Violence and Delusions: Data From the MacArthur Violence Risk Assessment Study

Methods of the study have been reported in detail elsewhere (18, 19). In brief, subjects were approached soon after acute psychiatric hospitalization at three study sites (Worcester, Mass.; Pittsburgh; and Kansas City, Mo.) and asked to give written informed consent to participate in the study. Eligibility was limited to English-speaking white, African American, or Hispanic patients 18–40 years of age who were civilly admitted and had a medical record diagnosis of schizophrenia, schizophreniform disorder, schizoaffective disorder, depression, mania, brief reactive psychosis, delusional disorder, alcohol or other drug abuse/dependence, or a personality disorder. Of 1,695 potential subjects approached in a stratified random sampling design, 71.0% (N=1,203) agreed to participate, and 1,136 (67.0%) completed the interview after admission and formed the baseline sample.

Baseline data were collected in the hospital during 4–6 hours of interviewing by a research interviewer and a research clinician. The latter made a research diagnosis by using the DSM-III-R Checklist (20). Subjects were then recontacted in the community after discharge by the research interviewers and were reinterviewed five times (every 10 weeks) over the subsequent year. At least one follow-up interview was obtained for 83.7% (N=951) of the subjects; three or more interviews were obtained for 72.0% (N=818). A collateral informant nominated by the subject as the person most familiar with his or her behavior in the community was also interviewed on the same schedule after having given written informed consent to participate. Three or more collateral interviews were obtained for 77.3% (N=878) of the collateral informants. Supplementary data were obtained from subjects’ hospital and arrest records.

To determine whether subjects were delusional, a set of questions taken mostly from the Diagnostic Interview Schedule (21) was administered at baseline and at each follow-up interview. Interviewers were then asked to judge, according to DSM criteria and all information available to them, whether the subjects were possibly or definitely delusional or whether subject responses reflected reality or some other nondelusional perception (e.g., drug dealers on the subject’s front steps really were watching [“spying on”] him; subject’s expressed belief that someone was trying to “control” her was based on ex-husband’s withholding of child support payments). Delusions that appeared to be related to substance use were not excluded. At the baseline interview, assessment of delusions was performed by research clinicians (one Ph.D. and two with master’s degrees), with consultation from an experienced psychiatrist available at each site. Delusions were classified by using a DSM-based typology, and additional note was made of whether the delusions involved other people or contained violent ideation; these determinations were reviewed for consistency by the first author.

For subjects rated as definitely or possibly delusional, a more detailed assessment of the delusion identified as having the greatest recent impact on the subjects’ lives was conducted. (In some cases, the interviewer had to make this choice when the subject could not.) The assessment was conducted by using the MacArthur-Maudsley Delusions Assessment Schedule, an adaptation of the Maudsley Assessment of Delusions Schedule (22), which generates scores on six non-content-related dimensions: conviction, negative affect, acting on belief, refraining from acting because of belief, preoccupation, and pervasiveness. Further details and psychometric data on the MacArthur-Maudsley Delusions Assessment Schedule have been presented elsewhere (23).

Incidents of violence were ascertained by interviewing subjects and collateral informants about the subjects’ behavior in the previous 10 weeks and by a review of subjects’ hospitalization and arrest records. Actions were considered to constitute violence if they were 1) batteries that resulted in physical injury or involved the use of a weapon, 2) sexual assaults, or 3) threats made with a weapon in hand. Acts judged to be in self-defense were not counted as violence. Batteries that did not result in injury or involve weapon use were coded separately as “other aggressive acts.” Violent acts reported by any information source were reviewed by two independent coders to obtain a single reconciled report of violence. A hierarchy of coding rules is available from the authors.

Table 1 compares the rate of violent acts reported at each follow-up evaluation for subjects who were and those who were not assessed as delusional at the previous interview. There were no significant differences in rates of violence between the two groups at any follow-up evaluation.

Next, we examined whether particular types of delusions might be associated with a higher likelihood of violent behavior. Here and in subsequent analyses we set the level of statistical significance at p<0.05 (two-tailed). Although an argument could be made that a more conservative value should have been employed because of the large number of comparisons made, we selected this significance threshold to provide the greatest opportunity for a possible relationship between violence and delusions to be demonstrated. The only significant findings came at the first two follow-up assessments (table 2 displays the findings from the first), at which body/mind control delusions—contrary to what might be predicted by previous research on threat/control override delusions—displayed a negative relationship to the incidence of violence. In other words, subjects with body/mind control delusions were significantly less likely to commit violent acts during the subsequent 10 weeks than were subjects without such delusions. Because of the multiple comparisons made, the significance of even these relationships is questionable.

Delusions were also categorized according to whether the delusional content involved other people in general, violence in general, or violence that was specifically directed toward other people. None of these categories was significantly related to violence at any follow-up evaluation. Moreover, none of the non-content-related dimensions of delusions assessed by the MacArthur-Maudsley Delusions Assessment Schedule was significantly related to violence at any follow-up evaluation. Self-reported duration of delusions similarly had no relationship to subsequent violence at any follow-up assessment.

Given that neither the presence of delusions per se nor any of the content-related or non-content-related descriptors of delusions was positively associated with the occurrence of violence, we explored the possibility that the presence of selected concomitant variables—generally believed to potentiate violence—might influence the likelihood of violent behavior. Delusional subjects with command hallucinations at baseline (N=145), a group thought by some to be at higher risk of violence (24), showed no greater rate of violence in the subsequent 10 weeks or at any follow-up evaluation than delusional subjects without such hallucinations (e.g., at the first follow-up assessment: χ²=2.30, df=1, p=0.13; N=843). As shown in table 3, the same was true for subjects with research diagnoses of alcohol or drug abuse/dependence, although a significant potentiating effect of substance abuse was seen across the first three follow-up evaluations in the nondelusional group. A Mantel-Haenszel summary odds ratio analysis showed that there was no significant difference at any follow-up evaluation in the likelihood of violent behavior for subjects with and without a substance abuse diagnosis, in the presence or absence of delusions (e.g., at the first follow-up evaluation: χ²=0.15, df=1, p=0.70).

Delusional subjects who reported at baseline having thought about committing violence (i.e., responding affirmatively to the question, “Do you sometimes think about hurting other people?”; N=260) were significantly more likely than delusional subjects without imagined violence to be violent at the first (χ²=14.89, df=1, p=0.03; N=849) and second (χ²=7.90, df=1, p=0.005; N=722) follow-up evaluations. However, comparison with the nondelusional group revealed this to be an effect of imagined violence per se; imagined violence had a significant effect on the rate of violence in the nondelusional group across all five follow-up evaluations (χ²=5.39–8.70, df=1, p=0.02–0.003). At the first follow-up assessment, for example, entering imagined violence into a logistic regression to predict violence yielded a significant effect (χ²=9.00, df=1, p=0.003), but there was no significant improvement from adding the delusions measure or the interaction term between imagined violence and delusions; similar findings were obtained at all subsequent follow-up evaluations.

The same was true for the concomitant presence at baseline of high psychopathy (score >12 on the screening version of the Revised Psychopathy Checklist [25]; N=174), which also had a significant effect on violence rates among delusional subjects at the first (χ²=12.20, df=1, p<0.001; N=785), second (χ²=9.20, df=1, p=0.002; N=688), and fifth (χ²=9.50, df=1, p=0.002; N=648) follow-up evaluations. But similar rates of violence were found among nondelusional subjects with high psychopathy scores. Entering psychopathy scores into a logistic regression yielded significant results (e.g., first follow-up: χ²=76.80, df=1, p<0.001), but there was no significant improvement from adding delusions or the interaction term, which made it clear that the effect represents the influence of psychopathy alone. This was true for all follow-up evaluations.

To test directly the effect of threat/control override delusions on violence in this sample, a subset of the questions used to screen for delusions was identified that best fit the threat/control override construct (appendix 1). Four questions addressed the threat dimension, and four addressed the control override dimension. The resulting scale had high internal consistency (Cronbach’s alpha coefficient=0.82), with essentially no variation if any item was deleted. When we compared rates of violence in the subsequent 10 weeks for subjects who were and were not rated as having threat/control override delusions (table 4), a significant difference was found between the two groups only at the first follow-up evaluation. At that point, subjects with threat/control override delusions, contrary to what was expected on the basis of previous studies, had a significantly lower rate of violence than those without such delusions. Also in contrast to previous studies, there was no significant relationship between the number of threat/control override delusions present and violent behavior at any follow-up evaluation.

Analyses were also performed separately to examine the relationship between threat/control override delusions and the four largest primary diagnostic groupings in the study: depression/dysthymia (N=355), schizophrenia/schizoaffective disorder (N=147), bipolar disorder (N=124), and alcohol or drug abuse/dependence (N=177). The only significant association occurred at the fourth follow-up visit for subjects with bipolar disorder (χ²=8.46, df=1, p=0.04; N=643). Given the number of analyses performed, this result is likely to represent a chance association.

To take advantage of data from all five follow-up evaluations to estimate the effect of threat/control override delusions on violence, the data were reanalyzed by using a discrete-time, event-history method (26). For each follow-up period during which a subject was known to be at risk of committing a first violent act, a separate observational record was created. Individuals for whom no violent act was recorded by the end of the study (i.e., censored individuals) thus contributed a maximum of five observational records to the analysis file. Explanatory variables included threat/control override symptoms for each subject at each follow-up, as well as a series of four indicator variables that collectively represented the five follow-up periods. Logistic regression was used to estimate the probability that an individual would commit a violent act given that one had not been previously committed (i.e., the hazard rate of violence). The results of these analyses (N=3,542) confirmed the absence of a significant predictive effect (odds ratio=0.71, 95% confidence interval=0.46 –1.10) of threat/control override delusions on violence (Wald test=2.35, df=1, p=0.13).

Because of the failure to confirm the results of earlier studies (10–12, 14, 15), several methodological explanations were considered. First, the possibility was examined that the different criteria used for threat/control override delusions in this study affected the outcome. Close approximations to the criteria used by Link and Stueve (10) and by Swanson et al. (13, 14) were constructed from our data. Our threat/control override criteria at baseline correlated moderately with those of Link and Stueve (r=0.64, N=1,136, p<0.0001) and highly with those of Swanson et al. (r=0.96, N=1,136, p<0.0001), which, in turn, correlated moderately with each other (r=0.68, N=1,136, p<0.0001). Although some of the variance between our findings and Link and Stueve’s may be accounted for by the difference in criteria, this cannot explain the failure to replicate the results of Swanson et al. Moreover, the correlation between our criteria and Link and Stueve’s was highly significant and roughly the same as the correlation between the criteria from the other two studies.

Second, we explored whether differences in measurement of the outcome variable (i.e., violence) might have accounted for the contrast in our findings. Previous studies did not restrict reports of violence, as this study did, to instances in which harm resulted, weapons were used, or sexual assaults occurred. However, even redoing the analyses in table 4 by adding “other aggressive acts,” which included all batteries without injury, to the outcome measure yielded no significant relationship with threat/control override delusions at any follow-up evaluation. The same was true when we redid the discrete-time, event-history analysis with the expanded outcome measure.

A third methodological difference involved the use of prospective versus retrospective determinations of the relationship between threat/control override delusions and violence. Previous studies all had asked about threat/control override symptoms and violence during some previous period. Our analyses focused on the prospective predictive effect of the presence of threat/control override delusions. As with all retrospective methods, the earlier studies may have introduced unspecifiable biases into the data. When data from the current study were reanalyzed retrospectively (i.e., examining the relationship between the occurrence during the previous 10 weeks of both threat/control override delusions and violence), there were no longer significant findings in the “unexpected” direction. Overall, there were no significant differences between subjects who did and those who did not report threat/control override delusions. Thus, this methodological issue, taken by itself, may have played some small role in producing different outcomes among the studies.

The final methodological issue involved the manner of determining the presence of threat/control override delusions. Earlier studies of Link and Stueve (10) and Swanson et al. (13, 14) relied on subjects’ answers to screening questions. (The study of Link et al. [11, 12] from Israel, however, also used psychiatrist-rated delusion items from the Schedule for Affective Disorders and Schizophrenia.) Thus, for the most part, if subjects responded affirmatively to a question regarding, for example, whether someone was trying to do them harm, they were rated as having a threat/control override delusion. (In a Dec. 1998 personal communication, Swanson reported that the lay ECA interviewers had the option of rating responses to two of the four items incorporated into his threat/control override measure as “plausible,” i.e., presumably nondelusional. If so rated, responses were not considered indicative of threat/control override symptoms in his studies. He reported that “not many” responses were excluded on this basis.) In contrast, the interviewers in our study were instructed to probe subjects’ responses and to assess, on the basis of all data available to them (including the subjects’ medical records at baseline), whether subjects actually were delusional. This process resulted in a reduction of the number of persons who could be rated at baseline as having threat/control override delusions, from the 532 subjects who responded affirmatively to at least one threat/control override screening question to the 230 who were judged to have a threat/control override delusion. When additional reanalyses that included all subjects with self-reported threat/control override symptoms were performed, a significant difference was found at the second follow-up evaluation in the expected direction (i.e., persons reporting threat/control override symptoms were more likely to be violent) (χ²=6.19, df=1, p=0.01; N=726). This methodological difference, therefore, may have accounted for some portion of the discrepancy between our results and those of earlier studies.

In an effort to duplicate the methods of the earlier studies as closely as possible, a set of analyses was run by using both a retrospective approach and relying on subjects’ self-report of symptoms. Persons with self-reported threat/control override symptoms had higher rates of violence in the previous 10 weeks at baseline (χ²=11.12, df=1, p=0.001; N=1,136) and at every follow-up evaluation. This essentially replicated the results of the earlier studies. Exploring the characteristics of this group, we found that self-reported threat/control override symptoms at baseline had significantly stronger correlations with all four subscales of the Novaco Anger Scale (27) (testing for differences between two nonindependent correlation coefficients [28], t=4.08–6.24, df=1130, p<0.001) and with all three subscales of the Barratt Impulsiveness Scale (29) (t=2.37–3.31, df=1121, 0.05>p<0.001) than did clinician-rated threat/control override delusions. Controlling for the effect of these anger and impulsivity measures eliminated the significant association between self-reported threat/control override symptoms and violence at every measurement point.