Academia.eduAcademia.edu

Figure from:Cyber-victimization Trends in Trinidad & Tobago: The Results of An Empirical Research

See full PDFdownloadDownload figure
Note. Area Under Curve (AUC) is calculated for every class against all other classes. The “importance” ranking of the variables in the random forest model for cyberbullying is illustrated in Figure 3. The graphic indicates that the cyberbullying class variable is best predicted by the following variables: posting personal information online, online shopping hours, social media hours, use of spam fil. ters, use of pop-up blockers, knowledge of victimization risk/awareness of cybercrime, age, and sex. Notably these variables represent all three aspects of the RAT tested in this study i.e. target exposure, target acces: sibility and capable guardianship.

Figure 3 Note. Area Under Curve (AUC) is calculated for every class against all other classes. The “importance” ranking of the variables in the random forest model for cyberbullying is illustrated in Figure 3. The graphic indicates that the cyberbullying class variable is best predicted by the following variables: posting personal information online, online shopping hours, social media hours, use of spam fil. ters, use of pop-up blockers, knowledge of victimization risk/awareness of cybercrime, age, and sex. Notably these variables represent all three aspects of the RAT tested in this study i.e. target exposure, target acces: sibility and capable guardianship.

Related Figures (13)

Table 1. Cronbach’s Alpha Coefficient for Reliability Analysis Demographics
Table 1. Cronbach’s Alpha Coefficient for Reliability Analysis Demographics
Table 2. Distribution of Participants by Gender International Journal of Cybersecurity Intelligence and Cybercrime, Vol. 4, Iss. 1, Page. 46-63, Publication date: March 2021.
Table 2. Distribution of Participants by Gender International Journal of Cybersecurity Intelligence and Cybercrime, Vol. 4, Iss. 1, Page. 46-63, Publication date: March 2021.
Figure 1. Sequential Analysis at 95% Confidence Interval with Robustness Check Note. 1)Respondents were all over 18, however, for comparability to national statistics grouping the age group was written as 15-24. 2) For the test, the alternative hypothesis specifies that the population mean is different from 34.3.
Figure 1. Sequential Analysis at 95% Confidence Interval with Robustness Check Note. 1)Respondents were all over 18, however, for comparability to national statistics grouping the age group was written as 15-24. 2) For the test, the alternative hypothesis specifies that the population mean is different from 34.3.
All age groups were represented in the sample, with the highest number of participants coming from the ‘35-44” age group (82 participants) and the lowest was the “55 and over” age group (7 participants). This information is summarized in Table 3. The mean age of the participants was found to be 34.8, and this was compared to the population mean of 34.3 using a Bayesian one-sample t-test. The results of the t-test are given in Table 4 and a graph showing the sequential analysis of the t-test with a robustness check (sensitiv- ity prior to selection) is also presented in Figure 1. The Bayes Factor (BF01) of 7.998 shows that the sample data is statistically similar to the population and the sequential analysis shows that the sample size was more than sufficient to accurately select the null hypothesis of Ho: p1 = p2.
All age groups were represented in the sample, with the highest number of participants coming from the ‘35-44” age group (82 participants) and the lowest was the “55 and over” age group (7 participants). This information is summarized in Table 3. The mean age of the participants was found to be 34.8, and this was compared to the population mean of 34.3 using a Bayesian one-sample t-test. The results of the t-test are given in Table 4 and a graph showing the sequential analysis of the t-test with a robustness check (sensitiv- ity prior to selection) is also presented in Figure 1. The Bayes Factor (BF01) of 7.998 shows that the sample data is statistically similar to the population and the sequential analysis shows that the sample size was more than sufficient to accurately select the null hypothesis of Ho: p1 = p2.
Table 5. Distribution of Cybercrime Victims Split by Demographic Classes For both unauthorized access and cyberbullying female participants were twice as likely to report vic- timization than their male counterparts. The intraclass analysis shows that this observation holds partic- ularly true for cyberbully as the prevalence among females is twice the prevalence among males. Cases of cyberbullying were particularly high among persons in the “15-24” age group. The lowest instances of cyber- bullying were witnessed in the “55 and over” age group. In relation to unauthorized access the occurrences were spread among the age groups observed. However, notably the age group of “45-54” showed the highest relative prevalence with 82% of the persons within that age bracket reporting victimization. Overall, from the initial analysis it appears that cyberbullying is most prevalent among young females. Unauthorized ac- cess is also most prevalent among females, however, age does not appear to be necessarily a key factor. Table 5 provides the distribution of cybercrime victims split by the classes of respective demograph- ic variables of sex and age for each type of cybercrime observed. The table provides the interclass percent- age — this is the percentage of cybercrime victims among/between all classes for that variable — and the intra-class percentage, which is the percentage of cybercrime within the class for that variable. The in- tra-class percentage is provided as a means of normalizing the data to compare risk in each class because the classes were not equally represented.
Table 5. Distribution of Cybercrime Victims Split by Demographic Classes For both unauthorized access and cyberbullying female participants were twice as likely to report vic- timization than their male counterparts. The intraclass analysis shows that this observation holds partic- ularly true for cyberbully as the prevalence among females is twice the prevalence among males. Cases of cyberbullying were particularly high among persons in the “15-24” age group. The lowest instances of cyber- bullying were witnessed in the “55 and over” age group. In relation to unauthorized access the occurrences were spread among the age groups observed. However, notably the age group of “45-54” showed the highest relative prevalence with 82% of the persons within that age bracket reporting victimization. Overall, from the initial analysis it appears that cyberbullying is most prevalent among young females. Unauthorized ac- cess is also most prevalent among females, however, age does not appear to be necessarily a key factor. Table 5 provides the distribution of cybercrime victims split by the classes of respective demograph- ic variables of sex and age for each type of cybercrime observed. The table provides the interclass percent- age — this is the percentage of cybercrime victims among/between all classes for that variable — and the intra-class percentage, which is the percentage of cybercrime within the class for that variable. The in- tra-class percentage is provided as a means of normalizing the data to compare risk in each class because the classes were not equally represented.
Table 6. Abbreviations Used to Represent the Variables Under Review in this Research Unauthorized access victimization. We investigated the RAT based are fundamental attributes that the existing studies mentioned above a accuracy, the classifier predicted victimization with about 71.4% accuracy ever, the accuracy score in the test data (hold-out data) was only 58.8% model wil gests that evaluation metrics scores for individual classes and overall predict flect the e el can predict all of the negative class but only approximately 50% of all t ffect of unbalanced datasets on these scalar measures and why so tried to predict. Looking at see Table 7). This means that he data instance predicted will predictions in detail because these the during the validation phase, how- the classify 58.8% of a dataset, which was not used to train the mode i.e. real-world usage. This sug- the model is a relatively poor predictor of unauthorized access victimization. Table 8. shows ions. The Precision and recall scores they may be misleading. The mod- the re- De- long to the negative class. Similarly, the model will accurately classify all instances of the positive class but will only detect 12.5% of the instances with a positive outcome in the given dataset. Therefore, the overall model based on the RAT performs poorly as is shown by the low F1-Score, MCC and AUC. Specifically, the AUC score of 0.5 suggests that the model and hence the RAT as operationalized in this study has no dis- criminating capacity to distinguish between the classes.
Table 6. Abbreviations Used to Represent the Variables Under Review in this Research Unauthorized access victimization. We investigated the RAT based are fundamental attributes that the existing studies mentioned above a accuracy, the classifier predicted victimization with about 71.4% accuracy ever, the accuracy score in the test data (hold-out data) was only 58.8% model wil gests that evaluation metrics scores for individual classes and overall predict flect the e el can predict all of the negative class but only approximately 50% of all t ffect of unbalanced datasets on these scalar measures and why so tried to predict. Looking at see Table 7). This means that he data instance predicted will predictions in detail because these the during the validation phase, how- the classify 58.8% of a dataset, which was not used to train the mode i.e. real-world usage. This sug- the model is a relatively poor predictor of unauthorized access victimization. Table 8. shows ions. The Precision and recall scores they may be misleading. The mod- the re- De- long to the negative class. Similarly, the model will accurately classify all instances of the positive class but will only detect 12.5% of the instances with a positive outcome in the given dataset. Therefore, the overall model based on the RAT performs poorly as is shown by the low F1-Score, MCC and AUC. Specifically, the AUC score of 0.5 suggests that the model and hence the RAT as operationalized in this study has no dis- criminating capacity to distinguish between the classes.
Table 7. Overview of Random Forest Classification Model for Unauthorized Acces
Table 7. Overview of Random Forest Classification Model for Unauthorized Acces
Figure 2. Evaluation Metrics Scores for Random Forest Classification Model for Unauthorized Access While the effectiveness of the model is not simply a cumulative sum of different parts, it is possible to identify the predictors that are functional and most important. This is done by observing how the over- all predictive power of the model drops if a specific variable is removed from the model i.e. mean decrease in accuracy. The “importance” ranking of the variables is given graphically in Figure 2. The shaded bars in Figure 2 represent the predictors, with those right of 0 being more important and causing a decrease in model performance. Therefore, a parsimonious model may contain those predictors given to the left of the 0 mark as removal of the other variables will not negatively affect model performance. To see how strongly and in which direction the predictors are associated with victimization, we further investigated the coefficients and odds ratios given by logistic regression. We used the logistic regression classifier because, unlike tree-based methods it not only shows the strength of the predictors, but also the sign/direction of their effect. Table 9 gives the summary of the hypothesis test for the Logistic Regression, which suggests that at the 95% confidence interval that the RAT is not associated with unauthorized ac- cess. This aligns with the output of the random forest analysis.
Figure 2. Evaluation Metrics Scores for Random Forest Classification Model for Unauthorized Access While the effectiveness of the model is not simply a cumulative sum of different parts, it is possible to identify the predictors that are functional and most important. This is done by observing how the over- all predictive power of the model drops if a specific variable is removed from the model i.e. mean decrease in accuracy. The “importance” ranking of the variables is given graphically in Figure 2. The shaded bars in Figure 2 represent the predictors, with those right of 0 being more important and causing a decrease in model performance. Therefore, a parsimonious model may contain those predictors given to the left of the 0 mark as removal of the other variables will not negatively affect model performance. To see how strongly and in which direction the predictors are associated with victimization, we further investigated the coefficients and odds ratios given by logistic regression. We used the logistic regression classifier because, unlike tree-based methods it not only shows the strength of the predictors, but also the sign/direction of their effect. Table 9 gives the summary of the hypothesis test for the Logistic Regression, which suggests that at the 95% confidence interval that the RAT is not associated with unauthorized ac- cess. This aligns with the output of the random forest analysis.
Note. Area Under Curve (AUC) is calculated for every class against all other classes. Table 8. Evaluation Metrics Scores for Random Forest Classification Model for Unauthorized Access
Note. Area Under Curve (AUC) is calculated for every class against all other classes. Table 8. Evaluation Metrics Scores for Random Forest Classification Model for Unauthorized Access
Table 9. Overview of Logistic Regression Analysis for Victimization by Unauthorized Access Cyberbullying victimization. The estimated classification accuracy of the best model found during the validation phase was 85.7%. However, the accuracy of the model on the test data was 94.1% (see Table 10). Table 11. shows the evaluation metrics scores for individual classes and overall predictions in relation (o cyberbullying. The model shows high recall and precision in relation to both classes. It should be noted that while the model will identify all instances of the positive class, the overall class predictive accuracy will be 66.7% i.e. 66.7% of person predicted to become a cyberbullying victim will have succumb to this form of criminality. Given the high overall accuracy of the model, the mid-range accuracy level suggests that the model is better at predicting true negatives than true positives 1.e. it is ‘easier to predict who will not be a victim rather than who will be one. This means avoiding set behaviours can decrease risk, however, engag- ing in risky online behavior does not guarantee a victimization i.e. there is a degree of randomness. Howev- 2r, controlled behaviours can decrease risk. The F1-Score, MCC and AUC scores are all high showing that the model for cyberbullying based on the RAT has strong predictive power. Specifically, the AUC score of ).815 suggests that the model performs well over a wide range of threshold values. Further, the MCC score of 0.789 shows that the model is good at accurately predicting both classes.
Table 9. Overview of Logistic Regression Analysis for Victimization by Unauthorized Access Cyberbullying victimization. The estimated classification accuracy of the best model found during the validation phase was 85.7%. However, the accuracy of the model on the test data was 94.1% (see Table 10). Table 11. shows the evaluation metrics scores for individual classes and overall predictions in relation (o cyberbullying. The model shows high recall and precision in relation to both classes. It should be noted that while the model will identify all instances of the positive class, the overall class predictive accuracy will be 66.7% i.e. 66.7% of person predicted to become a cyberbullying victim will have succumb to this form of criminality. Given the high overall accuracy of the model, the mid-range accuracy level suggests that the model is better at predicting true negatives than true positives 1.e. it is ‘easier to predict who will not be a victim rather than who will be one. This means avoiding set behaviours can decrease risk, however, engag- ing in risky online behavior does not guarantee a victimization i.e. there is a degree of randomness. Howev- 2r, controlled behaviours can decrease risk. The F1-Score, MCC and AUC scores are all high showing that the model for cyberbullying based on the RAT has strong predictive power. Specifically, the AUC score of ).815 suggests that the model performs well over a wide range of threshold values. Further, the MCC score of 0.789 shows that the model is good at accurately predicting both classes.
Table 10. Overview of Random Forest Classification Model for Cyberbullying Table 11. Evaluation Metrics Scores for Random Forest Classification Model for Cyberbullying
Table 10. Overview of Random Forest Classification Model for Cyberbullying Table 11. Evaluation Metrics Scores for Random Forest Classification Model for Cyberbullying
Figure 8. The Estimated Effect that Each Variable has on the Predictive Accuracy of the Cyberbullying Model Cyber-victimization Trends in Trinidad & Tobago Table 12 gives the summary of the hypothesis test for the logistic regression for cyberbullying which suggests that at the 95% confidence interval that the tenets of the RAT are associated with cyberbullying. In other words, routine activities are predictive of cyberbullying, and that by extension the RAT is capable of explaining cyberbullying victimization. Further, analysis of the regression coefficients show that posting personal information online, social media hours and of cybercrime victimization risk/awareness increase risk of victimization. The fact that cybercrime victimization risk/awareness increases risk is counter-intu- itive and may be related to the fact that the study is cross-sectional i.e. victims seeks knowledge about cy- bercrime after victimization. As expected from a theoretical understanding of the role of guardianship spam filters and pop-up blockers were found to be protective in relation to cyberbullying. Strangely, online shop- ping, which was expected to be a risk factor was found to have a protective effect. Examination of the odds ratio suggests that posting personal information online has the greatest effect on cyberbullying risk, in- creasing its probability approximately 6-fold. Further, being female increased risk by a factor of 4 and being in the age group of 25-34 increased risk by a factor of 5.
Figure 8. The Estimated Effect that Each Variable has on the Predictive Accuracy of the Cyberbullying Model Cyber-victimization Trends in Trinidad & Tobago Table 12 gives the summary of the hypothesis test for the logistic regression for cyberbullying which suggests that at the 95% confidence interval that the tenets of the RAT are associated with cyberbullying. In other words, routine activities are predictive of cyberbullying, and that by extension the RAT is capable of explaining cyberbullying victimization. Further, analysis of the regression coefficients show that posting personal information online, social media hours and of cybercrime victimization risk/awareness increase risk of victimization. The fact that cybercrime victimization risk/awareness increases risk is counter-intu- itive and may be related to the fact that the study is cross-sectional i.e. victims seeks knowledge about cy- bercrime after victimization. As expected from a theoretical understanding of the role of guardianship spam filters and pop-up blockers were found to be protective in relation to cyberbullying. Strangely, online shop- ping, which was expected to be a risk factor was found to have a protective effect. Examination of the odds ratio suggests that posting personal information online has the greatest effect on cyberbullying risk, in- creasing its probability approximately 6-fold. Further, being female increased risk by a factor of 4 and being in the age group of 25-34 increased risk by a factor of 5.
Tablel2. Overview of Logistic Regression Analysis for Victimization by Cyberbullying Discussion
Tablel2. Overview of Logistic Regression Analysis for Victimization by Cyberbullying Discussion