The filter-and-refine strategy is well-established as the basis for spatial join algorithms. In contrast to the filter step, the refinement step has received little attention, despite contributing significantly to the total cost of a join evaluation. This paper reports investigations of spatial join algorithms for z-ordering and R-trees, with particular emphasis on interactions between choices of algorithms for the filter, sequencing and refinement steps and on the effects of clustered and unclustered organization of full spatial descriptions of objects. Our experiments show that while it is in general desirable to introduce an additional housekeeping step to reduce I/O costs of the refinement step, it is not necessary in all cases. In addition, we propose a new caching strategy for spatial joins, called zig-zag, which outperforms its competitors in all but one case. These results suggest that spatial joins need caching strategies other than non-spatial ones. Furthermore, our experiments confirm that the choice of the sequencing strategy used is very important and that clustering has a significant influence on join performance.