GAVEL-TO-GAVEL CONGRESSIONAL TELEVISION COVERAGE AS POLITICAL ADVERTISING: THE IMPACT OF C-SPAN ON LEGISLATIVE SESSIONS.

MIXON JR, FRANKLIN G. ; Hobson, David L. ; UPADHYAYA, KAMAL P. 等

KAMAL P. UPADHYAYA [*]

This article examines the effect of television on the length of legislative sessions at the federal level in the United States. Data from the U.S. Congress during the period 1972-96 are employed, during part of which time each house of Congress received significant television coverage by C-SPAN and C-SPAN2. Evidence from a Parks regression suggests that the presence of C-SPAN has increased House sessions by 88-250 hours and the presence of C-SPAN2 has increased Senate sessions by a striking 252-431 hours, other things constant. Additional estimates suggest that House sessions are about two minutes longer per bill introduced under the eye of C-SPAN, and Senate sessions are about four minutes longer per bill introduced in the presence of C-SPAN2. Longer sessions, which represent low-cost forms of advertising for incumbents, are not without costs to taxpayers. We estimate that these costs lie somewhere between $16 million and $392 million in real terms per session of Congress. (JEL D72, H11)

I. INTRODUCTION

As Grain and Goff (1988) point out, the number of political "watchers" ("C-SPAN junkies") seems to be on the rise from a 1986 estimate of 40 million. Since 1986, the Grain and Goff (1988) study has been the only significant empirical study of the impact of television on representative democracy despite the potential availability of relevant data and statistical techniques. The purpose of the present article is to partially fill the void concerning the impact of television on the democratic process by concentrating on the effect of television on the length of legislative sessions in the U.S. Congress. This study makes use of a "voter shopping" construct that analyzes political services as search/experience goods and employs data from the U.S. House of Representatives and the U.S. Senate during the period 1972-96, during some of which time each branch of Congress received significant television coverage by C-SPAN or C-SPAN2.

II. PRODUCT CLASSIFICATIONS AND THE PROVISION OF POLITICAL SERVICES

As Grain and Goff (1988) point out, the theory of product advertising parallels the theory of information as in Stigler (1961) and Nelson (1970, 1974). These articles define search goods as goods for which consumer judgments about quality or performance can be made prior to purchase and experience goods as those for which consumer judgments about product quality/performance can be made only after purchase. Nelson points out that information regarding experience good (search good) quality can be obtained by consumers at high cost (low cost). For experience goods, ads contain cues ("indirect") on the reputation of the seller, which is verifiable at low cost and from which the buyer can make inferences regarding product quality. [1]

Crain and Goff (1988) point out that one point of view in the literature holds that political services are like search goods because candidates' records are available to voters for verification and evaluation at low cost. On the other hand, it might be costly for voters to gather information on candidates prior to an election and/or difficult for voters to draw inferences about the future behavior of politicians, which suggests that political services are experience goods (Crain and Goff, 1988, 11). In this regard, any technological advancement that changes the costs of information gathering by voters will alter the nature of political services as in Crain and Goff (1988, 19). Because televising legislative proceedings lowers the cost of evaluating a politician's "advertised qualities" against his or her "actual qualities," political services would be more like search goods.

It is within this framework that we analyze the impact of television on the length of legislative sessions. However, we argue that there are differences in this regard with state-level legislatures, which were analyzed by Crain and Goff (1988), and the U.S. Congress. Voting records for members of the U.S. Congress have always been more accessible than those at the state level, due in large part to coverage by nationally circulated periodicals, such as Time and Newsweek; widely circulated newspapers, such as the Wall Street Journal and the New York Times; and cable television networks, such as CNN, HNN, CNBC, and MSNBC. These are all in addition to Washington, D.C.-based information sources, such as Congress Daily, The Hill, and others. Also, federal legislators have significantly larger operating budgets than their state-level counterparts and often use them to provide voting records to their constituents. Think tanks and congressional watchdogs, such as the Progressive Policy Institute, the Cato Institute, Americans for Tax Reform, and the National Taxpayers Union, publish and distribute voting guides on all congressional members for practically every issue that comes before Congress. Finally, the advent of the Internet (and Web magazines, such as Slate) has made gathering these types of informational cues easier for voters concerned with federal politics. These examples suggest that federal legislative services have historically been more like search goods than state-level political services. [2]

One other difference between federal and state legislatures concerns the types of voting devices used in the U.S. House of Representatives. Voters cannot determine, through the C-SPAN medium, exactly how a legislator voted at the time of a bill's passage or failure. Similarly, although the U.S. Senate still uses a voice roll-call voting mechanism, C-SPAN2 almost always overlays the roll-call with classical music. On the other hand, the wide variety of techniques used to televise state-level legislatures permits some voters to record the votes cast by their own representatives. Unlike at the federal level, this may be their only low-cost source of information on votes. Given that some state legislators have diverse constituencies, such information could be harmful to their reelection bids, and they would work to prevent the televising of legislative sessions. If unsuccessful, they would attempt to conclude legislative sessions in a shorter time, as in Crain and Goff (1988).

At the federal level, because C-SPAN and C-SPAN2 do not serve as particularly unique sources of information concerning actual votes, opposition to televising would be less significant, even in the presence of diverse constituencies. In fact, more general support for televising would exist because of the numerous opportunities for legislators to "showcase" themselves by taking popular positions. In the U.S. House, these opportunities are afforded by "One Minute Speeches" and "Special Order Speeches" (which have been aired by C-SPAN since 1982). The latter begin in prime time, and live coverage often continues past midnight. [3]

In the Senate, Morning Business and filibustering offer similar avenues for senators to "grandstand" in front of television cameras. These examples are all in addition to the committee and subcommittee meetings that C-SPAN and C-SPAN2 cover in large measure. Therefore, legislators have the opportunity to use the C-SPAN cameras as a medium for low-cost advertising throughout each session, thereby creating longer sessions. The fact that such exposure would cost significant sums of money for challengers to replicate means that incumbents gain significant electoral advantage from legislative television.

III. ADVERTISING AND INCUMBENCY ADVANTAGE: PREVIOUS LITERATURE

A number of studies on incumbency advantage have appeared in the public choice literature. Mayhew (1974) points out that (1) incumbents in the House are made safer by the increased name recognition resulting from an increased number of franked mailings to district residents, (2) redistricting allows politicians to create districts that ensure incumbent success, (3) incumbents have access to more scientific polling data than challengers, and (4) increasing within-district government spending represents an advantage to incumbents. [4] All of the aspects listed above are endogenous to the federal legislative process. Such is also the case with legislative session length. [5] The extent to which longer sessions will occur depends in part on the perceived relationship between absolute and relative political advertising expenditures and election results.

A number of studies have examined the relationship between political advertising and electoral success, including Banaian and Luksetich (1991), Bender (1988), Lott (1991), Erikson and Palfrey (1993), Mueller and Stratmann (1994), and Nagler and Leighley (1992). Banaian and Luksetich (1991) and Lott (1991) challenge research that suggests that campaign spending lowers votes received by incumbents in congressional elections. [6] Erikson and Palfrey (1993) also present estimates suggesting that incumbent spending effects are highly significant on votes and greater in magnitude than challenger spending effects. [7]

Mueller and Stratmann (1994) categorize the types of political advertising. As they suggest, informative advertising by either candidate provides information about their own positions or those of their opponent. [8]

They also state, with regard to persuasive advertising--which is an attempt to convince voters to provide support regardless of platforms/positions--that both candidates are trying to convince voters that they are honest (and will bring funds to the district) and it is the candidate who is perceived to be more honest and will bring the most funds to the district who stands to win. This candidate is likely to be the one who has spent the most on persuasive ads, as in Mueller and Stratmann (1994, 65). Eventually, persuasive campaign advertising will crowd out informative advertising, as in Mueller and Stratmann (1994). This study not only provides a rationale for the advertising--votes relationship but also is consistent with the idea that the institutional arrangements in the U.S. House and Senate--One Minute Speeches, Special Order Speeches, and filibustering for a popular cause (e.g., the Bill of Rights, children's health, etc.)--are tailored to impacting voter beliefs in such a persuasive manner. Support fo r the advertising-votes relationship is important to the hypothesis that federal legislators will attempt to increase session length as a way of engaging in low-cost persuasive advertising in return for votes.

IV. ECONOMETRIC MODEL AND RESULTS

Based on the literature described above, the following statistical equation is proposed:

(1) SESSION [alpha]+ [[beta].sub.1]SENATE

+ [[beta].sub.2]RLBB + [[beta].sub.3]INCUMB

+ [[beta].sub.4]DEMOCRAT

+ [[beta].sub.5]CSPAN

+ [[beta].sub.6]CSPANSEN + [[beta].sub.7]POP

+ [[beta].sub.8]POPSQ + [[beta].sub.9]BILLS

+ [[beta].sub.10]BPOP + [[beta].sub.11]SPOP

+ [[beta].sub.12]LFPR + [[beta].sub.13]GGDP

+ [[beta].sub.14]PREZ + [[beta].sub.15]RINC

+ [[beta].sub.16]UNION + [epsilon].

The definitions of all the variables in equation (1) are included along with summary statistics and data sources in Table 1. The dependent variable above is equal to the length of each legislative session for the House and Senate (separately) in hours, for all legislative sessions from 1972 to 1996. The data set consists of House session observations from 1972 to 1996 and Senate session observations for the same time period (a panel data set). Figures 1 and 2 track movements in session length for the House and Senate from 1972 to 1996. [9] SENATE is a dummy variable equal to one for Senate observations and zero otherwise. SENATE is included to control for institutional differences across congressional branches. Because the U.S. Constitution and Senate rules create a more deliberate body in the Senate, it is our expectation that the parameter estimate for Senate will be positive, ceteris paribus. RLBB is the real value of the legislative branch budget for each session and is expected to be positively related t o SESSION.

INCUMB is the percentage of eligible incumbents seeking reelection at the end of each legislative session for each branch. DEMOCRAT is the percentage of each branch of Congress represented by Democrat legislators. It is expected that both variables will be positively related to SESSION. Incumbents running for reelection favor large spending packages, which will come from longer sessions, as in Kimenyi and Tollison (1995). Democrat legislators also generally prefer larger spending packages and will, therefore, prefer longer sessions. For INCUMB, however, it could be the case that shorter sessions are preferred, ceteris paribus, so that legislators can spend more time visiting their districts and debating their opponents.

CSPAN is a dummy variable equal to one for the advent and presence of C-SPAN (or C-SPAN2) coverage and zero otherwise. [10] It is expected that, other things constant, the presence of C-SPAN cameras will lengthen legislative sessions as legislators use the institutional infrastructure of each congressional branch as low-cost advertising. CSPANSEN is an interaction dummy between CSPAN and SENATE and is included to test for the presence of a Senate premium regarding session length and television cameras. The more deliberative body has greater latitude in lengthening sessions, and this variable accounts for that possibility. We also recognize that Senate "districts" are larger than House "districts," and that the Crain and Goff (1988) hypothesis suggests that larger, more diverse districts create incentives for politicians to oppose television coverage and longer legislative sessions when covered by television cameras. As pointed out above, incumbents' political positions at the federal level are obtained by vo ters at much lower cost than those of state-level legislators, and this process is not enhanced by the presentation offered by C-SPAN concerning actual votes. Therefore, it is likely to be the case that cameras receive greater support because they offer a significant medium for persuasive political posturing and grandstanding.

POP (population) and POPSQ (population squared) are included as regressors in equation (1). We do not offer statistical expectations for these; because our data set is partly time series, they are included to control for the size of the federal constituency and any nonlinearity regarding size. So far, the equation being estimated has related outputs to inputs and technology. We have previously considered what might be called "institutional inputs" as well as variables relating to the technological change of television. The number of bills (BILLS) introduced each session provides a more direct input measure and as such is expected to be positively related to session time in each congressional branch.

Additionally, SPOP (the percent of the U.S. population that is Hispanic), BPOP (the percent of the U.S. population that is black), LFPR (the U.S. labor force participation rate), RINC (real per-capita income in the United States) and UNION (the percent of the U.S. workforce that is unionized) are included as alternative measures of diversity (to POP and POPSQ). These not only track overall U.S. trends but also interstate diversity better than POP and POPSQ.

SPOP BPOP, and UNION track changes in traditionally liberal constituencies and as such estimate the demand for larger, more complex pieces of legislation and regulation. Growth rates in LFPR stem mainly from greater inclusion of women (another traditionally liberal constituency) in the workforce. These variables are all expected to be positively related to SESSION. No a priori judgment is made concerning RINC.

GGDP, or the percentage of U.S. gross domestic product accounted for by federal government spending, measures the size of the public budget and estimates the "complexity" of legislation at hand, as in Kimenyi and Tollison (1995). It is, therefore, expected to be positively related to SESSION. Finally, given that a possible tendency exists for the amount of legislation and thus session length to vary systematically across periods of a presidential term--with more legislation introduced by an executive branch in the first two years versus the last two years of a four-year term--we include the dummy variable PREZ, which is equal to one for legislative sessions coinciding with the first two years of a president's term and zero otherwise.

Applying Greene (1997, 612-47), we can expect that the variance will be quite different in the two time-series (1972-96) of our model, that the lengths of sessions in the House and Senate will be tied to macroeconomic and legislative chamber-specific factors, which leads to correlation of disturbances across congressional houses, and that the errors will be correlated over time. The Parks (1967) estimation technique assumes heteroscedasticity, a first-order autoregressive error structure, and contemporaneous correlation between cross-sections. Specifically, the random errors, [u.sub.it] (for i = 1,2,...,n, and t = 1,2,...,T) have the structure

(2) E([[u.sup.2].sub.it]) = [[sigma].sub.ii], E([u.sub.it], [u.sub.jt]) = [[sigma].sub.ij] and

[u.sub.ij] = [[rho].sub.i] [u.sub.i,t-1] + [[epsilon].sub.it].

In this model, the covariance matrix for the vector of random errors (u) can be expressed as V, which is estimated by a two-stage procedure, and [beta] is then estimated by generalized least squares. The first step in estimating V involves the use of ordinary least squares to estimate [beta] and obtain the fitted residuals (u = y - X[[beta].sub.OLS]). A consistent estimator of the first-order autoregressive parameter is then obtained in the usual manner, and the autoregressive characteristic of the data is asymptotically removed by the transformation of taking weighted differences (and without losing any observations). This system is written as

(3) [[Y.sup.*].sub.it] = [[[sigma].sup.P].sub.k=1] [[X.sup.*].sub.itk] [[beta].sub.k] + [[u.sup.*].sub.it]

(i = 1, 2, . . . , n; t = 1, 2, . . . , T).

The second step in estimating the covariance matrix is to apply ordinary least squares to the transformed model (obtaining [u.sup.*] = [y.sup.*] - [X.sup.*][[[beta].sup.*].sub.OLS]), from which the consistent estimator of [[sigma].sub.ij] is calculated. The Parks estimator is unbiased, consistent and efficient, with

(4) Var([beta]) = [(X'[V.sup.-1]X).sup.-1].

A restricted version of equation (1) was run utilizing the Parks regression technique detailed above to obtain an efficient panel estimator. These results are included as column (1) of Table 2 and include the first five regressors from equation (1). This includes all of the "indirect" measures of "institutional inputs" and the CSPAN dummy variable. Most of the coefficients retain their expected signs, and every variable is statistically significant in this version. [11] In column (1), CSPAN is quite large (250.18), suggesting that the presence of C-SPAN cameras leads to (or has led to) an additional 250 hours of session time in each branch of Congress, ceteris paribus. Other variables suggest that (1) the Senate sessions are about 352 hours longer; (2) larger populations reduce session length in both houses, but at a decreasing rate; (3) an additional $500 million in legislative budget per session increases session length in both houses by 226 hours; and (4) an additional 1 percentage point in INCUMB and DEM O reduce session time (in each branch) by about 15 and 25 hours, respectively.

Column (2) of Table 2 presents results from a restricted form of equation (1) that adds CSPANSEN. This version tests for any premium (positive or negative) in the Senate due to the presence of C-SPAN cameras. All of the variables retain their previous signs, and all remain statistically significant, including CSPANSEN, which is positive. The presence of the interactive dummy (CSPANSEN) points out that column (1) overestimates the institutional differences in the Senate that lead to longer sessions there. The parameter estimate for SENATE falls from 351.86 to 234.49 with the inclusion of the interactive term. Although it remains that the more deliberative body of Congress, the U.S. Senate, has longer sessions on average, the portion of the difference attributed to institutional factors is overstated in column (1). A large part of that estimated difference in column (1) is due to the presence of C-SPAN2 cameras. [12] In column (2), the parameter estimate for CSPAN is 187.81, which represents a decrease of abou t 68 hours from column (1). The interactive term CSPANSEN retains an estimate of 169.15. These results suggest that (1) the presence of C-SPAN cameras results in an additional 188 hours of session time in each house, and (2) that the premium due to the C-SPAN2 cameras in the Senate is an additional 169 hours there, for a total of 357 additional hours in the Senate. [13] The Senate premium is likely due to the Constitutional features and specific rules of the Senate (e.g., the filibuster, greater ability to offer and speak on amendments) that allow senators to take better advantage of the presence of cameras than their House counterparts. The Crain and Goff (1988) models, using mainly state-level data, suggest that, because Senate "districts" are more diverse than House districts, the opposite sign for the Senate "premium" is expected. However, it should be noted that it is the stated differences in federal and state legislative branches--the fact that the costs to voters of acquiring information about federal legislators and their votes has historically been lower than for state legislators, and the method of televising the vote-taking process by C-SPAN--which leads to such an expectation.

Columns (3) through (7) of Table 2 present the more unrestricted specifications of equation (1) by including alternate measures of constituent diversity and institutional inputs that may impact session length. As the results suggest, BPOP (columns [6] and [7], SPOP (columns [3], [4], and [5]), LFPR (columns [4], [5], and [6]), and GGDP (columns [5], [6], and [7]) are all positive. This finding is consistent with the notion that growth of traditional liberal constituencies results in greater levels of more complex legislation and thus greater session time in the U.S. Congress. Also, the number of bills introduced is positive and significant in columns (5) and (6) of Table 2. In column (7), an anomalous result is obtained for UNION, and RINC is positive and significant. In each of these columns (3)--(7), both CSPAN and CSPANSEN are positive and significant. Estimates suggest that the presence of the C-SPAN cameras in the House works to increase legislative sessions there from a low of 88 hours to a high of 148 hours. Also, the results point out that the presence of C-SPAN2 cameras in the Senate works to increase session length from a low of 252 hours to a high of 431 hours. These numbers are similar to those produced by column (2) of Table 2. Other results from these versions suggest that (1) an additional $500 million in legislative budget per session increases session length in both houses by 181 hours (column [7]), (2) Senate sessions are 276-338 hours longer, (3) every additional percentage point in INCUMB and DEMO reduce session time in each branch by 4-17 hours and 21-51 hours, respectively, (4) an additional 100 bills proposed increases session time in each house by 16-17 hours, and (5) every additional percentage point in BPOP and GGDP increases session time in both houses by 48-69 hours and 240-753 hours, respectively. [14]

Use of the variable BILLS potentially presents statistical problems, given that it is endogenous to the process. Therefore, an additional strategy is to estimate "productivity" as the dependent variable. By dividing SESSION by BILLS, SESBIL measures the amount of session time per bill introduced in each branch of Congress. In addition to several of the regressors listed in equation (1), we also include two new variables: (1) the number of personal legislative staff members per legislator for each branch (STAFF), and (2) the total number of committee staff members for each branch of Congress (COMMS). This latter measure is a proxy for the number of committees in each branch. It is expected that each of the two new measures will be negatively related to SESBIL because larger staff sizes/more committees allow legislators to dispense with a given number of bills more expeditiously. However, if our thesis is correct, the variable CSPAN will be positively related to SESBIL; greater levels of political "grandstandi ng" as a form of low-cost advertising will be reflected in greater amounts of time spent per bill introduced in each house.

Results from five versions of a Parks model employing SESBIL as the dependent variable are presented in Table 3. In all five versions, CSPAN is positive; it is significant in four of those. Also in all five versions CSPANSEN is positive; it is also significant in four of those. Using the parameter estimates reported in Table 3, the presence of C-SPAN cameras has increased session time per bill introduced by about two minutes in the U.S. House. This is quite remarkable given that the mean time per bill in the House is 11 minutes (from 1972-96). The results also suggest that the presence of C-SPAN2 cameras in the Senate has increased session time per bill by about four minutes. Again, this is significant given that the mean time per bill in the Senate is 14 minutes (from 1972-96). Other results in Table 3 suggest that (1) an increase of $500 million in real legislative budget per session will increase session time per bill by three to six minutes in each branch; (2) an additional percentage point in DEMO will reduce session time per bill by 11 seconds in each branch; (3) Congress spends two minutes less per bill during the first two years of a president's term compared with the last two years; (4) an additional percentage point of BPOP will increase session time by four to six minutes in each branch; (5) an additional ten personal legislative staff per legislator will reduce session time per bill in each branch by 22 seconds to 1.2 minutes; and (6) an additional 100 committee staff personnel will reduce session time per bill by two minutes in each branch. Finally, as with the results in Table 2, the Table 3 estimates produce sizable adjusted [R.sup.2] measures for the Parks models.

As a final test of the robustness of the difference between CSPAN and CSPANSEN, the data are split across the House and Senate. This creates two time-series data sets, and a test for unit roots is required, as suggested by Nelson and Plosser (1982). We conducted an augmented Dickey-Fuller test as in Dickey and Fuller (1979) and a PhillipsPerron test as in Phillips (1987) and Phillips and Perron (1988), which uses nonparametric correction to deal with any correlation in the error terms. All of these results are reported in Table 4. The results are quite mixed; several of the series are not stationary in either level form or first-difference. We therefore included SESSION, SESBIL, DEMO, INC UMB, BILLS, RLBB, POP, and RINC in first-difference form and the others in level form. Results from four versions each of a model using [delta]SESSION and [delta]SESBIL as the dependent variables are presented in Table 5. Although the CSPAN dummy generally retains a positive sign and relatively large value, it is never sign ificant. However, none of the eight specifications are jointly significant based on the F-statistics reported in the table. These discouraging results are due to the fact that when the pooled model is split into two time-series models, the number of observations is cut by one-half; use of a first-difference term causes the first observation in each data set to be lost as well [15] As an alternative, we interacted the SENATE variable with all of the other "institutional inputs" in column (2) of Table 2 to test for any difference between the House and Senate chambers not being accounted for in equation (1). The results are similar to those in column (2), and the interaction between RLBB and SENATE is only marginally significant among the interaction terms (the model produced an adjusted [R.sup.2] of 0.93). These results are supportive of the pooled models reported in Tables 2 and 3.

V. FURTHER INTERPRETATION OF RESULTS

Further interpretation of the results presented in column (2) of Table 2 yield some startling conclusions. First, if the C-SPAN cameras had been present at the beginning of our data series, the 1972 House and Senate sessions are predicted to have been 13% and 16% longer, respectively. Similarly, had C-SPAN cameras not been allowed in either branch, the 1996 House and Senate sessions are predicted to have been 8% and 12% shorter, respectively. We also examined the data set to see where the biggest differences regarding the presence/absence of television coverage were located. For the House, the 1972 session would have been (predictably) 13% longer in the presence of television, whereas the 1982 session (the first year of House coverage) would have been (predictably) 13% shorter without television's presence. For the Senate, the 1984 session (the year preceding extensive Senate coverage) would have been (predictably) 18% longer with C-SPAN2 coverage, and the 1990 session would have been (predictably) 16% short er without television coverage in the Senate chamber. These figures are quite remarkable.

On the 19 May 1999 episode of CNN's Crossfire, cohost Bill Press noted that each hour of House session costs about $7,000 (e.g., lighting costs, the costs of putting speeches into the record, etc.). We computed the real value of this figure using the consumer price index in our data series and estimated the incremental cost to taxpayers of the added legislative session time due to C-SPAN in each branch of Congress using our regression estimates. The Parks estimates suggest that, in real terms, C-SPAN raised the cost of each House session (since 1982) by $814,044, and C-SPAN2 raised the cost of each Senate session (since 1986) by $1,547,207. Therefore, the total incremental costs per session, since 1986, are approximately $2.4 million in real terms. [16] The total incremental costs in the House, in real terms, due to C-SPAN are $6,512,349 (since 1982), and the costs for the Senate are $9,283,245. These figures, when summed, amount to approximately $15.8 million in real terms.

As the Kimenyi and Tollison (1995) study points out, longer legislative sessions produce larger, more complex pieces of legislation. The creation of more complex legislation leads to legislative costs that go beyond those associated with the actual business that occurs within the walls on Capitol Hill. Longer, more complex legislation has cost implications at the bill drafting stage, which occurs inside congressional offices, located in Washington and throughout the various 435 federal districts and across the 50 states. The size and types of legislative staff involved are affected, as well as the sources of technology (e.g., polling, etc.) that legislators use. For each of the years in our data series, we divided the real legislative branch budget (RLBB in equation [1]) by total session hours for both the House and the Senate. This hourly exchange rate was multiplied by 187.81 (the estimate for CSPAN) for each of the years that the House was televised by C-SPAN (1982-96). Based on this computation, our mode l predicts that the increment in House sessions due to C-SPAN since 1982 has cost taxpayers approximately $1.147 billion in real terms. Additionally, this hourly exchange rate was then multiplied by (187.81+169.15), or 357 (which is the CSPAN parameter plus the CSPANSEN parameter), for each of the years that the Senate was televised by C-SPAN2 (1986-96). This computation predicts that the increment in Senate sessions due to the C-SPAN2 cameras since 1986 has cost taxpayers approximately $1.628 billion in real terms. The total real incremental cost to taxpayers of "legislative television" has been approximately $2.799 billion. On average, the costs are $143 million and $271 million for the House and Senate, respectively. The average total per session (since 1986) of $392 million is remarkable given that the real value of campaign expenditures by all candidates running for House and Senate seats during the 1993-94 election cycle was approximately $500 million. If the additional spending by legislators represent s "persuasive" advertising, then the true cost of campaigns in the United States is underestimated by about 44%, and the difference in incumbent-challenger efforts in this regard is vastly underestimated as well. This amount also supports Lott's conclusion that incumbents do not have to spend as much to win. [17] These figures and differentials are certainly consistent with the striking success rates of incumbent politicians. In sum, our estimations provide lower and upper bound estimates of $16 million and $392 million for the real incremental cost per legislative session for both branches of Congress due to the presence of television cameras.

VI. CONCLUDING COMMENTS

Although voters can perhaps make more informed decisions in the presence of legislative television, this article points out that there are hidden costs in the electoral and legislative processes that come from recent trends to televise legislative procedures. Against a theoretical construct in which political services are modeled as goods (search and experience), politicians often engage in persuasive advertising campaigns to win the production contract from their constituencies. Given the institutional rules of the federal legislative branch in the United States, the presence of cameras on the legislative floor allows for political grandstanding and posturing that would not likely take place otherwise. We provide evidence from a Parks regression that suggests that House sessions have become longer as a result of the presence of C-SPAN cameras. The predicted increase of 187 hours is significant, as is the predicted increase of 357 hours in Senate sessions due to the presence of C-SPAN2 cameras. Longer sessio ns are not without costs to taxpayers. In all, we estimate that those costs lie somewhere between $16 million and $392 million per legislative session of Congress. That such large hidden costs can emanate from a process that has received so little attention in the literature is surprising. This study partially fills that void.

Mixon: Associate Professor of Economics, University of Southern Mississippi, Box 5072, Hattiesburg, MS 39406-5072. Phone 1-601-266-5083, Fax 1-601-266-4920, E-mail [email protected]

Hobson: J. D. Candidate, School of Law, Emory University, Atlanta, GA 30322. Phone 1-812-331-0296.

Upadhyaya: Assistant Professor of Economics, University of New Haven, West Haven, CT 06516. Phone 1-203-932-7487, E-mail [email protected].

(*.) We thank two anonymous referees, Mark Crain, Steve Caudill, Troy Gibson, and James Wilkinson for helpful comments. This work also benefited from financial support through an SU2000 Grant from the University Research Council.

(1.) Empirical tests have expanded on the goods classifications developed by Nelson and have verified the hypothesis concerning "indirect" forms of advertising for experience goods in product markets and for an additional class of goods--credence goods--detailed by Darby and Karni (1973) later (see Mixon [1995, 1999]; Ekelund et al. [1995] for empirical tests).

(2.) Crain and Goff (1988) are correct to point out that very few goods, including political services, are entirely search or experience. Rather, most goods have characteristics of both types. A continuum exists between polar cases of search and experience. Most products fall somewhere along this spectrum (Crain and Goff, 1988, 8).

(3.) On 19 May 1999, Representative David McIntosh (R-IN) led a colloquy on the House floor with several other Republican representatives in early prime time during the Special Order session of House business. They were describing in conversational tone what they viewed as the radical environmental views of then-Vice President Al Gore. Passages from Gore's book on global warming were real aloud before the C-SPAN cameras. CNN's Crossfire repeatedly broke away to cover the colloquy, and liberal cohost Bill Press complained about such a use of taxpayer money on the House floor.

(4.) Mayhew (1974) points Out that the mail-flow curve closely tracks the incumbency advantage curve. Also, a recent study by Mixon and Upadhyaya (1997) suggests that the 1982 amendments to the Voting Rights Act of 1965 reduced turnover by predicted 10.3 percentage points in states that created minority-majority districts. Finally, the scientific polling data and techniques allow incumbents to achieve a proper alignment of political views within their own districts.

(5.) In a related study, Kimenyi and Tollison (1995) test the relationship between session length (in hours), the complexity of legislation, and federal spending. They hypothesize that the amount of government spending will be larger in bills that are long and complex (e.g., omnibus-type appropriations, etc.) and that complex legislation will emerge from longer sessions. Using data from 1947-82, they find that total session hours and the length of bills (in pages) enacted are directly related and that the ultimate result of longer sessions is greater levels of public spending at the federal level, ceteris paribus (Kimenyi and Tollison, 1995).

(6.) Banaian and Luksetich (1991) employ a model that simultaneously determines votes received and spending and find that both incumbent and challenger spending are significant determinants of the popular vote received. Lott (1991) goes further by pointing out that studies finding a negative relationship between campaign spending and votes received by incumbents have failed to account for the stock of reputational capital that the incumbent brought to the election, which biases the econometric results. Accordingly, "more well known politicians may not have to spend very much to win" (Lott, 1991, 91). A summary of several of the studies criticized by Banaian and Luksetich and Lott is included in Levitt (1995).

(7.) Their multi-equation approach accounts for the effect of anticipated closeness on spending and includes variables measuring district party strength and total spending by incumbents and challengers. Similarly, Nagler and Leighley (1992) find that anticipated closeness in presidential elections leads to increased advertising outlays and that presidential campaign spending significantly increases a candidate's vote share. Bender (1988) finds that a politician's likelihood of voting on campaign spending limits varies directly with the impact of such changes on the incumbent's reelection probabilities (using House votes on 1974 campaign spending limits legislation). This finding is also consistent with the view that the spending-votes relationship is significant.

(8.) Mueller and Stratmann (1994) suggest that informative advertising by Candidate A is a substitute for informative advertising by Candidate B. This proposition would need to account for the credibility problem that arises when one candidate provides information about his or her opponent (see Crain and Goff [1988], 12).

(9.) The simple Pearson correlation coefficients for session length and time are 0.60 and and 0.54 for the House and Senate, respectively. These are significant at the 0.03 and 0.07 levels, respectively. Additionally, the mean value for SESSION in the House is 1,695.6 before C-SPAN and 1,799.3 after C-SPAN. For the Senate, these mean values are 2,210.7 and 2,468.0 before and after C-SPAN2, respectively.

(10.) As the C-SPAN Web site points Out, C-SPAN began televising House proceedings in 1979, the year used by Grain and Goff (1988). We chose 1982 because it was the first year C-SPAN extended coverage to 16 hours per day. This allows for rising labor force participation rates, giving viewers the ability to watch outside of daily work schedules. Extensive Senate coverage began in 1986, so it was included as such in our model.

(11.) The deletion of POPSQ leads to similar results compared to inclusion, with the exception that POP is significant only in the presence of POPSQ.

(12.) The variable SENATE models the "fixed effect" of congressional chamber. Another such "fixed effect" is the session, each of which contains unique factors that influence session length, which the autocorrelation correction will not completely control. We attempted to reestimate column (2) with a series of dummies for session, with the final session serving as the omitted session. However, the estimation matrix was singular and many parameters were not estimable (the classic "dummy variable trap").

(13.) Several other specifications of column (2) were conducted to check for the robustness of our results. When DEMO is omitted, only SENATE and INCUMB remain significant. When column (2) is run without POP and POPSQ CSPAN and CSPANSEN become insignificant and the other variables remain significant. When DEMO, POP, and POPSQ are all omitted, only SENATE and RLBB remain significant (though CSPAN and CSPANSEN remain positive). However, we note here that columns (3) through (7), which contain better measures of "institutional inputs" and "constituent diversity," confirm the results in columns (1) and (2) concerning the impact of television. Finally, we note that when columns (1) and (2) are rerun in ordinary least squares the regressors in the model retain their signs, though some (including CSPAN and CSPANSEN) lose their significance. It remains true, however, that ordinary least squares estimation in the presence of either heteroscedasticity or autocorrelation produces inefficient estimates. The Parks model corrects for these problems, thus producing efficient estimates.

(14.) When column (1) of Table 2 is run with only CSPAN, BILLS (the direct measure of "institutional inputs") and some of the alternative measures of constituent diversity (BPOP, UNION, and GGDP), each variable retains its expected sign, but none are significant at any conventional level and the adjusted [R.sup.2] is 0.09. The parameter estimate for CSPAN here is 31.9 with a standard error of 174.8.

(15.) Splitting the pooled data into pre-C-SPAN and post-C-SPAN time periods would result in an even greater data limitation than the one created by time-series for each branch.

(16.) Similar calculations can be made using the models with SESBIL as the dependent variable. For instance, using the estimates from column (3) of Table 3, C-SPAN has raised the real cost of each House session (since 1982) by $985,138, and C-SPAN2 has raised the real cost of each Senate session (since 1986) by $2,038,715. Therefore, the total incremental costs per session, since 1986, are approximately $2.9 million in real terms. These estimates are strikingly similar to those reported above.

(17.) We do acknowledge the possibility that longer sessions may to some degree crowd out the time that an incumbent might be fundraising. That is, C-SPAN may not be a pure add-on to other campaign-related activities. Representatives and senators frequently complain that a longer session time cuts into their ability to travel during the campaign season. This notion, that C-SPAN time and some campaign-related activities are substitutes, is consistent with Lott's thesis.

REFERENCES

Banaian, K., and W. A. Luksetich. "Campaign Spending in Congressional Elections." Economic Inquiry, 29, 1991, 92-100.

Bender, B. "An Analysis of Congressional Voting on Legislation Limiting Congressional Campaign Expenditures." Journal of Political Economy, 96, 1988, 1005-21.

Cram, W. M., and B. L. Goff. Televised Legislatures: Political Information Technology and Public Choice. Boston: Kluwer Academic Publishers, 1988.

Darby, M. R., and E. Karni. "Free Competition and the Optimal Amount of Fraud." Journal of Law & Economics, 16, 1973, 67-88.

Dickey, D. A., and W. A. Fuller. "Distribution of the Estimators for Autoregressive Time Series with a Unit Root." Journal of the American Statistical Association, 74, 1979, 427-31.

Ekelund, R. B. Jr., F. G. Mixon Jr., and R. W. Ressler. "Advertising and Information: An Empirical Study of Search, Experience and Credence Goods." Journal of Economic Studies, 22(2), 1995, 33-43.

Erikson, R. S., and T. R. Palfrey. "The Spending Game: Money, Votes and Incumbency in Congressional Elections." Cal Tech Social Science Working Paper 851, 1993.

Greene, W. H. Econometric Analysis. Upper Saddle River, NJ: Prentice Hall, 1997.

Kimenyi, M. S., and R. D. Tollison. "The Length of Legislative Sessions and the Growth of Government." Rationality & Society 7(1), 1995, 151-55.

Levitt, S. D. "Policy Watch: Congressional Campaign Finance Reform." Journal of Economic Perspectives, 9, 1995, 183-93.

Lott, J. R. Jr. "Does Additional Campaign Spending Really Hurt Incumbents? The Theoretical Importance of Past Investments in Political Brand Name." Public Choice, 72, 1991, 87-92.

Mayhew, D. "Congressional Elections: The Case of Vanishing Marginals." Polity, 6(2), 1974, 295-317.

Mixon, F. G. Jr. "Advertising as Information: Further Evidence." Southern Economic Journal, 61, 1995, 1,213-18.

_____. "Customer Return Policies for Experience Goods: The Impact of Product Price and Consumer Search Costs on Seller-Provided Informational Cues." Applied Economics, 31, 1999, 331-36.

Mixon, F. G. Jr., and K. P. Upadhyaya. "Gerrymandering and the Voting Rights Act of 1982: A Public Choice Analysis of Turnover in the U.S. House of Representatives." Public Choice, 93, 1997, 357-71.

Mueller, D. C., and T. Stratmann. "Informative and Persuasive Campaigning." Public Choice, 81, 1994, 55-77.

Nagler, J., and J. Leighley. "Presidential Campaign Expenditures: Evidence on Allocations and Effects." Public Choice, 73, 1992, 319-33.

Nelson, C. R., and C. I. Plosser. "Trends and Random Walks in Macroeconomic Time Series." Journal of Monetary Economics, 10, 1982, 139-62.

Nelson, P. "Information and Consumer Behavior." Journal of Political Economy, 77, 1970, 311-29.

_____. "Advertising as Information." Journal of Political Economy, 81, 1974, 729-54.

Ornstein, N. J., T. E. Mann, and M. J. Malbin. Vital Statistics on Congress, 1991-1992. Washington, DC: Congressional Quarterly, 1992.

Parks, R. W. "Efficient Estimation of a System of Regression Equations when Disturbances are Both Serially Correlated and Contemporaneous Correlated." Journal of the American Statistical Association, 62, 1967, 500-09.

Phillips, P. C. B. "Time Series Regression with Unit Roots." Econometrica, 55(2), 1987, 277-301.

Phillips, P. C. B., and P. Perron. "Testing for a Unit Root in Time Series Regression." Biometrica, 75(3), 1988, 335-46.

Statistical Abstract of the United States. U.S. Department of Commerce, Bureau of the Census, various issues/years.

Stigler, G. J. "The Economics of Information." Journal of Political Economy, 68, 1961, 213-25.

Will, G. F. Restoration: Congress, Term Limits and the Recovery of Deliberative Democracy. New York: Free Press, 1992.

TABLE 1
Variables, Definitions, Means, and Standard Deviations
Variable Definition Mean
SESSION The length of each legislative 2,044.42
 session, in total hours (1972-96)
 for the U.S. House or the U.S.
 Senate.
SENATE Dummy variable equal to one for 0.50
 Senate observations and zero
 otherwise.
RLBB Real value of the legislative 2,930.90
 branch budget for each legislative
 session (1972-96). RLBB is measured
 millions and is a two-year sum.
INCUMB Percentage of eligible incumbents 0.86
 seeking reelection at the end of
 each legislative session. For
 House observations, INCUMB is
 x/435, where x is the number of
 incumbents seeking reelection. For
 Senate observations, INCUMB is y/33.33,
 where y is the number of incumbents
 seeking reelection.
DEMOCRAT The percentage of representatives or 0.56
 senators that are members of the
 Democrat party during each legislative
 session.
CSPAN Dummy variable equal to one for 0.54
 House or Senate sessions covered by
 one of the C-SPAN networks. For House
 observations, CSPAN = 1 in 1982 and
 thereafter. For Senate observations,
 CSPAN = 1 in 1986 and thereafter. House
 coverage is provided by C-SPAN, and
 Senate coverage is provided by C-SPAN2.
CSPANSEN A dummy interaction variable equal to 0.23
 the product [SENATE x CSPAN].
POP The size of the U.S. population in 236,822.81
 the middle of each legislative session.
 POP is measured in thousands.
POPSQ The size of the U.S. population in the 56,380,760,786
 middle of each legislative session
 squared.
BILLS The number of legislative bills 12,106.7
 introduced in each legislative session.
BPOP The ratio of the U.S. population 0.118
 accounted for by the black population
 at the end of each session.
SPOP The ratio of the U.S. population 0.073
 accounted for by the Hispanic population
 at the end of each session.
LFPR The U.S. labor force participation 64.20
 ratio at the end of each session (%).
GGDP Federal government spending as a 21.45
 percentage of U.S. gross domestic
 product during each session.
PREZ Dummy variable equal to one for 0.46
 sessions covering the first two
 years of a presidential administration,
 and zero otherwise.
RINC Real per-capita income (U.S.) during 13,253.3
 each session.
UNION The percentage of the U.S. workforce 17.85
 that is unionized.
SESBIL Session time spent per bill introduced 0.211
 in each session in total hours
 [SESSION/BILLS].
STAFF Personal staff per legislator for 45.14
 each session.
COMMS The number of committee staff 2,740.0
 personal for each session.
Variable SD
SESSION 381.78
SENATE 0.51
RLBB 420.33
INCUMB 0.07
DEMOCRAT 0.06
CSPAIN 0.51
CSPANSEN 0.43
POP 17,537.03
POPSQ 8,338,929,202
BILLS 5,854.4
BPOP 0.004
SPOP 0.020
LFPR 2.29
GGDP 1.23
PREZ 0.51
RINC 1,491.9
UNION 4.76
SESBIL 0.110
STAFF 30.40
COMMS 553.7
Sources: Ornstein et al. (1992); Statistical Abstract of the
United States: Will (1992); World Wide Web Site for C-SPAN,
www.cspan.org.
TABLE 2
Parks Regression Results (dependent variable:
SESSION; standrad errors in parentheses)
Regressor (1) (2) (3)
Constant 17,000 *** 17,996 # 81,076 #
 (7,904.2) (7,180.1) (19,571)
SENATE 351.86 # 234.49 # 278.59 #
 (67.48) (89.36) (82.18)
RLBB 0.452 # 0.498 # 0.502 #
 (0.172) (0.158) (0.194)
INCUMB -1,492.14 # -1,546.84 # -723.6 **
 (463.0) (412.3) (432.5)
DEMO -2,502.2 # -3,013.34 # -2,696.5 #
 (807.3) (775.0) (774.5)
CSPAN 250.18 *** 187.81 * 147.89 **
 (123.5) (119.1) (82.46)
CSPANSEN 169.15 ** 197.53 #
 (92.9) (61.44)
POP -0.105 * -0.108 * -0.63 #
 (0.067) (0.061) (0.16)
POPSQ 0.19e-6 * 0.19e-6 * 0.11e-5 #
 (0.1e-6) (0.1e-6) (0.3e-6)
BILLS -0.008
 (0.028)
BPOP 33,246
 (33,163)
SPOP 69,961 **
 (33,442)
LFPR
GGDP
PREZ
RINC
UNION
Parks adj. [R.sup.2] 0.8955 0.9192 0.9664
Regressor (4) (5) (6)
Constant 69,594 # 59,170 # 57,426 #
 (20,666) (17,598) (18,180)
SENATE 325.45 # 338.19 # 276.94 #
 (83.07) (69.44) (65.64)
RLBB 0.261 0.075 0.262
 (0.252) (0.222) (0.207)
INCUMB -549.81 [sim] -428.15 -686.62 **
 (409.14) (337.65) (346.13)
DEMO -2,161.9 # -2,109.4 # -2,793.3 #
 (810.1) (694.0) (648.7)
CSPAN 109.20 [sim] 87.93 [sim] 108.0 *
 (79.39) (63.47) (71.72)
CSPANSEN 171.67 # 163.59 # 185.75 #
 (59.70) (47.74) (51.03)
POP -0.62 # -0.87 # -0.92 #
 (0.16) (-.17) (0.17)
POPSQ 0.10e-5 # 0.16e-5 # 0.17e-5 #
 (0.3e-6) (0.4e-6) (0.3e-6)
BILLS 0.048 0.172 # 0.162 #
 (0.048) (0.060) (0.060)
BPOP -482.71 22,030 68,607 #
 (39,852) (35,565) (23,116)
SPOP 96,670 # 54,395 *
 (36,468) (33,755)
LFPR 240.47 * 752.64 # 737.73 #
 (168.01) (240.91) (241.83)
GGDP 222.37 *** 270.56 #
 (88.65) (80.06)
PREZ -53.55 -94.19
 (76.23) (76.27)
RINC
UNION
Parks adj. [R.sup.2] 0.9724 0.9838 0.9786
Regressor (7)
Constant 71,834 #
 (8,600)
SENATE -7.45
 (30.70)
RLBB 0.362 #
 (0.067)
INCUMB -1,676.6 #
 (229.37)
DEMO -5,142.5 #
 (301.0)
CSPAN 95.84 **
 (53.47)
CSPANSEN 335.28 #
 (33.75)
POP -0.62 #
 (0.08)
POPSQ 0.11e-5 #
 (0.2e-6)
BILLS -0.0007
 (0.009)
BPOP 48,354 #
 (8,070)
SPOP
LFPR
GGDP 207.07 #
 (72.51)
PREZ
RINC 0.497 #
 (0.041)
UNION -29.87 #
 (9.37)
Parks adj. [R.sup.2] 0.9929
(#) (***) (**) (*)([sim])denote significance at
the 0.01 (0.25) (0.05) (0.085) 0.10 level or
better for a one-tail test.
TABLE 3
Parks Regression Results (dependent variable: SESBIL; standard errors
in parentheses)
Regressor (1) (2) (3)
constant -1.027 # -0.932 # -0.843 ***
 (0.372) (0.368) (0.356)
SENATE 0.049 0.052 0.038
 (0.060) (0.060) (0.049)
RLBB 0.2e-3 # 0.2e-3 # 0.2e-3 #
 (0.3e-4) (0.3e-4) (0.3e-4)
INCUMB -0.093
 (0.100)
DEMO -0.305 # -0.309 # -0.320 #
 (0.110) (0.110) (0.107)
CSPAN 0.020 0.026 [sim] 0.028 *
 (0.196) (0.019) (0.018)
CSPANSEN 0.037 * 0.037 * 0.036 *
 (0.024) (0.025) (0.023)
PREZ -0.028 ** -0.029 ** -0.032 **
 (0.016) (0.016) (0.019)
BPOP 9.680 # 8.345 # 8.214 #
 (3.784) (3.540) (3.357)
GDDP 0.006 0.004
 (0.010) (0.010)
STAFF -0.2e-2 * -0.18e-2 -0.1e -2
 (0.1e-2) (0.3e-2) (0.1e-2)
COMMS -0.1e-3 # -0.1e-3 # -0.1e-3 #
 (-0.3e-4) (-0.3e-4) (-0.3e-4)
Parks adj. [R.sup.2] 0.9493 0.9495 0.9575
Regressor (4) (5)
constant -0.752 *** -0.715 ***
 (0.327) (0.316)
SENATE
RLBB 0.1e-3 # 0.1e-3 #
 (0.3e-4) (0.3e-4)
INCUMB
DEMO -0.311 # -0.308 #
 (0.111) (0.106)
CSPAN 0.036 *** 0.036 #
 (0.016) (0.014)
CSPANSEN 0.026 0.026 [sim]
 (0.021) (0.019)
PREZ -0.031 ** -0.032 **
 (0.017) (0.019)
BPOP 7.558 *** 6.900 ***
 (3.474) (2.903)
GDDP 0.001
 (0.009)
STAFF -0.6e-3 ** -0.6e-3 **
 (0.3e-3) (0.3e-3)
COMMS -0.1e-3 # -0.1e-3 #
 (-0.3e-4) (-0.3e-4)
Parks adj. [R.sup.2] 0.9489 0.9564
(#) {***} (**) [*] ([sim]) denote significance at the 0.01
{0.025} (0.05) [0.085] 0.10 level or better for a one-
tail test.
TABLE 4
Unit Root Tests
 Augmented Dickey-Fuller
 Level First-Difference
Variable (House Data)
 COMMS -2.94 -2.69
 STAFF -2.70 -5.48 **
 SESSION -1.69 -1.94
 SESBIL -4.00 ** -7.O8 **
Variable (Senate Data)
 COMMS -2.66 -3.50 *
 STAFF -2.18 -3.37
 SESSION -2.54 -2.84
 SESBIL -3.41 -4.02 **
 Phillips-Perron
 Level First-Difference
Variable (House Data)
 COMMS -0.89 -5.71 ***
 STAFF -3.56 * -5.49 ***
 SESSION -1.51 -2.31
 SESBIL -6.03 *** -7.87 **
Variable (Senate Data)
 COMMS -3.85 * -3.41 *
 STAFF -2.32 -2.69
 SESSION -2.23 -3.61 *
 SESBIL -5.22 *** -6.38 ***
(***)(**)[*] denote significance at the 0.01 (0.05) [0.10] level.
Variable series common to both branches of Congress (e.g., POP, RINC,
etc.) were either insiginificant in both level and first-difference or
significant in first-difference form.
TABLE 5
First-Difference Regression Results for Time-Series
(1974-96) Data (Standard errors in parentheses)
 Dep. Variable: [delta]SESSION
 House Senate
 (1) (2) a (1)
Constant 175.06 1,419.29 -109.69
 (308.31) (1,703.11) (181.65)
CSPAN -120.10 -105.77 167.31
 (297.40) (274.64) (207.03)
[delta]DEMO -1,456.5 -2,884.6 -1,446.8
 (1,893.4) (2,526.4) (2,294.2)
[delta]INCUMB -1,531.1 -2,912.2 -1,353.0 *
 (3,693.0) (3,751.5) (930.06)
GGDP -23.51 24.21
 (126.64) (94.46)
[delta]BILLS 0.014 0.020 -0.019
 (0.073) (0.064) (0.065)
[delta]RLBB -0.022 -0.171 0.444
 (0.598) (0.592) (0.480)
[delta]POP -0.270
 (0.361)
PREZ
[delta]RINC
F-statistic 0.140 0.221 0.742
[R.sup.2] 0.144 0.224 0.471
 Dep. Varibale: [delta]SESBIL
 House
 (2) b (1) (2) c
Constant 1,212.9 -0.072 0.039
 (1,232.55) (0.300) (0.038)
CSPAN 274.84 0.5e - 2 0.6e - 2
 (213.80) (0.045) (0.037)
[delta]DEMO -2,972.1 0.124 0.152
 (2,538.8) (0.519) (0.308)
[delta]INCUMB -1,574.4 ** 0.002
 (824.58) (0.732)
GGDP 0.007
 (0.020)
[delta]BILLS -0.033
 (0.051)
[delta]RLBB 0.458
 (0.433)
[delta]POP -0.303 2.1e - 5
 (0.280) (6.6 - 5)
PREZ -0.035
 (0.037)
[delta]RINC 4.8e - 5
 (2.4e - 5)
F-statistic 1.086 0.081 0.292
[R.sup.2] 0.566 0.063 0.143
 Senate
 (1) d (2) e
Constant 0.126 0.087
 (0.300) (0.310)
CSPAN 0.052 0.033
 (0.048) (0.048)
[delta]DEMO -0.148 -0.109
 (0.525) (0.545)
[delta]INCUMB
GGDP 0.022
 (0.017)
[delta]BILLS
[delta]RLBB
[delta]POP -2.8e - 5 -1.6e - 5
 (7e-5) (7e-5)
PREZ 0.010 -0.003
 (0.038) (0.039)
[delta]RINC
F-statistic 0.435 0.138
[R.sup.2] 0.266 0.073
(a) When [delta]DEMO and [delta]POP are omitted, CSPAN approaches
zero (-42.61) and remains insignificant.
(b) When two additional versions of this specification are
run--one without [delta]BILLS and the other without
[delta]RLBB--the CSPAN parameter is 226.9 and 228.8, but
marginally falls outside of the 0.10 level of significance.
(c) When [delta]BILLS is added to this version, the CSPAN
parameter is 0.035 and falls marginally outside of the 0.10 level
of significance.
(d) When [delta]BILLS is added to this version, the CSPAN
parameter is 0.074 and is significant at the 0.03 level.
(e) When [delta]BILLS is added to this version, the CSPAN
parameter is 0.073 and is significant at the 0.02 level.
(**)[*] denote significance at the 0.085 [0.10] level for a
one-tail test.