Are spatial models advantageous for predicting county-level HIV epidemiology across the United States?

Predicting human immunodeficiency virus (HIV) epidemiology is vital for achieving public health milestones. Incorporating spatial dependence when data varies by region can often provide better prediction results, at the cost of computational efficiency. However, with the growing number of covariates available that capture the data variability, the benefit of a spatial model could be less crucial. We investigate this conjecture by considering both non-spatial and spatial models for county-level HIV prediction over the US. Due to many counties with zero HIV incidences, we utilize a two-part model, with one part estimating the probability of positive HIV rates and the other estimating HIV rates of counties not classified as zero. Based on our data, the compound of logistic regression and a generalized estimating equation outperforms the candidate models in making predictions. The results suggest that considering spatial correlation for our data is not necessarily advantageous when the purpose is making predictions.