Risk Factors for Suicidality in Patients With Schizophrenia

A Systematic Review, Meta-analysis, and Meta-regression of 96 Studies

Ryan Michael Cassidy; Fang Yang; Flávio Kapczinski; Ives Cavalcante Passos

Disclosures

Schizophr Bull. 2018;44(4):787-797. 

In This Article

Methods

Guidelines

The present study was registered at PROSPERO (CRD42015027027). We utilized the guidelines described by the Preferred Reporting Items for Systematic Reviews and Meta-Analysis (PRISMA) statement.[13] The PRISMA checklist we employed can be found in supplementary table 1 and the flow diagram in figure 1. Further, since we examined observational studies, we also structured our selection criteria and statistical analysis in concordance with the MOOSE guidelines for meta-analyses and systematic reviews of observational studies.[14] There are 3 methodological categories of observational studies which we chose to include: prospective and retrospective cohort studies where patients with the risk factor (in this case, schizophrenia) are followed until an outcome occurs (suicidal behavior) within a certain time frame; case-control studies, where patients are included based upon an outcome and matched by demographic characteristics to another patient without that outcome; and, cross-sectional studies, where patients with a risk factor are assessed a single time for presence of the outcome.

Figure 1.

Flowchart of review process and study selection.

Search Strategy

We searched Pubmed, Web of Science and EMBASE with the following search terms: ("Schizophrenia" OR "Schizophrenias" OR "Schizophrenic Disorders" OR "Schizophrenic Disorder" OR "Dementia Praecox") AND ("Suicide" OR "suicides") AND ("Risk Factors" OR "Risk Factor") from January 1, 1960 to December 18, 2016. Medical Subject Heading (MeSH) terms were used in PubMed and Emtree terms were used in Embase. These were loaded onto EndNote online to remove duplicates. We also searched the reference lists of included studies. Two researchers (R.M.C. and F.Y.) independently screened and selected titles and abstracts for full-text inclusion, with disagreements mediated by I.C.P. who made the final decision. We sought out translated versions of articles in languages which none of the authors spoke. We did not search the grey literature, due to the large dataset we predicted acquiring with the above search strategy.

Selection Criteria

Inclusion criteria were met if an article reported the following: a dichotomous sample of patients with schizophrenia with suicidality (suicidal ideation, attempted suicide, or suicide) compared to schizophrenic patients without suicidality; at least 70% of the sample had a diagnosis of schizophrenia using either International Classification of Disease or Diagnostic and Statistical Manual of Mental Disorders criteria. All 3 kinds of observational studies (cohort, cross-sectional, and case-control) were included. Review articles and studies performed on children and adolescents were excluded. If 2 or more studies reported the same risk factor within the same data set, we included only the more recent one.

Data Extraction

Each article was reviewed and the information compiled into an Excel workbook. We used the online version of EndNote to remove duplicate data. The following variables were extracted; first author; publication year; primary psychiatric diagnosis; diagnostic tool; study design; suicidality (suicidal ideation, attempted suicide, or suicide); continent; latitude of city where study was performed; number of patients in each group; sex; age; race; antipsychotic use (typical vs atypical); physical comorbidity; rural inhabitance; employment status; child status; marriage status; whether the patient lived alone; family history of suicide, psychiatric disorders, alcohol abuse, depressive symptoms, and schizophrenia; personal history of attempted suicide, suicidal ideation, depressive symptoms, alcohol abuse, tobacco use, and illicit drug use; insight; current delusions; current hallucinations; presence of flat affect; psychomotor agitation; worthlessness; hopelessness; poor compliance with treatment; aggressiveness; age of onset of illness; duration of illness; number of psychiatric hospitalizations; intelligence quotient (IQ); years of education; PANSS scores; Hamilton Depression Rating Scale (HAM-D); Beck Depression Inventory (BDI); Newcastle-Ottawa Quality Assessment Scale (NOQAS) score. For illness length, data were converted to years if reported as months. Child status was ultimately excluded as only one study reported this variable.

Authors were contacted for missing information, for mean and SD of continuous variables if they were reported as median and interquartile range, and for separated outcomes data if the study grouped 2 outcomes together (suicide ideation and suicide attempts, or suicide attempts and suicide).

Statistical Analysis

We began with meta-analysis of any risk factor which was reported in 2 or more studies, utilizing the metafor package (Version 1.9–8) in R (Version R 3.3.1) and R Studio (Version 0.99.902). A random-effects model with restricted maximum-likelihood estimator was used to synthesize the effect size across studies. This model incorporates both within-study variability and between-study variability.[15,16] The OR was used to assess the effect size for categorical risk factors because of the inclusion of case-control studies in the analysis.[15] The standardized mean difference (SMD) was used to assess the effect size for continuous risk factors.[15] SMD was calculated by use of Cohen's d. The significance level for this meta-analysis model was 0.05.[15] For SMD, an effect size of 0.2 is considered a low effect, whereas 0.5 a moderate effect and 0.8 or more a large effect.[17]

Heterogeneity, Bias, and Quality Assessment

Two researchers (F.Y. and R.M.C.) rated each article with the NOQAS to evaluate the risk of bias and quality of the study.[18] Disagreements in score were resolved with discussion between F.Y., R.M.C., and I.C.P. Egger's linear regression test was used to assess for asymmetry of the funnel plot in any case where 3 or more studies were included.[19] If the resulting P value is less than .1, we assumed asymmetry which may indicate publication bias. To account for this, we then employed the Duval and Tweedie's trim and fill method and reported whether a significant effect persists. We also used the leave-one-out function[15] for sensitivity analyses. This method consists of the removal of one study at a time from the dataset to run the meta-analysis. This analysis tests if the effect size of the meta-analysis is driven by one study. We used the Q statistic to test the existence of heterogeneity and I2 to assess the proportion of total variability due to heterogeneity.[20] An I2 value of about 25% could be regarded as low, about 50% as moderate, and about 75% as high.[20] We used τ2 to estimate the total amount of heterogeneity.[20]

Meta-regression Analysis

We further explored sources of heterogeneity in studies using meta-regression analysis.[21,22] For each risk factor with a significant effect size associated with suicidality, we performed univariate meta-regression with each of the following moderators: mean age of the total sample, mean length of illness of the total sample, region where the study was performed, study design (cross-sectional, cohort, or case-control), latitude of the city where the study was performed, and NOQAS score. To correct for repeated sampling, we employed the Knapp & Hartung adjustment.[23] When 2 or more moderators accounted for a significant amount of heterogeneity, a multivariate meta-regression with these moderators was conducted.[15] The pseudo-R 2 statistic represents the variance accounted for by the model; this indicates what percentage of all heterogeneity is accounted for by the confounding effect of the covariates. Cohort studies explore the relationship between exposure and outcome and aid in identification of causal associations. Therefore, we ran a supplemental meta-analysis only with the prospective and retrospective cohort studies in order to provide further evidence of the robustness of the risk factors identified through this meta-analysis.

processing....