Campaign News and Vote Intentions

Abstract

This paper explores the relationship between campaign news and vote intentions, drawing on manual content analyses from the 2004 and 2006 Canadian federal elections. The content analysis is designed to capture, among other items, the ‘tone’ of coverage for parties and leaders. Combining time series of ‘tone’ and commercial polling results, econometric methods are then used to build a model of vote intentions. Media content appears to explain a good deal of the over‐time variance in vote intentions. Results are discussed as they pertain to two versions of the media‐opinion relationship: (1) media content captures and arranges in a readily quantifiable form the evolving mood of the campaign, or (2) media do not simply reflect, but affect vote intentions.

Acknowledgements

A previous version of this paper was presented at the Annual Meeting of the Canadian Political Science Association, Vancouver BC, 4–6 June 2008. We are grateful to Mark Pickup for providing the dataset of Canadian election polling results, to John Galbraith for comments on an early draft, and to the Journal’s anonymous reviewers. This work was funded in part by the Fonds québécois de la recherche sur la société et la culture, the Donner Canadian Foundation, the McGill‐Max Bell Strategic Initiative, and the McGill Institute for the Study of Canada.

Notes

1. Horserace coverage has received a particularly large amount of attention in the literature (see, e.g., Craig, 2000; Fletcher, 1981, 1991; Graber, 1976; Jamieson, 1992; Mendelsohn, 1993; Patterson, 1993; Wilson, 1980).

2. There are of course vast literatures on the link between media and opinion in campaigns (for work in Canada see, e.g., Blais & Boyer, 1996; Mendelsohn, 1994, 1996; Mendelsohn & Nadeau, 1997; Johnston et al., 1992; Wagenberg et al., 1988; for work elsewhere see, e.g., Brians & Wattenberg, 1996; Druckman, 2004; Krosnick & Kinder, 1990; for a more thorough review of the earlier US literature, see Weaver, 1996).

3. All polls are of course in the field for several days, so we have indexed polls for this dataset in a way that captures as accurately as possible the timing of shifts in opinion. For polls in the field over a three‐day period (most are), we use the middle day. For polls in the field for two or four days, we use the second or third day, respectively.

4. These data are distributed by the Media Observatory at the McGill Institute for the Study of Canada, available online at ⟨http://www.mcgill.ca/misc/research/media-observatory⟩.

5. All measures included in this analysis achieved an appropriate level of reliability. Detailed methodological information is available at the website for the Media Observatory at the McGill Institute for the Study of Canada.

6. We consider, for the purposes of this article, only stories mentioning at least one of the two major parties and their leaders. This includes the incumbent Liberal Party (Paul Martin), and the Conservatives (Stephen Harper), the largest opposition party represented in the House of Commons prior to each election campaign kickoff.

7. This is not the only conceivable measure of campaign’s tone or mood, of course (see Ridout & Franz, 2008, for an overview of campaign tone measurement in the media and politics literature.

8. The manually coded data used here capture horserace coverage generally, but actually do not record direct mentions of polls. That said, the proportion of coverage dealing with polls is relatively easy to assess using a simple automated search. In this case, we rely on an existing dataset of all election‐related stories in the five English‐language newspapers (available through the Media Observatory at the McGill Institute for the Study of Canada), and a search identifying all articles mentioning the word “poll” more than once.

9. And note that lagged poll results are included in the final prediction model (discussed in the Analysis section). If lagged media content matters to vote share predictions, above and beyond lagged vote shares themselves, then there is strong evidence that there is some other component of media content (that is, not just polling reports) that matters.

10. We have also conducted a series of analyses making use of Bayesian statistics, similar to recent work by Pickup and Johnston (2007) and Jackman (2005). Using vague priors centered at zero with a low level of precision, the performance of our predictions is for the most part unaffected.

11. Note also that there is another aspect of error in polling results that is rarely discussed but that is often implicitly accepted in public opinion research. Because respondents may be in a better position to accurately express vote intention later in the campaign, survey responses may be more reliable closer to the election date. As a consequence, time series of vote intentions may be prone to temporal heteroskedasticity – random error variance may not be constant across time. This violates a critical assumption of OLS, but we do not address it here. We simply assume temporal homoskedasticity.

12. Note that for “net tone” we lump together all articles from all newspapers – we do not give newspapers different weights based on audience reach, nor do we distinguish between the potentially different content in different newspapers. While newspapers differ in levels of tone for different parties (Soroka & Andrew, 2010), however, they follow very similar trends over the campaign. It is not clear that there is much to gain by looking at newspapers separately.

13. This choice was based on a series of cross‐correlograms exploring the relationships between (a) the two party vote share (dependent) variables, and (b) each of the four leader and party tone (independent) variables, in both 2004 and 2006. That said, the structure of our prediction models present some difficulties for diagnostics such as bivariate cross‐correlograms, since in each model a single party’s vote share is regressed on four independent variables: lagged values of both party and leader tone, for both parties. Our decision to use lags 4–6 was thus also based in large part on tests of various alternative forms of the final prediction models, e.g., lags 3–5, lags 5–7, a series of models using just two lags, and so on.

14. Preliminary tests confirmed that using a single lag for vote intentions was all that was required – once vote intentions at t–4 are included, vote intentions at t–5 and t–6 have no substantive effect.

15. Note that these results cannot rule out the possibility that media content is at least partly affected by prior opinion (see also the discussion of news stories about polls in the Data section above). Our focus on prediction makes this somewhat of a secondary issue. That said, relatively simple Granger causality tests in which each series is regressed on lagged values of itself and the other series suggest that media does in most cases “Granger‐cause” opinion. And the fact that media matter, above and beyond lagged opinion (see results below), also suggest a largely uni‐directional causal effect.

16. Note that much of the campaign tone during these campaigns was negative. Bad news typically outweighed good news stories for all the major parties and leaders, as we would expect. The negativity bias in election reporting does not change our expectation. But, in practical terms, it does means that we expect that less negative Conservative tone is positively related to Conservative vote intentions and vice versa.

17. On the value of the MAE and SEE (Standard Error of the Estimate) as goodness of fit measures in prediction and forecasting, see Krueger & Lewis‐Beck, 2005.

18. There exist many such dictionaries (see, e.g., Stone et al., 1966; Strapparava & Valitutti, 2004; Mergenthaler, 1996, 2008; Pennebaker et al., 2001; Martindale, 1975, 1990; Whissell, 1989; Bradley & Lang, 1999).

19. On proximity‐based automated assessments of tone, see, e.g., Pang et al., 2002; Mullen & Collier, 2004.

20. It is worth stressing that this finding is based on newspaper‐supplied election news. It is reasonable to expect that our results will actually strengthen as we incorporate news from other sources.

21. On “opinion leaders”, see work on the two‐step flow in political communications (esp. Lazarsfeld et al., 1944 and Katz & Lazarsfeld, 1955).

22. Assuming the movement is not just random, of course; that is, assuming that movement has something to do with the campaign.