The research portrayed in this dissertation aims to derive empirical means of predicting CBR values from index testing parameters and parameters calculated from them (e.g. shrinkage product). The project involved compiling a database of test results for a range of rock material types across moist and dry regions in southern Africa. The database was compiled in such a way that it represents natural gravels sampled (mostly) for construction or rehabilitation of road layer works. The database included a location description, material description, Weinert N-Value, Atterberg Limits, grading analysis and CBR values. In addition to this, the linear shrinkage product, shrinkage product, grading coefficient, grading modulus and dust ratio were calculated and also used in the analyses. All the samples were divided into two groups based on climate, as described by Weinert (1980). The data was then further sub-divided into compaction classes (95%, 98% and 100% Mod AASHTO) and within these compaction classes, each sample was assigned to a rock material group based on the classification proposed by Weinert (1980), but with minor alterations (e.g. further subdivision of pedogenic deposits). A total of 60 groups were created. Data processing was done using grading normalised to 100% passing the 37,5mm screen. In order to limit interdependency resulting from the cumulative grading, the sieve analysis results were converted to percentages retained on each sieve. This was necessary as statistical regressions often rejected datasets due to interdependency among input parameters (such as Atterberg Limits and cumulative grading). Based on the nature of the data, both stepwise linear regressions and Weibull regressions were performed. Though the Weibull regression is more suitable to the data (in theory) the linear regression could not be excluded, due to variable data. In addition, the existing model proposed by Kleyn (1955) which was derived empirically by Stephens (1988) was also retained for the analysis. In an attempt to refine Kleyn s model, the two parameters used by the method (i.e. grading modulus and plasticity index) were used in normal linear regressions in an attempt to adapt the model to specific material (and compaction) groups in the two climatic regions. More than 130 regressions were done for the final analysis (excluding experimental regressions, etc.), after restricting the predicted CBR ranges in an attempt to eliminate the prevailing data trend. The attempt proved futile, though, placing severe restrictions on the derived models. For each of the 60 groups all four methods were tested (i.e. stepwise linear regression, Weibull regression, Kleyn s model and a linear model adapted for each group based on Kleyn s model) and the most suitable model selected. A number of regressions were incomplete due to insufficient data, particularly in the groups associated with dry regions. Results proved poor and are ascribed to data variability rather than test methods. The data variability, in turn, is the result of test methods with poor reproducibility and repeatability. In short, the test methods (the CBR in particular) resulted in inconsistent data and subsequently poor results, making accurate predictions nearly impossible. / Dissertation (MSc)--University of Pretoria, 2010. / Geology / unrestricted