Spatial analysis Study Guide

📖 Core Concepts Spatial analysis – uses topological, geometric, or geographic properties of entities to study patterns. Spatial dependence – values at nearby locations are statistically related; the basis for interpolation and many models. Modifiable Areal Unit Problem (MAUP) – aggregating point data into arbitrary zones changes totals, rates, or densities; both shape and scale matter. Uncertain Geographic Context Problem (UGCoP) – bias from using aggregates without accounting for movement across zones; linked to MAUP and ecological fallacy. Tobler’s First Law – “everything is related to everything else, but near things are more related than distant things,” justifying spatial methods. Spatial autocorrelation – co‑variation of a variable at neighboring sites; can be positive or negative. Spatial heterogeneity – process parameters vary across space, so a single global model may be inappropriate. --- 📌 Must Remember MAUP → bias from arbitrary spatial partitions (shape + scale). UGCoP → bias when phenomena cross unit boundaries; watch for ecological fallacy. Positive spatial dependence → reduces sample‑mean accuracy vs. independent samples; negative dependence can improve it. Key autocorrelation statistics – Moran’s I, Geary’s C, Getis‑Ord G (global) and their local versions. Kriging – best linear unbiased predictor for spatial interpolation, relies on semivariograms. Inverse Distance Weighting (IDW) – simple interpolation that down‑weights values by distance. Gravity models – flows = (origin factor × destination factor) / distance (or travel‑time) factor. Geographically Weighted Regression (GWR) – produces locally varying parameter estimates. Cellular Automata (CA) – fixed grid, state updates by neighborhood rules; calibrated via Monte‑Carlo. Agent‑Based Modeling (ABM) – autonomous agents with goals; can move, interact, and modify environment. --- 🔄 Key Processes Constructing a spatial autocorrelation test Choose global statistic (Moran’s I, Geary’s C, Getis‑Ord G). Compute spatial weights (e.g., contiguity or distance‑based). Compare observed statistic to permutation distribution → assess significance. Kriging interpolation Build experimental semivariogram → fit theoretical model. Use covariance function to compute weights for unknown locations. Produce prediction and associated variance (uncertainty). Applying a gravity model Define origin variable (e.g., commuters) and destination variable (e.g., jobs). Select distance metric (Euclidean, Manhattan, cost‑distance). Estimate parameters (often via Poisson or log‑linear regression). Calibrating a cellular automaton Set initial land‑use map. Define neighborhood rules (e.g., transition probabilities). Run Monte‑Carlo simulations; adjust rules until output matches observed pattern. Running a GWR Choose bandwidth (fixed or adaptive). Fit local regressions weighted by distance to each target location. Map spatial variation of coefficients. --- 🔍 Key Comparisons MAUP vs. UGCoP – MAUP = bias from arbitrary spatial units; UGCoP = bias from ignoring movement across units. Spatial autocorrelation vs. Spatial heterogeneity – Autocorrelation = similarity of values at nearby sites; heterogeneity = variation of process parameters across space. Cellular Automata vs. Agent‑Based Modeling – CA: fixed grid, rule‑based state changes; ABM: mobile agents with purpose, can modify environment. Global Moran’s I vs. Local Moran’s I – Global gives overall clustering measure; Local identifies specific hot‑spot or cold‑spot units. Kriging vs. IDW – Kriging uses modeled spatial covariance (provides uncertainty); IDW simply weights by inverse distance, no explicit error estimate. --- ⚠️ Common Misunderstandings “Longer coastlines mean more coastline” – Measured length depends on scale; without context it can be meaningless. Locational fallacy – Assuming a person’s entire spatial behavior is captured by a single address. Atomic fallacy – Treating spatial elements as independent “atoms” and ignoring surrounding context. Ecological fallacy – Inferring individual‑level behavior from aggregated zone data; violates within‑unit variation. Ignoring spatial autocorrelation in regression – Leads to unstable coefficients and unreliable significance tests. --- 🧠 Mental Models / Intuition “Nearby = similar” – Visualize a smooth surface; points close together sit on similar heights (values). “Zoom lens” for MAUP – Changing the zoom (scale) or shape of the lens (zone) changes the picture you see; patterns can flip from dispersed to clustered. “Flow as gravity” – Like planets, larger masses (origins/destinations) attract more flow; distance acts as friction. CA vs. ABM analogy – CA = a chessboard where each square follows simple neighbor rules; ABM = a crowd of people each with their own agenda. --- 🚩 Exceptions & Edge Cases Negative spatial dependence – Rare but can improve sample‑mean accuracy; most methods assume positive dependence. Temporal MAUP (Modifiable Temporal Unit Problem) – Aggregating over time can bias results similarly to spatial aggregation; must choose appropriate temporal windows. Non‑point data – Most statistical techniques assume point representations; lines, areas, or volumes often need conversion to points or homogeneous elements first. --- 📍 When to Use Which Assessing overall clustering → Global Moran’s I or Getis‑Ord G. Finding local hot‑spots → Local Moran’s I or Anselin Local Moran’s I. Predicting values at unsampled sites → Kriging (if semivariogram can be modeled) or IDW (quick, no covariance model). Modeling flows between zones → Gravity model (top‑down, aggregate). Exploring emergent land‑use patterns → Cellular automata (grid‑based) or ABM (agent‑driven) depending on need for mobility and purpose. Dealing with MAUP/UGCoP concerns → Use finer, more homogeneous zones; incorporate ancillary data; avoid overly aggregated statistics. --- 👀 Patterns to Recognize Clustered residuals after a regression → likely omitted spatial autocorrelation → switch to spatial regression. Sharp changes in rates when crossing administrative boundaries → possible boundary problem or MAUP. Consistent hot‑spot locations across multiple local autocorrelation maps → robust spatial association. Fractal‑like landscape metrics that stay constant across scales → indicates scale‑invariant patterns. --- 🗂️ Exam Traps Choosing the “right” distance metric – The outline lists Euclidean, Manhattan, connectivity, direction, cost‑distance; exam may tempt you to pick Euclidean by default. Remember the context (e.g., road networks → cost‑distance). Confusing MAMA (Modifiable Areal Unit Problem) with UGCoP – Both involve aggregation bias, but UGCoP stresses movement across zones; watch for wording about “context” vs. “unit shape/scale.” Assuming Kriging always outperforms IDW – Kriging requires a valid semivariogram; if none can be fit, IDW may be the appropriate choice. Treating a global Moran’s I value as evidence of local hot‑spots – Global statistic tells you overall clustering, not where it occurs. Interpreting a significant Moran’s I as proof of causation – It only indicates spatial pattern, not why it exists. ---