Skip to main content
Log in

The role of slant and message consistency in political advertising effectiveness: evidence from the 2016 presidential election

  • Original Research
  • Published:
Quantitative Marketing and Economics Aims and scope Submit manuscript


We explore the relationship between the content of political advertising on television and ad effectiveness. Specifically, we investigate how slant – the extremeness of the message – and consistency with the candidate’s primary campaign messaging in national ad buys relate to two measures of voter behavior: online word-of-mouth (WOM) and voter preference (captured through daily polls) for the candidates. Using data from the 2016 presidential election, we find that ad messages that are more (1) centrist and (2) consistent with a candidate’s primary-election platform associate with increases in online WOM and voter preference for the candidate. We further find that consistency is more important in the early (pre-October) stages of the campaign. Our results suggest that while there may be a benefit to candidates moderating their message after winning the primary election, they need to be careful about shedding their messaging from the primary election during the early stages of the general election. Additionally, our results enrich our understanding of the use of extreme messaging in political advertising, a phenomenon that is on the rise, by showing that it may have a cost of decreased candidate-related WOM and voter preference for the candidate.

This is a preview of subscription content, log in via an institution to check access.

Access this article

Price excludes VAT (USA)
Tax calculation will be finalised during checkout.

Instant access to the full article PDF.

Fig. 1
Fig. 2
Fig. 3

Similar content being viewed by others


  1. The WOM measures are available only nationally. While there are state-level polls, these tend to be conducted very irregularly, and by a number of different pollsters, making the creation of a high-quality state-specific panel very difficult (and nearly impossible to do so at the even more granular DMA level).

  2. We probe the overlap of national versus local primetime political ads in our data using the name of the TV creative and find significant overlap. This suggest that the content of national ads likely reflect the campaigns’ overall message. Specifically, 100% of both Clinton and Trump’s national primetime ad creatives in our data were also aired locally in primetime. That said, the candidates do employ more ad creatives for their primetime local ad airings. Yet, the majority of their local primetime airtime consists of ad creatives that are aired both locally and nationally. Specifically, only 3.1% (9.2%) of Clinton’s (Trump’s) primetime local ad airings consist of ad creatives that were only aired locally.

    Additionally, we find that Clinton and Trump’s national primetime ads exhibit very similar airing patterns as their local primetime ads in terms of time, day of the week, and month aired (see Appendix 7).

  3. Among the words in the congressional records that remain after pre-processing, we focus on words that appear at least 2 times and fewer than 100 times in the candidates’ public speech documents and transcribed ad texts. We then select 1000 words for analysis that are found to be most asymmetrically used by the parties. We conduct sensitivity analysis and find that our results are robust to the number of words in the analysis.

  4. We consider the Republican candidate’s vote share in the 2012 presidential election in the state for senator and electoral district for representative.

  5. We consider unigrams for our implementation. Although higher-order n-grams capture richer information in general, this severely reduces both the total number and variety of n-grams that we can extract from ads due to their shortness.

  6. Similar to Wu et al. (2019), we conduct two types of validation checks. First, we treat the training documents (i.e., 62 documents that we use as inputs) as if they were new, and infer vectors for the documents using the trained model, and see whether the inferred documents are found to be most similar to themselves via cosine similarity. We find that all inferred documents are most similar to themselves. Second, in order to showcase the face validity of our approach to measuring consistency, we select a few documents and show documents that are most and least similar in terms of the cosine similarity, excluding the chosen document. See Appendix 2.

  7. We do not remove stop words, because dov2vec deals with frequent word, which are akin to stop words, by randomly down-sampling high frequency words.

  8. For parameters, we use vector size = 200, epochs = 300, window = 5, sub-sampling = 10−2, and negative = 5.

  9. While our approach offers many benefits, one caveat is that the ad texts are very short. This may lead to some measurement error in our measures of slant and consistency. It is likely that any coding mechanism would also have some measurement error. Given that slant and consistency are independent variables, whatever measurement error does exist would be expected, on average, to lead to attenuation of coefficients.

  10. For robustness, we also run the analysis using 2-min and 3-min time windows. See Appendix 4.

  11. This is done through the inclusion of a Clinton dummy variable. We consider only whether the ad is pro-Clinton or pro-Trump, and not whether the ad was sponsored by the candidate or by a supporting PAC because the vast majority of Trump’s ads were PAC ads while the vast majority of Clinton’s ads were run by the campaign. Thus, the Clinton coefficient captures both the difference between Clinton and Trump as well as the difference between candidate ads and PAC ads.

  12. As there are many networks in the data, we group networks with fewer than 7 ad airings together as “Other Networks.”

  13. We cluster at the candidate level to allow for the WOM residuals to correlate within candidates. However, we could instead cluster at the ad creative level, which is the level at which the slant and consistency measures vary. To guide our decision, we conduct statistical tests for the appropriate level of clustering (MacKinnon et al., 2020) and find support for clustering at the candidate level. This matches the recommendation of Cameron and Miller (2015) to cluster at a more aggregate level rather than a more disaggregate level. In Appendix 5, we document the results of the statistical tests for clustering.

  14. As there are many small networks with only a few ad airings, we select the top 10 networks by the number of ad airings and group the rest as “Other Networks.”

  15. The estimation code can be found at

  16. As noted in the introduction, it is also possible that the ads prompt a flurry of tweets that would have been sent at a later time in the absence of the ad.

  17. We add a handful of words to the existing list of stop words from NLTK in Python, such as madam, speaker, and thank. These words appear frequently but are not informative of one’s political ideology.

  18. The chi-square statistic measures the extent to which a given word is used with asymmetric frequencies by parties.

  19. Contrary to Gentzkow and Shapiro (2010) who used newspaper articles, ad texts are typically short. Therefore, in order to overcome the scarcity of words, we use both the candidates’ speeches and ad texts to select words to consider.

  20. We consider the Republican candidate’s vote share in the state for senators and congressional district for congresswomen and congressmen.

  21. This is equivalent to regressing \(\tilde{f}_{pn} - \alpha_{p}\) on βp.


  • Adgate, B. (2019). The 2020 elections will set (another) ad spending record. Forbes. Accessed at on Oct. 9, 2020.

  • Baer, J., & Sinagoga, D. (2018). Why TV still totally dominates when it comes to political advertising. AdAge. Accessed at on Oct. 9, 2020.

  • Bartels, L. (2016). The rise of presidential extremists, The New York Times, Accessed at on 9 Oct 2020.

  • Bass, F. M., Bruce, N., Majumdar, S., & Murthi, B. P. S. (2007). Wearout effects of different advertising themes: A dynamic Bayesian model of the advertising-sales relationship. Marketing Science, 26(2), 179–195.

    Article  Google Scholar 

  • Berger, J., Humphreys, A., Ludwig, S., Moe, W. W., Netzer, O., & Schweidel, D. A. (2020). Uniting the tribes: Using text for marketing insight. Journal of Marketing., 84(1), 1–25.

    Article  Google Scholar 

  • Berman, R., Melumad, S., Humphrey, C., & Meyer, R. (2019). A tale of two Twitterspheres: Political microblogging during and after the 2016 primary and presidential debates. Journal of Marketing Research, 56(6), 895–917.

    Article  Google Scholar 

  • Bermingham, A., & Smeaton, A. (2011). On using Twitter to monitor political sentiment and predict election results. Proceedings of the Workshop on Sentiment Analysis where AI meets Psychology (SAAIP 2011) (Asian federation of natural language processing, Chiang Mai, Thailand), 2–10.

  • Boush, D. M., & Loken, B. (1991). A process-tracing study of brand extension evaluation. Journal of Marketing Research, 28(1), 16–28.

    Article  Google Scholar 

  • Broniarczyk, S. M., & Alba, J. W. (1994). The importance of the brand in brand extension. Journal of Marketing Research, 31(2), 214–228.

    Article  Google Scholar 

  • Cameron, A. C., & Miller, D. L. (2015). A Practitioner’s guide to cluster-robust inference. Journal of Human Resources., 50(2), 317–372.

    Article  Google Scholar 

  • Comscore TV Essentials. (2020). Average audience, June 2016-November 2016, US.

  • Conger, K. (2019). Twitter will ban all political ads, C.E.O. Jack Dorsey says. The New York Times. Accessed at on Oct. 9, 2020.

  • Deng, Y., & Mela, C. F. (2018). TV viewing and advertising targeting. Journal of Marketing Research, 55(1), 99–118.

    Article  Google Scholar 

  • Feng, S. (2020). The proximity of ideas: An analysis of patent text using machine learning. PLoS One, 15(7), e0234880.

    Article  Google Scholar 

  • Fossen, B. L., Mallapragada, G., & De, A. (2020). Impact of political television advertisements on Viewer’s response to subsequent advertisements. Marketing Science. Forthcoming.

  • Fossen, B. L., & Schweidel, D. A. (2017). Television advertising and online word-of-mouth: An empirical investigation of social TV activity. Marketing Science, 36(1), 105–123.

    Article  Google Scholar 

  • Gentzkow, M., & Shapiro, J. M. (2010). What drives media slant? Evidence from US daily newspapers. Econometrica, 78(1), 35–71.

    Article  Google Scholar 

  • Gordon, B. R., & Hartmann, W. R. (2013). Advertising effects in presidential elections. Marketing Science, 32(1), 19–35.

    Article  Google Scholar 

  • Haggin, P. (2019). Google to restrict political ad targeting on its platforms. The wall street journal. Accessed at on 9 Oct 2020.

  • Huber, G. A., & Arceneaux, K. (2007). Identifying the persuasive effects of presidential advertising. American Journal of Political Science., 51(4), 957–977.

    Article  Google Scholar 

  • Hummel, P. (2010). Flip-flopping from primaries to general elections. Journal of Public Economics., 94(11–12), 1020–1027.

    Article  Google Scholar 

  • Jahanbakhsh, K., & Moon, Y. (2014). The predictive power of social media: On the predictability of us presidential elections using twitter. Available at [cs.SI].

  • Jennings, W., & Wlezien, C. (2018). Election polling errors across time and space. Nature Human Behaviour, 2(4), 276–283.

    Article  Google Scholar 

  • Joo, M., Wilbur, K. C., Cowgill, B., & Zhu, Y. (2014). Television advertising and online search. Management Science., 60(1), 56–73.

    Article  Google Scholar 

  • Katz, H. E. (2013). The media handbook: A complete guide to advertising media selection, planning, research, and buying. Routledge.

    Google Scholar 

  • Kaye, K. (2017). Data-driven targeting creates huge 2016 political ad shift. AdAge. Accessed at on 9 Oct 2020.

  • Keller, K. L., & Lehmann, D. E. (2006). Brands and branding: Research findings and future priorities. Marketing Science, 25(6), 740–759.

    Article  Google Scholar 

  • Kennedy, C., Blumenthal, M., Clement, S., Clinton, J. D., Durand, C., Franklin, C., McGeeney, K., Miringoff, L., Olson, K., Rivers, D., Saad, L., Witt, G. E., & Wlezien, C. (2018). An evaluation of the 2016 election polls in the United States. Public Opinion Quarterly., 82(1), 1–33.

    Article  Google Scholar 

  • Kotler, P. T., & Keller, K. L. (2016). Marketing Management (15th ed.). Pearson.

    Google Scholar 

  • Le, Q. V., & Mikolov, T. (2014). Distributed representations of sentences and documents. Proceedings of Machine Learning Research., 32(2), 1188–1196.

    Google Scholar 

  • Lewis, R. A., & Reiley, D. H. (2013). Down-to-the-minute effects of super bowl advertising on online search behavior. Proceedings of the Fourteenth ACM Conference on Electronic Commerce, 9(4), 639–656.

    Article  Google Scholar 

  • Liaukonyte, J., Teixeira, T., & Wilbur, K. C. (2015). Television advertising and online shopping. Marketing Science, 34(3), 311–330.

    Article  Google Scholar 

  • Liu, Y., Li, K. J., Chen, H., & Balachander, S. (2017). The effects of products’ aesthetic design on demand and marketing-mix effectiveness: The role of segment prototypicality and brand consistency. Journal of Marketing., 81(1), 83–102.

    Article  Google Scholar 

  • Lovett, M. J., & Shachar, R. (2011). The seeds of negativity: Knowledge and money. Marketing Science, 30(3), 430–446.

    Article  Google Scholar 

  • MacKinnon, J. G., Nielsen, M. Ø., & Webb, M. D. (2020). Testing for the appropriate level of clustering in linear regression models. Queen’s economics department working paper no. 1428.

  • McGranaghan, M., Liaukonyte, J., Fisher, G., & Wilbur, K. C. (2019). Lead Offer Spillovers. Marketing Science, 38(4), 643–668.

    Article  Google Scholar 

  • Mikolov, T., Sutskever, I., Chen, K., Corrado, G. S., & Dean, J. (2013). Distributed representations of words and phrases and their compositionality. Advances in Neural Information Processing Systems, 26, 3111–3119.

    Google Scholar 

  • Miller, J. (2015). Report: TV ad spending in 2016 election to reach $4.4 billion. CBS News. Accessed at on 15 Sept 2020.

  • OpenSecretes. (2020) Expenditures. Accessed at on 9 Oct 2020.

  • Park, C. W., Milberg, S., & Lawson, R. (1991). Evaluation of brand extensions: The role of product feature similarity and brand concept consistency. Journal of consumer research., 18(2), 185–193.

    Article  Google Scholar 

  • Passwaiter, S. (2018). 2018 campaign ad spend will be in the billions. Cook Political Report. Accessed at on 15 Sept 2029.

  • Passwaiter, S. (2020). Political ad spending this year reached a whopping $8.5 billion. Advertising age. Accessed at on 27 May 2021.

  • Silver, N. (2018). The Polls Are All Right. Accessed at on 9 Oct 2020.

  • Spenkuch, J. L., & Toniatti, D. (2018). Political advertising and election results. The Quarterly Journal of Economics., 133(4), 1981–2036.

    Article  Google Scholar 

  • Thomson, M. (2006). Human brands: Investigating antecedents to consumers’ strong attachments to celebrities. Journal of Marketing., 70(3), 104–119.

    Article  Google Scholar 

  • Tirunillai, S., & Tellis, G. J. (2017). Does offline TV advertising affect online chatter? Quasi-experimental analysis using synthetic control. Marketing Science, 36(6), 862–878.

    Article  Google Scholar 

  • Wang, Y., Lewis, M., & Schweidel, D. A. (2018). A border strategy analysis of ad source and message tone in senatorial campaigns. Marketing Science, 37(3), 333–355.

    Article  Google Scholar 

  • Wells, G., & Seetharaman, D. (2018). Coming soon to Facebook: Lots of extreme political ads. The wall street journal. Accessed at on 9 Oct 2020.

  • Wilbur, K. C., Xu, L., & Kempe, D. (2013). Correcting audience externalities in television advertising. Marketing Science, 32(6), 892–912.

    Article  Google Scholar 

  • Wu, L., Wang, D., & Evans, J. A. (2019). Large teams develop and small teams disrupt science and technology. Nature., 566(7744), 378–382.

    Article  Google Scholar 

  • Zhang, L., & Chung, D. J. (2020). The air war vs. the ground game: An analysis of multichannel marketing in US presidential elections. Marketing Science, 39(5), 872–892.

    Article  Google Scholar 

Download references

Author information

Authors and Affiliations


Corresponding author

Correspondence to Raphael Thomadsen.

Ethics declarations

Conflict of interest

The authors did not receive support from any organization for the submitted work. The authors have no relevant financial or non-financial interests to disclose.

Additional information

Publisher’s note

Springer Nature remains neutral with regard to jurisdictional claims in published maps and institutional affiliations.

Authors contributed equally and are listed in alphabetical order.


Appendix 1. Slant variable details

In this appendix, we present details on the calculations of the slant index.

1.1 Political slant

We first pre-process texts using the NLKT module in Python: we make words lower case and remove stop words,Footnote 17 punctuations, and numbers. We then tokenize each of the texts and stem the words in the text.

Following Gentzkow and Shapiro (2010) and their notations for consistency, we first derive the mapping between a vector of word counts a congressperson used and the political leanings of their district. For each word p in the 114th Congressional Record, we count the number of times the word p is used by each of the two parties and calculate a chi-square statistic, \({\mathcal{x}}_p^2\).Footnote 18 We restrict our focus to words that occur at least 2 times but fewer than 100 times in the candidates’ public speeches and transcribed ad texts.Footnote 19 This removes some of the most common and least common words, which are not useful for the analysis. We then select the 1000 words with the highest values of \({\mathcal{x}}_p^2\).

Among the 1000 selected words, we regress congressperson c’s relative frequency of word p, \(\tilde{f}_{pc}\), on their ideology, ideologyc, measured by the Republican vote share in the districtFootnote 20 from the 2012 presidential election (collected from and estimate an intercept parameter αp and slope parameter βp. A positive (negative) slope estimate suggests that the word p is associated with the Republican (Democratic) party.

We then compute the political slant for each of the ad creatives by applying the same mapping between the relative word frequencies and political slants of those words used in the ad. Specifically, the political slant of ad creative n is computed as \(\tilde{y}_{n} = \frac{{\mathop \sum \nolimits_{p = 1}^{p = 1000} b_{p} \left( {\tilde{f}_{pn} - a_{p} } \right)}}{{\mathop \sum \nolimits_{p = 1}^{p = 1000} b_{p}^{2} }}\).Footnote 21

Finally, we re-index the estimated political slant of ad creative n, \(\tilde{y}_{n}\) to \(\tilde{y}_{nc}\) to denote the candidate c that ad creative n supports. Our slant measure is then calculated as \(\hat{y}_{cn} = - \left( {\tilde{y}_{cn} - 0.5} \right)\) if c is Clinton and \(\hat{y}_{cn} = \tilde{y}_{cn} - 0.5\) if c is Trump. Thus, a greater (lower) slant measure always corresponds to a more politically extreme (centrist) message for both candidates.

Appendix 2. Doc2Vec validation

In this appendix, we provide additional evidence to confirm the validity of the performance of the doc2vec algorithm by presenting two ad creatives with the other ad creatives that are found to be the most similar, excluding the chosen ad creative in consideration, and the least similar.

Tables 8 and 9

Table 8 Example of an Ad that supports Clinton
Table 9 Example of an ad that supports Trump

Appendix 3. Sensitivity to size of vector space in dov2vec

In this appendix, we demonstrate that our results in the main text are robust to the dimension of the vector space, one of the most important hyper-parameters, by providing results for different vector sizes. In all cases, standard errors are clustered at the candidate level.

Tables 10, 11, 12, and 13

Table 10 Effects of political slant and message consistency on WOM (vector = 150)
Table 11 Effects of political slant and message consistency on WOM (vector = 300)
Table 12 Effect of slant and message consistency on voter preference (vector = 150)
Table 13 Effect of slant and message consistency on voter preference (vector = 300)

1.1 Appendix 4. Sensitivity of the WOM results to the time window used

In this appendix, we show that our results in the main text are robust to different time windows by providing results for two- and three-minute windows.

Tables 14 and 15

Table 14 Effects on WOM (time window = 2 min)
Table 15 Effects on WOM (time window = 3 min)

Appendix 5. Tests on the level of clustering

In this appendix, we conduct the statistical test for the appropriate level of clustering proposed by MacKinnon et al. (2020). MacKinnon et al. (2020) test the null hypothesis of a finer clustering level against the alternative hypothesis of a coarser clustering level. In our context, we can test whether clustering at the ad creative level (null hypothesis) against clustering at the candidate level (alternative hypothesis). The results, shown below, reveal that the no clustering case (the finest case) is rejected for both the ad creative level and the candidate level, but clustering at the ad creative level is rejected against the candidate level. Taken together, these results suggest that we cluster standard errors at the candidate level.


Estimated Model: Eq. (2)



Bootstrapped p value

N vs A



N vs C



A vs C



  1. Note: N denotes no clustering; A denotes clustering at the ad creative; C denotes clustering at the candidate level

Appendix 6. Influence of outliers on the WOM analysis

In this appendix, we show that our results remain very similar to removal of outliers. Specifically, we run our WOM analysis after both winsorizing and trimming the post-WOM volume at the 1 and 99% level.

Tables 16 and 17.

Table 16 Effects on WOM (winsorized Post-WOM at the 1 and 99% level)
Table 17 Effects on WOM (trimmed Post-WOM at the 1 and 99% level)

Appendix 7. Comparison of national primetime ads versus local primetime ads

In this appendix, we show that Clinton and Trump’s national primetime ads exhibit very similar airing patterns as their local primetime ads in terms of time, day of the week, and month aired. Note that the summary statistics for the local prime ads are generated from the raw Stradegy data on the candidate’s primetime advertising and have not been cleaned for data errors.

Tables 18, 19, 20.

Table 18 National versus local primetime ad airings by time
Table 19 National versus local primetime ad airings by day of the week
Table 20 National versus local primetime ad airings by month

Rights and permissions

Reprints and permissions

About this article

Check for updates. Verify currency and authenticity via CrossMark

Cite this article

Fossen, B.L., Kim, D., Schweidel, D.A. et al. The role of slant and message consistency in political advertising effectiveness: evidence from the 2016 presidential election. Quant Mark Econ 20, 1–37 (2022).

Download citation

  • Received:

  • Accepted:

  • Published:

  • Issue Date:

  • DOI: