Negative Advertising and the Dynamics of Candidate Support

It is also possible that negative messages fail to influence public opinion.

Doherty and Adler (2014) report evidence that campaign mailers produce similar short-term effects on public opinion.

Kahn and Kenney (1999) argue that campaigns should stimulate turnout as they become more negative, but only until the election environment becomes overly saturated with negative information. When this happens, they argue that turnout will decline.

That said, the results in this literature are decidedly mixed. See Lau et al. (2007) for a thorough meta-analysis on negativity’s effects on turnout and other topics.

The above argument centers on the support enjoyed by competing candidates relative to one another because this is the focus of our analysis. We could also consider the level of support enjoyed by individual candidates. Citizens, for example, may be less likely to support candidates who attack while their likelihood of supporting the targeted candidates may be unaffected. However the process unfolds, candidates who attack their opponents should suffer.

Because of the nature of our public opinion polling data—the Democratic candidate’s share of the two-party intended vote—that we discuss below, we are unable to test this theoretical possibility independently of the above hypotheses. Still, it provides an additional reason why we might observe higher or lower levels of support for one candidate relative to the other.

The data were obtained from a project of the University of Wisconsin Advertising Project includes media tracking data from TNSMI/Campaign Media Analysis Group in Washington, D.C. The University of Wisconsin Advertising Project was sponsored by a grant from The Pew Charitable Trusts. The opinions expressed in this article are those of the authors and do not necessarily reflect the views of the University of Wisconsin Advertising Project or The Pew Charitable Trusts.

We treated Paul Wellstone and Walter Mondale as a single candidate in Minnesota’s 2002 Senate race due to the former’s death late in the campaign. Excluding the race from our analysis does not alter the substantive findings we report in this research.

This eliminated four contests. The Democrat in the 2004 Senate election in Ohio aired 69 advertisements. In the other three contests, at least one of the candidates aired fewer than 10 ads. This choice did not alter the substance of the findings reported in this research.

We do not include advertisements sponsored by political parties or interest groups in our analyses because this research focuses only on the political impact of candidate strategy. We also do not include any advertisements aired by 501(c) groups, which did not become prominent parts of campaigns for federal offices until the 2004 election cycle.

An additional 28 and 22 % of Democratic and Republican-sponsored spots were coded as contrast ads. All but a very small number of the remaining advertisements were coded as positive.

National Journal data and Polling Report data are subscription based polling agencies. Website access can be found at www.nationaljournal.com and www.pollingreport.com.

WCalc can estimate daily, monthly, quarterly, annual, or multi-year series of polling figures.

See the Wcalc manual at http://​stimson.​web.​unc.​edu/​files/​2015/​08/​Wcalc6 for more detailed information. Stimson notes that WCalc “implements the Dyad Ratios algorithm for building a continuous regular time series from the scraps of dated survey results that are typically available for public opinion analysis. It emulates the logic of principal components analysis in most regards but, unlike principal components, it does not require that all variables have a complete set of cases or, indeed, anything close to it. Input is survey data expressed in a summary score, e.g., percent liberal responses, which is dated and includes a number of cases (although the program will accept 0 for N when it is unknown).”

Grant and Lebo (2016) argue that ECMs are often misunderstood in the literature and make several recommendations for practitioners. We address their concerns in an appendix.

Multiple equation VAR models are equivalent to SUR models as long as the models exhibit the same lag structure in each equation (Hamilton 1994, p. 314).

Many scholars refer to these coefficients as long-term effects because they can be transformed algebraically to produce estimated effects on the dependent variable across additional future time periods.

We calculate the standard errors of each long run multiplier using the Bewley (1979) transformation (see also DeBoef and Keele 2008).

Replication materials for all of the analyses presented in the manuscript and its appendix are available at https://​dataverse.​harvard.​edu/​dataset.​xhtml?​persistentId=​doi:​10.​7910/​DVN/​U1GGGQ.

Error correction models are algebraically equivalent to auto distributed lag models. Because the latter allow researchers to easily calculate the effect of a change in the independent variable on the value of the dependent variable in individual time periods, this can be done in an error correction context as well with a bit of simple algebra (DeBoef and Keele 2008). Researchers need to use three coefficients from the output of an ECM to calculate these effects: those of lagged Y, lagged X, and the first difference of X. Using the results of the equation predicting Democratic support from Table 3 as an example, that means that we use the coefficients generated for lagged Democratic polling advantage (\(-0.03\)), lagged Democratic attack advantage (\(-0.003\)), and differenced Democratic attack advantage (0.008). Assume that we are interested in calculating the distributed effects of a one unit increase in the Democratic attack advantage on Democratic polling advantage. The contemporaneous effect is captured by the estimated coefficient of the differenced Democratic attack advantage indicator: 0.008 \(\times\) 1 = 0.008. Thus Democratic support increases by 0.008 units when the attack environment increases by one unit during the week in which the change occurred. We can also calculate the effects of this change that are distributed over future weeks. In the week following the change in the Democratic attack advantage, this effect can be calculated by subtracting the coefficient of differenced Democratic attack advantage indicator from the coefficient of lagged Democratic attack advantage measure and multiplying that quantity by the change that we are interested in, in this case an increase of one unit. We then add that quantity to one minus the absolute value of the coefficient of lagged Democratic polling advantage, which is multiplied by the contemporaneous change. Algebraically, we can express that as follows: (\(-0.003 - 0.008\)) \(\times 1 + (1 - |-0.03|) \times 0.008 = -0.003\). The formula to calculate the effect of a change in the Democratic attack advantage on Democratic polling advantage in all periods after time t + 1 is simpler: \((1 - |-0.03|) \times\) g, where g is the distributed effect calculated for the previous time period. Thus the effect during the second week after the change in the Democratic attack advantage is (1 − |−0.03|) \(\times\) −0.003 = −0.0029, in the third week it is (1 − |−0.03|) \(\times\) −0.0029 = −0.0028, and so on. The the effect of a one unit increase in the Democratic attack advantage leads to an immediate increase in Democratic polling advantage followed by several small decreases in Democratic polling advantage over future weeks. Note that when the estimated coefficient of the lagged dependent variable is small, the effect of a change in an independent variable on the value of the dependent variable decay slowly over time.