Introduction

Comorbidity bias (also known as Berkson's bias) predicts that if the definition of a disease with multiple psychopathological dimensions is based on observations of patients in the hospital, spurious comorbidity of those psychopathological dimensions that independently increase the probability of hospital admission will become included in the disease concept [1, 2]. Accordingly, associations between the symptoms that occasion admission to the hospital will be much higher in in-patients than in out-patients. That comorbidity bias may operate in schizophrenia was suggested by an earlier general population study showing that the association between positive and negative symptom dimensions of schizophrenia was much higher in patients compared with non-patients with psychotic experiences [2].

Traditional concepts of phenomenology in schizophrenia and psychosis are generally based on in-patients, and therefore, may represent only a minor selection of the total phenotypic continuum[3–5]. Admission to the hospital is related to more complex psychopathology, increased levels of distress, and lower levels of functioning [6, 7]. Consequently, the recruitment of in-patients with a diagnosis of schizophrenia for the purpose of research will not be representative for the underlying population, and conclusions may lack external validity. Therefore, results should be interpreted carefully [8]. Similarly, psychopathological differences between in-patients and out-patients not only have implications for the schizophrenia concept but also for service provision [9]. If associations, correlations, and covariance between symptoms are significantly higher in in-patients than in out-patients, the cumulative effect of in-patient treatments on separate outcome dimensions will be interpreted as an enhanced symptom reduction and large intervention effects, while this interpretation may not be valid for out-patient settings. Anecdotal evidence suggests that mental health workers and policy makers are still unaware of these differences and their consequences.

On the other hand, some symptoms may preclude admission. For example, when depressive symptoms are prominent, a patient may not be perceived as a risk for his environment, and hospitalisation may not be indicated. By focussing on criteria generated in the current dominant concept of schizophrenia, important cues of prodromal or first-onset schizophrenia may be misconstrued as benign in light of the referenced clinical reality of severe psychopathology observed in hospital settings.

In the present paper, it was hypothesised that the clustering of symptom dimensions differs between in-patients and out-patients, with a higher level of clustering expected in in-patients. We studied the direction of any dimensional comorbidity bias in a longitudinal data set of the psychopathology of patients both inside and outside the hospital. Data were collected from a cumulative register of needs of in-patients and out-patients with psychotic illness in the defined geographical area of South Limburg, The Netherlands. First, associations between symptoms and current (cross-sectional) and future admissions (longitudinal) were studied. The cross-sectional analyses assessed whether psychopathology dimensions were susceptible to comorbidity bias, while the longitudinal analyses examined the possible causality of this association in terms of predictive power. Second, correlations between psychopathological dimensions in out-patients and in-patients were analysed, and to what degree these differed between the two groups.

Methods

Cumulative needs registration

The local Cumulative Needs for Care Register (CNCR, formerly called the Psychosis Protocol) [10] is an initiative of the mental health service providers, legislators, and local consumers to assess the match between clients' need for care and the level of service provision. All patients with a diagnosis of psychotic disorder according to the DSM-IV classification system [11] were included in the CNCR, and their psychiatrists, psychologists, and case managers were trained to administer various clinical scales (such as the Brief Psychiatric Rating Scale (BPRS), see below). Assessments are carried out routinely at intake as a yearly evaluation. In addition, an assessment is also carried out with every major change in treatment or setting (e.g. hospitalisation, start of a new psychiatric treatment, and discharge).

Hospitalisation and other measures

Admission status (in-patient or out-patient) was the dependent variable in the analyses. All interviews before 24 June 2002 were included.

Each CNCR assessment provides information on (1) demographic variables, (2) the BPRS [12, 13], (3) the General Assessment of Functioning (GAF) scales to assess levels of handicap and psychopathology [11], (4) the Camberwell Assessment of Need (CAN) [14], (5) a 5-item quality of life scale, and (6) quality of care (one item). The present paper focusses on dimensions of psychopathology from the BPRS as independent variables. Demographic variables were entered a priori as confounders in the analyses.

Based on previous BPRS research [15], a factor analysis was carried out. This confirmed the existence of four underlying constructs (hereafter, BPRS symptom dimensions): negative symptoms, positive symptoms, manic excitement, and depression/anxiety. Blunted affect, motor retardation, emotional withdrawal, and self-neglect loaded on negative symptoms; bizarre behaviour, unusual thought content, disorientation, hallucinations, and suspiciousness on positive symptoms; motor hyperactivity, elevated mood, excitement, distractibility, hostility, and grandiosity on manic excitement; and depression, anxiety, suicidality, and guilt on depression/anxiety.

Statistical analyses

All analyses were performed using Stata (version 7) [16]. First, logistic regression was used to study the association between symptom dimensions (included separately) and admission status. This was done both in cross-sectional analyses, using all observations at each time point in each person, and in longitudinal analyses linking BPRS symptoms at the first time point of each out-patient with hospitalisation at any future time point (yes/no). Second, Pearson correlation coefficients between the four symptom dimensions were calculated in the group of out-patients and in the group of in-patients, using the assessment at the first time point of each patient only. Finally, we performed linear regression analyses (using all observations at each time point in each patient), including an interaction term admission status × symptom dimension, to test the statistical significance of differences in associations between symptom dimensions in out-patients and in-patients.

Because patients were interviewed at more than one time point, yielding more than one observation per person (i.e. several records in the data set), the assumption of independence of the observations for standard linear and logistic regression analyses was invalid except for the longitudinal analyses (using a data set with baseline variables and future admission in the same record). Therefore, we performed linear and logistic regression methods ideally suited for the analysis of this data structure: multilevel regression analysis [17]. The parameters (βs obtained from multilevel linear regression analyses and odds ratios obtained from multilevel logistic regression analyses) can be interpreted identically to the estimates obtained from standard unilevel analyses.

Results

Descriptives

The data included 994 interviews in 480 persons, each person contributing between one and nine assessments between the start of the data collection in 1997 and June 2002. Sixty per cent of the population was male. At the first interview, 49.1% of the clients were hospitalised. The mean age of the out-patients was 39.8 years (men 36.3, women 45.3); the mean age of the in-patients was 43.8 years (men 40.3, women 49.2). BPRS symptom levels were 20–25% lower in out-patients, and these differences were statistically significant (Table 1).

Table 1 Descriptives
Full size table

Associations between symptom dimensions and hospitalisation

Both the univariate analyses and the analyses controlling for age and gender showed that all four BPRS symptom dimensions were associated with in-patient status (Table 2). Age was positively associated with hospitalisation (OR=1.02, p=0.02), but there was no association with gender (OR=0.99).

Table 2 Cross-sectional logistic regression analyses (multilevel)
Full size table

Longitudinal analyses showed that manic excitement in out-patients at baseline was a predictor for in-patient status at follow-up (Table 3). Positive symptoms also predicted future admission, albeit statistically imprecise by conventional alpha.

Table 3 Analyses using longitudinal data
Full size table

Associations between symptom dimensions in two settings

In in-patients, correlations between symptom dimensions were generally stronger than in out-patients (Table 4), reaching statistical significance for the association between positive symptoms and manic excitement (β=0.28, p<0.001; not in table). Notable exceptions were the correlation between positive symptoms and depression/anxiety (β=−0.16, p=0,002), and the correlation between manic excitement and depression/anxiety (β=−0.20, p=0.004), which was significantly lower in in-patients.

Table 4 Correlations between BPRS factors in out-patients and in-patients, first interview of each
Full size table

Discussion

Associations between symptom dimensions

Associations between symptom dimensions were generally higher in in-patients than in out-patients, suggesting the possible influence of dimensional comorbidity bias [2]. However, only one association (between positive symptoms and manic excitement) was significantly higher in in-patients than in out-patients. The correlation between positive symptoms and depression/anxiety, as well as the correlation between manic excitement and depression/anxiety, were higher in out-patients than in in-patients, although absolute symptom levels of both manic excitement and depression/anxiety were lower in out-patients. The results therefore suggest that the occurrence of mania and positive symptoms has a more depressive “colouring” in patients who are not admitted. Treatment trials and prevention studies should take into account this excess depression comorbidity associated with these symptom domains in out-patients. Thus, the concept of psychosis, in particular schizophrenia, needs cautious interpretation in non-clinical settings.

Previously, both positive and negative symptom dimensions were independent predictors of mental health care use, indicating higher associations between positive and negative symptom dimensions in patient groups as opposed to non-patients with sub-threshold psychotic experiences [2]. The present study, which did not compare patients and non-patients but in-patients and out-patients, did not show significant differences in associations between positive and negative symptoms in in-patients and out-patients. Instead, the associations between positive symptoms and manic excitement were significantly different in the comparison between hospitalised and out-patient groups. The different findings in the two studies were likely due to the fact that two different populations were studied. The general population study indicated that positive and negative symptom dimensions were most important for the initiation of any treatment [2], whereas the present study indicated that positive symptoms and manic excitement were important dimensions for the decision of admitting an already-treated psychotic patient to the hospital.

In addition, the present analyses showed that all four symptom dimensions were as relevant with regard to admission status. This makes the assumption that the concept of schizophrenia is a hospital-based diagnosis even more plausible. From a clinical perspective, it has been shown that differences in psychosocial needs are related to the dimensions of schizophrenia. More needs are associated with the disorganised and excitatory subtypes [18]. This is in agreement with our finding that manic excitement in subjects with a diagnosis of schizophrenia is associated with hospitalisation.

The longitudinal analyses also showed an association between manic excitement and future hospitalisation. In addition, there was also some evidence that a higher level of positive symptoms was associated with future hospitalisation status, albeit statistically inconclusive by conventional alpha. This can be due to a lack of power; only 30 of the 239 included out-patients were hospitalised during follow-up. Thus, results showed strong evidence that manic excitement is not only a reason for being admitted but also predicts future admission, and future analyses including more cases may also identify positive symptoms as a predictor of future admission.

Implications for intervention studies

Current research in psychosis focusses on the decrease of the duration of untreated psychosis, relapse prevention by early intervention programmes, medication compliance, cognitive therapy, and functional biological anomalies in patients with schizophrenia [19–22]. While many interventions focus on positive and negative symptoms, the present study suggests that comorbidity of positive symptoms and manic excitement may be relevant for admission to the hospital. Similarly, the assessment of change in positive and manic symptoms in a sample mixing out-patients and in-patients should take into account the comorbidity of these domains with depression/anxiety in out-patients, which may introduce underlying heterogeneity in response due to differential treatment effects across in-patient and out-patient groups. Prevention research and first-onset programmes should be particularly aware of the differences in diagnostic symptom patterns in clinical samples and the target population for these interventions.

Methodological issues

The longitudinal data collected with the CNCR interview provided more insight in the longitudinal admission risk of psychotic patients and the interaction between psychopathology and admission status. This is important both for treatment and research. However, this data set has some limitations. First, interviews should have been carried out at intake, as part of the yearly evaluation, and at every mutation in the patient's situation. However, due to pragmatic and logistic reasons, professional carers do not always comply exactly. In some patients, two interviews were more than 1 year apart. In addition, the CNCR assessment tended to be omitted when patients were admitted repeatedly or stayed for periods of less than 2 weeks. Since reminders are now routinely sent to all interviewers who do not turn in the yearly reassessment, the logistics of the CNCR have improved. This will affect the compliance rate for the yearly reassessments, but not for every change in treatment or setting.

Second, interviews were administered by different psychiatrists and psychologists in various settings. Although this can be regarded as a strength of the data set, it also introduced potential problems of reliability. To minimize assessment bias, (1) a manual was developed [10], (2) all participant interviewers were trained extensively, and (3) repeated booster training sessions were organised. Nevertheless, a subjective interpretation of the patients' complaints and morbidity can never be completely avoided. It is difficult to envisage, however, how this would have biased the current results.

[4], [20], [21]