Part II: How Can We Represent Processes? Toward A Theory Of Process Representation
7.4 Methodological Considerations of Grammatical Models
To apply the kind of grammatical model outlined here in empirical research on
organizational processes, there are a number of methodological considerations that need to be addressed. It is worth noting that an emphasis on processes as a unit of analysis implies a significant departure from conventional methodologies. We are much more accustomed to using individuals, organizations, or networks as the unit of analysis, treating them as
hypostatized objects, and formulating theories based on their variable properties (Mohr 1982;
Abbott 1992). Nonetheless, the growing interest in process analysis has given rise to a variety of methodologies for analyzing sequences of events (Hewes 1979; Procter and Abell 1985;
Bakeman and Gottman 1986; Gottman and Roy 1990; Abbott 1990; Abbott and Hrycak 1990; Van de Ven and Poole 1990). These methodologies are too numerous to review in detail, but they are generally designed to discover meaningful regularities in sequences of events that can be observed, coded, and compared. Grammatical models add to this growing family of tools for analyzing process data by providing a way to link what Poole, Folger, and Hewes (1987) call the ''syntagmatic''structure of a process to its global sequential structure. To the extent that grammatical models rely on sequential data, they share many of the same methodological considerations as sequential techniques in general, such as reliability of coding. Grammatical models, however, provide a rather different
approach to understanding the connections between sequences of events and the structural features that enable and constrain them. As a result there are several issues that deserve attention here.
7.4.1 Identifying a Lexicon and Syntactic Constituents
To perform a grammatical analysis of a class of organizational processes, the first step would be to identify the lexicon of moves and the appropriate syntactic constituents. The questions here are, What is the vocabulary of action in this process? What are the steps in the process? What are the different ways in which the steps can be accomplished? These questions bear a striking resemblance the ''structural questions''described by Spradley (1979, pp. 116-17) in his primer on ethnographic interviewing. The objective of Spradley's (1979) technique is to map out the semantic domain used by the members of a particular cultural group to describe some aspect of their work or lives. In process research, the two most relevant semantic domains would be sequence (''X is a step (stage) in Y '') and means- ends (''X is a way to Y '') (Spradley, p. 111). In these relationships, Y is called a ''cover term''and X is called an ''included term.''For each cover term, there is generally more than one included term, and there may be many. In terms of the grammatical metaphor outlined above, the included terms will tend to correspond to moves and the cover terms are likely to correspond to syntactic constituents. This nesting of lexical items is a distinctive feature of organizational processes that the grammatical metaphor encourages us to explore explicitly.
Other process models treat data as flat, with each element having roughly equivalent status (e.g., Holmes and Poole 1991).
In practice, one may need to abstract somewhat from an informant's talk to arrive at a set of syntactic constituents that can be generalized across settings. The necessity of creating more abstract categories of action to facilitate analysis raises a familiar question: Do we impose our own, etic terminology for the actions we observe, or do we use the emic
terminology of our informants (Spradley 1979)? This is essentially the same problem that we confront when we collect survey data that are presumed to mean the same thing to different respondents, so that their responses can be subjected to mathematical transformation and analysis (Cicourel 1964). The important issue here is not so much the use of member's own terminology but the semantic relationship between covering terms and included terms. The included terms must be ''steps in''or ''ways to''complete the action described in the covering term (e.g., paying is a step in checking out).
7.4.2 Identifying and Formulating Constraints
To formulate meaningful explanations, the critical problem is to identify the relevant sources of constraint on the lexicon and the ways in which its elements can be recombined. The most convenient way of formulating a constraint is as a rule, as in this kind of sequential
coordination constraint: ''A product must be designed before it can be manufactured.''It is interesting to note that there may be a large number of different steps in the process that intervene between design and manufacturing (dealing with strategy, marketing, finance, etc.).
A simple sequential constraint does not require adjacency (although one could formulate a more stringent constraint that would). This simple example points to one of the major advantages of this approach: syntactic models provide explicit ways of stating hypotheses about constraints on events in a process that are widely separated in the observed sequence. This is a special property of syntactic models that is not shared by statistical techniques such as Markov models (Chomsky 1956). In a supermarket, for example, you may have any number of iterations of 'select item'before you 'check out'. For this reason a Markov model would do poor job of capturing the sequential dependence between arriving at the store and checking out, though these events are structurally constrained to occur in this order in every complete transaction.
7.4.3 Comparison to Phasic Analysis and Other Sequential Techniques
Although a detailed review of sequential methods is beyond the scope of this chapter, there are some points of reference in the literature that might be helpful for some readers. For example, Holmes and Poole (1991) describe a method called phasic analysis that has some interesting similarities but also some important differences. The basic idea of phasic analysis is to code sequential data in terms of a set of events that mark a particular phase of activity.
For example, in a stage model of organizational development (e.g., Greiner 1972), there might be an early stage of creativity and leadership marked by certain kinds of behavior, followed by a stage of direction and autonomy, delegation and control, and so on. Using phase analysis, one can test the observed sequence of stages against a predicted model or analyze typical sequences of stages (Pelz 1985; Holmes and Poole 1991). Poole and Roth (1989), for example, used phasic analysis to develop a typology of group decision-making processes.
There are a number of terminological similarities between phasic analysis and the
grammatical models suggested here. Holmes and Poole (1991, p. 295) write about testing phase models by ''parsing of a sequence of phase markers into discrete
phases.''Grammatical models are also tested by parsing sequences of events into syntactic constituents, but these constituents can have a much more elaborate internal structure.
Furthermore a grammatical constituent is typically marked by a single event rather than a sequence of similar events. Holmes and Poole (1991, p. 293) also discuss the use of coding systems that include 'major categories'and 'subcategories,'which in some ways are like Spradley's (1979) covering terms and included terms. However, the semantic characteristics of the respective coding scheme are very different. In the coding schemes described by Holmes and Poole (1991), the sub-categories are indicators of the major categories, not steps used to accomplish it, and the categories are basically etic (which include things like 'denial and equivocation', 'noncommittal remarks', 'topic management'). In the grammatical approach, the included terms should be steps needed to accomplish the covering term (e.g., paying is a necessary part of checking out). In further contrast to the phasic approach, it is helpful if the coding scheme at the lowest level (moves) is basically emic. The use of emic categories is not essential, but it facilitates the collection of data by asking people to describe what they are doing. Furthermore, grounding the lexicon in ethnographically derived
categories helps keep the analysis more closely connected to the phenomenon.
7.4.4 Logic of Analysis
Testing grammatical models raises some interesting problems concerning what, exactly, should count as evidence of disconfirmation. Linguists can often disconfirm a grammatical rule by pointing to a single sentence construction that intuitively seems grammatical but violates a hypothesized constraint. As mentioned above, organization theorists are not so lucky, because we do not have introspective access to a hypothesized universal structure that guides our theorizing. Rather, we must collect data in the field, a procedure that is fraught with all kinds of possibilities for error in coding, sampling, and so on. While we might want to follow a strict rule of single case disconfirmation, it would tend to lead to spurious rejections of the hypothesized model.
Alternatively, we might also follow a statistical approach, similar to that used for testing traditional variance models. Unfortunately, no rigorous statistical tests have yet been developed to test the goodness of fit of grammatical models or other rule-based models (Olson, Herbsleb, and Rueter 1994; Simon 1992). Simon (1992) notes that in cognitive science, the general heuristic for the adequacy of rule-based models of human performance is that the model must explain many more cases than the number of rules it uses. In this context Simon's (1992) usage of the term ''explain''simply means that the behavior described by the model matches the observed behavior, either in functional form or in exact detail. In the case of our simple grammar of suburban shopping trips, this criterion is easily met; at the level of detail expressed in the grammar, I would expect it to fit nearly every supermarket in the United States. This seems like a reasonable basis for proceeding until the statistical properties of these models can be worked out.
7.4.5 Limits of Applicability
There are limits on the kind of processes and level of detail for which a grammatical approach may be appropriate. Levinson (1983) notes that conversation analysts have had limited success in formulating interaction grammars of the kind Hymes (1972) proposed. In essence, there are just too many possibilities and contingencies available in interaction, and too strong a tendency toward the strategic use of cultural constraints (e.g., Grice's 1975 maxims for conversational cooperation) to create irony and implicature. There is no point, as Abell (1987) points out, in attempting to unpack every little motion or inflection as a separate piece of data. The kinds of processes that seem more natural and appropriate for
grammatical analysis are more deeply embedded in organizational structures, less fine grained, and hence less subject to capricious variation. This suggests that it is important to limit the level of granularity with which one describes a process to those moves that can be easily observed and reliably coded (Folger, Hewes, and Poole 1984; Van de Ven and Poole 1990). The diffculty involved here should not be underestimated. We have a natural ability to parse sentences into recognizable words and constituents, but our ability to parse
organizational processes depends on artificial methods. The development of reliable coding schemes for moves and syntactic constituents that can be generalized across organizational settings will be an important research question in and of itself.
Furthermore the culturally and historically embedded quality of organizational processes implies that it is important to bound the scope of the data one is drawing upon in constructing a grammar so that it is relatively homogeneous. In their discussion of grammatical
techniques in cognitive anthropology, Colby, Kennedy, and Milanesi (1991, p. 383) note that grammars of folktales can only be developed for a culturally bounded group of people.
To reiterate the condition for analyzing plot grammars, it is necessary that the sample of texts be geographically bound to a particular language using group of people and that it consist of the same genre and same general time period. With these restrictions, and if the sample is sufficiently large (numbering at least over fifty and preferably twice as much) it should be
possible to eventually work out a plot grammar.
The same basic recommendations seem quite appropriate for the analysis of organizational processes. The critical issue here is one of sampling, and the scope of data that one can hope to meaningfully incorporate into a single grammar. One cannot expect processes operating within different institutional, cultural, and technological structures to fit the same grammar, unless that grammar is very abstract. While these limitations need to be taken seriously, they are not especially different in kind or severity than the limitations on traditional variance models. What is different, of course, are the kinds of phenomena that grammatical models can express, and the kinds of research questions they allow us to pursue.