‘Location Matters’: Incorporating Spatial Autocorrelation Into Modelling the Drivers of Cyanobacterial Biomass in Reservoirs of Northeastern Brazil

ABSTRACT

Spatial autocorrelation (SAC) is a critical yet often overlooked component of ecological analyses, which can significantly influence the accuracy and interpretation of statistical models. Therefore, this study aimed to evaluate the spatial distribution of cyanobacteria biomass among tropical reservoirs at the watershed scale and to identify the main environmental predictors associated with this distribution, explicitly incorporating SAC into regression models. Specifically, we investigated whether the inclusion of spatial structure improves model fit, modifies regression coefficient estimates and alters associated statistical inferences. The analysis was conducted using data from 32 reservoirs in northeastern Brazil, using nonspatial (ordinary least squares—OLS) and spatial (spatial error model—SEM; spatial lag model—SLM) regression models. SAC was assessed using Moran's Index and Local Indicators of Spatial Association (LISA). The results showed a significant positive SAC for total phytoplankton biomass and total and relative cyanobacterial biomass, especially at the watershed scale, although spatial effects were not significant across all basins. ‘High‐High’ and ‘Low‐Low’ clustering patterns were observed, indicating that geographically close reservoirs share similar environmental and anthropogenic conditions that either favour or inhibit cyanobacterial growth. Explanatory variables also showed positive or negative SAC, with patterns depending on the variable and spatial scale analysed. Furthermore, the manifestation of SAC was species‐specific, highlighting that spatial patterns are not generalizable and depend on the ecological and biogeographic context of each species. Predictor variables included pH, electrical conductivity, water temperature, dissolved oxygen, total phosphorus, nitrate, turbidity urban areas and agropastoral. Spatial models often yielded lower BIC values and higher R ² values than the nonspatial model. However, their advantage depended on the response variable and spatial scale, as ΔBIC values ≤ 2 indicated similar fit in several comparisons. They also altered the magnitude of the coefficients and, in some cases, reversed the direction of the predictor effects and increased the number of significant variables. These results reinforce the importance of incorporating SAC into ecological models to improve predictive accuracy and guide management strategies for cyanobacterial blooms in tropical reservoirs.