REP17182 - Laserfiche WebLink

Home Browse Search

1 1 1 LJ 3) Models based on strongly skewed or amodal distributions tend to yield unstable prediction equations that do not generalize well across samples. 9) Because of the exploratory nature of model building, predictive models are always in danger of capitalizing on chance, an effect that also results in an unstable prediction equation. Predictive models range from intuitive, verbalistic, and largely subjective designs, as presented by Greiser (1983), to sophisticated multivariate paradigms such as those used recently by Kvamme (1993) and Zier et al (1987). Models may be theory driven and incorporate concepts form ethnographic analogy and/or anecdotal sources, or based solely on a specific sample of measurable physical traits in the archaeological record and the natural landscape. Construction of a valid predictive model for the Dowe Flats project area was constrained by a number of technical considerations. Foremost among these was the small sample of sites available for analysis. Previous work by Burney, Gleichman, Nykamp and others, as well as inventories conducted by Grant concurrent with the production of this document, have yielded a combined sample of 24 prehistoric sites within and adjacent the study area. Sites included in this sample were located on Indian Mountain, Rabbit Mountain, and Dowe Flats proper. With so few sites available for analysis, prediciton methods were limited. Combinations of multiple predictor variables were used only in an exploratory and tentative manner.' Since the current site sample represented data accumulated by various researchers over a period of several years, information on specific site characteristics varied considerably. A total of 22 'Non-experimental multivariate methods generally require at least 20 cases for each independent variable. Violation of this assumption undermines the statistical power of any inferential tests. 80