Spatial Statistics and Geostatistics for Actuaries

Types of Spatial Data

Spatial statistics analyzes data that has geographic coordinates attached to each observation. Three types of spatial data arise in actuarial work, each requiring different analytical tools. Point pattern data consists of the locations of discrete events, such as the geographic coordinates of reported auto claims or policyholder addresses. The primary question is whether the points are randomly distributed or show clustering or regularity.

Geostatistical data consists of continuous measurements taken at specific locations, such as property values, soil contamination levels, or average claim costs sampled across a region. The goal is typically to predict values at unobserved locations. Lattice (areal) data consists of measurements aggregated over predefined regions, such as county-level claim rates, state-level average premiums, or ZIP code loss ratios. The data is associated with areas rather than precise points.

Spatial Autocorrelation

Spatial autocorrelation refers to the tendency for observations at nearby locations to be more similar than observations at distant locations. This is sometimes summarized as Tobler's first law of geography: "Everything is related to everything else, but near things are more related than distant things." In insurance data, claim frequencies and severities often exhibit spatial autocorrelation because neighboring areas share similar demographics, traffic patterns, weather exposure, and economic conditions.

Moran's I is the most common test for spatial autocorrelation. It is computed as I = (n / S_0) * (sum over i,j of w_{ij} * (x_i - x_bar) * (x_j - x_bar)) / (sum of (x_i - x_bar)^2), where w_{ij} are spatial weights (often 1 for neighbors and 0 otherwise) and S_0 = sum of all w_{ij}. Moran's I ranges approximately from -1 (perfect dispersion) through 0 (randomness) to +1 (perfect clustering). Geary's C is an alternative that focuses on squared differences between neighbors: C = ((n-1) / (2*S_0)) * (sum of w_{ij} * (x_i - x_j)^2) / (sum of (x_i - x_bar)^2). Values of C less than 1 indicate positive spatial autocorrelation.

The Variogram

The variogram (or semivariogram) characterizes the spatial dependence structure of geostatistical data. It is defined as gamma(h) = 0.5 * E[(Z(s+h) - Z(s))^2], where Z(s) is the value at location s and h is a separation distance (lag). The empirical variogram is computed by grouping pairs of observations by their separation distance and computing the average squared difference within each distance bin.

A typical variogram has three key features. The nugget is the variogram value at distance zero (or just above zero), representing measurement error and micro-scale variability. The sill is the variogram value at which it levels off, representing the total variance. The range is the distance at which the variogram reaches the sill, beyond which observations are essentially uncorrelated. Fitting a parametric model (such as exponential, spherical, or Gaussian) to the empirical variogram enables prediction through kriging.

Kriging

Kriging is the optimal interpolation method for geostatistical data, using the fitted variogram to predict values at unobserved locations. The kriging predictor at a new location s_0 is a weighted average of observed values: Z_hat(s_0) = sum of lambda_i * Z(s_i), where the weights lambda_i are chosen to minimize the prediction variance. The weights depend on the spatial configuration of the observed locations relative to s_0 and the variogram model.

Ordinary kriging assumes a constant but unknown mean. Universal kriging allows the mean to vary as a function of spatial coordinates or other covariates. A key advantage of kriging over simpler interpolation methods (like inverse distance weighting) is that it provides not only the predicted value but also the prediction variance, quantifying the uncertainty at each location. Locations far from observed data points have larger prediction variances.

Spatial Regression Models

Standard regression models assume independent observations. When data is spatially autocorrelated, this assumption is violated, leading to underestimated standard errors and potentially misleading inference. Spatial regression models address this by explicitly modeling the spatial dependence. The spatial lag model includes a spatially lagged dependent variable: Y = rho * W * Y + X * beta + epsilon, where W is a spatial weights matrix and rho measures the strength of spatial dependence.

The spatial error model instead assumes spatially correlated errors: Y = X * beta + u, where u = lambda * W * u + epsilon. This model is appropriate when the spatial pattern is in the residuals rather than in the response itself. Choosing between these models depends on the hypothesized mechanism: does the outcome in one area directly affect outcomes in neighboring areas (lag model), or are there unobserved spatially varying factors creating correlation in the errors (error model)?

Actuarial Applications

Territory rating is perhaps the most direct actuarial application of spatial statistics. Insurers divide their market into rating territories and assign a rate relativity to each. Spatial smoothing techniques (including kriging and spatial regression) help identify geographic patterns in loss experience and create more stable territory rate factors by borrowing strength from neighboring areas with similar experience.

Catastrophe risk assessment uses spatial models to estimate the geographic footprint and intensity of natural disasters such as hurricanes, earthquakes, and floods. These models combine historical event data with physical models to simulate potential future events and their impact on insured portfolios. The spatial correlation of losses within a catastrophe event is critical for estimating aggregate losses and determining appropriate reinsurance structures.

Common Pitfalls

A frequent mistake is ignoring spatial autocorrelation in standard regression and interpreting artificially small p-values as evidence of significant effects. Always test for spatial autocorrelation in residuals (using Moran's I) before trusting regression results with spatial data. Another error is confusing the variogram with the covariance function; they are related by gamma(h) = C(0) - C(h), where C(h) is the covariance function.

When interpreting variograms, students sometimes confuse the nugget (measurement error plus micro-scale variation) with the sill (total variance). A large nugget relative to the sill indicates that much of the variation is non-spatial, reducing the effectiveness of kriging predictions.

Study Tips

Exam MAS-I covers basic spatial concepts, variogram interpretation, and the rationale for spatial methods in actuarial applications. Focus on understanding the three types of spatial data and which methods apply to each. Be able to interpret an empirical variogram (identifying nugget, sill, and range) and explain how kriging uses this information. Know why spatial autocorrelation invalidates standard regression assumptions and how spatial models address the problem. Practice computing Moran's I for small lattice data sets.