Portable NIR spectrometers are effective in detecting authentic pharmaceutical products in intact capsule formulations, which can be used to screen for substandard or falsified versions of those authentic products. However, the chemometric models are trained on libraries of authentic products, and are generally unreliable for detection of quality problems in products from outside their training set, even for products that are nominally the same active pharmaceutical ingredient and same dosage as products in the training set. As part of our research directed at developing better non-brand-specific strategies for pharmaceutical screening, we investigated the impact of capsule composition on NIR modeling. We found that capsule features like gelatin type, color, or thickness, give rise to a similar amount of variance in the NIR spectra as the type of API stored within the capsules. Our results highlight the efficacy of orthogonal projection to latent structures in mitigating the impacts of different types of capsules on the accuracy of NIR chemometric models for classification and regression analysis of lab-made samples. The models showed good performance for classification of field-collected doxycycline capsules as good or bad quality when an NIR-based % w/w metric was used, identifying five samples that were adulterated with talc. However, the % w/w was systematically underestimated, so when evaluating the capsules based on their absolute API content according to the monograph standard, the classification accuracy decreased from 100% to 70%. The underestimation was attributed to an unforeseen variability in the quantities and types of excipients present in the capsules.