Human events are physical phenomena, but humans are not mindless organisms responding only to physical reality, and physical laws alone do not explain human action. Humans govern their own behavior through internal representations of reality, and to understand how human actions are structured we have to understand the limiting realities that people create for themselves. This paper is directed toward developing some methodology for studying subjective representations of reality-commonly called "knowledge." The methodology is not intended to discover new knowledge1 but to elicit knowledge from those who have it and to represent various forms it takes. The presumption right from the beginning is that different people have different knowledge, even about the same phenomenon (Kempton, 1986), and that variations in knowledge are of interest sociologically.
Subjective realities traditionally have been studied by field researchers or clinicians who make a substantial research investment in order to cultivate informants and attain a deep enough understanding of subjects' thinking to report the realities that the subjects cannot or would not articulate themselves (McCall and Simmons, 1969). The methodology presented here advances rather than replaces the traditional approach: subjects' personal realities still are the focus, and data still derive from cooperative informants and from investigators who understand the domain of interest. However, computer-based procedures are presented which extend methods in cognitive anthropology (Dougherty, 1985; Werner & Schoepfle, 1987) directed at focusing field work and systematizing collection and analysis of qualitative data. The extended methodology permits systematic analysis of qualitative data, yields qualitative models, and allows qualitative formulations to be tested empirically. The computer technology does not make qualitative research quantitative, though as noted by Conrad and Reinharz (1984: 7) "one of the long-range consequences of the integration of computers and qualitative data may be to break down the polarization of qualitative and quantitative research." As in Carley's similar work (1986), the aim is to apply computers and ideas from cognitive science in order to develop precise and powerful models of social knowledge2.
A basic assumption in all that follows is that subjective realities are logical. To some extent representations of reality must be logical because they account for a logical material world3. Also representations of reality account for social life, and humans collectivity try to shape their social relations so that they are logical, as if materially constrained. However, the premise is not that people so perfectly understand the material world which is logical or that social life is logical like the physical world and people understand it, too, but only that thinking is so permeated with logic that people rationalize nearly everything4. People reason about things, they organize their reasoning into knowledge, and whether their knowledge is right or wrong, scientific or mystic5, it can be analyzed by principles of logic6.
This paper proceeds as follows. First, I present a framework of theoretical propositions and methodological premises in order to delineate how models of qualitative events can be constructed from qualitative data. Second, analyses of specific incidents are described to illustrate the framework. The analyses consist of (1) developing a preliminary structural model from expert judgments about event prerequisites, (2) refining the structural model with observations of actual event sequences, (3) computing event priorities from observational data, and (4) conducting simulations in order to examine implications of a formulation and to discover its weaknesses. In the third and final part of the paper I initiate discussion concerning the utility and limitations of the procedures.
The specific aim is to develop a methodology for qualitative modeling of logical structures that guide human action in concrete situations. It is proposed that this can be achieved by having informants relate and interpret incidents while a computer assists and interprets their interpretations within a theoretical framework. This section discusses some of the theoretical ideas, and the next major section focuses on methodological principles.
An integrating framework is provided by a theory of rational action being developed in cognitive science--the theory of production systems7, introduced by Newell and Simon (1972) and elaborated and applied by psychologists (e.g., Anderson, 1983), sociologists (Axten and Skvoretz, 1977; Skvoretz, Fararo and Axten, 1980; Fararo and Skvoretz, 1984), and computer scientists (e.g., Waterman and Hayes-Roth, 1978). The bare essentials of this theory are as follows.
Action is governed by if-then rules: if a certain configuration of conditions arises, then a certain production occurs. A configuration of conditions is recognized through application of declarative knowledge, which is learned rapidly--e.g., in a single trial. Productions are carried out through application of procedural knowledge, which is attained slowly through practice and the development of habits.
Productions have natural consequences--they cause changes in conditions, and these results also can be phrased as if-then rules: if a given production occurs, then condition A changes from state x to y. Since the completion of a production causes a new configuration of conditions to arise, it is likely to activate a new production, and productions lead one to another in an orderly fashion.
When conditions for more than one production are fulfilled at once, a conflict exists and has to be resolved. Numerous principles for conflict resolution have been suggested, but the simplest is that productions are ordered in terms of priority, and the highest priority production is the one that is activated. In this case the system consists of an ordered list of productions that are appropriate in the situation, combined with a processor who goes down the list from the top until an instantiated configuration of conditions is found. This instigates the corresponding production, and when that production is completed, the processor returns to the top of the list and starts over, now with a new environment that ordinarily will fulfill conditions for a different production.
Production systems are hierarchically organized. Instigation of a production sets goals that are achieved by invoking more micro production systems, and operations of the micro systems produce outcomes needed by the higher order system. Sociologically, individuals provide the micro systems for higher order systems of roles and institutions.
References cited above provide more discussion of the theory, but this sketch is sufficient to say why production systems seem appropriate as a framework for modeling subjective realities. First, production systems with their if-then implication rules are intrinsically logical and appropriate for the study of rational frameworks. Second, the psychological interpretations (provided by Anderson) explicitly treat learning, memory, and forms of knowledge. Third, the interpretations of social roles and institutions (provided by Fararo, Skvoretz, and Axten) suggest that even complex notions of social reality can be represented by production systems. Fourth, despite the apparent simplicity of the framework, it is a powerful way of representing thought, as demonstrated by Newell and Simon's applications in the study of problem solving and by computer science applications in the development of artificial intelligence and expert systems. In general, the production system approach permits representation of knowledge about verbally-defined events8 in models that can generate new and meaningful sequences of events.
The if-then rules of a production system map configurations of states into a set of possible actions. A second set of principles maps actions into configurations of states that represent event consequences. The dual focus on states and actions directly represents declarative and procedural knowledge, and it models events with precision. However, there is a major practical disadvantage to the approach. The changes in states that result from events and that instigate new events are hard to discover and systematize.
Laymen and social scientists rarely describe happenings in terms of the dual representation. Rather, people tend to talk about events as following events9, employing a single mapping that is the product of the two separate mappings used in production systems (Skvoretz, 1984).
(events-->states) X (states-->events) = events-->events
The semantics of words involved in descriptions of events make this feasible. Verbs used to describe actions typically incorporate an understanding of consequences as part of the meaning of the verb (Hanson, 1958, Chapter 3), and a well-constructed description of an event implies the conditions of its occurrence (Turner, 1953).
Something may be lost in simply relating events to events. Events may influence states beyond the system under consideration, generating emissions that make non-system events possible; and system events may depend on exogenous states that cannot be manipulated from within the system. However, these facts do not necessarily create a descriptive problem for people trying to understand events. External consequences and exogenous conditions simply add an element of probability into what otherwise might be a determinate system10.
The focus here is on recovering production systems from the mind of individuals who are using them. Consequently, an analytic framework that is oriented toward event-event contingencies is most useful because it corresponds to the kind of information that people are best able to provide. Instead of the two kinds of rules in production systems relating states and actions, we seek a single kind of rule:
if [conjunction of prerequisite events] then [consequent event]
The preconditions for an event may be established by several events, so we allow multiple events to map into a single event. An event may establish preconditions for several other events, so each event also may map to multiple other events.
Viewed graphically, the event-event mapping is a directed graph representing logical relations among events. Each source node represents an event that is a necessary condition for the events to which it branches. Each target node is an event that implies events branching to it. Constructing an event model involves identifying the events involved in the system (the graph nodes) and defining the implications between events (the graph branches).
The interpretation of a sequence of events--or the generation of a sequence--within the framework of a production system requires several assumptions about how events occur and the conditions under which events do not occur. These are default assumptions: we assume they generally hold but allow for ad hoc adjustments during an analysis.
The production system approach proposes that an event remains latent until all of its conditions are fulfilled. An event has to be "enabled" by satisfaction of its preconditions11. Translated to the event-event framework, this means that an event should not occur until all of its prerequisites events have occurred. The event has to be primed by prior events.
This assumption gives structure to event sequences: in general, events in a system happen only after a combination of other events has occurred previously. Accordingly, the assumption provides a basis for testing how well a provisional model fits an observed event series--something is wrong if an event occurs before the conjunction of its prerequisites.
The assumption itself is one thing that might be wrong. In particular, the required configuration of states for an event, E, might be attained through alternative prerequisite events in which case E is disjunctively related to its prerequisites or to subsets of its prerequisites. We assume conjunction as the default condition, but permit disjunction if that appears to be necessary in order to make sense of data.
Another assumption in production-system theory is that an event's consequences typically include change of the states that instantiate the event, so the system does not get stuck in a loop producing the same event over and over (Skvoretz and Fararo, 1980; Skvoretz, 1984). Anderson (1983, p. 135) provided an alternative interpretation of the same idea. "If data are matched to one production, they cannot be matched to another, including another instantiation of the same production." Still another way of verbalizing this is to say that occurrence of an event depletes the conditions that prime it. In an event-event model this means that the prerequisites of an event are used up by occurrence of the event, and the event itself must be used up by one of its consequences before it can repeat.
This assumption also adds structure to an event series: an event cannot be repeated until some of its consequences have been implemented. Accordingly, the assumption provides another basis for testing a provisional model against series data. Something is wrong if we find two occurrences of an event without an intervening consequence.
This, too, is a default assumption, and one way of fixing a problem of non-depletion is to change the assumption. We may specify that depletion by a particular event does not occur along branches to some of its prerequisites, or that some events are repeatable without depletion12.
I assume that graphs representing event structures are acyclical--without loops. This seems to accord with common-sense thinking about event structures: incidents, rituals, tasks, etc, generally seem to progress from a beginning to an end. A practical reason for the assumption is that the methodology for eliciting event structures requires real-time processing of graphs, and acyclical graphs are easier to handle computationally13.
Real event systems do involve some degree of cycling, however. For example, enter-a-room is a prerequisite for leave-the-room. Leave-the-room depletes enter-the-room and also is a prerequisite for a repetition of enter-the-room. If the room is re-entered, then that entry depletes the last leaving. Thus, after an initializing first entry, leaving and entering are prerequisites for each other, they deplete each other, and only one event is instantiated at a time.
Such pairs of events can be represented in an acyclical graph by making the event that has to be initialized the prerequisite for the other (thereby according with common sense). Then the branch between the two is specially flagged so that the two events deplete each other. Moreover, one event in the pair always is treated as instantiated so it cannot occur again until depleted by the other. I call these specially-flagged branches commutations.
Ordinary branches involving one-way depletion are the appropriate default assumption in modeling event structures. Commutative relations are added only as they are needed to account for observed processes. The need usually is discovered when some event does not deplete one of its prerequisites, and we search for the event that does deplete that prerequisite. For example, enter-the-kitchen is a prerequisite for wash-the-dishes but wash-the-dishes does not deplete enter-the-kitchen. What does? Leave-the-kitchen depletes enter-the-kitchen, and these two events are commutative.
Constructing a model of qualitative events involves, first, listing the relevant events, and second, pairing events in order to ask whether one event implies the other.
Listing events that make sense together is not as simple as it seems. Words have to be chosen carefully in describing events so that the descriptions evoke background knowledge about preconditions and consequences, otherwise the logical implications between events will be obscure and perhaps impossible to define. In fact, sensible event descriptions may require a specialized lexicon that incorporates appropriate knowledge (Agar, 1974). The general problem is that there are no truly objective descriptions of events, only descriptions that emphasize certain aspects of behaviors and their consequences, and the biases have to be delicately matched across events if the events are to form a coherent logical system. At the same time, event descriptions that are phrased appropriately may be too esoteric for a non-expert to interpret in order to assess implications. (For example, does "booting gravy" imply "hitting a rope"? It does according to Agar's, 1974, description of the junkie lexicon.) Only a person who knows the lexicon can see the implications hidden in it.
The definitional problems mostly can be avoided if we rely on experts to define events in the domain of interest. Expert descriptions usually will be linguistically framed to represent required background knowledge. If the expert is a layman--an informant--then the hard work of inventing concepts and consistent projections between concepts (Katz, 1972) has been done within the informant's culture or subculture. If the expert is a social scientist, then this work is accomplished during the long process of coming to understand the community of interest and being able to articulate its happenings.
Experts also can identify readily the prerequisites and consequences of events, recovering this information semantically from the special words that they use or by reviewing their store of memories. The selective perceptions and conceptualizations that lead to logical inference in an event system are second nature for experts.
I am suggesting that production system models of social scenes be elicited from social experts using the language that the experts use to talk about and think about their experiences. Then production system representations are a kind of extension of verstehen methodology. They are objectifications of indigenous knowledge.
Pairing events to determine implications creates practical problems. Event structures may have fifty or more events (Clarke, 1983), and therefore questions would have to asked about thousands of pairs14. This might be impractical, especially with informants.
A partial solution to the practicality problem--and to several other problems as well--can be achieved by focusing always on specific incidents. I suggest that the appropriate way to objectify event structures is like the Supreme Court refines law: not in the abstract trying to figure out all contingencies, but in response to real cases. An analysis should be based on a record of events in an incident. As the incident unfolds, each new event is added to the set of system events, and logical relations between the new event and prior events are defined15.
Analysis of events in time-ordered sequence permits a drastic reduction in the number of relational questions that have to be asked. We never have to ask if the current event, E, is necessary for earlier events. If we are dealing with the first occurrence of E in an incident then we can assume it is not required for the earlier events, relying on the general metatheoretical principle that later events cannot be prerequisites for earlier events (McCullagh, 1984; Heise, 1975). If E occurred previously in the incident then we need not analyze it at all since it already has been analyzed. In this way time-ordering reduces by half the number of relational questions that have to be asked16.
Focusing on incidents also addresses another problem--that events in a system should be equally significant or relevant. For example, if a religious ritual includes "choir sings hymn", we presumably would not want to include the microactions involved in the priest moving from one place to another. Yet we have no general theory about how to chose relevant events. Indeed, even the example can be turned to prove the point: in some religions the microactions of priestly motion are critical parts of ritual, on a par with group singing.
The relevance problem is addressed when experts are analyzing real incidents because experts typically report just the events that are necessary to relate what happened17. Any single incident may not involve all of the events in the system, but multiple incidents can be pooled to approach the full set. At every step in the process we obtain a model that accounts for real happenings.
People are not very good at computing long chains of implications. They are prone to say that one event does not imply a much earlier one even though a transitive implication does exist through other events, and they accept their error when the chain of implication is made explicit. Consider this example. Does locate-car-key have to happen in order that eat-Peking-duck happens? An informant in a suburban community might offer a quick no because the two events seem too disparate, yet accept the argument that you have to locate the car key to drive the car, and drive the car in order to get to a Chinese restaurant, and get to a Chinese restaurant in order to eat Peking duck--so the answer should have been yes. Some people may handle chains of this length without difficulty. Yet the fact remains: a "no" merely means the informant does not see a relation between two elements; "no" does not necessarily mean that an implication is absent. With long chains of events, informants sometimes lose the logical thread--perhaps because short-term memory gets overloaded--and do not see a relation where one actually exists (Rips, 1983).
"No" has to be regarded as "maybe no, maybe yes" when eliciting implicational relations. The answer "yes" does not generally have this character (unless the informant is guessing). "Yes" indicates that the informant sees a specific connection between two elements, the implication is recognized within the reality constraints that the informant takes for granted. "Yes" answers generally occur in response to questions about elements that are logically close, linked by short chains of inference at most. The invalid "no" arises with elements that are logically distant in the sense that a chain of inferences is required to show that the right answer is "yes". Thus this might be called the principle of local logic. People can judge the truth value of propositions involving short chains of inference, but do not reliably report the truth value of propositions involving long chains of inference.
The principle of local logic means that if we pair all events in a system and ask about implications, the resulting set of answers might be logically inconsistent. The relative trustworthiness of "yes" and "no" answers suggests that the problem can be handled by treating the "no" answers as less valid. The logical structure should be developed predominately from "yes" answers, ignoring "no" answers when they conflict with the implications of "yes" answers.
In theory, the burden of pairwise questions about implications can be reduced by making use of implications already identified during the elicitation process. Given some answers, others can be computed, and many of the possible questions need not be asked at all.
Here is an example of how questions may be eliminated. An informant says event Y is essential for event X. She also tells us that event Z is essential for Y. If we question too mechanically, we could go on to ask if Z is essential for X. Yet that question need. not be asked because the answer can be derived from available answers--Z is essential for Y, and Y is essential for X, therefore Z is essential for X. This is simply the classical logical syllogism: X implies Y, Y implies Z, hence X implies Z.
Logic allows two kinds of calculations to be made. The first is like that in the example above. If an event is a prerequisite for event Y, then it also is a prerequisite for all of Y's consequences, and we need not ask questions about those relations.
Here is an example of the second kind of derivation. Given the ZYX chain from above, we want to add event W. Is W essential for X? The informant says, Yes. Is W essential for Y? The informant says, No. Now there is no need to ask if it is essential for Z because Z is a prerequisite of Y. If an event is not a prerequisite of Y then it cannot be prerequisite for any of the prerequisites of Y. Thus, logically, the answer "no" permits derivations of answers to unasked questions once we know some of the implicational structure.
However the principle of local logic means that we cannot generally assume that an event reported to be a non-prerequisite of Y is necessarily a non-prerequisite of Y's prerequisites. Indeed, my awareness of local logic arose from computerized elicitations that made use of the second kind of computed answer and yielded different graphs when element pairs were considered in different orders. Eliminating computed answers based on "no" answers eliminated the inconstancy in results.
In practice, the only answers that can be inferred validly instead of being obtained through questions are those that rely on relations established by prior "yes" answers. Thus employing the laws of logic can reduce the tedium of elicitations but not quite as much as one logically might suppose. Nevertheless, the reduction in questions achieved by using just the first principle, in conjunction with the reduction obtained by analyzing time-ordered events, makes the elicitation of qualitative event models practical.
Qualitative data in sociology sometimes are presumed to have an integrity beyond other social data. For example, a single inconsistent observation is enough to instigate revision of a model in the method of analytic induction (Robinson, 1951). Humans seem to be such sensitive measuring instruments that their qualitative perceptions are trusted implicitly. There are reasons not to deviate too far from this standard--especially the maxim that social reality is what people perceive, whether it is true or not. However, there also are reasons not to reify any data based on human observations. Even trained observers do not catch everything that is going on at a scene, people may infer events that "must" have happened, and one person's interpretations of actions may not agree with others'.
Historians have a fairly elaborate doctrine about error in qualitative data (McCullagh, 1984)--not a theory but a variety of heuristic tests, along with an acute consciousness that records can be wrong regardless of who the recorder was. An historical fact never is to be believed absolutely: it is simply an interpretation of a configuration of evidence that is subject to change, an inference that can be overwhelmed by logical counter-inferences.
In the spirit of historians' skepticism about records, I propose that the record is one thing that can be changed--selectively and thoughtfully--during an analysis in order to achieve consistency between a model and qualitative data. As we begin developing refined, highly specific representations of action systems, we may find that an event seems missing at a particular point in the record, and a sensible solution to this problem might be to suppose or remember or recognize that the event did occur in fact. Similarly, an event may seem out of place: having considered and rejected other ways of dealing with the problem, it may be reasonable to suppose or remember or recognize that the event did not occur at the point where it was recorded.
Quantitative researchers, with their elaborate theories of error (e.g., Lord and Novick, 1968) do something like this routinely: any particular observation is presumed fallible. A model is estimated from overall patterns in the data along with assumptions that specify the model, and any particular inconsistency in the data is no cause to reject the model: the analytic procedures allow a measurement's true value to have been different from what was recorded. I propose no more than this for qualitative data: that a model be developed from patterns in data along with assumptions about the event system, and when an observation is inconsistent with the general logical structure, its "true value" may be derived and substituted if that does not strain credulity.
One might try to measure event priorities in a production system directly by having experts rate the events in terms of importance or in terms of how motivated the events seem to be. Axten and Skvoretz (1980) took a different approach and tried to derive priorities by analysis of actual behavior traces. Their analyses employed guidelines like: events with fewer prerequisites should have lower priority; higher priority events will occur more often and with shorter intervening periods.
The Axten-Skvoretz approach relates to actual observations of events rather than someone's memory of events, and this seems preferable in eliminating possible biases18 and in eliminating an additional task for informants to perform. However, their heuristics for computing priorities might be improved by taking explicit account of event structuring and progress.
At a given moment relatively few events in a production system are practicable; many of the events are dormant because they are unprimed or because they occurred previously and have not been depleted. An event's occurrence is evidence that it has priority over other practicable events but provides no information about its priority with respect to dormant events.
Thus in a prioritization analysis we should start at the beginning of an event series, determine which events can happen on each event cycle, and then give a measure of priority to the event that happens--relative to the other events that were possible at that point. Information about priority builds up from order of occurrence among the practicable events. Event repetitions improve the statistical quality of analyses, but frequency of repetition does not directly reflect priority since a low-priority event might repeat often if everything else is dormant.
This approach can be operationalized as an event-event count matrix with cells reporting the number of times in an event series that the row event took precedence over the column event. Since some events may never have been possible at the same time, some of the cells remain undefined. The analysis of this matrix in order to obtain an overall ranking of events by priority presents interesting challenges to mathematical sociologists. Provisionally and crudely, the ranking can be derived as follows. Let C be the square of the count matrix with ones in the diagonal (squared in order to reduce the number of unrelated events). Then quantities [Cij / (Cij + Cji)] are averaged over all row entries, j, where Cij > 0 and Cji > 0; and these row indices are sorted to define event priorities.
The modeling and methodological assumptions will be illustrated by analyzing some interaction. The events are purposely problematic in order to serve as a medium for discussing issues and problems, with no goal beyond exposition. The topic--encounters between professors--is of no particular consequence, but I am a fair expert on it, and the reader may be also, which is useful in the exposition.
Analyses were conducted using a computer program for eliciting and analyzing implicational structures19. The program employs a modifiable verbal framework in order to elicit verbally-defined elements in a desired domain and the logical relations among elements. During elicitations the program employs past answers about implications to minimize further questions (with the presumption of local logic). The program analyzes an event series for consistency with an obtained structure and then again to compute priorities. The program also allows completed models of event structures to be used for simulations.
The analysis proceeds in three phases. (1) Entering the events and determining their logical relations. (2) Analyzing the series of events in order to make it consistent with the implicational structure and other modeling assumptions. (3) Reanalyzing the series in order to assess event priorities. The completed model is used for simulation.
The data concern two routine interactions between professors. The two interactions are presumed to have the same underlying structure, so they can be pooled in order to increase reliability of results.
In the discussion that follows, italics show the events as typed into the program; abbreviations for the events are in capitals; text in angular brackets elaborates event meanings; and text in square brackets offers other kinds of explanatory comments. The interaction between program and analyst is discussed after the entry where appropriate.
Professors encountering PRO
[Name of incident. This becomes the root node for the directed graph representing the implicational structure.]
A enters room ENT <Professor A enters the department office>
[No questions are asked about the first event. The program simply draws a graph on the computer screen showing PRO branching down to ENT.]
A engages in task ENG <Professor A sorts his mail>
This event has to be structurally located relative to ENT. There is no need to ask whether ENG implies ENT since ENG occurred later. The program asks "Does A engages in task require A enters room (or a similar event)?" The analyst answers "yes", and the program draws a revised graph: PRO-ENT-ENG.
Entering a department office with mailboxes is being viewed as necessary in order to engage in the task of sorting mail. More generally, it is being assumed that entering a room is necessary for the kinds of tasks that professors do. These kinds of background understandings by the "expert" analyst guide answers to the related questions.
The phrase "or a similar event" is included in the question-frame in order to allow that an event might have disjunctive prerequisites. If so, a person might be inclined to say one particular prerequisite is not essential because equally good alternatives exist. The explicit allowance for alternatives counteracts the inclination.
B enters room E1T <Professor B enters the department office>
[The digit in the abbreviation replaces a letter in order to distinguish this form ENT.]
The program asks if either of the prior events is essential for this one, and the analyst answers "no" to both questions. The program draws a new graph in which PRO branches to ENT (which branches to ENG) and additionally PRO branches to E1T.
A greets B GRE
The program asks whether "A engages in task" is essential for the greeting, and the analyst answers "no". The program also asks if the entries of A and B are essential for the greeting. These questions are answered "yes" from the perspective that both persons have to be in the same physical location for a greeting to occur. Once again, extra background knowledge is being added to the skeletal information provided by the phrasing of events in order to answer relational questions.
A new graph is drawn at this point and also at later points, but the graphs are becoming too complex to describe verbally.
B greets A G1E
The program asks if GRE and ENG are essential for this greeting, and the analyst says "no". However, B's greeting, like A's, is made dependent on ENT and E1T.
B positions for talk POS <B stands beside A, partially facing him>
The program asks if A's engaging in a task is essential for this, and the analyst says "no". Then the program asks if the two greetings are essential, and the analyst says "yes", taking the perspective of an expert who "knows" that professors do not try to strike up conversations without prior greetings.
The program asks no further questions. Since B's positioning requires the greetings, it logically requires the entries of A and B, and questions about ENT and E1T need not be asked.
A inquires about B INQ <A asks about B's recent overseas trip>
The program asks if A's engaging in a task is a prerequisite. and the answer is "no". Next the program asks if B's positioning is required, and the analyst answers "yes", reasoning that Professor A could have continued sorting his mail after the exchange of greetings, had not B positioned for talk. Thus POS becomes a prerequisite for INQ. (If enough incidents were studied, it probably would be found that positioning by either A or B is required for an inquiry--either person can instigate a conversation, so B's positioning is a disjunctive prerequisite.)
The program has sufficient information to place the new event structurally, and it asks no further questions.
B talks about self TAL <B commends his trip and his sabbatical>
The program asks if A engaging in a task is necessary for this; the answer is "no". The program asks if A's inquiry is essential, and the analyst answers "yes". No further questions are asked.
B inquires about A I1Q <B queries A about his upcoming sabbatical>
The analyst answers "no" to questions about whether this act depends on A engaging in a task, B talking about self, and A inquiring about B. The event is seen as requiring B positioning for talk. No further questions are asked.
A talks about self T1L <A affirms his sabbatical and his desire for it>
Again the analyst answers "no" to questions about whether this depends on ENG, TAL, or INQ. T1L is made dependent on I1Q.
B engages in task E1G <B turns to sorting his mail>
The program asks whether the events that have happened so far are essential for this and receives "no" answers for everything but "B enters room".
Logically, some of the questions could have been bypassed after the first few "no" answers, but the program presumes the analyst is subject to errors of local logic and therefore does not use the "no"s as a basis for inference.
A engages in task <A returns to sorting his mail>
[A repetition of ENG. The second occurrence is logged into the event series, but no questions are asked because the event already has been structurally positioned.]
B positions for talk <B pauses by door, facing A>
[A repetition of POS.]
B leaves room LEA
The program asks about every event as a possible prerequisite for this. Only "B enters room" is accepted.
Professors encountering
[The name of the incident is repeated to signal we are starting a new incident which presumably has the same structure as the first--a replication study.]
A enters room <Professor A enters the commons room>
[A repetition of ENT, now with a different room. The event is logged into the event series, but no questions are asked since ENT was positioned while analyzing the first incident.]
A engages in task <A sits down and begins grading student papers>
[A repetition of ENG with a different professorial task.]
B enters room <Professor B enters the room>
[A repetition of E1G. This is a different B than in the first incident, so B simply stands for "a second professor".]
A greets B
[A repetition of GRE.]
B greets A
[A repetition of GlE.]
B positions for talk <B sits down, facing A>
[A repetition of POS.]
A inquires about B <A asks B about his general welfare>
[Repetition of INQ.]
B talks about self <B says some problems are getting sorted out>
[Repetition of TAL.]
B inquires about A <B asks what A is doing>
[Repetition of I1Q.]
A talks about self <A describes his grading task>
[Repetition of TIK.]
B talks about self <B worries about upcoming lecture>
[Repetition of TAL.]
A supports B SUP <A expresses confidence in B's competence>
This is a new event and has to be positioned. Since it occurs during a replication series, the program discards assumptions about time-ordering and asks whether it is a prerequisite for other events in the system as well as whether other events are prerequisite for it. After all the questions are answered, SUP is seen to be directly dependent on TAL, and not essential for anything.
B positions for exit P1S <B stands up>
Another new event, this one is treated as depending on "B enters room" (which ultimately will be found erroneous).
A gives parting routine G1V <A says "So long" >
Another new event. During the questioning, a parting routine is judged to be necessary for leave-taking. A's "so long" is viewed as dependent on the combination of B's positioning for talk and B's positioning for exit. (These specifications will be modified during the series analysis.)
B positions for talk <B pauses by the door>
[Repetition of POS.]
A engages in task <A returns to grading papers>
[Repetition of ENG.)
B leaves room
[Repetition of LEA.]
Phase 1 results in a directed graph representing the implicational structure that was defined during the questioning process. Figure 1 shows the diagram printed by the computer program.
Some special conventions are employed in the computer's representation of the graph so that the program will operate with a variety of computer hardware. Instead of a separate line for each branch, a source node drops to a horizontal brace line, and vertical lines drop from the brace line to each target node. Arrowheads are not included in the diagram: they would be at the top of the vertical lines ascending to source nodes in order to represent implication.
The logical structure developed in Phase 1 may not account for the event series. For one thing, specifications added for the replication analysis may not fit the first series. More generally; the complete production model includes default assumptions about priming and depletion, and the defaults may not work with the structure that was specified. Thus, in Phase 2 we work through the two series of events to see whether the model could have produced them.
The computer automates this process. Each event is taken in sequence. If the event is possible according to the model, then the program computes how it depletes events and primes other events, and goes on to consider the next event. This continues until an event is supposed to occur which is either not primed or not depleted. At that point the program interrupts its processing, reports the problems, and offers the analyst some feasible ways to correct the difficulty.
The following analysis shows how a series of analysis can raise questions about the original specification of logical structure, about the default assumptions, and about the data being analyzed.
A enters room [No problems.]
A engages in task [no problems.]
B enters room [No problems.]
A greets B [Unprimed.]
The greeting requires the presence of both people, but A's entry got used up by his engaging in a task (according to the default assumptions). The program offers several possible solutions.
First, the program suggests that A entering and B entering--both requirements for a greeting--might be disjunctively related to the greeting: only one or the other is required. The analyst rejects this option.
Second, the program suggests that A entering might not be required after all for A engaging in a task or for A greeting B. If it is not required for the task, then the task would not have used up A's entry, and the greeting would be properly primed. If A's entry is not required for the greeting, then it would not matter that the entry is used up. The analyst rejects these solutions, too.
Third, the program suggests that maybe A entered the room again after engaging in his task; that is, a second unrecorded entry by A occurred that primes the greeting. This is rejected.
Now the program suggests that perhaps A can engage in a task without using up his entry into the room so the entry still is undepleted and available to prime the greeting. The analyst accepts this, and the program presents all consequences of A entering the room to find which one depletes the entry. In fact, none do because we never entered "A leaves room", but the program allows us to specify the root node, ENT, as a dummy depleter.
At this point, the problem is solved. The program restarts at the beginning (just to make sure that the change in the depletion assumption would not create problems earlier in the sequence), works through all the events again, and this time accepts A's greeting.
B greets A [Unprimed.]
Now we have a similar problem for B's greeting: B's entry got used up by A's greeting. Again the program presents some possible solutions, and it is decided that A's greeting does not deplete B's entry. "B leaves room" is defined as the depleter of B entering, and the program establishes commutation between these two events. After starting at the beginning and repeating the sequence, the program accepts B's greeting.
B positions for talk [No problem.]
A inquires about B [No problem.]
B talks about self [No problem.]
B inquires about A [Unprimed.]
B's inquiry requires proper positioning, but that got used up by the previous talk. The solution is that A inquiring about B does not use up B's positioning. The analyst specifies that B positioning for exit is what depletes B positioning for talk.
A talks about self [No problem.]
B engages in task [No problem.]
A engages in task [Unused.]
A engaging in task has not been used up from last time. The program suggests that maybe this event is repeatable without depletion, and the analyst accepts this solution20.
B positions for talk [Unused. Unprimed.]
B positioning for talk has not been used up because B has not positioned for exit. From the various solutions offered, the analyst chooses a change in data--really, B did position for exit from the conversation before B engaged in his task. Evidently "B positions for exit" should be renamed "B positions for termination of talk".
The interruption continues even after this correction because positioning for talk at this point is not primed--the earlier positioning used up the required greetings. The analyst decides that positioning for talk does not deplete greetings; that greetings are depleted only by a parting routine.
B leaves room [Unprimed.]
B leaves, but A hasn't given a parting routine. The analyst decides that B can leave without a parting ceremony, contrary to his opinion in Phase 1.
A enters room [Restarting on the replication series. No problem.]
A engages in task [No problem.]
B enters room [No problem.]
A greets B [No problem.]
B greets A [No problem.]
B positions for talk [No problem.]
A inquires about B [No problem.]
B talks about self [No problem.]
B inquires about A [No problem.]
A talks about self [No problem,]
B talks about self [Unused. Unprimed.]
B talks about self is undepleted. The analyst examines the various offered solutions and decides the best is a change in data. He recalls that he did offer support just before B talks about himself some more; thus B talks about self was depleted.
The event also is unprimed. The analyst decides again that the record is inadequate. A did also make an inquiry just before B talked about self again.
A supports B [Unused.]
Since "A supports B" was added as a solution to the last problem, we now have two of them. The only solution offered by the program is to make A's support repeatable, and this is accepted.
B positions for exit [No problem.]
A gives parting routine [Unprimed.]
A's parting routine needs "B positions for talk" undepleted, and it is not. The analyst accepts the program suggestion of revising the structure so that "B positions for talk" no longer is a prerequisite for A's parting routine.
B positions for talk [Unprimed.]
Positioning for talk is not primed at this point because the parting routine depleted the greetings. As the program offers suggestions, the analyst confirms that greetings are required for positioning, parting does deplete greetings, and second greetings didn't occur. That exhausts the program's ideas about a solution, and evidently the event record is wrong somehow, so the program switches to a routine that allows the analyst to change the series data. The analyst decides that this event did not really occur--he just imagined it as an observer. An alternative would be that the observer confused the order of events, and this event needs to be transposed with the last event (this would be a way of dealing with the last problem as well).
A engages in task [No problem.]
B leaves room [Unprimed.]
This event is not properly primed by "B engages in task"--B left without doing anything. The analyst decides that B saw talking to A as a professorial task, and E1G is inserted in the event record before B positions to leave.
That completes the serial analysis: all of the events now are accountable by the revised model given in Figure 2. Examination of the new model reveals that serial analysis of event records changed the implication structure, some of the default assumptions, and it changed portions of the events record as well. A more realistic model resulted--one better able to handle relevant events--and this became evident in the replication series with its lower proportion of problematic events.
Modifying the record of events may alarm some readers, but I attest that the serial record of events is unreliable at the level of specificity demanded by the model, even given the notes that I made promptly after each incident. Of course, the only safeguard against erroneously disfiguring data is to obtain sound-image records (Grimshaw, 1982) that can be consulted anew before changing a transcription of happenings.
Once model and data are consistent, the data can be reanalyzed to obtain priorities. We go through a series step by step to see which events were possible on each round and which event actually occurred, thereby accumulating evidence about how events take precedence over others. Though the computer program does this automatically, I will examine the process to elucidate what is happening. Table 1 presents the required information concerning the first incident of professorial encounter.
At round 1 only the entering-room events are possible because only they can be done without being primed by other system events. Consequently the occurrence of "A enters room" provides little information about its priority with respect to other events, except for "B enters room"21. The same is true at round 2. "A engages in task" happens, but this does not mean necessarily that it has high priority. In fact, looking across the row for ENG we see that "A engages in task" is often possible but seldom actualized once another professor is in the room, suggesting that the event has relatively low priority. Similar considerations apply at round 3, where only A's task activity and B's entry are possible.
More possibilities open up with the dual presence of both professors at round 4: either can work or greet the other. Thus the occurrence of a greeting at this point indicates that greetings take precedence over working. The counter-greeting in round 5 provides additional evidence of the priority of greetings over work.
On round 6, after greetings have occurred, the professors can work, they can instigate a parting routine, or they can maneuver for conversation. Preparing for conversation is what actually happens, indicating that this is more compelling than the other options. Similarly, round 7 indicates that initiating a conversation dominates working or positioning for a termination of interaction.
Round 8 indicates that talking about self occurs as soon as it is possible, taking precedence over working, querying the other, or terminating interaction. Round 9 indicates that querying has higher priority than working, terminating interaction, or giving social support. Round 10 shows again that talking about self has high priority--higher than working, terminating interaction, making inquiries, or giving support.
Round 11 is interesting because the same events are possible as in round 9, but this time terminating interaction takes precedence over other events, including querying--the dominant event in round 9. The conflicting information from the two rounds is combined statistically at present, but it might be better to allow for time-varying priorities--like time-varying utilities in rational choice models (Pugh, 1977).
Rounds 12 and 13 indicate that working takes precedence over giving social support, termination proceedings, or maneuvering for more talk. Again we have conflicting information, since maneuvering for talk took precedence over work in round 6. Round 14 continues to provide conflicting information about maneuvering for talk since it now does take precedence over other possible events.
Round 15 indicates that B leaving the room can take precedence over conversational events and over A working.
The computer analysis continues this kind of analysis through the second incident and keeps a quantitative tally of the information gained, then ranks the events in terms of observed precedences, as described previously in the section on methodology. The final results are as follows, with high priority events listed first and approximate ties shown in parentheses: (TAL T1L ENT) (INQ POS) (GRE I1Q P1S) (SUP G1E LEA) E1G GIV (ENG E1T). The results suggest, for example, that talking about self is very high priority in professorial encounters and engaging in tasks is not.
Some obvious flaws remain in the model--e.g., one professor's work has been defined as repeatable and the other's has not. We might try to clean up such problems by examination and direct modifications, or we could analyze more incidents until we are confronted with inconsistencies during series analyses and are forced to correct them. In general, a model always should be treated as provisional, able to account for the data that generated it but subject to change as new data are processed.
In principle, an adequately defined production system accounts for future events as well as events that happened in the past. Indeed, we can use the model to explore possible futures. A completed model--consisting of a logical structure, assumptions, and event priorities--produces a unique sequence when it generates events without outside interference. A model also can generate other sequences that are logically correct while not according with model priorities at one or more points. Such simulations of events can be conducted in order to explore a variety of happenings within the system22. Simulations also are useful in discovering troublesome features of a model.
The model for professorial encounters now can be employed to illustrate simulations. Abbreviations for the events that are possible at each point in time are listed on the left in order of priority; the event selected for "implementation" is italicized. Comments on the right describe what is happening according to the simulation and also discuss some problems with the model that are raised by the simulation analysis.
| POSSIBLE EVENTS | COMMENTS |
| ENT E1T | At the beginning either professor may enter the room. No other events are primed yet. I chose "A enters room." |
| ENG E1T | Now that A is in the room, he may engage in a task, or B may enter the room. I had A engage in a task. |
| ENG ElT | The same options appear again since ENG was declared repeatable. The lower priority event has to be chosen in order to make the simulation progress. I selected "B enters room." |
| GRE GIE E1G ENG | Either professor may greet the other, or else engage in a task. I had A greet B. |
| GlE E1G ENG | B may return A's greeting, or either may engage in a task. I had B return the greeting. Modeling problem: People probably do not engage in a task when a return greeting is due. Perhaps the conceptualization of greetings needs to be revised so that either A or B can "invoke a greeting ritual", where this is a single event--a subroutine that has to be done jointly. Then other events would not be possible while B greeting and return-greeting are in process. |
POS E1G GIV ENG |
B can position for talk, either professor can engage in tasks, or A can initiate a parting routine. I let B position for talk. |
| INQ P1S I1Q E1G ENG | After B positions for talking, he can decide not to talk and undo the positioning for talk. Also at this point, either can inquire about the other, or either can engage in a task. I had A inquire about B. Modeling problem: Engaging in a task is not polite when one party has positioned for talk. Possibly gaze events have to be included to deal with this--working requires "looking at work" and talking requires "looking at other". The gaze events would be commutative with each other, so embarking on talk would make working unprimed, and starting to work would make talking unprimed. Of course, we would have to observe interactions to see if gaze operates like this. Because of this problem, E1G and ENG persist in rounds devoted to social interaction. I will not discuss them until they become relevant again. |
| TAL P1S I1Q E1G ENG | After A 's inquiry, B can talk about self, signal the end of talk, or inquire about A. To accept the last two as options here, we have to understand that conventions of discourse require at least a glossed answer to A's inquiry. (For example, "How was your trip?" "Fine. Are you going to the faculty meeting?") I allowed B to talk about self. |
INQ P1S I1Q SUP E1G ENG |
Now either can inquire about the other, B can signal the end of talk, or A can offer support. I had A inquire about B again. |
| P1S I1Q SUP E1G ENG | After A's second inquiry, B can signal the end of talk, B can inquire about A, or A can give support. (Interestingly, B has to evade A's last inquiry because his talking about self is undepleted--perhaps corresponding to the uncomfortable sense that one is talking too much about oneself.) I had B position for exit. |
POS SUP E1G G1V ENG |
Now I had B act indecisive, repositioning for interaction after positioning for termination of talk. At this point A could give support or say goodbye. |
P1S I1Q SUP E1G ENG |
Here I had B engage in a task. Modeling problem: These options are the same as two events ago--a loop due to POS and P1S being commutative. A possible solution would be to find another prerequisite for B's positioning to talk besides B's entry into the room--something that is now depleted so that POS will not be primed. In any case, we have to go down in priorities to order to leave the loop. |
P1S I1Q SUP LEA ENG |
Again we have to drop to a low priority event in another line of action in order to progress beyond the loop. I selected "B leaves room." |
I1Q SUP ENG E1T |
Modeling problem: Now that B has left, all of the events involving him should be unprimed. Evidently we have to force directly dependency of events involving B (like I1Q and SUP) on "B enters room" (E1T), even though such dependency is derivable through transitive logic. A function would have to be added to the program in order to do this easily. |
The example simulation and prior analyses demonstrate how event structure
modeling focuses attention on complexities of interaction. Far from concealing
details, this methodology instigates a meticulous concern with particulars that
is uncommon--perhaps unprecedented--in qualitative research.
Event structure models go beyond ordinary verbal descriptions of incidents in two ways. First, the methodology is systematic. Event-structure analysis generates a model that accounts fully for the event sequences used to develop the model. Moreover the representation can be tested and improved by analyzing more data. These features may not obtain in ordinary ethnographic descriptions. In fact, after seeing how tedious and difficult event-structure analyses can be, one suspects that ordinary descriptions often might be incomplete, logically inconsistent, and untestable.
Second, event-structure models are quantifiable in that directed graphs representing logical relations among events can be converted to a matrix format subject to mathematical analysis. One possibility, for example, is measuring the degree to which different people have the same knowledge23--a topic that Carley (1986) has explored in interesting detail. In work along these lines by myself and students24, each respondent's structure is represented within a catalog matrix that includes rows and columns for cognitions used by any respondent. An index of similarity is computed by counting the number of cells that are identical across respondents, thereby measuring similarity of content and also similarity of logic within areas of shared content. Factor analyses of the similarity measures can identify clusters of people who think alike. (A variety of such procedures can be developed, and they need to examined methodically to determine their characteristics and relative benefits.)
Any area of social science that deals with knowledge--e.g., ethnology, sociology of science, cognitive science--gains. from a method that objectifies knowledge and allows it to be recorded (Colby, 1975), overcoming the ephemeralness and transiency of informal knowledge. Archives of procedural knowledge based on event-structure models may be useful for trainees who are trying to master an esoteric body of knowledge25, and the models have the benefit of permitting simulations, of allowing analysts to ask "What if".
From a different viewpoint, event-structure models record systems of social constraint that are controlling action and generating social institutions. They are records of the cybernetic systems of society (Fararo and Skvoretz, 1986), and archives of such models can provide a corpus of structures for comparison and analyses.
Popular culture is recorded largely in narratives, ranging from autobiographic statements to fiction and, within fiction, from folktales to novels. Such materials seem to provide useful insights about worldviews of other times and other groups, but our standard method of extracting the insights--readings and expositions by experts--is undisciplined, sensitive though it may be.
Event structure analysis offers new opportunities for the study of popular culture through written texts: narratives can be used to construct models that represent popular realities and popular reasoning. Such models of culture support an unusual claim to validity in that they can be shown to account for narratives, event by event. Moreover, event structure models are explicit and objective enough to focus scholarly debates on specific points of interpretation in order to limit the possibilities for misunderstandings. Furthermore, the models can be used by students--even students who do not know the culture very well--in order to explore the foreign reality and learn it. Of course, incorporating such a methodology into the study of popular culture can succeed only if experts who know a popular culture fashion the models to represent texts adequately.
Hoping to encourage such work, I presented elsewhere (Heise, 1988b) a model of a folktale common in French peasant society several hundred years ago (Darnton, 1984: 9-10). My model of the narrative is deficient because I am no expert on that culture, but even so the analysis suggests interesting hypotheses--that the tale dramatically communicates values by inverting values, displaying the good as abnormal and the bad as normal; and (based on analyses conducted after the 1988 publication) that the story of Little Red Riding Hood retains its emotional impact today because it can be interpreted at an abstract level that teems with contemporary concerns.
The procedures of event-structure analysis can be used to clarify and formulate one's own ideas as well as others'. Theoretical models result if one happens to be a social scientist operating in an area of specialization. Indeed; causal theories of qualitative events seem to be essentially the same thing as production-system or event-structure models.
McCullagh (1984) has provided a thoughtful analysis of what is involved in theorizing by historians. According to McCullagh, historians' ideas about the general structure of an historical process "are no more than descriptions of patterns of change which frequently have occurred in slightly different ways on several occasions" (p. 13). McCullagh's interpretation of cause corresponds to prerequisite in event-structure modeling26.
"Historians judge one event or state of affairs to have been a cause of another, I suggest, only if they believe the occurrence of the first to have been necessary in the circumstances, contingently necessary, for the occurrence of the second." (p. 176). "[O]ne event is contingently necessary for another if the occurrence of an event having at least one of the possible descriptions of the first is necessary in the circumstances for the occurrence of an event having at least one of the possible descriptions of the second" (p. 180).
In general,
"one event or state of affairs is causally related to another if and only if (1) the former is not identical to the latter, or part of it, and does not merely logically imply it; (2) the former is at least contingently necessary. ..for the latter; and (3) the former does not succeed the latter" (p. 185).
A similar orientation to causation arises in cognitive science; for example, Patel and Groen (1986, p. 95) define a causal network as "a network of antecedents linked to consequents. Equivalently, it can be viewed as a network of if-then rules or productions."
Thus the kind of explanations that qualitative researchers provide should be amenable to modeling by event-structure analysis, and the procedures discussed in this paper offer several benefits to qualitative researchers as they formulate their theories. First, the construction of an explicit model is facilitated by computer-assisted elicitation procedures. Second, event-structure analysis objectifies the logical relations in a theory and demands that they be complete and coherent for one's purposes. Third, records of observed events from field notes or historical archives can be used to test, evaluate, and revise formulations on the basis of empirical data. Fourth, the general process of creating a formal model provides a medium for thinking systematically and methodically about one's work.
Since the technology is new, little is known about the practical aspects of obtaining event-structure models. Some informants might be able to sit at a computer and develop a model on their own: undergraduates without prior computer experience have done this27. Alternatively, a researcher might take a portable computer into the field and use it for computer-assisted interviewing28, separating respondents from the technical activity (which is carried on by the interviewer instead). However, this requires respondents to concede the substantial amount of time required for constructing and testing event-structure models. The third approach is for investigators to reach verstehen-level understanding of their subjects and then construct models themselves by analyzing event sequences from their field notes29. Methodological research on the viability and reliability of these approaches--and related approaches (Axten and Fararo, 1977; Garfinkel, 1964)--would be useful.
Some speech acts have a ritualistic quality like those in the examples of professorial encounters--one cannot predict what will be said, but the timing, the speaker, and the general topics seem accountable in an event-structure model. Additionally, certain processes involved in talking like turn-taking (Duncan, 1972) and interpretive processes (Labov and Fanshel, 1977) are so rule governed that they could be represented in formal models.
Beyond this, discourse raises some perplexing problems when modeling action. On one hand, discourse is largely rational, conveying information and reasonings, so it seems that it must relate to rational action. On the other hand, many speech events--like asking, telling, suggesting, commanding--do not seem essential in general for other events. When they are included in event-structure analyses, particular vocalizations of this kind may seem necessary in the given incident, but they turn out to be unnecessary in general as similar incidents are analyzed.
Clarke's (1983, pp. 237-238) speculations about how discourse operates in interaction suggest some hypotheses that might be explored with event-structure models. In particular, instrumental vocalizations may occur when a high-priority act is primed but not enacted, which may be especially likely after one subgoal has been reached, and participants are trying to start progress toward other subgoals. In such a case, participants may confirm the state of priming events through questions and answers, then a primed act may be suggested or commanded. Justifications and excuses may arise when participants disagree over priorities, and negotiations may assemble a new structural model in which people's priorities are better aligned30.
Instrumental discourse in which people assemble shared conceptual models, keep each other updated on the current state of the environment, and air conflicts are intimately related to the structure of action, but the speech acts are auxiliary events that largely can recede wherever people share perspectives, perceptions, and motivations, and such speech acts generally should not be treated as prerequisites for other events31.
Both affective and rational factors are important determinants of human action. The limitations of reason can be appreciated by noting some contributions of affect.
Much human action is thoughtless and unreasoned, yet retains its sociocultural structuring all the same. Affect maintains social order by providing "acquired instincts" that allow people to sense what they should do without necessarily knowing why. Affect is especially important in the formation of creative responses which allow people to deal with a novel circumstances in a culturally-appropriate manner even though the circumstance is beyond acquired knowledge. Affect also controls expressive behavior and other free variations in action that have no instrumental value but which are important phenomenologically and socially. Additionally, affect underlies responses to deviance--sanctioning and labelling--and the moralities that justify social control. Rationalist frameworks may have some success in dealing with these matters (just as an affective model is peculiarly effective in specifying instrumental actions for various roles--Heise, 1979), but the topics are not within the natural domain of rational models.
I myself have not abandoned two decades of work on affect (Heise, 1979; 1985; Smith-Lovin and Heise, 1988) in order to do the work described in this paper. Rather I am trying to lay a foundation for empirical work on event structures that will yield the kind of synthesis between cognitive and affective control systems that Fararo and Skvoretz (1984) suggested. Fararo and Skvoretz speculated that affect-control might account for motivation in rational action. Besides that, I hope detailed studies of event structures will reveal how sentiments are shaped within systems of production and how affective factors adjust rational processes--e.g., by changing priorities or by invoking responses to deviance.
Event structure models are inherently limited by their focus on rational processes. Ultimately, a mature science of social action will have to offer models that mix reason and feeling in the human composition.
NOTES
* Randall Collins, William Corsaro, Donna Eder. Diane Felmlee, Thomas Fararo, and Douglas Heckathorn provided me with helpful responses to this paper. I hope, but do not assure, that I addressed their concerns adequately.
1. Ragin (1988; Ragin, Mayer, and Drass, 1984) is developing procedures to define models like those presented here through data reduction, inference, and generalization. Ragin is concerned with generating knowledge; this paper is concerned with recovering knowledge that people have.
2. My emphasis is on knowledge about events as compared to Carley's emphasis on knowledge about people.
3. Rational thinking about events emphasizes constraint and consequences: the world limits actions, actions shape the world, and prerequisite actions have to happen in order to make wanted actions possible. Rationalism fits material reality so nicely that the capacity for it might have evolved to register physical existence. Everyday we reason mundanely but effectively in order to make our way through life--for example, having entered a room, we know we have to leave before we can go elsewhere. In science we reason similarly about more exotic aspects of material reality and universally find constraint, consequence, and logical coherence.
4. Logic is a human universal. The most influential social science argument to the contrary--D.D. Lee's account of non-logical thinking among Trobrianders--was critiqued devastatingly by Hutchins (1981), who began by pointing out that Lee never visited the Trobriands.
While every human has the capacity for logic (like the capacity for language), human thinking and behavior are not always guided by the laws of logic--see the end of this article.
5. I have discussed the quality of knowledge elsewhere--Heise (1988a).
6. It is fortuitous that some of the control systems governing human action (i.e., people's knowledge) are developed via the same logic that we use to study them, but to the extent that it is true we can make use of the fact.
7. A competing orientation toward the representation of social events is provided by the script formulation of Schank and Abelson (1977), which also has led to a concern with event knowledge (Nelson, .1986). I prefer the production system approach because production systems can account for sequences of action that may not have been observed before, whereas scripts deal with known sequences of action and only the sequences which have been defined explicitly are rigorously accountable. Production systems have a generative power that scripts lack.
8. Verbal descriptions of events raise some issues that will have to be confronted in the future.
Axten and Skvoretz (1980: 550) note that events have to be defined "in a suitably general form", meaning that some distinguishable events are to be treated as the same. McCullagh (1984) observes that "only one set of all the possible sets of truth conditions of a description must correspond to part of reality for that description to be true" (pp. 74-75); for example: "There are various ways in which one country can insult another. ..the satisfaction of anyone set is sufficient for the correct use of that description, but the description is not warranted unless at least one set is satisfied" (79). Fararo and Skvoretz (1984: 163) offer some leads for addressing the level-of-generality issue in their discussion of functional equivalence.
Some anthropologists have argued that many actions hardly can be defined verbally (Dougherty and Keller, 1985; Gatewood, 1985) because they involve performance knowledge, and people are not articulate about performance knowledge. This could mean that action systems are beyond description unless they have been analyzed cognitively through cultural process or through the work of an interested and articulate expert.
9. Fararo and Skvoretz (1984) refer to this as the "functional structure".
10. Exogenous conditions add a probabilistic element to what is supposed to be a deterministic system: an event may not occur even though it is primed and highest in priority because the exogenous conditions are not fulfilled. Non-independence between emissions and exogenous conditions may make relations between system events probabilistic in that an event is sometimes primed by a prerequisite and sometimes not, depending on events beyond the system boundaries.
Some researchers have defined the structure underlying a sequence of qualitative events with Markov models. First-order Markov models have been found to be effective in describing animal behavior (Van Hooff, 1982) and second-order Markov models sometimes do a good job in describing human interactions (e.g., Brent and Sykes, 1979).
11. Petri nets (Peterson, 1981) provide an alternative formalism for representing processes involved in production systems. Markings of Petri nets represent instantiations of events, enablement corresponds to what I call priming, and next-state functions define what I refer to as depletions.
12. Places (events) in a Petri net (Peterson, 1981) may be marked with multiple tokens, and, if enabled, a place may fire repeatedly until all of its tokens are gone (see Peterson, 1981). A similar stacking of instantiations might be incorporated usefully into some production system models.
13. The computational problems involved in cyclical graphs are not totally intractable, and this assumption may be relaxed in the future. Petri nets (Peterson, 1981) routinely allow for cycles.
14. With the assumption of acyclical systems we do not need to ask if B implies A once we discover that A implies B. However, if we ask "Does A imply B" and the answer is "No" then we have to ask the reverse question as well. Thus the total number of questions will be somewhere between N(N-1)/2 and N(N- 1), where N is the number of events.
15. The focus on incidents is an analogue of Newell and Simon's (1972) use of protocols.
16. With time-ordering the upper limit on relational questions in an acyclical system is N(N-1)/2 instead of N(N-I).
17. Actually an expert description may carry this too far, glossing over events that are supposed to be inferable by a competent audience. Richer detail can be achieved by considering problematic incidents that preclude glossing.
18. However, expert ratings of event priorities might be superior in some cases. The expert presumably draws from a deep pool of experiences, and priority assessments based on many experiences might be better than priorities computed from events in a single incident, even allowing that an expert's judgment is subject to various biases because of personal attitudes and limitations in recall.
19. A complete description of program ETHNO (Heise, 1988c) is not possible here because ETHNO is so large--compiled from 6,000 lines of Pascal code. Implicational structures are graphed on the computer screen as the analysis proceeds; a data-management system for working on graphs is included in the program; and a windowing system presents help and error messages.
ETHNO runs on microcomputers with MS-DOS operating systems and either monochrome or color adapters; graphs can be printed on most printers. Implicational structures may have up to 100 nodes and 500 branches, and event series may be up to 500 units long.
20. This event is not really repeated but restarted after an interruption. In the future it might be desirable to include an interruption function that leaves an interrupted event primed and undone: "any conditions established or removed by the appearance of an action ... are established or removed only upon completion of the action" (Skvoretz, 1984: 91). The repeatable function serves in lieu of this.
21. A more sophisticated approach would treat the entries of the two professors as instances of the same event--"Professor [x] enters room", following the convention of Axten and Fararo(1977). A number of other pairs of events in the example similarly can be viewed as the same event with variable actors, and I will discuss them this way, even though the current program treats each event separately.
22. Simulation requires repeated reference to the logical structure and keeping a log of events that are primed and depleted, but those matters are handled automatically by the computer program.
23. Fararo (1980) discusses this as the problem of embodiment.
24. Kandi Stinson worked on this at the University of North Carolina; she is now at Lehigh University. Sharon Georgianna, an Indiana University graduate student, has compared the structure of "life forces" as viewed by Creationists and research biologists.
25. Impressing experts is not one of the capabilities of formal models of knowledge. Experts may enjoy constructing a model, but they find the final product boring. Mittal and Dym (1985, p. 33) observed: "while the experts acknowledged the diagnostic acumen of the MDX system, they did not find that particularly impressive because they were also doing the same task on a routine basis. Instead, they were more interested in a computer system which could help in performing tasks with which they had difficulty."
26. In contrast to the view of quantitative researchers in sociology (e.g., Heise, 1975), sufficiency is not involved in historians' interpretations of cause. "One can know that generally certain sorts of conditions, or disjunctions of conditions, are necessary for the occurrence of a certain kind of event, without knowing any set of conditions to be strictly sufficient for it, necessitating the occurrence of that kind of event. This is typically all that historians do know about the causes of the kinds of events they study. ...[S]ome causes do not even combine to make an event probable. .." (McCullagh, 1984, p. 178).
Causal thcory means different things to qualitative and quantitative researchers, and this can be a source of exasperation in their relations with one another.
27. Of course, these college students were technologically savvy by virtue of being middle-class, they received substantial instruction in using the event-structure program, and they were motivated to spend hunks of time because the assignment was a course requirement. Model building by indigenous informants in other kinds of populations might be infeasible.
28. Georgianna did computer-assisted interviewing with Creationists in the Indiana study mentioned in a note above. Biologists in the same study used a portable computer, which was taken to their offices, and they managed the program without difficulty after a little instruction. However, the task did not include a series analysis.
29. William Corsaro has experimented with this approach in collaboration with me, focusing on peer-group routines of three- and four-year-olds as reported in Corsaro (1985). The work reveals that models can be constructed to account for field observations, and the models can contribute to research goals--in this case showing that preschoolers' actions are consistent with the notion of a subculture. Developing models to account for field observations requires substantial time investment in order to analyze different incidents. Some technology has to be mastered--using a personal computer and using the event-structure program.
30. Sowa (1984) argued persuasively that people do not carry about complete scripts for every scene they encounter, but rather they assemble structures as needed from conceptual molecules. What Sowa's theory misses (and is missing generally in cognitive science formulations) is the sociological element: people help each other assemble conceptual structures, and they do this through conversation -see Carley (1986).
31. This is not to say that instrumental discourse should be ignored. Recorded dialogue can contribute to event-structure analyses in the same way that conversation helps participants in an incident--by clarifying what people are doing and what they think they are supposed to do. Moreover, instrumental discourse offers cues about the social structural relations of interactants to one another. People's writings--another kind of discourse--might be content-analyzed to identify conceptual models they use and want to promote, as has been done, for example, by Axelrod (1976).
REFERENCES
Agar, Michael, 1974. Talking about doing: Lexicon and event. Language in Society, 3: 83-89.
Anderson, John R., 1983. The Architecture of Cognition. Cambridge, MA: Harvard University Press.
Axelrod, Robert (Ed.), 1976. Structure of Decision: The Cognitive Maps of Political Elites. Princeton, NJ: Princeton University Press. .
Axten, Nick, and Fararo, Thomas J., 1977. The information processing representation of institutionalized social action. Sociological Review Monograph, No. 24.
Axten, Nick, and Skvoretz, John, 1980. Roles and role-programs. Quality and Quantity, 14: 547-583.
Brent, Edward, E., Jr., and Sykes, Richard E., 1979. A mathematical model of symbolic interaction between police and suspects. Behavioral Science, 24: 388-402.
Carley, Kathleen, 1986. An approach for relating social structure to cognitive structure. Journal of Mathematical Sociology 12: 137-189.
Clarke, David D., 1983. Language and Action: A Structural Model of Behaviour. Oxford: Pergamon Press.
Colby, Benjamin N., 1975. Culture Grammars. Science, 187: 913-919.
Conrad, Peter, and Reinharz, Shulamit, (Eds.), 1984. Computers and Qualitative Data. New York: Human Sciences Press. (Special issue of Qualitative Sociology, 7, Nos. 1 and 2.)
Corsaro, William A., 1985. Friendship and Peer Culture in the Early Years. Norwood, NJ: Ablex Publishing Corp.
Darnton, Robert, 1984. The Great Cat Massacre and Other Episodes in French Cultural History. New York: Basic Books.
Dougherty, Janet W.D., Ed., 1985. Directions in Cognitive Anthropology. Urbana: University of Illinois Press.
Dougherty, Janet W.D., and Keller, Charles M., 1985. Taskonomy: A practical approach to knowledge structures. Pp. 161-174 in J. Dougherty, (Ed.), Directions in Cognitive Anthropology. Urbana: University of Illinois Press.
Duncan, S.D., Jr., 1972. Some signals and rules for taking speaking turns in conversations. Journal ofPersonality and Social Psychology, 23: 283-292.
Fararo, Thomas J., 1981. Social activity and social structure: A contribution to the theory of social systems. Cybernetics and Systems: An International Journal, 12: 53-81.
Fararo, Thomas J., and Skvoretz, John, 1984. Institutions as production systems. Journal of MathematicalSociology, 10: 117-182.
Fararo, Thomas J., andSkvoretz, John, 1986. Action and institution, network and function: The cybernetic concept of social structure. Sociological Forum, 1: 219-250.
Garfinkel, Harold, 1964. Studies of the routine grounds of everyday activities. Social
Problems, 11: 225-250.
Gatewood, John B., 1985. Actions speak louder than words. Pp. 199-219 in J. Dougherty (Ed.), Directions in Cognitive Anthropology. Urbana: University of Illinois Press.
Grimshaw, Allen D., 1982. Sound-image data records for research on social interaction: Some questions and answers. Sociological Methods & Research, 11: 121-144. (In a special issue devoted to sound- image records, edited by A.D. Grimshaw.)
Hanson. Norwood Russell, 1958. Patterns of Discovery: An Inquiry into the Conceptual Foundations of Science. New York: Cambridge University Press.
Heise, David R., 1975. Causal Analysis. New York: John Wiley & Sons.
Heise, David R., 1979. Understanding Events: Affect and the Construction of Social Action. New York: Cambridge University Press.
Heise, David R., 1985. Affect control theory: Respecification, estimation, and tests of the formal model. Journal of Mathematical Sociology, 11: 191-222.
Heise, D.R., 1986. Modeling symbolic interaction. In S. Lindenberg, J. Coleman, and S. Nowak, (Eds.), Approaches to Social Theory. New York: Russell Sage Foundation.
Heise, David R., 1988a. Delusions and the construction of reality. In T.F. Oltmanns and B.A. Maher, (Eds.), Delusional Beliefs. New York: Wiley, 259-272.
Heise, David R., 1988b. Computer analysis of cultural structures. Social Science Computer Review, 6: 183-196. ,
Heise, David R., 1988c. ETHNO (program and documentation). Raleigh, NC: National Collegiate Software Clearinghouse, NCSU, Box 8101, Raleigh, NC 27695.
Hutchins, Edwin, 1981. Reasoning in Trobriand discourse. Ronald W. Casson (Ed.), Language, Culture, and Cognition: Anthropological Perspectives. New York: Macmillan Publishing Co. 481-489.
Katz, Jerold J., 1972. Semantic Theory. New York: Harper & Row.
Kempton, Willett, 1986. Two theories of home heat control. Cognitive Science. 10: 75-90.
Labov, William, and Fanshel, David, 1977. Therapeutic Discourse: Psychotherapy as Conversation. New York: Academic Press.
Lord, FredericM., and Novick, Melvin R.,1968. Statistical Theories of Mental Test Scores. Reading, MA: Addison-Wesley.
McCall, George J., and Simmons, J .L. (Eds.), 1969. Issues in Participant Observation: A Text and Reader. Reading, MA: Addison-Wesley.
McCullagh, C. Behan, 1984. Justifying Historical Descriptions. New York: Cambridge University Press.
Mittal, Sanjay, and Dym, Clive L., 1985. Knowledge acquisition from multiple experts. AI Magazine. 6: 32-36.
Nelson, Katherine, (Ed.), 1986. Event Knowledge: Structure and Function in Development. Hillsdale, NJ:
Lawrence Erlbaum. Newell, A., and Simon, H., 1972. Human Problem Solving. Englewood Cliffs, NJ: Prentice-Hall.
Patel, Vimla L., and Groen, Guy J., 1986. Knowledge based solution strategies in medical reasoning. Cognitive Science. 10: 91-116.
Peterson, James L., 1981. Petri Net Theory and the Modeling of Systems. Englewood Cliffs, NJ: Prentice-Hall.
Pugh, George Edgin, 1977. The Biological Origin of Human Values. New York: Basic Books, Inc., Publishers.
Ragin, Charles C., 1988. Between Complexity and Generality: The Logic of Qualitative Comparison. Berkeley: University of California Press.
Ragin, Charles C., Mayer, Susan E., and Drass, Kriss A., 1984. Assessing discrimination: A Boolean approach. American Sociological Review. 49: 221-234.
Rips, L.J., 1983. Cognitive processes in propositional reasoning. Psychological Review. 90: 38-71.
Robinson, W .S.~ 1951. The logical structure of analytic induction. American Sociological Review. 16: 812-818.
Schank, R.C., and Abelson, R.P., 1977. Scripts, Plans, Goals, and Understanding. Hillsdale, NJ: Lawrence Erlbaum.
Skvoretz, John, 1984. Languages and grammars of action and interaction: Some further results. Behavioral Science. 29: 81-97.
Skvoretz, John, and Fararo, Thomas J., 1980. Languages and grammars of action and interaction: A contribution to the formal theory of action. Behavioral Science, 25: 9-22.
Skvoretz, John, Fararo, Thomas J., and Axten, Nick, 1980. Role-programme models and the analysis of institutional structure. Sociology, 14: 49-67.
Smith-Lovin, Lynn, and Heise, David R., (Eds.), 1988. Analyzing Social Interaction: Advances in Affect Control Theory. New York: Gordon & Breach. (Reprint of a special issue of the Journal of Mathematical Sociology, 13 (1987), #1-2, edited by Smith-Lovin and Heise.}
Sowa, J.E., 1984. Conceptual Structures: Information Processing in Mind and Machine. Reading, MA:Addison-Wesley Publishing Co.
Turner, Ralph, H., 1953. The quest for universals in sociological research. American Sociological Review, 18: 604-611.
Van Hooff, J.A.R.A.M., 1982. Categories and sequences of behavior: Methods or description and analysis. Pp. 362-439 in K.R. Scherer and P. Ekman (Eds.), Handbook of Methods in Nonverbal Behavior Research. New York: Cambridge University Press.
Waterman, D.A., and Hayes-Roth, Frederick, 1978. Pattern-Directed Inference Systems. New York: Academic Press.
Werner, 0., and Schoepfle, G.M., 1987. Systematic Fieldwork: Foundations of Ethnography and Interviewing. (2 vols.) Beverly Hills, CA: Sage.