Climate change and victimization risk: A disaggregated look at NCVS data

Abstract

The current study uses pooled National Crime Victimization Survey data (1992–2015) to examine if the relationship between climate change and victimization risk is modified by victim and incident characteristics. Panel analysis yields interesting findings. First, results mirror those found in prior studies utilizing Uniform Crime Report data, providing another indication that the link between a warming climate and crime may be quite robust. Second, the results indicate that climatic effects may play out differently in different contexts. For example, outdoor victimizations, especially those near a person’s residence, appear increasingly elastic during anomalously warm temperatures. In addition, subpopulations (males and African Americans) are also at increased risk of victimization. Our results effectively suggest that at-risk populations are more vulnerable to climatic variability.

Keywords

Climate victimization risk victim characteristics incident characteristics

Introduction

Over the past decade, climate change has been increasingly recognized as a ubiquitous factor in shaping human behavior (see Carleton and Hsiang, 2016 for an overview). At the same time, the empirical relationship between climate change and violence has been uncovered in a variety of settings: from domestic violence in India (Sekhri and Storeygard, 2013) and Australia (Auliciems and DiBartolo, 1995), assaults in Nigeria (Badiora et al., 2017), assaults and homicides in the USA (McDowall and Curtis. 2015; Ranson, 2014) and Tanzania (Miguel, 2005), rapes and violent robberies in China (Hu et al., 2017), ethnic violence within Europe (Anderson et al., 2013), to civil conflicts throughout the world (Bergholt and Lujala, 2012; Hsiang and Burke, 2014; Hsiang et al., 2013). Effectively, crime increases as climates warm (Burke et al., 2009; Burke et al., 2015; Carleton and Hsiang, 2016; Hsiang et al., 2013; Ranson, 2014; Rinderu et al., 2018, and others).

Not surprisingly, with rapidly accumulating evidence of a global climate change, more and more studies focus on the effects of climate change on the rates of violent offending (Anderson and Delisi, 2011; Anderson et al., 1997; Hsiang et al., 2013; Mares, 2013a, 2013b; Mares and Moffett, 2016, 2019; Ranson, 2014; Rotton and Cohn, 2003; Williams et al., 2015). Previous US national-level studies have relied on Uniform Crime Report (UCR) data during their examination of the relationship between climate change and crime (see Ranson, 2014; Rotton and Cohn, 2003; Harp and Karnauskas, 2018; Mares and Moffett, 2019). Nonetheless, the general consensus from this burgeoning literature is that a warming climate results in higher levels of crime.

Though UCR data are well established and possess important strengths, they also come with serious limitations. One of the issues with the UCR is that it only contains information on crimes reported to the police, which makes it possible that results based on UCR data may be the outcome of a reporting bias. Another serious limitation of UCR data is that they contain only summary statistics for index offenses, although Supplementary Homicide Reports fill some gaps for homicides. The inability to disaggregate UCR data makes it difficult to examine contextual causal linkages between climate variables and crime and understand the impact of weather events across criminal contexts.

In the current study we employ National Crime Victimization Survey (NCVS) data from the United States to address the aforementioned concerns and answer whether data disaggregation may help us better understand the climate-crime relationship. To do so, we provide a comparison of UCR and NCVS data and explore several disaggregated crime incident dimensions, including the location of the incident, race, and gender. Results of analysis indicate NCVS data mirror temperature impacts found in UCR data, but, in addition, we find that climate change effects may disproportionally influence individuals already at risk for victimization. The key findings in this study have important implications for policymakers, practitioners, and the public, and may improve predictive, preventative, and reactive policing.

Literature review and theoretical framework

A long tradition of criminological research has produced compelling evidence to support the link between weather and crime (DeFronzo, 1984; Mares, 2013b; Quetelet, 1842; Ranson, 2014; Rotton, 1986; Rotton and Cohn, 2003). Prior studies have examined a broad array of meteorological factors including temperature, sunlight, precipitation, wind, and humidity (Cohn, 1990), although national level research typically limits its focus to the effects of temperature on crime (Anderson and DeLisi, 2011; Gamble and Hess, 2012; Harp and Karnauskas, 2018; Mares and Moffett, 2019; Ranson, 2014). Research typically finds stronger predicted effects for temperature than for other climatic variables, such as precipitation. Some scholars exploring daily data have found evidence to support the effects of precipitation on crime (Horrocks and Menclova, 2011; Hsiang et al., 2013; Jacob et al., 2007). Using monthly levels of aggregation, however, may wash out rarer events such as extreme heat and precipitation (Field, 1992; Mares, 2013a; Ranson, 2014). Robust findings in the literature thus connect higher temperatures with higher rates of offending, more police calls for service, and increases in other categories of aggressive and violent behavior (Lebeau, 1990; Lebeau and Langworthy, 1986; McDowall et al., 2011).

The seasonality of criminal activity is another well-established finding in the criminological literature (e.g. McDowall et al., 2011), and most recent studies examining the link between weather or climate and crime necessarily consider seasonality in some form (e.g. Ranson, 2014). For instance, Hu et al. (2017) examine the relationship between temperature and a broad range of violent and property offenses and found that strongest correlations can be attributed to seasonality. Mares (2013b) found strong positive relationships between crime and expected monthly temperatures. This is further confirmed by direct examinations of seasonality (e.g. Carbone-Lopez and Lauritsen, 2013; McDowall and Curtis, 2015).

A number of studies spanning a wide range of locations and temporal scales have independently uncovered persuasive evidence to support the relationship between temperature and violent crime. Investigations at the scale of individual cities, such as Dallas, TX (Gamble and Hess, 2012); St. Louis, MO (Mares, 2013a; 2013b); Philadelphia, PA (Schinasi and Hamra, 2017); Minna and Benin City, Nigeria (Badiora et., 2017); and Tangshan, China (Hu et al., 2017), uniformly confirm a positive relationship between temperature and crime. At the same time, other studies yield similar findings at a broader scale, either by examining pooled data at the city or county level in the United States (Hipp et al., 2004; Jacob et al., 2007; Ranson, 2014), by analyzing annual, nationally aggregated data for the United States (Anderson et al., 1997; Rotton and Cohn, 2003), or a variety of locations across the globe such as British Columbia, Canada (Linning et al., 2017), New Zealand (Horrocks and Menclova, 2011; Williams et al., 2015), South Africa (Bruederle et al., 2017), Finland (Tilhonen et al., 2017), Malaysia (Habibullah, 2017), England and Wales (Field, 1992), or a number of countries around the world (Mares and Moffett, 2016). A common conclusion is that the relationship strength varies across crime types, with aggravated assaults consistently yielding some of the strongest predicted temperature effects.

Several theories have been offered to explain the apparent relationship between temperature and crime, but most recent studies have converged on two potential causal mechanisms: the Temperature-Aggression Hypothesis and the Routine Activities Theory (Anderson, 1989; Anderson and Anderson, 1998; Anderson et al., 2000; Breeztke and Cohn, 2011; Carbone-Lopez and Lauritsen, 2013; Ceccato, 2005; Cohen and Felson, 1979; Cohn and Rotton, 2003, 2005; Harries et al., 1984; Hipp et al., 2004; Kenrick and MacFarlane, 1986; Lebeau, 1990; Lebeau and Corcoran, 1994; Lebeau and Langworthy, 1986; McDowall et al., 2011; Reifman et al., 1991; Rotton and Cohn, 2000).

The Temperature-Aggression Hypothesis, also known as The General Affective Aggression Model (Allen et al., 2018; Anderson et al., 1996), suggests that aggression increases during periods of hot temperatures because heat increases both aggressive motivation and, under certain conditions, aggressive behavior (Anderson, 2001). Following this logic, a warming climate should result in warmer seasons, which then produce more violent offenses (Anderson et al., 2000). Temperature-Aggression receives support from both laboratory studies (Anderson et al., 2000; Baron and Bell, 1976) and real-world examples, like the increased likelihood of baseball pitchers hitting batters during games played on hotter days (Krenzer and Splan, 2018; Larrick et al., 2011; Reifman et al., 1991) as well as higher incidences of frustration-inspired car honking during higher temperatures (Kenrick and MacFarlane, 1986). Seasonality research, which finds violent incidents peaking during the warmer months of the year, lends further support to the Heat – Aggression Hypothesis (Anderson and DeLisi, 2011; Bruederle et al., 2017).

Routine Activities Theory (Cohen and Felson, 1979) proposes an alternative framework for examining the relationship between weather and criminal behavior by suggesting that in order for a crime occur, the following three elements must converge in space and time: (1) a motivated offender; (2) a suitable target, and (3) the absence of a capable guardian. Cohen and Felson (1979) argue that patterns of routine activities are mirrored by the patterns of criminal opportunity. Routine activities are relatively habitual and make up patterns repeated over time. Any environmental shifts may lead to changes in everyday activities of individuals. Changing weather conditions could affect the likelihood of the convergence of the offender and target (Agnew, 2011). During periods of pleasant weather, individuals are more likely to be outside, mobile, and engaging in social interactions. For this reason, pleasant weather is likely to increase both the number of interpersonal interactions and victim availability (Hipp et al., 2004; Lebeau, 1990; McDowall and Curtis, 2015; McDowall et al., 2011; Rotton and Cohn, 2003).

Most studies give weight to both explanations, but no study has been able to dissect these mechanisms because such a task would necessitate the use of incident-level data, a level at which little generalizable data are produced. Most examinations of the relationship between climate and crime are characterized by some degree of aggregation bias, obfuscating the exact causal mechanisms driving the weather-crime link. One type of aggregation bias has to do with the fact that prior studies focus solely on aggregated crime categories. Theoretically, it stands to reason, for example, that disaggregating data by location of victimization should yield differential results as exposure to heat (heat-aggression), or exposure to potential offenders (routine activities theory) is likely to be different between indoor and outdoor locations.

Another limitation of law enforcement data is that it only includes information on crimes reported to the police. Therefore, it is impossible to separate the weather-crime relationship from reporting effects. Pleasant weather, for example, may very well increase reporting of offenses as more people venture outside and are more likely to witness criminality. By the same token, police may also be more proactive during pleasant weather.

In the current article, we address the aforementioned gaps in the literature by utilizing victimization data recorded in the NCVS. First, we explore similarities between NCVS and UCR data. Given the robust findings reported in the prior literature, results for victimization data should replicate those found in incident data. Second, we use the unique detailed aspects of victimization data to examine if the relationship between temperature and victimization risk is mediated by victim and incident characteristics, and if changes in environmental conditions affect some victims differentially. These results allow us to identify if certain population groups are at a greater risk of environmental change, which may have important policy implications.

Data and methods

Here we use data from the NCVS, a nationally representative household survey.¹ Each year, the NCVS captures a nationally representative sample of about 135,000 household interviews, composed of nearly 225,000 interviews of persons within these households. The NCVS sample design involves a two-stage stratified sampling of housing units and group quarters in 542 sample areas. The first stage consists in the definition, stratification, and selection of primary sampling units (PSUs). First-stage sampling occurs every 10 years. The second stage selects housing units and group quarters within selected PSUs (Bureau of Justice Statistics n.d.).

The NCVS target population is residents age 12 and older living in households or group quarters, such as dormitories, rooming houses, and religious group dwellings, within the United States and the District of Columbia. The NCVS has a longitudinal panel design, where each household selected into the sample is interviewed every 6 months for three and a half years. Each interview is limited to victimizations that occurred within the last 6 months. After the seventh interview the household leaves the panel and a new household rotates into the sample (Bureau of Justice Statistics n.d., National Archive of Criminal Justice Data n.d.).

The NCVS maintains a stable and moderately high response rate at both the household and individual levels. For instance, in 2011, the response rate for households was 90% and the person-level rate was 88% (National Research Council, 2014). Response rates have fluctuated by several percentage points over the period from 1992 to 2015, but not substantially (Bureau of Justice Statistics n.d.).²

For the purposes of the current study, one important strength of the NCVS is that it counts all victimization incidents, including those not reported to the police. The NCVS also has a number of important design features to facilitate examinations of temporal patterns in crime by reducing potential sources of error when measuring the month of occurrence. First, the relatively short 6-month reference period allows for better recall of the incident details and fewer instances of under-reporting of victimization due to fading memory and minimizes the number of cases where an incident would be allocated to the wrong month. Second, repeated interviews make it possible to partially control for potential forward telescoping by ‘bounding’ (using data from prior interviews to establish a reference point and prevent double counting of the same incidents in more than one reference periods). ‘Bounding’ procedure also helps more accurately allocate the month of the incident. Third, NCVS has the ‘rolling reference period’ which allows the collection of data throughout each month of the year. This design feature gives equal weight to the measurement of victimization over all months of the year and by doing so minimizes recency bias errors that may have occurred otherwise (Carbone-Lopez and Lauritsen, 2013).

NCVS data also have limitations for the purposes of assessing temporal or seasonal variations in victimization. First, the NCVS collects self-report data, which introduces a number of potential sources for non-sampling error. One such source is the inability of a respondent to clearly recall the incidents which occurred during the 6- month reference period. Assault is recalled with the least amount of accuracy of any crime measured by the NCVS. This may have to do with the fact that most assaults are committed by offenders who are not strangers to the victim, or if assaultive behaviors may be a part of everyday life, and therefore, not recalled and reported to the interviewer (National Archive of Criminal Justice Data (NACJD), n.d.). Another potential source for non-sampling error is related to the inability of the respondent to accurately recall the exact month when the incident took place. The previously mentioned bounding procedure helps minimize such errors. Other sources of non-sampling error may result from errors in reporting, coding, and processing the data. NCVS implements quality control and editing procedures to minimize the number of errors made by the respondents and the interviewers (NACJD, n.d.). It should be noted, however, that it is possible to assume that the magnitude of non-sampling error is constant from year to year and therefore should not impede the analysis of temporal variations in crime levels (NACJD n.d.).

Temperature and precipitation data are retrieved from the National Oceanic and Oceanographic Administration (NOAA). Admittedly climatic data are difficult to match to NCVS data as the geographic indicators in this data set are unsuitable for linking them to station or gridded climatic data. Nonetheless, because the NCVS is a nationally representative sample, using national level climatic data should provide a reasonable match. This was recently illustrated by Mares and Moffett (2019) who aggregated monthly UCR data by local agencies, metropolitan areas, states, and national levels. Their results show that while coefficient strength is somewhat reduced at higher levels of aggregation (state and national level), direction and significance of the estimates remain identical. The most likely reason why results remain largely consistent—despite spatial aggregation—is that climatic shifts occur across a wide area.

There is one caveat in linking NCVS data to national climatic data, however. NCVS data are effectively sampled on population density, whereas climatic averages reflect geographic averages. We tested this in our work by using Mares and Moffett’s (2019) data and compared their spatially weighted temperature data, versus geographically weighted NOAA data. Results indicated only a small drop in coefficient strength (around 10%) for the geographically weighted NOAA data. This would suggest that with a representative sample such as the NCVS similar results should be attainable.

It is important to stress that rather than raw temperature we use temperature anomalies (see Mares, 2013a; Mares and Moffett, 2019). The procedure for such data is quite straightforward and is simply the difference between temperatures in a specific month subtracted by the long-term monthly mean for that corresponding month. Employing such measures has two advantages. One, it creates a measure that stands independent of seasonality as anomalies are positive or negative only in relation to long-term averages for that specific month and is therefore a more effective measure of the actual temperature effect on crime. Two, because we use the long-term mean (from 1900–2000) we can effectively speak not just to the temperature effect but also to the plausible impact of climate change on crime (see Hsiang and Burke, 2014). Whereas temperature is considered an independent variable in our study, the same is not true for precipitation. While some studies have found negative associations between precipitation and crime at daily levels, this is not the case for monthly studies (see Mares, 2013a). Precipitation anomalies, however, occur more frequently in cooler months (i.e. spring and fall), which means high levels of precipitation are likely to suppress temperature effects during such months. In short, we believe precipitation anomalies have little inherent value in monthly studies, other than serving as a correction for temperature because months with more precipitation are cooler on average (cloud coverage restricts solar absorption of land) and precipitation is likely to suppress mobility, thereby creating conditions for spurious effects.

In order to explore if socioeconomic variation may alter the link between climatic indicators and crime we also include the consumer price index and the unemployment rate (seasonally unadjusted) collected from the Federal Reserve. In addition to the economic controls we also employ a ‘standard’ battery of binaries: year, month, number of weekend days and number of days in the month. The latter variables control for unobserved variation tied to specific time periods (Mares and Blackburn, 2019)

Analysis and results

Our analytical strategy is primarily influenced by the distributional nature of monthly NCVS data. Once weighted and aggregated by month, NCVS data take on an atypical distribution. While many of our aggregated variables appear to show a normal distribution, all exhibit positive skew and high levels of kurtosis. Histograms further confirm that weighing of the incidents does not fully transform the distributions from a Poisson-type distribution to a normal distribution. Rather, our dependent variables follow a Gamma distribution, characterized by positive skew and high kurtosis. In essence, weighted monthly NCVS data are fractionalized Poisson data as the underlying sample produces a Poisson distribution (count data), that is subsequently weighted. While the weighing of NCVS data is important because each raw incident does not equally reflect population distributions, weighing simply upscales the distribution, but on average does not fundamentally alter it.

Because serial correlation is present in the dependent variables we run the risk of artificially reducing standard errors; to counter this problem we ran diagnostic regression models in and excluding our time-dependent binaries (year month, weekend days, and days of the month). Including such binaries adequately contains serial correlation in our models (as measured by Durbin-Watson, statistic) and once included the use of heteroscedasticity and autocorrelation robust standard errors (such as Newey-West) produce little difference from default standard errors, an important marker of model robustness (King and Roberts, 2015).

In order to accommodate the distributional nature and have the ability to model multiple distributional variations as well as different standard error calculations we run GLM models (Gamma with log link presented here, additional models may be viewed by contacting the authors) in which we model the predicted exponentiated log odds of the sum of crime incidents on time t as the outcome of binary time controls (τ) and the concurrent predicted effects of temperature and precipitation anomalies as well as economic indicators and a population term.

L n E (C r i m e_{t}) = τ_{t} + β_{0} + β_{1} T e m p e r a t u r e_{t} + β_{2} P r e c i p i t a t i o n_{t} + β_{3} C P I_{t} + β_{4} U n e m p l o y m e n t_{t} + β_{5} P o p u l a t i o n_{t} + ε

We explored numerous variations of the model, including Gaussian (both log and identity link), Poisson and Negative Binomial as well as autoregressive–moving-average (ARMA) models and we also explored the stability of significance levels using various standard errors (Newey-West, robust, bootstrapped). In all instances the models return substantially similar results that would not change our general conclusions. In fact, the consistency of the model and standard error choices raises our confidence in the results presented.

We are hesitant to present models for individual crime types as many are too infrequently reported in the NCVS, causing erratic month-to-month levels which are likely to inflate standard errors, thereby underestimating significance levels, although coefficient sizes appear quite comparable with UCR data. In addition, not all NCVS crime categories can easily be compared to UCR crime categories as NCVS only counts victimizations against individuals. Therefore, we first compare aggregated categories of crime against monthly UCR offenses that broadly examine property and violent offenses.

Table 1 shows exponentiated coefficients (incident rate ratios, IRR) for all reported UCR (Part I and simple assaults) and NCVS offenses and further breaks them down by type (violent or property offense). There is little difference in the reported coefficients. For all offenses (which is mostly carried by the numerically largest group of property offenses), NCVS data report an average increase of 0.47% per degree Fahrenheit, whereas UCR data indicate a 0.53% average increase. For violent crimes, NCVS data indicate a 0.61% average increase per degree Fahrenheit and for UCR data this number equals 0.52%. For property offenses, NCVS data report slightly lower coefficients than the UCR (0.42% versus 0.53% average increase per degree Fahrenheit). In sum, NCVS and UCR data show comparable results for the association between anomalous temperatures and crime.

Table 1.

A comparison of UCR and NCVS data: exponentiated coefficients for all reported UCR (Part I and simple assaults) and NCVS offenses, further broken down by type (violent or property offense). Standard errors in parentheses.

	NCVS	UCR	Violence NCVS	Violence UCR	Property NCVS	PropertyUCR
CPI	0.9997 (0.006678)	0.9989 (0.002289)	1.0003 (0.01199)	0.9990 (0.002067)	0.9993 (0.006870)	0.9989 (0.002645)
Unemployed	1.0107 (0.02602)	0.9959 (0.008300)	1.0543 (0.04820)	1.0020 (0.007778)	1.0019 (0.02663)	0.9938 (0.009445)
Precipitation	1.0192* (0.009396)	1.0136** (0.003077)	1.0355* (0.01705)	1.0117** (0.002885)	1.0145 (0.009635)	1.0140** (0.003499)
Temperature	1.0047* (0.001898)	1.0053** (6.257e−04)	1.0061*** (0.003404)	1.0052** (5.840e−04)	1.0042* (0.001953)	1.0053** (7.128e−04)
Population	1.0000 (3.329e−05)	1.0000 (1.127e−05)	1.0000 (5.941e−05)	1.0000** (1.062e−05)	0.9999 (3.433e−05)	1.0000** (1.282e−05)
Constant	8.9692e+10** (7.6348e+11)	4,172.6** (12,024)	33,409 (507,506)	4.825e−04** (0.001310)	1.1360e+12** (9.9718e+12)	3.0186e+06** (9.8909e+06)
Observations (n)	283	263	283	263	283	263

UCR: Uniform Crime Report; NCVS: National Crime Victimization Survey; CPI: Consumer Price Index.

**p < 0.01, *p < 0.05, ***p < 0.1.

Because NCVS data include contextual location indicators, we next examine how different types of locations may be differentially sensitive to anomalous temperature. Table 2 explores differences between indoor and outdoor locations. Whereas indoor locations show no association with temperature, outdoor locations show a significant positive association with temperature, indicating that anomalous temperatures appear to primarily impact victimizations that occur in outdoor settings.

Table 2.

Indoor versus outdoor crimes: exponentiated coefficients (standard errors in parentheses).

	Indoor	Outdoor
CPI	1.0029 (0.01216)	0.9963 (0.008015)
Unemployed	0.9970 (0.04648)	0.9923 (0.02686)
Precipitation	1.0150 (0.01705)	1.0234* (0.01000)
Temperature	0.9993 (0.003440)	1.0056** (0.002131)
Population	1.0000 (6.058e−05)	1.0000 (3.644e−05)
Constant	3.4656e+08 (5.3681e+09)	6.2679e+07*** (5.8385e+08)
Observations	283	283

**p < 0.01, *p < 0.05, ***p < 0.1.

Moreover, Table 3 shows how distance from a victim’s home may impact victimization risks. The risk of victimization during anomalously warm weather appears to increase substantially when a victim leaves their residence, but only up to a point. Within the first mile around a residence, victimization risk increases 0.97% per average degree Fahrenheit increase in temperature, whereas in the direct vicinity of one’s home (model 1), or beyond 1 mile (models 3 and 4) victimization risk increases only about 0.3% per degree Fahrenheit increase.

Table 3.

Distance from home: exponentiated coefficients (standard errors in parentheses).

	Street	<1 mile	Between1–5 miles	>5 miles
CPI	0.9985 (0.007308)	0.9778 (0.01949)	0.9980 (0.01408)	1.0101 (0.01586)
Unemployed	1.0320 (0.02923)	1.0129 (0.06142)	0.9789 (0.05211)	0.9823 (0.04401)
Precipitation	1.0130 (0.01026)	0.9959 (0.02063)	1.0262 (0.01972)	1.0250 (0.01600)
Temperature	1.0037*** (0.002089)	1.0097* (0.004502)	1.0030 (0.003928)	1.0033 (0.003346)
Population	1.0000 (3.656e−05)	0.9999 (8.248e−05)	1.0000 (6.973e−05)	1.0000 (6.239e−05)
Constant	1.4910e+09* (1.3939e+10)	6.9228e+19* (1.4597e+21)	35,910 (640,202)	15,133 (241,305)
Observations	283	283	283	283

**p < 0.01, *p < 0.05, ***p < 0.1; CPI: Consumer Price Index.

For males, the risk of victimization during anomalous temperatures is slightly higher than for females (0.51% vs. 0.43% per degree Fahrenheit), showing that there may be some gendered effects (see Table 4). What is more, victimization risks appear more strongly associated with anomalous temperature for African Americans (0.90%) than for whites (0.57%), or Hispanics (0.10%).

Table 4.

Gender and race: exponentiated coefficients (standard errors in parentheses).

	Male	Female	White	Black	Hispanic
CPI	1.0015 (0.008624)	0.9979 (0.008133)	0.9975 (0.007389)	0.9887 (0.01852)	1.0143 (0.01453)
Unemployed	1.0383 (0.03444)	1.0041 (0.03002)	1.0337 (0.02954)	1.0563 (0.06487)	0.9441 (0.05078)
Precipitation	1.0248* (0.01216)	1.0154 (0.01074)	1.0197*** (0.01042)	1.0270 (0.02221)	1.0016 (0.01966)
Temperature	1.0051* (0.002452)	1.0043*** (0.002223)	1.0057** (0.002109)	1.0090+ (0.004617)	1.0010 (0.003994)
Population	1.0000 (4.285e−05)	1.0000 (3.792e−05)	0.9999 (3.696e−05)	1.0000 (7.869e−05)	1.0000 (7.043e−05)
Constant	1.1041e+10* (1.2098e+11)	5.0378e+09* (4.8838e+10)	8.2918e+12** (7.8361e+13)	6,039.7 (121,493)	4,682.9 (84,336)
Observations	283	283	283	283	283

**p < 0.01, *p < 0.05, ***p < 0.1; CPI: Consumer Price Index.

Control variables present a consistent picture with economic indicators showing no impact across the various categories examined. The latter is not too surprising; economic shifts do not necessarily impact crime levels as the most vulnerable populations in the United States are likely to see little impact from broader economic trends. Precipitation typically generates positive coefficients, but the actual predicted impact of precipitation is quite small compared to temperature, as a one-unit change (a one-inch increase in monthly precipitation levels) in precipitation equals over two standard deviations. Regardless, precipitation presents fairly consistent positive associations to crime. It is unclear why this relationship is positive, but it may be an outcome of the monthly level of aggregation where warmer than average months may spur more precipitation due to the ability of the atmosphere to hold more moisture. Year and month binaries show predictable coefficients, with year binaries generally showing negative coefficients (compared to year 1992) and monthly binaries showing positive coefficients throughout warmer months and negative during cooler months (compared to January).

Our analysis is quite robust, and alternative modelling strategies yielded no difference in the key conclusions from this study. For example, we ran all models presented using Gaussian (both log link and identity link), Poisson and negative binomial models, and resulting differences are negligible, especially given the relatively small number of monthly observations. ARMA models were also developed and again yielded substantially similar results. Different standard errors were calculated as well (including robust, Newey West and bootstrapped), but results showed a large degree of consistency, indicating that our modeling approach is likely to be robust (King and Roberts, 2015).

Discussion

Although it is impossible to reliably drill down NCVS data further due to the small sample size of finely disaggregated categories, it appears likely that victimization risk during warmer than normal months may be compounded for certain groups of citizens. Taken together, for example, our analysis indicates that African American males who frequent outdoor locations in their neighborhood are likely to be at greater risk than white females staying indoors. At a minimum, however, our approach supports the notion that victimization risks are increased by anomalously warm temperatures. Our results indicate similar temperature associations as provided by prior research using UCR data, suggesting that NCVS data can be used to model the impact of anomalous temperatures. In addition, UCR data are likely not to be biased by reporting difference under varying climatic conditions given the similarity in results.

Nonetheless, while the NCVS has on the whole a representative sample, its effective sample size is comparable to a city of under 100,000 residents. This certainly creates some issues, including the fact that infrequent crimes or disaggregated combinations of incident characteristics show greater variability and therefore greater standard errors than data sets that draw from a larger population such as the UCR.

Nonetheless, our results are theoretically intriguing. Like prior research we find evidence that violent crimes are more subject to temperature changes. More importantly, we find that much of the coefficient strength of both violent and property offenses is derived from victimizations that occur in outdoor settings, whereas victimizations in indoor settings show no elasticity with regards to temperature. This may mean that being indoors can be an effective way to avoid victimization during warmer than usual weather, but it could also indicate that people simply increase their presence outside during such weather. The finding that crime increases most in the direct vicinity of a victim’s home (under a mile, but not in the victim’s home) during anomalous temperatures is also interesting as it suggests that these victims were likely on foot and thus experienced more direct weather exposure. From a routine activities point of view this could additionally mean that not all forms of mobility matter equally in bringing victims and offenders together. In addition, we find slight gender differences in offending, with males more likely to be impacted by anomalous weather; this may mean that men are more likely to be exposed to direct weather conditions, or that men are more likely to increase their mobility during warmer than usual weather. For race, the reported differences are larger, especially for African Americans, who appear especially at risk during anomalously warm weather.

Combined our results indicate that victimization is not equally impacted by anomalous weather but varies by contextual factors. This finding is highly important as it indicates that climatic variation may impact different groups of people at different levels. The findings resonate with Mares’ (2013b) study showing that neighborhoods with higher levels of social disorganization are more affected by warmer than normal temperatures. Unlike Mares’ study, which was confined to one non-representative city, our results have greater generalizability as the NCVS is nationally representative. The fact that vulnerable populations will bear a greater burden of negative climatic impacts is certainly not surprising, but our study is the first that shows this for a representative sample.

Our results measure monthly temperature anomalies—in effect, how much or little a given month’s temperature deviates from the long-term mean—and not climate change directly. Nonetheless, our measure for climate change estimates how average increases in temperature would impact crime levels. Given that temperatures have increased nearly two degrees Fahrenheit over the last century, one may conclude that at present the average impact of climate change on crime can be estimated to be around 1%, depending, of course, on crime type and contextual factors. Given that prior studies used UCR data and found similar coefficient sizes, they were likely to underestimate the total impact on crime, because UCR data only include reported data. Using UCR and NCVS crime totals in conjunction with our findings from Table 1 allows us to extrapolate the predicted difference for a month with a 1ºF higher than average temperature. UCR totals for violent crime in 2015 were 1,197,704 and property offenses totaled 7,993,631 in 2015. NCVS totals for that same year equaled 5,006,620 for violent crime and 14,611,040 for property offenses. Using the coefficients in Table 1 would thus lead us to predict a 6,228 increase in violent UCR offenses and a 42,366 increase in property UCR offenses for each degree warmer than average (total 48,594). Using NCVS totals creates a substantially higher prediction, a 30,540 increase in violent offenses and a 69,766 increase in property crimes (total 100,306). It is true that the UCR totals do not include all Part II offenses; however, the UCR also misses crimes not reported. NCVS data, however, do not include homicides. Effectively, the NCVS predicts double the crime increase resulting from the same amount of climate change. Given that recent temperatures have been close to 2ºF above average, the United States is likely to have experienced about 200,000 additional crimes each year as a result of climate change, substantially more than reported in prior studies based on UCR data (see Mares and Moffett, 2019: 517).

Our study certainly has important limitations given the characteristics of the NCVS data. First, the sampling errors for important at-risk subpopulations, including racial ethnic minorities, can become quite large and unstable from year to year because there are few affirmative responses to questions about serious violent victimization in the sampled groups. With the total annual sample of approximately 225,000 persons, there is not much room to introduce sufficient number of respondents in each subcategory. Moreover, households, i.e. geographical locations and types of residences, and not persons with particular characteristics are targeted for inclusion into the sample.

Although the NCVS has a number of design features to reduce potential sources of error when measuring month of occurrence, including a 6-month reference period and ‘bounding’ procedure, such errors may still occur. A level of precision and accuracy associated with recorded month of occurrence of a victimization incident in NCVS is not the same as would be recorded in official police documents.

In addition, because the NCVS does not provide exact location information for the victims in the survey, temperature data cannot be matched geographically. Just the same, temperature data correlate highly across the United States, especially at the monthly level (Mares and Moffett, 2019). We believe, however, that finer grained data (such as the National Incident Based Reporting System [NIBRS]) could produce more robust findings and we would encourage researchers to explore data disaggregation with other data sets.

Our results for precipitation present somewhat of a quandary. In nearly all our models, precipitation anomalies are positively related to crime levels. These relationships persist even when temperature anomalies are excluded from the models. We have no immediate answer to this counter-intuitive finding. Prior research has found contradicting results with respect to precipitation, with daily studies more typically reporting negative relationships. It is possible that on small time scales precipitation suppresses crime, and some studies report that such results are washed out on monthly levels (Mares, 2013a). It is therefore puzzling to find positive associations. It would be possible that an unmeasured variable correlates with precipitation but given its highly erratic trend that seems unlikely. Because there is a weak negative relationship between temperature and precipitation anomalies, we modeled a few additional models in which we added an interaction effect between the two variables, but this uncovered nothing relevant (no significance or particular direction of the interaction effect). Our best guess is that precipitation anomalies suppress crime levels on the day they occur but may drive an overcompensation (pent up demand) in crime on those days when precipitation levels subside. Unfortunately, there is no way to test this idea with the current data.

Finally, our data are not well suited to uncover causality and thus speak to the theoretical implications. Because the NCVS reports monthly aggregates of victimization, the anomalously warm days may or may not produce higher crime levels. It is feasible that the cooler than average days of the month (which also occur in warmer than normal months) produce the effect. Given the prior literature on this topic, such a conclusion seems unlikely, however.

Conclusion

Our results show that NCVS data can be utilized to explore how temperature anomalies shape victimization. Results show that NCVS and UCR data produce similar results and coefficient sizes are comparable to other studies, favoring a greater impact on violent offenses than property offenses. We estimate that—at present—an additional 200,000 crimes occur annually as a result of the warming that has already occurred. This number is much higher than those produced by studies relying on UCR data. Once disaggregated, NCVS data show distinct patterns that indicate victimization during warmer than usual months does not affect everyone equally. Outdoor victimization, and victimization near a person’s home (up to 1 mile) appear especially vulnerable to increases during anomalous temperatures. In addition, subpopulations (males and African Americans) are also at increased risk of victimization. Our results effectively suggest that at-risk populations are more vulnerable to climatic variability. Such a finding is not surprising. In sum, our results are highly suggestive of a substantial impact of climate change on various types of victim characteristics, with an average impact of about 0.5% per degree Fahrenheit. Results indicate that African Americans may be at especially high risk as our climate warms, contributing to the further marginalization of this high-risk group.

It is key that our results are replicated using additional data from various locations and using various levels of aggregation. Not only are these results useful for understanding how climate change may impact future crime levels, they also hold relevance for criminal justice policy and practice. Understanding which contextual factors are mediated by climatic changes could help improve predictive policing, such as Risk Terrain Modelling, as well as preventative and reactive policing. If violence increases during warmer than usual times, and if this violence occurs in outdoor location among specific groups, evidence-based preventative policing strategies could assist. Understanding which population groups and locations are at greatest risk for victimization could help inform preventative and reactive policing strategies.

Footnotes

Notes

References

Agnew

(2011) Dire forecast: A theoretical model of the impact of climate change on crime. Theoretical Criminology 15: 1–22.

Allen

Anderson

Bushman

(2018) The general aggression model. Current Opinion in Psychology 19: 75–80.

Anderson

(1989) Temperature and aggression. Psychological Bulletin 106: 74–96.

Anderson

(2001) Heat and violence. Current Directions in Psychological Science 10(1): 33–38.

Anderson

(1998) Temperature and aggression: Paradox, controversy, and a (fairly) clear picture. In: Geen

Donnerstein

(eds) Human Aggression: Theories, Research, and Implications for Social Policy. San Diego: Academic Press, 247–298.

Anderson

Deuser

(1996) Examining an affective aggression framework: Weapon and temperature effects on aggressive thoughts, affect, and attitudes. Personality and Social Psychology Bulletin 22(4): 366–376.

Anderson

Dorr

DeNeve

Flanagan

KMM

(2000) Temperature and aggression. In: Zana

(ed), Experimental Social Psychology. New York: Academic Press, volume 32, 63–133.

Anderson

Bushman

Groom

(1997) Hot years and serious and deadly assault. Journal of Personality and Social Psychology 73: 1213–1223.

Anderson

Delisi

(2011) In: Krugianski

Williams

(eds.) The Psychology of Social Conflict and Aggression. New York: Psychology Press, 249–265.

10.

Anderson

Johnson

Koyama

(2013) From the persecuting to the protective state? Jewish expulsions and weather shocks from 1100 to 1800. MPRA Working Paper 44228 . Dearborn: University of Michigan.

11.

Auliciems

DiBartolo

(1995) Domestic violence in a subtropical environment: police calls and weather in Brisbane. International Journal of Biometeorology 39: 34–39.

12.

Badiora

Afon

Dada

(2017) Seasonality of violent and property crime in Nigeria: Some preliminary findings. International Journal of Criminology and Sociological Theory 10(2): 1–23.

13.

Baron

Bell

(1976) Aggression and heat: The influence of ambient temperature, negative affect, and a cooling drink on physical aggression. Journal of Personality and Social Psychology 33(3): 245–255.

14.

Bergholt

Lujala

(2012) Climate-related natural disasters, economic growth, and armed civil conflict. Journal of Peace Research 48: 147–162.

15.

Breetzke

Cohn

(2011) Seasonal assault and neighborhoods deprivation in South Africa: Some preliminary findings. Environment and Behavior 44: 641–667.

16.

Bruederle

Peters

Roberts

(2017). Weather and crime in South Africa (No. 739). Ruhr Economic Papers. https://doi.org/10.4419/86788859.

17.

Bureau of Justice Statistics (n.d.) Survey methodology for criminal victimization in the United States. Available at: http://bjs.gov/content/pub/pdf/ncvs_methodology.pdf (accessed 2 February 2020).

18.

Burke

Hsiang

Miguel

(2015) Climate and conflict. Annual Review of Economics 7: 557–617.

19.

Burke

Miguel

Satyanath

Dykema

Lobell

(2009) Warming increases the risk of civil war in Africa. Proceedings of National Academy of Sciences of the United States of America 106:20670–20674.

20.

Carbone-Lopez

Lauritsen

(2013) Seasonal variation in violent victimization. Journal of Quantitative Criminology 29: 399–422.

21.

Carleton

Hsiang

(2016) Social and economic impacts of climate. Science 353: 6304.

22.

Ceccato

(2005) Homicide in Sao Paolo, Brazil: Assessing spatial-temporal and weather variations. Journal of Environmental Psychology 25: 307–321.

23.

Cohen

Felson

(1979) Social change and crime rate trends: A routine activity approach. American Sociological Review 44: 588–608.

24.

Cohn

(1990) Weather and crime. The British Journal of Criminology 30(1): 51–64.

25.

Cohn

EG.

Rotton

(2003) Even criminals take a holiday: Instrumental and expressive crimes on major and minor holidays. Journal of Criminal Justice, 31: 351–360.

26.

Cohn

Rotton

(2005) The curve is still out there. Journal of Personality and Social Psychology, 89: 67–70.

27.

DeFronzo

(1984) Climate and crime: Tests of an FBI assumption. Environment and Behavior 16: 185–210.

28.

Field

(1992) The effect of temperature on crime. British Journal of Criminology 32(3): 8–23.

29.

Gamble

Hess

(2012) Temperature and violent crime in Dallas, Texas: Relationships and implications of climate change. The Western Journal of Emergency Medicine 13(3): 217–220.

30.

Habibullah

(2017) The effects of weather on crime rates in Malaysia. International Journal of Business and Society 18(2): 263–270.

31.

Harp

Karnauskas

(2018) The influence of interannual climate variability on regional violent crime rates in the United States. GeoHealth 2: 356–369.

32.

Harries

Stadler

Zdorkowski

(1984) Seasonality and assault: Explorations in inter-neighborhood variation, Dallas 1980. Annals of the Association of American Geographers 74: 590–604.

33.

Hipp

Bauer

Curran

Bollen

(2004) Crimes of opportunity or crimes of emotion? Testing two explanations of seasonal change in crime. Social Forces 82(4): 1333–1372.

34.

Horrocks

Menclova

(2011) The effects of weather on crime. New Zealand Economic Papers 45(3): 231–254.

35.

Hsiang

Burke (2014) Climate, conflict, and social stability. Climatic Change 123: 39–55.

36.

Hsiang

Burke

Miguel

(2013) Quantifying the influence of climate on human conflict. Science 341: 1212–1226.

37.

Chen

Sun

(2017) Impact of climate variability and change on crime rates in Tangshan, China. Science of the Total Environment 609: 1041–1048.

38.

Jacob

Lefgren

Moretti

(2007) The dynamics of criminal behavior. Journal of Human Resources 42(3): 489–527.

39.

Kenrick

MacFarlane

(1986) Ambient temperature and horn honking. Environment and Behavior 18: 179–191.

40.

King

Roberts

(2015) How robust standard errors expose methodological problems they do not fix, and what to do about it. Political Analysis 23: 159–179.

41.

Krenzer

WLD

Splan

(2018) Evaluating the heat-aggression hypothesis: The role of temporal and social factors in predicting baseball related aggression. Aggressive Behavior 44(1): 83–88.

42.

Larrick

Timmerman

Carton

Abrevaya

(2011) Temper, temperature, and temptation: Heat-related retaliation in baseball. Psychological Science 22(4): 423–428.

43.

Lebeau

(1990) The oscillation of police calls to domestic disputes with time and the temperature humidity index. Journal of Crime and Justice 17: 149–161.

44.

Lebeau

Corcoran

(1994) Changes in calls for service with changes in routine activities and the arrival and passage of weather fronts. Journal of Quantitative Criminology 6: 269–291.

45.

Lebeau

Langworthy

(1986) The linkages between routine activities, weather and calls for police services. Journal of Police Science Administration 14: 137–145.

46.

Linning

Andresen

Ghaseminejad

Brantingham

(2017) Crime seasonality across multiple jurisdictions in British Columbia, Canada. Canadian Journal of Criminology and Criminal Justice 59(2): 251–280.

47.

McDowall

Curtis

(2015) Seasonal variation in homicide and assault across large U.S. cities. Homicide Studies 19: 303–325.

48.

McDowall

Loftin

Pate

(2011) Seasonal cycles in crime and their variability. Journal of Quantitative Criminology 28: 389–410.

49.

Mares

(2013a) Climate change and levels of violence in socially disadvantaged neighborhood groups. Journal of Urban Health 90: 768–783.

50.

Mares

(2013b) Climate change and crime. Crime Law and Social Change 59: 185–208.

51.

Mares

Blackburn

(2019) Major league baseball and crime: opportunity, spatial patterns, and team rivalry at St. Louis Cardinals games. Journal of Sports Economics. https://doi.org/10.1177/1527002518822702

52.

Mares

Moffett

(2016) Climate change and interpersonal violence: A ‘global’ estimate and regional inequities. Climatic Change 135(2): 297–310.

53.

Mares

Moffett

(2019) Climate change and crime revisited: An exploration of monthly temperature anomalies and UCR crime data. Environment and Behavior 51(5): 502–529.

54.

Miguel

(2005) Poverty and witch killing. Review of Economic Studies 72:1153–1172.

55.

National Archive of Criminal Justice Data (NACJD) (n.d.) Resource guide: National Crime Victimization Survey. Available at: https://www.icpsr.umich.edu/icpsrweb/NACJD/NCVS (accessed 2 February 2020).

56.

National Research Council (2014) Estimating the Incidence of Rape and Sexual Assault. Panel on Measuring Rape and Sexual Assault in Bureau of Justice Statistics Household Surveys, Committee on National Statistics, Division on Behavioral and Social Sciences and Education. Washington, D.C.: The National Academies Press.

57.

Quetelet

(1842) A Treatise on Man and the Development of His Faculties. New York: Burt Franklin.

58.

Ranson

(2014) Crime, weather, and climate change. Journal of Environmental Economic Management 67: 274–302.

59.

Reifman

Larrick

Fein

(1991) Temper and temperature on the diamond. Personality and Social Psychology Bulletin 17: 580–585.

60.

Rinderu

Bushman

Van Lange

(2018) Climate, aggression, and violence (CLASH): A cultural-evolutionary approach. Current Opinion in Psychology 19: 112–118.

61.

Rotton

(1986) Determinism redux: Climate and the cultural correlates of violence. Environment and Behavior 18: 346–368.

62.

Rotton

Cohn

(2000) Violence is a curvilinear function of temperature in Dallas. Journal of Personality and Social Psychology 78: 1074–1081.

63.

Rotton

Cohn

(2003) Global warming and US crime rates. Environment and Behavior 35: 802–825.

64.

Schinasi

Hamra

(2017) A time series analysis of associations between daily temperature and crime events in Philadelphia, Pennsylvania. Journal of Urban Health 94(6): 892–900.

65.

Sekhri

Storeygard

(2013) Dowry deaths: consumption smoothing in response to climate variability in India. Working Paper . Charlottesville: University of Virginia.

66.

Tilhonen

Halonen

Tiihonen

Kautiainen

Storvik

Callaway

(2017) The association of ambient temperature and violent crime. Scientific Reports 7(1): 1–7.

67.

Williams

Hill

Spicer

(2015) The relationship between temperature and assault in New Zealand. Climatic Change 132: 559–573.