This paper evaluates Probabilistic Linear Discriminant Analysis (PLDA) scoring for robust speaker verification using the ECAPA-TDNN as speaker embedding extractor in both automatic speaker verification (ASV) and tandem spoofing-robust ASV (SASV). We tested ECAPA-TDNN with cosine similarity scoring and various PLDA-based backends. We propose a joint optimization approach, in which the ECAPA-TDNN embedding extractor and a discriminative PLDA backend are fine-tuned end-to-end. On the ASVspoof2019 evaluation set, cosine similarity achieved the best minimum normalized a-DCF across the tested SASV systems; however, several PLDA-based configurations reached comparable performance levels. Furthermore, certain PLDA-based configurations slightly outperformed the cosine baseline in terms of the minimum normalized t-DCF across the tested SASV systems.