Substance Use Disorder in Hospitalized Severely Mentally 111 Psychiatric Patients: Prevalence, Correlates, and Subgroups

Transcription

1 Substance Use Disorder in Hospitalized Severely Mentally 111 Psychiatric Patients: Prevalence, Correlates, and Subgroups Abstract by Kim T. Mueser, Paul R. Yarnold, Stanley D. Rosenberg, Chester Swett, Jr., Keith M. Miles, and Diane Hill The prevalence and demographic and clinical correlates of lifetime substance use disorders were examined in a cohort of 325 recently hospitalized psychiatric patients (53% schizophrenia or schizoaffective disorder). Alcohol use was the most common type of substance use disorder, followed by cannabis and cocaine use. Univariate analyses indicated that gender (male), age (younger), education (less), history of time in jail, conduct disorder symptoms, and antisocial personality disorder symptoms were predictive of substance use disorders. Lifetime cannabis use disorder was uniquely predicted by marital status (never married) and fewer psychiatric hospitalizations during the previous 6 months. Optimal classification tree analysis, an exploratory, nonlinear method of identifying patient subgroups, was successful in predicting 74 percent to 86 percent of the alcohol, cannabis, and cocaine use disorders. The implications of this method for identifying specific patient subgroups and service needs are discussed. Keywords: Severe mental illness, schizophrenia, substance abuse, dual diagnosis, prevalence. Schizophrenia Bulletin, 26(1): , 2. Patients with severe mental illnesses such as schizophrenia and bipolar disorder are more likely to have substance use disorders than the general population (Regier et al. 199). Substance use disorders in these patients are associated with a variety of negative outcomes, including relapses and rehospitalizations (Drake et al. 1989), homelessness (Caton et al. 1994), violence (Lindqvist and Allebeck 1989), and higher use of services (Bartels et al. 1993). The high rate of substance abuse and dependence, and its effects on the course of psychiatric illness, has made the identification and treatment of these individuals a high priority. One approach to characterizing which psychiatric patients are most prone to substance use disorders has been to examine the demographic and clinical correlates of substance abuse in this population. For example, younger age and male gender have been found to be related to substance use disorders in psychiatric patients across numerous studies (Mueser et al. 199,1992; Kozaric-Kovacic et al. 1995; Menezes et al. 1996). Identifying a profile of risk factors for substance use problems in psychiatric patients can be useful for several reasons. First, known risk factors may be used to heighten clinicians' "index of suspicion" regarding a comorbid substance use disorder in those patients who meet the profile, thereby improving detection. Second, patient characteristics related to substance use disorders may be of theoretical significance, leading to hypotheses about the mechanisms underlying the high rate of substance use disorders in psychiatric populations. And third, identifying demographic and clinical correlates may lead to interventions designed to prevent the development of substance use disorders in vulnerable patients. There is a general acceptance of the importance of discriminating which psychiatric patients are most vulnerable to substance use disorders, but studies vary in both prevalence rates and demographic and clinical correlates of substance use disorders (Mueser et al. 199). Some of the variability across studies is a result of differences in sampling procedures, such as the setting in which patients are assessed (Galanter et al. 1988). Other methodological problems have also contributed to inconsistent findings, including the failure to employ standardized instruments to determine diagnoses and a lack of information about refusal rates, problems that limit the possibility of generalizing findings to the broader population of psychiatric patients. A final limitation of earlier research has been that almost all studies have involved patients living in urban areas, and the few studies conducted in rural settings have suffered from small sample sizes. The present study was conducted to examine the prevalence, correlates, and subgroups of substance use disorder in a large study group of patients recently admitted to a psychi- Reprint requests should be sent to Dr. K.T. Mueser, New Hampshire Dartmouth Psychiatric Research Center, Main Building, 15 Pleasant St., Concord, NH

2 Schizophrenia Bulletin, Vol. 26, No. 1, 2 K.T. Mueser et al. atric hospital in a rural state. Because the focus of this article is on the relationship between patient characteristics and lifetime substance use disorders, we examined a wide range of patient variables not directly related to these disorders, including demographics, psychiatric diagnoses, hospital utilization, and involvement in the legal system (days in jail). This study has several methodological strengths compared with previous research in this area. First, psychiatric diagnoses were assessed using standardized, structured interviews, and substance use diagnoses were based on both structured interviews and clinician reports. Second, a large majority of the patients eligible for the study consented to participate. Third, the study was conducted in a setting in which patients living in rural, suburban, and urban areas were admitted for psychiatric treatment, improving the possibility of generalizing to the broader population of severely ill psychiatric patients. Methods The study was conducted at New Hampshire Hospital (NHH), the only State hospital in New Hampshire and the primary site for both acute and long-term care of psychiatric inpatients. The average length of stay for acute admissions is approximately 1 days. All patients admitted to NHH are at least 18 years old. By State law, admissions to NHH require an involuntary emergency admission procedure, which is based on the medical/psychiatric finding that the persons represent a danger to themselves or others. In rare instances, patients can also be admitted through preapproval of the hospital medical director, if such an admission is requested by the medical director of one of the ten community mental health centers in the State. These voluntary admissions are generally transfers from other hospitals. The inclusion criteria for the study were the following: (1) an Axis I psychiatric diagnosis; (2) a contact during the previous 6 months with a clinician in the community who was familiar with the patient's functioning (in order to obtain clinician ratings of substance use disorders); and (3) a provision of written, informed consent to participate in the research interviews. Over the study period, 882 unique patients were admitted to NHH, of whom 459 (52%) had an Axis I psychiatric diagnosis. Of these 459 patients, 352 (77%) met the criterion of contact with a clinician. Of these 352 patients, 325 (92%) consented to participate in the study. The 325 patients who participated in the study were quite similar in most respects to the 459 patients with an Axis I diagnosis, including age, psychiatric diagnosis, ethnic group, and marital status; the actual participants were more likely to be female than the group as a whole. The demographic and diagnostic characteristics of the study group are summarized in table 1. Table 1. Study group characteristics (n = 325) Age, mean yrs (SD) Education, mean yrs (SD) Mini-Mental State Score, mean (SD) Female gender, n (%) Caucasian race, n (%) Marital status, n (%) Ever married Never married Legal status, n (%) Involuntary admission Voluntary Diagnosis, n (%) Schizophrenia Schizoaffective Depression Bipolar Other Residential status, n (%) Living independently Living with family Supervised living situation Other 38.8(1.8) 12.1 (2.54) 23.4 (9.3) 171 (53) 316(98) 144(45) 178(55) 36 (95) 15(5) 88 (27) 83 (26) 73 (23) 61 (19) 18(6) 138(43) 13(4) 46(14) 9(3) Measures. Psychiatric and substance use diagnoses were assessed using the Structured Clinical Interview for DSM-III-R (SCID; Spitzer et al. 199). NHH serves only patients with a primary psychiatric disorder. SCID assessments confirmed the presence of an independent psychiatric disorder (i.e., persistent psychiatric symptoms in the absence of substance abuse) through routine diagnostic methods, including a review of records and information from other providers and significant others (mainly family members). SCIDs were performed by two research interviewers. To determine the reliability of the SCID diagnoses, independent ratings based on live or taped interviews were obtained for 15 patients. The average Kappa for primary psychiatric diagnosis and substance use disorder diagnoses was.94, indicating satisfactory reliability. The SCID was also used to elicit the number of symptoms of childhood conduct disorder (CD) and adult antisocial personality disorder (ASPD). Symptom counts for CD and ASPD were used in the statistical analyses to evaluate whether cutpoints other than those used to establish the diagnoses improved the discrimination of substance use disorders in this study group of severely mentally ill patients. In addition to the SCID, substance use diagnoses were assessed with case manager reports based on the Clinician Rating Scales for alcohol and drug use (CRS; Drake et al. 199). On the CRS, clinicians rated on 5- point scales the extent of alcohol, cannabis, and cocaine 18

3 Substance Use Disorder in Psychiatric Patients Schizophrenia Bulletin, Vol. 26, No. 1, 2 abuse over the past 6 months. Points 1 and 2 corresponded to no substance use and substance use without abuse, respectively, and points 3, 4, and 5 corresponded to substance abuse, dependence, and severe dependence, respectively. The CRS has been used in several studies and has been found to have good interrater reliability and validity (Mueser et al. 1995; Carey et al. 1996). Different methods of assessing substance use disorder often detect different patients, with the discrepancy between measures usually a result of nondisclosure, rather than the overdiagnosis of substance use disorders (Drake et al. 199; Shaner et al. 1993; Goldfinger et al. 1996). Therefore, we classified patients as having a lifetime substance use disorder based on either a positive SCID diagnosis (lifetime substance use disorder, including current disorder) or a CRS rating of 3-5 for a particular substance (current substance use disorder). For alcohol, the SCID identified 164 patients as having a lifetime (including current) alcohol use disorder, and the CRS identified 19 additional patients (or 1% of total) as having a current alcohol use disorder. For cannabis, the SCID identified 8 patients as having a lifetime cannabis use disorder, and the CRS identified 14 additional patients (or 15% of total) as having a current cannabis use disorder. For cocaine, the SCID identified 34 patients as having a lifetime cocaine use disorder, and the CRS identified 13 additional patients (or 28% of total) as having a current cocaine use disorder. Record reviews of patients with discrepant SCID and CRS ratings provided support for including patients who were identified as having a substance use disorder with either assessment method in the substance use disorder group. Demographic information, legal status (voluntary/involuntary), and number of hospitalizations in the past 6 months were recorded from patients' charts. The number of lifetime days incarcerated was determined by patient interview. Procedure. The charts of all patients admitted to NHH were screened by the project coordinator for a possible Axis I disorder. For patients who appeared to be eligible for the study, the treatment team leader was contacted for permission to approach the patient to obtain written, informed consent. Consent was usually obtained within 5 days of hospital admission. The clinical interview was completed within 1 week of the date the consent was obtained. Case managers were contacted who provided the CRS ratings. To enhance generalizability, patients were informed that the results of the interviews would be provided to their treatment teams. All patients were paid for participating in the research interview. Statistical Analyses. In our research on the relationships between patient characteristics and substance use disorders, we reported different findings as a function of substance type (Mueser et al. 199, 1992). In addition, in the present study group there was only modest overlap between different substance use disorders. Specifically, among patients with a cocaine use disorder, 37.5 percent had a cannabis use disorder and 4.3 percent had an alcohol use disorder. Among patients with a cannabis use disorder, 24.6 percent had an alcohol use disorder and 65.5 percent had a cocaine use disorder. Among patients with an alcohol use disorder, only 12. percent had a cocaine use disorder and 9.8 percent had a cannabis use disorder. Because of the low overlap between different types of substance use disorder, separate analyses were performed on the most commonly abused substances in this study group: alcohol, cannabis, and cocaine. Patients who abused two or three of the substances were included in each analysis as having a use disorder for that particular substance. The data analysis was divided into two broad parts. First, univariate relationships were examined between patient characteristics and each of the three types of lifetime substance use disorder (alcohol, cannabis, cocaine). Second, nonlinear models were developed to predict each type of substance use disorder from multiple patient characteristics. Univariate analyses were conducted using optimal data analysis (ODA). ODA is a nonlinear method of data analysis that optimizes discrimination without assuming distributional properties of either the predicted or predictor variables. For continuous predictor variables, a univariate ODA (or UniODA) model is of the following form: If score s cutpoint, then predict the observation is from class 1; otherwise, if score > cutpoint, then predict the observation is from class. For categorical predictors, a UniODA model is of the form: If score = categories, then predict the observation is from class 1; otherwise, if score = the other categories, then predict the observation is from class. Regardless of the metric of the predictor variable, by definition, no alternative assignment rule can result in greater (non)weighted classification accuracy for the training sample. Once the optimal model has been identified, type I error is computed as a permutation (exact) probability, and no assumptions regarding parent distributions are required (Yarnold and Soltysik 1991). Unlike models based on the general linear model (that maximize variance ratios) or maximum likelihood (that maximizes the value of the likelihood function) paradigms, only ODA-based models explicitly maximize the number of classifications that are correct in the sample. If observations are weighted (by time, money, importance, etc.), then only ODA-based models explicitly maximize the total return (or minimize the total cost) of the classifications. Multivariable models for the prediction of the substance use disorders were constructed using hierarchically 181

4 Schizophrenia Bulletin, Vol. 26, No. 1, 2 K.T. Mueser et al. optimal classification tree analysis, or CTA (Yarnold 1996; Yarnold et al. 1997; Feinglass et al. 1998; Ostrander et al. 1998). CTA explicitly maximizes mean sensitivity (i.e., the percentage of the observations in a given drug (non)use disorder category that was correctly classified by the model) at each hierarchical step, or node, of the classification tree model. At each node, the mean sensitivity of every potential attribute is computed based on the cutpoint that explicitly maximizes the mean sensitivity, and the attribute yielding the greatest mean sensitivity is retained. For each node, the permutation probability was assessed using 1, Monte Carlo experiments, and a sequentially rejective Sidak Bonferroni procedure ensured an experimentwise type I error rate of p <.5 (Yarnold and Soltysik 1991). In this study, all reported effects met the experimentwise criterion for statistical significance. A jackknife validity analysis ensured model stability at each node. Finally, a bootstrap analysis involving 1 experiments with a 5 percent resample (with replacement) was conducted to assess each model's sensitivity to sample size. Analyses were performed using ODA statistical software (Yarnold and Soltysik, in press). Several nonlinear classification methods have recently been developed, including neural networks and genetic algorithms (typically engineered using traditional discriminant analysis or logit functions); classification and regression tree analysis (CART; Breiman et al. 1984); chisquare automatic interaction detection analysis (CHAID; Magidson 1988); and combination strategies such as AnswerTree, which weds CART and CHAID. A problem with general linear model-based methods (CART, AnswerTree) is that they assume multivariate normality for estimated type I error to be valid, and a problem with maximum likelihood-based methods (CHAID, AnswerTree) is that their coefficients are unbiased only for very large samples. A common problem with all of these suboptimal methods is that they do not explicitly maximize (non)weighted classification accuracy hierarchically; no alternative hierarchical ordering of attributes can achieve greater classification accuracy than the CTA model. In contrast, hierarchical CTA does not require large samples or assumptions about parent distributions, and it alone explicitly maximizes (non)weighted classification accuracy hierarchically. Results The prevalence of different lifetime substance use disorders across the diagnostic groups is summarized in table 2. Different diagnostic groups had similar rates of substance use disorders. Alcohol was the most commonly abused substance; approximately half of the patients had a lifetime alcohol use disorder. Cannabis was the most commonly abused illicit drug, with one-quarter of the subjects having a lifetime use disorder. Relatively few patients had other types of substance use disorder; cocaine use disorder was present in 11 percent of the study group. Overall, 58 percent of the study group met criteria for at least one type of lifetime substance use disorder, 25 percent abused alcohol only, and 26 percent abused alcohol and drugs. Univariate Analyses. The univariate associations between patient characteristics and lifetime alcohol, cannabis, and cocaine use disorder, as identified by ODA, are summarized in table 3. Younger age was associated with cannabis and cocaine use disorder, but not alcohol use disorder. Patients with cannabis use disorder had significantly lower levels of education. Patients who (1) had been incarcerated, (2) had at least one symptom of conduct disorder, and (3) had one to three or more symptoms of antisocial personality disorder (depending on the specific type of substance) were more likely to have all three types of substance use disorder. Patients who had been hospitalized once or not at all in the past 6 months were more likely to have a cannabis use disorder than patients hospitalized more often, whereas recent hospitalizations were not related to alcohol or cocaine use disorder. Males were more likely to have alcohol and cannabis use disorder. Patients who had never married were more likely to have cannabis use disorder. Multivariate Models of Specific Substance Use Disorders. Multivariate CTA models for the prediction of lifetime alcohol, cannabis, and cocaine use disorder are described below. Lifetime alcohol use disorder. Figure 1 presents the CTA model for predicting alcohol use disorder from patient attributes. Using the model to classify individual patients is straightforward. For example, imagine a hypothetical 3-year-old patient with an antisocial personality disorder total score of 1, who was incarcerated for a total of 3 days. Starting with the first node, because the patient was incarcerated for 1 or more days, the right branch is appropriate. Considering next the second node, because the patient had an antisocial personality disorder total score of 1 or, the left branch is appropriate. Considering finally the third node, because the patient's age is 35 or younger, the left branch is appropriate. Thus, the patient is classified by the model as having an alcohol use disorder. Note that of the 26 patients who were classified into the same endpoint via this particular branch of the classification tree, 19 were correctly classified. Thus, the predictive value for this endpoint is 73.1 percent, and the empirical probability of having an alcohol use disorder is p (user) =.731. Had the hypothetical patient's age been 38 years, the third-from-the-right endpoint (corresponding to the 182

5 Total (n = 325) 12(4) 5(2) 9(3) 9(3) 5(2) 188(58) 82 (25) 22(7) 84 (26) 34(11) 8 c8 o t I a' Co to Table 2. Number (%) of patients with lifetime substance use disorder by diagnosis (n = 325) Schizophrenia (n = 89) Schizoaffective (n = 84) Bipolar Major depression Other (n = 73) (n = 61) (n=18) Alcohol 38 (43) 51 (61) 38 (52) 29 (48) 1(56) Cannabis 23 (26) 24 (29) 19(26) 1(16) 6(33) Cocaine 11 (12) 6(7) 6(8) 7(12) 4(22) 166(51) 82 (25) 34(11) Amphetamines 3(3) 3(4) 5(7) 1(2) () Narcotics () 1(1) 3(4) () 1(6) Sedatives () 5(6) 2(3) 2(3) () Hallucinogens 4(5) 1(1) 3(4) 1(2) () Inhalants 1(1) 2(2) 1(1) () 1(6) Any lifetime substance use disorder 46 (52) 57 (69) 42 (59) 31 (51) 12(66) Alcohol use disorder only 18(2) 29 (35) 18(25) 15(25) 2(11) Drug use disorder only 8(9) 6(7) 4(6) 2(3) 2(11) Alcohol and drug use disorder 2 (23) 22 (26) 2 (27) 14(23) 8(44) Polysubstance use disorder 9(1) 8(1) 1(14) 5(8) 2(11) I-

6 use disorder % Use disorder P< Q DO o H c Table 3. Univariate associations of attributes with alcohol, cannabis, and cocaine use disorder, for total study group 1 Attributes Age, yrs Education, yrs Model >4 ^4 l\\ Alcohol n use disorder % Use disorder P< Model >38 <38 l\\ Cannabis n use disorder % Use disorder P<.1.92 Model >42 <42 <H Cocaine n Days incarcerated > > > Hospitalizations, past 6 mos > ^ <! 19 2 Conduct disorder symptoms > > > Antisocial personality disorder symptoms > si > Race White Other White Other White Other Gender Female Male Female Male Female Male Marital status (married, or live with significant other?) Yes No Yes No Yes No Legal status Voluntary Involuntary Voluntary Involuntary Voluntary Involuntary Diagnosis 1,3,4, ,4 1,2, ,4 1,2, Living with family No Yes No Yes 1 Identified through univariable optimal discriminant analysis, the cutpoints/categories under "Model" maximized effect strength for sensitivity (Soltysik and Yarnold, 1993; Yarnold, 1996). Type I error is reported for illustrative purposes, but is not considered in Bonferroni alpha adjustment conducted for the three CTA models (figures 1-3). For diagnosis, 1 = schizophrenia, 2 = schizoaffective, 3 = bipolar, 4 = major depression, 5 = other. For each model, patients falling in the first listed cutpoint category are predicted to not have a use disorder, and patients falling in the second listed cutpoint category are predicted to have a use disorder No Yes

7 Substance Use Disorder in Psychiatric Patients Schizophrenia Bulletin, Vol. 26, No. 1, 2 Figure 1. CTA-based model for predicting alcohol use disorder Alcohol One or more One or none Antisocial Personality Disorder Total Score 9/18(5.) 13/15 (86.7) 35ye«r» / V or younger 1 Mis' / 36 to 47 / ye*" m \ /.54 Use Disorder No Use Use Disorder Disorder 19/26(73.1) 15/2(75.) 19/22(86.4) r \ 48 years \ or older Note. Circles represent decision points (nodes), arrows indicate decision paths (branches), and rectangles represent prediction endpoints (the model classifies patients with a given attribute profile as either "Use disorder" or "No use disorder"). Numbers beneath nodes give the generalized (per-comparison) type I error for the node (if more than one branch emanates from a node, then the left p is for the left and middle branches, and the right p is for the middle and right branches), and numbers adjacent to arrows indicate the cutpoint value for the node. Fractions beneath prediction endpoints give the number of correct classifications at the endpoint (numerator) and the total number of patients classified at the endpoint (denominator). Numbers in parentheses adjacent to the fractions give the predictive value of the model at the endpoint. middle branch of the third node) would have been appropriate and the patient would be classified by the model as not having an alcohol use disorder, with a predictive value of 75 percent. It is interesting that there is a U-shaped relationship between age and alcohol use for patients who were incarcerated and have a low antisocial personality disorder total score: younger (s 35 years) and older (a 48 years) patients tend to have had an alcohol use disorder, whereas patients at intermediate ages (36 to 47 years) have not. Of the 325 patients with data for alcohol use disorder status, this model classified 29 (89.2%), of whom 216 (74.5%) were classified correctly. Nonclassified patients had missing data on at least one of the attributes in the model. Table 4 presents the cross-classification table and summary classification performance indexes for the model, which yielded 49.8 percent of the theoretical classification improvement that it is possible to achieve beyond chance. Bootstrap analysis (1 replications of a 5% resample with replacement) indicated that with the predictive value estimated at each endpoint, corresponding predictive values based on the total sample analysis fell within the 95 percent certainty interval for the estimated endpoint. Thus, the total sample model was replicated using a random sample onehalf the size of the total sample. Table 5 presents the model-based staging system for predicting alcohol use disorder. Note that prediction endpoints from the model (figure 1) are ordered from lowest to highest p (user) (stage 1 to 7, respectively). Stage is an ordinal scale on which increasing integers reflect an increasing likelihood of having an alcohol use disorder: If a ratio scale is desired, then the corresponding p (use disorder) may be used rather than the ordinal value. Note that stages 1, 4, and 6 differ in terms of patient age. That is, of those patients who were incarcerated for at least 1 day and who had an antisocial personality disorder total score of 1 or, patients who were 48 years old or older (stage 6) have more than a threefold higher p (use disorder), and patients who were 35 years old or younger (stage 4) have nearly a threefold higher p (use disorder), compared with patients between 36 and 46 years of age (stage 1). And, of patients who have never been incarcerated and who have an antisocial personality disorder total score of two or greater, those living with family (stage 7) had nearly a twofold higher p (use disorder) than patients not living with family (stage 3). Lifetime cannabis use disorder. The CTA model for predicting cannabis use disorder is presented in figure

8 Classification performance index Alcohol use disorder Cannabis use disorder Cocaine use disorder Stage Days incarcerated Antisocial personality disorder total score Age, yrs Living with family (use disorder) Odds (use disorder) 1 Dashes indicate that the node was not relevant for the indicated stage. N is the number of patients having the indicated pattern (or profile) of responses. The empirical probability of alcohol use disorder, p (use disorder), is the proportion of the patients in the indicated stage who had used alcohol. The approximate odds of alcohol use disorder, odds (use disorder), reflects both n and p (use disorder) and is useful in assessing the incremental increase in p (use disorder) that occurs as one moves toward higher stages. User 5 1:3 1:2 1:1 3:1 5:1 19:3 13:2 Co S- i a S' 6B c to CTv O 2 Table 4. Classification performance summaries for CTA models 1 Overall classification accuracy Sensitivity (use disorder) Specificity (no use disorder) Positive predictive value (use disorder) Negative predictive value (no use disorder) Overall effect strength Cross-classification tables Patient's predicted use disorder status Alcohol use disorder Cannabis use disorder Cocaine use disorder Patient's actual use disorder status Nonuser User Nonuser User Nonuser Nonuser User Note. CIA = classification tree analysis. 1 Overall classification accuracy is the overall percentage of correctly classified patients; sensitivity and specificity (descriptive indices) give the percentage of actual users/nonusers that the model correctly classified; and positive and negative predictive value (prognostic indices) give the percentage of correct classifications into user/nonuser categories made by the model. Effect strength ranges between (chance) and 1 (perfect classification), and indicates the percentage of the theoretical possible improvement in classification accuracy beyond chance that the model achieves. All model performance indexes were stable in jackknife and bootstrap (5% resample) analysis. Table 5. Staging system for likelihood of predicting alcohol use disorder 1 1 > 1 1 oro 36 to oro *2 No oro s ^ oro a Yes 15.87

9 Substance Use Disorder in Psychiatric Patients Schizophrenia Bulletin, Vol. 26, No. 1, 2 Of the 325 patients with data for cannabis use disorder status, this model classified 299 (92.%), of whom 258 (86.3%) were classified correctly. Table 4 gives the crossclassification table and summary classification performance indexes for the model (bootstrap analysis indicated consistent performance for a 5% resample), and table 6 gives the corresponding staging system for predicting cannabis use disorder. For patients who were 38 years of age or younger and who had an antisocial personality disorder total score of 1 or less, those who did not graduate from high school (stage 3) have nearly a threefold greater p (use disorder) than patients who did graduate from high school (stage 2). Lifetime cocaine use disorder. The CTA model for predicting cocaine use disorder is presented in figure 3. Of the 325 patients with data for cocaine use disorder status, this model classified 312 (96.%), of whom 25 (8.1%) were classified correctly. Table 4 gives the cross-classification table and summary classification performance indices for the model (bootstrap analysis indicated consistent performance for a 5% resample). Table 7 gives the corresponding staging system for predicting cocaine use disorder. Note that, relative to patients with antisocial personality disorder total scores of 2 or less (stage 1), patients with antisocial personality disorder total scores of 7 or greater (stage 3) have at least a tenfold greater p (use disorder), and patients with antisocial personality disorder total scores of between 3 and 6 (stage 2) have nearly a threefold greater p (use disorder). Discussion The rates of lifetime substance use disorders in this study group of hospitalized psychiatric patients were remark- Figure 2. CTA-based model for predicting cannabis use disorder (see note to figure 1) Figure 3. CTA-based model for predicting cocaine use disorder (see note to figure 1) Two or Fewer No Use Disorder 21/223(9.1) Antisocial Personality Disorder Total Score No Use Disorder Seven or More Use Disorder 52/71 (73.2) 5/5 (1) ably similar to other large prevalence studies of similar populations (Mueser et al. 199, 1992; Regier et al. 199; Lehman et al. 1994). Consistent across all studies, alcohol was the most common type of substance use disorder; about half of the patients had a lifetime alcohol use disorder. Cannabis was the next most common substance use disorder, with about one-quarter of the patients meeting criteria for a lifetime use disorder. These rates are comparable to the rates of cannabis abuse among hospitalized schizophrenia patients in our first study, conducted in Philadelphia, PA, (Mueser et al. 199) and reported by Lehman et al. (1994) in Baltimore, MD, and slightly higher than the rate found for hospitalized psychiatric patients in our second study in Philadelphia (Mueser et al. 1992). In the present study, cocaine was the third most commonly abused type of substance; 11 percent of the study group had a lifetime use disorder. This rate is comparable to the rate of cocaine use disorder found in our earlier study in Philadelphia, conducted before the rise in cocaine use (Mueser et al. 199), and lower than the rate found in more recent studies in urban areas such as Philadelphia (Mueser et al. 1992) or Baltimore (Lehman et al. 1994). The lower rate of cocaine use disorder appears to reflect the relatively rural nature of New Hampshire compared with major cities. Other types of substances were less frequently abused; fewer than 5 percent of the study group had a use disorder for any one substance. The univariate analyses examining the associations between patient characteristics and substance use disorders (table 2) identified variables that were generally consistent with other studies (Drake and Brunette 1998). Demographic characteristics age, education, and gender were all related to lifetime substance use disorders. Young, unmarried male patients with lower levels of education were most likely to have a substance use disorder. Demographic characteristics were not consistently related 187

10 Schizophrenia Bulletin, Vol. 26, No. 1, 2 K.T. Mueser et al. Table 6. Staging system for likelihood of predicting cannabis use disorder 1 Stage Age, yrs a 39 ^38 s38 s38 Note. HS = high school. 1 See note to table 5. Antisocial personality disorder total score 1 oro 1 oro a 2 Education At least HS graduate Did not graduate HS Table 7. Staging system for likelihood of predicting cocaine use disorder 1 Stage Antisocial personality disorder total score n 1 s to * See note to table 5. to substance use disorders across all three major types of substances. For example, age was related to lifetime prevalence of cannabis and cocaine use disorder, but not alcohol use disorder. Illicit substance abuse tends to decline with age in the general population (Chen and Kandel 1995), but the tendency for younger patients to have lifetime drug use disorders, consistent with our previous studies (Mueser et al. 199, 1992), may reflect cohort or selection factors related to the inpatient setting. It is possible that as patients stop abusing drugs such as cocaine and cannabis over time, they are less likely to be hospitalized, and therefore less likely to be represented in inpatient samples. Another possibility is that drug abuse among persons with severe mental illness contributes to higher rates of incarceration and mortality, so that older inpatient and outpatient samples of patients are less likely to have ever had a drug use disorder. Of course, both explanations may be correct for different patients. The univariate analyses of the diagnostic and clinical correlates of substance use disorders also revealed several relationships consistent with previous research. Psychiatric diagnosis was not related to type of substance use disorder, in line with most other research (Regier et al. 199). The number of hospitalizations in the past 6 months was related to cannabis use disorder in a counterintuitive direction: patients with more than one recent hospitalization were less likely to have a lifetime cannabis use disorder than patients with one or no recent hospitalizations. It should be noted that the association between cannabis use disorder and fewer hospitalizations cannot be explained by age. Age was related to cannabis use disorder (younger patients more likely to have disorder), but age was also related to recent hospitalizations in the opposite direction, n p (user) p (use disorder) Odds (use disorder) 1:8 2:7 5:3 2:1 Odds (use disorder) 1:9 1:3 a 5:1 with younger patients more likely to have been recently hospitalized (r = -.16, p <.4). Therefore, while younger patients were more likely to have recent hospitalizations, patients with a lifetime cannabis use disorder, who tended to be younger than other patients, were apparently less likely to have recent psychiatric hospitalizations. Although surprising, this finding is consistent with a previous study we conducted on a similar population of acute inpatients using similar methods (Mueser et al. 199). In that study, patients with a lifetime history of cannabis use disorder also had significantly fewer lifetime psychiatric hospitalizations. In a second study, in an inpatient cohort where cocaine use disorders had replaced cannabis use disorders as the most common illicit drug diagnosis (Mueser et al. 1992), those with cocaine use disorders had fewer lifetime hospitalizations. It may be that among hospitalized psychiatric patients, certain styles of social functioning may increase vulnerability to developing illicit substance use disorders because of the social nature of drug use, as has been reported in two studies of patients with schizophrenia (Dixon et al. 1991; Arndt et al. 1992). Patients with more severe negative symptoms, and modes of premorbid functioning characterized by social isolation, may lack the exposure to illicit substances, or the social skills to obtain them regularly, in order to develop a drug use disorder (Cohen and Klein 197), but may nevertheless be prone to multiple hospitalizations. Conversely, patients with more antisocial, thrillseeking symptomatology may be more prone to involvement in social networks where substances are commonly available and use disorders more prevalent. In patients with severe mental illness, other research has found that current substance use disorders are related 188

11 Substance Use Disorder in Psychiatric Patients Schizophrenia Bulletin, Vol. 26, No. 1, 2 to more, not fewer, hospitalizations (reviewed by Drake and Brunette 1998). There are several possible explanations for the different associations between illicit substance use disorders and previous hospitalizations in our three studies (Mueser et al. 199, 1992, this study) compared with other reports. First, our studies have focused on the correlates of lifetime substance use disorders, whereas many others have examined current or recent substance use disorders. Second, we evaluated substancespecific associations (for alcohol, cannabis, and cocaine), whereas most other research has examined alcohol and mixed drug use disorders. Third, our studies were conducted with acutely ill inpatients, which may not generalize to the broader population of outpatients with severe mental illness. Further research is needed to evaluate the relationships between premorbid social functioning, illicit substance use disorders, and psychiatric hospitalizations in persons with severe mental illness. The finding that the number of conduct disorder symptoms, number of antisocial personality disorder symptoms, and days incarcerated were strongly related to lifetime alcohol, cannabis, and cocaine use disorders is consistent with research on antisocial personality disorder and substance use disorders in the general population (Alterman and Cacciola 1991). Several studies have also demonstrated that antisocial personality disorder in patients with schizophrenia is associated with higher rates of substance use disorders (Caton et al. 1994, 1995; Hodgins et al. 1997) and, among dually diagnosed patients, the severity of substance use disorder (Mueser et al. 1997). Because of the importance of ASPD in predicting substance use disorders, we examined the frequency of different ASPD symptoms in our study group. The most common ASPD symptoms were failure to conform to social norms with respect to lawful behaviors (28%), reckless disregard for safety of self or others (25.5%), and irritability and aggressiveness as indicated by repeated fights or assaults (17.2%). The least common ASPD symptoms were lacks ability to function as a parent or guardian (3.7%), has never sustained a totally monogamous relationship for more than a year (4.9%), and lacks remorse (9.2%). It is interesting to note that the univariate analyses indicated that there is a much lower threshold of CD and ASPD symptoms than is required for a diagnosis of substance use disorders (according to either DSM-IH-R [American Psychiatric Association 1987] or DSM-IV [American Psychiatric Association 1994]) in psychiatric patients with a history of these disorders. According to DSM-III-R criteria, three or more symptoms are necessary to diagnose CD, and an additional four or more ASPD symptoms to diagnose ASPD. DSM-IV requires three or more symptoms each from CD and ASPD to diagnose ASPD. Among the psychiatric patients studied here, the presence of only one or two CD or ASPD symptoms was needed to optimally discriminate patients with versus patients without a substance use disorder. These findings suggest that CD and ASPD may be more sensitive predictors of substance use disorder in the psychiatric population compared with the general population. A limitation of previous research on demographic and clinical correlates has been the use of statistical models that are unduly restrictive with respect to their assumptions about the distribution and linearity of measures hypothesized to be related to substance use disorders. In most studies in this area, only one predictor variable has been considered at a time. And, when multiple predictors have been evaluated simultaneously, statistical models were employed that assume that substance use disorders can be efficiently predicted by a linear composite of the predictor variables (e.g., multiple regression, logistic regression). While such approaches have been useful in identifying some correlates of substance use disorder, they are limited by distributional assumptions often not met by the data, and their inability to detect nonlinear associations between predictor variables and substance use disorders. One goal of this study was to evaluate the performance of a nonlinear method of combining demographic and clinical characteristics, based on CTA, in developing a parsimonious model to predict alcohol, cannabis, and cocaine use disorders. Analyses indicated a relatively high overall classification accuracy for all three substance types (range = 74.5%-86.3%; table 4). Furthermore, for two of the models (alcohol and cannabis), the models resulted in relatively high sensitivity (7.5% and 63.7%) and specificity (79.8% and 86.9%). The model for cocaine use disorder had very low sensitivity (1.9%), in part because of its low base rate. The models for alcohol and cannabis use disorder suggest some intriguing interpretations. However, caution is necessary until the findings are replicated on independent samples. It must also be noted that these interpretations apply to lifetime rather than current use disorders, and that they refer to a particular subset of all patients: those involuntarily committed for an acute inpatient treatment. The model for alcohol abuse indicated lifetime days incarcerated and total antisocial personality disorder symptoms as the most important predictors. Living with family and age were also important predictors on some branches of the tree, but not others. Among patients who had never been incarcerated, those who were living with their family and had two or more antisocial personality disorder symptoms were likely to have a lifetime alcohol use disorder (86.7%). There is evidence that loss of family support among dually diagnosed patients contributes to housing instability and homelessness (Clark 1996), although less research has examined 189

12 Schizophrenia Bulletin, Vol. 26, No. 1, 2 K.T. Mueser et al. the relationship between family support and criminal activity. It is possible that families serve to buffer patients with alcohol use disorders and antisocial features from illegal activities or from the consequences of such illegal behavior (e.g., charges being pressed, incarceration). Families, especially parents, may also serve as caregivers of "last resort" for patients with severe substance use disorders who have few other housing alternatives available in the community (Clark and Drake 1994). The clinical implications of these findings are that family work may need to address the substance use problems of dually diagnosed patients living at home, while more targeted efforts are needed to develop alternative housing options for these patients (Osher and Dixon 1996). Among patients with a history of incarceration but few antisocial personality disorder symptoms, older (48 years or older) and younger (35 years or younger) patients were more likely to have an alcohol use disorder than patients between 36 and 48 years old (figure 1). These results suggest that age is not monotonically related to lifetime alcohol use problems in psychiatric patients. It is interesting to note that in our two other studies (in which we used linear multivariate models) age was not related to history of alcohol use disorders (Mueser et al. 199, 1992). This finding is similar to several other studies of general psychiatric admissions (Crowley et al. 1974; Pulver et al. 1989), but different from studies at VA hospitals where younger patients are more likely to have alcohol use disorders (e.g., Alterman et al. 1981; Lambert et al. 1996). Although in need of replication, this finding suggests that the relationship between age and alcohol use disorder in recently hospitalized patients may be complex, with both cohort and aging effects having an impact on the relationship between alcohol use disorder and hospitalizations in people with severe mental illness. Menezes et al. (1996) reported a similar pattern of alcohol use disorder to be more common among younger and older men with severe mental illness than men of intermediate age. The CTA model predicting cannabis use disorder showed that, consistent with previous research (e.g., Mueser et al. 199, 1992), age was a strong predictor of use disorder status. However, among younger patients, antisocial personality disorder total score and educational level were also important predictors. Patients with higher antisocial personality scores were likely to have a cannabis use disorder, whereas among those with lower scores, only patients with lower education tended to have a cannabis use disorder. This finding appears to reflect the disruptive effects of cannabis use on completing education in this subgroup of patients. The CTA models developed to predict alcohol, cannabis, and cocaine use disorders need to be replicated with samples of patients in other geographical settings, including both inpatient and outpatient populations. The patients studied here differed from those in other settings in that a very high percentage were admitted on an involuntary basis, were white, and tended to live in rural environments. Despite these differences in patient characteristics from our two previous studies (Mueser et al. 199, 1992), there was considerable consistency in findings across the three studies in terms of most correlates, including gender, age, education, marital status, and number of previous hospitalizations. In summary, demographic variables were potent concurrent predictors of specific lifetime substance use disorders, consistent with previous research. These findings suggest that exposure and access to specific substances are important determinants of substance use disorders in the psychiatric population. These, in turn, are highly influenced by life-course developmental, cohort, and social network factors. The interaction among illness, substance abuse, and hospitalization must be understood in the context of an aging individual in differential and evolving social worlds. In addition, endorsement of even a few conduct disorder symptoms in childhood, and antisocial personality disorder symptoms in adulthood, as well as having spent time in jail, were strong predictors of lifetime history of substance use disorders, regardless of age. These same patient characteristics are strongly predictive of substance use disorders in the general population, indicating that these factors operate in the same manner in the psychiatric and general populations. Future research is needed to determine the replicability of the subgroups of patients with substance use disorders described here, and to evaluate whether the course of their disorders differs from that of other patients. Such research will help to determine whether these patients require different types of dual diagnosis treatment (e.g., more intensive) from those of other patients. References Alterman, A.I., and Cacciola, J.S. The antisocial personality disorder diagnosis in substance abusers: Problems and issues. Journal of Nervous and Mental Disease, 179:41-49, Alterman, A.I.; Erdlen, F.R.; and Murphey, E. Alcohol abuse in the psychiatric hospital population. Addictive Behaviors, 6:69-73, American Psychiatric Association. DSM-III-R: Diagnostic and Statistical Manual of Mental Disorders. 3rd ed., revised. Washington, DC: The Association, American Psychiatric Association. DSM-FV: Diagnostic and Statistical Manual of Mental Disorders. 4th ed. Washington, DC: The Association,

13 Substance Use Disorder in Psychiatric Patients Schizophrenia Bulletin, Vol. 26, No. 1, 2 Arndt, S.; Tyrrell, G.; Flaum, M.; and Andreasen, N.C. Comorbidity of substance abuse and schizophrenia: The role of pre-morbid adjustment. Psychological Medicine, 22: , Bartels, S.J.; Teague, G.B.; Drake, R.E.; Clark, R.E.;. Bush, P.; and Noordsy, D.L. Substance abuse in schizophrenia: Service utilization and costs. Journal of Nervous and Mental Disease, 181: , Breiman, L.; Friedman, J.H.; Olshen, R.A.; and Stone, C.J. Classification and Regression Trees. Pacific Grove, CA: Wadsworth, Carey, K.B.; Cocco, K.M.; and Simons, J.S. Concurrent validity of substance abuse ratings by outpatient clinicians. Psychiatric Services, 47: ,1996. Caton, C.L.M.; Shrout, P.E.; Dominguez, B.; Eagle, P.F.; Opler, L.A.; and Cournos, F. Risk factors for homelessness among women with schizophrenia. American Journal of Public Health, 85: , Caton, C.L.M.; Shrout, P.E.; Eagle, P.F.; Opler, L.A.; Felix, A. F.; and Dominguez, B. Risk factors for homelessness among schizophrenic men: A case-control study. American Journal of Public Health, 84:265-27, Chen, K., and Kandel, D.B. The natural history of drug use from adolescence to the mid-thirties in a general population. American Journal of Public Health, 85:41 47, Clark, R.E. Family support for persons with dual disorders. In: Drake, R.E., and Mueser, K.T., eds. Dual Diagnosis of Major Mental Illness and Substance Abuse Disorder II: Recent Research and Clinical Implications. New Directions in Mental Health Services. Vol. 7. San Francisco, CA: Jossey-Bass, pp Clark, R.E., and Drake, R.E. Expenditures of time and money by families of people with severe mental illness and substance use disorders. Community Mental Health Journal, 3: , Cohen, M., and Klein, D.F. Drug abuse in a young psychiatric population. American Journal of Orthopsychiatry, 4: , 197. Crowley, T.J.; Chesluk, D.; Dilts, S.; and Hart, R. Drug and alcohol abuse among psychiatric admissions: A multidrug clinical-toxicologic study. Archives of General Psychiatry, 3:13-2, Dixon, L.; Haas, G.; Weiden, P.J.; Sweeney, J.; and Francis, A.J. Drug abuse in schizophrenic patients: Clinical correlates and reasons for use. American Journal of Psychiatry, 148:224-23, Drake, R.E., and Brunette, M.F. Complications of severe mental illness related to alcohol and drug use disorders. In: Galanter, M., ed. Recent Developments in Alcoholism. Vol. 14. The Consequences of Alcohol. New York, NY: Plenum Press, pp Drake, R.E.; Osher, F.C.; Noordsy, D.L.; Hurlbut, S.C.; Teague, G.B.; and Beaudett, M.S. Diagnosis of alcohol use disorders in schizophrenia. Schizophrenia Bulletin, 16(l):57-67, 199. Drake, R.E.; Osher, F.C.; and Wallach, M.A. Alcohol use and abuse in schizophrenia: A prospective community study. Journal of Nervous and Mental Disease, 177:48-414,1989. Feinglass, J.; Yarnold, P.R.; Martin, G.J.; and McCarthy, W.J. A classification tree analysis of selection for discretionary treatment. Medical Care, 36:74-747, Galanter, M.; Castaneda, R.; and Ferman, J. Substance abuse among general psychiatric patients: Place of presentation, diagnosis and treatment. American Journal of Drug and Alcohol Abuse, 14: , Goldfinger, S.M.; Schutt, R.K.; Seidman, L.J.; Turner, W.M.; Penk, W.E.; and Tolomiczenko, G.S. Alternative measures of substance abuse among homeless mentally ill persons in the cross-section and over time. Journal of Nervous and Mental Disease, 184: , Hodgins, S.; Toupin, J.; and Cote, G. Schizophrenia and antisocial personality disorder: A criminal combination. In: Schlesinger, L.B., ed. Explorations in Criminal Psychopathology: Clinical Syndromes With Forensic Implications. Springfield, IL: Charles C. Thomas, pp Kozaric-Kovacic, D.; Folnegovic-Smalc, V; Folnegovic, Z.; and Marusic, A. Influence of alcoholism on the prognosis of schizophrenic patients. Journal of Studies on Alcohol, 56: , Lambert, M.T., Griffith, J.M.; and Hendrickse, W. Characteristics of patients with substance abuse diagnoses on a general psychiatry unit in a VA medical center. Psychiatric Services, 47: , Lehman, A.F.; Myers, C.P.; Corty, E.; and Thompson, J.W. Prevalence and patterns of "dual diagnosis" among psychiatric inpatients. Comprehensive Psychiatry, 35:1 5, Lindqvist, P., and Allebeck, P. Schizophrenia and assaultive behaviour: The role of alcohol and drug abuse. Acta Psychiatrica Scandinavica, 82: , Magidson, J. Improved statistical techniques for response modeling: Progression beyond regression. Journal of Direct Marketing, 2:6-18,

14 Schizophrenia Bulletin, Vol. 26, No. 1, 2 K.T. Mueser et al. Menezes, P.R.; Johnson, S.; Thornicroft, G.; Marshall, J.; Prosser, D.; Bebbington, P.; and Kuipers, E. Drug and alcohol problems among individuals with severe mental illnesses in South London. British Journal of Psychiatry, 168: , Mueser, K.T.; Drake, R.E.; Ackerson, T.H.; Alterman, A.I.; Miles, K.M.; and Noordsy, D.L. Antisocial personality disorder, conduct disorder, and substance abuse in schizophrenia. Journal of Abnormal Psychology, 16: , Mueser, K.T.; Nishith, P.; Tracy, J.I.; DeGirolamo, J.; and Molinaro, M. Expectations and motives for substance use in schizophrenia. Schizophrenia Bulletin, 21(3): , Mueser, K.T.; Yarnold, PR.; and Bellack, A.S. Diagnostic and demographic correlates of substance abuse in schizophrenia and major affective disorder. Acta Psychiatrica Scandinavica, 85:48-55, Mueser, K.T.; Yarnold, PR.; Levinson, D.F.; Singh, H.; Bellack, A.S.; Kee, K.; Morrison, R.L.; and Yadalam, K.G. Prevalence of substance abuse in schizophrenia: Demographic and clinical correlates. Schizophrenia Bulletin, 16(l):31-56, 199. Osher, F.C., and Dixon, L.B. Housing for persons with cooccuring mental and addictive disorders. In: Drake, R.E., and Mueser, K.T., eds. Dual Diagnosis of Major Mental Illness and Substance Abuse Disorder II: Recent Research and Clinical Implications. New Directions in Mental Health Services, Vol. 7. San Francisco: Jossey-Bass, pp Ostrander, R.; Weinfurt, K.P.; Yarnold, P.R.; and August, G. Diagnosing attention deficit disorders using the BASC and the CBCL: Test and construct validity analyses using optimal discriminant classification trees. Journal of Consulting and Clinical Psychology, 66:66-672, Pulver, A.E.; Wolyniec, P.S.; Wagner, M.G.; Moorman, C.C.; and McGrath, J.A. An epidemiologic investigation of alcohol-dependent schizophrenics. Acta Psychiatrica Scandinavica, 79:63-612, Regier, D.A.; Farmer, M.E.; Rae, D.S.; Locke, B.Z.; Keith, S.J.; Judd, L.L.; and Goodwin, F.K. Comorbidity of mental disorders with alcohol and other drug abuse. Journal of the American Medical Association, 264: , 199. Shaner, A.; Khalsa, M.E.; Roberts, L.J.; Wilkins, J.N.; Anglin, D.; and Hsieh, S.C. Unrecognized cocaine abuse among schizophrenic patients. American Journal of Psychiatry, 15: , Soltysik, R.C., and Yarnold, PR. ODA 1.: Optimal Data Analysis for DOS. Chicago, IL: Optimal Analysis, Soltysik, R.C., and Yarnold, PR. Univariable optimal discriminant analysis: One-tailed hypotheses. Educational and Psychological Measurement, 54: , Spitzer, R.L.; Williams, J.B.W.; Gibbon, M.; and First, M.B. Structured Clinical Interview for DSM-IIl-R Patient Edition. Washington, DC: American Psychiatric Association, 199. Yarnold, PR. Discriminating geriatric and non-geriatric patients using functional status information: An example of classification tree analysis via UniODA. Educational and Psychological Measurement, 56: , Yarnold, PR., and Soltysik, R.C. Theoretical distributions of optima for univariate discrimination of random data. Decision Sciences, 22: , Yarnold, PR., and Soltysik, R.C. Optimal Data Analysis, With Software for DOS and Windows. Washington, DC: American Psychological Association Books, in press. Yarnold, PR.; Soltysik, R.C; and Bennett, C.L. Predicting in-hospital mortality of patients with AIDS-related Pneumocystis carinii pneumonia: An example of hierarchically optimal classification tree analysis. Statistics in Medicine, 16: , Acknowledgments This research was supported by NIMH grants MH-594 and MH-839. The authors appreciate the comments of Robin E. Clark, Ph.D., and Robert E. Drake, M.D., Ph.D., on earlier drafts of this manuscript. The Authors Kim T. Mueser, Ph.D., and Stanley D. Rosenberg, Ph.D., are Professors of Psychiatry, Dartmouth Medical School, Hanover, NH. Paul R. Yarnold, Ph.D., is Research Professor, Division of General Internal Medicine, Northwestern University Medical School, Chicago, IL, and Adjunct Professor, Department of Psychology, University of Illinois, Chicago, IL. Keith M. Miles, M.P.A., is a research associate, New Hampshire- Dartmouth Psychiatric Research Center, Concord, NH. Chester Swett, Jr., M.D., is Clinical Professor of Psychiatry, Harvard Medical School, and Chief of Psychiatry, Brockton, MA, Medical Center, Brockton, MA. Diane Hill, R.N., is a Clinical Nurse Specialist, New Hampshire Hospital, Concord, NH. 192