partisan return gap: the polarized stock market in the
TRANSCRIPT
Partisan Return Gap:
The Polarized Stock Market in the Time of a Pandemic
Jinfei Sheng Zheng Sun Wanyi Wang*
February 2021
Abstract
We document sharp differences in stock price responses to COVID-19 related news between
public firms headquartered in blue counties (dominated by Democratic supporters) and those in
red counties (dominated by Republican supporters). Red-county stocks on average experience 18
basis points higher abnormal returns than blue-county stocks on days with important COVID-19
news. We call this “Partisan Return Gap”. The partisan return gap can be explained by a
behavioral channel: investors in red counties are less concerned about COVID-19 and give more
favorable interpretation to COVID-19 related news. Using smartphone app data tracking visits to
non-essential services (e.g. restaurants), we confirm that individuals in red counties conduct less
social distancing behavior in response to the surge of COVID-19 cases and government
lockdown orders. Moreover, we find that stocks in counties where investors conduct less social
distancing behavior have higher returns on COVID-19 news days. Exploiting Facebook
connections as an identification, we further show that the partisan return gap is unlikely fully
driven by local economic conditions and policies. Overall, the results are consistent with
investors’ partisanship affecting their attitude toward COVID-19, which leads to polarized stock
prices in the time of the pandemic.
Keywords: Stock market, COVID-19, Partisanship, Return gap, Polarization, Social finance,
Political finance
*All authors are with Merage School of Business at University of California, Irvine. Emails: [email protected],
[email protected], [email protected]. We thank Philip Bromiley, Diego Garcia, Jack Favilukis, David Hirshleifer,
Yukun Liu, Kelly Shue, Johannes Stroebel, and seminar participants at University of California Irvine for helpful
comments. We thank Zhiqi Rong, Diana He, Margaret Qu, Zoey Zhou, and Benlin Gan for excellent research
assistance. All errors are our own.
1
“The coronavirus crisis once seemed to be the kind of gut-wrenching shock that would pull
together a politically divided nation. Increasingly, though, it is pulling the nation apart along
familiar lines.”
– Wall Street Journal
1 Introduction
The global pandemic due to COVID-19 has hallmarked 2020 as one of the deadliest years
in human history. As of January 31, 2021, there are over 100 million cases and 2.2 million deaths
due to COVID-19 worldwide. In United States, there are over 26 million confirmed cases and
about 450, 000 deaths. The onset of the black swan event posed many questions: How deadly is
the virus? What should be the best response to curtail the disease? What will be the economic
impact of the pandemic? … Large disagreement persists regarding these issues, and the gap
seems to mainly arise along the line of partisanship. From elite leaders to news media, the
Republicans have consistently downplayed the severity of the disease and the necessity of the
social distancing measures. The Democratic Party, on the other hand, has emphasized the threat
of the disease and argued for strict lockdown policies. Studies show that a similar partisan
division exists among general households, where the Republicans are less likely to practice
social distancing measures (e.g., Allcott et al. 2020, Barrios and Hochberg 2020, Fan, Orhun and
Turjeman 2020, Gadarian, Goodman and Pepinsky 2020).
In this paper we study whether the partisan disagreement about COVID-19 could
generate a similar partisan gap in stock prices. In particular, we examine whether stocks of
companies headquartered in red and blue counties are priced differently during the first few
months of COVID-19, after controlling for the effect of fundamentals. Our study is premised on
the vast empirical evidence that investors tend to concentrate holdings in stocks to which they
2
are geographically close (e.g., Coval and Moskowitz 1999, Grinblatt and Keloharju 2001,
Huberman 2001, Hong, Kubik and Stein 2008). The implication is that the equilibrium prices of
stocks headquartered in red (blue) counties are more likely to be determined by the risk attitude
of Republican (Democratic) investors. Therefore, a divergence of their risk attitude during the
COVID-19 period could lead to polarized stock prices.
Despite the evidence that Americans’ health behaviors divide between partisanship, ex-
ante it is not obvious that such a gap carries over to the financial markets. First, scholars in
political sciences argue that apparent partisan gaps in beliefs can shrink substantially when there
are moderate incentives for accuracy (e.g., Bullock et al. 2015, Prior et al. 2015). Due to direct
consequences on their wealth, investors should have strong incentives to minimize biases from
their political beliefs. Second, even if differences of opinions exist among investors, stock prices
should reflect the average opinions of investors which could still be unbiased. Moreover,
arbitragers may serve as a correcting force that guard against any mispricing caused by non-
fundamental factors, including partisanship. In this paper, we ask if political gaps could persist in
asset prices, in the face of large incentives and arbitrage forces.
To study the effect of political polarization on stock returns during the pandemic, we first
identify the shocks of COVID-19 to the financial market by looking at aggregate stock market
movements. We find days on which the S&P 500 index moves up or down by more than 2.5%
and use news articles to identify whether the main reason for the swing is COVID-19. Our
research question is to examine whether responsiveness of the red-county stocks to COVID
shocks are different from that of the blue-county stocks.
Anecdotal evidence suggests that stocks in red and blue stocks reacted differently to
COVID shocks. For example, Range Resources Corporation and Montage Resources
3
Corporation are two companies in the Crude Petroleum and Natural Gas industry. Range
Resources Corporation is headquartered in Tarrant county of Texas and Montage Resources
Corporation is in the Dallas county of Texas. Tarrant and Dallas are neighboring counties with
similar size of population (2 million). The majority (54%) in Tarrant county voted for the
Republican party during the 2016 Election, while the majority (64%) of Dallas county voted for
the Democratic party. These two companies are in the same industry, located in adjacent counties
and have similar risk exposures.2 However, their stock price reactions to COVID-19 related news
are very different. The average risk-adjusted return of Range is 1.3% during the COVID-19
shocks days, while Montage’s average risk-adjusted return is -1.0%. 3 Figure 1 plots the
cumulative return of these two companies from January 2020 to June 2020 and it shows a big
gap between two stocks. The cumulative return of Range was higher than that of Montage since
early 2020 and the gap increased over time. This comparison suggests that there may be big
return difference in companies in red counties and companies in blue counties.
We examine all stocks in the United States and find that stocks of companies located in
red counties earn higher risk-adjusted returns than companies in blue counties on days when
COVID-19 related news triggered large market movements. In contrast, there is no statistical
difference in returns between the two groups of stocks outside of the Covid-19 news days. We
use risk-adjusted returns to alleviate the concern that our results are driven by the different risk
exposures of the two groups of companies. The economic magnitude is large. Firms in red
counties are associated with 18-21 basis points higher risk-adjusted returns on COVID-19 news
days. The result is robust to the inclusion of an extensive set of control variables such as severity
2 Betas for the Market, SMB, and HML factors are 1.01(1.05), 1.48(1.49), 0.93(0.81), respectively for Range
Resources Corporation (Montage Resources Corporation). 3 On positive COVID-19 news days, the average returns for Range Resources Corporation and Montage Resources
Corporation are 8.3% and 6.0%, respectively. On negative COVID-19 news days, the average returns for Range
Resources and Montage Resources are -4.7% and -7.1%, respectively.
4
of local COVID conditions, government lockdown orders, demographic information, and firm
characteristics. We call this difference “Partisan Return Gap.”
What could be an explanation of the Partisan Return Gap on the Covid-19 news days?
First, psychologists have long found that people display a tendency to search for and interpret
information in a way that confirms or supports their prior beliefs – “the confirmation bias”. (e.g.
Lord, Ross, and Lepper 1979, Taber and Lodge 2006, Westen et al. 2006). When encountering
Covid-19 related news, the confirmation bias could result in each partisan interpreting it as in
support of their existing attitudes, widening rather than narrowing the disagreement between
them. This combined with the well-established finding that investors are more likely to invest in
firms headquartered in the local areas (e.g., Coval and Moskowitz 1999, 2001), could lead to
more polarized stock prices between red and blue county stocks on Covid-19 news days.
Confirmation bias suggests that investors tend to believe in information that supports
their prior view and discount information that goes against it. Since Republicans are generally
less concerned about Covid-19 than Democrats, they will react more strongly to good Covid-19
news and less strongly to bad news. As a result, red county stocks are likely to experience higher
increases in prices on good news days and less price drops on bad news days. To test this
hypothesis, we look at days with positive and negative COVID-19 shocks separately. We find
that firms in red counties have higher risk-adjusted returns than firms in blue counties on good
news days (e.g., vaccine news), suggesting that they overreact to good news. Also, firms in red
counties have higher risk-adjusted returns than firms in blue counties on bad news days,
suggesting that they underreact to bad news. These results are consistent with the prediction of
confirmation bias. They also suggest that the big gap in stock price reactions to COVID-19
related news is unlikely due to missing risk factors. If the higher alpha on good news days by
5
red-county stocks relative to blue-county stocks is due to red-county stocks’ higher loadings on a
missing risk factor, then the same effect should generate a lower alpha for red-county stocks on
bad news days.
To provide more direct evidence that the partisan return gap is due to investors’ different
risk attitude toward COVID-19, we turn to individuals’ social distancing behavior during our
sample period. Using smartphone app data that tracks individuals’ visits to public places, we
confirm that, relative to blue county residents, people living in red counties pay more visits to
non-essential business (e.g. restaurants) in response to COVID-19 cases and lockdown policies,
manifesting their lower risk perceptions about the disease. More importantly, when replacing the
red county variable in our main specification with the change in visits to non-essential business,
we find that firms in counties with less social distancing behaviors earn higher risk-adjusted
returns on COVID-19 news days. Taken together, the results are consistent with that red county
residents perceive less risk in the face of COVID, as revealed by their less social distancing
behavior. Their less concern about COVID leads to a more favorable interpretation about
COVID news, contributing to a higher return earned by red county stocks during COVID-news
days.
Alternatively, the return gap we find can also be explained by the fundamental channel.
Despite extensive controls in the baseline specification, it is possible that our result is due to
omitted local factors that correlate with both local partisanship and local stock returns. To
address this issue, we employ an identification using the Social Connectedness Index (SCI)
introduced by Bailey et al. (2018). It is a county-level measure based on Facebook friendship
links, which captures the relative probability of residents in any two U.S. counties being
Facebook friends. Motivated by the finding that investors are more likely to invest in firms
6
located in counties where they have stronger social ties (Kuchler et al. 2020), we expect that
stock returns are more affected by the political beliefs of investors from more socially connected
counties. Thus, for each county, we construct a social-connection-based partisanship measure.
Because the measure only consists of the partisanship of geographically distant counties, it is
arguably exogenous to local factors of the focal county. We find that companies located in
counties with stronger social ties to Republican areas earn higher stock returns on COVID news
days than firms in counties with more social connections to Democratic areas. This finding
suggests that the difference in local economic fundamentals between red and blue counties
cannot fully explain the partisan return gap.
We also examine the fundamental channel at the firm level. It is possible that firms in
blue counties are fundamentally more negatively affected by COVID-19, and the return gap may
come from investors’ rational expectations of lower future earnings. However, we find no
evidence that companies in blue counties are hit harder by COVID-19. There is no statistical
difference in the changes in ROA between the two groups of companies. Regarding changes in
profitability, red-county firms are more negatively affected than blue-county firms, but the gap
becomes insignificant after we control for industry compositions. Thus, the difference in firm
fundamentals between red and blue counties cannot explain our observed pattern.
To gain further understanding of the sources of our partisan return gap, we conduct
several subsample tests. First, if the partisan return gap is due to home bias, it is likely to be more
pronounced in stocks that are more likely to be affected by local investors. Small firms and less
well-known firms (like non-S&P 500 stocks) are more likely to be held by local investors.
Indeed, we find the gap is concentrated among small firms and non-S&P 500 firms. Second, we
investigate the behavioral channel by comparing stocks that are more likely to be affected by
7
retail investors, because retail investors are more likely to be biased. We split companies by
institutional ownership and transaction cost. We find that the return gap only exists in firms with
low intuitional ownership and firms with high transaction cost. Third, presumably people with
higher income and higher education have more resources to learn about the disease so should be
less biased. Indeed, we find that the effects we document concentrate on companies
headquartered in low income and low education counties. Taken together, these findings support
a behavioral explanation that the difference in risk perceptions about Covid-19 between
Democratic and Republican investors leads to a gap in stock returns.
There may be concerns that our results are driven by unobservable differences between
companies in red and blue counties, and these differences have nothing to do with COVID-19
risk attitudes. To address the concern, we run a placebo test and repeat our procedures in 2018-
2019. During this earlier period, we do not observe the performance gap as we documented in
the main regression. Firms located in red counties do not have higher risk-adjusted returns than
firms in blue counties on stock market jump days.
Our results are robust to a number of alternative specifications, including alternative
benchmarks of risk-adjusted returns (CAPM; Fama-French Carhart four-factor model; Fama-
French five-factor model), alternative thresholds of market movement (1%, 3%, 5%), and
alternative measures of partisanship (county-level vs. state-level; continuous vs. discrete).
Our paper adds to a growing literature on how partisanship affects financial decisions and
outcomes. Kaustia and Torstila (2011) find that left-wing voters and politicians are less likely to
invest in stocks. Hong and Kostovetsky (2012) find that Democratic mutual fund managers hold
less of their portfolios in companies that are deemed socially irresponsible. Relatedly, Giuli and
Kostovetsky (2014) find that Democratic-leaning firms perform better in corporate social
8
responsibility. Hutton, Jiang, and Kumar (2014) demonstrate that personal political preferences
of corporate managers influence corporate policies. Hutton, Jiang, Kumar (2015) study whether
the political culture of a firm determines its propensity for corporate misconduct. Jiang, Kumar,
and Law (2016) show that analysts who contribute primarily to the Republican Party adopt a
more conservative forecasting style. Meeuwis, Parker, Schoar, and Simester (2019) show that
individuals in Republican areas invest their retirement assets more aggressively after the 2016
election. Like our paper, several papers link partisanship and economic decisions in the context
of COVID-19. For example, partisanship has been shown to affect risk preferences (Barrios and
Hochberg 2020) and stock market beliefs (Cookson, Engelberg and Mullins 2020). Unlike our
paper, most of the existing studies focus on micro-level data of individual preferences and
behaviors. While the individual data provide a more direct link between individual’s political
beliefs and their economic behavior, macro-level evidence is still needed to show that the
divergent political beliefs among different partisans do not end up netting each other in affecting
equilibrium outcomes. Our paper is one of the first to show that political beliefs could have an
important effect on asset prices, arguably one of the most important equilibrium variables in
finance.
The financial market experienced unprecedented volatility in the early weeks of the
COVID-19 pandemic. Several papers examine the causes of the price movement during this
period. Boudoukh, Liu, Moskowitz, and Richardson (2020) study the factor structure of returns
of different asset classes during COVID-19 times. Gormsen and Koijen (2020) use data from the
aggregate equity market and dividend futures to quantify investors’ expectations about economic
growth. Ding, Levine, Lin, and Xie (2020) study cross-firm stock price reactions to COVID-19
as functions of pre-shock corporate characteristics. These papers focus on the stock price
9
movement driven by fundamentals. On the other hand, Cox, Greenwald, and Ludvigson (2020)
use a dynamic asset pricing model to estimate the prices of the stock market risk and find that the
price fluctuation is mainly driven by shift in risk aversion or sentiment. Our paper shows that at
least part of the sentiment-driven price movement is due to investors’ political beliefs.
Out paper also contributes to the growing literature on social finance. Social finance
studies how social interaction affects financial decisions (Hirshleifer 2020, Han, Hirshleifer and
Walden 2020). Bailey et al (2018) show that social interaction measured by Facebook connection
can affect housing purchase decisions. Kuchler et al (2020) show that institutional investors are
more likely to invest in firms from regions to which they have stronger social connections. Our
paper shows that political belief can generate big impact on stocks that are geographically far
away through social networks. Our finding highlights the significant role of social interactions in
understanding asset pricing.
The rest of the paper is organized as follows. Section 2 offers a description for datasets,
measures and summary statistics. Section 3 shows the baseline results that there are significant
differences in asset prices between firms headquartered in blue counties and firms in red counties
during the COVID-19 period. Section 4 explores both behavioral and fundamental channels as
potential explanations of our baseline findings. Section 5 provides further discussions and
conducts robustness tests. Section 6 concludes.
2 Data and Measurement
2.1 Data
We start from a list of public companies from CRSP/Compustat Merged Database. We
restrict our sample to common share stocks listed on NYSE, Nasdaq and AMEX and exclude
10
companies that have no book value in the fiscal year ending in 2019 or no market value by the
end of 2019. We also exclude firms whose headquarter is outside the United States. Using tickers
of these companies, we download daily stock prices and trading volume from January 1, 2020 to
June 30, 2020 from CRSP. We then drop penny stocks whose price falls below $1 on any trading
day during the sample period. In total, there are 3,030 firms in our sample.
We merge firm headquarter information from Compustat with the HUD-USPS Crosswalk
File to convert zip codes to county FIPS codes. This allows us to link financial information with
geographic information such as partisanship. We obtain 2016 Presidential Election voting results
from MIT Election Data & Science Lab. We measure local partisanship with the proportion of
votes to the Republican and Democratic candidates in the area. A county is labeled as red (blue)
if the Republican candidate received more (less) votes than the Democratic candidate in the
county.
To measure individuals’ social distancing behavior, we use anonymized foot traffic data
provided by SafeGraph. Partnering with smartphone applications, SafeGraph obtains GPS
location data from 45 million smartphones and aggregates it into customer visits to public places.
There are over 6 million uniquely identified public places in the dataset, including shops,
restaurants, hotels, airports, etc. For each place, we observe its address, industry classification,
and the number of visits every day.
To measure social connections between counties in the United States, we use the Social
Connectedness Index developed by Bailey et al. (2018). Based on anonymized friendship links of
Facebook users, the county-level pairwise index captures the likelihood of residents in any two
U.S. counties being Facebook friends. Because Facebook has a large user base and requires
11
mutual consent to establish friendship links, Facebook friends provide a good proxy for real-
world social connections.
We obtain other state and county-level variables from various sources. Daily cumulative
COVID-19 cases are from the New York Times. Government lockdown orders are extracted
from the dataset collected by Keystone Strategy. Weekly unemployment claims are from U.S.
Department of Labor website. Demographics are from the 2012-2016 American Community
Survey (ACS) and the U.S. Census Bureau. Religiosity information is from “U.S. Church
Membership Data” collected by the Association of Religion Data Archives (ARDA).
2.2 Measurement
In order to examine the impact of COVID-19 on the stock market, the first and most
intuitive method that people usually think of is to use variables that directly measure the severity
of COVID-19 (e.g., cases, deaths, growth rate). However, in the United States, there is a huge
disconnect between the development of COVID-19 and the response of the stock market: the
market collapsed when there were only a few COVID-19 cases and rebounded and remained
high when COVID-19 cases grew rapidly. It turns out that the government and the Federal
Reserve’s policy responses to COVID-19 and investors’ prospects for the future development of
COVID-19 most affect the stock market. Therefore, we focus on days with COVID-19-related
news. For example, on March 12, 2020, the USA government declared a travel ban to Europe
and the S&P 500 index dropped 10%. In contrast, on March 24, 2020 and March 26, 2020,
medias reported that Congress and the government were about to reach a $2 trillion stimulus plan,
and the stock market rebounded strongly by 9% and 6%.
To identify important dates with COVID-19 related news, we focus on days when the
stock market moves up or down by more than 2.5%, which is about the threshold of top and
12
bottom 10% of S&P 500 returns during the sample period. The results are robust to using
alternative thresholds (see section 5.3). From January 1, 2020 to June 30, 2020, there are 33 days
when the stock market swings by more than 2.5%. We use news articles to identify whether the
main reason for the swing is COVID-19. Figure 2 plots some major events during the period.
The complete list of the 33 days with related news articles is in Appendix B. Among these 33
days, only 5 days are not related to COVID-19. The remaining 28 days are driven by COVID-19
news and we label them as COVID Shock days. We further define Positive (negative) COVID
Shock to distinguish market rises from market falls. There might be concern that COVID-19
related news occurs almost every day in 2020 and these 5 days might be also related to COVID-
19. In a robustness test in section 5.3, we also include these 5 days and the results are similar.
To measure firm-level stock responses to COVID-19 news, we use risk-adjusted daily
returns as the dependent variable in our main analysis. Specifically, we examine individual
stocks’ risk exposures to Fama-French three factors (Fama and French, 1993). We regress daily
excess returns on excess market return, SMB and HML from January 1, 2018 to June 30, 2020.
We then subtract the expected return based on Fama-French 3 factor model from the raw return.
We also calculate turnover as the daily trading volume divided by total shares outstanding. To
measure corporate earnings in the first two quarters of 2020, we define ROA as income before
extraordinary items divided by total assets. Following Novy-Marx (2013), we calculate gross
profitability as returns on gross profits (revenues minus cost of goods sold) scaled by total assets.
We also denote the year-over-year change of ROA and gross profitability as ROA and
Profitability.
To measure individuals’ social distancing behavior as a proxy for COVID-19 risk
perception, we calculate the change in visits to non-essential services compared to pre-pandemic
13
levels. Specifically, we define non-essential services as places whose 2-digit NAICS code is 71
(Arts, Entertainment, and Recreation) or 72 (Accommodation and Food Services).4 We then
count the total number of non-essential visits for each county on each day and calculate its 5-day
moving average to adjust for weekly seasonality.5 To measure the degree of social distancing, we
define visits as the 5-day moving average on day t divided by the visits at the beginning of 2020:
∆𝑣𝑖𝑠𝑖𝑡𝑠𝑖,𝑡 =(∑ 𝑣𝑖𝑠𝑖𝑡𝑠𝑖,𝜏)/5𝑡+2
𝜏=𝑡−2
(∑ 𝑣𝑖𝑠𝑖𝑡𝑠𝑖,𝜈)/55 𝑣=1
− 1
To measure local economic conditions and policy responses to COVID-19, we define
new cases as the state-level new COVID-19 cases per 1000 residents every day. % Unemp is the
unemployment claim rate during a week in a state. NPI indicates whether there is a state-level
“shelter-in-place”, “non-essential services closure” or “closing of public venues” order in effect
on a given day. We focus on the three types of lockdown orders as they have direct impact on
business operations. Besides, we define % Female as the percentage of female in a county’s total
population and HH Income as the median household income in the past 12 months. We measure
local religiosity as the proportion of a county’s total population that attends church, using the
survey conducted by the Association of Religion Data Archives (ARDA) in 2010. The total
religiosity ratio (TRR) is calculated as the number of adherents of all 217 religious
denominations divided by the total population of the county.
2.3 Summary statistics
Table 1 presents summary statistics of key variables in our main analysis. The average
risk exposure to Rm − Rf, SMB and HML is 0.93, 0.78 and 0.31, respectively. The average FF3
risk-adjusted return is 0.045%, and the average daily turnover is 1.28%. During the sample
4 Specifically, they are: theaters; sport centers; museums; historical sites; zoos; amusement parks; casinos; golf
courses; hotels and inns; RV parks and campgrounds; bars; restaurants; cafeterias. 5 There are only trading days in the data (Monday-Friday), so 5-day moving average eliminates weekly seasonality.
14
period, 22% of trading days are labeled as COVID Shock days, among which 10% are positive
shocks and 12% are negative shocks. Out of 3,030 firms in the sample, 20% reside in red
counties (i.e., counties dominated by Republican supporters), and the average county-level
voting shares to the Republican candidate in the 2016 Presidential Election is 37%. Regarding
COVID-19 severity, the average daily new COVID-19 cases are 0.074 per 1000 residents.
During Jan 1, 2020 – Jun 30, 2020, people reduces their visits to non-essential services by 38%.
The average unemployment claim rate is 13.9% at the state-level, and 43% of firm-date pairs are
associated with at least one of “shelter-in-place”, “non-essential services closure” or “closing of
public venues” orders. For counties in our sample, the average median household income is $68k
per year. The average proportion of women and proportion of residents attending church are both
51%. For firm characteristics, the average firm in the sample has a market value of $10.4 billion,
a book-to-market ratio of 0.57, and an institutional ownership of 54%. In the first two quarters of
2020, the average return on assets (ROA) and gross profitability is -2.4% and 4.9%, a decrease of
0.7 and 1.1 percentage points compared with the same period in 2019.
3 Partisan return gap
In this section, we document that there are striking differences between firms
headquartered in blue counties and firms headquartered in red counties. We examine both the
characteristics of firms and their stock prices during COVID-19 period.
3.1 Firms with color: red vs. blue
We first examine whether there are systematic differences in characteristics between
firms in blue counties and firms in red counties. Table 2 Panel A presents the comparison of
several firm characteristics. We find that firms headquartered in blue counties are bigger in terms
of market capitalization and have lower book-to-market ratio. Given the focus of this paper is on
15
stock returns, we also examine their risk exposures to Fama-French three factors. We find that
while there is no big difference in firms’ exposures to market factor and size factor, firms in red
counties have significantly higher exposure to the value risk factor. Given this finding, it is
important to control risk exposures when comparing stock returns of firms. Therefore, we use
Fama-French 3-factor alphas as a measure of stock performance.
We also look at the industry distributions of these firms based on Fama-French 12
industry classification. Table 2 Panel B shows that firms in red counties are concentrated in
industries like manufacturing, wholesale, retail, and some services, while firms in blue counties
are mainly in industries like healthcare, medical equipment, and drug, business equipment. Given
the important difference in industry distribution among these firms, we include industry fixed
effects in our regression to control for that.
3.2 Partisan return gap: result
We now examine how stock returns behave for firms in blue counties, compared to firms
in red counties. Motivated by the fact firms have different risk exposure to common risk factors,
we use the abnormal return adjusted by Fama-French 3 factors, 𝐴𝑏𝑛𝑅𝑒𝑡𝑖,𝑡 , as the dependent
variable. We estimate the following regression:
𝐴𝑏𝑛𝑅𝑒𝑡𝑖,𝑡 = 𝛼 + 𝛽1𝐶𝑂𝑉𝐼𝐷_𝑆ℎ𝑜𝑐𝑘𝑡 + 𝛽2𝑅𝑒𝑑𝑖 + 𝜷𝟑𝑪𝑶𝑽𝑰𝑫_𝑺𝒉𝒐𝒄𝒌𝒕 × 𝑹𝒆𝒅𝒊
+ ∑ 𝛾𝑖𝑋𝑖,𝑡
𝑛
𝑖=1+ ∑ 𝜃𝑖𝐶𝑂𝑉𝐼𝐷_𝑆ℎ𝑜𝑐𝑘𝑡 × 𝑋𝑖,𝑡
𝑛
𝑖=1+ 𝐹𝐸𝑠 + 𝜖𝑖,𝑡 (1)
where 𝐶𝑂𝑉𝐼𝐷_𝑆ℎ𝑜𝑐𝑘𝑡 is a dummy variable that equals to 1 if COVID-19 related news on day t
triggered S&P 500 index to move up or down by more than 2.5%. 𝑅𝑒𝑑𝑖 is a dummy variable that
16
equals to 1 if firm i is headquartered in a republican county.6 We define a county as a republican
county if Donald Trump received more share of votes in 2016 Presidential Election. Figure 3
displays the geographic distribution of red and blue counties.
Table 3 presents the results of this test. In column (2), the coefficient on 𝐶𝑂𝑉𝐼𝐷_𝑆ℎ𝑜𝑐𝑘
is negative and significant, suggesting that the abnormal returns on these COVID-19 shock
related days for blue-county stocks are negative. Although it may appear to be mechanical since
more of these shock days are days with big drops in the stock market, these are abnormal returns
that are adjusted for risk exposures. The variable Red itself is close to 0 and statistically
insignificant, suggesting that on the days with no Covid-19 news, there is no significant
difference in returns between the red and blue county stocks. The main variable of interest is the
interaction term 𝐶𝑂𝑉𝐼𝐷_𝑆ℎ𝑜𝑐𝑘 × 𝑅𝑒𝑑. Its coefficient is positive and significant, suggesting that
the abnormal returns of firms in red counties are greater than those of firms in blue counties on
COVID-19 news days. The economic magnitude is also large. The coefficient on the interaction
term is 0.19, suggesting that compared with blue-county stocks, red-county stocks earn 0.19%
higher abnormal returns on COVID-19 news days. We call this return difference as “Partisan
Return Gap”.
One concern is that the stock underperformance of firms located in blue counties is due to
the fact that blue counties are more affected by COVID-19 in the early stage of the pandemic. As
Figure 4 shows, Democratic-dominated areas are associated with more coronavirus cases, earlier
6 In our main specification, we use Red as an indicator variable instead of the percentage of Republican votes
because the effect of election results on stock return is likely to be nonlinear. It is more likely for the smaller county
to receive a more united election outcome, resulting either a very high or very low percentage of Republican votes.
However, the local investors in smaller counties are less likely to have an impact on the stock prices. Thus, the effect
of Republican votes on stock prices is likely to become weaker when the vote percentage is either very high or very
low. In the robustness section, however, we use the continuous percentage of Republican votes as our independent
variable, and show that our results are robust to this specification.
17
government lockdown orders, and higher unemployment rates. Korniotis and Kumar (2013)
show that when the local economy experiences a stronger recession relative to the national
economy, local investors are likely to become more risk averse and sell local stocks, which could
generate lower returns on the Covid-19 news days. We thus include state-level daily new
COVID-19 cases, whether a government order affecting business operations is in effect, weekly
new unemployment claims and their interactions with 𝐶𝑂𝑉𝐼𝐷_𝑆ℎ𝑜𝑐𝑘𝑡 as control variables. We
also control for county demographics such as gender, income and religiosity. Besides, Daniel and
Titman (1997) shows that firm characteristics provide additional explanatory power of returns
beyond corresponding factors. Thus, we also include market value and book-to-market ratio in
the regression. We further control for firm characteristics and the aggregate time trend with firm
and date fixed effects. Column (3)-(5) present the result. Although firms in areas with more
COVID-19 cases are associated with lower abnormal returns on shock days, our main effect, the
coefficient on 𝐶𝑂𝑉𝐼𝐷_𝑆ℎ𝑜𝑐𝑘 × 𝑅𝑒𝑑, remains unaffected and positively significant.
Another concern is that the stock underperformance of firms in blue counties may come
from the difference in industry compositions. In column (6), we include Fama-French 12
industries by date fixed effects to control for the industry difference. Our main coefficient of
interest remains significant under 5% significance level. Taken together, these findings suggest
that there are striking performance differences along partisanship in the cross-section of stock
returns.
4 Explanations
Why do stock prices of firms headquartered in red counties behave differently from those
headquartered in blue counties on Covid-19 news days? In this section, we explore both
behavioral and fundamental channels.
18
4.1 Behavioral channel
The behavioral bias of investors is one possibility that drives the big difference in the
market reactions to COVID-19 shocks between firms in red counties and firms in blue counties.
Psychologists find that people display a confirmation bias in that they tend to search for and
interpret information in a way that confirms or supports their prior beliefs. One effect of the
confirmation bias is belief polarization, a phenomenon in which disagreement becomes more
extreme as the different parties consider evidence on an issue. Psychologists found that the effect
is particularly strong when people encounter ambiguous evidence. Being a brand-new disease,
Covid-19 generated a lot of news that are subject to interpretation. When encountering these
ambiguous Covid-19 news, the confirmation bias could result in each partisan interpreting it as
in support of their existing attitudes, widening rather than narrowing the disagreement between
them. This, combined with well-established finding that the investors are more likely to invest in
firms headquartered in the local areas (e.g., Coval and Moskowitz 1999, 2001), could lead to
more polarized stock prices between red and blue county stocks on Covid-19 news days.
Confirmation bias suggests that investors tend to believe in information that supports
their prior opinions and discount information that goes against it. It is well-reported that residents
in Republican areas are, on average, more optimistic about Covid-19 than Democrats. Therefore,
Republicans will likely react more (less) strongly to good (bad) Covid-19 news than Democrats.
As a result, red county stocks are likely to experience higher price increases on good Covid-19
news days, and less price drops on bad news days. To test this hypothesis, we examine positive
and negative COVID shocks separately. Specifically, we run a similar regression as equation (1)
but with two interaction terms: Positive COVID Shock × Red and Negative COVID Shock × Red.
Table 4 presents the result. The coefficient on Positive COVID Shock × Red is positive and
19
significant for all specifications, suggesting that firms in red counties experience significantly
higher abnormal returns on days with good COVID-19 news. The coefficient on Negative
COVID Shock × Red is also positive and significant, suggesting that firms in red counties
experience less price decrease with bad COVID-19 related news. The results are consistent with
the confirmation bias hypothesis. Moreover, the positive coefficients on both good and bad news
days make it less likely for missing risk factors to be an alternative explanation. If the higher
alpha on good news days by red-county stocks relative to blue-county stocks is due to red-county
stocks’ higher loadings on a missing risk factor, then the same effect should generate a lower
alpha for red-county stocks on bad news days. 7
To provide further evidence that the difference in risk perception between Republicans
and Democrats plays a role in explaining the partisan return gap, we turn to individuals’ social
distancing behavior. Barrios and Hochberg (2020) find that as Trump voter share rises,
individuals engage in less social distancing behavior, as measured by lower drop in visits to
public places compared to the pre-pandemic level. They believe it is the manifestation of lower
perceptions of risk during the COVID-19 pandemic. If that’s true, we should find firms located
in counties with less social distancing behavior earn higher returns. Of course, social distancing
behavior could also be an indicator of local economic activities, which could be related to both
partisanship and local stock returns. Thus, we only focus on visits to non-essential services, such
as bars, restaurants, and movie theaters. We also exclude these industries in analyzing stock
performances, so that we can break the link between economic activities and stock returns.
7 Interestingly we find stronger results on good news days than bad news days. The stronger response to the good
news is consistent with the behavioral literature: People in good moods process information less critically and less
accurately, potentially leading to more biased asset prices (Schwarz 1990; Petty, Gleicher and Baker 1991; Sinclair
and Mark 1995).
20
To understand the relation between partisanship and social distancing behavior, we first
present the time trend of non-essential visits in red and blue counties, following Allcott et al.
(2020). As Figure 5 Panel A shows, the two groups are similar prior to the pandemic, but the
Democratic counties experience a sharper drop since mid-March, and the gap continues to widen
over time. However, as Panel B shows, there are also more COVID-19 cases in blue counties.
Thus, we consider the following regression specification to formally examine the relationship:
∆𝑉𝑖𝑠𝑖𝑡𝑠𝑖,𝑡 = 𝛽1 log(𝐶𝑂𝑉𝐼𝐷 𝑐𝑎𝑠𝑒𝑠𝑖,𝑡 + 1) + 𝛽2 log(𝐶𝑂𝑉𝐼𝐷 𝑐𝑎𝑠𝑒𝑠𝑖,𝑡 + 1) × 𝑅𝑒𝑑𝑖
+𝛽3𝑁𝑃𝐼 𝑒𝑛𝑓𝑜𝑟𝑐𝑒𝑖,𝑡 + 𝛽4𝑁𝑃𝐼 𝑒𝑛𝑓𝑜𝑟𝑐𝑒𝑖,𝑡 × 𝑅𝑒𝑑𝑖 + 𝛽5𝑁𝑃𝐼 𝑙𝑖𝑓𝑡𝑖,𝑡 + 𝛽6𝑁𝑃𝐼 𝑙𝑖𝑓𝑡𝑖,𝑡 × 𝑅𝑒𝑑𝑖
+𝜇𝑖 + 𝛾𝑡 + 𝜖𝑖,𝑡 (2)
where ∆𝑉𝑖𝑠𝑖𝑡𝑠𝑖,𝑡 is the change in non-essential visits in county i on day t compared to the
beginning of the year, 𝐶𝑂𝑉𝐼𝐷 𝑐𝑎𝑠𝑒𝑠𝑖,𝑡 is the cumulative COVID-19 cases in county i on day t,
and 𝑁𝑃𝐼 𝑒𝑛𝑓𝑜𝑟𝑐𝑒𝑖,𝑡 and 𝑁𝑃𝐼 𝑙𝑖𝑓𝑡𝑖,𝑡 are dummy variables indicating whether “shelter-in-place”,
“non-essential services closure” and “closing of public venues” orders have been enforced or
lifted in state i as of date t. We control for county and date fixed effects, and standard errors are
double clustered at the county and the date level.
Table 5 presents the result. Total COVID-19 cases and the enforcement of a government
lockdown order are negatively related to individuals’ visits to non-essential services, but the
effect is largely muted in red counties. On the contrary, residents in blue counties do not respond
to the lifting of lockdowns, while red-county residents increase their visits immediately. Taken
together, these evidences suggest that Republicans engage in less social distancing behavior in
response to COVID-19 cases and lockdown policies, manifesting their lower risk perceptions
about the disease. This is consistent with the findings of Barrios and Hochberg (2020).
21
Next, we analyze the relation between social distancing behavior and local stock returns.
We run a similar regression as equation (1), but we replace Red with the demeaned change in
non-essential visits (∆𝑉𝑖𝑠𝑖𝑡𝑠𝑑𝑒𝑚𝑒𝑎𝑛). We demean each observation by the average change in
visits on date t across all counties to remove the time trend and only focus on cross-sectional
comparisons. Importantly, we exclude companies operating in non-essential service industries
(NAICS code = 71 or 72) from the left-hand stocks to break the link between returns of these
companies and their daily operations. Table 6 presents the result. The coefficient on COVID
Shock × ∆𝑉𝑖𝑠𝑖𝑡𝑠𝑑𝑒𝑚𝑒𝑎𝑛 is positive and significant across all specifications, suggesting that firms
located in counties with less social distancing behaviors earn higher risk-adjusted returns on
COVID-19 news days.
Taken together, the results in Table 5 and 6 suggest that red county residents perceive
less risk in the face of COVID, as revealed by their less social distancing behavior. Their lower
concern about COVID contributes to a higher return earned by red county stocks during COVID-
news days.
4.2 Alternative channel
One concern is that firms’ stock prices may be affected by local economic conditions and
government policies. Indeed, prior literature show that local economic factors can affect stock
returns of local firms (Tuzel and Zhang 2016, Jin and Li 2020). To address this concern, we
exploit an identification strategy which utilizes the social connection information from Facebook.
Social Connectedness Index (SCI) is a county-level measure based on Facebook friendship
links.8 It captures the relative probability of residents in any two U.S. counties being Facebook
8 The data is available here: https://dataforgood.fb.com/tools/social-connectedness-index/.
22
friends. It is calculated as the number of Facebook friend pairs between two counties divided by
the product of the two counties’ population and then scaled up by a factor of 1012.
Figure 6 presents two examples. Panel A plots the Social Connectedness Index of New
York County, NY. It is the county with the largest number of firm headquarters (184 companies)
among all Democratic counties in our dataset. Although geographically close areas are more
socially connected, we also observe strong connections with distant regions (e.g. Florida and
California). This allows our identification to separate the effect of partisanship from other local
factors. In panel B, we also plot the social connectedness to Maricopa County, AZ, the county
with the greatest number of headquarters among all Republican counties (42 companies).
Based on Social Connectedness Index, for each county, we construct a new measure of
partisanship, Social-Connection-based Partisanship (SCP). Following Kuchler et al. (2020), we
calculate a log weighted sum of Republican voting shares in the 2016 Presidential Election,
where the weight is the SCI between the county of interest and other counties (excluding own
county or own state), as follows:
𝑆𝐶𝑃𝑖 = 𝑙𝑜𝑔 ( ∑ (𝑆𝐶𝐼𝑖,𝑗 × 𝑅𝑒𝑝𝑉𝑜𝑡𝑒𝑗)
𝑗(𝑗≠𝑖)
)
Because the measure only consists of the partisanship of geographically distant counties,
it is arguably exogenous to local factors of the county in which a firm locates. Figure 7 presents
the geographic distribution of the SCP measure. The correlation between our main partisanship
measure (whether a firm locates in a red county) and the social-connection-based measure is 0.51.
Kuchler et al. (2020) show that institutional investors are more likely to invest in firms
from regions to which they have stronger social ties. Moreover, this effect of social proximity on
23
investment behavior is distinct from the effect of geographic proximity. These findings make the
social-based partisanship measure an ideal case for identification. On the one hand, the social
connection based partisan measure are likely to be correlated with our main partisanship measure.
On the other hand, the social-connection-based measure is less likely to be correlated with other
local factors such as local economic conditions or policies. Using social connection partisanship
(SCP), we test whether firms located in counties with strong social ties to red counties (likely
investors) tend to have different stock reactions to COVID shock. We run a similar regression as
in Equation (1) and replace Red with SCP. Table 7 shows the result of this test. The coefficient
on the interaction term COVID Shock × SCP is positive and significant, suggesting that stocks in
the red counties have much greater abnormal returns to COVID shocks relative to that of the blue
county stocks. The economic magnitude is also large. A one standard deviation increase in the
SCP is associated with a 9.75% percentage point increase in the return on COVID news days.
This result is similar to what we find in Table 3. This finding indicates that it is unlikely that the
polarization in the stock market is fully driven by local economic conditions.
Alternatively, it is possible that firms headquartered in red counties are fundamentally
doing better than firms in blue counties during the COVID-19 period, result in higher stock
returns. To test this conjecture, we examine whether firms in red counties have stronger
fundamentals. Specifically, we look at changes in earnings, measured by changes in ROA and
profitability. Table 8 presents the results. The coefficient on Red is insignificant and close to zero,
suggesting that firms in red counties do not have stronger fundamentals. As for profitability, red-
county firms are in fact more negatively affected than blue-county firms, but the coefficient
becomes insignificant after we control for industry fixed effects. This suggests that the gap is
largely explained by industry composition difference: the energy industries (coal, natural gas,
24
and oil) are hit the hardest during the pandemic, and they mainly locate in red counties; the
commercial equipment and healthcare industries are the least hit, and they mainly locate in blue
counties. Taken together, it is unlikely that fundamentals are driving the different responsiveness
of the red county stocks to COVID shocks relative to that of the blue county stocks.
5 Discussion
5.1 Subsample analysis
To further explore the behavioral channel, we conduct several subsample analyses. First,
we look at firms that are more likely to be affected by local investors. In particular, we look at
small firms and firms not in the S&P 500 list. These firms are less known and hence are more
likely to be held by local investors. Table 9 presents the results. We find indeed the effect is
concentrated among small firms and non-S&P 500 firms.
Second, while retail investors are more likely to over-react or under-react to COVID-19
news, institutional investors are more professional. Thus, we expect that the return gap is larger
for firms that are traded more by retail investors. Table 10 Panel A shows the result of this test.
We find the effect is mainly concentrated among firms with low institutional ownership, which is
consistent with our conjecture.
Third, if the return gap is due to behavioral bias, the mispricing should be stronger for
stocks with severer limits to arbitrage such as higher transaction costs. We examine firms with
different levels of transaction cost, as measured by turnover. Table 8 Panel B shows that the
partisan return gap is concentrated in firms with low turnover.
Finally, we examine how partisan return gap varies with income and education levels.
Presumably, people with higher income and higher education have more resources to learn about
25
the disease so should be less biased. In Table 11 Panel A and B, we examine the partisan return
gap for high and low income and education counties, respectively. We find that the effects
concentrate on companies headquartered in low income and low education counties. The
stronger partisan return gap for low-income counties is particularly interesting yet concerning.
The result suggests the low-income investors are more likely to be subject to the behavioral
biases in their trading, potentially leading to more trading losses and higher income inequality.
5.2 Placebo test
There may be concerns that our results are driven by unobservable differences between
companies in red and blue counties, and these differences have nothing to do with COVID-19
risk attitudes. For example, Miller (1977) argue that differences of opinion and short sale
constraints could lead to higher stock prices. If there is larger disagreement among investors in
red counties than in blue counties, we could observe higher returns for stocks headquartered in
red counties on news days. In this case, our result should hold on any big market movement day,
regardless of the trigger. To address the concern, we run a placebo test and repeat our procedure
in 2018-2019. We are able to identify 13 days on which the market moved by more than 2.5%.
Table 12 presents the result. The coefficient on the interaction term is very close to zero and not
significant across all specifications. Thus, firms located in red counties do not earn higher risk-
adjusted returns than firms in blue counties on stock market jump days. During this earlier period,
we do not observe the performance gap as we documented in the main regression.
5.3 Robustness
We conduct several robustness tests. First, we consider alternative measures of abnormal
returns. In the main specification, we use Fama-French three factor model as the benchmark.
Alternatively, we adjust daily returns with the Capital Asset Pricing Model, Carhart four factor
26
model (Carhart 1997) and Fama-French five factor model (Fama and French 2015). As Table 13
Panel A shows, our results are robust to all three measures. Our results also hold if we exclude
firms in the finance and utilities industry from our sample.
Next, we consider alternative thresholds of market movements. In the main body of the
paper, we identify COVID shocks by using days on which 1) the market fluctuates by more than
2.5% and 2) the movement is triggered by news related to COVID-19. To alleviate the concern
that our attribution of the trigger may not be comprehensive as there is news about COVID-19
almost every day, we include all days on which the market fluctuates by more than 2.5%,
regardless of the reason. As column (1) in Table 13 Panel B shows, our results continue to hold.
To ease the concern that our result is subject to the use of a 2.5% threshold, we also consider
days on which S&P 500 index moves up or down by more than 1%, 3%, and 5%. There are 68,
26, and 10 trading days identified respectively under these thresholds. Table 13 Panel B column
(2) – (4) present the result. The coefficient on the interaction term is significant across all
specifications. The results also show that the magnitude of the coefficients increases with the
threshold, suggesting that the partisanship return gap is bigger among days with greater market
movements.
Finally, we consider alternative measures of partisanship. In our main regression, we
divide companies into two groups based on whether they locate in a red county. One may also
consider investors in the same state but from different counties as local investors. Thus, in Table
13 Panel C, we show that our results are similar if we measure partisanship at the state level. Our
results also hold if we use the percentage of Republican votes as a measure of partisanship.
27
6 Conclusion
This paper studies the impact of political polarization on the stock market during the
COVID-19 pandemic. We document sharp differences in asset prices in response to COVID-19
related news between public firms headquartered in blue counties and those in red counties. Red-
county stocks have higher abnormal returns than blue-county stocks on both good and bad
COVID-19 news days.
We further examine the underlying channels through which political polarization affects
stocks’ reactions to COVID-19 shocks. We find that there is no significant difference in firm’s
fundamentals. The effect we find is more likely to be attributed to the behavioral channel:
investors in red counties are less concerned about COVID-19 and give more favorable
interpretation about COVID related news, which affects their trading on stocks. Given the
importance of local investors (i.e., home bias), local investors’ attitude towards COVID-19 can
explain the striking effect we find. We explore the behavioral channel by looking people’s social
distancing behavior during COVID as well as how it affects local stock returns on COVID news
days. Using smartphone app data tracking visits to non-essential services (e.g. restaurants), we
confirm that individuals in red counties conduct less social distancing behavior in response to the
surge of COVID-19 cases and government lockdown orders. Moreover, we find that stocks in
counties where investors conduct less social distancing behavior have higher returns on COVID-
19 news days.
One concern is that the effect may be due to the difference in local economic conditions
and policies since there are dramatic difference in responses to COVID-19 across different states
and counties. Exploiting Facebook connections as an identification, we find that companies
located in counties with stronger social ties to Republican areas earn higher returns on the
28
COVID news days than firms in counties with more social connections to Democratic areas.
Overall, the results are consistent with investors’ partisanship affecting their risk attitude toward
COVID-19, which leads to polarized stock prices in the time of the pandemic.
Our findings help us to better understand the effect of political polarization through the
lens of the stock market. The politicization of COVID-19 not only affects residents’ health
welfare, but also has significant impact on their financial investments.
29
References
Allcott, H., Boxell, L., Conway, J., Gentzkow, M., Thaler, M. and Yang, D., 2020. Polarization and
public health: Partisan differences in social distancing during the coronavirus pandemic. Journal
of Public Economics, 191, p.104254.
Bailey, M., Cao, R., Kuchler, T. and Stroebel, J., 2018. The economic effects of social networks:
Evidence from the housing market. Journal of Political Economy, 126(6), pp.2224-2276.
Barrios, J.M. and Hochberg, Y., 2020. Risk perception through the lens of politics in the time of the
covid-19 pandemic. Working paper. National Bureau of Economic Research.
Boudoukh, J. Liu, Y., Moskowitz, T., and Richardson, M., 2020. Risk, Return and Diversification in
Times of Crisis: (How) Is COVID-19 Different? Working paper. Yale University.
Bullock, J., Gerber, A., Hill, S., Huber, G., 2015. Partisan bias in factual beliefs about politics. Quarterly
Journal of Political Science, 10(4), pp. 519–578.
Carhart, M.M., 1997. On persistence in mutual fund performance. Journal of Finance, 52(1), pp.57-82.
Cookson, J.A., Engelberg, J.E. and Mullins, W., 2020. Does partisanship shape investor beliefs? Evidence
from the COVID-19 pandemic. Review of Asset Pricing Studies, 10(4), pp.863-893.
Coval, J.D. and Moskowitz, T.J., 1999. Home bias at home: Local equity preference in domestic
portfolios. Journal of Finance, 54(6), pp.2045-2073.
Coval, J.D. and Moskowitz, T.J., 2001. The geography of investment: Informed trading and asset
prices. Journal of Political Economy, 109(4), pp.811-841.
Cox, J., Greenwald, D.L. and Ludvigson, S.C., 2020. What Explains the COVID-19 Stock
Market? Working paper. National Bureau of Economic Research.
Daniel, K. and Titman, S., 1997. Evidence on the characteristics of cross-sectional variation in stock
returns. Journal of Finance, 52(1), pp.1-33.
Di Giuli, A. and Kostovetsky, L., 2014. Are red or blue companies more likely to go green? Politics and
corporate social responsibility. Journal of Financial Economics, 111(1), pp.158-180.
Ding, W., Levine, R., Lin, C. and Xie, W., 2020. Corporate immunity to the COVID-19 pandemic (No.
w27055). Journal of Financial Economics, forthcoming.
Fama, E.F. and French, K.R., 1993. Common risk factors in the returns on stocks and bonds. Journal of
Financial Economics, 33(1), pp.3-56.
Fama, E.F. and French, K.R., 2015. A five-factor asset pricing model. Journal of Financial
Economics, 116(1), pp.1-22.
Fan, Y., Orhun, A.Y. and Turjeman, D., 2020. Heterogeneous actions, beliefs, constraints and risk
tolerance during the covid-19 pandemic. Working paper, National Bureau of Economic Research.
Gonzalez-Uribe, J. and Wang, S., 2020. The effects of small-firm loan guarantees in the UK: insights for
the COVID-19 pandemic crisis. Working paper. Available at SSRN 3382280.
Gormsen, N.J. and Koijen, R.S., 2020. Coronavirus: Impact on stock prices and growth
expectations. Review of Asset Pricing Studies, 10(4), pp.574-597.
30
Grinblatt, M. and Keloharju, M., 2001. How distance, language, and culture influence stockholdings and
trades. Journal of Finance, 56(3), pp.1053-1073.
Han, B., Hirshleifer, D. and Walden, J., 2020. Social transmission bias and investor behavior. Journal of
Financial and Quantitative Analysis, forthcoming.
Hirshleifer, D., 2020. Presidential address: Social transmission bias in economics and finance. Journal of
Finance, 75(4), pp.1779-1831.
Hong, H., Kubik, J.D. and Stein, J.C., 2008. The only game in town: Stock-price consequences of local
bias. Journal of Financial Economics, 90(1), pp.20-37.
Hong, H. and Kostovetsky, L., 2012. Red and blue investing: Values and finance. Journal of Financial
Economics, 103(1), pp.1-19.
Huberman, G., 2001. Familiarity breeds investment. Review of Financial Studies, 14(3), pp.659-680.
Hutton, I., Jiang, D. and Kumar, A., 2014. Corporate policies of Republican managers. Journal of
Financial and Quantitative Analysis, 49(5-6), pp.1279-1310.
Hutton, I., Jiang, D. and Kumar, A., 2015. Political values, culture, and corporate litigation. Management
Science, 61(12), pp.2905-2925.
Jiang, D., Kumar, A. and Law, K.K., 2016. Political contributions and analyst behavior. Review of
Accounting Studies, 21(1), pp.37-88.
Jin, Z. and Li, F.W., 2020. Geographic Links and Predictable Returns. Working paper. Available at SSRN
3617417.
Kaustia, M. and Torstila, S., 2011. Stock market aversion? Political preferences and stock market
participation. Journal of Financial Economics, 100(1), pp.98-112.
Korniotis, G. and Kumar, A., 2013. State-Level Business Cycles and Local Return Predictability. Journal
of Finance, 68(3), pp. 1037-1096.
Kuchler, T., Li, Y., Peng, L., Stroebel, J. and Zhou, D., 2020. Social proximity to capital: Implications for
investors and firms. Working paper. National Bureau of Economic Research.
Kushner Gadarian, S., Goodman, S.W. and Pepinsky, T.B., 2020. Partisanship, health behavior, and
policy attitudes in the early stages of the COVID-19 pandemic. Working paper. Available at
SSRN 3562796.
Lord, C., Ross, L. and Lepper M., 1979. Biased assimilation and attitude polarization: The effects of prior
theories on subsequently considered evidence. Journal of Personality and Social Psychology, 37
(11), pp. 2098–109
Meeuwis, M., Parker, J.A., Schoar, A. and Simester, D.I., 2018. Belief disagreement and portfolio choice.
Working paper. National Bureau of Economic Research.
Memon, S.A., Razak, S. and Weber, I., 2020. Lifestyle disease surveillance using population search
behavior: Feasibility study. Journal of Medical Internet Research, 22(1), p.e13347.
Miller, E., 1977. Risk, uncertainty, and divergence of opinion. Journal of Finance, 32(4), pp.1151-1168.
MIT Election Data Lab, Science, 2018. County Presidential Election Returns 2000-2016. V42. Harvard
Dataverse. https://doi.org/10.7910/DVN/NH5S2I.
31
Novy-Marx, R., 2013. The other side of value: The gross profitability premium. Journal of Financial
Economics, 108(1), pp.1-28.
Petty, R. E., Gleicher, F. & Baker, S. M. 1991. Multiple roles for affect in persuasion. In: FORGAS, J. P.
(ed.) International series in experimental social psychology. Emotion and social
judgments. Elmsford, NY, US: Pergamon Press.
Prior, M., Sood, G., Khanna, K., 2015. You cannot be serious: the impact of accuracy incentives on
partisan bias in reports of economic perceptions. Quarterly Journal of Political Science, 10 (4),
p.p. 489–518.
Schwarz, N., 1990. Feelings as information: Informational and motivational functions of affective states.
The Guilford Press.
Sinclair, R.C. and Mark, M.M., 1995. The effects of mood state on judgemental accuracy: Processing
strategy as a mechanism. Cognition & Emotion, 9(5), pp.417-438.
Taber, C. and Lodge M., 2006. Motivated Skepticism in the Evaluation of Political Beliefs. American
Journal of Political Science 50 (3), pp. 755-769.
Tuzel, S. and Zhang, M.B., 2017. Local risk, local factors, and asset prices. Journal of Finance, 72(1),
pp.325-370.
Westen, D., Blagov, P., Harenski, K., Kilts, C. and Hamann S., 2006. Neural Bases of Motivated
Reasoning: An fMRI Study of Emotional Constraints on Partisan Political Judgment in the 2004
U.S. Presidential Election. Journal of Cognitive Neuroscience 18 (11), pp. 1947–1958
Wright, W.F. and Bower, G.H., 1992. Mood effects on subjective probability assessment. Organizational
Behavior and Human Decision Processes, 52(2), pp.276-291.
32
Figure 1. Cumulative returns of Range and Montage
This figure plots the cumulative stock returns of Range Resources Corporation and Montage Resources
Corporation between January 1, 2020 and June 30, 2020. The blue line indicates Montage Resources
Corporation (MR), and the red line proxies for Range Resources Corporation (RRC).
33
Figure 2. Major events triggering large market movements
This figure plots the historical price of S&P 500 index and major events that triggered large market movements between January 1, 2020 and June 30, 2020.
Mar 2, +4.6%.
Expectation on
Fed to cut rate.
Mar 12, -9.5%
Trump declares
travel ban on Europe
Mar 13, +9.4%.
Trump declares national
emergency; Pelosi says
House will pass a relief bill.
Mar 11, -4.9%
WHO declares
global pandemic
Mar 9, -7.6%
Oil price crashes
Mar 16, -12.0%
FED cuts rate by
100bp and plans to
buy $700B bonds.
Feb 24/25/27,
-3.4/-3.0/-4.4%
Surge of cases in
Europe and signs of
spread in the U.S.
Mar 3, -2.8%
Fed cut rate by 50bp.
Mar 17, +6.0%.
Fed to launch funding
facility; Trump seeks $1
trillion to fight covid.
Mar 18, -5.2%
U.S. - Canada
border close
Mar 24/26, 9.4%/6.2%
Agreement on $2T
coronavirus bill.
Apr 6/8, 7.0%/3.4%
Slower case growth in
NY and lower death
rate in Europe.
Apr 21, -3.1%
Crude oil contract
dropped to negative.
Apr 17/29, +2.7%/2.7%
Gilead’s breakthrough
on Remdesivir.
May 18, +3.2%
Moderna’s
progress on
vaccine.
Jun 11, -5.9%
COVID-19
infections resurge.
34
Figure 3. 2016 Presidential Election Result
This figure plots the share of votes received by Donald Trump in the 2016 Presidential Election. Red indicates more votes to the Republican party. Blue indicates
more votes to the Democratic party. The darker the color, the greater the difference in votes between the two parties.
35
Figure 4. COVID-19 cases, government orders, and unemployment across states
This figure plots the geographic distribution of COVID-19 cases, government lockdown orders and unemployment
rate across U.S. states. Panel A shows the cumulative covid-19 cases as of 6/30/2020; panel B presents the earliest
start date of three types of government orders: non-essential business closure, closing of public venues, and shelter-
in-place; panel C plots the unemployment claim rates as of 7/4/2020.
Panel A: Cumulative COVID-19 cases (as of June 30, 2020)
Panel B: Government NPI start date
Panel C: Unemployment claim rate (as of July 4th, 2020)
36
Figure 5. Time trends of non-essential visits and COVID-19 cases
This figure plots the time trend of visits to non-essential services and cumulative COVID-19 cases for Republican
and Democratic counties. Panel A shows the non-essential visits pattern. We define non-essential services as places
whose 2-digit NAICS code is 71 (Arts, Entertainment, and Recreation) or 72 (Accommodation and Food Services).
For each county on each day, we count the total number of visits to non-essential services, calculate its 5-day
moving average to adjust for weekly seasonality, and then normalize it with the visits at the beginning the year.
Panel B shows the cumulative COVID-19 cases. The red line indicates Republican counties, and the blue line
indicates Democratic counties.
Panel A: Visits to non-essential services
Panel B: cumulative COVID-19 cases
37
Figure 6. Examples of Facebook Social Connectedness Index
This figure shows two examples of Facebook Social Connectedness Index. Social Connectedness Index measures
the relative probability of people in two counties being Facebook friends. It is calculated as the number of Facebook
friend pairs between two counties divided by the product of the two counties’ population and then scaled up by a
factor of 1012. Panel A presents the social connectedness to New York County, NY. It is the county with the greatest
number of firm headquarters (N=184) among Democratic-dominated counties. Panel B shows the social
connectedness to Maricopa County, AZ, the county with the greatest number of firm headquarters in our dataset
(N=42) among Republican-dominated counties.
Panel A: New York County, New York
Panel B: Maricopa County, Arizona
38
Figure 7. Social-Connection-based Partisanship
This figure plots the Social-Connection-based Partisanship (SCP) measure across U.S. counties. SCP is the
logarithm of a weighted average of the share of votes to the Republican party in the 2016 Presidential Election,
where the weight is the Social Connectedness Index between the focal county and other counties (excluding own
county). Social Connectedness Index measures the relative probability of people living in two counties being
Facebook friends. It is calculated as the number of Facebook friend pairs between two counties divided by the
product of the two counties’ population. It is then scaled up by a factor of 1012 to become an integer. A county with
a high SCP means that it is more socially connected to the Republican party. Thus, high SCP counties are colored
red, and low SCP areas are colored blue.
39
Table 1. Summary statistics
This table presents the summary statistics of key variables at the firm-date level. The sample period is from January
1, 2020 to June 30, 2020. Panel A shows variables related to stock performances. Raw return measures daily stock
returns winsorized at the top/bottom 1% on COVID shock days. Excess return equals raw return minus risk-free rate.
FF3 α, CAPM α, FFC4 α, and FF5 α are daily returns adjusted by Fama-French three-factor model, CAPM model,
Fama-French-Carhart four-factor model, and Fama-French five-factor model, respectively. 𝛽𝑅𝑚−𝑅𝑓, 𝛽𝑆𝑀𝐵 and 𝛽𝐻𝑀𝐿
are risk exposures on excess market return, SMB and HML factors. All factor loadings are estimated using daily
returns from Jan 1, 2018 to Jun 30, 2020. Turnover is calculated as the daily trading volume divided by total shares
outstanding. Panel B presents variables related to COVID-19 shocks. COVID Shock indicates days on which
COVID-19 related news triggered S&P 500 index to move up or down by more than 2.5%. Positive (negative)
COVID Shock indicates days on which COVID-19 related news triggered S&P 500 index to move up (down) by
more than 2.5%. Panel C summarizes measures of partisanship. Red is a dummy variable that equals to 1 if the firm
is headquartered in a republican county (i.e., counties where Trump received more votes). Red (state) indicates
whether the firm is headquartered in a republican state. % Rep Vote (county /state) is the percentage of votes the
Republican party at the county/state level. Social-Connection-based Partisanship (SCP) is the logarithm of a
weighted average of the Republican voting shares, where the weight is the Social Connected Index between the
focal county and other counties based on Facebook friendship links (excluding own county or own state). All voting
shares are measured using the 2016 Presidential Election data. Panel D displays summary statistics of local variables.
Visits is the change in non-essential visits compared to the beginning of 2020. For each county, we replace non-
essential visit with its 5-day moving average to eliminate weekly seasonality. New Cases is the number of new
COVID-19 cases per 1000 residents in a state on a day. % Unemp is the state-level new unemployment claim rate
during a week. NPI indicates whether there is a state-level “shelter-in-place”, “non-essential services closure” or
“closing of public venues” order in effect on a day. % Female is the percentage of female in the county’s total
population. HH Income is the median household income in the past 12 months. Total Religiosity Ratio (TRR) is the
proportion of a county’s total population that attends church. Panel E shows firm characteristics. ME is the market
value in millions of dollars on December 31, 2019. BE is the book value in millions of dollars for the fiscal year
ending in 2019. B/M is the book-to-market ratio. S&P 500 is a dummy variable that equals to 1 if the stock is a S&P
500 index constituent as of December 31, 2019. Institutional Ownership equals the shares of stocks held by
institutional investors divided by total shares outstanding. Corporate earnings measures are at the firm-quarter level.
ROA is income before extraordinary items divided by total assets. Following Novy-Marx (2013), gross profitability
is calculated as returns on gross profits (revenues minus cost of goods sold) scaled by total assets. ROA is the year-
over-year change of ROA, and Profitability is the year-over-year change of gross profitability.
N Mean STD Min P25 P50 P75 Max
Panel A: stock performance (%)
Raw return 384,425 -0.0038 5.42 -23.8 -2.35 0 2.19 22.1
Excess return 384,425 -0.0068 5.42 -23.8 -2.35 -0.006 2.19 22.1
FF3 α 384,416 0.045 4.04 -37.0 -1.62 -0.037 1.51 37.1
CAPM α 384,416 -0.033 4.37 -37.0 -1.88 -0.13 1.59 38.3
FFC4 α 384,416 0.039 4.00 -35.5 -1.62 -0.035 1.51 33.1
FF5 α 384,416 0.051 3.98 -37.6 -1.60 -0.032 1.51 36.1
Turnover 384,427 1.28 1.97 0.001 0.36 0.72 1.35 14.0
𝛽𝑅𝑚−𝑅𝑓 384,435 0.93 0.33 -0.63 0.77 0.97 1.15 1.87
𝛽𝑆𝑀𝐵 384,435 0.78 0.61 -1.19 0.34 0.78 1.20 2.95
𝛽𝐻𝑀𝐿 384,435 0.31 0.59 -1.65 -0.074 0.30 0.72 2.21
Panel B: COVID Shocks
COVID Shock 384,445 0.22 0.42 0 0 0 0 1
Pos. COVID Shock 384,445 0.10 0.30 0 0 0 0 1
Neg. COVID Shock 384,445 0.12 0.32 0 0 0 0 1
40
(Table 1 continued)
N Mean STD Min P25 P50 P75 Max
Panel C: partisanship measures
Red 381,295 0.20 0.40 0 0 0 0 1
Red state 383,437 0.41 0.49 0 0 0 1 1
% Rep Vote (county) 381,295 37.0 16.4 4.30 23.8 37.2 45.8 91.4
% Rep Vote (state) 383,437 45.5 9.60 4.30 35.3 47.2 52.7 75.7
SCP (ex. county) 382,303 15.7 0.75 14.5 15.1 15.6 16.3 18.8
SCP (ex. state) 382,303 15.3 0.51 14.4 14.9 15.3 15.7 17.4
Panel D: local variables
Visits 361,059 -0.38 0.30 -0.86 -0.66 -0.44 -0.073 0.13
New Cases 380,390 0.074 0.15 -0.19 0 0.028 0.086 1.85
% Unemp 383,437 13.9 14.4 0.34 1.80 10.3 20.8 84.2
NPI 384,445 0.43 0.50 0 0 0 1 1
% Female 381,295 51.0 0.91 45.8 50.5 51.0 51.6 54.3
HH Income ($1000) 381,295 68.2 18.0 29.5 54.4 64.4 79.9 114.3
TRR 380,917 0.51 0.11 0.17 0.44 0.50 0.60 1.22
Panel E: firm characteristics
ME ($ million) 384,445 10,249.3 50,437.4 0 249.0 1,057.1 4,322.3 1,304,764.8
B/M 384,194 0.57 0.82 -10.5 0.20 0.45 0.81 17.5
S&P 500 384,445 0.14 0.35 0 0 0 0 1
Institutional
Ownership 384,445 0.54 0.32 0 0.27 0.52 0.85 1.15
ROA (%) 6,213 -2.41 7.83 -46.3 -2.63 0.11 1.03 9.02
Profitability (%) 5,489 4.85 7.59 -23.3 1.44 4.87 8.45 33.0
ROA (%) 6,124 -0.66 5.69 -23.5 -1.55 -0.19 0.42 31.9
Profitability (%) 5,402 -1.13 4.17 -18.2 -2.09 -0.44 0.27 16.7
41
Table 2. Firms in Republican vs. Democratic counties
This table presents summary statistics for firms headquartered in republican counties and democratic counties.
Republican (democratic) counties are defined as counties where Trump received more (less) votes than Clinton in
the 2016 Presidential Election. The sample period is from Jan 1, 2020 to Jun 30, 2020. We restrict our sample to
stocks listed on Nasdaq, NYSE and Amex. We exclude companies that have no book value in the fiscal year ending
in 2019 or no market value by the end of 2019. We also exclude stocks whose price falls below $1 during the sample
period. Panel A shows firm characteristics. ME is the market value in millions of dollars as of Dec 31, 2019; BE is
the book value in millions of dollars for the fiscal year ended in 2019; BE/ME is the book-to-market ratio; 𝛽𝑅𝑚−𝑅𝑓,
𝛽𝑆𝑀𝐵 and 𝛽𝐻𝑀𝐿 are risk exposures estimated using daily returns from Jan 1, 2018 to Jun 30, 2020. Panel B presents
the number of firms by Fama-French 12 industry. Panel C lists the distribution of firm headquarters by state.
Panel A: firm characteristics Republican Democratic Difference (Dem.- Rep.)
Variable Frequency Mean Frequency Mean Diff. T-stat
ME ($ mil) 608 4,482.49 2,422 11,561.73 7,029.24*** 3.11
BE ($ mil) 608 1,767.43 2,421 3,559.30 1,791.87** 2.54
BE/ME 608 0.67 2,420 0.55 -0.12*** -3.25
𝛽𝑅𝑚−𝑅𝑓 608 0.88 2,421 0.95 0.065*** 4.39
𝛽𝑆𝑀𝐵 608 0.77 2,421 0.78 0.019 0.67
𝛽𝐻𝑀𝐿 608 0.49 2,421 0.26 -0.23*** -8.49
Panel B: industry distribution Rep. Dem. Total
Fama-French 12 industry Freq. Pct. Freq. Pct. Freq. Pct.
Consumer Non-Durables 26 4.28 91 3.76 117 3.86
Consumer Durables 15 2.47 50 2.06 65 2.15
Manufacturing 69 11.35 186 7.68 255 8.42
Oil, Gas, and Coal Extraction and Product 25 4.11 64 2.64 89 2.94
Chemicals and Allied Products 11 1.81 53 2.19 64 2.11
Business Equipment 72 11.84 407 16.80 479 15.81
Telephone and Television Transmission 7 1.15 48 1.98 55 1.82
Utilities 17 2.80 58 2.39 75 2.48
Wholesale, Retail, and Some Services 45 7.40 202 8.34 247 8.15
Healthcare, Medical Equipment, and Drug 58 9.54 521 21.51 579 19.11
Finance 179 29.44 458 18.91 637 21.02
Other -- Mines, Constr, BldMt, Trans, Hotel, etc. 84 13.82 284 11.73 368 12.15
Total 608 100.00 2,422 100.00 3,030 100.00
Panel C: headquarter distribution
State Freq. State Freq. State Freq. State Freq.
California 536 Minnesota 60 Iowa 17 Idaho 8
New York 295 Connecticut 58 Kentucky 17 Maine 8
Texas 274 Washington 54 Louisiana 17 Mississippi 8
Massachusetts 209 Maryland 52 South Carolina 17 West Virginia 7
Pennsylvania 146 Michigan 49 Kansas 16 District of Columbia 6
Illinois 135 Wisconsin 49 Alabama 15 South Dakota 5
New Jersey 119 Arizona 45 Oregon 14 North Dakota 4
Florida 116 Indiana 44 Arkansas 13 New Mexico 4
Ohio 99 Tennessee 42 Delaware 13 Vermont 3
Virginia 86 Missouri 36 Nebraska 12 Montana 2
Georgia 81 Nevada 30 Hawaii 10 Wyoming 1
Colorado 71 Utah 25 Rhode Island 10
North Carolina 64 Oklahoma 19 New Hampshire 9 Total 3,030
42
Table 3. Partisan return gap
This table presents the partisan difference of firms’ risk-adjusted return on COVID-19 shock days. The sample period is from
January 1, 2020 to June 30, 2020. We restrict our sample to stocks listed on Nasdaq, NYSE and Amex. We exclude companies
that have no book value in the fiscal year ending in 2019 or no market value by the end of 2019. We also exclude stocks whose
price falls below $1 during the sample period. The dependent variable is daily return adjusted by Fama-French 3 factor model (3-
factor α). Factor loadings are estimated using daily returns from Jan 1, 2018 to Jun 30, 2020. COVID Shock indicates days on
which covid-19 related news triggered S&P 500 index to move up or down by more than 2.5%. Red is a dummy variable that
equals to 1 if the firm is headquartered in a republican county (i.e., counties where Trump received more share of votes in the
2016 Presidential Election). NPI indicates whether there is a state-level “shelter-in-place”, “non-essential services closure” or
“closing of public venues” order in effect on that day. % Unemp is the state-level new unemployment claim rate during the week.
New Cases equal the number of new COVID-19 cases per 1000 residents in a state on that day. % Female is the percentage of
female in the county’s total population. HH Income is the median household income in the past 12 months. TRR is the proportion
of a county’s total population that attends church. Log(1+ME) is the logarithmic form of one plus the market value in millions of
dollars on December 31, 2019. BE is the book value in millions of dollars for the fiscal year ending in 2019. B/M is the
corresponding book-to-market ratio. Standard errors are clustered at the date level. T-statistic is reported in parentheses. ∗, ∗∗,
∗∗∗ indicate significance at the 10%, 5%, and 1% level, respectively.
(1) (2) (3) (4) (5) (6)
COVID Shock -0.15 -0.19* -2.52 -2.50 (-1.65) (-1.88) (-1.30) (-1.28) Red -0.03 -0.02 (-0.92) (-0.78) COVID Shock × Red 0.19** 0.21*** 0.21*** 0.20*** 0.18** (2.51) (2.87) (2.84) (2.68) (2.60) NPI 0.02 0.03 0.03 0.03 (0.54) (0.69) (0.69) (0.95) % Unemp 0.00 0.00 -0.00 -0.00 (0.52) (0.37) (-1.30) (-1.50) New Cases 0.02 0.01 0.13 0.16** (0.10) (0.08) (1.37) (2.45) % Female 0.00 (0.17) HH Income 0.00 (0.76) TRR -0.11 (-1.25) Log(1+ME) -0.04*** (-3.06) B/M -0.03 (-1.62) COVID Shock × NPI 0.38*** 0.38*** 0.07 0.11 (3.18) (3.13) (0.81) (1.44) COVID Shock × Unemp -0.00 -0.00 0.01** 0.00** (-0.24) (-0.23) (2.03) (2.27) COVID Shock × New Cases -0.48* -0.47* -0.48*** -0.46*** (-1.92) (-1.84) (-3.23) (-3.62) COVID Shock × % Female 0.02 0.02 0.01 0.00 (0.54) (0.53) (0.33) (0.03) COVID Shock × Income -0.00 -0.00 -0.00 -0.00* (-1.14) (-1.13) (-1.22) (-1.70) COVID Shock × TRR 0.24 0.23 0.22 0.28** (1.11) (1.09) (1.09) (2.03) COVID Shock × log(1+ME) 0.15*** 0.15*** 0.15*** 0.15*** (4.31) (4.28) (4.31) (4.38) COVID Shock × B/M 0.09* 0.08* 0.09* 0.07 (1.74) (1.72) (1.74) (1.47) Constant 0.08*** 0.08*** 0.22 0.05* -0.37 -0.25 (3.00) (2.78) (0.29) (1.79) (-0.83) (-0.73)
Firm FE N N N Y Y Y
Date FE N N N N Y Y
FF12 × Date FE N N N N N Y
R2 0.000 0.000 0.002 0.006 0.011 0.044 Observations 384416 381266 377363 377363 377363 377363
43
Table 4. Positive vs. negative COVID news days
This table presents the partisan difference of firms’ risk-adjusted return on days when COVID-19 related news
triggered large market movements (up and down respectively). The sample period is from Jan 1, 2020 to Jun 30,
2020. We restrict our sample to stocks listed on Nasdaq, NYSE and Amex. We exclude companies that have no
book value in the fiscal year ending in 2019 or no market value by the end of 2019. We also exclude stocks whose
price falls below $1 during the sample period. The dependent variable is Fama-French 3-factor alpha. Positive
(negative) COVID Shock indicates days on which covid-19 related news triggered S&P 500 index to move up (down)
by more than 2.5%. Red is a dummy variable that equals to 1 if the firm is headquartered in a republican county.
Control variables include NPI, % Unemp, New Cases, % Female, HH Income, TRR, log(1+ME), B/M and their
interactions with Positive and Negative COVID Shock. NPI indicates whether there is a state-level “shelter-in-place”,
“non-essential services closure” or “closing of public venues” in effect on that day. % Unemp is the state-level new
unemployment claim rate during the week. New Cases represent the number of new COVID-19 cases per 1000
residents in a state on that day. % Female is the percentage of female in the county’s total population. HH Income is
the median household income in the past 12 months. TRR is the proportion of a county’s total population that attends
church. Log(1+ME) is the logarithmic form of one plus the market value in millions of dollars on December 31,
2019. BE is the book value in millions of dollars for the fiscal year ending in 2019. BE/ME is the corresponding
book-to-market ratio. Standard errors are clustered at the date level. T-statistic is reported in parentheses. ∗, ∗∗, ∗∗∗
indicate significance at the 10%, 5%, and 1% level, respectively.
(1) (2) (3) (4) (5) (6)
Positive COVID Shock -0.11 -0.17 -5.23* -5.22*
(-0.75) (-1.03) (-1.87) (-1.85)
Negative COVID Shock -0.18* -0.20* -0.58 -0.56
(-1.74) (-1.82) (-0.24) (-0.23)
Red -0.03 -0.02
(-0.92) (-0.78)
Positive COVID Shock × Red 0.30** 0.34** 0.34** 0.32** 0.26**
(2.21) (2.58) (2.56) (2.31) (2.06)
Negative COVID Shock × Red 0.09* 0.11** 0.11** 0.11* 0.10**
(1.69) (2.06) (2.02) (1.95) (2.14)
Constant 0.08*** 0.08*** 0.22 0.05* -0.39 -0.26
(3.00) (2.78) (0.29) (1.78) (-0.88) (-0.78)
Controls, Interactions N N Y Y Y Y
Firm FE N N N Y Y Y
Date FE N N N N Y Y
FF12 × Date FE N N N N N Y
R2 0.000 0.000 0.002 0.007 0.012 0.044
Observations 384416 381266 377363 377363 377363 377363
44
Table 5. Partisanship and social distancing behavior
This table presents the effect of partisanship on individuals’ social distancing behavior in response to COVID-19
cases and government lockdown orders. The dependent variable is Visits, the change in non-essential visits
compared to the visits at the beginning of 2020. For each county, we replace non-essential visits with its 5-day
moving average to eliminate weekly seasonality. Ln(1+cases) is the log form of one plus the cumulative COVID-19
cases in the county. NPI enforce is a dummy variable that equals to 1 if any of a state-level “shelter in place”, “non-
essential services closure”, or “closing of public venues” order has taken into effect as of that day. NPI lift is a
dummy variable that equals to 1 if all of the state-level “shelter in place”, “non-essential services closure”, or
“closing of public venues” orders have been lifted as of that day. Red is a dummy variable that equals to 1 if the firm
is headquartered in a republican county. County fixed effects and date fixed effects are controlled across all columns.
Standard errors are double clustered at the county and the date level. T-statistic is reported in parentheses. ∗, ∗∗, ∗∗∗
indicate significance at the 10%, 5%, and 1% level, respectively.
(1) (2) (3) (4) (5) (6)
Ln(1+cases) -0.03*** -0.03*** -0.03*** -0.03***
(-9.14) (-9.87) (-9.26) (-7.26)
Ln(1+cases) × Red 0.02*** 0.02***
(11.15) (4.63)
NPI enforce -0.04*** -0.12*** -0.03** -0.04**
(-2.69) (-6.81) (-2.23) (-2.21)
NPI enforce × Red 0.13*** 0.02
(10.66) (0.80)
NPI lift 0.05*** -0.01 0.05*** 0.01
(4.96) (-0.79) (5.39) (0.69)
NPI lift × Red 0.08*** 0.05***
(5.65) (3.88)
Constant -0.20*** -0.23*** -0.29*** -0.28*** -0.19*** -0.22***
(-17.56) (-20.91) (-34.03) (-34.54) (-14.06) (-17.76)
County FE Y Y Y Y Y Y
Date FE Y Y Y Y Y Y
R2 0.911 0.924 0.900 0.919 0.913 0.926
Observations 59260 58903 59260 58903 59260 58903
45
Table 6. Social distancing behavior and stock performance
This table presents the relation between social distancing behavior and firms’ risk-adjusted return on COVID-19
shock days. The sample period is from Jan 1, 2020 to Jun 30, 2020. We restrict our sample to stocks listed on
Nasdaq, NYSE and Amex. We exclude companies that have no book value in the fiscal year ending in 2019 or no
market value by the end of 2019 and stocks whose price falls below $1 during the sample period. We also exclude
stocks operating in industries (NAICS code = 71 or 72) that are used to calculate visits to non-essential services. The
dependent variable is FF3 α. COVID Shock indicates days on which covid-19 related news triggered S&P 500 index
to move up or down by more than 2.5%. Visitsdemean is the demeaned change in visits to non-essential services. We
demean by subtracting the average change in visits from each observation on a daily basis. Control variables include
NPI, % Unemp, New Cases, % Female, HH Income, TRR, log(1+ME), B/M and their interactions with COVID
Shock. NPI indicates whether there is a “shelter-in-place”, “non-essential services closure” or “closing of public
venues” order in effect on that day. % Unemp represents state-level new unemployment claim rate during that week.
New Cases represent the number of new COVID-19 cases per 1000 residents in a state on that day. % Female is the
percentage of female in the county’s total population. HH Income is the median household income in the past 12
months. TRR is the proportion of a county’s total population that attends church. Log(1+ME) is the log of one plus
the market value in millions of dollars on December 31, 2019. BE is the book value in millions of dollars for the
fiscal year ended in 2019. B/M is the corresponding book-to-market ratio. Standard errors are clustered at the date
level. T-statistic is reported in parentheses. ∗, ∗∗, ∗∗∗ indicate significance at the 10%, 5%, and 1% level,
respectively.
(1) (2) (3) (4) (5) (6)
COVID Shock -0.14* -0.15* -2.29 -2.33
(-1.67) (-1.78) (-1.15) (-1.17)
Visitsdemean -0.14 -0.02 0.11 0.08 0.00
(-0.96) (-0.14) (0.63) (0.44) (0.00)
COVID Shock × Visitsdemean 1.01** 1.04** 1.10** 1.10** 1.03**
(2.45) (2.36) (2.48) (2.25) (2.49)
Constant 0.08*** 0.09*** 0.04 0.05** -0.32 -0.21
(3.03) (3.26) (0.05) (2.06) (-0.64) (-0.59)
Controls, Interactions N N Y Y Y Y
Firm FE N N N Y Y Y
Date FE N N N N Y Y
FF12 × Date FE N N N N N Y
R2 0.000 0.000 0.002 0.006 0.011 0.042
Observations 374966 352111 350564 350564 350564 350564
46
Table 7. Identification based on Facebook Social Connectedness Index
This table presents the partisan difference of firms’ risk-adjusted return on COVID-19 shock days using an
identification strategy. The sample period is from Jan 1, 2020 to Jun 30, 2020. We restrict our sample to stocks
listed on Nasdaq, NYSE and Amex. We exclude companies that have no book value in the fiscal year ending in
2019 or no market value by the end of 2019. We also exclude stocks whose price falls below $1 during the sample
period. The dependent variable is daily return adjusted by Fama-French 3 factors (3-factor α). COVID Shock
indicates days on which covid-19 related news triggered S&P 500 index to move up or down by more than 2.5%.
Social-Connection-based Partisanship (SCP) is the logarithm of a weighted average of the voting shares to the
Republican party in the 2016 Presidential Election, where the weight is the Social Connected Index between the
focal county and other counties based on Facebook friendship links (excluding own county in Panel A and excluding
own state in Panel B). Control variables include NPI, % Unemp, New Cases, % Female, HH Income, TRR,
log(1+ME), B/M and their interactions with COVID Shock. NPI indicates whether there is a “shelter-in-place”,
“non-essential services closure” or “closing of public venues” order in effect on that day. % Unemp represents state-
level new unemployment claim rate during that week. New Cases represent the number of new COVID-19 cases per
1000 residents in a state on that day. % Female is the percentage of female in the county’s total population. HH
Income is the median household income in the past 12 months. TRR is the proportion of a county’s total population
that attends church. Log(1+ME) is the logarithmic form of one plus the market value in millions of dollars on
December 31, 2019. BE is the book value in millions of dollars for the fiscal year ended in 2019. B/M is the
corresponding book-to-market ratio. Standard errors are clustered at the date level. T-statistic is reported in
parentheses. ∗, ∗∗, ∗∗∗ indicate significance at the 10%, 5%, and 1% level, respectively.
Panel A: excluding own county
(1) (2) (3) (4) (5) (6)
COVID Shock -0.15 -2.24** -4.48* -4.45*
(-1.65) (-2.38) (-1.75) (-1.74)
SCP (ex. county) -0.03 -0.02
(-1.65) (-1.26)
COVID Shock × SCP 0.13** 0.12** 0.12** 0.13** 0.12**
(2.37) (2.05) (2.03) (2.05) (2.06)
Constant 0.08*** 0.61* 0.68 0.05* -0.86 -0.72
(3.00) (1.82) (0.78) (1.83) (-1.37) (-1.44)
Firm FE N N N Y Y Y
Date FE N N N N Y Y
FF12 × Date FE N N N N N Y
R2 0.000 0.000 0.002 0.006 0.011 0.044
Observations 384416 382274 377363 377363 377363 377363
Panel B: excluding own state
(1) (2) (3) (4) (5) (6)
COVID Shock -0.15 -2.50** -3.99 -3.95
(-1.65) (-2.30) (-1.64) (-1.62)
SCP (ex. state) -0.03 -0.01
(-1.27) (-0.51)
COVID Shock × SCP 0.15** 0.11* 0.11* 0.13* 0.13**
(2.28) (1.78) (1.75) (1.93) (2.05)
Constant 0.08*** 0.59 0.37 0.05* -0.77 -0.68
(3.00) (1.42) (0.39) (1.81) (-1.28) (-1.41)
Firm FE N N N Y Y Y
Date FE N N N N Y Y
FF12 × Date FE N N N N N Y
R2 0.000 0.000 0.002 0.006 0.011 0.044
Observations 384416 382274 377363 377363 377363 377363
47
Table 8. Earnings
This table presents the partisan difference in corporate earnings during COVID-19. The sample period is 2020Q1
and 2020Q2. We restrict our sample to stocks listed on Nasdaq, NYSE and Amex. We exclude companies that have
no book value in the fiscal year ending in 2019 or no market value by the end of 2019. We also exclude stocks
whose price falls below $1 during the sample period. We calculate ROA as income before extraordinary items
divided by total assets. Following Novy-Marx (2013), we calculate gross profitability as returns on gross profits
(revenues minus cost of goods sold) scaled by total assets. The dependent variable in column (1)-(3) is ROA, the
year-over-year change of ROA. The dependent variable in column (4)-(6) is Profitability, the year-over-year
change of gross profitability. We include log(1+ME), B/M, Past 12-month return, ∆𝑅𝑂𝐴𝑞−4 and ∆𝑃𝑟𝑜𝑓𝑖𝑡𝑎𝑏𝑖𝑙𝑖𝑡𝑦𝑞−4
as control variables. Log(1+ME) is the logarithm of one plus the market value in millions of dollars on December 31,
2019. BE is the book value in millions of dollars for the fiscal year ending in 2019. B/M is the corresponding book-
to-market ratio. Past 12-month return is the cumulative stock return for the 12 months prior to the end of the quarter.
∆𝑅𝑂𝐴𝑞−4 and ∆𝑃𝑟𝑜𝑓𝑖𝑡𝑎𝑏𝑖𝑙𝑖𝑡𝑦𝑞−4 are one-year lags of ROA and Profitability. Robust standard errors are applied
in all columns. T-statistic is reported in parentheses. ∗, ∗∗, ∗∗∗ indicate significance at the 10%, 5%, and 1% level.
ROA Profitability
(1) (2) (3) (4) (5) (6)
Red 0.07 0.06 0.19 -0.31** -0.31** -0.19
(0.50) (0.40) (1.30) (-2.23) (-2.27) (-1.38)
∆𝑅𝑂𝐴𝑞−4 -25.04*** -25.75***
(-6.80) (-7.03)
∆𝑃𝑟𝑜𝑓𝑖𝑡𝑎𝑏𝑖𝑙𝑖𝑡𝑦𝑞−4 -14.22*** -15.87***
(-4.73) (-5.25)
Past 12-month return 0.16 0.08 0.25*** 0.24***
Log(1+ME) -0.10** -0.07 -0.07** -0.06**
(-2.29) (-1.63) (-2.45) (-2.10)
B/M -0.49*** -0.31** -0.12 0.04
(-3.04) (-2.02) (-1.61) (0.59)
(1.17) (0.57) (3.67) (3.51)
Constant -0.68*** 0.25 0.34 -1.08*** -0.57** -1.37***
(-7.92) (0.63) (0.68) (-16.92) (-2.13) (-3.57)
Quarter FE N N Y N N Y
FF12 FE N N Y N N Y
R2 0.000 0.065 0.096 0.001 0.029 0.083
Observations 6124 5818 5818 5402 5108 5108
48
Table 9. Subsample analysis: by size and index constituents
This table presents subsample analyses based on firm size and S&P 500 index constituents. The sample period is
from Jan 1, 2020 to Jun 30, 2020. We restrict our sample to stocks listed on Nasdaq, NYSE and Amex. We exclude
companies that have no book value in the fiscal year ending in 2019 or no market value by the end of 2019. We also
exclude stocks whose price falls below $1 during the sample period. Panel A splits the sample based on the median
firm market value on December 31, 2019. Panel B partitions the sample by whether the stock is a S&P 500 index
constituent on December 31, 2019. The dependent variable is Fama-French 3-factor alpha. COVID Shock indicates
days on which covid-19 related news triggered S&P 500 index to move by more than 2.5%. Red is a dummy
variable that equals to 1 if the firm is headquartered in a republican county. Control variables include NPI, %Unemp,
New Cases, % Female, HH Income, TRR, log(1+ME), B/M and their interactions with COVID Shock. NPI indicates
whether there is a “shelter-in-place”, “non-essential services closure” or “closing of public venues” order in effect
on that day. % Unemp represents state-level new unemployment claim rate during that week. New Cases represent
the number of new COVID-19 cases per 1000 residents in a state on that day. % Female is the percentage of female
in the county’s total population. HH Income is the median household income in the past 12 months. TRR is the
proportion of a county’s total population that attends church. Log(1+ME) is the log of one plus the market value in
millions of dollars on December 31, 2019. BE is the book value in millions of dollars for the fiscal year ended in
2019. B/M is the corresponding book-to-market ratio. Standard errors are clustered at the date level. T-statistic is
reported in parentheses. ∗, ∗∗, ∗∗∗ indicate significance at the 10%, 5%, and 1% level, respectively.
Panel A: by size High ME (large stock) Low ME (small stock)
(1) (2) (3) (4) (5) (6)
COVID Shock 0.06 0.77 -0.44*** -6.36***
(0.57) (0.27) (-3.15) (-3.00)
Red -0.01 -0.01 -0.05 -0.02
(-0.35) (-0.59) (-1.24) (-0.53)
COVID Shock × Red 0.10 0.08 0.10 0.30*** 0.27*** 0.18**
(1.47) (1.19) (1.19) (2.93) (2.88) (2.28)
Constant 0.03 0.12 0.21 0.14*** 0.65 -0.83**
(0.86) (0.16) (0.41) (3.13) (0.65) (-2.14)
Controls N Y Y N Y Y
Firm, FF12×Date FE N N Y N N Y
R2 0.000 0.001 0.081 0.001 0.004 0.045
Observations 190435 188673 188673 190831 188690 188690
Panel B: by S&P 500
S&P 500 Non-S&P 500
(1) (2) (3) (4) (5) (6)
COVID Shock 0.09 2.88 -0.24** -3.93**
(0.93) (0.79) (-2.20) (-2.19)
Red 0.04* 0.03 -0.04 -0.02
(1.66) (1.01) (-1.22) (-0.82)
COVID Shock × Red -0.12* -0.09 -0.09 0.24*** 0.23*** 0.19**
(-1.69) (-1.27) (-1.23) (2.79) (2.76) (2.52)
Constant 0.00 -0.71 0.54 0.10*** 0.50 -0.53
(0.05) (-0.69) (0.91) (2.99) (0.62) (-1.54)
Controls N Y Y N Y Y
Firm, FF12 × Date FE N N Y N N Y
R2 0.000 0.001 0.154 0.000 0.002 0.043
Observations 54180 53625 53625 327086 323738 323738
49
Table 10. Subsample analysis: by institutional ownership and turnover ratio
This table presents subsample analyses based on institutional ownership and turnover ratio. The sample period is
from Jan 1, 2020 to Jun 30, 2020. We restrict our sample to stocks listed on Nasdaq, NYSE and Amex. We exclude
companies that have no book value in the fiscal year ending in 2019 or no market value by the end of 2019. We also
exclude stocks whose price falls below $1 during the sample period. Panel A divides the sample based on the
median institutional ownership as of 2019Q4. Institutional Ownership is calculated as the shares of stocks held by
institutional investors divided by total shares outstanding. Panel B partitions the sample based on the median
turnover on each trading day. Turnover is calculated as the daily trading volume divided by total shares outstanding.
The dependent variable is Fama-French 3-factor alpha. Control variables include NPI, % Unemp, New Cases, %
Female, HH Income, TRR, log(1+ME), B/M and their interactions with COVID Shock. NPI indicates whether there
is a “shelter-in-place”, “non-essential services closure” or “closing of public venues” order in effect on that
day. %Unemp represents state-level new unemployment claim rate during that week. New Cases represent the
number of new COVID-19 cases per 1000 residents in a state on that day. % Female is the percentage of female in
the county’s total population. HH Income is the median household income in the past 12 months. TRR is the
proportion of a county’s total population that attends church. Log(1+ME) is the logarithmic form of one plus the
market value in millions of dollars on December 31, 2019. BE is the book value in millions of dollars for the fiscal
year ended in 2019. B/M is the corresponding book-to-market ratio. Standard errors are clustered at the date level. T-
statistic is reported in parentheses. ∗, ∗∗, ∗∗∗ indicate significance at the 10%, 5%, and 1% level, respectively.
Panel A: by institutional ownership High institutional ownership Low institutional ownership
(1) (2) (3) (4) (5) (6)
COVID Shock -0.12 -2.59 -0.26** -2.18
(-0.93) (-0.90) (-2.50) (-1.24)
Red 0.03 0.04 -0.07** -0.06*
(1.02) (1.27) (-2.11) (-1.83)
COVID Shock × Red 0.10 0.12 0.12 0.28*** 0.27*** 0.19**
(1.21) (1.42) (1.37) (2.88) (2.87) (2.47)
Constant 0.06* 0.32 -0.28 0.11*** 0.20 -0.17
(1.83) (0.33) (-0.51) (3.34) (0.27) (-0.47)
Controls N Y Y N Y Y
Firm, FF12×Date FE N N Y N N Y
R2 0.000 0.001 0.061 0.001 0.003 0.042
Observations 190510 188747 188747 190756 188616 188616
Panel B: by turnover ratio
High turnover (low transaction cost) Low turnover (high transaction cost)
(1) (2) (3) (4) (5) (6)
COVID Shock -0.30* -5.79* -0.06 0.83
(-1.72) (-1.79) (-0.61) (0.55)
Red 0.01 0.03 -0.01 0.00
(0.25) (0.89) (-0.17) (0.06)
COVID Shock × Red 0.04 0.05 0.06 0.27*** 0.25*** 0.20**
(0.46) (0.66) (0.71) (2.83) (2.79) (2.53)
Constant 0.23*** 1.17 -0.93 -0.07*** -1.20** 0.19
(4.36) (0.92) (-1.43) (-2.68) (-2.31) (0.56)
Controls N Y Y N Y Y
Firm, FF12×Date FE N N Y N N Y
R2 0.001 0.004 0.074 0.000 0.003 0.050
Observations 190601 188586 188490 190665 188777 188715
50
Table 11. Subsample analysis: by income and education
This table presents subsample analyses based on income and education level. The sample period is from Jan 1, 2020
to Jun 30, 2020. We restrict our sample to stocks listed on Nasdaq, NYSE and Amex. We exclude companies that
have no book value in the fiscal year ending in 2019 or no market value by the end of 2019. We also exclude stocks
whose price falls below $1 during the sample period. Panel A divides the sample based on household income in the
past 12 months. A county is classified as High Income if its median household income is above the cross-sectional
median. Panel B partitions the sample based on the percentage of population with an education of less than a
bachelor's degree. The dependent variable is Fama-French 3-factor alpha. Control variables include NPI, % Unemp,
New Cases, % Female, HH Income, TRR, log(1+ME), B/M and their interactions with COVID Shock. NPI indicates
whether there is a “shelter-in-place”, “non-essential services closure” or “closing of public venues” order in effect
on that day. % Unemp represents state-level new unemployment claim rate during that week. New Cases represent
the number of new COVID-19 cases per 1000 residents in a state on that day. % Female is the percentage of female
in the county’s total population. HH Income is the median household income in the past 12 months. TRR is the
proportion of a county’s total population that attends church. Log(1+ME) is the logarithmic form of one plus the
market value in millions of dollars on December 31, 2019. BE is the book value in millions of dollars for the fiscal
year ended in 2019. B/M is the corresponding book-to-market ratio. Standard errors are clustered at the date level. T-
statistic is reported in parentheses. ∗, ∗∗, ∗∗∗ indicate significance at the 10%, 5%, and 1% level, respectively.
Panel A: by income High Income Low Income
(1) (2) (3) (4) (5) (6)
COVID Shock -0.20* -2.79 -0.18* -1.77
(-1.80) (-1.39) (-1.79) (-0.58)
Red 0.05 0.04 -0.05 -0.04
(1.44) (1.19) (-1.47) (-1.16)
COVID Shock × Red -0.09 -0.01 -0.02 0.29*** 0.29*** 0.25***
(-1.00) (-0.12) (-0.23) (2.92) (2.77) (2.70)
Constant 0.09*** 0.51 -0.24 0.08** -0.06 -0.29
(2.66) (0.67) (-0.66) (2.42) (-0.05) (-0.49)
Controls N Y Y N Y Y
Firm, FF12×Date FE N N Y N N Y
R2 0.000 0.002 0.041 0.000 0.002 0.056
Observations 192801 190519 190519 188465 186844 186844
Panel B: by education High Education Low Education
(1) (2) (3) (4) (5) (6)
COVID Shock -0.18* -2.01 -0.19* -1.87
(-1.83) (-0.89) (-1.82) (-0.69)
Red 0.05 0.04 -0.05 -0.05
(1.45) (1.07) (-1.58) (-1.35)
COVID Shock × Red -0.02 0.05 0.04 0.24*** 0.25*** 0.21**
(-0.15) (0.43) (0.35) (2.71) (2.70) (2.45)
Constant 0.08** 0.14 -0.06 0.09*** 0.07 -0.24
(2.44) (0.16) (-0.15) (2.74) (0.07) (-0.44)
Controls N Y Y N Y Y
Firm, FF12×Date FE N N Y N N Y
R2 0.000 0.002 0.045 0.000 0.002 0.051
Observations 183706 181497 181497 197560 195866 195866
51
Table 12. Placebo test
This table presents the partisan difference of firms’ risk-adjusted return on days with large market movements in
2018 and 2019. The sample period is from Jan 1, 2018 to Dec 31, 2019. We restrict our sample to stocks listed on
Nasdaq, NYSE and Amex. We exclude companies that have no book value in the fiscal year ending in 2017 and
2018 or no market value by the end each year. We also exclude stocks whose price falls below $1 during the sample
period. The dependent variable is daily return adjusted by Fama-French 3 factor model (3-factor α). Factor loadings
are estimated using daily returns from Jan 1, 2018 to Jun 30, 2020. Jump day indicates days on which S&P 500
index rises or falls by more than 2.5%. Red is a dummy that equals to 1 if the firm is headquartered in a republican
county. Control variables include % Unemp, % Female, HH Income, TRR, log(1+ME), B/M and their interactions
with COVID Shock. % Unemp represents state-level new unemployment claim rate during that week. % Female is
the percentage of female in the county’s total population. HH Income is the median household income in the past 12
months. TRR is the proportion of a county’s total population that attends church. We merge stock prices in year t
with the market value at the end of year t-1 and the book value of the fiscal year ending in year t-1. Log(1+ME) is
the logarithmic form of one plus the market value in millions of dollars. BE is the book value in millions of dollars.
B/M is the corresponding book-to-market ratio. We do not control for NPI and New Cases because there are no
COVID-19 cases and lockdown orders prior to the pandemic. Standard errors are clustered at the date level. T-
statistic is reported in parentheses. ∗, ∗∗, ∗∗∗ indicate significance at the 10%, 5%, and 1% level, respectively.
(1) (2) (3) (4) (5) (6)
Jump day 0.009 0.011 -1.591 -1.653*
(0.13) (0.15) (-1.62) (-1.66)
Red 0.014* 0.015**
(1.90) (2.22)
Jump day × Red -0.008 0.015 0.016 0.018 0.017
(-0.14) (0.29) (0.30) (0.33) (0.33)
% Unemp 0.009 0.007 -0.006 -0.002
(1.31) (0.56) (-0.94) (-0.28)
% Female -0.001
(-0.14)
HH Income -0.000
(-0.73)
TRR 0.000
(0.01)
Log(1+ME) 0.009** -0.300*** -0.317*** -0.320***
(2.10) (-9.58) (-10.17) (-10.65)
B/M 0.006 -0.005 0.001 0.004
(1.22) (-0.38) (0.10) (0.35)
Jump day × Unemp -0.036 -0.034 -0.029 -0.022
(-0.54) (-0.51) (-0.75) (-0.97)
Jump day × % Female 0.030* 0.032* 0.033* 0.028**
(1.83) (1.88) (1.95) (2.35)
Jump day × HH Income 0.001 0.001 0.001 0.001*
(0.96) (0.88) (0.82) (1.82)
Jump day × TRR -0.349** -0.333** -0.329** -0.228*
(-2.20) (-2.11) (-2.08) (-1.73)
Jump day × Log(1+ME) 0.032 0.032 0.032 0.028
(0.88) (0.88) (0.89) (0.75)
Jump day × B/M -0.004 -0.002 -0.005 -0.011
(-0.12) (-0.04) (-0.14) (-0.29)
Constant -0.002 -0.005 -0.042 2.055*** 2.143*** 2.166***
(-0.30) (-0.65) (-0.21) (9.61) (9.91) (10.46)
Firm FE N N N Y Y Y
Date FE N N N N Y Y
FF12 × Date FE N N N N N Y
R2 0.000 0.000 0.000 0.003 0.007 0.034
Observations 1469619 1457295 1440298 1440298 1440298 1440298
52
Table 13. Robustness
This table presents several robustness tests. The sample period is from Jan 1, 2020 to Jun 30, 2020. We restrict our sample to
stocks listed on Nasdaq, NYSE and Amex. We exclude companies that have no book value in the fiscal year ending in 2019 or no
market value by the end of 2019. We also exclude stocks whose price falls below $1 during the sample period. Panel A shows
results using alternative return benchmarks and industries. The dependent variable in column (1) – (3) is daily returns adjusted by
the CAPM model, Fama-French-Carhart four-factor model, and Fama-French five-factor model, respectively. Factor loadings are
estimated using daily returns from Jan 1, 2018 to Jun 30, 2020. Column (4) excludes firms in finance and utilities industry.
COVID Shock indicates days on which covid-19 related news triggered S&P 500 index to move up or down by more than 2.5%.
Panel B presents results with alternative thresholds of market movements. In column (1), we consider all days on which S&P 500
index moves by more than 2.5% regardless of the reason. In column (2) – (4), we consider days on which S&P 500 index moves
by more than 1%, 3% and 5%. Under these thresholds, there are 68, 26, and 10 trading days that are identified as shock days. Red
is a dummy variable that equals to 1 if the firm is headquartered in a republican county. Panel C presents results of alternative
partisanship measures. In Column (1), the measure is Red (state), a dummy variable that equals to 1 if the firm is headquartered
in a republican state. In column (2), Rep Vote (county) is the share of votes to the Republican party at the county. Control
variables, firm fixed effects and industry × date fixed effects are included in all columns of all panels. Control variables include
NPI, Unemp, New Cases, % Female, HH Income, TRR, log(1+ME), B/M and their interactions with COVID Shock. NPI indicates
whether there is a “shelter-in-place”, “non-essential services closure” or “closing of public venues” order in effect on that day.
Unemp represents state-level new unemployment claim rate during that week. New Cases represent the number of new COVID-
19 cases per 1000 residents in a state on that day. % Female is the percentage of female in the county’s total population. HH
Income is the median household income in the past 12 months. TRR is the proportion of a county’s total population that attends
church. Log(1+ME) is the logarithmic form of one plus the market value in millions of dollars on December 31, 2019. BE is the
book value in millions of dollars for the fiscal year ended in 2019. B/M is the corresponding book-to-market ratio. Standard errors
are clustered at the date level. T-statistic is reported in parentheses. ∗, ∗∗, ∗∗∗ indicate significance at the 10%, 5%, and 1% level,
respectively.
Panel A: alternative return benchmarks & industries (1) (2) (3) (4)
CAPM Fama-French-Carhart Fama-French 5 No finance & utilities
COVID Shock × Red 0.17** 0.18*** 0.15** 0.16**
(2.49) (2.63) (2.41) (2.44)
Constant -0.54 -0.28 -0.19 -1.00*
(-1.50) (-0.86) (-0.60) (-1.88)
Controls, Firm FE, FF12 × Date FE Y Y Y Y
R2 0.125 0.041 0.039 0.044
Observations 377363 377363 377363 288556
Panel B: alternative thresholds (1) (2) (3) (4)
All ≥ 2.5% ≥ 1% ≥ 3% ≥ 5%
COVID Shock × Red 0.17*** 0.09** 0.19*** 0.28**
(2.82) (2.05) (2.66) (2.15)
Constant -0.29 -0.07 -0.41 -0.15
(-0.83) (-0.13) (-1.32) (-0.61)
Controls, Firm FE, FF12 × Date FE Y Y Y Y
R2 0.043 0.043 0.044 0.044
Observations 377363 377363 377363 377363
Panel C: alternative partisanship measures (1) (2)
Rep Vote (county) Red (state)
COVID Shock × Partisan 0.0049** 0.1032*
(2.13) (1.90)
Constant -0.4445 -0.0947
(-1.06) (-0.29)
Controls, Firm FE, FF12 × Date FE Y Y
R2 0.044 0.044
Observations 377363 377363
53
Appendix A. Variable definition
Variable Definition
Panel A: variables related to daily stock performance
Raw return Daily stock returns winsorized at the top/bottom 1% on COVID shock days.
Excess return Raw returns in excess of daily risk-free rate.
FF3 α Daily excess returns adjusted by Fama-French three factors (RM−Rf, SMB, HML).
Factor loadings are estimated using daily returns from Jan 1, 2018 to Jun 30, 2020.
CAPM α Daily excess returns adjusted by CAPM. Market beta is estimated using daily
returns from Jan 1, 2018 to Jun 30, 2020.
Carhart4 α Daily excess returns adjusted by Carhart four factors (RM−Rf, SMB, HML, MOM).
Factor loadings are estimated using daily returns from Jan 1, 2018 to Jun 30, 2020.
FF5 α Daily excess returns adjusted by Fama-French five factors (RM−Rf, SMB, HML,
RMW, CMA). Factor loadings are estimated using daily returns from Jan 1, 2018
to Jun 30, 2020.
Turnover Daily trading volume divided by total shares outstanding.
Panel B: variables related to COVID-19 shocks
COVID Shock A date-level dummy variable that equals to 1 if news related to COVID-19 triggers
S&P 500 index to move up or down by more than 2.5% on the day.
Pos COVID Shock A dummy variable that equals to 1 if news related to COVID-19 triggers S&P 500
index to rise by more than 2.5% on the day.
Neg COVID Shock A dummy variable that equals to 1 if news related to COVID-19 triggers S&P 500
index to drop by more than 2.5% on the day.
Panel C: variables related to partisanship
Red A dummy that equals to 1 if a firm is headquartered in a republican county where
the Republican candidate received more votes in the 2016 Presidential Election.
Red (state) A dummy that equals to 1 if a firm is headquartered in a republican state where the
Republican candidate received more votes in the 2016 Presidential Election.
Rep Vote (county) County-level share of votes to the Republican presidential candidate in the 2016
Presidential Election.
Rep Vote (state) State-level share of votes to the Republican presidential candidate in the 2016
Presidential Election.
Social Connectedness Index The number of Facebook friend pairs between two counties divided by the product
of the two counties’ population and then scaled up by 1012. It measures the relative
probability of people in two counties being Facebook friends.
Social-Connection-based
Partisanship (SCP)
The logarithm of a weighted average of the Republican voting shares in the 2016
Presidential Election, where the weight is the Social Connected Index between the
focal county and other counties based on Facebook friendship links (excluding
own county).
54
Appendix A (continued)
Panel D: local variables
Visits The change in non-essential visits compared to the beginning of 2020.1 Raw
visits are replaced with its 5-day moving average to eliminate weekly seasonality:
∆𝑣𝑖𝑠𝑖𝑡𝑠𝑖,𝑡 =(∑ 𝑣𝑖𝑠𝑖𝑡𝑠𝑖,𝜏)/5𝑡+2
𝜏=𝑡−2
(∑ 𝑣𝑖𝑠𝑖𝑡𝑠𝑖,𝜈)/55 𝑣=1
− 1
Visitsdemean The demeaned change in visits to non-essential services. Demeaning is done by
subtracting the average change in visits from each observation on a daily basis:
𝑉𝑖𝑠𝑖𝑡𝑠𝑑𝑒𝑚𝑒𝑎𝑛 𝑖,𝑡 = 𝑉𝑖𝑠𝑖𝑡𝑠𝑖,𝑡 − 𝑉𝑖𝑠𝑖𝑡𝑠̅̅ ̅̅ ̅̅ ̅̅ ̅̅𝑡
New Cases The number of new COVID-19 cases per 1000 residents in a state on a day.
Unemp State-level new unemployment claim rate at the weekly frequency.
NPI A dummy that equals to 1 if there’s a state-level “shelter-in-place”, “non-essential
services closure” or “closing of public venues” order in effect on that day.
% Female Female as a percentage of total county population.
HH Income The median household income in thousands of US dollars in the past 12 months.
Total Religiosity Ratio (TRR) The proportion of a county’s total population that attends church. It is calculated
as the number of adherents of 217 religious denominations divided by the total
population of the county using the survey results of the Association of Religion
Data Archives (ARDA).
Panel E: variables related to firm characteristics
ME ($ million) The market value of a firm in millions of dollars on December 31, 2019.
B/M The book-to-market ratio, where BE is the book value in millions of dollars for
the fiscal year ending in 2019.
S&P 500 A dummy variable that equals to 1 if a stock is a constituent of S&P 500 index on
December 31, 2019.
Institutional Ownership The shares of stocks held by institutional investors divided by total shares
outstanding as of December 31, 2019.
Past 12-month return The cumulative stock return for the 12 months prior to the end of a quarter.
ROA Income before extraordinary items (Compustat item: IBQ) divided by total assets
(Compustat item: ATQ).
ROA The year-over-year change of ROA (𝑅𝑂𝐴𝑡 – 𝑅𝑂𝐴𝑡−4).
Profitability Returns on gross profits (revenues minus cost of goods sold) scaled by assets
(Compustat item: (REVTQ – COGSQ)/ ATQ).
Profitability The year-over-year change of profitability (𝑃𝑟𝑜𝑓𝑖𝑡𝑎𝑏𝑖𝑙𝑖𝑡𝑦𝑡 – 𝑃𝑟𝑜𝑓𝑖𝑡𝑎𝑏𝑖𝑙𝑖𝑡𝑦𝑡−4).
1 Visits are measured using smartphone location patterns. Non-essential services are defined as places whose 2-digit
NAICS code is 71 (Arts, Entertainment, and Recreation) or 72 (Accommodation and Food Services). Specifically,
they are: theaters; sport centers; museums; historical sites; zoos; amusement parks; casinos; golf courses; hotels and
inns; RV parks and campgrounds; bars; restaurants; cafeterias.
55
Appendix B. List of days with large market movements
This appendix lists all days on which S&P 500 index move up or down by more than 2.5% between January 1, 2020 and June 30, 2020. We also summarize the
explanation provided by mainstream newspapers, whether the explanation is COVID-19 related or not, and the link to related media reports.
Date Return Explanation COVID-19
related? Media Coverage
2/24/20 -3.35% Surge of cases outside China; fear of global impact of
coronavirus. Yes
WSJ: Dow Industrials Close 1,000 Points Lower as Coronavirus Cases Mount
2/25/20 -3.03% Health officials warned that coronavirus will likely spread in
the U.S. Yes
abcNEWS: Dow Jones plunges for 2nd straight day on coronavirus fears
2/27/20 -4.42%
Confirmation of the first U.S. community spread case;
growing fear that the coronavirus outbreak could cause a
recession.
Yes NYT: Coronavirus Fears Drive Stocks Down for 6th Day and Into Correction
3/2/20 4.60% Investors' hope on central banks to lower interest rates to
boost the market. Yes
WSJ: Dow Industrials Rally 5.1% on Central-Bank Stimulus Hopes
3/3/20 -2.81% FED cut rate by 50bp, signaling the US economy could be in
serious trouble because of the virus outbreak. Yes
CNNbusiness: Dow drops nearly 800 points after the Fed's surprising news about the economy
3/4/20 4.22% The strong Super Tuesday performance by Joe Biden boosted
the market. Health care stocks led the gains. No NYT: Stocks Surge as Biden Leads Super Tuesday Results
3/5/20 -3.39% California and Washington declared state of emergencies. Yes FOXbusiness: Dow falls nearly 1,000 points as coronavirus whipsaws markets
3/9/20 -7.60% Oil price plummeted amid price war between Saudi Arabia
and Russia and collapsed demand due to coronavirus. Yes
FOXbusiness: Dow plunges over 2,000 points, oil collapses amid price war and coronavirus
3/10/20 4.94% Trump propose economic relief proposals, including a payroll
tax cut and loans for small businesses. Yes
abcNEWS: Dow spikes more than 1,000 points on hopes of coronavirus economic relief
3/11/20 -4.89% World Health Organization declares global pandemic. Yes WSJ: Dow Jones Industrial Average’s 11-Year Bull Run Ends
3/12/20 -9.51% Trump declares travel ban on Europe; NBA suspended its
season; colleges suspended in-person classes. Yes WSJ: Stocks Plunge 10% in Dow’s Worst Day Since 1987
3/13/20 9.29% Trump declares national emergency; Pelosi says House will
pass a relief bill. Yes
NYT: Stocks Rally as Trump and Business Leaders Pledge Support
3/16/20 -11.98% FED cuts rate by 100bp and plans to buy $700 billion in
bonds to support economy. Yes
WSJ: The Day Coronavirus Nearly Broke the Financial Markets
56
(Appendix B continued)
Date Return Explanation COVID-19
related? Media Coverage
3/17/20 6.00% Fed plans to launch commercial paper funding facility;
Trump seeks $1 trillion stimulus to fight covid. Yes WSJ: Stocks Rise Sharply in Volatile Trading
3/18/20 -5.18% Trump announced U.S. and Canada close border to non-
essential traffic. Yes
CBSNews: Dow closes below 20,000, wiping out nearly all the gains of Trump's presidency
3/20/20 -4.34%
Fresh measures to contain the coronavirus pandemic spooked
investors; Andrew Cuomo ordered the New York state’s
workforce to stay home.
Yes WSJ: Stocks Wrap Up Tough Week With Another Fall
3/23/20 -2.93%
The Senate failed for a second time to vote through the
coronavirus economic relief package; FED to extend loans
and purchase government debt.
Yes WSJ: U.S. Stocks Drop Despite Fed’s Latest Stimulus Move
3/24/20 9.38% Congress and the White House near a deal of the stimulus
package after late-night negotiations. Yes
WSJ: Dow Soars More Than 11% In Biggest One-Day Jump Since 1933
3/26/20 6.24% Senate approves $2.2 trillion coronavirus bill. Yes WSJ: Dow Rallies 6.4% After Stimulus Vote
3/27/20 -3.37% Stocks pulled back after a furious three-day rally. No WSJ: Stocks Drop, But Finish the Week With Big Gains
3/30/20 3.35%
Trump extended the timeline for social distancing guidelines
to April 30, which many believe will reduce economic
damage in the long run.
Yes CNBC: Dow drops 400 points as stocks close out their worst first quarter ever
4/1/20 -4.41%
The White House warned that the U.S. could face as many as
240,000 deaths; Trump asking Americans to brace for an
unprecedented crisis in the days ahead.
Yes FOXbusiness: Stocks stumble as US coronavirus cases top
200,000
4/6/20 7.03% Slower case growth in NY and lower death rate in Europe
indicate progress in the fight against coronavirus. Yes
WSJ: Dow Industrials Surge About 1,600 Points at Start of Challenging Week
4/8/20 3.41% Hospitalizations and intensive care admissions slow down in
NY; declining new coronavirus infections in Italy. Yes WSJ: Stocks Close Higher After Bout of Volatility
4/14/20 3.06% Concerns over the coronavirus ease and early-stage plans of
re-opening some pockets of the economy take shape. Yes
FOXbusiness: Dow jumps 558+ points, Nasdaq exits bear market as coronavirus concerns ease
4/17/20 2.68%
Trump told governors they could begin reopening businesses;
Boeing planned to bring 27k employees back to work; Gilead
Sciences made breakthrough in Remdesivir.
Yes NYT: Stocks Rally After Talk of Reopening Economy
57
(Appendix B continued)
Date Return Explanation COVID-19
related? Media Coverage
4/21/20 -3.07% Oil price plummeted as a historic selloff pushed West Texas
Intermediate crude oil May contract to negative. No
FOXbusiness: Dow tumbles 631 points as oil selloff deepens
4/29/20 2.66%
FED promised to use more tools to aid the economic
recovery; Gilead Sciences got positive results evaluating
remdesivir on coronavirus patients.
Yes FOXbusiness: Stocks surge on Gilead’s coronavirus drug and Fed’s pledge to keep rates near zero
5/1/20 -2.81% Tech giants' revenues fall below expectation. No WSJ: Tech Giants Pull Stocks Lower as Dow Falls More Than 600 Points
5/18/20 3.15%
Drugmaker Moderna announced progress toward a COVID-
19 vaccine; FED pledge for further stimulus; lockdowns
continued to ease nationwide.
Yes FOXbusiness: Dow surges over 900 points amid coronavirus vaccine progress as lockdowns ease
6/5/20 2.62% The unemployment rate unexpectedly fell to 13%, compared
with estimates of 20%. No
WSJ: Stocks Close Sharply Higher on Surprisingly Upbeat Jobs Report
6/11/20 -5.89% COVID-19 infections resurge as more states reopened; FED
warned a slower economic recovery. Yes
WSJ: U.S. Stocks End Sharply Lower as Coronavirus Worries Return
6/24/20 -2.59% New coronavirus cases have surged in several states. Yes WSJ: Stocks Fall as Coronavirus Infections Surge