Dear Authors,

Two reviewers have now submitted their evaluation of the paper. Both saw great potential and were generally positive about the manuscript. However, both reviewers raise excellent points where the paper could be improved. I suggest to rework the paper using these constructive comments, along the lines of your response. I especially would encourage clarifying some parts of the methods section as referee 2 points out.

However, in addition to the reviewer comments, I have some additional thoughts/concerns that deserve the authors' attention while reworking the manuscript. Based on referee 1's call for an in-depth check of the paper's statistics, I have checked the methods section of the paper myself...

- I nowhere read that the model assumptions (residual normality and homoscedasticity to start with) were met.

- What approximation method was used to estimate the mixed effects models P-values? Kenward Rogers? Satterthwaite?

- The R2 table at the end (Table B8) is currently confusing. It reports only 4 model fits, where there is little reference to these in other parts of the paper. Inline you state that land cover (L413)/ clay+fine silt (L404) explain 10% of the variation (conditional R2), but the table currently says 0.01 for both.

- A major thing to consider: fixed effects selection in regression analyses is debated, and it has been shown to lead to spurious claims on the data, depending on what method you use for the selection. It is simply not the best practice, although I acknowledge that it is commonly used. I suggest you address this by either explaining more in-depth why you did a predictor selection this way, or use a more robust method (splitting up data in model building - model validation, or using shrinkage via a penalizing function, or others). In this specific case, I think it is actually very important: as the authors note themselves in the manuscript, but also as both referees noted: the land cover not being retained in the final model will lead the readership to conclude that land cover is not important in predicting SOC%. This is a potentially important finding and one of the main conclusions of the manuscript. I think that you could consider stressing that this conclusion is based on a regional-scale analysis where some other predictors simply range in about two orders of magnitude (for some predictors). Secondly, you fully rely on the variable selection procedure for that claim. Your full model seems to show that forest as a land use significantly increases SOC by 0.26 units across the dataset (Table B2), which is about half of what is predicted for a unit increase in Caex (0.54). For Caex, as a continuous standardized variable, 1 unit increase means increasing CAex by 1 SD (derived from the Caex distribution from the whole dataset), which is about 11 cmol kg-1 (while the mean is 10 cmol kg-1!). This is almost comparing two different things. To me, that forest effect size would even seem interesting enough to run parallel models for the different land-use classes, much like you do for the pH and MAP classes. Add to that that land use is a categorical predictor (the only one, between all your continuous predictors), and that you lose more degrees of freedom for the inclusion of land use as a categorical predictor in the model, compared to any continuous predictor. Obviously, this would be penalized by the AIC. If you would have used direct likelihood testing (using ML, and not REML estimation) of your models, land use might even have been retained in the model, and than the message would have been completely different? Bottom line: I think variable selection in mixed effects is complicated, add to that that you mix one categorical with multiple continuous predictors. I'm not necessarily saying that you need to change the methods, but these things need to be better founded, evaluated and discussed in the paper. You now 'let the data speak' through your analyses, which is great, but you need to be aware of the potential pitfalls of your choices as an analyst. In this case, this could even affect the main message you give to the readership.

- in line 163 you state that the slope of the regression varies across the sites, but if I understood your methods section well, than you fitted random intercept-only models? Needs clarification

Dear authors,

One of the original referees has reviewed your new manuscript version. He is happy with all the changes, and has only one minor comment left. Because of my (perhaps unusual) comments on your work, I have also re-evaluated the manuscript. If you address just that minor comment from the reviewer, this manuscript is ready for publication.
Congratulations on this work, and on the clear and qualitative rebuttal letter. I'm convinced this will be a good contribution to literature!

Dear authors, following the recommendation of the Guest Editor after the review process, I am glad to communicate that your manuscript is accepted for publication in SOIL.

Thanks for consider SOIL as the journal where submit your interesting research.

Sincerely

Jorge