For sensitivity analyses, we used inverse variance methods under fixed and random effects models for the outcomes with the largest number of treatment events; random effects models can be problematic for meta-analyses of rare events.21 We report subgroup analyses for the primary outcomes by age (<40 v ≥40), sex (<50% male v ≥50% male), ethnicity (<50% white v ≥50% white), presence or absence of psychiatric illness, smoking status (smokers including smokeless tobacco users v majority non-smokers (>60% non-smokers)), and whether or not the study was sponsored by a pharmaceutical company. Studies were not categorised as sponsored by a pharmaceutical company if the drug was provided at no cost by the manufacturer and/or if the research was investigator initiated—that is, the drug and some funding was provided by the manufacturer although there was no other involvement in study conduct or publication and data were independently held by the researchers. Tests for subgroup differences were performed. Funnel plot asymmetry was assessed for two outcome—depression and insomnia—with Harbord’s modified test for small study effects with the “metafunnel” and “metabias” commands in Stata.22 23
