Zombie Idea: Creditanstalt Did Not Cause the Depression

Brad DeLong and Barry Eichengreen have written the preface to a new edition of Kindleberger's The World in Depression. It is a very good introduction, except for this part:

Kindleberger’s second key lesson, closely related, is the power of contagion. At the centre of The World in Depression is the 1931 financial crisis, arguably the event that turned an already serious recession into the most severe downturn and economic catastrophe of the 20th century. The 1931 crisis began, as Kindleberger observes, in a relatively minor European financial centre, Vienna, but when left untreated leapfrogged first to Berlin and then, with even graver consequences, to London and New York. This is the 20th century’s most dramatic reminder of quickly how financial crises can metastasise almost instantaneously. In 1931 they spread through a number of different channels. German banks held deposits in Vienna. Merchant banks in London had extended credits to German banks and firms to help finance the country’s foreign trade. In addition to financial links, there were psychological links: as soon as a big bank went down in Vienna, investors, having no way to know for sure, began to fear that similar problems might be lurking in the banking systems of other European countries and the US.

In the same way that problems in a small country, Greece, could threaten the entire European System in 2012, problems in a small country, Austria, could constitute a lethal threat to the entire global financial system in 1931 in the absence of effective action to prevent them from spreading.

I've covered this before, so rather than restate it all I'll just point you to that and mention the gist here. Regarding the first paragraph, Creditanstalt was the largest and most well-connected bank in the Austro-Hungarian empire. Following World War I, it remained one of the most important banks in continental Europe. It was not "relatively minor". More importantly, the Depression was already underway before the Viennese institution went under. The New York Bank of the United States had collapsed several months before along with more than 600 other American institutions. It is just not the case that everything was fine right up until Creditanstalt went under. It is much more likely that the Depression caused the collapse of the Austrian bank and not the other way around.

Why is this important? Because contagion cannot emerge from anywhere. So the ramifications for the present day are not that Greece could destroy the entire European system, as Thomas and I wrote last year in Foreign Policy. The underlying research which motivated that article has now been released in Perspectives on Politics. We were right then, and the same intuition helped us to understand why the Cyprus meltdown was going to remain localized while others were talking about how it could drag down the entire global economy.

This matters because very smart people keep saying that contagion can emerge from anywhere at any time. At the recent International Studies Association annual meeting I heard one of the most prominent scholars in IPE say to a large audience that financial contagion worked like it did in the movie Contagion: anyone can become infected at any time. This was based on no research, just an intuition. Here are some other recent examples (1, 2).

But the intuition is false. This is not how the world works. The fact that the claim keeps being made is evidence that we in the social sciences really do not grasp dynamic complexity well at all. This clearly has major consequences not only for how we view the world, but how we govern it. We need to do better.

All Networks Are Not Equal (and Financial Crises Are Not Viruses)

Many thanks to Henry Farrell for discussing some research co-written by a decent chunk of this blog's contributors, which was just released (and is currently ungated, thanks!) by Perspectives on Politics as part of an issue on inequality and the global financial crisis. It's been kicked around the internet a bit already, and already I've come across a major misinterpretation* of the central argument from Mark Thoma:

Are highly interconnected networks better at dispersing risk? It depends upon the type of risk. Suppose a toxin hits a network. If diluting the toxin across the network also dilutes its effects to practically nothing, then we want the network to be as large and interconnected as possible. When shocks hit they will be quickly diluted and rendered relatively harmless. But for toxins that are deadly in minute doses, toxins that kill whatever they touch even when they are highly diluted, we want the infected node on the network to be isolated as much as possible.

This is, we believe, the dominant view of financial contagion in the social sciences and in particular in international economics. In the paper we cite several different formulations of this view in the academic and policy literatures. If we reiterated this view it would not be noteworthy, and it probably would not be publishable. We think our article is noteworthy (and was published) because it argues that this conceptualization of risk in networks is fundamentally misguided: it places undue focus on the strength of the shock and the density of the network, rather than the location of the shock and the topology of the network.

To see the difference consider two shocks of equal strength which hit two networks of equal density. The only difference in the two networks is in the distribution of that density: in one of the networks the connections are distributed more or less equally -- most nodes in the network have about the same number of connections to other nodes -- but in the other network the connections are distributed very unequally -- a few nodes have a lot of connections, while most nodes have few.

We believe that we should expect very different outcomes from the same shock and the same overall density because of different distributions of connections. All networks are not equal. Outcomes do not just depend on the strength of the toxin, but whom it contaminates.

We show empirically that different crises have have different impacts on the global system: crises originating in the US have adverse consequences throughout the entire network, while crises that hit other places do not. We show empirically that the global financial network is highly unequal: it is centered around the US (as Farrell notes in the bit Thoma quotes). And we argue that it is this variation in the distribution of connections, which we call "topology",  which made the subprime crisis so severe from a global (i.e. "systemic") perspective. Or, as Farrell put it in his useful discussion:

Oatley et al. argue that you get two kinds of financial crisis in this kind of world. First, you get financial crises in the periphery, which tend to be limited to a particular region because few other countries are directly exposed to the countries undergoing crisis, and to fizzle out. Here, US dominance serves as a dampener – since it is large enough to absorb shocks itself, it can prevent financial contagion from spreading. In contrast, when a crisis occurs within the US, it tends to spread everywhere, since every other country is heavily linked to the US. When US mortgage markets sneeze, everyone catches cold.

I'd say that when the US sneezes everyone catches pneumonia. So in our view the question isn't whether the toxin is "diluted"*. Nor is it whether a denser network might be more or less capable of absorbing a shock. In our view the performance of the system in the face of a shock depends on the structural properties of the system, such as its topology, and the location of the shock within that structure: if it hits the periphery, the impact is narrow and remains in the periphery; if it hits the core, the impact is broad and emanates throughout the system.

This may seem obvious. We believe it is obvious, after you've read the article. Before you've read it (as Thoma obviously has not) you may end up writing things like this (as Thoma has):

We have been told that problems in places like Cyprus have been walled off -- nodes in the network have been isolated -- but so long as a few isolated connections still exist that are difficult to cut, highly toxic shocks can pollute the rest of the network. In addition, as we saw today when "Jeroen Dijsselbloem, the current head of the Eurogroup, held a formal, on-the-record joint interview with Reuters and the FT today, saying that the messy and chaotic Cyprus solution is a model for future bailouts" and financial markets reacted negatively (the statement is being walked back), some connections -- those involving expectations -- cannot be severed in any case.

Highly interconnected networks are highly desirable so long as (1) we can quickly identify trouble, and (2) nodes can be quickly and effectively isolated. But when those conditions are not present, the occasional highly toxic shock will cause quite a bit of damage.

Despite being the conventional view (here's another example, also from yesterday) we think this is totally wrong. We think that the ongoing collapse in Cyprus is unlikely to have a major effect on the global economy, just as the collapses in Iceland and Ireland did not have a major effect on the global economy: the effects were devastating for those economies, and had some impact on the few countries which were strongly tied to them (mostly regional partners), but did not advance outside of that. Indeed, as the eurozone crisis has deepened over the past few years, the world economy has gone from recession to growth and global financial markets have posted strong gains (esp in the West; less in the "Rest").

We don't think this is a coincidence. We don't think we have just gotten lucky. We don't think that we were saved by wise and prudential crisis management (does anyone?). We think crises in peripheral nodes are very unlikely to spread to the core of the system because of the structural properties of complex networks. We think, in other words, that the global financial network is not some abstract quantity, but something that can be modeled and understood.

Thoma says that markets "reacted negatively" yesterday. The S&P was off 0.33% yesterday -- a totally normal fluctuation -- after increasing by 4.3% over the past month, during which time the botched Italian election and worsening situation in Cyprus were supposed to send financial markets into turmoil. As I write this, European markets are up today.

So financial turmoil from Cyprus hasn't happened yet, just as it didn't happen last summer when the Greek crisis flared up. We wrote about that too, and said the same thing then as we're saying now, just as many economists and pundits argued then that we may be on the brink of doom just as they are now. We were right then and we're right now. Greece has defaulted at least twice since last May, yet the global economy hardly even notices. The Cypriot financial sector has essentially disappeared overnight (from a network perspective the links connecting that node have been effectively severed) and financial markets are up.

I'm being pedantic about this because the message doesn't seem to get across. The belief that a crisis anywhere can lead to a crisis everywhere is so ingrained that very intelligent people don't even recognize contradictory arguments with supporting evidence when they are quite literally staring them directly in the face, as Farrell's precis of our research (and link to the article) was peering through the monitor right into Thoma's cornea.

So I'm afraid I might have to be boring on this point until folks start internalizing it: all networks are not equal.

*Thoma doesn't explicitly assign this view to us, but he quotes part of Farrell's summary of our article -- which says something different from what Thoma says -- and then goes on to the bit I quite as if we were making the a similar case.

**In fact, we refrain from discussing viral contagion at all, as we believe it is not an appropriate analogy for financial contagion. We spent a bit of time discussing this in a previous draft, but as it was somewhat tangential to our main argument we eliminated it in the final version for reasons of space. The general point is that a virus can infect anyone it comes into contact with regardless of who that person is: a king is no less vulnerable than a peasant. Our argument is that not all financial crises are capable of infecting all nodes with which they come into contact. Most crises are not contagious at all, in fact. Our theory provides an explanation for that. But the virus language was referenced by both Farrell and Thoma, so I'll run with it for the purposes of this post.

The Reductionist Gamble

The Reductionist Gamble (RG) appeared in IO two years ago this spring. It has been met with some puzzlement and it has been misunderstood. I can understand both reactions, as the paper asks people to think differently about the world, and yet it does so by using terms and concepts in ways that depart from more typical usage. I say reductionism, and people hear Waltz. I say system, people here system level. That's fine, to a point. Yet, when people begin to offer solutions to the problems the RG identifies as criticisms of the RG itself, it becomes apparent that some clarification of the argument may be not only useful but essential. I offer the following in that spirit. I apologize for the length of the post; the paper was 13,000 words.

The RG begins with a concern about the theoretical orientation of IPE and then highlights the consequences of this theoretical orientation for empirical work. Here, I reverse the logic and highlight the empirical manifestation of the problem using statistical theory and then work back to the theoretical problem from there.

Assume a system that is constituted by n units over which we collect data. Assume further that two distinct data generating processes (“DGPs”) coexist in this system. DGP ‘A’ generates observations for each unit that are independent of one another. DGP ‘B’ generates observations for each unit that are not independent of one another. The relative frequency of A and B in the system is unknown.

Assume that the only statistical tool employed to analyze observations collected on the units generates unbiased estimates of the effect of x on y if and only if the observations are generated by DGP ‘A.’ Statistical analysis of the data collected on this system will thus generate biased estimates at a rate proportional to the relative frequency of DGP ‘B’ in the system.

Within this framework, the problem RG claims to identify, therefore, is that OEP scholarship published in IO and APSR between 1996 and 2006 (and thus at the center of mainstream American IPE for the last 15-20 years) has relied upon statistical techniques (the general linear model implemented in a TSCS framework) that generate unbiased estimates if the observations for each country are independent of one another. In fact, however, the observations often are not independent of one another. Hence, the estimates these articles report are biased at a rate proportional to the relative frequency of DGP ‘B’ in the contemporary global economy. We do not know the frequency of DGP ‘B’ but  suspect that it increases with the density of interaction among the units--a density which we call interdependence. 

The RG develops this argument as problem of theory rather than as a statistical problem. Read my discussion of Herbert Simon (RG pages 317-19) and then map that discussion back on to the discussion here of DGPs. Here is the concise version. Simon suggests that the denser the web of interactions that link units to one another in a complex system, the more the outcomes in each unit are driven by systemic mechanisms and the less they are driven by unit-specific mechanisms. Hence, as the density of cross-unit relationships increases, the ability to theorize about the units independent of the system diminishes. It seems evident that Simon is a theoretical analog to the two DGPs highlighted above, though he conceives of the problem in continuous rather than discrete terms.

The “complex” characterization in Simon's conception of systems is an important one; in a complex system “more is different.” What this means is that one cannot understand how a complex system works by studying the units that comprise it, and one cannot understand the units by examining them (singly or in large groups) as if they are independent of one another or the system. A complex system is irreducible; one must study it as a system. Complex systems are thus characterized by DGP 'B' and therefore must be modeled empirically as systems.

The RG focuses on the theory problem rather than the statistical manifestation of the problem for a number of reasons.

• The problem is not a statistical problem because statistical solutions to (some of) the challenges posed by dependence among observations exist (or might be fashioned). We have spatial regression, ERGMS and latent space techniques. Consequently, the problem isn’t necessarily that we lack statistical solutions (though problems do exist for which we currently lack solutions); the problem is that until quite recently, we rarely implement them (or even test whether we need to). These techniques are now being applied with greater frequency (see, e.g., Cao for a recent application; see Nexon on Zeev Maoz's Kahler's network book). I note some of these statistical solutions in the conclusion (RG, 334-5).
• The problem is a theory problem because in order to implement the appropriate statistical solutions, one needs theory that recognizes the potential importance of DGP 'B'.
• OEP as a theoretical enterprise doesn't encourage us to recognize the potential importance of DGP 'B'. Consider David Lake’s characterization of OEP. "OEP begins with individuals, sectors, or factors of production as the units of analysis and derives their interests over economic policy from each unit's position within the international economy. It conceives of domestic political institutions as mechanisms that aggregate interests (with more or less bias) and structure the bargaining of competing societal groups. Finally, it introduces, when necessary, bargaining between states with different interests. Analysis within OEP proceeds from the most micro- to the most macro-level in a linear and orderly fashion, reflecting an implicit uni-directional conception of politics as flowing up from individuals to interstate bargaining"(Lake 2009, 225).
• The core assumption here is that every sub-system can be examined in isolation from the rest (this is what I mean by methodological "reductionism," by the way). The assumption applies to vertical disaggregation (we don’t need to factor in domestic institutional structure to understand individual preferences) and horizontal disaggregation (we don’t need to factor in bargaining between states to understand domestic aggregation and policy outcome).
• In Simon’s terms, OEP assumes the international system is nearly decomposable. In the language of statistical theory, OEP assumes the system contains a single DGP: DGP 'A'. In practice, this assumption has led to the modal empirical research design in which observations are assumed to be independent of one another. The implication is that this approach generates biased estimates in proportion to the relative frequency of DGP 'B' in the international economy.
• If the prevailing theory held that the global political economy was a complex system in which developments in one sub-system were dependent upon developments in other sub-systems, vertically and horizontally, then we would be more inclined to design research that incorporated those relationships and thus less likely to implement statistical techniques that assume observations are independent.
• In short, it is a theory problem not a statistical problem because our theory drives our choice of statistical tool. We need theory that encourages us to think about the potential presence of DGP 'B'. We don't presently have such theory.

So what is the theory? I don’t have fully developed answers to this question. But here is where I am.

• Theory could usefully move beyond classical IR thinking that remains wedded to the ontology of Newtonian mechanics.
• IPE could make this move by drawing a bit more from complexity science than it has done to date (yes, I know, Axelrod and Cederman, but also, I think one could count the number of IPE articles on the international financial system that cite Sornette or Gabaix on less than one hand).
• One way to begin moving down this path is to draw on the science of complex networks as potentially useful theoretical models of the international system.
• We have a paper forthcoming in Perspectives on Politics that applies this approach to the international financial system. The paper explores how the network topography of international financial relationships shapes the stability of the global financial system (the global spread of local crises; the stability of the topography in the face of changes in the underlying distribution of economic “power.”
• Some of Kindred's dissertation moves in this direction by applying network models to the global financial system, but I will let him write about that when he is ready to do so.

In short, RG constitutes the first step in a call to theorize in terms of (complex) systems, because OEP's theoretical orientation encourages an approach to empirical research that is likely to produce biased estimates systematically. And while it is true that statistical solutions to these problems exist, we need theories that encourage us to employ them. Multiple critiques of the argument exist. The assertion "but, we have spatial regression" isn't one of them.

Thanks for listening.