Previous studies of colourless Romano-British vessel glasses have suggested that, regardless of vessel type, they show considerable compositional homogeneity. Intriguing differences in variability (as opposed to mean composition) have, however, also emerged. This paper reports on a compositional study of 243 vessels, that is larger and more carefully controlled than in previous studies of this kind. Unexpected compositional differences have been found both between and within the four vessel types studied. We discuss the implications of these results in the context of different models that have been proposed for glass-making and glass-working in the Roman world.

SUMMARY Principal component analysis is commonly used in archaeometric applications to identify or display structure in the chemical composition of archaeological artefacts. A recurring topic of debate is whether, and how, data should be transformed and whether, after transformation, standardization should be used. Most discussion has focused on the use of logarithmic transformations. The merits of different approaches are investigated empirically in the paper, using 20 published data sets showing different degrees of structure. The opportunity is taken to examine the merits of the rarely used rank transformation, which has potential attractions when outliers occur or the variables are unusually distributed.

Compositional data arise commonly in archaeometry, in the study of artefact compositions where the variables measured either sum to 100%, or can be viewed as a subset of such a set of variables. There has been debate in Archaeometry about the appropriate way to analyse such data statistically, which amounts to argument about how the data should be transformed prior to statistical analysis. This paper reviews aspects of the debate and illustrates, using both simulated and real data, that what has been proposed as the 'correct' theoretical approach-log-ratio analysis-does not always work well. The reasons for this are discussed.

Principal component, cluster and discriminant analysis are multivariate statistical methods that are widely used in archaeometry. They are examples of what are known in some literatures as unsupervised and supervised learning methods. Over the past 20 years or so, a wide variety of other learning methods have been developed that take advantage of modern computing power and, in some cases, have been designed to handle data sets more complex than those often used in archaeometric data analysis. To date, these methods have had little impact on archaeometry. This paper reviews, in a largely non‐technical manner, the ideas behind these newer methods; illustrates their use on a variety of data sets; and attempts to assess their potential for future archaeometric use.

14 The amplitudes of the independent signals form a multi-dimensional time-varying 15 vector which was logged continuously for eight months. We found that combined 16 with specifically tailored weighting factors, this vector provides a signature highly 17 specific to the swarming process and its build up in time, thereby shedding new 18 light on it and allowing its prediction several days in advance. 19The output of our monitoring method could be used to provide other signatures 20 highly specific to other physiological processes in honey bees, and applied to better 21 understand health issues recently encountered by pollinators.22 23