In this paper the authors analyse a large dataset containing time series of plant abundance. They test how commonly used functional traits relate to population stability. I think this is an interesting and important question because we still have relatively little data on how functional traits relate to stability and understanding the drivers of population stability is important to understand ecosystem stability. The dataset is very impressive and extensive, allowing general conclusions to be drawn about the role of the traits. I think the analysis is generally solid, I just have one question about it, and I find the paper well written and very clear. I have two main comments on the discussion of the results.

The major issue that I see with the analysis is that while the traits do significantly affect population stability, they explain a very small amount of the variance (a pseudo R2 of 0.06-0.07). This is obviously a very heterogeneous dataset and so, to some extent, finding any general pattern is useful, but I do still feel the authors should comment on what might explain the majority of the variance in population stability. Differences between species and sites are clearly important as the random effects together with the fixed effects give a pseudo R2 of around 0.2. However, I do wonder what else might be important? Other unmeasured traits? Interactions between traits? Interactions between traits and the environment? Intraspecific trait variation? I don’t think the authors can test for all these possibilities, but I would appreciate some discussion of the issue and acknowledgement that the models cannot explain most of the variation in population stability.

Related to this, did you try to fit the categorical traits together with the continuous ones? It seems that some of the categorical traits have quite a large effect and whilst I fully agree that continuous traits are more useful, it might be interesting to check how much more variance you can explain if you combine the categorical and the continuous traits.

One of the most interesting results to me was that both LDMC and SLA had significant effects on population stability and leafN also has a tendency to affect stability as well. The authors nicely discuss why slow species (high LDMC, low SLA) might have more stable populations but why would both traits be significant, as both indicate resource use strategy? I think it would be interesting to have some discussion on what the separate effects of LDMC and SLA are, as it suggests that there is not just an overall difference in stability between fast and slow species but that different resource economics traits are linked to stability in different ways. Similarly, it would be interesting to discuss why it is LDMC that comes out as the strongest predictor? Why would LDMC in particular be linked to population stability, rather than, for instance, leafN?

I don’t understand why the N values for the full model and the reduced (final) model differ in Table 1? It almost looks as if you had more datapoints for the reduced model because you drop traits with a lot of missing values and therefore include more species. However, if this is the case then I can’t see how you could have done the model simplification because you can’t compare models fitted to different datasets. How would you be able to tell whether dropping traits makes the model significantly worse? If it is really true that you have more data for the reduced model then you should at least refit it to the same dataset as the full model, in order to be able to compare the two models, i.e. do the model simplification with the full dataset and then refit the reduced model to an extended dataset. The same applies any other time models are compared: in table S1 it also looks as if models with PCoA axes and single traits are fitted to different datasets.

In addition, I think it is very interesting that models with PCoA axes are worse, or at least no better, than those with single traits. However, it would be more solid to compare models using AIC rather than using only the marginal and conditional R2 values.

Minor comments
In Fig. S1 the codes for life form (in c) are not explained

This new version of the ms is much improved and I appreciate the thoughtful and thorough revision the authors have carried out. The analysis is much clearer and I found the new discussion section very interesting.

I only have one small comment on the analysis. I now understand what the authors did to compare models and I think it is OK. However, R2 should not generally be used to compare models because it will typically favour the most complex model as there is no penalty for adding more terms. I therefore think that it is important that the authors address this in the methods section by stating that they also did the AIC comparison of full and reduced models, fitted to the same dataset. This is mentioned in the response letter but should also go in the paper. Second it is important to state that only terms that were significant in the reduced model were considered.

On line 223, you could also mention that using R2 to compare models with the PCoA axes and the single traits is not problematic because the models have the same number of degrees of freedom.

Line 270-274: this sentence should be revised, the last part does not make sense. It is also very long, so could probably be split.