.
SEM requires a large sample size.However, there is no agreement on how large a sample size is needed.Anderson and Gerbing (1988) consider sample sizes between 100 and 150 as the minimum for SEM.Kline (1998) recommends that sample sizes below 100 could be considered small, between 100 and 200 cases as medium size and samples that exceed 200 cases could be considered as large.However, models with more parameters require a larger sample.Mueller (1997) suggests that the ratio of the number of cases to the number of variables is recommended to be at least 10:1.
The sample size used in this study meets these recommendations.The sample size is 322.In addition, as there were 50 free parameters and 18 variables in the hypothesis structural model, the ratio of the number of cases to the number of observed variables was 17.8:1.Therefore, the SEM could be conducted without a further problem.

Measurement Models
The measurement model is a confirmatory factor analysis (CFA).It provides an assessment of the reliability and validity of variables for each factor.The CFA was conducted on five factors and 18 items.The results indicated that two items (SN3 and PBC3) had very poor reliabilities as their squared factor loadings were less than 0.15.Thus, the initial model was modified by deleting the two items.The results of the modified model are shown in  146 .25 .28 .29 .46 .31 .18Note.PBC= Perceived Behavioral Control, all correlations are significant at the .001level. ies.ccsenet.

Structu
Structural TPB (see F the modifi subjective perceived Figure 3. significant sensitive t were with acceptable results ind   (Beck & Azjen, 1991;Harding et al., 2007;Mayhew et al., 2009;Passow et al., 2006;Whitley, 1998).This underscores the power of high school behaviour in predicting college behaviour.Research has shown that certain behaviors during college can be predicted by a person's having engaged in them during high school, behaviors that students bring with them to college and that remain unchanged by the college experience (Astin, 1993, Pasceralla & Terenzini, 2005, cited in Mayhew et al., 2009).
Another finding of this study was that subjective norm had a significant direct effect on attitude.This effect was not suggested by the TPB but is consistent with other research (Chang, 1998;Shepherd & O'Keefe, 1984;Vallerand, Deshaies, Cuerrier, Pelletier, & Mongeau, 1992).This implies that a student's attitude toward cheating is affected by what others think about it.According to the TPB, attitude towards a specific behavior is affected by beliefs about the positive and negative consequences of engaging in the behavior.If a student has positive beliefs towards cheating behavior, then the student will form a positive attitude toward cheating behavior.These beliefs are influenced by family, friends, and teachers.
This study has several implications for further research.First, few studies on cheating behavior have used structural equation modeling.It is recommended that future research use this method because it allows complex phenomena to be modeled and tested.Second, consistent with previous research, this study found that the TPB explained only a small proportion of the variance in student cheating behavior.This indicates that at least some important predictors of cheating behavior may not be properly identified by the theory.More research is needed to identify such predictors.The current study focused on factors drawn from the TPB with the addition of cheating in high school.Future research might investigate additional factors such as the role of religion.This might then increase the proportion of variance explained in any model of cheating behavior.Third, findings from previous studies and this current study indicate that a large number of students admitted to cheating and the future research could be directed to examine the efficacy of strategies to prevent cheating such as multiple grading opportunities, spaced seating and monitoring, multiple testing forms, and banning digital technologies.Finally, previous research and the current study found that men were more likely to cheat than women.This may indicate that the process leading to cheating behavior varies for men and women.Future research could examine the efficacy of the TPB across gender.
This study has limitations.It took place at one university; findings may not be generalizable to other populations.Also, the TPB variables were collected using self-report measures which are intrinsically vulnerable to social desirability bias.However, it is likely that this bias was minimal as the participants were assured complete anonymity and confidentiality.Additionally, there is evidence that self-report measures of cheating behavior can yield accurate information (Beck & Ajzen, 1991;Becker, Connolly, Lentz, & Morrision, 2006;Cizek, 1999).Finally, this study focused on cheating on tests.Future research could examine other types of cheating such as cheating on homework or plagiarism.

Table 1 .
Percentages of male and female students who cheat

Table 2 .
The number of missing values by cases Number of cases Number of missing in each cases Percentages of missing in each cases

Table 3 .
The number of missing values by variables

Table 4 .
Although the chi-square of 201.147 with 94 degree of freedom was statistically significant at p<0.001, all other fit indices were within acceptable values (χ²∕df = 2.139; GFI= 0.93; AGFI=0.90;CFI=0.96;RMSEA=0.06).All factors loadings were significant at p<0.001 and ranged from 0.62 to 0.95, indicating that each item was well represented by the factors.Alpha coefficient reliabilities for all factors were well above the cut-point of .70 as suggested byNunnaly and Bernstein (1994).Means, standard deviations and correlations between variables are provided in Table5.All factors were significantly correlated with cheating behavior in college (p<0.001).

Table 4 .
Reliabilities and standardized confirmatory factor loadings for factors