Simulating “Stage” Theories: An Exploration with Applications Edmund Chattoe-Brown1 1 Department of Sociology, University of Leicester, University Road, Leicester, LE1 7RH ecb18@le.ac.uk Abstract. This paper presents a simulation of so called “stage” theories (ST), systems in which agent states are constrained to follow sequences. Put more simply, agents can only reach stage x from stage x-1. A general example is provided by education (where one must understand arithmetic before algebra) but stages are often found in social theories, for example, those of Becker (marihuana use) or Rambo (religious conversion). Despite their ubiquity, ST have received little attention as a class. This paper intends to clarify their properties using simulation prior to empirical research. Keywords: Stage theories, religious conversion, simulation, social theory. 1 Introduction It is a commonplace of social science that certain processes occur in sequence. A general example is offered by education in which new knowledge must build on old. Extending this idea, we can view knowledge as a “tower” where the solidity of the foundation determines the maximum height. 1 However, we also find stages in specific theories. For example, in a famous paper, Becker [1] identifies stages in becoming a marihuana user. First the potential user must learn to inhale effectively, then they must learn to recognise the state of being high and finally they must learn to enjoy that distinctive state. At each stage, effective use may by prevented by a failure to learn the appropriate skill. However, the potential user always moves forward from the last stage they have mastered rather than having to go “back to the start” as one might after “dying” in computer games. This idea of stages is simple enough but although widely used in social science, the foundations and formal properties of such systems seem to have received little consideration. I shall postpone most foundational discussion until after a simulation has been presented and analysed but, in the next two sections, I shall define the kind of ST that will be discussed and examine their relation to existing simulations. 1 This is a more humanist view than the (often implicit) hypothesis that individuals have an inbuilt “limit” but, to my knowledge, educational research maintains the ideology that everyone can learn forever, thus neglecting such issues. To a practising teacher, this ideology is obviously wrong even if limitations actually reside in teaching and not learning. 2 Micro and Macro ST This paper will not explore what might be called macro ST, those in which whole societies supposedly proceed in stages. An example is provided by Comte’s hypothesis (the so called “law of three stages”) that all societies pass through theological, metaphysical and positive phases [2]. It can be seen that (from the ABM perspective) such theories are as unconvincing as many social scientists find them for other reasons. If we regard stages as particular sets of social arrangements then macro ST require implausible determinism without clear cause. For example, in undermining theology, it is “necessary” that metaphysics must arise and prove its value. 2 This observation shouldn’t be over stated. Certain social arrangements may be attractors or capable of reproducing themselves so macro societal patterns should not be ruled out but the logic of robust sequences in these patterns is unclear. Thus, this paper will only consider micro ST where sequences are supposed to exist in attributes of individual agents. Apart from the interest arising in this general class of theories as a result of their use in social science, they are also not so selfevidently unlikely. To foreshadow later argument, it is relatively easy to see how some states of an agent can be logically prior to others. For example, it is hard to see how agents can make choices when they don’t know the alternatives available and especially when they don’t know choices are available [3]. 3 Existing Simulations (and Other Research) The ABM implementation of ST can be extremely simple. A sequence of states is defined (for example as integers 1-10 constituting one attribute of an agent) and social processes produce ordered transitions between states for each agent: 1 to 2 to 3 and so on. It is the nature of these social processes that situate ST within existing simulation research. For example, in social learning models, an agent may advance one stage when it meets another at that later stage. This system is very close to a simple social influence simulation [4] except that “influence” only moves agents one way. This might be interpreted as an absence of “forgetting”. Once agents reach stage 5 they cannot revert to stage 4. This can be contrasted with a standard social influence model where an agent’s opinion may move from 4 to 5 and later reverse. 3 Micro ST can also be distinguished from systems like Granovetter’s [5] threshold model. It isn’t the number of other agents at any stage that directly determines “willingness to adopt” that stage (although it does indirectly through the ecology of agents at other stages). 2 In the simpler case of technology, the development the wheel does not necessarily occur. This might be seen as a weakness. An agent could lose credibility or be unwilling to display constant variation of opinions. Returning to a previous opinion might imply “mistakes” or lack of justification. Memory/self-consistency in opinion dynamics is usually ignored in simulation. 3 4 Preliminary Analysis Something can be understood about ST before simulation occurs. In a social influence model where agents can only achieve stage x if they meet other agents already at that stage, no agent can pass the latest stage in the initial population. A ST of this kind lacks what might be called “self starting”.4 One strength of evolutionary theories (for example) is that they only involve variation and selection but no quality requirement on system initialisation. However awful the initial population, as long as some instances are less awful, it is these that form the basis of subsequent populations and gradual improvement will occur from whatever starting point. By contrast, social influence cannot explain how initial populations achieve later stages. This seems unsatisfactory. The obvious solution, that agents can progress through stages without help seems to defeat the point of social influence. Why do we postulate social influence when agents can proceed alone? Should we apply Occam’s Razor? In fact, the situation isn’t quite so bad. Innovation and diffusion are conceptually distinct and we can still ask two interesting questions of “pure” social influence. The first is (given agents with infinite life times) what is the distribution of stages? The second is (given agents with finite life times) what is the latest stage that is sustainable in the population?5 In an educational context might we worry that, under certain social arrangements, the latest stages would simply fail to reproduce themselves effectively? This raises a third issue. The system dynamics will depend significantly on assumptions about the ability of those at later stages to adjust the stages of others. If agents can only do this one stage at a time then, as already discussed, the system is limited by latest initial stage. However, it is also potentially limited by the distribution of stages. It is easy to see how an agent at stage 1 might be left behind by the population. If all stage agents had moved on before they encountered the unfortunate stage 1 agent there would be no subsequent opportunity for it to change stages. 6 By contrast, this problem would not arise if agents at later stages could influence all those at earlier stages rather than just those adjacent to them in stage terms. In a teaching context, arguments exist for making different assumptions. On one hand, teachers might be able (at least) to bring pupils up to their own stage. On the 4 This is also true of many innovation diffusion models. Interestingly, the fact that all individuals may be “progressing” does not imply that societies progress. In fact, in a pure social influence model, the best a society can hope for is stable social reproduction and regression is perfectly possible. This in an interesting side light on the relation between micro and macro ST. 6 An interesting application involves the interface between educational institutions. It is one thing to bewail “dumbing down” in education as a whole but how much more damaging if gaps arise between schools and universities such that most pupils simply cannot benefit from what is traditionally regarded as “university level” education despite formal eligibility. How much more difficult for each sector to close any such gap without co-ordination even assuming they are willing. (In maintaining standards unilaterally, universities may simply make education unproductive while believing they are being scrupulous.) Ironically, one way in which such gaps occur may arise from ST. If teachers complete university education without really understanding its goals, they cannot teach these and thus the gap is likely to be reproduced, although individuals may transcend it at considerable effort through individual learning. 5 other, teachers might be most effective teaching somewhat below the limits of their own knowledge. Too close and they might not understand themselves (and what is understood and can be well explained may differ), too far and they might not be able to recognise the state of the learner effectively. Such issues cannot be resolved by “armchair theorising” (and it is possible that educational research has never cast the issue of learning systems in quite this way) but it is important to recognise the potential significance of the exact social influence assumptions for system dynamics. The final issue to be considered is that of reproduction, both social and biological. If the stage a child has reached when it starts school is determined by the stage of its parents and the level at which teaching is “pitched” is determined by those who are, themselves, at a particular level (teachers/policy makers), it is clear how a relatively subtle but nonetheless divisive problem arises. Following Bourdieu and Passeron [7], the “objective” educational failure of groups can actually be attributed to arbitrary preconceptions and an explicit reliance on abilities not provided by the system itself. Unless the state undertakes to rear all children from birth, a “handover” from parents to schools is inevitable. However, it is far from inevitable that the consequences of the relation between the stages of pupils, parents and teachers should be blamed exclusively on the pupil nor (to return to the practical use of simulation) that the consequences of such a potentially complex system should be susceptible to “linear” policy. Returning to the issue of interfaces between educational sectors, we can see how the ability of the whole system to deliver a particular distribution of stages in the population might depend crucially on the ability to maintain the latest stages. 7 Failing this, social “regression” may occur by a kind of “unravelling” process. 5 A Case Study: Rambo’s ST of Religious Conversion In this section, I discuss a significantly different application of ST. The purpose is firstly to demonstrate (within the relatively small compass of a single paper) that these issues are of general relevance and secondly to analyse the detailed foundations of ST. By this I mean that it is easy enough to specify a ST but less clear why this general class of theories exists and when it applies. For example, how likely is strict logical priority of stages in a typical social system? In what social settings are stage processes likely to be found? What kinds of social processes (collective or individual) are likely to produce movements between stages? The example is Rambo’s [6] ST of religious conversion. This involves the following stages: Context: All the things that brought the person to their current state where the study of the conversion process is supposed to start. [A biography of person x.] Crisis: What happens to destabilise their belief system. [The death of a loved one.] Quest: The search for an alternative belief system resolving the crisis. [A need to find a supernatural rationale for death absent from an atheist or agnostic standpoint.] 7 There is also an issue about the law of large numbers. Probablistic “gaps” in the learning ladder are unlikely to occur in big populations. However, they are presumably more likely in the small populations found at later stages. Encounter: A meeting with a person (or perhaps artefact?) which embodies the belief system resolving the crisis. [A conversation with a sympathic Anglican priest.] Interaction: Based on the encounter, the quester interacts with the relevant religious community and establishes that the faith does in fact meet his/her needs and that s/he is willing and able to commit to its requirements. [Visits to a local Anglican church.] Commitment: The seeker joins the religious community and commits to a role within it. [Tithing/other moral actions arising from membership, attendance, church duties.] Outcomes: The convert does (or doesn’t) find what they sought. [The death is at least accepted or the supernatural rationale doesn’t provide lasting comfort.] There is a lot for foundational understanding of ST here. Firstly, while context doesn’t look like a stage, it explains why other stages do in fact have that property. For example, different events will destabilise different belief systems (a death may destabilise an agnostic while encounters with evil may destabilise a Christian) and quest, interaction, commitment and outcome stages depend on the psychology, experience and so on that the person carries with them. The mistake is perhaps only to think of context as an undifferentiated “heap” of variables rather than (in ABM terms) time dependent specifications of agent state determining environmental interaction. Secondly, several stages involve external events but the nature of these can vary considerably. Crisis might involve a losing a loved one, “putting two and two together” about beliefs one previously thought made sense, witnessing how other people treat each other and so on. The same diversity can be imagined in the encounter and interaction stages. From an ABM perspective we need to allow for shifts that may occur cognitively without outside cause, on the basis of social encounters and through encounters with “artefacts” or any combination of these. Thirdly, choice is an important element in ST despite their apparently “deterministic” character. Context initially implies no choice. Crisis creates the need for choice. Quest identifies alternatives. Encounter and interaction sketch out potential costs, benefits and moral conflicts. Each of these stages seems strongly (if perhaps not logically) prior to the others. Without crisis, psychology urges “If it ain’t broke don’t fix it”. Without quest, there is no resolution to the crisis and no set of alternatives over which to decide. Thus a ST is deterministic in regard to the logic of the choice process and that, given a crisis, there will be quest till resolution (or death) but there is no determination of what faith will be chosen or how long the process will take. There may be other common social actions that “create” stages in this way. Fourthly, this kind of ABM specification can be used to identify stages which lack structural significance or have only “narrative” importance. For example, the first encounter with the faith that eventually resolves the crisis will be an important relief but little follows about its stage status. The encounter need not be a person, it might be as mundane as a story in a magazine bringing back memories of comfort from a neglected faith. In this case, while we can “define” the encounter as an event, it isn’t clear how this changes the dynamic of subsequent events. Given the definition of the ST, it isn’t clear that someone can “stick” at the encounter stage in the same way they can stick at the quest stage. From the ABM perspective, there is perhaps a requirement that a stage needs to involve elements outside the agent’s control. Questing doesn’t guarantee finding but a positive encounter makes it extremely unlikely that an agent wouldn’t search extremely actively for the corresponding interaction. (One could tell a contrived story about meeting a priest from an obscure faith and a seeker discovering that there was no church of that faith in his/her country but generally encounters strongly imply the possibility of realising interaction.) Finally, the example of reinvigorated faith in the face of crisis and the need to evaluate a faith (potentially resulting in further quest) draw attention to the possibility of loops. These should be distinguished from branches (to be discussed later) in that they do not violate the assumption of deterministic sequence used to define ST. 6 A Basic Simulation, Results and Discussion As befits an exploratory simulation, the model is extremely simple. Each agent has an attribute definining its stage (a positive integer) and, during interaction (based on random mixing), one agent at a higher stage may change the stage of another (at a lower one) by “teaching” it in a small number of distinct ways defined below. User controlled features are the population of agents, the number of those agents who are “starters” (not initialised at the earliest stage and thus serving as potential “teachers” in the system) and the number of stages.8 Different assumptions can be made about the distribution of stages for “starter” agents. These may all be at the same stage (call this the “homogeneous starters” condition, HoS) or distributed randomly up to that stage (“heterogeneous starters”, HeS). Different assumptions can also be made about how one agent may learn from another. Agents may only learn from those at the next stage (“Adjacent Learning”, AL9), from those at all subsequent stages (“Inclusive Learning”, IL) or from those up to x stages later than them (“Nearby Learning”, NL). Finally, the extent to which agents are able to “learn for themselves” can be varied (“Self Learn”, SL, and “No Self Learn”, NSL conditions). Even for this simple simulation 12 experimental conditions exist in addition to variation introduced by number of stages and proportion of agents who are starters. I shall only look at outcomes of the experimental conditions and user controls separately. For experimental conditions, I used a standard setting of 200 agents and 30 starters in a 20 stage process. In each case the variable for comparison is the “steady state” average stage in the population after 1000 periods.10 All agents live for 80 periods and spend the first 10 of these being “socialised”: receiving learning mainly from their (single) “parent” rather than from the community at large. Figure 1 shows a typical run in HoSALSL condition.11 The starters clearly fail to transmit later stages to the wider population before they die and there is a drop to a steady state which can reproduce itself (societal “regression”). Even with SL, there is little variation in the latest stage attained by the most “advanced” agent and thus little opportunity for the population to move further through the stages. 12 8 Some stage processes may be open ended (have no latest stage) but I neglect this possibility. Here, for a 20 stage process, “nearby” is defined as within 4 stages in all conditions. 10 Because many runs show “regimented” behaviour, this value depends on when sampling occurs. To minimise this, samples are averaged over a “lifetime” from 920 to 1000 ticks. 11 The title reflects an ordered sequence of stages as a kind of “learning ladder”. 12 It is possible that under conditions yet to be identified endogenous “renaissance” may occur. 9 Fig. 1. Latest, Average and Earliest Levels for a Typical Run in the HoSNLSL Condition. Figure 2 shows the other common regime in the system (HoSILSL condition). Here the latest stage attained by starters is maintained and the average stage of the population is enhanced (societal “progress”). Interestingly, as usually occurs in this regime, the pattern of average stage is quite “regimented” despite the probabilistic nature of the system. This surprising result will be investigated further. Fig. 2. Latest, Average and Earliest Levels for a Typical Run in the HoSILSL Condition. Table 1 shows the twelve possible experimental conditions. The first finding is that the system is relatively stable, without large variations in steady state values. However, allowing for variation, 3-4 distinct output regimes exist. The first is “high learning” (average stage values around 15 out of 20 available) for conditions 2, 5, 8 and 11. Here the effect of IL, largely independent of SL/NSL is to ensure that later stages are rapidly transmitted to the population. In terms of process, the IL condition most rapidly blurs the initial distinction between starters and “others”. The second is “intermediate learning” for conditions 3, 6, 9 and 12. Here NL does not allow transmission as rapid as IL but it is still effective largely regardless of SL/NSL. The third condition is “low learning” for conditions 1, 7 and 10. Here, the potentially limiting effects of AL (agents can only learn if there happens to be an agent at the next stage in the population) is meliorated by SL but still severe. Condition 10 is the only one that looks like it may be significantly different from other runs in its output regime (with values around 0.9 rather than 1.2 – this will be investigated further). Finally, there is a unique “no learning” regime (condition 4). Here, with NSL, starters simply cannot transmit to other agents as they are too many stages apart. Table 1. Average Stage of Population Under Full Set of Experimental Conditions. Starter Condition HoS HoS HoS HoS HoS HoS HeS HeS HeS HeS HeS HeS Social Learning AL IL NL AL IL NL AL IL NL AL IL NL Self Learning SL SL SL NSL NSL NSL SL SL SL NSL NSL NSL Average Stage 1.25, 1.37, 1.25 15.1, 15.39, 15.44 11.62, 11.41, 11.64 0, 0, 0 15.23,15.27, 15.16 11.19. 11.83, 11.35 1.21, 1.31, 1.4 15.4, 15.59, 15.36 11.45, 11.65, 11.61 0.99, 0.99, 0.99 15.18, 15.36, 15.12 11.58, 10.94, 11.57 These condition combinations do not lead to outcomes in a simple linear way. Different “factors” dominate the outcome in different combinations making the system unsuitable for reliable narrative or statistical characterisation. In retrospect, this justifies the decision to simulate. Recognisable regression occurs in conditions 1 and 4 and progress in the other conditions. The sort of regimented outputs shown in Figure 2 made me want to try and “break” these results. I have not done this for all twelve conditions but for a single condition (HeSNLSL chosen randomly), it proved encouragingly hard. Table 2 shows the outcome of exploring the parameter space. 13 Columns 1 and 2 show the effect of changing the number of stages. Column 1 gives the number of stages. In column 2, the steady state average is expressed as a fraction of the number of stages “available”. The system reproduces short stage sequences much better than long ones. This outcome is non linear and draws attention to the importance of “demographic” and “structural” issues like lengths of socialisation, schooling and learning “lifetime” in the social reproduction of “hard” skills (those involving many stages). Columns 3 and 4 show the effect of varying the population size (keeping the fraction of starters constant at 10%). This effect is fairly small and relatively linear. Columns 5 and 6 show the effect of number of starters (out of 200 agents) on average stage level. This effect is very small until the fraction of starters gets very high and even then it is not particularly large. This suggests both that it is the behavioural regime that dominates and that, fairly soon, the clear distinction between starters and others in the population at large breaks down as knowedge spreads and the population replaces itself. 13 Because of the relative reproducibility of runs shown in Table 1, these are the results of single runs. Table 2. Effects of Various Parameters on Average Stage in the Population. Stages 1 10 20 30 40 50 60 70 80 Fraction 0.99 0.8 0.575 0.39 0.32 0.28 0.21 0.17 0.17 Population 10 30 50 80 120 150 180 200 Average 11.7 11.4 11.2 10.9 11 10.4 9.6 9.4 Starters 1 5 10 30 50 80 120 150 200 Average 11.8 11.8 11.6 11.7 11.5 11.9 12.1 12.3 13.1 Finally, there is a concern that this system may be sensitive to initialisation assumptions about the age distribution of starters and the stages of the other agents at the start of the simulation. In fact, at least for the condition chosen for further study, the steady state shows almost no sensitivity to these factors. These findings provide encouragement that any results discovered are not intolerably brittle (as the “regimented” output might have suggested). What can be learned? At best, these simple (but I believe novel) simulations have shown the possibility of societal regression (failure of social reproduction) and the potential hazards of “gaps” (especially in the AL condition) informally raised earlier. Also, this approach casts a different light on stage systems, revealing their connection to macro phenomena of social reproduction in an interesting way. This kind of simulation steps back from the individualised approach to learning and sees teachers and learners as operating within a macrosocial context with a potential for collective action failure. If any one group fails to pass learners up the “ladder of learning” effectively, this affects the system as a whole and not just the learners concerned. It may also feed back unpredictably to the structure of the learning ladder itself. However, given the simplicity of the model and lack of calibration/validation, it would be unwise to draw conclusions about when progress or regression occur. (Comfortingly, regression appears to be rare, occurring mainly in ALNSL conditions.) However, because (to my knowledge), there are no other models of this kind in the literature, even a preliminary analysis still adds value in opening up this set of ideas. 7 Possible Extensions and Conclusion There are several ways in which this research could usefully be developed: 7.1 Modelling “Real” ST It would be relatively straightforward to simulate Rambo’s theory. 14 In addition to understanding limitations of that theory by simulation, deepening understanding of 14 The core element might be some sort of bit string representation of religious faiths which “matched” (or did not) the current state of the agent. Mismatch would constitute a crisis. the “building blocks” of ST15 and comparing real and simulated data (for example the distribution of times spent in various stages based on biographical data), this exercise might also illuminate an important foundational issue. Is the existing social scientific notion of “stage” simply sloppy thinking for the precise specification of process practiced in ABM? As suggested above, it may be that some stages turn out to have no substantive “underpinning” in the ABM. Conversely, others may serve as workable (and measurable) summaries of particular stages real social actors go through. We can see how a social actor could realistically be put in the “quest” stage given its behaviour and reports of a past crisis. Finally, by recognising the distinction between the agent and environment, we may be able to provide a formal definition of a stage in terms of the agents inability to move to the next without some external input. It may be the existence of exogenous processes which defines stages. 7.2 “Solving” the Problem of Regression The notion that (even with personal and social learning operating) social reproduction may fail is a disturbing one. It is too early to say much about when and why this might happen but even the thought that it might is salutary. This way of thinking about learning systems suggests topics for both empirical research and further simulation. Empirically, it would be interesting to look at classroom dynamics in terms of information transfer between teachers and pupils. The ability of education systems to “pass on” students and avoid “learning gaps” would also make a good topic of study. In particular, can “mental models” of students and their congruence with the goals of learning serve as predictors of success? What happens to students at university who believe their task is to fill a “bucket of facts” as they did at secondary school? Such explicit attempts to match the state of the learner with the state of the teaching might even have important policy implications for educational success. 16 On the simulation side, there are also interesting issues to be explored. The role of artefacts in social simulation is often neglected, perhaps on the grounds that “things” don’t matter to social interaction. In this simulation, however, artefacts (in this case “books”) have a distinct role to play. In a sense, what is “wrong” with the simulation above (which allows regression to occur) is that everyone learns only from each other or from themselves and the system lacks an anchor in the non social world. If a gap occurs in the knowledge ladder, this is a bottleneck until it is filled. How likely self learning is to fill such gaps is a matter for conjecture. At low levels (getting the back off a computer), it seems reasonable. At high levels (learning calculus), much less so.17 On the other hand, the presence of durable artefacts at each stage of the learning process is an obvious resolution to the bottleneck problem and, furthermore, gives artefacts a distinct role in the system as things that do not have constantly changing 15 In what sense, for example are simulated versions of Rambo’s stages processually “similar” to any of Becker’s stages? Is there an atomic “set” of common kinds of stage from which all extant ST can be reconstructed? If so, this would be a useful theoretical generalisation. 16 My own direct experience suggests both that it is ineffective to “take for granted” any obvious skills in modern students and that it is actually quite hard to figure out what their model of the university educational process must be from their behaviour! 17 Another refinement is that higher levels of education explicitly try to turn students into more effective self learners. stages. On the other hand, this clearly isn’t a panacea. Even with a book on calculus, it would be unreasonable to assume that everyone could teach themselves that skill! 7.3 Details of the Learning Process For simplicity, this model involves the typical “random wander on a flat plain” assumption. In practice, education is obviously much more structured. Pupils are put in classes, classes in educational institutions and educational institutions into educational systems. What consequences do these kinds of structures have for the overall distribution of stages? What effect does it have if teachers and institutions only operate over a fixed range of stages/ages under a division of labour? 18 What happens if these stages somehow fail to join up? One example is provided by the concern with class origins in education. The simple simulation only looks at average stage, earliest and latest. Because there is random mixing, we are unlikely to see polarisation of learning with successful and disaffected groups [8]. On the other hand, without random mixing and with homophily in class groups and selection into schools, we can much more easily see how the learning dynamics of good schools might become a virtuous circle while that of “sink” schools becomes a vicious one. In the real world, it is known that children from some backgrounds arrive at school “behind” and are much less likely to catch up. Add to this the “comprehensive” notion (captured by this simulation) that the better students may “draw up” the less able and it is possible to see how phenomena like selective education, streaming and changes in the “stages” of schooling (and the corresponding testing regime) may have important effects on the overall efficacy of the system that are very hard to predict. That would be a classic opportunity for social simulation in a new area of social science!19 7.4 Formal Developments to ST We can think of ST as a hypotheses with varying degrees of “strength”. The strongest possible is that all cases follow a single sequence barring measurement error. Weaker is that a certain number of stages can be “missed” or occur in non standard order. Weakest of all is just that a sub set of stages should occur more often than any other sub set of possible stages. Simulations of ST (and their extensions) can be used to explain why such strong or weak patterns might exist in terms of underlying (and potentially less observable) behaviour [9]. What will happen to the robustness of a sequence when there is probabilistic transition between some or all stages? For example, in the Rambo model, what impact does the choice of different possible religions have on the overall properties of the sequence? Once such multiple routes 18 It is possible that as a first approximation such issues could be explored by making agent interactions conditional on things like the age and “type” of the agent. At present, everyone teaches everyone. It could be this assumption that might be relaxed in the first instance rather than focusing on the details of the spatial and administrative details of education systems. 19 There is, of course, a large literature on organisational learning but the structures of firms and schools are so different that the findings would be unlikely to be comparable. In addition, these models (to my knowledge) deal with the efficacy of individual firms rather than systems of firms. exist, it may also be possible to carry out some of sort of Markov chain analysis of the resulting systems and there is also an interesting theoretical question about when it is useful to view learning as a “tower”, a “network” or some other structured object. 7.5 The Problem of Data and Self Delusion This paper deliberately seeks to meet the criteria of the conference in describing a simulation and its outcomes. However, the comparison with evidence is more problematic and needs further work. Governments have a considerable incentive not to draw attention to failing education systems. There is also an argument that says that it makes no sense to talk about quality independent of particular measures of it. Nonethless, there is anecdotal evidence (for example from individuals informally administering the same test year on year) that ability may be going down while measures of that ability go up. Further, it is possible (though not easy) to think of ways that quality might be observed unobtrusively regardless of its formal assessments. For example, A level markers might covertly be given essays to mark that were actually written several years ago or students might be asked to take old exam papers without knowing they were old.20 At the same time, there is also a potential problem of self delusion. We tend to be uncomfortable with the idea of “Ages of Gold and Lead”. Who wants to be the best social scientist in a deeply mediocre era? Although it is a commonplace that older people always think they live in the worst of times, it is possible (and worthy of empirically grounded social scientific study) to consider whether and when that might actually be true! Acknowledgments. I am grateful to Asiyah Kumpoh whose interest in religious conversion in Brunei drew my attention to this class of theories. References 1. Becker, H.: Becoming a Marihuana User. Am. J. Soc. 59 (1953) 235-42 2. Comte, A.: The Positive Philosophy of Auguste Comte. John Chapman, London (1853) 3. Chattoe-Brown, E.: The Social Transmission of Choice. Mind and Soc. (forthcoming) 4. Hegselmann, R., Flache, A.: Understanding Complex Social Dynamic., J. Art. Soc. and Soc. Sim. 1 (1998) <http://www.soc.surrey.ac.uk/JASSS/1/3/1.html> 5. Granovetter, M.: Threshold Models of Collective Behaviour. Am. J. Soc. 83 (1878) 1420-43 6. Rambo, L.: Understanding Religious Conversion. Yale UP, New Haven, CT (1995) 7. Bourdieu, P., Passeron, J.: Reproduction in Education, Society and Culture. Sage, London (1990) 8. Willis, P.: Learning to Labour. Saxon House, Farnborough (1977) 9. Abbott, A., Hrycak, A.: (1990) Measuring Resemblance in Sequence Data: An Optimal Matching Analysis of Musicians’ Careers. Am. J. Soc. 96 (1990) 144-185 20 Both approaches have obvious problems but are better than nothing.