Diagnostic Criteria for Complicated Grief Disorder

Respondents who met the criteria for inclusion in the study on the basis of a telephone screening were mailed preliminary consent forms and a self-report battery. When the packet was returned, the structured interview for diagnostic evaluation was scheduled. The SCID-NP, supplemented by the analogous module of 30 grief symptoms, was conducted by a clinical interviewer who made all relevant axis I DSM-IV diagnoses, rated each of the grief symptoms, and gave a global rating of the presence or absence of a pathological degree of grief response. The grief symptoms are listed in table 1. Each symptom and the global rating of complicated grief were scored on 7-point Likert scales: 0–1=absent, false, or minimal presence; 2–4=subthreshold or subclinical manifestation; 5–7=threshold or true, with a maladaptive manifestation. These definitions correspond to the SCID-NP rating system: our 0–1 ratings to SCID-NP rating 1; our 2–4 ratings to SCID-NP rating 2; and our 5–7 ratings to SCID-NP rating 3.

The SCID-NP interviews were videotaped. Twenty-five were reviewed and scored by a clinical interviewer. Interobserver reliability on rating the 30 grief symptoms was satisfactory (kappa=0.78). The 30 symptoms were divided into three clinically derived categories: avoidance, intrusion, and failure to adapt. These were found, upon analyses of the data, to have satisfactory levels of internal consistency: 1) avoidance of thoughts and experiences related to the deceased (alpha coefficient=0.58) included episodes of numbing emotion and inhibiting themes that could remind the subject of the death; 2) intrusion (alpha coefficient=0.52) included undesired, unbidden images of the deceased and frequent nightmares; and 3) failure to adapt to the loss (alpha coefficient=0.49) included extreme avoidance of new relationships and difficulty in returning to a daily routine of life activities.

The self-report battery included the following two instruments.

1. The Texas Revised Inventory of Grief (31) is an established self-administered questionnaire that measures grief response in reference to a specified deceased person. A 5-point scale is used to indicate degree of endorsement of each of 13 items.

2. The Reaction to Loss Inventory, developed by Horowitz and Field, is a new and as-yet unpublished self-administered questionnaire with 34 items that assess current symptoms, functional capacity, sense of identity, sense of relationship to the person who has died, and control of emotion related to the loss. The Reaction to Loss Inventory reduces many experiences, such as the intrusive and avoidance items found in the earlier Impact of Event Scale (32), to fewer items. The subject endorses each item on a 5-point scale ranging from “no experience” to “a high level of experience” within the last week.

The first step in identifying possible diagnostic criteria for complicated grief was to select a criterion against which to compare each of the 30 newly developed candidate items. This was accomplished by applying latent class analysis. Since there are various applications for latent class models, and these are only recently coming into use in psychiatric studies, we review briefly the methodological advantages and procedures that influenced us to proceed with data analysis in this fashion. Latent class models have been used previously to examine the congruity among potential diagnostic “gold standards” for other psychiatric disorders, such as schizophrenia (33, 34), mania (35), and, most recently, autism in the DSM-IV field trials (36). Estimates of sensitivity and specificity for various sets of potential diagnostic criteria are obtained by modeling the data with the use of maximum likelihood technique computed relative to the latent standard or classification. An expression is obtained for the probability of an individual’s having any of the combinations of observed results on the possible tests of the diagnosis. These expressions are made up of components associated with each of the latent classes that might characterize the subject (i.e., really having, or really not having, the diagnosis). The first component of the probability is associated with a true positive; the second with a false positive. The two components add up to an overall probability of a positive result for a subject of latent positive status. Parallel probability calculations are made for individuals with negative test results. The probability calculations make up the likelihood for the whole data set. The prevalence of the disorder and the sensitivity and specificity of the criteria are estimated from the data. Iterative numerical maximization of the likelihood for all the data leads to the maximum likelihood estimates of these parameters. The latent class model method used in this study is reviewed by Walter and Irwig (37); the program used for the calculations was developed by Szatmari and Walter (unpublished).

The data analysis assumes that the multiple measures of complicated grief being used should overlap for subjects identified as having positive positive, false positive, negative negative, and false negative diagnoses, that iterative convergence will estimate the latent class, and that the actual measure which most closely (singularly) corresponds to this latent class is the best estimator of the phenomenon. After the latent class model is applied to empirical selection of a “gold standard” against which a set of diagnostic criteria may be calibrated, signal detection analyses can be applied to select specific items to serve as diagnostic criteria.

The mathematics of the signal detection method has been described for psychiatric applications (38) and has already been used to model DSM-like sets of diagnostic criteria (39–41). In the signal detection method, it is possible to enter a series of predictors (possible diagnostic criteria) simultaneously and then to pick the predictors that correspond to the highest total predictive value, while monitoring the need to identify further predictors for cases that have not yet been diagnosed by the existing algorithm. Signal detection simultaneously selects diagnostic criteria that yield the highest total predictive value while continuing to select further diagnostic criteria that keep sensitivity and specificity best balanced. In addition, signal detection yields an algorithm consisting of structured “and/or” rules that models procedures typically used in the DSMs and easily used by clinicians.

Signal detection was selected for this study because it enables selection of “whole” items, which is clinically much more useful than a logistic regression approach, for example, which would produce a list of weighted predictor variables. It should be noted, however, that both the latent class model and signal detection approaches are forms of exploratory data analysis, and as such they do not allow testing of “trend” significance but, rather, serve to develop models. In this case, the goal was to produce a model set of criteria for complicated grief disorder that may be tested in future research on independent samples.

Separate latent class model analyses were constructed for data on the subjects 6 months and 14 months after bereavement to determine whether indicators of complicated grief more than a year after loss were stable over that time period. Data entered into the latent class models were four potential “gold standards” for measuring pathological grief. The 6-month data were 1) an above-median total score on the Texas Revised Inventory of Grief (median=50); 2) an above-median total score on the Reaction to Loss Inventory (median=64); 3) an above-median number of symptoms scored in the “complicated grief” range (i.e., ≥5) by clinical interviewers (median=6); and 4) the clinician’s global rating of the presence or absence of pathological grief after the diagnostic interview. Data were divided in this way to reflect nonuniform distributions suggesting a “high” cluster of subjects who could be hypothesized clinically to have markedly more severe symptoms. Data were dichotomized because of the requirements of latent class model analysis for nominal data. Table 2 shows how each of these possible “gold standards” compared to the “latent” standard 6 months and 14 months after bereavement. The comparison is expressed in terms of false negative and false positive rates, as well as the sensitivity and specificity of each standard to the latent model.

The results at 6 months indicated that the number of clinician-rated (SCID-NP) severe grief symptoms (five symptoms or more in a subject) was the closest reflection of the latent model, misclassifying only 3% of cases. A parallel analysis was conducted for the 14-month data (although median scores, and therefore cutoffs, varied from the 6-month data [Texas Revised Inventory of Grief, median=41; Reaction to Loss Inventory, median=57; severe grief symptoms, median=5]). At 14 months, the number of severe grief symptoms also proved to be the best reflection of the latent class model, accurately classifying 100% of cases.

Using the above-median versus below-median cutoff for the number of SCID-NP interview severe grief symptoms as the “gold standard” for detecting complicated grief, we then calculated the sensitivity and specificity of each item 6 months and 14 months after bereavement. The results (table 3) show that most of these symptoms had lower sensitivity and higher specificity, indicating that while any one symptom was relatively infrequent, most symptoms, when endorsed, tended to be endorsed by the subjects who reported multiple symptoms of complicated grief. Between 6 and 14 months after the loss, the sensitivity of items generally decreased and the specificity increased further, indicating that when symptoms were endorsed at 14 months, they were even more exclusively endorsed by those with prolonged problems of complicated grief, as indicated by the “gold standard” derived by the latent class model.

The latent class model analysis identified having an above-median number of severe grief symptoms according to the SCID-NP interview as our “gold standard” for assessing complicated grief. We then used signal detection to identify which specific criteria from among the 30 canvassed by the clinician interviewers with the SCID-NP might best define a diagnosis. These signal detection analyses can yield a set of criteria that are linked by “and/or” rules.

In the analysis 6 months after loss, the signal detection algorithm produced a diagnostic algorithm for complicated grief that consisted of either of the two symptoms with the highest specificity plus either of the two symptoms with the highest sensitivity. Thus, a subject with symptoms that (at least) included a low interest in important activities or strong spells of emotion, as well as unbidden memories or feelings of aloneness or emptiness, would be diagnosed by the computer on the basis of these symptoms as having complicated grief 6 months after loss (sensitivity=0.80; specificity=0.85; total predictive value=0.88). That is, 88% of the subjects identified by the latent class model analysis as having the most severe grief symptoms 6 months after loss experienced one of the above-mentioned combinations of symptoms. It is interesting that these criteria-like symptoms stem from each of the clinically derived symptom groupings of avoidance, intrusion, and failure to adapt. Different individuals and individuals in different phases of response to stress may exhibit different signs of maladaptive grief processing, as suggested by the clinical impressions and empirical findings on subtypes of complicated grief cited earlier. At 14 months, subjects who experienced a strong yearning for the deceased, continued feelings of aloneness or emptiness, trouble sleeping, or strong avoidance of reminders of the deceased would be found by a computer following this algorithm to have complicated grief (sensitivity=1.00; specificity=0.71; total predictive value=0.86). This pattern indicates that the symptoms of grief that were most severe at 6 months were not necessarily the ones most manifest at 14 months, but, as indicated below, the group of individuals showing severe symptoms remained quite stable over time.

Although our results suggest that it may be possible to detect complicated grief at 14 months by the presence of just one or two severe grief symptoms, all subjects who met this criterion experienced a multiplicity of symptoms, with the most prominent symptoms being significantly intercorrelated. It was felt that while it was statistically justified to base the diagnosis on one of the four severe grief symptoms identified by signal detection at 14 months, a conservative, clinician-based, and DSM-like approach would be needed to formulate a set of diagnostic criteria with a minimum of three of these symptoms. The fact that these symptoms are intercorrelated justifies such an approach statistically.

We propose that three or more of seven possible symptoms must be currently intensely present for a diagnosis of complicated grief disorder to be made, as shown in appendix 1. The seven symptoms are those identified in the 6-month and 14-month signal detection studies. The exploratory nature of the data analyses conducted here is another reason to tend toward overinclusion rather than exclusion of potentially informative diagnostic criteria. In addition, the inclusion of more rather than fewer symptoms to reach the diagnostic threshold is likely to further improve specificity (which was 78%) while likely lowering sensitivity only marginally because of the intercorrelation among criteria.

We applied these criteria to arrive at the frequency of the proposed diagnosis. We found that 41% (N=29) of the 70 subjects met the criteria for complicated grief disorder, and 21% (N=6) of these 29 subjects—about one-fifth—had a concurrent diagnosis of major depressive disorder, while 79% did not. Only a slightly higher proportion, 31% (N=9), of these 29 subjects had had a diagnosis of current major depressive disorder at the 6-month evaluation. However, as we hypothesized earlier, many of these subjects with complicated grief disorder (79%, N=23) met the SCID=NP criteria for a previous lifetime diagnosis of major depressive disorder. Of the 41 subjects (59%) without complicated grief disorder at 14 months, 41% (N=17) met the SCID-NP criteria for a lifetime diagnosis of major depressive disorder. We concluded that previous vulnerability to depression predisposed these subjects to formation of complicated grief disorder. Among the 41 subjects not diagnosed as having a complicated grief disorder at 14 months, none received a current diagnosis of major depressive disorder, and only 10% (N=4) had received a current diagnosis of major depressive disorder 6 months after loss.

The rate of lifetime major depressive disorder among the 70 subjects, judged by clinician interviewers at 14 months after loss, was significantly associated with the current diagnosis of complicated grief disorder made by these interviewers (χ²=1.8.0, df=1, p≤0.0001). These data replicate earlier research showing that higher grief distress occurs in subjects with a history of depressive disorders (6). Our subjects were also assessed by SCID-NP criteria for anxiety disorders. The anxiety disorders were few and had no relevant associations with complicated grief disorder. Although 40% (N=28) of our subjects gave a lifetime history of an anxiety disorder, these past experiences failed to be significantly associated with a current diagnosis of complicated grief disorder (χ²=3.0, df=l, p≤0.10).