### INTRODUCTION

### METHODS

### Protocol registration and guidelines

### Eligibility criteria

### Information sources and search strategy

### Data management and selection process

### Data collection process

### Outcomes

### Data synthesis

^{2}statistic >50%), or if the Q test for heterogeneity was significant, the random effects results were considered. Network meta-analysis models were ran considering both random and fixed effects, the assumption of transitivity was explored in network graphs with at least one closed loop by estimating the inconsistency of each model using the Q statistic and the netsplit techniques (i.e., comparing the difference between indirect and direct estimates where closed loops were available in the network graph) if a significant degree of inconsistency was determined the results of the random effects model were reported. No assessment of inconsistency was possible for treatment comparisons without direct estimates. Treatment ranking was performed using the P-score technique and was represented in a forest plot with the pooled effect sizes of each treatment estimated with the network meta-analysis.

*Doing Meta-Analysis with R: A Hands-On Guide*was used for technical guidance throughout the analysis [23].

### Risk of bias of individual studies and certainty of the evidence

### RESULTS

### Search

### TEAEs

^{2}=63.7%), thus, the random effects model estimates were used for building the network. Network grafts are presented in Figure 2. Thirteen comparisons were available for the network model, and the resulting graph (Figure 2A) had direct comparisons between all interventions except for the one between aripiprazole-LAI (AriLAI) and RisLAI.

^{2}= 64.4%), the inconsistency of the network model (Q=11.9; p=0.003), or the treatment ranking. Excluding studies combining the use of OAPs and LAIs resulted in a fixed effects network model with a nonsignificant degree of inconsistency (Q=1.8; p=0.18), but with no significant change in the treatment ranking.

### STEAEs

^{2}=60.4%), thus, the random effects model estimates were used for building the network. Sixteen comparisons were available for the network model, the resulting graph (Figure 2B) did not have direct comparisons for AriLAI vs. RisLAI and haloperidol-LAI (HDLAI) vs. placebo.

^{2}= 62.3%), the inconsistency of the network model (Q=2.27; p=0.32), or the treatment ranking. However, excluding studies combining the use of OAPs and LAIs lowered the heterogeneity of the effect size estimation (I

^{2}=29.3%), had no significant effect on the inconsistency of the network model, but had a significant impact on the treatment ranking with HDLAI showing the best profile across treatments (OR=0.44, 95% CI=0.25-0.77).

### Deaths

^{2}=0%), thus, the fixed effects model estimates were used for building the network. Eleven comparisons were available for the network model, the resulting graph (Figure 2C) did not have direct comparisons for HDLAI vs. placebo, HDLAI vs. RisLAI, HDLAI vs. AriLAI, and AriLAI vs. RisLAI.

^{2}=0%), the inconsistency of the network model (Q=0.71; p=0.70), or the treatment ranking. Similarly, excluding studies combining the use of OAPs and LAIs had no effect on the effect size heterogeneity (I

^{2}=0%), the inconsistency (Q=0.08; p=0.78), or the treatment ranking.

### All-cause discontinuation

^{2}=82.1%), thus, the random effects model estimates were used for building the network. Seventeen comparisons were available for the network model, the resulting graph (Figure 2D) did not have direct comparisons for HDLAI vs. placebo, HDLAI vs. RisLAI, HDLAI vs. AriLAI, and AriLAI vs. RisLAI.

^{2}= 83.8%), the inconsistency of the network model (Q=0.88; p=0.64), or the treatment ranking. Excluding studies combining the use of OAPs and LAIs decreased the heterogeneity for the effect size estimation (I

^{2}=53.5%) and the inconsistency of the network model (Q=0.07; p=0.79), nonetheless, no relevant changes in the treatment ranking were found. Considering drug formulations the inconsistency of the network model increased slightly while remaining nonsignificant (Q=0.58; p=0.75 vs. Q=0.84; p=0.84), AriMhLAI was associated with lower rates of discontinuations compared to AriLxLAI (OR=0.40, 95% CI=0.31-0.52 vs. OR=0.45, 95% CI=0.33-0.60), similarly, RisRsLAI was superior to RisISMLAI (OR=0.48, 95% CI=0.39-0.60 vs. OR=0.65, 95% CI=0.43-0.98). Further inclusion of the different injection intervals had little effect on the inconsistency of the network model (Q=0.21; p=0.89), the administration of AriLxLAI every two months was associated with a lower rate of discontinuation compared to its administration every month (OR=0.37, 95% CI=0.20-0.66 vs. OR=0.48, 95% CI=0.34-0.67) and also when compared to AriMhLAI given every month (OR=0.37, 95% CI=0.20-0.66 vs. OR=0.40, 95% CI=0.31-0.52), similarly, the use of RisLAI given every two weeks was superior to the monthly RisLAI formulation (OR=0.48, 95% CI=0.39-0.61 vs. OR=0.65, 95% CI=0.43-0.98), however, only one study reported the use RisLAI monthly.

### Discontinuation due to AEs

^{2}=58.7%), thus, the random effects model estimates were used for building the network. Seventeen comparisons were available for the network model, the resulting graph (Figure 2E) did not have direct comparisons for HDLAI vs. placebo, HDLAI vs. RisLAI, HDLAI vs. AriLAI, and AriLAI vs. RisLAI.

^{2}=53.7%), the inconsistency of the network model (Q=1.86; p=0.40), or the treatment ranking. Excluding studies combining the use of OAPs and LAIs also had no effect on the effect size heterogeneity (I

^{2}=53.6%), the inconsistency (Q=1.45; p=0.23), or the treatment ranking.