Reproducing Geocomplexity Explains Spatial Errors

1 Method Overview & Reproduction Scope

1.1 Core Idea

Spatial models rest on an assumption they rarely state aloud: that nearby places behave similarly, so a value can be borrowed from its neighbourhood. Where that holds, models work. Where it breaks — a coastline, a city edge, a mining town in a pastoral district — the model keeps borrowing anyway, and gets it wrong.

Geocomplexity measures where it breaks. At each location it asks how far the local arrangement of a variable departs from smooth spatial dependence: low where the neighbourhood is orderly, high where it is not. The paper's claim follows directly, and is testable: if geocomplexity marks the places where the assumption fails, then it should predict where a model's errors are — and it does, without ever seeing the model.

• Research problem: spatial model errors are not random, they cluster, and the usual diagnostics report only that they cluster rather than why.
• One-line contribution: geocomplexity turns the local breakdown of spatial dependence into a number computed from the explanatory variables alone, and that number explains a measurable share of the errors of ordinary, non-linear and locally varying models.
• Why it matters: a diagnostic computed before fitting tells you where the model will struggle; and because it is just another variable, feeding it back in improves the model — R² rises from 0.585 to 0.622 for a linear model and 0.802 to 0.836 for geographically weighted regression in the demo below.

Published reference · geocomplexity

1.2 Method Logic

Geocomplexity sits at the root of a family of methods on this site: once you can measure where spatial dependence breaks down locally, you can use that measure as a diagnostic, as an explanatory variable, or as a weight.

Key concepts and equations

• Spatial local complexity (the moran measure, package default). The decomposition in Fig. 1 above:
  ρ_i = −(1/m) Z_i Σ_j W_ijZ_j − (1/m) Σ_j W_ijZ_j (1/v_k) Σ_k W_jkW_ikZ_k
  The first term is local Moran — the location against its neighbours. The second is what geocomplexity adds — the neighbours against each other. A place can agree with its neighbours while those neighbours disagree among themselves, and only the second term sees it.
• Two alternative measures, for different questions. spvar is the spatial variance, Γ = Σ_iΣ_j≠iω_ij(y_i−y_j)²/2 ÷ Σ_iΣ_j≠iω_ij — the raw local fluctuation. shannon is the entropy of the neighbourhood. They are not interchangeable, and Step 2 measures how far apart they are.
• One variable or the whole configuration. geocd_* gives the complexity of each variable separately; geocs_* gives one number for the whole geographical configuration, built from the similarity between locations.
• The claim under test. With e_i the error of some fitted model at location i, geocomplexity should explain |e_i|. Because the strength of that relationship varies over space, the explanation itself is fitted with geographically weighted regression — as in the paper.

Published reference · geocomplexity

1.3 Reproduction Scope

Everything tagged Demo run regenerates from this folder; everything tagged Published reference is a static image from the article. The demo runs end to end with R 4.6.0 and geocomplexity 0.3.0.

How to use this page

• To reproduce the demo: run Rscript run-all.R test, then Rscript run-all.R. Every number and figure below regenerates.
• To use geocomplexity on your own data: §2.3 gives the data contract — one layer with a response and explanatory variables — and §4.1 the porting steps.
• To see geocomplexity used as a building block: the LPI tutorial treats the geocomplexity of every variable as an extra explanatory variable in a local stratified-power model.

2 Setup, Structure & Data Contract

2.1 Environment & Dependencies

The method is the CRAN geocomplexity package, written by the authors of the article. The pipeline adds the two model families whose errors are examined, and each step reports and continues if an optional package is missing.

# The method and its spatial-weights helper
install.packages(c("geocomplexity", "sdsfun", "sf", "spdep"))

# The models whose errors are explained
install.packages(c("e1071", "GWmodel"))

# Figures
install.packages(c("ggplot2", "scales"))

2.2 Project Structure & Configuration

One entry script runs the five pipeline steps, the figures, or the tests. Each step writes to results/, tables/, figs/ or data/derived/ and reads what the previous step wrote, so any step can be re-run in isolation.

GC/
├── gc.html                        # this guide
├── run-all.R                      # entry point: full run, single steps, figures, test
├── assets/                        # figures shown in this guide (web copies)
├── config/project-config.R        # the ONLY file to edit for a new domain
├── R/
│   ├── 00-config.R, 01-helpers.R  # paths, IO, spatial weights
│   ├── 10-prepare-data.R          # read the layer, build the weight matrix
│   ├── 20-geocomplexity.R         # compute GC; compare moran / spvar / shannon
│   ├── 30-baseline-models.R       # MLR, SVR, GWR and their errors
│   ├── 40-explain-errors.R        # GWR of |error| on geocomplexity  <- the claim
│   ├── 50-improve-models.R        # GCMLR and GeoCGWR
│   ├── 90-tables.R                # LaTeX tables + session info
│   └── p01 … p03                  # figure scripts
├── data/econineq.gpkg             # 333 SA3 regions, the published case
├── tests/                         # test-01 published values; test-02 properties
├── env/                           # requirements.md, session-info.txt
├── results/  tables/  figs/       # generated output, never hand-edited

Main configuration file

INPUT_FILE <- "data/econineq.gpkg"   # one layer: response + explanatory variables
RESPONSE   <- "Gini"
PREDICTORS <- NULL                    # NULL = every remaining column

# The neighbourhood definition. Geocomplexity is a statement about a
# neighbourhood, so this is a modelling choice, not a detail.
WEIGHTS_TYPE  <- "contiguity"        # "contiguity" or "knn"
WEIGHTS_QUEEN <- TRUE                # queen (TRUE) or rook (FALSE)
WEIGHTS_STYLE <- "B"                 # "B" binary, "W" row-standardised

GC_METHOD           <- "moran"                            # the working measure
GC_METHODS_COMPARED <- c("moran", "spvar", "shannon")     # compared in Step 2

# Models whose errors are explained, and the geocomplexity that explains them
SVR_COST <- 1.4; SVR_GAMMA <- 0.2      # the paper's cross-validated optimum
GWR_APPROACH <- "AIC"                  # bandwidth criterion
ERROR_EXPLAIN_VARS <- c("Income", "Indemp")

2.3 Data Contract

Geocomplexity needs a geometry, because it is computed from neighbourhoods. One layer — polygons or points — carrying the response and the explanatory variables is the whole requirement.

3 Reproduction Pipeline, Results & Validation

3.1 Pipeline Execution

cd GC                      # this folder
Rscript run-all.R test     # 1. check against the published values
Rscript run-all.R          # 2. full pipeline

Rscript run-all.R 30 40    # re-run single steps (10 20 30 40 50 90)
Rscript run-all.R figures  # only the figure scripts

3.2 Reproduce the Published Values

Two different things are checked here, and it is worth being precise about which is which.

The package's published values — reproduced exactly

run-all.R test locks the numbers published in the geocomplexity package vignettes, computed by the method's authors on this same layer. All twelve checks agree to four decimals, including the adaptive bandwidth the GWR search selects.

The article's tables — mechanism reproduced, magnitudes not

The article's Tables 3–4 are computed after a variable-selection step that it does not fully specify, while the shipped layer carries all eight explanatory variables. The absolute values therefore differ. What survives — and what the paper actually argues — is the ordering:

3.3 Core Analytical Steps

Five stages: purpose → code → output → result → how to read it. Step numbers match the script numbers (Step 4 is R/40-explain-errors.R).

x  <- sf::read_sf("data/econineq.gpkg")
wt <- sdsfun::spdep_contiguity_swm(x, queen = TRUE, style = "B")

Step 2 · the measure

Compute geocomplexity — and check the measure against its alternatives

What this step does

Computes the geocomplexity of every explanatory variable, then recomputes it under all three measures so the choice is visible rather than buried in a default.

Code

R — from R/20-geocomplexity.R Copy

gc <- geocomplexity::geocd_vector(xp, wt = wt, method = "moran",
                                  normalize = TRUE, returnsf = FALSE)

# and, for the comparison, the same call with method = "spvar" / "shannon"

Example output

Demo run

Demo run

Published reference · geocomplexity

◆How to read this result

• Complexity is a property of a variable, not of a place. Income is complex through the arid interior and simple across the settled south-east; industrial employment is the reverse. There is no single map of “complex regions” to draw — each variable carries its own.
• Read the maps in quantile classes, not on a linear ramp. Geocomplexity is strongly right-skewed, so an equal-interval ramp paints almost every region the same shade and hides the structure Step 4 goes on to use.
• Check the measure before trusting the pattern. The comparison in fig04 exists so the choice of moran, spvar or shannon is a stated decision; the warning below shows how little the three agree.

!The three measures are not interchangeable

Across the eight variables, moran and spvar correlate at a median of −0.09 — essentially unrelated — while moran and shannon reach +0.63 and spvar against shannon is −0.60. They answer different questions: local autocorrelation, raw fluctuation, and neighbourhood entropy. Name the measure in the Methods section; a result under one is not a result under another. The LPI tutorial uses spvar deliberately, because there geocomplexity enters as an explanatory variable describing pattern intensity rather than as a diagnostic of dependence.

Step 3

Baseline models — and the precondition that makes the claim worth testing

What this step does

Fits three models of increasing spatial awareness — a global linear model, a non-linear but aspatial support-vector regression, and a locally varying GWR — and keeps their errors. Then it tests whether those errors are spatially structured at all.

Code

R — from R/30-baseline-models.R Copy

f <- stats::as.formula(paste(RESPONSE, "~", paste(preds, collapse = " + ")))

# Three models of increasing spatial awareness; each leaves errors behind
m_mlr <- stats::lm(f, data = d)                          # global, linear
m_svr <- e1071::svm(f, data = d, kernel = "radial",      # non-linear, aspatial
                    cost = SVR_COST, gamma = SVR_GAMMA)
bw    <- GWmodel::bw.gwr(f, data = sp, approach = GWR_APPROACH,
                         kernel = GWR_KERNEL, adaptive = GWR_ADAPTIVE)
m_gwr <- GWmodel::gwr.basic(f, data = sp, bw = bw,       # locally varying
                            kernel = GWR_KERNEL, adaptive = GWR_ADAPTIVE)

errs[["MLR"]] <- y - stats::fitted(m_mlr)                # the subject of Step 4

Example output

Model	R²	RMSE	Errors: Moran's I
MLR — global linear	0.5846	0.0306	+0.422 (p ≈ 10⁻³⁰)
SVR — non-linear, aspatial	0.8030	0.0211	+0.234 (p ≈ 10⁻¹¹)
GWR — locally varying	0.8024	0.0211	+0.105 (p = 0.002)

Published reference · geocomplexity

◆How to read this result

• Clustered errors are the precondition, and they are tested rather than assumed. If Moran's I on the errors were near zero there would be nothing spatial left to explain, and Step 4 would be measuring noise.
• Spatial awareness shows up in the errors, not just the fit. Moran's I falls from +0.42 to +0.23 to +0.11 as the model becomes more spatially explicit. GWR has already absorbed most of the spatial structure — which is exactly why it has least left for geocomplexity to explain.
• SVR and GWR fit almost identically here (0.803 vs 0.802) but differently. The same R² conceals very different error geographies, and the Moran's I column is what reveals it.

Step 4 · the claim

Geocomplexity explains the errors

What this step does

Regresses each model's absolute error on the geocomplexity of income and industrial employment, using GWR so the strength of the explanation is allowed to vary over space — as the article does. Magnitude, not sign, is what is being explained: a model can be wrong in either direction.

Code

R — from R/40-explain-errors.R Copy

dat$abs_error <- abs(err[[nm]])          # what the model got wrong
f  <- abs_error ~ GC_Income + GC_Indemp  # explained by geocomplexity alone
bw <- GWmodel::bw.gwr(f, data = sp, approach = "AIC", adaptive = TRUE)
g  <- GWmodel::gwr.basic(f, data = sp, bw = bw, adaptive = TRUE)

Example output

Errors of	Bandwidth	R²	RSS	AIC
MLR	37	0.1806	0.1101	−1685
SVR	55	0.0888	0.0763	−1816
GWR	114	0.0220	0.0654	−1884

Demo run

Demo run

Published reference · geocomplexity

◆How to read this result

• The ordering is the finding. 0.181 → 0.089 → 0.022. The better a model already handles space, the less geocomplexity has to say about its mistakes. That is what makes this a claim about spatial dependence rather than a coincidence: it behaves the way the theory says it must.
• Nothing about the model was used. Geocomplexity is computed from the explanatory variables and the geometry alone, before any model is fitted. That it still predicts where the errors fall is the whole point — and it means the diagnostic is available before committing to a model.
• The relationship is local, which is why the global scatter looks flat. Pooled over the country the association is weak; mapped, it is concentrated in the dense eastern cities where regions are small and neighbourhoods change quickly. A global regression would have concluded there was nothing here.
• Bandwidth tells the same story from another angle. It widens from 37 to 114 regions as the model improves: the remaining error structure gets broader and vaguer until there is little local signal left to chase.

✓Checkpoint

results/error-explanation.csv holds one row per model, and the R² column should descend with the model's spatial awareness. If it does not, check that the errors were actually clustered in Step 3.

Step 5

Improve the models — the practical consequence

What this step does

If geocomplexity knows where a model is wrong, giving the model that information should help. Two ways: add it as explanatory variables (GCMLR), or build it into the geographical weighting itself (GeoCGWR).

Code

R — from R/50-improve-models.R Copy

# geocomplexity as extra explanatory variables
m2 <- lm(Gini ~ ., data = cbind(d, gc))

# geocomplexity inside the geographical weighting
g2 <- geocomplexity::gwr_geoc(Gini ~ ., data = x, bw = "AIC", adaptive = TRUE)

Example output

Model	Variant	R²	Adjusted R²
MLR	baseline	0.5846	0.5743
GCMLR	+ geocomplexity	0.6215	0.6024
GWR	baseline	0.8024	0.7409
GeoCGWR	+ geocomplexity	0.8360	0.8319

Demo run

◆How to read this result

• The adjusted R² is the honest column. GCMLR adds eight variables, so a naive R² would rise anyway. Adjusted R² rises too (0.574 → 0.602), which means the gain survives the cost of the extra parameters.
• GeoCGWR gains more than it appears. R² moves 0.802 → 0.836, but adjusted R² moves 0.741 → 0.832 — because weighting by geocomplexity buys the fit with far fewer effective parameters than widening the bandwidth would.
• Improvement and explanation are the same fact. Step 4 showed geocomplexity knows where the linear model errs; here that knowledge is handed back and the error shrinks. A method that explained errors but could not reduce them would be far less interesting.

3.4 Reading the Results Together

4 Adaptation, Writing & Reproducibility

4.1 Port to Your Domain

Geocomplexity needs only a geometry and some variables, so it travels wherever spatial models do. Within this tutorial series it appears in three different roles:

What to edit

1. config/project-config.R — INPUT_FILE and RESPONSE, then the neighbourhood: WEIGHTS_TYPE ("knn" for points), WEIGHTS_QUEEN and WEIGHTS_STYLE.
2. GC_METHOD — "moran" when geocomplexity is a diagnostic of spatial dependence, as here; "spvar" when it stands for pattern intensity, as in the LPI tutorial; "shannon" for neighbourhood diversity. Step 2 prints the correlation between them so the choice can be defended.
3. ERROR_EXPLAIN_VARS — which variables' complexity carries the explanation. NULL uses all of them, which is a reasonable starting point before selecting.
4. Nothing else: Rscript run-all.R test, then Rscript run-all.R.

4.2 Write the Paper

The published study has a clean architecture: define the measure, show it on the data, fit conventional models, then use the measure to explain their errors and to improve them.

Manuscript skeleton

1  Introduction        — spatial models assume dependence; errors cluster;
                          the gap: which places, and why (3 aims)
2  Geocomplexity       — 2.1 Local and surrounding features (all formulas)
                          2.2 Calculation methods and the weight matrix
                          2.3 Design: explaining errors with GWR
3  Case & data         — study area, response, explanatory variables
4  Results             — 4.1 geocomplexity across the study area
                          4.2 baseline models and the structure of their errors
                          4.3 error explained, by model
                          4.4 where it is explained (local coefficients)
                          4.5 improvement when geocomplexity is fed back
5  Discussion          — the ordering across models; the local geography;
                          use as a pre-fit diagnostic; limits of the measure
6  Conclusions         — 6-8 sentences answering the aims

Results-to-manuscript map

4.3 Final Reproducibility Package

References & credits

Demo outputs were generated by the scripts in this folder using the CRAN geocomplexity package. Reference figures are reproduced from the article, which is open access, for comparison.

1. Zhang Z, Song Y, Luo P, Wu P (2023). Geocomplexity explains spatial errors. International Journal of Geographical Information Science 37(7):1449–1469. doi:10.1080/13658816.2023.2203212 · PDF (source of the concept, workflow, Fig. 5, Fig. 6 and Fig. 8 reference images, and of the case reproduced here)
2. Lyu W, Song Y, et al. geocomplexity: Geographical Complexity. R package. CRAN · calculation vignette · application vignette (the reference implementation, the source of the demo data, and of the values locked in tests/test-01)
3. Anselin L (1995). Local indicators of spatial association — LISA. Geographical Analysis 27(2):93–115. doi:10.1111/j.1538-4632.1995.tb00338.x (the local Moran term geocomplexity extends)
4. Sun Y, Wu G, Song Y, et al. (2026). Local effects of pattern interactions in driving urbanization. Int J Appl Earth Obs Geoinf 146:105072. doi:10.1016/j.jag.2025.105072 · LPI tutorial (geocomplexity used as an explanatory variable)