Media Sentiment, 2024 Presidential Election Campaign
[last update Sept 19] For a recent discussion of these data see Good Authority.
The following graphic shows the balance of sentiment in television news coverage of the Democratic and Republican candidates for president during the lead-up to the 2024 election. When the line is to the left of the graphic, the sentiment of news coverage is more positive for the Democratic candidate than it is for the Republican candidate. When the line is to the right of the graphic, the sentiment of news coverage is more positive for the Republican candidate than for the Democratic candidate. Methodological details are included below, and I will be updating this page — correcting errors, revising analyses, and adding new media content — semi-regularly leading up to the 2024 election. [Click image for a larger version.]
This figure is based on all sentences mentioning either the Democratic presidential candidate or the Republican presidential candidate in all archived broadcast transcripts from Nexis Uni, for four major broadcasters: ABC, CBS, Fox News and CNN. (NBC and MSNBC transcripts are not currently available.) Most television transcripts are available within a day, but some are delayed in reaching the archive. So the data here are preliminary, and when the figure is updated there may changes in estimates from earlier in the campaign. Past results may shift accordingly, but typically only in very small ways.
Current gaps in the data are as follows: ABC, missing data on Sept 7, 12-18; CBS, missing data on Sept 14; Fox, partially missing data on Aug 23-25.
Archived transcripts include morning and evening news and newsmagazine programs on ABC and CBS. They include most programming on Fox News and CNN. There is of course more content on the cable networks than the broadcast networks. All sentences, regardless of network, are given equal weight in this measure. Measures for the Democratic candidate rely on Biden sentences until he steps down, and Harris sentences thereafter.
Once sentences are extracted from news transcripts, each sentence is scored for positive and negative sentiment using the Lexicoder Sentiment Dictionary. Each sentence is assigned a net sentiment score, which is the difference between the number of positive words and the number of negative words. [Following past work, it is calculated as follows: log ( (positive words + .5) / (negative words + .5)).] The daily result is then the mean of net sentiment in all Republican-candidate sentences minus the mean of net sentiment in all Democratic-candidate sentences. That daily measure, capturing the difference in the net sentiment of Republican- and Democratic-candidate sentences, is what is shown in the above figure.
Note that this kind of net sentiment measure is often correlated with electoral preferences, both in the US and cross-nationally. Recent work highlights the value of sentiment analysis (and content analysis more generally) in recent US campaigns as well. That said, the measure is far from perfect:
There is always some degree of error in automated content analysis.
There are alternative strategies for aggregating results that take the sum rather than the mean of sentiment, or weight sentences differently, averaging across networks or weighting by viewership.
This measure does not distinguish between things that are being said about a candidate and things that are said by a candidate.
Note also that the predictive capacity of the measure is likely not because news content actually predicts attitudes; in fact, media coverage may often follow electoral preferences. This is most likely because (a) journalists are following the campaign carefully, and their output tends to mirror the direction of the campaign, and (b) news content is measured relatively easily, and quickly, and so sometimes reflects shifts that are only just emerging in public opinion polling. These findings are not exclusive to campaigns; thinking more generally, media coverage may not affect public attitudes so much as it reflects them.
It nevertheless is true that analyses of media coverage can provide a valuable indictor of the partisan ‘tone’ of the campaign.
Diagnostics & Alternative Specifications
Volume. There are meaningful changes not just in the sentiment but also the volume of coverage that each candidate receives over the campaign. Volume is not reflected in the graphic of net sentiment used above, as it is based on daily averages of sentiment (independent of volume). But it is certainly possible to capture the volume of coverage for each candidate. The figure below shows the total number of sentences that mention Harris or Trump (regardless of sentiment). Rather than show the advantage for one candidate or the other, I’m showing the total volume for each — with Harris sentences on the left and Trump sentences on the right. [Click image for a larger version.]
Adding volume to sentiment. It is also possible to produce a measure of sentiment that takes the volume of coverage into account. The difference is straightforward: instead of using the daily mean of net sentiment for each candidates’ sentences, this measure uses the daily sum of the net sentiment of candidates’ sentences. The result is a measure that is moved not just by the tone of content but the volume of that content. The figure below shows this measure. Note that these estimates push sentiment more strongly in Harris’ favor just after Biden steps down and during the DNC — because during these times Harris receives a lot of coverage, and that coverage is mostly positive. [Click image for a larger version.]
Differences across networks. All of the above graphics show results combining all sentences from ABC, CBS, CNN and Fox. But there are some meaningful differences in the sentiment of coverage on each of these networks. The figure below shows the predicted value of net sentiment for each candidate, across each network. The figure is based on models for each candidate from August 1st onwards. Whiskers show 95% confidence intervals. [Click image for a larger version.]
