George Kunos, Tony Jourdan, Joseph Tam
Sophie Lotersztajn, Ariane Mallat
Simeng Wang, Qingzhang Zhu, Guosheng Liang, Tania Franks, Magalie Boucher, Kendra K. Bence, Mingjian Lu, Carlos M. Castorena, Shangang Zhao, Joel K. Elmquist, Philipp E. Scherer, Jay D. Horton
Rory D. de Vries, Marieke van der Heiden, Daryl Geers, Celine Imhof, Debbie van Baarle, RECOVAC-IR Collaborators
Anthony T. Tan, Nina Le Bert, Antonio Bertoletti
Dear Editor, We read with interest the article by Li and colleagues on the association between ACE inhibitors/angiotensin receptor blockers (ACE-I/ARB) use and in-hospital mortality among COVID-19 patients (1). The authors concluded that the use of ARB was associated with a significant reduction in in-hospital mortality among African American (AA) patients but not non-AA patients. However, we believe this conclusion is not per statistical principles and potentially misguiding readers. As noted by Altman and Bland (2), statistical analysis should be targeted to the clinical question: is the association between ARB use and in-hospital mortality different between AA and non-AA patients? To answer this question, one should directly compare the estimates (interaction test) (2), performed and reported by the authors. Here we argue that they did not accurately interpret this analysis. The authors showed an odds ratio (OR) of 0.196 (95% confidence interval [CI], 0.074 – 0.516) in the AA population and an OR of 0.687 (95% CI, 0.427 – 1.106) in the non-AA population. Accordingly, the interaction term was non-significant (95% CI, 0.185–1.292; P = 0.149). As the authors stated that “Statistical significance was defined as a 2-sided P value less than 0.05”, the correct interpretation of this result would be: the association of ACEi/ARB use and in-hospital mortality was not significantly different between these two populations (2). In contrast to this interpretation, the authors concluded that the association was only present in the AA population, which is not compatible with their analysis. The potential association between ACEi/ARB use and COVID-19 in-hospital mortality is of great interest to the medical community. Further, the ability to provide reliable subgroup analyses is vital in clinical decision-making (3). Interaction analyses are essential to answer the clinically relevant question of whether a specific subgroup of patients can benefit more from an intervention. However, we believe the correct interpretation of these results does not support the author’s conclusion.
Arthur M. Albuquerque, Carolina B. Santolia, Ashish Verma
Dear Editor, We appreciate Albuquerque, et al.’s interest in our paper [1,2], who raised the concern that we did not accurately interpret the interaction test, noting that “one should directly compare the estimates (interaction test)” and “the authors concluded that the association was only present in the AA population, which is not compatible with their analysis.” We would like to clarify that our primary clinical question is whether ACE-I/ARB use is associated with the COVID-19 outcomes in each sub-group. We used stratified analysis to answer the question because when race/ethnicity serves as a non-specific proxy for numerous (confounding) factors, these can be (partially) controlled for through stratification . Joint modeling of multiple groups is often used to gain power, but one needs to assume certain coherent distributions across different groups, which is not always true. Additionally, testing the interaction term is to assess association heterogeneity between groups, but it does not directly address whether the treatment is effective in each group. Specifically, we would like to elaborate on the following: 1) Our conclusion: “the use of ARB was associated with a significant reduction in in-hospital mortality among African American (AA) patients but not non-AA patients” was based on results from the stratified analysis. We reported that ARB in-hospital use was associated with reduced mortality in the AA stratum (OR=0.196; 95%CI:0.074-0.516; P=0.001) with statistical significance. On the other hand, the association in the non-AA stratum is not statistically significant (OR=0.687; 95%CI:0.427-1.106; P=0.122). As stated previously, our primary objective is to assess whether ACE-I/ARB use among AA patients is associated with COVID-19 mortality, rather than the difference between AA and non-AA patients. We were also aware that the estimated ORs across different stratum were not comparable as noted in [4-6]. 2) We performed the joint modeling of AA and non-AA patients as suggested by . Here, ARB in-hospital use was associated with reduced mortality in entire study population (OR=0.560; 95%CI:0.371-0.846; P=0.006). The interaction term added to the model was not significant (95%CI:0.185-1.292; P=0.149). Interpreting interaction terms in logistic regression is complex and a significant interaction term in log-odds may not be significant in difference-in-differences for probability. Furthermore, the assumption of the additive effects and imbalanced sample sizes could impact the inference. We believe these results and the interpretation are appropriate. We acknowledge that there are cases where comparing the interaction term in greater detail would be an important next step for understanding the association between COVID-19 mortality and race/ethnicity.
Shilong Li, Pei Wang, Li Li
Manami Tsunoi, Sunao Iyoda, Tadayuki Iwase
Inès Ambite, Ulrich Dobrindt, Catharina Svanborg
To the Editor: Povysil G. et al. report that “rare loss-of-function (LOF) variants in type I interferon (IFN) immunity genes are not associated with severe COVID-19” (1). We disagree with the authors’ interpretation of our data and their own (2), for six reasons: 1) Only predicted LOF (pLOF) variants are relevant for comparison between the two studies, because, unlike us, these authors did not test variants experimentally. The relevant proportion in our data is therefore not 24/659=3.5%, but 9/659= 1.36%, whereas theirs is 1/713=0.14%. 2) Our definitions of ‘severe/critical’ patients are different: we defined critical disease as severity grades 6-10 of the WHO scale (3), whereas they restricted their recruitment to grades 7-10 (i.e., excluding patients on high-flow oxygen, considered in our study). Their cohort of ‘mild’ cases may therefore include ‘severe’ COVID-19 cases (grade 6), such as perhaps their ‘mild’ TLR3 pLOF carrier. 3) Their ‘controls’ are subjects from the general population, without depletion of COVID-19 genetic risk factors, whereas we used pauci-/asymptomatic infected subjects (grades 1-3) as ‘controls’. Consequently their power computation in Figure 1 is based on an incorrect hypothesis about the odds ratio, which would be expected to be lower when using general population controls (as they did), than when using pauci- and asymptomatic infected individuals (as we did). 4) The ethnic origin of the patients differs between the two studies: 58% of our 659 patients (and 8 of our 9 pLOF carriers) were European, versus only 10% of their 713 patients with severe disease (and their pLOF carrier is East Asian). 5) Age is a key factor neglected in their comparison: our sample was much younger (mean age: 51.8 years) than theirs (mean: 65.9 years), and seven of our nine pLOF carriers were < 60 years old. We performed a comparison stratified by age (<60/≥60 years), and no significant difference in pLOF proportion was found between the two studies, even ignoring the only patient carrying a pLOF they found (of unknown age): 7/458 in our sample vs. 0/192 in their sample (p=0.11, Fisher’s exact test) for patients <60 years old, and 2/201 vs. 0/521 (p=0.07) for patients ≥60 years old. 6) Finally, and crucially, the authors did not exclude patients with autoantibodies against type I IFN, which account for at least 10% of critical cases and are much more frequent in patients > 60 years of age, particularly men (4).
Qian Zhang, Aurélie Cobat, Paul Bastard, Luigi D. Notarangelo, Helen C. Su, Laurent Abel, Jean-Laurent Casanova
No posts were found with this tag.