Time between Symptom Onset, Hospitalisation and Recovery or Death: Statistical Analysis of Belgian COVID-19 Patients

Round 1

Reviewer 1 Report

The authors have appropriately addressed all my comments, and I suggest the paper published as it is now if other reviewers' comments have also been addressed.

Reviewer 2 Report

I am happy with the revised versions, and the authors have adequately addressed my original comments. 

This manuscript is a resubmission of an earlier submission. The following is a list of the peer review reports and author responses from that submission.

Round 1

Reviewer 1 Report

The paper models time (days) from symptom onset to diagnosis and hospitalization or the length of stay (LoS) in the hospital among COVID patients information COVID-19 patients admitted to 114 Belgian hospitals between March 14 and June 12, 2020. Weibull and log-normal distributions are used in the analyses accounting for interval censoring and right truncation of the time intervals. 

The paper reads very well, and it adds base knowledge into this exciting work on the epidemiology of the COVID 19, using an example Belgium patient cohort. I have few comments that the authors could consider.

• There is no justification on the use of 31 days as truncation time? Is this from some clinical data elsewhere?
• Some categories are arbitrary say for age: 20-60, 60-80?
• I am not sure if its necessary to assess effects of covariates of shape (Weibull) and variance (log-normal), what purpose do the results from these results serve?
• Arbitrary perturbations of 0.5 and 1 are used to constrain the count data (days) to somewhat continuous variables that are amendable to the Weibull or log-normal distribution. Why not just use Poisson log-linear, given that 0.5 wont much different to x of say 10 and above? I would think of subtracting more days in term of onset of symptoms. It is subject to recall bias, how does know really the actual time they started?
• Given that the data are clustered at hospital level (on LoS), could you think of something on the lines of multilevel modeling? Conditional or marginal (this looks more like here as you more concerned with “population averages” effect
• Only COVID patients hospitalized are included? Apparent selection bias? I did not get it about those still in hospital, are these not censored right?
• what happneed when age or gender was missing?

 

Author Response

We would like to thank the reviewer for the detailed comments.  In attachment the answers to your questions.

Author Response File: Author Response.docx

Reviewer 2 Report

The authors collected data on 14,618 hospitalized patients with COVID-19 admissions from 114 Belgian hospitals between March 14 and June 12, 2020 and developed regression models to analyze the time between symptom onset and diagnosis, hospitalization or length of stay (LoS) in the hospitals in Belgium. The models accounted for interval censoring and right truncation of the time intervals, as well as the possible under-reporting in the online survey despite the high reporting coverage (> 70% of all hospitalized COVID-19 cases). One of the strengths of the analysis was that two distributions (Weibull and lognormal) were assumed for the time interval data and Bayesian Information Criterion (BIC) was evaluated to select the best fitting models. The models revealed significant differences in delayed time and LoS between various covariate groups, and the authors conducted sensitivity analyses and showed that the results were robust to the assumptions made in the modeling. Upon review of the manuscript, I find that the study samples were amply collected with no obvious selection bias, the statistical analysis was carefully conducted, and the evidence presented generally support the conclusions. A few more clarifications and minor analyses would be helpful to solidify and confirm the author’s findings.

Major Comments:

1. The methods need more clarifications. Were the ML estimators expressed in explicit form or did the authors use an optimization algorithm? If an optimization algorithm, such as the gradient descent method, was adopted, how fast does it converge? More importantly, the consistency and asymptotic properties of the estimators need to be mentioned, especially that a post-stratification weight was added to the likelihood function. It should be as trivial as simple applications of the law of large numbers and Slutsky’s theorem since the consistency and asymptotic normality of MLE have been fully explained, however some audience might not realize this conspicuousness.
2. In the Results section, the authors mentioned a fair amount of between-group differences stratified by another covariate, while only the marginal differences independent of other covariates were inferenced in the models. I wonder if interaction terms can also be explored to statistically support the results. The sample size is surely sufficient for an interaction effect analysis.
3. In Discussion Line 213-218, the authors compared the LoS between Belgium and other countries. It is essential to mention the criteria of hospital release for COVID patients in these countries, without which the comparisons of LoS would be pointless.
4. The discussions should be enriched regarding the mechanism for differences of delay time and LoS between various covariate groups. General audiences are more interested in the clinical aspects of what caused the differences and what can be done to improve the situation.

Minor Comments:

1. In Table 2 and 3, some of the estimations were noted as “/”, and the authors did not provide any explanations.
2. When referring to tables and figures in the supplementary materials, sometimes Table Sx and Figure Sx were noted, while sometimes Table Ax and Figure Ax were noted.
3. In Discussion Line 204, I think it should be (in the age group < 20), not > 20. The authors should check whether other typos exist.

Author Response

We would like to thank the reviewer for the detailed comments.  In attachment the answers to your questions.

Author Response File: Author Response.docx