Effectiveness and safety of janus kinase inhibitors in hospitalized patients with COVID-19: Systematic review and individual patient data meta-analysis of randomized trials

Primary analysis:All patients will be analyzed in the study group to which they were randomized (intention-to-treat principle). We plan to both apply an IPDMA two-stage and one-stage approach since we probably have a rare event situation for the primary outcome. Two-stage approach: In the first stage, we fitt appropriate regression models for each outcome (logistic binomial regression for binary, ordinal regression for ordinal, poisson regression for count, and cox regression for time-to-event outcomes), obtaining a separate treatment effect estimate and its variance for each trial. We adjust all models for age and respiratory support, using restricted maximum likelihood (REML) estimation, and Firth's penalisation correction in case of sparse data bias concern as per IPDMA recommendation. In the second stage, we combine the estimates across trials in a random-effects model, using the inverse-variance method, and apply the REML estimator for τ2, and the Hartung-Knapp Sidik-Jonkman approach to derive 95% confidence intervals. One-stage: We use an IPDMA one-stage model that most closely mirrors the described two-stage model applying a random treatment effect, stratification of the intercept and the prognostic factors by trial including trial-specific centering of these variables.To investigate potential effect modification solely based on within-trial information to avoid aggregation bias, we follow a two-stage approach. In each trial separately we added the effect modifiers in turn to the models as an interaction term with the treatment group while keeping the adjustment variables in the models, and then synthesise the treatment-covariate interaction estimates and their variances in the second stage as described above. Continuous effect modifiers are added as linear treatment interaction terms and we use the multivariable fractional polynomials interaction approach to explore non-linear relationships. We will plot treatment effects in each participant subgroup for each trial alongside a forest plot of the within-trial interactions. The credibility of sub-group effects, for which we find an interaction p-value smaller than 0.1, will be assessed, independently and in duplicate, using the Instrument for assessing the Credibility of Effect Modification Analyses (ICEMAN).Missing data:Missing data will be addressed with the corresponding study teams and, where possible, retrospectively collected. For the remaining missing data of important covariates (i.e., used for adjustment), we will use multiple imputations chained equation techniques. If an outcome is missing for an entire trial, we exclude this trial from the corresponding analyses. If an important covariate (i.e., used for adjustment) is missing for an entire trial, we will not impute any data for this covariate in the trial dataset.For the imputation of important covariates, we will create and analyze 100 multiply imputed datasets, separately by trial, using the default settings of the mice 3.0 package in R. The parameters of substantive interest will be estimated in each imputed trial dataset separately. According to the nature of the variable, linear regression, logistic regression, or ordinal regression will be used in the respective equations. We will pool the repeated estimates into the final estimate, by trial, using Rubin's rule.