Figure 2

Model selection results from Vitoria data. Each bar represents the DIC (left-side plots) or Spearman’s correlation coefficient between the model-predicted and observed data (right-side plots) for a given model. Each covariate was lagged and scaled to the best values (i.e., from model selection shown in Additional file 2: Figure S1 and Additional file 3: Figure S2). “Null” indicates a model with only the neighborhood population size as an offset and random effects of neighborhood. (A) Within-neighborhood effects. Lags are indicated. No scaling factors were used. Full i is M _i,t-13  + Y _i,t-1  + M _i,t-13 * Y _i,t-1 . (B) Comparison of nearest-neighbor (local) versus all between-neighborhood (global) effects. Lags and scales, respectively, were: 13, 0.5 (local, Mj), 12, 0.1 (global, Mj), 1, 0.5 (local and global Yj), 13 & 1, 0.5 (local M _j *Y _j ) and 1 & 1, 0.1 (local M _j *Y _j ). Structure of full j was ∑ _j M _j,t-x ^α1 + ∑ _j Y _{j, t-y} ^α2 + ∑ _j (M _j,t-z Y _j,t-y ) ^α3. Full i covariates (as specified in panel A) were included in each model (i.e., there are 6 covariates in “Full j”). (C) Effects of the type of approximation (f(x _j )) used for weighting global connectivity. Form used for f(x _j ) is indicated under the bars (d, distance; E, economy; D, density). Full i covariates (as specified in panel A) were included in each model (i.e., red and blue bars have 4 covariates). Lags and scales, respectively, were: 1, 0.5 (Y _j f(x _j )), 1 & 1, 0.1 (M _j Y _j f(x _j )). Structure of full j was ∑ _j M _j,t-12 ^0.1 + ∑ _j(Y _j,t-1 f(x _j )) ^0.5 + ∑ _j (M _j,t-1 Y _j,t-1 f(x _j )) ^0.1.