Empirical specification
The objective of the empirical exercise is to identify the effect of exposure to Fox on social distancing after the National Emergency was announced on March 13th. The pre-March 13th observations are used to test pre-trends. Our main specification is the following county-date panel regression:
$$ \mathit{SD}_{i(s)t} = \beta_{1}\mathit{FNCP}_{i(s)} \times \text{Before}_{t} + \beta_{2}\mathit{FNCP}_{i(s)} \times \text{After}_{t} + X_{i(s)}{\Gamma} + \mu_{s} + \lambda_{t} + \epsilon_{i(s)t}, $$
(1)
where SDi(s)t is a measure of social-distancing in a county i located in state s on date t, FNCPi(s) is the 2005 Fox position in channel lineup, and Beforet (Aftert) is a dummy equal to one for dates before (after) the national emergency. FNCPi(s) is normalized to have a mean of zero and a standard deviation of one, and a larger FNCPi(s) is associated with a smaller exposure to Fox News. We control for state (μs) and date (λt) fixed effects. Vector Xi(s) includes a set of county-level demographic and economic controls such as population density and poverty rate. Standard errors are clustered at the state level.Footnote 38
The coefficient of interest β2 captures the effect of Fox on social distancing after the National Emergency was announced. We expect it to be negative: counties with larger Fox lineup positions have a larger decrease in the daily distance traveled relative to their pre-COVID baseline. β1 represents the effect of Fox on social distancing before the National Emergency was announced, and we expect it to be zero, indicating that counties with difference Fox exposure did not exhibit differential social distancing behaviors in the pre-period.Footnote 39 Since our Fox News Channel lineup is measured in 2005, which is earlier than our study period (2020), we are eventually studying the heterogeneous effect of built-up Fox exposure on people’s behavioral choices, given that there is a national policy advocate.Footnote 40
In addition to the demographic and economic controls mentioned earlier, we take into account the potential confounding effect of the industrial composition on the relationship between Fox exposure and social distancing behaviors. Regions with more Fox exposure may have a particular employment mix, and as shown in Dingel and Neiman (2020), industries and occupations differ in their workability-at-home. We control for it directly in our regressions.
Fox exposure can affect the degree of conservatism, which can directly affect the social distancing behavior if a conservative population has different preferences or different constraints, and indirectly, through the interpretation of the COVID-related messages conveyed by Fox. These effects are on top of direct information feeds by Fox that may affect all of its audience, irrespective of their ideology. To separate (1) conservatism, (2) information, and (3) the interaction of the two, we also experiment with directly controlling for the county-level Republican vote shares in the 2012 and 2016 presidential elections and the 2016 turnout rate, which act as proxies for built-up conservatism. In this case, the remaining effect of Fox on social distancing should be either through the information feed or through the interaction of information and conservatism.Footnote 41,Footnote 42
One might be concerned that similar to the Fox exposure, other county-level characteristics also affect the social distancing behaviors differentially before and after the declaration of national emergency. If the Fox exposure is correlated with these characteristics, omitting this differential impact may bias our estimate of the Fox coefficient. To address this concern, we also consider a specification where we add the interaction of the controls Xi(s) with the After dummy.
Although we include a wide variety of county characteristics, there might be unobserved variables that are correlated with Fox exposure and affect the social distancing outcome. Thus, we also present a specification where instead of the state fixed effects, we use county fixed effects. Here, both the Fox interaction with the Before dummy and the levels of county characteristics Xi(s) will be absorbed by the fixed effects.
Main results
Table 2 shows the effect of Fox exposure on social distancing using various specifications. In panel A column 1, we estimate Eq. 1 with only state and date fixed effects. The estimand \(\widehat {\beta }_{1}\) is statistically insignificant, indicating that counties did not have differential patterns in social distancing before the National Emergency. The point estimate of interest \(\widehat {\beta }_{2}\) is negative and significant. It indicates that a one-standard-deviation increase in Fox News Channel lineup led to a 0.6-percentage-point larger decline in average distance traveled in a county.
Table 2 Effects of Fox News Channel position on reductions in mobility To put the magnitude of our results in context, the biggest decrease in distance traveled per person after March 13 happened in the District of Columbia (59%), and the smallest one—in Nevada (13%). According to the estimates of Martin and Yurukoglu (2017), moving Fox from channel 10 to channel 40 (approximately, two standard deviations) is associated with a 5-min reduction per week per person in time spent watching Fox. According to our results, when Fox is moved 30 positions higher in the cable lineup, it decreases social-distancing by one percentage point. Assuming linear marginal effects and comparing the effect size with the state-level reduction in mobility, we can roughly estimate that this effect can explain 2% and 8% of the total reduction in population movement in DC and Nevada, respectively.
In columns 2–5, we sequentially add controls for demographic and socio-economic variables. As documented in various other studies, higher education levels and higher incomes at the individual and region levels are positively associated with practicing of social distancing (Brzezinski et al. 2020, Fan et al. 2020, Mongey et al. 2020, Wright et al. 2020, among others). We control for the unemployment rate, urban dummies, economic-dependence county indicator, poverty rate, median income, population, the share of the population with a high-school education, county’s land area, the share of the nonwhite population, and net domestic migration rate. Column 6 further adds the employment share in workable-at-home jobs.Footnote 43 The coefficient estimate for the Fox effect remains almost identical compared to column 1.
As Fox can affect the general level of conservatism of the local population, it can potentially affect people’s response towards recommendations for social distancing. Column 7 adds controls for the turnout in 2016 and Republican vote share in the 2012 and 2016 elections. We find that controlling for these conservatism proxies does not affect the coefficient estimate of Fox exposure.Footnote 44 It suggests that our results are not driven by the accumulated Fox effect but by its immediate reaction to COVID-19, and possible by the interaction of the two.
We also want to test if the Fox effect comes from crowding out viewership of other media. If people watch less Fox and at the same time watch more of other channels such as CNN and MSNBC, our coefficient estimates may reflect the positive effect of other media instead of the negative effect of Fox. Panel B replicates panel A but adds controls for the channel positions of CNN and MSNBC. Neither of them appears to be significant and the coefficient for the Fox News Channel position lineup remains unchanged.Footnote 45
We show the robustness of our results using alternative specifications in panel C and panel D. Panel C adds the interaction terms of the controls with the After dummy to take into account differential effects of socio-economic and political characteristics on social distancing. Panel D uses county fixed effects instead of state fixed effects to account for additional unobserved factors. The results are very similar to panel A.
We also check if our results are driven by some specific regions. (Ananyev et al. 2021) find that urban and rural areas did not respond differentially to the Fox exposure (columns VI–VII of Table A.1). In addition, it was not driven by some particular state (Figure A.3).
An alternative way to define the start of people’s awareness of the policy recommendation of social distancing is using states’ shelter-in-place orders rather than the national emergency. Suppose that states where voters had been more exposed to Fox also voted for the government that was later in issuing stay-at-home order. In addition, people follow these state-level shelter-in-place orders. Then our effect can be explained by people with more exposure to Fox decreasing their movement less because of the lagged timing of shelter-in-place policies. Ananyev et al. (2021) find similar results of Fox using the shelter-in-place order timings, suggesting that people are paying attention to both federal and state recommendations and that the state order timings are not endogenous with respect to Fox News Channel positions (Table A.3).
Ananyev et al. (2021) also consider heterogeneous effects in Table A.4. We find some evidence that exposure to Fox News had smaller effects in the locations with a higher share of the population with a high-school education and a higher share of the population employed in workable-from-home industries (columns IV and V). We find no differential effects in urban locations in column III or locations with a higher number of Christian churches in column VI (that we consider as a proxy to conservatism).Footnote 46 We also find some suggestive evidence in columns VII and VIII, that counties that already reported first COVID cases and deaths also experienced smaller effect from the exposure to Fox News, suggesting that first-hand experience alleviated Fox mislead messages.
A natural extension would be to test the impact of Fox exposure on COVID cases and mortality rates. In Table A.5, Ananyev et al. (2021) document that locations more exposed to Fox experienced larger mortality rates from COVID-19, consistent with (Bursztyn et al. 2020). This suggests that Fox exposure can have important public health consequences through behavioral responses.
Event study evidence
In the previous Section, we show results for non-dynamic specifications, where there is only one coefficient estimate for the Fox exposure for all dates after the National Emergency was announced.Footnote 47 Alternatively, we allow separate point-estimates for weeks from February 24th to April 14th as follows:
$$ \begin{aligned} SD_{i(s)t(w)} = \underbrace{\sum\limits_{l=-4}^{-1} \gamma_{l} \cdot FNCP_{i(s)} \cdot \text{D}(w = l)}_{\text{pre-event period}} + \underbrace{\sum\limits_{l=0}^{4} \gamma_{l} \cdot FNCP_{i(s)} \cdot \text{D}(w=l)}_{\text{post-event period}} + \\ + X_{i(s)}{\Gamma} + \lambda_{t(w)} + \mu_{s} + \varepsilon_{i(s)t(w)}, \end{aligned} $$
(2)
where SDi(s)t(w) is social-distancing outcome of county i in state s at date t in week w. Week w = 0 is the week of March 13 to March 20. Week indices run from − 4 to 4 and represent the position of weeks relative to week w = 0. D(w = l) is a dummy equal to one if week w = l. Here, λt(w) are date fixed effects and μs are state fixed effects. Coefficients γl with l ≥ 0 capture the Fox exposure effect in the post national emergency period, and the ones with l < 0 capture pre-trends.
Figure 4 plots the resulting coefficients of Eq. 2 for the specification without controls (panel A) and with the full set of controls (panel B).Footnote 48 The first noteworthy feature is that neither specification exhibits pre-trends. There is an increase in the coefficient for the week prior to March 13th; however, the point estimate is insignificant. We fail to reject the joint F-test that the pre-event γls are zero. This suggests that the exact timing of the national emergency is not related to trends in social distancing in more-Fox-exposed counties and that social distancing behaviour did not start to change before the national emergency was announced.Footnote 49
The second noteworthy feature is that while we do not observe any effect at the week zero (γ0), four point estimates for four weeks after March 13th have almost the same magnitude as the point estimate of \(\widehat {\beta }_{2}\) from the baseline specification in Table 2. Thus, the effect is constant across all weeks and our baseline specification (1) captures the full time path of the effect. Ananyev et al. (2021) also show that the results also hold if one adds county fixed effects (Figure A.5, which is a similar specification to one in panel D of Table 2), using week t = − 1 as the baseline.
Ananyev et al. (2021) also replicate similar event-study graphs for the shelter-at-home orders in Figure A.4. Here, each state had its own relative time as week 0 started at the date when the state issued the order. While we see negative effects of the Fox News Channel position in the post-period, there are evident (while insignificant) downward pre-trends. This suggests that people might have started to decrease their mobility after the national emergency was announced but before their state officially ordered them to stay home.
Zip-code-level results
Thanks to Facebook’s “Data for Good” project, we are able to investigate the effect of slant media on zip-code-level data for the subsample of 14 states and DC. Since the channel positions are initially on the zip-code level, we decrease potential measurement error.
We first confirm that county-level social distancing measures using Facebook data are highly correlated with measures using UNACAST data. Ananyev et al. (2021) report in Figure A.6 the residual plots of the regression of UNACAST’s changes in distance traveled on Facebook’s distance traveled (panel A) and Facebook’s probability of staying at home (panel B). In both graphs, the measures are strongly correlated. (Ananyev et al. 2021) also show that our baseline results in Table 2 hold if we use county-level Facebook measures (Table A.7).
Because Facebook’s data start on March 10th, we can’t estimate pre-trends as we did in the baseline specification. In addition, instead of the changes in mobility, we observe the levels of mobility in the Facebook data. Thus, we control for the pre-COVID mobility more flexibly using the following equation:
$$ M_{j(s)t} = \beta FNCP_{j(s)} +\phi M_{j(s)t-45} + X_{j(s)}{\Gamma} + \mu_{s} + \lambda_{t} + \epsilon_{j(s)t}, $$
(3)
where Mj(s)t is the mobility measure of zip code j in state s and date t and Mj(s)t− 45 is the corresponding mobility measure in the 45 days before March 10. FNCPj(s) is the Fox lineup position in zip-code j in state s. We again control for state and time fixed effects. Vector Xj(s) now contains zip-code-level controls, including the number of Facebook bing tiles covered, number of Facebook users, population, population density, number of housing units, and land area.
Here, we use two measures of mobility: (i) probability of staying at home (panel A of Table 3) and (ii) daily distance traveled (panel B). In panel A column 1, we only control for baseline probability of staying at home, number of tiles, Facebook’s population, and date fixed effects. Fox News Channel lineup position has positive effects on staying home: one standard deviation increase in channel position results in a 0.1-percentage-point larger probability of staying at home. Columns 2 and 3 add controls for Facebook’s measure of population density and state fixed effects. Column 4 allows for state-and-date fixed effects. Finally, columns 5–7 add controls for population, number of housing units, and land area. The coefficient of interest remains unchanged and highly significant throughout all columns.Footnote 50 According to our results, among the 14 states (plus DC) where we have zip-code-level data, the 30-positions change in Fox increases the probability of staying at home by 0.2 percentage points. This explains 2% and 33% of the increase in the probability of staying at home in DC and West Virginia, respectively, which had the biggest and smallest changes.
Table 3 Zip-code-level evidence: more Fox News exposure, longer distance traveled, and smaller probability of staying at home Panel B reports results for the distance traveled. We, also find results consistent with our findings on the county-level: a one-standard-deviation increase in channel lineup explains 2.5% of differences in distance traveled between crisis and baseline measures. Overall, we find consistent evidence that Fox negatively affected social distancing responses both at the county and at the zip-code level.