Please refer to the supplement file attached here.

Thank you for the opportunity to review this manuscript. The major concern I have on the manuscript is centered on the analysis of the data. These emissions data are multivariate, strictly positive and relative which means they are compositional data and should be analyzed as such. The data can be transformed into log-ratios which places them on the real number line thus enabling application of many familiar statistical tools. References for this approach to the analysis of emissions data are presented in the comments below. Failure to use this approach can result in spurious correlations between the emissions and errors in interpretation of results. Another commonly used technique in emissions analysis and source apportionment is positive matrix factorization (Sekimoto, K., Koss, A. R., Gilman, J. B., Selimovic, V., Coggon, M. M., Zarzana, K. J., Yuan, B., Lerner, B. M., Brown, S. S., Warneke, C., Yokelson, R. J., Roberts, J. M., and de Gouw, J.: High- and low-temperature pyrolysis profiles describe volatile organic compound emissions from western US wildfire fuels, Atmospheric Chemistry and Physics, 18, 9263–9281, https://doi.org/10.5194/acp-18-9263-2018, 2018). This multivariate technique does not consider the relative nature of emissions data composition.

1. 81-83 This is a 1 sentence paragraph. Either expand the text or include it in the preceding or subsequent paragraph.
2. 99 Does the VOCUS identify the characteristic compound or does post-sampling analysis by an investigator identify a “characteristic” compound? What is meant by the term “characteristic” compound for a fuel type?
3. 111 Are you measuring the solid fuel combustion emissions (which would be from char) or the emissions produced from the combustion of gaseous products produced by the pyrolysis of solid fuels?
4. 117 Was either proximate or ultimate analysis performed on the fuels? I would expect a significant difference in N in the cow dung compared to the other fuels. If such a difference exists in the unburnt fuel, it would seem that it would translate through the combustion and into the emissions and identification of the characteristic compounds. The fuel composition is also compositional data and should be analyzed accordingly.
5. 125 What was used to represent agricultural waste? Was agricultural was straw only? Please clarify the difference between the agricultural waste and the fuels used to simulate “forest fires”? Was there a difference between the fuel arrangement or the burning conditions?  I recommend that you don’t use these fuels to characterize “forest fires” as there is a wide range of fuels which burn in forest and bush fires ranging from peat soils to coniferous and hardwood forest fuels to grasses to various shrub fuels.
6. 133-143 The burning of the logs is described. How did this differ from the straw burning?  Straw will ignite and burn more quickly than wooden logs. What was the moisture content of the various fuels? Was a constant heating rate used? These pyrolysis and combustion characteristics will affect time to ignition as well as the composition of the emissions.
7. 180-181 While Andreae and Merlet used the carbon mass balance approach in 2001, it was first used as early as 1969 (Boubel, R. W., Darley, E. F., and Schuck, E. A.: Emissions from burning grass stubble and straw, Journal of the Air Pollution Control Association, 19, 497–500, https://doi.org/10.1080/00022470.1969.10466517, 1969) and was well-established by the mid-1980s (Nelson, R. M., Jr.: An evaluation of the carbon balance technique for estimating emission factors and fuel consumption in forest fire, USDA Forest Service, Southeastern Forest Experiment Station, Asheville, NC, 1982.)
8. 193 Please make sure that the subscripts for C, O, and N are consistently italicized (or not). Make sure that the subscript for oxygen is O and not zero.
9. 202 It has been recently shown that smoke emissions data are multivariate, not independent and are relative values that are dependent on the compounds present in the mixture (Gibergans-Baguena, J., Hervada-Sala, C., and Jarauta-Bragulat, E.: The quality of urban air in Barcelona: a new approach applying compositional data analysis methods, Emerg Sci J, 4, 113–121, https://doi.org/10.28991/esj-2020-01215, 2020; Jarauta-Bragulat, E., Hervada-Sala, C., and Egozcue, J. J.: Air Quality Index revisited from a compositional point of view, Math Geosci, 48, 581–593, https://doi.org/10.1007/s11004-015-9599-5, 2016; Weise, D. R., Palarea‐Albaladejo, J., Johnson, T. J., and Jung, H.: Analyzing wildland fire smoke emissions data using compositional data techniques, J. Geophys. Res. Atmos., 125, e2019JD032128, https://doi.org/10.1029/2019JD032128, 2020). These characteristics of the data apply whether the emissions data are expressed as emission factors, emission ratios, mole ratios or mass ratios (van den Boogaart, K. G. and Tolosana-Delgado, R.: Analyzing compositional data with R, Springer, Heidelberg, 258 pp., 2013.).

It has also been shown that MCE as an index describing the completeness of combustion is not independent of the quantities of other emissions and that it should not be used as a predictor for the other gases in the composition.

How is the Mann-Whitney test and other techniques used in this manuscript affected by the statistical characteristics of your data? The Mann-Whitney test is a univariate test. You should consider using the generalized multivariate version if it has been applied to compositional data (https://doi.org/10.1016/j.jmva.2022.104946). As compositional data analysis has been used more extensively in Europe, recommend reaching out to the statisticians listed in the various publications above. You should also consider a global test (instead of pairwise comparisons) that controls the experiment-wise probability of committing a Type 1 error (such as false discovery rate-Benjamini, Y. and Hochberg, Y.: Controlling the false discovery rate: a practical and powerful approach to multiple testing, Journal of the Royal Statistical Society. Series B (Methodological), 57, 289–300, 1995).

1. 232 What does 0.99 +/- 0.02 mean? Is this arithmetic mean and standard error or standard deviation? Since MCE is a proportion that can not exceed 1, the correct formula for the confidence interval of this proportion should not exceed 1. The geometric mean is the appropriate measure of central tendency for relative (proportional data). EFs are expressed as gm pollutant/gm fuel burned which is a rate (and a relative value so a geometric mean should be used as in Butler, B. M., Palarea-Albaladejo, J., Shepherd, K. D., Nyambura, K. M., Towett, E. K., Sila, A. M., and Hillier, S.: Mineral–nutrient relationships in African soils assessed using cluster analysis of X-ray powder diffraction patterns and compositional methods, Geoderma, 375, 114474, https://doi.org/10.1016/j.geoderma.2020.114474, 2020.
2. 257 Correlation is not an appropriate measure for compositional data as the value of the correlation coefficient is dependent upon the other compounds in the composition. Dropping a gas from the composition changes the pairwise correlations (Aitchison, J.: A concise guide to compositional data analysis, 2003. Available at http://ima.udg.edu/activitats/codawork03/; Weise, D. R., Fletcher, T. H., Safdari, M.-S., Amini, E., and Palarea-Albaladejo, J.: Application of compositional data analysis to determine the effects of heating mode, moisture status and plant species on pyrolysates, Int. J. Wildland Fire, 31, 24–45, https://doi.org/10.1071/WF20126, 2022). Proportionality has been suggested as an appropriate measure of the association between two components of a composition (Lovell, D., Pawlowsky-Glahn, V., Egozcue, J. J., Marguerat, S., and Bähler, J.: Proportionality: a valid alternative to correlation for relative data, PLoS Comput Biol, 11, e1004075, https://doi.org/10.1371/journal.pcbi.1004075, 2015). Drawing conclusions based on the measure of association between two variables without determining the significance of the measure by a statistical test of some sort is not recommended.
3. 290 In compositional data analysis, the effect of fuel type on the log-ratios between different groups of compounds can be tested in an analysis of variance context to determine differences. These log-ratios are known as balances (Egozcue, J. J. and Pawlowsky-Glahn, V.: Groups of parts and their balances in compositional data analysis, Mathematical Geology, 37, 795–828, https://doi.org/10.1007/s11004-005-7381-9, 2005; Weise, D. R., Fletcher, T. H., Safdari, M.-S., Amini, E., and Palarea-Albaladejo, J.: Application of compositional data analysis to determine the effects of heating mode, moisture status and plant species on pyrolysates, Int. J. Wildland Fire, 31, 24–45, https://doi.org/10.1071/WF20126, 2022).
4. 308 see comment below for figures S4, S5 regarding error bars.
5. 348 You are discussing differences in relative terms which is appropriate. The statistics used to describe and test hypotheses should also recognize the relative nature of the data.
6. 370 See the earlier comment regarding the Mann-Whitney test, the multivariate nature of the data and the probability of committing a Type 1 error.
7. 398 Which supplementary table contains the characteristics compounds?
8. 423 Where is the chemical composition of unburnt cow dung presented?

Table 1 Are the values arithmetic mean +/- standard deviation?  Please provide more information on values. Should use geometric mean and present a confidence interval (or the standard error of the mean). Consider including complete fuel composition (CHNSO). Also proximate analysis because cows have ability to digest cellulose which make affect the burning characteristics or the relative amounts of cellulose, hemicellulose and lignin in the fuels which will affect both the pyrolysis and combustion processes as well as emissions production.

Figure 2-5           Recommend making the axes and other information larger fonts (relative to titles).

Figure 5               What do the different sized circles labeled 0.1, 0.5, 1 and 2 indicate?

Figure S4, S5     Why is ½ of 1 standard deviation used as an error bar? Is this is based on the assumption made for normally distributed data that 1 standard deviation captures about 68 percent of the data and 2 standard deviations capture about 95% of the data? This is based on the population and not the sample. A confidence interval should be calculated.  Also, these data are not normally-distributed. They are proportions which are constrained between 0 and 1 (or 0 and 100).

Figure S13          Caption needs to be fixed.