WiscAds uses a technology that monitors the transmission of 35 national networks in the top Designated Market Areas (DMA). A DMA is a geographical region where individuals 60These data are gathered from Bonica (2013), Goldstein, Franz, and Ridout (2002), Goldstein and Rivlin
receive the same TV content and it is the smallest geographical unit in which a politician can buy air time. Every time there is a political advertisement in these markets, WiscAds captures it. A team of students research assistants then analyzes the storyboard of the advertisement to code it into the dataset. I therefore have detailed information on each advertisement: tone (i.e. whether it is positive or negative), exact date and time, station, and ad sponsor, among other things. It also importantly includes the candidate, party, or group for which the ad was aired in support. The dataset also contains an estimated cost variable. There are three ad tone types in the data: positive, contrast, and attack. I follow the convention in the literature and define negative ads as those classified as either contrast or attack ads.
The sample is limited to the geographical borders of WiscAds for each year. I merge the
counties covered by WiscAds with the counties in each district.61 Over the span of the three
election cycles I consider, WiscAds should in principle cover 1,390 races. However, none of the candidates running in 814 of these elections purchased airtime and hence are not in
the WiscAds dataset.62 I therefore have no information on the campaign strategies of the
candidates in those 814 elections. I also drop the 20 elections in which ads were purchased that were held in Louisiana, since this state employs a runoff system, and the 23 elections where a third-party candidate was a winner or a runner-up due to the method by which I estimate the supporters of each candidate. Finally, I drop 183 elections for which at least one candidate received positive contributions, but did not advertise, since I have no way to infer overall campaign strategies without observing advertisements. This leaves us with 361 elections over the three cycles.
Details of the type of elections covered in the final sample are given in Table 6. I have between 20 and 23 Senate elections for each year, and about 85 House elections for 2000 and 2004. For 2008, there are 126 elections included. Among the 814 elections in which neither candidate had a television advertisement, 758 were House elections. Since these 61I do not observe this directly from WiscAds. I obtained the list of counties in each DMA and year from
Table 6
Sample Elections by Race Type and Year
YEAR RACES Senate House 2000 23 82 2004 20 90 2008 20 126 Total 63 298
elections tend to be less competitive, advertising in general is less common. Hence, I
observe a large fraction of the House elections where at least one candidate purchased TV ads. Within the final sample, I have 200 Republican wins versus 161 Democrat wins. The average winning margin for both parties is almost 18%. The summary of election results is available in Table 7. Table 8 shows the distribution of incumbency status in the sample. There are 16 elections for an open seat in the Senate and 61 in the House. The remaining 47 Senate races and 237 House races involve an incumbent. Given that the sample period covers a relatively successful period for Republicans, there are 180 races with a Republican incumbent, and 104 with a Democratic incumbent.
Table 7
Election Data - All Years
Democrats Republicans Winning Margin 17.75 17.82 (14.87) (11.61) Winner 161 200 Total Races 361 Table 8 Incumbency Status TYPE RACES Senate House Open Seat 16 61 Democrat Incumbent 19 85 Republican Incumbent 28 152 Total 63 298
all of the 210 DMAs, hence why the 2008 sample includes more races than previous years. Note that in each case, the DMAs cover a very large portion of the U.S. population: in 2000, the top 75 DMAs accounted for 78% of the population. In 2008, nearly the entire population is covered. However, since the observed DMAs do not exactly cover the entire U.S. population, I only partially observe the campaigns for some elections – that is, there are
some races where I observe political television advertisements only in some of the counties
within the district or state in which the election is held. To quantify the degree to which this occurs, for each race, I compute the size of the population in the intersection of the DMAs I observe and the Congressional district (for House races) or the state (for Senate races) of the election, and divide by the district or state size. I find that, on average, the dataset contains 91% of the population in a district or state. The boxplot for this measure is displayed in Figure 16. For House races, the 75th percentile is above 90% whereas for the Senate, it is around 53%. That is, for 75% of House races, at least 90% of the population is in a DMAs I observe. The median coverage for Senate races is 93.5%, while for the House it is 100%.
Figure 16
Although the dataset covers a large portion of each race, the incompleteness of the data could still be problematic for empirical implementation. The potential issue is the fact that I expect the top DMAs to contain more populous urban areas which may be more Democratic than the rest of the country. Hence, the areas I observe might have a Democratic bias, which could potentially affect the strategies of the candidates. In carrying out the empirical analysis, I assume that the candidate has the same campaign strategy across the district.
To investigate the degree to which campaign strategies may differ across different popula- tions, I analyze the variance in campaign strategy for elections in which advertising occurs
in more than one DMA. In particular, let ndi,e be the cost of all negative advertisements
aired in DMAdby candidate iin election e, and lettdi,e be the cost of all advertisements in
daired by this candidate. I denote the campaign strategy in this DMA for candidate iin
electioneas: Ni,ed = n d i,e tdi,e. Letting Ni,e = ni,e
ti,e denote the campaign strategy for candidate i in election e across all
DMAs. Finally, I compute for each DMA the absolute deviation from the mean,|Ni,ed −Ni,e|.
Since air time is purchased in bulk, I consider campaigns that placed more than 500 ads in at least two different DMAs. Among these campaigns, the median absolute deviation is 0.034 for Republicans and 0.021 for Democrats. The 75th percentile is 0.093 for Republicans and 0.088 for Democrats. While there may be systematic differences between DMAs in the sample and outside the sample, this evidence is suggestive of the idea that strategies do not change dramatically across different populations.