3. Difficulties in Temporal Predictive Modelling

There are several practical difficulties which must be taken into account when doing a temporal version of predictive modelling. While these problems would not necessarily prevent a temporal model from being created, they must be solved before the modelling can proceed.

The majority of the problems are with the site database itself, including problems with spatial aspects of the database (e.g. see Duncan and Beckman 2000; Ebert 2000), as well as with temporal aspects, which are most salient here. With the expansion of the cultural resource management industry in North America, the site database has grown over the last fifty years. The locations of sites have been collected using a variety of different methods, from map orienteering to global positioning system coordinates. Also, since the database has been created by numerous individuals with varying levels of experience and competence, there are varying levels of error, and therefore confidence. However, most of the keepers of the database, primarily in provincial government departments in Canada, have also invested time in improving site database quality. Since there is little that can be done in terms of positional accuracy, one must accept the database, be cognizant of the errors and proceed accordingly. Positional errors will be a problem with any type of predictive model and should be considered when interpreting and applying such models.

However, there are other problems with the site database, mainly relating to the nature of the data collected by archaeologists. The majority of the problems encountered in temporal predictive modelling concern the cultural affiliation of the site. Specifically, problems arise with regard to archaeologists' ability to identify the appropriate archaeological culture based on the evidence at the site. We can consider these as 'sites with problematic cultural affiliations'.

The first problem concerns sites with no presently known cultural affiliations. Often the evidence collected at an archaeological site is bereft of diagnostic artefacts to make an assignment to an archaeological culture. If it is not possible to make such a determination, the sites are recognised as archaeological, especially when a number of non-diagnostic artefacts are found, but labelled as an 'indeterminate' cultural affiliation. These sites are clearly problematic, as they are important variables spatially, but not temporally, leaving the model with only half of the required information.

The second of the major problems concerns those sites in the database with vague cultural affiliations. In some cases, sites are listed with affiliations such as 'pre-contact', meaning that it could be any time period from the first occupation to the day before the Europeans made contact. While these sites contain a small amount of temporal information, and are slightly better than indeterminate cultural affiliations, the advantage is small. For example, does 'pre-contact' mean that this label should be assigned to all of the temporal periods for modelling, or is it an indication that only non-diagnostic artefacts were found? Without consulting individual site reports from each of these vaguely identified sites, it is impossible to tell.

The last of the problems with cultural affiliations concerns sites with unexpected cultural affiliations. In this case, sites have diagnostic artefacts, but are likely misidentified. For this project, one site was identified as belonging to a boreal forest complex (Ebell 1982), despite the fact the area would have been under Glacial Lake Agassiz at the time.

While these problems present a challenge for temporal predictive modelling, they are not insurmountable. In the case of all three types of cultural affiliation difficulties, the question becomes 'how are such sites to be handled'? There are several options available, including: 1) their exclusion from consideration, 2) assignment to other time periods based on proportions or 3) the creation of a separate model for these sites. Each of these solutions has its own advantages and disadvantages.

Exclusion is the simplest of solutions; however, the exclusion of these data may eliminate variation from the sample of sites. Furthermore, exclusion may eliminate certain functional site categories more often, leading to a lesser understanding of pre-contact land-use patterns. For example, resource extraction sites such as lithic sources may only rarely have diagnostic artefacts, which would mean no method of determining cultural affiliation, and therefore such sources would be excluded from the database more often than not.

Assigning sites to time periods based on proportions is the 'polling' solution. This approach would mimic the way 'undecided' individuals are assigned to opinions based on the proportion of individuals in the opinion classes. In a temporal predictive model, this might mean that sites with a known cultural affiliation to the Woodland period sites, which represent 30% of the total sites in the study region, are randomly proportionally allocated 30% of the sites of indeterminate cultural affiliation. The problem with this approach is the dilution of temporal patterns. If undetermined affiliation sites are assigned randomly, it would increase the probability that sites would be included from other time periods, thus decreasing the chances of determining difference in temporally distinct land-use patterns.

Alternatively, one could treat these sites of indeterminate cultural affiliation as if they were a unique time period. This approach would create a separate model for these sites. However, by making a model for these sites, one could be blending results from multiple time periods. It would also raise the question of what to do with the predictions from this type of model. Should the results be used as supplementary information or should they be incorporated in some way with the other models from known time periods? While this approach would be superior in some ways to either exclusion or allocation, the utility of the predictions must be in question.

A second major problem for temporal models is the resolution of the time slices, with very fine temporal resolutions perhaps resulting in time periods having few or no sites. Statistically, it is undesirable to do many statistical tests with small sample sizes. If the sample is extremely small one must consider how useful the examination of patterns may be, even using those tests which are not proscribed for small sample sizes. Unfortunately, the solution to this problem is to employ very coarse time periods, in order to group together as many sites as possible into larger time periods. While it would clearly be more desirable to use finer temporal divisions, such as specific archaeological cultures, often the site database will not support this finer resolution, and somewhat coarser temporal resolutions become necessary.

The final problem in temporal predictive modelling occurs at the end of the temporal modelling process; specifically what to do with the resultant predictions. Traditional predictive modelling techniques return a single map of archaeological potential. A temporal modelling process results in the creation of multiple predictive models, varying in number depending on the number of divisions of time periods. Since most archaeological predictive models are used in a consulting capacity, most clients would not be interested in receiving a series of models, each of which need to be consulted during planning. Ultimately a plan for integrating the predictions, yet not losing the unique temporal patterns, must be devised. While this solution lies outside the scope of the current research, as it is more of a management problem rather than a research problem, it is an area that requires careful consideration.

While each of these issues have some effect on the ultimate approach to modelling, and need some consideration when creating a temporal modelling program, these considerations by no means prevent the creation of temporal predictive models.