Title: Politics Gets Personal: Effects of Political Partisanship and Advertising
on Family Ties
Authors: M. Keith Chen1*†, Ryne Rohla2*
Affiliations:
1

Anderson School, University of California, Los Angeles, CA 90095, USA.

2

School of Economic Sciences, Washington State University, Pullman, WA 99164, USA.

*Both authors contributed equally to this work.
†

Corresponding author. Email: keith.chen@anderson.ucla.edu.

Abstract: Research on growing American political polarization and antipathy primarily studies
effects on public institutions and political processes, ignoring private effects such as damaged
family ties. Using smartphone-tracking data and precinct-level voting, we show that politicallydivided families shortened Thanksgiving dinners by 20-30 minutes following the divisive 2016
election. This decline survives comparisons with 2015 and extensive demographic and spatial
controls, and more than doubles in media markets with heavy political advertising. These effects
appear asymmetric: while Democratic voters traveled less in 2016, political differences shortened
Thanksgiving dinners more among Republican voters, especially where political advertising was
heaviest. Partisan polarization may degrade close family ties with large aggregate implications;
we estimate 27 million person-hours of cross-partisan Thanksgiving discourse were lost in 2016
to ad-fueled partisan effects.
One Sentence Summary: Cell-tracking shows that mixed-party families had shorter 2016
Thanksgivings, an effect exacerbated by political advertising.
Introduction: American partisan antipathy has risen sharply over the past twenty-five years. Over
55% of Democrats and Republicans described “very unfavorable” feelings toward the opposing
party in 2016, up from 17-21% in the mid-1990s (1). Spatial sorting along partisan lines produces
increasingly homogeneous electoral “bubbles” at both state and local levels (2, 3). Political
minorities within these bubbles become more reticent to participate in or reveal their party as
polarization increases (4).
Animosity toward political rivals is not limited to the ballot box; implicit partisan biases manifest
in discriminatory decisions at a rate higher than racial or gender biases (5). In surveys, parents
have become less tolerant of their children dating and marrying across partisan lines (6), and
observed dating and marital behavior segregates on politics more strongly than on physical
attributes or personality characteristics (7). Political polarization impacts economic decisions in
the public sphere, including where to work and shop, at higher larger than those caused by race,
ethnicity, or religion (8).
We study whether politics strains close family ties by measuring family gathering durations. After
the historically divisive and stressful (9) 2016 presidential election, 39% of American families

 avoided political conversations during the holidays (10). Aversion to family political discussions
largely spans both party and socioeconomic lines (10). In this context we study Thanksgiving,
which in US election years, brings together family members with differing political views at a time
of partisan salience. Anecdotal evidence suggests that many altered or canceled Thanksgiving
plans in the wake of the 2016 election (11). Other families cut short their dinners if political
arguments arose. Most political-personal studies rely on survey data, and lack the ability to broadly
measure inflamed partisan antipathy and its effect on real-world behaviors such as spending time
with friends and family.
This study analyzes how political differences affected 2016 Thanksgiving dinners through the
merging of two novel datasets. A unique collection of smartphone location-tracking data from
more than ten million Americans allows observation of actual (not self-reported) movement
behavior, at extremely precise spatial and temporal levels. We combined this with a database of
the national precinct returns for the 2016 presidential election to impute individual political
leanings at the finest spatial resolution legally possible. By comparing the partisan bend of an
individual’s home location and where they attended Thanksgiving dinner, we can test the
relationship between political disagreement and time expenditure.
To further isolate the particular effect of election-year political partisanship, we compare time
spent at Thanksgiving in 2016 with the Thanksgiving of the year before. That comparison suggests
that our measured effect really is the result of heightened political rhetoric, and not an artifact of
politically-correlated demographic or spatial sorting. Finally, since political advertising polarizes
opinions (12) and media coverage of polarization heightens dislike for opposing parties (13), we
compare partisan rifts between families who live within a few miles of each other but on opposite
sides of media-market boundaries, and find political advertising more than doubles our measured
Thanksgiving effects.
Data Collection and Validation: We collected precinct-level results for the 2016 presidential
election through internet scraping and by contacting Secretaries of State, Boards of Election, and
other statewide and county-wide electoral authorities. For states that do not centrally compile
precinct-level election results, we contacted individual County Clerks by email, phone, fax, or in
person. The resulting dataset covers over 172,000 precincts in 99% of counties and over 95% of
all votes cast in the race. Finally, we matched these vote totals to precinct polygonal shapefiles
using Geographic Information Systems (GIS) software.
Our data on political advertising come from Kantar Media’s Campaign Media Analysis Group as
reported by the Center for Public Integrity (14). These data count every US presidential-level
general election television ad aired in each of the 210 Nielsen Designated Market Areas after June
12, 2016, including ads purchased either directly by campaigns or by outside groups such as
Political Action Committees. Additional demographic data from the Census Bureau’s 2012-2015
American Community Survey formed the basis of other control variables.
Location tracking data comes from Safegraph, a company that aggregates location information
from numerous smartphone apps. The data consist of “pings”, each of which identify the location
(latitude and longitude) of a particular smartphone at a moment in time. Safegraph tracks the
location of more than 10 million Americans’ smartphones, and our core analysis focusses on the

 more than 17 trillion pings Safegraph collected in the continental United States in November of
2016.
To merge these data, we first form a proxy for each smartphone owner’s “home”, which both links
that person to a voting precinct and determines whether a person ate Thanksgiving dinner away
from home. Home locations are determined by looking at where each person in our sample is most
frequently between 1 and 4am. If a person has a consistent early morning location over the three
weeks before Thanksgiving, we use this location as a simple proxy for their “home”. This
procedure identifies the home location of over 5 million people in the November Safegraph
sample, and we link these locations with their corresponding voting precinct, two-party vote share,
and census demographics using GIS software.
An individual’s location for Thanksgiving is constructed based on their modal location between 1
and 5pm that day. If this location differs from the individual’s home, we assume that they traveled
for Thanksgiving, and measure the amount of time that they spent at their Thanksgiving location
for our duration analysis.
By construction, this sample is only representative of the estimated 77% of Americans who own
smartphones, which opens the question of whether our sample is politically representative of the
American electorate as a whole. We test this by assigning to each person in our sample a voting
ratio proportional to the 2016 two-party vote share of their home precinct. If a person lives in a
precinct which recorded 150 Clinton and 50 Trump votes, for example, that person would be
assigned 0.75 Clinton and 0.25 Trump votes. Figure 1(B) compares these votes against actual 2016
two-party vote shares for each state. The degree to which these points track the 45 degree line
represents a joint test of the political representativeness of the Safegraph data, and the unbiasedness
of our home-to-precinct likely-vote imputation. Our imputed votes are accurate to within 1 percent
in 33 states and within 5 percent in all states. At a national level the data add up to a Democratic
vote share of 50.3%, compared to the actual share of 51.1%.
Empirical Analysis: Our first analysis examines whether conditional on traveling for
Thanksgiving dinner, the partisan distance between the traveling and hosting households affects
the dinner’s duration. For this we restrict our sample to households who were home both in the
morning and the night of Thanksgiving, but who traveled for Thanksgiving dinner. This restricts
our analysis to families who were likely able to control the duration of their visits. In Table 1 we
estimate the following equation:
durationij = α + β mismatchij + γ Fij + εij
where
mismatchij = Pi (1 – Pj) + (1 – Pi) Pj
In this specification, durationij is the minutes traveler i spent with host j on Thanksgiving, and Fij
is a set of fixed-effects that form groups of people for comparison defined by the pair of home /
travel locations for persons i and j. Pi and Pj are the two-party vote shares of the precincts
associated with home precincts for i and j, Pi = (demi / (demi + repi)). Using Pi and Pj as imputed
vote probabilities, mismatchij is the imputed probability that persons i and j voted for different
candidates in 2016. In all tables, regressions running from left to right control for progressively

 finer (i,j) location pairs culminating in 5-digit geohash boxes, a global grid of rectangular areas
roughly 1.5 miles per side.
To control for likely confounds such as demographics, distance, and travel time, our regressions
compare Thanksgiving durations only between families that live and eat in the same pair of
locations. For example, regression 3 compares two families if and only if they both live in zipcode
X and visit zipcode Y, or vice versa. β is the coefficient of interest, measuring the difference in
Thanksgiving durations between families that fall within the same Fij comparison groups, but some
of whose votes matched and some of whose votes mismatched. Standard errors are clustered at the
level of these comparison groups. We choose to use progressively tighter spatial controls to control
for both demographics and travel distance simultaneously, since individual characteristics of
people in our sample are inferred from spatial characteristics such as residential location. In online
tables (available upon request) we report statistically identical results when demographics such as
race, age, education, income, and employment are controlled for separately.
Results in Table 1 indicate that despite fixed effects, which aggressively control for both travel
distance and location-correlated demographics, families that were likely to have voted differently
spent between 20 and 30 fewer minutes with each other; with Thanksgiving dinner averaging 4.1
hours. Tellingly, as we add finer spatial controls, our estimate of β increases in magnitude,
suggesting that the measured decline in dinner duration is, if anything, an underestimate of the true
effect of partisan mismatch.
Table 2 separates the effects of the two components of mismatchij, [Pi (1 – Pj) and (1 – Pi) Pj]. This
can be interpreted as separately measuring the effect of voting disagreement on Democratic voters
visiting Republican households, and vice versa. Intuitively, this compares families within the same
Fij pair, and asks if whether voting disagreement asymmetrically affects liberal travelers going
from i to j, or more conservative travelers going from j to i. Table 2 suggests that conditional on
traveling, travelers from Democratic precincts do not significantly shorten their visits to
Republican hosts, while Republican-precinct travelers shortened their visits by over 40 minutes.
Table 3 investigates whether these effects interact with local political advertising, and finds that,
on average, Thanksgiving dinners are further shortened by around 1.5 minutes for every thousand
political advertisements aired in the traveler’s home media market. In swing states such as Florida,
media markets saw over 25 thousand ads over the course of the campaign, implying a 1.2 hours
shorter Thanksgiving for vote-mismatched families.
Consistent with our finding that advertising appears to heighten the partisan effect of the 2016
election on Thanksgivings, Table S1 gives the same analysis and regressions a year prior and finds
no Thanksgiving-shortening effect, presumably at a time when political antipathy was less salient
and polarizing. Continuing this comparison of 2015 and 2016, Figure S1 examines whether
conditional on choosing to travel, families appear to have biased their choice of destinations to
lower the political divide between guest and host, between 2015 and 2016. Figure S1 demonstrates
that for families that traveled in both 2015 and 2016 —the strongest possible control for
demographic and spatial confounds— there is no appreciable difference in the distribution of
likelihood of political mismatch.

 Finally, tables S2 and S3 estimate linear probability models for the choice whether to travel for
Thanksgiving dinner, in both 2015 and 2016. When comparing matched families living within a
mile and a half of each other, Democratic voters reduced their likelihood of travel between 2015
and 2016 by 5 percentage points more than comparable Republican voters, an effect that increases
significantly in areas with heavy political advertising.
Following the 2016 election, anecdotal media reports (11) and online social media behavior (15)
demonstrated an avoidance of political confrontations among Democratic voters, findings our
study corroborate. Republicans, however, were more sensitive to partisan differences at
Thanksgiving dinners. Aggregating across the 77% of American adults who own smartphones
(16), our results suggest partisan differences cost American families 62 million person-hours of
Thanksgiving time, 56.8% from individuals living in Democratic precincts and 43.2% from
Republican precincts. Political advertising eliminated an additional 3.3 million person-hours,
52.8% from Democratic precinct residents and 47.2% from Republican residents. We estimate 27
million person-hours of cross-partisan discourse were eliminated, which may provide a feedback
channel by which partisan antipathy reduces opportunities for close cross-party conversations.

References and Notes
1. Pew Research Center, “Partisanship and political animosity in 2016.” (June 22, 2016);
people-press.org/2016/06/22/partisanship-and-political-animosity-in-2016.
2. D. Wasserman, “Purple America has all but disappeared.” (FiveThirtyEight, March 8,
2017); fivethirtyeight.com/features/purple-america-has-all-but-disappeared.
3. A. Agadjanian, “The decline of competitive precincts from 2008 to 2016.” (Decision
Desk HQ, March 30, 2017); decisiondeskhq.com/data-dives/the-decline-of-competitiveprecincts-from-2008-to-2016.
4. R. Perez-Truglia, G. Cruces, Partisan interactions: evidence from a field experiment in
the United States. J. of Pol. Econ. 125, 1208-1243 (2017).
5. S. Iyengar, S. Westwood, Fear and loathing across party lines: new evidence on group
polarization. Am. J. of Pol. Sci. 59, 690-707 (2015).
6. S. Iyengar, G. Sood, Y. Lelkes, Affect, not ideology: a social identity perspective on
polarization. Pub. Opin. Quart. 76, 405-431 (2012).
7. J. Alford, P. Hatemi, J. Hibbing, N. Martin, L. Eaves. The politics of mate choice. J. of
Pol. 73, 362-379 (2011).
8. C. McConnell, Y. Margalit, N. Malhotra, M. Levendusky, The economic consequences
of partisanship in a polarized era. Am. J. of Pol. Sci., in press, (available at
web.stanford.edu/~neilm/Economic_Consequences_Final_Identified.pdf).
9. American Psychological Association, “Stress in America: U.S. presidential election
2016.” (October 13, 2016); apa.org/news/press/releases/stress/2016/presidentialelection.pdf
10. Pew Research Center. “How Americans are talking about Trump’s election in 6 charts.”
December 22, 2016. pewresearch.org/fact-tank/2016/12/22/how-americans-are-talkingabout-trumps-election-in-6-charts.

 11. S. Tavernise, K. Seelye, “Political divide splits relationships—and Thanksgiving, too.”
(The New York Times, November 15, 2016); nytimes.com/2016/11/16/us/politicaldivide-splits-relationships-and-thanksgiving-too.html
12. S. Ansolabenhere, S. Iyengar, Going Negative: How Political Advertisements Shrink &
Polarize the Electorate (Free Press, 1996)
13. M. Levendusky, N. Malhotra, Does media coverage of partisan polarization affect
political attitudes? Pol. Comm. 33, 283-301 (2016).
14. C. Zubak-Skees, “Tracking TV ads in the 2016 presidential race.” (The Center for Public
Integrity, October 25, 2016); publicintegrity.org/2016/01/21/19164/tracking-tv-ads-2016presidential-race
15. D. Cox, R. Jones, “’Merry Christmas’ vs. ‘Happy Holidays’: Republicans and Democrats
are polar opposites.” (Public Religion Research Institute, December 19, 2016);
prri.org/research/poll-post-election-holiday-war-christmas
16. L. Rainie, A. Perrin, “10 facts about smartphones as the iPhone turns 10.” (Pew Research
Center, June 28, 2017); pewresearch.org/fact-tank/2017/06/28/10-facts-aboutsmartphones

Acknowledgments: The authors would like to thank seminar participants at UCLA,
Northwestern, Simon Fraser, and Stanford Universities for helpful comments, and A. Hoffman,
R. Squire and N. Yonack at Safegraph for data access and technical assistance. All data and code
are available from the corresponding author by request. Neither author has any competing
interests or funding sources related to this study.
List of Supplementary Materials:
Figure S1
Tables S1-S3

 B
1

DC

0.9

Sampled Clinton Vote Share

Fig. 1. Sampling and imputation validation.
(A) Shows the results of the 2016 US
presidential election, by precinct. (B) Jointly
tests the representativeness of our celltracking data and the success of our simple
vote-imputation process using voter precinct
matching. For each state, the x-value is the
actual two-party vote share in the 2016
election, and the y-value is the vote share
obtained by assigning to each tracked
smartphone the vote share of its imputed
precinct. This produces a predicted national
share of 0.503 for Clinton, while the actual vote
share was 0.511.

0.8
0.7

California
Hawaii

0.6

Nevada

0.5

New Mexico

0.4

Alaska
Mississippi
Alabama

W. Virginia

0.3

Wyoming

0.2
0.1
0
0

0.1

0.2

0.3

0.4

0.5

0.6

0.7

0.8

2016 Clinton Vote Share

0.9

1

 Variable:
Prob. of political
mismatch
Observations:
R-squared:
Fixed-Effects:
Num. of Groups:

1
Thanksgiving
Duration (min)
-13.02***
(1.990)

2
Thanksgiving
Duration (min)
-20.39***
(3.643)

3
Thanksgiving
Duration (min)
-22.18*
(8.657)

4
Thanksgiving
Duration (min)
-26.94♦
(15.35)

583,071
0.0001
none

583,071
0.0552
county pairs
23,375

583,071
0.436
zip-code pairs
245,004

583,071
0.647
geohash-5 pairs
356,479

Table 1. Effect of political mismatch on Thanksgiving duration. Each column is an estimate of the
effect of voting disagreement on the length of Thanksgiving dinner in 2016. All regressions are fixedeffect linear regressions, where fixed effects control for the pair of locations where an individual lives
and ate Thanksgiving dinner. Regressions running from left to right control for progressively finer
pairs of areas, culminating in 5-digit geohash boxes, a grid of boxes roughly 1.5 miles per side. For
example, regression three compares families that live and ate in the same pair of zip codes, and can
be interpreted as saying that within these comparison groups, opposite-voting families had 22minute shorter Thanksgiving dinners. The number of comparison groups these fixed-effects entail is
listed for each regression. The mean duration of Thanksgiving dinner was 245 minutes, and the
average probability of opposite-voting political mismatch was 0.44 with a SD of 0.10. Standard errors
are clustered at the level of the fixed-effect and reported in parentheses, with significance levels: ***
p<0.001, ** p<0.01, * p<0.05, ♦ p<0.1.

 Variable:
Prob. D → R
trvlr → host
Prob. R → D
trvlr → host
Observations:
R-squared:
Fixed-Effects:
Num. of Groups:

1
Thanksgiving
Duration (min)
2.882
(2.290)
-29.63***
(2.316)

2
Thanksgiving
Duration (min)
-5.024
(3.807)
-36.27***
(3.963)

3
Thanksgiving
Duration (min)
-3.097
(9.067)
-41.04***
(9.050)

4
Thanksgiving
Duration (min)
-12.03
(15.91)
-43.95**
(16.34)

583,071
0.0004
none

583,071
0.0555
county pairs
23,375

583,071
0.436
zip-code pairs
245,004

583,071
0.647
geohash-5 pairs
356,479

Table 2. Asymmetric effects of political mismatch. Each column is an estimate of the effect of
voting disagreement between travelers and hosts (Dem to Rep and Rep to Dem) on the length of
Thanksgiving dinner in 2016. All regressions are fixed-effect linear regressions, where fixed effects
control for the pair of locations where an individual lives and ate Thanksgiving dinner. For example,
regression three compares families that live and ate in the same pair of zip codes, and can be
interpreted as saying that within these comparison groups, Republican-voting travelers leave
Thanksgiving dinner over 40 minutes sooner if their hosts voted Democratic, while Democrat-voting
travelers did not significantly shorten their stays at Republican-hosted dinners. The number of
comparison groups these fixed-effects entail is listed for each regression. The mean duration of
Thanksgiving dinner in 2016 was 245 minutes (4.1 hours), and the average probabilities of Democratvoting individuals eating at a Republican-hosted dinner (and vice-versa) were 0.221 and 0.216,
respectively, both with SD = 0.10. Standard errors are clustered at the level of the fixed-effect and
reported in parentheses, with significance levels: *** p<0.001, ** p<0.01, * p<0.05, ♦ p<0.1.

 Variable:
Prob. D → R
trvlr → host
Prob. R → D
trvlr → host
Political ads
(1K ads / mrkt)
Prob. D → R
× Pol. ads
Prob. R → D
× Pol. ads
Observations:
R-squared:
Fixed-Effects:
Num. of Groups:

1
Thanksgiving
Duration (min)
2.882
(2.290)
-29.63***
(2.316)

2
Thanksgiving
Duration (min)
-5.024
(3.807)
-36.27***
(3.963)

3
Thanksgiving
Duration (min)
13.53***
(2.643)
-17.76***
(2.658)
0.734***
(0.171)
-2.386***
(0.403)
-2.825***
(0.412)

4
Thanksgiving
Duration (min)
-1.821
(4.671)
-30.97***
(4.756)
0.136
(0.319)
-0.931♦
(0.562)
-1.491*
(0.632)

583,071
0.001
none

583,071
0.0555
county pairs
23,375

583,071
0.001
none

583,071
0.0556
county pairs
23,375

Table 3. Political advertising heightens partisan effects. Each column is an estimate of the effect
of voting disagreement between travelers and hosts (Dem to Rep and Rep to Dem) on the length of
Thanksgiving dinner in 2016, and whether political advertising heightens these effects. All regressions
are linear regressions, where fixed effects control for the pair of locations where an individual lives
and ate Thanksgiving dinner. Media markets in swing states like Florida saw over 25 thousand ads
in 2016; in regression four this implies a doubling of the main 31-minute effect, and resulting in a
Thanksgiving dinner of over an hour shorter for Republicans eating with Democratic hosts. The mean
duration of Thanksgiving in 2016 was 245 minutes (4.1 hours), and the average probabilities of
Democrat-voting individuals eating at a Republican-hosted dinner (and vice-versa) were 0.221 and
0.216, respectively, both with SD = 0.10. Standard errors are clustered at the level of the fixed-effect
and reported in parentheses, with significance levels: *** p<0.001, ** p<0.01, * p<0.05, ♦ p<0.1.

 Supplementary Materials:

Fig. S1. The (non)effect of partisanship on Thanksgiving destination choice. This compares the
distributions of voting mismatch for Thanksgivings 2015 and 2016 for individuals who traveled in
both years. (A) shows the probability distribution that a person who traveled in both years voted
differently from their Thanksgiving hosts in 2015 and 2016. (B) and (C) show the two ways mismatch
can occur; a Democratic traveler eating with a Republican host (B), or vice versa (C). T-tests confirm
that conditional on traveling for Thanksgiving dinner, the partisan difference of travelers and hosts
did not change significantly between 2015 and 2016.

 Variable:
Prob. of political
mismatch
Observations:
R-squared:
Fixed-Effects:
Num. of Groups:

1
Thanksgiving
Duration (min)
38.05***
(8.178)

2
Thanksgiving
Duration (min)
16.05
(12.02)

3
Thanksgiving
Duration (min)
-66.20
(111.3)

4
Thanksgiving
Duration (min)
-61.86
(182.5)

25,969
0.001
none

25,969
0.163
county pairs
4,578

25,969
0.841
zip-code pairs
22,301

25,969
0.915
geohash-5 pairs
24,030

Table S1. The (non)effect of partisanship on Thanksgiving 2015. Each column is an estimate of
the effect of opposite-party voting on the duration of Thanksgiving dinner as measured in celltracking data in 2015. All regressions are fixed-effect linear regressions, where fixed effects control
for the pair of locations where an individual lives and ate Thanksgiving dinner. Regressions running
from left to right control for progressively finer pairs of areas, culminating in 5-digit geohash boxes,
a grid of boxes roughly 1.5 miles per side. While the sample of tracked smartphones is much smaller
in 2015, these regressions suggest that unlike the post-election Thanksgiving of 2016, there was no
systematic relationship between political mismatch and the duration of Thanksgiving dinner in 2015.
The mean duration of Thanksgiving dinner in 2015 was 200 minutes, and the average probability of
opposite-voting political mismatch was 0.42 with a SD of 0.11. Standard errors are clustered at the
level of the fixed-effect and reported in parentheses, with significance levels: *** p<0.001, ** p<0.01,
* p<0.05, ♦ p<0.1.

Variable:
Post election
(year = 2016)
Home 2-party
vote (Dem)
PE × H2PV
Observations:
R-squared:
Fixed-Effects:
Num. of Groups:

1
Thanksgiving
not at home
0.146***
(0.00354)
-0.000771
(0.00580)
-0.121***
(0.00599)

2
Thanksgiving
not at home
0.145***
(0.00431)
0.0287**
(0.00933)
-0.116***
(0.00891)

3
Thanksgiving
not at home
0.151***
(0.00345)
0.0695***
(0.00703)
-0.122***
(0.00585)

4
Thanksgiving
not at home
0.156***
(0.00374)
0.0586***
(0.00781)
-0.128***
(0.00633)

2,025,855
0.005
none

2,025,855
0.0114
county
1,838

2,025,855
0.0335
zip code
30,237

2,025,855
0.0886
5-digit geohash
116,901

Table S2. Partisanship and Thanksgiving travel: 2015 & 2016. Each column is an estimate of the
effect of political leanings on the choice to eat Thanksgiving dinner at home or away, and how this
effect differs between the pre-election 2015 and post-election 2016 Thanksgivings. All regressions
are fixed-effect linear probability regressions, where fixed effects control for the location of a person’s
home. Standard errors are clustered at the level of the fixed-effect and reported in parentheses, with
significance levels: *** p<0.001, ** p<0.01, * p<0.05, ♦ p<0.1.

 Variable:
Post election
(year = 2016)
Home 2-party
vote (Dem)
PE × H2PV
Political ads
(1K ads / mrkt)
PE × H2PV
× Pol. ads
H2PV
× Pol. ads
PE × Pol. ads
Observations:
R-squared:
Fixed-Effects:
Num. of Groups:

1
Thanksgiving
not at home
0.139***
(0.00417)
-0.0113♦
(0.00683)
-0.109***
(0.00705)
-0.00178***
(0.000529)
-0.00303**
(0.000963)
0.00266**
(0.000927)
0.00170**
(0.000545)

2
Thanksgiving
not at home
0.139***
(0.00505)
0.0198♦
(0.0111)
-0.104***
(0.0106)
-0.00180**
(0.000641)
-0.00295**
(0.00114)
0.00231*
(0.00104)
0.00164**
(0.000582)

3
Thanksgiving
not at home
0.145***
(0.00410)
0.0606***
(0.00833)
-0.111***
(0.00695)
-1.65e-05
(0.00116)
-0.00265**
(0.000954)
0.00223*
(0.00108)
0.00152**
(0.000525)

4
Thanksgiving
not at home
0.150***
(0.00446)
0.0499***
(0.00930)
-0.118***
(0.00753)
-0.000514
(0.00162)
-0.00233*
(0.00104)
0.00213♦
(0.00122)
0.00135*
(0.000574)

2,025,855
0.005
none

2,025,855
0.0114
county
1,838

2,025,855
0.0335
zip code
30,237

2,025,855
0.0886
5-digit geohash
116,901

Table S3. Political advertising and Thanksgiving travel: 2015 & 2016. Each column is an estimate
of the effect of political leanings on the choice to eat Thanksgiving dinner at home or away, and how
this effect differs between the pre-election 2015 and post-election 2016 Thanksgivings, and between
areas that saw more or less political advertising in 2016. All regressions are fixed-effect linear
probability regressions, where fixed effects control for the location of a person’s home. For example,
regression three can be interpreted as saying that in 2016, Democratic households significantly
reduced their propensity for Thanksgiving travel relative to Republicans living in their same zipcode,
and that this effect was much more pronounced for Democratic households that were exposed to
large amounts of political advertising. Standard errors are clustered at the level of the fixed-effect
and reported in parentheses, with significance levels: *** p<0.001, ** p<0.01, * p<0.05, ♦ p<0.1.