5. Methodology

The first task in creating the temporal predictive models is to create the time periods which will be an integral part of the modelling. However, once divided into time periods, the same methodology is employed as if it were a traditional predictive model. In a traditional approach to predictive modelling, sites are correlated with aspects of the modern environment. Therefore, archaeologists do not really consider predictive models to be explanatory of past land-use or settlement patterns, but rather represent correlations between site locations and the proxy variables of the modern environmental data. Proxy variables do not explain why a site is located where it is, but rather only explain statistical correlations with site locations. This approach may be an obvious weakness in a temporal approach, where locational decisions changed along with changing biotic conditions. However, this approach is directly comparable to traditional predictive modelling approaches, and examines whether taking a temporal approach offers any advantages over traditional approaches, by identifying temporally sensitive correlations.

For this project three temporal models were created for each of the study areas (Archaic, Woodland, Historic for the Winnipeg study area; PaleoIndian, Archaic and Woodland for the MbMF study area), as well as a traditional predictive model for comparative purposes. Sites were allocated to the appropriate time slice and collected on an Excel spreadsheet. The ArcView 3.3 command 'Summarize Zones' was used to determine the values of each of the environmental variables at site locations. These data were collected in a spreadsheet for all sites, and for the time periods of interest. The number of cells for each value of the environmental variables was collected by creating queries in Map Calculator in ArcView 3.3. This information was also transferred, in summarised format, to the site spreadsheets. Once this procedure was complete, the statistical testing for significance could proceed.

Table 3: Significance 'Drop-off'

Variable testing followed the format as detailed in Kvamme (1990). One of the first practical problems of creating temporal models was encountered during the statistical testing. When site locations were tested against the background environment, there was variability in significance. There was a 'significance drop-off', where certain variables were significant for the non-temporal model, yet were non-significant in the individual time slices. Since some of the sample sizes were relatively small, the Dmax values grew very small, since the critical value, in part, is dependent on sample size. Because of the very high bar set for some of these values, despite there being as high as 10% differences in distributions, as shown in Table 3, there was not a significant value. In the case of slope, only when sites were collectively compared to slope across the background environment was there a significant result. Otherwise, the individual time periods were found to be non-significant. This phenomenon creates a problem for this project in that it raises the question of model comparability, as well as the potential need to eliminate some variables from consideration for some of the models. If models are to be comparable, then all factors, other than the temporal factor, should be held constant, in order that no factors, other than the temporal factor, could cause differences in the model. Therefore, the decision was made that all of the principal environmental variables would be employed in the modelling for all time periods, despite some non-significant tests. While this decision may create sub-optimal models, the point of this project is to evaluate temporal aspects of site location modelling, therefore justifying this decision.

After the statistical testing was completed, a master spreadsheet for each of the time periods and for all sites was created. Logistic regression requires non-site data to be included in the logistic regression calculation. In order to generate these sites, a random point generator was employed. The random points were filtered for those that were at least five kilometres from a known archaeological site, to ensure maximum differentiation. The location of these random point locations is shown in Figure 17 for the Winnipeg study area and Figure 18 for the MbMF study area. These were also queried for values of the environmental variables of interest at that location, which were also transferred to a master spreadsheet. With all the data for sites and non-sites incorporated into a single database, logistic regression analysis was done using SPSS version 13. Resultant formulae were entered into the Map Calculator in ArcView 3.3 to complete a predictive model.

© Internet Archaeology URL: http://intarch.ac.uk/journal/issue20/2/5.html
Last updated: Wed Aug 9 2006