We searched for surveys that estimated the prevalence of psychotic illness, major depression, personality disorder, alcohol dependence, and substance dependence in homeless people, published between January 1966 and December 2007. We searched computer-based literature indexes (EMBASE, MEDLINE, PsycINFO), scanned relevant reference lists, searched relevant journals by hand, and corresponded with authors. For the database search, we used combinations of keywords relating to psychiatric illnesses (e.g., mental*, psych*, depress*, substance/drug*/alcohol* abuse/dependence, personality) and being homeless (e.g., homeless*, roofless, shelter*). Non-English articles were translated. MOOSE guidelines were followed (Text S1).
For inclusion into the systematic review, the studies had to meet the following criteria: (1) A clear definition of homelessness was included; (2) standardized diagnostic criteria for psychiatric disorders using the International Classification of Diseases: Classification of Mental and Behavioural Disorders (ICD) or Diagnostic and Statistical Manual of Mental Disorders (DSM) were used; (3) psychiatric diagnosis was made by clinical examination or interviews using validated diagnostic instruments; (4) prevalence rates for psychiatric disorders within the previous six months were included, except for personality disorder where lifetime diagnoses were used; (5) study location was in North America, Western Europe, Australia, and New Zealand.
Surveys with less than a 50% response rate were excluded, as were surveys on selected populations (for example, a sample of homeless people referred to a psychiatric outpatient clinic or single mothers [24
]) or where diagnosis of psychiatric disorders was not obtained by direct interviews (for example, by self report or case note review) or was only reported as 12-mo or lifetime diagnoses [16
]. Studies that selected solely elderly or juvenile people were excluded [30
]. All the included reports were based on interviews with individuals. We identified one study that interviewed families, but this was excluded as it was based on a selected sample [32
Information on geographical location; year of interview; definition of homelessness; method of sample selection; sample size; average age; diagnostic instrument; diagnostic criteria, type of interviewer, participation rate; and numbers diagnosed with psychotic illness, major depression, personality disorder, alcohol dependence, and drug dependence was extracted independently from every eligible study. If required, further clarifications were sought by correspondence with authors of relevant studies.
Prevalence estimates were calculated using the variance stabilising double arcsine transformation [33
], because of the use of the inverse variance weight in fixed-effects meta-analysis is suboptimal when dealing with binary data with low prevalences. In addition, the transformed prevalences are weighted very slightly towards 50% and studies with zero prevalence can thus be included in the analysis. Confidence intervals (CIs) around these estimates were calculated using the Wilson method [34
] since the asymptotic method produces intervals which can extend below zero [35
]. Heterogeneity among studies was estimated using Cochran's Q
(reported with a χ2
-value and p
-value) and the I2
statistic, the latter describing the percentage of variation across studies that is due to heterogeneity rather than chance [36
], and presented with a 95% CI [37
, unlike Q
, does not inherently depend upon the number of studies considered, with values of 25%, 50%, and 75% taken to indicate low, moderate, and high levels of heterogeneity, respectively. Where heterogeneity was high (I2
> 75%), random effects models were used for summary statistics [36
]. In situations with high between-study heterogeneity, the use of random effects models (where the individual study weight is the sum of the weight used in a fixed effects model and the between-study variability) produces study weights that primarily reflect the between-study variation and thus provide close to equal weighting. The use of the arcsine-transformed prevalence estimates consequently had little material difference on the value of the overall random effects estimates, which were themselves found to be notably different (closer to 50%) from the fixed effects estimates in which smaller prevalences have smaller standard errors and thus greater weight. Potential sources of heterogeneity were investigated further by arranging groups of studies according to potentially relevant characteristics and by meta-regression analysis. Factors examined both individually and in multiple variable models were instrument (semistructured instrument versus clinical examination only), interviewer (conducted by mental health clinician or not), period (decade of study: 1970s, 1980s, 1990s, and 2000s), study size (both a continuous variable and as a categorical variable in increments of n
= 50 and n
= 100, and also as n
< 200 versus n
≥ 200), sex (as appropriate, e.g., male versus female; mixed versus male versus female), geographical region (as appropriate, e.g., mainland Europe versus rest of Western countries), and participation rates (classified into those >85% and those ≤85%). Because of low prevalence and small sample size for some studies, only those factors significant (p
< 0.10) individually were entered into a multiple regression model to avoid model instability. The regression coefficients for each study characteristic on individual analysis were provided to enable comparison across diagnoses. All analyses were done in STATA statistical software package, version 10 (Statacorp, 2007) using the commands propcii (to calculate Wilson CIs), metan (for random effects meta-analysis, specifying either two or three variables: double arcsine transformed prevalence and its variance, or double arcsine transformed prevalence and Wilson CIs), and metareg (for meta-regression).