Pre-Modeling Diagnostics for Multi-Source Urban Sustainability Indicators: A Five-Stage Framework for Arid and Semi-Arid Cities

Multi-source urban indicator datasets are frequently used in clustering, machine learning, and regression analyses without prior statistical auditing, potentially distorting results through non-normality, multicollinearity, and cross-source measurement disagreement. This study applies an integrated five-stage pre-modeling diagnostic protocol, comprising coverage auditing, normality assessment, outlier detection, correlation analysis, and VIF screening, to a 71-variable dataset compiled from six international sources for 106 arid and semi-arid cities across nine world regions. The protocol evaluates coverage completeness, distributional normality, outlier structure, inter-variable correlation architecture, and multicollinearity severity. Results show that 89% of continuous variables are non-normal, 17 of 23 predictors exceed the conventional Variance Inflation Factor threshold of 10, and 2 major international sources exhibit near-zero agreement for electricity access (Pearson r = −0.02). Gulf Cooperation Council cities form a structurally coherent outlier cluster requiring robust scaling rather than exclusion. The study produces an eight-action priority matrix to guide downstream analytical decisions and provides a replicable protocol for multi-source urban sustainability datasets.