Effect of distance measures and feature representations on distance-based accessibility measures

Distance-based accessibility measures are often built using vector representations of origin and destination features, and Euclidean or network-based distances. There are few comparisons of how the choice of feature representations and distance types affects results. Existing comparisons often use Spearman’s rank correlation coefficients. This study seeks to understand the effect of using different types of distance and feature representation on accessibility measures by comparing accessibility measures using the Bland-Altman plot, which measures the agreement between two variables. Accessibility measures for recreational areas in Malta's Grand Harbour Area (GHA) were calculated. Two distance-based measures were compared: distance to nearest recreational area (DNRA) and Nearest Recreational Area ID (NRAID), measured from all residential blocks within the study area. Each of these two measures was calculated using two different vector representations for destinations (the recreational areas), access points and internal geometric centroids, and three different ways of measuring distance, Euclidean, network and full-network distance (network distance plus distance from feature representation to network). The combinations were compared for each measure. DNRA results were compared using the Bland-Altman plot and NRAID results were compared using percentage overlap. Results and conclusions Analysis showed that NRAID and DNRA results are especially affected by the selection of Euclidean distances versus network and full-network distances, and to a lesser extent the selection between centroids and access points, especially when using network or full-network distances. An important assumption in this study was that full-network distance and access points are the most realistic alternatives for their respective groups. They are also the most demanding in terms of data collection, pre-processing, computation and analysis. One can conclude that the type of data and distance measure used and their possible effects on distance-based results must be taken into account. Some of the methods used to compare results were not fully applicable to the case and led to generic results. The conclusions drawn are therefore indicative and more work needs to be done to achieve more statistically valid results for comparison with other studies in the field.