Conservative News Media and Criminal Justice: Evidence from Exposure to Fox News Channel

Transcription

1 Conservative News Media and Criminal Justice: Evidence from Exposure to Fox News Channel Elliott Ash and Michael Poyker January 26, 2019 Abstract Exposure to conservative news causes judges to impose harsher criminal sentences. Our evidence for this claim comes from an instrumental variables analysis, where randomness in television channel positioning across localities induces exogenous variation in exposure to Fox News Channel. These data on news viewership, paired with data on almost 7 million criminal sentencing decisions in the United States for the years , are used in the regressions. Higher Fox News viewership increases incarceration length, and the effect is stronger for black defendants and for drug-related crimes. The effect is observed for elected, and not appointed, judges, consistent with voter attitudes as a potential mechanism. The effect becomes weaker as judges get closer to election, suggesting a diminishing marginal effect for judges who are already politically engaged. PRELIMINARY FOR CONFERENCE SUBMISSION Elliott Ash: ETH Zürich, elliott.ash@gess.ethz.ch; Michael Poyker (corresponding author): Columbia University, mp3780@columbia.edu. We are grateful to Sergio Galletta and Greg Martin for help with understanding the datasets. We thank David Cai and Romina Jafarian for helpful research assistance. All errors are ours. 1

2 1 Introduction A recent literature has documented that greater exposure to partisan television news has an impact on voting in presidential elections (DellaVigna and Kaplan, 2007; Martin and Yurukoglu, 2017) and congressional position-taking (Clinton and Enamorado, 2014; Arceneaux et al., 2016). An unexamined question is whether partisan news would have an effect on judge decision-making. The goal of this paper is to provide the first evidence on this issue. If judges are apolitical and make their decisions without regard to outside influences, partisan news exposure should have no effect (see, e.g., Posner, 2008; Epstein, Landes and Posner, 2013). But recent empirical work has documented that judges do respond to nonlegal influences, political and otherwise (Berdejó and Yuchtman, 2013, Ash and MacLeod, 2015, 2017; Chen, Moskowitz and Shue, 2016; Ash, Chen and Naidu, 2017; Berdejó and Chen, 2017). In addition, there is evidence suggesting that the judiciary has become more conservative over time (e.g., Ash, Chen and Naidu, 2017; Ash, Chen and Lu, 2017). This research asks whether we can attribute a causal influence to partisan news media in this trend. The empirical context is criminal courts in U.S. states for the years 2005 through We use combined microdata on criminal sentencing decisions from the National Corrections Reporting Program and Dippel and Poyker (2018), paired with data on cable news viewership at the county level. The measure of conservative news exposure is Fox News viewership (relative to other cable news networks), where exogenous variation comes from the channel positioning of Fox News across counties. As demonstrated in Martin and Yurukoglu (2017), this channel-number variation can be used as an instrument for TV viewership across channels. We replicate the strong first stage from Martin and Yurukoglu (2017) at the county level in our sample of states. We document that current Fox channel position is unrelated to preexisting markers for conservative policy, such as historical Republican vote shares, past crime rates, or past sentencing rates. We use the first-stage prediction for Fox News viewership to estimate the impact on criminal sentencing outcomes in a two-stage-least-squares framework. We find that an exogenous increase in Fox News exposure is associated with an increase in criminal sentence length. The result is robust to the inclusion of rich demographic controls and case controls, and to including controls for viewership of other cable news networks (CNN and MSNBC). In heterogeneity analysis, we find that effect is larger for black defendants than for white or Hispanic defendants. The effect is larger for drug-related crimes. 2

3 To understand the mechanism underlying the effect, we looked at the effects separately for elected and appointed judges. The appointed judges have tenure, and therefore face minimal political pressures once in office. We found that Fox News increases sentencing only for elected judges, and not for appointed judges. This is consistent with voter attitudes providing a possible mechanism for our effects. As voters become more conservative due to Fox News exposure, that increases electoral pressures on judges to be more conservative in their criminal justice decisions. These results will be of interest to scholars in empirical political economy, and in particular for those those who study courts and the mass media. The findings are relevant to recent debates on how judges should be selected, retained, and compensated (Epstein, Landes and Posner, 2013; Ash and MacLeod, 2017), along with recent debates on polarization and media regulation (Boxell, Gentzkow and Shapiro, 2017; Allcott and Gentzkow, 2017). More generally, this research is related to the literature on partisan slant, media coverage, and political accountability (Gentzkow and Shapiro, 2010; Snyder and Stromberg, 2010). 2 Background This paper is motivated by previous evidence that Fox News is conservative, and the ongoing discourse on how conservative media impact social attitudes and policy outcomes. Figure 1 shows three pieces of evidence on this point. First, the top panel, from Martin and Yurukoglu (2017), shows that Fox News tends to use politicized phrases associated with Republican politicians. Second, in the middle panel, we see that for the years , Fox speakers mention crime more often than speakers on CNN and MSNBC. 1 Third, in the bottom panel, we show in our data that places with higher Fox News ratings share tend to impose longer criminal sentences. An additional piece of cross-sectional evidence on how Fox News is related to criminal justice is reported in Figure 2. To make this graph, we produced average sentence length metrics by court. We then plotted the trends in sentence length separately by quintiles in Fox News viewership (for the years ). We can see that in the places with more Fox News viewership, there was a much larger jump in sentencing lengths starting in This is some striking descriptive evidence that places with more Fox News exposure had harsher criminal justice outcomes. The question, for this paper, is whether this correlation in the courts is due to a causal link. 1 These are counts of crime, criminal, murder, and homicide, divided by the number of spoken sentences, in transcripts for prime time shows for each network. 3

4 Figure 1: Fox News is Conservative Proportion of Speakers Mentioning Crime Year CNN MSBC FOX Notes: Illustrations for Fox News conservatism. Top panel is predicted ideology based on political phrases used by Republicans and Democrats. Middle panel is the number of references to crime per sentence spoken in cable news transcripts. The bottom panel is a binscatter for the OLS correlation between incarceration length and Fox Nielsen rating. 4

5 Figure 2: Fox News and Average Sentencing Length Notes: Illustrations for correlation of Fox News consumption and average sentence length. The figure shows average sentence length in top 25%, bottom 25%, and middle 50% of the counties by Fox News consumption. To better understand the crime-related discourse of Fox News, we used NLP tools to understand the language associations in the show transcriptions. We trained word2vec, a popular word embedding model (Mikolov et al., 2013), on transcripts for Fox, CNN, and MSNBC, for the years 2001 through This model works by reading through sentences and locating words close to each other in a vector space if they tend to occur in similar contexts (that is, windows of neighboring words). Similarity between words can then be measured using the cosine of the angle between the vector representations of each word. In the transcripts data, the most similar words to crime were crimes, murder, homicide, perpetrator, felonies, and other synonyms or closely related terms. What is most interesting for our purposes is the differences in word associations across the networks. To get the crime words most associated for Fox News, for example, we take the Fox cosine similarity and divide by the average of the similarities for CNN and MSNBC. We computed a symmetic measure for CNN and MSNBC. We then ranked the most associated words for each network. Word clouds illustrating the most crime-associated words for each of the three networks are reported in Figure 3. The results are striking. One can see immediately in Panel A that at Fox, discourse on crime is racialized. The highest-associated term is black-onwhite, and white-on-black is also highly ranked. Other words seem to personalize crime victimization: victimize, muggings. They also arguably demean the accused: perps and priors. 5

6 Figure 3: Crime Discourse in U.S. Cable News Channels A. Most Similar Words to Crime : Fox News B. Most Similar Words to Crime : CNN C. Most Similar Words to Crime : MSNBC Notes: Most closely related terms to crime, in Fox, CNN, and MSNBC, respectively. Similarities computed from word2vec models trained separately on the transcript corpora for each network. Larger words mean the word has higher similarity for the indicated network and lower similarity for the other two networks. 6

7 The word clouds for CNN (Panel B) and MSNBC (Panel C) have very different flavors. One can see that CNN focuses on criminal organizations and conspiratorial language: terrorism, mobsters, and underworld. The top terms from MSNBC are for a rapper ( Little John ) and a particular sensational murder case from 2006 ( Imette [St. Guillen] ). 3 Data 3.1 Sentencing Data The data on sentencing come from the National Corrections Reporting Program (ICPSR 36373, hereafter NCRP). This is a standard dataset for the literature and it contain data for all prison admissions in the United States from 2000 to This dataset poses several important characteristics that make it crucial for our study. As it spans all United States it gives us more variation in our main explanatory variable that varies of county-level. In addition, we have eight years of overlap for our explanatory variable (2005 to 2008 and 2010 to 2014). Our main outcome variable is the length of sentence imposed. We also use defendant and case characteristics. The seriousness of a crime is one of the main features of the judgment of a court, and the classification of offenses in the NCRP is standardized. Therefore we include in our regressions matrix of 180 fixed effects for the offense with the longest sentence length. We also include criminal history (recidivism), education, military background, and demographic characteristics, including age at conviction, gender, and race (Asian, Black, Hispanic, Indian, White and Other). We supplement our NCRP data with the sentencing data from Dippel and Poyker (2018). Its superior to the NCRP s dataset in a sense that it has (i) case-level information on each sentencing decision; (ii) information on the defendants that were fond not guilty or did not go to prison (e.g., went on probation); (iii) has information on judges. Here, we only use data from the 9 states that has judges information: Alabama, Arkansas, Colorado, Georgia, Kentucky, Minnesota, North Carolina, Tennessee, Virginia, and Washington. The years covered by the convictions data vary from state to state but range from 1980 to To construct the length of sentence imposed we assign zero for all cases in which the defendant is found not guilty or put on probation. In the case of consecutive sentences, those are summed. In the case of concurrent sentences, we take the max. The classification of offenses varies across states and trying to harmonize them would be complex and require many subjective decisions. Therefore we include in our regres- 7

8 sions a separate set of offense class fixed effects for each state. Data also contains information on recidivism and basic demographic characteristics, including age at conviction, gender, and race (Asian, Black, Hispanic, Indian, White and Other). Finally, we have information on judicial elections for a subset of states. 3.2 Media Data The data on channel positions and ratings come from Nielsen. This is an expanded version of the dataset in Martin and Yurukoglu (2017). The data includes channel listings by system and year, with associated zip codes, for the years 1998 through It includes zip code level viewership for Fox, CNN, and MSNBC for the years 2005 through It includes DMA (media market) level data on viewership for the years 2010 through For the text illustration (shown previously), we downloaded full-text transcripts of prime time shows on Fox News, CNN, and MSNBC, from LexisNexis. 4 Empirical Specification and Identification The identification strategy adopts an instrumental variables approach based on Martin and Yurukoglu (2017). The instrument relies on exogenous variation in where Fox News Channel appears in the channel number lineup across counties relative to other cable news channels. Counties are the lowest-level geographical unit for the sentencing data; in most cases judicial districts are composed of multiple counties. 2 Throughout the paper, an observation is a sentencing decision i that took place in county c of state s at year t SLS Specification The first-stage estimating equation is: T ct = α s(c)t + γz ct + X i(c)t β + η i(c)t, (1) where T ct is the time spent watching Fox News as a share of total television watched (share) for county c at time t, α s(c)t includes state-year (interacted) fixed effects, Z ct is the channel number for Fox News, and η i(c)t is an error term. X i(c)t includes other covariates describing demographics and cable system characteristics. From Martin and Yurukoglu (2017) we expect a negative and significant estimate for γ. 2 For the exception of Alaska, Delaware, Maine, and Texas. 8

9 The second-stage estimating equation models an outcome Y i(c)t (e.g., criminal sentencing harshness in courthouse/county c at time t) as: Y i(c)t = α s(c)t + ρt ct + X i(c)t β + ɛ i(c)t, (2) where the terms are the same as in equation (1) and ɛ i(c)t is the error term. The identification assumption is that conditional on the fixed effects and covariates, the channel position Z ct affects the outcome Y i(c)t only through Fox News viewership T ct. As treatment is on the county level, we clustering standard errors by county. 3 We require instrument relevance. Figure 4 Panel A shows graphical evidence of the first-stage variation we are using. There is a clear downward trend, with higher channel numbers having lower viewership. In the regression tables below, we report the F-statistic of excluded instruments for each regression, and they are consistently greater than 10. Figure 4 Panel B shows a binscatter for the reduced form. The vertical axis is the outcome (log incarceration length) and the horizontal axis is Fox channel position. We can see a negative relation, reflecting that in counties with a lower Fox channel position judges tend to be harsher in sentencing. The reduced form is identified if the channel position is exogenous. Martin and Yurukoglu (2017) provide a lengthy discussion and set of checks along these lines. In our data, as in theirs, the instrument is not related to past Republican vote shares. It is not related to past crime rates, and it is not related to past sentencing harshness. These points all support the exogeneity assumption. Still, in Robustness Section 5.4 we include these covariates as controls. Under exogeneity, two-stage least squares procures consistent estimates for ρ if the instrument satisfies an exclusion restriction. That is, the channel position affects sentencing decisions only through its affect on Fox News viewership. We feel this is a reasonable assumption in our context. In particular, we include state-year fixed effects in all specifications to control for the changes in state-level legislation: it allows us to rule-out the effect of conservative media on sentencing through changes in legislation. 4 A final assumption is monotonicity. That is, a lower Fox News Channel would not decrease Fox News viewership. Again, we feel this is a reasonable assumption in our institutional context. Still, we have performed a series of checks to see that the first stage is satisfied in subsets of the data. 5 3 Our results hold if we cluster by state or by county and year. These results are available upon request. 4 While this mechanism may exist we can t identify it separately. 5 We report a histogram of the first-stage coefficients in Figure A8. 9

10 Figure 4: First Stage and Reduced Form: Graphical Results Panel A (first stage) Panel B (reduced form) Notes: Binscatter diagrams for the first stage (Panel A) and reduced form (Panel B). 4.2 Estimating Heterogeneous Effects We are also interested in estimating heterogeneous effects. In particular, we check whether media slant is disproportionately affecting minorities, women, or certain nonviolent types of crime. To do so we adapt the 2SLS specification. The second stage is: Y i(c)t = α s(c)t + ρ 1 T ct + ρ 2 T ct µ i(c) + µ i(c) + X i(c)t β + ɛ i(c)t, (3) where µ i(c) is a characteristic for defendant i (e.g., race category or crime category). Here, we are interested in the baseline effect of media consumption, ρ 1, plus the interaction effect with the defendant s characteristic that might be targeted by the media, ρ 2. There are two endogenous variables. The first stage consists of T ct = α s(c)t + γ 1 Z ct + γ 2 Z ct µ i(c) + µ i(c) + X i(c)t β + η 1 i(c)t and (4) T ct µ i(c) = α s(c)t + γ 1 Z ct + γ 2 Z ct µ i(c) + µ i(c) + X i(c)t β + ηi(c)t, 2 (5) where the two excluded instruments are the channel position Z ct, plus the channel position interacted with the defendant characteristic, Z ct µ i(c). These first stages are used to predict the endogenous regressors. The assumptions for consistency are the same as those articulated above. Again, we need an F-stat above 10. As long as the channel position is randomly assigned, so is the 10

11 channel position interacted with defendant characteristics. 4.3 Judicial Elections As a final analysis, we test whether the effect is different in states where judges are elected versus states with appointed judges. We hypothesize that if judges are facing election or reelection they may change their sentencing behavior in accordance with political media messaging. Thus we split the sample in two and estimate the effect of Fox news consumption on each subsample of states separately. Besides the sample split, the estimation approach is the same as that in Subsection 4.1. In order to dig into the election channel, we also look at dynamic effects of the electoral cycle. For this purpose we need judge identifiers, which are not available in the NCRP data. Therefore, we use data from Dippel and Poyker (2018), where we have nine states with elected judges. Following Berdejó and Yuchtman (2013) and Dippel and Poyker (2018) we construct a variable that measures proximity to judge s reelection in quarters. In the case of four-year terms, for example, we assign the value of 16 for the quarter of election, and 1 for the first quarter after an election. The estimation strategy is the same as Subsection 4.2, but the instrument is interacted with a judge s proximity to election, rather than with defendant characteristics. In addition, we include tenure s length and judge fixed effects. 5 Results This section reports the results of the regression analysis. Subsection 5.1 reports the main results. Subsection 5.2 reports key heterogeneity analysis. Subsection 5.3 looks to mechanism by analyzing judicial elections. 5.1 Main Results The main results are presented in Table 1. We report OLS estimates for the specification without controls in Column I. As we already saw in the binscatter from the bottom panel of Figure 1, the share of Fox news viewers is positively associated with the sentence length. One percent increase in a share of Fox News consumption increases incarceration length by 1.2 percentage points or 21 days. 6 Columns III and V add demographic controls and case controls, respectively, which make the coefficient about twice as large. 6 A county with an increase in Fox news viewer share of 1.9% would experience approximately 2.3 percentage points longer sentences, or 39 more days in prison. 11

12 Table 1: Fox News and Sentencing Decisions Dependent variable: Log sentencing length in months I II III IV V VI VII VIII IX X Nielsen share OLS 2SLS OLS 2SLS OLS 2SLS OLS 2SLS OLS 2SLS Fox news 0.012** *** 0.050* 0.023*** 0.071** 0.022*** 0.070** [0.0134] [0.1722] [0.0002] [0.0897] [0.0000] [0.0113] [0.0000] [0.0155] Fox - (CNN+MSNBC)/ *** 0.074** [0.0000] [0.014] CNN 0.011** [0.0163] [0.6629] MSNBC [0.4972] [0.6369] Demographic controls X X X X X X X X Case controls X X X X X X R-squared Partial R-squared F-stat. of excl. inst Observations 4,971,052 4,960,189 4,971,052 4,960,189 4,971,048 4,960,185 4,950,290 4,940,401 4,950,290 4,940,401 All columns include state-year FEs. The dependent variable is the inverse hyperbolic sine of the sentencing length. The following variables are used as controls: age, age squared, and race dummies (Black, Hispanic, Asian, and Native American), dummy for recidivists and state-specific set of crime severity dummies. Standard errors clustered by FIPS. P-values are in brackets. 12

13 We report 2SLS estimates without controls in Column II. While the first stage is strong (F = 32.5), the coefficient is insignificant. In Columns IV and VI, we again add demographic and case controls. As with OLS, adding these controls increases the effect; with controls, the 2SLS effect is statistically significant. Defendant characteristics and case characteristics are correlated with Fox News viewer share, so they soak up the residual variation and strengthen estimates. These estimates are economically significant. According to the estimate from Column VI, increasing Fox News share by 1 percent would increase average sentencing length by 7.1 percentage points (about 4 months). A potential confounder in our analysis are the ratings for other cable news networks, CNN and MSNBC. To zero in on Fox News, Columns VII and VIII provide results with an alternative specification for viewership which is normalized relative to CNN and MSNBC. 7 Finally, Columns IX and X include CNN and MSNBC viewer share as (non-excluded) controls. In each of these alternative specifications, the 2SLS estimates are comparable to Column IV: positive and statistically significant Heterogeneous Effects by Defendant Characteristics In Table 2, we assess possible heterogeneous effects by providing estimates for Eq. 3. Column II shows that our effect goes almost entirely through black defendants. The coefficient for ρ 1 becomes insignificant, while ρ 2 is large, suggesting racial bias in the effect of Fox News on sentencing behaviour. 9 We find no evidence for a disproportionate media effect toward Hispanics (Column IV) or toward female defendants (Columns VI). The empirical literature on U.S. criminal justice has emphasized that minorities are often disproportionately prosecuted for non-violent crimes (Fagan and Ash, 2017). And anecdotally, conservative discourse often concerns itself with the risks posed by illicit drugs and the associated informal economy. Motivated by these points, we estimate the relative effects of Fox News exposure for drug-related crimes (Column VIII) and a nonillegal-drug placebo, driving under the influence of alcohol (DUI, Column X). We find that while there is no relative effect of Fox News on DUI crimes, there is a large interaction effect for drug crimes. The 2SLS estimates in Column VIII suggest that the effect operates almost entirely through drug-related crimes. 7 It is constructed as share_fox = share_fox (share_cnn + share_msnbc)/2. 8 The appendix reports graphical evidence for effects of CNN and MSNBC on sentencing. We do not see the same effects for these networks. 9 Interestingly, we don t find a positive relationship between being black and sentence length, which is somewhat different than other work in this literature. It could be that blacks are more often arrested with less serious crimes (even within observed charge categories). 13

14 Characteristic Table 2: Fox News and Sentencing Decisions: Heterogeneity by Defendant Characteristics Dependent variable: Log sentencing length in months I II III IV V VI VII VIII IX X Black Hispanic Female Drug-related crimes DUI crimes OLS 2SLS OLS 2SLS OLS 2SLS OLS 2SLS OLS 2SLS Nielsen share 0.018*** *** 0.073** 0.022*** 0.064** 0.019*** *** 0.071** [0.0000] [0.6956] [0.0000] [0.0154] [0.0000] [0.0221] [0.0000] [0.6119] [0.0000] [0.0122] Nielsen share x 0.015** 0.154** 0.016** ** 0.202*** Characteristic [0.0255] [0.0280] [0.0382] [0.6945] [0.3702] [0.1710] [0.0203] [0.0038] [0.8589] [0.9752] Characteristic ** *** *** [0.1688] [0.0274] [0.2994] [0.3856] [0.0000] [0.0004] R-squared Partial R-squared F-stat. of excl. inst Observations 4,971,048 4,960,185 4,971,048 4,960,185 4,971,048 4,960,185 4,971,048 4,960,185 4,971,048 4,960,185 All columns include state-year FEs. The dependent variable is the inverse hyperbolic sine of the sentencing length. The following variables are used as controls: age, age squared, and race dummies (Black, Hispanic, Asian, and Native American), dummy for recidivists and state-specific set of crime severity dummies. Characteristic coefficients in Columns VII X are absorbed by crime type fixed effects. Standard errors clustered by FIPS. P-values are in brackets. 14

15 5.3 Mechanisms: Judicial Elections We test judicial elections as a possible mechanism for the observed effect in Table 3. Column I includes the baseline specification with controls (Table 1, Column VII) for comparison. Column II reports results for the sample of states with appointed judges: the coefficient is negative and insignificant. Column III reports the coefficient for the subsample of states with elected judges. The coefficient of interest is significant and is not statistically different from the baseline coefficient. This result is consistent with the effects being driven by the election channel. 10 Table 3: Fox News and Sentencing Decisions: Elected vs. Appointed Judges Dependent variable: Log sentencing length in months I II III IV V VI VII Sample NACJ NACJ (appointed) NACJ (elected) NACJ (rest.) Full sample of 9 states Nielsen share 0.071** *** * 0.093* 0.106*** [0.0113] [0.2107] [0.0085] [0.4635] [0.0736] [0.0590] [0.0096] Proximity-to-election 0.017*** [0.0062] Proximity-to-election *** x Nielsen share [0.0061] Tenure X X Judge FE X X R-squared Partial R-squared F-stat. of excl. inst Observations 4,960,185 1,025,716 4,444, ,319 2,074,313 2,071,631 1,852,833 All columns include state-year FEs. The dependent variable is the inverse hyperbolic sine of the sentencing length. The following variables are used as controls: age, age squared, and race dummies (Black, Hispanic, Asian, and Native American), dummy for recidivists and state-specific set of crime severity dummies. Standard errors clustered by FIPS. P-values are in brackets. As mentioned before, we need judge identifiers to analyze dynamic election incentives, which shrinks our dataset to the sample of nine states with elections for which we have sentencing decisions matched to judges. The results with NCRP data for this sample of states is reported in Column IV. The number of observations decreases by five fold, and while the coefficient is comparable in magnitude, it is not statistically significant. Next, Column V shows the equivalent regression using the data from Dippel and Poyker (2018) 10 Here, we do not discard the legislation channel, as it is conditioned by the state-year fixed effects. While we can t rule it out it does not affect our results. 15

16 (rather than NCRP), which has more years and about double the sample size. In this data set there are positive and significant effects of Fox News (at 10% level), comparable in magnitude to Column I. Results hold if we add judge fixed effects and tenure (judge experience) in Column IV. We are interested in the potential channel that Fox news affects sentencing through the judicial election cycle. Using heterogeneity across judges in proximity to elections as an interaction term, we report results in Column VII. The coefficient for Fox News remains positive and significant. Judges become harsher closer to reelection, consistent with Berdejó and Yuchtman (2013) and Dippel and Poyker (2018). However, we find a negative coefficient for the interaction: the effect of election proximity is smaller for judges with greater Fox News exposure. 5.4 Additional Results and Robustness Checks In Table 4, we also test whether Fox News consumption affected other sentencingrelated outcomes. In Columns I and II, we found no evidence in the effect of Fox News on increase in probability of incarceration or conviction (i.e., found guilty but assigned a probation). We also find no effect on the number of distinct cases per sentence, suggesting, that Fox News does not affect sentencing through policing (i.e., police does not create more cases that potentially can increase the sentence). In Column IV, for a small subsample of states we estimate the effect of Fox News on the suggested sentence length for that crime; however, the first stage is not strong enough to draw any conclusions. Finally, in Column V, we check whether Fox News consumption affects the sentence length, conditional on any sentence. We find, that the resulting coefficient for the subsample of convicted to prison term is similar in magnitude to our baseline coefficient. Finally, we provide additional robustness checks to address possible exclusion restriction violation. First, in Columns II V of Table 6 we include additional control variables that may correlate with channel position: past Republican vote share, share of rural population, past crime rates, and past sentencing. While some of these variables appear to be significant and have expected signs, our results hold. They also hold when we include all of these controls together in Column VI. Since our Fox News data come in two pieces ( and ), in Table 5, we also check whether our results hold on the subsamples. Both OLS and 2SLS hold for the sample; however, the first stage on the subsample of is weak (Fstatistics equal 3.5) and the second-stage coefficient being insignificant (OLS coefficient is positive and significant). Our results are not driven by a particular statistical artifact and 16

17 hold if we omit any combination of 5 states. 6 Discussion and Concluding Remarks This paper has shown evidence that conservative television media exposure has a causal effect on judge decision-making. When Fox News has higher viewership due to lower channel numbers, that makes judges harsher in their sentencing. This result adds to previous work showing that Fox News has an effect on voter attitudes (Martin and Yurukoglu, 2017); here we have established that it also has an influence on judges (in the high-stakes decision of how long to incarcerate a person). Adding to work showing that judges respond to political incentives (Ash and MacLeod, 2016), we have established that politicized information (and not just incentives) matter for judge decision-making. In the heterogeneous effects section, we showed that the effect of Fox News is focused on black defendants, and on drug-related crimes. This result adds to the large literature on racial discrimination in the U.S. criminal justice system (e.g., Fagan and Ash, 2017), and specifically in the context of the war on drugs (Banks, 2003). We establish a racial bias in the effect of conservative discourse on criminal justice decisions, and this is linked to drug crimes. As Blacks are disproportionately arrested for non-violent drug-related offenses, the effect could be driven by racial bias in media messaging. Alternatively, it could be that tough-on-drugs rather than tough-on-crime rhetoric matters in this setting. In future work we will try to distinguish which types of rhetoric are more distinctive of Fox News. We see an effect for elected, but not appointed, judges. This is consistent with an electoral mechanism, where Fox News affects judge decisions by shifting voter attitudes, rather than shifting the policy preferences of judges directly. These regressions work by subsampling states, however, so we cannot rule out selective differences in responsiveness across states. In future work, we can do a within-state analysis using data from Kansas, where some judges are elected, and some are appointed (Gordon and Huber, 2007; Lim, 2013). Finally, the effect of Fox News on elected judges becomes weaker in the run-up to the election date. One interpretation of this result is that politicized information and politicized incentives are substitutes, rather than complements. As electoral pressures become stronger, media effects are reduced. Another possibility is that Fox News content becomes more election-focused, and less devote to crime, in the run-up to elections. We will try to distinguish these explanations in future work. More generally, in our follow-up work we plan to use text-analysis methods to recover which ideas in Fox News are driving our effects. Using recent developments in high- 17

18 dimensional instrumental-variables methods (e.g. Belloni et al., 2012; Ash, 2016), we will ask what features of cable news discourse have an impact on sentencing. The goal is a greater understanding of the political economy of media and criminal justice. The evidence produced by this research program will be useful to judges, policymakers, and the public. 18

19 References Allcott, Hunt, and Matthew Gentzkow Social media and fake news in the 2016 election. Journal of Economic Perspectives, 31(2): Arceneaux, Kevin, Martin Johnson, René Lindstädt, and Ryan J Vander Wielen The influence of news media on political elites: Investigating strategic responsiveness in congress. American Journal of Political Science, 60(1): Ash, Elliott The political economy of tax laws in the U.S. states. Ash, Elliott, and W Bentley MacLeod Intrinsic motivation in public service: Theory and evidence from state supreme courts. The Journal of Law and Economics, 58(4): Ash, Elliott, and W. Bentley MacLeod The performance of elected officials: Evidence from state supreme courts. NBER. Ash, Elliott, and W Bentley MacLeod Elections as Incentive and Selection Device: The Case of State Supreme Courts. Ash, Elliott, Daniel L Chen, and Suresh Naidu Ideas have consequences: The impact of economics on American justice. Ash, Elliott, Daniel L Chen, and Wei Lu Polarization of US Circuit Court Judges: A Machine Learning Approach. Banks, R Richard Beyond profiling: Race, policing, and the drug war. Stanford Law Review, Belloni, Alexandre, Daniel Chen, Victor Chernozhukov, and Christian Hansen Sparse models and methods for optimal instruments with an application to eminent domain. Econometrica, 80(6): Berdejó, Carlos, and Daniel L Chen Electoral cycles among us courts of appeals judges. The Journal of Law and Economics, 60(3): Berdejó, Carlos, and Noam Yuchtman Crime, punishment, and politics: an analysis of political cycles in criminal sentencing. Review of Economics and Statistics, 95(3): Boxell, Levi, Matthew Gentzkow, and Jesse M Shapiro Greater Internet use is not associated with faster growth in political polarization among US demographic groups. Proceedings of the National Academy of Sciences, Chen, Daniel L, Tobias J Moskowitz, and Kelly Shue Decision making under the gambler s fallacy: Evidence from asylum judges, loan officers, and baseball umpires. The Quarterly Journal of Economics, 131(3): Clinton, Joshua D, and Ted Enamorado The national news media s effect on Congress: How Fox News affected elites in Congress. The Journal of Politics, 76(4): DellaVigna, Stefano, and Ethan Kaplan The Fox News effect: Media bias and voting. The Quarterly Journal of Economics, 122(3): Dippel, Christian, and Michael Poyker Do Private Prisons Affect Court Sentencing? Epstein, Lee, William M Landes, and Richard A Posner The behavior of federal judges: a theoretical and empirical study of rational choice. Harvard University Press. Fagan, Jeff, and Elliott Ash New policing, new segregation? From Ferguson to New York. Georgetown Law Journal. Gentzkow, Matthew, and Jesse M Shapiro What drives media slant? Evidence from US daily newspapers. Econometrica, 78(1): Gordon, Sanford C, and Gregory A Huber The Effect of Electoral Competitiveness on Incumbent Behavior. Quarterly Journal of Political Science, 2(2):

20 Lim, Claire SH Preferences and incentives of appointed and elected public officials: Evidence from state trial court judges. The American Economic Review, 103(4): Martin, Greg, and Ali Yurukoglu Bias in Cable News: Persuasion and Polarization. American Economic Review. Mikolov, Tomas, Ilya Sutskever, Kai Chen, Greg S Corrado, and Jeff Dean Distributed representations of words and phrases and their compositionality Posner, Richard How Judges Think. Harvard University Press. Snyder, James, and David Stromberg Press Coverage and Political Accountability. Journal of Political Economy, 118(2):

21 Online Appendix to Conservative News Media and Criminal Justice: Evidence from Exposure to Fox News Channel 21

22 A Results for CNN and MSNBC This appendix includes the main graphical results for CNN and MSNBC. Figure A5 shows that higher CNN and MSNBC viewership are associated with higher sentencing cross-sectionally, although the relationship is weaker than with Fox News. Figure A6 shows that we can get a first stage for CNN, but not for MSNBC. Figure A7 shows that there are weak reduced form effects, which is negative for CNN (the same as Fox) but positive for MSNBC (the opposite of Fox). Figure 5: OLS relation: CNN/MSNBC Viewership and Incarceration Length Panel A (CNN) Panel B (MSNBC) Figure 6: First Stage: CNN/MSNBC Channel Positions and Viewership Panel A (CNN) Panel B (MSNBC) 22

23 Figure 7: Reduced Form: CNN/MSNBC Channel Positions and Incarceration Length Panel A (CNN) Panel B (MSNBC) 23

24 Figure 8: First-Stage Coefficients (by State) 24

25 Table 4: Fox News and Additional Outcomes Dependent variable: I II III IV V 1(Incarceration) 1(Conviction) # cases Predicted sentence length based on charges Sentence length, conditional on any sentence Nielsen share x x 0.071*** [0.8843] [0.1734] x x [0.0095] R-squared x x Partial R-squared x x F-stat. of excl. inst x x Sample 8 states NCRP 6 states Minnesota NCRP Observations 1,218,359 4,962,918 4,911,323 All columns include state-year FEs. The dependent variable is the inverse hyperbolic sine of the sentencing length. The following variables are used as controls: age, age squared, and race dummies (Black, Hispanic, Asian, and Native American), dummy for recidivists and state-specific set of crime severity dummies. Standard errors clustered by FIPS. P-values are in brackets. 25

26 Overlap and Table 5: Fox News and Sentencing Decisions: Subsample Analysis Dependent variable: Log sentencing length in months I II III IV Sample Nielsen share OLS 2SLS OLS 2SLS Fox news 0.018*** *** 0.086*** [0.0000] [0.8885] [0.0010] [0.0040] R-squared Partial R-squared F-stat. of excl. instrument F-stat. p-value Observations 2,387,252 2,375,967 2,586,946 2,586,946 All columns include state-year FEs. The dependent variable is the inverse hyperbolic sine of the sentencing length. The following variables are used as controls: age, age squared, and race dummies (Black, Hispanic, Asian, and Native American), dummy for recidivists and state-specific set of crime severity dummies. Standard errors clustered by FIPS. P-values are in brackets. 26

27 Table 6: Fox and Sentencing: Robustness to Inclusion of Additional Controls Dependent variable: Log sentencing length in months I II III IV V VI Nielsen share 0.075*** 0.074** x 0.073*** x x [0.0066] [0.0112] x [0.0011] x x 1996 republican vote share x [0.2589] x Log population x x x x t-1 avg. sentencing length 0.011*** x [0.0000] x t-1 crime rates x x R-squared Partial R-squared F-stat. of excl. inst Observations 4,962,918 4,889,187 3,828,688 All columns include state-year FEs. The dependent variable is the inverse hyperbolic sine of the sentencing length. The following variables are used as controls: age, age squared, and race dummies (Black, Hispanic, Asian, and Native American), dummy for recidivists and state-specific set of crime severity dummies. Standard errors clustered by FIPS. P-values are in brackets. x x 27