RDA is not generally used for predicting fitness variables using a species dataset as a predictor.

I would recommend trying other techniques, such as Canonical Correlation Analysis (CCA) or Partial Least Squares Regression (PLSR). Either would help you identify relationships between species and fitness variables and could make the necessary predictions.

In CCA, you can use a species dataset as one set of variables and a fitness dataset as another set of variables. CCA can identify linear combinations of these two sets of variables that are maximally correlated with each other. The resulting canonical variates can then be used to predict the values of the fitness variables based on the species variables.

In PLSR, you can use a species dataset to predict the values of fitness variables. PLSR identifies a set of latent variables that capture the common variation in both the species and fitness datasets. These latent variables can then be used to predict the values of the fitness variables based on the species variables.

I think either of these approaches would work better than RDA. If you have further questions, please feel free to reach out.

Thanks Noah! Unfortunatelly I cannot use CCA as one of my fitness variables has negative values for does not agree well with the chi-square distribution prerequisite. However, even if I were able to use it, CCA makes the same assumption as RA for the soecies being the y dataset, no?

Sofie

Sofie Spatharis It actually doesn't make the same assumption as RDA. In CCA, both species and the fitness variables are considered multivariate responses and the goal is finding linear combinations of both variable sets that are maximally correlated with each other. This is going to result in canonical variates, which are then used to understand the relationship between the two variable sets.

In comparison, in RDA, the species dataset is the response variable (y) and the fitness variables are considered explanatory variables (x). Thus, the goal is identifying the set of explanatory variables that best explain the variation in the response variable (y).

Neither RDA or CCA assumes that the species dataset must be the response variable, though it is fairly likely that this would be the case.

Let me know if this needs further clarification.

I think this is somehow troubling because RDA assumes a multivariate set of species abundances, that is, all response variables must be measured with same unit, which is not the case for environmental variables. Have you considered the OMI analysis? https://doi.org/10.2307/177351

I am not even sure if it fulfil your needs but it does look into species niche breadth which should somehow describe their fitness.

How about using SDMs to predict environmental variables?

Article Stream diatom assemblages as predictors of climate