PREVIOUS   NEXT   CONTENTS   HOME 

2.2. Aims and methods

The aims of intra-site analysis in this project were two-fold. The first was to provide a basic account of changes in pottery assemblage composition at Elms Farm over time, which could be compared with changing patterns of pottery consumption at similar sites in the region and period. Second, and more critical in understanding the articulation of social practice, was to investigate changing patterns of ceramic consumption and deposition, with particular emphasis on patterning within and between deposits, in terms of the contextual linkages both between different pot types and the deposit 'types' themselves. Would it be possible to identify recurring contextual associations and combinations of pottery types, indicative of culturally specific forms of consumption and depositional practice? Theoretically this would be extended to other artefact classes as well, although this possibility was beyond the scope of the present study. Furthermore, it was also desirable to investigate the functional zoning ascribed by the excavators at Elms Farm, and whether or not these spatial entities had any distinct characteristics in terms of ceramic consumption.

For the first aim, to document basic changes in ceramic assemblage composition, simple tables were produced showing the changing proportions of different pottery forms over time. In placing primary emphasis on pottery form over fabric, this approach has obvious value in providing clues about consumption and identity, with vessel form (rather than fabric) having a direct bearing on the social practices involving pottery. A two-tier form classification was adopted (Table 2), taking into account general perceived function (e.g. dining vessels, drinking vessels etc.), and individual form categories (e.g. beakers, platters, cups etc.) of the pottery. Imported fine ware products and their local imitations are indicated by the prefixes GB (Gallo-Belgic wares, usually terra rubra, terra nigra and imported white wares), I (general imported), TSG (samian ware, mostly Gaulish) and PR (Pompeian red wares). Tables were produced to account for general patterns by phase, area and individual feature.

In order to examine the second aim, the investigation of multiple contextual associations of broken pottery, a more complex approach was taken. The data were interrogated using the multivariate technique of correspondence analysis (hereafter referred to as CA), which presents a means of displaying trends in complex datasets in two dimensions for visual display (using Minitab). Whereas other forms of multivariate analysis such as multi-dimensional scaling and cluster analysis were applied to Romano-British pottery assemblages as early as the 1970s (e.g. Millett 1979), the application of correspondence analysis has been a more recent phenomenon, largely owing to the proliferation of faster home computers and relevant software since the early 1990s. CA has already had a number of useful applications to finds data in archaeology (e.g. Allison et al. 2004; Barclay et al. 1990; Biddulph 2005; Cool et al 1995; Cool and Baxter 1999; 2002; Lockyear 2000; Orton et al. 1993; and Pitts 2004; 2005a; 2005b; 2005c), not to mention its popularity in other academic disciplines where large cross-tabulations of data are unwieldy for more basic statistical analysis, such as sociology (e.g. Bourdieu 1984, 262), and market research (Hoffman and Franke 1986). Correspondence analysis is related to the more widely used multivariate method of principal components analysis (PCA), with the main difference being that CA is more suited to the analysis of categorical variables (i.e. counts of pots) rather than numeric measurements (i.e. the dimensions of pots) (Greenacre and Hastie 1987; Shennan 1997). CA produces a pair of plots with patterns in the first set of categoric variables or rows (e.g. different assemblages) directly corresponding to the respective patterns in the second set of categoric variables or columns (e.g. different types of pot). This can be done for specific phases and assemblages, within and between sites, so the potential for revealing detailed consumption patterns is huge.

The axes of the CA plots essentially measure the amount of variation from the average, with the most typical assemblages and the most commonly found pots occurring closest to the point where the graph axes cross, and the most unusual occurring at the plot extremes (see Greenacre 1993 for the mathematical underpinning of this technique). By default the components or axes selected by the computer software in CA are usually the first and second, which together account for the most inertia (a measure of the variability described by each axis). However, it is sometimes necessary to look at other components (e.g. the first and third) if over-clustering of the first and second components renders visual interpretation problematic (e.g. Pitts et al. forthcoming). Alternatively, it is also possible to remove dominant values or outliers in successive stages of analysis and examine less dominant trends in the data, in a process described as 'peeling the onion' (Cool and Baxter 1999; Pitts 2005c), which is the method employed in the present study. The most obvious example of this practice in the present study was the removal of generic jar forms prior to CA, in all plots. As jar forms comprised a dominant component (usually in excess of 50% by EVE) of most ceramic assemblages throughout the period of interest, it was necessary to remove them in order to identify other less striking patterns that would otherwise be obscured by the 'gravitational pull' of the jars towards the CA plot centres. The removal of further outliers from CA plots is indicated in Table 24.

The problem of amphorae not being quantified by EVE was addressed by plotting the occurrence of amphorae fabrics in CA by weight, as supplementary points. As the direct comparison of pottery quantified by two different methods (i.e. weight and EVEs) would introduce too much bias to be of any statistical value and skew the whole process of correspondence analysis, it was not within the interests of good practice to compare the amphorae weights directly with the rest of the pottery quantified by EVE in CA. However, if the EVE values are plotted first and the principal axes are established, it is possible to plot the additional values as supplementary points (Greenacre and Hastie 1987, 441), 'scoring' the extra data according to the results of the primary dataset, without causing any distortion to original values. In this case, additional columns of data containing the amphorae weight values were included in analysis as supplementary points, which would not influence the distribution of the principal assemblage data, yet could be plotted to give an indication of amphorae deposition in terms of the rest of the pottery forms and assemblages. Although not perfect, this method seemed reliable enough to indicate the assemblages containing high proportions of amphorae, with the added benefit of not having to consider the amphorae as a separate entity. Unfortunately only the pottery fabrics and not vessel forms had been quantified by weight, so it was not possible to conduct CA by both weight and EVE to note any biases. Basic quantification of amphorae by weight is given elsewhere according to period (Table 5) and assemblage (Table 7).

Interpretation of the CA plots shown here is as follows. Each run of CA produces two plots (relating to the rows and columns of the original tables of data), which can either be presented separately or superimposed. In this study, one plot displays the different assemblages or areas according to their similarities and differences in pottery composition (e.g. Figure 2a), whereas the other presents the individual pottery forms according to their deposition in different assemblages or areas (e.g. Figure 2b). Assemblages with similar compositions will cluster in the first plot, while pottery forms deposited in similar assemblages will cluster in the second. The area of the first plot directly corresponds to the same area on the second, hence the term 'correspondence analysis'. Ideally, the plots should be presented superimposed on one another to aid interpretation (i.e. to spot which pots characterise particular assemblages and vice versa). However, in this study, where there are large numbers of categories or excessive clustering, it is preferable to display the row and column plots separately (e.g. Figures 2a and 2b). All the CA plots here are symmetrically scaled. This means that the relationships between directly corresponding pottery forms and assemblages can only be assessed relatively, not absolutely. It is important to stress that CA is not intended as a form of absolute statistical testing. To account for this, any patterning of note was verified by reference to the original datasets.

 PREVIOUS   NEXT   CONTENTS   HOME 

© Internet Archaeology URL: http://intarch.ac.uk/journal/issue21/2/2_2.html
Last updated: Tue May 08 2007