Comparison of Associations Between Glucocorticoids Treatment and Mortality in COVID-19 Patients and SARS Patients: A Systematic Review and Meta-Analysis


 BackgroundThe response to glucocorticoids treatment may be different between Covid-19 and SARS. MethodsIn this systematic review and meta-analysis, we searched studies on Medline, Embase, EBSCO, ScienceDirect, Web of Science, Cochrane Library, ClinicalTrials.gov, ICTRP from 2002 to October 7, 2020. We used fixed-effects and random-effects models to compute the risk ratio of death in the group receiving glucocorticoids treatment and the control group for COVID-19 and SARS, respectively.ResultsTen trials and 71 observational studies, with a total of 45935 patients, were identified. Glucocorticoids treatment, was associated with decreased all-cause mortality both in COVID-19 (risk ratio, 0.88; 95% confidence interval, 0.82 to 0.94; I2=26%) and SARS (0.48; 0.29 to 0.79; 10%), based on high quality evidence, as well as decreased all-cause mortality-including composite outcome of COVID-19 (0.89; 0.82 to 0.98; 0%). In subgroup analyses, all-cause mortality was significantly lower among COVID-19 patients being accompanied by severe ARDS but not mild ARDS, taking low-dose or pulse glucocorticoids, being critically severe but not only severe, being of critical severity and old but not young, being of critical severity and men but not women, non-early taking glucocorticoids and taking dexamethasone or methylprednisolone; but for SARS, lower mortality were observed among those who were taking medium-high dose glucocorticoids, being severe or critically severe, early taking glucocorticoids, and taking dexamethasone or prednisolone. ConclusionsGlucocorticoids treatment reduced mortality in COVID-19 and SARS patients of critical severity; however, different curative effects existed between the two diseases among subpopulations, mainly regarding sex- and age-specific effects, optimal doses and use timing of glucocorticoids.


Introduction
The coronavirus disease 2019 (COVID-19) and severe acute respiratory syndrome (SARS) are both coronavirus-caused pandemics that broke out in 2019 and 2003, respectively. The SARS epidemic has passed, but COVID-19 is ravaging the world. As of early November 2020, nearly 50 million people have been con rmed with COVID-19, and 1.2 million have been dead of this disease. 1 Glucocorticoids, one of the non-speci c immunosuppressants, were the mainstay of immunotherapy during the SARS epidemic, but its role has been a lasting controversy, same in the COVID-19 epidemic. Till July 2020, RECOVERY trial, a large clinical randomized controlled trial (RCT) from the UK, involving 6425 patients, reported an absolute decrease in all-cause mortality of 2.8% in COVID-19 patients receiving a low dose of dexamethasone compared with those receiving usual care. 2 Lately, a prospective metaanalysis involved 7 RCTs concluded systemic glucocorticoids were associated with decreased all-cause mortality in critically ill patients with COVID-19. 3 However, scholars pointed out that removing the RECOVERY trial, the result of that meta-analysis changed to no statistical signi cance. 4 Therefore, to obtain a robust product, more ongoing RCTs should be included.
Although evidence has shown severe dysregulated host in ammatory and immunity response is deadly both in COVID-19 and SARS patients [5][6][7][8] , their epidemiology and clinical manifestations are substantial differences due to the subtle differences in pathogenesis. 9,10 We speculate that the response to glucocorticoids treatment is different between the two diseases. Moreover, which doses, timing, and types of glucocorticoids treatment are more effective, and whether there are any other survival bene t subgroups besides critically ill patients have not been determined. Therefore, we conducted this systematic review and meta-analysis to provide more clues in these undetermined aspects.

Guidance and Protocol
We followed the standards developed by the meta-analysis of observational studies in epidemiology (MOOSE) 11 and preferred reporting items for systematic reviews and meta-analyses (PRISMA) 12 for reporting our study. The protocol for this review and meta-analysis was registered with PROSPERO (CRD42020193823).

Eligibility criteria
We considered studies to be eligible if they met the following PICOS criteria (participants, interventions, comparators, outcomes, and study design). The participants of interest included patients (adults/children) who were diagnosed with COVID-19 or SARS. The intervention included any type of glucocorticoids, including but not limited to hydrocortisone, prednisolone, methylprednisolone, prednisone, dexamethasone, and triamcinolone, compared with placebo or usual care (which may have included antiviral, antibiotic and antifungal therapy, intravenous immunoglobulin, or respiratory support, if needed). The primary outcome was all-cause mortality (including 28-or more-day mortality, 14-to 28day mortality, and 7-day mortality). All-cause mortality rates were used to compute the pooled analysis on 7-day mortality if actual 7-day mortality rates failed to be neither extracted from the published studies nor obtained from study authors. The secondary outcome was the mortality-including composite outcome, including death and mechanical ventilation or intensive care unit (ICU) admission. The de nitions of outcomes are presented in Supplementary Table 1. Both RCTs and observational studies (including cohort studies, case-control studies, case series) were included.
Studies were excluded if they were before-after studies, case series with no more than ten patients, overlapping/repeated cohort data, and no or contaminated control design.

Literature search
Two of the authors (JB.L. and XL.L.) developed and executed the search strategy of several databases: Medline (Ovid), Embase (Ovid), EBSCO (H.W. Wilson: OmniFile Full-Text Mega), ScienceDirect, Web of Science (All database), Cochrane Library, from 2002 to October 7, 2020. We also searched ClinicalTrials.gov and consulted the WHO International Clinical Trials Registry Platform (ICTRP) to identify any ongoing or unpublished eligible studies. To maximize the search for relevant articles, we reviewed reference lists of identi ed studies, systematic reviews, and review articles on the same topic.
We did not apply language or publication status restrictions. Supplementary Table 2 presents the details of the search strategy.

Study selection
After removing duplicates, two independent groups of four authors (H.Y. and W.Z.; Y.Z. and LP.W.) screened the titles and abstracts to determine whether the citation met the participants and intervention eligibility criteria. They obtained full texts and then further screed when studies met all the rest eligibility criteria. Disagreements between groups were resolved by consensus, and if necessary, consultation with a third author (ZW.Z.).

Data collection process
Two independent groups of four authors (H.Y. and W.Z.; Y.Z. and LP.W.) extracted data from included studies into standard data collection forms and created tables for quality assessment evidence and information. Evidence of severity of illness and administration of glucocorticoids reported in the eligible studies were collected and categorized according to the pre-de ned (Supplementary Table 1).

Assessment of risk of bias
Two independent groups of four authors (H.Y. and W.Z.; Y.Z. and LP.W.) performed the risk of assessment using the Revised Cochrane risk-of-bias tool for randomized trials (RoB 2) and using the Newcastle-Ottawa-Scale (NOS) 13 for observational studies. The included randomized trials were assessed for (1) randomization process, (2) deviations from intended interventions, (3) missing outcome data, (4) measurement of the outcome, (5) selection of the reported result. Each domain was assessed as low risk, some concerns, or a high risk of bias. The study's overall risk of bias was determined by the highest risk of bias for any criteria. The included observational studies were assessed for (1) selection, (2) comparability, (3) exposure for case-control or outcome for cohort studies. Each domain is composed of 2 to 4 items of criteria, and each criterion was scored in the form of stars. A total score of 8 or 9 was assessed as low risk of bias, 6 or 7 as some concerns, and ≤5 as high risk. Disagreements between groups were resolved by consensus, and if necessary, consultation with a third author (ZW.Z.).

Data synthesis
We performed statistical analyses used the meta package in R (version 4.0.1; The R Project for Statistical Computing). Analyses were conducted for the intention-to-treat-based outcomes reported from RCTs and conducted for the narrowest-95%-con dence-interval-based data of risk for outcomes reported from observational studies. We equivalently transformed hazard ratios (HRs) 14 and odds ratios (ORs) 15 into risk ratios (RRs) and log-transform these effect sizes as well as their standard errors rst before they were pooled. We used RRs and their associated 95% con dence interval (CI) to assess outcomes, as well as a prediction interval (PI) for the effect of future studies based on present 16 , and we considered a 2-tailed P value of less than 0.05 to be statistically signi cant. We examined heterogeneity using the I 2 test. If nonsigni cant heterogeneity present (I 2 <50%), we used both xed-effects and random-effects models to pool outcomes; we used random-effects models when signi cant heterogeneity was present (I 2 ≥50%).
The possibility of publication bias was qualitatively visualized through contour-enhanced funnel plots and quantitatively tested through the Egger test. Duval & Tweedie's trim-and-ll procedure was also adopted to estimate the actual effect size when had the "missing" small studies been published. 17 Trial sequential analysis We performed a trial sequential analysis to explore whether RCTs' cumulative data were adequately powered to evaluate outcomes. Trial sequential analysis 18 (a Java-based software application, version 0.9.5.10) was used to calculate the required information size, and the O'Brien-Fleming approach was adopted to compute trial sequential monitoring boundaries. An optimal information size was calculated based on the event proportion of outcomes in the control arm, 2-sided 5% risk of type I error, 20% risk of type II error (power 80%), and an anticipated relative risk reduction of more than 10% in this study.

Multimodel inference
We performed multimodel inference to examine which possible predictor combinations provide the best t, and which predictors are the most important ones overall. 19 The "dmetar" package in R was used to calculate the multimodel inference coe cients for each predictor and all predictors and obtain modelaveraged importance of the predictors' plot. A total of 7 factors were considered as input predictors: incidence of events in control, the severity of illness, doses of glucocorticoids, types of glucocorticoids, the timing of glucocorticoids, primary data type, and time frame of mortality. No prior knowledge was available on how these predictors are related to effect sizes in our meta-analysis.

Sensitivity analyses
We conducted sensitivity analyses by (1) excluding studies with high, some concerns, or both risks of bias, (2) excluding the largest studies, (3) excluding studies reporting multiple glucocorticoids use, (4) excluding studies identi ed by clustering-algorithms-dependent Graphic Display of Heterogeneity (GOSH) plots 20-23 , (5) combining data from RCTs and observational, (6) using odds ratios to assess outcomes.

Eligible studies and study characteristics
Of the 6136 records, 81 studies 2,24-103 involved a total of 45935 patients were included in the nal metaanalysis ( Figure 1). Table 1 shows a summary of the included studies. The studies comprised 10 RCTs for COVID-19, including 7898 patients, 59 observational studies for COVID-19 including 26735 patients, and 12 observational studies for SARS including 11302 patients. These studies for COVID-19 reported median mortality of 17% and a median age of 57.4 years among their included patients, and studies for SARS reported median mortality of 6.3% and a median age of 46 years among their included patients. Supplementary Table 3 gives details of those studies.
Supplementary Figures 1 and 2 show a risk bias for RCTs. Seven trials had a low risk of bias, and three trials had a risk of some concerns. Supplementary Tables 4 and 5 show risk bias for observational studies. In case-control studies, two studies reported outcome data with low risk or risk of some concerns, three studies reported outcome data with risk of some concerns, two studies reported outcome data with high risk or risk of some concerns, and 29 studies reported outcome data with high risk. In retrospective cohort studies, 1 study reported outcome data with low risk, 13 reported outcome data with low risk or risk of some concerns, 13 studies reported outcome data with some concerns, four studies reported outcome data with high risk or risk of some concerns, and four studies reported outcome data with high risk.
Primary outcome: all-cause mortality All 10 RCTs reported all-cause mortality for COVID-19. The overall RR (0.88; 95% CI, 0.82 to 0.94; I 2 =26%; PI, 0.81 to 0.95) ( Figure 2) revealed an association between glucocorticoids treatment and improved allcause mortality in COVID-19 patients. There was no statistically signi cant difference in that association between 7-day, 14-to 28-day, and 28-or more-day mortality. The trial sequential analysis con rmed that the overall required information size was met for all-cause mortality (Supplementary Figure 3). Similar results were observed for all-cause mortality in all conducted sensitivity analyses (Supplementary Table  6). Funnel plot analysis showed no asymmetry (Supplementary Figure 12), and the Egger test (P=0.747) detected no signi cant small-study effects. A total of 52 observational studies for COVID-19 reported allcause mortality, and the pooled RR from studies of all risk levels showed an association between glucocorticoids treatment and increased all-cause mortality in COVID-19 patients (RR, 1.32; 95% CI, 1.08 to 1.61; I 2 =99%) (Supplementary Figure 5) while result from 11 studies with low risk of bias was similar with that from RCTs (RR, 0.68; 95% CI, 0.5 to 0.94; I 2 =67%) (Supplementary Table 6). Funnel plot analysis showed no asymmetry among observational studies reporting all-cause mortality for COVID-19 (Supplementary Figure 14).  Table 6). Funnel plot analysis showed no signi cant asymmetry (Supplementary Figure  13), the Egger test (P=0.403) detected no signi cant small-study effects. Additionally, Duval & Tweedie's trim-and-ll procedure (with 2 added studies) showed similar results with the initial ones.
Evidence from observational studies of all risk levels showed no signi cant association between glucocorticoids treatment and all-cause mortality; however, evidence indicated a signi cant reduction in 28-or more-day mortality-including composite outcome in COVID-19 patients receiving glucocorticoids treatment (RR, 0.53; 95% CI, 0.43 to 0.66; I 2 =83%) (Supplementary Figure 6). Sensitivity analyses excluding studies with a high risk of bias con rmed the signi cant reduction both in 28-or more-day and all-cause mortality-including composite outcome (Supplementary Table 6). No signi cant small-study effects were observed among these observational studies reporting the mortality-including composite outcome of COVID-19 (Supplementary Figure 15). Only one observational study reported con icting data of mortality-including composite outcomes of SARS and pooled RR showed no signi cant association between glucocorticoids treatment and all-cause mortality in SARS patients (RR, 0.84; 95% CI, 0.17 to 4.3; Figure 8).

Multimodel inference and subgroup analysis
Multimodel inference found that incidence of events in control, doses of glucocorticoids, and types of glucocorticoids were the top three important predictors for all-cause mortality in COVID-19 patients (Supplementary Figure 9). The former two were also the most important predictors for the mortalityincluding composite outcome of COVID-19 (Supplementary Figure 10). Differently, the types of glucocorticoids became the only most important predictor for all-cause mortality in SARS patients (Supplementary Figure 11).
Subgroup analysis revealed that all-cause mortality was signi cantly lower in COVID-19 patients with characters of being accompanied by ARDS, taking low-dose or pulse glucocorticoids, being severe or critically severe, being man, and taking dexamethasone or methylprednisolone ( Figure 3). Similar results occurred when the mortality-including composite outcome was considered ( Figure 4). Some differently, all-cause mortality was signi cantly lower in SARS patients with characters of taking medium-high dose glucocorticoids, being severe or critically severe, early taking glucocorticoids, and taking dexamethasone or prednisolone ( Figure 5). Because 2 of 3 studies for mild or moderate SARS patients reported zero events in both arms, there was no enough power to draw a convincing conclusion in these subgroup patients. Moreover, early taking glucocorticoids seemed to have no signi cant association with improved mortality in COVID-19 patients until when mechanical ventilation or ICU admission outcome was additionally considered (Figure 4 and 5).
Importantly, pooled effects for COVID-19 and SRAS patients' outcomes strongly correlated to the incidence of events in control. The bene t was only seen among studies that reported more than 20% of the incidence of events in control but was not seen for the incidence of less than 20% (Figure 3, 4, and 5). Pooled effects even showed a signi cant association between glucocorticoids treatment and worsen mortality reported from studies with the incidence of events in control from 0% to 5% (Figures 3 and 5). We strati ed studies according to these incidences to avoid the potential bias caused by the difference of outcome incidence reported in control (Supplementary Figure 17-24). Interestingly, results implied that COVID-19 patients who were critically severe (including severe ARDS) but not only severe (including mild ARDS), and COVID-19 patients who of critical severity were old but not young, or men but not women were more likely to obtain survival bene t from glucocorticoids treatment (Supplementary Figure 17-23,  25).

Discussion
In this meta-analysis of 10 RCTs (including 7898 COVID-19 patients) and 71 observational studies (including 26735 COVID-19 patients and 11302 SARS patients), glucocorticoids treatment was signi cantly associated with reduced all-cause mortality and mortality-including composite outcome both in COVID-19 and SARS patients. Subgroup analyses demonstrated that among COVID-19 patients, the bene cial effect in our outcomes was associated with being accompanied by severe ARDS but not mild ARDS, taking low-dose or pulse glucocorticoids, being critically severe but not only severe, being of critical severity, and old but not young, being of critical severity and men but not women, non-early taking glucocorticoids and taking dexamethasone or methylprednisolone; among SARS patients, that bene cial effect was associated with taking medium-high dose glucocorticoids, being severe or critically severe, early taking glucocorticoids, and taking dexamethasone or prednisolone. Our ndings suggested glucocorticoids treatment's survival bene ts both in COVID-19 and SARS patients and its association with the severity of illness; however, there were differences in the other aspects, mainly regarding sex-and agespeci c effects, doses, and timing of glucocorticoids treatment.

Principal ndings and comparison with other studies
As of writing this manuscript (early November 2020), 16 meta-analyses 3,104-118 have examined the use of glucocorticoids in patients with COVID-19 or coronavirus infections. Regarding mortality, 4 metaanalyses 3,106,111,112,117 included RCTs (1 including 7 RCTs, 3 including 1 or 2 RCTs) and the rest only included observational studies. Of these meta-analyses, 5 3,104-107 found that glucocorticoids treatment decreased all-cause mortality, 7 [108][109][110][111][112][113][114] found no association between glucocorticoids treatment and allcause mortality, and the others [115][116][117][118] found glucocorticoids treatment increased all-cause mortality. In general, meta-analyses that supported glucocorticoids treatment mostly identi ed studies on COVID-19 patients who were severely-ill or with ARDS. The latest meta-analysis 3 was a prospective meta-analysis involving seven clinical trials of 1703 critically ill patients with COVID-19. Findings in this meta-analysis showed that administration of systemic glucocorticoids was associated with lower 28-day all-cause mortality (OR based on a xed-effect meta-analysis, 0.66; 95% CI, 0.53 to 0.82; I 2 =15.6%). Results of this prospective meta-analysis without considering other ongoing trials were limited owing to its risk of selective reporting or publication bias. Moreover, the RECOVERY trial 2 exerted the most in uence (comprising 59% patients and contributing 57% of the weight) in that meta-analysis, and if the RECOVERY trial was excluded, results changed to no statistical signi cance. 4 Before that study, a singlearm meta-analysis 105 investigated the mortality in COVID-19 patients who were accompanied with ARDS and received glucocorticoids treatment and found that compared to 39% of all-cause mortality among COVID-19 patients with ARDS, glucocorticoids treatment decreased all-cause mortality to 28%. However, the limitation of the single-arm meta-analysis was obvious due to its non-pairwise design. There was one meta-analysis 118 investigating glucocorticoids use on mortality in SARS patients alone. The metaanalysis found no association between glucocorticoids treatment and all-cause mortality in SARS patients (RR, 2.56; 95% CI, 0.99 to 6.63; I 2 =77.4%). However, results from that meta-analysis were restricted owing to inconsistency (signi cant heterogeneity across involved studies), published bias (only seven studies), and risk of bias (including observational studies of all risk levels).
The ndings of our meta-analysis of the association of glucocorticoids administration with improved allcause mortality were in line with the recently published results on COVID-19. We further con rmed these ndings by including additional 3 RCTs and conducting the sensitivity analysis by excluding the largest trial (RECOVERY trial). Moreover, the results of the trial sequential analysis also reinforced our ndings.
Besides, we also investigated the effect of glucocorticoids treatment on the mortality-including composite outcome. We found a protective role of glucocorticoids treatment in 28-or more-day mortality-including composite outcomes in COVID-19 patients. Furthermore, results from observational studies with a low risk of bias were like those from RCTs.
Our ndings of the survival bene t of glucocorticoids treatment in SARS patients differ from previously published results on SARS. This difference could be explained by identifying more eligible studies and stratifying included studies according to the risk of bias in our meta-analysis. The high risk of bias mainly came from studies reporting multiple glucocorticoids use, which was also con rmed by multimodel inference. After removing these studies of a high risk of bias, survival bene t from glucocorticoids treatment was evident.
Another important difference in our meta-analysis is that we systematically compared the differences in the effects of glucocorticoids treatment on mortality between COVID-19 and SARS patients across subgroups de ned by patient characteristics before receiving glucocorticoids treatment, by disease severity, and by details of glucocorticoids regimen. We assessed optimal doses, types and timing of treatment, and other potential bene t subgroups by separately showing evidence across different levels of risk of bias from RCTs and observational studies. Results indicated that the old or men being critically severe or with severe ARDS were more likely to obtain survival bene t from glucocorticoids treatment among COVID-19 patients; however, there was no evidence to support sex-or age-speci c effect in SARS patients. Moreover, medium-or long-term types of glucocorticoids were associated with improved mortality both in COVID-19 and SARS patients. Regarding doses, COVID-19 patients were likely to bene t from low-dose or pulse glucocorticoids treatment; however, SARS patients may be more likely to bene t from medium-high dose glucocorticoids treatment. As for treatment timing, bene ts likely preferred to early taking glucocorticoids in SARS patients; however, no strong evidence supported early taking glucocorticoids in COVID-19 patients. In brief, optimal glucocorticoids regimens were different between COVID-19 and SARS patients, especially regarding sex-and age-speci c effects, doses, and treatment timing.

Strengths And Limitations
This systematic review and meta-analysis have several methodological strengths. We followed the recommendations of the MOOSE statement and PRISMA statement, including a priori protocol. This meta-analysis of all-cause mortality on COVID-19 met the optimum size in trial sequential analysis and was robust despite sensitivity analyses. We collected evidence from RCTs and observational studies, provided prediction intervals, and explored other predictors or confounding factors for all-cause mortality by multimodel inference. Through the systematic comparison of COVID-19 and SARS, we assessed potential bene t subgroups and the optimal regimens of glucocorticoids treatment. Considering potential bias caused by the various outcome incidence in control reported by studies, systematic subgroup analyses were also conducted across different outcome incidence levels in control.
Our study has limitations. First, the results of this meta-analysis on SARS were all from observational studies. Clinical heterogeneity is inevitable in these studies, including doses, types, timing, and therapy duration. Regarding statistical heterogeneity, the results of studies on SARS included in our analysis were variable, with a high degree of detected heterogeneity for the primary outcome of all-cause mortality (I 2 =98%), justifying by use of random-effects models when pooling data of all levels of risk of bias. By only the inclusion of studies with a low risk of bias, heterogeneity could be resolved but with a signi cant change in the primary outcome. Most studies with a high risk of bias reported multiple glucocorticoids use, which was the most important contributor to heterogeneity in the meta-analysis on SARS. There was a little heterogeneity (I 2 =26%) when pooling data of RCTs on COVID-19. This heterogeneity came from only one trial 32 , which reported pulse use of glucocorticoids. GOSH plots identi ed this trial, and by the exclusion of this trial, heterogeneity could be eliminated without signi cant change in the primary outcome.
Second, though funnel plots showed no asymmetry no matter among RCTs or observational studies, no matter for primary or secondary outcomes, publication bias may exist when subgroup analyses were conducted. Some subgroup analyses failed to be conducted due to limited data. To ensure the quality of results, we analyzed RCTs and observational studies separately; however, this may increase publication bias in subgroup analyses.
Third, prevalent time from 16 years, treatments and diagnostic techniques for virus pneumonia have evolved. Therefore, the medical and technical background should be considered when comparing treatment effects on SARS and COVID-19. For example, the severity of illness or clinical classi cation differed between SARS and COVID-19. For SARS, there was no strict distinction between severe and critically severe, which brings us di culty in interpreting related results.

Implications for practice
For many years and even today, glucocorticoids have aroused much controversy in treating viral pneumonia, including H1N1, SARS, MERS, and COVID-19. Historically, there was almost no hard evidence to support glucocorticoids treatment in this pneumonia. Even reviews and meta-analyses of this treatment indicated increased mortality, delayed viral clearance, and increased risk of superinfection. 119,120 Therefore, WHO suggests glucocorticoids use only under special circumstances i.e., septic shock or bronchoconstriction, however, based heavily on empirical evidence. Until the completion of the RECOVERY trial of COVID-19 therapy, the largest relevant RCT announced at present, from which evidence supported glucocorticoids treatment in critically ill COVID-19 patients who were on mechanical ventilation or oxygen support, we're looking at glucocorticoids therapy again. Though at present much evidence supporting its use in critically ill COVID-19 patients, we cannot blindly extend it to other subgroups of COVID-19 or other viral pneumonia. Our meta-analysis further revealed the massive impact of severity of illness on the role of glucocorticoids. It con rmed its e cacy in reducing mortality and the mortality-including composite outcome for COVID-19 patients being critically severe or with severe ARDS but not those being only severe or with mild ARDS. By comparing COVID-19 and SARS, we found differences between these diseases and further determined the optimal treatment regimens regarding sex-and age-speci c effects, doses, and treatment timing. These ndings appear to indicate that glucocorticoids should be prescribed at a low dose or pulse use but not at early taking (within ten days from illness onset to glucocorticoids therapy) for COVID-19 patients; however, this situation is complementary to that of SARS. Furthermore, sex-or age-speci c effect in COVID-19 patients should be noted when considering this treatment. Old age or man was usually related to severe illness and results may be affected by confounding effects of disease severity. However, when strati ed according to severity, sex-or age-speci c effects still existed across different severity levels. Thus, we infer that sex and age may be severity-of-illness-independent factors for the risk of death. We believe our ndings would bring light to the current clinical practice in glucocorticoids treatment of COVID-19.

Conclusion
The ndings suggest that glucocorticoids treatment reduced all-cause mortality and mortality-including composite outcome both in COVID-19 and SARS patients of critical severity. However, further bene t subgroups for COVID-19, compared with SARS, should be those who were critical but not only severe, with severe ARDS but not mild ARDS, of critical severity and man but not women, of critical severity and old but not young, taking low-dose or pulse but not medium-high dose, and non-early but not early taking.

Declarations
Ethics approval and consent to participate Not applicable Consent for publication Not applicable Availability of data and materials Additional data available from the corresponding author on reasonable request at kangyan@scu.edu.cn.

Competing interests
All the authors declared no competing interests.

Funding
This work is supported by the Project of Novel Coronavirus Pneumonia in West China Hospital (HX2019nCoV027). The study's funders had no role in the design and conduct of the study, data collection, data analysis, data interpretation, or writing of the report.    Tables   Due to technical limitations, table 1 is only available as a download in the Supplemental Files section. Figure 1 Preferred Reporting Items for Systematic Reviews and Meta-Analyses (PRISMA) for the Article Selection Process