The Influence of Communication Context on Political Cognition in Presidential Campaigns: A Geospatial Analysis

Abstract

Due to targeting strategies employed by political campaigns, campaign intensity is not uniform across the whole country. This study investigates how an individual's communication context, defined by geospatial characteristics created by campaigns, would influence his or her political learning. Data for this study come from three separate studies conducted during the 2000 U.S. presidential election. The results from a series of multilevel modeling analyses indicate that contextual-level political advertising and candidate appearances moderate the relationship between newspaper use and political knowledge, and the relationship between political discussion and political knowledge. This study not only demonstrates that conditional communication effects hinge on geospatial factors but also helps to develop contextual theories of communication that specifically address effects of contextual factors and cross-level interactions.

ACKNOWLEDGMENTS

I thank the two anonymous reviewers and the editor Dr. Perry for their constructive comments on previous versions of this article. I also thank Daron Shaw for providing a portion of the data for the study. The NAES, the Wisconsin Advertising Project, and Professor Shaw bear no responsibility for the findings, opinions, and conclusions in this article.

Notes

¹According to The Nielsen Company, there were 211 media markets in 2000. The Wisconsin Advertising Project indicated that more than 80% of the U.S. population lives in these 75 media markets.

²Some may be concerned about the reliability statistics for the measures of news use variables. This study chose to use the combined scales of news exposure and attention because this is guided by not only communication theories but also an established method in political communication research. Conducting a study to specifically assess this issue, Eveland, Hutchens, and Shen (2009) concluded that the best approach is the one used in this study.

³The following formula was used to calculate the values of political advertising (the total spots aired and the estimated cost of ads) of a state with n media markets. A state's total ad spending = (total spending in the 1st media market ∗ proportion of population in the 1st media market) + (total spending in the 2nd media market ∗ proportion of population in the 2nd media market) + … + (total spending in the nth media market ∗ proportion of population in the nth media market). A single state may contain multiple media markets, and a media market may cover an area across state lines. Thus, it should be noted that (a) media markets within a state may have different levels of advertising magnitude, and cost for ad broadcast varies from media market to media market because some places are more expensive than others. Population adjustments can be made to better reflect statewide advertising intensity, whose effects on voters can then be more accurately detected and compared from state to state. The information about population of residents in each media market in 2000 was obtained from Polidata Demographic and Political Guides (n.d.). (b) The value of a media market was used in all calculations of states which that market covers. (c) The media markets that are not among those studied by the Wisconsin project were treated as missing values in the calculation. Similar formulas for generating statewide advertising values from media market data with population adjustments can be found in other studies (e.g., Hill & McKee, 2005; Shaw, 1999a, 1999b).

Note. DMA = designated market area.

⁴During the process of combining the individual-level and the campaign-level variables, listwise deletion was performed to delete the missing values in the data files. To conserve cases, each campaign-level variable was combined with the data file of the individual-level variables, respectively, by either state or media market. Therefore, there were six data files to be analyzed by the multilevel modeling method.

⁵All the independent and control variables at the individual level, and the independent variables at the campaign level, were grand-mean centered in the models. T. Snijders and R. Bosker (as cited in Luke, 2004) argued that the differences between full maximum likelihood estimation and restricted maximum likelihood estimation are very small, especially when the number of level-2 units is large (i.e., 30 or more). Thus, all the analyses were conducted by the use of full maximum likelihood estimation because the numbers of the campaign-level units in all six data files are more than 30. Moreover, all the results reported here are with robust standard errors because the number of the campaign-level units is relatively large. Robust standard errors are appropriate when the number of units at the higher level is large (Raudenbush & Bryk, 2002). Besides, in all the analyses, there was no significant discrepancy between model-based standard errors and robust standard errors; therefore, it can be said that there is no indication of model misspecification (Raudenbush & Bryk, 2002).

⁶The intraclass correlation coefficient (ICC) ρ = τ ₀₀ /(τ ₀₀  + σ²) was calculated to see how much variance in political knowledge could be attributed to between-state or between-media market differences. It is found that about 1 or 2% of the explainable variance in political knowledge can be attributed to differences across communication contexts. Kreft and de Leeuw (1998) argued that a small ICC can increase the alpha level (the assumed Type I error probability 0.05), and this increase is significant, particularly when the number of lower level units within a higher level unit is large. They suggested that multilevel modeling should be used even in the case of small ICCs. Hayes (2006) also argued that it is beneficial to use multilevel modeling “even when the ICC is near zero” (p. 394).

Note. Values are HLM coefficients of fixed effects with standard errors in parentheses. Although not reported here, statistics of random effects of all the models tested are available upon request from the author. DMA = designated market area; γ ₀₀  = grand mean. γ ₀₁  = effect of campaign-level predictor (i.e., ad frequencies, ad expenditures, and candidate appearances).

*p < .05. ***p < .001.

⁷When all the individual-level variables were controlled in the model, an additional main effect of designated market area ad spending (γ = .0003, SE = .0001, p < .01) was found along with the main effect of designated market area candidate appearances (γ = .2018, SE = .0865, p < .05).

⁸The hypothesis testing method of this study should have a sufficient power to catch effects when they do exist. If the principles of single-level power analysis are followed, this study is like a census because the samples cover almost the entire population, and therefore the standard errors should be small. Scherbaum and Ferreter (2009) argued that “in multilevel models, power is not a simple monotonic function of the sample sizes at either level when holding the other factors constant” (p. 352). Researchers have proposed general rules for statistical power in multilevel modeling. I. Kreft (as cited in Scherbaum & Ferreter, 2009) suggested a 30/30 rule—at least 30 groups with 30 individuals within each group. Hox (1998, 2002) advocated a 50/20 rule—50 groups with 20 individuals in each group. It can be argued that the sample sizes in this study almost conform to these rules of thumb. Power computations were conducted for simple main effects (H1a and H1b) to further check whether the tests performed in this study have a sufficient power. All the computations used the formula Z _1 − β ≤ (Effect Size/Standard Error) − Z _1 − α/2 (Scherbaum & Ferreter, 2009) and assumed a medium effect size (ES = .50) and a two-tailed alpha of .05. The results show that the statistical power is pretty high (about 0.99) across all the tests.

Note. Values are HLM coefficients of fixed effects with standard errors in parentheses. Although not reported here, statistics of control variables (i.e., age, gender, education, and income) and random effects of all the models tested are available upon request from the author. DMA = designated market area; γ ₀₀  = grand mean; γ ₀₁  = effect of campaign-level predictor (i.e., ad frequencies, ad expenditures, and candidate appearances); γ ₁₀ , γ ₂₀ , γ ₃₀ , γ ₄₀  = effects of individual-level predictors (i.e., NP, network/cable TV, local TV, and discussion); γ ₁₁ , γ ₂₁ , γ ₃₁ , γ ₄₁  = effects of cross-level interactions.

*p < .05. **p < .01. ***p < .001.

⁹Due to the concern of multicollinearity, the cross-level interaction terms were entered into the models one at a time to see if results would be the same. The same four statistically significant interaction effects were found: NP Use × State Ad Spending (γ = .0003, SE = .0001, p = .010), NP Use × State Candidate Appearances (γ = .0329, SE = .0163, p = .050), Discussion × DMA Ad Spending (γ = .0001, SE = .0000, p < .01), and Discussion × DMA Candidate Appearances (γ = .0867, SE = .0403, p < .05).

¹⁰Due to the concern of the marginally reliable scales of news use, only the attention measures instead of the combined indexes of news exposure and attention were used in the models to see if results would be the same. An additional main effect of DMA ad spending (γ = .0003, SE = .0001, p < .05) was found along with the main effect of DMA candidate appearances (γ = .3150, SE = .1431, p < .05). Two of the same statistically significant interaction effects were found: Discussion × DMA Ad Spending (γ = .0002, SE = .0001, p < .01), and Discussion × DMA Candidate Appearances (γ = .1236, SE = .0589, p < .05). In addition, there was some evidence for H3a. Network and cable TV news use interacted with state ad frequencies (γ = −.0006, SE = .0003, p < .05), DMA ad frequencies (γ = −.0009, SE = .0003, p < .01), state ad spending (γ = −.0006, SE = .0003, p < .05), and DMA ad spending (γ = −.0006, SE = .0003, p < .05). Given that attention, particularly to electronic media, is the more important aspect of the measures, future research is needed to further assess these relationships.

¹¹Table 3 shows that newspaper use, network and cable TV news use, and political discussion are positively related to political knowledge across all the models, whereas local TV news use is negatively related to knowledge in all the models.

Additional information

Notes on contributors

Yung-I Liu

Yung-I Liu (Ph.D., The Ohio State University, 2008) is an Assistant Professor in the School of Communication at Cleveland State University. Her research interests include political campaign communication, strategic communication, and advertising.