-
Note that spatial neighborhood graphs may change meaning after subsetting.
-For example, for a k nearest neighbor graph, after subsetting, some cells
-might no longer have all k nearest neighbors from the original
+
The SFE method has special treatment for the spatial graphs. In listw,
+the neighbors are indicated by indices, which will change after subsetting.
+The SFE_graph_subset option determines whether the graphs are
+subsetted or reconstructed. In the default (options(SFE_graph_subset =
+TRUE)), the graphs are subsetted, in which case singletons may be produced.
+For options(SFE_graph_subset = FALSE), which is the behavior of
+versions earlier than Bioc 3.20, the graphs are reconstructed with the
+parameters recorded in an attribute of the graphs. This option can result
+into different graphs. For example, suppose we start with a k nearest
+neighbor graph. After subsetting, cells at the boundary of the region used to
+subset the SFE object may lose some of their neighbors. In contrast, when the
+graph is reconstructed, these same edge cells will gain other cells that
+remain after subsetting as neighbors in the new KNN graph.
+
Details
+
The option SFE_graph_subset was introduced because subsetting is
+usually faster than reconstructing and in some cases such as distance-based
+neighbors and Visium spot adjacency give the same results. It was introduced
+also because of the development of alabster.sfe for a
+language-agnostic on-disk serialization of SFE objects and some parameters
+used to construct graphs have special classes whose alabaster methods
+have not been implemented, such as BPPARAM and BNPARAM, so when
+reconstructing, the defaults for those arguments will be used.
+
The edge weights will be recomputed from the binary neighborhood indicator
+with the same normalization style as the original graph, such as "W" for row
+normalization. When distance-based edge weights are used instead of the
+binary indicator, the edge weights will be re-normalized, which is mostly
+some rescaling. This should give the same results as recomputing the distance
+based edge weights for styles "raw", "W", and "B" since the distances
+themselves don't change, but the effects of other more complicated styles of
+re-normalization on spatial statistics should be further investigated.
+