Campaign Effects in Theory and Practice

CAMPAIGN EFFECTS IN THEORY AND PRACTICE

American Politics Research, 2001 (vol. 29, pps. 419-437)

CHRISTOPHER WLEZIEN University of Oxford

ROBERT S. ERIKSON

Columbia University Authors’ note: An earlier version of this manuscript was presented at the Annual Meeting of the Southern Political Science Association, Atlanta, 2000. Portions of the research also were presented at the 1999 Conference on the Design of Election Studies, Houston. We thank Bruce Carroll and Jeff May for assistance with data collection and Pat Lynch, Tim Nokken, and especially Tom Holbrook for comments and suggestions. The research has been supported by a grant from the Institute for Social and Economic Research at Columbia University and forms part of a project supported by a grant from the National Science Foundation (SBR-9731308).

Abstract

While scholars debate the influence of election campaigns on electoral decision-making,

they agree that campaigns do have effects. That is, there is broad agreement that campaign

events can cause voters’ preferences to change. This is straightforward. Empirically identifying

the effects of the campaign is much less so. We simply do not have regular readings of voter

preferences over the election cycle, and the readings we do have are imperfect. Clearly, then, an

important question is: Can we actually detect the effects of election campaigns? This is a

fundamental empirical question. It forms the subject of this essay.

In the essay, we outline the primary theoretical perspectives on campaign events and their

effects. We then turn to the practice of empirically identifying these effects, focusing

particularly on survey error and its consequences for empirical analysis. Using selected poll data

from the 2000 presidential election cycle, we illustrate how the various forms of survey error

complicate the study of campaign effects. We also offer certain solutions, though these take us

only part of the way. Indeed, given the available data, it appears that all we can hope to offer are

fairly general conclusions about the effects of election campaigns.

1

Scholars debate the influence of election campaigns on voting behavior and election

outcomes (see, e.g., Alvarez, 1997; Campbell, 2000; Finkel, 1993; Gelman and King, 1993;

Holbrook, 1996; Johnston, et al., 1992; Lodge, Steenbergen, and Brau, 1995; Shaw, 1999a;

1999b). They nevertheless agree that campaigns do have effects. Put directly, but very

generally, events that happen over the course of the campaign cause voters’ preferences to

change. This is fairly straightforward. The problem is empirically identifying these effects. We

simply do not have readings of voter preferences at regular (relatively short) intervals over the

election cycle, and the readings we do have are imperfect. Clearly, then, an important question

is: Can we actually detect the effects of election campaigns? This is a fundamental empirical

question, and forms the subject of this essay. We begin with a discussion of theoretical

perspectives.

The Events Perspective

When studying the effects of campaigns, political scientists typically focus on the effects

of particular events. Most commonly, we examine the effects of very important events, such as

nominating conventions and general election debates. (For fairly recent treatments, see

Holbrook, 1996; Shaw, 1999a.) This focus is understandable for a number of reasons. First, we

know that conventions and debates are very visible, where a large number of people watch on

television and/or acquire information about them in other ways. Second, we can anticipate these

events, so our interpretation of their effects is not subject to the post hoc ergo propter hoc

reduction that characterizes interpretations of the seeming effects of many other campaign

events. Third, there already is evidence that they matter a lot more than other events, or at least

2

that they can. This is a good place to start the study of campaign effects. However, it only tells

us about the effects of certain events.

We know that campaigns really represent a series of many events, of which the highly

visible conventions and debates are a small number. It is fair to wonder about the effects of the

other events. Indeed, there is reason to think that they actually matter more, taken together. A

case study of polls during the 2000 presidential election campaign is suggestive. For this

analysis, we rely on the series of poll readings from Wlezien (2001). Figure 1 displays the data.

It shows Gore’s percentage share of the two-party vote intention (ignoring all other candidates)

in national trial-heat polls from the beginning of the year through Election Day. The

observations in the figure are the daily poll-of-polls. They represent Gore’s share for all

respondents aggregated by the mid-date of the reported polling period. The procedure used

exactly follows Erikson and Wlezien (1999).1 It allows readings for 173 separate days during

2000, 59 of which are concentrated in the period after Labor Day.

— Figure 1 about here —

In the figure we can see some pattern over time and also a lot of noise. Gore clearly

began the year well behind Bush and gained through the spring, where his support settled at

around 47%. This held until the conventions. We then observe the predictable convention

bounces, out of which Gore emerged in the lead—with about a 52% share—heading into the fall.

The polls bounce around a lot during the fall, though Gore’s support appears to drop fairly

continually until just before Election Day, when it rebounded sharply. The mean daily variance

of the polls is a modest 8.92 percentage points throughout the year, 5.67 percentage points after

Labor Day. How much of this variance is due to the conventions and debates?

For this exercise, we consider convention effects to include anything that happens from

3

the opening of the Republican convention through Labor Day and debate effects to include

anything that happens from the day of the first debate until 3 days after the last debate.2 We also

attribute all movement in the polls during these two seasons to separate daily convention and

debate events. To be perfectly clear, we estimate the unique effect of each day for which we

have poll readings during the two periods. Our analysis of variance thus provides very liberal

estimates of their effects. The results are shown in Table 1.

In the first column of the table, we can see that the daily change in the polls during the

convention season is quite meaningful. Indeed, the collective impact of this daily change is

statistically significant (F24,133=2.01, p<.01). This is as one might expect. The same is not true

for the period of the debates. What is most important for our analysis is that the conventions and

debates together account for at most 29 percent of the variance in the polls over the course of the

election year. At least 71 percent of the variance thus reflects something else entirely. Focusing

only on the fall campaign, in the second column in the table, we see that the full debate season

accounts for up to 14.4 percent of the poll variance after Labor Day. The debates evidently had

minor effects, seemingly less than what we would expect by chance (the adjusted R-squared is

negative). These findings are telling about the effects of the numerous, small events that occur

over the course of the campaign. These events collectively can have a much greater impact than

the handful of very visible events that scholars typically examine.3

— Table 1 about here —

Of course, it is difficult to identify the various things candidats do and the various other

things that happen over the campaign—on each day we have a cluster of events. In one sense,

the same is true of conventions and debates. The former really are complexes of different events

and the effects of the latter can reflect other ensuing events, including the effects of media

4

coverage and deliberation. How do we separate out the different events that occur over the

campaign? Conceiving of the campaign in terms of specific events is itself difficult enough; a

full-fledged empirical analysis seems well nigh impossible. This suggests to us the need for a

more general modelling approach.

A General Perspective

In other research (Wlezien and Erikson, n.d.), we have proposed various models of

campaign dynamics. The different models offer general characterizations of campaign effects.

To begin with, imagine an election campaign from beginning to end. Campaign events occur

over this timeline. These events include the full range of campaign behavior as well as the net

result of voter discussion and evaluation. Now, let us assume that we have trial-heat polls at

regular intervals over the course of the campaign and that the polls are perfect, i.e., no bias,

sampling error, and the like. Consider the time series of aggregate voter preferences (Vt) to be a

time series of the following form:

(1) Vt = α + β Vt-1 + gt,

where Vt is one candidate’s percentage share in the polls and g is a series of independent

campaign shocks drawn from a normal distribution.4 What can we learn from this equation?

What does it tell us about the effects of the campaign?

To begin with, if campaign events do have effects, we would observe poll movement

over the course of the campaign. Events would change voters’ preferences. The variance of g in

equation 1 would be greater than 0, and the larger the variance the greater the effects of events.

If events do not have any effects, conversely, preferences would remain constant over time. This

5

is straightforward, though do notice that the characterization really offers an upper bound on

campaign-induced change. That is, some of the movement in preferences that we observe will

be do to exogenous factors from outside the campaign, e.g., an oil shock.

Poll movement may tell us that campaigns have effects. It does not tell us whether the

effects last, however. Whether they do last or else decay is directly evident from the coefficient

β in equation 1. If 0<β <1, campaign effects decay, and the smaller the β the quicker the decay.

Preferences tend toward the equilibrium value of the series, which equals a / (1-β). This

equilibrium does not represent the final outcome, however; what happens on Election Day also

will reflect late campaign effects that have not fully dissipated by the time voters go to the polls.

This characterization is implicit in most forecasting models of election outcomes (see Campbell

and Garand, 2000), where the economic and political “fundamentals” remain fairly constant, at

least for much of the campaign. Now, if β equals 1.00, campaign effects do not decay, but

cumulate instead. Each shock makes a permanent contribution to preferences, and the election

outcome thus represents the sum of all shocks that occur over the campaign. This is the

characterization implied by “on-line” processing models of candidate preferences (see, e.g.,

Lodge, Steenbergen, and Brau, 1995).

Of course, we might expect that neither one of these models applies to all campaign

events. It may be that some events have temporary effects and others have permanent ones. Yet

other events may produce both effects, where preferences move and then bounce back but to a

level different from before. These possibilities imply the following “error correction” model of

preferences:

(2) Vt = V*t-1 + β (Vt-1 - V*

t-1) + gt, + ut,

6

where 0<β <1 and ut represents the series of shocks to the fundamentals. In this model, some

effects (ut) persist—and form part of the moving equilibrium V*t—and the rest (gt) decay.5 This

is the model implied by empirical analyses of selected campaign events, especially Shaw

(1999a).6

As we have noted, these different models of campaign dynamics formally represent

general arguments from the literature itself. They directly capture what we ultimately want to

know: (1) whether the campaign has real effects on electoral preferences; and (2) whether these

effects matter on Election Day. They apply generally, in any campaign and across levels of

analysis, e.g., in a presidential election, at the national level, the state level, and so on.7 They

also imply a strategy of analysis, one that allows for analysis of the effects of particular events,

such as conventions and debates. That is, one can directly model their specific effects. In

theory, then, the study of campaign effects is straightforward. In practice, it is not so easy.

Survey Error and Campaign Effects

Thus far, we have been assuming perfect polls and that any movement in aggregate poll

results indicates a campaign effect of some sort. We do not have perfect polls, however. Trial-

heat preferences from sample surveys represent a combination of true preferences and survey

error. We thus can ask: How much of the observed variance in measured preferences is real?

How much is simple survey error? Can we detect the dynamics of underlying voter preferences?

Let us consider survey error, its forms and consequences for the study of campaign effects.

Sampling Error

Survey error comes in many forms, the most basic of which is sampling error. All polls

7

contain some degree of sampling error. Thus, even when the division of candidate preferences is

constant and unchanging, we will observe changes from poll to poll. This is well known. The

implications for the study of campaign effects are less obvious but quite powerful, even when

sample sizes are large (say, greater than 1000). We simply cannot separate sampling error from

reported preferences.

The implications of sampling error are particularly pronounced for analyses of effects of

specific events. These are straightforward at the aggregate level.8 For a sample of a given size,

we know the errors for various levels of confidence associated with various sample sizes. Thus,

with random samples of 1,000 voters at regular fairly short intervals, we are about 95 percent

certain to detect effects of 3 percentage points or more. Holbrook’s analysis of presidential polls

during the three elections between 1984 and 1992 is useful here. Based on our reading of his

results, Holbrook finds statistically significant effects of nominating conventions in 5 out of 6

cases and presidential debates in 4 of 7 cases. The mean convention effect is about 6.5

percentage points and the mean debate effect is 2.2 points.9 If we assume that these averages are

indicative of the true effects of each convention and debate, the pattern of his results should

come as little surprise. We cannot always detect the seemingly large bumps associated with

conventions and the smaller effects of debates are particularly elusive. The effects of most other

events presumably would be even moreso.10

Sampling error also limits analysis of the more general time-serial characteristics of

aggregate voter preferences. Specifically, it increases the variance of reported preferences and

decreases evident persistence, by biasing the autoregressive coefficient downward, away from

1.0. We can to some extent adjust these analyses, however. First, we can fairly easily estimate

the proportion of variance of observed preferences that is due to sampling error.11 It is worth

8

noting then that, based on our own analyses of polls in the last 15 presidential elections between

1944 and 2000, most of the movement that we see during the fall campaign—the last 60 days—

is the result of mere sampling error (Wlezien and Erikson, n.d). On average, 52% is sampling

error. Second, while we cannot actually separate this sampling error from reported preferences,

we can adjust estimation of time-serial properties in particular election years using the statistical

reliability of the polls themselves (see note 11). That is, the observed correlation between poll

readings on different days is the simple product of the true correlation (r) and the reliability (rel).

The true correlation thus is the observed correlation r divided by rel. The point is that we can

recover the true, underlying pattern from the statistical noise, at least in theory (also see Erikson

and Wlezien, 1999). Of course, this presumes that we have poll data at regular, relatively short

intervals. It also presumes that there is no other survey error.

Design Effects

All survey results also reflect design effects. These effects represent the consequences of

the departure in practice from simple random sampling that results from clustering, stratifying,

weighting, and the like (Groves, 1989). When studying the effects of election campaigns, the

main source of design effects surrounds the polling universe. Obviously, it is not easy to

determine who will vote on Election Day.12 When we draw our samples, all we can do is

estimate the voting population. How we do this has consequences both for the cross-sectional

poll margins at each point in time and for the variance in the polls over time.

Historically, polling organizations relied on the registered voting population. While an

imperfect approach, it nevertheless provided us a fairly reliable measure of preferences over a

campaign. After all, 80-85 percent of the registered voting population actually voted, at least in

9

presidential elections. Since the passage of motor voter registration, however, this approach may

be less reasonable. (We are not entirely convinced, however.) Regardless, polling organizations

now rely on likely voter screens and various weighting schemes, both of which have substantial

design effects. Indeed, there is reason to think that the differences among likely voter screens

have much greater effects than the general differences between object universes, e.g., adults,

registered voters, and likely voters.13

As hinted at above, this is a problem even when polls are drawn using the same design

over time. Take the Gallup Poll during the 2000 presidential campaign. Gallup identified a

likely voter based largely on the “attention” a person was paying to the campaign. While this

screening device may have had little effect on the mean reported preference, it appears to have

had truly startling effects on the variance in poll results during the campaign. In Figure 2, we

display the Gallup polls after Labor Day, including each daily (3-day or 2-day) tracking poll.14

The numbers represent the Gore share of the two-party vote intention dated by the middle day of

the reported polling period. For purposes of comparison, we also display the Gore share for all

respondents from all other polls aggregated by the mid-date of the reported polling period.

— Figures 2 and 3 about here —

In the figure, we can see that the Gallup poll differs quite a lot from the average poll

during the fall campaign. Notice the well known flip-flop during the middle of the period, about

35 days before the election. On October 1, Gallup reported a dead heat in its three-day tracking

poll. Three days later, the poll showed Gore ahead by 11 points. Two days after that, where

one-third of the respondents in the two polls were the same, Bush was ahead by 8 points.15 Party

identification in the three polls shifted from 35-34% Democrat to 38-30% Democrat and then to

38-30% Republican. Using the set of polls available from pollingreport.com over the same days

10

indicates that Gore’s average share ranged between 50.6 and 52.9 percent. These patterns imply

that the attention people pay to the campaign can change a lot from day to day and that the

change can have meaningful consequences for poll results. The polls can vary from day to day

simply because of the screen itself.

The problem is even worse when we must rely on surveys from different organizations

using different screens and weighting schemes. This is the unfortunate plight of most

researchers. In order to get a time series of separate polls over a fairly long period at relatively

short (1-2) intervals, there is little alternative to piecing together surveys from different

organizations.16 Part of the problem with combining these data into a single series is that we

cannot tell what exactly the different organizations are doing. Putting aside Gallup, are all

survey organizations using the same screens? Are they using the same weighting schemes? We

know only a little, especially about weighting, and what we do know indicates that the

differences matter. In 2000, Voter.com equalized Democrats and Republican whereas the

Washington Post weighted by the distribution of party identification in the electorate, which was

about 4-5 points more Democratic. As expected, given the differences in their weighting

strategies, their poll results differ consistently at each point in time, by slightly more than 2

percentage points on average. This is clear in Figure 3, which shows Gore’s share in the two

tracking polls during the last 20 days of the campaign.17 Of course, the screens and weighting

procedures used by the many other survey organizations also differ, if in a less pronounced way.

The important point is that when we use polls from different survey organizations, results will

vary from one day to the next partly because the universes of the reported polls themselves

differ. This has important implications for any analysis of campaign effects.

11

House Effects

When combining polls from different survey organizations, house effects also are a

problem. These effects represent the consequences of survey houses employing different

methodologies. Even apart from polling universe (and question wording), results can differ

across houses due to data collection mode, interviewer training, procedures for coping with

refusals, and the like (see, e.g., Converse and Traugott, 1986; Lau, 1994; also see Crespi, 1988).

As for design effects, poll results will vary from day to day because the polls on different days

are conducted by different houses. Of course, as the discussion of the polling universe implies, it

is difficult to fully separate design effects from house effects. If design is not entirely clear but

varies by house, then what we will primarily observe in practice is house effects. If systematic,

as in the case of Voter.com, then these are relatively easy to correct.18 If they are not systematic,

as was apparently the case with the Gallup Poll in 2000, then the effects are much more difficult

to correct. This clearly complicates the analysis of campaign effects.

The Individual and the Aggregate

For the moment, let us assume that we can fully correct for the different types of survey

error. (Of course, this is not entirely possible.) That is, let us assume that we can produce a

series of true aggregate preferences. Let us further assume that the series moves. What does this

movement mean?

The standard interpretation of poll movement is that the mean support for the candidates

has changed. In other words, something happened to shift all voters, so that the distribution of

voters moved one way or the other, or that a segment moved and the rest remained largely in

place. It also may be, however, that the mean of the distribution has not changed but the

12

variance has. The possibility reflects the fact that we do not measure relative utilities for the

candidates; we measure their by-product, percentage point outcomes. (For further details, see

Erikson and Wlezien, 1998).

Consider a normal distribution of net voter preferences for two candidates, say, the

Democratic and Republican candidates for president. Zero is the neutral point and thus

represents the threshold between Democratic and Republican voting. People scoring positive on

this scale of relative preferences vote Democratic; those scoring negative vote Republican. If the

relative utility is greater than zero the voter prefers one candidate and if it is less than zero the

voter prefers the other. The aggregate indicator summarizes dichotomous individual

preferences; specifically, it represents one candidate’s percentage, say, the Democrat’s, of total

Democratic and Republican preferences. This has meaningful consequences for an analysis of

polls. Even if the mean underlying support for the candidates remains the same, shifts in the

variance in support will cause poll margins to move. Specifically, an increase in the variance

will cause the leader’s share to decrease; a decrease in variance will cause the share to increase.

For an illustration, see Appendix A. The point is that change in aggregate poll results can reflect

very different patterns of individual level movement. This clearly limits our ability to draw

conclusions about the effects of campaign events based solely on analysis of aggregate poll

results.

On Detecting Campaign Effects

Detecting campaign effects is anything but straightforward, even in the best of worlds.

The many facets of survey error complicate every analysis of polls, whether focused on the

specific effects of particular events or the more general effects of events over the course of a

13

campaign.19 The problems are particularly pronounced for analyses that combine polls from

different survey organizations using different polling universes. The main problem here is not

with the general differences, say, between polls of registered and likely voters, but the more

specific and less obvious differences within the broad category of likely voters. There is

substantial variation across survey organizations in the use of screens and weighting procedures

and it has meaningful implications for reported poll results. Thus, when we use polls from

different survey organizations, results will vary from one day to the next partly because the

universes of the reported polls themselves differ. Moreover, if recent practices are any

indication, the problem seems to be getting worse. There also is no real alternative to combining

different surveys, since no single organization offers regular readings at fairly short (1-2 day)

intervals over the course of the campaign.20

All is not lost, however. It is possible to at least some extent adjust for survey error,

especially in analyses of general campaign dynamics. To begin with, since most polls are

conducted over multiple days, we can date each poll by the middle day of the reported polling

period, as we did for our summary of presidential polls in Figure 1. Of course, it is necessary to

first must remove the rolling component from multi-day tracking polls (see note 20). We then

can directly estimate the effects of general differences in polling universes that survey

organizations use and the effects of survey houses themselves. Finally, we can adjust the polls

based on this analysis and aggregate respondents for each day. This provides a fairly clean time

series of aggregate preferences.21 Indeed, all that remains are variable design effects that we

cannot identify and sampling error itself. We cannot eliminate sampling error. But, as we have

seen, we can adjust our analyses of variance and general time serial characteristics using the

estimated sampling error, which is relatively easy to calculate. This is the approach used in our

14

analysis of polls from the 1996 presidential campaign (Erikson and Wlezien, 1999).

While the approach takes us a long way toward addressing survey error, we still cannot

be sure what the resulting patterns reflect. As we already have discussed, we cannot detect the

underlying pattern of individual level change using aggregate poll results. A change in the polls

may reflect a shift in mean preferences or a shift in the variance of underlying preferences. One

simply cannot tell. The obvious solution, it would seem, is a panel design with daily readings

over the course of the campaign. Assuming samples of about 1000 respondents and a reasonable

screen, we would be able to uncover the general characteristics of electoral preferences and also

fairly large, e.g., three-point, effects of events, both at the individual and aggregate levels. Even

this ideal would not allow us to detect the effects of most events, however, and it also comes

with its own, perhaps more serious, problems. (Consider sample attrition over time and the

likely Hawthorne effects.) Regardless, this seeming ideal is not very real.

There are other more reasonable possibilities available. One particularly useful strategy

is the rolling cross-section (RCS) design (see Johnston and Brady, N.d.). The RCS is essentially

a very large cross-section where the day on which a respondent is interviewed is selected

randomly. When conducting surveys over the course of campaigns, there seems to be little

reason not to adopt such an approach. Moreover, by embedding a panel at selected points in

time, we also can at least to some extent capture the underlying pattern of individual level

change.22 We are, of course, largely dependent on others—the pollsters themselves—to adopt

such a strategy. Short of this, we will be left piecing together cross-sections from different

houses employing different methodologies. Even if we can control for these differences, all that

the data allow is an analysis of aggregate results and, to the extent individual level data is

available, basic characterizations of the underlying dynamics. All we can hope to offer,

15

therefore, are very general conclusions about the effects of political campaigns. This is an

unfortunate fact of life.

16

Appendix A:

The Distribution of Individuals’ Preferences and Aggregate Poll Results: An Illustration

Consider the normal distribution of net voter preferences for the Democratic and

Republican presidential candidates discussed in the text. Recall that zero is the neutral point and

thus represents the threshold between Democratic and Republican voting, so that people scoring

positive on this scale of relative preference vote Democratic and those scoring negative vote

Republican. The poll margin at any particular point in time represents the sum of individuals’

dichotomous preferences. Now, let us assume that the mean preference is unchanged over time

but that the variance of underlying preferences increases, i.e., that preferences polarize. If this is

true, we will observe declining poll margins.

– Figure A1 about here –

Figure A1 illustrates the process. It shows two hypothetical normal distributions of net

voter preferences, where zero represents the threshold between Democratic and Republican

voting. (Note the overlay of the cumulative normal distributions in the figure.) The electorate

starts with Distribution 1, with a mean score of .5 standard deviation units, implying that 69

percent are to the right of zero, and vote Democratic. New information comes along to expand

the distribution of preferences but without altering the mean score, as in Distribution 2. Here,

the standard deviation is doubled so that the constant mean is only .25 of a standard deviation

from zero. With the electorate newly polarized in this way, the proportion Democratic shrinks to

60 percent. This shift occurs with a simple expansion of the variance and no change in mean

preferences. For a more technical exposition, see Erikson and Wlezien (1998).

17

Figure A1: Hypothetical Distributions of Voter Preferences

References

Alvarez, R. M. (1997). Information and Elections. Ann Arbor: University of Michigan

Press.

Box-Steffensmeier, J., & Franklin, C. (1995). “The Long Campaign.” In Herbert Weissberg, ed.

Democracy’s Feast. Chatham Hills, NJ: Chatham House.

Box-Steffensmeier, J., & Smith, R. (1998). "Investigating Political Dynamics Using

Fractional Integration Models." American Journal of Political Science 42:661-689.

Campbell, A., Converse, P.E., Miller, W.E., & Stokes, D.E. (1960). The American Voter.

New York: Wiley.

Campbell, J.E., (2000). The American Campaign: U.S. Presidential Campaigns and the

National Vote. College Station, Texas: Texas A&M University Press.

-----, & Garand, J.C. (2000). Before the Vote: Forecasting American National Elections.

Thousand Oaks, Calif.: Sage Publications.

Converse, P.E., & Traugott, M.W. (1986). “Assessing the Accuracy of Polls and

Surveys.” Science 234:1094-1098.

Crespi, I. (1988). Pre-Election Polling: Sources of Accuracy and Error. New York:

Russell Sage.

DeBoef, S. (2000). “Persistence and Aggregations of Survey Data Over Time: From

Microfoundations to Macropersistence.” Electoral Studies 19: 9-29.

Erikson, R.S., & Wlezien, C. (1999). “Presidential Polls as a Time Series: The

Case of 1996.” Public Opinion Quarterly 63:163-177.

-----. (1998). “The Timeline of Political Campaigns.” Paper presented at the Annual Meeting of

the American Political Science Association, Boston.

-----. (1996). “Of Time and Presidential Election Forecasts.” PS: Political Science and Politics

29:37-39.

Finkel, S.F. (1993). “Reexamining the ‘Minimal Effects’ Model in Recent Presidential

Elections.” Journal of Politics 55:1-21.

Franklin, M., & Wlezien, C. (N.d.). “Reinventing Election Studies.” Electoral Studies,

forthcoming.

Gelman, A., & King, G. (1993). “Why are American Presidential Election Polls so

Variable When Votes are so Predictable?” British Journal of Political Science 23:

409-519.

Groves, R.M. (1989). Survey Errors and Survey Costs. New York: Wiley.

Heise, D.R. (1969). “Separating Reliability and Stability in Test-Retest Correlations.”

American Sociological Review 34:93-101.

Holbrook, T. (1996). Do Campaigns Matter? Thousand Oaks, California: Sage.

Jacoby, W.G. (1997). “Public Preferences During a Presidential Campaign: Sources of

Stability and Change.” Presented at the Annual Meeting of the Midwest Political

Science Association, Chicago.

Johnston, R., Blais, A., Brady, H.E., & Crete, J. (1992). Letting the People Decide: Dynamics

of a Canadian Election. Kingston, Canada: McGill-Queen's Press.

Johnston, R., & Brady, H. (N.d.). “The Rolling Cross Section Design.” Electoral Studies,

forthcoming.

Lau, R. (1994). “An Analysis of the Accuracy of ‘Trial Heat’ Polls During the 1992

Presidential Election.” Public Opinion Quarterly 58:2-20.

Lazarsfeld, P. F., Berelson, B.R., & Gaudet, H. (1944). The People’s Choice.

New York: Columbia University Press.

Lebo, M.J., Walker, R.W., & Clarke, H.D. (2000). “You Must Remember This: Dealing With

Long Memory in Political Analyses.” Electoral Studies 19:31-48.

Lodge, M., Steenbergen, M., & Brau, S. (1995). “The Responsive Voter: Campaign

Information

and the Dynamics of Candidate Evaluation.” American Political Science Review 89:309-

326.

Popkin, S. L. (1991). The Reasoning Voter: Communication and Persuasion in

Presidential Campaigns. Chicago: University of Chicago Press.

Shaw, D.R. (1999a). “A Study of Presidential Campaign Event Effects from 1952 to 1992.”

Journal of Politics 61:387-422.

-----. (1999b). “The Effect of TV Ads and Candidates Appearances on Statewide Presidential

Votes, 1988-96.” American Political Science Review 93: 345-362.

Timpone, R. (1998). “Structure, Behavior, and Voter Turnout in the United States.”

American Political Science Review 92:145-158.

Wlezien, C. (2001). “On Forecasting the Presidential Vote.” PS: Political Science

And Politics 34:25-31.

-----. (2000). “An Essay on ‘Combined’ Time Series Processes.” Electoral Studies 19:77-93.

Wlezien, C., and Erikson, R.S. (N.d.). “The Timeline of Presidential Election Campaigns.”

Journal of Politics, forthcoming.

Zaller, J. (N.d.). “Assessing the Statistical Power of Election Studies to Detect

Communication Effects in Political Campaigns.” Electoral Studies, forthcoming.

Table 1: An Analysis of the Effects of Conventions and Debates on the Variance Of Presidential Election Polls, 2000 ------------------------------------------------------------------------------------------------------------------ Variable Election Year After Labor Day ------------------------------------------------------------------------------------------------------------------ Convention Season 2.08 --- (24 degrees of freedom) (0.01) Debate Season 0.30 0.48 (15 degrees of freedom) (0.99) (0.94) R-squared 0.29 0.14 Adjusted R-squared 0.08 -.15 Mean Squared Error 8.19 6.54 Number of cases 173 59 ------------------------------------------------------------------------------------------------------------------ Note: The numbers corresponding to the variables are F-statistics. The numbers in parentheses are p values.

Figure 1: Trial-Heat Presidential Polls Aggregated by Date, 2000

Perc

ent G

ore,

Tw

o-C

andi

date

Pre

fere

nces

Days Before Election-300 -200 -100 0

40

50

60

Figure 2: Results of Gallup Polls and All Other Trial-Heat Polls, Labor Day to Election Day, 2000

Perc

ent G

ore,

Tw

o-C

andi

date

Pre

fere

nces

Days Before Election

Gallup Other

-60 -40 -20 0

45

50

55

Figure 3: Results of Voter.com and Washington Post Tracking Polls, Final 20 Days of the Campaign, 2000

Perc

ent G

ore,

Tw

o-C

andi

date

Pre

fere

nces

Days Before Election

Voter.com Washington Post

-20 -10 0

45

50

Christopher Wlezien is University Lecturer and a Fellow of Nuffield College at the University of Oxford. His research and teaching interests encompass a range of fields in American and comparative politics, and his articles have appeared in various journals and edited volumes. He recently finished editing a special issue of Electoral Studies (forthcoming) on “The Future of Election Studies” and has begun writing a book with Robert Erikson on The Timeline of Political Campaigns. Robert S. Erikson is Professor of Political Science at Columbia University. He is coauthor of Statehouse Democracy (Cambridge University Press) and American Public Opinion (Allyn and Bacon). His research on American elections has been published in a wide range of scholarly journals, including the American Political Science Review, American Journal of Political Science, Electoral Studies, Journal of Politics, Legislative Studies Quarterly, and Public Opinion Quarterly. He is the former editor of the American Journal of Political Science.

1 The data were drawn from pollingreport.com. For the exact procedures used, see Wlezien (2001). 2 Specifically, the period of the conventions encompasses all days between July 31 and September 3, inclusive; the period of the debates includes all days between October 3 and October 20. 3 Analysis of polls in 1996 offers similar results. Using the same procedure, conventions and debates account for up to 26.0 percent of the poll variance over the full year; debates account for less than 14 percent of variance during the fall. 4 Of course, Vt is an aggregation of preferences across individuals, and equation 1 summarizes dynamics across individuals i:

vit = ai0 + Bi vit-1 + eit,

where the lower case vit signifies the preference of individual i at time t. This is of consequence for the study of campaigns effects, which we consider very generally below. For specifics, see Erikson and Wlezien (1998). 5 Of course, the rate of decay may vary across different shocks and individuals. In the aggregate this may produce

fractional integration, where effects decay but much more slowly than a stationary series (see Box-Steffensmeier and Smith, 1998; DeBoef, 2000; Lebo, Walker, and Clarke, 2000). Our series of polls then would represent the sum of an integrated series and a fractionally integrated series. 6 Holbrook’s (1996) characterization is similar, though the effect of events is dependent on the error correction component, i.e., whether and the extent to which reported preferences differ from an underlying equilibrium. Put differently, events induce equilibration. 7 Of course, analysis at one level may conceal campaign effects at lower levels, that is, if these effects cancel out. It may be, for example, that presidential candidates campaign most heavily in different states and that this activity has meaningful effects in those states (see Shaw, 1999b), but that they have little net consequence for national preferences. What happens in the states still is of obvious importance. 8 The consequences of sampling error for individual level analyses are much more complex and sobering. See

Zaller (n.d.). 9 Shaw (1999a) provides similar estimates using a larger set of elections. 10 It is important to note that detecting the aggregate effect of an event is more likely if one has regular readings of

preferences, that is, where it is possible to pool results from different polls both before and after the event occurs. This pooling increases statistical power. The degree to which this is true depends on a number of things, however, including the size of the effect and the pooled samples themselves as well as the permanence of the effect and the effects of other events. For example, if the effect of an event decays and other events impact preferences over the period, it may be difficult to detect a fairly large effect even with large pooled pre- and post-event samples. 11 For each poll the expected sampling error is: p (1-p) / N, where p is the proportion voting for, say, the Democratic candidate rather than the Republican. This gives us the estimated error variance for every poll. The error variance for a series of polls is simply the average error variance. The arithmetic difference between the total variance of the poll results themselves and the error variance is the estimated true variance. The ratio of (estimated) true to total variance is the statistical reliability. 12 We nevertheless do know quite a lot, at least for presidential elections. See, e.g., Timpone (1998).

13 Erikson and Wlezien (1999) show that these general differences in the polling universe did not matter in 1996.

14 We should be clear that these numbers are presented for expository purposes. The tracking polls cannot

effectively be used for actual data analysis because the polls reported on consecutive days are not independent: They literally share respondents. Also see Erikson and Wlezien (1999).

15 A less well-known but similar flip-flop occurred right between the Republican and Democratic conventions (not

shown in Figure 2). Here, Bush’s lead dropped from 16 points on one day to 1 point two days later and back up to 16 points five days after that. This accounts for the spike in Gore’s poll share about 90 days before the election in Figure 1. 16 There is one notable exception, namely, Annenberg’s rolling cross-section that accumulated more than 90,000 respondents over the 2000 presidential campaign. Alas, it appears that these data won’t be in the public domain for some time. 17 There is a benefit to weighting, at least under some circumstances. If the variables used—whether party identification or something else—are exogenous to the campaign and actually do structure the vote on Election Day, weighting will reduce sampling error. 18 Note, however, that Voter.com changed its design at least once during the 2000 campaign. 19 This characterization applies to analyses conducted at both the aggregate and individual levels. 20 It is tempting to turn to the multi-day tracking polls now conducted by various organizations, particularly during the fall campaign. Even assuming a reasonable likely voter screen, this is not an appropriate solution unless one has access to the daily readings from which the moving averages are constructed. As noted earlier, the results of multi-day polls reported on consecutive days are not independent—they not only share polling periods, they literally share respondents themselves. Thus, one only can use results for every nth day, e.g., with a three-day tracking poll, using results for every third day. 21 Note that this approach does not provide perfectly independent readings, since the results on consecutive days still will include polls with overlapping reporting periods. 22 For a broader consideration of this other issues, see Franklin and Wlezien (N.d.).