After the variables were selected, data cleaning methods were applied were the missing values deleted. Then, the data splitted to training set and test set. The training set are 80% of the total data and the rest are for test. Fig.5 present side byside histograms for the total tension loss of the strings in the original scale and in the logarithmic scale. Orlglnal Scale Iogarithmlc Scale 1000 800 Frequency 100 200 300 400 500 800 Frequency 400 600 200 0 l_'Tr"1'TTl rfT'_T'I 5 1 O 15 20 25 30 35 1 .5 2.0 2.5 3.0 3.5 The Total Tension Loss In {ha The logarithmic Total Tension Luca in lba Fig. 5. histograms for the total tension loss of the strings in the original scale and in the logarithmic scale Original scale would be used as the response. This is because the logarithmic scale shows that the data is skewed. However, when same data used in the original scale, the data becomes close to normal distribution which helps because the entire inferential statistics is favourable for the parametric tests only if the data is normally distributed. Moreover, the original scale represents larger number of total tension loss. The data range from about 5 to about 35 is larger in original scale in terms of orders of magnitude which makes it easier to deal with than the logarithmic scale. TABLE 2. GENERALIZED VARIANCE INFLATION FACTOR variables GVIF DF GVIF 2xDF OT 3.671 2 1.384 SS 14.181 2 1.941 M 6.894 8 1.128 PT 16.183 4.023 S 6.838 2.615 SSF 6.738 2.596 SSF 4.063 2.016 SP 10.623 1 3.259 To make GVIFs comparable across dimensions, Fox and Monette suggested using GVIF^(1/(2*DF)), where DF (degrees of freedom) is the number of coefficients in the subset to reduces the GVIF to a linear measure [9]. As a rule of thumb, if the value of GVIF^(1/(2*DF) ) exceed 10 that means there is high multicollinearity between the variables. As shown in TABLE. 2 we did check the multicollinearity of that generalized variance inflation factor measured by GVIF (1/(2*DF)) that's indicate none of them exhibiting high multicollinearity