Multi‐way prediction in the presence of uncalibrated interferents

The second‐order advantage states that predictions are possible for new samples even if they contain new interferents not taken into account in the calibration model. Traditionally generalised rank annihilation has been widely used for second‐order calibration, but alternative models and algorithms such as PARAFAC is known to often be more accurate when many samples are available. While the calibration step of second‐order calibration has received considerable attention in the literature, the procedure of how to predict has not been investigated in detail when PARAFAC is used. In this paper we test the second‐order advantage using different calibration approaches based on PARAFAC with both simulated and real datasets, focussing on prediction quality as a function of the size of the calibration set, the number and degree of overlap of new interferents and the type and magnitude of noise. Guidelines are given on how to implement predictions in PARAFAC‐based second‐order calibration.