Minnesota Department of Transportation

511 Travel Info

MnModel

Minnesota Statewide Archaeological Predictive Model

MnModel Survey Bias

Map of negative survey points and archaeological sites

Map of negative survey points and archaeological sites

MnModel is the first archaeological predictive model to make survey bias explicit in the final results so that the model's value for any given place can be assessed.

As we evaluated the Site Probability Models, it became apparent that they predicted surveyed places almost as well as they predicted site locations. This implied a high degree of survey bias and reduced our confidence in the interpretation of the predictive models, posing the question of whether areas were categorized as low probability because no sites were there or because there had been no surveys there. This led to the development of the Survey Probability Model, which might be thought of as a model of survey bias. This model has no precedent

Reasons for survey Bias

Archaeological field survey is very labor intensive, time-consuming, and, ultimately, expensive. Naturally, archaeologists are more interested in finding archaeological features and artifacts than in spending many hours surveying and finding nothing at all. Consequently, they tend to focus their survey efforts on places where they expect sites to be - usually places near water. Even so-called 'probabilistic' surveys usually are stratified so that more locations are surveyed near water than away from water. Although the goal of the 1995 and 1996 MnModel field surveys was to 'provide data on site location and non-site locations based on random sampling,' the archaeologists' stratified survey design gave 'rarer landforms... priority over more common ones to insure they be represented in the sample strategy.' A truly random survey, such as the 1997 MnModel field survey, would result in a sample representing each landform in proportion to its occurrence in the landscape.

Detecting Survey Bias

In Phases 1 and 2 of MnModel, we used 'negative survey points' as non-site locations. These were points mapped in sections that were surveyed, at least in part, but where no sites were found. By comparing the distribution of surveyed locations within the early models' high, medium, and low probability areas, it became apparent that survey distributions reflected archaeologists' notions of where sites were most likely to be found. Consequently, MnModel's predictive models are biased by the archaeologists' own intuitive models of probable site location. The map (above right) shows negative survey points in red and archaeological sites in black. The relationship of both to water features is apparent.

Modeling Survey Bias

To account for this bias, survey locations were modeled using the same methods and environmental variables that were used to model site locations. The resulting Survey Probability Model indicates which parts of the landscape have been adequately surveyed (darker purple in the map in the banner above).

Incorporating Our Understanding of Bias into the Predictive Models
Finally, the Site Probability and Survey Probability models were combined to create the Survey Implementation Model. This model qualifies the values of site probability by reference to the values from the survey probability model.

  • Where survey probability is high, site probability values may be high, medium, or low. There is a high degree of confidence in these values, since these are the kinds of environments that have been well surveyed in Minnesota.
  • Where survey probability is medium, site probability values may be possibly high, possibly medium, or possibly low. There is less certainty about site potential in these environments because they have not been as well surveyed.
  • If survey probability is low, we know there has been very little survey; site probability is classified as suspected high, suspected medium, and unknown. Suspected high and suspected medium imply that these areas have similarities to places where sites are found, though they also have characteristics of places archaeologists have not adequately surveyed. These areas need to be surveyed to confirm whether the site potential expressed by MnModel is valid. In areas classified as unknown, both survey probability and site probability are low. Archaeologists, using their expert systems models, have assumed that sites will not occur here. Yet sites are occasionally found in the 'unknown' areas. We need more surveys to help us understand the environmental factors associated with these sites.