DAACS Seriation Method
DAACS staff aims to produce a seriation-based chronology for each site using the same methods (see Neiman, Galle, and Wheeler 2003 for technical details). The majority of sites in the archive are comprised of data derived from deposits within quadrats. On these sites, only assemblages from features or stratigraphic groups with more than five ceramic sherds are included in these ceramic-based seriations. Plowzone contexts do not contribute to a DAACS seriation-based chronology.
The DAACS Caribbean Initiative focuses on exploring large-scale change on slave villages in the Caribbean through the use of shovel-test-pit surveys. For villages with extensive STP coverage, including the New River villages, a variation on our site-based seriation method is employed. This is because each STP is small (50 cm. in diameter) and provides a small artifact sample. As a result, STP assemblages are rife with sampling error. The samples from individual STPs are so small that variation among STPs is almost entirely statistical noise.
Successfully analyzing STP data, without first aggregating those pits into counting units called sites, requires methods to suppress sampling error. Here we use empirical-Bayes methods. They offer a smart way to smooth both artifact density surfaces and relative frequencies of artifact types. To understand how these methods work, consider an STP - let's call it STP 12. The number of artifacts found in STP 12 is likely to be similar to the number of artifacts in the STPs within a certain distance of it. The information contained in the neighborhood of pits is combined with the actual number of artifacts from STP 12 to arrive at an estimate of artifact counts that are less influenced by sampling error (Neiman et al. 2008).
We use two forms of Bayesian smoothing in succession. First, to smooth counts of ceramic ware types in individual STPs, we use a gamma-Poisson model. The gamma-Poisson algorithm highlights positive STPs that are near other positive STPs. We then use a beta-binomial model to estimate relative frequencies (percentages or proportions) of ceramic ware-types in individual STPs. Together two forms of Bayesian smoothing provide smoothed, stable estimates of artifact-type frequency variation in individual STPs, allowing us to see overall site patterning that may otherwise be distorted using raw data (Neiman et al. 2008).
To infer a chronology from the STPs we used correspondence analysis (CA) of ware type frequencies. We employ CA because with the numbers of STP assemblages in the hundreds, a traditional manual frequency seriation is completely impractical. CA converts a data matrix of ware-type frequencies into a set on scores which estimate the positions of the assemblages on underlying axes or dimension of variation. MCD's are weighted averages of the historically documented manufacturing date for each ware type found in an assemblage, where the weights are the relative frequencies of the types. Measuring the correlation between CA axis scores and MCDs offer an indication of whether the CA scores capture time (Ramenofsky, Neiman and Pierce 2009).
Dating The Spring's Village
The CA results for The Spring indicates that there is no temporal trend within the village (Figure 1) nor is there a significant correlation between the CA dimension-1 scores and the MCDs (Figure 2). However, a village-wide Mean Ceramic Date of 1787 points to the village's occupation the late 1700s. Two other measures that are less sensitive to excavation errors and taphonomic processes that might introduce a small amount of anomalously late material into an assemblage were used. They are TPQp90 and TPQp95. The TPQp95 of 1797 provides a robust estimate of the site's TPQ based on the 95th percentile of the beginning manufacturing dates for all the artifacts comprising it. The TPQp90 of 1775 provides a more robust estimate of the site's TPQ based on the 90th percentile of the beginning manufacturing dates for all the artifacts comprising it.