(s)PLS regression: negative Q2 but better for later components?

1

I am using both PLS and sPLS: sPLS always select 1 component using CV, although the Q2 score is negative. My aim is not really to build a predictive model (indeed, the RSS and PRESS are extremely large), but to understand the possible relationships between X and Y.
What I cannot understand, is that the Q2 score increases (but still negative) with later components (i.e., variates) when using PLS:

tmp
tmp2304×1344 44.4 KB

In the figure, the numbers in parenthesis represent the mean, min and max values. Shouldn’t the mean Q2 score decrease?

2

hi @lorenzoF92,

This is in fact a noticeable behaviour. Can you please send us an email with any data that could reproduce this? You can follow the step 5 at Reproducible example to clarify issues to do so.

Thanks

Al