Sage Journals: Discover world-class research

Abstract

Google Trends (GT) data could potentially supplement traditional crime measurement strategies, allowing criminologists to better understand crime statistics on a macro level. This study assesses the validity of GT crime estimates. The findings indicate that GT data are reliable for estimating MVT, larceny, and rape. Additionally, we illustrate how to use GT to identify places with high rates of unreported offenses. The results of this study demonstrate the feasibility of leveraging open-source big data such as GT to supplement traditional sources of crime data, particularly for categories of crime with substantial underreporting rates. Results suggest the GT rape measure may be a more accurate estimate of the true incidence of rape than the measure drawn from the Uniform Crime Report (UCR). The limitations associated with the use of GT to generate estimates of crime are also discussed.

Keywords

Google Trends crime measurement victimization big data dark figure of crime

Introduction

“Measuring crime has always been one of the most difficult challenges facing criminal justice researchers” (Maxfield, 1999, p.119). One key issue which has been identified is that most crime estimates underestimate the total crime committed each year. Prior research has demonstrated that a variety of factors, including gender, race, wealth, education, the nature of the crime, and the police response to previous victimization, influence victims’ decisions to file a police report (Avakame et al., 1999; Black, 1970; Fyfe et al., 1997; Hindelang, 1974; Laub, 1981; O’Brien, 1996; Skogan, 1984; Xie et al., 2006). In addition, certain categories of crime victims may not wish to contact law enforcement, out of concern for their own legal status, such as undocumented immigrants (Gutierrez and Kirk, 2017), sex workers (McBride et al., 2020), and drug dealers (Topalli et al., 2002).

The underreporting phenomenon extends to, and has varying effects on, distinct types of victimization. Existing research suggests homicide and auto theft are two of the most frequently reported crimes to the police. This may be due to the severity of the crime (loss is greater in homicide and motor vehicle theft), easier to discover than other crimes (body and the plate), and the requirement for insurance coverage plays a crucial part in the reporting behavior (Black, 1970; Goudriaan et al., 2004; Maxfield, 1999; Skogan, 1977, 1984). On the other hand, rape is known to suffer from the largest degree of underreporting. For example, according to the Bureau of Justice Statistics (2019), only 24.9% of rape victims reported the crime to police in 2018, compared to 78.6% for motor vehicle theft (MVT), 47.9% for burglary, and 28.6% for larceny. The reasons for underreporting in regard to rape include the loss of privacy, the risk of recrimination and reprisal, the familiarity between offender and victims (e.g. victimizied by relatives or friends), feeling shameful of being known by others, difficulty in accessing the evidence, and the risk of being “no crimed” by police (Allen, 2007; Gregory and Lees, 1996; Hine et al., 2021; Thomas and Wolff, 2021). Studies have shown that Native Americans are more likely to experience sexual assault than others due to the unfortunate combination of colonization, patriarchy, political status, and racism that exists in tribal communities (Bohn, 2003; Tjaden and Thoennes, 2006; Quasius, 2008). As a result, the Federal Bureau of Investigation’s Uniform Crime Report (UCR), which contains information about offenses “known to the police,” has inherent limitations in estimating the true incidence of crime.¹

Despite these shortcomings, scholars have continued to investigate the dependability and validity of UCR data. Prior research has compared UCR data to the National Crime Victimization Survey (NCVS) and other local crime data, such as arson statistics from local fire departments. They discovered that estimates of homicide, MVT, robbery, and burglary may be considered valid metrics of criminal behavior in UCR (Blumstein et al., 1991; Gove et al., 1985; O’Brien et al., 1980). On the other hand, rape, larceny, aggregated assault, arson, and serious violence are measures that they were unable to validate (Blumstein et al., 1991; Gove et al., 1985; Jackson, 1988; Lauritsen et al., 2016; O’Brien et al., 1980; Skogan, 1977). O’Brien (1996) and Lauritsen et al. (2016) examined the validity of violent crimes and concluded that UCR homicide is a more valid indicator of violent crime. They contended that if one indicator of violent crime is more strongly connected with UCR homicide rates than another, the former is more valid.

The global prevalence of the internet enables the use of big data in social science research. According to research, these internet-based and user-generated data are more accessible and less expensive than traditional social surveys (Salganik, 2017). Over 89% of adults in the United States reported using the internet (Pew Research Center, 2019). Additionally, internet users are distributed similarly to the broader population in terms of ethnicity, wealth, gender, and age (Pew Research Center, 2019; Stubbs-Richardson et al., 2018). Furthermore, 91% of internet users use search engines, with the vast majority (87.7%) of users using Google to perform their searches (Pew Research Center, 2012; Statcounter, 2021). Rowlands et al. (2008) research on the impact of Google on education concluded that search engines like Google have not only impacted the youth or students, but also extended to the parents and teachers. The evidence showed that people across all ages has changed the way they seek information, “We are all the Google Generation.”

Due to the extensive use of Google Search, a massive amount of data is recorded during this process. As a result, this data may be a precious resource for researchers interested in a variety of subjects. Google Trends (GT) is a Google Search extension that collects data on users’ time, location, and keywords. GT generates data on the search rates for keywords based on the acquired data to predict their users’ search interest. First, users enter a term on the GT website. Next, GT estimates the term’s “search interest” in a geographic area by dividing the number of search queries for that term by the total number of Google searches in that geographic area. From this data, GT calculates a number (scaled from 0 to 100) that represents a certain geographic area’s search interest relative to all other geographic areas. The area with the highest search interest is assigned a score of 100. The values for the other areas (99, 81, 76,…25) are based on the ratio of each region’s search rate to the area with the most searches using the designated phrases (for further discussion of GT method, see Arora et al., 2019; Mavragani and Ochoa, 2019).

Although its utility has not been thoroughly explored, some studies have demonstrated the utility of GT data in social science research. For instance, academics discovered that internet users with a practical goal in online searching help obtain certified GT data to match with real-world survey data (Ginsberg et al., 2009; Mellon, 2013). Moreover, scholars have discovered a strong association between GT data on religion and existing data on religious composition in the United States and Africa (Adamczyk et al., 2019, 2021a, 2021b; Scheitle, 2011). Other scholars have found that GT data accurately predicts future heroin emergency room visits, influenza patterns, methamphetamine-related crimes, premature deaths from alcohol, narcotics, and suicide, as well as racial animus toward a black candidate (Gamma et al., 2016; Parker et al., 2016; Stephens-Davidowitz, 2014; Xu et al., 2017; Young et al., 2018). Most recently, GT data has been used to estimate the prevalence of COVID-19 over time in Iran (Ayyoubzadeh et al., 2020) and to assess the effect of COVID-19 related lockdowns on individuals’ wellbeing (Brodeur et al., 2021). Importantly however, research leveraging GT data to probe criminal justice issues is limited at this time. Gross and Mann (2017) created a measure of concern about police violence using Google search queries and discovered that this measure is positively associated with the violent crime rate. Stubbs-Richardson et al. (2018) calculated the popularity of search phrases such as “vehicle alarm system” and “house alarm system” to represent individuals’ crime prevention searches and discovered a positive correlation between the relative volume of these queries and the rates of MVT and burglary captured in the UCR. More recently, Piña-García and Ramírez-Ramírez (2019) employed GT data and official crime data to explore the incidence criminal activity in Mexico City. Hoehn-Velasco et al. (2021) examined whether COVID-19 stay-at-home orders affected crimes targeting women (sexual offenses and domestic violence). Anderberg et al. (2022) observed a 40% surge in internet search-based domestic violence during the apex of the lockdown period in Greater London, which was seven to eight times more than the rise in police-recorded statistics and much closer to the increase in helpline calls reported by victim support organizations. From this review of existing research, it becomes clear that while researchers have begun to assess the potential utility of Google Search data in criminal justice research, much is left to be explored.

Current study

Despite the existing efforts to incorporate GT data into social science research, the feasibility of using GT to provide information about crime victimization and to generate crime estimates has not been evaluated. As a result, the current study examines the feasibility of using GT data to estimate crime victimization rates in metropolitan areas throughout the United States. The current study’s principal objectives are to assess the potential utility of GT data to estimate the occurrence of crime and investigate how these estimates might augment official crime reports and inform criminal justice policies and practices.

To our knowledge, this is the first study to use big data from GT to estimate criminal victimization at the metropolitan area level (i.e. designated market area [DMA]) across the United States. Additionally, this is the first study to examine the relationship between widely reported structural attributes and reported crime rates using Google Trends data.

Methods and data

Comparing two measures of crime—GT and UCR crime rates

We conduct several statistical tests of concurrent validity for the GT estimates of crime to determine their utility as a measure of crime victimization (Churchill, 1979). These statistical tests are Correlation analysis (Churchill, 1979; O’Brien et al., 1980; Scheitle, 2011), exploratory factor analysis (EFA), confirmatory factor analysis (CFA), and multiple linear regression with common crime covariates (Bergkvist and Rossiter, 2007; Churchill, 1979; O’Brien et al., 1980; Strauss and Smith, 2009).

To begin, we assume that DMAs have a “true” crime rate denoted by the variable ${C r i m e}_{T r u e}$ . In an ideal world, if there was a crime measurement $i$ that could accurately represent the genuine crime rate within DMAs, we would assume that ${C r i m e}_{i}$ = ${C r i m e}_{T r u e} .$ Unfortunately, there are no perfect measures for ${C r i m e}_{T r u e}$ due to the reporting errors, systematic recording errors and unknown random errors stemming from a multitude of sources. Ideally, we would expect a new measure drawn from GT (denoted as ${C r i m e}_{G T}$ ) to approximate the more traditional measure of crime drawn from the UCR (denoted as ${C r i m e}_{U C R}$ ). However, the existing accountable measures of ${C r i m e}_{T r u e}$ from ${C r i m e}_{U C R}$ are few. Thus, if a type of crime measure from ${C r i m e}_{G T}$ had a low level of correlation with that same type of crime measure of ${C r i m e}_{U C R}$ , it could be that either ${C r i m e}_{G T}$ or ${C r i m e}_{U C R}$ is not close to ${C r i m e}_{T r u e}$ . In this study, the previously validated UCR measures—homicide, MVT, robbery, and burglary (Blumstein et al., 1991; Gove et al., 1985; O’Brien et al., 1980; Skogan, 1974)—are used as criteria for the validation process. These criteria are designated as ${H o m i c i d e}_{U C R}$ , $M V T_{U C R}$ , ${R o b b e r y}_{U C R}$ , and ${B u r g l a r y}_{U C R}$ .

For the correlation analysis, it would make sense that $M V T_{G T}$ is a valid measure if $M V T_{G T}$ is highly correlated with $M V T_{U C R} .$ Similarly, we regard ${B u r g l a r y}_{G T}$ to be a reasonable estimate of criminal activity if the measure is highly correlated with ${B u r g l a r y}_{U C R}$ . This is not the case, however, for the validity test of rape and larceny measures. It is because that $L a r c e n y_{U C R}$ and ${R a p e}_{U C R}$ are not valid measures. If a tested measure (e.g. ${R a p e}_{G T}$ ) is not highly correlated with an non-validated UCR measure (e.g. ${R a p e}_{U C R}$ ), this could indicate that the tested measure is not close to, or that the non-validated UCR measure is not close to ${C r i m e}_{T r u e}$ , or both of them are not close to ${C r i m e}_{T r u e}$ .

Without previously valid rape and larceny data from UCR, we cannot validate rape and larceny crime measures from GT. Lauritsen et al. (2016), as well as O’Brien (1996), argued that because the homicide numbers at UCR are exceptionally well-measured, we can use ${H o m i c i d e}_{U C R}$ to access rape measures. Thus, during the assessment of ${R a p e}_{G T}$ and ${R a p e}_{U C R}$ , when one rape crime measure is more associated with ${H o m i c i d e}_{U C R}$ than another, it is deemed to be more valid. Similarly, in order to test the validity of property crime, because MVT data from UCR is also exceptionally well-measured, we can use $M V T_{U C R}$ to access $L a r c e n y_{G T}$ and $L a r c e n y_{U C R}$ . When one larceny crime measure is more highly associated with $M V T_{U C R}$ , it is considered to be a better estimate of the true rate of crime.

Regarding EFA and CFA tests, we suggest that a need-to-be-valid measure of violent/property crime should be able to be grouped with other already-valid measures of violent/property crime in UCR, such as homicide, MVT, robbery, and burglary. As scholars often divide crimes into two categories: property and violent crime (Fox et al., 2010; Matthews et al., 2011; Zhao et al., 2002), we propose that there are two latent variables: property crime rates and violent crime rates in the EFA and CFA test.² Following this logic, we assess the need-to-be-valid property crime measure from ${C r i m e}_{G T}$ (e.g. MVT_GT, ${B u r g l a r y}_{G T}$ , and $L a r c e n y_{G T}$ ) and ${C r i m e}_{U C R}$ (e.g. $L a r c e n y_{U C R}$ ) with the already-validated property crime criteria (e.g. $M V T_{U C R}$ and ${B u r g l a r y}_{U C R}$ ). Similarly, we also evaluate the novel violent crime measures from ${R a p e}_{G T}$ and ${R a p e}_{U C R}$ with the existing violent crime criteria (e.g. ${H o m i c i d e}_{U C R}$ and ${R o b b e r y}_{U C R}$ ).³

Additionally, we use multiple linear regression (MLR) analysis to determine the nomological validity (O’Brien et al., 1980) of the crime estimates. The outcome variables are drawn from ${C r i m e}_{G T}$ and ${C r i m e}_{U C R}$ , whereas the predictor variables are drawn from a pool of commonly used criminological covariates. The resulting estimates should be comparable in magnitude and direction to previously validated UCR violent/property measures (e.g. $M V T_{U C R}$ or ${H o m i c i d e}_{U C R}$ ). Additionally, if GT represents a valid measure of criminal activity, these resulting associations should be consistent with theoretical expectations. Finally, the indications of model fit⁴ suggest which crime measure best fits the included predictor variables.

Designated market area as the unit of analysis

The nature of the GT data determines the unit of analysis in this study. GT’s smallest geographic unit across the United States is the designated market area (DMA). In GT, 210 DMAs cover all 3142 US counties. Because the analytical unit is DMA, it is necessary to aggregate other county-level data sources to DMA to align them with GT data. Also, by aggregating the UCR data from county-level to a larger geographic level of DMA, we can avoid the drawbacks Maltz and Targonski (2002) mentioned in the county-level UCR data. We use Sood’s (2016) county-DMA crosswalk file because it is the only publicly available county-DMA crosswalk file and it has already been widely cited in peer-reviewed research (Obermeier et al., 2022; Sides et al., 2022; Sood and Iyengar, 2016).

Two sets of crime estimates—GT and UCR

The study’s focal measures come from two unique sources: first, GT data, then UCR crime data. The current study uses GT and UCR data from 2010 to 2019. We omit data before 2010 since internet availability reached over 75% of the US population in 2010 (Pew Research Center, 2019). The end year is 2019 because it is the final year for which county-level UCR data is provided. The steps we took to get criminal victimization estimates from GT are detailed below.

We need to use words that victims used in their online searches to obtain the “estimate of victimization” utilizing GT. We start with a simple term like “my car was stolen” to get more similar or often-used keywords that victims usually enter on GT. There is a “related queries” section on Google Trends that lists the top 25 queries related to the keyword entered. For instance, a search for “my car was stolen” will return similar phrases such as “my car was stolen, what do I do?”, “find my car,” and “report stolen car.” Then, we combine these terms if they are associated with a victim searching for information online (using “+”) or exclude them if they are not (using “−”). Thus, each type of crime will have a single representative search phrase on GT.

There are two main purposes for combining “related queries” into a single phrase, as opposed to searching for different terms each time and combining the separated GT data into a single dataset. First, if we employ the latter method, we will lack sufficient data points for analysis. It is because the greater the number of input keywords in a phrase, the more DMA data points we can acquire from GT. Second, the use of related queries can mitigate potential bias caused by the researcher’s own selection of arbitrary words.

The keywords we use to estimate multiple crime types from GT are mentioned below. The negative symbol (−) excludes unrelated terms from “related queries” in GT. The plus sign (+) is also used to link terms. These terms are used to gather estimates of ${C r i m e}_{G T}$ from GT.⁵

Terms to pull $M V T_{G T}$ : “car stolen + find stolen car + report police stolen car + insurance car stolen − dream − check.”

Terms to pull ${B u r g l a r y}_{G T}$ : “burglary + home burglarized + my house broken into + my home broken into + burglary report police − dream.”⁶

Terms to pull $L a r c e n y_{G T}$ : “my stolen + find my stolen + theft report police + someone stolen my − car − heart − dream − number.”

Terms to pull ${R a p e}_{G T}$ : “I raped + raped me + raped report police + being raped + been raped − dream − kobe − trump − porn.”⁷

UCR crime measures are derived from the Inter-University Consortium for Political and Social Research (Kaplan, 2021).⁸ MVT, burglary, larceny, homicide (combination murder and manslaughter), robbery, and rape are the crime types that we employ in the current study. We aggregated the counts of each category of crime in each county to the DMA level and then divided the total number of crimes by the DMA’s total population multiplied by 100,000 to determine the crime rate in each DMA using the UCR data. Finally, in order to align UCR crime rates with GT estimates (scaled 0–100), we employ a similar procedure to convert crime rates to an estimate of relative crime rates for each DMA. A number, 100, is assigned to the DMA with the highest crime rate as measured by the UCR. Other DMAs are relative numbers (e.g. 93, 80, 68…13) based on the ratio of each area’s crime rate to the highest crime rate seen in the UCR data.⁹

Notably, the GT data collection has certain missing information that must be addressed. When a DMA’s volume of a search query falls below a threshold (in terms of total searches for a specific term), GT removes that geographical unit from the results generated. This is because the volume of total searches in a region for the keywords is insufficient to produce credible estimates (Trends Help, 2023). When we enter a list of victimization keywords into GT, the GT rape estimates provide the sample size of 107 DMAs, which limits the size of our sample to these 107 DMAs. While the 107 included DMAs account for only 51% of all 210 DMAs, these 107 areas cover over 86.87% of the United States' population (see Supplemental Table 1).

Common covariates of crime rates

In the early 20th century, the Chicago School of Sociology, led by Robert Park, examined the relationship between crime and the urban environment (Park et al., 1967 [1925]). Their research revealed that social disorganization, represented by economic disadvantage, social mobility, and the percentage of immigrants, was associated with criminal rates in the inner city (Park et al., 1967 [1925]; Shaw and McKay, 1972 [1942]). Following their path, criminologists continue to work on the association between crime and place found that common crime covariates of crime rates are concentrated disadvantage (Sampson et al., 2008; Sampson and Groves, 1989; Wilson, 1987), social mobility (Lane and Meeker, 2004; South and Messner, 2000; Xie and McDowall, 2008), racial and ethnic heterogeneity (Avison and Loring, 1986; Blau, 1977; South and Messner, 1986), the percentage of foreign-born individuals (Butcher and Piehl, 1998; Reid et al., 2005; Wadsworth, 2010), the percentage of divorced individuals (Cáceres-Delpiano and Giolito, 2012; Sampson, 1986), the percentage of young males (Farrington, 1986; Moffitt, 1993), and drug mortality rate (Rajkumar and French, 1997).

The multiple linear regression (MLR) models presented below are used to assess the association between the GT and UCR data and the common covariates of crime rates in order to further investigate these relationships. The aim of MLR is not to determine whether the social disorganization theory is better suited to explain crime rates, but to utilize common crime covariates to evaluate the validity of GT and UCR. The independent variables in the MLR models are previously studied covariates of crime rates. These measures are drawn from three other sources of publicly accessible data: the American Community Survey (ACS), the National Center for Health Statistics (Centers for Disease Control and Prevention, 2020), and the Simply Analytics database (Simply Analytics, 2019).

The concentrated disadvantage (CD) index created by combining four standardized measures drawn from the ACS (Census Bureau, 2020): the percentage of the population aged 16 years and older who are unemployed, the percentage of families whose income is below the poverty level, the percentage of female-headed households with children aged under 18 years, and the percentage of the population aged over 25 years with less than high school education.¹⁰ The mobility index is created by standardizing and combining the percentage of people who move within the last year and the proportion renters.¹¹ To create a heterogeneity index, we calculate the probability that two randomly selected individuals in a DMA are of different races¹² (Blau, 1977). The other three predictor variables (the percentage of foreign-born individuals, the percentage of divorced individuals, and the percentage of the young males aged 15–24 years) are also derived from ACS county-level data, which are aggregated to DMA level. The measure of drug mortality rates is drawn from National Center for Health Statistics’ (NCHS) by County (Centers for Disease Control and Prevention, 2020).

Finally, in order to account for differences in the underlying population and potential differences in access to the internet we include an estimate of the total population drawn from the ACS, and an estimate of the percentage of homes with internet access and the median number of vehicles per household drawn from the MRI Consumer Survey (Simply Analytics, 2019). Table 1 provides summary statistics for all measures used in the current study.¹³

Table 1.

Descriptive statistics.

	M	SD	Minimum	Maximum
Validated crime measures
$M V T_{U C R}$	34.09	17.37	8.48	100.00
${B u r g l a r y}_{U C R}$	49.24	15.62	17.64	100.00
${H o m i c i d e}_{U C R}$	39.99	15.05	7.27	100.00
${R o b b e r y}_{U C R}$	19.61	12.01	5.65	100.00
Crime measures needs to be validated
$M V T_{G T}$	58.17	13.72	23.87	100.00
${B u r g l a r y}_{G T}$	62.52	13.53	26.48	100.00
$L a r c e n y_{G T}$	53.10	10.21	31.29	100.00
${R a p e}_{G T}$	62.88	11.26	38.90	100.00
$L a r c e n y_{U C R}$	59.83	13.23	35.28	100.00
${R a p e}_{U C R}$	50.97	15.42	21.37	100.00
Common covariates of crime
Concentrated disadvantage index	0	3.34	−6.11	14.42
Mobility index	0	1.68	−2.82	5.65
Heterogeneity index	0.45	0.15	0.11	0.71
% Foreign-born	9.63	7.26	1.13	32.43
% Divorced	9.26	1.21	5.58	11.70
% Young males	7.10	0.77	5.32	10.20
Drug mortality rate	173.79	60.42	44.26	397.77
Control variables
Population (log)	14.39	0.82	10.80	16.87
% Internet usage HH	81.64	0.96	79.31	84.19
Median vehicle HH	2.15	0.13	1.63	2.41

N = 107 DMAs.

HH: Household.

Results

Correlation analysis

As aforementioned, the previously validated measures of ${H o m i c i d e}_{U C R}$ and ${R o b b e r y}_{U C R}$ serves as criteria in violent crimes, as $M V T_{U C R}$ and ${B u r g l a r y}_{U C R}$ does in property crimes. The results are shown in Table 2 and Figure 1; Figures 2 and 3 represent the heatmaps of property crime and violent crime distribution of GT and UCR among DMAs for a better understanding of the data, as well as their similarities and differences.

Table 2.

Correlation of crime measures.

		(1)	(2)	(3)	(4)	(5)	(6)	(7)	(8)	(9)
$M V T_{G T}$	(1)
${B u r g l a r y}_{G T}$	(2)	0.293**
$L a r c e n y_{G T}$	(3)	0.811**	0.408**
${R a p e}_{G T}$	(4)	−0.081	0.241*	0.111
$M V T_{U C R}$	(5)	0.811**	0.265**	0.758**	−0.005
${B u r g l a r y}_{U C R}$	(6)	0.416**	0.456**	0.563**	0.485**	0.521**
$L a r c e n y_{U C R}$	(7)	0.323**	0.328**	0.398**	0.260**	0.448**	0.705**
${H o m i c i d e}_{U C R}$	(8)	0.159	0.194*	0.266**	0.421**	0.233*	0.623**	0.595**
${R o b b e r y}_{U C R}$	(9)	0.223*	0.191*	0.231*	0.449**	0.308**	0.497**	0.445**	0.525**
${R a p e}_{U C R}$	(10)	0.077	−0.035	0.189	0.267**	0.162	0.306**	0.288**	0.318**	0.061

N = 107 DMAs.

p < 0.05. **p < 0.01.

Figure 1.

Correlation matrix of crime measures.

Figure 2.

Heatmap of violent crimes from UCR and GT.

Figure 3.

Heatmap of property crimes from UCR and GT.

Our first focus is on the result of property crimes. Table 2 shows that $M V T_{G T}$ has a high correlation (r = 0.811, p < 0.01) with $M V T_{U C R}$ , and a moderate correlation with ${B u r g l a r y}_{U C R}$ (r = 0.416, p < 0.01). On the other hand, ${B u r g l a r y}_{G T}$ is only moderately correlated with ${B u r g l a r y}_{U C R}$ (r = 0.456, p < 0.01), and has a low correlation with $M V T_{U C R}$ (r = 0.265, p < 0.01). Finally, $L a r c e n y_{G T}$ has a high correlation with $M V T_{U C R}$ (r = 0.758, p < 0.01) and a moderate correlation with ${B u r g l a r y}_{U C R}$ (r = 0.563, p < 0.01). The results show that there is a moderate-to-strong correlation between $M V T_{U C R}$ , ${B u r g l a r y}_{U C R}$ , $M V T_{G T}$ , and $L a r c e n y_{G T}$ , and there is a low-to-moderate correlation between ${B u r g l a r y}_{G T}$ , ${B u r g l a r y}_{U C R}$ , and $M V T_{U C R}$ .

Related to violent crime, Table 2 shows that ${R a p e}_{G T}$ has a higher correlation with ${H o m i c i d e}_{U C R}$ (r = 0.421, p < 0.01) than the correlation between ${R a p e}_{U C R}$ and ${H o m i c i d e}_{U C R}$ (r = 0.318, p < 0.01). Similarly, ${R a p e}_{G T}$ has a higher correlation with ${R o b b e r y}_{U C R}$ (r = 0.449, p < 0.01) than the correlation between ${R a p e}_{U C R}$ and ${R o b b e r y}_{U C R}$ (r = 0.061, p > 0.05). The results indicate that ${R a p e}_{G T}$ is more strongly associated with violent crime criteria such as ${H o m i c i d e}_{U C R}$ and ${R o b b e r y}_{U C R}$ than ${R a p e}_{U C R}$ is.

Exploratory factor analysis (EFA)

The two-factor analysis results of EFA test are shown in Table 3. The latent variable Factor 1 has high positive loadings from items for $M V T_{G T}$ (0.97), $M V T_{U C R}$ (0.837), and $L a r c e n y_{G T}$ (0.819). The latent variable Factor 2 has high positive loadings from items for ${H o m i c i d e}_{U C R}$ (0.798), ${B u r g l a r y}_{U C R}$ (0.795), ${R a p e}_{G T}$ (0.692), $L a r c e n y_{U C R}$ (0.655), and ${R o b b e r y}_{U C R}$ (0.609). We find two nonsignificant items in EFA: ${R a p e}_{U C R}$ and ${B u r g l a r y}_{G T}$ .¹⁴ This finding again indicates that ${R a p e}_{U C R}$ and ${B u r g l a r y}_{G T}$ are not closely related to the other measures being examined.

Table 3.

Exploratory factor analysis.

	Factor 1	Factor 2
$M V T_{G T}$	0.970	−0.092
$M V T_{U C R}$	0.837	0.087
$L a r c e n y_{G T}$	0.819	0.134
${H o m i c i d e}_{U C R}$	−0.086	0.798
${B u r g l a r y}_{U C R}$	0.238	0.795
${R a p e}_{G T}$	−0.276	0.692
$L a r c e n y_{U C R}$	0.182	0.655
${R o b b e r y}_{U C R}$	0.012	0.609
${R a p e}_{U C R}$	−0.007	0.343
${B u r g l a r y}_{G T}$	0.247	0.279

N = 107 DMAs; χ² = 30.29, p < 0.01; RMSR = 0.06; RMSEA = 0.107; TLI = 0.894; Loading cutoff = 0.6; The grey text color represents the loadings under 0.6.

Confirmatory factor analysis (CFA)

The results of CFA models are shown in Table 4 (property crime) and Table 5 (violent crime). Among the first four property crime CFA models in Table 4, except for ${B u r g l a r y}_{G T}$ (0.482), the other tested measures ( $M V T_{G T}$ : 0.805, $L a r c e n y_{G T}$ : 0.906; and $L a r c e n y_{U C R}$ : 0.779) are considered associated with the criteria of property crime. For the other violent crime CFA models, ${R a p e}_{G T}$ (0.601) is more strongly associated with violent crime criteria while ${R a p e}_{U C R}$ (0.193) is not. The results of CFA again suggest that ${R a p e}_{U C R}$ and ${B u r g l a r y}_{G T}$ are inappropriate for inclusion in the latent variables, violent crime, and property crime, respectively.

Table 4.

Confirmatory factor analysis—property crimes.

Indicators	Standardized loadings	p-Value	RMSEA	TLI
Property crime model 1
$M V T_{U C R}$	1.007	0.000	0	1
${B u r g l a r y}_{U C R}$	0.517	0.000
$M V T_{G T}$	0.805	0.000
Property crime model 2
$M V T_{U C R}$	0.838	0.000	0	1
${B u r g l a r y}_{U C R}$	0.622	0.000
$L a r c e n y_{G T}$	0.906	0.000
Property crime model 3
$M V T_{U C R}$	0.575	0.000	0	1
${B u r g l a r y}_{U C R}$	0.905	0.000
$L a r c e n y_{U C R}$	0.779	0.000
Property crime model 4
$M V T_{U C R}$	0.550	0.000	0	1
${B u r g l a r y}_{U C R}$	0.946	0.000
${B u r g l a r y}_{G T}$	0.482	0.000
Property crime model 5
$M V T_{U C R}$	0.876	0.000	0.278	0.676
${B u r g l a r y}_{U C R}$	0.597	0.000
$L a r c e n y_{U C R}$	0.478	0.000
$M V T_{G T}$	0.888	0.000
${B u r g l a r y}_{G T}$	0.394	0.000
$L a r c e n y_{G T}$	0.896	0.000

N = 107 DMAs; Loading cutoff = 0.6; The grey text color represents the loadings under 0.6. Due to the degree of freedom is 0 from model 1 to 4, the results are “just identified models,” and we cannot access the model fit (that’s why the RMSEA is 0 and TLI is 1). Thus, we perform an extra CFA test in model 5 (OARC Stats, n.d.). Overall, when the RMSEA is smaller than 0.05, and the TLI is close to or greater than 0.95 (Harrington, 2009, p.52) shows a good-fit for CFA models. However, the purpose of this CFA test is to see how $L a r c e n y_{U C R}$ , $M V T_{G T}$ , ${B u r g l a r y}_{G T}$ , and $L a r c e n y_{G T}$ are associate with other types of property crimes criteria. The goodness-of-fit is not our main concern here.

RMSEA: root mean square error of approximation; TLI: Tucker–Lewis index.

Table 5.

Confirmatory factor analysis—violent crimes.

Indicators	Standardized loadings	p-Value	RMSEA	TLI
Violent crime model 1
${H o m i c i d e}_{U C R}$	0.702	0.000	0	1
${R o b b e r y}_{U C R}$	0.748	0.000
${R a p e}_{G T}$	0.601	0.000
Violent crime model 2
${H o m i c i d e}_{U C R}$	1.653	0.152	0	1
${R o b b e r y}_{U C R}$	0.317	0.188
${R a p e}_{U C R}$	0.193	0.244
Violent crime model 3
${H o m i c i d e}_{U C R}$	0.761	0.000	0.195	0.681
${R o b b e r y}_{U C R}$	0.675	0.000
${R a p e}_{U C R}$	0.325	0.003
${R a p e}_{G T}$	0.607	0.000

N = 107 DMAs; Loading cutoff = 0.6; The grey text color represents the loadings under 0.6. Due to the degree of freedom is 0 from model 1 to 2, the results are “just identified models,” and we cannot access the model fit (that’s why the RMSEA is 0 and TLI is 1). Thus, we perform the model 3 of CFA test (OARC Stats, n.d.). Overall, when the RMSEA is smaller than 0.05, and the TLI is close to or greater than 0.95 (Harrington, 2009, p.52) shows a good-fit for CFA models. However, the purpose of this CFA test is to see how ${R a p e}_{U C R}$ and ${R a p e}_{G T}$ associate with other types of violent crimes criteria. Thus, the goodness-of-fit is not our main concern here.

RMSEA: root mean square error of approximation; TLI: Tucker–Lewis index

Multiple linear regression analysis

The results of multiple linear regression analysis are shown in Table 6,¹⁵ and a summary of the results is shown in Table 7. To present the results of our multiple linear regression models more clearly, we display the ${C r i m e}_{G T}$ models alongside the ${C r i m e}_{U C R}$ models in Tables 6 and 7.

Table 6.

Multiple linear regression models for common crime covariates on crime measures.

	Dependent variables
	$M V T_{G T}$	$M V T_{U C R}$	${B u r g l a r y}_{G T}$	${B u r g l a r y}_{U C R}$	$L a r c e n y_{G T}$	$L a r c e n y_{U C R}$	${R a p e}_{G T}$	${R a p e}_{U C R}$	${H o m i c i d e}_{U C R}$	${R o b b e r y}_{U C R}$
	(1)	(2)	(3)	(4)	(5)	(6)	(7)	(8)	(9)	(10)
CD index	1.05** (0.39)	1.95** (0.46)	1.03* (0.49)	2.82** (0.41)	1.46** (0.26)	1.61** (0.44)	1.66** (0.30)	0.24 (0.56)	2.29** (0.47)	1.46** (0.25)
Mobility index	2.15* (1.02)	3.61** (1.20)	0.59 (1.27)	0.59 (1.07)	0.68 (0.69)	1.65 (1.16)	1.20 (0.78)	0.47 (1.46)	0.75 (1.22)	2.01** (0.66)
Heterogeneity index	28.36* (10.81)	34.23** (12.68)	13.00 (13.23)	37.95** (11.18)	7.45 (7.18)	19.27 (12.12)	−0.08 (8.12)	−4.01 (15.28)	30.19* (12.72)	46.70** (6.95)
% Foreign-born	0.38 (0.21)	0.49 (0.25)	−0.15 (0.26)	−1.03** (0.22)	0.33* (0.14)	−0.69** (0.24)	−1.02** (0.16)	−0.33 (0.30)	−0.99** (0.25)	−0.98** (0.14)
% Divorced	2.09 (1.39)	−0.35 (1.63)	3.05 (1.71)	4.41** (1.45)	3.70** (0.93)	2.18 (1.57)	−0.65 (1.05)	5.20* (1.98)	2.09 (1.65)	−3.91** (0.90)
% Young males	−2.11 (1.88)	−6.11** (2.20)	1.68 (2.23)	1.13 (1.89)	3.11* (1.21)	−1.09 (2.05)	−1.12 (1.37)	5.68* (2.58)	−1.69 (2.15)	−6.77** (1.17)
Drug mortality rate	0.04 (0.02)	0.04 (0.02)	−0.04 (0.02)	−0.01 (0.02)	0.02 (0.01)	−0.03 (0.02)	−0.002 (0.01)	−0.02 (0.03)	−0.03 (0.02)	0.04** (0.01)
Population (log)	2.64 (1.87)	0.21 (2.19)	−1.61 (2.33)	1.35 (1.97)	1.15 (1.27)	0.68 (2.14)	1.07 (1.43)	−0.32 (2.69)	−0.32 (2.24)	−4.07** (1.22)
% Internet usage HH	−0.33 (2.09)	−2.15 (2.45)	0.43 (2.12)	1.80 (1.79)	3.00* (1.15)	1.69 (1.94)	−2.90* (1.30)	2.12 (2.45)	0.82 (2.04)	−1.28 (1.11)
Median vehicle HH	38.87** (11.24)	74.74** (13.18)
Constant	−63.87 (160.43)	65.44 (188.06)	13.87 (174.55)	−171.41 (147.52)	−274.61** (94.77)	−97.37 (159.87)	308.27** (107.10)	−196.43 (201.60)	−28.89 (167.85)	247.94** (91.64)
RMSE	9.39	11.01	11.82	9.99	6.42	10.82	7.25	13.65	11.36	6.2
MAE	7.05	8.36	8.86	7.84	4.79	8.31	5.52	10.01	8.39	4.57
Observations	107	107	107	107	107	107	107	107	107	107
Adjusted R²	0.48	0.55	0.16	0.55	0.56	0.26	0.54	0.14	0.37	0.71
F Statistic	10.71**	14.09**	3.23**	15.34**	16.27**	5.17**	14.98**	2.86**	7.95**	29.22**

N = 107 DMAs.

HH: household; CD: concentrated disadvantage; RMSE: Root Meas Squared Error; MAE: Mean Absolute Error.

p < 0.05. **p < 0.01.

Table 7.

Summary of multiple regression results.

	$M V T_{G T}$	$M V T_{U C R}$	${B u r g l a r y}_{G T}$	${B u r g l a r y}_{U C R}$	$L a r c e n y_{G T}$	$L a r c e n y_{U C R}$	${R a p e}_{G T}$	${R a p e}_{U C R}$	${H o m i c i d e}_{U C R}$	${R o b b e r y}_{U C R}$
CD index	+	+	+	+	+	+	+		+	+
Mobility index	+	+								+
Heterogeneity index	+	+		+					+	+
% Foreign-born				−	+	−	−		−	−
% Divorced				+	+			+		−
% Young males		−			+			+		−
Drug mortality rate										+

“+” = significant positive association; “−” = significant negative association.

CD: concentrated disadvantage.

For the first set of comparisons, the results indicate that $M V T_{G T}$ model fits better than $M V T_{U C R}$ model because the Root Mean Square Error (RMSE) of the former is 9.39 for the latter is 11.01. A similar result can be found in the Mean Absolute Error (MAE) comparison. The MAE of $M V T_{G T}$ is 7.05, which is lower than 8.36 of $M V T_{U C R}$ . However, the adjusted $R^{2}$ value of $M V T_{G T}$ model is 0.48, lower than the $R^{2}$ value of 0.55 for the $M V T_{U C R}$ model. It indicates that the common crime covariates can explain 48% of the variance in $M V T_{G T}$ , and 55% of the variance in the $M V T_{U C R}$ . We also find that $M V T_{G T}$ is significantly related to the same common covariates as $M V T_{U C R}$ . The common covariates significantly related to both crime estimates are the CD index, mobility index, and heterogeneity index. These results indicate that $M V T_{G T}$ is significantly associated with many of the theoretical constructs, along with $M V T_{U C R}$ which serves as the criteria.

${B u r g l a r y}_{U C R}$ has a better model fit than ${B u r g l a r y}_{G T}$ . The RMSE of ${B u r g l a r y}_{U C R}$ (9.99) is lower than it of ${B u r g l a r y}_{G T}$ (11.82). Similarly, the MAE of ${B u r g l a r y}_{U C R}$ (7.84) is lower than it of ${B u r g l a r y}_{G T}$ (8.86). The adjusted $R^{2}$ for the ${B u r g l a r y}_{G T}$ model is only 0.16. It is substantially lower than the adjusted $R^{2}$ (0.55) for the ${B u r g l a r y}_{U C R}$ model. Additionally, except for the CD index, we find that none of the common covariates are significantly associated with ${B u r g l a r y}_{G T}$ . In contrast, ${B u r g l a r y}_{U C R}$ is significantly related to the CD index, heterogeneity index, the percentage of foreign-born, and the percentage of divorced. These results indicate that ${B u r g l a r y}_{G T}$ is not associated with many of the hypothesized constructs.

For the measures of larceny, $L a r c e n y_{G T}$ has a better model fit than $L a r c e n y_{U C R}$ . The RMSE for the $L a r c e n y_{G T}$ model is 6.42, which is almost half lower than 10.82 for the $L a r c e n y_{U C R}$ model. Similarly, the MAE of $L a r c e n y_{G T}$ (4.79) is lower than it of $L a r c e n y_{U C R}$ (8.31). The adjusted $R^{2}$ for the $L a r c e n y_{G T}$ model is 0.56, which is twice higher than 0.26 for the $L a r c e n y_{U C R}$ model. Moreover, $L a r c e n y_{G T}$ is significantly associated with more crime covariates than that in $L a r c e n y_{U C R}$ . These crime covariates are the CD index, the percentage of foreign-born, the percentage of divorced, and the percentage of young males. $L a r c e n y_{U C R}$ is significantly associated only with the CD index and the percentage of foreign-born. These results indicate that $L a r c e n y_{G T}$ is significantly associated with many of the theoretical constructs.

For rape, the model fit of ${R a p e}_{G T}$ is better than ${R a p e}_{U C R}$ . The RMSE for the ${R a p e}_{G T}$ model is 7.25, which is substantially lower than 13.65 for the ${R a p e}_{U C R}$ model. Similarly, the MAE of ${R a p e}_{G T}$ (5.52) is also lower than it of ${R a p e}_{U C R}$ (10.01). Additionally, the adjusted $R^{2}$ for the ${R a p e}_{G T}$ model is 0.54, which is almost quadruple higher than 0.14 for the ${R a p e}_{U C R}$ model. This indicates that ${R a p e}_{G T}$ accounts for 54% of the variance in the model while ${R a p e}_{U C R}$ only accounts for 14% of it. Although ${R a p e}_{G T}$ and ${R a p e}_{U C R}$ do not share the same significant common covariates, we find that both the ${R a p e}_{G T}$ model, ${H o m i c i d e}_{U C R}$ and ${R o b b e r y}_{U C R}$ model identify the CD index and the percentage of foreign-born as significant covariates. Whereas ${R a p e}_{U C R}$ has none identified common covariates with ${H o m i c i d e}_{U C R}$ , and opposite direction on the percentage of divorced and the percentage of young males with ${R o b b e r y}_{U C R}$ . These results indicate that ${R a p e}_{G T}$ is significantly associated with many of the theoretical constructs, along with the violent crime criteria, ${H o m i c i d e}_{U C R}$ .

Finding DMAs which may suffer from underreporting in UCR

Finally, we aim to demonstrate how ${C r i m e}_{G T}$ could inform policy. From the previous correlation analysis, EFA, CFA, and multiple linear regression analysis, we find evidence that ${R a p e}_{G T}$ may represent a more valid measure than ${R a p e}_{U C R}$ . If we have reasons to believe that ${R a p e}_{G T}$ is more valid than ${R a p e}_{U C R}$ , we can identify DMAs with underreporting, that is, DMAs with low reporting rape crimes in the UCR but with high victimization search volumes in GT.

First, we correlate ${R a p e}_{G T}$ with ${R a p e}_{U C R}$ , and the names of the data points (names of the DMAs) are shown in the same geometric graph. The median of the ${R a p e}_{G T}$ served as the origin point of the x-axis, and the median of the ${R a p e}_{U C R}$ served as the origin point of the y-axis. Therefore, the fourth (IV) quadrant is the underreported quadrant. It contains the metropolitan areas wherein ${R a p e}_{G T}$ appear in the top 50% of rape victimization searches. At the same time, it also contains the DMAs wherein ${R a p e}_{U C R}$ fall in the bottom 50% of criminogenic areas in rape crime. Simply knowing the DMAs have higher rape estimates in ${R a p e}_{G T}$ but lower estimates in ${R a p e}_{U C R}$ alone does not provide enough evidence of underreporting because this result may be the result of random errors. We still need another indicator to identify the DMAs where their ${R a p e}_{U C R}$ is significantly/extremely underreported compared with the ${R a p e}_{G T}$ . Therefore, we use Cook’s distance to find the outliers of the regression of ${R a p e}_{G T}$ and ${R a p e}_{U C R}$ . The high influence¹⁶ of data points in the underreported quadrant is identified as DMAs with underreporting ${R a p e}_{U C R}$ . The result is shown in Figure 4. We find that ${R a p e}_{U C R}$ is heavily underreported in the areas of Jackson, MS, and Charleston-Huntington, WV. This analysis, albeit simple, demonstrates the potential utility of examining the prevalence of crime using ${R a p e}_{G T} .$

Figure 4.

Underreported areas of ${R a p e}_{U C R}$ .

Discussion and conclusion

Our correlation analyses (see Table 2) suggest that $M V T_{G T}$ and $M V T_{U C R}$ are highly correlated in MVT, the type of crime that approximately 80% of victims report to the police. We also find that ${R a p e}_{G T}$ may be considered a better estimate of the true estimate of rape because it has a higher correlation with ${H o m i c i d e}_{U C R}$ (r = 0.421, p < 0.01) and ${R o b b e r y}_{U C R}$ (r = 0.449, p < 0.01) than with ${R a p e}_{U C R}$ does with Homicide_UCR and Robbery_UCR (r = 0.318, p < 0.01; r = 0.061, p > 0.05, respectively).

The findings of EFA, and CFA provide additional evidence that measures from ${C r i m e}_{G T}$ remain reliable and closer to the criteria of crime measures except ${B u r g l a r y}_{G T}$ . Our search for latent variables (property crime and violent crime from UCR and GT) confirms that ${R a p e}_{U C R}$ is not eligible for combining into the latent variable “violent crime,” while ${R a p e}_{G T}$ is. We conclude that ${R a p e}_{G T}$ may be more valid than ${R a p e}_{U C R}$ , as the latter is not significantly associated with violent crime criteria in the EFA, and CFA.

Our MLR model results further reaffirm our reservations about the use of official crime statistics in assessing common variables of crime when the dark figure of crime is high. It is problematic to use official rates of underreported crime types to determine their association with broadly accepted structural factors. We discovered the significant predictors of ${R a p e}_{U C R}$ are distinct from those of ${R a p e}_{G T}$ and ${H o m i c i d e}_{U C R}$ . Furthermore, the regression model with ${R a p e}_{G T}$ explains 54% of the overall variance, but the regression model with ${R a p e}_{U C R}$ explains only 14% of the whole variance. Across the three forms of crime, MVT, larceny, and rape, the regression models using ${C r i m e}_{G T}$ consistently retain a high adjusted $R^{2}$ of 0.5, however, the adjusted $R^{2}$ in the regression models using ${C r i m e}_{U C R}$ decreases as the underreporting percentage increases.

For GT data to be useful in crime estimation, our study makes certain underlying assumptions. To begin, the internet has permeated practically every adult’s life in the United States, and the majority of internet users rely on Google. As a result, victims of crime use Google to look up solutions to crime situations. Second, regardless of the reasons for not reporting a crime to the police, these reasons would not prohibit victims from “googling” solutions. Offense victims may opt to “google” remedies to a variety of crimes, from MVT to rape, regardless of whether they report the crime to the police. If researchers can develop a consistent and effective approach to extract ${C r i m e}_{G T}$ using victimization terms, we would anticipate that the reporting error for ${C r i m e}_{G T}$ would be more consistent across crime types than the reporting error for ${C r i m e}_{U C R}$ . Third, non-victims may use victimization terms in Google Search, but there are no significant regional differences. Fourth, there are no regional disparities in victims’ search and reporting behaviors when it comes to different forms of crime. Fifth, there are no systemic changes in the way crimes are recorded across different crime categories or regions. For future research to confirm these assumptions, a national survey with a sufficient sample size will be required.

With these assumptions, we may argue that ${C r i m e}_{G T}$ could serve as an additional source of crime estimates for various types of crime. It may be a valuable source for estimating crime when there is substantial concern about reporting loss in official crime data. Criminological investigations attempting to find factors influencing crime levels may choose to utilize ${C r i m e}_{G T}$ to produce crime measurements. This article proposes a novel method for exploring the relationship among official crime data, GT data, and common crime covariates. By clarifying the dark figure of crime and reassembling these missing pieces of the puzzle, we illustrate the feasibility of this new approach for obtaining potentially more accurate crime estimates.

Overall, we can conclude that $M V T_{G T}, L a r c e n y_{G T}$ , and ${R a p e}_{G T}$ can be used to supplement insights of crime measures, and ${R a p e}_{G T}$ is a better measurement than ${R a p e}_{U C R}$ . This conclusion also aligns with our assumption that victims use online search as a help-seeking behavior so that when the official reporting rate decreases, the data discrepancy increases between GT and UCR across different crime categories. McCart et al. (2010) noted that just a small fraction of crime victims seek assistance from formal networks (reporting to police or mental health service); however, many of them turn to informal sources for help (family and friends), and informal support has been linked with positive feedback and psychological adjustment, especially in sexual assault and intimate partner violence. If we assume victims will search online for useful information, this can be viewed as a kind of informal help-seeking behavior and thus can explain the difference between UCR and GT, especially when it comes to rape.

Although these findings are encouraging in addressing the dark figure of crime, we should consider several limitations in the use of GT as a tool for studying crime. First, although GT’s “topic” keyword search includes multiple languages in the data, we used “terms” to obtain “victim-oriented” data. Our data are limited to English users in the United States and cannot be generalized to populations of non-English users. According to data from Census Bureau (Dietrich and Hernandez, 2022), approximately one in five (67.8 million) Americans speak a language other than English at home in 2022. Researchers who wish to use GT data without a “topic” keyword search should keep this limitation in mind.

Second, there are victims who cannot engage in search behavior due to the nature of the victimization. For instance, if the victims lose their laptop or cell phone, they are unable to perform an online search. It also applies to victims of severe attacks or mental disorders who are incapable of typing or searching. Similarly, we acknowledged that there are regional differences in people’s trust in law enforcement organizations (Taylor et al., 2015), the classification of crimes (Wells and Weisheit, 2004), and the application of the law (Beattie, 1960); therefore, additional research is required to determine how these regional differences affect victim’s internet search help-seeking behaviors.

Third, ${C r i m e}_{G T}$ may be easily affected by search interests in movies, video games, events, or slang with keywords that are similar to our search terms. For example, there is a popular video game named Grand Theft Auto, and it would be misleading if we use “auto theft” as the only keyword to obtain the $M V T_{G T}$ . However, this limitation may be avoided if the researcher looks up the “related topics” category on the GT website. This function provides information about popular search phrases related to our queries in GT, and we use “−” to exclude unwanted phrases. The word “dream” was used as an exclusion in every search term. It appears most frequently in rape-related terms in GT, and it occasionally appears in other contexts as well. We excluded “dream” because people who search “dream” with crime terms do not necessarily want to discover a solution for a victimization experience but for the implication of that “dream.” The frequent occurrence of the phrase “dream” in GT and its association with “fear of crime” merits further investigation and research.

Fourth, the ${C r i m e}_{G T}$ used in this study are at the metropolitan level. A DMA is a relatively large geographic unit compared to street-, neighborhood-, or even city- or county-level analyses, which are most common in studies of criminological theory. Thus, a DMA includes both rural and metropolitan areas and cannot address the underreporting problems in rural areas. However, DMA-level data is the lowest national-level data that we can obtain from GT. This result is because GT data do not provide information at lower levels of aggregation due to privacy issues. Moreover, owing to the geographic level limitation on DMA of this study, the publicly accessible NCVS, which only provides state-level data, cannot be used here. The future direction would be to use state-level data from both GT and NCVS for a national comparison of crime victimization. In future research, we suggest gaining access to the restrictive data from NCVS and aggregating the data to the DMA level for further comparisons to GT.

Fifth, the cross-sectional approach in the current study cannot confirm a causal relationship between the covariates and crime outcomes, and we do not assess whether GT data captures fluctuations in crime over time. Sixth, not all types of crime victimization may be found in GT data. For example, homicide victims cannot perform Google searches. In the future, when the GT data is available at the DMA level for a time-series analysis, the authors will continue to investigate the validity of GT.

Seventh, GT data are sensitive to the keywords that researchers use to obtain data. A subtle difference in keywords may lead to different output data. For example, the present study uses a single keyword, “burglary,” in combination with other victimization terms to obtain ${B u r g l a r y}_{G T}$ . We do not combine any victim-oriented words, such as “my,”“I,” and “been” with this single term “burglary” because solely incorporating these terms in GT would result in a smaller viable sample in ${B u r g l a r y}_{G T}$ . However, our ${B u r g l a r y}_{G T}$ , which adds the single keyword “burglary,” is the only item that is not valid among all of the measures in ${C r i m e}_{G T}$ . Similarly, our multiple linear regression analysis finds no common covariates to explain ${B u r g l a r y}_{G T}$ , and the model’s adjusted $R^{2}$ is much lower than the other models using ${C r i m e}_{G T}$ that contain victim-oriented search terms. Not using “victimization” terms in searching for ${C r i m e}_{G T}$ violates the second assumption that the search patterns of terms should be consistent. One limitation of using “I” in search terms is the possibility of combining perpetrator and victim searches (e.g. “I raped”). It is also possible that during the “MeToo” movement, the perpetrators will use Google search for legal advice. Future research could focus on how the victim’s or perpetrator’s online search behavior influences the results of GT.

Eighth, although it is plausible to use the 0–100 range to represent the data in the search, the percentage cannot represent the total count of search volumes or crime rates. Therefore, we only know the ranking and the relative seriousness of crime among DMAs, and we do not know the actual incidence of crime in a given area. Our concept of ${C r i m e}_{T r u e}$ is the relative crime rates among areas. Therefore, ${C r i m e}_{G T}$ needs to be coupled with ${C r i m e}_{U C R}$ to estimate the actual counts of crimes. The other problem of this 0–100 range comparison is that the numerators and denominators of ${C r i m e}_{U C R}$ and ${C r i m e}_{G T}$ are different. The numerator of ${C r i m e}_{U C R}$ is the actual count of crimes in a DMA, and the denominator is the population in a DMA. For ${C r i m e}_{G T}$ , however, the numerator is the search volume of specific victimization keywords in a DMA, and the denominator is the total search volume from the population in a DMA. These differences may represent a different source of error, especially if the use of Google Search varies across geographic areas in any systematic fashion.

Finally, GT should not and cannot be considered a scientific survey (Trends Help, 2023). The primary issue is that the method used for data collection is not for scientific investigation. Additionally, GT only contains the online search data of the people with internet/computer/mobile phone access, and the data is likely to be skewed towards the population that utilizes internet/search engines more than other types of population and increases the systematic errors in the data. To address the issue of discrepancy in internet access, we incorporated “the percentage of internet usage” as a control variable into our MLR model, finding its inclusion did not impact the results presented (Supplemental Table 4).

This paper is a bold attempt to use GT as an alternative source to generate crime data, and the exploratory nature of this paper generates hope and doubts. Despite the limitations and unsolved issues present in the current study, we are convinced that GT, if effectively utilized, may be a valuable tool for understanding the reality of crime problems. As ${C r i m e}_{G T}$ remains underexplored in the field of criminology, we think it is worth borrowing from research in the public health field to discuss the value of big data within the field of criminology. Lazer et al. (2014) noted that GT flu data were inaccurate at the time of the highest flu peak because the estimates from GT doubled the actual number of flu cases from CDC data. They concluded that GT data still cannot represent/replace the official estimate. This is true when the official data are reliable and valid. However, this is not the case for some measures of crime drawn from the UCR, such as rape, a type of crime statistic which is heavily underreported. At the end of their article, Lazer et al. (2014) also appealed to scientists to “use big data to understand the unknown.” If a method can better estimate the true volume of rape occurring in metropolitan areas, we should not fail to recognize its importance and utility. Therefore, we believe there is great potential in the use of GT to measure crime, especially for those which are underreported. This application of GT requires further studies to explore its potential in innovative ways.

We do not claim that ${C r i m e}_{G T}$ should replace UCR or NCVS data. Our suggestion is to use ${C r i m e}_{G T}$ to supplement, not supplant traditional crime measurements. Both ${C r i m e}_{U C R}$ and ${C r i m e}_{G T}$ have their own limitations, and they can supplement each other. ${C r i m e}_{G T}$ demonstrated the potential to detect areas with high crime rates and profound underreporting in ${C r i m e}_{U C R}$ . However, ${C r i m e}_{U C R}$ can serve as criteria to validate ${C r i m e}_{G T}$ when the dark figure of crime is not a problem for the relevant crime types, such as MVT.

This study explored how ${C r i m e}_{G T}$ may be used to identify areas of underreporting and may be more comparable to ${C r i m e}_{T r u e}$ than the traditional crime assessment. Governments and law enforcement agencies can utilize this approach to understand/examine the underlying meaning behind official crime statistics, enabling them to develop policies and conduct measures based on the GT victimization search information and pay more attention to high-search but low-reporting districts, providing further information and assistance to victims in the areas, and taking additional preventative measures against crime.

Research Data

sj-csv-2-mio-10.1177_20597991231183962 – Supplemental material for Big data in crime statistics: Using Google Trends to measure victimization in designated market areas across the United States

Supplemental material, sj-csv-2-mio-10.1177_20597991231183962 for Big data in crime statistics: Using Google Trends to measure victimization in designated market areas across the United States by Yu-Hsuan Liu, Kevin T Wolff and Tzu-Ying Lo in Methodological Innovations

Supplemental Material

sj-docx-1-mio-10.1177_20597991231183962 – Supplemental material for Big data in crime statistics: Using Google Trends to measure victimization in designated market areas across the United States

Supplemental material, sj-docx-1-mio-10.1177_20597991231183962 for Big data in crime statistics: Using Google Trends to measure victimization in designated market areas across the United States by Yu-Hsuan Liu, Kevin T Wolff and Tzu-Ying Lo in Methodological Innovations

This article is distributed under the terms of the Creative Commons Attribution 4.0 License (http://www.creativecommons.org/licenses/by/4.0/) which permits any use, reproduction and distribution of the work without further permission provided the original work is attributed as specified on the SAGE and Open Access pages (https://us.sagepub.com/en-us/nam/open-access-at-sage).

Footnotes

Acknowledgements

The authors express their gratitude to Dr. Amy Adamczyk, Dr. Michael Maxfield, and Dr. Chongmin Na for their valuable advice during the initial phase of this manuscript.

Declaration of conflicting interests

The author(s) declared no potential conflicts of interest with respect to the research, authorship, and/or publication of this article.

Funding

The author(s) received no financial support for the research, authorship, and/or publication of this article.

ORCID iD

Yu-Hsuan Liu

Supplemental material

Supplemental material for this article is available online.

Notes

Author biographies

Yu-Hsuan (Sean) Liu is an Assistant Professor at St. Johns University, the Department of Criminal Justice and Homeland Security & The Lesley H. and William L. Collins College of Professional Studies in 2023. He is also a doctoral candidate in Criminal Justice at John Jay College/The Graduate Center, City University of New York. He worked as the Chief of the New Taipei City Police Department in Taiwan for twelve years. He was the recipient of the Study Abroad Scholarship from the Ministry of Education in Taiwan, the Early Research Initiative Pre-Dissertation Research Award from The Graduate Center, CUNY, and the James Fyfe Award from John Jay College of Criminal Justice. His research interests include terrorism, crime statistics, and the application of data science techniques to the social science research. His works have appeared in Social Science Research, Methodological Innovations, the Police Practice and Research: An International Journal, and Journal of Human Trafficking.

Kevin T Wolff is an associate professor at John Jay College of Criminal Justice in New York City. He also serves as the research advisor to the doctoral program. He earned his PhD from the College of Criminology and Criminal Justice at Florida State University. Professor Wolff's research interests include adverse childhood experiences, dynamic risk assessment, program evaluation, and quantitative methods. A number of his research projects have been funded by the National Institute of Justice. He recently received The Feliks Gross Award from The City University of New York and the Tory J. Caeti Memorial Award from The Academy of Criminal Justice Sciences in recognition of his contribution to the discipline.

Tzu-Ying Lo is an assistant professor at St. John's University in the Department of Criminal Justice & Homeland Security, Collins College of Professional Studies. She earned her Ph.D. from John Jay College of Criminal Justice, The City University of New York. Professor Lo's research interests include community policing, police legitimacy, victimization and misconduct in prisons, victim-offender overlap, and juvenile justice.

References

Adamczyk

Barrera-Vilert

LaFree

(2019) Using Google and Twitter to measure, validate and understand views about religion across Africa. Society 56(3): 231–240.

Adamczyk

Liu

Scott

(2021a) Understanding the role of religion in shaping cross-national and domestic attitudes and interest in abortion, homosexuality, and pornography using traditional and Google search data. Social Science Research 100: 102602.

Adamczyk

Scott

Hitlin

(2021b) Using Internet-derived data to measure religion: Understanding how Google can provide insight into cross-national religious differences. Sociology of Religion 83(2).

Allen

(2007) The reporting and underreporting of rape. Southern Economic Journal 73(3): 623–641.

Anderberg

Rainer

Siuda

(2022) Quantifying domestic violence in times of crisis: An internet search activity-based measure for the COVID-19 pandemic. Journal of the Royal Statistical Society Series A: Statistics in Society, 185(2): 498–518.

Arora

McKee

Stuckler

(2019) Google trends: Opportunities and limitations in health and health policy research. Health Policy (Amsterdam) 123(3): 338–341.

Avakame

Fyfe

McCoy

(1999) “Did you call the police? What did they do?” An empirical assessment of Black’s theory of mobilization of law. Justice Quarterly 16(4): 765–792.

Avison

Loring

(1986) Population diversity and cross-national homicide: The effects of inequality and heterogeneity. Criminology 24(4): 733–749.

Ayyoubzadeh

Zahedi

, et al. (2020) Predicting COVID-19 incidence through analysis of Google trends data in Iran: Data mining and deep learning pilot study. JMIR Public Health and Surveillance 6(2): e18828.

10.

Bagozzi

(1988) On the evaluation of structural equation models. Journal of the Academy of Marketing Science 16: 74–94.

11.

Beattie

(1960) Criminal statistics in the United States—1960. The Journal of Criminal Law, Criminology, and Police Science 51(1): 49–65.

12.

Bergkvist

Rossiter

(2007) The predictive validity of multiple-item versus single-item measures of the same constructs. Journal of Marketing Research 44(2): 175–184.

13.

Black

(1970) Production of crime rates. American Sociological Review 35: 733–748.

14.

Blau

(1977) Inequality and Heterogeneity: A Primitive Theory of Social Structure. New York: Free Press.

15.

Blumstein

Cohen

Rosenfeld

(1991) Trend and deviation in crime rates: Comparison of UCR and NCS data for burglary and robbery. Criminology, 29(2): 237–264.

16.

Bohn

(2003) Lifetime physical and sexual abuse, substance abuse, depression, and suicide attempt among Native American women. Issues in Mental Health Nursing 24(3): 333–352.

17.

Brodeur

Clark

Fleche

, et al. (2021) COVID-19, lockdowns and well-being: Evidence from Google Trends. Journal of Public Economics 193: 104346.

18.

Bureau of Justice Statistics (2019) Criminal victimization, 2018. Available at: https://www.bjs.gov/content/pub/pdf/cv18.pdf (accessed 28 December 2020).

19.

Butcher

Piehl

(1998) Cross-city evidence on the relationship between immigration and crime. Journal of Policy Analysis and Management 17(3): 457–493.

20.

Cáceres-Delpiano

Giolito

(2012) The impact of unilateral divorce on crime. Journal of Labor Economics 30(1): 215–248.

21.

Census Bureau (2020) Selected demographic characteristics, 2005–2019 American Community Survey 5-year estimates. Available at: https://www.socialexplorer.com/explore-tables (accessed 28 September 2020).

22.

Centers for Disease Control and Prevention (2020) NCHS-Drug Poisoning Mortality by County: United States. National Center for Health Statistics. Available at: https://healthdata.gov/dataset/nchs-drug-poisoning-mortality-county-united-states-1 (accessed 1 March 2021).

23.

Churchill

(1979) A paradigm for developing better measures of marketing constructs. Journal of Marketing Research 16(1): 64–73.

24.

Costello

Osborne

(2005) Best practices in exploratory factor analysis: Four recommendations for getting the most from your analysis. Practical Assessment, Research & Evaluation 10(7): 1–9.

25.

de Winter

JCF

Dodou

Wieringa

(2009) Exploratory factor analysis with small sample sizes. Multivariate Behavioral Research 44(2): 147–181.

26.

Dietrich

Hernandez

(2022) What Languages Do We Speak in the United States? Census Bureau. Available at: https://www.census.gov/library/stories/2022/12/languages-we-speak-in-united-states.html (accessed 31 March 2023).

27.

Fabrigar

Wegener

(2012) Exploratory Factor Analysis. New York: Oxford University Press.

28.

Farrington

(1986) Age and crime. Crime and Justice 7: 189–250.

29.

Fox

Lane

Akers

(2010) Do perceptions of neighborhood disorganization predict crime or victimization? An examination of gang member versus non-gang member jail inmates. Journal of Criminal Justice 38(4): 720–729.

30.

Fyfe

Klinger

Flavin

(1997) Differential police treatment of male-on-female spousal violence. Criminology 35(3): 455–473.

31.

Gamma

Schleifer

Weinmann

, et al. (2016) Could Google trends be used to predict methamphetamine-related crime? An analysis of search volume data in Switzerland, Germany, and Austria. PLoS ONE 11(11): e0166566.

32.

Ginsberg

Mohebbi

Patel

, et al. (2009) Detecting influenza epidemics using search engine query data. Nature 457(7232): 1012–1014.

33.

Goudriaan

Lynch

Nieuwbeerta

(2004) Reporting to the police in western nations: A theoretical analysis of the effects of social context. Justice Quarterly 21(4): 933–969.

34.

Gove

Hughes

Geerken

(1985) Are uniform crime reports valid indicator of the index crimes an affirmative answer with minor qualifications? Criminology 23(3): 451–502.

35.

Gregory

Lees

(1996) Attrition in rape and sexual assault cases. The British Journal of Criminology 36(1): 1–17.

36.

Gross

Mann

(2017) Is there a “ferguson effect?” Google searches, concern about police violence, and crime in U.S. cities, 2014–2016. Socius 3: 2378023117703122.

37.

Gutierrez

Kirk

(2017) Silence speaks: The relationship between immigration and the underreporting of crime. Crime & Delinquency 63(8): 926–950.

38.

Harrington

(2009) Confirmatory Factor Analysis. New York: Oxford University Press.

39.

Henson

Roberts

(2006) Use of exploratory factor analysis in published research: Common errors and some comment on improved practice. Educational and Psychological Measurement 66(3): 393–416.

40.

Hindelang

(1974) Decisions of shoplifting victims to invoke the criminal justice process. Social Problems (Berkeley, Calif.) 21(4): 580–593.

41.

Hine

Murphy

Yesberg

, et al. (2021) Mapping the landscape of male-on-male rape in London: An analysis of cases involving male victims reported between 2005 and 2012. Police Practice and Research 22(1): 109–126.

42.

Hoehn-Velasco

Silverio-Murillo

de la Miyar

JRB

(2021) The great crime recovery: Crimes against women during, and after, the COVID-19 lockdown in Mexico. Economics & Human Biology 41: 100991.

43.

Jackson

(1988) Assessing the validity of official data on arson. Criminology 26(1): 181–195.

44.

Kaplan

(2021) Jacob Kaplan’s Concatenated Files: Uniform Crime Reporting Program Data: Offenses Known and Clearances by Arrest, 1960–2019. Ann Arbor, MI: Inter-university Consortium for Political and Social Research. Available at: https://doi.org/10.3886/E100707V16 U.S. (accessed 16 January 2021).

45.

Lane

Meeker

(2004) Social disorganization perceptions, fear of gang crime, and behavioral precautions among Whites, Latinos, and Vietnamese. Journal of Criminal Justice 32(1): 49–62.

46.

Laub

(1981) Ecological considerations in victim reporting to the police. Journal of Criminal Justice 9(6): 419–430.

47.

Lauritsen

Rezey

Heimer

(2016) When choice of data matters: Analyses of U.S. crime trends, 1973–2012. Journal of Quantitative Criminology 32(3): 335–355.

48.

Lazer

Kennedy

King

, et al. (2014) The parable of Google flu: Traps in big data analysis. Science (American Association for the Advancement of Science) 343(6176): 1203–1205.

49.

Maltz

Targonski

(2002) A note on the use of county-level UCR data. Journal of Quantitative Criminology, 18: 297–318.

50.

Matthews

Johnson

Jenks

(2011) Does religious involvement generate or inhibit fear of crime? Religions 2(4): 485–503.

51.

Mavragani

Ochoa

(2019) Google trends in infodemiology and infoveillance: Methodology framework. Public Health and Surveillance 5(2): e13439.

52.

Maxfield

(1999) The national incident-based reporting system: Research and policy applications. Journal of Quantitative Criminology 15: 119.

53.

McBride

Shannon

Bingham

, et al. (2020) Underreporting of violence to police among women sex workers in Canada: Amplified inequities for im/migrant and in-call workers prior to and following end-demand legislation. Health and Human Rights 22(2): 257.

54.

McCart

Smith

Sawyer

(2010) Help seeking among victims of crime: A review of the empirical literature. Journal of Traumatic Stress: Official Publication of The International Society for Traumatic Stress Studies 23(2): 198–206.

55.

Mellon

(2013) Where and when can we use Google trends to measure issue salience? PS, Political Science & Politics 46(2): 280–290.

56.

Mellon

(2014) Internet search data and issue salience: The properties of Google trends as a measure of issue salience. Journal of Elections, Public Opinion & Parties 24(1): 45–72.

57.

Moffitt

(1993) Adolescence-limited and life-course-persistent antisocial behavior: A developmental taxonomy. Psychological Review 100(4): 674–701.

58.

OARC Stats (n.d.) Confirmatory Factor Analysis (CFA) in R With Lavaan. UCLA Advanced Research Computing, Statistical Methods and Data Analytics. Available at: https://stats.oarc.ucla.edu/r/seminars/rcfa/#s2a (accessed 12 January 2023).

59.

Obermeier

Berry

Klockow-McClain

, et al. (2022). The creation of a research television studio to test probabilistic hazard information with broadcast meteorologists in NOAA's Hazardous Weather Testbed. Weather, Climate, and Society 14(3): 949–963.

60.

O’Brien

(1996) Police productivity and crime rates: 1973–1992. Criminology 34(2): 183–228.

61.

O’Brien

Shichor

Decker

(1980) An empirical comparison of the validity of UCR and NCS crime rates. The Sociological Quarterly 21(3): 391–401.

62.

Park

Burgess

McKenzie

(1967) The City (1925). Chicago: Chicago UP, p. 1.

63.

Parker

Cuthbertson

Loveridge

, et al. (2016) Forecasting state-level premature deaths from alcohol, drugs, and suicides using Google trends data. Journal of Affective Disorders 213: 9–15.

64.

Pew Research Center (2012) Search engine use over time. Available at: https://www.pewresearch.org/internet/2012/03/09/main-findings-11/ (accessed 4 February 2021).

65.

Pew Research Center (2019) Internet/broadband fact sheet. Available at: https://www.pewresearch.org/internet/fact-sheet/internet-broadband/ (accessed 4 February 2021).

66.

Piña-García

Ramírez-Ramírez

(2019) Exploring crime patterns in Mexico City. Journal of Big Data 6: 1–21.

67.

Quasius

(2008) Native American rape victims: Desperately seeking an Oliphant-fix. Minnesota Law Review 93: 1902.

68.

Rajkumar

French

(1997) Drug abuse, crime costs, and the economic benefits of treatment. Journal of Quantitative Criminology 13(3): 291–323.

69.

Reid

Weiss

Adelman

, et al. (2005) The immigration–crime relationship: Evidence across US metropolitan areas. Social Science Research 34(4): 757–780.

70.

Rowlands

Nicholas

Williams

, et al. (2008) The Google generation: The information behaviour of the researcher of the future. Aslib Proceedings 60(4): 290–310

71.

Salganik

(2017) Bit by Bit: Social Research in the Digital Age. Princeton, NJ: Princeton University Press.

72.

Sampson

(1986) Crime in cities: The effects of formal and informal social control. Crime and Justice 8: 271–311.

73.

Sampson

Groves

(1989) Community structure and crime: Testing social-disorganization theory. American Journal of Sociology 94(4): 774–802.

74.

Sampson

Sharkey

Raudenbush

(2008) Durable effects of concentrated disadvantage on verbal ability among African-American children. Proceedings of the National Academy of Sciences of the United States of America, 105(3): 845–852.

75.

Scheitle

(2011) Google’s insights for search: A note evaluating the use of search engine data in social research. Social Science Quarterly 92(1): 285–295.

76.

Shaw

McKay

(1942) Juvenile Delinquencies and Urban Areas, 3rd edn. Chicago: University of Chicago Press.

77.

Sides

Vavreck

Warshaw

(2022) The effect of television advertising in united states elections. American Political Science Review 116(2): 702–718.

78.

Simply Analytics (2019) MRI Consumer Survey US 2011–2019. Simply Analytics Database. Available at: https://simplyanalytics.com/ (accessed 28 September 2020).

79.

Skogan

(1974) The validity of official crime statistics: An empirical investigation. Social Science Quarterly 55(1): 25–38.

80.

Skogan

(1977) Dimensions of the dark figure of unreported crime. Crime and Delinquency 23(1): 41–50.

81.

Skogan

(1984) Reporting crimes to the police: The status of world research. The Journal of Research in Crime and Delinquency 21(2): 113–137.

82.

Sood

(2016) Geographic Information on Designated Media Markets. Harvard Dataverse Database. Available at: https://dataverse.harvard.edu/dataset.xhtml?persistentId=doi:10.7910/DVN/IVXEHT (accessed 1 December 2020).

83.

Sood

Iyengar

(2016) Coming to dislike your opponents: The polarizing impact of political campaigns. https://ssrn.com/abstract=2840225.

84.

South

Messner

(1986) Structural determinants of intergroup association: Interracial marriage and crime. American Journal of Sociology 91(6): 1409–1430.

85.

South

Messner

(2000) Crime and demography: Multiple linkages, reciprocal relations. Annual Review of Sociology 26: 83–106.

86.

Statcounter (2021) Search engine market share United States of America. Available at: https://gs.statcounter.com/search-engine-market-share/all/united-states-of-america (accessed 31 March 2021).

87.

Stephens-Davidowitz

(2014) The cost of racial animus on a black candidate: Evidence using Google search data. Journal of Public Economics 118: 26–40.

88.

Strauss

Smith

(2009) Construct validity, advances in theory and methodology. Annual Review of Clinical Psychology 5(1): 1–25.

89.

Stubbs-Richardson

Cosby

Bergene

, et al. (2018) Searching for safety: Crime prevention in the era of Google. Crime Science 7(1): 1–13.

90.

Taylor

Wyant

Lockwood

(2015) Variable links within perceived police legitimacy: Fairness and effectiveness across races and places. Social Science Research 49: 234–248.

91.

Tjaden

Thoennes

(2006) Extent, nature, and consequences of rape victimization: Findings from the National Violence Against Women Survey. Washington, DC: National Criminal Justice Reference Service.

92.

Thomas

Wolff

(2021) Crime distortion within the NYPD: A potential method for estimating crime misclassification within CompStat statistics. Police Practice and Research 22(4): 1390–1407.

93.

Topalli

Wright

Fornango

(2002) Drug dealers, robbery and retaliation. Vulnerability, deterrence and the contagion of violence. British Journal of Criminology 42(2): 337–351.

94.

Trends Help (2023) FAQ about Google Trends data. Available at: https://support.google.com/trends/answer/4365533?hl=en&ref_topic=6248052 (accessed 31 March 2023).

95.

Wadsworth

(2010) Is immigration responsible for the crime drop? An assessment of the influence of immigration on changes in violent crime between 1990 and 2000. Social Science Quarterly 91(2): 531–553.

96.

Wells

Weisheit

(2004) Patterns of rural and urban crime: A county-level comparison. Criminal Justice Review 29(1): 1–22.

97.

Willmott

Matsuura

(2005) Advantages of the mean absolute error (MAE) over the root mean square error (RMSE) in assessing average model performance. Climate Research 30(1): 79–82.

98.

Wilson

(1987) The Truly Disadvantaged: The Inner City, the Underclass, and Public Policy. Chicago: University of Chicago Press.

99.

Wolf

Harrington

Clark

, et al. (2013) Sample size requirements for structural equation models: An evaluation of power, bias, and solution propriety. Educational and Psychological Measurement 73(6): 913–934.

100.

Xie

McDowall

(2008) Escaping crime: The effects of direct and indirect victimization on moving. Criminology 46(4): 809–840.

101.

Xie

Pogarsky

Lynch

, et al. (2006) Prior police contact and subsequent victim reporting: Results from the NCVS. Justice Quarterly 23(4): 481–501.

102.

Gel

Ramirez Ramirez

, et al. (2017) Forecasting influenza in Hong Kong with Google search queries and statistical model fusion. PLoS ONE 12(5): e0176690.

103.

Young

Zheng

Chu

, et al. (2018) Internet searches for opioids predict future emergency department heroin admissions. Drug and Alcohol Dependence 190: 166–169.

104.

Zhao

Gibson

Lovrich

, et al. (2002) Participation in community crime prevention: Are volunteers more or less fearful of crime than other citizens? Journal of Crime & Justice 25(1): 41–61.

Supplementary Material

Please find the following supplemental material available below.

For Open Access articles published under a Creative Commons License, all supplemental material carries the same license as the article it is associated with.

For non-Open Access articles published, all supplemental material carries a non-exclusive license, and permission requests for re-use of supplemental material or any part of supplemental material shall be sent directly to the copyright owner as specified in the copyright notice associated with the article.

0.00 MB

0.06 MB

0.04 MB