STOCHASTIC PROCESSES AND PUNCTUATED CHANGE

The methods we use to study the general properties of policy change are stochastic process approaches. In stochastic process models, randomness is fundamental to the model rather than assuming the position of an “error term” in a model that otherwise would exactly predict particular outcomes. Stochastic process models try to ascertain what kinds of probability distributions could have accounted for an observed frequency distribution of outcomes. They are most useful when the exact specification of a model is not possible, either because of the essential complexity and uncertainty or because of the current state of scientific knowledge. While political scientists have tended to gravitate more to point prediction models and regression studies, stochastic process approaches have become increasingly popular in the natural and biological sciences. A major reason is the recognition of the key importance of extreme values in the determination of the time paths of complex systems (Sornette 2000). The more traditional regression studies often relegate these critical values to the error structure of “unexplained variance.”

There is stochastic process work in political science, however. For the purposes of this paper, the most important studies are the budgetary studies by John Padgett (1980, 1981). Padgett derives and then estimates empirically the probability distributions associated with boundedly rational budgetary decision-making, showing that traditional incremental approaches will yield a normal distribution but that a sequential search model in which actors serially search for options (such as bigger and bigger budget cuts) until one satisfies the constraints they face will yield a double or two-sided Paretian distribution (because decision-makers can enjoy growth as well as suffer cuts). Also called the power distribution, it is characterized by a slender peak and heavy tails; it is the most famous of the punctuated distributions in science (Bak 1997; Mandelbrot 1997, 1999).

An idea of how frequency distributions can be used to study punctuated change is illustrated in Figure 1. There we plot the frequency distribution of inflation-adjusted annual percentage changes in U.S. Congressional Budget Authority from FY 1947 through FY 2000, pooled across Office of Management and Budget (OMB) subfunctions (Jones, Baumgartner, and True 1996).³

The clear pattern of stability coexisting with punctuations is striking. We have noted the Social Security expenditure changes discussed above in Figure 1 as an illustration of the pattern. The empirical distribution exhibits great stability around the mean; most changes are very small indeed. But the distribution also exhibits the signature “heavy tails” pattern: There are a considerable number of cases that are very far from the mean, indicating very large annual changes. As might be expected, there are more large budget increases than decreases—note the extraordinarily large positive tail. Most of the quantitative changes in the time series underlying the data come not from the incremental changes but from the extreme values.⁴

Frequency distributions with this general shape are characterized as having positive kurtosis. Kurtosis is defined statistically as the fourth moment around the mean; the variance and skew are the second and third moments. A unit-free measure of kurtosis somewhat analogous to the coefficient of variation measure of variance (the standard deviation divided by the mean) was developed by Karl Pearson; it is the raw fourth moment divided by the square of the variance (Anscombe and Glynn 1983).⁵

This standardized kurtosis has a value of three for the normal distribution; distributions with positive kurtosis (termed leptokurtosis) tend to have slender peaks and heavy tails as in Figure 1. Focusing on kurtosis offers a way of directly comparing the distributions we study, but in practice it encounters some difficulties in interpretation and application. As a consequence, we first examine kurtosis tests and then move to more accurate direct parameter estimates of the underlying probability distributions that are hypothesized to underlie the frequency distributions we study.⁶

COSTS IN POLITICAL INSTITUTIONS

An institution may be defined as a set of individuals acting according to common rules resulting in collective outcomes. Institutional rules are not neutral, in the sense that different rules often lead to different outcomes (Jackson 1990, 2). These aggregations of individuals interacting according to rules react to information from the environment and come to a collective response (even if the collective response is simply the sum of individual actions, as it is for markets and elections and roll-call voting in legislatures).

Decision-making systems impose four kinds of costs in making decisions in response to a changing environment: decision costs, transaction costs, information costs, and cognitive costs. Decision costs are costs that actors trying to come to agreement incur. They include bargaining costs and institutionally imposed costs, such as those built into a separation of powers governing arrangement (Bish 1973; Buchanan and Tullock 1962). Transaction costs are costs that parties incur after they come to agreement (North 1990). In market transactions, these involve such items as the cost of ensuring compliance to contractual agreements and other payments to third parties to complete the transaction. It ought to be clear that in advanced democracies decision costs in the policymaking process heavily outweigh transaction costs. The costs of bringing relevant parties to agreement in a system of separated powers (decision costs) generally outweigh the costs of holding hearings, enacting statutes, or changing budgetary allocations once agreement has been reached (transaction costs). In any case, we combine these costs in our analysis, terming them together “decision costs.”

Information costs are search costs—costs of obtaining information relevant to making a decision. These are costs that exist when a person (or an organization) wants to make a decision. Cognitive costs are costs associated with the limited processing capacity of any social institution comprised of human beings. These are costs that occur because people do not know they need to make a decision. If one is not attending to a key component of the environment, then he or she cannot decide whether or not to incur search or information costs.

We posit here that these costs act similarly on the capacity of political institutions to process the flow of information. Institutional costs in politics may approximate the manner in which friction operates in physical models. When friction is introduced into idealized physical models, nonlinear systems result (Bak 1997). Such open systems result in an output pattern that is episodic and punctuated, with extraordinary difficulty in making point predictions. Earthquakes are an example. Predicting a particular earthquake is not currently possible, but the patterning of earthquakes follows a lawlike pattern—the power function of the Guttenberg–Richter law (Bak 1997; Rundle, Turcotte, and Klein 1996; Schroeder 1991; Sornette 2000; West and Deering 1995).

AGENDA-SETTING EFFECTS

A hypothetical fully efficient decision-making institution that imposed no costs would respond seamlessly to the world around it. That is, it would incorporate all relevant aspects of the information it encountered and would “use up” all the information in its decision-making process. The outputs of such a system would perfectly reflect the information flows coming from its environment (Simon 1996). The major example of such a cost-free system is the classical model of a competitive economy.

In such a pure system,

where

• R = response = ΔO = change in output,
• S = information (signal),
• β = benefits derived from the information flow (<1).

The system reacts directly to the input flow by changing its output. If costs are assumed to act linearly on the system, then

Our hypothetical system continues to respond directly to the input flow. Now, however, it will not act until it recovers the costs that must be invested in reacting to the flow of information.

In idealized frictionless systems, the output (decision) result is entirely a function of the information received. This is a standard against which other processes may be compared. In such a system, if the information flow were distributed in a normal or Gaussian fashion, then the output flow would be normal, as the normal distribution is invariant under a linear transformation. There is good reason to expect the Gaussian as an input distribution. If participants in an institution receive information from independent diverse streams and weight and sum these diverse streams in an index, then the resulting distribution would be normal by the central limit theorem, at least in the limit. This is an important standard, because in most situations we cannot observe the flow of information into a policymaking system with any precision. Without a complete model of the policy system, something we lack at present, there is no way of knowing which aspects of a complex situation will be judged relevant (of benefit).⁷

The input flow, however, may not be normal because whatever processes generate that flow do not approximate the central limit theorem.⁸

Let us now turn to less-than-perfect human systems. If individual decision-makers rely on a limited set of indicators to monitor their environments, and update them or include newly salient aspects of the environment in the decision-making calculus episodically, the result will be a flow of “news” (that is, the information flow to which the decision-maker attends) that is not normal (Jones 2001, chap. 7). This kind of decision-making underlies agenda-setting studies. In effect, the decision-maker locks choice into a set of facts based in the past and must update in a punctuated manner in the face of change that cannot be ignored. The “news” is leptokurtic. If the news is leptokurtic, outputs in completely efficient institutions will be leptokurtic. Since collective outputs from markets and elections are simple aggregates of individual decisions, outputs will follow news flows.

Institutional decision costs will add to the kurtosis of output distributions. Difficulty in changing the status quo results in incremental decision-making rather than reform. This shows up in output distributions as the tall central peak associated with leptokurtic distributions. When change occurs, it requires substantial mobilization to overcome the stasis associated with the workings of political institutions and the tendency of humans to adopt rules of action that are difficult to change. As a consequence, when change occurs, it tends to be relatively extreme. This results in the characteristic “heavy tails” and “weak shoulders” of leptokurtic distributions. This leads to our first key hypothesis.

If our reasoning is correct, we expect always to find positive kurtosis in the outputs of human institutions—elections, lawmaking, budgeting, media coverage, scheduled hearings on bills, etc.

INSTITUTIONAL FRICTION

Some institutions impose higher decision and transaction costs than others. We hypothesize that the higher the institutional costs imposed, the more leptokurtic the output distribution.

There would exist considerable friction in policymaking systems even if they were informationally efficient, because American political institutions are not designed to be fully responsive. Existing formal studies of institutions are equilibrium based, and hence static, but they clearly imply that continuous preference changes would result in discontinuous policymaking processes (Hammond and Miller 1987; Krehbiel 1998). Generally the costs that have been studied are decision costs; transaction costs are relatively low in the U.S. lawmaking process. Aspects of the policy process, however, vary in their susceptibility to decision and transaction costs. For example, holding a hearing on a policy-relevant issue entails decision costs, but these are far less onerous than actually passing a law.

We can study only a sampling of social processes that might be relevant to policymaking. Broadly speaking, in this paper we examine three classes of processes: the U.S. stock market (because of the long history of study in finance and economics of these processes); U.S. Presidential, Congressional, and Senatorial elections (because of their relevance to political choice); and selected aspects of the U.S. national-level public policymaking process (because they represent collective choice most directly).

Can we rank this diverse set of institutions in order of the costs they impose on collective decision-making? To do so we first have to assume that, in effect, cognitive costs are about the same in each institutional setting. Then we need to think through how each institutional setting would increase decision costs relative to a hypothetically fully efficient system (that is, one that translated informational inputs into institutional outputs directly and seamlessly).

We have already noted that modern stock and bond markets impose low transaction costs (and no decision costs). As a consequence, in the United States at least, they fairly closely approximate the fully efficient institutional mechanism.

Elections are more problematic. In the past, high transaction costs have been imposed on some voters: non-property holders, blacks, women. The 2000 Presidential election, as conducted in Florida, indicates that transaction costs may still be imposed on some groups, even though the systematic exclusion practiced in the past no longer exists. The long periods we use for elections in the United States make the changing nature of the cost structure problematic. In the modern era, however, transaction costs have been relatively low, and theories of electoral choice have emphasized cognitive friction (in the form of partisan identification and, more recently, heuristic reasoning) to the virtual exclusion of institutional costs. In any case, whatever transaction costs have been imposed in the past should be imposed similarly for the three kinds of elections we study, and they should differ among themselves only in cognitive stickiness. We hypothesize that Congressional elections, because of their low visibility, will be more subject to leptokurtosis than either Senatorial or Presidential elections and that Senatorial elections will be more leptokurtic than Presidential ones.

Now we turn to the policymaking distributions. Policy process scholars find it useful to divide the policymaking process into stages such that an issue must access one stage before moving to the next. For example, an issue must access the agenda before it is enacted into law, and it must be enacted before a budgetary allocation (deLeon 1999). The process looks roughly like this:

The distributions we study essentially monitor the process at the agenda-setting, enactment, and resource commitment stages. The existing literature indicates that institutional costs increase as a proposal moves through the policy stages. It clearly is more difficult to get a statute passed than to get a hearing scheduled on a proposal. These institutional difficulties reflect decision costs—the blocks to collective policy action on the parts of the legislative and executive branches. Policy-stage theory allows us to rank these distributions by the stages they assess, under the assumption that decision and transaction costs will be higher as we move through the policy cycle.

In sum, we rank the various social processes we study in terms of institutionally imposed costs (least to most) as follows.

1. Markets. Pure markets have low transaction and decision costs; modern stock and bond markets come very close to this ideal. Information is freely available. (Appendix 2 details a smattering of the long history of the empirical study of financial markets.)
2. Elections. Voting in modern times has low transaction and decision costs, but elections themselves differ in the extent to which they impose information costs. In Presidential elections, information is more available (but can cost cognitive resources in organizing it). Less information is available in most House races
3. Policy processes. Policy-stage theory implies the following ranking, from the least to the most costly.

• a. News coverage. We assess the systemic agenda via general news coverage. Changing from one topic of coverage to another is not cost-free, but it is relatively easy compared to lawmaking.
• b. Congressional hearings. We assess the policy or governmental agenda through the scheduling of hearings. The topic of scrutiny must attract attention (a cognitive cost) and incur some institutional costs (the minority party cannot schedule hearings alone).
• c. Congressional quarterly coverage. The coverage of inside-the-beltway coverage of issues likely mimics the salience of issues on the governmental agenda. The specialized press has no independent agenda-setting power; rather it indexes the lawmaking activity of Congress.
• d. Executive orders. The president alone acts, but executive orders often have major policy impacts (Mayer 2001). Decision costs are high but are concentrated within a single branch of government.
• e. Statutes. Both houses of Congress and the President must cooperate in the lawmaking process; decision costs imposed by the structure are very high.
• f. Budgets. The literature indicates a complex interaction between cognitive and institutionally imposed costs. Institutionally, large changes are subject to the same dynamics as statutes. Cognitively, budget routines and operating procedures dominate the proceeding, and these tend to be highly resistant to change.

Aspects of this ranking may be debatable, but at least it provides a starting point for comparing the outputs of diverse institutional arrangements. In any case, it is clear that political institutions impose decision and transaction costs because of their deliberate design and operation and that this facet of political life can be examined empirically.

Table 1 indicates the nature and sources of the data we employ. We noted above that what we term cognitive costs interact with institutionally imposed costs, so that it is somewhat artificial to try to disentangle them. However, in cases where institutionally imposed costs are very low, leptokurtosis in output distributions very likely indicates that cognitive costs are operative. It is always possible, of course, that the distribution of information is leptokurtic in such circumstances, so in this case we rely on the argument above concerning implicit indicators. That implies that, in the limit at least, information is normally distributed, and hence deviations from normality in such institutions are due mostly to cognitive costs. Elections and asset markets (stock and bond markets) are both characterized by very low decision costs (generally the buyer or voter has control over the decision) and transaction costs.

FINDINGS

All of the distributions we studied are first-difference change scores or percentage changes; our theory applies to changes. For all distributions, we pool the time series change scores across the basic units. (See Appendix 1 for a discussion of measurement and analysis strategies employed here.)

There is no ideal method for comparing distributions regarding their punctuations, so we settled on a three-stage approach. First, we ran Kolmogorov–Smirnov (K-S) tests on all the distributions; then we examined the sample kurtosis for the distributions; finally, we moved to the direct parameter estimate approach.

The K-S test compares the values of a distribution of data to a theoretical probability distribution; significance implies that the observations do not follow the distribution of interest. This test is not particularly powerful; it is fairly easy to reject the null hypothesis. Nevertheless, the null hypothesis that the underlying probability distribution is either normal or lognormal is easily rejected for all the distributions we studied.

Armed with this information, we examined the standardized kurtosis values for each of the distributions we studied. These results are presented in Table 2. Kurtosis has suffered from difficulties in interpretation (Finucan 1963; Moors 1988) and is flawed as an empirical measure (Balanda and MacGillivray 1988; Groeneveld 1998). In effect, it is too sensitive to the extreme values to estimate the full shape of the distribution properly. But it is the best direct comparative measure we have.

Table 2 generally supports our institutional friction hypothesis. First, all distributions suffer from what finance economists term the “excess kurtosis problem” (Mills 1995)—each and every one of the values exceed that expected from the normal distribution. We can test the statistical hypothesis that the observed sample kurtosis exceeds the expected three⁹

We can say more. Let us divide the distributions into three groups: those assessing final output commitments (the budget distributions), those indicating public policy processes, and those independent of policymaking institutions (here termed “input distributions”). We see great differences in the average sample kurtosis—over 72 for the budget distributions and almost 23 for the policy process distributions, but only 6.37 for the input distributions.¹⁰

There are, however, several “out-of-order” distributions within these groups. The most glaring of these is the kurtosis for executive orders, which falls closer to the values characteristic of the input distributions than to the policy distributions. Also noteworthy is that the kurtosis for statutes is less than that for hearings. For the input distributions, elections seem to exhibit less kurtosis than markets. Again, however, these results cannot be viewed as solid without confirming evidence.

We now move to estimate directly the form of the probability distributions underlying the frequency distributions. It is technically not possible to study all types of possible probability distributions by directly estimating parameters, but it is appropriate for two important classes of distributions—the double exponential (or Laplace) and the double Paretian.¹¹

We present below selected frequency distributions and their associated scatterplots used to estimate the type of probability distribution with the highest likelihood of generating the empirical frequency distributions. Paretian and exponential distributions may be estimated by examining the fit of the frequency distribution (frequencies versus category midpoints) in log–log and semilog plots, respectively.¹³

All of the distributions we studied except for government budgets were best approximated by a double-exponential probability distribution. In almost all cases, the fits were exceptionally good, although the right (positive-change) tails for the policy distributions display better fits than the left (negative) tails.

Policy Process

We expect to find more pronounced punctuations in policy commitment distributions than in agenda-setting distributions. Early in the policy process, when proposals struggle to gain agenda access, cognitive costs are high but institutional costs are reasonably low. The scheduling of a policy topic for a hearing is indicative that policymakers are taking the topic seriously—the topic has accessed the policy or governmental agenda, but holding the hearing requires paying only limited decision and transaction costs. Figure 3 shows the frequency distribution for Senate hearings, along with the log–log and semilog plots for the distribution. The hearings data are skewed toward the positive side. The slopes for the left side of the scatterplots have been reversed in sign for purposes of comparison. The fit for the log–log plot displays a distinct downward curvature, indicating that the fit is not good; basically the distribution is less ‘wild’ than the Paretian. The semilog plot fits much better. The slope estimates for the two tails are distinctly of different absolute magnitudes; this indicates that punctuations are more likely in a positive direction than a negative one.

Figure 4 depicts the frequency distribution and the associated scatterplots for lawmaking. The frequency distribution is not as well structured as the other policy distributions, but a right long tail is clearly present, and the right tail displays an excellent fit. The left tail fits less well, a phenomenon that characterizes all of the policy frequency distributions. Moreover, the strong central peak is not as much in evidence in the distribution—it is overshadowed by the extended tail. This is characteristic of exponential distributions as the parameter increases in absolute magnitude. One interpretation of this finding is that when major statutory change occurs, it tends to shift activity from the existing set of policy concerns to new ones.

Table 3 presents the exponential fits for policy data (except government budgets), as well as for the elections data. Because of differences in measurement, the market, elections, and policy data are not directly comparable, but they are comparable within sets. The policy output data can be ranked by using the parameter value—essentially lower values indicate more extended tails.¹⁵

There are difficult statistical issues associated with direct parameter estimates. First, cumulated plots will not yield independent errors, but the error seems to be trivial (Sornette 2000, 141). Second, the number of cases is not properly represented by the number of points plotted. We report “psuedo-standard errors” to indicate that we adjust these for the actual number of cases studied. This is not entirely correct either, because the extreme values are estimated by fewer points than the points closer to the center of the distribution.

As hypothesized, accessing the systemic agenda, assessed by the New York Times coverage of politics and government, displays more moderate change—essentially more flexibility—in moving from policy topic to policy topic than distributions assessing later stages in the policy cycle. Accessing the policy or governmental agenda, assessed by the scheduling of hearings, comes next, with House hearings slightly more punctuated than Senate hearings. This can be due to constitutional facets that require hearings in certain categories—appropriations, for example. In any case, the difference is not large, and both fit the general pattern.

Inside-the-beltway specialized coverage by the CQ displays a higher kurtosis than press coverage or the scheduling of hearings. This may be a surprise at first glance, but it became clear that CQ's coverage is indexed to the lawmaking activities of Congress. As hypothesized, statutes are the most punctuated of all the exponential policy distributions. Lawmaking requires the cooperation of three separate policymaking institutions and, as a consequence, should be exceptionally sticky. Presidential executive orders occupy a position between the agenda-setting variables and the statutes. Presidential attention must be engaged for him to issue an executive order, but he commits the nation to a policy change with his signature (Mayer 2001). While he need not secure the cooperation of other branches of government in the process, he must deal with intrabranch cooperation among affected agencies, the OMB, and the Justice Department.

Public Budgets.

Next we examine resource commitment. Figure 5 presents scatterplots for U.S. Congressional Budget Authority, pooled across OMB subfunctions, for 1947–96 (Figure 1 presents the frequency distribution). We have focused on policy subfunctions, eliminating financial transactions, and have adjusted for inflation. Figure 6 presents the full set of graphs for U.S. Budget Outlays, 1800–1994. Table 4 describes the log–log fits for the two budget series we examined. In each case, the Paretian fit better than the exponential. For U.S. Budget Authority, available only for the modern era, an exponential fit the left, negative change tail better than the Paretian. That is, program curtailment is less punctuated than program growth since the Second World War.

Elections

Elections in general should be characterized by less institutional friction than policymaking processes, and our distributional analysis shows that is indeed the case. Figure 7 presents the results for House elections; similar results were obtained from the analyses of Senate and Presidential elections. As is the case for the policy distributions, the exponential is the proper fit for this distribution. The slopes for each side of the double exponential are similar in absolute value, indicating that change is about as likely in one direction as the other.

Within elections, Presidential contests are the least punctuated, while House contests are the most (see Table 3). House elections are characterized by long periods of very little deviation in the voting patterns for the incumbent, punctuated by occasional large swings. Missing is the moderate adjustment that is expected with the normal. It is of course well-known that legislators may make themselves safer than might be predicted by the policy attitudes of the represented, but our approach offers a method of comparison not available when we examine distributions in isolation. It is likely that information in Presidential elections leads to more frequent updating, and hence more moderate change, than in House elections.

Markets

Markets, in theory, should be the most friction-free of the major political and economic institutions in America. Figure 8 presents the results for U.S. stock market returns as a composite diagram across 100 years of the Dow–Jones Industrial Average. We estimated exponential fits for daily returns for the Dow–Jones Industrial Average for four separate periods and for the entire series, 1896 through 1996. Table 5 presents these results.

Financial economists have studied the kurtosis associated with stock returns (beginning with Mandelbrot 1964). Direct empirical estimates of the distributions have been rarer than time series studies, and occasionally analysts have rushed to conclude that stock markets follow power functions (Peters 1991, 1994). Table 5 shows that stock markets, like much political data, follow an exponential probability distribution. The upward curvature is suggestive of the stretched exponential.

SOME CAVEATS IN STUDYING DYNAMICS

The distributions we have studied order themselves clearly into three groups: the “input distributions” (elections, markets, and news coverage), with relatively low kurtosis and shorter tails; the “policy process” distributions (hearings, CQ coverage, executive orders, and statutes), with higher kurtosis and somewhat more extended tails; and the budget distributions, with very high kurtosis and both relatively high stability and thick tails. Within the policy process distributions, there are relatively more extreme values later in the policy cycle (statutes) compared to earlier (hearings).

The finding of so many exponential distributions among those we studied is intriguing but causes some modification of the institutional friction theory. Technically statisticians categorize the exponential distribution as (relatively) thin-tailed, sometimes using the exponential as the dividing point between thin- and thick-tailed distributions. In addition, in the case of the exponential, peakedness and stretched tails (a better term than “thick tails” for the exponential) are inversely related. Empirically this means that, within the exponential family of distributions, major changes or punctuations must be compensated for by “giving up” some stability or peakedness. That is not the case for the Paretian, where both stability and punctuations can increase at the expense of the shoulders.

As the parameter for the exponential increases, then, the tail becomes more “stretched out” and there are relatively more extreme changes. But this is at the cost of stability. Both stability and punctuations occur relatively less in any exponential distribution than in Pareto distributions, but the exponential parameter is an important governor of the extensiveness of punctuations within the exponential family. As the exponential parameter increases, there are relatively more punctuations and relatively less stability. This seems to be the dynamic characteristic of most policymaking distributions, even statutes. It suggests that policymaking systems following an exponential path are more adaptive to incoming information than systems following a Paretian path.

A second issue relates to the causal dynamics that underpin the distributions we have studied. We have relentlessly pursued the notion that political change is affected by institutional friction, and it is clear that there is considerable evidence for the hypothesis. This cannot be the whole story of punctuated political distributions, however. Most importantly, heavy tails in frequency distributions can be caused by cascades—bandwagons and fads, for example. In such situations, there exists interdependency among the cases. Such “chain reaction” dynamics is most in evidence in the budget data, where clearly program budgets are not completely independent, nor are they dictated solely by macro budgeting concerns (which would induce case dependency). More generally, any process that induces interdependency (such as the familiar monitoring and mimicking behaviors observed in markets and in lobbying) result in “thick-tailed” distributions. We cannot address just how these cascading processes are intertwined with friction-related institutional costs here, but we believe that this ought to be a major avenue of future research in policy processes.

While our direct parameter tests indicate that the double exponential fits both the elections data and the stock market data, the distribution for elections is more centrally peaked than that for stock market data. As a consequence, the punctuations for the stock market are relatively more extreme and the central stability is relatively less rigid. Election swings are bounded, while stock market returns are not, which can account for the difference. But other factors may be at work. Both institutions aggregate individual choices in a straightforward manner, but Tip O'Neil's dictum, “All politics is local,” suggests that elections for the House of Representatives, at least, are less subject to national election trends than stock markets. As markets become more linked, they could be more subject to cascades even as information is more and more widely available. One fruitful line of inquiry would be to examine more thoroughly the “anchoring” effect of local forces in Congressional elections. It stands to reason that election distributions will be more strongly anchored by the incumbency effect, something that does not exist for markets.

The general point here is that it is difficult to isolate the effects of institutional friction, which leads to punctuated outputs, from cascade dynamics, which also leads to punctuated outputs. And of course there are other potential causal processes that could account for the distributions we analyze in this paper. A disadvantage of the stochastic process approach is that it is less amenable to ruling out alternate causal models than the more familiar regression-based structural modeling approach. On the other hand, examining distributions can lead to radically different understandings of causal mechanisms, to fresh tests of causal processes, and to new directions of research.