1 Introduction Edelman and Eeuwens (1959) proposed that the landscape of south Limburg (fig. 1) reveals the effects of a Roman centuriated land survey. This idea attracted some support (Lambert 1971: 48), but it is not generally accepted. 1 Despite this, we should keep an open mind. The hypothesis is difficult to dismiss on theoretical grounds, and it is supported by empirical results which show anomalies in the distribution of Roman sites, similar to those observed in other areas of centuriation. The centuriation grid (fig. 2) can be located accurately by calculation (Peterson 1993: 43-47). The module is 711.61 m and the orientation is N 42.064° E. One point is located at the Limbricht St-Salviuskerk (186680, 336320) which, according to Edelman and Eeuwens (1959: 53), stands ‘precies aan een hoekpunt’ (precisely at a corner). Their evidence for the centuriation is of five sorts: firstly a large number of existing boundaries have a consistent orientation; secondly major boundaries or roads are spaced at multiples of 2400 Roman feet (hence they could represent remnants of major divisions, or limites, of the grid); thirdly several medieval churches are positioned on these hypothetical limites; fourthly the orientation of some of these churches accords with the proposed grid, and fifthly Roman villas are positioned in a non-random way near the limites. Some of their views can be supported by inspection. Maps show that existing roads, paths and boundaries coincide with the hypothetical limites of the centuriation, and on the ground it is clear that several of these features do not conform locally to natural topography. Quantitative approaches may also be used, and are likely to provide a more secure basis for judgement. An earlier study was that of J.A. Brongers, B.M. Hilwig-Sjöstedt and E. Milikowski, who conducted a numerical analysis of the distribution of the orientation of boundaries. They concluded that the dominant orientations, which vary from place to place, are better related to the morphology of different parts of the landscape than to any overall general Roman influence on the parcelling in the whole region. However, they do not say that there is no centuriation, but that the information cannot be extracted solely from John W.M. Peterson A computer model of Roman landscape in South Limburg an analysis of modern parcel boundaries (Brongers, pers. comm.). 2 Quantitative study of site distribution Since this earlier quantitative study was inconclusive, and since, in any case, undateable boundaries may not be seen as a good source of evidence, another approach is adopted here. This measures the claimed association between the grid and Roman sites of all types, including villas, using a database already independently assembled by Martijn van Leusen (1993: 105), using information from the Nether- lands State Archaeological Service (ROB). In 1992 it held about 1300 records, of which 491 referred to Roman sites, including villas. This is a large data set which had not been collected together to suit Edelman and Eeuwens’ hypo- thesis. It may therefore be used to test their claim. Given that many Roman (and later) sites are expected to be associated with the limites 2 , we can examine the distribution of distances of sites from the grid lines, when compared to the distribution of distances which would be expected if the points are scattered uniform randomly with respect to the grid. It seems reasonable to assume that, for a large grid, this latter distribution would arise. The sites may be non- randomly related to natural features, but there is, in many places, very little relationship between these features and the grid (fig. 5). The Kolmogorov-Smirnov single sample test may be used. The test statistic, D+, is the largest positive difference between the number of points observed at a given distance from the lines of the grid, and the number of points which would be expected on the basis of the null hypothesis (Lapin 1973: 422). In this case it is the maximum value of i | - (1 - (1 - x i ) 2 ) | n where x i is the distance of the i th point in order of distance from the grid lines (Peterson 1993: 69). Tables of critical values of D + show with what confidence we can reject the null hypothesis. One such table, giving values for sample sizes up to 100 was first presented by Miller (1956). For larger samples the critical value, D + a for a given probability, a, can be calculated