Money, Politics, and the Post-War Business Cycle

Board of Governors of the Federal Reserve System International Finance Discussion Papers Number 572 November 1996 MONEY, POLITICS AND THE POST-WAR BUSINESS CYCLE Jon Faust and John Irons NOTE:InternationalFinanceDiscussionPapersarepreliminarymaterialscirculated to stimulate discussion and critical comment. References to International Finance Discussion Papers (other than an acknowledgment that the writer has had access to unpublished material) should be cleared with the author or authors.

Abstract While macroeconometricians continue to dispute the size, timing, and even the existence of e(cid:11)ects of monetary policy, political economists often (cid:12)nd large e(cid:11)ects of political variables and often attribute the e(cid:11)ects to manipulation of the Fed. Since the political econometricians often use smaller information sets and less elaborate approaches to identi(cid:12)cation than do macroeconometricians, their striking results could be the result of simultaneity and omitted variable biases. Alternatively, political whims may provide the instrument for exogenous policy changes that has been the Grailofthepolicyidenti(cid:12)cationliterature. Inthispaper, welayoutandapplya framework for distinguishing these possibilities. We (cid:12)nd almost no support for the hypothesis that political e(cid:11)ects on the macroeconomy operate through monetary policy and only weak evidence that political e(cid:11)ects are signi(cid:12)cant at all.

Money, politics and the post-War business cycle 1 Jon Faust and John Irons Suppose you were looking for a simple economic rule of thumb for remembering what party held the White House in each post-War presidential term. You would be hard pressed to do better than the following rule: if a recession starts in the (cid:12)rst 6 quarters of term, its a Republican; otherwise its a Democrat. This rule calls 12 of 13 presidential terms|its one faux pas is calling Reagan’s second term Democratic. Political economists have uncovered many such striking associations between political variables and headline measures of macroeconomic well-being such as economic growth, unemployment, and in(cid:13)ation. Evaluating the size and cause of such 2 e(cid:11)ects has generated a (cid:13)ourishing literature. One leading explanation for the as- 3 sociations in U.S. data seems to be the political manipulation of monetary policy. These political results are all the more striking from the perspective of the venerable and growing macroeconometric literature on the e(cid:11)ects of monetary policy. This literature continues to show little agreement on the size, timing, and even the 4 existence of e(cid:11)ects of monetary policy. The con(cid:13)ict between the political econo- 1 Faust is a sta(cid:11) economist at the International Finance Division of the Board of Governors of the Federal Reserve System. Irons is in the Economics Department at MIT. The authors thank Olivier Blanchard, Mike Gibson, Beth Ingram, Ed Leamer, Torsten Persson, Lars Svensson, the econometricslunchgroupandworkshopattheFederalReserveBoardaswellasseminarparticipants at Duke, Indiana U., MIT, Northwestern, and North Carolina State for useful comments. Part of this work was completed while Faust was a visitor at the Institute for International Economic Studies in Stockholm. Irons thanks the National Science Foundation for Financial Support. The views in this paper are solely the responsibility of the authors and should not be interpreted as re(cid:13)ectingtheviewsoftheBoardofGovernorsoftheFederalReserveSystemorofanyotherperson associatedwith the Federal ReserveSystem. 2 Friedlaender, 1973; Nordhaus, 1975; Hibbs, 1977, 1987; MacRae, 1977; Frey and Schneider, 1978;McCallum,1978;Tufte,1978;GoldenandPoterba,1980;Beck,1982;Browning,1985;Allen, 1986; Chapell and Keech, 1978; Richards, 1986; Soh, 1986; Alesina and Sachs, 1988; Haynes and Stone, 1989,1990; Bizer and Durlauf, 1990; Ellis and Thoma, 1991; Allen and McCrickard, 1991; G(cid:127)artnerandWellersho(cid:11),1991;HavrileskyandGildea,1991a,b;AlesinaandRoubini,1992;Alesina, Londregan and Rosenthal, 1993; Hess, 1993; Klein, 1993; Alesina and Rosenthal, 1995; Bange, Bernhard,andGranato, 1995. 3 e.g., Alesinaand Sachs,1988; Alesina and Rosenthal,1995. 4 Recent contributions include Bernanke and Blinder, 1992; Christiano and Eichenbaum, 1992; Leeper and Gordon, 1992, 1994; Christiano, Eichenbaum, and Evans, 1994; Hoover and Perez 1994a,b;Leeper,1993;RomerandRomer,1989,1990,1994;Sims,1992;SimsandZha,1994;Strongin1995.

metric results showing strong and consistent e(cid:11)ects of politically driven monetary policy and the less de(cid:12)nitive macroeconometric results almost certainly stems from blind spots in each approach. The macroeocnometric literature pays elaborate attention to questions of identi(cid:12)cation, attempting to sort out the e(cid:11)ects of policy from among the myriad interactions in a simultaneous system. Political variables are seldom if ever explored. In contrast, virtuallyallofthe worklinkingpoliticsto economicoutcomesis bi-variate, linking one economic variable with one political variable (many such bi-variate rela- 5 tions are examined). Further, the political variable|either a dummy variable for thepartyinpower,orsomeothermeasureofthepoliticalstate|isgenerallytreated as exogenous, despite the fact that endogeneity of political variable is clear a priori 6 and well documented. There is also a literature relating politics to instruments of 7 monetary policy, rather than macroeconomic outcomes. In this literature, some indicator of policy stance is modelled as a function of several economic and political variables. The link from instrument to economy is generally not investigated, and once again the political variables are often taken to be exogenous. In light the blind spots in the literatures, we see two obvious possibilities for resolving the con(cid:13)icting results. The macroeconometricians could be correct if the political results were all due to simultaneous equations bias and omitted variable bias. The dramatic results advanced by the political economists, however, make it plausible that more sophisticated econometric techniques will not eliminate the e(cid:11)ects. In this case, the formal interpretation is that the political variables provide the instrument needed for identi(cid:12)cation: if the party in power determines the path of monetary policy, and the party is itself chosen (at least in part) based on noneco- 5 Amongthe papers listed above, Friedlaender,1973, is the clearest exception to the bi-variate norm. Evenwhentheinformationsetincludesmorethanonemacroeconomicvariable,asinGa(cid:127)rnter and Wellersho(cid:11), 1991, it seldomincludesthe major determinantsfound in typical macroeconomic models. 6 Results on either side of this issue includes Stigler, 1973; Tufte, 1978; Hibbs, 1987; Fair,1978,1982; Gleisner, 1992; Alesina, Londregan and Rosenthal, 1993; Haynes and Stone, 1994; Alesinaand Rosenthal,1995; GranatoandSuzuki, undated. 7 e.g., Golden and Poterba, 1980; Beck, 1982; Allen, 1986; Richards, 1986; Allen and Mc- Crickard,1991;Ellis andThoma,1991;Bange, Bernhard, andGranato,1995. 2

nomic factors, then political variables may be the instruments for exogenous policy changes that has been the Grail of the policy identi(cid:12)cation debate. Thispaperprovidesatheoreticaleconometricframeworkfor addressingissues of causalityfrompoliticstotheeconomy. Thisframeworkshedssomelightonstrengths and weaknessesofearlierworkandprovidesthebasisforanewlookat theevidence. Whilethebasicapproachisverygeneral,theparticularapplicationshouldbeviewed asa(cid:12)rststepandhasseverallimitations. Weconsideronlyonepoliticalvariable,the party in the White House, and ignore Congressional issues raised, e.g., by Alesina, Londregan and Rosenthal [1993]. We do not allow for systematic di(cid:11)erences across administrations of a given party, which have been found to be important in some work [e.g., Beck, 1982]. Because we are focussing on special issues that arise under a (cid:12)xed, long election cycle, we consider only U.S. data. For the most part, we consider only the major macroeconomic variables: output, employment, interest rates, money, and in(cid:13)ation. In focussing on the links from politics to the economy, we do not model the reaction functions as carefully as work focussing on reaction functions exclusively. Further, we limit ourselves largely to linear e(cid:11)ects. Finally, 8 we consider only data since 1948. We (cid:12)nd most support for the view that political e(cid:11)ects on the economy, if they exist,aresmallanddi(cid:14)cultto measurewithcon(cid:12)dence. Thus, omittedvariableand simultaneousequationbiasappeartobealargeprobleminthepoliticaleconometric work (cid:12)nding large and systematic e(cid:11)ects. We (cid:12)nd some evidence that the party in power a(cid:11)ects output growth|the party in power seems to help some in accounting for the recession that has historically followed the election of Republicans. This evidence is very weak and not robust to various reasonable alterations of the speci(cid:12)cation. We (cid:12)nd almost no support for the view that any political e(cid:11)ects operate through monetary policy. Thus, the reaction of interest rates, money, and in(cid:13)ation to the party in power appear economically small and statistically insigni(cid:12)cant. The results cast strong doubt on earlier results in the political econometric lit- 8 Good a priori reasoning and econometric work support the view that there is little hope of (cid:12)tting a stable,linearmacroeconomic modelfor the entire century. 3

erature that ignore the sources of bias discussed above. Because of the limitations outlinedabove,theseresultsshouldbeinterpretedcautiously. Areasonableinterpretation is that political e(cid:11)ects are not large and systematic enough to be measured using relatively standard methods and macroeconomic information sets. Finding those e(cid:11)ects will either require putting more structure on the problem in the form of a priori assumptions, using more powerful statistical methods, or bringing di(cid:11)erent information to bear. In section 1, we review the theories and present a preliminary look at the data. Section 2 lays out our theoretical framework; Section 3 presents evidence from multivariate systems, and Section 4 concludes. 1 A preliminary look at the theories and data 1.1 Political business cycle theories There is a wide range of theories about the links between politics and the economy. Alesina and Rosenthal [1995] provide a good summary. For our purpose, a coarse characterization will be su(cid:14)cient. The (cid:12)rst important distinction is whether the theories focus on partisan di(cid:11)erences in the control of instruments or on behavior that should be common to the parties. For example, Nordhaus’s [1975] theory posits that either party in power should stimulate the economy before elections to enhance its election prospects by a myopic electorate. Partisan theories [such as Hibbs, l977] posit that the parties should control the instruments according to di(cid:11)erent objective functions, implying that the instruments and the economy should behave di(cid:11)erently under the two parties. The second distinction is whether the stochastic nature of election outcomes is a crucial part of the theory. Under both Nordhaus’s and Hibbs’s theories, the party in power is stochastic, but this feature plays no central role in the predictions of the theory. In contrast, in Alesina’s [1987] rational partisan theory, crucial implications turn on the e(cid:11)ects of surprise changes in policy that come from election surprises. 4

In particular, the parties implement di(cid:11)erent preferred average rates of in(cid:13)ation. A surprise election outcome in favor of Democrats leads to a surprise increase in in(cid:13)ation, which has the textbook stimulative e(cid:11)ects on the economy. From this characterization, a few gross facts are clear. We wish to have an approach that allows us to identify systematic changesin instruments and outcomes that vary with the deterministic election cycle. These e(cid:11)ects may or may not vary by party. We also want to have an approach that allows us to identify the e(cid:11)ects of election surprises, where these e(cid:11)ects may also vary by party. Below, we derive an econometric framework that will accommodate these desiderata. Althoughourframeworkwouldallowit,wedonotspecifyformalrepresentations of the theories and test them directly. Formal models must abstract from myriad issues presentin reality,and any direct test of those modelmust, thereby, be viewed asatestofthe jointhypothesesimbeddedinthecoreoftheformalmodelalongwith the auxiliary simplifying assumptions. Such tests shed little light on whether the core insights of the model are valid. Below we derive and test hypotheses associated with core implications of models. Further work sharpening these hypotheses might well be warranted. 1.2 A preliminary look at the real variables We begin our examination of the data with some graphs that highlight the associations between political and economic variables and suggest pitfalls in trying to identify political e(cid:11)ects on the economy. We focus on the period 1948q1 to 1995q2, which includes thirteen presidential terms (the (cid:12)rst and last terms are incomplete). Of the thirteen terms, 7 are Republican; the parties came in the order D D R R D D R R D R R R D. Throughout, we take no account of the fact that Kennedy and Nixon served incomplete terms. Figure 1 shows the average unemployment rate by party beginning 3 quarters before taking o(cid:14)ce through the end of the 4-year cycle. The data are averaged over 5

9 the presidential terms in our sample. On entering the White House, Republicans are greeted by an unemployment rate about one percentage point lower than are Democrats; the rate begins to rise sharply in the third quarter in o(cid:14)ce and, by the sixth quarter, has begun to stabilize at a level that is more than a percentage point higher than on inauguration day. The opposite pattern is shown under Democrats, the rate falling by more than a percentage point before stabilizing. Because averages can be dominated by outliers, we examine the median and interquartile range of the unemployment rate under the two parties (Figure 2). For each party, the time-series line connects the median, and the vertical bar gives the interquartile range. The interquartile range for the Democrats is displaced slightly to the left of the relevant quarter, while that for the Republicans is displaced to the right. The di(cid:11)erences in the unemployment rates under the two parties are quite broad based. The median Republican shows a lower unemployment rate than th the 25 percentile Democrat until the cross-over point in quarter 5, after which th the median Republican is above the 75 percentile Democrat. In 7 of the (cid:12)nal 9 th quaters of term, the 75 percentile Democrat has a lower unemployment rate than th the 25 percentile Republican. As the unemployment rate moves sharply in opposite directions in quarters 3{5 under the two parties, growth in hours of employment diverges widely (Figure 3). This widely divergent hours growth implies widely divergent output growth (Figure 10 4). Averageoutputgrowthduringquarters3and4is5.3percentunderDemocrats and -1.6 percent under Republicans|a di(cid:11)erence of about seven percentage points. The negative growth early in Republican administrations generally marks an NBER business cycle peak. We noted above that 6 of 7 Republicans and no Democrats had recessions start in the (cid:12)rst 6 quarters of term. Only two of 6 Democrats had recessions begin in their terms: 1948q4 came at the end of Tru- 9 Throughout, the quarters of the term are numbered from zero in the quarter of regularly scheduled inaugurations. 10 Analogous pictures show asimilarlyrobust patterninconsumption(with amediandi(cid:11)erence about half as large as for income) and investment (with a median di(cid:11)erence over twice as large), but not in government spending. 6

man’s (cid:12)rst term; 1980q1 began Carter’s credit control recession. 1.3 Nominal variables Republicanscomeintoo(cid:14)cetohigherinterestratesthanDemocrats; ratesfallunder Republicans as the economy slows, and rise under Democrats as it accelerates. In the second half, the di(cid:11)erences between the parties are not robust (Figure 5). Note that this pattern in the interest rates is about what one might have expected from looking at the real variables, without knowing that the sample has been sliced by party. Of course, this fact would be missed in a univariate system. The median in(cid:13)ation rate is higher when Republicans come into o(cid:14)ce and falls slightly over the term; the rate under Democrats rises during the term (Figure 6). After quarter 5, however, the di(cid:11)erences are not robust. While in(cid:13)ation is rising steadily under Democrats in quarters 0 through 6, the M2 growth rate is falling (Figure 7). 1.4 Implications for more formal analysis While these results are tantalizing, they both demand careful statistical analysis and portend complex problems for the analyst. Two facts are clear: First, indicator variables for the party in power and stage in the presidential term are correlated with a wide range of nominal and real economic variables. Second, economic variables|particularly unemployment and interest rates|at the time of the election are correlated with the subsequent election outcome. These two facts can be summed up in the claim that there is a rich simultaneity with both rich lead and lag relations among the macroeconomic and political variables. Thisrichsimultaneitysuggeststhatomittedvariablebiasandsimultaneousequationsbias willbe major concernsin anyworkrelating endogenous politicalvariables to macroeconomic variables. While the association between party in power and the unemployment rate might be due to political manipulation of policy, it might just as well be due to causality from the economy to the party in power. Suppose the 7

electorate chooses a Republican to run the country when the unemployment rate is belowthe naturalrate and Democrats whenitisabove. If theunemploymentrate is stationary,belowaverageunemploymentrates are naturallyfollowedbyrisingrates. ThismightaccountforFigure1withoutrelyingonanycausalityfrompoliticstothe economy. On the other hand, if this reverse causation from economy to politics isto account for Figure 1, it must be the case that, in e(cid:11)ect, the electorate can forecast one percentage point rises in the unemployment rate about a year in advance with great accuracy. Otherwise, they could not reliably put Republicans in o(cid:14)ce when such a rise is coming. It is not clear that unemployment rate rises are su(cid:14)ciently predictable. Asomewhatricherillustrationofproblemsmeasuringpoliticale(cid:11)ectscomesfrom recent empirical work by Alesina and Rosenthal [1995] explaining in(cid:13)ation in terms of its own past and the party in power. In support of the claim that Democrats implement higher rates of in(cid:13)ation than Republicans, Alesina and Rosenthal cite the following regression: It= 1:2 + 1:3 It(cid:0)1(cid:0) 0:3 It(cid:0)2+ 0:1 It(cid:0)3 (5:4) (17:4) ((cid:0)2:9) (0:7) (cid:0) 0:2 It(cid:0)4(cid:0) 0:6 BW(cid:0) 0:3 Repubt(cid:0)3+ 0:7 oilt ((cid:0)2:4) ((cid:0)4) ((cid:0)2:5) (2:9) (We produced these results, which very nearly replicate those of Alesinaand Rosenthal.) In this regression, I is annual CPI in(cid:13)ation, BW is a dummy variable that is one under the Bretton Woods system (through 1971) and zero otherwise, Repub is a variable that is one under Republicans and zero otherwise, and oil is a dummy variablethatisoneduring1973q3-1974q4and1979q4{1980q4. Thesampleperiodis 1949q1{1991q4; t-statistics are under the coe(cid:14)cients. Alesina and Rosenthal interpretthecoe(cid:14)cientonRepubasevidencethatRepublicansimplementlowerin(cid:13)ation rates than Democrats, and that the e(cid:11)ect begins to operate after three quarters in o(cid:14)ce. Since the Repub dummy is endogenous we might worry that it is standing in for other macroeconomic variables that are important in determining in(cid:13)ation. For 8

example, if the evolution of in(cid:13)ation depends on labor market tightness, as it does in a wide range of theoretical models, then Figures 1 and 2 suggest that the Repub dummy may enter this regression signi(cid:12)cantly acting as a crude proxy for current 11 and past labor market tightness. If we suspect that the Repub dummy is a proxy th for economic variables then we might expect 16 order serial correlation in this regression|the macro variables move smoothly, but the dummy is constant for 16 th quarters at a time. The LM test for absence of 16 order serial correlation rejects at less than the 0.01 percent level. We add enough lags of in(cid:13)ation to eliminate the symptoms of serial correlation (13 lags) and include one lag of growth in hours to allow for an association between labor market tightness and in(cid:13)ation. In this regression, hours enters the regression very signi(cid:12)cantly (coe(cid:14)cient 0.04, t-statistic 3.8) and the e(cid:11)ect of the political dummy falls to one-sixth its previous magnitude and becomes statistically insigni(cid:12)cant (coe(cid:14)cient -0.05, t-statistic -0.7). We present this brief result not because this is our preferred in(cid:13)ation equation, but to illustrate a general and important point: Due to simultaneous equations and omitted variable bias, there is little a priori reason to expect robustness of political business cycle results that are based on a small macroeconomic information set and that treat political variables as exogenous. Inthefollowingsection,welayoutaframeworkfordealingwiththeseproblems. 2 Speci(cid:12)cation andidenti(cid:12)cation of asimultaneoussystem ofmacroeconomic and presidential cycle variables The political variable considered here provides some special opportunities and specialproblems,eachofwhichstemfromthefactthatthepartyinpowertakesononly two values and changes only every 16 quarters. The framework we adopt is quite general, but the particular implementation we choose is designed to be the simplest 11 Note that using lagged values as instruments might solve the simultaneity problem (if the instruments are valid), but will not solve the problem stemming from omitting serially correlated variables. 9

possible generalization of the identi(cid:12)ed vector autoregression (VAR) approach. We choose this implementation because it provides both a simple and a natural starting point given the current prominence of VAR work for identifying the e(cid:11)ects of monetary policy. 2.1 Deriving and identifying the standard VAR Take one political variable zt and an (n(cid:2)1) vector of macroeconomic variables Xt, and posit that these variables are determined by 2 zt 3 = f(z~t(cid:0)1;X~t(cid:0)1;et) (1) 6 Xt 7 4 5 0 0 0 0 where X~t = (Xt;Xt(cid:0)1;:::;Xt(cid:0)p), z~t is similarly de(cid:12)ned, and et is a vector of exogenous shocks. For any plausible macroeconomic model, f is nonlinear, but we 12 typically take a linear approximation to the model to give 2 zt 3 2 z~t(cid:0)1 3 = (cid:11)+B +"t (2) 6 Xt 7 6 X~t(cid:0)1 7 4 5 4 5 Equation (2) is a vector autoregression (VAR) and is often written, 2 zt 3 B(L) = (cid:11)+"t (3) 6 Xt 7 4 5 where Lzt = zt(cid:0)1 and B(L) is a matrix polynomial in L. The nonlinear model is (generically) identi(cid:12)ed [McManus, 1992], but in linearizing we take on an identi(cid:12)cation problem arising from the fact that (3) is observationally equivalent to 0 0 (cid:3) (cid:3) C(L)[zt;Xt] = c+"t where C(L) = CB(L), c = C(cid:11) and "t = C"t, and C is full rank. Each such C gives a di(cid:11)erent identi(cid:12)cation of the system, and the response of the economy to a shock, say, in the political variable is di(cid:11)erent in each. 12 We attempt to choose the approximation and some transform of the variables so that "t has constantmeanandvariance,andisseriallyuncorrelated. Theadequacyofthelinearapproximation is testable, and the evidence for nonlinearities in U.S. macroeconomic data has not been strong enough to justifyabandoningthe linear framework. 10

In this paper, we only attempt to identifythe e(cid:11)ects on Xt ofan exogenous shift in the zt equation. Thus, we only identify the e(cid:11)ects of an election shock. If we ignore the special properties of the zt variable, two assumptions will be su(cid:14)cient for identi(cid:12)cation. First, assume that in each quarter, the value of the political variable zt isdetermined before any other variable. Thissuggests a block recursive structure of the economy as in Sims [1980a], in which Xt does not enter the zt equation contemporaneously. Second, assume that "1t is orthogonal to all the other "s at all leads and lags. That is, we place zt (cid:12)rst in a block recursive ordering of the VAR and assume that "1t is orthogonal to the remaining "jt at t. Under these assumptions, we can estimate the model and calculate the dynamic e(cid:11)ects of a one-period exogenous shift in the zt equation. The dynamic e(cid:11)ects of th suchachangeonthej X variablearesummarizedinanimpulseresponsefunction, which is the sequence of numbers, aji = (cid:20)@Xjt+i=@"1t i= 0;1;::: (4) 13 where (cid:20) is an arbitrary scaling re(cid:13)ecting the size of the presumed exogenous shift. The scheme laid out in the following section is the natural generalization of this scheme to allow for the special nature of zt. 2.2 Taking account of the presidential cycle variable Our zt is an indicator for party in power and is equal to 1 if a Democrat is in the White House at t and is zero otherwise. Thus, zt is discrete and changes only every 16 periods in quarterly data. To deal with these special features, de(cid:12)ne q(t) 2 f0;:::;15g as the time t quarter of the presidential term, numbering from zero. Now assume that zt evolves according to: 13 th The coe(cid:14)cient aji (up to a proportionality factor) is given by (1;j) element of the matrix 1 m (cid:0)1 Ai, where A(L)= m=0AmL and A(L)=B(L) . P 11

i) Et(cid:0)1zt = zt for all t;q(t); ii) zt = zt(cid:0)1 if q(t)6= 0, and iii) Et(cid:0)2[zt] = (cid:8)(X~ t(cid:0)2;zt(cid:0)2) if q(t)= 0: Thus, the party in power may change every 16 quarters; it is picked based on the economy 2 quarters earlier; and the outcome of the election is known one quarter before the president takes power. Further, the probability that a Democrat will win, based on the macroeconomic variables and the current party in power, is (cid:8)(X~t(cid:0)2;zt(cid:0)2). These assumptions would be exactly appropriate if the election were held October 1 and the president took power January 1. We will model as if this slightly modi(cid:12)ed timing were correct. Now replace the general representation of the economy in (1) with, 8 zt(cid:0)1 if q(t) 6=0 zt = > (5) < (cid:8)(X~t(cid:0)2;zt(cid:0)2)+ut(cid:0)1 otherwise > : (cid:3) Xt = G(q(t);zt+1;zt;X~t(cid:0)1;(cid:23)t) In G, we allow the presidential party this quarter and next quarter to a(cid:11)ect this quarter’s economy. Of course, both of these values are always known, and the two 14 can only di(cid:11)er in the (cid:12)nal quarter of term. Why is q(t) an argument of G? It will obviously be relevant under Nordhaus-style theories. It may also be relevant under rational partisan theories since political uncertainty is not homoskedastic through time. Macro variables will fail to be homoskedastic if political uncertainty has important e(cid:11)ects on the economy. Of course, some form of time homogeneity must be imposed. Notice that (zt+1;zt;q(t)) form a triplet of discrete variables. It is natural to view this triplet as the state of the political economy at t and to assume that the economy is homogeneous conditional on the political state. Call the set of the 34 possible values of 14 Further,lagsofzt could beincluded without changingthe analysis. 12

15 this state variable (cid:10). We can re-write the X equation as (cid:3) Xt = Gs(t)(X~t(cid:0)1;(cid:23)t) where s(t) 2 (cid:10) is the political state at t. The evolution of the state variable is governed by (5) and the exogenous quarter-of-term variable. After linearizing, the analog of the macroeconomic VAR in (2) is now Xt = (cid:11)s(t)+Bs(t)X~t(cid:0)1+(cid:23)t This expression says that the VAR representation for Xt has di(cid:11)erent (cid:11) and B coe(cid:14)cients depending on the value of the political state. One can also write system as, Xt = (cid:11)!d!;t+ B!d!;tX~ t(cid:0)1+(cid:23)t (6) X X !2(cid:10) !2(cid:10) where d!;t = 1 if ! = s(t)2 (cid:10) and d!;t is zero otherwise. Thus, d!;t is an indicator that isone ifthe stateatt is! andzerootherwise. Equation(6)saysthatwesimply need to take the standard macroeconomic VAR and augment it with political state dummy variables interacted with the intercept and with all the slope coe(cid:14)cients. These dummy variables are not exogenous, but since their values at t are known at t(cid:0)1, they are predetermined. Of course, the system still involves the nonlinear equation for zt. We show below that we can do the inference we wish to do simply by estimating the augmented VAR and need not estimate the nonlinear part of the model. The form of the augmented VAR is similar to the model estimated by Ellis and Thoma [1991]. Ellis and Thoma derived no structural interpretation on the form, however, and considered a very limited speci(cid:12)cation, and, hence, face the 16 identi(cid:12)cation and omitted variable problems common in the rest of the literature. 15 Forthe(cid:12)rst15quartersofthetermzt =zt+1 andcantakeontwovalues;inthe(cid:12)nalquarter, there are four possiblevalues. 16 Morespeci(cid:12)cally,EllisandThomatreatedthedvariablesasexogenous;theyexaminedimpulse responsesderivedfromanarbitraryrecursiveorderingoftheVARattributingnothingtotheshock inthedvariables. TheirVARincludedinstrumentsofpolicy,butnomacroeconomicvariablesthat may be importantin determining those instruments. 13

It is clearly impossible to estimate the coe(cid:14)cients of (6) in the post-War macro data. The summation over(cid:10) involves 34 possible values of the state variable. Thus, the political model involves 34 times the (already large) number of coe(cid:14)cients in a standard macroeconomic VAR. This parsimony problem is similar to the one faced in most macro applications, and we deal with it in a standard manner as outlined below. 2.3 Hypotheses of interest The sketch of political theories above suggests three nested hypotheses regarding the augmented VAR model. At the coarsest level, we are interested in whether any political variables need to be included in macroeconomic models. In the augmented VARthisisamatter ofwhethercoe(cid:14)cientscanbeconstrainedtobeconstantacross the political state, s(t). Thus, the hypothesis that the political state is irrelevant is: H0 : (cid:11)m = (cid:11)n and Bm = Bn for all m;n 2 (cid:10) IfH0 doesnothold,wecanconsidertheNordhaustheoryinwhicheconomicactivity mayvarywithquarteroftermbutnotwiththevalueofthepartyinpower. De(cid:12)ning Q(m) as the quarter of term in state m, the hypothesis is H1 :(cid:11)m = (cid:11)n and Bm = Bn whenever Q(m)= Q(n) Under H1, we (cid:12)nd quarter-of-term e(cid:11)ects, but no partisan di(cid:11)erences. Rejection of H0 and H1 is evidence of partisan e(cid:11)ects: H2 :(cid:11)m 6= (cid:11)n or Bm 6= Bn for some m;n with Q(m)= Q(n) Sorting out whether these e(cid:11)ects are due to surprise e(cid:11)ects, as in the rational partisan theory, or due to di(cid:11)erences in policy that do not have their e(cid:11)ects through surprises does not involve simple restrictions on the coe(cid:14)cients. We can, however, look at the impulse response to election shocks to see if the pattern favors a partic- 17 ular interpretation. 17 Even stronger evidence would come from seeing e(cid:11)ects that di(cid:11)er with the degree of election surprise. Since we donotestimate the z equation,we cannotpursuethisangle here. 14

2.4 Identi(cid:12)cation in the augmented model Before we cantest the hypothesesabove, we mustbeable toform estimatesthatare arguablyfreefromsimultaneitybias. Wenowhaveamodelcomprisedbyanonlinear model for zt, which determines the dm;t variables, and (6), an augmented VAR. As in the discussion of the standard case outlined above, we rely on a block recursive structure of zt and Xt and an orthogonality assumption. Under the assumed timing of events, zt belongs (cid:12)rst in a causal ordering of the model. We must also assume that the election shock ut is uncorrelated with (cid:23)s for all s;t. Economically, this amounts to the decision to attribute to the political variable any macroeconomic outcome that is correlated with the election surprise. This assumption would be inappropriate if variables other than X~t that determine the outcome of the election at the end of period t also directly a(cid:11)ect the economy. Suppose, for example, that news arrives in the third quarter of an electionyear of the breakdown in some peace talks. The breakdown might well alter the re-election prospects of the president 18 and also directly a(cid:11)ect the economy through its e(cid:11)ects on military procurements. While it is easy to come up with examples like this, it is more di(cid:14)cult to come up with such examples that systematically favor one party or the other, thus inducing a spurious correlation between party and economy. By considering a large range of economic variables, and by considering the sensitivity of the results to changes in the information set, we hope to minimize this risk. Validity of these assumptions is su(cid:14)cient to allow us to estimate consistently the augmented VAR for Xt while including zt and zt+1 (through the d variables) in the equations. The (cid:12)nal issues concerns calculating the response of macro variables to the election shock. 18 FaustandLeeper[1993],buildingonHansenandSargent[1991b]andMarcet[1991],question assumptions about the correlation of structural shocks based on problems with time-aggregated data. Given the nature of the zt variable, these problems would not existhere if the election were in fact held on Oct. 1, andwe believe that thesee(cid:11)ectswill be smallin practice. 15

2.5 The impulse response to election shocks Because the zt equation is nonlinear, several complications arise regarding what we mean by an impulse response function [see, e.g., Gallant, et al., 1993]. Several of these are simpli(cid:12)ed by limiting ourselves to a 16 quarter impulse response, starting in quarter 15 of a presidential term, which we also call quarter -1 of the following term. Electionshocks happen only at the beginningof quarter-1, implyingthatthe process for zt and zt+1 (and, hence, for the dt variables) is conditionally deterministic for quarters 0 through 14 of the term. The next shock happens at the beginning of quarter 15 (=-1). Thus, the 16 quarter impulse response of the political state to an election shock at quarter -1 is trivial and depends on no unknown parameters. Because no variables feed back on the zt equation for 16 quarters, we can also compute the 16 quarter impulse response of Xt to an election shock without estimating the nonlinear zt equation. What we report below for the impulse response of Xjt+i to an election shock at the beginning of time t is the sequence of numbers, @Et (cid:0)1Xjt+i aji = 0:10 i= 0;1;:::;15: (7) @Et (cid:0)1ut where ut is the exogenous shock to the election outcome. Note that a rise in Et (cid:0)1ut by 1 percentage point equals a 1 percentage point change in the probability that a Democrat will be elected. Thus, while the standard impulse response gives the response of Xj to a one standard deviation change in some shock, the modi(cid:12)ed impulse response gives the response of Xj to a 10 percentage point change in the 19 public’s subjectiveprobabilitythataDemocrat willtake theWhite Houseatt+1. The impulse response is easily calculated (see the Appendix). 19 Much of the nonlinearity of the problem has disappeared in this formulation. It remains in the zt equation,however. Thus, given the historyup throught(cid:0) 1, only oneof two shocks ut can happen: (cid:0) (cid:8)or1 (cid:0) (cid:8). In orderto estimatetheactualshocksthatcouldhave happenedat(cid:0) 1,we mustestimatethezt equation. Further,thereisnosimplenotionofvariancedecompositioninthis setting. 16

3 Application: A multivariate model with political e(cid:11)ects In this section, we estimate the augmented VAR. Our goals are to assess the importance of political variables in traditional macro models and to assess the role of the monetarypolicy channelin accounting for any political e(cid:11)ects. Thus, we begin with amuchstudiedVARincludingfourmacroeconomicvariables: three-monthTreasury bill rate, M2, the CPI,and GNP. These variables will be referred to as R, M, P, and Yrespectively. TheRMPYsystemisanaturalstartingpoint. Someversion ofthese variables is at the core of the Friedman and Schwartz [1963] analysis and variations of this basic VAR have been investigated extensively in assessing monetary policy e(cid:11)ects [e.g., Sims 1980b, Sims 1994]. The baseline VAR for RMPY has all variables in levels and all variables except the interest rate in logarithms. The VAR includes a constant and three seasonal dummies. The theoretical framework suggests that we augment the standard VAR withslopeandinterceptdummiesforeachquarterofthetermundereachparty. This modelistoopro(cid:13)igatelyparameterizedtobeofinterestandiseventoogeneraltobe the starting point of a Hendry-style general-to-speci(cid:12)c model search [Hendry, 1995]. Instead we start with a fairly general model, from which we derive a parsimonious baseline model. After demonstrating the baseline results, we consider a number of changes to the baseline model in order to assess robustness. The estimation period is 1953q2 to 1995q2, and the lag length of six for the macro variables was selected by serial correlation tests. For our starting point we considered only political intercept dummies. From the possible 34 dummies for th th the -1 through 15 quarters of term under each party, only a subset are linearly independent given the constant and seasonal dummies. We began with a maximal set of these dummies, which were not jointly signi(cid:12)cant at the 10 percent level, and very few of the individual coe(cid:14)cients were signi(cid:12)cant at the 10 percent level. Becauseofthelargenumberofdummyvariablesinthisregression,thetest probably has low power to reject the hypothesis of no e(cid:11)ects. Thus, we tried replacing the quarter-of-termvariablesforoneofthepartiesatatimewithyear-of-termdummies. 17

Using Republican year-of-term variables, the the political variables were not jointly signi(cid:12)cant. UsingDemocraticyear-of-termvariables,alltheofyear-of-termvariables and most of the Republican quarter-of-term variables were signi(cid:12)cant at the 1 or 5 percent level. We took this as our baseline model. Thus, the baseline augmented VAR includes in four intercept dummies, dy1 through dy4, that are zero except in the subscripted year of a Democratic administration; 15 dummies,rq0;:::;rq14, that are zeroexceptinthesubscripted quarterof Republican administrations; and dq(cid:0)1 andrq(cid:0)1 whicharezeroexceptin the quarter before a Democrat or Republican take o(cid:14)ce, respectively. The model passes serial correlation tests (for (cid:12)rst, fourth, and sixteenth order correlation) and heteroskedasticity tests (Table 1, column i). Given the minimal attempttodatamineforthebestsetofpoliticalvariables,wetakethebaselineresults as relatively strong statistical evidence against H0|the irrelevance of politics|in thismodel. Fiveofthe 16Republicandummiesaresigni(cid:12)cantatthe 1percentlevel; an additional seven are signi(cid:12)cant at the 5 percent level. The political variables are jointly signi(cid:12)cant at the 7 percent level. The test of H1, the absence of partisan e(cid:11)ects, rejects at the 3 percent level. As the point estimates of dummy variables are almost impossible to interpret in a dynamic system, we turn now to the impulse response functions, which provide a much more interpretable picture of the economic signi(cid:12)cance of the political e(cid:11)ects (Figure8). Thepointestimateoftheimpulseresponseisindicatedbythecircle;and the lines are empirical 95 percent coverage intervals calculated using the Sims-Zha 20 [1994] Bayesian bootstrap. The 10 percentage point rise in Democratic election prospects generates a rise of over half a percentage point in output growth over the (cid:12)rst year in o(cid:14)ce, which dissipatesbytheendofthesecondyear. Thisseemslikealargee(cid:11)ect,say,ofelecting a Democrat when the ex ante probability of doing so was 90 percent. The expected 20 The procedure was implemented in Gauss by the authors. The procedure uses an ignorance priordescribedbySimsandZha. As SimsandZhanote,there iscurrentlynoentirelysatisfactory way to calculateclassical con(cid:12)denceintervalson impulse responses. 18

version i ii iii variable rq(cid:0)1 * rq0 ** ** rq1 *** ** * rq2 *** * rq3 * *** rq4 *** *** ** rq5 *** ** rq6 ** ** rq7 ** ** rq8 ** rq9 *** * ** rq10 *** ** rq11 ** rq12 * rq13 *** ** rq14 ** dq(cid:0)1 dy1 ** dy2 ** dy3 ** dy4 *** BW ** zt(cid:2)BW *** D1980q2 *** *** D1980q3 D1980q4 *** *** AR(1{16) ** *** Hetero Notes: ***, **, *,indicatesigni(cid:12)canceatthe1, 5,and10percentlevels,respectively. The signi(cid:12)cance tests are F tests for inclusion of all terms involving the variable in the model. The AR test reported is the standard LM test. The heteroskedasticity test essentially involves a regression of the squared errors from the model on the levels and squares of all the variables and provides a test of signi(cid:12)cance of the variables in this second stage regression. The signi(cid:12)cance levels are all for the F form of the statistic. Tests were reported by PC-FIML and are further explained in Doornik and Hendry [1994]. Table 1: Signi(cid:12)cance of political state variables in the baseline model 19

3-month Treasury bill rate follows the pattern that might be expected if the output e(cid:11)ectwereduetoapersistentsurpriseriseinmoneygrowth: interestrates fallforat least a year and then rise by a similar amount and stay higher. The e(cid:11)ect, however, is quite small, less than 25 basis points in each direction. M2 growth shows only a small and brief rise, and in(cid:13)ation shows similarly small e(cid:11)ects and no economically signi(cid:12)cant pattern at all. Thus, while the basic pattern of interest rates is roughly consistent with the election shock generating a monetary policy surprise, the e(cid:11)ects on output are probably larger than one would expect for such a small change in interest rates, and money growth and in(cid:13)ation are not very supportive of the story. These results are much less supportive of the partisan model operating through moneysupplyshocksthan those ofAlesinaandSachs[1988]and Alesinaand Rosenthal [1995]. Those results involved interpreting political dummy variables from univariate regressions directly without resolving the dynamic interactions among variables or attempting to isolate the surprise element that is central to the theory. Takenatfacevalue,theresultspresentapuzzle. Figures1through9andTable1 show evidence of strong e(cid:11)ects of party on aggregate measures of economic activity. The channels through which policy causes these e(cid:11)ects, however, are not clear from any of these sources of information. One possibility is that the e(cid:11)ects operate through (cid:12)scal policy. This possibility seems unlikely to account for the recession that quickly follows Republican elections, since (cid:12)scal policy is slow to change. In any case, we investigate this channel in a separate paper [Faust and Irons, 1996], (cid:12)nding no clear evidence that (cid:12)scal policy accounts for the observed output e(cid:11)ects. An alternative possibility is that the political variables are proxying for some other source of variation in output. If this is the case, we might expect the results to change when the speci(cid:12)cation of the model is altered. Thus, we turn to the robustness of the baseline results. We investigate 5 basic VARs: 1) the baseline model, 2) substitute M1 for M2 in the baseline, 3) add log of hours of work to the baseline, 4) add the log of the producerprice index for intermediatematerialsto the 20

21 baseline, 5) add the producer price index for sensitive materials to the baseline. Model2iswarrantedbecauseresultsareoftensensitivetowhichmonetaryaggregate ischosen. Model3issuggestedbythee(cid:11)ectofhoursgrowthintheAlesinaRosenthal regression. Models 4 and 5 are motivated by work [e.g., Sims, 1992, Sims and Zha, 1994]arguingthatcommoditypricescansubstantiallyaltertheresultsfromRMPYtype VARs. Three versions of each of the 5 models are run. Version i involves no economic dummy variables. Version ii is motivated by the fact that in the baseline model, 1980q2 is a large outlier. This is the quarter in which credit controls were enacted 22 leadingtothesharpest,shortestrecessioninU.S.history. Itisclearthatwecannot model the process leading to the imposition of credit controls, and it is important to investigate the sensitivity of the results to this episode. Thus, version ii adds separate dummies, labelled D1980q2 through D1980q4, for the (cid:12)nal three quarters of 1980. Version iii includes the credit control dummies and a dummy variable, BW, that is one up until 1972. Alesina and Rosenthal use this variable because the Bretton Woods system may have constrained in(cid:13)ation behavior before 1972. Many other important changes in the macroeconomy occurred at about this time: the onset of oil shocks and the productivity slowdown are two of the most important. To allow for the possibility that the parties had di(cid:11)erent mean in(cid:13)ation rates in the post-Bretton Woods system, but the same rate in under Bretton Woods, we also include the BW dummy interacted with the zt dummy. The importance of these dummy variables is clear from their e(cid:11)ects on the signi(cid:12)cance of the political variables in the baseline model (Table 1, columns ii and iii). Forexample, inversionsii andiii,noneoftheDemocratvariablesissigni(cid:12)cant at the 10 percent level and only one of the Republican variables is signi(cid:12)cant at the 21 Wealsotriedavarietyofothermodels. Weinteractedalloftheslopecoe(cid:14)cientsinthemodel with the zt dummy. Only two of the 24 slope dummies made a signi(cid:12)cant contribution to the systematthe(cid:12)vepercentlevel. Variousothercon(cid:12)gurationsofslopedummieswerealsotriedwith noclearsupportfortheirinclusion. Wealsoconsideredavariablemeasuringtherelativepriceofoil and the Alesina-Rosenthal oil shock dummy described above. Neither of these added signi(cid:12)cantly to the explanatory power of the model. 22 This quarter involved the largest quarterly drop in R and Y in the sample. See the 1981 Economic Report ofthePresident for an account of thisperiod. 21

1 percent level. Table 2 summarizes the statistical evidence for important political e(cid:11)ects in the 23 15 models. TheDemocraticandRepublicanpoliticalvariables|thesamepolitical variables as in Table 1|are statistically signi(cid:12)cant at the 10 percent level or less in only about 1/3 of the models (rows labelled H0). In version ii (all models) and model 3 (all versions), the political e(cid:11)ects are never signi(cid:12)cant. Wesuspectthatthepoliticale(cid:11)ectsmaystillbeoverparameterizedinthismodel, and that in particular, the only important political e(cid:11)ects come in the (cid:12)rst half of Republican administrations. The rows labelled HX test the hypothesis that the all the Democratic variablesand the Republican variablesforquarter 9and beyond are zero. This hypothesis is only rejected once at the 10 percent level. Thus, it appears that Democratic presidential terms and the second half of Republican terms are similar and that the evidence of political e(cid:11)ects comes from the recessions that coincide with the election of Republicans. After removing the political variables for all but quarters -1 through 8 for Republicans, we can test whether the remaining Republican e(cid:11)ects are signi(cid:12)cant. This is a test of both H0 and H1, since any e(cid:11)ect that is found involves a partisan di(cid:11)erence. We label this hypothesis H0;1. After limiting the setof included political dummies based on prior tests, the remaining political dummies for Republican for quarters -1 through 8 are jointly signi(cid:12)cant at the 10 percent level or better in 10 of 15 models. Of course, since we have tailored the set of variables to maximize the measured e(cid:11)ects, we should make our standards more stringent. In only 2 cases are the dummies signi(cid:12)cant at the 1 percent level. The (cid:12)rst half Republican e(cid:11)ects are also insigni(cid:12)cant in all versions of model 3 (involving hours). Thus, even after tailoring the set of dummy variables, the evidence in favor of political e(cid:11)ects is mixed, at best. A more economically interpretable view of the VARs is given in Table 3, which summarizes impulse response functions for the 15 models. For each model and 23 In all of the models that include economic dummies, the test of the joint hypothesis that all the economic dummies arezero rejectsatless than the 1 percent level. 22

model hypothesis 1 2 3 4 5 i: No economic dummies H0 0.07 0.30 0.26 0.11 0.04 H0;1 0.04 0.09 0.12 0.01 0.07 HX 0.35 0.70 0.55 0.66 0.15 ii : Credit Control H0 0.20 0.41 0.39 0.25 0.12 H0;1 0.05 0.12 0.20 0.03 0.13 HX 0.63 0.79 0.63 0.82 0.28 iii : Credit Control & Bretton Woods H0 0.04 0.18 0.22 0.05 0.01 H0;1 0.03 0.04 0.16 0.01 0.05 HX 0.24 0.70 0.43 0.58 0.04 Notes: H0: all political dummies are jointly signi(cid:12)cant. H0;1: after imposing HX, all remaining political dummies (rq(cid:0)1;:::;rq8) are zero. HX: all political dummies exceptrq(cid:0)1;:::;rq8 arezero. Thenumbersreportedaremarginalsigni(cid:12)cancelevels, the tests are as described in the notes to Table 1. , , , indicate signi(cid:12)cance at the 1, 5, and 10 percent levels, respectively. Table 2: Signi(cid:12)cance of political state variables 5 models 23

version, the Table gives the averaged impulse response of Y, R, M, and P to an election shockoverthe (cid:12)rst and second halves ofthe presidential term. Ifsigni(cid:12)cant output e(cid:11)ects operate through monetary policy, we would expect to see positive output growth e(cid:11)ects in the (cid:12)rst half, positive money and in(cid:13)ation e(cid:11)ects in both halves, and perhaps negative then positive interest rate e(cid:11)ects. In model 3, the e(cid:11)ects are statistically insigni(cid:12)cant and the point estimates are 24 not consistent with the rational partisan story. The (cid:12)rst-half output e(cid:11)ects are smaller than the second-half e(cid:11)ects which may be positive or negative, depending on the verison. The money growth e(cid:11)ects are negative, and the in(cid:13)ation e(cid:11)ects are also negative in two versions. Only the interest rate e(cid:11)ect is as expected. The (cid:12)rst half output e(cid:11)ects are largest and most statistically signi(cid:12)cant in versions i and iii. In these versions, however, the interest rate and money growth e(cid:11)ects are quite small, relative to the output e(cid:11)ects. Further, the in(cid:13)ation e(cid:11)ects are small, with point estimates that are often negative. Thus, while versions i or iii may support important (cid:12)rst-half output e(cid:11)ects, they provide no support for the view that these e(cid:11)ects operate through monetary policy. Thepatternofpointestimatesinversionii(especiallymodel1)isprobablymost supportiveoftherationalpartisantheory. Thesecond-halfinterestrateandin(cid:13)ation e(cid:11)ects are in the right direction and the (cid:12)rst-half output e(cid:11)ects are generally large. The money growth e(cid:11)ects tend to be negative in the second-half, however. Overall, the point estimates in Table 3 and the statistical signi(cid:12)cance (cid:12)gures in Table 2 show no consistent pattern of support for important output e(cid:11)ects or for output e(cid:11)ects caused by surprise changes in monetary policy. 4 Discussion This paper highlights several questions regarding standard macroeconometric and politicaleconometricresultsand sheds some lighton theanswersto thosequestions. 24 In Table 3, the measure of statistical signi(cid:12)cance comes from the Bayesian bootstrapped coverageintervals, andnot F tests. 24

model 1 2 3 4 5 variable half average response i: No economic dummies Output growth 1 0.24 0.18 0.10 0.18 0.18 2 0.04 0.08 0.10 0.06 0.06 3-Month T-Bill 1 -0.08 -0.12 -0.14 -0.14 -0.16 2 0.16 0.06 0.06 0.06 0.08 Money growth 1 0.10 -0.10 0.00 -0.06 -0.04 2 0.08 0.04 -0.08 0.02 -0.04 CPI in(cid:13)ation 1 -0.06 -0.10 -0.10 -0.12 -0.16 2 0.06 0.00 -0.06 0.00 0.06 ii : Credit Control Output growth 1 0.16 0.12 0.00 0.12 0.16 2 -0.10 0.00 -0.06 -0.06 -0.04 3-Month T-Bill 1 0.04 0.00 -0.02 -0.02 -0.04 2 0.40 0.28 0.30 0.26 0.42 Money growth 1 0.00 -0.28 -0.16 -0.16 -0.24 2 0.16 -0.02 -0.12 0.06 -0.22 CPI in(cid:13)ation 1 0.20 0.14 0.18 0.08 0.10 2 0.40 0.34 0.32 0.28 0.46 iii : Credit Control & Bretton Woods Output growth 1 0.22 0.16 0.08 0.16 0.16 2 0.08 0.12 0.16 0.12 0.12 3-Month T-Bill 1 -0.04 -0.08 -0.10 -0.10 -0.12 2 0.16 0.06 0.04 0.06 0.10 Money growth 1 0.08 -0.16 -0.06 -0.10 -0.10 2 0.12 0.10 -0.04 0.08 -0.02 CPI in(cid:13)ation 1 -0.02 -0.06 -0.08 -0.10 -0.14 2 0.04 -0.02 -0.10 -0.04 0.04 Notes: Half 1 is quaters -1 through 8; half 2 is quarters 9 through 15. The impulse response is averaged over the relevant half. The symbols are as in Table 2. In this table, a quantity is signi(cid:12)cantly di(cid:11)erent from zero at, say, the 10 percent level if th th the interval between the 5 and 95 percential points from the Bayesian bootstrap (described in the text) does not cover zero. Table 3: Summary of impulse response to election shock 25

On the negative side, it is clear that correlations among political variables and a wide range of macroeconomic variables at both leads and lags renders unreliable results from univariate work treating political variables as exogenous. We provide a framework for sorting out these interactions. Applying this framework, we conclude that there is some weak and fragile evidence in favor of important political e(cid:11)ects. The strongest evidence seems to come from the (cid:12)rst half of Republican administrations: recessions have followed the election of Republicans and macroeconomic factors alone may not account for this fact. There is little evidence, however, that the causal explanation of any political e(cid:11)ects on the economy operates through surprise changes in monetary policy. These results should be viewed as a challenge to those who believe that we have clear evidence of political e(cid:11)ects, but not proof that there are no such e(cid:11)ects. A balanced conclusion is that the political e(cid:11)ects are not large and systematic enough to be easily and precisely measured. Finding such e(cid:11)ects will either require putting more structure on the problem in the form of a priori assumptions, using more powerful statistical methods, or bringing di(cid:11)erent information to bear. 26

Appendix Themethodforcalculatingtheimpulseresponsecanbeseenfromtheaugmented VAR. The expected path of the economy at the end of t(cid:0)1 can be written, Et (cid:0)1Xjt+i = Et (cid:0)1[Xjt+ijzt+1 = 1]Et (cid:0)1[zt+1]+ Et (cid:0)1[Xjt+ijzt+1 = 0](1(cid:0)Et (cid:0)1[zt+1]) Since @Et (cid:0)1[zt+1]=@Et (cid:0)1ut = 1, @Et (cid:0)1Xjt+i = Et (cid:0)1[Xjt+ijzt+1 = 1](cid:0)Et (cid:0)1[Xjt+ijzt+1 = 0] (8) @Et (cid:0)1ut which is simply the di(cid:11)erence in the expected path of the economy under the two parties. For i < 15, this expression can be evaluated simply from the augmented VAR and does not depend on (cid:8) or the zt equation. Expression(8)iseasytoevaluate. Similartothecalculationofimpulseresponses in a standard VAR,the impulse response to election shocks can be calculated as the di(cid:11)erence in a dynamic simulation of the model under the two possible outcomes for the election. Data All economic data (except M1 and M2 for the period before 1959) are from the Federal Reserve’s U.S. database. All growth rates are computed as the annualized quarterlychange inthelogarithmofthevariable. The data seriesareGNP,personal consumption expenditures, gross private domestic investment, and federal government purchases of goods and services from the GNP accounts, all in 1987 dollars. The unemployment rate is for civilian males 20 years or older; hours are for all persons in the nonfarm business sector. M1 and M2 are not seasonally adjusted; the pre-1959 data were constructed by Robert Rasche. The Three-month T-Bill rate is a yield from the secondary market. The in(cid:13)ation rate is for the CPI, all urban consumers. The two commodity price indices are the producer price index for intermediate materials and sensitive materials. All of the dummy variables were constructed by the author. All data are available on request. 27

References Alesina,A.,1987. Macroeconomicpolicyinatwo-partysystemasarepeatedgame, Quarterly Journal of Economics, 102, 651{678. Alesina, A., 1991. Evaluating Rational Partisan Business Cycle Theory: A Response. Economics and Politics, 3, 63{71. Alesina, A. and N. Roubini, 1992. Political Cycles in OECD Economies. Review of Economic Studies, 59, 663{688. Alesina, A. and H. Rosenthal, 1995. Partisan politics, divided government and the economy, Cambridge University Press: New York. Alesina, A., J. Londregan and H. Rosenthal, l993. A Model of the Political Economy of the United States. American Political Science Review, 87 (1), 12-33. Alesina, A., and J. Sachs, 1987. Political Parties and the Business Cycle in the United States, 1948-1984. Journal of Money, Credit, and Banking, 20, 63{82. Allen, S., 1986. The Federal Reserve and the Electoral Cycle. Journal of Money, Credit, and Banking, 18, 88{94. Allen, S. and D. McCrickard, l991. The In(cid:13)uence of Elections on Federal Reserve Behavior, Economics Letters, 37, 51{55. Bange, M., W. Bernhard, and J. Granato, 1995. Partisan Monetary Policy, In(cid:13)ation, and Economic Growth. Michigan States University’s Institute for Public Policy and Social Research Working Paper 95-04. Beck, N., 1982. Presidential In(cid:13)uence on the Federal Reserve in the 1970s. American Journal of Political Science, 26, 415{445. Beck, N., 1982. Parties, Administrations, and American Macroeconomic Outcomes. The American Political Science Review, 76, 83{93. Bernanke, B. and A. Blinder, l992. The Federal Funds Rate and the Channels of Monetary Transmission. American Economic Review, 82, September, 681-21. Bizer, D., and S. Durlauf, 1990. Testing the positive theory of government (cid:12)nance. Journal of Monetary Economics, 26, 123{141. Browning,R.,l985. Presidents,Congress, andPolicyOutcomes: USSocialWelfare Expenditures, l949-77. American Journal of Political Science, 29, 197-216. Chapell, H., and W. Keech, 1986. Party Di(cid:11)erences in Macroeconomic Policies and Outcomes. American Economic Review, 76, 71{74. 28

Christiano, L.and M. Eichenbaum, 1992. Identi(cid:12)cation andthe Liquidity E(cid:11)ect of a Monetary Policy, in Political economy, growth, and business cycles, (Cukierman, A., et al. eds.), MIT press, Cambridge and London, 335{70. Christiano,L.,M.Eichenbaum,andC.Evans,1994. Thee(cid:11)ectsofmonetarypolicy shocks: some evidence from the (cid:13)ow of funds, NBER Working paper #4699. Doornik, J. and D. Hendry, 1994. PcFiml 8.0. Interactive Econometric Modelling of Dynamic Systems. International Thomson: London. Economic Report of the President, 1981. U.S. Government Printing O(cid:14)ce. Ellis, C., and M. Thoma, 1991. Causality in Political Business Cycles. Contemporary Policy Issues, 9. 39{49. Faust, J.andJ.Irons. 1996. Politics, policy, and the economy,manuscript,Federal Reserve Board. Faust, J. and E. Leeper. 1993. When do long-run identifying restrictions give reliable results? Manuscript Federal Reserve Board. Fair, R., 1978. The E(cid:11)ect of Economic Events on Votes for President. The Review of Economics and Statistics, 60, 159{173. Fair, R., 1982. The E(cid:11)ect of Economic Events on Votes for President: The 1980 Results. The Review of Economics and Statistics, 62, 322{325. Frey, B., and F. Schneider, 1978. An Empirical Study of Politico-Economic Interaction in the United States. The Review of Economics and Statistics, 60, 174{183. Friedlaender, A., 1973. Macro Policy Goals in The Postwar Period: A Study in Revealed Preference. Quarterly Journal of Economics, 87, 25{43. Friedman, M., and A. Schwartz, 1963. A Monetary History of the United States 1867{1960, Princeton, NJ: Princeton University Press, 1963. Gallant, R, P. Rossi, and G. Tauchen, 1993. Nonlinear dynamic structures, Econometrica, 61, 871{907. Gartner. M., and K. Wellersho(cid:11), 1991. Theories of Political Cycles: Lessons From the American Stock Market. Paper delivered at the January 1992 Annual Meetings of the American Economic Association, New Orleans. Gleisner,R.,l992. Economicdeterminantsofpresidentialelections: theFairmodel, Political Behavior, 14, 383{394. Golden, D., and J. Poterba, 1980. The Price of Popularity: The Political Business Cycle Reexamined. American Journal of Political Science, 24, 696{714. 29

Gordon, D. and E. Leeper, l994. The Dynamic Impacts of Monetary Policy: An Exercise in Tentative Identi(cid:12)cation. Journal of Political Economy, 102, December. Granato, J.,andW.West, 1994. WordsandDeeds: SymbolicPoliticsandDecision Making at the Federal Reserve. Economics & Politics, 6, 233{255. Granato, J., and Suzuki. Undated. The use of the encompassing principle to resolve empirical controversies invoting behavior, manuscript,Michigan State University. Hamilton, H., 1985. Historical Causes of Postwar Oil Shocks and Recessions. The Energy Journal, 6, 97{116. Hansen, L. P., and Sargent, T. J., 1991b. Identi(cid:12)cation ofContinuous Time Rational Expectations Models from Discrete Time Data, in Rational Expectations Econometrics, L. P. Hansen and T. J. Sargent (eds.), Boulder, CO: Westview Press, 219{238. Havrilesky, T., and J. Gildea, 1991a. The Policy Preferences of FOMC Members as Revealed by Dissenting Votes. Journal of Money, Credit, and Banking, 23, 130{137. Havrilesky, T., and J. Gildea, 1991b. Screening FOMC Members for their Biases and Dependability. Economics & Politics, 3, 139{149. Haynes, S, and J. Stone, 1989. An Integrated Testfor Electoral Cycles in The U.S. Economy. The Review of Economics and Statistics, 71, 426{434. , 1990. Political Models of the Business Cycle Should Be Revived. Economic Inquiry, 28, 442{465. , 1994. Why Did Economic Models Falsely Predict a Bush Landside in 1992? Contemporary Economic Policy, 12, 123{130. Hendry, D., 1995. Dynamic Econometrics, Wiley and Sons: New York. Hess, G., l993. Are Tax Rates Too Volatile? Southern Economic Journal, 60 (1), 72-88. Hibbs, D., 1977. Political Parties and Macroeconomic Policy. The American Political Science Review, 71, 1467{1487. Hibbs, D., l987. The American Political Economy: Macroeconomics and Electoral Politics. Harvard University Press: Cambridge. Hoover, K. and S. Perez, l994a. Post Hoc Ergo Propter Hoc Once More: An Evaluation of \Does Monetary Policy Matter?" in the Spirit of James Tobin. Journal of Monetary Economics, 34, August. 30

, l994b. Money May Matter, But How Could You Know? Journal of Monetary Economics 34, August. Klein, M., 1993. Timing is All: Elections and The Duration of United States Business Cycles, 1993. Manuscript, Tufts University. Leeper, E., l993. Has the Romers’ Narrative Approach Identi(cid:12)ed Monetary Policy Shocks? Manuscript, Federal Reserve Bank of Atlanta. Leeper, E. and D. Gordon, 1992. In Search of the Liquidity E(cid:11)ect,Journal of Monetary Economics, 29, 341{69. Lehmann, E., 1986. Testing statistical hypotheses. John Wiley: New York. Marcet, A., 1991, Temporal Aggregation of Economic Time Series, in Rational Expectations Econometrics, L. P. Hansen and T. J. Sargent (eds.), Boulder, CO: Westview Press, 237{282. MacRae, C., 1977. A Political Model of the Business Cycle. Journal of Political Economy, 85, 239{263. McCallum, B., 1978. The Political Business Cycle: An Empirical Test. Southern Journal of Economics, 44, 504{515. McManus, D., 1992. How common is identi(cid:12)cation in parametric models? Journal of Econometrics, 53, 5{23. Nordhaus, W., 1975. The Political Business Cycle. Review of Economic Studies, 42, 169{190. Richards,D.,1986. UnanticipatedMoneyandthePoliticalBusinessCycle. Journal of Money, Credit, and Banking, 18, 447{457. Romer, C. and D. Romer, l989. Does Monetary Policy Matter? A New Test in the Spiritof Friedmanand Schwartz, inBlanchard,OlivierJ.andStanleyFischer, eds., NBER Macroeconomics Annual l989 (Cambridge: MIT), 121-70. , l990. New Evidence on the Monetary Transmission Mechanism, Brookings Papers on Economic Activity 1, 149-98. , l994. Monetary Policy Matters. Journal of Monetary Economics 34, August. Runkle, D., 1987. Vector Autoregressions and reality, Journal of Business Economics and Statistics, 5, 437{442. Sims, C., 1980a. Macroeconomics and Reality, Econometrica, 48, 1{48. 31

Sims,C.1980b. ComparisonofInterwarandPostwarBusinessCycles: Monetarism Reconsidered, Papers and Proceedings of the American Economic Association 70, 250-57. Sims, C., 1992. Interpreting the macroeconomic time series facts, European Economic Review 36, 975{1011. Sims, C. and T. Zha, 1994. Does Monetary Policy Generate Recessions?: Using Less Aggregated Price Data to Identify Monetary Policy, manuscript, Yale University. Sims, C. and T. Zha, 1994. Error bands for impulse responses, manuscript, Yale. Soh, B., 1986. Political Business Cycles in Industrialized Democratic Countries. KYKLOS, 39, 31{46. Stigler, G., l973. General economic conditions and national elections, American Economic Review, 63, 160{167. Strongin,S.,1995. TheIdenti(cid:12)cationofMonetaryPolicyDisturbances: Explaining the Liquidity Puzzle, Journal of Monetary Economics, 35, 463{97. Tufte, E., l978. Political Control of the Economy, Princeton University Press: Princeton. 32

33

34

35

36

37

38

39

40