 OK, can you hear me? Good. Thanks very much. Thanks for the warm introduction. I'll try and stop being so productive. I think it's time for me to slow down a bit. Some of you may know that I nearly didn't get here thanks to the Chilean volcano, Ash Plume, which spread across Australia and delayed me for two days in getting here. So I'm very pleased to be here, actually. Carter just suggested that I should reenact a volcano to start the thing off. But this is not going to be a volcanic keynote. What I want to talk about today, and I hasten to add, that this is very much my own personal perspective on a number of issues. And I'm not claiming that this is the only way to view the world. And I certainly do not have a theory of networked social systems, nor do I have a universal method of analysis for networked social systems. So please don't expect that. What I want to do three things, really, today. The first thing is this. To argue and explain why, network theory and network analysis are different. Maybe I don't need to do that with this group, but that's what I intend to do. Secondly, to illustrate how a network theory and analysis go hand in hand with an actual case study that we've done. And thirdly, to talk about representations of network system and how the classic representation that we use so often may not be always the right one. In the end, it's a plea for a package deal between theory and analysis. The term networked social systems already carries theory with it, I put to you. And this is my take on what counts as a system. There are, of course, plenty of other definitions, and I don't claim primacy here. But I think you need elements in a system, and they need to be interdependent. And I emphasize that word because I see that as I will explain as the core of network theory, interdependence in some form. And also, automatically, this carries with it a multi-level concept. You have elements, and you have a system. So you have outcomes at both the element level and the system level. Social. I think there are many ways of construing social, of course. Thinking about this, I thought, what's important here is that the elements in the system exhibit some form of intentionality towards each other. And I've listed a number of contrasts just to think about the type of intentionality that people can or actors of some form can do to each other, can relate to each other. And so I think that that's a really important element that these are actors in a social system, these elements. And networked is, in many ways, the most controversial of all I put to you, even though for us here, it might seem very obvious. We're not just saying that there are relationships among the elements or among the people or the actors. We're saying that from a networked perspective, the relational ties are the sinews of the system. They're really at the heart of it. They're not abstract. They're not like Kurt Lewin's interdependence of fate, whereby you might be, because you are part of one social category, you might share the same fate as other people in the same social category, whether you know them or not. That means that, in fact, the social network ties I think are, at least in principle, knowable by the individuals in the system. So you know whether you have relationships, you might be wrong, but you can know. The social ties are dyadic. So this is not a group view of the world. And they are the vehicles through which the social actors exhibit that intentionality. Now, lest you think this is obvious, I hasten to add that there are other theoretical conceptions. Those of you who know anything about social psychology will know social identity, self-categorization theory. And these theorists argue that it is the social categories that count. Yes, people have relationships, but those ties are outcomes of other more basic processes. They're not the sinews of the system, as I said before, whereas we claim the opposite. So there's a lot of theory in using the term networked social system. There's a lot of network theory output to you. Also, when we're dealing with political networks of various types, indeed, any sort of networks, I guess, there's also an issue of scale and of level. And this was apparent right from the very beginning of this network discipline. This is the New York Times page 17, April 3, 1933. And it reports Jacob Moreno, who was interviewed, the founder of social network analysis. And you may not be able to read that. I realize it's a bit blurry. If we ever get to the point of charting a whole city or a whole nation, Dr. Moreno added, we would have an intricate maze of psychological reactions, which will present a picture of a vast solar system of intangible structures, powerfully influencing conduct, as gravitation does bodies in space. No wonder he went on to invent psychodrama. Such an invisible structure underlies society. So at the very beginning, Moreno had this idea that we should not just, you know, we should chart a city or a nation, heaven forbid. And that's ambitious. In another life, when I was doing some rather poor criminology at the Australian Embassy in Moscow, we had this as our domain of interest. Now, is it really what is it going to give us to do a network chart in some sense of the whole of the Soviet Union? There's an issue of level here, and maybe it is important. Maybe it is important. And certainly the individual actions at a local level explained a lot of what was going on. But certainly, there were aspects of the Soviet Union that were importantly networked and needed to be understood. And so in a sense, we have this, if you like, two systems at different levels entrained in some sense with one another. I don't plan to explain how we might come to analyze this, although I will give you a few hints about multilevel networks later. Interestingly, in the system, there's a really interesting feedback effect. So the gerontocracy tried to preserve itself and kept the system, the large scale system, running slowly and poorly. But eventually, it had to be replaced by a more radical Politburo, which unleashed social forces that fed back to change the entire structure, entire country. So what we end up with is a kind of multilevel structure where something from the top feeds down to the bottom, which feeds back and changes the top. The other example I like is the People's Action Party of, this is the first PAP government in Singapore. And the thing I particularly like about this is that these men, Lee Kuan Yew, Go Keng Sui, and others, were not big, strong men, as you can see. They did not have AK-47s and rocket launchers in their hands. And yet, they cleaned out the Chinese triads. They did it through a systemic method of feedback across levels whereby they could influence the social structure and then that fed back to outcomes at the system level and the individual level. Now that's a really interesting point, because it's a systemic explanation that's required here, not whether these individuals had certain qualities, like rocket launches and so on. It's a very non-Rambo view of the world, but it's the only explanation that's applicable. And so, it's not just what the individuals have that counts. When we do, and I'm working in a psychology department, which is very individualized, and we do what I would call individualized social science by and large. We search for factors that explain whatever individuals have, the properties they have, the qualities, the issues, the problems. We look for associations among variables or differences among individuals. I think that networks are different things. Here, our search is more for the patterns that explain the structure so that we can look for outcomes. So what counts as a good network? Why is the Singapore leadership structure workable and long lasting, whereas the Soviet structure was not? At least in the last 20 or 30 years. Now, various of our members have, in this community, have had explanations for this. And I see John Padgett up the back, and I think he's got the most profound and interesting of those explanations. But it's a very different thing from individualized social science. This is what we do in standard, what I submit, we often do in standard social science. And there's absolutely nothing wrong with it. It just depends upon your conceptualization of the world. You know, psychology department. We often have two groups of people. We put them in two boxes, so to speak. Psychologists never think of them as nodes of different sizes, but I do. And so, you know, they might have different qualities. They might have rocket launches or IQ or whatever it is that they have. And we can look for differences between them. The statistical analysis to do that depends on independent observations. However, if there are relationships that count in some way among these units, that assumption becomes untenable. And the boxes, in a sense, start to disappear. We've still got colors on the nodes. But the simplistic analysis fails if there are these complex interdependencies. And I started out by saying that complex interdependence is the essence of a networked social system. So if we have a networked view of the world, if we think that networks count and are important, if this is our world rather than the two boxes, then we actually can't just run regressions or analyses of variance, even though they're very convenient things to do because they are relatively simple. So interdependence is not the problem here. It's not a problem at all. It is the substance of what we do. It is the point about networks. Dependence is not something to wish away. It is the content of our work if this is how we construe the world. And structural patterns I put to you are the essence of interdependence. So if we have red people and blue people, that's fine. And the general linear model can cope with that quite well. But once we start putting relationships between them, that in some sense count, we start to undermine the system, that an analytic system. And that's a theoretical point. It's not an analytical point. So for instance, if we theorize that being connected to a red person can make a blue person turn red, which we often do, then we can't assume independence because the heart of the theoretical argument is of dependence. Of course, there's another way of getting to this pattern and that's two red people might form a relationship because they like each other because they're similar to one another. That's a selection hypothesis. These are what I call structural patterns. Here they involve attributes on the nodes and relationships between the nodes. The last example, however, is about the formation of a network tie. So are there other ways in which ties can be formed? And I put to you that there are, and they don't always involve actor attributes. So dependence among network tie variables can lead to patterns in the data, irrespective of actor attributes. There are a couple of simple examples, one that we know full well. A two path like this might be closed. We see triadic closure very commonly in human social networks. So that's a pattern that if we observe frequently implies that ties create the presence, create the circumstances for other ties to come into being. This is an issue of dependence, not among actors as we might think is the easy and simple way of thinking about that issue, but among the ties themselves. Another example is that popularity may be self-creating, a kind of preferential attachment hypothesis. This is a patterning which arises because certain ties, certain patterns of ties create the presence or the preconditions for further ties to come into being. So there are many possibilities here. Those are not the only type of ties. Things that we often look for also reflect dependence. Reciprocation, I mentioned triangulation. Denser regions of the network in terms of multiple triangulation, what Mark Newman referred to as community structure. These higher degree know it's hubs in the network. Preferential attachment type processes. These things can occur irrespective of the actor attributes. And this is a very different world from that individualized social science that I referred to before. So here are some questions, not the only questions that I think are at the heart of network theory. What are the structural patterns and their outcomes? So here's a good question for us all because I don't think anyone really knows the answer. What counts as an effective network in a particular context? I'll talk a little bit about that in a minute. What are the factors that affect structure? Is does the network, do the network ties self-organize? Are there exogenous factors that come into play? The dynamics are important, but there's a lot of work being done on dynamics but I'm not going to talk much about that in this presentation. And then I put this last, but I see this as most fundamental and I'm going to come back to it at the end of the talk. The examples I've shown you so far have been of networks with one type of node and one type of tie. That's a particular network conceptualization. Is that the best representation of a network social system? That these are couched as empirical questions. You could imagine looking, except perhaps for the last one, you could imagine going into some data and looking for that. But I put it to you that, sure, you can do that analysis, but these are theoretical questions. We need theory to guide us as to what to look for in this sort of analysis. I'm going to show you an example of what I think is a clearly network theory. It's actually a network theory of network governance. And I'm going to show you how we might use it to build hypotheses about patterns and how we might analyze it. Now I know there's been a lot of work on network governance in recent years and many of the people here know much more about it than I do. So again, I'm not making great claims but I'm using this as an illustration to show you the kind of way that I think about these particular problems. It's a great pleasure to present this with Steve in the audience because it comes from Jones, Hestely and Bugatti, 1997, where they present what I think is not all theories of network governance are actually, to my mind, network theories of network governance. The network governance literature itself often says that, or sometimes notes that the networks in the network governance are loosely conceptualized, sometimes used even just as metaphor. This, however, is quite clear. It can be quite specific in this theory about what to look for as I'll show you. Jones et al theorized that an effective network governance would have some properties. What they call relational embeddedness, structural embeddedness and macro culture. And we've just written about this. It'll appear shortly in public administration so if people want more details, they can look out for it there. Now I need to go through these three terms. They'll be familiar enough, I think, to most if not all of you. Relational embeddedness is about strong relationship within diets, within pairs of network partners. Now this makes perfectly good sense. You would expect an effective, this is about an effective system of network governance. You would expect that there would be some strong cooperative relationships between some partners. Now how do we actually consider that? Well, one of the important components on Brian Uzi's more recently written extensively about that is reciprocity. So I'm gonna present to you a very simple hypothesis that in a well-functioning governance network, we expect to see the presence of reciprocated network ties. I don't expect every tie to be reciprocated or not that naive, but I expect to see more reciprocated ties than I would otherwise expect to see by chance. Structural embeddedness, according to Jones et al. Related to the extent to which the dyad shared partners as the extent to which those partners themselves were connected. So this is a closure argument. Closure has many interpretations in the network world and has been talked about a lot. It goes back to social capital arguments of Coleman. We've obviously got Granavetta, which I've cited up the top there. Speaking about network embeddedness more generally. Closure, a closure triadic pattern is a kind of embryonic or a kind of archetypal pattern for collaboration and cooperation. But also for scrutiny for the enforcement of norms. You can, if you're connected to everyone else in a little group, you can see them all and you can see whether they're doing what they should do. This is an important element that I think makes perfectly good sense in an effective system of network governance. In this community, more recently, Burrata and Schultz have talked about the importance of closure as a mechanism for handling risk, managing risk in these types of governance systems. So, here's the second hypothesis. In a well-functioning governance and network, we expect to see the presence of triangulated exchanges, triads. Macroculture is, in many ways, the most interesting of the three. And it's talking about how, yes, you need to have some sharedness of goals, agreement about how they should be implemented. That makes perfectly good sense. In a governance system, if it's gonna be effective, you expect there to be some sharedness about what's to be achieved and how it is to be achieved. Now, if I were to measure this fully, I would go to my institutions or informants or respondents and I would ask them about their aims and their values and their goals. But actually, the important point here is the sharedness, the agreement. And so, I put to you then that in a well-functioning governance network, we expect to see fewer negative ties that involve contestation and conflict. So, for this particular hypothesis, I don't need the actor attributes about what sort of culture should we have and what sort of aim should we have. I can simply see whether there's disputation among the ties. So, this is a network theory of network governance because it presents me, my argument why it is so, is it presents me with three hypothesis, each of which are a patterned network presentation configuration that I can investigate. Let me show you a case study of what we did. This is environmental governance of the Swan River in Western Australia. Probably very few of you have been to Perth, Western Australia, it is possibly the most isolated city in the world. I kid you not, it takes two days, three days to drive to another city of similar size. It sits in a small fertile area at the southwest corner of the state and is otherwise surrounded by desert. It's not desert that, it's desert with advantages. It's desert that has tons of iron ore and other minerals that are currently in being dug up at a great rate and shipped off to China. So Perth is like a gold mining town, it's a boom town. It's very wealthy, but without this water catchment, there is no Perth. Per Capita, this is one of the wealthiest cities in the world, all of that money goes through Perth. So the importance of this river, this system to Perth, made the state government invent something a few years ago in the mid 2000 called River Plant, which I'm not gonna go through, but it consists of pages and pages of this sort of stuff which identifies the lead organization and partners and the key actions and all that sort of thing. It is actually a highly, highly top down network governance system. Of course, it doesn't work like that, it's impossible. We studied it by, of course, the advantage of River Plant is that you've got the key players already identified. So we started with 21 key organizations and snowballed out asking about their relationships with other organizations. This is the network, the yellow nodes are the key 21s and I'm going to show you analysis based solely on the yellow node. And these are the important, we did more than this, but these are the important things for this presentation, the important questions. First of all, we've identified the partners that they work with, the other institutions, the other organizations and we ask how important is that working relationship? And we also asked how easy is that working relationship. I don't care how easy it is, I wanna know how difficult it is, but there's a certain kind of face validity to asking easy that doesn't trouble the participants or ethics committees. So what I'm going to show you is an analysis of crucial ties and difficult ties, difficult or extremely difficult ties. And I'm going to interpret the difficult and extremely difficult ties as these ties of contestation, disputation. These are the crucial ties, these are the difficult ties, to some degree they overlap. I'm going to show you, now I've got two types of ties, so I'm going to show you a bivariate exponential random graph model. There's been plenty spoken about Ergums in the last days, so I'm gonna say very little, except that it is for those of you who don't know what it is, it is a statistical model that identifies prominent patterns, network patterns in the network data. And so what are the patterns I'm gonna look for? Well, I've got two types of networks, I'm gonna look for certain patterns that involve only one type of tie at a time. So these are directed networks, I'm looking for density and reciprocity. I'm gonna look for in and out degree parameters, centralization effects, and closure effects. These are standard parameters now for directed networks, but it's a bivariate model, so it's interesting to think about how the two ties might connect with one another. So I'm going to have a parameter for when they line up together, and I'm gonna have a parameter for when they are reciprocated. Now, I'm gonna show you the results for a model, which is what I call a reduced model just for the purposes of presentation. What I've done is I've kept the density and the density and the degree distribution effects in there because they're good for controls, and other than that, I've only kept in parameters that are significant. Just for ease of presentation, I could present the whole model. And this is all the effects with parameters and standard errors and asterisks indicate significance. So I'll go through it bit by bit. We have low density, not surprising, and we have some centralization on the in-degree distribution, but it's also interesting what is not there, what is not significant. There's no reciprocation and there's no closure in this system of network governance. And crucial working ties. These are the difficult ties. Low density, some centralization on popularity. I'm not sure that difficult popularity quite makes sense, but you know what I mean. But no reciprocation and no closure. And then we have the bivariate effects. These are very strong effects. Crucial ties are difficult. That's what entrainment means. They line up together. Exchange means a kind of bivariate reciprocation. You think I'm crucial, I think you're difficult. I think it's fair to say that these institutions occupy a highly contested political domain. And in fact, other analysis that we have in the paper supports that argument as well in more detail. So what can I say then about the Jones et al theory of network governance? Well, in relation to this data, I can say there's no reciprocity effect for either crucial or difficult ties. There's no tendency for relational embeddedness. There's no tendency for structural embeddedness. And crucial ties are difficult, et cetera. So there's disputed macro culture. So Jones et al identified three conditions for effective network governance. None of these conditions hold in this system of network governance. So I've got two choices. Either the theory is wrong, heaven forbid, Steve, or this is not an effective system of network governance. What can I say? Within two years of the study being conducted, the state government had closed down the entire system and changed the legislation and rebuilt something else. One case study does not prove a theory. However, this is the type of work that I think that we need to do across many different instances to see whether these types of situations and effects show up, to see whether this pattern that I've identified here is prominent in all network governance systems or is it just in poorly functioning network governance systems? It's a lot of work potentially. We need to do it context by context. But this is a way that shows that it can be done and that we can have a network theory which presents network hypotheses and can be analyzed in a way which respects the dependencies, in fact builds on the dependencies in the data. So finally, in the third part of my talk, if we're gonna have a network social system, you will have noticed that already I have moved away from the one type of node, one type of tie. The theory of Jones et al. implied that there needed to be two types of ties to study. So this kind of one node, one type of tie is not necessarily the right conceptualization. What is the right conceptualization? Well, that's a theoretical question. What we need to do is to think about what are the important elements that we theoretically see in a network social system. And we need to match our conceptualization with those elements. And there are many things that could be done. Many things that could be included. I'm not saying that we need to include more, but we need to think about such possibilities. And interestingly, all of these things are within methodological reach. Some of them have been quite well established. In fact, within methodological reach is the prospect of putting all of these things together in the one model if we had the data and if we could get the data and if we thought that was a sensible thing to do. So the fact that we can do it doesn't mean that it's sensible. The only answer to that comes from the theory that drives us. I'm not going to go through all of these, but I want to talk a little bit about multi-level because I started out talking about multi-level, although this is a rather different take on it. It's not surprising, is it, that we should think about top-down and bottom-up processes as relevant in these sorts of social systems. And in fact, in some sense, we already do that. We use bipartite network structures. This is, Mark Lubell is going to be talking about this tomorrow. So I invite you to come and hear about another take on network governance based on the San Francisco Bay Area water management. Rather different from the analysis I've just shown you, but still it explicitly takes into account that institutions come together and cooperate within organizations come together and cooperate in larger-scale institutions. That's familiar. That's a bipartite network and there's many things that can be done with that alone. What I'm more interested in for these immediate purposes and the things that we've just started to work on have been something a little bit more complex than that. So Emmanuel Lasega has some interesting data on French cancer researchers. And each of these cancer researchers works within a medical laboratory of some type. The researchers have links among themselves as do the laboratories. Now this is, I put to you, a more general multi-level network where you have two levels, expressly two levels, and you have networks at both levels and networks connecting the levels. Now here the networks connecting the levels are not particularly interesting because it's a basic nesting structure. I'd like to be even more general than this exogenous cross-level structure. What I would like, I'll just do it this way. What I would like to do, what we are working on, is having a two-level system where at one level, which we can call level A, doesn't have to be levels, but it's a nice way of thinking about it, could be other sorts of structures, but this is the basic data structure that we're playing with. We can have network ties amongst A ties, we have another B tie network, and then we have a bipartite network as well. So the entire structure comprises two types of nodes which here I've talked about levels, but they could be other things, they could be divisions or something like that, I'll show you an example of that. And we have three types of ties. We don't suppose that all the ties are the same, they need to be modelled separately because they're involved in different things. Now this I think is really interesting for all sorts of purposes. We're interested in applying this to climate change adaptation where individuals might work together at the one level, but also be associated with various institutions or political parties or whatever at another level which might also have connections of some sort amongst themselves. The possibilities are really interesting. A structure like this actually enables, this is a very big and pompous statement so you can correct me if I'm wrong, but a structure like this actually enables us to start thinking about macro, micro and meso arguments in a fully empirical way. You can actually devise an ergonom for those, for that particular data structure with meso effects, that's the bipartite effects, and within level and meso interactions which I think are particularly interesting because that's the kind of cross level macro micro thing that's of interest. Let me show you some possible effects. The simplest cross level effects involve these simple star like patterns at different levels, and here, oh I've got a pointer haven't I, oh great. Wonderful, the talk's nearly finished, he finally discovers the technology. Here is, so in this point, this node, if this were a positive effect, nodes that are popular within one level actually have lots of connections across levels. Very simple, very simple. We could have these kinds of complicated closure effects, we could have cross level triangulation, shared affiliation if you like, and this one is really, really nice I think to my way of thinking, very simple again, but if there's a collaboration, sorry, if there are ties across level then perhaps there's collaboration within levels. It is cross level entrainment. This really is a kind of macro micro, or at least my take on a kind of macro micro effect if you take one level as micro and another as macro. I know some of you will have objections to that and I'm happy with that. Let me show you an example of a simulation, that particular point here. I'm gonna show some very simple, the only effects in the model, very simple. These two cross level parameters, one is positive so the blue nodes if they're popular at one level are popular across level and the other thing is reverse, where it's negative. So if the red nodes are popular at one level, they're not popular across level. So if all I look at are the red nodes, the A network, I get a network that looks like this and I could go off and I could do a, study that in whatever way I do and I could say, oh, this looks like there's a bit of triangulation and so on. And I could look at the degree distribution and if that's all I did, I would miss out on something really important, the fact that the other network is very highly structured and it's got these two hubs. It's very centralized, not surprising when you think about the parameters and that the bipartite network is also extremely hub-like so the overall network looks like that. So my point here is quite simple. If the world really is like a multi-level network and we go and look at one network and not the other, we might be very wrong in our conclusions about what's driving the structure here. How do I know if the multi-level network is the appropriate picture? Well, that's a theoretical question. Let me show you an empirical example. This is not a multi-level example. This was just data I have to handle. It's a two-division organization. The 150-odd managers and I fitted a model and let's not sit here. Let's just quickly go through it. I'm nearly finished. Let me show you that there are some cross-level effects. There are other effects as well but these are cross-level effects. So I have one of the negative star parameter, cross-level star parameter. I've got shared affiliations and I've got this cross-level entrainment effect. So the point here is that in this empirical data, if I only analyze within the networks and not between the networks, I miss out on some quite important structuring effects. So my conclusions are, we are quickly coming to the point where our network methods can handle some quite highly complex representations of social systems. Where different types of effects can be examined and tested against each other simultaneously. So we really are, I think, methodologically at an important point in time. But I don't think that network theory is well-developed. Partly because it's, maybe I hesitate to say this, but I will nevertheless, partly because it's grown out of this individualized social science. And social theory is not always well-adapted to translation into network terms. So my call is for us to start thinking networks in a network way, not just as an addition to individualized social science to try and prove our R squared or something, and to think about how, what are the best representations of our social system? These are ultimately theoretical questions, although they can be tested empirically once they've been developed. Yes, it's complex. What we need to do is to understand the level of complexity necessary to explain our social system. This is something we have no idea about, I suspect. But because we can now put many effects in models simultaneously, we can always push the boundaries out a bit further until we reach the point where we don't need, we know we don't need the extra level to explain something. So I guess my call here is to think about what's the level of complexity that we think is necessary, and then add a bit more to see whether that adds something to the explanation. Thank you. Thank you very much. John. Policy piece here. And it just points out, it seems to me, the challenge that the next decade has for people in the policy studies to try to get at these relational type of things, taking your study as showing these certain structural factors that are there, and that according to one theory, at least are ill-advised in general, I would say that there's really two types of issues that we really need to look at and need to be able to translate into these network structures. One is, how do we get to that mess? That is, what are the processes? I would say one of the institutional structures and the individual utility functions and the individual characteristics that drive a system to that kind of point, and that's certainly the kind of thing that we can study fairly well, and I think that may be easier than the other challenge, which is to ask, under what circumstances would that kind of a system in fact be an optimal system? Because clearly there's not going to be a single governance structure that's going to work on them, so again, thinking of the macro and the micro level as contributing under, what are the factors on both of those levels for which certain kinds of structures work well and lead to pleasing outcomes and which ones do not? So I see this as really having, as a challenge, thinking one of how to actually interpret these funny little things you guys have already come up with, and secondly, come up with our own perhaps that may actually encompass some factors that have not been, and I suspect that's the challenge that you've been putting forth for us in any case. I think you put it very eloquently, more eloquently than me, certainly more succinctly. Yeah, I think on the first point, I think, of course, what we need to do is to, what we need to do is to think about process and then either model the process directly, that is dynamic, because sometimes it's hard to get the data, or to think about what the implications for that would be in terms of cross-sectional data, if that's what we require. And I think we can do those things for better or for worse. Maybe not as well as we might be able to in five or ten years' time, but we can start to do it. The second point is a big challenge. I think this is one of the big challenges for the network theory in whatever domain. What counts as a good network, whatever we mean by a good and effective network, or an ineffective network? How do we know, how do we tell? We can't just suppose we get something's network there or it's, ooh, the Soviet example shows that that is not necessary or so. So what are the conditions? Now, of course, conditions will change depending on context. But if we are to have theories and results, we need some level of generalizability. So we have to be able to search for a level that enables us to generalize, even if it's to say, we have this type of context that looks like that and we have this color card context that looks like this. We have to find a way to do it. This is the scientific challenge I think that maybe faces in the next ten years' time. And you're absolutely right. Look, it's easy for me to sit there and invent data structures and do little simulation studies and show them off in talks like this. But we need to name a little theory that is network theory. I think that's exactly the case. So that involves university. Hi, Gary. So, nice talk. If I understood you correctly, one of the points you were making is that you get imprints wrong. You can't take into account multiplexity. When you made that point that this one network, and we don't think about the second network and how it interacts with the first network, that we can end up having a flawed inference, I was expecting you to provide an example. So I'm wondering if you could provide an example of an inference that through simulated data or real data that would get wrong by artificially drawing the boundary around that first network. And could you characterize the conditions that we should be most worried about excluding that second or third or fourth level network? Ah, yes, that's a really good question. Look, I... Yeah, well, what about the other things you've done? And with this not a little work that you've only just started on, I can't answer that right now. But in a certain other context, I know, for instance, if you leave out, there are certain other models. If you don't include certain network structures in the model, then you draw the model inference. Any inference about attribute inference, I think these things are reasonably well known. And they're not just common to network science. They're common. Generally, if you don't put in a factor in standard regression, then you may well draw the model inference. It's always a problem about how to decide what's the right way to proceed. So, I think I think we can find examples. The question of when or where we should do it and set up the boundary conditions. I guess we can do that methodologically, but I guess what I'm saying is that I'd like to see that theoretically thinking of that. So, you know, I want someone to tell me, no, you don't need to worry about that. That doesn't make any sense theoretically. Because we've got a plethora of possibilities and I think we need to think about how to constrain that empirically. Now, of course, once you do that you can then look at the things empirically and you might decide that you do need something more when you actually go out to the empirical world. So, there's a kind of trade-off here between analysis and theory and empirical work and so on. It's a package deal. So, basically I think an early model of method of analysis is to try to explain method of structure with a function of nodal attribute and then the method theory. But sometimes I'm pretty confused sometimes or confused because there is a kind of some indogeneity issue. So method of structure can be a function of nodal attribute and the method of attribute. But also this kind of nodal attribute can be influenced by the structure. So how do we deal with this? Well, I think the most compelling way to deal with that is to have more of a tuned balance and to use a stochastic recto-oriented model or some other sort of similar model. Because that enables you to pull apart the influence and selection of cross-sectional data. I think that in some circumstances it might be possible to do this with cross-sectional data. With cross-sectional data we face the same problems in the network world as we do in standard social science. You can't defer causality very compellingly from cross-sectional data unless you have a very strong research design that enables you to control certain elements of the survey data you took and we don't have that. So there are ways of dealing with whether it's selection or influence. But I think the data demands are high because you're asking more. What we can say with an order is that there's some form of association and that there are more in terms of doing the same thing. It's just a matter of being careful about our causality claims. Please join me in thanking Gary Robert. I'll give this great conversation in the great hall.