Molalakgotla, Mamoloko Portia2021-04-242021-04-242020https://hdl.handle.net/10539/30988A research report submitted to the Faculty of Science, University of the Witwatersrand, in partial fulfillment of the requirements for the degree of Masters of Science (by Coursework and Research) School of Statistics and Actuarial Science, 2020Reducing poverty remains a central priority in South Africa. Exploring poverty multidimensionally has been a recent topic in literature. There has been a growing demand for measuring poverty multidimensionally. The four dimensions agreed on by Statistics South Africa and treated as predetermined dimensions for measuring poverty in South Africa are: Education, Health, Economic Activity and Standard of living. This study has explored the use of Nonlinear Principal Component Analysis on a sample of the 2011 population census data (focusing on Limpopo Province) to see if it generates the same grouping of indicators as the pre-determined dimensions. The same dataset has also been subjected to the K-modes clustering analysis and Latent Class Analysis (LCA). Results from the Nonlinear PCA have shown that some dimensions contain different indicators as compared to the pre-determined dimensions. Clustering of households in Limpopo was done based on the Kmodes and Latent Class Analysis Of Polytomous Outcome Variables (poLCA) algorithms. Findings reveal that the K-modes and LCA methods generated the same number of groups (3 groups). The results obtained from poLCA algorithm put the households in 3 clusters with the dominating cluster/group containing households that are multidimensionally poor. The second dominating cluster contains households that are not mired in poverty and the third cluster has households which are deprived of only 1 dimension. The advantage of LCA over K-modes is that it makes use of objective statistical measures such as BIC and AIC to determine the ideal number of groups. Its downturn is that it has problems when it comes to handling huge datasetsenThesis