Statistical analyses
We conducted propensity score matching analysis to compare outcomes
between the two groups. A multivariable logistic regression model with
the following variables as covariates was used to estimate propensity
scores for receiving glucocorticoids on the day of admission: age, sex;
body mass index category, Japan Coma Scale score (alert, drowsy,
somnolent, and comatose)10; smoking status (never,
past and current smoker, and missing); diagnoses (T78.0, T78.2, T88.6);
Charlson Comorbidity Index score (0, 1, 2, and ≥3); history of asthma,
atopic dermatitis, and atopic rhinitis; use of histamine 1 blockers,
histamine 2 blockers and beta 2-adrenergic receptor stimulants; hospital
volume (very low, low, high, and very high); and teaching hospital. We
performed one-to-four nearest-neighbor matching with replacement for
estimated propensity scores, using a caliper width set at one fifth of
the standard deviation of the estimated propensity scores. To assess the
accuracy of the matching, we compared the covariates before and after
propensity-score matching using absolute standardized differences,
absolute standardized differences ≤ 10% being considered to denote
negligible imbalances between the two groups.11 After
propensity score matching, we assessed the outcomes through generalized
linear models, accompanied by cluster-robust standard errors with
hospitals as the clusters. We calculated odds ratios and their 95%
confidence intervals (CIs) with generalized linear models using the
logit link function.