Choose your neighborhood wisely: implications of subsampling and autocorrelation structure in simultaneous autoregression models for landscape ecology

Context

Large datasets that exhibit residual spatial autocorrelation are common in landscape ecology, introducing issues with model inference. Computationally intensive statistical techniques such as simultaneous autoregression (SAR) are used to provide credible inference, yet landscape studies make choices about autocorrelation structure and data reduction techniques without adequate understanding of the consequences for model estimation and inference.

Objectives

Our goal is to understand the effects of misspecification of neighborhood size, subsampling, and data partitioning on SAR estimation and inference.

Methods

We use remotely sensed burn severity for a large wildfire in north-central Washington State as a case study. First we estimate SAR for remotely sensed burn severity data at multiple subsampling intensities, data partitions, and neighborhood distances. Second, we simulate landscape burn severity data with SAR errors and calculate type I error rates for SAR estimated at the simulation neighborhood distance, and at misspecified neighborhood distances.

Results

Subsampling and misspecification of the neighborhood result in spurious inference and modified coefficient estimates. Type I error rates are close to the specified α-level when the model is estimated at both the simulation neighborhood and the distance that minimizes AIC.

Conclusions

By evaluating the effectiveness of pre-burn fuel reduction treatments on subsequent wildfire burn severity, we demonstrate that misspecification of the neighborhood distance and subsampling the data compromises inference and estimation. Using AIC to choose the neighborhood distance provides type I error rates near the stated α-level in simulated data.