The World Wide Web and the U.S. Political News Market: Online Appendices

Transcription

1 The World Wide Web and the U.S. Political News Market: Online Appendices Online Appendix OA. Political Identity of Viewers Several times in the paper we treat as the left- most leaning TV station. Posner (2005) explains the process in which became liberally - biased: The current tendency to political polarization in news reporting is thus a consequence of changes not in underlying political opinions but in costs, specifically in the falling costs of new entrants. The rise of conservative Fox News Channel caused to shift to the left. was going to lose many of its conservative viewers to Fox anyway, so it made sense to increase its appeal to its remaining viewers by catering more assiduously to their political preferences. The data from the Pew s biennial media consumption surveys from shows a process very similar to the one described by Posner, at least in terms of the mean selfidentified liberal-conservative position 1 of and Fox viewers. As shown in figure OA.1, as late as 1998 the mean position of viewers was close to the mean position of Fox viewers and the mean position of National Broadcast news viewers. At later years, as Fox news attracted more conservative viewers, s audience became more liberal. 1 The liberal-conservative scale in the Pew data takes the following values: 1= Extremely liberal, 2=Liberal, 3=Moderate, 4=Conservative, 5=Extremely conservative 1

2 Unweighted Mean Lib/Con National Survey year Weighted Mean Lib/Con National Survey Year Figure OA.1: Mean self-identified liberal-conservative position of viewers of different TV news sources (on a 5 point scale) in different survey years. 2

3 Online Appendix OB. KN and Pew Survey Methodology We use two main sources of data in this work: surveys conducted by Knowledge Networks, and as corroborating data we use the Pew biennial media consumption survey. We use this appendix to discuss the relevant survey methodological issues of both surveys, and explain the different methods of correcting for different sampling problems. 1. Knowledge Networks Public Affairs Profiles The main data source we use is a collection of public affairs surveys collected by Knowledge Networks (KN), a survey research company that uses its nationally representative online KnowledgePanel SM of United States residents. The national sample for recruiting KnowledgePanel SM is constructed using random-digit dial (RDD) methodology. The RDD sample frame is updated quarterly to be most inclusive of all residential landline telephone numbers. All telephone numbers have a known probability of selection. For purposes of refreshing KnowledgePanel SM, on-going sampling (without replacement) and on-going recruitment is conducted. The members of Knowledge Networks panel are required to fill out one survey a week. The surveys are filled online; respondents who own a computer get their internet connection paid by Knowledge Networks, and respondents who do not have a personal computer at home receive a Microsoft Web-TV, set- top appliance who enables them to fill out surveys online and to have an account. The company provides survey services for both commercial and academic purposes, and administers clients surveys to sub-samples of their panel, when each sub sample is constructed to be nationally representative. When there are not enough clients surveys to administer to all panel 3

4 members, they are asked to fill surveys that were constructed internally (by Knowledge Network), in order to keep panel members active and engaged (respondents do not know whether the survey they fill is an internal one or a client s survey). The public affairs survey used in our research was conducted as such an internal filler survey, part of Knowledge Networks efforts to maintain panel members engagement. This sampling scheme, which for lack of a better name we call left over sampling, makes it so that populations that are under (over) represented in Knowledge Networks panel will be even more under (over) represented in the left over population. This under (over) representation is corrected by using post- sampling weights. Weighting: To correct for sample composition, Knowledge Networks uses rim weighting to match Census distributions of age, gender, ethnicity, income and education. Response Rates: The surveys were conducted over a period of 7 years, using an on- going rotating panel, which makes it difficult to categorize the types of non- response in the traditional way (currently, there is no standard to compute response rates for online or on- going panels, nor has AAPOR issued a document to do so). The main issue is categorizing those who have declined to join the panel, or have dropped from the panel; for any specific survey, we may want to categorize them as non- contacts instead of nonresponse. Scholars using online panels typically report cumulative response rate (Couper, 2007; Huggins & Eyerman, 2001; Schlengen et al., 2002; Schonlau, Van Soest, Kapteyn, & Couper, 2006; Tourangeau, 2003), which is a product of recruitment rate (RR), profile rate (PR) and completion rate (CR). These rates pertain to three stages that take place from recruitment of online panel members to survey completion. RR, which signifies the 4

5 people who indicate that they are willing to participate in the panel when they are first contacted by RRD, was 38% for the KN panel. PR, which indicates people who respond to the initial socio-demographic survey out of all who agreed to become panel members in the initial phone interview, was 61%. Those who complete the survey become active panel members, and they start getting invitations to complete various surveys. Survey CR is the percentage of those who completed a particular survey out of all people invited to do so. CR for the public affairs profiles we use was between 70% and 75% over the 7 year period in which they were conducted, yielding cumulative response rates of 16% - 17%. 2. Pew Biennial Media Consumption Surveys Detailed information about the sampling scheme, response rates, and the questionnaires themselves can all be found on the Pew Research Center s webpage at We report here the bare essentials, for an easy comparison to the Knowledge Networks data. The Pew survey was conducted over the phone. The sample was drawn using standard list-assisted random digit dialing (RDD) methodology. As many as 10 attempts were made to contact every sampled telephone number. Calls were staggered over times of day and days of the week to maximize the chance of making contact with potential respondents. See Table A.1 for a comparison of the demographics of the sample population in the two data sets we use, and the comparable statistics from the Census Bureau s 2005 Annual Social and Economic Supplement (ASEC). Weighting: Pew weights are calculated to balance the interviewed sample of all adults by form to match national parameters for sex, age, education, race, Hispanic 5

6 origin, region (U.S. Census definitions), and population density. The White, non- Hispanic subgroup was also balanced on age, education and region. These parameters came from a special analysis of the Census Bureau s 2005 Annual Social and Economic Supplement (ASEC) that included all households in the continental United States that had a telephone. Weighting was accomplished using Sample Balancing, a special iterative sample weighting program that simultaneously balances the distributions of all variables using a statistical technique called the Deming Algorithm. Response Rates: the Contact Rate (proportion of working numbers where a request for interview was made) was 80%-81% (based on the 2004 and 2006 surveys, respectively). The Cooperation Rate (proportion of contacted numbers where a consent for interview was at least initially obtained, versus those refused) was 45% - 39%. The Completion Rate (proportion of initially cooperating and eligible interviews that were completed) was 94% - 92%. Overall, the cumulative response rates for the 2004 survey was 34%, and for the 2006 survey 29%. 3. Comparing Sample Compositions, Correcting for Sample Selection Table OB.1 compares the population reference parameters (taken from the Census Bureau s 2005 ASEC) to weighted and unweighted sample means of both the Knowledge Networks data and the Pew data (using the 2006 survey population). As is apparent from the table, both sample populations (KN and Pew) are older, more educated and not as ethnically diverse (with smaller representation for minorities) as the Census Bureau s ASEC. One way to correct for sample selection is to use post- sampling weights, and as Table OB.1 shows, using the weight gets the calculated means closer to the reference 6

7 population values. These sampling weights are the ones used in the body of the text, when we calculate weighted means of the different dependent variables. - Table OB.1 about here - An alternative way to correct for the sample selection problem is to run OLS regressions of the dependent variables, using the different demographic variables as regressors. The relationships uncovered in these regressions (the coefficients on the variables of interest) are then unbiased, as long as the selection into the sample is made on the independent variables (the various demographic controls we use), and not on the dependent variable (political views, strength of party id, or the level of interest in issues not covered by mainstream media). The interpretation of the coefficients then is slightly different then weighted population means; the coefficients in the OLS regressions are then the mean difference between two hypothetical respondents who are similar in their demographics, (or the residual difference between them, taking into account all the differences that can be attributed to observable demographic differences). 7

8 Table OB.1. Comparison of Sample Populations Demographics to The Census Bureau s 2005 ASEC sample Census KN Pew Reference Unweighted Weighted Unweighted Weighted Gender Male 48.1% 45.0% 48.2% 43.9% 47.9% Female 51.9% 55.0% 51.8% 56.1% 52.1% Age % 3.4% 5.2% 8.3% 12.5% % 15.9% 20.0% 12.6% 16.6% % 22.7% 19.0% 18.0% 19.3% % 25.4% 21.7% 20.6% 19.3% % 18.5% 17.2% 17.2% 13.7% % 13.9% 16.9% 20.8% 16.3% Education Less than HS Graduate 15.0% 6.1% 12.9% 7.5% 12.0% HS Graduate 36.1% 21.5% 31.7% 32.6% 35.8% Some College 23.1% 38.4% 27.3% 25.2% 24.2% College Graduate 25.8% 34.0% 28.0% 33.7% 27.3% Race/Ethnicity White/not Hispanic 71.2% 78.9% 73.6% 77.5% 72.5% Black/not Hispanic 10.9% 8.5% 9.5% 10.4% 10.8% Hispanic 12.1% 4.8% 6.6% 5.5% 9.4% Other/not Hispanic 5.8% 7.8% 10.3% 5.2% 5.9% 8

9 Online Appendix OC. Full OLS Regression Results Table OC.1. - OC.3 in this appendix include the full results of the OLS regressions that were shown in the body of the text, in Appendix A. Table OC.1. includes the full OLS regressions using the Knowledge Networks data, which correspond to lines 1 and 2 of the third column in tables A1 and A2 in the body of the text. Table OC.2. includes the full OLS regression using the Pew data, which corresponds to line 3 of the third column in tables A1 and A2 in the body of the text. Table OC.3 presents the results for hypothesis H1, which corresponds to the results in Table A3. In interpreting the results, recall that the hypotheses tested in the paper are all about correlations; we used regression to potentially control for selection into the sample (particularly in the Knowledge Networks data) that might affect the results. Table OC.1. The Ideological/Partisan Difference between Internet News Users and Non-Internet News Users Using Knowledge Networks Data Among Fox News Viewers Among News Viewers Party ID Lib-Con Position Party ID Lib-Con Position Internet 0.60*** 0.76*** -0.42*** -0.72*** Std. Err Female -0.22*** -0.26*** -0.24*** -0.21*** Std. Err Married 0.20*** 0.29*** 0.12*** 0.19*** Std. Err White 0.45*** 0.39*** 0.25*** 0.22*** Std. Err Black -0.99*** -0.25*** -0.79*** Std. Err Hispanic ** -0.18** Std. Err Religious 0.38*** 0.55*** 0.26*** 0.45*** Std. Err Age -0.02*** *** 0.01 Std. Err

10 Age Squared *** Std. Err Highest Degree Earned: High School 0.18*** 0.19*** Std. Err College 0.25*** 0.23*** Std. Err Advanced Degree *** -0.35*** Std. Err Household Income: $5,000 to $7, *** ** Std. Err $7,500 to $9, ** Std. Err $10,000 to $12, *** ** Std. Err $12,500 to $14, Std. Err $15,000 to $19, Std. Err $20,000 to $24, Std. Err $25,000 to $29, Std. Err $30,000 to $34, ** Std. Err $35,000 to $39, Std. Err $40,000 to $49, Std. Err $50,000 to $59, Std. Err $60,000 to $74, Std. Err $75,000 to $84, ** Std. Err $85,000 to $99, *

11 Std. Err $100,000 to $124, *** 0.18** Std. Err $125,000 to $149, ** Std. Err $150,000 to $174, *** 0.22** Std. Err $175,000 or more -0.42*** 0.20** Std. Err Year * Std. Err Year ** Std. Err Year ** Std. Err Year * -0.23*** 0.20** 0.13 Std. Err Year * Std. Err Year ** Std. Err Constant 3.09*** 3.54*** 3.18*** 3.35*** Std. Err R N 17, , (*) - significant at 10% level. (**) significant at 5% level (***) significant at 1% level. 11

12 Table OC.2. The Ideological Difference between Internet News Users and Non- Internet News Users Using Pew Data Among Fox News Viewers Among News Viewers Lib-Con Position Lib-Con Position Internet Std. Err Female -0.15* Std. Err Married 0.49*** 0.39*** Std. Err White Std. Err Black Std. Err Asian Std. Err Hispanic -0.44** Std. Err Religious 0.40*** 0.48*** Std. Err Age Std. Err Age Squared Std. Err Highest Degree Earned: Grades Std. Err High School Std. Err Vocational Std. Err Some College Std. Err College Graduate Std. Err Advanced Degree Std. Err

13 Household Income: $10,000 to $20, Std. Err $20,000 to $30, Std. Err $30,000 to $40, Std. Err $40,000 to $50, Std. Err $50,000 to $75, Std. Err $75,000 to $100, Std. Err $100,000 or more Std. Err Year ** Std. Err Year * Std. Err Constant 1.92*** 1.98*** Std. Err R N (*) - significant at 10% level. (**) significant at 5% level (***) significant at 1% level. 13

14 Table OC.3. The Likelihood of choosing Other Issue as the most important Issue for Internet News Users and Non-Internet News Users Using Knowledge Networks Data Likelihood of Choosing Other Issue Internet 0.014*** Std. Err Female -0.02*** Std. Err Married ** Std. Err White -0.02** Std. Err Black -0.03*** Std. Err Hispanic Std. Err Religious -0.02*** Std. Err Age Std. Err Age Squared Std. Err Highest Degree Earned: High School 0.13** Std. Err College 0.02*** Std. Err Advanced Degree 0.03*** Std. Err Household Income: $5,000 to $7, Std. Err $7,500 to $9, Std. Err $10,000 to $12, Std. Err $12,500 to $14,

15 Std. Err $15,000 to $19, Std. Err $20,000 to $24, Std. Err $25,000 to $29, Std. Err $30,000 to $34,999 Std. Err * 0.01 $35,000 to $39,999 Std. Err $40,000 to $49,999 Std. Err $50,000 to $59,999 Std. Err * 0.01 $60,000 to $74,999 Std. Err. 0.02** 0.01 $75,000 to $84,999 Std. Err ** 0.01 $85,000 to $99,999 Std. Err * 0.01 $100,000 to $124,999 Std. Err ** 0.01 $125,000 to $149,999 Std. Err $150,000 to $174,999 Std. Err $175,000 or more Std. Err Year 2005 Std. Err *** 15

16 0.003 Constant Std. Err. 0.09*** 0.02 R 2 N ,477 (*) - significant at 10% level. (**) significant at 5% level (***) significant at 1% level. 16

17 Online Appendix OD. Testing the Hypotheses using Subgroups for Comparison follows: Our hypotheses about the ideological position of Internet news users are given as H1: Of those who watch the News Channel, those who also get news from the Internet will be more conservative than all other News viewers. H2: Of those who watch news, those who also get news from the Internet will be more liberal than all other news viewers. In the body of the paper we test these hypotheses by comparing those who use the Internet and the extreme television new source to all others who are using the same extreme television news source. In other words, our comparison group in the tests, the non-internet news users, included those who are just using the extreme television news source, those who are using the extreme television news source and watching national network news broadcasts, and those who are using both extreme television news sources. A stronger version of these hypotheses is that the Internet news users should be more extreme when looking at each of these three sub-groups separately. The results of performing the analysis for each of the three different sub-groups separately are presented in Tables D.1 and D.2. The estimated difference is always in the predicted direction. Further, the results using the Knowledge Networks data are almost always statistically significant at conventional levels and the results based on using the Pew data often are despite the very small sample size being used in the tests. These findings provide further support to the claim that those using the Internet to supplement their news consumption from or by searching for news on the Internet are more ideological extreme and identify more strongly with the their respective party than their counterparts who do not. 17

18 Table D.1. The Ideological/Partisan Difference between Internet News Users and Non-Internet News Users Among News Viewers: Divided by Subgroups DV = Party Identification (Knowledge Networks data) Comparison Subgroup: +National + +National + +National + Difference 0.39*** 0.94*** 1.00*** 0.29*** 0.90*** 1.00*** 0.14*** 0.63*** 0.64*** Std. Error N DV = Liberal-Conservative Position (Knowledge Networks data) Comparison Subgroup: +National + +National + +National + Difference 0.42*** 0.87*** 1.21*** 0.31*** 0.91*** 1.14*** 0.18*** 0.73*** 0.75*** Std. Error N DV = Liberal-Conservative Position (Pew data) Comparison Subgroup: +National + +National + +National + Difference 0.31** * 0.25** 0.26** 0.38** ** Std. Error N (*) - significant at 10% level. (**) significant at 5% level (***) significant at 1% level. See Appendix B for exact definitions of each of the variables and their coding schemes. Control variables included in the OLS regression include dummies for marital status, race / ethnic groups, levels of education, levels of religious participation, Income levels and survey years. 18

19 Table D.2. The Ideological/Partisan Difference between Internet News Users and Non-Internet News Users Among News Viewers: Divided by Subgroups DV = Party Identification (Knowledge Networks data) Comparison Subgroup: +National + +National + +National + Difference -0.24* *** -0.28** -0.16* -0.65*** -0.24** -0.26*** -0.77*** Std. Error N DV = Liberal-Conservative Position (Knowledge Networks data) Comparison Subgroup: +National + +National +FO X +National + Difference -0.75*** -0.63*** -0.74*** -0.74*** -0.68*** -1.06*** -0.56*** -0.57*** -1.10*** Std. Error N DV = Liberal-Conservative Position (Pew data) Comparison Subgroup: +National + +National +FO X +National + Difference 0.31** * * -0.48*** Std. Error N (*) - significant at 10% level. (**) significant at 5% level (***) significant at 1% level. See Appendix B for exact definitions of each of the variables and their coding schemes. Control variables included in the OLS regression include dummies for marital status, race / ethnic groups, levels of education, levels of religious participation, Income levels and survey year. 19

20 Online Appendix OE. Testing the Hypotheses when the sample is broken down by level of Political Interest We also checked to see whether the results varied by the respondent s level of political interest. Our expectation is that since the politically interested are more likely to be using Internet to look at political content, the magnitude of the estimated effects to be larger among the more politically interested. For the estimation we divided the sample into those with low and high levels of political interest. We reran the three main analyses, broken down by the level of the respondent s political interest and present these results in tables E1-E3. In all cases, the magnitude of the effect is larger among the respondents with high levels of political interest. For the less politically interested, the results are less strong than originally estimated and often insignificant, although always in the right direction. 20

21 Table E.1. The Ideological/Partisan Difference between Internet News Users and Non-Internet News Users Among News Viewers: Divided by Political Interest (A) Respondents with High Levels of Political Interest DV = Party Identification (Knowledge Networks data) Difference 0.972** 0.927** 0.647** Std. Error (0.047) (0.030) (0.029) N 14,047 14,047 14,047 DV = Liberal-Conservative Position (Knowledge Networks data) Difference 0.994** 1.007** 0.828** Std. Error (0.052) (0.032) (0.031) N 13,812 13,812 13,812 (B) Respondents with Low Levels of Political Interest DV = Party Identification (Knowledge Networks data) Difference 0.386** 0.297** Std. Error (0.105) (0.079) N 3,420 3,420 3,420 DV = Liberal-Conservative Position (Knowledge Networks data) Difference * Std. Error (0.090) N 3,328 3,328 3,328 (*) significant at 10% level (**) significant at 5% level (***) significant at 1% level. The sample is limited to Fox News viewers; the point estimates indicate the difference between the weighted means of Internet and non-internet news users in terms of their party identification and liberal-conservative position. See Appendix B for the exact definitions of each of the variables and their coding schemes. 21

22 Table E.2. The Ideological/Partisan Difference between Internet News Users and Non-Internet News Users Among News Viewers: Divided by Political Interest (A) Respondents with High Levels of Political Interest DV = Party Identification (Knowledge Networks data) Difference * ** ** Std. Error (0.129) (0.083) (0.082) N 6,463 6,463 6,463 DV = Liberal-Conservative Position (Knowledge Networks data) Difference ** -0.82** ** Std. Error (0.150) (0.102) (0.102) N 6,073 6,073 6,073 (B) Respondents with Low Levels of Political Interest DV = Party Identification (Knowledge Networks data) Difference Std. Error N 1,233 1,233 1,233 DV = Liberal-Conservative Position (Knowledge Networks data) Difference Std. Error N 1,125 1,125 1,125 (*) significant at 10% level (**) significant at 5% level (***) significant at 1% level. The sample is limited to news viewers; the point estimates indicate the difference between the weighted means of Internet news and non-internet news users in terms of their party identification and their liberal-conservative position. See Appendix B for the exact definitions of each of the variables and their coding schemes. 22

23 Table E.3. Difference between Internet News Users and Non-Internet News Users in Terms of the Diversity of Issues Considered Important: Divided by Political Interest (A) Respondents with High Levels of Political Interest DV = Likelihood of Identifying Other as Most Important Issue (Knowledge Networks data) Difference 0.017** 0.017** 0.012** Std. Error (0.003) (0.003) (0.003) N 25,974 25,974 25,974 DV = Likelihood of Identifying Less Salient Issue as Most Important Issue (Knowledge Networks data) Difference 0.031** 0.031** 0.015** Std. Error (0.004) (0.004) (0.005) N 25,974 25,974 25,974 (B) Respondents with Low Levels of Political Interest DV = Likelihood of Identifying Other as Most Important Issue (Knowledge Networks data) Difference Std. Error N 8,415 8,415 8,415 DV = Likelihood of Identifying Less Salient Issue as Most Important Issue (Knowledge Networks data) Difference Std. Error N 8,415 8,415 8,415 (*) significant at 10% level (**) significant at 5% level (***) significant at 1% level. The point estimates indicate the difference between the weighted means of Internet and non- Internet news users for the given dependent variables. See Appendix B for the exact definitions of each of the variables and their coding schemes. 23