Genuine versus Simulated Suicide Notes: An Issue Revisited Through Discourse Analysis

That language or symbolic behavior is an indicator of certain cognitive states in an individual seems to be a well accepted tenet of psychological research. Osgood (1960) has argued that messages produced by people generally covary with a variety of psychological states including intelligence, associations, attitudes, and other emotional states. For some time, researchers concerned with suicidal behavior have been interested in assessing what language indicants may be valuable in determining when an individual may be suicidal and in what ways these indicants may be used to diagnose and treat such states. The writing of a note presents a particularly intriguing problem as this behavior turns what is in many ways an individual act into one that is social. In the late fifties and early sixties, classic research studies were produced which sought to distinguish genuine from simulated suicide notes. These studies essentially sought to: (1) determine whether or not content analytic methods could be used to distinguish notes and (2) determine what specific language indices could be used to indicate a suicidal state. In this article, we wish to: (1) assess these studies in terms of their success in dealing with these issues, (2) present a contemporary method of discourse analysis which provides a unified theoretic view of language as an indicant of cognitive states, (3) present research in which this method is applied to the discrimination of genuine from simulated suicide notes, and (4) present a predictive multivariate model and profile of the suicidal individual based upon their language behavior.

Studies

The three major studies which have dealt with the issue of distinguishing genuine from simulated suicide notes are those of Osgood and Walker (1959), Gottschalk and Gleser (1960), and Olgivie, Stone, and Shneidman (1966). In each of these studies, various content analytic measures were used including the type-token ratio, distress-relief quotient, noun-verb/adjective-adverb ratio, mands, allness terms, and other methods utilized in the General Inquirer (Stone, Olgivie, Dunphy, et al, 1966).

Osgood and Walker found that such indices as stereotyping, use of distress words, use of mands, and use of allness terms were significantly different between genuine and simulated notes. Their research seemed to support the idea that suicidal persons use more nouns and verbs, and that especially male subjects tend to use shorter sentence segments in genuine notes. Gottschalk and Gleser found that genuine notes could be distinguished from simulated by a higher frequency of references to people, things, and places as well as a lower percentage of references to cognitive processes in the genuine notes. Olgivie, Stone, and Shneidman found similar results. They further found that specific words such as “love,” “know,” and “woman” were used in the genuine notes. They also found that specified instructions to others were more often found in the genuine than the simulated notes.

In all of the studies, criteria were generated which would permit the researcher or clinician to analyze a particular piece of discourse to assess a suicidal state. In particular, the Gottschalk and Gleser article provided numerical criteria for assessing such states. They found varying percentages of successful classification depending upon the particular content category or selection process used.

We find three basic problems with these studies. First, though significant differences are found between genuine and simulated notes on a variety of content measures, no attempt is made to provide a statistically significant multivariate model which would provide a quantitative and predictive tool for accurately and consistently making discriminations. For example, no quantitative assessment is made with regard to the combined predictive nature of variables such as the TTR, DRQ, use of nouns, or others. Second, though several of the measures have a theoretical base in either psychological research or traditional grammar, no unified theory of discourse analysis methodology is generated or used. Again, discriminations are made in a somewhat piecemeal fashion where the subjective judgment of the researcher or clinician must be used to make decisions both in terms of what measure to use and when to use it. Third, in none of the studies does a consistent language profile merge which would provide a unified theoretical view of the suicidal person.

In order to adequately deal with these problems, it would seem necessary to first have a language analysis system which would provide a reliable and coherent view of the relations among its measures theoretically, and second, to form research strategies around the idea of uncovering the multivariate predictiveness of these measures in their ability to make discriminations between genuine and simulated notes. In recent times, psychologists and communication scholars have given increased attention to discourse analysis as a viable tool for analyzing the underlying structure of verbal messages. The primary aim of most research in this area is to provide a coherent system of language categories or structural variables which are indicative of certain states of the organism (Freedle, 1977). Such systems generally go beyond the structural notion of “rules” such as in those of Piaget (1972) or Chomsky (1975), and instead concentrate on what is often termed “manifest structure” (Pepinsky, 1979). Modern approaches to discourse analysis thus concentrate on those aspects of language which are by their nature indicants of certain states within the organism rather than the states themselves. Further, advances in the use of multivariate techniques such as multiple discriminant analysis have made it possible to statistically assess the joint predictive ability of many variables in discriminating such states as suicidal versus non-suicidal discourse. Together, we feel that such advances provide a viable alternative for making decisions.

Syntactic Language Computer Analysis

The discourse analysis system which we have utilized in our research was developed by Cummings and Renshaw (1976, 1979). Termed Syntactic Language Computer Analysis, Version III (SLCA-III), the system consists of a unified set of langauge categories into which words of a message are placed. The system utilizes traditional-grammatical categories, but concentrates on differing aspects and functions of these categories within the analysis of the message as a whole rather than specific content categories such as the TTR or DRQ. SLCA-III attempts to derive a “map” of an individual's cognitive states by examining that individual's verbal behavior.

In the SLCA-III system, verbal behavior is viewed systemically as “interdependent with perception and cognition, as a symbolic identifier or marker of discriminable elements and their relations perceived to exist in the real or imaginary environment of a living organism” (Cummings & Renshaw, 1979). An important aspect of this system is the notion of “language markers” which are said to be in essence “mirrors” of perception and cognition. Cummings and Renshaw (1979) further discuss the import of this:

While a full description of the SLCA-III system is beyond the scope of this article, we shall attempt to give the reader a general idea of the terms used, their derivation, and use. Further, where we speak of certain of the categories generated by SLCA-III later in the article, we shall also indicate the nature of that category by describing it in terms of traditional grammar as far as possible.

In SLCA-III, information units correspond to nouns, relational units to verbs, and qualifiers to adjectives and adverbs. In addition to these traditional classifications, each one is said to have certain aspects or qualities. There are basically eight qualities which apply in differing ways to these three classifications. They are: (1) social perception, which refers to a person's ability to perceive and attribute characteristics to animate or inanimate objects; this is generally applied to nouns and is found in terms of the use of proper nouns, different types of pronouns, reference to a thing or person, or to the general “it”; (2) sensation, which refers to what the individual perceives as either concrete or abstract; this generally applies to nouns and adjectives or adverbs and may be illustrated by the use of the term “book” (concrete) or “truth” (abstract); (3) existence, which refers to what an individual perceives to exist or not exist; it is applied to all three basic classifications and is found in the use or non-use of the word “not”; (4) motion, which is generally a classification of verbs in terms of whether the verb indicates an activity or a state of some entity; (5) disposition, which concerns primarily verbs and is found in whether the verb indicates an assertion or a condition (“I will” as opposed to “I could”); (6) time, which refers to past, present, or future and is of course applied to verbs; (7) symmetry, which indicates the intentionality of the language; it is assessed in verbs and basically is found in the transitive-intransitive distinction; (8) definition, which concerns primarily nouns and verbs and indicates whether these units are modified by a qualifier or not.

The SLCA-III system is a computerized discourse analysis technique which takes the theoretic categories above and analyzes a message according to how these categories are applied to each word. The system generates for each message the total number of words plus an additional 36 densities (ratios ranging from 0 to 1 based upon the number of words in a category divided by the total). The analysis uses a combination of dictionary “look-up” plus positional analysis to provide a reliable and consistent set of unified language measures for each message. An example of the output of a SLCA-III analysis of a message may be seen in Figure 1. We feel that the SLCA-III system provides a coherent and workable system for: (1) providing a unified theoretic base of content analytic method which can discriminate genuine suicidal messages from simulated ones, and (2) providing a cohesive system in which a general profile of this pathology may be discerned.

Research

In order to assess SLCA-III in its ability to both discriminate and provide a profile of the suicidal state, we chose a strategy involving basically two levels of research. The first involved our attempt to replicate as far as possible the findings of other researchers. We did this by basically looking at specific SLCA-III variables as they related to categories previously used. Our second strategy was to take all SLCA-III variables and attempt to develop a multivariate model which would both show us which SLCA-III variables were the best predictors of genuine versus simulated notes and which would provide the best profile of suicidal individuals who write notes.

In the first strategy, we felt that basically three aspects of the SLCA-III system would discriminate notes based upon their relation to previous studies. We felt that the basic category of information units (nouns) would show significant differences based upon all three previous studies. Due to the findings in both the Gottschalk and Gleser (1960) and Olgivie, Stone, and Shneidman (1966) studies, we felt that the category of sensation as applied to both nouns and modifiers would show significant differences, and we also felt that the category of what Cummings and Renshaw call “authority other” (use of proper nouns) would show differences.

Method

The notes used in this study may be found in Shneidman and Faberow (1957). In this collection, there are 33 genuine and 33 simulated suicide notes. Each genuine note was matched with a simulated one on the basis of such demographics as age and occupation. All notes were entered into the SLCA-III system using appropriate computer encoding procedures. In order to assess our first research strategy, we followed a similar procedure as in previous research. We conducted a series of t-tests using a two-tailed probability level of 0.05 for rejection of the null hypothesis that no differences existed. In addition, tests of homogeneity of variance were conducted, and in instances where significant variance differences were found, corrected t-tests were used. These t-tests were run on all SLCA-III variables as an exploratory exercise.

In order to develop a multivariate model and a means of assessing a language profile of suicidal persons, a multiple discriminant analysis was performed using all SLCA variables as predictors and the nominal distinction of genuine versus simulated as the dependent variable. The analysis used was a stepwise multiple discriminant analysis using Wilk's minimum lambda as the criteria for entry into the model. A criterion level of 0.05 was set as both the F to remove and the F to include in the analysis. A chi-square value for evaluation of the model eigenvalue(s) was also set at 0.05.

Results

Results of the t-test analysis confirmed our speculations concerning the relation of SLCA variables to those similar in previous research. In addition significant differences were found in seven other SLCA variables. Results of these analyses may be found in Table 1. In this table, we have listed the variable name along with a description in traditional grammatical terms where possible, means for both genuine and simulated groups, and the significance values.

After entry of all 36 SLCA variables into the multiple discriminant analysis, 9 variables were selected on the criteria outlined in the method section. The chi-square value for analysis which produced one discriminant function was 39.173 (p < 0.0001). This discriminant function may be found in Table 2. We have included the variable name, the F to remove, and the standardized discriminant coefficients. This function was of course used to discriminate between the genuine and simulated notes. Following use of the discriminant function to classify the notes, it was found that 27 of the genuine notes were classified correctly and 26 of the simulated notes. This provided a percentage of 80.3 percent correct classification. A Z-test was utilized to see if this percentage was significantly different from a 50-50 chance of correct classification. The Z value produced was 4.9 (p < .0001).

Conclusions

There are basically two areas in which the results of our research may be analyzed. The first concerns the ability of SLCA—III to verify previous findings and provide a discriminatory tool for assessing language indicants in the pathological state of suicide. The second concerns the development of a general language profile of the cognitive state of a person who writes a suicide note prior to committing the act.

In terms of the ability of SLCA-III to make accurate discriminations of genuine versus simulated suicide notes, it would appear that SLCA can be a very valuable tool for use in diagnosis when researchers or clinicians are seeking to determine the probability of suicidal states given a sample of writing made by a patient. Based upon simple analyses similar to those found in Gottschalk and Gleser (1960), these discriminations could be made reliably with a probability level near theirs. The discriminant function produced in our research further provides a quantitative tool for making such decisions. The researcher or clinician concerned with making a probabilistic decision concerning a particular patient may by hand, if necessary, compute the SLCA-III densities for the nine variables, weight these values by the appropriate coefficient in equation form, and have a reasonably accurate and reliable assessment of whether the individual is suicidal or not.

Apart from the conclusions reached above, we feel that an even greater significance is found in our research. It will be noted in the SLCA categories, that fine distinctions, frequently in the nature of opposites (positive as opposed to negative, modified as opposed to unmodified), are made. In the discriminant function, the significant variables give us what we feel is a much more precise language profile as an indicant of the suicidal state than is available in previous studies. As we look at that profile, the reader should refer to both Tables 1 and 2. It should be noted that just because significant differences are found between genuine and simulated notes in a category such as negative relation density (negative verbs), this does not mean that a significant difference is found in the complementary category positive relational density (positive verbs). These distinctions provide an interesting view of the suicidal person who writes a note.

According to Cummings and Renshaw, total words are an indicant of cognitive energy. The finding of significant difference here would seem to indicate that suicidal persons experience a heightened state of cognitive energy. It is not that they necessarily refer to cognitive issues, but that they experience more activity. It would seem that as this energy increases, noticeable changes in general verbal behavior emerge. First, an increase is seen in cognizance of objects, and these objects appear to be concrete. Further, modification of both objects and actions or states of objects increase and these also seem to be concrete. As the modification takes place, it appears to be negative. It would seem that the cognitive preoccupation with concreteness is accompanied by negations of those objects and the actions or states they have. Second, specific people, places, or things are negated as well as the concept “you” if the person speaks of them. Along with this, however, generalized others (they, them) are cognized in a more positive manner. It is as though the suicidal person cognizes the world as very concrete, the people he or she knows are negated, and it is only the “unknown they out there” that are viewed as positive. The sense of alienation from the world in general seems to be strongly indicated in the verbal behavior of the suicidal person. Third, relations among people and objects appear to be viewed with a specific notion of time. Decreases in the use of future references emerge, and conditionality of objects increase. We did not find that references to the past or the present increase for a suicidal person, but only that future references decrease. It is as though a finality of life is cognized so that processing of events in the future become considered less and less.

On the whole, the suicidal person seems to have a need to structure the world in concrete terms and in static action. If the suicidal person is experiencing unresolved or hidden resentment concerning specific others, or if they need to reinforce both the feelings that drive them to the act or their decision to commit the act, making the world concrete may be the only means of assessing reality.

We feel that our findings have shown that discourse analysis systems such as SLCA-III provide a viable alternative to more traditional content analytic measures by producing a cohesive and unified system of language analysis which can be extremely valuable in both research and nragmatic diagnosis of suicidal behavior. We further believe that we have only broken the ground in terms of the implications for application of discourse analysis to psychological issues, not only suicidal behavior, but a variety of pathological areas.