Antipsychotic-Induced Weight Gain

Dose-Response Meta-Analysis of Randomized Controlled Trials

Hui Wu; Spyridon Siafis; Tasnim Hamza; Johannes Schneider-Thoma; John M Davis; Georgia Salanti; Stefan Leucht


Schizophr Bull. 2022;48(3):643-654. 

In This Article


Search Strategy and Selection Criteria

We followed the PRISMA guidelines (checklist in supplementary appendix 1) and registered a protocol in the International Prospective Register of Systematic Reviews (PROSPERO), registration number CRD42020181467 (supplementary appendix 2).[6]

We included all fixed-dose studies that compared the following drugs with placebo or at least one different dose of the same drug in adult patients with schizophrenia or schizoaffective disorder: amisulpride, aripiprazole [oral and long-acting injectable (LAI)], asenapine (oral and transdermal patch), brexpiprazole, cariprazine, clozapine, haloperidol (oral and LAI), iloperidone, lumateperone, lurasidone, olanzapine (oral and LAI), paliperidone (oral and LAI), quetiapine [immediate-release (IR) and extended-release (ER)], risperidone (oral and LAI), sertindole, ziprasidone, and zotepine. We planned separate analyses for the following patient subgroups: (i) chronic with acute exacerbation (ii) first-episode, (iii) elderly, (iv) predominant negative symptoms, but there were only eligible studies for the first and last subgroup. We excluded maintenance studies in stable patients (relapse prevention studies). In these studies patients are all pretreated. Often there are even run-in phases during which patients are stabilized on the drug in question before they are randomized to staying on the drug or switching to placebo/another drug. This procedure would limit the additional weight gain in the randomized phase. The inclusion of such studies would thus lead to methodological and clinical heterogeneity.

We searched the Cochrane Schizophrenia Group's study-based register of trials for studies comparing at least two doses of SGAs or haloperidol until March 9, 2020 and ran a final PubMed search on June 14, 2021. We inspected the reference lists of our previous systematic reviews on the acute efficacy of antipsychotics.[2,7–10] For these reviews, we had undertaken exhaustive searches including multiple electronic databases, the medical reviews that pharmaceutical companies must submit to the FDA, the reference lists of other meta-analyses of second-generation antipsychotic drugs (SGA),[11–18] Cochrane reviews comparing SGAs and haloperidol versus placebo,[19–21] and Cochrane reviews on optimum SGA doses,[22,23] and we had sent requests to the manufacturers of the SGAs (Search strategy in supplementary appendix 3). There were no language restrictions except studies from mainland China for which major quality concerns have been raised.[24] Two reviewers (SL, HW) examined reports independently. Risk of bias was assessed with the Cochrane Risk of Bias tool 1.[25] All data were extracted in duplicate and independently by HW and SL.

Data Analysis

We conducted a one-stage dose-response meta-analysis[26] in a frequentist framework using restricted cubic splines with the R package "dosresmeta" developed by Crippa and Orsini.[27] We investigated the relationship between dose (independent variable) and weight gain (dependent variable or response). In pharmacology, such a relationship is usually called dose-response, although the term could be confusing, because we actually address whether weight gain is dose-related. The primary outcome was mean weight change from baseline to endpoint using mean differences (MD in kilogram, kg) as the effect size measure. The secondary outcome was the number of patients with weight gain, preferably defined as at least 7% increase from baseline, analyzed with odds ratios (OR). When the numbers of participants with weight gain were not reported, we imputed them from mean scores using a validated imputation method and a cut-off of 7% increase from baseline.[28] Considering that there was no clear difference between long-acting and oral formulations of antipsychotics regarding the risk of weight gain,[29] we pooled the available different formulations for each drug by converting them to oral equivalents (supplementary appendix 5). This allowed a more comprehensive synthesis of evidence and increased the precision of the results. Nevertheless, we conducted sensitivity analyses for the different formulations. As in our previous analyses,[2] knot locations at the 25th, 50th, and 75th percentiles were used. For asenapine and sertindole, knot locatons at the 10th, 50th, and 90th percentiles were used, because the former quantiles could not form three unique knot points for these two drugs. For drugs with enough data, we used the Wald test to assess overall dose-response association and reported the P-values. Alpha was set at two-sided 0.05.

Sensitivity analyses of the primary outcome were performed excluding studies that compared one single-dose of an antipsychotic with placebo and studies in treatment-resistant patients. Posthoc we also analyzed: different formulations of drugs separately; standardized mean difference (SMD) as effect size for risperidone LAIs and lurasidone (two studies[30,31] with Body Mass Index (BMI) data instead of weight change couldn't be analyzed otherwise); pooled data in the 2009 FDA sertindole clinical review from three individual studies[32–34] that included 8 mg/d data (not reported in the individual studies); dose-response relationship of weight gain rate (weight gain divided by study duration, kg/week). In a posthoc sensitivity analysis of the secondary outcome, we excluded studies with imputed number of patients with weight gain.

Heterogeneity in the dose-response meta-analysis was quantified with the variance partition coefficient (VPC) which is a multivariate extension of the I2 value suggested by Crippa et al.[26] Small-study effect and possible publication bias were explored with contour-enhanced funnel plots of the pairwise comparison of an antipsychotic versus placebo when at least 10 studies were available.[25] Pairwise meta-analysis was conducted using the R package meta v4.12–0.[35] Data analysis was conducted in R statistical software v4.0.3.[36]