This study investigates the influence of spectral measurement errors on the accuracy and reliability of Near-Infrared (NIR) spectroscopy in predicting cannabinoid content, specifically examining the variability across multiple NIR instruments of the same model and virtual instruments. Through a detailed case study using NeoSpectra miniaturised spectrometers, we explore the sources and structures of measurement errors, their covariance and correlation patterns, the implications on preprocessing, and subsequent model performance. This study also introduces the Integral Error Correlation Index (IECI), a novel metric designed to objectively quantify measurement error correlation, as meeting the independent and identically distributed (iid) error assumption is critical for Partial Least Squares (PLS) regression models. This metric is proposed for aiding in the systematic exploration of preprocessing methods through their impact on error correlations, and their subsequent model performance. The results underscore that preprocessing methods yielding lower IECI values lead to simplified, more accurate PLS models, demonstrating the potential for improved prediction reliability. This research contributes to the optimisation of NIR spectroscopy in cannabinoid determination or other applications, offering a robust framework for managing measurement errors coming from different sources and refining multivariate predictive models in analytical methods.