Bayesian panel modeling of spatiotemporal dynamics in intraseasonal meteorological drought persistence

Meteorological drought is one of the most complex and severe natural phenomena, with far-reaching impacts across nearly all aspects of life. Understanding its persistence and underlying factors is crucial for mitigating its adverse effects on vulnerable regions. Building on this, the research investigates the factors influencing intraseasonal drought persistence in Punjab. Before applying the Bayesian Panel Data Model for spatiotemporal analysis, spatial and temporal correlation tests were conducted to confirm the presence of significant dependencies. For spatial correlation, the dataset was first converted into a spatial object (sf format) with geographic coordinates. A spatial neighborhood structure was created using k-nearest neighbors (k = 5), followed by the construction of a spatial weights matrix (W) using row-standardized weighting. Moran’s, I test was then performed, revealing significant spatial clustering of drought conditions, with a positive Moran’s I value (0.0133), a high Z-score (3.2375), and a significant p-value (0.0006). These results were further validated using spatial autocorrelation maps and empirical variograms, highlighting strong spatial dependence in drought persistence. For temporal correlation, Binary Autocorrelation Function (BACF) and Binary Partial Autocorrelation Function (BPACF) plots of drought conditions were analyzed, confirming significant temporal dependencies. Given these findings, a Bayesian Panel Data Model incorporating both fixed and random effects was applied for a comprehensive spatiotemporal analysis. To classify drought conditions, SPI-3 was computed, and binary logistic regression and ridge regression were applied separately for each season to identify key determinants of drought persistence and derive informative priors for the Bayesian framework. Bayesian modeling was conducted using the RStan package in R, leveraging MCMC simulations to generate posterior samples and estimate model parameters. The findings revealed that logistic regression priors are more effective for modeling drought persistence in warmer months, while ridge regression priors perform better in colder months. Meteorological variables influence drought persistence differently across seasons. Precipitation and topsoil wetness consistently reduce drought persistence, while surface pressure, temperature, humidity, wind speed, and root soil wetness exhibit varying seasonal impacts. Spatially, drought persistence trends increase eastward during autumn-winter and northward during spring-summer. Since 2000, the frequency and intensity of drought persistence have risen across all seasons, with significant drought events recorded in 2000, 2004, 2014, and 2015. Specific months, including January, March, June, October, and November, were found to be particularly prone to drought persistence. The results demonstrate that the Bayesian mixed model approach, coupled with panel data analysis, is a powerful tool for fitting complex statistical models, particularly in contexts where incorporating spatiotemporal heterogeneity and prior information about covariates is essential. It also provides policymakers with data-driven insights for informed decision-making to aid in the design of more effective drought mitigation and climate adaptation strategies in Punjab, Pakistan.