2.4 Townships: Within-State Variation and Poverty Clustering The state/region quadrant analysis in Section 2.3 characterizes poverty at an administrative level, but the averages underlying each quadrant position conceal substantial variation within states. Townshiplevel estimates reveal that the within-state range of MPI headcounts is often as analytically significant as the between-state differences, and that the quadrant assignment of a state reflects the balance of its internal distribution rather than a uniform poverty profile. Township MPI headcounts are estimated using a model-based small area estimation approach6 that combines PPS 2025 household data with township-level auxiliary information and benchmarks results to state totals. These estimates should be interpreted as indicative comparisons that support geographic prioritization rather than precise measurements of poverty levels. The full township MPI estimates are presented in Appendix Table 5. Figure 3 Township-level distribution of multidimensional poverty (MPI Headcount) Source: PPS 2025 (UNDP) 6 Township-level MPI headcounts are estimated using a unit-level small area estimation (SAE) approach. A generalized linear mixed model (GLMM) with a binomial logit link is fitted to PPS 2025 household data, with MPI poverty status as the binary outcome, township-level random intercepts, and five township-level auxiliary covariates: night-time light intensity (proxy for economic activity), a Vulnerability to Conflict Index derived from ACLED event data, flood risk, forest cover (proxy for remoteness and geographic isolation), and mining area. All covariates are log-transformed and standardized. For townships with 30 or more sampled households, the predicted headcount is the weighted mean of household-level predicted probabilities from the full model. For townships with fewer than 30 sampled households or no sample, a synthetic prediction based on fixed effects only is used. All township estimates are then benchmarked to the directly observed state/region MPI from PPS 2025 by applying a proportional scaling factor, ensuring internal consistency between the township and state/region levels. Uncertainty is estimated via a 500-iteration parametric bootstrap that propagates uncertainty in both fixed effects and random intercepts, with benchmarking applied in each replicate. L AYE R S O F DE P R I VAT I O N : M u l t i di m en si on a l Pov er t y i n M y anmar 15