This paper presents a MLR statistical attribution of the 1985-2018 PM2.5 trends in three megacity clusters in China, using visibility data as proxy for pre-2013 PM2.5 data. It finds a large meteorological (non-emission) contribution to the trend, and argues that previous MLR analyses of the 2013-2018 trend using the actual PM2.5 data starting in 2013 and attributing the trend to emissions are not robust. The paper makes some good points about the difficulty of sorting out meteorological effects when interpreting short (post-2013) trends. However, I believe that it may be (1) flawed in its reconstruction of the 1985-2018 PM2.5 record which is the basis for most of the argumentation, (2) mistaken in claiming that attribution of recent PM2.5 trends to emissions is not based on mechanistic knowledge, and (3) annoying in belaboring trivial statistical points that are well known to any trained scientist.  I don’t think that this paper is publishable in ACP in current form. Specific comments are below.

1. The ‘bulge-2013’ feature in Figure 1 (line 88) anchors much of the argumentation in the paper but it is very weird. It seems caused by the switch from the visibility proxy to the actual PM2.5 data in 2013. The methods are buried in Supplementary Material. Is this ‘bulge-2013’ seen in the consistent long-term satellite AOD data record? I think that the authors would have to show that it is present in the AOD data in order to have credibility.
2. What is the ‘emission’ in Figure 1? Of what species?
3. Line 91: the Mao et al. 2019 reference which is intended to provide support for the authors’ meteorological attribution of the trend is in fact grey literature involving some of the same authors.
4. Line 132, etc.: the mechanistic meteorological connection of ASI to PM2.5 is not clear, and as the authors point out any meteorological variable with a suitable long-term trend would do the trick. But there is in fact a strong mechanistic argument for emissions to be related to PM2.5 (line 154), and there is strong independent evidence that Chinese emissions have decreased over the 2013-2018 period (emission inventories, satellite data). To claim that the connection of PM2.5 to emissions has no mechanistic support strikes me as obviously wrong. In fact the authors cite Chen et al. 2019 in demonstrating the mechanistic connection in WRF-CMAQ but argue that the analysis is flawed because it did not consider the effect of the bulge-2013 (line 158). As pointed out above, I am very suspicious of this bulge-2013.
5. There is a lot of trivial stuff about the non-mechanistic basis of statistical models, correlation not implying causality, more years increasing the credibility of the model, etc., that is repeated again and again and does not rise above the level of a basic course in statistics.

This is a very interesting paper analyzing the causes of PM2.5 trends observed in eastern China. The research topic is highly important from air pollution control point of view and, although this topic has been studied quite intensively during the recent years, this paper manages to provide new insight into it. The paper is clearly organized and relatively well written. I could not find any scientific errors, even though I do feel being an expert on trend analysis. I have a few minor issues to be considered before accepting this paper for publication:

Please explain how ASI is defined. Based on Figure 1 it seems to be dimensionless variable but this has not been explained anywhere. This is particularly important because related to Arctic sea ice, most often the concept “Arctic sea ice area” is used in a scientific literature.

Similarly, please explain explicitly what is meant by “emissions” appearing in Figures 1 to 5. Are they simply primary PM emissions taken from the emission inventory, or do they also include precursors that form secondary aerosol matter?

Describing what was done or observed, is usually written in a past tense. Please check out this throughout section 2. Past tense is also preferred in the following places: … were crucial (line 35), … was the most (line 37), … carried out (lines 111 and 128).

Section 3.2. Please reformulate the title of this section (e.g. “Comparing the MLR results…). Starting the section by referring to “the second question” is not a good practice, as the reader need to find out from the earlier text what is this question. I would recommend repeating this section e.g. by writing “The answer to the question whether Table 1 or 2 is correct is that neither of them is correct, for the following ….”

This work proposes a different method for the MLR analysis of PM2.5. Based on the new interpretation and the comparison with previous studies, the MLR results among different studies were found to be more consistent. In addition, the authors also pointed out that the relationship constrained by long-term data is more reliable. Overall, this is an interesting study and it provides some useful information for other researchers when choosing MLR for air quality trend analysis. However, more explanations, especially for the methodology, are still needed. Major revision is suggested and specific comments are listed as follows.  

(1) Line 60, the resolution of the PRD emission inventory is three degrees, which is rather coarse. 

(2) This work mainly focuses on the PRD, YRD and Jing-Jin-Ji regions. For YRD and Jing-Jin-Ji regions, the authors combined the MEIC and PKU emission inventories to do the analysis. While for PRD, they combined PKU and PRD-EI to do the scaling. MEIC and PRD-EI are different emission inventories and the methods that used to derive these two emission inventories should be not consistent. Based on the literatures, the MEIC emission inventory should have already covered the PRD region, why not also using the MEIC emission inventory to analyze the PRD region?  

(3) Please label the scaling factor and Ei equations. 

(4) Line 71, please use data or reference to support this assumption. 

(5) The PRD scaling factor was calculated by the emission sum from 2006 to 2013, while the scaling factors for the other two regions were calculated by the emission sum from 2010 to 2013. Please explain why using different emission sum to derive the scaling factors.  

(6) The authors applied the nonlinear exponential fitting to retrieve the long-term PM2.5 concentration before 2013, because China began to release the air quality observation data since 2013 and it is unlikely to acquire long-term observation data in this nation before 2013. However, based on the figures in the supplemental material, some of the fittings are not acceptable for further analysis, such as BTH-RH (40, 60) and YRD-RH (90, 100). The authors need to analyze and discuss whether such errors can influence their conclusion. 

(7) For the PM2.5 concentration retrieval, I suggest the authors use the data of 2014-2018 for the fitting and the 2013 data for the verification, this can help to verify whether the methods implemented by the authors are reliable or not. 

(8) Please introduce about the data source of RH and visibility in section 2.2. Generally the locations of the meteorological stations are not the same with those of the air quality stations. Did the authors use the nearest matching to pair the data? If so, what is the mean distance between the meteorological station and air quality station? 

(9) Line 120, combined with other studies and this work, we understand that the emission is the major factor that influences the PM2.5 trend when compared to the meteorological variables. However, in Chen et al. (2019), the meteorological factors can still account for 21% of the contribution, which is much larger than the values reported by the authors in Line 117. I do not think this is an ‘agreement’.  

(10) Lines 162-164, based on the analysis performed by the authors, if there exists any method that can compensate the shortcomings of the MLR and prognostic model?  

(11) Lines 180-185, whether this means that previous studies that focused the ASI harbor relatively large uncertainty? 

(12) Lines 166-169, I suggest the authors to provide some theoretical foundations to support this interpretation.