General comment:

The paper addresses an analysis of reprocessed GPS/MET data using a single frequency retrieval technique which is compared to the dual frequency processing product of CDAAC 2007 and the MERRA-2 analysis. With this new and partly not overlapping dataset they can extend the CDAAC 2007dataset and hence can provide more (RO) measurement data for climate trend analysis and climate model validation in a time where globally even distributed measurement data is sparse.

The paper is structured in a good way, the title clearly reflects the contents of the paper, and the scientific methods and assumptions are valid and clearly outlined. The results are sufficient to support the interpretations and conclusions. Some minor questions, comments and typos I will provide in the specific comments.

Specific comments:

1. Section 2 page 4 line 121: I would change "... and underwent reprocessing in 2007." to "and underwent a first reprocessing in 2007." since there were more than this reprocessing in the last years.

2. Section 2.1, page 5 line 143/144: I would add the cite to Liu et al. 2020 https://doi.org/10.3390/rs12213637 ("New Higher-Order Correction of GNSS RO Bending Angles Accounting for Ionospheric Asymmetry: Evaluation of Performance and Added Value") on the bi-local higher-order RIE here and also in Section 4.1 page 15 line 466.

3. Section 2.1 page 6 last paragraph: Which orbit/orbit processing software did you use? are there any updates to the CDAAC 2007 processing? Could you add a short sentence to specify this in your paper?

4. Section 2.1 page 7 line 216ff: Could you please explain here a little bit more detailed what you mean by this sub-interval processing. What I think I understand from the following paragraphs and Figure 2 you're doing this on each "ray" for the 1 Hz pseudo-range data and then interpolating this to the 50 Hz CA measurements. But I'm not sure that I've understand that correctly.

5. Section 2.1 page 8 line 233: How do you extrapolate below 15 km? constant, linear, ...?

6. Section 2.1 page 8 line 254/255: The sentence "The altitude range for the fits in 15-60 km, or whatever the upper altitude of the occultation happens to be." is finishing without being finished ...

7. Section 2.1 page 8/9 line 258 - 260: I think these two sentences can be removed since in the next paragraph you're telling the same only in more detail.

8. Section 2.1 page 9 line 275: typo: "...AS in on." -> "... AS is on."

9. Section 3 page 11 line 331f: "Other climate-related work that ...": please add a cite which work you mean.

10. Section 3 page 12 360ff and Figure 6: To compare the CDAAC AS on AS off data with the new JPL data it would be good to separate the JPL data according to AS on and AS off too, although there is no difference in the processing in the JPL case.

11. Section 3 page 12 line 371: typo: remove the "." after 30 km.

12. Section 4 page 13 discussion on Figure 7 lines 406ff: The bias of the 12 km level is larger at higher (shown) latitudes (JPL dataset but also both CDAAC datasets). This could be a problem of the MERRA-2 analysis there. The approximate height of the mid-latitude tropopause should be there. Could these biases at this height level be related to the mid-latitude tropopause? You only mention the tropopause with respect to the lower latitudes and the tropical tropopause. Please discuss this. The 17 km level shows an opposite effect than the 15 km level for the JPL data. This could indicate that there is a possible mislocation of the tropopause.

13. Section 5 page 16 line 503 typo/auto-correction error: "...radio location ..." -> "radio occultation"

14. Figure 4, Figure 5, and Figure 6: no x unit.

15. Figure 7: caption: you could interchange AS-on and AS-off in the caption since then it corresponds to the panel order.

General comments:

The paper compares radio occultation GPS/MET data retrievals from UCAR/CDAAC dual-frequency processing in 2007 with a newly processed GPS/MET dataset from JPL using single-frequency processing. The datasets partly overlap, and both contain periods with AS-on and AS-off. Together the datasets provide more data than before in the GPS/MET period of 1995-1997.  Since the JPL dataset relies on L1 single-frequency processing, it is insensitive to the L2 encryption (AS-on).  On the other hand, the single-frequency processing is affected by psudorange noise.  All datasets are compared to MERRA-2 reanalyses in the 10-30 km range, concluding that errors at some altitudes and latitudes in MERRA-2 are likely larger than the errors from either approaches of retrieving the refractivity from the GPS/MET data.  Thus, it is likely that future reanalysis projects could benefit from single-frequency processed GPS/MET data.

At times I found it difficult to understand exactly how the single-frequency processing is done, and I think a few additional equations may help (see specific comments below). Also, parts of the explanations seem to be contradictory, but I think it is just a question of being more precise in the language.

At the end, a possibly use of single-frequency processing is discussed, suggesting that it may provide some information on the magnitude of the residual ionospheric error due to raypath separation in dual-frequency processing. I have some doubt on that, partly because I don't think the single-frequency processing avoids raypath-induced residual ionospheric errors (see specific comment below), and partly because I think errors from pseudorange noise may be too large to be able to conclude much about residual ionospheric errors. However, I think it would be relatively easy to estimate the errors in bending angle due to the pseudorange noise, and I suggest a simple way to do that at the end of this review. This could then be compared to the expected size of residual ionospheric errors.

Specific comments:

There are many acronyms in the abstract that are not defined. I assume they should be defined, but I'm not sure of the journal's policy on this.

Page 2, line 35-37:

The order and logic seems reversed here. Wouldn't it make more sense to say that "... combining observations from multiple missions ... must be addressed". And then "Therefore, ... measurement accuracy needs to be characterized ..."?

Page 4, line 125:

Would it be more correct to say "A benefit of the single-frequency data set is that we can assess the impact on the dual-frequency data set ..."?  Was the CDAAC processing different (e.g., different filtering methods) in AS-on periods than in AS-off periods?

Page 5, line 153: 

"... are of equal but opposite sign". I suppose it should be 'size' after 'equal'.

Equation 1: 

What is actually meant by 'delay' here? Is L (disregarding noise) the same as the optical path length minus the distance between transmitter and receiver?

Equation 3:

I think the signs on mu_1 and nu_1 should be opposite of what they are in this equation. Otherwise the math does not work out.

Page 6, line 170:

Why not just add (1) and (2) and divide by two? Wouldn't that give the same without the intermediate calculation of I? Or is there another reason for the intermediate calculation? I think it becomes clear later why, but that should perhaps be mentioned already here.

Page 6, line 175:

I do not understand this argument about the bandwidth. I am fine with that there in principle is no raypath separation at all between L1 and P1 measurements. These come from the same signal, and in the geometrical optics formulation at L-band frequencies there is only one ray path for that signal. Mathematically, L1 phase and P1 pseudorange relate to the integral of the refractive index and group refractive index, respectively, along this path.  However, it is unclear to me if this avoids raypath-induced residual ionospheric errors in the further processing. When using eq.(3) to remove the ionospheric delay from L1 in eq.(1), you still have the integration along the original L1 path embedded in the eta term (as I understand the equations and the notion that this is a calibrated GPS signal, eta is the integral along the L1 path of the neutral atmospheric part of the refractive index minus the distance between transmitter and receiver). Thus, a part of eta depends on the ionospheric gradients because the path of the L1 signal depends on the ionospheric gradients. I don't think that is negligible in the context of residual ionospheric errors. How do you deal with this in the further processing?  What are the equations that you use? I think these equations needs to be given, so that the results can be reproduced by others, and so that it can be understood to which degree raypath-induced residual ionospheric errors are present.

Page 7, line 191-195:

Here it becomes clear that the sampling rate is not the same for L and P. Is smoothing of I in eq.(3) and/or interpolation there not necessary? It is mentioned that smoothing is applied directly to P_1.  Is that really correct? Don't you have atmospheric variation in P_1 that you don't want to smooth (eta in eq.(2))? Why not apply the smoothing to I in eq.(3) where the atmospheric variation has been removed? Well, I think it is when I read on. Revision in the text here is probably needed.

Page 8, line 225-226:

Is it correct to use the word bias here? If it is different from profile to profile (with different noise), it is not a systematic error, and thus not a bias. The approach seems good, although I am not an expert in filters and smoothing. There will of course be a residual error after the smoothing no matter how you do it, and there will be error correlations between adjacent points, but I don't think you can call that a bias.

Page 8, line 234-235:

Here it is explained that the smoothing is applied to the ionospheric estimate in (3). I think that contradicts the information on page 7, where it was stated that smoothing is applied directly to P_1. It makes more sense if it is applied to I at 1-second intervals in (3). But I think the text on page 7 needs to be revised to be consistent with this.

Page 8, line 247:

I think I understand the approach, but I don't understand the sentence here. What is 'formal variance'?

Page 9, line 262:

This is basically the same as just mentioned two lines before.

Page 9, line 274-275:

Is it really smoothing of L2, or is it rather smoothing of L1-L2? In any case it is not the frequency that is smoothed, but the phase.  And not the lower SNR that is mitigated, but rather the negative effects of it. The language in the paper could be more precise here.

Page 9, line 315-317:

The sentence here does not really make sense to me. Could it be rephrased?

Page 11, line 332:

Could you provide one or more references to support "Other climate-related work" here?

Figure 4:

The figure shows median values. Would it be similar with mean values? Or are there a number of outliers that makes the median and mean significantly different?

Page 12, line 365:

Common profiles in three data sets? I suppose there are not common profiles between the CDAAC AS-on and CDAAC AS-off data sets. Either AS was on or it was off. It cannot be both at the same time for the same profiles. So should it rather be common profiles in two datasets here?

Page 12, line 366-369:

Could the reason for the differences also be that these are median values? With mean values one would expect to be able to see consistency between differences in Figure 4 and Figure 6, since this is more or less linear algebra. But with median values it is a more complex and different story, and one cannot generally expect such consistency.  It is difficult to see the differences in altitudes without a grid in the figures, but they seem quite similar (I do see a small offset at 10 km).

Page 13, line 390:

ECMWF-Interim? Do you mean ERA-Interim?

Page 14, line 449-450:

I was not able to find this estimate (1% near 30 km) in (Danzer et al., 2013).  The number seems at least an order of magnitude too large.  Danzer et al. (2013) show mostly errors/differences in bending angle, but at much higher altitudes. They show also the effect on temperature profiles (in their Fig. 8), but biases in temperature at 30 km can be very different from biases in refractivity, because of downward error propagation via the hydrostatic integration and large biases in the retrieved pressure. In the introduction of Danzer et al. (2013), they cite error estimates in previous works (Schreiner et al., 2011) of 0.045% at 30 km for refractivity, which sounds much more reasonable to me.

A couple of questions related to this: What would be the size of errors in bending angle (in micro-radians), that typical pseudorange noise could create in single-frequency processing? And how would this compare to expected residual ionospheric errors in dual-frequency processing?  I think you would be able to answer these questions with the data that you have: You could take the derivatives with respect to impact altitude (in m) of the differences between the L1-L2 and CA-L1 fits in figure 3 (examples c, d, e, and f). That should give you four examples of bending angle errors (in radians) due to pseudorange noise between 15 and 60 km. I don't know the answer myself, but me feeling is that it will be difficult (even when averaging over many profiles) to say anything conclusive about the residual ionospheric errors using the single-frequency processing because of the pseudorange errors. In any case, it would be very interesting and very relevant to see example estimates (and perhaps also ensamble averages) of the bending angle errors from the single-frequency processing with this straight-forward approach. I strongly suggest such estimates to be included in the paper.