Spatial analysis - Analytical Methods and Geostatistics

Understanding Spatial Analysis Methods Introduction to Spatial Analysis Spatial analysis is the study of phenomena across geographic space. The methods in this section address different questions: Do nearby areas tend to be similar? How do people and goods move between locations? Can we estimate values at unobserved locations? These questions are fundamental to geography, urban planning, epidemiology, and many other fields. Spatial analysis methods can be organized along two dimensions: Individual vs. Aggregate: Some methods analyze data at individual locations or unit level, while others work with aggregate patterns across regions. Exploratory vs. Explanatory: Some methods describe spatial patterns, while others model relationships between variables. As shown in the figure above, individual methods examine spatial patterns at specific locations (like three address points showing illness events), while aggregate methods work with broader regional patterns (like comparing illness rates across larger zones). Spatial Data Analysis: Factor Analysis Factor analysis is a dimensionality reduction technique that simplifies complex spatial datasets. In urban geography and social analysis, researchers often collect dozens of census variables describing neighborhoods: income, education, employment type, housing costs, family structure, and more. These variables are typically correlated—for example, areas with higher education tend to have higher incomes. Factor analysis identifies the underlying patterns in these correlated variables and reduces them to a smaller number of independent factors (also called principal components). Each factor represents a dimension along which areas vary. The Dominant Socioeconomic Factor In spatial social analysis, the first and most dominant factor almost always represents socioeconomic status. This single factor captures the primary distinction between wealthy and poor neighborhoods—it accounts for more variance than any other factor. A second factor often represents family structure or racial composition, depending on the variables included. The key insight is that instead of trying to understand 50 different census variables, you can understand neighborhoods by looking at their position along a few key factors. This makes patterns easier to see and interpret. Choice of Distance Metric The method you use to measure "distance" between observations in variable space affects which factors emerge. Three common choices are: Euclidean distance: The standard Cartesian distance, treating all variables equally Chi-square distance: Weights variables by their variance, useful for frequency data Mahalanobis distance: Accounts for correlations between variables, adjusting for non-spherical data distributions Different metrics may emphasize different patterns in your data, so the choice matters for interpretation. Spatial Autocorrelation Statistics Spatial autocorrelation measures whether nearby locations are more similar to each other than we would expect by chance. If neighborhoods with high poverty tend to cluster together, that indicates positive spatial autocorrelation. If high-poverty and low-poverty neighborhoods alternate in a checkerboard pattern, that indicates negative spatial autocorrelation. Global Measures Global spatial autocorrelation statistics describe the overall pattern across your entire study area: Moran's I: Tests whether values at nearby locations are more similar than expected. It ranges from -1 (perfect negative autocorrelation) to +1 (perfect positive autocorrelation), with 0 indicating no spatial autocorrelation. Geary's C: Similar to Moran's I but emphasizes differences between neighboring observations. It ranges from 0 (positive autocorrelation) to 2+ (negative autocorrelation). Getis-Ord G statistic: Identifies global clustering of high or low values without requiring assumptions about variance. These measures help you determine whether spatial patterns exist at all, which is a prerequisite for other spatial analyses. If spatial autocorrelation is zero, your data lacks spatial structure worth analyzing further. Local Measures Local spatial autocorrelation statistics evaluate whether autocorrelation exists at individual locations or neighborhoods: Local Moran's I: Shows which specific neighborhoods are significant hot spots (surrounded by similar high values) or cold spots (surrounded by similar low values), and which neighborhoods are outliers. Local Geary's C: Identifies local pockets of dissimilarity or similarity. These local measures are particularly useful for spatial epidemiology (finding disease clusters) and urban analysis (identifying neighborhoods that are unusual given their surroundings). Spatial Interaction: Gravity Models Spatial interaction refers to flows of people, goods, services, or information between locations. Examples include commuter flows between residential and employment areas, migration patterns, or trade between cities. Gravity models estimate these flows based on the principle that interaction between two locations depends on their "size" and their distance apart. Basic Components A gravity model has three essential components: Origin variables: Factors at the sending location that generate flows. For commuting, this might be the resident population or number of workers. For migration, it's the population at risk. Destination variables: Factors at the receiving location that attract flows. For commuting, this might be office space, number of jobs, or employment growth. For migration, it might be job availability or climate amenities. Distance or friction variable: A measure of the impedance to travel between locations. This could be Euclidean distance, travel time, or actual transportation costs. Closer destinations attract more flow than distant ones. The basic model has this form: $$T{ij} = k \frac{Oi Dj}{d{ij}^b}$$ Where $T{ij}$ is the flow from origin $i$ to destination $j$, $Oi$ is the size of the origin, $Dj$ is the attractiveness of the destination, $d{ij}$ is the distance between them, and $b$ is a parameter controlling how strongly distance decays the interaction (typically estimated between 1 and 3). Why Distance Matters Notice that distance appears in the denominator: as distance increases, flows decrease. This captures the distance decay effect—people prefer to travel short distances. The parameter $b$ controls how steep this decay is. If $b = 2$, distance has a very strong effect (doubling distance reduces flows to one-fourth). If $b = 1$, distance has a gentler effect. Competing Destinations Real spatial behavior is more complex than this simple model. Competing-destination effects capture the idea that your choice of destination depends not just on that destination's attractiveness but on alternatives. For example, when deciding where to shop, you don't just consider how far the nearest supermarket is—you consider all nearby supermarkets. Accounting for competition produces better predictions than simple distance decay alone. The space-time prism visualization above illustrates another refinement: the time dimension. In modern gravity models, accessibility depends on how far you can travel in available time, not just physical distance. Spatial Interpolation: Inverse Distance Weighting Spatial interpolation estimates values at unobserved locations based on observations at nearby locations. You have measurements at some locations but need estimates everywhere. Inverse distance weighting (IDW) is a simple interpolation method: the estimated value at an unmeasured location is a weighted average of nearby observations, where weights decrease with distance. A nearby observation contributes more to the estimate than a distant one. This is intuitive—you trust nearby data more than distant data. The estimate at location $p$ is: $$\hat{z}p = \frac{\sum{i=1}^{n} \frac{zi}{di^a}}{\sum{i=1}^{n} \frac{1}{di^a}}$$ Where $zi$ is the observed value at location $i$, $di$ is the distance to location $i$, and $a$ is a power parameter (typically $a = 2$). Higher values of $a$ make the method more local—distant points have almost no influence. IDW assumes that nearby values are more similar to the target location than distant values, which is often reasonable for geographic phenomena but isn't always true. Other interpolation methods like kriging make more sophisticated assumptions about spatial patterns. Spatial Regression: Geographically Weighted Regression Standard regression (OLS) assumes that the relationship between variables is the same everywhere. You fit one line to all your data. But in geography, relationships often vary by location. For example, the relationship between education and income might be strong in some regions but weak in others. Geographically weighted regression (GWR) relaxes this assumption. Instead of fitting one global regression model, GWR fits local regression models at each location, using observations from nearby areas with greater weight. The key insight is that parameter estimates become locally varying: different locations have different regression coefficients. A variable might have a strong effect in the northern region but weak effect in the southern region. This produces a much richer understanding of spatial relationships than assuming uniform effects. GWR is particularly valuable when you suspect spatial non-stationarity—that is, when geographic processes operate differently in different areas. This is very common in real-world data. <extrainfo> Simulation and Modeling Beyond analyzing existing data, spatial analysis can generate patterns through simulation. Cellular Automata Cellular automata use simple rules on a fixed grid to generate complex spatial patterns over time. Each cell has a state (on/off, occupied/vacant, disease-infected/susceptible), and rules determine how cells transition based on their current state and their neighbors' states. Despite simple local rules, the global pattern that emerges can be remarkably complex. This bottom-up approach models how local interactions produce large-scale spatial structures—useful for understanding urban sprawl, ecosystem dynamics, or disease spread. Agent-Based Modeling Agent-based modeling (ABM) represents individuals or entities explicitly, each with goals and decision-making rules. Agents interact with their environment and other agents, and the researcher observes what large-scale spatial patterns emerge from these interactions. ABM offers great flexibility because it can represent heterogeneity (different agents behave differently) and complex adaptive behavior. Comparing Perspectives Gravity models are top-down: they specify aggregate flows directly. Cellular automata and agent-based models are bottom-up: they specify individual behavior and let aggregate patterns emerge. Each perspective offers value depending on your research question and available data. </extrainfo> <extrainfo> Multiple-Point Geostatistics Multiple-point geostatistics is a specialized spatial analysis technique used primarily in geology and resource estimation. Purpose and Method The fundamental goal is to analyze spatial patterns in a training image—a conceptual model of the geological phenomenon being studied. The method extracts multiple-point statistics (patterns involving three or more points) from this training image, which are more flexible than traditional two-point statistics (like the variogram used in kriging). The algorithm then uses these extracted patterns to generate multiple realizations—alternative possible maps of the phenomenon that are consistent with the observed data and the patterns shown in the training image. Each realization represents one plausible outcome given the constraints. Uncertainty Quantification A single realization is simply one possibility. The power of the method emerges when you generate many realizations. Together, they map out the space of uncertainty—the range of plausible scenarios. By computing statistics across all realizations, you can quantify spatial uncertainty, answer "what if" questions about uncertain scenarios, and communicate the limits of your knowledge. This is valuable in resource estimation and environmental prediction where decisions depend on uncertain spatial patterns. </extrainfo>