You are on page 1of 17

A Fuzzy Future? Fuzzy Logic Potential for Institutional Analysis1 Sue E. S.

Crawford DRAFT November 2005

Abstract: In Understanding Institutional Diversity, Elinor Ostrom challenges us to recognize the configurational and complex nature of interactions between institutions and their contexts in institutional analysis (Ostrom 2005b). Fuzzy logic appears intuitively to be a useful tool for complex and configurational analysis, particularly when that analysis is geared toward diagnostics. This paper explores the possible use of fuzzy logic methods for institutional analysis based on the Institutional Analysis and Development Framework (IAD) and the grammar of institutions. It discusses potential use of fuzzy logic tools for static analysis and for analysis of institutional change. The static analysis section focuses primarily on tasks associated with coding cases and analysis of coded cases, with particular emphasis on uses for common pool resource case analysis. The section on analysis of change proposes ways in which fuzzy logic may help to unpack the feedback and evaluative components of the IAD framework. This section focuses in particular on fuzzy logic tools to model the learning advantages of farmer managed irrigation systems as compared to government managed systems.

Prepared for Elinor Ostrom Festschrift, Bloomington, IN November 22-24, 2005

2005 by author
Grateful thanks to John Mordeson from the Creighton Math Department for his assistance with math resources. Thanks also to Terry Clark for helpful discussions about fuzzy math issues in political science and for initiating collaborative discussion on the topic at Creighton. Neither reviewed this paper. All errors in interpreting fuzzy math principles are my own.
1

In Understanding Institutional Diversity, Elinor Ostrom (2005) challenges institutional analysts to recognize the configurational and complex nature of interactions between institutions and between institutions and their social and environmental contexts. This call reflects a theme in many of her writings concerning danger of policy analysis that places too much emphasis on a single (or small number) of institutional components without attention to the assumptions of the analysis concerning other relevant institutional components. For example, she and colleagues have spent considerable time illustrating the importance of default conditions for rules in game models and default components of rules (Crawford and Ostrom 1995; Ostrom 2005a; Ostrom 2005b; Ostrom, Gardner, and Walker 1997). Similarly, Understanding Institutional Diversity (and many other published pieces) stresses the danger of simplistic policy advice that ignores possible innovative solutions assuming instead, for example, that the only choices are market solution or public solution. Another concern is analysis that assumes that a working institution in one situation can be implemented successfully in other locations (blueprint thinking) (Ostrom 2005b). This essay explores possible uses of fuzzy logic methods for complex institutional analysis based on the Institutional Analysis and Development Framework (IAD) and the grammar of institutions. Can fuzzy logic tools facilitate more complex and configurational institutional analysis? My exploration of fuzzy logic tools discusses their potential use for static analysis and analysis of institutional change. The static analysis section focuses primarily on tasks associated with coding cases and analyzing coded cases for comparative analysis, with particular emphasis on uses for common pool resource case analysis. It discusses the basic logic of Ragins fuzzy set methods for social science (Ragin 2000; Ragin and Rihoux 2004) with attention to some extensions and alternatives possible given a broader set of fuzzy logic tools. The section on change proposes ways in which fuzzy logic may help us to unpack feedback and learning components of the IAD framework. This section also considers the usefulness of fuzzy logic tools in analyzing adaptive advantages of farmer managed irrigation systems as compared to government managed systems (see Ostrom 2005). An important caveat is in order first. This analysis should be seen as only an initial step in a careful discussion of the potential usefulness of fuzzy logic methods. The arguments reflect initial hunches that will need to be further tested with empirical and modeling applications and more careful attention to mathematical proofs. A key point of this exercise is to help institutional analysts consider where the effort of pursuing these methods might be worthwhile. I focus on discrete fuzzy applications only since most of our current methods for analyzing institutions and action situation components tend to be discrete category oriented measures. Political scientists with analyses more geared to continuous methods have an even broader range of fuzzy math tools at their disposal. Fuzzy logic appears intuitively to be a useful tool for complex and configurational analysis, particularly when that analysis is geared toward diagnostics. Fuzzy logic is a form of soft computing currently used to improve diagnostic abilities and to improve the performance of tools (such as cameras) that must react to complex environmental conditions. It is also used to improve control and diagnostics in other situations in which

treatments or adjustments must be made based on information related to multiple criteria. Moreover, fuzzy logic allows predictive models that combine fuzzy elements that might reflect institutional components with well-established mathematical models of physical systems (Bardossy and Duckstein 1995). Thus, it seems, at least on the surface, to be a potential source of tools for institutional analysis that allow for complex configurations of causes and effects. Fuzzy logic tools rely on set function logics that differ from traditional variable and statistics logics. The first big question to consider, then, is whether or not it is appropriate for a given analysis to shift from variable and probability assumptions to set assumptions. Introduction to Set Logic Traditional quantitative tools focus on measurement of variables at the highest reasonable level of precision. It assumes the value of capturing as much measurable differences between cases as possible. Statistical tests tend to assume additive (often linear) functions between the independent variables and the dependent variable. Interactive terms can be used to model possible ways in which two variables interact, but the underlying assumption is independence of the independent variables. The statistical tests use linear distance or probability functions to estimate the impact of each independent variable on the dependent variable. So, important underlying assumptions are that differences between cases are to be measured with as much precision as possible and that additive and probability functions are appropriate to determine the influence of independent variables (institutional components) on the dependent variable (system performance). Set logic focuses on whether or not cases belong to particular sets or not, rather than the placement of a case along an established ordinal or interval dimension. Nominal or dichotomous variables differ little from crisp set assignments (belonging = 1 and not belonging = 0) in terms of measurement here. For example, either approach would code irrigation cases as belonging or not to the set of systems with a membership boundary rule. Those systems with a clear membership boundary rule are coded as a 1 and those without such a rule are coded as 0 with both set and variable logic. A shift from crisp set membership (0 and 1 only) to fuzzy set membership allows analysis to capture more variation and allows measurement that reflects typical linguistic assessments of the cases. Cases may now differ in the extent which each condition in the configuration is met. For example, we can assign membership scores in the set strong boundary rule with values of 1 for very strong boundary rules, 0 for cases where no boundary rule clearly exists, and values between 0 and 1 to reflect categories like: mostly strong boundary rule or almost strong boundary rule. Fuzzy logic provides mechanisms for manipulation of assigned set scores such as negation (translating the scores into scores for no strong boundary rule), concentration (translating the scores into scores for a set condition of very strong boundary rule) and dilation (translating the scores into scores for a weaker set condition of boundary rule).

Using fuzzy membership requires the analysts to clearly define a membership function for each set. The first step is identifying the two endpoints. What does a case that clearly falls into the set look like? What does a case that clearly does not fall into the set look like? The next step is to identify the meaningful variation between full membership and no membership. There are different ways to do this. Ragin (2000) stresses the importance of developing a clear cross-over value first, which serves as the meaning for the .5 set membership value. What do we understand as the characteristics or the point at which we tend to move from assuming that a case is out of the relevant set to assuming that it could fall into the relevant set. His discussion of set measurement focuses more on set assignments to fuzzy sets based on an underlying continuous variable (e.g., GNP, union membership), so he stresses the need to identify a membership function that translates the interval measure of a variable into meaningful fuzzy set membership values. A membership function may use a crossover point and a slope or step function assumption on each side to translate existing interval data (such as national GNP) into a fuzzy set (such as rich countries). He stresses that developing the membership function requires the analyst to explicitly decide what variation in an interval variable holds substantive and theoretical meaning. Is the theory about rich countries one that assumes a substantive meaningful difference between the US and the UK, for example, or does the logic of the theory really only address differences between mostly rich and mostly poor countries (not assuming a linear relationship between wealth and an outcome, but rather assuming a logical relationship between sufficient wealth and an outcome) (Ragin 2000, p. 157-159). Other proponents of fuzzy logic stress linguistic measurement assumptions and emphasize fuzzy logics strength in manipulating intuitive linguistic values (e.g., Arfi 2005). The goal here is to develop membership categories that mimic intuitive linguistic evaluations of experts, with categories like weak, strong, very strong or very vague, vague, mostly clear, and very clear. Both the Ragin approach and the linguistic approach require researches to use expertise to develop the membership assignments for the cases. If using fuzzy logic for meta-case analysis, one would likely see a coding process that uses expertise experience from qualititative analyses and then refines the membership assumptions to improve substantive meaning and inter-coder reliability in the first iterations of coding. Whether a set of well-developed intermediate or linguistic membership scores would improve or weaken inter-coder reliability on coding qualitative cases for meta-analysis is an interesting methodological question to test. While the fuzzy membership assignments of .8 and .2 for the set clear boundary rule may look like values for an interval variable like degree of membership boundary clarity, they are not. There is no assumption that the nearly clear rule (.8) is four times as clear as the mostly vague (.2) rule. The assignment also looks like a probability, but it does not operate like one. Coding a case as .8 membership on clear boundary rule does not mean that one assumes a .8 probability that a clear boundary rule exists. Instead it means that one assumes that the case definitely has a boundary rule that falls short of a clear boundary rule. The difference between fuzzy membership scores based on discrete variables and probabilities becomes even clearer when we move to analysis of combinations. Probability combinations are assumed to be the product sum of

the individual probabilities. Set combinations of fuzzy scores for discrete sets use logical combination methods instead (AND, OR, XAND, and combination functions of these functions), which produce different results. The data simplification illustration discussed below illustrates this difference. Set assumptions apply not only to the assumptions about measurement of cases and the manipulation of set measurements (negation, dilation, and concentration), but also to the analysis of combinations and relationships. The relationship analysis emphasis shifts to an IF-THEN logic with no prior assumptions about linearity or variable independence. The object of analysis is to identify an appropriate set of IF-THEN rules that predict an outcome set effectively. Ragins (2000) set logic method for social science assumes that the IF-THEN types of rules that are most useful for social science are those that can be understood as necessary and sufficient conditions. These can be interpreted in IF-THEN logic as: IF (outcome met) THEN (necessary condition exists) and IF (sufficient condition met) THEN (outcome met). The analysis recognizes that the absence of a component is also relevant, so the absence of some condition (e.g., residential mobility) may be a necessary or sufficient condition for sustainability of an irrigation system. It also assumes that a combination of set conditions may operate together as a necessary or sufficient condition. So, no residential mobility AND weak boundary rule might be sufficient for sustainability. Ragins method works best with a configuration of a small number of independent sets. While the fuzzy membership scores create more variation, one can still use the fuzzy cutoff value to sort cases into sets of logically possible combinations of the presence and absence of the configurations to see how many cases fall into each basic type of configuration and to see which types of configurations are missing. For example, one would treat a case with a boundary rule membership of .8 as boundary rule present for sorting into logically possible configurations. Each logically possible combination present reflects an INTERSECTION (AND) of the particular set conditions. (A case with no residential mobility and weak boundary rules is a case with an intersection of those two set conditions.) So, one can use INTERSECTION to establish a measure of the extent to which each case in the analysis reflects each logically possible combination of independent sets. Fuzzy logic INTERSECTION of fuzzy set scores takes on the value of the minimum set membership score in the combination. In our above example, the relevant intersection would be ~residential mobility AND weak boundary rule. Other fuzzy logic analysts would call the score that results a degree of fulfillment (Bardossy and Duckstein 1995). Another way of understanding that number is as the level of truth that the case reflects the intersection of those characteristics. So, a case with very low scores on residential mobility (which by negation would be very high scores on ~residential mobility) and very high scores on weak boundary rules (assuming an at least weak boundary rules assignment here) has a high degree of fulfillment (DOF) or a high level of truth. This case matches the configuration conditions well. Similarly, cases can have low degrees of fulfillment and thus low configuration scores reflecting that the

conditions do not describe the case as well (the intersection of the conditions are not as true about that case). So, one can develop a table with all the logically possible combinations and calculate a degree of fulfillment score for the intersection of those conditions for each case. Those cases with high DOF scores are cases that reflect that combination of characteristics more closely. Cases with a value of 1 would be those cases that exactly reflect those conditions. Each case also has a membership score on the relevant outcome (similar to a dependent variable). The analysis of necessary and sufficient conditions uses subset logic to draw conclusions from the comparison of the membership score for the outcome for each case to the degree of fulfillment score for the intersection of each particular combination for each case. The relevant comparison for his method is less than or equal to. The first step is to look for combinations that operate as necessary conditions. To do this, we see if we consistently have cases for a particular configuration with outcome scores that are less than or equal to the configuration scores. Cases that meet this condition for each configuration support the argument that the cases with the outcome operate as a subset cases with that combination of conditions. If all (or using probability assumptions enough cases to meet our level of confidence) cases have outcome measures lower than the configuration measure for a particular configuration, then we can say, for our population of cases, that the outcome set condition is a subset of the intersection of those conditions. Ragin (2000) argues that when this is true, this configuration operates as a necessary condition. This follows the IF-THEN logic for necessary conditions outlined above. IF (the outcome occurs) THEN (the configuration exists). The configuration is a necessary condition because we do not get high values on the outcome UNLESS we have even higher values on that configuration of conditions. If there are necessary conditions, then those conditions must be in any configuration of possible sufficient conditions. No sufficient configuration could lead to the outcome if it did not contain a necessary condition, if a necessary condition exists. The next step is to look for sufficient configuration conditions for those cases with nonzero values for the necessary conditions. The relevant set analysis here is that the sufficient conditions are those that are found to be a subset of the outcome for the set of cases. These are configurations for which all (or a sufficient number) of cases have configuration scores lower or equal to the outcome score. If this is true, then the configuration is understood as a subset of the outcome. This is argued to be consistent with the IF (sufficient conditions) THEN (outcome) logic. The configurations found as sufficient conditions are then simplified by eliminating subsets of conditions that are covered by simpler conditions (as long as the simplification does not eliminate a condition found earlier to be a necessary condition). The simplified sufficient conditions statement, then will be a set of configurations joined by UNION that could lead to the outcome. So, (~residential stability AND weak boundary rule) OR (farmer managed) might be sufficient conditions that appear for an analysis of sustainability of a system.

The less than or equal to analysis of subsets for sufficient conditions still strikes me as weak. A configuration with a low level of truth on the conditions would have a low DOF value, if it is then lower than the membership score on the outcome, how does this make a convincing argument that the conditions comprise a meaningful set of sufficient conditions? The discussion of sufficient conditions emphasizes that sufficient conditions are those cases that fall above the diagonal when mapping the conditions as the horizontal axis and the outcome as the vertical axis. Any case that falls above the diagonal is counted as evidence that the configuration is a sufficient condition. Ragin (2000) argues for an even looser standard when he argues that it makes sense to allow a margin for error and accept cases that barely fall below the diagonal. There are three possible ways to think about strength that might move us beyond less than or equal to standard. Bardossy and Duckstein (1995) discuss setting a threshold value that can be used to determine whether to include a rule in ones model based on the extent to which one can conclude whether or not the outcome holds as a consequence of knowing the conditions (p. 155). One could perhaps argue that conclusions about sufficient conditions would need to be made relying on cases with DOFs over .5, since that would at least mean that the combination more clearly fits the configuration conditions than not. Another possible tool for pushing beyond subset analysis alone is a measure of inclusion. There are different definitions and formulas for the measure of inclusion, but basically it serves as a strength measure for subsets. So one can go beyond simply subset or not. The inclusion measure would indicate the sufficient conditions that are stronger subsets of the outcome. Analysis could use a threshold measure of inclusion and only consider as sufficient conditions those configurations that yield at least particular measure of inclusion on the outcome set. One strength of set analysis of configurations that pays careful attention to the existence or absence of each logically possible combination of conditions is the ability to analyze biases in case selection systematically and adjust results accordingly. Given the challenges of possible selection bias when conducting meta-analysis of cases (Poteete and Ostrom 2005), this feature could lend extra credibility to analysis of patterns in the cases . Ragin (2000) argues that this feature provides more careful analysis of the possible problems from missing diversity in the data. Combinations of conditions might be missing because of the selection of cases process, in which case identifying missing cases and explicitly stating the case conditions for which the analysis cannot speak. Combinations might be missing not because of case selection bias, but simply because a particular combination rarely exists (or rarely exists long enough to be studied). If a missing set condition could reasonably be a reason for success, it can incorporated into the findings using intersect set logic. For example, if almost all of the cases analyzed fell into the set small, then the set condition not small is missing from the analysis. If there is plausible theoretical reason to expect that smallness contributes to success, the analyst can intersect small into the combinations of necessary and sufficient conditions developed for well-maintained systems to explicitly recognize that smallness cannot be ruled out as a component of the configurations that lead to success.

Analysis of institutional configurations to outputs using fuzzy logic need not use necessary and sufficient condition logic. The broader emphasis is on developing IFTHEN statements that are theoretically meaningful with predictive power. So, Ragins method of identifying necessary and sufficient conditions is one variant of the possible model construction methods that could be used. Neural networks analysis is frequently mentioned as an inductive method that can be used to develop IF-THEN models. The neural network approach has three disadvantages for developing theoretical fuzzy logic models for institutional analysis. It can overlearn and create explanations that are too specific to the training cases (also a possible problem if the Ragin method is used inductively (Lieberson 2004)), training neural networks requires large training sets, and the neural networks can result in models with serious black box limitations (Bardossy and Duckstein, p. 3, 124). We may be able to predict with a neural net model, but may not be able to explain the theoretical relevance of the model. Bardossy and Duckstein (1995) argue that the more IF-THEN rules an analyst can specify based on expertise the more the analysis uses fuzzy logic strengths. They indicate that one can specify part of a model that is better understood and then have neural network analysis to help specify a part that is less understood. A fuzzy model of an outcome is really a set of IF-THEN statements. A fuzzy logic model can also be represented as a matrix if the analytical question is not whether or not a particular outcome is met (like sustainability), but instead which of a set of possible outcomes is met. For example, a common fuzzy logic example is a fuzzy logic matrix model that maps symptoms to possible diseases. So the cause and effect model developed focuses on which disease(s) relate to each configuration of symptoms, not on simply predicting whether the patient has a single particular disease. The matrix, then, represents a model that allows multiple sources of causality and does not assume linear or additive relationships. Arfi (2005) illustrates the use of fuzzy logic for predicting rational choices for political actors given assumptions about preferences or criteria for choices. Fuzzy logic is used to allow for linguistic preference orderings. This piece illustrates aggregation set functions for linguistic matrices. It also demonstrates the use of a weighted aggregation function that allows for different levels of ORness and ANDness when joining sets.

Fuzzy Data Simplification The Ragin method, and arguably other methods discussed thus far, work best with a reasonably small number of premises. Nearly 10 years ago, we tried the crisp set logic method and quickly found that it did not handle long rule strings well. However, institutional analysis, especially analysis attentive to the Institutional Analysis and Development model and institutional diversity, rarely lends itself to a small number of simple premises. One of the possible strengths of fuzzy logic assumptions for institutional analysis is its assumption about the meaningful nesting of IF-THEN rules. Fuzzy logic discussions from math, computer science or engineering often tout the ability of fuzzy logic to develop linguistically interpretable models at multiple levels (Bardossy

and Duckstein 1995). Lower level, or local simple models can solve for a configuration fuzzy membership score, that then can become one of the pieces in a model of a configuration at a broader level of generalization. So, just as intersection or union can be used to create a single degree of fulfillment score for a simple configuration for subset analysis, so too, logical aggregation could be used to create a single score that reflects multiple institutional characteristics about a case. Fuzzy logic proponents stress that fuzzy logic allows the development of these nested models in which each level remains tied to meaningful linguistic statements (unlike neural networks black box conditions). Fuzzy logic aggregation, then, could be used to create a single numerical representation of a complex set of institutional conditions, while still retaining the linguistic meaning behind that number. The example below illustrates this potential as well as an important difference between discrete fuzzy set logic and probability logic. Consider a set logic analysis of rules-in-use and a set logic analysis of culturally competent rules. Assume that cases were coded according to their membership in the set has a clear membership boundary rule. Now, assume that cases are also coded into membership in the set boundary rules enforced. (Put aside for the purpose of this example the fact that a workshop coder would likely only code an enforced rule as a rule.) A case with a membership boundary rule in use would be modeled in set logic as the intersection of the set assignments for clear boundary rule and boundary rule enforced. The extent to which a membership boundary rule is understood as a rule in use depends on the extent to which the rule exists AND is enforced. In crisp set logic, if either is not true, then the potential rule would not be considered a rule in use and would be assigned a 0. Fuzzy set logic allows more variation by allowing different levels of membership (or truth) to the statement that a case contains a membership boundary rule in use. A case with a .9 membership in clear boundary rule and a .7 membership in boundary rule enforced has a .7 membership in the intersection of those sets. The intersection (AND) assumes that the minimum membership is the critical one for assessing joint membership. In other words, the institution is only as strong as its weakest link. If the scores were probabilities, to assess clear boundary rule AND enforced, one would multiply the scores to get a probability of a rule in use of .63. The difference is not just a mathematical one, but also a difference in what the number means. The number .63 means the likelihood that a clear boundary rule exists and is enforced (the likelihood that the case has a rule that meet the conditions for full membership), this is different from the .7 membership number which means that the extent to which the case has a rule in use is .7. Now, lets look at a more complex intersection case. Lets assume that were measuring cultural competency of each rule by assigning the case a membership score to the set consistent with religious norms. Assume that this same case as a membership of .7 for their membership boundary rule. The membership score for rule in use AND culturally competent would be .7. The minimum membership score is .7 and so the intersection membership score is .7. If these were probabilities the probability would be .63 * .7 = .44, which would mean a .44 probability that a culturally competent rule in use exists. Notice the difference in results and logic. A combination of institutional elements that all have strong membership scores would yield an intersection result that also has a strong

membership score. There is no assumption that it is necessarily harder to belong to a complex configuration in set theory whereas probability logic assumes that the more complex the configuration, the harder it should be to expect that configuration to exist. Intersection is not the only way to model a configuration in set theory. One might argue that a model of some configuration of institutional elements is better represented as a union. This assumes that the institutional configuration being analyzed is as strong as its strongest link (rather than as strong as its weakest link). Union logic can be understood as a model of compensatory capacity (Arfi 2005). When applied to an institutional configuration, the union logic assumes possible compensatory qualities to the institutions; that a strong rule in one area can make up for a weaker rule in another. It may be the case that one strong boundary rule can make up for other weaker boundary rules. If so, it may be reasonable to simplify a string of membership scores for boundary rules by using a union function. Again, the assumption here would be that the rule string is as strong as its strongest link. The intersection and union functions also apply to crisp sets. So, one could analyze a rule string with zeros and ones (such as those in (Ostrom 2005a; Ostrom, Gardner, and Walker 1997)) using either. The logic is the same. An intersection is as strong as its weakest link and a union is as strong as its strongest link. With crisp sets and simple intersection or union functions, though, the intersection would get pushed to zero in configurations missing any of the rules and would be pushed to one in unions that include any of the relevant rules. Fortunately, fuzzy logic provides more flexibility than a choice between intersection and union, fuzzy logic formulas can be developed that combine different levels of andness and orness by incorporating weights into the aggregation formulas (Arfi 2005, Bardossy and Duckstein 1995). So, for example, one might argue that boundary rules are mostly as strong as the strongest link, with the level of mostly related to how weak that weakest component is. Fuzzy logic allows for a wide variety of combination formulas. The key would be to develop formulas that have a theoretically sound justification given the analysis at hand. The development of these aggregation formulas provides a way to develop simple set assignments that reflect a complex set of assumptions. Fuzzy aggregation formulas also provide a means to develop set assignments for configurations of rule strings that begin as crisp sets on a rule string that do not simply get pushed to 0 or 1. A fuzzy aggregation of a string of boundary rules will yield a single boundary rule configuration value with a formula that indicates whether the configuration is assumed as strong as its strongest link, as strong as its weakest link, mostly as strong as its strongest link, or some other assumption. Likewise, one could develop aggregation functions for action situation components or for any institutional part of analysis that can be understood as a combination of set conditions. These functions would provide a single measure for the institutional component with a summary of the assumptions behind the combination. The creative possibilities for assigning numerical values to configurations of institutional and contextual conditions appears to be one of the strongest potential benefits of adopting set assignment features of fuzzy logic. Fuzzy set logic provides a mechanism for developing a numeric representation of a complex configuration such as a portion of a

10

rule string, or a combination of some rule string portion and some physical world and cultural components. The configuration score can be developed with a theoretical model in mind. Fuzzy set membership, then, can be developed to reflect membership in a configuration of sets. To illustrate how this might look for institutional analysis, lets return to a boundary rule example. Imagine that we have boundary rule string data for a set of cases. This boundary rule string data has zeros and ones to indicate which of the commonly found types of boundary rules exist for each case. Now for each boundary rule that does exist, assume that we also have some indication of whether that rule appears to be supported by norms (perhaps a fuzzy membership set like: very much, mostly, barely, not). That norm assignment might itself be the result of a lower more precise model of norm support created by a weighted combination of the assumed shared minimal external delta and assumed minimum internal delta for that rule in that community. The combination of rule existence and normative support might then be combined with a measure of enforcement support (that might itself be pulled from analysis of a configuration of other elements in the action situation that tend to lend themselves to effective rule enforcement) to create a configurational assignment that would represent the strength of that boundary rule. If a boundary rule configuration only included a membership boundary rule, then the result would just be a strength measure for that one rule. In situations with multiple boundary rules in place, one could develop a boundary rule strength measure for each one that exists. In some instances the boundary rule results could then be combined with expert knowledge about the effectiveness of particular kinds of boundary rules in particular kinds of settings (Im thinking about a fishery example here, but I dont remember the specifics) to allow for a calculation of strength AND context appropriateness of the boundary rules. These examples illustrate the kinds of nestings that you could be developed. A likely goal would be to create nested models that use case codings of single characteristics to allow analysis at whatever level is appropriate for the question at hand. In some analysis it may simply be whether a membership boundary rule exists, in other analysis we may want to develop models that provide rich nested interpretations of the much broader boundary condition described as one of the seven characteristics of successful long-lasting commons (Ostrom 1990). The fuzzy logic not only simplifies the data (resulting in one fuzzy assignment number reflecting combinations of multiple assignment numbers), but also preserves explicit statements of the institutional analysis assumptions that yield that number. One could then go back and test the implications of changes in those specific assumptions. One can also assess the logical soundness of those explicit assumptions. Analysis with feedback loops and learning could then allow these expert assumptions to be refined. For example, the analysis might be aiming to clarify what makes boundary rules in action situations lead to conditions that one would characterize as meeting the boundary commons condition. This analysis might use feedback to keep refining the fuzzy combination rules to get the configuration scores closer and closer to the expert assigned membership score for each case for whether or not the case meets that commons boundary condition. This feedback adjustment for data simplification does not run into the issues raised when using feedback to change assumptions if we were conducting a

11

cause and effect analysis (for example, of whether strong institutional boundary conditions are a necessary condition for sustainability).i The fuzzy set nesting models provide a way to connect many specific rules and institutional configuration conditions to broad conditions such as the seven design principles for commons governance (Ostrom 1990) or to conditions for adaptability (Ostrom 2005a). It also provides a mechanism to refine simplifications of broad conditions such as the conditions for effective adaptability. Fuzzy set reduction could be used in conjunction with fuzzy set configuration analysis. One configuration condition (or more) could be a complex one (like the extent to which the case meets the broad boundary condition described as one of the seven design principles of effective commons (Ostrom 1990), as calculated from nested analysis of components assumed to create that condition). Fuzzy simplification along with configurational case analysis then may allow us methods to test the effects of changes in specific institution while keeping other parts of the configuration constant that does not rely on independence and additive assumptions. For example, one could compare the effects of strong membership boundary rules to those for strong residential requirement boundary rules for cases with attention at the same time to the presence of graduated sanctions and farmer management and a broad other measure that measures the extent to which the case meets a weighted combination of four of the design principles. Analysis could also examine the necessary and sufficient conditions with complex institutional premises such as an overall measure of boundary rule strength, an overall measure of payoff rule strength, an overall measure of rule compliance, and a water scarcity measure. Is it possible to use fuzzy logic for data simplification and then use traditional statistical methods to test additive or probability models? For example, could we test for a correlation between (strong AND enforced membership boundary rules OR ~residential mobility) and (strong labor contribution). Does it make sense to test for a correlation between (strong AND enforced membership boundary rules OR ~residential mobility) and an interval variable such as an interval measure of agricultural productivity? [I dont know the answer yet!] All of the discussion thus far has assumed that the smallest level of analysis for institutions is at the rule level with the relevant set membership functions indicating whether a particular kind of rule exists or not. The grammar of institutions pushes analysis down another level to components within rules (Crawford and Ostrom 1995; Ostrom 2005b). Fuzzy logic set assignment could work at this level too. For example, one might analyze the effects of inclusiveness of decision making with analysis that includes membership scores that reflect the inclusiveness of the ATTRIBUTE components of the aggregation rules for the action situation. If one is willing to move from a crisp set assumption about rule existence (it clearly exists or it clearly does not), then one could allow a fuzzy set membership on the OR ELSE component, which then, because the grammar assumes a rule is only a rule to the extent it has an OR ELSE, could be translated directly into a rule existence fuzzy score.

12

Fuzzy Dynamic Analysis When we move to analysis of institutional change, the most direct application of the methods discussed above uses the nested configuration models. The nested configurational models would allow an analyst to plug in a small change in a single rule and then see how that change would be expected to affect broader configurational characteristics of the case. One could trace the path of nested configurational statements that change (if any) to see the different ways in which the rule change would be expected to change the dynamics of the situation. If possible, we could compare the logical assumptions of the dynamics of change from our nested model to qualitative analysis of stories told by those in that setting about how the rule change changed the politics of the situation. The emphasis of fuzzy logic on a dense set of nested IF-THEN rules that have linguistic meaning encourages the development of models that should lend themselves well to shifting from static analysis of an existing rule configuration to dynamic analysis of the impact of a change in that rule configuration. When we shift to analysis that attempts to model adaptability and intentional institutional change, it makes sense to consider fuzzy logic tools for control and diagnostics. Control and diagnostic analysis get at the heart of the evaluative and feedback elements of the IAD model. Modeling control involves modeling the way that a system responds to feedback in an effort to improve performance. A traditional control model would be a sophisticated formula that calculated precise changes necessary in a system based on precise measures of a small number of input variables. Fuzzy control focuses on a dense network of feedback loops rather than precision or the elaboration of a single model of the system. The strong performance of fuzzy control systems in many situations speaks to the value of dense feedback loops and mechanisms that allow small localized changes for control. One cut at the question of why farmer managed systems have adapted better than government-managed systems would be to consider whether the farmer managed system more closely resembles a fuzzy control system while the government managed system more closely resembles a traditional control model. The farmers likely have more ways in which they gather information about small changes in the conditions of the system and in most cases more incentive to gather the input as well. The emphasis in fuzzy control on dense local mechanisms that take measurements that feed into the next level, and then the next offers another way to possible advantage of nested institutional systems with meaningful local incentives to pay attention to feedback and have some ability to influence resulting changes alongside instutitutions at other levels that can also receive feedback and make changes. Fuzzy logic diagnostic models offer other tools for modeling institutional change. These models would focus particularly on the evaluative criteria component of the IAD model. Diagnostic analysis with fuzzy logic relies upon expertise matrices. For example, a fuzzy medical diagnostic system would rely on a matrix that maps symptoms to possible diseases related to those symptoms. A doctor would evaluate a patient, develop beliefs about the symptoms of the patient and then aggregate the patients symptoms with the expertise matrix to get the likely disease (or possibly a set of diseases to treat). The doctors prescription, then, would reflect both his beliefs about the patients symptoms and the expertise within the medical diagnostic matrix. Models of patterns in the beliefs

13

generated in evaluations of a situation (the equivalent of the vector of symptoms for the patient) and models of the beliefs in the expertise matrix may provide tools to explore the roles of beliefs in intentional institutional change consistent with models of change such as those discussed by North (2005), those central to a punctuated equilibrium model of policy change (Jones 2001) or those consistent with an advocacy coalition model of policy (Sabatier 2003; Sabatier and Jenkins-Smith 1993; Sabatier, Loomis, and McCarthy 1995). One could use information about the beliefs of participants in action situations who engage in evaluation with any power to make changes and attempt to model the dominant expertise matrix that comes out of those beliefs. It would also be reasonable to assume that beliefs shape the nature of the evaluative vector for participants that results from their assessment of the interactions in the action situation. These models could reflect assumptions about selective perception and bounded rationality. For example a strong set of free trade beliefs might make it difficult for a participant to see trade-related problems and so one would expect that the participant would not see certain symptoms even if they exist. Likewise, strong free-trade beliefs would result in a matrix that would link negative economic situations (the symptoms) to shortfalls related to nationalistic or protectionist trade policies still in place (the diseases to be treated). Models of the beliefs and resulting evaluative and diagnostic matrices of participants could, for example, illustrate how strong shared beliefs maintain stability (as predicted by the punctuated equilibrium model) and then how a shift in beliefs can then lead to shifts in the vector or matrix that result in dramatic changes (or conditions in which they do not). A regime change might replace one belief system with another, ushering in a different set of diagnostic models. In situations with multiple participants with matrices that can be aggregated an intersection of those matrices reveals common ground. A weighted aggregation function might be used to reflect power differentials in the situation and predict the nature of the diagnoses of the situation that will be most politically relevant. In political analysis of diagnostics for institutional control one would assume that interests also shape diagnostics and control. So, we might establish an interest vector that reflects the extent to which a participant believes that each possible diagnosis in the matrix furthers their own interests. The expected diagnosis, then, would be an aggregation of the evaluation of the situation, the assessment of interests, and the diagnostic matrix. Models such as this could be used to model the negative dynamics of powerful participants with interests that sabotage diagnostics as well as the positive dynamics of participants with interests consistent with effective diagnostics. This illustrates just one possible analysis of adaptive capability. Other matrices could be devised to model the impact of other conditions on diagnostics for participants with power to make feedback changes. Fuzzy logic diagnostic and control tools, on the surface at least, seem appropriate for modeling intentional institutional change. They provide ways to model these processes that can reflect beliefs, bounded rationality, and interests. Intentional institutional change means change resulting from explicit attempts to change the structure. It does not imply that the resulting change in the action situation is necessarily what the proponents intended. This contrast with an emphasis on unintentional change in some evolutionary

14

models. The unintentional change in institutions occurs through errors or forgetfulness rather than through attentive diagnostics and control. In the end, we probably need models of institutional change that accommodate both intentional and accidental changes and that recognize that evaluation of the resulting change then solidifies or rejects those changes. Conclusion This cursory tour of a fuzzy logic toolbox is written in honor of an avid tool collector. Overall, the tools have certain strengths that lend themselves well to institutional analysis in the Ostrom tradition. They incorporate configurational assumptions, they provide a possible tool for simplifying rule strings for analysis, they incorporate bounded rationality assumptions, and they provide tools for analysis of institutional change. The tools also bring careful attention to the assumptions that guide analysis, which also seems fitting since much of Lins work has been uncovering and clarifying assumptions that others often unknowingly adopt. However, she has also always stressed to her students that tools should be used only for jobs for which they are appropriate. This essay highlights a broad set of potential uses, but it should not be read as an argument for jumping into fuzzy logic for all analysis. Much more work remains to be done in order for us to understand which of the tools work well and the institutional analysis tasks most appropriate to those tools. The tools need to be tried and adapted and the instruction manual needs to be developed. This essay just lays the tools on the table and invites others to join in the fun.

15

References Arfi, Badredine. 2005. Fuzzy Decision Making in Politics: A Linguistic Fuzzy-Set Approach (LFSA). Political Analysis 13 (1):23-56. Bardossy, Andras, and Lucien Duckstein. 1995. Fuzzy Rule-Based Modeling with Applications to Geophysical, Biological, and Engineering Systems. Boca Raton: CRC Press. Crawford, Sue E. S., and Elinor Ostrom. 1995. A Grammar of Institutions. American Political Science Review 89:582-600. Jones, Bryan D. 2001. Politics and the architecture of choice : bounded rationality and governance. Chicago: University of Chicago Press. Lieberson, Stanley. 2004. Comments on the Use and Utility of QCA. Qualitative Methods: Qualitative Methods Section of American Political Science Association. Ostrom, Elinor. 1990. Governing the Commons: The Evolution of Institutions for Collective Action. Cambridge: Cambridge University Press. Ostrom, Elinor. 2005a. Developing a Method for Analyzing Institutional Change. Paper read at Who Should Do What in Environmental Governance: Institutions and Constraints, at Sardinia, Italy. Ostrom, Elinor. 2005b. Understanding Institutional Diversity. Princeton: Princeton University Press. Ostrom, Elinor, Roy Gardner, and James Walker. 1997. Rules, Games, & Common Pool Resources. Ann Arbor: University of Michigan Press. Poteete, Amy R., and Elinor Ostrom. 2005. Bridging the Qualititative-Quantitative Divide: Strategies for Building Large-N Databases Based on Qualitative Research. Paper read at American Politcal Science Association, at Washington, DC. Ragin, Charles C. 2000. Fuzzy-Set Social Science. Chicago: University of Chicago Press. Ragin, Charles C., and Benoit Rihoux. 2004. Qualitative Comparative Analysis (QCA): State of the Art and Prospects. Qualitative Methods: Qualitative Methods Section of American Political Science Association. Sabatier, Paul A. 2003. Beliefs and Institutions in Water Resources Management: Advocacy Coalitions and Collaborative Governance. Paper read at Symposium on Adaptive Governance and Florida's Water Conflicts, November 21-22, 2003, at Florida State University. Sabatier, Paul A., and Hank C. Jenkins-Smith. 1993. Policy change and learning : an advocacy coalition approach, Theoretical lenses on public policy. Boulder, Colo.: Westview Press. Sabatier, Paul A., John Loomis, and Catherine McCarthy. 1995. Hierarchical Controls, Professional Norms, Local Constituencies, and Budget Maximization: An Analysis of U.S. Forest Service Planning Decisions. American Journal of Political Science 39:204-242.

16

Ragin (2000) argues that it is not inappropriate to go back and adjust the set membership functions as long as we are staying true to our theory assumptions, that the refinements provide a means to identify the appropriate sensitivity of the measures. However, it is easy to see how feedback from relationship testing to change the condition assumptions of the independent variables to get better relationship results leads to criticism of manipulation of data and poor to nonexistent measurement discipline. (Thanks to Terry Clark for discussions of this issue with me.) For analysis of cause and effect IF-THEN rules, Bardossy and Duckstein (1995) recommend use of training set and a test set to avoid this problem. Obviously, this requires a large enough set of cases. Simulations analysis could produce ample case data that could be used for training perhaps to ease this problem.

17

You might also like