6 RESULTS 82
6.3 Statistical Results: pXRF 98
A number of elements had to be removed from the data prior to using GAUSS because certain elements had readings that were so low they are below the limits of detection. Therefore, the data were limited to only 24 elements.
6.3.1 pXRF Bi-Plots
Before bi-plots of the samples were created, principal components were generated and plotted to determine the significant elements for the pXRF data. Although sulfur (S) and manganese (Mn) are displayed on the bi-plots, other combinations of prominent elements were incorporated during the analysis to compare and determine any differences between the bi-plots. The principal components that were used were determined based on the cumulative percent of variance after Principal Component Analysis (PCA). Similar to the LA-ICP-MS data, PC01 through PC08 were compared. PC01 was positively loaded on S and P. PC02 was positively loaded on Pb. PC03 was positively loaded on Ca. PC04 was positively loaded on Cl. PC05 was positively loaded on Sr, Mn, and Nb. PC06 was positively loaded on Ba and Mn. PC07 was positively loaded on Ti, P, and Ca. Lastly, PC08 was positively loaded on Cu.
Figure 19 Bi-Plot of logged Principal Component 1 (PC01) and Principal Component 6 (PC06) that shows the prominent elements
Once the prominent elements were determined, a bi-plot of the pXRF measurements for all the samples was created to get a general picture of how the samples’ chemical compositions related to each other. At a first glance, the pXRF bi-plots appeared to be somewhat similar to the LA-ICP-MS data, but do not exactly replicate those plots. Figure 20 displays a large cluster of samples rather than three distinct groups, however, there is one similar outlier (T034). Unlike in Figure 9, Figure 20 shows the Tarija and La Paz samples in the pXRF data appear to correlate with the heartland sites.
Figure 20 Bi-Plot of logged S and Mn concentrations consisting samples divided into Group 1 and Group 2
When examined more closely, a similar pattern can be established in Figure 20 that was present in the LA-ICP-MS bi-plots. Two different groups were visible, but still unclear. This was because there were various groups represented by different colors so it was difficult to ascertain exactly what samples were where. Therefore, Groups 1 and 2 were created. Group 1 included samples from Tiwanaku, Chiripa, Lukurmata, Tarija, and La Paz. Group 2 consisted of the Chen Chen and Cerro Baúl samples. When Group 1 and Group 2 were plotted the pattern became much more evident. The results were slightly similar to the LA-ICP-MS, however, only two groups were able to be established through pXRF data rather than three. Further, pXRF data
x Group 2
showed that the Tarija and La Paz samples correlated with heartland samples. This was not the case in the LA-ICP-MS results. Another major difference involved the Cerro Baúl samples. Unlike in the LA-ICP-MS results, the Cerro Baúl ceramics were more compositionally similar to the Chen Chen samples compared to the heartland samples.
These inconsistencies are some of the issues that are encountered when using multiple techniques. LA-ICP-MS has been noted to be more accurate and precise in its measurements and it can access the sample deeper than the surface eliminating chances of surface contamination, as discussed in Chapter 3. When using pXRF, not only are there increased chances of surface contamination, but its levels of accuracy and precision are much lower when compared to LA- ICP-MS (Pollard et al. 2007; Rice 1987). However, the use of multiple techniques is beneficial because certain samples may not be able to be measured using certain techniques. For example, some samples were unable to be measured using LA-ICP-MS, but were measured using pXRF. Without the use of pXRF, certain samples may not have been incorporated in this research.
6.3.2 pXRF Hierarchical Cluster Analysis
HCA for the pXRF data derived from all the samples indicated two major compositional groups. From the bi-plots, two major groups were able to be established (Group 1 and Group 2). Very similar results are indicated by the HCA of the pXRF samples (Figure 21).
Figure 21 Hierarchical Cluster Analysis of samples measured by pXRF
According the HCA, Group 1 consists of mostly Tiwanaku, Lukurmata, and Chiripa, and La Paz samples, however a few outlying samples from Chen Chen and Tarija appear as well. Group 2 contains mostly samples from Chen Chen and Cerro Baúl, as well as some from Tiwanaku, Chiripa, Tarija, and Lukurmata. With the exception of some samples, Group 1 and 2 shown in the HCA are similar to Group 1 and 2 exemplified in the bi-plots.
6.3.3 pXRF Mahalanobis Distance Measurement
When Mahalanobis distance measurements were run on the pXRF data using the prominent elements that were listed above. Groups 1 and 2 from the pXRF bi-plots were
measured. Only a few samples were out of place in the Mahalanobis distance measurement. For
Outlier
Group 2
instance, sample T002 from Tiwanaku was originally in Group 1, the measurement established that it closely correlated with Group 2 instead. Samples CB003 from Cerro Baúl and M009 and M013 from Chen Chen were originally in Group 2 but were similar in composition with Group 1.