The Lifetime Risk of Suicide in Schizophrenia: Title and subTitle BreakA Reexamination FREE

In 1977, Miles¹ reviewed 34 studies that observed people with schizophrenia and estimated a 10% lifetime risk of suicide. Cited more than 270 times, this study has informed the writers of most major psychiatry and suicidology texts.^{2 - 6} In addition, a widely cited 1990 review by Caldwell and Gottesman⁷ references 2 long-term schizophrenia follow-up studies by Bleuler⁸ and Tsuang,⁹ wherein 12.8%⁸ and 10.1%⁹ of the deceased had died by suicide, to argue that the rate is 10% to 13%.

In his chapter on suicide in TheNew Harvard Guide to Psychiatry, Tsuang¹⁰ cites his 10% figure⁹ but notes that these suicides compose only 4% of his total sample of 195 Iowans with schizophrenia because less than 40% of the sample had died. The difference in these numbers exemplifies the difference between proportionate mortality (PM) (the percentage of the dead who died by suicide, 10% in this example) and case fatality (CF) (the percentage of the original sample who died by suicide, 4% in this example). Proportionate mortality provides information only about the dead. In most of the cohort studies we reviewed, a large proportion of the subjects are still alive, having survived the initial period of high suicide risk and having not yet entered the period of increased risk of death from other causes. Thus, the direct use of PM rates, as in the review by Caldwell and Gottesman,⁷ assumes a constant rate of suicide over a lifetime and will therefore overestimate suicide risk.

Miles¹ attempted to correct for this fact by fitting a curve to the declining percentages of deaths by suicide over time but was limited, in part, by the technology of the time.¹¹ However, even more sophisticated statistical modeling cannot account for the fact that beginning with PM calculations requires that both the decreasing rate of suicide over the schizophrenic illness¹² and the increasing rate of other causes of death (ie, heart disease and cancer) must be corrected for. That is, both the numerator and the denominator (suicides and deaths) change over the duration of follow-up.

Case fatality rates look at the number of suicides as a fraction of the total sample. While these will increase over a lifetime of follow-up, extrapolation requires correction of only the numerator, based on timing of suicides within the illness course. Particularly in schizophrenia, where suicides are known to concentrate early in the course of the illness,^{12 - 14} CF is a preferable starting point. Such an approach has precedent in the literature.¹⁵

In addition to a consideration of statistical approaches,cohort composition will also affect an estimate of lifetime suicide risk. In his analysis of excess mortality inschizophrenia, Brown¹⁶ notes that cohorts of first-episode schizophrenics have 2.7 times the excess mortality of samples of chronic schizophrenics and that suicide is the largest contributor to this excess. First-episode cohort studies thus provide the most accurate suicide risk assessment, as other cohorts will have already survived the period of greatest mortality.

What is needed, then, is a meta-analysis that distinguishes cohorts observed from illness onset or first admission from cohorts taken from chronic samples or samples of all-comers. It must also begin with CF calculations. We believe the combination of these 2 approaches yields the most accurate estimation of suicide risk in schizophrenia to date.

Our analysis begins with strict inclusion criteria based on rigorous standards that parallel those of Harris and Barraclough.¹⁷ Studies included for meta-analysis have a mean follow-up time of at least 2 years because of the known increase in suicide risk following hospitalization.^{18 - 19} Consistent with other analyses, we require less than 10% loss to follow-up.^{15 ,17} This minimizes the chance that unaccounted-for subjects also represent unaccounted-for suicides²⁰ and increases the confidence with which conclusions can be drawn from the data.²¹

Included studies document the number of suicides during the observation period and provide an overall cohort size; by this standard, studies documenting no suicides can be included.^{22 - 23} Studies with a systematic age bias are excluded, so studies limited to adolescents are not considered. Studies of persons diagnosed as having schizoaffective illness are excluded, although it is possible that persons in the included cohorts had an undocumented comorbid affective illness. Finally, a distinction is made between studies that observed chronic or mixed chronic and new-onset patients (all-comers) and those that specifically observed patients from first hospital admission, documented illness onset, or first psychiatric contact.

The articles we cite cover nearly 90 years and include diagnostic criteria expressed by Bleuler,²⁴ Feighner et al,²⁵ and multiple iterations of the International Classification of Diseases and DSM systems. Studies used the accepted criteria of their time.

Studies were obtained through review of the bibliographies of the relevant review articles^{7 ,16 - 17 ,26} and textbooks^{2 - 6 ,10} as well as a computer search of the MEDLINE (1966-present) and PsychInfo (1984-present) databases. We exhaustively searched the bibliography of each article obtained, searching for additional references to cohorts of schizophrenics observed for any reason and repeated this process with each subsequent set of studies until all bibliographic references of interest were obtained. In this manner, we reviewed well over 600 articles. Several of the articles included would seem to have nothing to do with suicide but still fulfilled inclusion criteria.²⁷ Data from each article selected for inclusion were extracted independently by 2 of us (B.A.P. and J.M.B.).

Thirty-two studies of schizophrenics composed of all-comers and 29 studies observing schizophrenics from first admission or illness onset met inclusion criteria. Of note, no studies consisting exclusively of outpatient schizophrenics met inclusion criteria.

Data extracted from each study include only the number of patients, number of deaths, number of suicides, and length of follow-up. Survival status and length of follow-up for each subject would be preferable but are unavailable. Only deaths classified as suicides by the study in which they are reported are categorized as such.

For each of these 61 studies, 2 probability estimates were calculated. The first probability estimate is the CF prevalence, the probability that a subject will die by suicide during the course of the study (number of suicides divided by number of subjects). The second is the PM prevalence, the probability that a subject will die by suicide given that the subject will die during the course of the study (number of suicides divided by number of deaths in the study).

Each of these 2 probabilities, the PM prevalence and CF prevalence, along with 95% confidence intervals (CIs), were calculated using the generalized estimating equations²⁸ capabilities found in PROC GENMOD, a part of the SAS statistical package.²⁹

The suicide mortality prevalence was determined by assuming that CF rates increase over time to an upper limit and that PM rates decrease to a lower limit that is the same as the CF upper limit. This common limit is an estimate of the theoretical point at which CF will equal PM—when all subjects have died—and is an estimate of the lifetime risk of suicide.

The regression model for the relationship between PM and length of follow-up was log_e[PM / (1 − PM)] =a −e × exp(−f × years of follow-up). This describes an exponential decay to the limit a, on a logit scale, with e and f positive constants describing the baseline value and the rate of decay, respectively. The regression model for CF as a function of follow-up length was similar: log_e[CF / (1 − CF)] = a + b × exp(−d × years of follow-up). This describes a limited exponential growth to the limit a, on a logit scale, with b and d positive constants describing the baseline value and the rate of growth, respectively. The parameter a was fit to the 2 models simultaneously and represents the limit to which the PM and CF measures converge as follow-up time approached infinity.

These nonlinear regression models were fit to 2 clusters of data. The first was extracted from articles considering only subjects who were identified early in the course of their illness. The second was extracted from articles describing collections of patients identified at any point in their illness. A modified regression model was required for the CF curve approximating the data collected from the second set of studies. The CF data from these articles would not support a 3-parameter model. In this alternate model, the b parameter was constrained to be equal to the a parameter: log_e[CF / (1−CF)] = a × [1 + exp(−d × years of follow-up)]. This had the effect of maintaining the same general form of the regression model, a limited growth from a baseline value to an upper limit, by constraining the approximate baseline CF estimate to be smaller than the final upper limit.

The 2 nonlinear regression curves were fit to the PM and CF estimates extracted from the various articles using weighted least squares, as implemented in SAS’s PROC NLIN.²⁹ The weights used in the estimation procedure were equal to the number of subjects for the CF measurements and the number of deaths for the proportionate measurements.

The Table summarizes the studies selected for inclusion and documents the CF, PM, follow-up time, and sample size of each study.

For first-admission and new-onset samples, the generalized estimating equation yielded a CF estimate of 4.9% (95% CI, 4.3%-5.6%) and a PM of 30.6% (95% CI, 19.0%-49.1%). Among mixed samples of chronic and new-onset subjects, the CF rate was 2.3% (95% CI, 1.5%-3.5%), and the PM was 5.6% (95% CI, 1.4%-20.4%).

Nonlinear regression curves included data points weighted by sample size as shown in Figure 1 for illness onset/first-admission cases and Figure 2 for studies of mixed samples. Because point sizes are proportional to the log of the sample sizes, small samples are given disproportionate visual weight, although this artifact does not affect regression equation calculations. For the CF portion of the model, the weights were equal to the number of cases, and for the PM component, the weights were equal to the number of deaths (the denominators to each of the individual estimates). For the samples of all-comers, the originally proposed regression model for the CF data was overparameterized, as there was little evidence from the data for increasing CF rates as a function of time. As described in the “Methods” section, a slightly modified version of the model was employed in this case. The estimates of suicide mortality prevalence from the regression models (a transformation of parameter a) were 5.6% (95% CI, 3.7%-8.5%) for first-admission and new-onset samples and 1.8% (95% CI, 1.4%-2.3%) for samples of all-comers.

To analyze the impact of disparate diagnostic systems, we compared the CF rates of samples diagnosed exclusively with recent criteria (DSM-III⁸³ or later, International Classification of Diseases, Ninth Revision⁸⁴ or later, or Research Diagnostic Criteria⁸⁵ ) with those diagnosed with older diagnostic systems. Of the 12 first-admission studies that used the new criteria (n = 10 375), the mean ± SD CF rate was 4.9% ± 3.2%; this is not significantly different from the remainder of the first-admission studies (n = 12 223) with a mean ± SD CF rate of 4.8% ± 2.4%. Among samples of all-comers, 15 studies used new criteria (n = 10 382) and showed a mean ± SD CF rate of 2.3% ± 3.1%, identical to the mean ± SD CF rate of 2.3% ± 1.8% seen in the remainder of the studies (n = 15 196).

This study is a meta-analysis of suicide risk in subjects with schizophrenia that draws on pooled data from diverse sources. Two points emerge robustly. First, it is possible to obtain an approximation of lifetime suicide risk by estimating PM and CF curves as a function of average follow-up time and evaluating the level to which the 2 curves converge. Second, first-admission and new-onset studies more accurately estimate suicide risk because they include the initial years of the illness when death by suicide is most likely.

Proportionate mortality estimates of lifetime prevalence are accurate only when (1) all subjects are observed until death or (2) suicides occur at the same rate over time, relative to the total number of deaths. The first condition is difficult to meet; no studies in our review observed all subjects until death. The second defies what we know about suicide in schizophrenia and causes of death in general. Young patients, early in the course of schizophrenia, are more likely to commit suicide than older patients; moreover, older patients are much more likely to die of other medical conditions. Thus, estimating suicide risk relative to total deaths returns an artificially high estimate for all subjects, particularly young persons. Such errors are particularly true for studies with a relatively short length of follow-up. Seven such studies showed a PM prevalence of suicide of 100%.^{31 ,51 ,62 ,65 - 67 ,71}

Case fatality estimates do not suffer from this bias, since they allow individuals who have not yet died to provide information concerning the probability of suicide during follow-up. To be sure, CF prevalence risks missing future suicides and might thus provide a low estimate of lifetime prevalence. However, PM prevalence misses both future suicides and future deaths, and because suicides tend to occur earlier on in the course of follow-up, PM prevalence suffers from a more severe bias.

The difference in suicide risk between the group of studies restricted to first-admission and new-onset cohorts and the group composed of all-comers is striking. Indeed, the rate among cohorts observed from a beginning illness point (5.6%) is more than 3 times that of cohorts observed from any illness point (1.8%). Not only does this fact necessitate that future cohort studies begin as close to illness onset as possible to ensure accuracy and completeness, it also highlights the significant effect of suicides concentrated early in the course of the illness. The finding is more dramatic when considering the fact that among first-admission and new-onset samples, 30.6% of the deaths were due to suicide, while only 4.9% of deaths were suicides in samples of all-comers. This fact confirms the findings of several studies^{32 ,53 - 55 ,86 - 88} and implies that suicide-prevention interventions are best directed at persons early in the course of their schizophrenic illness.

A number of potential sources of bias and several limitations must be noted. The principal limitation is that the estimates of lifetime risk of suicide presented here are based on composite data from collections of studies. The lack of specific follow-up times for individual subjects in the various studies makes it impossible to answer the research question using the preferred analytical approach, Kaplan-Meier, or other explicit survival analysis techniques. Our method of combining the only data that are available, CF and PM rates, across the various studies attempts to resolve this limitation, which unfortunately cannot be fully resolved in such a meta-analysis. Therefore, the estimates reported here are best viewed as approximations to the risk of suicide in subjects with schizophrenia. A second limitation to the modeling used to obtain the estimates of suicide risk is that, while the data conformed reasonably well to the mathematical models, the estimate of lifetime suicide risk in individuals identified soon after their diagnosis is based on an extrapolation beyond the observed data. The impact of these 2 analytical limitations is impossible to evaluate without a large long-term study, or studies, of subjects newly diagnosed as having schizophrenia.

Additional limitations are evident in this analysis. Only 2 of 61 studies^{22 - 23} reported the number of suicides as zero. Several other published cohorts of people with schizophrenia made no mention of suicides occurring during follow-up. Without an explicit statement about suicide number, we elected not to include them in the meta-analysis. This decision may have artificially increased the suicide prevalence we calculated. Moreover, this analysis applies only to cases that came to clinical attention in certain treatment systems. It is not clear how many persons with schizophrenia spectrum psychoses fail to come to clinical attention (or deny symptoms or refuse research participation), but it is clear that the rate of suicide among these persons cannot be known. We also do not address in this meta-analysis such diagnoses as schizoaffective disorder and the ongoing debate in the descriptive psychiatry literature about whether this condition is an affective variant of schizophrenia or an entity in its own right. Future research must attend to not only the question of whether schizoaffectives are different from schizophrenics but also whether the group of schizophrenics itself is composed of subgroups with varying degrees of risk.

The analysis includes only suicides classified explicitly as such, excluding accidental deaths and other deaths of uncertain cause. Excluding these ambiguous cases likely only slightly diminishes suicide counts. Coroners over the last century have been shown to apply consistent standards of classification⁸⁹ and to underestimate suicide by exclusion of accidental/ambiguous deaths at a consistent rate of 15% to 20%.⁹⁰

While the grouping together of first-admission and new-onset studies implies uniformity in identifying subjects at illness onset, this is probably not the case. When the schizophrenic process actually begins has been a subject of intense academic scrutiny for at least a century; first admission to psychiatric care is certainly an artifact of convenience. In the era of de-institutionalization, many persons may be managed exclusively as outpatients, and at least 1 study suggests that these persons commit suicide at a higher rate than their admitted counterparts.³⁶ First hospital admission is the closest approximation presently extractable from the literature as a surrogate for illness onset.

Diagnostic criteria vary in the different samples, although included samples typically meet the diagnostic criteria used at the time of the study. All were published in peer-reviewed journals as studies of schizophrenics. Studies of schizoaffective patients or patients with other mood psychoses were not included. At least 1 study has shown no relative change in suicide rates with changing diagnostic criteria,¹⁵ although this remains an obvious source of potential bias. To evaluate this, we compared CF rates among studies utilizing newer diagnostic criteria (DSM-III or later, International Classification of Diseases, Ninth Revision or later, or Research Diagnostic Criteria) with those using older or inconsistent criteria and found that the rates were nearly identical in each group.

The most accurate estimate of suicide risk would be obtained by maintaining follow-up on all individuals under study, rather than the aggregate data available in the articles. One approach to account for the nonconstant risk of suicide during follow-up is to apply regression models to describe the relationships between PM and CF prevalences as a function of the length of follow-up. With information about individual cases, it would be possible to clarify the association between age at schizophrenia onset and age at suicide. Do schizophrenics diagnosed at younger ages, for example, have a higher risk of suicide? Given that schizophrenia is typically diagnosed in the third decade of life, our data with its preponderance of suicides soon after diagnosis would suggest that suicide in this population is skewed toward a younger age.

In summary, this analysis reveals 3 central ideas. First, CF rates more accurately approximate suicide prevalence than PM rates, and an even more accurate prevalence estimate may be approximated with statistical reconciliation of these 2 statistics. Second, suicide in schizophrenia will occur in approximately 5.6% of subjects over their lifetimes. Third, first-onset and first-admission studies give the most accurate picture of the natural history of the disease, including the fact that suicides are concentrated early in the illness course. Intervention and prevention efforts are therefore best directed toward the early stages of the illness.

Correspondence: John Michael Bostwick, MD, Department of Psychiatry and Psychology, Mayo Clinic, 200 First St SW, Rochester, MN 55905 (bostwick.john@mayo.edu).

Submitted for Publication: December 19, 2003; final revision received June 4, 2004; accepted July 7, 2004.

Funding/Support: This study was supported by a grant from the American Foundation for Suicide Prevention, New York, NY.

Previous Presentations: Earlier versions of this study were presented at the American Academy of Suicidology; April 10-13, 2002; Bethesda, Md; and the Ninth European Symposium on Suicide; September 14-17, 2002; Warwick, England.

Acknowledgment: We thank R. Theresa Steele and Frances Cockrum, MLS, and Patricia Erwin, MLS, of the Mayo Clinic library staff for assistance in identifying and locating studies used in this analysis.