Browsing by Author "Banjoko A. W."

Now showing 1 - 3 of 3
  • Thumbnail Image
    Item
    A Test Procedure for Ordered Hypothesis of Population Proportions Against a Control
    (Turkiye Klinikleri Journal of Biostatistics, 2016) Yahya W. B.; Olaniran O. R.; Garba M. K.; Oloyede I.; Banjoko A. W.; Dauda K. A.; Olorede K. O.
    Objective: This paper aims to present a novel procedure for testing a set of population proportions against an ordered alternative with a control. Material and Methods: The distribution of the test statistic for the proposed test was determined theoretically and through Monte-Carlo experiments. The efficiency of the proposed test method was compared with the classical Chi-square test of homogeneity of population proportions using their empirical Type I error rates and powers at various sample sizes. Results: The new test statistic that was developed for testing a set of population proportions against an ordered alternative with a control was found to have a Chi-square distribution with non-integer values degrees of freedom v that depend on the number of population groups k being compared. Table of values of v for comparing up to 26 population groups was constructed while an expression was developed to determine v for cases where k > 26. Further results showed that the new test method is capable of detecting the superiority of a treatment, for instance a new drug type, over some of the existing ones in situations where only the qualitative data on users' preferences of all the available treatments (drug types) are available. The new test method was found to be relatively more powerful and consistent at estimating the nominal Type I error rates (α), especially at smaller sample sizes than the classical Chi-square test of homogeneity of population proportions. Conclusion: Conclusion: The new test method proposed here could find applications in pharmacology where a newly developed drug might be expected to be more preferred by users than some of the existing ones. This kind of test problem can equally exist in medicine, engineering and humanities in situations where only the qualitative data on users' preferences of a set of treatments or systems are available.
  • Thumbnail Image
    Item
    Multiclass Feature Selection and Classification with Support Vector Machine in Genomic Study
    (Professional Statisticians Society of Nigeria (PSSN), 2017) Banjoko A. W.; Yahya W. B.; Garba M. K.; Olaniran O. R.; Amusa L. B.; Gatta N. F.; Dauda K. A.; Olorede K. O.
    This study proposes an efficient Support Vector Machine (SVM) algorithm for feature selection and classification of multiclass response group in high dimensional (microarray) data. The Feature selection stage of the algorithm employed the F-statistic of the ANOVA–like testing scheme at some chosen family-wise-error-rate (FWER) to control for the detection of some false positive features. In a 10-fold cross validation, the hyper-parameters of the SVM were tuned to determine the appropriate kernel using one-versus-all approach. The entire simulated dataset was randomly partitioned into 95% training and 5% test sets with the SVM classifier built on the training sets while its prediction accuracy on the response class was assessed on the test sets over 1000 Monte-Carlo cross-validation (MCCV) runs. The classification results of the proposed classifier were assessed using the Misclassification Error Rates (MERs) and other performance indices. Results from the Monte-Carlo study showed that the proposed SVM classifier was quite efficient by yielding high prediction accuracy of the response groups with fewer differentially expressed features than when all the features were employed for classification. The performance of this new method on some published cancer data sets shall be examined vis-à-vis other state-of-the-earth machine learning methods in future works.
  • Thumbnail Image
    Item
    Partial Least Squares-Based Classification and Selection of Predictive Variables of Crimes against Properties in Nigeria
    (Professional Statisticians Society of Nigeria (PSSN), 2017) Olorede K. O.; Yahya W. B.; Garuba A. O.; Banjoko A. W.; Dauda K. A.
    In this study, the state-of-the-art Partial Least Squares (PLS) based models (PLS-Discriminant analysis (PLS-DA), Sparse PLS-DA (SPLS-DA) and Sparse Generalized PLS (SGPLS)) were employed to model and classify the rate of crimes (low or high) committed against properties across the 36 states in Nigeria and the Federal Capital Territory (FCT). The core variables that are predictive of this crime type in Nigeria were identified using the LASSO penalty method via the PLS. Data on occurrences of cases of offences against property obtained from the data base of Nigerian Police Force were utilized in this study. The missing values due to non-occurrence or non-reportage of crime cases were imputed, using the techniques of multivariate imputation by chained equation. The complete data set were partitioned into training and test sets using 80:20 holdout scheme. The 80% training set was used to build the PLS-based models that were in turn used to predict the overall crime rates of Nigerian cities in the 20% held out test data over 200 Monte-Carlo cross-validation runs. All the PLS-based models yielded good classification of unseen test samples into either of two qualitative classes of high and low crime rates with average Correct Classification Rate (CCR) of 94%. Other performance metrics including sensitivity, specificity, positive and negative predictive values, balance accuracy and diagnostic odds ratio were estimated to further examine their classification efficiencies. The SGPLS identified fewer (just 3 out of 12) core relevant crime variables that are predictive of the overall crime rates in Nigerian states with highest CCR than the SPLS which selected 9 such variables to achieved about the same feat.