The Dynamics of Criminal Behavior

Evidence from Weather Shocks

Brian Jacob
Lars Lefgren
Enrico Moretti

ABSTRACT

While the persistence ofcrimiruil activity i.s well documented, this may be due
to persistence in the unobserved determinants of crime. There are good
reasons to believe, however, tluit there may actually be a negative
relationship between crime rales in a particular area due to temporal
displacement. We e.xploit the correlation between weather and crime to
examine the short-run dynamics of crime. Using variation in lagged crime
rates due to weather shocks, we find that the positive serial correlation is
reversed. The.te findings suggest that the long-run impact of temporary
crime-prevention efforts may be .smaller than the short-run effects.

I. Introduction

The persistence of criminal activity is well documented. Higher

crime today in any particular area is associated with higher crime tomorrow. This
serial correlation is illustrated in Figure I. A 10 percent increase in violent crime in a
typical city this week is associated with 1.6 percent more violence the following
week, conditional on jurisdiction-year fixed effects. The serial correlation for property

Brian Jacnh Is assistant pmfessnr at the Jotin F. Kennedy School of Government at Hanard Vuiverxity,
Lars Lefgren is as.^istatit professor of ecnnomirs iil Brighum Young Univerxiiv. and Enrico Moreiti is
associate professor of economics at the University of California al Berkeley. The authors iluink Alberto
Ahadie, Dick Butler, Jne Doyle. Steve Levitt. Lance Loctmer. Erzo Lullinei; Frank Mclntyre. Hruee
Sacerdote. two referees, and seminar participants at the 2004 NtiER Summer Institute, BYU, Wisconsin.
Florida, and 201)4 AEA Meetings for useful suggestions. Email: Jacob can he contacted at
hrianjacoh@hurvard.edu. Leffjren can he contacted at l-lefgren'^hyu.edu. Moretti can he contacted at
moretti@econ.herkeley.edu. The datu used in thi.-i article can he obtained beginning January 2(X)fi
through December 2010 from Enrico Moretti, 549 Evans Halt, Berkeley. CA 94720-iliSO. e-mail:
Moretli(^ econ.herkeley.edu.
ISubmiiled February 2(H)(i; accepted October 2I)O6|
ISSN 022-166X E-ISSN 1548-8004 © 2007 by ihe Board of Regenis of ihe University of Wisconsin System

THE JOURNAL OF HUMAN RESOURCES • XLII • 3

490 The Journal of Human Resources

Lagged Crime (in weeks)

Coeftlcicni on Lagged Vjtilent Crime

Coet'fideni on lagged Property Crime

Figure 1
The Persistence of Crime
Note: This figure shows the coefficieni estimates from two separate regressions of currenl crime (vi-
olenl or property) on ten lags of crime (violent or property). The unit of observalion is a jurisdiclioti-
week. All crime rates are computed by dividing the actual number of crimes during a week hy the
average number of weekly crimes wiihiti the jurisdiclion. Other covariales include jurisdiclion fixed
effects. Standard errors in parenthesis are clustered at the state*year*month level. Observations are
weighted by the mean number of all crimes within the jurisdiction.

crime is even higher—10 percent more property crime this week is associated with
over 3.1 percent more property crime the following week.'
One of the most common explanations for this autocorrelation is that potential
offenders are influenced by the criminal behavior of others. In other words, a crime
committed by one individual increases the likelihood that other individuals in the
same locality will engage in criminal activity. In addition to social interactions that
occur over a long time period through, for example, social learning, criminologists
have long argued that forces such as imitation and revenge may lead to social mul-
tiplier or contagion effects that operate over very short time horizons such as days or
weeks.^ However, the persistence in crime rales over time also could be explained by
the persistence of unobserved factors that determine the costs and benefits of criminal
activity such as police presence and poverty levels. Moreover, there are good reasons
to believe that, particularly over a short lime horizon, Iherc actually may be a nef>alive
relationship between crime rates in a particular area due to displacement^in other
words, the shifting of criminal activity from one time or location to another. A study

1. Correlation is positive for further lags. For example, a 10 percent increase in violent and property crime
this week is associated with 1.1 and 1.6 percent higher crime two weeks later respeetively.
2. See Cook and Goss (1996) and Braga ei al. (2001).
 Jacob. Lefgren, and Morelti 491

of juvenile curfews in Detroit, for example, found that afternoon crime nearly dou-
bled after the introduction of the curfew (Hesseiing 1994).
Understanding the dynamics of criminal activity is of interest for both practical
and thcorciical reasons. If the timing of criminal behavior is more elastic to tempo-
rary changes in the costs of crime than is ihe tolal anmunt of criminal activity, then
the effects of a police crackdown in a given week may be partially offset by increases
in criminal activity in subsequent weeks. From a theoretical perspective, understand-
ing the dynamics of criminal activity is important because it sheds light on the max-
imizing behavior of criminals. Beginning with Becker (1968). economic models of
criminal behavior bavc generally been constructed and tested using a static frame-
work. While these models have been very useful in understanding some features
of criminal behavior, they are not well suited to explaining how criminal behavior
changes over time in response to changes in various costs and benefits. The economic
literature on crime dynamics is limited.'*
in this paper, we exploit the correlation between weather and crime to analyze the
short-run dynamics of criminal activity. Our aim is to determine the true persistence
of criminal activity by estimating the causal relationship between crime rates in dif-
ferent time periods within the same locality. In other words, does more crime today
lead to more or less crime tomorrow?
In order to eliminate any spurious serial correlation arising from persistent unob-
served heterogeneity, we use an instrumental variable strategy that is based on
weather shocks. Criminologists have long recognized that weather is strongly corre-
lated with short-run fluctuations in crime, with hotter weather generally associated
with more crime and inclement weather associated with less crime. Drawing on
crime-level data from the FBI's National Incident-Based Reporting System (NIBRS),
we construct a panel of weekly crime data for 116 jurisdictions from 1995-2001.
Simple OLS estimates contirm that violent and property crimes are highly correlated
over time within localities. Weeks with above (below) average crime rates are typi-
cally followed by weeks with above (below) average crime rates, even after control-
ling for a rich set of jurisdiction-specific seasonality effects.
However, when we instrument for lagged crime with lagged weather conditions,
we actually find the opposite result for both property and violent crime. The 2SLS
estimates reveal that weeks with above average crime rates are followed hy weeks
with below average crime rates. Notably, our results do not appear to be driven by
the persistence in weather conditions over time, or displacement of legal economic
activity. Our models control for a series of jurisdiction-specific seasonality measures
so that our identification essentially relies on deviations in expected weather patterns
that influence crime rates in a particular locale.
The magnitude of the displacement is substantial. A 10 percent increase in violent
crime due tu a weather shock reduces violent criminal activity by about 2.6 percent

3. Similarly, prior work diKunierits Ihal juvenile violence peaks in the afier-school hours on school days
and ill the evenings on nonschool days (Jacob and Lefgren 2003).
4. Otie notable exception is Lochner (19991.
5. ScvLTal channels are potentially resp<jnsilile for this correlation. Psychologists have shown ihai higher
icmperatiircs increase aggression directly (.'Xnderson 2{K11). Allcmalively. adverse weather conditions
may iiffeci the cost of execuling a paniciilar crime and/or Ihe availability and actions of potential victims.
We discuss the implicutitms of these various channels in later sections.
492 The Journal of Human Resources

in the following week. Moreover, there is evidence that additional displacement

occurs over a longer time horizon. The estimated reduction in violent crime over a
month is 5.4 percent, more than double the estimated displacement for one week.
These findings are consistent with a model in which the marginal utility (cost) of vi-
olence is decreasing (rising) in the amouni of violence committed during the prior
week. This would be true if, for example, an assailant who "settles a score" in
one period feels less need to do so in a subsequent period.
The estimates for property crime are similar, though somewhat smaller. A 10 per-
cent increase in property crime reduces property crime by about 2 percent the follow-
ing week. Interestingly, while displacement of violent crime is apparent across
multiple crime categories, displacement of property crimes is concentrated among
crimes that involve highly valuable property, like car theft.'' These results are consis-
tent with a simple model in which transitory fluctuations in the costs of crime create
an income effect that is manifested for multiple periods. This dynamic is similar to
the one observed in a standard labor supply model with transitory shocks in the wage.
Our findings suggest that criminal labor supply re.sponds to transitory fluctuations in
the wage in a manner that is similar to that of other self-employed individuals^ and is
also consistent with evidence on the presence of some forward-looking behavior of
low-income populations.** At the same time, our findings are not consistent with the
case of permanent-income offenders who face no liquidity constraints and have very
long time horizon.
While we believe that these results provide interesting insight regarding the dy-
namic optimization behavior of criminals, the relevance of our results for crime-
prevention policy depends on the nature ofthe policy. Like all instrumental variable
estimates, our findings reflect a particular local average treatment effect—namely,
the impact of an exogenous increase in those crimes that are elastic lo weather con-
ditions. To the extent the changes in crime due to weather shocks are similar to
changes induced by police interventions; our findings suggest thai the long-run im-
pact of short-term police crackdowns may be smaller than the initial effects of these
policies.'*

II. Simple Dynamic Models of Property

and Violent Crime
We present two simple models of property and violent crime. In these
models, utility-maximizing criminals respond to transitory changes in the price of
crime by shifting the time of their criminal activity, inducing a negative relationship
between current and future crime. The goal of these models is to clarify the condi-
tions under which temporal displacement may occur.

6. The value of property stolen also displays signilicanl displacement: weeks where ihe value of siolen
properly Is high are followed by weeks where the value is low.
7. See, for example, the Mudies of [axi drivers by Farber (2005) and Camerer, Babcock. Lowenstien. and
Thaler (1997).
8. Sec studies of welfare recipienis—for example. Grogger and Michalop<»ulos (2003).
9. Note ihut we are not suggesting ihat the effects of such crackdowns •"decay" over time, but rather than
criminals shift criminal behavior to a later date.
 Jacob. Lefgren, and Moretii 493

A. Property Crime
Given the financial motivations underlying property crime, a standard labor supply
Irainework provides considerable insight regarding the potential displacement of
property crime. Displacement, if it occurs, would plausibly come about through an
income effect—in other words, a transitory increase in the benefits of crime generates
a positive Income effcci which reduces the incentive to commit crime in subsequent
periods. This suggests that displacement will not occur when individuals urc unable
to borrow and save. During lucrative periods, for example, nothing is saved so the
next period's choice of crime is unaffected. Likewise during lean times, offenders
are unable to affect available income in subsequent periods by borrowing. This type
of period-by-period maximization also would occur if individuals were completely
myopic (Case I). If. on the other extreme, criminals are farsigbted and have access
to good credit markets, there will also be no displacement because a transitory
change in the price of criminal activity will have a negligible effect on lifetime in-
come (Case 2). In order to generate a linkage between lagged and current criminal
activity, it is necessary lo construct a model that allows individuals to cither save
or borrow across periods. On the other hand, the credit market must be sufficiently
imperfect or the lime horizon must be short enough for a transitory shock in the ben-
efits of criminal activity to have a meaningful income effect (Case 3).
We'll now formalize this intuition in the context of a simple model. We assume
that an individual's utility each period is defined over consumption, c. and leisure,
/, in the following way: u, = u{c,J,). We will assume that u(c,.l,) is increasing in
both arguments and strictly concave. Each period a criminal must allocate a single
unit of time between leisure and property crime, s. Because this time constraint must
hold with equality, it must be the case that /, = 1 - s,. Each period, an individual
earns MV.V, from criminal endeavors, where w, is the net wage from criminal endeav-
ors. Note that w, reflects both the abundance of criminal opportunities and the period
specific costs of engaging in criminal activity. In the context of our empirical anal-
ysis, Huctuations in the weather generate variation in vtv^perhaps due to weather-
related changes in the supply of targets or in the disutility of committing crime
outdoors.'"

Case I
The first case worth discussing is one in which individuals are unable or unwilling to
save or borrow. In this case, the agent faces the following budget constraint: c,'^w,s,
each period. We assume that the criminal maximizes discounted lifetime utility sub-
ject to the budget constraints he faces. It is trivial to show that the first-order condi-
tions are equivalent to those obtained from the period-by-period maximization
problem. This is because each period's utility and budget constraint is unaffected
by that which has gone before or that which will occur later. For this reason, a tran-
sitory shock to the benefit (wage) of property crime can have no impact beyond the
current period. '

10, Similar results could be obtained by assuming that weaiher affects the utility cost nf engaging in crim-
inal activity.
494 The Journal ot" Human Resources

Case 2
Another extreme case is that criminals are farsighted and have access to perfect cap-
ital markets. While it is unlikely that most offenders have access to sophisticated
capital markets, it still may be useful as a benchmark. Abstracting from uncertainty,
each offender chooses a series of consumption and criminal activity to maximize his
lifetime utility. The Lagrangian for the offender's optimization problem is:

max

where r is the interest rate at which an offender can borrow or lend, and Tis the num-
ber of periods and is assumed to be large. In order to understand how a transitory
change in the wage of crime affects .subsequent criminal behavior, it is helpful to ex-
amine the first-order conditions:

1y

+r

together with the budget constraint. Equations 2 and 3 must hold for all t. It is now no
longer true that the first-order conditions are equivalent to those consistent with
period-by-period optimization. Instead, the ability to borrow and save implies that
the marginal utility of lifetitne income. \ . is a function of the wages of crime in every
period. Thus, while an increase in w, can induce a substitution effect that causes
s, to rise, the only mechanism through which it can affect .s,+ \ is through \ . This
corresponds to an income effect. In dynamic models of labor supply, it is generally
assumed that the lifetime income effects of a transitory wage shock are minimal—
limiting the temporal displacement of property crime.

Case 3
Consider a model identical to the one above except with 7=2. For simplicity, assume
that the discount rate and the interest rate are both zero, and the utility function is
separable in consumption and leisure. Under these assumptions, the first-order con-
ditions are identical to those in Equations 2 and 3. What happens to crime at Time 2
when there is an exogenous shock to the net benefit of committing crime in Period 1 ?
The comparative statics are straightforward:

(4) ^ = - ^ i i < o
dw flj dw

where dX/dW] <0 and the number in the subscript denotes the time period. Equation 4
indicates that an increase in the first-period wage of crime will reduce the amount of
criminal activity committed in Period 2. This occurs because a transitory increase in
 Jacob, Lefgren, and Moretti 495

lhe benefits of crime generates a positive income effect (lowering the marginal utility
of wealth), which reduces the incentive to commit crime in subsequent periods. As-
suming that the substitution effect dominates the income etTect in the first period,'' a
transitory increase in the wage of crime will initially lead to higher levels of crime
Eint! will then reduce subsequent criminal activity. In this case, we will observe tem-
poral displacement of property crime.
Even in this third case, it is interesting to question what type of property crime
would generate the strong income effects necessary to observe temporal displace-
ment. Clearly, it is unlikely that stealing a small inexpensive item could have any
significant income effect. If any displacement occurs, we should observe it only
for crimes that involve goods with substantial monetary value. Notably, this is
exactly what we observe in the data. We find that car theft is characterized by toial
displacement, while all other property crimes are characterized by virtually no
displacement.

ti. Violent Crime

While the motives underlying violent crime are less often financial, temporal dis-
placement still may occur. In the framework outlined below, temporal displacement
may occur for two rea,sons. First, the benefits of violence may persist over time. This
would be true if injuring an individual in the lirst period "settled a score" or "taught
a lesson," reducing the need to do so again in the second period. Second, the costs of
violence in one period may depend on the level of violence in the previous period.
For example, alcohol and/or drugs are sometimes a cofactor in violent crimes (bar
fights, domestic violence, etc.). It is possible that the violence that was precipitated
by substance use in the first period may cause the offender to feel guilty, taming al-
cohol consumption in the second period. Alternatively, a violent act in the lirst period
may result in arrest and/or greater police supervision in the second period.
Consider a simple two-period model. Assume that tirst-period utility is given by
the following:

where I'l is violence in the first period, g\ {v\) is an increasing but concave function
of V|, and fl| is an exogenous per-unit cost of violence. Second-period utility is given
by the following:

(6) W2 = g 2 ( v 2 + 8V|) -62(V1)V2,

where i'2 is violence in Period 2, S is the fraction ofthe benefits of violence that carry
over to the next period, 52(i'2+Sv'i) is an increasing and concave function, and
02(v[) is the per-unit cost of committing violence in the second period.

11. If we thought of the amount of properly crime as the dolhirs jjenerated from crimimil activily. M^,.V,. the
nmouril criminal aclivity in Period 1 necessarily increases. In oiher words, the ainouni stolen would rise-
even if the iiinuuni of time spent on criminal activity declined. In ihis case, a rise in lhe Perit>d I wage of
crime would inerea.se Peritxl 1 erime and reduce Period 2 crime. This would lead to the temporal displace-
ment of property crime.
496 The Journal of Human Resources

In our model, the per-unit cost of violence in the second period. 92(vi). is an in-
creasing and convex function of first-period violence. This assumption seems plau-
sible.'^ The criminal's optimization problem involves choosing Periods 1 and 2
violence to maximize utility over the two periods. The first-order conditions of this
problem are the following:

Oi'i dv] ovj dv

and

These conditions define the equilibrium level of crime in the two periods.'^
What happens to violent crime if the cost of first-period violence exogenously
increases (for example, because of a weather shock)? It is not surprising that that
an increase in first-period violence results in a decrease in violent crime in the first
period: ^ < 0 . ' " * The comparative static for second-period crime is more relevant for
our analysis. In particular, what happens to violent crime in Period 2 when there is an
exogenous shift in the cost of violent crime in Period I?

The denominator of Equation 9 must be positive in order for the second order con-
ditions to hold. Both terms in the numerator are positive suggesting that second-
period violence is likely to increase in the first-period cost of violence if either the
benefits of violence are durable (8>0) or the marginal cost of second-period vio-
lence is rising in first-period violence. These findings suggest that the displacement
of violent crime can occur under plausible conditions.

III. Empirical Strategy

The previous section suggests that crime in one period may affect
crime in the next period either positively or negatively. The basic empirical frame-
work therefore relies upon estimating the following simple equation:
(10) crimeu = p,, +

12. Violent acts in Che lirst period may result in arTest or increased police supervision: injuries sustained in
the first period could hamper ihe ability to commil violent crime in the second period: or for crimes such as
domestic violence, it is possible Ihat guill over a violeni acl in Period I would increase the cost of com-
mitting a similar acl in Period 2.

14. In particular, ^ = -^ where A is the determinant of the matrix of second order conditions and must be
positive in order for ihe second order conditions to hold.
 Jacob. Lefgren, and Moretti 497

where crimeij reflects the level of crime in jurisdiction ( in period t and p, reflects
the causal effect of an exogenous increase in criminal activity on the next period's
crime. Following the models presented in Section II, p , is the net effect of social
interactions and displacement. The empirical challenge is that crimcj,- \ is almost
certainly positively correlated to the error term r.,,. since factors affecting the costs
and benefits of crime are likely persistent over time.
The ideal instrumental variable is correlated with criminal activity in period /-1 but
uncorrelated to the error term. One candidate instrument is the weather. The corre-
lation between criminal activity and weather conditions has been well documented."
It has been hypt)thesized that higher temperatures might increase aggression directly
(see Anderson 2001), thus providing a transitory increase in the net benefit of crim-
inal activity. Adverse weather conditions may affect the cost of executing a particular
crime, due to changes in the ease of transportation or the likelihood of witnesses to
the crime, which may influence the chance of arrest.
In this model, the structural equation is given by:

where i indexes jurisdictions, and / indexes time period (in our analysis, a week).
weather,, is a vector of weather variables, and the vector X,, includes jurisdiction
*year fixed effects, jurisdiction-specific fourth order polynomials in day-of-year
and fixed effects for each month. The first stage is given by:

(12) crimeij-] = TXii + y^weatherij- i +72H'ea/fter,,-i-Ti,,."'

To increase the efficiency of our estimates, we weight each observation by the av-
erage number of (violent or property) crimes committed each week within the juris-
diction." The standard errors are cluster corrected at the statc*year*month level as
weather and criminal activity may be spatially and temporally correlated- In some
models we include more than one lag in crime. When Equation 11 includes crime
at r-1. t-2, r-3. and r-4. we instrument t-2 crime using t-2 weather conditions. /-3
crime using /-3 weather conditions, and so forth.
The key identifying assumption in this model is that, conditional on weather at
time / and other covariates. weather at time t-\ cannot directly influence crime at
time t. More formally. cov(K/f. weatherjj - i \weaihen,,Xj,) = 0. In asse.ssing this as-
sumption, it is important to highlight .several potential issues. A first concern is that
weather, when measured at high frequency, is serially correlated. Therefore, in the

15. See Cohn's (1990) extensive literature review on the subject. More reecm work on ihe issue by Rotiiin
and Cohn (2()00) and Field (1992) conHrms tbe tintling.
16. One might t>e concerned [hat thi.s instrumenting siraicgy docs not work tieciiu^ Ihe firsi ,stage seems
inconsisieni wiiti tbe recursive structure of the data-generating process. However, it is trivial to show tbat,
il' the exclusion restriction holds, our strategy yields an unbiased estimate of p , .
17. To understand wby we do this, suppose tbat criminal aciivity is generated by a Poisson process, the
mean and variance or crime with a jurisdiction both equal c,. We discuss below that we normalize crime
rates by the mean weekly incidence over the sample—doing so yields point estimates that can be inter-
preted as percentage effects, tn this case, the variance of the normalized crime measure is l/C,. By weight-
ing each jurisdiction by c,. wtiich is the approximate inverse of the variance of its residual, we increase the
precision of our estimates.
498 The Journal of Human Resources

absence of good controls for current weather, lagged weather may be directly cor-
related with current crime because it will contain information regarding current
weather. While we do control for current weather conditions in our models, it is pos-
sible that weather within a period is measured imperfectly, in which case lagged
weather still may provide information regarding the unobserved aspects of current
weather. For example, one day's maximum temperature likely provides informa-
tion regarding the weather conditions at 1 a.m. of the next day. The combination of
serial correlation in weather, along with imperfect measures of weather conditions
in any one period, will violate the assumptions necessary for satisfactory identifica-
tion and result in a positive bias in the coefficient on lagged crime. We can show that
this bias decreases as the length of the time window expands. (For a fonnal proof, see
Appendix A in Jacob. Lefgren. and Moretti 2004.) Additionally, we later show the
results of a "reverse experiment" in which we estimate Ihe effect o\ future crime
on current crime, and instrument for future crime with future weather. The instru-
ment coefficient on future crime is insignificant, suggesting that this type of bias
in insignificant.
A second concern is that weather may lead to the displacement of noncriminai
economic activity. To the extent that criminal activity is follows noncrimina! eco-
nomic activity, apparent displacement of criminal activity may not generalize to set-
tings in which noncriminai activity remains constant. To assess whether empirically
this is an important source of bias, we examine three pieces of evidence, all of which
are discussed in detail in Section VI. First, we provide evidence that some noncrim-
inai activities—namely, traffic patterns—are not significantly displaced by weather
conditions. Second, we compare lhe impact of lagged crime on indoor and outdoor
crimes separately. Third, we explore whether the relationship between the victim and
offender of violent crime is associated with the magnitude of displacement. Taken
together, these three pieces of evidence indicate that our finding of temporal dis-
placement is unlikely to be driven only by noneriminal economic activity.
Finally, in interpreting our estimates it is important to realize that our strategy
yields a particular local average treatment effect. LATE (Imbens and Angrist
1994). The estimates presented bere reflect the impact of an exogenous increase
in those crimes that are elastic to weather conditions. We will provide evidence in
Section VI that weather affects a broad range of criminal behaviors. It is possible,
however, that the set of criminals whose behavior is affected by the weather is
different from the set of criminals who vary their behavior in response to transitory
law enforcement activity. Thus, our findings might not fully generalize to other
contexts.

IV. Data
Crime data are from the FBI's National Incident Based Reporting
System (NIBRS), which was established as a voluntary crime reporting system in
January 1989. The unit of observation in NIBRS is each criminal incident reported
to police. NIBRS is organized into a system of files that gives information regarding
type of offense, time, date, and location ofthe incident, characteristics ofthe victims
and offenders, as well as the value and amount of any stolen property or illegal drugs.
 Jacob. Lefgren, and Moretti 499

The massive detail of the database allows us to aggregate data into categories and
time periods of our chtwsing. Unfortunately, many of the largest jurisdictions do
not to participate in NIBRS. The 2001 data file contains information on over three
million incidents reported in 3.611 different jurisdictions including Austin. Texas,
and Cincinnati. Ohio. Most jurisdictions in the sample, however, are very small with
more than 70 percent reporting fewer than 500 crimes for the whole year. Earlier
sample years contain information on even fewer juri,sdictions.
Because the most severe crimes occur infrequently in the jurisdictions that we ob-
serve, we focus our analysis on broad aggregates such as violent and property crime.
Violent crimes include simple and aggravated assault, intimidation, homicide, man-
slaughter, and sex crimes. Property crimes include extortion, counterfeiting, fraud,
larceny, vehicle theft, robbery, and stolen property offenses.'** Data on weather are
from the National Climatic Data Center (NCDC). They contain daily readings of
minimum and maximum temperature, and inches of precipitation for 24,833 weather
stations in the United States. We average these weather measures across stations
within each county to construct a county-level panel of daily weather conditions.
As discussed in Section IV, we aggregate the daily data lor both crime and weather
to the weekly level in order to reduce any bias due to the autocorrelation of
weather.'"^ Our primary analysis sample includes 116 jurisdictions for the period
1995-2001 with a total of 26.338 jurisdiction-week observations. Tn order to maxi-
mize the statistical power of our estimates, we have cho.sen the largest jurisdictions
for which NIBRS data is available. While NIBRS is by no means a nationally rep-
resentative sample of police jurisdictions in the United States, and the largest juris-
dictions in the country tend not to participate, our sample does include a relatively
diverse set of cities and counties. (For a complete list of the jurisdictions included
in our sample, see Table 6.)~"
Because some jurisdictions are larger than others, it is helpful to normalize the
crime rate across jurisdictions to reduce heteroskedasticity. This specification also
allows weather to have the same percentafie effect on the crime rate—regardless
of jurisdiction size. Researchers often address these concerns by using the natural
log of the crime rate. This specification is not well suited for the current paper be-
cause, when examining specific types of crimes, it is sometimes the case that no

18. Robbery is included as a propeny erime because the underlying motive is financial. Vandalism is ex-
cluded for tbe same reason.
19. In .some jurisdictions, multiple weeks in a single montb contain noeriminalactivity due tomisreporting
by tiic police agency. In these cases, we restrict the sample in following manner. We drop jurisdiction-
monih observations in whicb the monthly crime rate is more (ban twice the jurisdiction-specific interquar-
lik range below the median. We use the monthly crime rate because in some jurisdictions, it might be the
case tbat no crime occurred over several day.s or weeks. By using a longer time periixl. we are more con-
lidcnt (hat we are excluding observations with gross underreporling. We choose lo restrict the sample on the
basis of median and interquartile range becau.se these measures are less sensitive to periodic massive under-
reporting than are the mean and standard deviation. However, if monthly crime is normally distribuied. our
exclusion re.strittion corresponds to months in which criminal activity is more ihan 2.67 standard deviations
below the mean.
20. For example, the jurisdietions in our sampte span t7 states, witb 3 percent of jurisdictions from the
Northeast. Til percent from the Midwe.st. 63 perceni from the South, and 18 percent from tbe West, The
live jurisdictions in our sample witb the largesi number of crimes reported areChatlan(X)ga. Tenn.: Cincinnati.
Ob. Austin. Tex; Nashville, Tenn.; and Memphis. Tenn. Tbe jurisdictions with the live smallest nuniher of
crime reported are R<K:k Hill. S.C; towa City. Ia,; Burlington. Vt.; Riley County. Kan.; and PtKatello. Id,
500 The Journal of Human Resources

crimes of a particular type occur during the course of a given week. Our large sample
size and massive number of covariates complicate the use of count models. Thus, in
our primary specifications our measure of criminal activity is the number of crimes
committed during the week divided by the average weekly incidence in the jurisdic-
tion during the sample period.
In our sample, the mean weekly temperature in our sample is 58 degrees Fahren-
heit, in the average week, there is only 0.11 inches of daily precipitation. In the 18
percent of weeks that experience some snowfall, there is an average of 0.40 daily
inches of snowfall. Note that a number of jurisdictions appeared to be somewhat in-
consistent in reporting snowfall. For this reason, we include only measures of tem-
perature and precipitation in our preferred specifications, though our estimates are
robust to the inclusion and use of snowfall.

V. Empirical Findings
A. Graphical Relationship between Lagged and Current Crime
Figures 2a and 2b illustrate the baseline serial correlation of crime in our data. To do so,
we regress the violent or property crime rate in Period t on ten lags of crime within

0.1

0.08
o
U
ja 1).U6

If 0.04

¥ 0,02

H 1 I ' I
2 3 4 5 fi ' 9 I)

-0.02
Lag (in weeks)

Figure 2a
The Estimated Relationship between Current and Lagged Violent Crime
Notes: Each panel figure shows the coefficient estimates from a regression of current violent crime on
10 lags of violent crime. The unit of observation is a jurisdiction-week. All crime rates are computed
by dividing the actual number of crimes during a week by the average number of weekly crimes
within Ihe jurisdiction. Other covariates include average temperature and total precipitation in tbe
current period; jurisdiction-year fixed effects, jurisdiction-specific fourth order polynomials in day-
of-year. and fixed effects for month. Standard errors in parenthesis are clustered at tbe state*year
*month level. Observations are weighted by the mean number of all crimes witbin tbe jurisdiction.
 Jacob. Lefgren, and Moretti 501

9 I

Lag (in weeks)

Figure 2b
The Estimated Relationship between Current and Lagged Property Crime
Notes: Each panel figure shows the coefticieni estimales from a regression of current property crime
on 10 lagsot property crime. The utiil of obser\alion is a jurisdiction-week. All crime rates are com-
puted by dividing the actual number of crimes during a week by the average number of weekly
crimes wilhin Ihe jurisdiclion. Other covariates include average temperature and imal precipita-
tion in the current period; jurisdiction-year fixed effects, jurisdiction-specific fourth order polyno-
mials in day-of-year, and fixed effects for month. Standard errors in parenthesis are clustered at
the state*year*month level. Observations are weighted by the mean number of all crimes within
the jurisdiction. ,

the same jniisdiction, controlling for the same set of covariates that will be used in
our primary estitnation—namely, jurisdiction*year fixed effects, jurisdiction-specific
fourth order polynomials in day of year as a smooth control for seasonality. month
fixed effects, and average temperature and total precipitation in Period / as controls
for current weather conditions. Standard errors are clustered by state-year-month to
account for spatial and temporal correlation. The figures present the coefficients on
all lagged crime variables. Figures 2a and 2b differ from Figure 1 in that it includes a
detailed set of controls."'
The top panel shows that violence during each of the past five weeks is a statisti-
cally significant predictor of violence in the current period. The coefficients on the
lags start at 0.065 and generally fall with distance from the reference week. By Week
6. the coefficients are statistically insignificant. In the bottotn panel, we see that the

21. On average, about 93 violent crimes are reported each week in our jurisdictions; ^b of these are simple
assauli (for example, violence noi involving a weapon or serious injury), and 17 are aggravated assault. In
iiboui 70 percent of cases, the viciim knows ihe offender—indeed about 23 percent of all violence is K--
iween family mtmhers. Propeily crime is more common than violent crime wiih nearly 240 reported inci-
dcTils per week. About 60 percenl of ihese incidents involve larceny of some type. Other types of property
crime are far less common. All of Ihese means are calculated weighting each jurisdiction by the average
number of weekly crimes in the jurisdiction.
502 The Journal of Human Resources

serial correlation for property crime is even higher than that for violent crime (for
example, the coefficient on the first lag is 0.22), but the autocorrelation appears to
decay in a similar way. While these conditional correlations are statistically signif-
icant, it is interesting to note that they are smaller than the unconditional correlations
presented in Figure I. This suggests that factors associated with jurisdiction-specific
time and jurisdiction-specific seasonality effects (for example, factors such as sea-
sonal changes in the economy, or police interventions) have a strong influence on
crime rates.

B, The Effect of Weather on Crime

Table 1 examines the relationship between weather and violent as well as property
crime using the baseline set of controls described above. The dependent variable
here is the number of incident.s in a jurisdiction-week divided by the average number
of weekly incidents in that jurisdiction over the entire sample period, so that the
coefficients on the explanatory variables can be interpreted roughly as a percent
change in the outcome. Looking first at Column I, we see that weather—particularly
tetnperature—is strongly correlated with violent crime. A ten degree increase in the
average weekly temperature is correlated with about a 5 percent increase in violent
criminal activity. Precipitation, on the other hand, is associated with reductions in
criminal activity. An increa.se in average v^'cckly precipitation of one inch is a.ssoci-
ated with a 10 percent reduction in violence. These effects are highly statistically
significant—the /^-statistic of joint significance is over 200.^"
We see similar patterns for property crime, although weather appears to be less
predictive of property than violent crime. As the average weekly temperature rises
by 10 degrees, property crimes fall by about 3 percent. The coefficient on precipita-
tion is not .statistically significant. The F-statistic of joint significance for all weather
variables is 51.
Columns 2-10 show the effect of temperature and precipitation on a variety of dif-
ferent types of crimes. This not only provides additional insight regarding the overall
weather-crime relationship but. perhaps more importantly, allows one to better inter-
pret the local average treatment effect ofthe estimates presented below. The results
indicate that the effect of weather is quite consistent across all types of violent crime.
Weather has a similar effect on domestic violence and violence against strangers;
across crimes of varying levels of .seriousness; regardless of whether a weapon
was used; and for violent crimes involving juvenile as well as adult offenders. The
results for various types of property crimes are roughly similar—higher temperature
is always associated with more property crime—although the effect of precipitation
varies more for property than violent crime."

22. Tbere is a convex relationship between temperature and crime—very hoi temperatures re.'iult in a more
than proponional increase in violence—but a concave relationship tietween precipitation and violence. (.See
Table 3 in Jacob. Lefgren. and Moretti 2C)04,)
23. Higher precipitation appears to be associated with increases in the incidence of burglary. We speculate
that precipitation may reduce the probability of detection, perhaps because there are fewer people arouttd to
serve as potential witnesses or tbere is reduced visibility.
 Jacob, Lefgren, and Moretll 503

— — r - tf-j
• ^ r ^ i n r^i
u-i O — O
.-. p r-i
o o o o -o o d

* #
yD O d *—•
-S .5 5 3 — 00 fN <n
p p
2 ^ 4O •^ — — p
d d d d d d

^O O (^-l
p q 5 30 p Tf
d c d K d d

c * o ON — — 00
" D. iO ^P r t —
"5 Ii « ! 10 o —' p r- p r-i
d d d £

•c O ^ ON U-1
S — rt Q
cs I p p ON ON

•^ u =^ d d d d

m — p
o o o o 00 o_ o
O ^ £ 1 r^ d d
*
*
o
f—• 1—1
1 H

ON
OC
o o <n o
o o o o tN o

*
* *
o , ,
.00
.02
.47
.17

s
0 0 0 0
a.
o
x:
P 9
d d

* —
-s:
I1
bo <
— o c
p p "*
00 o d
— ri

"^ '—'

oo 00 r-'
r- (N (y.
•* O 6
'OP'*.
o o o o god

s
3 -i
2 2
sr I
504 The Journal of Human Resources

8p p
a. o O

p P
O d

Cu

go
m d d

a. p p
o
C/3
do
•s •=

O O ^ •—•tn
(N — O Q "*
p p in p 1/^
d d o^ d d

o o —o o

1^
3 -S
 Jacob, Lefgren, and Moretii 505

0.05

O.IM

0.0.1

[),02

0.01

-0.01

-0,02

-0.03
Week
Pereem Violeni Cnme above Normal
Degrees above Ntirmat

Figure 3a
Heat Waves and Violent Crime
Notes: For each panel, we regressed temperature on month fixed effects. Jurisdiction*year fixed
effects and jurisdiction-specific fourth order polynomials in day of year. We included periods in
our sample in which !he temperature in week 0 was more ihan 6 degrees Fahrenheit above predicted
(using our controls) and the temperature in Week I wiis within 3 degrees of predicted. For each week,
we show ihe average temperature residual—weighting each jurisdiction by the average number of
crimes committed during a week. We also show the average violent crime residual, which is obtained
hy running a regressions with controls discussed above.

C. The Effect of Heat Waves on Crime

Here we illustrate the intuilion behind our instrumental variable strategy using an ex-
ample of extreme weather conditions—heat waves. We identify a set of unusually hot
weeks that were followed by relatively normal weather."'' If displacement occurs, we
should see relatively high crime during ihe hot week and relatively low crime during
Ihe subsequent weeks.
Figure 3a shows the results for violent crime. The solid line shows average tem-
perature during the hot week {Time 0) along with the temperature during the subse-
quent ten weeks. The dashed line shows how the average deviation of the violent
crime rate from the predicted rate. In Week 0, both temperature and violeni crime
are higher than expected. Indeed, the violent crime rate is 4.5 percentage points
higher than normal. The following week, temperature is close to its predicted value.

24. We define unusually hoi weeks as those in which ihe average weekly temperature is 6 degrees Fahrenheit
warmer ihan predicted after conirolling for jurisdiclion-year fixed effects and jurisdiciion-specific fourth
order polynomials in day of year. To he included in our sample, the temperature of the following week must
be within three degrees of the predicted lemperature. There are 1,1 OS such week.s in our sanfiple.
506 The Journal of Human Resources

0.04

-0.04
Week
PerL-ent F^operty Crime aht)ve Normal
Degrees above Nomial

Figure 3b
Heat Waves and Property Crime
Notes: For each panel, we regressed temperature on month lixcd etTects, jurisdicti()n*year fixed
effects and jurisdiclion-specific fourth order polynomials in day of year. We included periods in
our sample in which ihe temperature in week 0 was more than 6 degrees Fahrenheit above predicted
(using our controls) and the temperature in Week 1 was witliin three degrees of predicted. For each
week, we show the average temperature residual—weighting each jurisdiction by the average number
of crimes eommitted during a week. We also show the average property crime residual, which is
obtained by running a regressions with controls discussed above.

Violent crime, however, is nearly 2 percentage points lower than normal, suggesting
displacement of about 40 percent over ihe subsequent week. Over the next ten weeks,
both temperature and the rate of violent crime bounce around the predicted level,
though violent crime is unusually high eight weeks after the hot week. Figure 3b
shows the analogous results for property crime. During the initial hot week, property
crime is more than 2 percentage points higher than expected. The following week it
is about 0.5 percentage points lower. This is again consistent with displacement,
though the drop in Week 1 is smaller than for violent crime, in the subsequent weeks,
we again see a small amount of variation around the predicted crime rate.

D. IV Estimates of the Impact of Lagged Crime on

Current Criminal Activity
The analysis of heat waves—an extreme weather shock—casts doubt
on the positive serial correlation of crime that one observes in the raw data. Indeed,
these results indicate that, over a short-time horizon, exogenous increases in crime
 Jacob. Lefgren, and Moretti 507

may be followed by decreases in crime—suggesting temporal displacement in criminal

activity. To more tomially examine this for the full data sample, Tables 2 and 3 pre-
sent OLS and IV estimates of the relationship between lagged and current crime. By
way of reminder, ihe first and second stage specifications are given by Equations 11
and 12 respectively. As explained above, all models include Jurisdiction*year Mxed
effects, month effects and the jurisdiction-specific fourth-order polynomials in
day-of-year to control for seasonality. In order to account for the persistence of
weather over time, they also control for current weather conditions including the
weekly average of daily mean temperature, inches of precipitation. Our instruments
are lagged average temperature and total precipitation."^ To take into account that the
error terms are not independent across jurisdictions or over time, we cluster the stan-
dard errors at the state*year*month level."''
Looking first at the violent crime results in Table 2. the OLS estimates show the
strong positive correlation documented in Figures 2a and 2b. However, when we in-
strument for lagged violent crime using lagged weather conditions. Ihe results are
acttially reversed. The IVestimale in Column 2, for example, indicates that a 10 per-
cent increase in criminal activity in one week is associated with a 2.6 percent de-
crease the following week. Note that the first stage F-statistic is 213. indicating
that our instruments are quite strong (which is also reflected in the precision of
our estimates). In Columns 4, 6. and S. we see that the effects of more distant lags
are generally smaller than the first lag. though mo.st estimates remain statistically sig-
nificant and negative. The sum of the lags provides a measure of the total displace-
ment over an extended period of time. For example, in Column 8. the sum of the four
lags is -0,536. indicating thai a 10 percent increase in violent crime during a partic-
ular week is associated with a reduction of roughly 5.4 percent over a one-month
period—roughly double our estimate for the one-week period. Note that the magni-
tude of the implied displacement is quite large. For example, these results suggest
that the actual impact of a violent crime-prevention program is less thiiii half lUe
magnitude of its contemporaneous impact. Note that these results are stalisticalty
significantly different from zero as well as one (which would indicate complete dis-
placement).
The results for property crime in Table 3 reveal a similar pattern. In stark contrast
to the positive correlations documented in OLS. the IV estimates suggest that lagged
crime has a statistically significant negative eftect on crime in the current period. The
estimate in Column 2. for example, indicates that a 10 percent increase in property
crime in one week will lead to a 2,0 percent decline in property crime the follow-
ing week. Note that, like in violent crime, the IV estimates arc not only statistically
significantly different than zero, but also statistically significantly different from
the OLS estimates. In all models, the sum of the lags is negative and statistically

25, We do not use snowfall due to concerns about the consistency of data collection. (.)ur estimates arc
robust Hi the inclusion of snowfall measures.
26. In theiiry. one limitalion of this type of clustering is that we allow for arbitrary auiocorrelation beiween
two weeks in the same month, siate. and year, but not between Iwd weeks in different months, even if they
are consecutive weeks. We assume thai this is not a major problem in our context. We have expcrimenied
with different clusledng, including ones that include shorter and longer periods, and found our standard
errors to be robu.si.
508 The Journal of Human Resources

q — p
OCiocio OO

5 o o o o o o o
o o d o d o o o

•a
o *
•E CT^ v^ C^ O^
^ q q q
o o o o

•c
u rj
[^
O r- 00
Q ^ O
rj
f i
o o o o o
d d d d d
>
O^ ^ ^ rN C"!
O lo r- w-i
rj q — q
d d d d

r- o^ fN 00
q qq w~.
t^
q qQ
d d d d

r^ q
d d
•S-

*
m O
00 —
q q
d d

r.1
•c c
u u u u
 Jacob. Lefgren. and Moretti 509

q
O O^ d •3 *

*
* --^
NO O c>5 [/;
00 —I OO
—q
d d O "-1

-C- oo O c^ wi in V)
S r- A 00 .!> .0) .oj

Sj -a -5 =
o o — o
I •—OC ' - ' .2 s -j -a

t u n

— q >-
d d

o o
— oo _
^ s =^
<N
io 1/: 1/1

d d r-i 9d rM
(111

c a
C
C3
£ .£ 1 u
p
— q
ent!

d d o Q
.a
*
* —.
O m
sD <n 00
rj q lo
r*^ in 1/1 W5 1/1
gu'§
^ ^ ^ ^
d d ro
_ : =o
1 rM f^-

00 —
o o
1-1
E u *

_ y 5 o
E ,CJ

TVI
D -t! -a
c v: u ^ (J 2 ^
•n o -c O "> JJ •? u
510 The Journal of Human Resources

r o O f o c X o r l O
— p — o q q o q r n f N
oododdodcio

Q\ O ^ f^^ d r^ ^ 00 O^ 'O
(NO — q q q q p - ^ q
d d d d d d d d d d

-o
o
•c
a. > o ^
— q — q q q
c ^
[^ 00 m 00 — oc
_c d d d d d d d d
u w w w

U
E ni o m o OS —
o rn q
5 o
fvi q _ o q q
d d d d d d

I 3
o o r-
!!: o Hi g R —

d d d d d a>
^ 1
I *
*
*
o *I N 0 0 OO
fN o
o o
d d d d do
» *
*
o oc
so 00
(N
o

r
CN q (N q
d d d d

"fc^

:^
I
u a
E B
OLS

3 •c
U
•c
U
•c
u
•c
u
I
 Jacob, Lefgren. and Moretti 51

^3 C 1. n
u u o u
3 T3 ' ^ £

ill o ~

•op
alue

rf
'ifhi
ags

c
2 >
o 13
-g
tio

1^

O^ •a
3 u *3
CO c 3
u • r ;

<N

= i^
.iJ J^ .!> ,.<u

c= S U 5 S
ft ^ ^ ^

•a c '5 S

.I 2
t o a 'S
'^ 3 'J e "
g =•
o ,Q. 3
3£ fS6 "3 •

yj i/j i« i/i

- >•>->•>•

6 5 «2
c E 5 *
1
l-s ^ '^ I i
E ^ •- ^

<U O
-I S^ "> Cl. C
- 33
^_
on .C
" X
C ^ S c^ — r. = ..
;3 -D
["3

P 7, " - •=
•f ^ "^ 1 S t^
•.
bji
C
1^
:3
—
.^
to
u
^ "O tJ ^ K
k. ± 3 O CL, Q ;> u ''^ fl>
Z =6 S k, 5
512 The Journal of Human Resources

significantly different than both zero and one. The effects are somewhat smaller for
property than violent crime, but are still substantial. In Column 8. the sum of Ihe four
lagged crime measures is - 0 . 3 3 , which implies that tor property crime, the displace-
ment that occurs over one month is roughly 50 percent more than that which occurs
over one week.
To this point, we have examined up to four lags. However, it is possible that dis-
placement could operate over an even longer time horizon. To examine this possibil-
ity, we calculate IV estimate for both violent and property crime in which we include
up to ten lags and present the results in Figures 4a and 4b. Note that these results are
directly comparable to the OLS estimates shown in Figures 2a and 2b. The top panel
shows the results for violent crime. As we saw earlier, the first, second, and fourth
lags are statistically significantly different from zero. The others hover around zero
but are not consistently of one sign or another. At -0.37, the sum of the ten lags is
negative and statistically significant at the 10 percent level, suggesting that all of
the displacement in violent crime occurs within a month. The bottom panel shows
the results for property crime. In contrast to the results we found earlier, none of the
lags is statistically different from zero. Furthermore, the sum of lags is only -0.14
and statistically insignificant. Though the point estimates suggest that little displace-
ment occurs over 10 weeks for property crime, the standard error on the sum of co-
efficients is too large to rule out substantial displacement. For both violent and

1
1

1
5 ] ;)

Lag (in weeks)

Figure 4a
IV Estimates of the Relationship between Current and Lagged Violent Crime
Notes: Panels show the (V estimates from a regression of current violent crime on 10 lags. The unit of
observation is a jurisdiction-week. All crime rates are compuled by dividing the actual number of
crimes during a week by ihe average number of weekly crimes within the jurisdiction. Other cova-
riates include average temperature and total precipitation in the current period: jurisdiciion-year lixeii
effects, jurisdiction-specific fourth order polynomials in day-of-year, and fixed effects for month.
Lagged crime is instrumented using lagged average temperature and total precipitation. See the text
for more details. Standard errors in parenthesis are clustered at the state*year*month level. Observa-
tions are weighted by the mean number of all crimes within the Juri.sdiction.
 Jacoh, Lefgren, and Morelli 513

0.3

•€ 0.2
IC
B
O
u 0-1
•£ 1

* "5 0
1
| | •
1)

1"
•
•

-O.I
• 1

a -0.2

1 -0.3

-0.4
Lag (in weeks)

Figure 4b
IV Estimates of the Relationship between Current and Lagged Property Crime
Notes: Panels show the IV estimates from a regression of curretit property crime oti 10 lags. The unit
of observalicn is a jurisdiction-week. All crime rates are computed hy dividing the actual nutnber of
crimes duritig a week by the average number of weekly crimes withiti the jurisdiction. Other cova-
riates include average temperature and lotal precipitation in Ihe current period, jurisdiction-year fixed
effccis. jurisdiction-specihc fourth order polynomials in day-of-year, and fixed effects for month.
Lagged crime is instrumented using lagged average temperature and total precipitation. See the text
for more details. Standard errors in parenthesis are clustered at the state*year*month level. Observa-
tions are weighted by the mean number of all crimes within the jurisdiction.

property crime, however, the results do stand in stark contrast to the positive OLS
estimates.

E, Estimates by Specific Crime Categories i

It is of great interest to establish how our results vary by type of offense. This is cotn-
plicated. however, by the fact that temporal displacement may occur across different
types of crime. For example, an aggravated assault that is prevented by adverse
weather conditions may result in a simple assault in a later period." For this reason,
a regression of simple assault on lagged simple assault may violate the assumptions
necessary for valid instrumental variables identification. In particular, tbe instru-
ments (lagged temperature and precipitation) tnay influence simple assault in the cur-
rent period not only through its effect on lagged simple assault, but also through its
effect on lagged aggravated assault or some other type of related crime in the prior
period. To overcome this issue, when examining the effect of lagged critne on par-
ticular types of crime, we estimate models in which the left-hand-side variable is the

27, In practice, the difference between simple and aggravated assault tan be fairly small. Indeed, relatively
minor differences in the seriousness of injury can lead to a dilTerent categorization.
514 The Jounial of Human Resources

rale of the specific type of crime under examination while the right-hand-side vari-
able is a lagged measure of all violent or property crime.""
Table 4 shows the findings for violent crime. The results suggest that an exogenous
increase in violent crime leads to subsequent reductions in most violent crime cate-
gories. In particular, a 10 percent increase in all violent crime reduces simple and
aggravated assaults by 3.5 and 2.9 percent respectively. Similarly, a U) percent
Increase in all violent crime reduces violent crime against family members and indi-
viduals known to the offender by nearly 3.0 percent over one week. We see similar
effects for crimes with and without weapons. The IVe.stimate for violent crime by
strangers is not statistically significant, though the point estimate suggests some dis-
placement.~ !n all cases, when examining a four-lag model, we observe substantial
displacement for all types of violent crime.^" While our specification allows us to
examine whether or not displacement occurs for particular crimes, it is not possible
to compare tbe magnitudes of the coefficients across crime types. Still, these models
allow us to conclude that temporal displacement of violence appears to operate for a
variety of different types of violent crime.
In Table 5, we examine the findings for different types of property crime. Remem-
ber that the simple labor supply framework outlined in Section II predicts that dis-
placement should only operate for property crimes involving a fairly large monetary
value, since the predicted income effect of stealing small items (for example, a candy
bar) is likely trivial. Consistent with that prediction, the point estimates for the IV
results suggest substantial displacement only for burglary and vehicle theft, although
only Ihe effects for vehicle theft are statistically significant. Interestingly, the 2SLS
estimates for vehicle theft for four lags show enormous displacement. Because many
property crimes—particularly those in the categories of larceny, shoplifting or
robbery—involve relatively small amounts of money, it is not surprising that we find
little effect for these crimes.
To more accurately capture property crime displacement, we estimate a model in
which we measure property crime by the total value of the property stolen during a
particular period.^' Interestingly, the results in Column 7 indicate statistically sig-
nificant displacement over a one-week period. A 10 percent increase in the value
of property stolen in one week is associated with a decline in value of property stolen
in the following week by nearly 6 percent. The results for the four-lag model are not
precise enough to be informative. Overall, these results for property crime suggest
some displacement over a short time period, although the results are not as robust
as for violent crime.

28. Note thai these models assume ihal there is no displacement from violeni to property crime, or vice
versa. While ihis is probably not strictly true, it is likely that the magnitude of this type of displacement
is second-order. Estimates from models where the right-hand-side is the specific crime under consideration
are available upon request. In general. Ihey are qualitatively similar to the ones shown here.
29. We expeci gang violcnt-c might be disproponionately coneentrated in the category of violent crime by
strangers. To the exteni thai one expects the violent crime displacemeni to refleci gang activity, the insig-
nificance of the displacement result t'nr violence committed by strangers is a bit surprising. This could re-
flect, however, the imprecision of the point estimates.
30. These estimates arc computed from separate regressions in which we incltjde one or lour lags.
31. As with the other models, we normalize by diving by the average for the jurisdiction over the entire
sample period. We obtain comparahle results if we use the log of the total value of stolen property.
 Jacob, Lefgren, and Moretti 515

o q q
d d
— rn
d d
— q
d d ^ >

§Is a

010

.272

209

506
165

Yes
094

Yes
066
(N

> U '5 ^ o o o o
o
o o o o
* *

036**

268**
*

033)
ll
017)

(0.107)

(0.244)
Yes
Yes
o o o d o d
>u
090**

219**
*
5

212)
029)
015)

104)
021

Yes
Yes
.2 I o o o o o o d
> "c i3

498**
080**

280**

180**
c 03
063)

022)
Oil)

140)
=3

Yes
Yes
U
o o o o o o o o

I CQ
^ o — (N
o o OO 00

I ^ .£ .« *" (N q
d d
— o
o o

u
c

lis
CO <
00
q q
d d
—

o o
S (NCN
<N q d d
d d

a *
00 —
(N q — q
_ u q q
d d d d d d
d d

« O
I
t/1
516 The Journai of Human Resources

c u t: o
c u =•
E i:!

S 2

i/1 [/I •i I
-i i/i ^ 2
•2 r^ ^ I c ,2

I ££ a 1/5 l/l • ^ •=

" . - | XI
>,

III

is the
that
ou
E
•3 2 i « Sg_^
•c OE I
3 .o
•S'^'BJi
^ U ^

B
•c .?s ^
u C ^= C Ml
> 'C ^ ^
u £
e 1
is.- •5 S
— »: .ii C
v: DTJ
.2 E I S nil-
T3
> 'S [1. 5 5I g

*• OJ >» C
'-- E. — 3
Is O > . «J O

Xn s s E ^-
00
^ s= = ^ = u

O §
^ o c P
ssau

"a.
i- E !>•? u o
« c I- u
lent

o •E ^I
c 5 i-'S •.s :-5
u
i cor
;sdi
lenti

>
_i
a- a •£
u
cri

•a
.d
•s
n -^
O
Si 3
(J — ^ ir; w ra "'

•u 'B a is 5 Cl.,

I I I I 4.1
T po

6.=o
ear
ef

I O 'C O
•3 3
i^ Oj U ;r Q
 Jacob. Lefgren, and Morelti 517

0.

•a X)

0 0 0 0
S o

* — —.
O w"i (^ r-
(^ rT) ^ '•C
(N O O f^ mp p ^
o c; d o d d d d d

— f:n 2; 00
fN p u~, rj 'J; p p Tt
d d d d d d —d
o

— S 3 — '^ '^
00 • * fn p rj —
d d d d d d d d
u

•c sO io r ; tN
r^, f. O >ri
— q ' r'j p
p <N rn
O

00 — ^
p p
rJ OO Tt q — (N
d d d d d d d d

r- — O 00
tN p ri p •jj; p rn
d d d d d d d

*I
O
11
518 The Journal of Human Resources

•O JO
U 'C

E
i3 'i
w- •a
"p
Q.
•y. u
J= <T1
X) 5 U

o 5
t-
• ~
on 2

S •= c i i

5 iJ r ^ c
> H CQ Oil ^ — EA

'C cm ,p o .D
S- J5 "• •= O

p In
_ S

CQ
1
U E y •- =5
.? 5 y «

o
r? b « t: s;
s •£ £ S ii

I ^ T3 -g " O g^

•B E 5 - fc •S

ii K~
k- 3 OJ 4j ^'J tn
•E ;., ^ 5 g ^

f-3
^•2

I CJ
<U
C
O

I
m
a. c
5 Og CL
11
.^ E
I I -f «
~ ^II^ '
it c
3 o
 Jacob, Lefgren. and Moretti 519

VI. Robustness Checks

A. Temporal Displacement of Noncriminal Economic Activity
The main identifying assumption in our empirical strategy is that lagged weather
conditions only influence current period crime through their influence on lagged
crime. While high frequency variation in weather is unlikely to be correlated with
many of the unobserved factors that determine the persistence of crime over time
(for example, income levels, crime-prevention policies, etc.). weather may affect
the intensity of noncriminal activity which, in turn, could influence the cost and/or
benefit of crime. If inclement weather causes people to stay home in one period,
for example, it tiiay result In greater than expected economic activity in the following
period, which could increase the benefits of crime {by increasing the availability of
victims, for example). More generally, if weather displaces noncriminal activity and
this activity influences the cost/benetit of criminal activity, our estimates tnay be biased.
To assess tbe empirical importance of this concem, we first provide evidence re-
garding the extent to which noncriminal activity is displaced by weather conditions.
Though few measures of economic activity are reported at a sufficiently high fre-
quency to examine this issue, we have collected information from the Federal High-
way Admini.stration (FHWA) that includes dally measures of traffic from in-road
tiionitors in over 20.000 locations throughout the United States for 2000-2001.
We regard traffic as a good summary measure for noncriminal economic activity. "
We aggregate this data to the county-week level, so that our primary outcome mea-
sure is the number of vehicles counted in a particular county during a given week.
Using data on weekly vehicle traffic for states represented in our sample, we find
ihat weather conditions have only a small correlation with vehicle traffic. This is ex-
tremely informative itself because if weather does not have a strong impact on this
type of economic activity over a one-week period, it is unlikely that weather would
induce substantial amounts of displacement to bias our estimates. Table 7 shows that
when we instrument lagged traffic with lagged weather conditions, we find no evi-
dence of temporal displacement over one or four weeks, though the standard errors
are in some case.s quite large." This is also inconsistent with widespread temporal
displacement of economic activity.

32. Tbe traffic volume daia includes hourly traffic counts for each traffic station provided by permanent in-
road traffic monitors. Geographic identifiers allow one to link each station to states and counties. Traffic
data arc available for 66 of the 92 counties included in the crime analysis. Recall that our primary analysis
sample for crime includes 116 jurisdictions in *)2 counties in 17 states. For more infonnation on the traffic
data, see: Office of Higbway Policy Information at http://www.fhwa.dot.gov/policy/ohpi.
33. To be consistent wiih the crime estimates, we normalize these measures by dividing weekly counu by
ihf average for that counly over tbe sample periixJ.
34. These results are reassuring, although ideally, one would like lo look at a broader set of measures of
economic activity to draw more general conclusions. A recent paper looks at the effect of weather on retail
sales (Starr-McCluer 2(XM)). The author finds some evidence o!" displacement. tJnusually cold tem[>eratures
in a given month lend to depress sales in ihat month, but they lead to higher sales in the following month.
The current and lugged effects do not completely offset each other. Moreover, unusually warm weather is a
given month increa.ses sales, but reduces sales in the month after. The current and lagged effects roughly
offset each other. Unfortunately, data on retail sales are available only at the monthly level, which makes
the comparison with our analysis difficult.
520 The Journal of Human Resources

ounty
CO X
$ $^ U H <:
Q CO 1^ * •"•
"! sz H
u •d -^ -
i CO

oanoke.

ioux Cil
iley Cm
3

aUna. K
oseville

andy, U
ock Hil

aginaw.
3
re a O c C
o I/:
Si _re E d ,o _re E
>
3
•o
aj
o c
o
o X Xu
i cu 5 1 a: a: t o CO to

D.
n o o o o o o o o o o
CO
o ooooo
fN fN fN fN tN fN
o o o o
f - l fN (N f^J r>l
c I I ! 1 ^ J
I J I I J

O
u X u
o HH •*<'

> o O U 0
33
oh d
: o
aho

nc iq i j j D
a jUi ju gC CE B
h PC : C^
U O O O X X X X X •o O T

a.
C/5 N f N N rjrJfNfN
I I I I I I
c c

•c
 Jacoh. Lefgren. and Moretti 521

r-i N fi <N r-i r4 (N

1 I 1 1 I I I I

u 5 $
u
I j c/j U SB
00 oij.2 l< O C -3

te
C C — •3 C
- -;^ ^
•4_i 'V^ <y:

> ^

J u
z z c
nty,

o £^
ou I/;

OJ H c U
o , H "XI - * Oti
L.
rt
P u= 5 •= O, CQ - ii .n
U
^"^ ^^ 3
'c '
X)
> s g —, CD
ran

>^ X 3 C c o
s aj o >,
_1 HJ -J -J Z Z ooO

u < O
> U u
Q !^ < c/1
z 3 c c
< - ?
Vlen

luff:
Spri

U O H
iy
Cou

VA

H
> T3 X
rfie]
ttes

o ni
bia.

CQ .
c C >
C I'; 'o ^~ S= o
.2 S o 3 2 E £ c >
re <u (jj P b > <U c 3 C
x: .c o o o o o O (5
U U u U U Uu U U U U U U Q
522 The Journal of Human Resources

ododocJdodo
o

*w-i *
— r j — i - g - — r- — m —
i^poppppoooo
d d d d d d d d d d
o
1 ^

OS ( ^
O —
p —
> 3" d d
r-i

00 O
d d
CN

I — (NOfNp — — fNpTf
r^
CN
dddddddddd

* * *
— fO 3^ fN (N f . C-, r-t GO r^,
Oppppppp CN
(N

o o
p (N 3 rj— w-1 C
(^ p
d d d d cr! d
•5-

r-1
-J —
o - O C do
'—' <n
cr-

BA

4.
S 2
o
 Jacob, Lefgren, and Moretti 523

As a second piece of suggestive evidence on temporal displacement in economic

activity, we estimate sepanite models for indoor and outdoor crime. If weather indu-
ces displacement in noncriminal activity, we would expect to see more (less) outdoor
activity after a bad (good) spell of weather. If crime follows economic activity, we
also should expect more outdoor crime following bad weather (and vice versa) but
less indoor crime. This suggests that our displacement results would be focused
on outdoor as opposed to indoor crime. For violent crime, we find that an exogenous
increase in violent crime leads to statistically significant reductions in both outdoor
and indoor violence over a four-week period, although the one-week effect is only
siatlstically significant for indoor crime. The property crime estimates are uniformly
negative and similar in magnitude for indoor and outdoor crime, but only signifi-
cantly different from zero in one of four cases (see Table 10 in Jacob. Lefgren.
and Moretti 2004).
A third suggestive piece of evidence regarding the potential bias from the dis-
placement of economic activity can be obtained by examining the relationship be-
tween offenders and victims for violent crime. It seems plausible that interactions
with family members would be least sensitive to the degree of noncriminal economic
activity. In Table 7, we showed that an exogenous increase in violent crime leads to
statistically significant reductions in the amount of violence against family members
during the following week. Although by no means definitive, taken together these
results suggest that our finding of temporal displacement is not driven by displace-
ment of noncriminal economic activity. ^

B. Endogenous Police Staffing levels

A second concern is that police may departments may change staffing levels in re-
sponse to weather conditions or small fluctuations in the crime rate. For example,
police may know that on hot days there is more crime and ihus increase the number
of working officers on those days. Such behavior amplifies or attenuates the correla-
tion between weather and crime but ultimately acts simply as another cost factor.
Thus the consistency of our IVestimates is unaffected by such behavior.
A more serious concern would be if current staffing levels were correlated to
lagged weather conditions. This would violate the exclusion restriction necessary
for satisfactory identification. Such a situation might exist if police departments ad-
justed staffing levels going forward in response to small fluctuations in crime rates
as.sociated with weather condilions. To reduce concerns on this front, we contacted
several police departments in our sample. The responding agencies indicated thai
they did not adjust staffing levels in response to nonemergency weather conditions.
Furthermore, changes in staffing levels were proactive—not reactive. In other words,
departments might increase staffing levels in response to a parade but not in response
to a temporary and unexpected increase in number of reported offenses. These
responses suggest that the endogeneity of police staffing levels is unlikely to be a
meaningful source of bias.

C. Persistence in Weather
A third potential problem is that weather, when measured at high frequency, is seri-
ally correlated. While we control for current weather conditions in our models, it is
524 The Journal of Human Resources

possible that weather within a period is measured imperfectly, in which case lagged
weather still may provide information regarding the unobserved aspects of current
weather which would, in turn, directly impact current crime. As discussed earlier,
this will violate the assumptions necessary for satisfactory identification and result
in a positive bias in the coefficient on lagged crime. It is possible to show that this
bias decreases as the length ofthe time window expands. For this reason, we aggre-
gate to the weekly level.
To examine whether this sonrce of bias is empirically unimportant in our current
framework, we conduct a '"reverse experiment" where we estimate the effect of
fuinre crime on current crime, and instrument for future crime with future weather.
Specifically, we estimate a specification in which the dependent variable is crime in
Period t, but the right-hand-side variable is crime in Period (+\ and the instrument
is weather conditions in Period f+\. If measurement error in weather conditions is
an important source of bias, the bias would be similar whether we look at lagged
or future crime. Thus, we would expect the estimated impact of future crime on cur-
rent crime to be positive and statistically significant. This is not the case (see Jacob,
Lefgren, and Moretti 2004). In particular, the estimated impact of future crime is
small in magnitude and statistically insignificant for both violent and property
crime. This suggests that our results are not materially affected by imperfect weather
measures.

D. Other Robustness Checks

We have experimented with a variety of alternative specifications. For a full list, see
Tables 10 and 11 in Jacob. Lefgren, and Moretti (2004). We find that our results are
robust to the exclusion of observations with violent crime rates more than twice the
interquartile range above the median wiihin the jurisdiction. This suggests that our
findings are unaffected by periodic over-reporting of criminal activity. We also find
that our results are comparable if we measure crime using the log number of violent
or property crimes rather than scaling by the average number of crimes in the juris-
diction. Results using only temperature as an instrument are nearly identical lo our
baseline specitication. When we limit our instruments to precipitation alone, we have
very little statistical power, particulariy in the case of property crime, although the
four-lag violent crime results are comparable to baseline.
In our primary specification, we implicitly assume symmetry between the effect of
an exogenous reductit>n in criminal activity and an exogenous increase in criminal
activity. However, this need not be the case. For example, individuals prevented from
committing a criminal act may do so in a subsequent period while individuals who
engage in opportunistic crimes due to favorable weather conditions may not reduce
their criminal activity during the following week. To examine the impact of positive
weather shocks, we construct a variable that is the average temperature during a
week if the temperature was above the average in the state during the month and zero
otherwise. We use this as our only instrument in the first stage. To examine the im-
pact of negative shocks, we do the analogous analysis using variation in temperature
below the average. Over a one-week period, the displacement results appear to be
larger for positive shocks than negative shocks. Over four weeks, however, the dis-
placement effects appear roughly consistent across positive and negative shocks.
 Jacob, Lefgren, and Moretti 525

VII. Conclusion
In this paper, we exploit the correlation between weather and crime
to examine the short-run dynamics of criminal activity. In sharp contrast to the po.s-
itive serial correlation in crime rates reported in most studies, we find that higher
crime in one week is followed by less crime in subsequent weeks. These lindings
suggest that the positive serial correlation in crime commonly reported is not an en-
dogenous process driven by the optimization of oftenders, but likely reflects persis-
tence of unobserved factors that influence of criminal activity.
The magnitude of displacemenl in violent and property crime over a short-time
horizon is substantial. A 10 percent increase in violent crime in one week leads to
a 2.6 percent reduction in violent crime the following week. Over the course of four
weeks, over half of the initial increase in violence will be mitigated through displace-
ment. The results for property crime are somewhat smaller—a 10 percent increa.se in
property crime results in a decrease of 2.0 percent the following week—and appear
lo be limited to high value property crimes, vehicle theft in particular.
The propeny crime results suggest that criminal behavior is consistent with a stan-
dard labor supply framework, where the observed displacement can be explained by
an income effect. Criminals who are prevented from committing propeny offenses in
a given week engage in higher levels of criminal activity during the subsequent
weeks to make up for lost income. This behavior is consistent with a model in which
offenders have at least some foresight and/or are some liquidity constraints. Consis-
tent with the predictions of the model, we find that displacement arises only for
crimes that involve highly valuable propeny. such as vehicle Iheft.
In the case of violent crime, a possible interpretation of the results is that the ben-
efits of violence may be durable—^in other words, the marginal utility (cost) of vio-
lence is decreasing (rising) in the amount of violence committed during the prior
week. This would be true if. for example, an assailant who "seules a score" in
one period feels less need to do so in a subsequent period. Similarly, a husband
who abuses his wife in one period may be less inclined to do so in the next period,
perhaps because of a sense of guilt or because he has received a warning from the
police.
Besides its theoietical implications, the uncovering of substantial temporal dis-
placement also may have important policy implications. Our findings suggest that
the long-run impact of temporary crime-prevention efforts may be smaller than the
shon-run effects. In the case of violent crime, the shon-run impact of a one-week

35. Incarceration may in theory be a factor behind the displacement of property and violent crime. H<tw-
L'ver. given the low arresi and conviclion rates, it is unlikely thai incarceration could explain a substantial
ponitin of ihe obser\'ed displacemenl. Consider, for example, that each incident of violence reported pro-
duces only about 0.4 arrests. It is likely that ai least half of ihese are either nol charged with an offense or
released almost immediately on bail (Dilulio 19%). Furtheniiore, even if the remaining individuals are in-
volved in a reported aci of violence 10 limes per year, incarceration could only gener.ile displacement of
about 4 percent over one week, a small traction of the total effect. This is even truer for property crime,
where clearance rates arc much lower \ related possibility is that the cost being caught is an increasing
and convex function of ihc number of crimes committed. Because we are looking at week-io-week variation
in crime rates, we suspect that in our context, this is not empirically very relevant. For most crimes, II is
unlikely that ofienders are caughl, arresled. tried, sentenced, and freed in one week.
526 The Journal of Human Resources

crime-prevention effort will be twice as large as the impact over one month. To the
extent that policy makers engage in short-term policy experiments or anti-crime
crackdowns, standard policy evaluations may prove misleading.
It should be noted, however, that the estimates presented here reflect a particular
local average treatment effect—namely, the impact of lagged crime that is elastic lo
weather conditions. The relevance of our results for policy depends upon the extent
to which the set of criminals and crimes whose behavior is affected by the weather is
similar to the sets which respond to transitory law enforcement activity. Although we
show that weather affects a broad range of criminal behaviors, it is important to rec-
ognize that our findings might not fully generalize to other contexts, particularly
those that involve longer-term interventions such as improving economic opportuni-
ties for potential offenders or short-run interventions that involve the incapacitation
of many potential offenders. It is also important to note that changes in weather con-
ditions have no lasting impact on the perceived costs of criminal activity. This differs
from extended police crackdowns, in which offenders may only gradually realize that
the program has ended. The effects of such programs have been shown in .some ca.ses
to extend even beyond the duration of the crackdown (see Sherman 1990). To the
extent that police engaged in shorter and more predictable crackdowns, our findings
would be more relevant
Finally, it is useful to consider how these results inform the literature on social
interactions. To the extent that social interactions operate at the week-to-week fre-
quency we examine, our estimates should be interpreted as the net effect of these so-
cial interactions and displacement. In this case, our results suggest that the total
temporal displacement more than offsets social interactions operating at the weekly
level. Given the relatively high frequency of our data, our estimates cannot speak to
the possibility of social interactions that operate in the long run through, for example,
social learning. Thus, these findings do not rule out the importance of social interac-
tions for explaining long-run differences in crime rates across k^alities.

References

Anderson. Craig A. 2(X)1, "Heat and Violence." Current Directions in Psychological Science
IO(l):33-38.
Becker. Gary. 1968. "Crime and Punishment: An Economic Approach." Journal of Political
Economy 76(3): 169-217.
Braga. Anthony A.. David M. Kennedy, Elin J. Waring, and Anne M. Piehl. 2001. "Problem-
Oriented Policing. Deterrence and Youlh Violence: An Evaluation of Boston's Operation
Ceasctire." The Journal of Research in Crime ami Delinquency 38:195-225.
Camerer. Colin. Linda Babcock. George Lowenstien. and Richard Thaler. 1997. "Labor
Supply of New York Cily Cab Drivers: One Day at the Time." Quarterly Journal of
Economics 112(2):407-l L
Cohn, Ellen G. 1990. "Weather and Crime." British Journal of Criminology 30(1): 51-64.
Cohn. Ellen G., and Jame>; Rolton. 20(X). "Weather. Seasonal Trends and Property Crimes in
Minneapolis, 1987-88: A Moderator-Variable Time-Series Analysis of Routine Activities."
Journal of Environmental Psychology 2O(3):257-72.
Cook. Philip J.. and Kristin A. Goss. 1996. "A Selective Review of Ihe Social-Contagion
Literature," Working paper. Raleigh: Duke University.
 Jacob, Lefgren, and Morelti 527

Dilulio, John J. ! 9 % . "A New Crime Wave." Investor's Business Daily, March 6.
Farber, Henry. 2005. "Is Tomorrow Another Day? The Labor Supply of New York City Cab
Drivers." Journal of Political Economy 113(l):46-82.
Field. Simon. 1992. "The Effect of Temperature on Crime." British Journal of Criminology
32(3):34O-5!.
Grogger, Jeffrey, and Charles Michalopoulos. 2003. "Welfare Dynatnics under Time Limits."
Journal of Political Economy 11 U3):530-54.
Hesscling, Rene B. R 1994. "Displacement: A Review of the Empirical Literature." In Crime
Prevention Stiulie\. ed. R.V. Clarke. 2. Monsey, N.Y.: Criminal Justice Press.
Jacob. Brian. Lars Lefgren, and Enrico Moreiti. 2004. "The Dynamics of Criminal Behavior:
Evidence tmm Weaiher Shocks." NBER Working Paper No. 10739.
Jacob, Brian, and Lars Lefgren. 2003. "Arc Idle Hands [he Devil's Workshop? lncapacitation.
Concentration and Juvenile Crime." American Ectmtmiic Review 93(.'));I360-77.
Lochner, Lance. 1999. "Education, Work, and Crime: Theory and Evidence." RCER Working
Papers No. 465. New York: University of Rochester.
Rotton, James, and Ellen G. Cohn. 2000. "Violence Is a Curvilinear Function of Temperature
in Dallas: A Replicalion." Jotinuil of Personality and Social P.sychology 78(6): 1074-81.
Sherman. Lawrence W. 1990. "Police Crackdowns: Initial and Residual Deterrence." Crime
ami Justice 12:1-48.
Starr-McCluer, Martha. 2000. The Effects of Weather on Retail Sales. Federal Reserve Board
Discussion Series.