Issues and Democratic Primaries 2004

What Influences the News Coverage that Candidates Receive on Issues?
A look at Education Coverage in the 2004 Presidential Election

Results

To analyze this data, I ran several regression models. As mentioned above, regression analysis seeks to explain variation in the dependent variable using one or more independent variables. Further, predicted values allow one to examine the way one specific independent variable affects the value of the dependent variable. My regression models seek to explain the nature of the variation of candidate coverage on education, based on different variables. Tables six, seven and eight present the results of my regression models. Tables nine, ten and eleven present predicted levels of coverage.

I initially ran separate models using each of the three primaries for which we had data. However, the Iowa and New Hampshire primary variables did not show a significant difference from the base model that included the simple passage of time. Therefore, I included only the results from the model that included the South Carolina primary variables, which did present some interesting comparisons.

Model One

My first model, which I call the base model, used the variables of whether a candidate issued a press release, the number of real candidates in the race, the days of the week, whether there was a debate, the day before, day of and day after the State of the Union, and the time variable beginning with one. I also added a condition that restricted this model to the time during which the candidates were in the race. This presents a more realistic picture of how candidate behavior influences coverage, since candidates essentially stop agitating after they drop out of the race. Table six presents the results of my regression model and table nine presents the predicted levels of coverage.

Seven of the variables I included were significant for one or more candidates. I expected that candidates would not be able to affect their coverage on education by issuing press releases, and this expectation held for five of the six candidates. However, Dean was able to affect his coverage by issuing a press release. The coefficient for the effect of his press release on coverage was 0.53 and was significant at the p<.05 level. This means that as he issued more press releases, he gained more coverage. For example, if Dean did not release a press release on a particular day, he would receive on average 0.59 stories. However, if he issued one press release, 1.12 stories published mentioned him and education. I understand this variation due to Dean’s status for most of our time period as front-runner, which caused him to have the greatest number of stories published about him. Consequently, it was easier for him to affect coverage than the other candidates, since many in the news media focused on him.

For Clark, Kerry and Lieberman the number of candidates in the race had a significant effect on their coverage. For each man, as the number of candidates decreased, their coverage increased. However, since Lieberman dropped out second, this translated from an average of 0.08 stories published when all six candidates were in the race to 0.38 after Gephardt dropped out. Clearly, this boost did not prove effective enough to keep him in the race. Clark received a similar boost to that of Lieberman. With all six candidates, 0.04 stories mentioned him and education, while with five 0.35 did and with four 0.67 stories mentioned Clark and education. While Clark and Lieberman received a similar change based on the number of candidates in the race, Clark ended with more stories about him since he was in the race longer. For Kerry the results are quite impressive. Kerry jumped from a low of 0.37 stories published per day with all six candidates (note that this is much higher than either Clark or Lieberman) to a high of 1.37 articles published per day. With each successive candidate leaving the race, Kerry received a 0.2 increase in the stories published about him per day. This regression coefficient was significant at the p<.01 level. These results matched my hypothesis that coverage would increase as the number of candidates decreased.

Concerning the days of the week, Monday and Thursday were significant. Dean and Kerry both received a sizeable increase in coverage on Mondays. Both received approximately 0.6 more stories on a Monday than any other day, but Dean had more average stories in general. This means that, on any given Monday, Dean received 0.96 stories and Kerry received 0.80 stories whereas on days other than Monday 0.33 stories mentioned Dean while 0.15 stories mentioned Kerry. Thursday was significant for Dean, Edwards and Kerry. They received, on average, slightly less than 0.60 more stories on a Thursday than other days of the week. For Dean, this translated into a jump from 0.33 stories on days other than Thursday to 0.86 on Thursday. Edwards moved from 0.24 stories on days other than Thursday to 0.76 stories on Thursday. The stories mentioning Kerry increased from 0.15 on days other than Thursday to 0.76 on Thursday. Interesting, Kerry and Edwards, the two last candidates in the race, still received fewer stories than Dean did. As mentioned above, I believe this is due to the fact that Dean was the front-runner for the majority of the time we collected data. These findings near my hypothesis that the beginning and ending days of the week would be important. However, I failed to consider the Congressional workweek from Tuesday to Thursday. This could explain the fact that Dean and Kerry received more coverage on Monday (the day before the beginning of the Congressional workweek), following the indexing hypothesis and the idea of “beats” for reporters. I suspect that the three candidates received more coverage Thursday because it is the last day of the Congressional workweek. Of course, other explanations could exist for these variations.

The State of the Union variables proved significant for two candidates. For Lieberman, coverage increased on the day of the State of the Union address. He received 1.15 stories on the day of the address, while on any other average day he received only 0.26 stories. An explanation for this increase in coverage is difficult to find without more in-depth study of his coverage. The day after the State of the Union address proved significant for Kerry. The day after the address, Kerry received 3.63 stories, considerably more than his average of 0.48 stories on any day other than the State of the Union address. This increase in coverage could be explained by the fact that Kerry had just won the Iowa and New Hampshire primaries and began to unseat Dean as the front-runner.

The passage of time proved a significant variable for four of the six candidates. Dean received the greatest boost, moving from 0.16 stories forty days after the beginning of our study to 1.12 stories 147 days (January 25, 2004, two days before the New Hampshire primary) in the study. Edwards received a similar increase in coverage, moving from 0.19 stories per day to 0.62. Stories mentioning Gephardt increased from 0.06 to 0.26 from the fortieth day to the last day on which he contested the race (January 20, 2004, the 142nd day of our tabulation). Kerry received a larger increase, from 0.22 stories per day to 0.76. It is interesting to note that Dean increased the most of any candidate, but Kerry began with slightly more stories than Dean or Edwards.

Model Two

My second model differs from the first only in that I removed the condition and ran the model for the entire race, and added whether the candidate was in the race as an independent variable. Many of the results are similar, so I will focus only on the differences. One important difference is that I was able to run a regression that modeled all coverage concerning education during the entire race. Table seven presents the results of the regression and table ten presents the predicted levels of coverage.
Three important variables were not significant in both models. In the model concerning the entire race, the number of candidates is no longer significant for Lieberman. I believe this is because measuring the effects of the independent variables over the course of the entire race adds a complete month that Lieberman did not contest. However, measuring over the entire race changes the day after the State of the Union address from insignificant to significant for Lieberman. In the model for the entire race, 0.83 stories were predicted to be published on the day after the State of the Union address, while only 0.10 stories were predicted on any other day. The candidate’s participation variable proved significant for Clark at the p<.01 level. When he was in the race, 0.36 stories mentioned about him and education, while zero stories appeared after he exited the race. This follows my hypothesis that candidates would receive less coverage after they exited the race.

For Dean, considering the entire race decreases the predicted number of stories published in each of his significant variables, while shifting the regression coefficients in various directions. Interestingly, Edwards’s regression coefficients do not change, while his predicted values increase slightly. This change is quite unexpected, since removing the condition that Edwards be in the race should not have any affect on the results because he was in the race every day we collected data. Yet, as we lagged the coverage a day behind our other values, perhaps the coverage on Thursday, March 4, 2004 affected these variables (particularly since Thursday is significant for Edwards). The regression coefficients do not vary in any consistent pattern for Clark. The values for Gephardt and Lieberman decrease, probably due to the fact that by considering the entire race, we add over a month in which those candidates were not in the race.
For the cumulative regression I ran on all candidates, three variables proved significant, the number of candidates in the race, Monday and Friday. The regression coefficient for the number of candidates in the race was -1.02, and was significant at the p<.01 level. As I hypothesized, the fewer candidates in the race, the more stories published. In fact, when there were six candidates in the race, the predicted level of stories mentioning education or No Child Left Behind was 5.91. When three candidates remained, 8.96 stories published mentioned one or the other and when only one candidate remained, an impressive 10.99 stories appeared with education or No Child Left Behind. From table two, we see that the frequency distribution of stories indicates that, on 22% of the days that we studied, ten or more stories appeared mentioning education or No Child Left Behind. Thus, while these results are surprising, they are not implausible. The other significant variables were Monday and Friday. Monday was significant at the p<.001 level with a regression coefficient of 9.75. This translates into a difference of 5.10 stories published on days other than Monday to 14.85 on a Monday. This large increase could be due to the media starting the week with more stories than the rest of the week. Also, this finding was not surprising, since Monday was significant for several of the individual candidates. Friday was significant too, but not for any of the individual candidates. Friday, like Monday, was significant at the p<.001 level with a coefficient of 6.36. This means that on days other than Friday, 5.63 stories mentioned education or No Child Left Behind, while on Fridays nearly twelve stories did. Note that the negative variable for both Monday and Friday predict similar levels of coverage. I believe that Friday produced more stories as the media issued more stories to begin the weekend.

Model Three

My third model is also similar to the first. I used the same main independent variables and the same condition that the regression run only the dates on which the candidate participated in the race. However, I removed the variable that captured all the days on which we collected data and replaced it with the South Carolina primary variables. The first variable captured the days remaining until the primary and the second captured the days after the primary. As with the previous analysis, I will focus primarily on the differences between this model and the first. Table eight presents the regression model results and table eleven presents the predicted levels of coverage.

The most important change is that the independent variable representing number of candidates in the race became significant Dean and Edwards and all candidates combined. Since the variable is not part of the Gephardt model, this effectively means that the variable is significant for every candidate. For Clark, the variable became more significant and increased in absolute value. This translated to an increase in the predicted levels of coverage of 0.15 stories per day. Since this variable was not significant for Dean or Edwards in the first model, comparisons are not possible. Dean ranged from a low of 0.04 stories per day with six candidates to a high of 2.22 stories when only three candidates remained. With those three candidates, Edwards actually received fewer stories than Dean: 2.01. However, Edwards began the race of six candidates with 0.07 stories per day and ended with 2.66 when only he and Kerry remained. Lieberman’s predicted levels of coverage changed from 0.38 with five candidates and 0.08 with six in the first model to zero articles, regardless of the number of candidates in the race. Kerry received the largest increase of any candidate, moving from zero stories with six candidates to a high of 4.75. This is markedly more stories than predicted by the base model that included all the days on which we collected data. Recall, in that model, Kerry ranged from a low of 0.37 to a high of 1.37.

The change in these results, particularly the increase in Kerry’s coverage (since we know that four or more articles published mentioned Kerry on only six days), was surprising. Thus, I ran a correlation between the number of candidates and South Carolina primary variables. These results can be found in table twelve. From this test, we see that the days until the South Carolina primary and the number of candidates in the race correlate 62% of the time. The days after the South Carolina primary independent variable and the number of candidates in the race correlate at the 94% level. This correlation could be because five candidates remained in the race after the South Carolina primary and candidates dropped out successively after February 3, 2004. Similar correlation tests run on the days until the Iowa and New Hampshire primaries prove that the South Carolina primary variables were the optimum choice because the days after Iowa correlated with the number of candidates at the 99% level and days after New Hampshire correlated at 98%. This too, makes sense because the six candidates dropped out after Iowa, and five after New Hampshire. The high level of correlation between the number of candidates and the days after South Carolina variables could skew the results in this model and result in unreliable measures.

Concerning the days of the week variables, using the South Carolina primary variables increased regression coefficients slightly on Monday for Dean, Kerry and all candidates combined. The variable change also increased the coefficient for Edwards and Kerry on Thursday. This translated into an increase of less than 0.50 stories per day in each category. Interestingly, Tuesday and Wednesday became significant at the p<.05 level for Kerry in the third model. The regression coefficients were approximately 0.5 for both days. This is near the regression coefficients for both Monday and Thursday, although Monday’s coefficient is the largest at 0.68. These coefficients indicate a predicted level of coverage of approximately one story per day. This finding is interesting, since we know that Kerry actually received coverage on only about 27% of the days on which we collected data, while this finding implies that he would have a story approximately 57% of days (every Monday, Tuesday, Wednesday and Thursday). This discrepancy could be due to factors not included in this model, or the imprecision of predicted values. The regression for all candidates, however, the inclusion of the South Carolina variables produced almost no change, though it was in the same direction as the changes for individual candidates, meaning more predicted stories.

The dummy variable reflecting debates became significant for Kerry in this model. The regression coefficient was 0.42, which meant that 0.91 stories the day of a debate mentioned Kerry and education, while on any other day 0.49 stories mentioned him. I believe this is because most of the debates were held in 2004, when Kerry’s candidacy gained momentum. However, an explanation as to why the debate variable was significant for only Kerry, and not the other candidates, is difficult to find without more study.
Concerning the State of the Union address, the regression coefficient for Lieberman did not change greatly but this translated into interesting predicted values. Instead of 1.15 on the day of the address and 0.26 on other days, both values moved to zero. Since no stories mentioned Lieberman for 90% of the days on which we collected data, this second set of results seems more plausible. However, since the President mentioned education in the State of the Union address, it is probable that at least one story mentioned education and Lieberman on the day of the State of the Union address. The day after the State of the Union address, which was significant for Kerry, did not change in a specific pattern.

The South Carolina primary variables were significant for several candidates. The days until the primary variable was significant for Dean and Gephardt. When there were 116 days until the primary, Dean had 0.14 predicted stories, while Gephardt had 0.09. Closer to the primary, at fifteen days (which happens to be the day on which the Iowa caucuses were held), Dean moved to 0.82 stories and Gephardt to 0.34. The variable capturing days after the South Carolina primary was significant at the p<.001 level for Dean, Edwards and Kerry. This is not surprising since Clark dropped out of the race only eight days after the primary, Gephardt exited the race before the primary and Lieberman the day of the contest. The regression coefficients for Edwards and Kerry were about -0.10 and for Dean -0.23. This means that as more days pass after the South Carolina primary, fewer stories mentioned these three candidates. One day after the primary, 0.82 stories are predicted to mention Dean, 0.61 mention Edwards and 0.72 mention Kerry. This finding is curious, since Edwards won the South Carolina primary, with Kerry finishing second, and Dean coming in fifth behind Sharpton and Clark. Further, Dean did not do markedly better in any of the other primaries that were held on February 2. But, given the fact that more stories mentioned Dean than any other candidate, this is not wholly unexpected. Even more curious is the predicted level of coverage nine days after the South Carolina primary (February 15). For Dean, Edwards and Kerry, the predicted level of coverage is zero articles. One could speculate that this is because February 15 was a Sunday, yet the same level of coverage is predicted for all days after about the seventh. Essentially, the predicted levels of coverage for all days after February 9, 2004 are zero. Since the regression coefficient is negative (and about 0.15) for all three candidates, it logically follows that their levels of coverage, since they begin at less than one one day after the primary, eventually their coverage will reach zero and stay there.

By: Beth Daniel '04

© Davidson College, 2004, Department of Political Science, Davidson College, Davidson, NC 28035
Send comments, questions, and suggestions to Patrick Sellers
Created: 4/27/2004. Last updated: 5/2/2004.