Reply.

We thank the authors Duan et al for the interest in our study and their comments. The concerns raised primarily arise from the generalizability of results and the risk of bias from unmeasured confounders. We fully agree that our results should be interpreted with respect to the study population. As we have noted in our article, the heterogenous study populations from the studies included in the meta-analysis may partly explain their conflicting results.

First, we acknowledge that obesity is a well-established risk factor for poor outcomes including death in patients with COVID-19. Unfortunately, we do not have clinical data on height and weight from the applied registries. Possible confounding by obesity could be present because of an observed higher body mass index (BMI) among individuals with gastroesophageal reflux disease, a common indication for prescription of proton pump inhibitors (PPI). To explore the direction, magnitude, and uncertainty caused by obesity on our estimate of 30-day mortality we have conducted a quantitative bias analysis.

In our analysis we first assign reasonable values to the bias parameters including (1) the association between the confounder (obesity) and the outcome (mortality in COVID-19 patients), (2) the prevalence of the confounder in the exposed group (current PPI use), and (3) the prevalence of the confounder in the unexposed group (never PPI use). As the first bias parameter we used the pooled odds ratio of mortality from multivariate analysis in a meta-analysis on obesity (BMI ≥30) as a risk factor for severe COVID-19 outcomes: pooled odds ratio of 1.49 (95% confidence interval, 1.20–1.85). The proportions of individuals with obesity in the exposed and unexposed group were assumed based on data from another Danish cohort study on hospitalized patients with COVID-19, reporting a proportion of 8% of individuals with BMI >35. To align with the meta-analysis that used a BMI of 30 as cutoff, we set the prevalence proportions higher than the 8% reported with cutoff of 35 (ie, to median 0.20 with a range from 0.10 to 0.30 among PPI users and median 0.10 with a range from 0.05 to 0.15 among PPI nonusers, both with trapezoidal distributions).

Using these bias parameters, and assuming a true null-association between PPI use and adverse outcome, we performed 10,000 probabilistic bias analysis simulations. We found a median relative risk of 0.84 (95% simulation interval, 0.67–1.04). The simulation interval incorporates the uncertainty in the bias parameters and the random error from the original study estimate, adjusted relative risk of 0.88 (95% confidence interval 0.72–1.08). The quantitative bias analysis suggests that the estimate is biased slightly away from the null, thus supporting our conclusion of a null association from the original study.

Second, it would be interesting to obtain data on socioeconomic status and test whether the findings by Mena et al could be replicated in our cohort. However, we would assume that the impact of socioeconomic status on mortality would be less pronounced because Denmark is a less segregated society where all residents have tax-funded universal access to health care.

Third, based on the findings in the survey study by Perez-Araluce et al it is not possible to infer what effect a Mediterranean diet would have on COVID-19 outcomes because this was not examined. The authors observed a protective effect concerning risk of acquiring SARS-CoV-2 infection in a cohort of well-educated Spanish residents and only when restricting the analyses to non–health professionals. The generalizability of these results to our or other study populations is probably limited, being based on self-reported test results with unknown test type and by using a design inherently prone to selection bias. Therefore, we welcome more studies before considering diet a plausible confounder.

Fourth, the study by Burchill et al addresses the impact of the COVID-19 pandemic on the gut microbiome mainly in relation to the design of ongoing and future microbiome studies, accounting for direct and indirect interactions. Although we agree that the behavioral changes attributable to the pandemic are important to consider, we do not find these nor any gut microbiome changes within the scope of our study.

Despite the possibility of additional unknown confounders we believe our study results are relevant because they rely on a large sample size with nationwide data and adjustments are made for a wide range of known confounders. Because of the register-based design we do not have clinical data. However, when we assessed the impact of obesity as an unmeasured confounder the quantitative bias analysis showed that it did not influence our result greatly, thus supporting our null result.