The Iowa Record-Linkage Experience

The study of suicide has bedeviled clinicians and researchers alike. Because suicide is important from both personal and economic perspectives, particularly for psychiatric patients, multiple strategies have been developed over the past decades for its investigation. The ultimate goal of this research is to enable clinicians to predict—and, one hopes, to prevent—suicide. Some investigators have focused on retrospective “psychological autopsies” of suicide victims, including the now classic reports by Robins, Murphy, Wilkinson, Gassner, and Kayes (1959) and Barraclough, Bunch, Nelson, and Sainsbury (1974). Other strategies have included case-control designs (Roy, 1982; Winokur & Black, 1987) and prospective study designs (Helgason, 1964; Martin, Cloninger, Guze, & Clayton, 1985a,b; Rorsman, Hagnell, & Lanke, 1982). The latter are relatively uncommon because of the large numbers of patients who must be followed in order to witness the relatively rare suicide outcome, and the huge expense required to conduct such studies. On the other hand, another strategy has evolved to evaluate high-risk populations, such as psychiatric patients, using a historical prospective design (Black, Warrack, & Winokur, 1985a,b,c,d; Eastwood, Stiasny, Meier, & Woogh, 1982; Tsuang, 1978). This involves determining the outcome of persons at risk in an earlier decade. The determination of outcome can be performed in a variety of ways, such as making personal or telephone contact, or using the follow-up material found in available case notes. If the objective of follow-up is to study a predetermined outcome such as death, then the only concern may be whether the patient is deceased, and, if deceased, when the death occurred and what caused it. This type of study can be accomplished through a relatively simple record-linkage process, wherein the roster of patients is compared electronically with death certificates for a given period. This is merely one example of record-linkage study; other uses of record linkage are unlimited.

Historical Overview

Studies over the past 50 years have documented that psychiatric patients are at risk for excess mortality, particularly from suicide. These findings have been noted in many countries, in different decades, and in different groups, including both inpatients and outpatients (Babigian & Odoroff, 1969; Black et al., 1985a; Eastwood et al., 1982; Helgason, 1964; Innes & Millar, 1970; Martin et al., 1985a; Morrison, 1982; Roy, 1982). There has been a decline in mortality among psychiatric patients, due primarily to the decline in premature deaths from natural causes (Black, Winokur, & Nasrallah, 1987a; Martin et al., 1985b). In fact, premature death from natural causes may no longer be as significant a problem among psychiatric patients as it once was many years ago, particularly among the institutionalized. This may be due to improvements in the efficacy and availability of psychiatric treatments, better care of intercurrent medical problems, and the rarity of long-term hospitalization nowadays. Suicides and accidents continue to contribute to excess mortality, and there is recent evidence that suicide rates may be climbing, particularly in young white males (Weed, 1985).

The Iowa Record-Linkage Study

Few historical prospective studies of mortality in psychiatric patients have made age- and sex-adjusted comparisons with death rates in the general population. My coinvestigator, George Winokur, MD, had participated in the Iowa “500” study many years earlier, in which careful mortality studies were carried out with schizophrenic, manic, and depressed patients (Tsuang & Woolson, 1977). Because of the study design, patients in other diagnostic groups were not followed. We wanted to enlarge upon this experience and see how patients in all categories fared. Furthermore, because of secular trends noted by other investigators, we became interested in studying mortality in a relatively recent sample.

We began by obtaining a list of all patients admitted to our hospital over a 10-year period from 1972 through 1981. After duplicate names resulting from rehospitalizations were removed, our list consisted of 5,412 patients, including 2,933 (54.2%) women and 2,479 (45.8%) males. The record included multiple identifiers, including name, sex, date of birth, and hospital number. This information was then linked by computer with all Iowa death certificates for the same period. A total of 331 deaths were identified; all matches were confirmed by manually checking the list of deceased individuals against hospital admission records. The cause of death listed on the death certificate was recorded. Each patient was then assigned a single psychiatric diagnosis from a hierarchical and shortened list of 10 psychiatric categories based on the International Classification of Diseases, ninth edition (U.S. Department of Health and Human Services, 1980):

The diagnostic hierarchy was determined by Winokur, based on his clinical and research experience. We developed a computer program that reviewed each patient's clinical diagnosis (or diagnoses) made by the ward psychiatrist and picked the one with the highest priority in the hierarchy. For example, a patient with both schizophrenia and alcoholism would be assigned a research diagnosis of schizophrenia. In a case of multiple admissions, we used only information from the patient's last psychiatric hospitalization.

Our statistician (Giles Warrack, PhD) then constructed mortality tables adjusted for age, sex, and follow-up time among the patients, allowing us to compute expected numbers of deaths. We did not adjust for race, because so few Iowans (<3%) are nonwhite. The computations were based on Poisson's distribution, and the method has been fully described (Black et al., 1985a). Standardized mortality ratios (SMRs), representing the ratio of observed to expected mortality, were calculated and tested for significance using the chi-square statistic modified by Freeman-Tukey deviates. The Freeman-Tukey transformation was used because it is more conservative than the regular chi-square test, and many of our observed and expected numbers were quite small.

Findings from the Iowa Record-Linkage Study

Our findings are presented in Tables 6.1, 6.2, and 6.3. Relative risk for premature death was greatest among women and the young, particularly patients younger than 40 years. Risk was associated with all psychiatric diagnoses and was significantly higher among patients of either sex with an organic mental disorder or schizophrenia; women with acute schizophrenia, depressive neuroses, alcoholism, drug abuse, and psychophysiological disorders and special symptoms; and men with neuroses. Death from natural causes, especially from heart disease, was significantly excessive among women, while death from accidents and suicides was excessive for both men and women. The overall SMR was 1.65 (p < .001). Furthermore, and most importantly, we found that the greatest excess of mortality occurred within the first 2 years after hospital discharge. In fact, 188 (56.8%) of all deaths occurred during the early follow-up period, when only 59.8 (29.7%) had been expected. The data on excess deaths (observed death minus expected) showed that 98.7% occurred within the 2-year period; after that, the observed number of deaths approximated those expected.

Among findings specific to suicide, both men and women in all age categories except the senium (over 69 years) were at significant risk for suicide. Patients were at great risk for suicide throughout the entire follow-up period, although 54 (79.4%) suicides occurred during the first 2 years of follow-up. Patients in all categories were at significant risk for suicide, except for patients with organic mental disorders, adjustment disorders, psychophysiological disorders, and special symptoms; men with depressive neuroses; and women with neuroses or personality disorders. The overall SMR (men and women combined) was 21.20 (p < .001). Thus, we were able to demonstrate that risk of mortality in general, and risk of suicide specifically, differed according to age, sex, diagnosis, and portion of the follow-up.

Advantages and Disadvantages of the Iowa Record-Linkage Approach

Few research designs are infallible. The Iowa Record-Linkage Study is no exception, and its drawbacks involve patient sampling, death ascertainment, psychiatric assessment, and statistical inference. First, our sample included former adult inpatients hospitalized at a tertiary care facility. Although our sample was random in the sense that all patients in a given 10-year period were included (it was, in fact, a population), it was nonrandom in that we chose a highly select group of patients to study. Therefore, it is difficult to generalize our findings beyond the confines of our sample.

Next, death was ascertained using a record-linkage process that did not allow for personal follow-up. It is likely that our method underestimated the true number of deaths, as many patients would have emigrated from Iowa, changed their names, or died in other states, and follow-up data on them would thus not have been readily available. We determined that about 6% of our patients were from other states, but had no estimate on the number of persons who emigrated, changed their names, or died elsewhere, although it is generally recognized that Iowa's population is stable. Past studies (e.g., the Iowa “500”) have demonstrated that even after 30 or 40 years, most Iowans remain in Iowa and are traceable. We relied on death certificates for death determination—a practice often fraught with difficulty, as there is no standardization in assignment of death cause, or even in who fills a death certificate out (e.g., physician, mortician). The difficulty of this assignment is illustrated by the case of a living cancer patient dying of pneumonia. Should the death be recorded as due to cancer or pneumonia? Information about death from unnatural causes (e.g., accidents, homicides, suicides) is probably less idiosyncratic, but still rests on the coroner's verdict. A coroner's expertise may vary from place to place, and verdicts often reflect administrative, economic, or other constraints, not merely medical fact.

Third, our psychiatric assessment was not based on an interview with each subject or on careful chart review, but rested on chart diagnosis. Each patient was assigned a single diagnosis; multiple diagnoses, however important they may have been, were not allowed. Also, diagnoses were assigned in a hierarchical fashion, so that certain disorders preempted others. For example, organic mental disorders were ranked first and were followed by schizophrenia, acute schizophrenia, affective disorders, and so on (see list above). Naturally, a hierarchy of this nature is arbitrary and minimizes the contribution to mortality that multiple diagnoses may make. For example, a patient with both affective disorder and alcoholism would have received a study diagnosis of affective disorder, yet alcoholism clearly contributes to excess mortality. Other rankings may have produced other findings.

Fourth, although our total sample was large, many subsamples were quite small after the total group was divided into different categories (e.g., sex, diagnosis), and both observed and expected numbers of deaths within the subsamples were even smaller. Naturally, drawing a statistical inference from such small numbers may not be warranted, even with the conservative Freeman-Tukey modified chi-square. For example, two females with acute schizophrenia committed suicide when 0.03 were expected. This is statistically significant, yet one wonders whether such small numbers merit any inference at all. Lastly, because there were no chart reviews or personal interviews, relevant risk factors (e.g., demographic, clinical treatment, etc.) could not be evaluated for their possible influence on mortality.

Several advantages can be cited in defense of our approach. Our study sample included all persons within a 10-year period who were hospitalized, and in this sense, it was unbiased. These adults comprised a group at high risk for early mortality—a group of great concern to all hospital-based clinicians. Therefore, the findings from this study are relevant to that select group of patients. Death ascertainment may have been incomplete, but because Iowa's population is relatively stable, it was probably not too far off the mark. We had to rely on death certificates for assignments of death causes, and, regardless of errors, not much could be done about them. One strategy we used was to look at death from either natural or unnatural causes. By creating two categories, we probably minimized the influence of any idiosyncratic practices on determining specific natural causes of death, or of the administrative, legal, and political concerns that plague ascertainment in cases of unnatural death. For example, a suicide may have been misclassified as accidental for political reasons. By classifying it as unnatural, we were able to ignore such concerns. (On the other hand, this method ignores the differences between suicides and accidents.) Other investigators have found this dichotomy useful as well (Martin et al., 1985a, 1985b). Although our use of a single, hierarchical diagnosis has been criticized (Martin, 1985), most patients did have a single psychiatric diagnosis.

Advantages of our approach are apparent. Record-linkage studies are simple and economical; a large staff is unnecessary. Using this approach allowed us to work with relatively large sample sizes not possible with prospective follow-up study designs. We used a statistically sound, actuarial approach for calculating expected numbers of deaths based on U.S. census and Iowa vital statistics data. Our computer programs could easily be updated for future studies. Furthermore, the programs were adjusted for age, sex, and duration of follow-up. The last of these is essential in studies such as ours, where some patients are followed for 1 year and others may be followed for 10 years. Naturally, a person followed for 10 years would have a greater cumulative risk for mortality than someone followed for 1 year.

Iowa Depression Study

As an offshoot of our original study, we became interested in learning about specific risk factors associated with mortality in affective disorders. In our original study, we had created a broadly defined affective disorders category. We were unable to look at subtypes, because we lacked specific diagnostic information. Furthermore, our original design did not allow us to look at risk factors associated with mortality other than age, sex, and duration of follow-up. Therefore, we initiated the still ongoing Iowa Depression Study. We reviewed the charts of all patients admitted to our hospital over a 12-year period (1970 through 1981) who had an affective disorder. Included were all patients with chart diagnoses of major depression, bipolar disorder, schizoaffective disorder, depressive neurosis, secondary depression, and so on. Our final sample consisted of 2,054 charts. Each chart was individually reviewed by a master's-level psychologist (Amelia Nasrallah). Each chart was systematically evaluated for demographic, diagnostic, clinical, treatment, and outcome data. In all, over 120 bits of information were obtained on each patient. Next, we conducted a record linkage, comparing the roster of affective disorder patients with Iowa death certificates through 1983, allowing at least a 2-year and at most a 14-year follow-up. Each patient was carefully diagnosed according to three sets of operationalized criteria: DSM-III, Research Diagnostic Criteria, and the St. Louis criteria. Multiple diagnoses were allowed, and medical illnesses were coded as well. Thus, we were able to correct some of the deficiencies of the Iowa Record-Linkage Study, such as lack of chart review, enabling us to evaluate multiple risk factors and look at specific diagnostic groups.

Findings of the Iowa Depression Study

As the study is still in progress, only preliminary data are available (Black, Winokur, & Nasrallah, 1987a,b,c). We have evaluated death rates in unipolar and bipolar affective disorders. A total of 157 deaths occurred in the sample, but only 97.5 were expected, yielding an SMR of 1.61 (p < .001). We found that overall mortality risk differed between subtypes (see Table 6.4), but much of the difference was attributable to physical illnesses. Patients with bipolar disorder and secondary depression were more likely to have a serious physical disorder than primary depressives (partly because of the definition we used). After deleting patients with these conditions from the analysis, we found death rates between the subtypes to be similar.

We found little evidence for a difference in risk for suicide among depressive subtypes (see Table 6.5). Among patients with primary unipolar depression, 31.2% of the deceased died from suicide; corresponding percentages among secondary unipolars and bipolars were 26.3% and 16.7%, respectively. Although it appears that bipolars were at less risk for suicide, it became clear after further study that this finding was simply a result of combining the manic and depressed bipolar patients into a single group. As manies have a very low risk for suicide, and depressed bipolars have a risk for suicide similar to that of primary and secondary unipolar patients, the result is the appearance of lowered risk for bipolars. Our findings suggest that depression, regardless of specific subtype, leads to a high risk for suicide, although a diagnosis of mania does not. These findings appeal to common sense. Mania is not traditionally associated with risk for suicide, but depression is. Thus, a person with bipolar disorder who has both manic and depressive episodes may spend less time altogether at risk for suicide than a unipolar depressive who experiences only depressive episodes. At least in a short-term follow-up such as ours, bipolars appear to have a lower risk of suicide.

We also noted a trend for suicides to cluster near the beginning of the follow-up period. Nearly three-quarters of the suicides occurred during the first 2 years of follow-up. These findings are consistent with the Iowa Record-Linkage Study and suggest that the relative risk for suicide may be increased early in the course of illness, but continues throughout follow-up.

Future Directions

Studies of suicide risk continue to be important, because prevention of suicide is the ultimate goal of treatment. A knowledge of suicide and of factors influencing its risk is necessary before successful interventions can be systematically applied. Suicide remains a multidimensional problem and requires research designs and analytic methods suited for a multidimensional analysis. The method we have used in the Iowa Depression Study has allowed us to look at risk for suicide in patients with specific subtypes of affective disorder. With demographic, diagnostic, clinical, and treatment data gathered from the chart review, a multivariate study of correlations between these risk factors in suicide can be accomplished. We are now in the process of carrying this out.

The Iowa Record-Linkage Study has made clear the need for large-scale studies when investigating mortality, and has pointed to the need for careful diagnostic assessment. Our economical method has allowed us to gather a large number of data on many patients. We are now planning additional studies on other diagnostic groups, including patients with schizophrenia, sexual deviancy, and antisocial personality.