I wrote a paper that focuses on this question (Anglim & Grant, 2014; pre-print is https://osf.io/g8kbj/download).
In short, if you're interested in estimating how well facets predict an outcome, then you should include all facets as predictors. Likewise, if you want to estimate the incremental prediction of facets over domains (e.g., 30 facets over the Big 5), then you should include all domains in the domain regression model and all facets in the facet regression model.
If you're doing research, then presumably, you don't know which facets predict the outcome. Therefore, it's an empirical question which do predict.
An important point is that you use an estimator that corrects for biases when including many predictors. A reasonable approach is to use adjusted r-squared, although when you have 30 facets, large samples help a lot.
I also have more recent primer that is designed to be fairly accessible (Anglim & O'Connor, 2018, see preprint https://psyarxiv.com/a78g2/download )
If you want more detail, post further questions in the comments.
- Anglim, J., & Grant, S. L. (2014). Incremental criterion prediction of personality facets over factors: Obtaining unbiased estimates and confidence intervals. Journal of Research in Personality, 53, 148-157. https://osf.io/g8kbj/download
- Anglim, J., & O'Connor, P. (2018). Measurement and Research Using the Big Five, HEXACO, and Narrow Traits: A Primer for Researchers and Practitioners. Australian Journal of Psychology. https://psyarxiv.com/a78g2/download