Introduction

Previously, both individual-level socioeconomic status and neighbourhood-level social environment have been linked with schizophrenia incidence rates. However, findings were inconsistent [1–3] which may be related to the fact that even though individual and neighbourhood characteristics are highly interrelated, the two concepts are rarely included in the same set of analyses.

For example, individual-level socioeconomic status has been associated with schizophrenia rates [1–5]. The direction of this association was inconsistent with associations in the direction of either low socioeconomic status [2, 3, 5–8] or high socioeconomic status (before the onset) [4, 9]. These early studies, however, did not differentiate between individual and neighbourhood socioeconomic status. Instead, socioeconomic status at the area level was considered a proxy for individual-level socioeconomic status and studies referred to each other without distinguishing between individual-level and area-level measures.

More recently, the importance of area characteristics for schizophrenia risk was demonstrated in the study of, for example, urbanicity in relation to schizophrenia [10–15]. However, the mechanisms whereby urbanicity impacts on schizophrenia risk remain unknown. Seven decades ago, social isolation was hypothesised to have a causal influence on the development of psychotic symptoms [5], and this is supported by recent research [15]. A related mechanism has been proposed involving social capital, hypothesising that reduced trust, cohesion and social control between neighbours may be associated with higher risk of schizophrenia [16]. Social capital can be seen as the “glue that holds society together” [17]. It has been defined as “those features of social organisations—such as networks of secondary associations, high levels of interpersonal trust and norms of mutual aid and reciprocity—which act as resources for individuals and facilitate collective action” [18–20]. The measurement of neighbourhood social capital relies largely on the subjective assessments of the residents of the neighbourhood social environment, but the objective measure of residential instability can be seen as a proxy, because social capital is lower when neighbours move more frequently [21–23].

There is a growing interest in disentangling effects of individual and neighbourhood characteristics, using new statistical methods such as multilevel regression analyses. Previously, analyses including both individual and neighbourhood factors have been performed using both subjective (quality-of-life) and objective (mental health service use, crime) mental health outcomes [21, 24–27]. For example, one multilevel study including individual socioeconomic status showed that socioeconomic deprivation as well as lower levels of social capital were associated with higher rates of treated psychiatric disorders, but only before adjustment for individual-level confounders [25]. In addition, this study reported that social capital, given the presence of a psychiatric disorder, was associated with more contacts with mental health services. However, because differentiating between psychiatric disorders is essential [28], results may be different when studying schizophrenia. To our knowledge, only one recent study on neighbourhood factors and schizophrenia included a measure of individual-level socioeconomic status [29], while many others did not [30–35].

For the present study, data on patients with a diagnosis of schizophrenia were selected from a Psychiatric Case Register [36]. These data and data pertaining to population controls (unmatched case–control design) were matched with one measure of individual socioeconomic status and four neighbourhood measures: one measure of socioeconomic deprivation and three measures of social capital (informal social control, social cohesion & trust, residential instability). The present paper addressed four research questions:

  1. 1.

    Is individual-level socioeconomic status associated with treated incidence of schizophrenia?

  2. 2.

    Are neighbourhood variables associated with treated incidence of schizophrenia?

  3. 3.

    Is individual-level socioeconomic status associated with the level of outpatient service consumption and number of inpatient days in patients diagnosed with schizophrenia? (Quantity of service consumption).

  4. 4.

    Are neighbourhood variables associated with the level of outpatient service consumption and number of inpatient days in patients diagnosed with schizophrenia? (Quantity of service consumption).

Methods

Maastricht is a relatively small Dutch city with a population of 122,000 inhabitants (60 km2) living in 36 residential neighbourhoods (administrative units) of approximately 3000 inhabitants each (range 300–8500; all ages). The boundaries of these neighbourhoods follow main roads and are ecologically meaningful.

Cumulative incidence of schizophrenia

Since 1981, the Psychiatric Case Register South Limburg (PCR; formerly known as MHCR) has cumulatively collected data on all mental health contacts in Maastricht and surrounding areas (i.e. contacts with psychiatric hospital, community mental health centre, psychiatric department of university hospital, community psychiatric emergency outreach team, psychogeriatric nursing homes, sheltered housing, child psychiatric services, services for the mentally impaired or alcohol and drug misuse services). These collected data were chronologically ordered per patient using a probability linkage procedure involving five variables: gender, date of birth, first letters of the (maiden) name, place of birth and place of residence. Individual demographic (gender, age, marital status, education, employment and living conditions) and diagnostic data (e.g. schizophrenia, alcohol addiction and affective disorders) were also registered. For the present analyses, all patients diagnosed with schizophrenia (DSM IV 295, 297) living in the city of Maastricht were selected. All patients who were diagnosed with schizophrenia were included, both as first-ever diagnosis and after a diagnostic change (e.g. from early diagnosis of depression to later diagnosis of schizophrenia; see [30]). Within the city of Maastricht, all distances can easily be covered by bicycle and, therefore, effects of the distance to psychiatric services were negligible. Mental health services are covered by a national insurance and a general practitioner referral is not necessary for attending a community mental health centre. Individuals with no known or erroneous address were excluded from the analyses. In order to provide descriptive statistics per neighbourhood, neighbourhood-level 10-year cumulative incidence (1993–2002) of all cases aged 15 years and over was determined. Incident cases were individuals who did not have registered contacts with mental health services for at least 5 years. The mean size of the population, the denominator of the incidence fraction, was estimated by averaging the population sizes at 1-1-1995 and at 1-1-2001 (≥15 years) [37].

Relative risks of neighbourhood factors on schizophrenia

In order to analyse treated schizophrenia incidence rates, all incident cases (between 1-1-1998 and 31-12-2002) with the diagnosis of schizophrenia, aged between 20 and 65 years, were compared with a population control group (case–control design, see below for details of analysis).

In order to obtain measures of social capital from each neighbourhood, approximately 200 inhabitants aged between 20 and 65 years were randomly selected using the municipal database [38]. These residents received a questionnaire on social capital (described in detail below), but the questionnaire also included demographic data so that the respondents (48%, n = 3469) could serve as a control group for the current analyses according to an unmatched case–control design.

Quantity of service consumption

All incident cases, aged between 20 and 65 years at first contact between 1-1-1988 and 31-12-1997, were followed for 5 years to determine service consumption: number of days in hospital and number of outpatient contacts. Since the service consumption variables were highly skewed to the left, the scores were transformed using the natural logarithmic function.

Neighbourhood variables

The address at the first contact was used to define patients’ neighbourhood of residence. The measures of neighbourhood socioeconomic deprivation and residential instability were based on various neighbourhood socioeconomic characteristics obtained from the Maastricht Statistics Department and Statistics Netherlands (CBS). In addition, the above-mentioned population control group originally served to collect neighbourhood-level measures of social capital in all neighbourhoods. Social capital was measured using two dimensions: informal social control (ISC) and social cohesion and trust (SC & T). These two collective efficacy scales were developed by Sampson and colleagues [22] and adapted to the Dutch situation (see appendix). The two sum scores were obtained from individual answers and aggregated to the neighbourhood level. The ISC scale measures the willingness to intervene in hypothetical neighbourhood-threatening situations, for example in the case of children misbehaving. The SC & T scale measures bonds and trust among neighbourhood residents. All four neighbourhood variables were standardised to unity standard deviation. Higher scores indicated more socioeconomic deprivation, more residential instability and lower levels of social capital.

Individual-level socioeconomic status variables

Individual demographic variables included in the PCR data, such as marital status, education, employment and living conditions, are proxies for individual socioeconomic status. In addition, an extra variable on socioeconomic status was added to the PCR data, using postal code house-price registrations [25]. Postal code areas (as used by the postal services) are defined by 6 characters, and the Maastricht residential neighbourhoods each contain between 27 and 186 postal codes (mean 82), and thus a postal code area includes only a few addresses. Therefore, measures at postal-code-level can be used as a proxy for individual-level variables. To estimate postal code socioeconomic status, house prices per postal code between 1998 and 2002 were collected from the website of the Dutch land registry organisation. The prices per year were divided by the mean price per year (to control for increase of the house prices over the years) and these were used as an index of postal code socioeconomic status. This index was not available from postal codes with rented houses only (43% of all residential postal codes). In order to obtain one proxy for both owner occupied and rented houses, house price index was estimated in postal codes with rented houses only using a regression model with all available data on rented houses, such as rent of the houses, mean number of floors and percentages of houses with a garden.

Statistical analysis

Data were grouped according to neighbourhood and were, in statistical terms, part of a multilevel structure with level-one units (individuals) structured into level-two units (neighbourhoods). These hierarchically structured data were subjected to multilevel regression analysis [39] in order to investigate neighbourhood effects while controlling for individual effects. Multilevel or hierarchical linear and logistic modelling techniques are a variant of the more often used unilevel linear and logistic regression analyses and are ideally suited for the analysis of clustered data, in this case consisting of multiple persons clustered within a single neighbourhood. The βs (linear) and the odds ratios (logistic) are the regression outcomes of the predictors in the multilevel model and can be interpreted identically to the estimates in unilevel analyses.

First, the case–control data were subjected to multilevel logistic regression analyses using MLwiN (second order Penalised Quasi Likelihood method; PQL) [40–42]. Since we sampled the same number of controls in every neighbourhood, data of controls were weighted to reflect the distribution of the general population per neighbourhood. The individual variables gender, age group (20–30, 30–40, 40–50, 50–65), marital status, educational status (high, low), employment status (employed, unemployed), living conditions (not single, single, other) and individual socioeconomic status index (in quintile groups) were included in the analyses. Socioeconomic status index as well as the other individual-level variables had between 8 and 30% missing values. Therefore, the missing values of all individual-level variables were imputed, based on regression scores of the other individual-level variables, using the Stata statistical programme [43].

All categorical variables were recoded into dummies with the following reference categories: aged 20–30 years, married, high education, employed, not single and high socioeconomic status. This resulted in the following fixed effects multilevel logistic regression model:

$$ \eqalign{\ &{\rm In} (\rm odds) \cr &= {\beta_{0} \, + \, \beta _{1}} \, {{\rm neighbourhood \ variable}_{j}} \cr &+ {\beta _{2} \ {{\rm gender}_{ij}}} \, + \, {\beta _{3 - 5} \ {\rm age \ dummy}} \ {(1 - 3)_{ij}} \cr &+ {\beta _{6 - 8} \ {\rm marital \ status \ dummy}} \ {(1 - 3)_{ij}} \cr &+ {\beta _{9} \ {\rm low \ education}_{ij}} \, + \,{\beta_{10}} \ {{\rm unemployed}_{ij}} \cr &+ {\beta _{11 - 12} \ {\rm living \ conditions \ dummy}} \ (1 - 2)_{ij} \cr &+ {\beta _{13 - 16} \ {\rm socioeconomic \ status \ dummy}} \ (1 - 4)_{ij} \cr &+ {\mu _{j}}} $$

The ln(odds) was the odds of being registered with the diagnosis of schizophrenia in the PCR (treated incidence). Due to collinearity (e.g. Pearson correlation between socioeconomic deprivation and SC & T: 0.89, n = 35), neighbourhood variables (socioeconomic deprivation, residential instability, ISC and SC & T) were entered separately except for socioeconomic deprivation and informal social control (Pearson correlation 0.65, n = 35).

Second, analyses on level of inpatient and outpatient service consumption of patients diagnosed with schizophrenia were performed using Stata (version 8) [43]. All the above-described confounders were again added to the models and again the missing values were imputed. This resulted in the following fixed effects multilevel linear regression model (maximum likelihood method):

$$ \eqalign{&{\rm Outcome} \cr &= {\beta_{0} \, + \, \beta _{1}} \, {{\rm neighbourhood \ variable}_{j}} \cr &+ {\beta _{2} \ {{\rm gender}_{ij}}} \, + \, {\beta _{3 - 5} \ {\rm age \ dummy}} \ {(1 - 3)_{ij}} \cr &+ {\beta _{6 - 8} \ {\rm marital \ status \ dummy}} \ {(1 - 3)_{ij}} \cr &+ {\beta _{9} \ {\rm low \ education}_{ij}} \, + \,{\beta_{10}} \ {{\rm unemployed}_{ij}} \cr &+ {\beta _{11 - 12} \ {\rm living \ situation \ dummy}} \ (1 - 2)_{ij} \cr &+ {\beta _{13 - 16} \ {\rm socioeconomic \ status \ dummy}} \ (1 - 4)_{ij} \cr &+ {\varepsilon _{ij}} \, + \, {\mu _{j}}} $$

The outcome was the logarithmic function of inpatient days, or outpatient contacts, and the multiple error terms reflect residual variation at each level: individual (ε ij ) and neighbourhood (μ j ). Again, the four neighbourhood variables were added to the model separately, socioeconomic and ISC were entered jointly.

Results

Cumulative incidence and neighbourhood variables

Per 1000 Maastricht residents aged 15 years and over, 2.9 were diagnosed with schizophrenia in the last 10 years (1993–2002) and had contact with a mental health institution (thus 10-year total schizophrenia incidence was approximately 29 per 100,000 person years). The cumulative incidence per neighbourhood ranged between 0.7 and 6.0 per 1000 residents. Figures 1–3 show variation in neighbourhood socioeconomic deprivation, ISC, and neighbourhood 10-year cumulative incidence of schizophrenia, respectively. Table 1 presents descriptives of all neighbourhood variables.

Fig. 1
figure 1

Neighbourhood variation in socioeconomic deprivation

Full size image
Fig. 2
figure 2

Neighbourhood variation in informal social control (ISC)

Full size image
Fig. 3
figure 3

Neighbourhood variation in incidence rates of schizophrenia (DSM IV diagnoses: 295 and 297)

Full size image
Table 1 Means, standard deviations and Pearson correlation coefficients of neighbourhood variables
Full size table

Relative risk of neighbourhood factors on schizophrenia rate

Table 2 presents descriptives of cases and controls. Crude analyses showed that one standard deviation increase in one of the neighbourhood variables was associated with a 13–35% increase in treated incidence rates of schizophrenia (Table 3). The associations between social cohesion & trust and residential instability on the one hand and treated incidence on the other (OR = 1.29 and 1.35, respectively) were statistically significant, and the association between socioeconomic deprivation and treated incidence (OR = 1.19) was also apparent but statistically imprecise by conventional alpha. However, when controlling for individual-level variables, the associations between socioeconomic deprivation, ISC and SC & T on the one hand and treated incidence rates on the other were reduced and non-significant. The association with residential instability was statistically imprecise by conventional alpha (p < 0.1).

Table 2 Description of individual-level variables in cases and controls (controls weighted to reflect distribution of the populations per neighbourhood)
Full size table
Table 3 Odds ratio’s (OR) and 95% confidence Intervals (CI) of treated incidence of schizophrenia; multilevel logistic regression
Full size table

Measures of individual socioeconomic status and schizophrenia rate

Persons who were not married (divorced, single or widowed) and persons who were unemployed had a higher risk of schizophrenia treated incidence than married persons and employed persons, respectively. On the other hand, low education was negatively associated with treated incidence of schizophrenia (also when not including other individual-level variables in the model). The results presented in the table including all socioeconomic status variables showed that lower socioeconomic status based on postal code was associated with a lower schizophrenia risk in a non-linear way, but this could be the result of collinearity. When only age, gender and socioeconomic status based on postal code were included, the association was linear, and the risk of schizophrenia was highest in the lowest three categories (OR = 2.5, 2.2 and 2.0, respectively, p < 0.05, cf. the highest category, data not shown).

Quantity of service consumption

Between 1988 and 1997, 343 incident cases were diagnosed with schizophrenia. On average, these cases were 150 days in an inpatient setting for psychiatric care (range 0–91% of the 5-year period) and accumulated 46 outpatient contacts (range 0–277 contacts) in the 5-year follow-up period (i.e. 260 weeks).

Crude analyses and analyses controlling for individual-level confounders showed that none of the neighbourhood variables were associated with quantity of inpatient or outpatient service consumption in patients diagnosed with schizophrenia (Table 4). However, ISC was associated with quantity of inpatient service consumption when controlling for neighbourhood socioeconomic deprivation albeit statistically imprecise by conventional alpha (β = −0.45, p = 0.05): patients living in low ISC neighbourhoods (i.e. higher scores on the variable) had fewer inpatient days.

Table 4 Associations between social variables (at neighbourhood and individual-level) and service consumption (natural logaritm) of cases; multilevel linear regression, βs and 95% Confidence Intervals (CI)
Full size table

Patients aged above 40 years had fewer outpatient contacts than patients aged between 20 and 30 years. None of the individual socioeconomic status variables were associated with quantity of care consumption (Table 4).

Discussion

Treated incidence rates

Although crude analyses showed statistically significant associations between neighbourhood factors and treated incidence, all these associations were apparently explained by individual-level confounders. The results of socioeconomic disadvantage are in agreement with previous studies including a schizophrenia outcome [29, 32]. Studies on social capital and schizophrenia are scarce. An aggregate level study [33] reported a strong concentration of persons with schizophrenia in areas characterised by both socioeconomic disadvantage and weak social integration (social disorganisation). This is in agreement with results of our crude analyses, but aggregate level studies cannot take individual characteristics into account. A US multilevel study including individual socioeconomic status reported an association between neighbourhood residential instability and schizophrenia incidence rates [29], while the present results did not show such an association. This may be due to the fact that differences between Maastricht neighbourhoods are smaller [44, 45]. Unfortunately, studies on schizophrenia rates did not assess subjective measures of neighbourhood social capital. A previous Dutch multilevel study including all psychiatric service use (of which schizophrenia only represents a very small proportion) also did not show any effect of these social capital variables on treated incidence after controlling for confounders [25]. Thus, the present results do not provide evidence that neighbourhood social climate impacts on treated incidence of schizophrenia.

Individual-level variables were associated with treated incidence of schizophrenia: most factors related to the concept of socioeconomic status were risk factors, except for education that displayed an association in the opposite direction. Once patients are in contact with mental health services, there is no association between individual-level variables and the quantity of service consumption.

Thus, the present results suggest that it is the individual socioeconomic status that impacts on schizophrenia, while the neighbourhood social environment is only associated with treated incidence because it is a proxy of individual socioeconomic status. This suggests that the use of neighbourhood and individual-level socioeconomic status as interchangeable entities as was done in the early studies may be warranted [2–5]. In addition, associations with individual-level variables of socioeconomic status are inconsistent, even within the present study. Not only, low socioeconomic status, but also higher educational levels were associated with higher treated incidence. This may indicate a premorbid drift down the social ladder as a consequence of the developmental manifestations of the disease, as has been suggested previously [46]. However, schizophrenia is also associated with premorbid cognitive deficits that prevent individuals from reaching their potential [47, 48] making a primary association between schizophrenia and higher educational level improbable. One explanation may be that educational level may be particularly important in accessing services, such that lower educated patients may experience more barriers in accessing mental health care services [49].

Quantity of service consumption of patients diagnosed with schizophrenia

The results showed that the number of inpatient contacts was higher in neighbourhoods high in ISC (after controlling for socioeconomic deprivation). Previously, ISC has been associated with quantity of outpatient service consumption of psychiatric patients (all diagnoses) [25]. When the analyses of this previous paper were post hoc repeated excluding the patients diagnosed with schizophrenia, this yielded very similar results. Thus, while informal social control may favour more outpatient service consumption in patients not diagnosed with schizophrenia (who have little inpatient service consumption), it appears to contribute to more inpatient service consumption when the patient is diagnosed with schizophrenia (associated with more inpatient service consumption). These associations may be the result of interventions by neighbours, residents of neighbourhoods with higher levels of social control possibly seeking greater levels of resolution of psychiatric disorder in patient-residents, and by consequence inducing greater levels of service consumption before community reintegration may be attempted [25]. The present results support this hypothesis, with the type of care being dependent on whether the disorder is less severe (more outpatient service consumption) or more severe (more inpatient service consumption).

Methodological issues

The two domains for analysis included in the present paper, treated schizophrenia incidence rates as well as quantity of in- and out-patient service consumption of patients diagnosed with schizophrenia were measures of service consumption and as such cannot be taken to represent the full spectrum of clinical outcome measures. Therefore, the presented associations can be the result of an increase in psychopathology or of an increase in service consumption due to other reasons, such as changes in social circumstances. The present paper cannot disentangle these two possibilities; nevertheless, service consumption by itself is an important dimension of outcome.

Although the MQoL neighbourhoods are widely used and ecologically meaningful geographical units, neighbourhood residents may perceive different boundaries to their neighbourhood. Therefore, methods of a previous study [50] were adapted and respondents of a relatively small convenience sample of individuals living in Maastricht (n = 23), were asked to draw what they believed were the boundaries of their neighbourhood. Results showed that boundaries differ per person, and therefore perceived neighbourhoods cannot be used when studying the neighbourhood [51]. However, the not directly measurable, perceived neighbourhoods could be aggregated to the larger defined neighbourhoods, because boundaries of perceived neighbourhoods generally do not cross the boundaries of defined neighbourhoods. Associations in smaller areas will be even larger than the reported associations [51].

Based on previous research [38, 52], an interaction effect between neighbourhood socioeconomic deprivation and residential instability was hypothesised, and a recent study in which patients with all classes of psychiatric diagnoses were studied did show an association between socioeconomic disadvantage and quantity of service consumption in residential stable neighbourhoods only [25]. In the present analyses, analysing the subgroup of patients diagnosed with psychosis, there was no evidence for such an interaction.

The present paper has some limitations. First, because controls were interviewed in the year 2000, only cases that were incident between 1998 and 2002 (a 5-year period) were included in the analyses on treated incidence (i.e. 98 patients). Therefore, not finding any effect may be the result of low power. However, a previous case–control study in children including 56 cases and 206 controls did show effects of neighbourhood socioeconomic deprivation [53]. When calculating sample sizes in multilevel analyses, sample sizes of unilevel analyses need to be multiplied by a factor which depends on the intra class correlation and the number of observations per macro unit [39]. Using this multiplication factor, the unilevel equivalent of the sample size of the present study (3467) was calculated as 980. Unilevel analyses with a sample size of 980 and a prevalence in the data of 0.028 (98/3467) have a power of 90, 85, 80 and 70% to show an association when the odds ratio is 2.15, 2.0, 1.9 and 1.75, respectively [54, 55].

Second, incident cases were defined as the persons who did not have (known) contacts with mental health services for at least 5 years. Therefore, patients who had been in care somewhere outside the region who recently moved to the city of Maastricht were wrongfully classified as incident cases. This could have resulted in a slight overestimation of the incidence rates in all neighbourhoods. In addition, new Maastricht residents are more likely to settle in a neighbourhood with a higher residential turnover. However, it is unlikely that results are substantially biased, because migration rates are relatively low. In 2000, 5% of the population were new residents of the city of Maastricht. Since approximately one-fifth [56] of Maastricht new residents originate from the surrounding areas of Maastricht, the population that could have been misclassified as incident cases was approximately 4%.

For the first research question MLwiN was used because Stata has no routines to handle different weights within one macro level unit (controls were weighted to reflect the distribution of neighbourhoods, and cases all had a weight of 1). The MLwiN multilevel logistic regression procedure can model four different estimation methods. We chose second order PQL (Penalised Quasi Likelihood), because this method has been reported to produce unbiased estimates [41, 42].

Variance at the neighbourhood-level was low in all analyses. However, neighbourhood researchers tend to analyse neighbourhood effects, even when the neighbourhood variation is low, and it is generally held that this is warranted [57].

Finally, socioeconomic and demographic variables had missing values. The percentage of missing values was highest in the cases, because the care-givers who should supply this information were not always accurate, with the result that missing data status depended on the case–control status. However, we assumed that the data were missing at random (MAR). Therefore, we could impute the missing data using information on case–control status and all other individual-level variables available. Analyses where missing values were imputed and analyses were missings were deleted list-wise, yielded similar results. Additionally, sensitivity analyses were conducted using the Stata HOTDECK procedure. This procedure is used several times within a multiple imputation sequence since missing data are imputed stochastically rather than deterministically. For both research questions, five imputation sequences were run, yielding five data sets in which the analyses presented in the current paper were repeated. For the first analyses, the HOTDECK imputation was stratified by case–control status and age category; for the second research question the HOTDECK imputation was stratified by gender and marital status. Results were similar to the results presented in the current paper.