Panel Local Indicators of Spatial Association (Panel LISA)

Running LISA separately for each year is exactly what Panel LISA does at the computation stage; the 'panel' extension lies in the comparative analysis across years — testing persistence, computing stability indices, or using the resulting local statistics as inputs to a panel regression. Using a consistent weights matrix and applying multiple-testing corrections across all periods together are best-practice panel-specific refinements.

Contiguity matrices (queen or rook) are common for polygonal areal units (e.g., administrative regions). Distance-band or k-nearest-neighbour matrices suit point or irregular data. The choice should be theoretically motivated by the diffusion or interaction mechanism you expect. Always test sensitivity to at least two alternative specifications and report differences.

Apply a false-discovery rate (FDR) correction (Benjamini-Hochberg) or a Bonferroni correction to the permutation p-values before mapping clusters. FDR is generally preferred for spatial data because it is less conservative. Some software packages (PySAL, spdep) offer built-in FDR corrections for LISA.

No. Panel LISA is exploratory — it identifies where spatial clusters exist over time but does not control for covariates or estimate structural parameters. Use it to motivate and specify a spatial panel model (SAR, SEM, SDM), not to replace one.

LISA can still be computed for each period independently using the units observed in that period, but direct comparison across periods becomes harder. If missingness is systematic rather than random, spatial clusters in underrepresented periods may reflect data gaps rather than true spatial patterns. Report the number of observations per period and check for spatial selectivity of missingness.