RESEARCH RESULTS
6.2 DIMENSIONALITY ANALYSIS
6.2.1 Integrated discussion of the dimensionality analysis results over the three ethnic group samples three ethnic group samples
Tables 6.2 to 6.4 provide an overview of the principal axis factor analyses for the three ethnic groups. The Kaiser-Meyer-Olkin (KMO) and Bartlett’s test were used to examine the factor analyzability of the observed inter-item correlation matrices. The KMO measures sampling adequacy as an index expressing the ratio of the sum of the squared inter-item correlations and the squared inter-item correlation plus the sum of the squared partial inter-item correlation coefficients (Sricharoena &
Buchenrieder, 2005). The KMO measure varies from unity to zero; values closer to unity are regarded as better values. If items reflect a common underlying factor the value will approach unity. Where KMO approaches at least .60 the correlation matrix is considered to be factor analyzable (Moyo, 2009). With regards to the results in Table 6.2 to Table 6.4 the values of the KMO range between .65 and .89. This indicates that that all the correlation matrices were factor analyzable.
137 The null hypothesis that the inter-item correlation matrix is an identity matrix in the parameter was tested by the Bartlett test of sphericity. An identity matrix is one in which all items only correlate with themselves and not with each other (Moyo, 2009).
This can be seen when all the diagonal elements are 1’s and all off diagonals are 0’s. The results for all 16 subscales across the three ethnic groups revealed that the null hypothesis could be rejected. This further indicated the factor analyzability of the correlation matrices.
The results of the KMO and Bartlett tests suggested that it would be meaningful to conduct factor analysis on the 16 inter-item correlation matrices across the three ethnic groups.
Table 6.2
SUMMARY OF THE RESULTS OF THE PRINCIPAL AXIS FACTOR ANALYSES FOR THE WHITE SAMPLE GROUP
No. of
% Variance Factors Subscale Determinant KMO Bartlett x² Explained Extracted
FA .17 .86 7923.83 22.14 2
FB .17 .81 7937.66 20.17 3
FC .14 .87 9074.72 24.16 3
FE .23 .84 6723.11 19.80 3
FF .11 .85 10124.14 24.07 2
FG .13 .89 9387.06 25.01 2
FH .06 .89 13098.72 30.03 2
FI .18 .82 7814.01 20.39 3
FL .17 .82 8027.95 20.35 3
FM .29 .75 5687.26 15.11 4
FN .13 .83 9329.87 22.71 3
FO .18 .88 7833.31 22.53 2
FQ1 .17 .79 8035.18 18.82 3
FQ2 .16 .86 8422.73 22.43 2
FQ3 .29 .80 5546.64 16.75 3
FQ4 .10 .89 10344.69 26.00 2
FA - Factor A; FB - Factor B; FC - Factor C; FE - Factor E; FF - Factor – F; FG - Factor G; FH - Factor H; FI - Factor I; FL - Factor L; FM - Factor M; FN - Factor N; FO - Factor O; FQ1 - Factor Q1;
FQ2 - Factor Q2; FQ3 - Factor Q3; FQ4 - Factor Q4
138
Table 6.3
SUMMARY OF THE RESULTS OF THE PRINCIPAL AXIS FACTOR ANALYSIS FOR THE BLACK SAMPLE GROUP
No. of
% Variance Factors Subscale Determinants KMO Bartlett x² Explained Extracted
FA .53 .76 2806.39 11.82 3
FB .33 .77 4883.4 14.96 3
FC .25 .81 6127.204 18.10 3
FE .57 .73 2534.7 10.50 3
FF .20 .81 7118.1 18.814 4
FG .30 .85 5317.32 18.06 2
FH .20 .85 7354.54 21.236 3
FI .36 .69 4496.18 12.22 4
FL .33 .74 4883.29 14.19 3
FM .57 .65 2506.77 8.26 4
FN .46 .75 3469.48 12.57 3
FO .37 .79 4387.34 15.32 4
FQ1 .43 .69 3755.09 11.48 4
FQ2 .40 .77 4084.65 13.99 3
FQ3 .62 .72 2144.13 9.75 4
FQ4 .50 .74 3128.74 11.38 3
FA - Factor A; FB - Factor B; FC - Factor C; FE - Factor E; FF - Factor – F; FG - Factor G; FH - Factor H; FI - Factor I; FL - Factor L; FM - Factor M; FN - Factor N; FO - Factor O; FQ1 - Factor Q1;
FQ2 - Factor Q2; FQ3 - Factor Q3; FQ4 - Factor Q4
139 Subscale Determinants KMO Bartlett x² Explained Extracted
FA .34 .80 1125.42 16.05 4
The results for the Aloof – Empathic subscale for the White sample revealed that two clear factors emerged. Two factors obtained eigenvalues greater than unity. The rotated factor matrix (pattern matrix11; see Appendix 4) revealed that factor 1 had three items (Q52, Q76 and Q101) with loadings greater than .50 and four items (Q51, Q77, Q151 and Q176) with loadings greater than .30. Factor 2 indicated three items with substantial negative loadings. One item (Q2) obtained a loading of less than -.50 and two items (Q27 and Q151) obtained loadings of less than -.30. The negative loading reveals a negative correlation between the factor and the item.
Three items (Q2, Q26 and Q126) did not load on any of the two factors. As indicated in the results one item showed itself as a complex item (Q151) because it simultaneously loaded on both factors. No meaningful identity could be determined
140 for the two extracted factors based on common themes shared by the items that loaded on them.
Due to the confirmatory nature of this study a single factor was forced on the scale as per the a priori model. It is evident from Table 6.5 that the loadings for the single extracted factor were reasonable. Four items (Q1, Q52, Q77 and Q151) obtained loadings greater than .50 and seven items (Q26, Q27, Q51, Q76, Q101, Q126 and Q176) obtained loadings greater than .30. Only one item (Q2) did not load on the single extracted factor.
The residual correlations were calculated for both the two-factor and one-factor solutions. The two-factor solution showed a small percentage (9%) of non-redundant residuals with absolute values greater than .05. The one-factor solution’s percentage (15%) of large non-redundant residuals was larger than for the two-factor solution, signifying that the one-factor solution provided a less credible, but still plausible explanation, for the observed correlation matrix.
The dimensionality analysis results for the Black sample revealed a three-factor structure based on the eigen-value-greater-than-unity rule. The pattern matrix (Appendix 4) revealed that factor 1 had one item (Q151) with a loading greater than .50 and three items (Q1, Q52 and Q77) with loadings greater than .30. There was only one item (Q26) with a loading greater than .30 on factor 2. Factor 3 indicated one item (Q101) with a loading greater than .50 and two items (Q76 and Q176) with loadings greater than .30. Four items (Q2, Q27, Q51 and Q126) did not load on any of the three factors. No meaningful identity could be determined for the three extracted factors based on common themes shared by the items that load on them.
Due to the confirmatory nature of this study a single factor was forced on the scale as per the a priori model. Table 6.5 revealed that two items (Q52 and Q151) obtained loadings greater than .50 and five items (Q27, Q76, Q77, Q101 and Q176) loadings greater than .30. Five items (Q1, Q2, Q26, Q51 and Q126) did not load on the single extracted factor.
The residual correlations were calculated for both the factor solutions. The three-factor solution indicated a zero percentage of non-redundant residuals with absolute values greater than .05. The one-factor solution’s percentage (12%) of large
non-141 redundant residuals was larger than the three-factor solution, signifying that the one-factor solution provided a less credible, but still acceptable explanation for the observed correlation matrix.
The results of the analysis for the Coloured sample once again revealed factor fission in that a four-factor structure underlied the subscale (based on the eigen-value-greater-than-unity rule). The rotated factor structure revealed that factor 1 had one item (Q151) with a loading greater than .50 and four items (Q1, Q27, Q52 and Q77) with loadings greater than .30. Two items (Q2 and Q51) with loadings greater than .30 loaded on factor 2. Factor 3 indicated two items (Q26 and Q176) with loadings greater than .30 and factor 4 also indicated two items (Q76 and Q101) with loadings greater than .30. One item (Q126) did not load on any of the four factors.
Again no meaningful identity could be determined for the four extracted factors based on common themes shared by the items that loaded on them.
Fairly low item loadings were obtained when a single factor was forced. Table 6.5 revealed that three items (Q52, Q77 and Q151) obtained loadings greater than.50 and five items (Q1, Q76, Q51, Q27 and Q101) had loadings greater than.30. Four items (Q2, Q26, Q126 and Q176) did not load significantly on the single extracted factor.
Further to this the residual correlations were calculated for both the four-factor and one-factor solutions. The four-factor solution showed a zero percentage of non-redundant residuals with absolute values greater than .05. The one-factor solution’s percentage (13%) of large non-redundant residuals was larger than the four-factor solution, signifying that the one-factor solution provided a less credible but still plausible explanation for the observed correlation matrix.
The dimensionality analyses results for this subscale revealed two factors for the White group, three factors for the Black group and four factors for the Coloured group when the eigen-values-greater-than-unity rule was applied. The overall results, therefore, revealed more than one factor underlying the structure of this subscale in every one of the three groups. This signified the need for more than one factor to satisfactorily explain the observed correlations between the items in the subscale.
Strictly speaking the unidimensionality assumption was therefore not corroborated.
Item Q2 did not load effectively on the White and Black groups. Item Q126 also
142 revealed an insignificant loading on the factors of the Coloured and Black groups.
The item analysis results also indicated item Q2 as a problematic item. When the extraction of a single factor was forced the majority of items in the three groups obtained relatively good loadings. Therefore it could be deduced that the majority of the items represent the underlying latent variable well, with the exception of items Q2 and Q126. The percentage of large residual correlations obtained for the single-factor solution was still sufficiently small to regard the single single-factor solution as a credible explanation for the observed correlation matrix. Interpreted somewhat more leniently the assumption of essential unidimensionality can therefore be regarded as not altogether without merit.
Table 6.5
FACTOR MATRIX WHEN FORCING THE EXTRACTION OF A SINGLE FACTOR (FACTOR A) OVER THE THREE ETHNIC GROUP SAMPLES
White Sample Black Sample Coloured Sample
15FQ+_FA_Q1 .50 15FQ+_FA_Q1 .30 15FQ+_FA_Q1 .40
15FQ+_FA_Q2 .10 15FQ+_FA_Q2 -.00 15FQ+_FA_Q2 .00
15FQ+_FA_Q26 .30 15FQ+_FA_Q26 .16 15FQ+_FA_Q26 .20
15FQ+_FA_Q27 .30 15FQ+_FA_Q27 .30 15FQ+_FA_Q27 .30
15FQ+_FA_Q51 .50 15FQ+_FA_Q51 .28 15FQ+_FA_Q51 .40
15FQ+_FA_Q52 .60 15FQ+_FA_Q52 .52 15FQ+_FA_Q52 .60
15FQ+_FA_Q76 .50 15FQ+_FA_Q76 .33 15FQ+_FA_Q76 .50
15FQ+_FA_Q77 .70 15FQ+_FA_Q77 .47 15FQ+_FA_Q77 .60
15FQ+_FA_Q101 .40 15FQ+_FA_Q101 .34 15FQ+_FA_Q101 .40
15FQ+_FA_Q126 .30 15FQ+_FA_Q126 .17 15FQ+_FA_Q126 .10
15FQ+_FA_Q151 .70 15FQ+_FA_Q151 .52 15FQ+_FA_Q151 .60
15FQ+_FA_Q176 .40 15FQ+_FA_Q176 .35 15FQ+_FA_Q176 .30
1 factor extracted. 5 iterations required.
The items that have been highlighted can be considered satisfactory in terms of the proportion of item variance that can be explained by the single extracted factor.
6.2.1.2 Factor B
The results for the Intellectance subscale for the White group returned a three-factor structure. Examination of the pattern matrix (see Appendix 4) revealed two items (Q102 and Q152) with loadings greater than .50 and three items (Q127, Q153 and Q178) with loadings greater than .30 (on Factor 1). Substantial negative loadings of less than -.50 for two items (Q53 and Q177) were evident on Factor 2. Factor 3 indicated one item (Q78) with a loading greater than .50 and two items (Q3 and Q28) with loadings greater than .30. Two items (Q103 and Q128) did not load on any of
143 the three extracted factors. The identity of the three extracted factors could not be inferred from the items loading on them.
It was evident from Table 6.6 that upon forcing a single factor, reasonable item loadings emerged. Two items (Q102 and Q153) obtained loadings greater than .50 and ten items (Q3, Q28, Q53, Q78, Q103, Q128, Q127, Q152, Q177, and Q178) obtained loadings greater than .30. All items loaded greater than .30 on the forced single extracted factor.
The residual correlations were calculated for both the three-factor and one-factor solutions. The three-factor solution showed a small percentage (4%) of non-redundant residuals with absolute values greater than .05. The one-factor solution’s percentage (45%) of large non-redundant residuals was large, signifying that the one-factor solution was a less credible explanation for the observed correlation matrix.
The dimensionality analysis results for the Black sample also revealed three factors.
Three factors had eigen values greater than unity. The rotated pattern matrix (see Appendix 4) indicated that factor 1 had one item (Q153) with a loading greater than .50 and five items (Q3, Q28, Q78, Q128 and Q178) with loadings greater than .30.
Factor 2 had two items (Q53 and Q177) with loadings greater than .50 and factor 3 had three items (Q102, Q127 and Q152) with loadings greater than .30. Only one item (Q103) did not load on any of the three extracted factors. No meaningful identity could be determined for the three extracted factors based on common themes shared by the items that loaded on them.
Next, a single factor was extracted. It was evident from Table 6.6 that the loadings for the single extracted factor were fairly low. Only one item (Q153) had a loading greater than .50 and eight items (Q3, Q28, Q78, Q102, Q178, Q127, Q128, and Q177) obtained loadings greater than .30. Three items (Q53, Q103 and Q152) did not load on the single extracted factor.
The results of the calculated residual correlations for the three-factor solution showed a small percentage (3%) of non-redundant residuals with absolute values greater than .05. The one-factor solution’s percentage (27%) of large non-redundant residuals was larger than the three-factor solution signifying that the one-factor
144 solution provided a less credible, but still plausible, explanation for the observed correlation matrix.
Similar to the previous two analyses, the results for the Intellectance subscale for the Coloured sample also revealed that a three-factor structure best explained the observed correlation matrix. Three factors obtained eigenvalues greater than unity.
The rotated pattern matrix (see Appendix 4) revealed that factor 1 indicated three items (Q28, Q78 and Q127) with loadings greater than .50 and one item (Q3) with a loading greater than .30. Factor 2 indicated two items (Q53 and Q177) with negative loadings less than -.50 and one item (Q102) with a negative loading less than -.30.
Factor 3 indicated one item (Q178) with a loading greater than .50 and two items (Q128 and Q153) with loadings greater than .30. Two items (Q103 and Q152) did not load on any of the three extracted factors. Upon forcing a single factor, reasonable factor loadings emerged. Table 6.6 revealed one item (Q177) with a loading greater than .50 and ten items (Q3, Q28, Q53, Q78, Q102, Q128, Q127, Q152, Q153, and Q178) with loadings greater than .30. Only one item (Q103) did not load on the forced single extracted factor. Again no meaningful identity could be determined for the three extracted factors based on common themes shared by the items that load on them.
The three-factor solution showed a small percentage (1%) of non-redundant residuals with absolute values greater than .05. The one-factor solution’s percentage (28%) of large non-redundant residuals was substantially larger signifying that the one-factor solution was a less credible, but still plausible explanation for the observed correlation matrix.
Overall the dimensionality analyses results indicated three factors with eigenvalue-greater than unity for this subscale across the three samples. This signifies the need for three factors to satisfactorily explain the observed correlations between the items in the subscale. Strictly speaking the unidimensionality assumption was therefore not corroborated. Item Q103 was flagged as a problematic item as it did not load on any of the factors across the three groups. When the extraction of a single factor was forced the majority of items in the three groups obtained relatively good loadings.
This phenomenon indicated that the majority of the items represent the underlying latent variable well. Attention should be given to item Q103. The percentage of large
145 residual correlations obtained for the single-factor solution was still sufficiently small (especially for the Black and Coloured samples) to regard the single factor solution as a credible explanation for the observed correlation matrix. When the results are interpreted somewhat more leniently the position that a single common factor underlies the 12 items of the Intellectance subscale therefore is not altogether untenable.
Table 6.6
FACTOR MATRIX WHEN FORCING THE EXTRACTION OF A SINGLE FACTOR (FACTOR B) OVER THE THREE ETHNIC GROUP SAMPLES
White Sample Black Sample Coloured Sample
15FQ+_B_Q3 .40 15FQ+_B_Q3 .40 15FQ+_B_Q3 .40
15FQ+_B_Q28 .40 15FQ+_B_Q28 .40 15FQ+_B_Q28 .50
15FQ+_B_Q53 .40 15FQ+_B_Q53 .20 15FQ+_B_Q53 .40
15FQ+_B_Q78 .50 15FQ+_B_Q78 .50 15FQ+_B_Q78 .50
15FQ+_B_Q102 .60 15FQ+_B_Q102 .40 15FQ+_B_Q102 .50
15FQ+_B_Q103 .40 15FQ+_B_Q103 .30 15FQ+_B_Q103 .30
15FQ+_B_Q127 .50 15FQ+_B_Q127 .40 15FQ+_B_Q127 .40
15FQ+_B_Q128 .40 15FQ+_B_Q128 .40 15FQ+_B_Q128 .40
15FQ+_B_Q152 .50 15FQ+_B_Q152 .30 15FQ+_B_Q152 .40
15FQ+_B_Q153 .50 15FQ+_B_Q153 .50 15FQ+_B_Q153 .50
15FQ+_B_Q177 .50 15FQ+_B_Q177 .30 15FQ+_B_Q177 .50
15FQ+_B_Q178 .50 15FQ+_B_Q178 .40 15FQ+_B_Q178 .50
1 factor extracted. 5 iterations required
The items that have been highlighted can be considered satisfactory in terms of the proportion of item variance that can be explained by the single extracted factor.
6.2.1.3 Factor C
The results for the White sample revealed that the Affected by feelings – emotionally stable subscale split into three factors, based on the eigen-value-greater-than-unity rule. Examination of the rotated pattern matrix (see Appendix 4) revealed that two items (Q104 and Q129) with loadings greater than .50 and two items (Q29 and Q55) with loadings greater than .30 loaded on Factor 1. One item (Q5) with a loading greater than .50 and two items (Q30 and Q54) with loadings greater than .30 was evident for Factor 2. Factor 3 indicated three items (Q80, Q154 and Q179) with negative loadings more than -.50 and two items (Q4 and Q79) with negative loadings more than -.30. All items loaded at least on one of the extracted factors. However, no meaningful identity could be determined for the three extracted factors based on common themes shared by the items that loaded on them.
146 It was evident from Table 6.7 that when forcing a single factor, all items loaded reasonably on the single extracted factor. Six items (Q4, Q54, Q55, Q80, Q104 and Q179) obtained loadings greater than .50 and six items (Q5, Q29, Q30, Q79, Q129 and Q154) obtained loadings greater than .30. Hence, all items load greater than .30 on the forced single factor.
The results of the residual correlations calculations revealed that the three-factor solution obtained a small percentage (4%) of non-redundant residuals with absolute values greater than .05. For the one-factor solution a larger but still acceptably small percentage (18%) of large non-redundant residuals was evident. Therefore it was deduced that the one-factor solution provided a less credible albeit still acceptable explanation for the observed correlation matrix than the three-factor solution.
The results for the Black group also revealed that three factors should be extracted.
Examination of the pattern matrix (see Appendix 4) revealed that for Factor 1 five items (Q4, Q5, Q30, Q79 and Q179) obtained loadings greater than .30. Factor 2 indicated two items (Q104 and Q129) with negative loadings of more than -.50 and two items (Q29 and Q55) with negative loadings of more than -.30. Factor 3 indicated one item (Q154) with a negative loading of more than -.50 and one item (Q80) with a negative loading of more than -.30. Only one item (Q54) did not load on any of the three extracted factors. No meaningful identity could be determined for the three extracted factors based on common themes shared by the items that load on them.
Upon forcing a single factor reasonable factor loadings emerged. Table 6.7 revealed two items (Q104 and Q179) had loadings greater than .50 and eight items (Q4, Q29, Q54, Q55, Q79, Q80, Q129 and Q154) had loadings greater than .30. Two items (Q5 and Q30) did not load on the forced single extracted factor.
The residual correlations were calculated for both solutions. The three-factor solution obtained a small percentage (4%) of non-redundant residuals with absolute values greater than .05. The one-factor solution indicated a larger but still acceptably small percentage (31%) of large non-redundant residuals. Therefore the one-factor solution provided a less credible, but nonetheless still plausible, explanation for the observed correlation matrix.
147 The results for the Affected by feelings – emotionally stable subscale’s dimensionality analysis for the Coloured sample indicated three factors with eigenvalues greater than unity. The result suggested factor fission. Factor 1 contained one item (Q179) with a loading greater than .50 and four items (Q4, Q79, Q80 and Q154) with loadings greater than .30. Factor 2 had two items (Q129 and Q104) with negative loadings of more than -.50 and factor 3 indicated three items (Q5, Q30 and Q54) with loadings greater than .30. Two items (Q29 and Q55) did not load on the extracted factors. Again no meaningful identity could be determined for the three extracted factors based on common themes shared by the items that loaded on them.
Table 6.7 revealed that when forcing a single factor, all items loaded in a reasonable manner. One item (Q179) obtained a loading greater than .50 and ten items (Q4, Q5, Q29, Q54, Q55, Q79, Q80, Q104, Q129 and Q154) obtained loadings greater than.30. Only one item (Q30) did not load on the single extracted factor.
Results of the residual correlations for the three-factor solution showed a small percentage (6%) of non-redundant residuals with absolute values greater than .05.
The one-factor solution obtained a larger, but still acceptably small percentage (21%) of large non-redundant residuals. Therefore the one-factor solution provided a less credible but still permissible explanation for the observed correlation matrix.
Overall the dimensionality analyses results indicated three factors with eigenvalues greater than unity for this subscale across the three samples. This signified the need for three factors to satisfactorily explain the observed correlations between the items in the subscale. Item Q129 and Item Q104 both had significant negative loadings in the Coloured and Black group. Strictly speaking the unidimensionality assumption was therefore not corroborated.
When the extraction of a single factor was forced the majority of items in the three groups obtained relatively good loadings. This phenomenon indicated that the majority of the items represent the underlying latent variable well. The percentage of large residual correlations obtained for the single-factor solution was still sufficiently small for all three samples to regard the single factor solution as a permissible explanation for the observed correlation matrix. When the results were interpreted somewhat more leniently, the position that a single common factor underlies the 12