Ramundo, Red Light Cameras in Chicago 1 Evaluation of Causality between the Installation of Red Light Cameras and the Frequency of Accidents at Major Intersections in Chicago Max Ramundo 14.36 Advanced Econometrics, Spring 2014 16 May 2014 .
Oct 02, 2015
Ramundo, Red Light Cameras in Chicago 1
Evaluation of Causality between the Installation of Red Light Cameras and the Frequency of
Accidents at Major Intersections in Chicago
Max Ramundo
14.36 Advanced Econometrics, Spring 2014
16 May 2014
.
Ramundo, Red Light Cameras in Chicago 2
1. INTRODUCTION
The primary aim of this paper is to determine whether or not the installation of red light
cameras reduces accident frequency at the intersections at which cameras are placed in the City
of Chicago. In 2012, the Insurance Institute for Highway Safety (IIHS) reported 683 fatalities
and an estimated 133,000 injuries nationally in incidents proceeding from drivers disregarding
traffic signals. For all incidents where the primary cause is attributed to a driver running a red
light, it is estimated that the annual cost to society is $14 billion (Federal Highway
Administration, 2008). This includes losses in market and household productivity, property
damage, the cost of medical and emergency services, travel delays, legal and workplace costs, as
well as insurance administration costs (Blincoe, et al, 2002).
In an effort to combat the detrimental effects of intersection accidents, many
communities across the United States have introduced programs installing red light cameras at
intersections. These devices allow for the targeting and identification of any vehicle crossing an
intersection after the signal directing the vehicles direction of traffic has turned red. The
application of these cameras first saw use in the United States in the early 1990s in response to
an increasing trend of accident fatalities at intersections. Reported incidents jumped by 19%
nationwide between 1992 and 1996 (Retting, et al, 1999). According to the IIHS, 503 U.S.
communities were operating red light camera programs in 2013. Chicago launched such a
program in 2006 with a scale up covering many major intersections around the city occurring in
2008. Chicago will then be used as a case study in the identification of causality between the
implementation of red light cameras and the safety of intersections.
Numerous studies attempting to answer this question have been conducted in the past
both independently and through communities operating red light programs. Results have
generally been conflicting. As a key example, a Virginia Department of Transportation report in
Ramundo, Red Light Cameras in Chicago 3
2007 finds that red light cameras are not universally effective in improving intersection safety.
The report goes on to detail that the level of intersection crashes actually increased overall during
the period of observation (Garber, et al, 2007). Conversely, a previous study commissioned by
the Virginia Transportation Research Council and conducted by the same lead author found a
significant reduction in violations and incidents following the implementation of intersection
cameras, making a final recommendation to continue and expand their use (Garber, et al, 2005).
Programs have been abandoned in communities where findings on the effectiveness of cameras
in achieving their intent have been inconclusive or negative, and the IIHS reports that the number
of programs, though rising rapidly over the past two decades overall, has declined since 2012. If
cameras are proven ineffective, it can be argued that they pose an unnecessary tax on the citizens
of the community in which they are installed.
Determining the true effect of red light cameras is important because of their purported
potential to improve safety at intersections. Accidents that proceed from drivers disregarding
traffic signals represent a significant cost to society. With the myriad of conflicting reports
previously completed on the subject, the only prudent course of action is to continue studying the
question in an attempt to help foster a more coherent and comprehensive consensus. If red light
cameras do, in fact, contribute positively to intersection safety, corroborating this hypothesis
would be important to facilitate a more widespread adoption of technology that could save
billions of dollars and, more importantly, motorists lives. On the other hand, if the cameras can
be proven to be ineffective, communities will be more inclined to pursue alternative solutions
that might be more effective and be less inclined to sink taxpayer money into a deficient
technology.
Two separate studies were conducted by the City of Chicago in 2010 and 2011. The first
considered a small set of cameras that were installed in 2006 and 2007 and compared the average
Ramundo, Red Light Cameras in Chicago 4
number of accidents at each intersection two years before the camera installation to the average
taken two years after the installation. The second report incorporated cameras installed in 2008,
representing the majority of the intersections in the study. It compared the average number of
crashes at each intersection in 2005 with that in 2010. Both studies found a decrease in the total
number of crashes over all intersections considered, with both reporting an increase of lower
magnitude in the number of rear-end crashes over time. Neither study uses a regression model,
relying on changes in averages over time exclusively. They also do not include a control group
for reference. In evaluating Chicagos red light program, I expand on these studies by
implementing a fixed effects regression model to determine statistically if there is causality
between the cameras and the safety of the respective intersections. I compiled a panel data set
that contains intersection crash data for seven years (2005-2011) for intersections that saw a
camera installed in 2008 (the treatment group) and for intersections where a camera has never
been implemented (the control group). I use this data set to determine differences-in-differences
(DD) estimates characterizing the two sets of intersections. Finally, knowing the primary cause
and severity of each crash observed, I calculate the likelihood of each type of crash and its
associated cost to society. Estimating the fraction of red light violations that result in a crash, I
determine the distributed cost of each violation, comparing this value to the penalty associated
with a camera-issued citation to see whether or not the ticket adequately represents the expected
cost of the violation.
Using a formal econometric model in analyzing the crash data garnered from
intersections around Chicago will serve as a more comprehensive evaluation of the citys red
light camera program. While the previous studies commissioned by the city detail a strong
association between the program and augmented intersection safety, they do not try to establish
causality, which I attempt to do in this report. The results of the fixed effect regression analysis
Ramundo, Red Light Cameras in Chicago 5
indicate a correlation between the treatment and a decreased number of intersection crashes;
however, the results are not statistically significant, so causality cannot be assumed. Overall, the
model predicts a negative correlation between crashes and time starting in 2009 (and a positive
correlation before 2009). This is expectedthe cameras were installed at the intersections under
consideration in 2008, and an associated effect is apparent in 2009 onwards. Given the
apparently statistical insignificance of the treatment and the fact that the time effect is universal
among the treatment and control group, it appears that any effect that the installation of cameras
at treatment intersections might have on crash rates spills over to control group intersections. A
discussion of this halo effect is expanded in the next section. Similar results were obtained for
both a linear and Poisson fixed effects regression. I attribute the ambiguity of the results to the
selection of the control group. Using control sites in the same jurisdiction as that of the treatment
sites may be inadequate due to the aforementioned halo effect of camera enforcement. I wanted
to restrict the evaluation to be within the city of Chicago such that the site-dependent fixed effect
could be at the intersection level rather than the jurisdiction level.
The DD construction also yields ambiguous results. The average number of crashes at
intersections follows a similar trend for both the treatment and control groups, increasing up to
2007 and then declining past 2008. Summary statistics also show a significant increase in the
number of rear-end crashes after the treatment year, accompanied by a mild decrease in the
number of angle crashes. Calculating the expected cost to society of disregarding a traffic signal,
I find that the expected cost far outweighs the current penalty for a violation.
2. LITERATURE REVIEW
The studies previously conducted by the City of Chicago use a similar albeit more limited
version of the dataset that I constructed for this research. They employ no formal econometric
Ramundo, Red Light Cameras in Chicago 6
techniques or analysis and rely only on observed trends in year-on-year summary statistics to
draw conclusions on the effectiveness of the citys red light camera program. The research of
Richard A. Retting is often cited supporting the benefit of cameras. His two most prominent
reports include studies on programs in Oxnard, California, and Fairfax, Virginia (Retting, et al,
2002 and 1999, respectively). Retting implements a generalized linear regression model in both
studies to evaluate the average change in intersection crashes, further employing analysis of
variance to test the statistical significance of the results. For comparison, he runs the same
analysis on data collected from another city in the same state, one without a red light camera
program. The assumption is made that the cities are similar enough to establish unit
homogeneity. This assumption is based solely on the fact that the cities saw a similar number of
intersection crashes in the first pre-treatment year of observation. While Rettings research is
well-established, I maintain that the city-dependent fixed effect could be powerful enough to
discredit his assumption of unit homogeneity and therefore his results.
A comprehensive review of literature on this subject was published by the IIHS in 2002
(Retting, et al, 2002). The review included reports from six different countries where
communities were operating red light cameras. The reports make conclusions based on evidence
proceeding from both observations of correlation and from linear regression analysis. Again,
fixed effects and non-linear regression techniques are not employed. The review concludes that
camera enforcement is highly effective in reducing signal violations and the dangerous accidents
associated with them. For the studies that also take into account intersections without cameras, a
halo effect is generally reported, whereby reductions in crash rates are also seen at these
control intersections. The idea is that the installation of cameras at select intersections impact the
behavior of motorists in general; that is, the behavior modification causing them to exercise more
caution when approaching an intersection is not restricted to the local area around the treatment
Ramundo, Red Light Cameras in Chicago 7
intersection. It can therefore be difficult to build a control group from intersections in the
jurisdiction where the program is implemented. This subsequently makes it difficult to conduct
an accurate DD estimate without making assumptions on unit homogeneity at levels above that
of the intersection. Comparisons between a treatment and control group where the control group
is subject to this halo effect will likely underestimate the effect of the treatment.
The report also details a number of other issues that specifically pertain to the evaluation
of this subject. There is almost always an unavoidable selection bias with respect to the treatment
group. Communities are inclined to install cameras at intersections that see a historically high
comparative level of crashes. In effect, the treatment group is more often than not composed of
outliers. Because of this, there is a significant risk of overestimating the treatment effect. Galton
observes that outliers will tend to move toward the average in successive periods (regression to
the mean) (Galton, 1889). It is therefore easy to contribute a natural regression to the mean (with
respect to intersection crash rates) to the installation of a red light camera. Finally, while the
review reports a decrease in the overall number of crashes at treatment sites in most instances, an
increase in the number of rear-end accidents is nearly universal. Red light cameras have the
unintended effect of increasing erratic stopping behavior, heightening the risk of a collision from
behind. These accidents are often attributed to the rear driver following too closely or to the front
driver operating the vehicle recklessly.
Through this report, I will be expanding on the previous studies conducted by the City of
Chicago while utilizing a similar but more comprehensive dataset. I have collected panel data
that covers seven years, where Chicagos previous studies only utilized data from two years (at
some point before and after the installation of a camera). Moreover, because the previous studies
only compared select year-on-year summary statistics, they are only valid in the sense that they
draw attention to an associative relationship between camera implementation and intersection
Ramundo, Red Light Cameras in Chicago 8
crashes. By utilizing regression techniques, I will attempt to prove causality between the two
parameters, rather than just a correlation. Going even further with the data available, I take
advantage of the availability of information on primary and secondary causes of crashes.
Limiting the scope of data used to instances where the cause of an accident is attributed to the
disregard of a traffic signal, I then adapt a method of cost-benefit analysis used by Kriz, et al
(Kriz, et al, 2006). This will allow me to estimate the expected cost to society of a red light
violation and, in combination with my regression analysis, the potential savings that could be
garnered from a successful red light camera program. This will also facilitate the determination
of an adequate citation accompanying a violation, equating the ticket with the expected cost to
society of an accident distributed across all violations regardless of whether they result in a
crash.
I will also be using a non-linear regression method in my analysis, something that I did
not see as common in my literature review but think will be useful in more accurately modelling
the data. If this method proves effective in attaining a valid result, it could be applied in future
studies or in revisiting past research on this issue. The manner in which I construct a DD
estimate is also relatively unique. Few studies that I came across in my literature review included
in the control group intersections that were in the same jurisdiction as those in the treatment
group. I pursue the same strategy so that the fixed effect in the regressions can be at the level of
the intersection. Studies that build a control group from intersections in a different city account
for differences between the cities as a whole, rather than differences between individual
intersections. The former approach restricts the researcher to conclusions about the effect of
cameras on the city as a whole. This approach could potentially permit the evaluation of how
cameras affect the individual intersections at which they are installed. While there is an apparent
risk of a halo effect obstructing an observation of the true treatment effect, the alternative (using
Ramundo, Red Light Cameras in Chicago 9
data from a second city as the control) is difficult because cities of a comparable size that could
be considered (a shortlist containing Los Angeles, Houston, Philadelphia, Phoenix, and San
Antonio) all run red light camera programs as well. The method of using separate cities to build a
control group has been implemented in previous studies. While this would theoretically mitigate
the halo effect entirely, that all cities similar to Chicago have experienced some form of the
treatment did not allow me to employ this strategy. Nevertheless, recognizing that the method by
which I am analyzing Chicagos program differs from the citys previous studies, the results of
my analysis will be novel if not different as will be the conclusions I detail at the end of this
report.
3. THE MODEL
The question asked in this report is looking for a causal link between red light cameras
and intersection safety. The hidden parameter that facilitates this relationship and benefits from
being explained in a formal model is the behavior of motorists. An intermediate question might
be: Do red light cameras change motorists behavior? If the answer is yes, the manner in which
they do determines the answer to the original question. Red light cameras in Chicago are evident,
and information regarding their locations is available to the public. If motorists know an
intersection is patrolled by a camera, they should be less inclined to disregard traffic signals
there. Ignoring the assumption of universal knowledge, we can make more specific predictions
about how the trends in accidents and violations are likely to change over time after the
installation of a camera. If motorists that are accustomed to running a light at an intersection are
not aware that a camera has been implemented there, they may not change their behavior
immediately. Upon receiving a citation, however, they gain knowledge and an incentive to start
exercising more caution. We may therefore see a delayed onset in the treatment effect. Along the
Ramundo, Red Light Cameras in Chicago 10
same lines of information deficiency, if motorists learn about the installation of a camera as they
approach the physical intersection, the sudden realization may cause them to act rashly in order
to avoid a ticket for running the light. Previous studies have shown an increase in rear-end
accidents associated with camera installationthis effect should also be apparent; however, it
should also taper off as knowledge among motorists becomes universal and the effect of a
surprise of an intersection camera is diminished. I use a model with more than two periods to try
and capture these nuances in behavior. While crash statistics are the best metric to evaluate the
effectiveness of red light cameras, understanding the changing behavioral considerations that
directly impact these statistics is also important. This could have been formalized by using a
panel data set of citations issued at each intersection; however, adequate data was not available
to facilitate this experiment. Controlling for year-on-year variation in traffic volume, citations
could have been used as an outcome that may be superior to the number of crashes in measuring
the impact of the cameras after installation. However, a DD estimate similar to the one used with
crashes as the outcome would not have been appropriate, because pre-treatment data would have
been inadequate. Enforcement at intersections prior to camera installation would have been
subject to an uncontrollable level of variation. The empirical model therefore might not be able
to elucidate behavioral changes as well as the theoretical model can. The former is nevertheless
essential for establishing causality and running a cost-benefit analysis.
4. DATA DESCRIPTION
For this report, I gained access to Illinois Safety Data Mart that is a maintained by the
states Department of Transportation. The database makes available crash data that is submitted
by law enforcement agencies throughout the state. Information on the exact location of crashes
(the unit of observation) was available for the years 2005 through 2011. This information (along
Ramundo, Red Light Cameras in Chicago 11
with a binary that indicated whether or not the crash happened at an intersection) allowed me to
collect a panel data set with year-on-year observations for each intersection considered. While
the data included all information relevant for this research, the database was a limiting factor.
The crashes were mapped with a latitude and longitude value; however, they were never
attributed to a specific intersection, so I could not query the database using that mode of
identification. I was therefore limited to searching all crashes within the vicinity of an
intersection and confirming the correlation with the intersection using the aforementioned binary
indicator. Furthermore, results returned by the database were limited to 300 observations per
query. This also complicated the collection of the data. Nevertheless, in total, I collected 11,469
observations from 104 intersections (54 treatment, 50 control) covering a span of seven years
(2005-2011). All treatment intersections had cameras installed sometime in 2008, so I begin
looking for the effect of the treatment in 2009.
The data for the treatment group was collected at all intersections where a red light
camera was installed in 2008. Intersections for the control group were selected randomly with
the restrictions that they were composed of two thoroughfares and had a similar level of
accidents to the intersections in the treatment group in the first pre-treatment year of observation
(2005). Instead of using all intersections at which a camera was not installed as the treatment
group, I limited data collection to these 50 intersections for three reasons. First, the query
restrictions posed by the database would have made the compilation of a broader array of
intersection data impractical. Second, I excluded intersections that were adjacent to a treatment
intersection to try and mitigate the expected halo effect. All intersections considered in the
control group were separated from treatment intersections by at least one major junction in all
directions. Again, were it not for the ubiquity of red light camera programs in major cities in the
United States, a better control group could have been constructed using data from a similar but
Ramundo, Red Light Cameras in Chicago 12
entirely separate city in order to completely mitigate the halo effect. Finally, using intersections
that were similar to the majority of treatment intersections should have minimized the extent to
which the amount of exposure affected the count data. In this case, the key parameter that
would affect crash statistics is the intersection traffic volume, which here the underlying
population size. Furthermore, this controls for the level of law enforcement at each intersection
in pre-treatment years. I assume that intersections with similar traffic volumes would have seen a
similar police presence that could have deterred red light violations before and after the
installation of cameras.
Selected summary statistics are presented in Tables 1-3 below. Table 1 presents the
average number of crashes (and associated standard deviation) for the treatment and control
group intersections considered as a group. The shaded column in 2008 indicates that the
treatment (camera installation) was applied in this year. In line with predictions based on the
previous studies, a drop in the average number of crashes is observed in the year following the
treatment. Congruent with the prediction of a halo effect, this decline is seen for both the
treatment and control groups.
Table 1. Average Intersection Crashes by Year
Crashes 2005 2006 2007 2008 2009 2010 2011
Treatment
Mean 20.8 21.6 24.4 21.2 17.7 17.7 15.5
Std Dev 9.0 9.1 8.6 8.5 6.9 6.7 6.8
Control
Mean 12.4 16.1 16.9 15.3 12.2 11.9 11.4
Std Dev 5.9 7.5 7.6 7.2 5.4 5.0 6.4
Ramundo, Red Light Cameras in Chicago 13
Table 2 gives the fraction of crashes that involve an angle collision, rear end collision, or
are attributed to a driver disregarding a traffic signal (running a red light) as a percentage of all
intersection crashes in the city for a given year. This takes into account a random sample of
incidents not restricted to the specific intersections selected for the treatment and control groups
(the data from which is presented in Table 1). From a pre- and post- treatment perspective, there
is a slight average decrease in angle crashes after 2008 when the majority of red light cameras in
Chicago were installed. The more significant change is in the number of rear end crashes that
occur after the treatment. The figure nearly doubles from 2008 to 2009 and continues to be
elevated through to 2011. Finally, while the level of crashes attributed to a red light violation
decreases following the treatment, it follows a pre-existing trend starting in 2005 making change
harder to attribute to the implementation of cameras.
Table 2. Types of Intersection Crashes by Year
Fraction of Crashes 2005 2006 2007 2008 2009 2010 2011
Angle 0.22 0.26 0.25 0.21 0.20 0.24 0.23
Rear End 0.25 0.22 0.24 0.24 0.40 0.33 0.32
Disregarding Signal 0.12 0.10 0.08 0.06 0.05 0.12 0.06
Table 3 also considers all intersections in Chicago. Taking again a two period perspective
(pre- and post- treatment), the average number of intersection crashes resulting in a fatality
declines following the treatment in 2008. The same is true for crashes that see an incapacitating
injury, and the declining trend is more apparent year-on-year (instead of just when comparing
two aggregations). These two types of consequences represent the worst (and most expensive to
Ramundo, Red Light Cameras in Chicago 14
society) classifications, so significant decreases in the cost to society proceeding from red light
violations will most likely be a product of decreases in these two categories.
Table 3. Consequences of Intersection Crashes by Year
Total Crashes 2005 2006 2007 2008 2009 2010 2011
Fatal 15 19 22 13 15 13 13
Incapacitating
Injury 351 399 319 182 226 187 170
Although a complete panel data set for red light camera citations by intersection was not
available, I was able to obtain aggregate data for the total number of tickets issued by year. The
data included is from 2009 to 2013 to control for an increasing number of cameras in the city up
through 2009. No cameras were installed after 2009 and before the fourth quarter of 2013. This
data is presented in Figure 1 below.
Figure 1. Citations Issued by Red-Light Cameras throughout Chicago
500
550
600
650
700
750
2009 2010 2011 2012 2013
TIC
KE
TS
ISSU
ED
(X
10
00
)
Red Light Camera Tickets Issued
Ramundo, Red Light Cameras in Chicago 15
The data shows a consistent and significant year-on-year decline in the number of tickets issued
for red light camera violations throughout the city. While this in itself is not an indicator of
improved intersection safety, it does indicate that the cameras are achieving the goal of deterring
drivers from running red lights, which is itself a means to the end of improving intersection
safety.
5. ECONOMETRIC MODEL
For this model, I assume that the treatment is singular and binarythe installation of a
camera at an intersection is the only change that occurs at the time of the treatment and there is
only one level of intensity at which the camera can be implemented. For the econometric
specification, I adopted a fixed effects strategy in two forms, linear and non-linear. Both have
seen implementation in select previous empirical work; however, these techniques have not been
utilized in analyzing Chicagos red light camera program. The first is a fixed effects linear
regression model that uses the natural logarithm of the crash counts as the dependent (response)
variable. The fixed effects serve to control for the range of discrepancies between the
observation points and those that are time dependent. Even apparently similar intersections in the
same city will vary widely based on a number of factors (traffic volume, visibility constraints,
signal timing, etc.). For this reason, when evaluating crashes across all intersections, omitted
variable bias will come from both the intersection and year level. To account for this bias, I
include in my regression a time effect that is common among all intersections (incorporating
year-on-year trends common throughout the city) and a time-invariant effect that captures
constant differences between intersections. These effects act as coefficients on dummies for the
year in which an observation occurs and the intersection where the crash occurred, respectively.
Because Chicagos intersections are well-established and constant in structure over the periods of
Ramundo, Red Light Cameras in Chicago 16
observation, I assume that the discrepancies between the intersections are constant in time. The
specification for the linear model is given below.
(log ) = + + +
E(int | n, t) = 0; n indicates whether an intersection is in the treatment or control group; t gives
the time period; and camera is a binary for camera installation. The regression takes the form
below when run in practice.
log = 1 2011
=2005+ 2
Year is a binary that indicates the period; the second item is an interaction term, where post is a
binary indicating whether or not the data was taken in a post-treatment year (after 2008) and
camera is second binary, again for camera installation. For this specification, a statistically
significant and non-zero coefficient for the interaction term would point to causality between the
installation of red light cameras and a decline in crashes at the respective intersection.
The fixed effects Poisson specification utilizes the same set up as the linear model;
however, the dependent variable crash is kept as linear instead of logarithmic. Because the
dependent variable is a count, a Poisson regression is appropriate. I make the assumption that the
observed events are independent in that the instance of one crash does not make the next crash
more or less likely to occur. The use of this type of regression is also prudent because, like the
data, the Poisson distribution is discrete and positive. The model is therefore restricted to
predicting rational outcomes (that is, the model will not be able to predict negative crashes at an
intersection for some year). Assuming the data follows a Poisson distribution, the model predicts
a value for crashes y based on the expected number of crashes to occur with a probability mass
function as given below.
( = | ) =
!
Ramundo, Red Light Cameras in Chicago 17
A drawback to this specification is that it is limited by the assumption that, with only one free
parameter, the mean and the variance of the data are equal, which is not the case for this data set.
Although subsequent overdispersion is a negative consideration, the specification is still useful
for its ability to effectively model count data. The standard error in this regression will be
adjusted for clustering on the individual intersections to account for the fact that multiple
observations are assigned to a single intersection identification.
To obtain a causal estimate of the camera parameter, the population DD will be
calculated as shown below. This method is a basic theoretical set up (Angrist and Pischke, 2008).
[E(yint | n = treat, t = 2) E(yint | n = treat, t = 1)]
[E(yint | n = control, t = 2) E(yint | n = control, t = 1)] =
The DD specification compares four meanstwo each from the treatment and control group at
two different time periods. The difference in these differences is taken as , the causal parameter.
This contrasts the change in the number of crashes at intersections where a camera has been
installed with the same parameter change at untreated intersections around the timeframe when
the cameras were installed. To construct this model, the assumption must be made the trend in
intersection crashes would be the same among all intersections if the treatment was never
applied. The intent of the control group is to exhibit this constant trend in order to better
visualize the deviation to be seen in the treatment group.
As an addendum to the primary specifications in this report, I also set out to estimate the
cost to society of a red light violation and, subsequently, the potential savings that could be
generated by reductions in violations as a product of red light cameras. To do this, I make several
assumptions based on findings in previous studies. From Rettings Fairfax study, I take that 0.2%
of all violations result in an accident (Retting, et al, 1999). I take the cost of each type of accident
from Krizs report on red light cameras in Milwaukee (Kriz, et al, 2006). These estimates include
Ramundo, Red Light Cameras in Chicago 18
losses in market and household productivity, property damage, the cost of medical and
emergency services, travel delays, legal and workplace costs, as well as insurance administration
costs. Citation costs are excluded because they are a transfer from individuals to the government
and therefore do not represent a loss to society. The fraction of each type of accident is adjusted
for the overall probability of an accident happening in the first place. This probability is then
multiplied by the expected cost of the accident to yield a figure distributed over all violations.
The values are summed for each type of accident, and the final figure is compared to the average
citation in Chicago to see if the penalty exceeds or falls short of the expected cost to society.
6. RESULTS
The results of the log-linear fixed effects regression and the Poisson fixed effects
regression are presented below in tables 4 and 5, respectively.
Table 4. Summary of Estimated Effects from Log-Linear Regression
Log(crashes) Coef. Std. Err. t P > | t | 95% Conf. Interval
2006 .151 .043 3.48 .001 .066 .236
2007 .252 .043 5.81 .000 .167 .337
2008 .100 .043 2.29 .022 .014 .185
2009 -.069 .050 -1.38 .169 -.166 .029
2010 -.078 .050 -1.56 .119 -.175 .020
2011 -.214 .050 -4.30 .000 -.312 -.116
post*treat -.012 .047 -0.26 .797 -.104 .080
constant 2.68 .031 87.23 .000 2.62 2.74
Ramundo, Red Light Cameras in Chicago 19
Table 5. Summary of Estimated Effects from Poisson Regression
crashes Coef. Robust
Std. Err. z P > | z | 95% Conf. Interval
2006 .120 .035 3.46 .001 .052 .188
2007 .214 .038 5.68 .000 .140 .287
2008 .089 .036 2.46 .014 .018 .160
2009 -.101 .049 -2.07 .039 -.197 -.005
2010 -.109 .048 -2.29 .022 -.202 -.016
2011 -.208 .056 -3.74 .000 -.318 -.099
post*treat -.014 .058 -0.23 .816 -.127 .100
The most important variable to consider in the tables above is the post*treatment
combination. Again, this represents the expected change that would be seen at a treatment
intersection after the installation of a camera there. In both regressions, however, this is the least
statistically significant variable (in that the result cannot be considered statistically significant).
The corresponding coefficient is negative, as expected, but the test statistics do not support this
as a significant result. I do not believe that this result proceeds from an inadequate econometric
specification; I believe it is a product of the control group data. The summary statistics show a
decrease in the number of crashes at treatment intersections following the date of the treatment;
however, this is also apparent for the control group. The coefficient of the treatment parameter
and the significance of the variable are determined through a comparison with the control group.
It would appear that there is a general downward trend in accidents at all intersections
throughout Chicago over the period of observation. Whether or not this is a product of camera
installation is difficult to discern because the control group (against which I am comparing the
treatment group) experiences a similar trend to the treatment group. This is evidence of the halo
Ramundo, Red Light Cameras in Chicago 20
effect reported in previous literature on the subject. It is apparent that a change in motorist
behavior (be it as a product of the camera program or not) is not constrained to the treatment
intersections. The solution to this problem would be to select a control group outside Chicago
and set the time-independent fixed effect at the level of the city rather than the intersection.
However, the aforementioned difficulties with finding a comparable control group to resemble
Chicagos untreated intersections is a limiting factor in this research. Shortcomings of the final
regressions notwithstanding, the coefficients can be interpreted. The log-linear regression
predicts that the presence of a camera at an intersection will decrease the number of crashes seen
at that intersection by 1.2%. The coefficient on the explanatory variable in the Poisson model is -
0.0135; therefore, the number of crashes at an intersection decreases by a factor of e-0.0135 =
0.9866 given the presence of a red light camera at that intersection. This translates to a 1.34%
decrease, on par with what was predicted by the log-linear model.
The dummies for the observation period in both regressions are more statistically
significant. We see the coefficients on these dummies switch from positive to negative in 2009
this is expected with the treatment occurring in 2008. However, this can only be taken as
associational evidence given the fact that the treatment dummy in each regression comes out to
be insignificant. A visualization of the DD estimate is shown below. This better represents the
trend followed by both the treatment and control group intersections.
Ramundo, Red Light Cameras in Chicago 21
Figure 2. DD Estimate of Intersection Crashes
The two groups follow very similar trends throughout all observation periods. There is a
noticeable decline in the expected number of crashes after 2007. Some cameras were installed as
a trial in Chicago as early as late-2006; the decline before 2008 could be an artifact of a city-
wide behavior change starting in anticipation of the widespread implementation of the camera
program. In a DD estimate, one would expect to see the control group maintaining a steady trend
from the first observation period while the treatment group deviates from this trend at the time of
the treatment. Furthermore, the treatment group, on average, sees a higher number of accidents.
This is not unexpected, as there is a significant level of selection bias in the treatment group that
is due to the city installing cameras at historically dangerous intersections.
Going further into the data parameters, a breakdown of the trends in the types of crashes
is shown below.
0
5
10
15
20
25
30
2005 2006 2007 2008 2009 2010 2011
AV
ER
AG
E C
RA
SHE
S
Intersection Crash Trends
Treat Control
Treatment
Ramundo, Red Light Cameras in Chicago 22
Figure 3. Crash Types as a Percentage of Total Intersection Crashes
The decline in angle crashes following the treatment is apparently nominal. One might expect a
greater decrease, as angle crashes are most often the result of one driver disregarding a red light.
An expected spike in rear end crashes following the treatment is observed, however. Motorists
previously unaware of the installation of a camera and realizing its presence only when
approaching an intersection can panic and decelerate rapidly to avoid a citation. This facilitates a
hazardous situation for a following driver who can collide with the front driver if he or she fails
to anticipate the rapid deceleration of the latter. This spike tapers off in later periods as drivers
learn where cameras have been installed and can better anticipate so as to not have to act rashly
when approaching an intersection.
The trend in intersection crashes by type is further broken down to control for whether or
not the intersection is a treatment or control site. These results are presented below.
0
0.05
0.1
0.15
0.2
0.25
0.3
0.35
0.4
0.45
2005 2006 2007 2008 2009 2010 2011
FR
AC
TIO
N O
F T
OT
AL
CR
ASH
ES
Intersection Crash Types
Angle Rear End
Treatment
Ramundo, Red Light Cameras in Chicago 23
Figure 4. Angle Crashes as a Percentage of Total Intersection Crashes
Figure 5. Rear End Crashes as a Percentage of Total Intersection Crashes
0
0.02
0.04
0.06
0.08
0.1
0.12
0.14
0.16
0.18
2005 2006 2007 2008 2009 2010 2011
FR
AC
TIO
N O
F T
OT
AL
CR
ASH
ES
Angle Crashes
Treatment Control
Treatment
0
0.05
0.1
0.15
0.2
0.25
0.3
0.35
0.4
0.45
0.5
2005 2006 2007 2008 2009 2010 2011
FR
AC
TIO
N O
F T
OT
AL
CR
ASH
ES
Rear End Crashes
Treatment Control
Treatment
Ramundo, Red Light Cameras in Chicago 24
Comparing control and treatment intersections, a relative decline in angle crashes is seen at
treatment crashes as would be expected and desired. The increase in rear end accidents reflects
an overall adjustment in driver behavioras cameras become more ubiquitous throughout the
city, drivers are more likely to stop suddenly at all intersections if they were not previously
aware of the status of the intersection.
Table 6 below breaks down intersection crashes by results ranging from property damage
to fatality. It gives the probability of each event, the associated cost, and the cost distributed
among all red light violations that occur. Crash data on which this is based is taken evenly from
all observation periods considered.
Table 6. Estimated Expected Cost of a Red Light Violation
Result Probability Cost Cost per Violation
Fatal (K) 8.94E-05 $ 4,803,555.00 $ 429.37
Incapacitating Injury (A) 8.94E-05 $ 484,773.00 $ 43.33
Non-Incapacitating Injury (B) 0.00027 $ 202,605.00 $ 54.33
Possible Injury (C) 0.00025 $ 79,519.00 $ 19.55
Property Damage Only (0) 0.00131 $ 21,923.00 $ 28.66
No Accident 0.998 $ - $ -
Total $ 575.24
This method is derived from a study making the same estimate for the city of Milwaukee. The
final cost per violation here is relatively high when compared with the literature value. This is
largely due to the fraction of fatal accidents being higher for Chicago (where fatal accidents are
by far the most costly to society). The final cost per violation should set a target for citations
Ramundo, Red Light Cameras in Chicago 25
issued for a red light violation, as the cost of the violation should represent its cost to society as
accurately as possible. A ticket in Chicago, however, is only $100.00 for first-time offenders.
This research suggests that the value could be higher with ample justification as a fair penalty.
7. CONCLUSIONS
The regression analysis in this report failed to prove causality between the installation of
red light cameras at intersections and the crash statistics at those intersections. Because the
control group was constructed from intersections within the same jurisdiction as those in the
treatment group, a halo effect mitigating the differences between the groups was unavoidable. A
future study should consider a control group outside Chicago while setting the fixed effects to
the level of the city rather than the intersection. In hindsight, a better question to answer would
have been, Do red light cameras impact accident frequencies at intersections throughout the
associated community? Nevertheless, based on the summary statistics and DD estimate, there
appears to be a correlation between the rollout of the red light camera program in Chicago and
crash rates at intersections across the city. This study also confirms an expected spike in rear end
accidents at intersections that proceeds from the implementation of the cameras. Finally, it
estimates the expected cost to society of an individual red light violation, such that the
recommendation can be made that Chicago is within the rights of a fair penalty to increase the
value of a ticket issued with each citation. This research should be continued, as red light
cameras have the potential to improve road safety and mitigate significant societal costs. This
report is not conclusive on Chicagos program; although causality was not proven here, this was
a product of the data collection method, not the econometric specification or the shortcomings of
the program itself. Therefore, further efforts should be taken to evaluate this program and
establish a formal relationship between red light cameras and intersection safety.
Ramundo, Red Light Cameras in Chicago 26
8. REFERENCES
Angrist, J., Pischke, J. S. (2008) Mostly Harmless Econometrics: An Empiricists Companion.
Princeton University Press. pp. 169-181
Blincoe, L., Seay, A., Zaloshnja, E., Miller, T., Romano, E., Luchter, S., and Spicer, R. (2002)
The Economic Impact of Motor Vehicle Crashes, 2000. U.S. Department of
Transportation. Washington, D.C.
Federal Highway Administration (2008) NHTSA Traffic Safety Facts 2008. U.S. Department of
Transportation. Washington, D.C.
Galton, F. (1889) Natural Inheritance. Macmillan, London
Garber, N., Miller, J., Abel, E., Eslambolchi, S., Korukonda, S. (2007) The Impact of Red Light
Cameras (Photo-Red Enforcement) on Crashes in Virginia. Virginia Department of
Transportation. Charlottesville, VA.
Garber, N., Miller, J., Eslambolchi, S., Khandelwal, R., Mattingly, K., Sprinkle, K., Wachendorf,
P. (2005) An Evaluation of Red Light Camera (Photo-Red) Enforcement Programs in
Virginia. Virginia Transportation Research Council. Charlottesville, VA.
Kriz, K., Moran, C., Regan, M. (2006) An Analysis of a Red-Light Camera Program in the City
of Milwaukee. Public Affairs 869. Robert M. La Follette School of Public Affairs,
University of Wisconsin-Madison.
Retting, R., Kyrychenko, S. (2002) Reductions in Injury Crashes Associated With Red Light
Camera Enforcement in Oxnard, California. American Journal of Public Health.
November, 2002, v. 92, n. 11, pp. 1822-1825.
Retting, R., Williams, A., Farmer, C., Feldman, A. (1999) Evaluation of Red Light Camera
Enforcement in Fairfax, Va., USA. ITE Journal. August, 1999, pp. 30-34.
Ramundo, Red Light Cameras in Chicago 27
9. APPENDIX: ORIGINAL STATA OUTPUT
Ramundo, Red Light Cameras in Chicago 28