DETERMINANTS OF EXTORTION COMPLIANCE: EMPIRICAL EVIDENCE FROM A VICTIMIZATION SURVEY

This article focuses on the situational-, victimand area-level determinants of extortion compliance. Extortion, a quintessential organized crime, is one of the most common crimes in Mexico. However, compliance with extortion demands is relatively rare. Previous research suggests that compliance with extortion depends on the perceived risk of punishment for non-compliance. However, most research has been theoretical or experimental. The article offers empirical evidence of patterns of extortion compliance based on data from a large commercial victimization survey conducted in Mexico. Findings suggest that situational factors (extortion type, the presence of weapons and number of offenders) are the main determinants of extortion compliance. Victimand area-level variables have comparatively smaller effects. Implications for research and practice are discussed.

However, despite its high prevalence rate, statistics suggest that compliance with extortion demands is relatively rare. According to Mexico's 2014 commercial victimization survey (the Encuesta Nacional de Victimización de Empresas, INEGI 2014c), victims complied with extortion demands in only about 13 per cent of incidents.
The relatively low compliance rate contrasts with the public perception of extortion in the country as a 'feudal regime' (Perez 2018) with gangs dominating large swathes of territory and extorting all businesses within them. Evidence from Italy (Frazzica et al. 2013;Savona and Sarno 2014) suggests that compliance with extortion demands is common where organized crime groups exert a strong territorial control, which would give grounds to assume that extortion compliance was widespread in Mexico. Similarly, given anecdotal evidence of the dramatic consequences faced by those who refuse to comply with extortion demands, such as the episode described above and the cases described by Guerrero-Gutiérrez (2011) and Hale (2016), one would expect refusals to comply to be the exception, rather than the norm.
Nonetheless, the relative rarity of extortion compliance does not diminish the gravity of the extortion phenomenon; using data from a different survey, Locks (2015) estimated that illicit revenues from extortion in Mexico ranged between US$2.2 and 7.4 billion in 2012. However, it does raise a relevant question of academic and practical importance: Why are most extortion incidents in Mexico not complied with?
The literature on organized crime-particularly on Italian mafias-suggests that, in addition to avoiding fear of reprisals, compliance with extortion can be attributed to social and cultural factors related to the vulnerability of particular regions to mafia control (e.g. La Spina et al. 2014Spina et al. , 2016. Some communities see paying protection money as 'normal' or 'natural' due to long-standing organized crime governance arrangements (La Spina et al. 2016). However, such research is mostly focused on sustained compliance in the context of systematic extortion rackets, 1 and does not explore the situational characteristics that explain why some incidents in the same context lead to compliance while others do not.
In contrast, research on coercion and decision theory (e.g. Nacci and Tedeschi 1973;Luckenbill 1982;Gambetta 1994;Tedeschi and Felson 1994;Smith and Varese 2001) provides a suitable framework to understand the situational determinants of extortion compliance. From this perspective, target compliance is the result of a rational choice: victims choose to comply when the costs of doing so are lower than of not complying. Thus, this literature points towards the situational characteristics that help participants in the extortion interaction weigh the costs and benefits of compliance. However, as most research concerning extortive interactions from this perspective has been theoretical or based on experimental data (e.g. Konrad and Skaperdas 1997;Smith and Varese 2001;Elsenbroich and Badham 2016), there is a need for studies that assess extortion compliance empirically using real-world interactions.
From a practical perspective, identifying the situational determinants of extortion compliance can provide more nuanced characterizations of extortion incidents-a crucial step to design more effective crime prevention interventions (Clarke 2009). Furthermore, being more crime specific not only helps improve the targeting of such interventions, but it can also reveal 'pinch-points' (Read and Tilley 2000;Bullock et al. 2010) in the sequence of events involved in extortions-i.e. the crime script (Cornish 1994)-which can point to the mechanisms that could underpin successful interventions.
Thus, using novel incident-level data from Mexico's 2014 commercial victimization surveyone of the largest victimization surveys of its kind-this study aims to identify the situational 1 Elsenbroich and Badham (2016) define extortion rackets as 'the continuous, regular and systematic extortion of several victims' . Researchers use various terms to refer to similar phenomena: racketeering (McIntosh 1973), extortion racketeering (Savona and Zanella 2010; Savona and Sarno 2014), extortion racket systems (Frazzica et al. 2013;La Spina et al. 2014), private protection (Gambetta 1993;Varese 2001) and violent entrepreneurship (Volkov 2002), among others.

DETERMINANTS OF EXTORTION COMPLIANCE • 3
determinants of victim compliance in extortion incidents. The article proceeds as follows: In the next section, I review the literature to inform the hypotheses to be tested in the study. Then I describe the data and analytical approach used. This is followed by the results and discussion.

FACTOR S AFFECTING E XTORTION COMPLIANCE
As noted above, it is generally assumed that victims choose to comply with an extortion demand when doing so is less costly than not complying. However, as the true costs of non-compliance are uncertain-threats may not materialize-game theoretical models of extortion note that the main determinant of compliance is the victim's estimation of the likelihood of punishment for non-compliance (e.g. Gambetta 1994;Skaperdas 1997, 1998;Smith and Varese 2001). Given that this likelihood is unknown, Konrad and Skaperdas (1997) argue that victims consider threat credibility (the rate at which the extortionists punished noncompliant victims in the past) (see also Konrad and Skaperdas 1998). On the other hand, Gambetta (1994) and Smith and Varese (2001) broaden this to include more subjective perceptions, and consider that it is the reputation groups have for their willingness to use violence, rather than actual retaliation for non-compliance, which matters most in influencing the likelihood of compliance. However, this allows 'pirates' (Gambetta 1994) and 'fakers' (Smith and Varese 2001) to exploit someone else's reputation spuriously, e.g. by pretending to be a member of an organized crime group (an example of Felson's 'mimicry' principle, 2006).
One of the central issues determining the victim's perception of the likelihood of punishment for non-compliance-and hence of their decision to comply-is the offender's ability to convince the victim of the authenticity of the threat. Gambetta (1994) argues that extortionists establish their 'authenticity' using symbols and signals that communicate their belonging to a particular organized crime group. However, as actors in an extortive strategic interaction (Goffman 1970;Best 1982) have implicit incentives to deceive their opponents, explicit signals and symbols can still be mimicked. Therefore, victims may be forced to rely on additional cues gleaned from the interaction to determine whether the threats should be believed (Luckenbill 1982).
In a communicative interaction, the medium used is itself a source of information that can deeply influence how the message being exchanged is interpreted (McLuhan 1964). Thus, in the context of extortion, the communication medium or channel used by the threat's sender (the extorter) to convey the message (the actual avowed threat) to the receiver (the extorted) can have a strong bearing in believability. As O'Hair et al. (2011) note, 'those who threaten others have a number of communication channels available to them… Channel selection is sometimes a spontaneous and convenient choice, whereas in cases of predation the choice of channel can be quite strategic' (57).
According to media richness theory (Daft and Lengel 1986;Lengel and Daft 1989), communication channels can be classified based on the amount of information (verbal, non-verbal, visual, etc.) they can convey. Lengel and Daft (1989) classify face-to-face interactions as the richest form of media, while other interactive media, such as telephone and other technologymediated channels, are considered relatively leaner, as they 'lack the element of "being there"' (226). Senders strategically select rich media when they aim to reduce uncertainty and equivocality (the possibility of deriving several meanings) (Daft and Lengel 1986: 555).

Types of extortion and communication channels
According to the media richness of the channels used to convey threats, extortion incidents in Mexico can be classified into 'remote' (lean media) and 'in-person' (rich media) extortion. Remote extortion relies on the use of technology-mediated channels to convey the threat. In the most common type of remote extortion, threats are communicated over the telephone.
According to ONC (2014), there are several variations of how telephone extortion is carried out. Incidents generally begin by offenders cold-calling victims and attempting to convince them to pay an amount into a bank account or mobile phone number. To achieve this, offenders use advanced-fee scams, 2 'virtual kidnappings', 3 or claim to be a member of an organized crime group 4 and threaten to carry out severe punishments if victims do not comply with the demands (ONC 2014: 30). Particularly for the last two types, offenders use personal details obtained on social media, through data breaches or in previous calls, to convince victims of the authenticity of the threats (ONC 2014).
The internet is another common channel used in remote extortion. Internet extortion incidents rely on the same tactics as telephone extortion, the difference being that offenders contact victims via email, social media or electronic means other than a telephone (ONC 2014: 32). An exception is 'ransomware' extortion, which relies on malware-a computer virus-that encrypts the victim's computer or infrastructure until a ransom is paid, usually using a cryptocurrency such as bitcoin (Darrel 2013, 'ransomware'). Whereas in ransomware incidents the threats are levelled against digital assets (the data or applications under ransom), the threats in internet extortion incidents are usually aimed at the victims' personal safety.
On the other hand, in-person extortion incidents rely on face-to-face communication to convey threats. In-person incidents are also known as cobro de piso, 5 and are thought to be carried out by 'authentic' members of an organized crime group. In these incidents, offenders threaten victims with damage, assault, death or other harms if they refuse to pay a fee (or provide some requested service) (Mugellini 2013b;ONC 2014). Mugellini (2013b) notes that offenders can also offer 'protection' from other criminal groups in these types of incident (34). Furthermore, ONC (2014) considers that cobro de piso extortions involve periodic payments at a set frequency-e.g. monthly, weekly. However, in-person extortion incidents can also be committed by non-organized criminals who demand one-off payments.
Thus, given that use of leaner media has been associated with a higher likelihood of engaging in deceptive behaviour, and that receivers are less likely to trust messages sent using leaner media (Rockmann and Northcraft 2008), it is reasonable to expect that victims would be more likely to believe in-person extortion threats are authentic, when compared to remote extortion threats, and would therefore be more likely to comply with the former than the latter. The first hypothesis in this study is: H1: The likelihood of compliance with an extortion threat is higher in cases of in-person extortion incidents, when compared to remote extortion incidents.

Other factors affecting extortion compliance
In addition to threat believability, Luckenbill (1982) suggests that threat compliance is also affected by the severity of the potential punishment, the offender's capacity to inflict such punishment, and the victim's capacity to oppose or resist the threat (811-2).
2 An advanced-fee scam is 'a form of fraud … in which the victim is invited to pay financial fees in the hope of sharing in a much greater reward' (Daintith and Wright 2008). For example, the extortionist claims the victim has won a prize from a contest or raffle, but requires the victim to pay a sum before receiving the reward. Sometimes, the scams are used to obtain personal details that will be used in subsequent calls for virtual kidnappings or threatening calls (ONC 2014: 30).

DETERMINANTS OF EXTORTION COMPLIANCE • 5
Anecdotal accounts of punishments inflicted on noncompliant victims-which include homicide, assault, arson and other extensive criminal damage (e.g. Guerrero-Gutiérrez 2011;Wilkinson 2011;Hale 2016)-suggest that the punishments promised in an extortion interaction are probably quite severe. However, as the incident-level dataset used in this study does not contain precise information on the severity of punishment for non-compliance, it is not possible to ascertain its effect on compliance patterns.
The effect of the offender's capacity to inflict punishment on the likelihood of compliance cannot be understood in isolation, but must also be considered with respect to the victim's capacity to resist such punishment. As compliance is assumed to be the result of a rational calculus, victims are more likely to comply if they perceive that the offender's capacity to punish is greater than their own capacity to resist, i.e. when there is a perceived power asymmetry in favour of the offender (Michener et al. 1973;Bacharach and Lawler 1976;Luckenbill 1982).
However, as Bacharach and Lawler (1976) note, 'power capabilities are typically ambiguous; hence conflicting parties must use situational cues to form subjective power estimates' (3). Common situational factors that clearly signal power asymmetry in favour of the offender are the presence of lethal resources (i.e. weapons, Luckenbill 1982: 814) or of multiple offenders. Thus, the second set of hypotheses is: H2a: The likelihood of compliance with an extortion demand is higher when offenders use weapons. H2b: The likelihood of compliance with an extortion demand is higher when there is more than one offender involved.
Furthermore, research on organized crime suggests that contextual factors can also have an effect in determining the likelihood of extortion compliance (Gambetta 1994;Smith and Varese 2001;La Spina et al. 2014. Such contextual factors are not unique to each incident and instead represent area-level characteristics related to the perceived costs of using violence and the reputation of organized crime groups in a victim's area. The perceived costs of violence can be captured using a general measure, such as the strength of the rule of law. On the other hand, the reputation of organized crime groups can be captured by their readiness to use violence (e.g. the amount of crimes involving weapons), and by the type of illicit markets they are involved in (e.g. groups involved in drug-trafficking are usually less likely to be involved in extortion) (Estévez-Soto et al. 2020). Thus, the third set of hypotheses is: H3a: The likelihood of compliance with an extortion demand is higher in areas where the rule of law is weaker. H3b: The likelihood of compliance with an extortion demand is higher in areas with more weapon-related crimes. H3c: The likelihood of compliance with an extortion demand is higher in areas with fewer drug crimes.
Victim vulnerability can similarly be classified into situational and contextual measures. At the situational level, victim characteristics may have a part to play. For example, research suggests that some business types are inherently more susceptible to intimidation (e.g. restaurants, Schelling 1971: 646), and empirical studies confirm that some business types are more likely to comply with extortion demands (Chin et al. 1992: 641;Estévez-Soto et al. 2020). Business size could also be indicative, as smaller businesses are inherently more vulnerable than larger businesses. Lastly, the number of years that a business has been in operation could be negatively associated with compliance, as older businesses can be expected to have more social capital-a source of power to resist extortion demands (Anzola 2016).
H4a: The likelihood of compliance with an extortion demand is associated with business type. H4b: Small businesses are more likely to comply with extortion demand, when compared to larger businesses. H4c: Newer businesses are more likely to comply with an extortion demand, when compared to older businesses.
The literature on repeat victimization suggests that, in some crimes, the probability of suffering a repeat is associated with how the victim responds to a previous offence (Farrell et al. 1995: 396). In particular, a study on repeat extortion victimization found that the number of repeated extortions suffered is not likely to be explained by victim or area characteristics, suggesting that event dependence may play an important role in determining future risk (Estévez-Soto et al. 2020), meaning that an initial event could entice further attempts, as the victim is known to be acquiescent. Thus, in the case of extortion, it is reasonable to expect an association between the likelihood of compliance and the amount of extortion incidents suffered by a business. Furthermore, Estévez-Soto et al. (2020) also found strong associations between corruption victimization and extortion. While it is not yet clear why this association exists, it is possible that businesses that suffer more corruption victimization are inherently more vulnerable to extortion. Thus, it is reasonable to expect an association between business-level experiences of corruption and the likelihood of compliance with extortion.
H5a: The likelihood of compliance with an extortion demand is positively associated with the amount of extortion demands a victim receives. H5b: The likelihood of compliance with an extortion demand is positively associated with the amount of bribes victims are asked to pay.

DATA AND ME A SURES
The study uses the 2014 sweep of Mexico's nationally representative commercial victimization survey, ENVE. The survey is conducted biennially, sampling all business sectors-except those in agriculture and the public sector. As is common in other victimization surveys (e.g. UNODC/UNECE 2010), the instrument is divided into two parts. First, a screening questionnaire records prevalence (whether a respondent was victimized) and incidence (how many crimes victims experienced) measures for crimes that took place during the previous calendar year (in this case 2013), as well as gathering business characteristics. The second section-the victim form-is used for victimized businesses only, capturing details on each crime incident reported in the screening questionnaire-however, there is a cap of 7 incidents per crime type per business (INEGI 2014c). As compliance with extortion demands is captured at the incident level, the study uses information primarily from the victim forms, with business-level data coming from the screening questionnaire (for a detailed review of the ENVE, see Jaimes Bello and Vielma Orozco 2013), and area-level data from other sources (detailed in the following sections). The survey has nationwide coverage and is representative at the national and subnational scale (state level). In 2014, a stratified sample of 33,479 premises 6 was drawn from a sampling frame comprising 3.8 million units (INEGI 2014a(INEGI , 2014b. Interviews were conducted through face-to-face interviews, with computer-assisted telephone interviews to follow-up ( Jaimes Bello and Vielma Orozco 2013). The response rate was around 85 per cent (INEGI 2014b).

DETERMINANTS OF EXTORTION COMPLIANCE • 7
To protect anonymity, access to the disaggregated incident-level responses is restricted by the data provider. Thus, analyses were carried out remotely, using custom-written R scripts 7 (R Core Development Team 2015) processed by INEGI staff in Mexico City.

Dependent variable
The dependent variable, compliance with extortion, is captured in the victim forms after businesses have indicated that they suffered at least one extortion incident in 2013. 8 For each incident, compliance was coded as '1' when respondents responded 'yes' to the question 'did you comply with the extortionist's demands?' ('¿Entregó lo que le exigió el extortsionador? ' INEGI 2014c), and '0' if otherwise. The survey captured 3,369 extortion incidents (among 2,259 victimized businesses). Compliance was observed in only 425 incidents (12.6 per cent), whereas compliance was not observed in the remaining 2,944 incidents (87.4 per cent).

Independent variables
This section describes the independent variables selected to test hypotheses. Incident-level variables are presented first, followed by victim-and area-level measures respectively.
Categories with very small number of observations were recategorized to avoid complete and quasi complete separation, which occur when a categorical variable perfectly (or almost perfectly) predicts the value of the dependent variable (i.e. when all or nearly all observations of a particular category have the same value in the dependent variable). The presence of complete and quasi complete separation means that estimations using maximum-likelihood estimation will be unreliable (see Zeng and Zeng 2019).
Extortion type (H1) was recorded as 'telephone extortion' , 'by internet/email' , 'on the street' , 'on the premises', 'cobro de piso' and 'other' . Incidents categorized as 'telephone' and 'internet' extortion were recategorized as 'remote' extortion, while incidents classified as 'other' were dropped from the analysis. 9 According to INEGI (2014a), 'on the street', 'on the premises' and 'cobro de piso' incidents are considered to be 'in-person' extortion incidents, though there is no precise distinction provided for cobro de piso and other in-person extortions. Nonetheless, the distinct categories were retained to explore if they are associated with different patterns of compliance.
Moving on to business-level variables, business type (H4a) was captured by the survey according to the North American Industrial Classification System (SCIAN, INEGI 2007). However, using this classification system, there were some categories with few or no observations. Thus, only the following 11 categories were kept in a compromise between avoiding separation and maintaining theoretical relevance: 'Retail', 'Wholesale' , 'Hotels, restaurants and bars' , 'Transport', 'Other services' and 'Industry' . Business size (H4b) categories were defined by the 7 Available upon request. 8 The specific question in the screening questionnaire is: Did the business suffer in 2013 'any kind of threat or coercion committed against the local unit's owner or staff for the purpose of obtaining money, goods or forcing them to do or stop doing something?' ( Jaimes Bello and Vielma Orozco 2013: 172).
9 There were only 8 (0.2 per cent) incidents of internet extortion and 9 (0.3 per cent) incidents categorized as 'other' . 10 Unless otherwise noted, missing values for independent variables were classified as 'dk/da' . 11 Categories with few observations were aggregated into the higher-order classification offered by the SCIAN. The British Journal of Criminology, 2021, Vol. XX, No. XX survey according to the number of employees. 12 The business age (H4c) category was calculated by subtracting the year respondents reported that their business started operations from the survey reference year (2013). Then, businesses were grouped in quintiles from the 20 per cent youngest to the 20 per cent oldest.

•
The number of extortion incidents (H5a) suffered by businesses-henceforth extortion concentration-was taken from the uncapped extortion victimization experiences reported in the screening questionnaire. Similarly, the amount of bribes (H5b) demanded from businesses-henceforth corruption incidence-was taken from the uncapped figure captured in the screening questionnaire in response to the question 13 : 'In total, how many separate acts of corruption did you suffer during 2013?' (INEGI 2014d). As the estimates for these variables were overdispersed, a log transformation was used. 14 State-level variables measure variation at the state level. 15 The strength of the rule-of-law (H3a) was measured using a revised index calculated by IMCO (2016); a composite 100 point score composed of kidnapping incidence, vehicle theft, costs of crime, total personal and household crime incidence, the crime underreporting rate, fear of crime, availability of notaries, and contract enforcement (higher scores represent a stronger rule of law). For this study, homicide rates were excluded from the index, as these were collinear with other crime covariates used. Measures for weapon-related crimes and drug-related crimes (H3b and H3c) were taken from the Executive Secretariat of the National System for Public Security (Secretariado Ejecutivo del Sistema Nacional de Seguridad Pública, SESNSP 2015) as reported in 2013 by the Attorney General's Office (Procuraduría General de la República, PGR). Lastly, the area-level corruption prevalence, an economic competitiveness index, 16 the population and the number of businesses surveyed in each state were used as controls.
The state-level variables weapon crimes, drug crimes, corruption prevalence, population and the number of surveyed businesses were log-transformed to reduce overdispersion. All statelevel variables were centred around the national mean 17 to facilitate interpretation.
Descriptive statistics for the data used are presented in the Appendix.

ANALY TICAL METHOD
In order to mitigate confounding variables and to estimate the partial effect of each variable, the relationship between compliance and the selected independent variables must be evaluated using a multiple regression method. As the study is concerned with testing the effects of several independent variables on the likelihood of compliance-a dichotomous dependent variable with responses taking either 0 or 1 values-a multiple logistic regression was used. However, while this model controls for different extortion types, it ignores the fact that other predictors may operate differently in remote versus in-person extortions.
As the cross-tabulations in Table 1 indicate, compliance varies dramatically according to extortion type. While 12.6 per cent of all incidents led to compliance, only 5.4 per cent of remote 12 There are four categories: Micro businesses have 10 employees or fewer; small businesses have between 11 and 50 employees (11-30 in the commerce sector); medium businesses in industry employ between 51 and 250 people, 31 and 100 in commerce, and 51 and 100 in services; large businesses are those with 101 or more employees (251 or more in industry). 13 An act of corruption refers to a situation where a public servant-or a third party acting on their behalf-directly asked for, suggested, or set the conditions for the payment of a bribe by the business ( Jaimes Bello and Vielma Orozco 2013; INEGI 2014d). 14 As corruption incidence includes 0, the function log(x + 1) was used for this variable. 15 Mexico is divided into 32 autonomous states. 16 The index used was a slight revision of IMCO's competitiveness index (2016), based on 9 subindices measuring sustainable development, social development and health, political stability, government effectiveness, labour productivity, economic stability, infrastructure and international connections. 17 Log-transformed variables were centred around the log of the national mean (log(x) − log(x)).

DETERMINANTS OF EXTORTION COMPLIANCE • 9
extortion incidents led to compliance. For in-person extortion incidents, compliance was observed in between 49.6 per cent and 66.7 per cent of events. Moreover, a Pearson's χ 2 test of independence indicated that the differences in compliance rates according to the type of incident were statistically significant ( p < 0.001). Furthermore, considering the differences in modus operandi between remote and in-person extortion, it is reasonable to expect that some predictors may play a bigger role in one type of extortion when compared with the other. Thus, two additional models were estimated, one restricting incidents to remote extortion, while the other used in-person extortions only.
However, an additional complication is that the data have a hierarchical structure, as some businesses suffered more than one incident and businesses are grouped within states (see Table 2). This is a violation of the assumption of independence for logistic regressions. To mitigate this violation, clustered standard errors (Zeileis 2006;Berger et al. 2017) with victim-and state-level clusters were used. Table 3. The ' All incidents' model estimates the conditional odds of complying with an extortion demand for all the incidents in the data, whereas the 'Remote' and 'In person' models estimate the conditional odds for subsets of incidents where the extortion attempts took place remotely or in person, respectively. Wald χ 2 goodness-of-fit statistics suggest that the three models are significantly different from a null specification. Generalized variance-inflation factors (Fox and Monette 1992) indicated that multicollinearity was not present.

Results of the models estimated can be found in
As coefficient estimates are in the log-odds scale, interpretation of the exponentiated coefficients (e B ), also known as odds ratios, is more straightforward (see OR columns). The odds ratio is interpreted as the multiplicative effect on the odds of observing 1 in the dependent variable, for a one-unit increase in the independent variable. For categorical independent variables, the odds ratio is the multiplicative change in the odds in reference to a base category.    log(x + 1) was used.
Odds ratios (OR) were calculated by exponentiating the coefficient estimates (e B ). However, odds ratios for log-transformed variables were calculated for a 10% change in the predictor (1.10 B ). Standard errors were calculated using a robust variance-covariance matrix with business and state clusters.

Continued
In what follows, I describe the partial effect of each variable, thus the effect sizes refer to the expected change in the dependent variable after controlling for all other variables. In the ' All incidents' model, the odds ratios for extortion type categories were significant at the 99.9 per cent confidence level and greater than 1, suggesting that in-person extortion incidents are more likely to involve compliance than remote extortion (the reference category). In this model, street and in-premises extortion incidents were 7.67 and 8.33 times more likely to involve compliance than remote extortion incidents. Similarly, cobro de piso incidents were 16 times more likely to lead to compliance than remote extortion incidents.
The estimates from the 'In person' model further characterize the relationship between compliance and extortion type. According to this model, the likelihood of compliance with an extortion incident in a business's premises is not significantly different from the likelihood of compliance with an extortion incident that takes place on the street (the reference category for this model). In contrast, the odds ratio for cobro de piso incidents was greater than 1 and significant at the 95 per cent confidence level, meaning that these type of incidents were associated with greater rates of compliance than street extortion incidents. Specifically, cobro de piso incidents were 2.48 times more likely to lead to compliance than street extortion, when considering inperson incidents only.
The effect of most other independent variables appears to be more muted; however, the models fitted to different subsets of extortion incidents suggest that the partial effects on compliance of these independent variables are different for remote and in-person extortion.
The number of offenders involved in an extortion incident appear to be significant and positive in the ' All incidents' model, however, the estimates from the 'Remote' and 'In person' models suggest that the relationship is only significant for remote extortion incidents, as the coefficients for the number of offenders are not significant in the 'In person' model. For remote extortion, incidents with 2, 3 and 4 or more offenders are 2.95 ( p > 0.05), 4.51 ( p > 0.05) and 4.54 ( p < 0.01) times more likely to lead to compliance than incidents with only one offender, respectively.
In contrast, weapon use was significant ( p < 0.001) and positive in all three models, which suggests that incidents in which a weapon was used are more likely to lead to compliance for all extortion types. However, the magnitude of the coefficient was different for remote and inperson extortion incidents. The use of a weapon in a remote extortion incident was associated with a 12.7 times greater odds of compliance, whereas for in-person incidents weapon use was associated with 2.96 times greater odds of compliance.
Regarding business-level variables, only extortion concentration, corruption incidence and business type had a significant effect on compliance, though again the effects varied by extortion type.
Extortion concentration was negative and significant ( p < 0.001) for all incidents, however, the 'Remote' and 'In person' models suggest that the relationship was only significant ( p < 0.001) for in-person incidents. A 10 per cent increase 18 in the number of extortion incidents suffered by a business was associated with an 8 per cent decrease in the odds of compliance with an in-person extortion incident. 19 Similarly, corruption incidence was significant ( p < 0.001) and positive for all incidents, but the secondary models suggested that the relationship was only significant ( p < 0.01) for in-person incidents. According to the 'In person' estimates, a 10 per cent increase in the number DETERMINANTS OF EXTORTION COMPLIANCE • 13 of bribes experienced by a business was associated with a 10 per cent increase in the odds of complying with an in-person extortion incident.
Most business types coefficients were not significantly associated with higher odds of compliance. The only exceptions are the 'Other services' and 'Industry' categories, which were significant at the 95 per cent confidence level in the ' All incidents' model. However, the estimates from the 'Remote' and 'In person' models suggest that only the coefficient for 'Industry' is robust, and only in the case of remote extortions. The estimates suggest that businesses in the industrial sector are 77 per cent less likely to comply with a remote extortion incident, when compared with retailers (the reference category). In contrast, the estimates for the 'In person' model suggest that all business types are as likely to comply with in-person extortion demands.
Business size and age were insignificant in all models, meaning that the likelihood of compliance was not affected by these variables in either remote or in-person extortion.
Area-level variables were mostly insignificant. Only the amount of weapon-related crimes showed a significant and positive association with extortion compliance ( p < 0.05) in the ' All incidents' model; however, the coefficients for weapon-related crimes in the 'Remote' and 'In person' secondary models were not statistically significant, albeit both were positive.

DISCUSSION AND CONCLUSIONS
This study sought to answer why-despite a very high prevalence rate, and associations with violent punishments-extortion compliance is relatively rare in Mexico. Using incident-level data from Mexico's commercial victimization survey-one of the largest exercises of its kindthe study tested whether situational-, victim-and area-level factors influenced victims' decision to comply, using multiple logistic regression.
The first hypothesis tested was that the likelihood of compliance with extortion demands would be higher in cases of in-person extortion, when compared to remote extortion incidents, as it was assumed that threats conveyed by richer media channels (in-person extortion) would be more believable than those conveyed via leaner channels (remote extortion). The findings strongly support this hypothesis, as all in-person extortion categories (street, in-premises and cobro de piso) were associated with substantially higher likelihoods of compliance when compared with cases of remote extortion. It is unclear what specific characteristics distinguish cobro de piso incidents from other in-person incidents, as the survey does not provide a precise definition. However, the fact that the odds of compliance in cobro de piso incidents were significantly higher than the odds of compliance in street extortion incidents (in the 'In person' model), suggests that the distinction is relevant and should be considered further.
Hypotheses 2a and 2b tested whether power asymmetry in favour of the offender-operationalized as the presence of weapons and multiple offenders-increased the likelihood of observing compliance. The findings strongly supported Hypothesis 2a, as the presence of a weapon was significantly associated with higher odds of compliance in remote and in-person incidents. The marginal effect was much larger in the case of remote extortion, though this disparity can be explained by the much smaller baseline odds for this extortion type. In practice, the presence of a weapon increases the predicted probabilities for remote and in-person incidents to a similar level (57 per cent for remote extortion and 52 per cent for in-person extortion). In contrast, the findings supported Hypothesis 2b only in remote extortion, as incidents with more than one offender had consistently higher odds of compliance. This was not the case for in-person incidents, where the number of offenders had no effect on compliance. The findings suggest that after controlling for threat believability as captured by the in-person/remote distinction, additional markers of power asymmetry can have a substantive effect on a victim's decision to comply with an extortion demand, especially for remote extortion. However, further research-particularly of a qualitative nature-is needed to better understand how victims infer the presence of weapons or the number of offenders involved in remote extortion incidents.
Hypotheses 3a, 3b and 3c tested contextual factors that speak to the perceived costs of violence in the area where extortion incidents took place-the assumption being that compliance would be more likely in areas where the costs of violence are lower. The findings did not support Hypotheses 3a and 3c: I failed to find any relationship between extortion compliance and the strength of the rule of law (H3a), or the amount of drug crimes in the state where businesses operate (H3c). In contrast, the findings partly supported Hypothesis 3b: incidents in areas with more weapon-related crimes, and hence organized crime groups with more demonstrated readiness to use violence, were more likely to lead to compliance. However, this relationship was not significant in the secondary models, which weakens the evidence supporting Hypothesis 3b. It is unclear why the relationship between weapon crimes and compliance was not significant in the secondary models, however, a potential explanation may be the reduced statistical power of the 'Remote' and 'In person' models, as they use smaller samples than the ' All incidents' model.
Hypotheses 4a, 4b and 4c related to whether business characteristics were associated with extortion compliance, under the assumption that some businesses are more inherently vulnerable to intimidation. The findings suggested that most business types (H4a) have the same likelihood of complying with extortion demands for all extortion types. The main exception was businesses in the industrial sector, which were less likely to comply with remote extortion incidents. Furthermore, there was no evidence of a relationship between extortion compliance and business size (H4b) or business age (H4c).
On the other hand, Hypotheses 5a and 5b related to whether dynamic characteristics that speak to business vulnerability-extortion concentration and corruption incidence-had an effect on extortion compliance. In contrast to what was predicted, the more extortion incidents a victim experienced, the less likely they were to comply, though this was only significant for in-person extortion. Given the cross-sectional nature of the data, it is not possible to establish the direction of the causal effect; it may be that suffering more extortion incidents helps victims properly assess risks and avoid complying, or it could reflect repeated attempts by offenders to harass victims into compliance after being refused. However, establishing the direction of the effect would require longitudinal data that are not available. On the other hand, the relationship between corruption incidence and compliance was consistent with what was expected: the amount of bribes that victims were asked to pay was positively associated with the likelihood of compliance, though only for in-person incidents.
There are limitations to the findings reported here. Extortion against businesses is notoriously difficult to measure: on the one hand, statistics based on crimes reported to the police rarely disaggregate crimes by victim type; on the other, extortion incidents are usually underreported, as victims fear reprisals. While, commercial victimization surveys can overcome such limitations to an extent (for a review, see Mugellini 2013c), the estimates and patterns captured by surveys suffer from well-known limitations involving memory decay, telescoping effects and victims' reticence to report certain experiences (Skogan 1986;UNODC/UNECE 2010;Mugellini 2013a). Due to these limitations, Mugellini (2013a) notes that victimization estimates tend to underestimate the 'true' prevalence and incidence of crimes, though they do represent an improvement over other crime statistics. However, such underestimates notwithstanding, the large sample size and high response rate help assuage fears of any systematic biases affecting the reliability of the patterns observed.
The study has important academic implications. It contributes to the literature on organized crime by highlighting the role of situational characteristics of extortion incidents in determining compliance and suggesting that contextual factors play a less relevant role, in contrast with existing research (see La Spina et al. 2014Spina et al. , 2016.

DETERMINANTS OF EXTORTION COMPLIANCE • 15
The research also contributes to the literature on decision theory (e.g. Gambetta 1994;Skaperdas 1997, 1998;Luckenbill 1982;Smith and Varese 2001) by empirically testing theoretical predictions and experimental findings regarding the role of threat believability. Furthermore, the study introduces the role of the medium through which extortion threats are conveyed as an important factor affecting threat believability, using the theoretical framework of media richness theory (Daft and Lengel 1986;Lengel and Daft 1989).
Lastly, the research contributes to the literature on victimization, specifically the research on repeat extortion victimization (see Estévez-Soto et al. 2020). The different compliance patterns observed for remote and in-person incidents suggest that these are quite distinct types of offences. Thus, future studies on extortion victimization should analyse the patterns of concentration by extortion type, as they may be associated with different opportunity structures. To facilitate this, crime surveys should measure remote and in-person extortion as different crimes, rather than as categories of the same crime type. Doing so would allow capturing uncapped measures of extortion victimization per type, and would allow capturing more meaningful follow-up questions in the victim forms that take into account the distinct modus operandi associated with each extortion type. Such data could then be used to deepen our understanding of the risk factors associated with extortion victimization and compliance, and inform more effective interventions to control these crime types.