Tam et al. analyse a set of 16 CMIP models to study di7erences in marine low-cloud radiative feedback as well as the pattern e7ect contribution. To this end they analyse three di7erent experiments, AMIP, historical and 4xCO2. From their analysis they find the pattern e7ect to be mostly driven by Southern Hemispheric EIS changes and changes between the models to be due to di7erences in cloud sensitivity rather than di7erences in cloud controlling factor changes.

I think this is a very well done analysis tackling an important issue. However, at times I found some arguments and conclusions hard to follow. Extending some of the discussions of the results would help explain the authors reasoning. Also, a deeper discussion of previous results seems necessary to place the current results into context. While these points are important to address, I have no fundamental objections to the analysis and can therefore recommend publication after minor revisions.

Introduction

Could do with a gentler start (see below my comments under “Pattern e7ect”). Also the jump to the last paragraph is very abrupt. You didn’t really map out why what you are trying to do is important and novel. Readers that are less familiar with the field will appreciate this.

I think the introduction would also benefit from expanding the discussion on what exactly this manuscript sets out to do. This is currently only covered in one sentence in the last paragraph, but seems to be covered in more detail in L. 114-117.

Pattern effect

I think the manuscript would benefit from a deeper introduction to what the pattern e7ect is (possibly in or before the current first paragraph), to help readers that are slightly less familiar with the topic.

This extends to the discussion about quantifying the pattern e7ect. You mention in the abstract and in the conclusions (L.252) that “The pattern e7ect is defined as the di7erence in feedbacks between transient and long-term simulations.”. However nowhere in the introduction is this definition mentioned or discussed. A more in depth discussion on this definition and explanation of the use should be given in the introduction or method section.

Building on the previous paragraph, in 3.1.1/ 3.1.2 you argue that the change from AMIP- >historical->4xCO2_fast->4xCO2 in the model mean as well as in individual models is a measure of the pattern e7ect. I think this needs more discussion/justification as it is at the core of your analysis of the pattern e7ect. I agree with your analysis, given that the sensitivities used for the calculations are the same throughout the experiments, the changes have to come from changes in the CCFs pattern. However, this is quite some leap for readers to make, especially for those less familiar with the field. I think your reasoning should be discussed explicitly, why the di7erences in the experiments are an indication of the pattern e7ect, maybe even refer to eq. 1 for quantitative explanation.

L. 147 In line with all the above comments, this sentence is very vague.

Previous literature

The authors should incorporate more discussion on previous literature. For example, one of the main points, that the inter-model spread in feedback is mostly caused by di7erences in cloud sensitivity rather than di7erences in CCF changes is a somewhat know result (e.g. Klein et al. 2017). The same is true to an extent for the EIS changes. The authors analysis is very thorough and has (to my knowledge) not been done in this way before, but there are several related studies and it should be clarified what findings are agreeing (or disagreeing) with previous results and what findings are novel.

As a side note, given that the authors have done all this work, an extension of the analysis to provide insight into what drives the di7erences across experiments in the models seems an interesting, novel and important question. I think this would be a not too time consuming additional analysis, but I leave it up to the authors if they want to pursue this idea.

Minor

L. 12-14: Less negative over time under which conditions? 4xCO2, historical, 1%CO2 increase. All of those?

L. 21: This needs a source in my opinion, also as far as I am aware not all models show this behaviour.

L.28: Again a source seems needed here

L. 67 In my mind, sensitivities are partial derivatives, since it tells me how x changes with y, all else being equal. Personally I wouldn’t use the word here. However, I don’t think there is a fixed definition, so this is just a suggestion.

L. 77 Are all these studies on these six specifically?

L. 81 While I think I know what the authors mean by “local large scale environment”, I find this expression a bit self-contradictory

L. 111 I wouldn’t call epsilon an error term. It doesn’t come from measurement errors or similar things, but from inherent covariances in the system, as the authors point out themselves.

L. 114-117 This information seems to belong (at least in parts) into the introduction rather than the method section

L. 117 Why is there a citation here? Aren’t those the goals you are stating?

L. 168 I assume ‘pattern’ refers here to the relative importance of the terms of Fig. 1b compared to the relative importance in the subfigures of Fig. A1? In that case I would use a di7erent word, as it might be misleading.

L. 173 as in L.168

L. 179 or a covariance term (epsilon). Was this checked for in strength in this analysis, similar to what was done for the pattern (Fig. 4).

L. 178-184 I would have expected you to do the decomposition according to eq.3, but you replaced the term using model average sensitivity with observed sensitivity in figure b. I understand that you probably wanted to “recycle” figure 1b and not create an almost identical figure again, but I still found it initially confusing. Also, it makes the 1b and 1c not apples to apples comparable, since the spread will be larger/smaller if the sensitivities are larger/smaller. I fully acknowledge that the result seems pretty robust in terms of where the spread is coming from, but it still initially confused me.

L. 191 Following up from the above comment (on L. 178-184), I see you address this here by pointing to A1d (btw, the d is missing). Why not cite this in the paragraph above and adjust the discussion accordingly to avoid confusion?

L. 221 I would put ‘in historical and 4xCO2’ instead of ‘across experiments’. I stumbled upon this part because I stopped reading and looked at the maps and found there to be some significant di7erence in AMIP to the other experiments (Fig. 3a compared to 3b and 3c). That threw me o7 until I continued reading to see your caveat in the next sentence.

L. 236.-238. I don’t follow how you arrive at this conclusion from the before presented results. Against what is the historical estimate biased high? Against the AMIP? And why is it biased high? Also, why is the pattern e7ect biased low? I think this part needs significant expansion to explain your reasoning.

L. 239 Maybe add a transition sentence here to help the reader understand what you now intend to have a look at.

Technical

L. 51 refer to

L. 62 R-> R_low (twice)

L. 63 being a bit picky, but i is the index, the individual CCF would be CCF_i L. 69 CCF_i

L. 89 calculating -> calculated

L. 120 would cut sub-component

L. 207 to->do

Figures

Fig.1: exp-> Exp.

Fig.1: Maybe put a dot between the two terms in the titles and remove the parenthesis. Also, you could consider shortening model->m and CERES-> C to make things more readable

Fig.1: Personally, I would remove the bold font in the caption.

Fig.2/3: The maps are relatively small and the label for the zonal mean plots almost not readable in the printed version. For the digital version this is of course not a problem. The authors could consider rearranging the plots to make them bigger.

Fig. 4: I would suggest to call the last column epsilon, to stay consistent with eq. 3. Fig. 4: Again, personally I would remove the bold font in the caption.