The Role of Personal Availability and Gender in Negative Online Congressional Campaigning

It is worth noting, however, that evidence of the extent to which candidates are using their social and digital media platforms more for personalizing strategies than for traditional campaign content is mixed (McGregor et al., 2017, pp. 266–267).

As Kahn (1996, p. 53) notes in the context of how voters perceive candidates: “If voters are provided with a great deal of information about a candidates’ personality characteristics, then voters are likely to consider this information when developing impressions of candidates.” Such is also the case with campaign opponents, with the added threat that opponents can then spin this information in their own constituent messaging strategies.

Concerning selection biases in terms of candidates having personal campaign websites, Druckman and colleagues note that “[they] were able to identify almost all Senate candidate Web sites and nearly 95% of House sites in ...[the] sample. This suggests that while not all candidates had Web sites, clearly the overwhelming majority did” (2010b, p. 92).

More information on the data collection procedure can be found in Druckman et al. (2010b, pp. 92–93).

Indeed, though politically involved or engaged actors are more likely to visit campaign websites, “voters in general” and “undecided voters” remain the “primary target audiences” on these web platforms (Druckman et al., 2009, p. 346; Druckman et al., 2010b). For instance, there is a strong association between levels and likelihoods of negativity in both website- and television-based campaigning. As Druckman and colleagues note: “Forty-eight percent (351 of 732) of candidates went negative on the Web, compared to 55% in their television advertisements (128 of 232). Candidates are not more likely to go negative on the Web” (2010b, p. 95).

It is interesting to note, however, that campaign strategists reported going negative less often on their websites for the 2016 election cycle than they did from 2008 to 2014 election cycles (Druckman et al., 2018b, p. 400).

Results from the estimated regression models were similar when constructing this variable using principal components analysis and as a simple additive index. Cronbach’s alpha for the six-item additive scale was low (α = .36), which might reflect the fact that the underlying factor space between these six items is a higher dimensional one. The first two dimensions of the PCA space, for example, each have eigenvalues greater than one, and the second dimension in the MCA space accounts for a non-trivial 18% of the total inertia (\(\frac{\chi ^2}{n}\)) in the indicator matrix.

Multiple correspondence analyses were carried out using the FactoMineR package (Husson et al., 2013) in the R statistical computing environment (Team et al., 2013). All visualizations in this paper were made using ggplot2 (Wickham, 2016).

Though Druckman et al. (2010b) found that candidates were much more likely to go negative in their online campaigning in highly competitive races. However, in their study of negative campaigning in Senate races from 1988 to 1998, Lau and Pomper (2001) found that the competitiveness of a race was not a significant predictor of going negative after adjusting for other factors.

Competitiveness was measured using The Cook Political Report (Druckman et al., 2010a, p. 347). Race competition ratings from these reports were used to classify each race according to whether it was “solid Democratic or Republican” (0), “likely Democratic or Republican” (.33), leaning Democratic or Republican” (.67), or a “toss-up” (1).

The issues addressed on each candidate’s website were coded and organized into the following higher-order typology: defense, jobs and the economy, healthcare, education, group advocacy, environment, government reform, immigration, crime, moral/ethical issues, social security, taxes, and government spending (Druckman et al., 2010a, p. 9). These data were then compared to national public opinion polls to create “issue ownership” scores for each issue area. These scores were the estimated percent of the public who believed an issue was “owned” by Republicans or Democrats in that year. These issue scores were then summed per candidate depending on how many times they brought up a particular issue on their website. Each issue score in the summation was also signed according to the candidate’s party affiliation (where, e.g., a Democratic candidate gets a + for an issue they address that is owned more by Democrats and a - for an issue that is owned more by Republicans). The summation was then normalized by the total number of issues brought up on the candidate’s website. The resulting “Issue Ownership” variable reflects the extent to which a candidate in a given race addresses issues more associated with their party, where larger and more positive values indicate more party-aligned issue engagement and more negative values indicate less party-aligned issue engagement (Druckman et al., 2009, p. 349). To index issue salience, each issue area per year was scored as the percent of the public (in that year) that saw that issue area as one of the two most important issues facing the country using public opinion data on issue importance derived from Harris Interactive (Druckman et al., 2009, p. 361). These percentages were then summed for each candidate in a race depending on the number of times the candidate engaged the issue on the site and normalized by the number of issues discussed (Druckman et al., 2009, p. 361). The resulting “Issue Salience” variable is the mean public importance (in %) of the issues addressed by the candidate in that year.

Logit models give similar results to the probit models.

A random effects probit model with campaigns nested within candidates produced virtually identical results to the one-level probit model for the primary coefficients of interest. Lastly, to account for the possibility that the associations of interest here are primarily candidate-level associations rather campaign-level associations, I ran a linear between-effects model—where the proportion of a candidate’s total campaigns wherein they received negativity is regressed on candidate-level means for all covariates. The main coefficients of interest are the same in terms of direction, though the effect size is slightly attenuated when compared to the one-level probit model. However, an unconditional variance decomposition probit model (Singer et al., 2003, p. 75) suggested that very little of the total variance in the negativity reception variable is accounted for at the candidate level—about 10%. Taken together, a simple one-level probit model was chosen over random effects and between effects specifications. These multilevel results are available upon request.

A version of the model with constituency fixed effects—states for Senate candidates and state districts for House candidates—provided similar coefficient estimates. The model fit, however, was poorer with constituency fixed effects (\(AIC_c\) = 796.761 and BIC = 1345.944, though AIC = 625.761) when compared with the reported model (see Table 2). This was also the case for the model 1.

The \(AIC_c\) statistic—a version of AIC that adjusts for the tendency of AIC to suggest overfitting models in the presence of small sample sizes (Hurvich & Tsai, 1989)—was calculated as follows for each model: \(AIC_c = AIC + \frac{2k(k+1)}{n-k-1}\), where k is the number of estimated parameters (including the intercept) and n is the sample size (Burnham & Anderson, 2003, p. 66).

The \(\Delta BIC_{a-m}\) does not suggest any sort of meaningful increase in model fit (\(\Delta BIC_{a-m}\) = − 2.131).

Table versions of the models are available in Table 6 of Appendix 3.

Special thanks to the anonymous reviewer who suggested this example.

An incumbent was defined as a candidate who was neither a challenger nor an open-seat candidate.