Bounded Rationality in Budgetary Research Author(s): John F. Padgett Source: The American Political Science Review, Vol. 74, No. 2 (Jun., 1980), pp. 354-372 Published by: American Political Science Association Stable URL: http://www.jstor.org/stable/1960632 Accessed: 25/08/2008 16:28

Your use of the JSTOR archive indicates your acceptance of JSTOR's Terms and Conditions of Use, available at http://www.jstor.org/page/info/about/policies/terms.jsp. JSTOR's Terms and Conditions of Use provides, in part, that unless you have obtained prior permission, you may not download an entire issue of a journal or multiple copies of articles, and you may use content in the JSTOR archive only for your personal, non-commercial use.

Please contact the publisher regarding any further use of this work. Publisher contact information may be obtained at http://www.jstor.org/action/showPublisher?publisherCode=apsa.

Each copy of any part of a JSTOR transmission must contain the same copyright notice that appears on the screen or printed page of such transmission.

JSTOR is a not-for-profit organization founded in 1995 to build trusted digital archives for scholarship. We work with the scholarly community to preserve their work and the materials they rely upon, and to build a common research platform that promotes the discovery and use of these resources. For more information about JSTOR, please contact [email protected].

http://www.jstor.org BoundedRationality in BudgetaryResearch

JOHN F. PADGETT HarvardUniversity

Two bounded rationalitytheories of federal budgetarydecision makingare operationalizedand tested within a stochastic process framework.Empirical analyses of Eisenhower,Kennedy and Johnson domestic budgetdata, compiledfrom internalOffice of Managementand Budgetplanning documents, support the theory of serialjudgment over the theory of incrementalismproposed by Davis, Dempster and Wildavsky.The new theory highlightsboth the structureof ordered search through a limited number of discrete alternativesand the importanceof informaljudgmental evaluations.Serial judgment theory predicts not only that most programsmost of the time will receive allocations which are only marginallydifferent from the historical base, but also that occasional radical and even "catastrophic"changes are the normal result of routine federal budgetary decision making. The methodological limitations of linear regression techniques in explanatorybudgetary research are also discussed.

The lure of the bounded rationality para- and was subsequently developed in a budgetary digm for researchers in organizational decision context by Wildavsky (1964, 1975) and by making has been strong because many of the Fenno (1966). As operationalized by Davis, policy problems such researchers are most Dempster and Wildavskyin 1966, however, the fascinated by are characterized by a degree of theory of incrementalismposited that budget- complexity which exceeds human decision ary decision makers simplify very complex makers' cognitive constriants. Whenever this allocational problems by relying on a few degree of complexity is present, and whenever "decision rules," which can be expressed in selection environments are not so restrictive as terms of linear regressions. Since these incre- to impose unique solutions, the paradigm of mentalist "decision rule" models were first bounded rationality suggests that we can im- proposed, their influence on quantitativebudg- prove our understanding by examining the etary research has been remarkable.Variations decision heuristics or "aids to calculation" on the basic Davis, Dempster and Wildavsky which decision makers use to make sense of regressions have been used to analyze and to their difficult problems. First advanced by replicatebudgetary decision processesin 56 and Simon (1957), the bounded rationality para- then in 116 federal domestic bureausand in the digm encompasses those choice "theories which Congress (Davis et al., 1966, 1971, 1974), in incorporate constraints on the information- the U.S. Department of Defense (Stromberg, processing capacities of the actor" (Simon, 1970), in U.S. state governments(Sharkansky, 1972, p. 162). More particularly, such theories 1968), and in municipal governments in the emphasize limited alternative sets, systematic U.S. (Larkey, 1975), in Britain (Danziger, search routines, and the cognitive difficulty of 1974), and in Norway (Cowert, Hansen and marginal value tradeoffs in the face of "incom- Brofoss, 1975). In all of these applications, mensurable" objectives. regression R2 's have been observed to exceed One prominent application of this paradigm .9. in the area of budgetary research has been the Despite the diffusion and apparentstatistical theory of incrementalism. This theory was first success of these incrementalistmodels, a num- proposed verbally by Lindblom (1961, 1970), ber of researchers have remained skeptical (Natchez and Bupp, 1973; Williamson, 1966; Ripley et al., 1973; Gist, 1974; Wanat, 1974; LeLoup, 1978b). In this same skeptical spirit, I am very gratefulfor the supportand for the many the present article has two purposes. First, the helpful suggestions and criticisms given to me at methodological foundations of quantitativein- various stages of this research by Robert Axelirod, crementalist research are examined and Ronald L. Breiger, Michael D. Cohen, Mimi Fahs, criti- Donald Kacher,Mark Kamlet, HarrisonC. White,and cized. Traditional time series linear regression two anonymous referees. Particularlystrong is my of the incrementalistvariety, I will argue, is a debt to John P. Crecine. This research has been relatively nonfalsifiable approach for distin- supported by NSF grants SOC72-05488, SOC76- guishing among plausible decision process mod- 01052, SOC76-24394, and by HUD grantH-2368G. els. In its stead, I propose the use of distribu- 354 1980 BoundedRationality in BudgetaryResearch 355 tional tests based on cross-sectional program as reliable information indices. On the basis of allocation change data. This alternativetesting such a reliable "base," decision makers were procedureis groundedin the modeling method- hypothesized to rely on simple decision rules of ology of stochastic processes. the form: "Grant to an agency some fixed Secondly, in a more substantive vein, an mean percentageof that agency's base, plus or alternativebounded rationality theory of budg- minus some stochastic adjustment to take etary decision making, to be labeled "the account of special circumstances"(p. 532). The theory of serial judgment," will be developed. cognitive emphasis on fixed mean percentage The perspective to be adopted here is that, change imposed a linear structure to such despite specific problems with the incremental- decision rules. This structural framework was ist models, the paradigmof bounded rationality postulated to be temporallystable, except for a remains a useful approach to the analysis of few discrete "shift points" in which parameters budgetary decision making. Bounded rationali- and/or functional forms changed due to pres- ty has spawned a rich assortmentof alternative sure from a variety of exogenous events. Per- decision theories,l all of which are consistent centage parameters,in turn, reflect past learn- with behavioralconstraints on information pro- ing about the strategic biases inherent in the cessing, alternative sets, value integration, and roles of advocate and guardian. the ability to predict consequences. And serial Within this linear decision rule framework, judgment is another such theory, which differs Davis et al. (1966, p. 537) operationalized a from incrementalismin part because of a more variety of simple regressionequations, each of explicit and contextually dependent search which embodied a different approachto strate- process, and in part because of the more gic "gaming."But in 87 percent of the bureau ambiguous character of its final choice selec- cases, the empirically dominant incrementalist tion. model of executive request behaviorwas Both the theory of incrementalismand the theory of serialjudgment are operationalizedas REQUESTt=al APPROPRIATIONt.I+ et. stochastic process models which predict the distribution of percent allocation change. These And in 80 percent of the bureau cases, the two distinct predictions are then evaluated empirically dominant model of congressional empirically, using domestic programallocation appropriationbehavior was data from the Eisenhower (F.Y. 1957), the Kennedy (F.Y. 1964), and the Johnson (F.Y. APPROPRIATIONt=a2 REQUESTt+Ot@ 1966) administrations.The empiricalemphasis is on various stages within the executive half of This theory of budgetary decision making, the budget planningcycle. while both influential and straightforwardin and of itself, has unfortunately been plagued by confusion generatedby the multiple uses of Processversus Outcome Incrementalism the term "incrementalism"(LeLoup, 1978a). The term "incrementalism"in its broadest form The incrementalist "Theory of the Budget- has been used to refer to concepts as disparate ary Process," first described and applied to as marginal change, linear decision rules, fair federal budget data by Davis, Dempster and shares, stability in advocate and guardianrules, Wildavskyin 1966, emphasized the importance and "the strategy of disjointedincrementalism" of "aids to calculation" in substantively com- propoundedby Lindblom (1970). plex, yet institutionally stable problem environ- To clarify the meaning of "the theory of ments. Institutional stability was said to engen- incrementalism"to be analyzed in this article, der stable mutual expectations, which markedly therefore, I will adopt a convention proposed reduce the burden of calculations for the by numerous authors (Wanat, 1974; Bailey and participants.Given stable expectations, budget- O'Connor, 1975; LeLoup, 1978; Crecine, per- ary decision makers treat each others' decisions sonal communication)."Incrementalism" in the descriptivesense refers to a pattern of marginal change in final allocation outcome relative to 1Besides Simon's original pioneering work on satis- is the March's some base, which frequently previous ficing (1957), see, for example, Cyert and outcome. Or, as Davis, Demp- "adaptive rationality" (1963), Steinbruner's "cyber- year's allocation netic and cognitive process paradigms" (1971), ster, and Wildavsky themselves put it, "This Newell's and Simon's hierarchical search heuristics year's budget is based on last year's budget, (1972), Winter's evolutionary theory of the firm with special attention given to a narrow range (1971), and Radner's "putting out fires" (1975). of increases or decreases"(1966, pp. 529-30). 356 The AmericanPolitical Science Review Vol. 74

This marginalchange use can be labeled either underlie incrementalist regression results. Le- as "outcome incrementalism"or as "marginali- Loup (1978b) has pointed out that budgetary ty" and will not be the focus of this article. decision processes appear quite differently de- Incrementalismin the explanatory sense, on pending upon time period measurement,level the other hand, refers not just to the fact that of aggregation,and dependent variable format. some reasons for observedmarginality are given And, in perhapsthe most damagingcritique of (Wanat, 1974), but rather to a particular all, Wanat (1974) has demonstrated that the process of decision making which underliesthe empirical goodness-of-fits of Davis et al. can observed allocation choices. From this explana- also be generated by a purely random decision tory point of view, the crucial process hypothe- model in which percent cuts or increases are sis embedded in the Davis, Dempster and drawn each time period from a Uniform [0, .1 Wildavsky models is the linear decision rule distribution. hypothesis mentioned above. Other aspects of Taken individually, the significance of these incrementalist theory, such as institutional sta- disparate critiques can be challenged (Padgett, bility, mutual expectations and strategic calcu- 1978). However, at least two more general lations, serve the important theoretical function themes underlie these arguments. First is the of increasing the a priori plausibility of this methodological point, to be developed in a later formalization. However, it is the linear decision section, that the incrementalistregression find- rule aspect which is tested by quantitative ings are very susceptible to multiple process goodness-of-fit measures.2 This hypothesis, interpretations. Second is a more substantive therefore, may be labeled "processincremental- concern with the rigidity implicit in the tem- ism," and will be the focus of this article's porally fixed, linear decision rule formulation. analysis. Linear rules appear to relegate the impact of Viewed in this manner, the decision process bureaucratic, political and technical dynamics models of Davis, Dempster and Wildavskyhave either to annual "stochastic adjustments" to a clear affinity with the broader bounded fixed percent changes or to very rare "shift rationalityparadigm. Process incrementalism, in point" alterations in fundamental program this sense, is simply one particularform of the parameters.The dominant image of incremen- "simple decision heuristics" or "standard op- talist theory is one of a very inertial and erating procedures"emphasized by Newell and buffered bureaucratic system which extrapo- Simon (1972) and by Cyert and March(1963). lates deterministicallyfrom t to t+l. For many Alternativeformal operationalizationsof "stan- researchers, an "as if" justification of formal dard operating procedure" theory which exist models built upon this image is sufficient. For in the budgetary literature include Crecine others, however, the image seems unacceptably (1969) and Gerwin(1969). far removed from qualitative accounts of what Even with this context of "process incre- actually transpireswithin the federal budgetary mentalism," however, Davis', Dempster's and process (Natchez and Bupp, 1973).3 Wildavsky'sdecision rule interpretationof their One particular qualitative observation mo- empirical regression results have been chal- tivating the present researchis based upon the lenged by a number of their critics. Natchez Office of Management and Budget and the and Bupp (1973), for example, have argued Department of Housing and Urban Develop- that aggregate bureau level statistics can tap ment internal budget memoranda which were only "massive administrative stability" and available to me,4 The contextually rich infor- completely miss the purportedlymore political policy choices which occur at the program level. Williamson(1966) has derived the iden- 3The question of appropriate criteria for the tical incrementalist"mark up" functional forms evaluation of formal models is of course complex and from radically divergent optimization princi- subject to dispute. However, for one articulation of ples. Gist (1974) has the position that plausible process correspondence is allegedthat incrementalist one patterns are such criterion, see Simon (1968). For the oppos- artifacts of uncontrollability in ing viewpoint, see Friedman (1953). agency allocations. Ripley et al. (1973) have shown that appropriation-expenditurelinkages 4The Crecine archive is a rich collection of internal OMB budgetary planning documents and memoranda, focusing primarily on the OMB Office of Budget Review. This archive spans the period from the 20f course,this is not to implythat otheraspects Truman through the Nixon administrations, and is of incrementalisttheory cannot be investigatedwithin described in greater detail in Crecine (1977). The the regressionframework. For example,estimated Crecine archive was assembled by J. P. Crecine, G. regressioncoefficients can be interpretedin termsof Galloway, M. Kamlet, D. Mowery, J. F. Padgett, and "strategiccalculations." C. Stolp. The Padgett archive was assembled from the 1980 BoundedRationality in BudgetaryResearch 357

mation detail alluded to by Daviset al. does not starting point, however, the serial judgment appear to play, on an annual basis, only the decision maker next makes a conscious choice subsidiary role of "stochastic adjustments to about "direction of search"-namely, whether take into account special circumstances."Ra- to search for alternativesrepresenting increased ther, complex information on a wide variety of budget levels or whether to search for alterna- substantive, administrative, political and eco- tives representing decreasedbudget levels. Pro- nomic dimensions appears to be the heart of pensities to search in one direction or the other decision making and of disputes over program may be influenced by institutionalized role allocations. Indeed, in the executive branch at biases, by aggregatefiscal policy "climate,"and least, one never observes literally a two-stage by decision makers'opinions about the substan- decision processin which first fixed percentages tive merits of the program in question. Once of base are calculated and then "specialadjust- serial judgment decision makersstart searching, ments" made.5 they perceive a number of salient discrete Despite this apparent absence of literal budgetary alternatives,based upon their knowl- incrementalist decision rules, Davis, Dempster edge of a wide variety of program-specific and Wildavskyare justified in their emphasison contextual detail.6 The distribution of alterna- the cognitive limitations of budgetary decision tives so perceived will of course be influenced makers faced with exceedingly complex and by the program's legal and technical con- voluminous information. Executive budgetary straints, which are often discussed in the decision makers do experience great difficulty budgetary literature under the rubric of "con- when confronted with the task of making trollability." omnisciently rational tradeoffs across a wide Within this salient alternative framework, variety of issues and programs,as is indicated the serialjudgment process of choice is simple- by the demise of PPBS. And use of "base" one just starts cycling sequentially through figures as comparativereference points is wide- alternativesencountered along the direction of spread. search, either adding discrete increases or sub- tracting discrete decreases, until an alternative The Theory of SerialJudgment is encountered which is deemed acceptable both on the grounds of program merit and The theory of serial judgment is an alterna- within the context of fiscal policy constraints.7 tive hypothesis about the behavioral structure These serial accept/reject decisions are based on of program or bureau level federal domestic the application of informal "judgment,"which budgetary decision making. Like process incre- is the mixture of cognitive predisposition and mentalism, the theory of serial judgment re- contextual sensitivity to a variety of disparate mains in the bounded rationality tradition; programmatic detail. In other words, each however, it implies greater temporal flexibility than the linear decision rule formulation. This alternative theory derives its name from two 6The contextual determinants of such salient alter- distinctive features: (1) a sequential search natives are almost too large to enumerate. However, through an ordered set of discrete budgetary examples include the following: the size of Boston's latest Urban Renewal application, the amount of alternatives, and (2) a non-deterministicfinal unobligated balance carry-over from the previous year, selection based upon the slightly ambiguous the remaining dollars left in congressional authoriza- application of "informedjudgment." tion, the sizes of new construction starts which could Like the incrementalist decision maker, the be deferred, the amount of mortgage assets which serial judgment decision maker begins the could be sold at a discount, the dollar implications of choice process with a fixed reference point or inflation in per unit costs, the budgetary cost of base, which is historically given in the form of projected demographic change in populations eligible prior budget estimates. From this historical for transfer payments, the dollar implications of new presidential or agency proposed legislation, and so forth. The diversity of these contextual cues is one major reason for a stochastic formal representation. files of the HUD budget office, and spans the fiscal 7The role of "fiscal climate" is specifically singled years 1957-1970. out because of the Crecine (1975) finding that 5The only cases of executive decision behavior I aggregate constraints on the overall size of the federal know of which might be interpreted within this budget are crucial to the structure of OMB and framework are occasional across-the-board percent executive budgetary decision maldng. "Merit" con- cutting exercises, used to generate a list of alternatives siderations, moreover, are meant to be interpreted for further evaluation. These exercises were observed broadly to include political, strategic and administra- during the Eisenhower administration, and also during tive evaluations as well as substantive policy prefer- the very last year of the Johnson administration. ences. 358 The AmericanPolitical Science Review Vol. 74

accept/reject decision may be justified by refer- formed into a percent change format. The goal ence to a number of incommensurable and of the modeling sections which follow will be frequently unmeasurableissues which are at- to use the methodology of stochastic processes tached to the program.However on average,the to derive competing probability distribution probability of accepting salient alternativesis predictions. Since this approach and the meth- determined by the decision maker's cognitive odology upon which it is based are unorthodox predisposition towards the merits of the pro- for budgetary decision research, a few words of gram and by the decision maker'sevaluation of justification are in order. the aggregatefiscal climate. These probabilities As Wanat (1974) and others have already can be interpreted as stochastic analogues of argued, a linear decision rule interpretation of "aspiration levels." Final choice is treated as the Davis et al. regression results is open to non-deterministicin part because of an assump- dispute. Incrementalist regressions certainly de- tion that no unambiguouslycorrect budgetary monstrate that federal budgetary decisions alternativeis perceivedto exist. usually differ in only a marginal, temporally Serial judgment decision theory has its roots stable, and approximately linear manner from in three broader behavioral choice traditions. earlier decisions. Whether such descriptive pat- First, the affinity of this decision process with terns have actually been generated by the Simon's theory of satisficing (1957) is obvious. application of incrementalist decision rules is Serial judgment theory presumes that only a more problematic. limited number of alternativesare considered;it To appreciate the potential severity of this postulates that the first "acceptable" alterna- methodological issue, consider the following list tive encountered is chosen; and it highlightsthe of conceivable alternative decision processes: notions of search and, implicitly, aspiration 1. an economic cost-benefit process in which levels. Secondly, serial judgment theory is the measurements of benefits and costs consistent with one of the three heuristics of change in a marginal and stable manner over Tversky and Kahneman for "judgment under time (classical public finance approach); uncertainty" (1974)-namely, the heuristic of "anchoring and adjustment." The serial judg- 2. an overt bargaining process in which non- ment decision maker starts with an historically incrementally motivated decision makers in given initial value, in the form of prior budget conflict jointly determine budgetary alloca- estimates, and then systematically adjusts this tions within a system in which "power initial value by cycling through neighboring relationships" remain approximately stable alternatives,which are taken here to be discrete (political conflict of interest approach); in character. Finally, the theory of serial 3. a decomposition process in which aggregate judgment builds upon March's and Olsen's fiscal policy based totals, which themselves emphasis (1976) on ambiguityin organizational change only marginally, are sequentially choice. Informal judgment is required because suballocated into smaller and smaller sets of omnisciently rationalcalculations are infeasible. categories in an at least approximately pro- Serial judgment, I argue,is not an unreasonable portional manner (Simon's "nearly decom- way for decision makers to cope with federal posable systems" approach); policy complexities. 4. a selection In OMB and HUD budget plan- process in which randomly gen- qualitative erated proposals are serial frequently is rejected by hostile ning memoranda, judgment political environments which do manifest either in the form of HUD "priority not change radically in composition or selection criteria bands" of program cuts or in the form of over time (evolutionary itemized OMBlistings of programincreases and approach); decreases, which are then cycled through and S. an exogenous determinants process in which tabulated until some aggregatefiscal target is "service levels" are set in response to clien- achieved, tele pressures which build and diminish relatively smoothly over time (municipal finance's public expenditure determinants LinearModels in approach); or even Decision MaldngResearch 6. a null model in which nothing but inflation In the empiricalanalyses which follow, I will drives the system. evaluate the two theories of process incremen- All of these schematically described pro- talism and of serial judgment primarily upon cesses could, under plausible circumstances, the basis of cross-sectionalprogram or bureau generate allocational decisions which differed in level budgetary data which have been trans. only a marginal, temporally stable, and approxi- 1980 BoundedRationality in BudgetaryResearch 359 mately linear manner from earlierdecisions. Of "net causal effects' can be made whether or course, this is not to say that all of these not one understands the detailed mechanisms processes necessarily lead to patterns of deci- by which such effects are transmitted. sions which would be well trackedby the Davis On the other hand, the very adaptability et al. models, nor that one cannot imagine which makes linear models useful in these circumstances in which these processes would predictive and estimation contexts makes linear lead to differing predictions. However, the models indiscriminant in an explanatory con- examples illustrate the point that process inter- text in which the goal is to understand the pretation of incrementalistregression results is underlying process mechanisms by which final not straightforward. decision outcomes are produced. Generallinear I would suggest, however, not only that this models are all too robust in such a context for problem of multiple process consistency with essentially two reasons. First, as Dawes and incrementalist empirical results is severe, but Corrigan (1974) demonstrate through simula- also that this problem is exacerbated by reli- tion, linearmodels are quite robust in goodness- ance on linear regressionmethodology. Whileto of-fit under deviations from optimal regression some extent such multiple process consistency coefficient estimates. This appears to explain problems are inevitable for any formal model, thg high R2's obtained in Wanat's"uniformly linear regression analysis of raw-dollar, time- distributed percent cuts" simulation exercise. series allocation data seems a peculiarlyinsensi- Perhapsmore importantly, however, behavioral tive strategy for making "criticaltest" empirical decision data are very commonly characterized distinctions among competing decision process by conditionally monotone relationships be- theories. tween information inputs and choice outputs, My pessimism in this regard stems in part for the simple reason that information is a cue from the experience of the discipline of psycho- for dimensions of value. These relationships logy with linear models of behavioral choice insure that such data will to a first approxima- theories (Dawes and Corrigan,1974). There, a tion have strong linear components, especially number of simulations have been performed if the range of dependent variable variation is which show that a wide variety of nonlinear not great. In the face of competing process decision process models can generate final explanations, however, the demonstration of choices which are well tracked by simple linear such linear components is frequently not very regressions of those choices on information discriminating(Green, 1968). Instead of a futile inputs. For example, Rorer (1971) has simu- search for marginal improvements to already lated graduateschool admissionschoices by ten high R2's, therefore, I recommend a change in different nonlinear decision rule models, rang- methodological focus-a reaffirmation of the ing from quadratic models to multiplicative "critical test" philosophy, in which the goal is models to various sequential branchingmodels to identify dimensions of data in which the to categorical or pattern recognition models to predictions of competing theories differ mar- lexicographic models. Multiple regressions of kedly. simulated candidate ratings on information inputs, however, produced an average correla- tion coefficient of .85, with a maximum of .96 and a minimum of .71. Thus, even though the Stochastic ProcessMethodology "true" underlying decision processes were radi- cally nonlinear in character, the general linear The alternative stochastic process approach model tracked aggregate patterns of final to be employed in this article is designed to choices fairly well. alleviate some of these methodological prob- One can draw one of two conclusions from lems. The primaryjustification for a focus on this type of result. On the one hand, linear predicted probability densities of cross-sec- functional forms are often quite useful for tional percent allocation change is the fact that modeling and predicting final decision out- such predictions are clearly distinguishableand, comes even when the underlyingchoice process hence, falsifiablerelative to one another. is not understood. In the context of estimation, Focusing explicitly on the analysis of chang- moreover, linear regressionmodels are valuable es in allocations rather than on the analysis of (as long as prescribed precautions are taken absolute dollar levels is one straightforwardfirst against standardeconometric problems such as step towards the elimination of dominant but heteroscedasticity, etc.) for investigating rela- indiscriminant monotonic patterns from data. tive magnitudesof "net causaleffects" of some Models which predict change obviously predict independent variables on some dependent absolute levels as well, but such models are variable of interest. Such relative assessmentof more likely to diverge in statistical evaluation 360 The AmericanPolitical Science Review Vol.74 due to typically lowered goodness-of-fits.Stan- This level of analysis mitigates "compositional dardization of changes in terms of percentages effect" problems, which stem from the possi- insures comparability of programs of various bility that departmentaltotals reflect in aggre- sizes within a cross-sectional grouping frame- gate the mixed outcomes of radically different work. programdecision processes. Cross-sectionalgrouping of programsat sin- As stated above, the objective of the stochas- gle points in time offers some promise for tic process technique is to derive probability decision process research because it tends to density predictions for observable phenomena bring into focus the object of study better than such as percent allocation change. The use of grouping across time for individual programs. this technique will be illustrated in the follow- Cross-sectional grouping of data across pro- ing sections, but one introductory comment is grams tends to hold constant sets of actors and in order. Stochastic process derivations of organizational procedures, even as it increases predicted program choice distrib"lions can be the diversity of the program information cues carried out under two different assumptions- being evaluated within the system. Time-series density parameters are homogeneous across grouping, on the other hand, tends to hold programs, and density parameters are hetero- constant the types of information cues being geneous across programs. The heterogeneity focused on, even as it increases variability in assumption is clearly the more plausible of the actors and organizationalprocedures. However, two, since it permits decision makers to hold while cross-sectional budgetary analysis is ar- varioussubstantive opinions about the merits of guably superior to time-seriesanalysis for large- different programs. Heterogeneity, therefore, ly methodological reasons, simulation tech- will be introducedinto the following deviations niques also will be used in this article to via the technical device of postulating mixing compare the time-seriesbehavior of serialjudg- distributions of parameters across programs. ment with process incrementalism. Both the homogeneous and the heterogeneous Proper level of data aggregationis a complex versions of the two process models, however, issue for budgetary analysis, since quite differ- will be evaluatedempirically. ent considerationsmay govern decision making The actual data to be analyzed here were at the various levels of overall budget total, drawn from the Crecine archive of internal departmental (e.g., HEW)allocations, and pro- Office of Management and Budget (OMB) gram (e.g., Vocational Rehabilitation) alloca- planning documents and memoranda,and from tions. Indeed, as pointed out by Crecine(1975, the Padgett archive of internal Departmentof 1977), in the executive branch higher-level Housing and Urban Development (HUD) plan- fiscal policy outcomes may successively con- ning documents and memoranda.These execu- strain less aggregatelevels of budgetarydecision tive branch data were supplementedwith con- making. gressional data drawn from the Senate Appro- Formally, this hierarchicaldecision structure priations Committee publication Appropria- presents no problem, since it may be repre- tions, Budget Estimates,Etc. sented by a decomposable model in which the Budget stages coded included the following: of one level of analysis are taken as parameters = the time of derivingin part from more aggregateconsidera- CEi OMB's current estimate,9 at tions. In this article, however, parameterswill be taken as exogenously fixed, and the focus of attention will remain on the programlevel of Services or AFDC program) and sometimes to sub analysis for three reasons: (1) Theoretically,the bureau entities, such as PPBS categories. In this article, primary concern here is with behavioral deci- the term "program" refers operationally to that set of sion theory, which correspondsmost naturally sub-departmental units which appears in OMB internal to the more micro level of budgetary decision planning tabulations. Typically, these units are equiva- making. (2) This programlevbl correspondsto lent to what Davis et al. refer to as "agencies" or the Daviset al. bureaulevel of aggregation.8(3) bureaus. However, occasionally but consistently OMB will group small budget line items together into categories such as HEW's Community Health pro. gram(s). A listing of the "program allocation" data 8Semantic ambiguities sometimes confuse aggrega- employed here can be found in Padgett (1978). tion descriptions in budgeting. For example, the term "'agencies" refers in some instances to departments 9The "Current Estimate" is the most recent OMB (e.g., HEW) and in others to bureaus (e.g., Bureau of evaluation of the NOA and Expenditure implications Indian Affairs). Likewise, the term "programs" some- of congressional appropriations, after taldng into times refers to entities very comparable to bureaus account supplementals and uncontrollable adjust- (e.g., HUD's Urban Renewal program, HEW's Family ments. 1980 BoundedRationality in BudgetaryResearch 361

Preview ceilings, of the previous years independent of which particular linear "dec- New Obligational Authority (NOA) for sion rule" is chosen for analysis. program i; Under these assumptions, the process incre- Be = OMB's spring or summer Preview NOA mental hypothesis that allocation decisions are ceiling for program i; produced "as if" they were generated by the Ai= Domestic agencies' NOA fall submission application of simple and temporally stable or request to OMB for program i; linear decision rules may be operationalized as B1 = OMB's Director's Review NOA allowance follows: for program i; Pi= Presidential NOA budget submission to Bi= aol1 CE, + e Congress for program i; Al = t2i - hi + C, Congressional NOA appropriation for Ol program i. B1- a31 Bg+61 =a4 These raw-dollar budget data were then trans- Pi Bji+Qi formed into the following "percentage change" and C= a5i Pi + Pi. formats: A simple transformation yields: Bi- CE A ,-B B___i Ab = h CCEi B B A =i= ai -1 + 1/E1i

= + sA =PiB; ACT_ Ci-Pi Ai a3i -1 oIlA B. Pi AP! = a4i - I + fi/B; More complete institutional descriptions of the and Ac,=a51j-I +vpg/1. federal budgetary planning cycle can be found in Crecine (1977), LeLoup (1977), and Padgett Therefore, under incrementalist decision rule (1978). specifications, the problem of finding the pre- dicted distribution of standardized change re- The Process Incrementalism Model duces to the problem of finding the distribution of decision rule parameters and the distribution To operationalize the theory of process of the standardized error terms. The latter of incrementalism within a stochastic process these two distributions is straightforward-since framework, we first have to specify linear all previous decisions are fixed at the time of "decision rules" for each of the above budget choice and hence are not random variables stages. This specification involves both a choice themselves, the Davis et al. hypothesis that about the relevant "base" and a choice about stochastic disturbances are Normally distrib- appropriate incrementalist functional forms. uted implies that the distributions of standar- Here, the "base" will be taken in most cases dized stochastic disturbances are also Normal. to be simply the most recent prior decision. The main theoretical question, therefore, The only ambiguity in this operationalization centers around the distribution of oa's which is comes in the case of the OMB allowance stage, implicit in incrementalist theory. Davis, Demp- where it is unclear whether the "base" should ster and Wildavsky said little in their original be the most recent decision made by agencies- 1966 paper about the determination of such agency requests-or the most recent decision parameters, except to note that they should made by OMB itself-preview ceilings. Here, the change only occasionally during "shift points." preview ceiling will be the allowance "base" A later (1974) paper, however, presents a more under the assumption that this choice is more explicit exogeneous determinants theory about likely to be closer to OMB's allowance prefer- how decision rule parameters are fixed during ences than is the typically higher agency re- shift points. That theory may be labeled a quest. "cumulative pressure" theory. Hence, The decision rule functional form to be examined will be the simple linear specification While the incremental behavior specified by these models ... to be the with one independent variable. This was the appeared general rule, a major finding concerned the nature of specification which Davis, Dempster and Wil- the exceptions. For many agencies, epochs in davsky found to be empirically dominant. In which the underlyingrelationships appeared to fact, however, it will be seen that the distribu- change were identified statistically.... We in- tional implications of incrementalist theory are vestigated a subset of these epochs by docu- 362 The AmericanPolitical Science Review Vol. 74

mentary analysis and classified the major influ- Davis, Dempster and Wildavsky themselves ences at work... Although some of these adopt, then the distributionof ai is identical to influences were essentially random and non- the distribution of the regression stochastic recurring, most could be seen to be due to disturbanceterm, which of course is Normal. If, specific political, or general economic or social on the other hand, one presumes the z's to be events. This suggests that, although it is basical- then, under the supplementaryas- ly incremental, the budget process does respond stochastic, to the needs of the economy and society, but sumption that the z's are independent, the only after sufficient pressure has built up to Central Limit Theorem implies that cia again cause abrupt changes precipitated by these approaches Normality, regardless of the dis- events (1974, p. 3). tribution of the z's. Two comments can be made about the Thus, the process image which Davis et al. independence assumption re- in supplementary present regarding "shift point" changes quired to make the Central Limit Theorem of decision rules is one in which a variety argument hold in the stochastic z approach. "the needs of exogeneous events influencing First, Davis et al. themselves implicitly adopt exert pressure on the the economy and society" this very common methodological assumption internal operation of the budgetary process. of imme- in their stochastic disturbancerepresentation This pressure in turn cumulates, without Moreimportantly, or excluded variables. however, diate effect, until decision makers recognize of the Central Limit Theo- at advanced versions are forced to recognize the need for change, rem (Loeve, 1955, p. 377) have shown that which point an abrupt transformation of deci- even -sums of dependent random variablescon- sion rules occurs which reflects the aggregate verge to Normals under certain technical condi- level of all of this cumulated pressure. This tions.10 In less precise terms, these advanced budgetary theory is consistent with more gen- results imply that the sum of even dependent (Cyert eral standard operating procedure theory random variables will converge to a Normal 1974), which and March, 1963; Steinbruner, as as one one summand is decision heuristics shift distribution long presumes that simple in the presence of an infinite when a sufficiently severe "prob- "dominant" abruptly only number of other summands. This fact appears the updating and revision of past lem" forces to account for the robustness of Normal ap- learning. proximations of sums even when common p. 9) Davis, Dempster and Wildavsky (1974, independenceassumptions are violated. operationalize this verbal image of cumulated Hence, under at least two different interpre- pressure caused by exogenous events in the tations, the Davis, Dempster and Wildavsky following manner: "shift point" theory implies that all of their decision rule parameters should be at least AO1 po0 +P1liz + 321Z2 + ^ * , + PImiZnm approximatelyNormally distributed.For exam- ple, The z's here represent all those political, eco- nomic and social events or variables which determine the perceived "need" for a program p(au) = O4Iri2 alJ. and which, hence, put pressure on the budget- ary process regarding that program. Davis et al. actually studied 18 such variables, but they This result in turn, coupled with the linear leave little doubt that an even larger number of decision rule specification listed above, implies variables are likely to be operative. The joint that the density of standardized allocation effect of excluded variables was grouped into choices predicted by process incrementalism Normal regression "error" terms. should also be Normal, since the convolution of This cumulative pressure theory of "shift any number of Normals remains Normal. For point" changes, therefore, states that (a) there example, exists a wide variety of possible exogenous event determinants of the ai's, and that (b) these determinants are cumulative or additive in 10These technical conditions are (a) that the their effects. expected differences between finite conditional and Given these two substantive hypotheses, one finite unconditional means and variances of zm's to generate of two approaches can be taken approach zero as m -* ?, and (b) that the so-called the desired predicted distribution of ai. Both, Lindeberg-Feller condition holds. This last condition however, lead to the same answer. If one implies that the ratio of any z's variance to the presumes the z's to be nonstochastic, as is variance of the sum approaches zero as m -+ 0 implicit in the regression framework which (Woodroofe, 1975, p. 256). 1980 BoundedRationality in BudgetaryResearch 363

mate most unimodal patterns. 11 p(Abi) = 1l[Wi, Ul'12], The implications of these two heterogeneity where assumptions, taken one at a time for the illustrative case of Abi, are as follows: (a) A Normal distribution of across programs, [All - I + (p/CEi)J, jli's pj1t with ai's temporarily held constant, implies the and following Normal distribution of AbI's:

aP2.= [a2 + (a, ICE1)2] where

a = [p2 + c2 + (aCefCEI)2. This convolution property accounts for the earlier remark that distributional implications (b) However, if (u72)-l 'in turn is assumed to of process incrementalism are independent of be distributed as Gamma (at, 3) across programs, the particularlinear decision rule specification. then the predicted distribution of percent Increasing the number of terms in a linear change program choices becomes a type of decision rule simply increases the number of Student's t distribution: Normally distributed standardized parameters terms which must be convoluted. But such an ______(Ai _-j)2 ..(a+%) increase in no way affects the final distribu- p(A) = (2)- [ 1 + 2j3 tional form. This distributionalresult holds for individual programs. That is, process incrementalismim- where BG(, a) is the symbol for the beta plies that the probability density of allocation function. choice for any programat a fixed point in time Thus under an assumption of parameter is Normal. This individual programprediction, homogeneity, process incrementalism implies a however, will be identical to the prediction that Normal distribution of appropriately standar- observablecross-program distributions are Nor- dized allocation change. However, under the mal only under the supplementaryassumption more plausible assumption of parameter hetero- that all program density parameters are con- geneity, process incrementalism implies approx- stant across programs. In fact, however, a imately a Student's t distribution of standar- heterogeneity assumption, reflecting different dized allocation change. Not surprisingly, these information inputs and substantive policy pri- two distributions are not qualitatively very orities, is more plausible. different from one another, the main difference Incorporatingheterogeneity into the analysis being the slightly "fatter tails" of the Student's requiresthe postulation of mixing distributions t. of parametersacross programs.Particular such hypotheses are hard to justify on theoretical grounds, so I will simply adopt the strategy of The Serial Judgment Model presumingmixing distributions which are tract- able but which also are sufficiently malleableto To operationalize the competing serial judg- at least approximate virtually any unimodal ment theory of budgetary decision making, first reality. The only constraints in this Normal posit that the historical "starting points" re- case, of course, are that gi parameters lie quired for each budgetary stage are identical to between -o and +oo and that a2 parameters lie the "bases" specified for the process incremen- between 0 and +oo. talism model. Given these fixed historical refer- Two mixing distributionswhich appear rea. ents, however, the serial judgment decision sonable under this strategy are the following: maker then decides upon a direction of search (a) The distributionof means across programsis -namely, upon whether some increase or some Normal (, J2). (b) The distribution of the decrease is warranted both "on the merits" and inverse of variancesacross programsis Gamma (a, 1). The first of these two heterogeneity assumptionsis straightforwardin the absence of 1For example, Gammas can assume a J pattern any theoretical reasons to the contrary. The with peak at 0, or unimodal patterns with peak second of these two assumptionsis convenient anywhere from 0 to ??. Moreover, this class subsumes since Gammas are sufficiently malleable under within it the Chi-Square and Negative Exponential different (a, f) selections to at least approxi- distributions as special cases. 364 The AmericanPolitical Science Review Vol. 74

"under the current fiscal climate." salient alternativesperceived in the interval [0, This first "direction of search" phase can be Abe)," and k representsparticular integer sam- represented, for the illustrative case of OMB ple outcomes of Xi. Xi is a parameterwhich ceilings, as a binary choice between the two equals the expected value of Xi. This measure alternative sets Abi>O and Abi01 = P. Pi P[Ab,a

IL The probability of perceiving two or more Ab=s Asi., salient alternatives in the interval dA bi is where o(dAbi), o(dAbi) is a technical term Ab which can be interpreted as "negligible proba- wherethe i's representthe successivesizes of bility." salient alternativesconsidered, and n represents It is shown in many probability tests (e.g., the total number of alternativesaccepted. Both Feller, 1968, pp. 446-48), that these two Ab11and n are random variables. Hence, the hypotheses are sufficient to imply that the desired p(Abi) is a compound distribution of resultant full distribution of perceived salient p(Abig)and p(n). alternatives over the large interval [0, A?d) is Poisson. That is, 12Thatis, the morecontrollable the program,the P[Xi-k] = Lke-XiAbi. larger,on average,discrete dollar or percentalterna- tivesare likely to be. Moreover,the largerthe average size of individualalternatives, the fewerof themare Here, Xi is the random variable "number of likelyto be perceivedwithin any fixed interval. 1980 BoundedRationality in BudgetaryResearch 365

The distribution P(Abiq) follows from the that 7pi and (113jare constrainedto lie between above Poisson "unbiased"perception of salient 0 and 1 and that Xti is constrained to lie alternatives hypothesis. For it can be shown between 0 and +0: (Karlin and Taylor, 1975, p. 121) that the (a) is distributedas Beta (uv).13 distribution of distances between Poisson pi* events is Exponential (Xi), and n-fold convolu- (b) Pi*is distributedas Beta (v,>). tions of such Exponentials are Gamma (n, Xi). (c) Xi is distributedas Gamma(a,6). Moreover, it can be shown (Feller, 1968, pp. 164-68) that the distribution of number of Under these heterogeneity assumptions,with evaluations until the first acceptance "stopping the added constraint that a = v+v, it can be point" is Geometric(13.1i). Therefore, fixing the shown that the cross-programempirical dis- direction of search to be positive, and using the tribution of choices produced by a serial notation P(Abhin* to represent the n-fold judgment maker will be a Double Pareto: convolution of p(Abhi) with itself, the serial judgment prediction for the distribution of P percent allocation change is Exponential (1-tqf)u2*(4;[- )A -(P+1) RX(l-3131)Jwith a singularity at Abi=O, as is demonstratedin Figure 1. v Ib <0 Coupling this result with first "direction of search"stage is straightforward: * Il-pi* )B *eV' Ab6, AbiO In this case uh and u Y-., where Pi* Api, pi*-!33 and ?r7 Xi(1-1i). This Double Exponential prediction is analo- gous to the process incrementalismprediction of a Normal Density of allocation choice for Thus, the serialjudgment process producesa any one individualprogram. Double Exponential empirical distribution of As was the case in the last section, however, I also wish to explore the implications of the 13The Beta distributionis very malleablesince, plausible assumption that parametersare heter- undervarious selections of u and v parameters,this ogeneous across programs. Using the same distributioncan assume J shapeswith peak at either0 strategy of exploring mixing distributions or 1, U shapeswith peaks at 0 and1, unimodalshapes, which are both tractable and malleable, the triangularshapes, or eventhe rectangularshape of the following hypotheses seem reasonable, given Uniformdistribution.

P(Abi) = 2 In) n=op~n)p(Abq;

00

n=1

'?1 1NzO ji=o0F 0(1-31i)(3) { 0 , 4vbjO oAb> =1 (ni!(bifbii~i ,~,>0

(-ppli)O t~l0

(pi)wa i a 4pJiw-k\l-l~M~)A\g A.bi>O

Source:Compiled by the author.

Figure1. SerialJudgment Derivation for the IndividualProgram Level 366 The AmericanPolitical Science Review Vol. 74 percent change allocation choices for the case is relatively autonomous, institutionally, from of homogeneous programs. And it produces a decision making for domestic programs. Also, Double Pareto empirical distribution of percent very small programs, whose Preview Current change allocation choices for the more plausible Estimate totalled less than $1 million, were case of heterogeneous programs. excluded from the analysis. In all, complete Graphically, the competing distributional data series for 62, 98 and 94 domestic programs predictions of process incrementalism and serial could be reconstructed for FYs 1957, 1964, judgment are illustrated in Figure 2. and 1966, respectively. These programs repre-. Substantively, the characteristic sharp peak sented 83, 91, and 86 percent of the total NOA near zero of the Double Pareto implies that domestic14 President's budget for FYs 1957, most programs most of the time receive budget 1964, and 1966. A more complete description allocations which are only marginally different and listing of the program allocation data from the historical referent or base. The equally analyzed can be found elsewhere (Padgett, characteristic "fat tails" of the Double Pareto, 1978). however, imply the occasional occurrence of In particular, the "major budget stages" very radical changes. What Davis, Dempster and coded are listed in Table 1. More than the usual Wildavsky describe as a "shift point," induced number of stages were coded for FY 1957 by exogenous intervention into a routinized because of the greater emphasis placed on these budget system, the serial judgment strategy early planning figures by the Eisenhower ad- generates as the normal outcome of its more ministration. An additional FY 1964 ceiling flexible decision process. revision stage was coded because of Kennedy's unanticipated midyear decision to cut the Data Analysis budget total by $2 billion in order to increase the chances of congressional passage of the first As mentioned above, the program allocation major Keynesian tax-cut measure. data to be analyzed here were drawn from the As has already been discussed, the data were Crecine OMB and the Padgett HUD archives, first standardized into a percent change format. and were supplemented with publically avail- Besides the Ab, Aa, Ab; Ap, and Ac transforma- able congressional data. The years chosen for tions already mentioned, the standardizations analysis were fiscal years 1957, 1964, and employed were the following: 1966. These three particular years were selected within data availability constraints to represent one year each from the Eisenhower, the Ken- nedy, and the Johnson administrations. Only 14The term "domestic" is meant to exclude the domestic program data were coded because of Department of Defense-Military,the Atomic Energy the Crecine (1975) finding that allocational Commission, the National Aeronautics and Space decision making for National Security programs Administration,and ForeignMilitary Assistance.

Process Incrementalism Serial Judgment

p( Ab) p(Ab)

Normal ~~~~~Exponential Student's t Pareto

LAb A Source:Compiled by theauthor,

Figure2. GraphicalDisplay of Incrementalismversus SerialJudgment Distributional Predictions 1980 BoundedRationality in BudgetaryResearch 367

A 1957 Projection - 1956 Current Estimate) Table 2 reports finally derived Kolmogorov- 1956 Current Estimate Smirnov goodness-of-fit statistics, along with critical values for the .05 level of significance. 1957 Preview Request - 1957 Projection) Results for both constrainedand unconstrained versions of the serialjudgment model are given. 1957 Projection The criticalvalues reportedare the Monte Carlo values cited by Bickel and Doksum (1977, p. (1957 Ceiling - 1957 Projection 381 ), which are exact for the Normalcase with 1957 Projection parametersestimated and which are approxima- tions otherwise.16 1964 "Ratcheted" Ceiling - 1964 Ceiling The results are remarkablyconsistent over 1964 Ceiling all years and over all budget stages. In 17 out of the 18 budget stages analyzed, serialjudgment The primary statistical procedureemployed predictions are superior to processincremental- for the relative evaluation of the process incre- ism predictions when numbers of parameters mentalism and the serial judgment theories was are constrained to be comparable. Moreover, the Kolmogorov-Smirnov one-sample test. The even this one discrepant Abr stage vanishes Kolmogorov-Smirnov statistic is the maximum when the unconstrainedserial judgment model deviation over all cases of the empirical cumula- is estimated, since almost half of the program tive distribution from the predicted cumulative cases in this stage were observed to equal zero probability distribution (Bickel and Doksum, percent change. When the full serialjudgment 1977). Hence, the lower the statistic, the better model is estimated, all Double Pareto statistics the goodness-of-fit. drop below the critical values cited. It is shown One problem with the direct calculation of elsewhere (Padgett, 1978) that these general Kolmogorov-Smirnov statistics in the present comparative findings are not altered if one research context is the fact that the full serial employs averageabsolute error,sum of squared judgment models, as developed above, possess errors, or x2 as alternativegoodness-of-fit mea- one extra free parameter than do the corre- sures. sponding process incrementalism models. This Thus on the basis of more refined cross-sec- problem was solved by constraining Exponen- tional analysis of percentageallocation change, tial Pi* and Pareto u* parameters to equal 1.0. the theory of serial judgment appears more With or without this comparability constraint, consistent with program-levelfederal budgetary however, all relevant density parameters were decision making than does the Davis, Dempster estimated by the method of maximum likeli- and Wildavsky theory of process incremental- hood, except for Student's t parameters which ism. It can be shown, moreover, that this were estimated by the method of moments.15 superiority is not an artifact of uncontrollabil- ity in the budget process (Padgett, 1978). One oversimplification of the serial judg- ISMaximum likelihood estimation for the Normal, Exponential and Pareto distributions is described in Johnson and Kotz (1970, Vol. 1). The method of 16Thestandard Kolmogorov-Smirnov critical values moments (Johnson and Kotz, 1970, Vol. 2, p. 116) reportedin manystatistics text tablesare not appro- was employed for the Student's t due to convergence priatehere due to parameterestimation. Unfortunate- difficulties experienced in the numerical analysis of ly, relevantMonte Carlo simulations for non-Normal the three relevant log likelihood derivative functions. casescould not be foundin theliterature.

Table 1. Budget Stages Coded

FY 1957 FY 1964 FY 1966

OMB Preview Current Estimates OMB Preview Current Estimates OMB Preview Current Estimates OMB May Projections OMB Preview Ceilings OMB Preview Ceilings Agency Preview Requests OMB "Ratcheted" Ceilings Agency Requests OMB Preview Ceilings Agency Requests OMB Director's Review Allowances Agency Regular Requests OMB Director's Review Allowances President's Budget OMB Director's Review Allowances President's Budget Congressional Appropriations President's Budget Congressional Appropriations Congressional Appropriations

Source: Compiled by the author. 368 The American Political Science Review Vol. 74

ment model, however, should be noted. It is respective roles in the budgetary process. apparent from visual inspection of Double Despite the observed superiority of serial Pareto residuals that, over the range of typically judgment's distributional predictions over pro- less common decrease choices, Double Pareto cess incrementalism's distributional predictions, predictions for agency requests are higher than it is of course always possible that the Double observed cumulative distributions, and that, Pareto distribution could also be consistent over the same range, Double Pareto predictions with some other decision process theory not for OMB ceilings and allowances are lower than yet examined.17 While no definitive answer to observed distributions. This suggests that the this problem exists, some suggestive informa- simplifying assumption that X* parameters are tion can be generated by examining the more symmetric for both increase and decrease direc- "fine-grained" predictions of serial judgment tions of search is somewhat in error. Apparent- ly, once agencies decide that some decrease from ceiling is necessary, they are much more reluctant to grant sizable decreases than they 171n fact, I report elsewhere (1978) that the serial are to grant comparable-size increases. On the judgment model is also superior in its distributional other hand, once OMB decides that some predictions to two other stochastic models of budget- decrease for a program is warranted, OMB ary decision maldng-a tacit bargaining model, which appears much more eager to make such a is a member of the rational choice tradition, and an information bombardment model, which is a member decrease sizable than it does to grant compar- of the recent "organized anarchy" tradition of Cohen, able increases. These two observed biases, of March, and Olsen (1972). The fact remains, however, course, are quite consistent with our substan- that possible alternative process consistency is a tive preconceptions about agencies' and OMB's serious issue.

Table 2. Kolmogorov-SmirnovStatistics

SerialJudgment Incrementalism (Unconstrained) (Constrained) Double Double Double Double Student's Expon. Pareto Expon. Pareto Normal t 1957 (n=62) OMBProjections (IWe) .1938 .1029 .2104 .1003 .2366 .2307 PreviewRequests (hap) .1572 .0547 .1891 .0596 .3148 .3020 OMBCeilings (Ab') .1460 .0544 .1510 .0691 .2310 .2292 RegularRequests (Oa) .1502 .0474 .1530 .0441 .2821 .2703 OMBD.R. Allowances (fb) .1147 .0597 .1174 .0637 .2448 .2115 President'sBudget (Ap) .1342 .0723 .2674 .2452 .4329 .4271 CongressionalApprop. (Ac) .0936 .0528 .2198 .1850 .2513 .2572 CriticalK-S value at p=.05 .1123 1964 (n=98) RegularOMB Ceilings (fib) .0794 .0611 .0812 .0547 .1666 .1472 RatchetedCeilings (Abr) .0566 .0564 .4490 .4490 .2539 .2733 Agency Requests (Aa) .1439 .0682 .1547 .0711 .2571 .2422 OMBD.R. Allowances (Afb) .1010 .0559 .1134 .0575 .2398 .2267 President'sBudget (Ap) .1638 .0667 .2037 .1392 .3111 .2817 CongressionalApprop. (Ac) .1209 .0671 .1473 .0767 .2574 .2210 CriticalK-S value at p=.05 .0897 1966 (n=94) OMB Ceilings (A 5) .1900 .0787 .1945 .0745 .2771 .2680 Agency Requests (Aa) .1535 .0772 .1725 .0867 .2781 .2596 OMB D.R. Allowances (Ab) .1118 .0366 .1368 .0515 .2271 .2128 President'sBudget (Ap) .1972 .0472 .2327 .0642 .3414 .3297 CongressionalApprop. (Ac) .1783 .0514 .2170 .0785 .3484 .3410 CriticalK-S value at p=.05 .0916

Source: Data coded from CrecineOMB archive and PadgettHUD archiveof U.S. governmentinternal memoranda and budgetaryplanning documents. 1980 BoundedRationality in BudgetaryResearch 369

theory. Perhaps one of the most critical of such Table 3. X2 Test of Compound Hypothesis predictions is that observed Double Pareto that All Individual Departments' distributions in fact derive from a mixture of Distributions are Exponential Double Exponential distributions. This expecta- tion suggests that if one could somehow group 1957 (n=53) programs, on an a priori basis, into relatively OMB Projections (Abp) 19.144 more homogeneous parameter subsets, the ob- Preview Requests (Aan) 23.898 served distributions should be close to the more OMB Ceilings (Ab') 25.723 fundamental Double Exponential. Regular Requests (Aa) 13.289 OMB D.R. Allowances (Ab) 16.724 One obvious such grouping procedure is to President'sBudget (Ap) 16.649 subset programs according to their departmen- Congr. Appropriations (1c) 9.024 tal memberships. It is at least plausible to speculate that, because of organizational struc- Critical X23 at p.05 35.172 ture, strategic or other reasons, program param- eters are more homogeneous within depart- 1964 (n=93) OMBRegular Ceilings (Aib) 30.866 ments than they are across departments. If this RatchetedCeilings (Abr) 38.984 were the case, then the distributions of alloca- Agency Requests (Aa) 21.221 tion choices for only HUD or HEW programs, OMBD.R. Allowances (ji) 35.873 for example, could be examined separately, via President'sBudget (Ap) 27.136 the Kolomogorov-Smirnov test or some other Congr.Appropriations (AC) 26.569 procedure, to see if such distributions are CriticalX25 at p=.05 49.802 consistent with Double Exponential predic- tions. 1966 (n=91) Unfortunately, for the data analyzed here, OMBCeilings (Ab) 40.790 this straightforward procedure is unreliable due Agency Requests (Aa) 28.513 to the small sample sizes of programs within OMBD.R. Allowances (Ab) 21.663 departments. (The data set used here had an President'sBudget (Ap) 31.199 26.511 average of only eight programs within depart- Congr.Appropriations (Ac) ments.) Under this situation, therefore, a x2 CriticalX2s at p=.05 49.802 significance test of the more complicated com- pound hypothesis that all departments' distri- Source:Data coded from CrecineOMB archive and PadgettHUD archiveof U.S. governmentinternal butions are Exponential simultaneously was memorandaand budgetary planning documents. constructed along the lines suggested by Bickel and Doksum (1977, p. 320).' The procedure employed was the following: Using absolute grained" expectation that observed value choices, individual department param- Double MI Pareto distributions actually derive from eters first were estimated using the Newton a heterogeneous mixture of underlying Double iterative procedure suggested by Bickel and Exponentials. Doksum.18 The X2statistic was then calculated for each budget stage over all departments on One final empirical point is worthy of note. the basis of the four fixed-interval groupings The consistency of serial judgment theory with [.00, .05), [.05, .10), [.10, .20) and [.20, oo). the high regressionR 2's reported by Daviset al. The results are reported in Table 3, along is easily demonstrated.Monte Carlosimulations with the critical X2 (p=.05) appropriate for of a sequence of Double Exponential choices m-s-i degrees of freedom. (s here is the number have been performed, and incrementalist"deci- of parameters estimated, equal to the number sion rule" regressions have been run on the of departments in each stage; and m=4-s is the reconstructed time series dollar outcomes. For total number of intervals over which the X2 purposes of illustration, 56 programallocation statistics were calculated.) As can be seen, the "cases"lwere generated over 17 years19 using compound hypothesis that all departments' pi*, IBtand MIparameters which were randomly distributions are Exponential cannot be re- selected from Beta and Gamma mixing distribu- jected for any budget stage. These results give tions. The parametersof these mixing distribu- some support to the serial judgment "fine- tions, in turn, were taken to equal the average, over the three years studied, of the uncon- strained parametersestimated for each Ab, AR, 189? was constrainedto equal 1, and departments with only three programsor less were excluded from the analysis.I am gratefulto MarkKamlet for bringing "9These were the number of bureaus and years this procedureto my attention. studied by Davis et al. (1966). 370 The AmericanPolitical Science Review Vol. 74

-p, and Ac budget stage. Dollar time-seriesdata superiority of this serial judgment theory over were reconstructed from simulated percent the competing Davis, Dempster and Wildavsky change choices by means of the identities theory of process incrementalism. In -some respects, these results support both PRES.BUDGETi,t = CONG. APPROP.it-(1+Abit) the incrementalists and their critics. Certainly, (l+Ai,t)(1+APi,t), the importance of the bounded rationality paradigm's emphasis on cognitive constraints CONG.APPROP.i, t = PRES&BUDGETi (1 +Aci, t), and simplifying decision heuristics, which the incrementalists share, has been reaffirmed. In where the Preview Current Estimate here is addition, the incrementalist notion of "base" taken to equal the simulated congressional has been given support in the context of serial appropriation. The incrementalist regressions judgment's reinterpretation of that concept as a run were of the simple form listed in the "reference point." However, when the focus beginning of this article, without using any shifts from absolute dollar levels to the analysis 'shift points." of allocational change, the rigidity implicit in The resulting simulated distributions of R's the incrementalist temporally fixed decision are reported in Table 4, along with the analo- rule formulation is undercut. In its stead, serial gous distributionsreported by Davis, Dempster judgment theory substitutes a systematic, yet and Wildavsky (1966, p. 537). It is apparent very contextually dependent process of search that the distribution of serial judgment simula- through and informal evaluation of a limited tion R's is not substantially different from the number of salient alternatives. distribution of actually estimated R's. Hence, This serial judgment process of decision serial judgment theory is consistent with gross making generates budgetary outcomes in which time-series patterns of federal budgetary data, most program allocations most of the time as well as with more refined cross-sectional differ in only a marginal, but temporally patterns. However, these simulationresults also variable, manner from the historical base. How- illustrate concretely the methodological argu- ever, occasionally, as the normal outcome of ment presented above about the insensitivity of serial judgment decision making, more radical regression analysis in process-orientedbudget- and "catastrophic" changes are also produced. ary decision-makingresearch. Serial judgment theory implies that the federal budgetary system is much more responsive to Conclusions political, bureaucratic and technical dynamics, on a routine even if constrained basis, than the This article has examined the methodolog- theory of process incrementalism would lead ical foundations of incrementalist quantitative one to believe. research and has proposed a new bounded Somewhat paradoxically, this formal serial rationality theory of budgetary decision mak- judgment theory is quite consistent with Wil- ing. The empirical results support the relative davsky's verbal description of The Politics of

Table4. Frequenciesof RegressionCorrelation Coefficients

A. SerialJudgment Simulation R's (n=56) 1 - .995 - .99 - .98 - .97 - .96 -.95 - .94 - .93 - .90 - .85 -0 Executive 3 5 8 6 4 5 3 4 6 3 9 Congressional 28 8 7 1 2 2 2 1 3 1 1

B. Davis,Dempster and WildavskyReported R's (Exec. n=61; Cong.n=67) 1 - .995 - .99 - .98 - .97 - .96 - .95 - .94 - .93 - .90 - .85 -0 Executive 9 2 2 8 5 2 4 3 5 11 10 Congressional 21 8 15 4 5 2 2 1 5 2 2

Source: Compiledby the author;Otto A. Davis,M. A. H. Dempster,and AaronWildavsky (1966), "A Theory of the BudgetaryProcess," American Political Science Review 60: 529-47. 1980 BoundedRationality in BudgetaryResearch 371

the Budgetary Process. Certainly, Wildavsky's Crecine,John P. (1969). GovernmentalProblem-Solv. generalizationsthat "budgeting is experiential; ing: A ComputerSimulation of MunicipalBudget- - . . budgeting is simplified; . .. budgeting offi- ing. Chicago:Rand McNally. cials 'satisfice'; ... budgeting is [frequently (1975). "Fiscal and OrganizationalDeter- outcome] incremental" are embedded in the minants of the Size and Shape of the U.S. Defense serial judgment Budget." In Appendices: Commissionon the Or- formalization. Also the numer- ganization of the Governmentfor the Conduct of ous strategic cues itemized by Wildavskymay ForeignPolicy, Vol. 4. be important components of the contextual (1977). "Coordinationof Federal Fiscal and information detail which lies behind serial BudgetaryPolicy Processes:Research Strategies for judgment parameters.Indeed, one implication Complex Decision Systems." Presented at the of the serialjudgment model is that Wildavsky's annual meeting of the AmericanPolitical Science original highly contextual and political empha- Association,Washington, D.C. sis need not be abandoned in the search for Cyert, Richard M., and James G. March (1963). A bounded rationalityprocess regularities. BehavioralTheory of the Firm. EnglewoodCliffs, Future research should be N.J.: Prentice-Hall. directed toward Danziger,James N. (1974). "Budget-Maldngand Ex- an analysis of the actual substantive issue penditureVariations in EnglishCounty Boroughs." content which flows through this serial judg- Ph.D. Dissertation,Stanford University. ment process. In model terms, this suggests an Davis, Otto A., M. A. H. Dempster and Aaron investigation into the cross-programstructure Wildavsky (1966). "A Theory of the Budgetary of estimated serial judgment parameters.One Process." American Political Science Review 60: such extension which seems promising is to 529-47. model serial judgment parametersin terms of (1971). "On the Process of BudgetingII: An the hierarchicalorganizational structure empha- EmpiricalStudy of CongressionalAppropriations." sized In R. F. Byrne et al. (eds.), Studies in Budgeting. by Crecine. Indeed, it can be shown that Amsterdam:North-Holland. formally embedding the serial judgment pro- (1974). 'Toward a Predictive Theory of gram-level process into a departmental-level GovernmentExpenditures: U.S. Domestic Appro- attention-focusing process, in which salient priations." British Journal of Political Science 4: alternativesare sequentiallyallocated to various 1-34. programsaccording to a stationary probability Dawes, Robyn M., and Bernard Corrigan (1974). "strategy vector" until a fixed aggregate fiscal "LinearModels in Decision Making,"Psychological target is reached, is also consistent with the Bulletin 81: 95-106. empirically confirmed Double Pareto distribu- Feller, William(1968). An Introductionto Probability tion (Padgett, 1979). This type of extension Theory and Its Applications, Vol. 1. New York: might be one approach to reconciling bounded John Wiley. rationality theories, such as serial Friedman,Milton (1953). "The Methodologyof Posi- judgment, tive Economics."In Essayson PositiveEconomics with more aggregate organization theory and Chicago:University of ChicagoPress. with historically oriented studies of budgetary Gerwin, Donald (1969). Budgeting Public Funds. managementstrategies. Madison:University of WisconsinPress. Work in this direction is currently in pro- Gist, John R. (1974). MandatoryExpenditures and gress. Serial judgment theory as it now stands, the Defense Sector: Theory of BudgetaryIncre- however, permitssuch organizationaland politi- mentalism. Pacific Palisades,Calif.: Sage Publica- cal extensions to proceed on a more firm tions. behavioraldecision theory foundation than has Green, Bert F. (1968). "Descriptionsand Explana- previouslyexisted. tions: A Comment on Papers by Hoffman and Edwards."In B. Kleinmuntz(ed.), FormalRepre- sentation of Human Judgment. New York: John References Wiley. Johnson, Norman L., and SamuelKotz (1970). Con- Bailey, J., and R. O'Connor (1975). "Operationalizing tinuous UnivariateDistributions, Vols. 1 and 2. Incrementalism: Measuring the Muddles." Public New York: John Wiley. Administration Review 35: 60-66. Karlin,Samuel, and HowardM. Taylor (1975). A First Bickel, Peter J., and Kjell A. Doksum (1977). Mathe- Course in Stochastic Processes. New York: Aca- matical Statistics. San Francisco: Holden-Day. demic Press. Cohen, Michael D., James G. March and Johan P. Larkey,P D. (1979). EvaluatingPublic Programs: The Olsen (1972). "A Garbage Can Model of Organiza- Impact of GeneralRevenue Sharingon Municipal tional Choice." Administrative Science Quarterly Governments.Princeton, N.J.: PrincetonUniversi- 17: 1-25. ty Press. Cowart, A. T., T. Hansen and K Brofoss (1975). LeLoup, Lance (1977). Budgetary Politics: Dollars, "Budgetary Strategies and Success at Multiple Deficits, Decisions. Brunswick,Ohio: King'sCourt. Decision Levels in the Norwegian Urban Setting." (1978a). Review of Aaron Wildavsky,Budget- American Political Science Review 69: 543-58. ing. American Political Science Review 72: 372 The American Political Science Review Vol. 74

351-52. Personality Measurement in Medical Education. (1978b). "The Myth of Incrementalism:Ana- Washington, D.C.: Association of American Med- lytic Choices in Budgetary Theory," Polity 10: ical Colleges. 488-509. Sharkansky, I. (1968). "Agency Requests, Guberna- Lindblom, Charles E. (1961). "Decision-Makingin torial Support, and Budget Success in State Legisla- Taxation and Expenditures."In Public Finance: tures." American Political Science Review 62: Needs, Sources, and Utilization, National Bureau 1220-31. Committee for Economic Research. Princeton, Simon, Herbert A. (1957). "A Behavioral Model of N.J.: PrincetonUniversity Press. Rational Choice." In Models of Man. New York: Lindblom,Charles E. and DavidBraybrooke (1970). A John Wiley. Strategyof Decision. New York: Free Press. (1968). "On Judging the Plausibility of Theo- Love, Michel (1955). Probability Theory. Toronto: ries." In J. van Rootselaar and H. Staal (eds.), Van Nostrand. Logic, Methodology and Philosophy of Sciences Natchez, Peter B., and IrvinC. Bupp (1973). "Policy III. Amsterdam: North-Holland. and Priority in the BudgetaryProcess." American (1972). "Theories of Bounded Rationality." In PoliticalScience Review 67: 951-63. C. B. McGuire and Roy Radner (eds.), Decision Newell, Alan, and H. A. Simon (1972). Human and Organization. Amsterdam: North Holland. Problem Solving. EnglewoodCliffs, N.J.: Prentice- Steinbruner, John D. (1974). The Cybernetic Theory Hall. of Decision. Princeton, N.J.: Princeton University Padgett, John F. (1978). "Coping with Complexity: Press. StochasticModels of BudgetaryDecision Making in Stromberg, J. L (1970). "The Internal Mechanisms of O.M.B. and Domestic Agencies." Ph.D. Disserta- the Defense Budget Process." Rand Publication tion, Universityof Michigan. No. RM-6243-PR. (1979). "A Sequential Attention Model of Wanat, John (1974). "Bases of Budgetary Incremental- Federal Budgetary Decision Making: Preliminary ism," American Political Science Review 68: Historical, Mathematicaland EmpiricalAnalyses." 1221 -28. RIAS WorldngPaper No. 19, HarvardUniversity, rilliamson, Oliver E. (1966). "A Rational Theory of unpublished. the Federal Budgetary Process." In Gordon Tul- Radner, Roy (1975). "A BehavioralModel of Cost lock (ed.), Papers on Non-Market Decision Making. Reduction." The Bell Journal of Economics 6: Charlottesville, Va.: Thomas Jefferson Center for 196-215. Political Economy. Ripley, R., Grace Franklin, William Holmes, and Winter, Sidney G. (1971). "Satisficing, Selection and WilliamMoreland (1973). Structure,Environment, the Innovating Remnant." Quarterly Journal of and Policy Actions: Exploring a Model of Policy- Economics 85: 237-61. Making.Beverly Hills, Calif.: SagePublications. Woodroofe, Michael (1975). Probability with Applica- Rorer, Leonard G. (1971). "A CircuitousRoute to tions. New York: McGraw-Hill. Bootstrapping."In Harold B. Haley et al. (eds.),