This study deals with the problem of explaining the variance observed for speech-in-noise performances across individuals. Here, this is approached by using tasks indicating peripheral and central auditory processing. The study introduces 1) a novel Auditory Figure Ground paradigm (a low and high frequency version), and 2) SEMs to better understand the underlying structure of factors contributing to the SIN variance. In a larger pool of participants, the new tests were evaluated by comparison to other tasks via correlation showed significant similarity to SIN task, while regression results showed that adding these new AFG tests can explain more variance with significant coefficients. The AFG and other tests were then used to build 3 SEMs with latent variables for AFG (and SIN in one model). The results indicate that SIN perception requires age, PTA, and the AFG measures of grouping/central auditory system to describe the observed variance for SIN across partictipants.
The study is straightforward including two novel aspects (AFG task and SEM) that are both of high relevance to the community and potential for the clinical field. Still, some concerns regarding the description and methodology require attention before publication.

Main points:
1. A confusion matrix of tests (SIN, AFG, Age, PTA, SSQ [could be removed if n.s.]) seems motivated as a first exploration, in particular for describing the new AFG tests hearing tests. Please report the additional correlations in text (l. 556/7 hints for example at the relationship of Age and PTA), add these to figure 1, or report them in a table.
2. SEMs: I enjoyed reading about the use of SEMs for formalizing the knowledge/insights from the different hearing tests into an underlying model. I think it’s an innovative approach to aggregate hearing tests to suggest the underlying functioning and links of the hearing system. Overall, the authors provided good descriptions and discussed important points for interpretation of the results. Still, the creation of such models is difficult given the vast search space of how to set the SEMs up. The approach taken, to make use of prior knowledge and the hierarchical regression results, seems the most feasible and likely necessary. I think some points still require attention
a. Model 3 was chosen as the best model and described. I agree that the reported statistics hint at this. However, this conclusion requires a test as the fit will be better given the larger number of parameters. Please test whether adding the two parameters increases the model variance using difference tests.
b. Related to a. In tables 2 and 3, the Rsquare is indicated. In the caption of Table 3, it says ‘adjusted Rsq also reported in last row per model’. If the Rsq in Table 3 is the adj Rsq, please say this in the table itself. Providing the adj Rsq seems desirable, also for the hierarchical regression results (Table 2).
Minor points:
1. l.321: The authors say that the inclusion of SIN tasks as well as AFG tasks was determined by a prior analysis and a subset of participants. I wonder the reasoning behind using a subset that was afterwards still included when computing the full models: either one uses a fully independent set of data to determine the structure or, as the subset enters the computation of the full models, use the full set of participants. The first would be preferred but I understand that the number of participants necessary for training the full models makes this difficult. A very elegant solution would be to keep the data fully independent, but I understand if this not possible given the data size requirement.
2. L.392. The statement of lower variability seems strong as it doesn’t take the absolute value into account (which is higher for AFG and lower for SIN tasks). For a statement like this, a test seems warranted.
3. L.577-584. The authors worry about the small number of data (participants/observation) vs estimated coefficients, and I agree with their concern. To get a better idea on the robustness and correctness, bootstrapped SEMs could provide confidence intervals of estimated parameters. This should be possible with the lavaan package.
4. L.235 & 267. ‘final score was calculated by averaging the dB SNR …’: later on (e.g., l. 292), it is described that the median was used, please clarify.
5. Fig.1. Please specify the hearing threshold units (dB_HL?)
6. Fig.4. Please indicate the unit for demographics (years, db_HL). Readers might wonder whether these were transformed (which was probably not the case) given the use in the same graph.
7. It might be helpful to refer to the supplemental data in the methods section. It was a bit hidden in the manuscript. Please add the code for the analysis as well. If possible, sharing the material would be great given the general interest from the auditory neuroscience field as additional data point for describing the participants’ hearing.