*2.5. Data Synthesis and Statistical Analyses*

Pooled estimates and their confidence intervals (CIs) were obtained with a randomeffects meta-analysis model (DerSimonian–Laird estimator) [21] based on the assumption of substantial clinical heterogeneity. Rate ratios or relative risks (RRs) and 95% CIs were used to evaluate the efficacy and safety outcomes. If the desired effect estimates were inadequate for data synthesis, the corresponding authors were contacted to obtain relevant information. Statistical significance was indicated by *p* < 0.05 and a CI not containing 1. All statistical analyses were conducted using Stata v.16 (StataCorp, College Station, TX, USA).

A pairwise meta-analysis comparing the effect of JAK inhibitors against placebos was conducted. However, when a multiple-arm RCT was enrolled, a unit-of-analysis error may arise if the same group of participants is included twice in the same meta-analysis (for example, if "dose 1 vs. placebo" and "dose 2 vs. placebo" are both included in the same analysis, with the same placebo patients in both comparisons) [25]. Therefore, when an enrolled RCT included more than two arms, the reported results were combined to create a single pairwise comparison against placebos [21]. This method had been utilized in previous studies [26,27].

Between-study heterogeneity was quantified using Cochran's Q and I<sup>2</sup> statistics [28]. *p* < 0.01 and I<sup>2</sup> ≥ 50% indicated substantial heterogeneity. To identify the potential sources of heterogeneity, several subgroup analyses were conducted according to the administration route, AD severity, participant age, mode of action, and different JAK inhibitors at different time points. Meta-regression analyses were also performed to explore potential effect modifiers, which refer to the explanatory variables of the overall effect estimates that may contribute to the heterogeneity [21]. Visual inspection of the funnel plot and Egger's regression test were used to assess publication bias [21].

Concerning the possibility of producing false-positive results when using the DerSimonian–Laird method, we applied the Hartung–Knapp–Sidik–Jonkman (HKSJ) method for sensitivity analyses [29]. This method widens the CIs to reflect uncertainty in estimating between-study heterogeneity, especially when the study number is less than 20 [21,30]. The HKSJ method had been widely used in previous meta-analyses [31–33].

The GRADE approach was adopted to summarize the CoE at the outcome level [34] and judged by all authors. In the event of disagreement, discussions were undertaken to arrive at a consensus for each outcome. CoE, classified as high, moderate, low, or very low, refers to the confidence that the true effect lies in a particular range [34]. The CoE was downgraded by one level if a serious flaw was present in the domains of risk of bias, inconsistency, indirectness, imprecision, and publication bias [34]. The NNT and NNH were calculated to evaluate the evidence-based efficacy and safety of JAK inhibitors in the management of AD [34].

#### **3. Results**
