Research Replication:
“Does movie violence increase violent crime?”
Purpose
Motivations:
● A 2000 report by the American Academy of Pediatrics, American Psychological
Association, and other medical groups raised concerns about violent media’s
influence on youth behavior.
● The report explicitly acknowledged that it did not establish a causal link between
exposure to media violence and actual violent crime.
● The gap in evidence motivated the authors to explore whether short-term
exposure to violent films correlates with real-world violent outcomes.
Source of Original Data and Methodology:
● Panel Data
● County and Date fixed effects (limited variation in crime around movie releases)
● The paper exploits movie attendance that varies daily and the fact that more
people go to theaters on weekends, especially to see violent films.
● The National Incident-Based Reporting System (NIBRS) is a program sponsored
by the FBI, reporting daily assault incidents across 573 cities from 1995 to 2004.
(Crime Data)
● National box office data from Variety and Nielsen EDI, merged with content
classifications (violent vs. nonviolent) (Movie Data)
● U.S county population estimates, day-of-week, holidays, and weather data
(Demographic and environmental data)
● This data allows the author to examine within a U.S county changes in assault
rates on days with higher exposure to violent content.
General Findings:
● Violent crime decreases on days when more people attend violent movies
● Assaults drop about 1.1 ~ 1.3% on days with high violent movie attendance
● The drop is largest during hours when people are most likely to be in theaters in
the evenings.
● There is some evidence of a small increase in assaults the next day, but this does
not fully offset the same-day reduction.
● The authors suggest that people being occupied in theaters and reduced alcohol
consumption are key mechanisms.
● On days with high attendance of violent movies, same-day assaults decrease,
particularly between 6 p.m. and midnight. The authors attribute this to an
incapacitation effect; people watching movies are not out in places where assaults
might occur.
Replication Goal:
● We aim to replicate the finding that violent movie attendance is negatively
correlated with violent crime.
● Using monthly and yearly exposure to violent movies is associated with a
reduction in violent crime, particularly assault rates.
Replication Data and Methods
➢ talk about original data: how many subjects, observations, where data is coming from
➢ talk about the variables used and what they did
➢ check and see if paper gives a summary statistics of their dataset and input table
➢ then talk about our dataset: describes the more general of counties, attendance by
millions, how many observations
➢ input table of our summary statistics
➢ explicitly compare summary statistics from OG data with replication data: can compare
observations, means, proportions
Original Data Description
● The original dataset in Dahl and DellaVigna (2009) contains data at the monthly and
yearly level for the U.S counties.
● The authors use data on all assaults reported to the police from the National
Incident-Based Reporting System (NIBRS), a program sponsored by the FBI.
● The main independent variable is movie attendance by county, derived from national box
office data merged with county-level population data and estimates of movie attendance
rates. (Such as weather, day-of-week, holidays, year)
● Variables:
○ Assaults
○ Movie attendance (violent, mild, normal)
○ Month, year (fixed effects)
○ Weather controls (rain, snow)
● Dataset Overview:
○ 1,563 weekend observations (Friday ~ Sunday) from January 1995 to December
2004.
○ 2,272,999 total assaults
○ 1,781 reporting agencies
● Assault statistics
○ The average of 1,454 assaults per weekend day.
○ Assaults are most frequent:
■ Evening (6 P.M. ~ 12 A.M)
■ Also common inthe afternoon (12 P.M. ~ 6 P.M.) and night (12 A.M ~ 6
A.M)
● Highest on Fridays and Saturdays, lower on Sundays and weekdays
● Assaults:
○ Three times more common among males than females.
○ Decrease with offender age (above 18)
○ 17% involve suspected alcohol or drug use, especially during night hours.
● Movie attendance:
○ Average weekend day audience: 6.29 million
○ Peak attendance on Saturdays.
○ Audience by movie type:
■ Strongly violent movies: 0.87 million
■ Mildly violent movies: 2.43 million
● Additional Data:
○ Summary statistics also include VHS and DVD rentals
variable
Original dataset value
Replication dataset value
assaults
2,272,999 (this is the total)
8,719 (this is the total)
4,352.66 (this is the mean)
assaults for weekend
1,454
N/A
assaults for weekday
1,293
N/A
assaults for Friday
1,589
N/A
assaults for Saturday
1,564
N/A
Assaults for Sunday
1,209
N/A
Violent movie attendance
0.87 million (this is the avg,
but it is for the weekends
only)
2.59 million (this is the mean)
Mildly violent movie
attendance
2.43 million (this is the avg,
but it is for the weekends
only)
4.28 million (this is the mean)
Nonviolent movie attendance
2.99 million (this is the avg,
but it is for the weekends
only)
8.95 million (this is the mean)
Average weekend attendance
6.29 million
N/A
Average weekday attendance
2.00 million
N/A
Average Friday attendance
5.74 million
N/A
Average Saturday attendance
7.90 million
N/A
Average Sunday attendance
5.24 million
N/A
–If we want to make the numbers match the same generalization as the replication, we can
maybe make a footnote saying we calculated the values ourselves?
Replication Results
➢ we first want to talk about a single variable regression of attend_v, attend_m, attend_n to
assaults. We know that these coefficients are not very accurate as there is a lot of room
for error from “hidden causation and interaction between variables” and are bad models
➢ Display a table of the single variable regressions
➢ Display a graph of the correlation between attend_v, attend_m, attend_n with assaults.
Results are either no correlation(attend_v) or very weak positive ones (attend_m/n)
➢ so we decide to add attend_v and attend_m to get the results of the paper?
➢ from here start following with the paper does: first they do regressions with only year
controls “bc counties in the sample vary year-to-year”
○ paper results: exposure to v movies increases crime, exposure to n movies
increases crime significantly suggesting that at least part of this correlation is due
to omitted variables
○ Our results: exposure to v (2.49) and m (1.49) movies increases crime
➢ Second they do regressions with year and months
➢ Third they do regressions with year, months, holiday
➢ Fourth they do regressions with year, months, holiday, weather
○ Paper results: as they kept adding variables, R^2 increased, and signs flipped on v
movies attendance
○ Our results: as we kept adding variables, R^2 increased, and sign did flip on v and
m movies attendance
○ “This suggests that the seasonality in movie releases and in crime biases the
estimates upward”
➢ Insert table of these 4 different “basic” regressions, base it off of Table II
➢ “Negative correlation may still be due to other unobserved variables: rainy days assaults
are lower but attendance higher. So add weather controls for hot/cold temps, snow/rain”
➢ Here we decide to start doing the regressions with the added violent movies
➢ Now start playing around with different weather and holiday controls
○ Paper results: “attendance usually increases for holidays, different holidays have
different impacts on attendance, weather just talked about where they got the data
and are categorized”
○ Our results: assaults increase during summer months (6, 7, 8) with ALLv but
decrease for nonviolent, assaults are bigger w_rain > 0.3 for ALLv (113.51) and a
little bit bigger as well for nonviolent (38.65) [different from paper, but maybe not
adding the same control variables],
➢ Maybe include a table of difference in attendance for certain holidays, weather?
➢ Do some sort of t/f test to confirm that our findings somewhat match the findings from
the paper (what Kellen did at the end of his paper)
➢ Talk about maybe why findings aren’t exactly the same bc of the difference in data
○ “OG dataset included far more variables to include in its regression, decreasing
the likelihood of omitted variable bias”
➢ In total…
Potential Extension
➢ Talk about what the original study can expand on
➢ Talk about confounding factors that were ignored
➢ Talk about how these results might change in current times, OG paper was from 2009 and
tech has improved today so….