The sample enrichment strategy basically consists of the creation of a sample with many people at risk by selectively filtering them out over a range of consecutive referral processes starting in the general population, through to general practioners (GPs) , mental health services, and the early detection clinic. The sample enrichment strategy is undoubtedly the most widely used approach in the early intervention literature, but possibly also the worst understood, in that the high predictive values obtained in sample enrichment studies are often wrongly attributed to some instrument with supposedly high predictive values, whereas in reality they are a function of the sample enrichment strategy itself. For example, several authors have suggested that the high “transition rate” to psychotic illness in individuals exhibiting psychosis-like symptoms is between 40% and 70%, thanks to the use of some prodromal-rating instrument,34,45,46,62,63
advocating the use of such instruments in order to reduce transition to full-blown illness. However, closer inspection of these data is required, as illustrated by the following example.
In a recent publication, Klosterkotter et al63
reported a follow-up study of 160 young individuals who were considered to be at risk of developing psychotic illness. The signs and symptoms used to predict transition to schizophrenia were from a list of “basic symptoms.”64
The presence of any of the baseline basic symptoms was used as a test to predict the onset of psychosis over a mean follow-up period of 9.6 years. The main results are presented in Table IV:
the risk of developing schizophrenia, given the presence of a basic symptom described by Huber et al,64
was 77/110 (70%). Therefore, these data apparently predicted the onset of schizophrenia over a 9-year period with 70% accuracy! The question, however, is whether this high predictive accuracy can be completely attributed to these basic symptoms, or whether instead other factors are more important. In reality, only a minor proportion of the predictive value can be attributed to the basic symptoms, because most can be ascribed to the very high baseline rate of schizophrenia in this sample. As can be seen in Table IV,
the final rate of schizophrenia in this sample was 79/160 (49%). The conclusion from this is that by chance alone, any subject in this study had a nearly 50% probability of developing schizophrenia anyway. In other words, the predictive value of any factor, whether it be basic symptoms, sex, color of the eyes, or any other random variable, would have been at least 50%. Thus, the role of the basic symptoms was to raise the predictive value modestly from 50% to 70%, not to fully cause a predictive value of 70%. Instead of the predictive value of basic symptoms, the real noteworthy element of the remarkable study by Klosterkotter et al63
lies in the fact that the authors succeeded in creating a series of sophisticated selection processes that led to a final enriched sample of individuals with a 50% probability of developing schizophrenia over the 9.6-year period. This selective enrichment process involved the existence of special interest groups at German university psychiatry departments, with an interest in young people who posed a challenge with regard to a possible diagnosis of schizophrenia. Such a center “attracts” a highly enriched sample of individuals at risk of schizophrenia through a series of selection processes, as illustrated in Table V
Individuals in the general population developing illness behavior visit the GR The GP refers those with suspected severe mental disorder to the general mental health services. The general mental health services refer those with suspected schizophrenia onto the specialist university department. With each referral from one level to the next, a selection process takes place creating “enriched” samples that are progressively more likely to contain individuals who are likely to develop schizophrenia. Other groups wishing to replicate the German findings in their own setting, must therefore not only use the basic symptom scale, but, much more importantly, replicate exactly the same sample enrichment strategy to yield a sample with a 50% probability of developing schizophrenia. In addition, rather than a posteriori, any additional contribution of basic symptoms to the predictive value needs to be replicated prospectively in a fresh sample at the start of sampling enrichment procedure. If the results hold after these replications, the basic symptoms,64
or instruments used in Melbourne, Australia,46
or New Haven, Conn,45
may possibly be used to modestly raise the predictive value from 50% to 70% in samples enriched with schizophrenia risk. In Table V
, the effect of using measures such as Basic Symptoms or other high-risk instruments at other levels in the sample enrichment procedure with more diluted samples and therefore lower rates of (future) schizophrenia is shown. The predictive values were calculated using the DIAGTEST procedure in the STATA statistical program, version 8,65
at various levels with their corresponding best estimate rates of schizophrenia. The DIAGTEST procedure in STATA provides the predictive values based on Bayes' theorem. As can be seen in Table IV
in the general population, the basic symptoms would not yield a positive predictive value (PPV) of 70%, but of only 1.4%, and at other levels in the sample enrichment procedure the predictive values were also much lower.
Thus, reports of high PPVs in schizophrenia on the basis of mental states or other risk factors can be attributed to the high baseline rates of schizophrenia in the samples used, created through a series of selection procedures in the sampling process34,45,46,63
or, sometimes, the statistical62
These selection procedures contribute to exciting and clinically relevant findings. However, almost invariably a large proportion of the predictive values are wrongly attributed to the various predictors used rather than to the selection procedures that resulted in higher prevalence or incidence rates of schizophrenia and hence a higher baseline predictive value. The consequence of this confusion is that clinicians attempting to use such findings in their practice may erroneously focus on the predictors rather than the sampling enrichment selection procedure itself. In addition, the high predictive values reported in the literature are often based on a posteriori optimization algorithms in the sample at the end of the sample enrichment filtering procedure. The correct procedure would be to demonstrate its validity a priori in a prospective investigation at the beginning of the sample enrichment procedure.
Of course, the sample enrichment strategy, similarly to the other strategies to raise the prevalence of schizophrenia that were discussed earlier, suffers from the limitation that only a tiny proportion of all detectable schizophrenia cases in the general population will be identified for early treatment in this way. The great majority of prodromal schizophrenia will never make it through the various selection procedures from the decision to visit the GP, subsequent GP referral to mental health outpatients, and from there to the specialized prodromal clinic- they will only come to the attention of mental health services after
having developed the first acute psychotic episode, not before.
Although it is true that prevention of even 1% of all transitions to schizophrenia would constitute an important result from the clinical viewpoint, one may nevertheless question whether specialized early intervention clinics can ever be made cost-effective, given competing demands for funding.67