| Dissertation |
Thesis (M.Sc.) --NUI, 2013 at Department of Statistics, UCC. |
| Summary |
Chapter one used text mining on social media data about student’s university course choices. Linear regression analysed student sentiment against CAO points and annual changes in student sentiment against annual changes in CAO points for 2007-2012 for undenominated science, arts, law, commerce and engineering courses. Science had a positive, significant relationship for student sentiment. Law had a significant, positive relationship for the annual changes in student sentiment. Other results were not significant. Sentiment measures showed potential for tracking larger trends and annual changes showed potential for tracking smaller CAO points changes. However, limited data meant results were not conclusive. Chapter two investigated information criteria’s performance in selecting within-subject covariance matrices for linear mixed models comparing two treatments for restoring elderly dentition. Two patient well-being measures were assessed one, six and twelve months after treatment. Selection used pre-analysis consideration of suitable covariance structures, assessment of the unstructured matrix for the maximal model and five information criteria: AIC, BIC, AICc, CAIC and HQIC. Information criteria and other assessments all agreed on choice of covariance matrix for each well-being measure. Results indicated information criteria are suitable for covariance structure selection. Chapter three assessed a 12 variable binary logistic model in predicting subscriber non-renewal for a monthly magazine. Parsimonious models were selected by a novel individual candidate variable and best subsets procedures. Predictive models built using 70% of subscribers then predicted churn for the remaining subscribers. The full model had a 88% correct classification rate, a sensitivity of 95%, a PPV of 91% and an AUC of 93%. Best subsets selected a 7 variable model with a 89% correct classification rate, a sensitivity of 95%, a PPV of 91% and an AUC of 93%. Shorter and less valuable subscriptions indicated a higher likelihood of non-renewal. Results indicated logistic models are suitable for predicting non-renewal. |
| Subject |
Data mining -- Statistical methods.
|
|
Knowledge acquisition (Expert systems).
|
| Collection |
Theses Masters
|
|
Theses Statistics Department
|
| Description |
191 pages ; 30 cm. |
|
