Abstract
From 1882 to 1926, lynch mobs in Georgia killed 514 victims in 410 separate events. Based on a new comprehensive dataset, this study examined characteristics of the mobs related to the level of the atrocity of the violence in the lynchings. Consistent with prior research, the size of the crowd was a stable predictor of level of atrocity. However, in contrast to two prior studies, results did not uniformly support the self-attention perspective. Instead, the findings were more consistent with the concept that situational norms were the important mechanism behind collective violence, particularly because the violence used in a lynching event reflected the violence used in nearby lynchings. These results suggest that, rather than losing all norms of behavior through a process of deindividuation, lynchers adopted norms they viewed as appropriate for the situation.
A 1934 New Yorker cartoon by Reginald Marsh was captioned “This is her first lynching.” 1 In this night scene, White men in overalls and wide-brimmed hats and White women, some wearing bonnets, are in front of a farmhouse looking to the left at a fire-illuminated unseen murdered Black man. In the right front of the drawing, a pensive and perhaps confused little girl is held aloft by an older woman, probably a relative, who shouts to the woman behind her that this event is the child’s first lynching. The clear import of Marsh’s focus on the White audience is that lynching was “a communal entertainment” (Apel, 2004, p. 90).
The portrayed mob, with blurred faces and bodies melding into a single mass, depicts deindividuation, which is the most common explanation for the atrocity of lynchings (Leader, Mullen, & Abrams, 2007; Mullen, 1986). According to deindividuation, individuals are more likely to break norms and commit atrocious acts when they are anonymous and not self-attentive (Diener, 1980; Festinger, Pepitone, & Newcomb, 1952). Rather than breaking norms, though, the anonymity provided by crowds has also been suggested to increase the salience of group norms (Reicher, 1984; Reicher, Spears, & Postmes, 1995). In this study, we explored whether the atrocities that occurred during lynchings exemplify breaking or following norms. We suggest that there are rules for violent events and, like the little girl in Marsh’s drawing, these rules are learned by observing others (Surette, 2013).
Lynchings are extra-judicial killings of individuals by three or more people, generally consisting of White mobs and African American victims. Not all lynching events were the same (Smångs, 2016). Mobs ranged in size (Brundage, 1993), level of crowd excitement (Miller & Dollard, 1941), and the degree of social distance from the victim (Senechal de la Roche, 1996). In this study, our focus was on the level of atrocity in the lynchings.
Atrocity
Lynching atrocity is the severity of violence surrounding the death of the victim. A lynching event in which the mob hangs, shoots, and burns the victim seems more atrocious than a lynching event in which the mob only shoots the victim (Mullen, 1986). More than just the act, severe violence often connotes a moral wrongfulness (Gromet & Darley, 2009) that is often considered in contemporary law, as in the upward departure provisions of the U.S. Sentencing Commission (2014, p. 467) guidelines for “torture of a victim, gratuitous infliction of injury, or prolonging of pain or humiliation.” In prior research, severity of lethal violence has been conceptualized as related to both the number of victims and the amount of violence. For instance, the upper-end of Cornell’s (1996) Severity of Violence scale distinguishes homicides from extreme homicides involving multiple killings or mutilation. Considering multiple killings as more severe is also in line with Costalli and Moro’s (2012) use of number of deaths as a proxy for severity of violence in their examination of violence during the Bosnian civil war. Rather than just the number of victims, the excessive aspects of killings, like mutilating a lynching victim, can distinguish “extra-lethal” violence from other types of lethal violence (Fujii, 2013).
Anonymity and Deindividuation
A key characteristic used to explain varying levels of violence is anonymity, which is the inability of others to identify an individual (Marx, 1999). Anonymity can be influenced by factors such as group size, disguises, and darkness, and under such conditions individuals can get “lost” in a crowd. Consistent with this notion of anonymity, larger groups have been found to be more violent than smaller groups (McGloin & Piquero, 2009). And, related to group size, Mullen (1986) and Leader et al. (2007) examined the atrocity of lynching events through group composition in the form of the Other-Total Ratio, which is calculated as the number of victims in an event divided by the total number of victims and perpetrators at the event. As lynch mobs grew in size, relative to the victims present, mob members committed more atrocious lynchings. Both Mullen (1986) and Leader et al. (2007) found that lynch mobs containing more people in relation to the number of victims committed higher atrocity lynchings.
In addition to group size, disguises increase anonymity. Anonymity provided by disguises leads to increased levels of violence among warriors (Watson, 1973), among terrorists (Silke, 2003), and in experimental laboratory conditions (Johnson & Downing, 1979; Zimbardo, 1969). Darkness can also provide anonymity, encouraging individuals to act deviantly (Gergen, Gergen, & Barton, 1973; Zhong, Bohns, & Gino, 2010). In sum, as anonymity contributes to varying levels of violence, the variation of group size, disguises, and darkness should relate to varying levels of violence committed by a crowd: groups that are large, disguised, and organized at night should be more likely to act violently.
It is clear that increasing anonymity corresponds with increasing violence, but the mechanism connecting the two is under some debate. The key question seems to be whether anonymity leads individuals to drop general social norms and act violently, or whether anonymity leads individuals to adopt different, contextual norms and act violently. While data limitations prevented the current study from arbitrating between these two sides of the debate, a review of the potential mechanisms is appropriate, especially since the two quantitative studies that examined the level of violence in lynchings accounted for only one side of the debate.
Mullen (1986) and Leader et al. (2007) based their lynching analyses on the concept of deindividuation, the psychological process in which individuals lose their sense of self. Deindividuation theory suggests that groups affect the psychological state of their members, reducing individual agency and increasing behaviors that break social norms, including violence (Diener, 1980). According to the theory, rising deindividuation decreases adherence to social norms. As individuals submerge themselves into a crowd identity, they lose both their individual identities and their self-control (Festinger et al., 1952; Le Bon, 1895). With individuals having lost norms that emphasize order, the group becomes increasingly violent. The more deindividuated the members of a lynch mob are, the greater the atrocities of the lynch mob.
This deindividuated state is achieved through mob characteristics affecting the anonymity of its members. In the work of Mullen (1986) and Leader et al. (2007), a small Other-Total Ratio represented a less self-attentive mob, and higher atrocity lynchings among these mobs was interpreted as being consistent with the process of deindividuation and reduced restraints against breaking norms (Mullen, 1983).
Norms
While past work on lynching atrocity advocated the deindividuation perspective, Leader et al. (2007) mentioned that their results were also consistent with individual mob members adopting violent group norms—rather than the dropping of norms that occurs with deindividuation and reduced self-attention. In fact, elements of deindividuation—group size, anonymity, and self-awareness—also seem to help establish local group norms (Postmes & Spears, 1998). Postmes and Spears (1998) found little evidence for either a deindividuated state or deindividuated behaviors in their meta-analysis of the deindividuation literature.
Instead of making individuals lose self-awareness, the anonymity of groups can encourage individuals to consciously break broader social norms due to reduced accountability (Prentice-Dunn & Rogers, 1989). Further, as individuals partly define themselves by which groups they belong to, being in a group can strengthen that group identity and encourage adherence to group norms (Reicher et al., 1995). This social identity model of deindividuation differs from classic deindividuation theory by emphasizing that crowds lead individuals to conform to (sometimes violent) local group norms rather than cause individuals to lose self-awareness and break broader norms (Spears, 2014).
Lynch mobs may have had norms of violent behavior just as other groups have norms of violent behavior. For example, among larger groups, attacks on property characterize disorderly campus gatherings, whereas attacks on authorities and civilians better characterize urban riots—differences likely due to the different expectations individuals have when entering the situation (Martin, McPhail, & McCarthy, 2009). People model their norms and behaviors on the norms and behaviors of others, as is the case with copycat crimes (Surette, 2013). A lynch mob, then, should have based its norms and behaviors on the norms and behaviors of other lynch mobs. Norms about lynching events could be learned by observing lynchings directly, reading about lynchings in a newspaper, or hearing about lynchings via word of mouth, thus encouraging individuals in a lynch mob to imitate the actions of earlier lynch mobs in copycat lynchings.
In addition to these situational, group norms, other broader social norms could be relevant during lynching events. For example, gender norms could have influenced the atrocity of a lynching. Female perpetrators are generally viewed as less to blame than male perpetrators for deviant behavior (Rye, Greatrix, & Enright, 2006). Due to gender stereotypes, judges sentence female defendants more leniently than male defendants, seeing females as less blameworthy for their actions (Doerner & Demuth, 2010). Female lynching victims were relatively rare (Bailey & Tolnay, 2015), perhaps due to gender norms at the time that were less likely both to view females as blameworthy and to view a lynching as just punishment for them. If this were the case, one would also expect the atrocity of lynching events with female victims to be lower. Bailey and Tolnay (2015) wrote that women were more likely than men to be lynched as secondary—rather than central—targets of the mob. This perspective casts female victims as less criminal than their male counterparts, and hints that the atrocity of events with female victims may have been different.
Finally, it should be mentioned that broad temporal norms regarding lynching atrocity may also have existed. Lynchings clearly displayed an overall trend of growth and fall at the turn of the century (Tolnay & Beck, 1995). The varying frequency of lynchings over time is consistent with broader research on violence and punishment (Pinker, 2011). In the United States, there have been fluctuations over time in homicide (Baller, Anselin, Messner, Deane, & Hawkins, 2001) and imprisonment rates (Jacobs & Carmichael, 2001). Given that the frequency of lynchings changed over time, norms regarding the method of lynching may also have changed over time.
Social Distance
Aside from deindividuation and norms that might have governed lynchings, as mob members felt more removed from the lynching victim, there should have been greater social distance from the victim and greater violence toward the victim. Social distance increases between people who do not (a) participate in each other’s lives, (b) share a common culture, (c) rely on each other, or (d) hold equal status (Senechal de la Roche, 1996). Dissimilarity between mob and victim increases social distance, decreases sympathy, and allows for greater violence to occur. The importance of social distance may be evidenced by local communities having been more likely to lynch strangers than members of the community (Brundage, 1993).
Black (1993) suggested that individuals favor people who are socially close. This bias also reflects Allport’s (1954) concept of individuals favoring their ingroups. And while ingroup bias does not necessarily lead to outgroup hostility (Brewer, 1999), individuals conform to the behaviors of their ingroups (Crandall & Stangor, 2005). In the context of lynching, the closer the crowd feels toward the victim of the alleged crime (and further from the target), the more likely lynching becomes (Senechal de la Roche, 2001). Considering the social distance between mob and victim allows for victim characteristics to affect the actions of the mob. For example, a White mob may have felt distant from a Black target, and a “moral” mob may have felt distant from a target accused of sexual assault. When threatened, individuals overidentify with their ingroups (Aviram, 2009). Thus, threatening crimes like murder should have increased the social distance between lynch mobs and their victims.
Most discussion of how social distance affected the lynching event, however, is educated conjecture. While they did not write explicitly about social distance, some of the variables in the work of Mullen (1986) and Leader et al. (2007) could be interpreted as measuring social distance between mob and victim. Mullen (1986) categorized five types of crimes that mobs accused victims of committing, and Leader et al. (2007) ranked alleged crimes on a 7-point scale based on the federal sentencing guidelines. Lynching victims accused of both rape and murder (Mullen, 1986) and more severe crimes in general (Leader et al., 2007) suffered greater atrocity at the hands of the lynch mob. While neither of these studies theorized why this relationship exists, it could be that more severe crimes increased the social distance between victims and mob, allowing mobs to commit greater atrocities.
Social distance and ingroup literatures would suggest that greater social distance encourages more atrocious types of violence (Senechal de la Roche, 1996). An alternative explanation for moral social distance would be to suggest that serious crimes have serious consequences. Therefore, graver crimes—such as murder and sexual assault—should correspond to more retributive justice.
Data
Besides updating the theoretical perspective of Mullen (1986) and Leader et al. (2007) to include group norms and social distance, this study analyzed a new dataset. While this new dataset has some limitations, the data used by Mullen (1986) and Leader et al. (2007) likely suffered from selection bias, and were limited in their size, scope, and reliability.
In Mullen’s (1986) original study, Ginzburg’s (1962) collection of newspaper articles in 100 Years of Lynching served as the data source, and Mullen (1986) was able to analyze 60 of the 300 lynching events. In addition to using these lynching events, Leader et al. (2007) captured 22 additional lynchings (from 98 photographs) from Allen, Als, Lewis, and Litwack’s (2000) Without Sanctuary: Lynching Photography in America. While both of these efforts should be applauded for the detailed variables they were able to obtain, both Mullen (1986) and Leader et al. (2007) conducted their analyses with fairly small samples.
Despite the small size of their datasets, Mullen (1986) and Leader et al. (2007) used data with a broad temporal and geographic scope. Scope matters because lynchings varied over time and place (Corzine, Creech, & Corzine, 1983; Reed, 1972; Tolnay & Beck, 1995). The 60 lynchings identified by Ginzburg’s (1962) articles spanned the period 1899 to 1946, the 22 lynchings identified by Allen et al.’s (2000) photographs spanned the period 1890 to 1935, and both articles and photographs came from across the country. Using so few cases with such a large temporal and geographic scope likely limited the possibility of detecting any temporal or geographic patterns in the lynchings.
Finally, the reliability of the measures used by Mullen (1986) is unclear, as he did not provide any indicators of reliability. Mullen’s (1986) lynching atrocity index could not be recreated in the present study because certain variables (e.g., the mutilation of the victim and duration of the event) used to create the index could not be reliably coded. It is possible that Mullen’s (1986) coding of these variables was reliable, whereas ours was not. But it seems at least as likely these unreliable variables were too rare and/or too subjective to be reliably coded by either Mullen (1986) or our raters. Although Leader et al. (2007) did include information on the reliability of their measures from coding Allen et al.’s (2000) photographs, they did not report the reliability of their coding of the Ginzburg (1962) data, meaning that some of Leader et al.’s (2007) measures may also lack reliability.
Besides the unreliability of coded variables, it is possible that Ginzburg’s (1962) newspaper articles and Allen et al.’s (2000) photographs were not generally representative of lynchings. Ginzburg (1962) and Allen, Als, Lewis, and Litwack (2000) may have chosen lynching events that were particularly sensationalistic, or in other ways not representative of all lynching events in order to tell a story about the horror and tragedy of lynching events. If the books included a disproportionate number of events with large mobs committing high-atrocity lynchings, then analyses of these data become more likely to identify patterns connecting larger mobs to higher atrocity lynchings. If Mullen (1986) and Leader et al. (2007) selected their samples from sources biased in such a way, their findings may not hold when examining a different sample of lynching events. Below, we present evidence that these studies may have oversampled hangings.
Hypotheses
While the present study could not test whether anonymity is related to either deindividuation or group norms, anonymity should have affected the violence of lynching events. The violence employed by lynch mobs should have covaried with a number of characteristics of the lynching events related to the anonymity of lynch mob members, leading to the following hypotheses:
Besides anonymity, the literature on norms would suggest that both broad societal norms—such as gender—and situational, group norms—such as the characteristics of prior lynching events—would affect the violence of lynching events.
The final theoretical perspective, on social distance, proposes the relationship between the lynch mob and victim as contributing to the severity of violence during a lynching event.
Method
Data Collection
Newspaper articles from The Atlanta Constitution, collected by Mattias Smångs (2016), composed the data to be analyzed. Articles were limited to The Atlanta Constitution because, being situated in the biggest city in Georgia, the paper had the most extensive coverage over the entire state. It is also one of the only major Southern newspapers available in full-text online. Georgia is an appropriate state for this study because it was second only to Mississippi in the number of lynching victims (Tolnay & Beck, 1995), and has been regarded as representative of the Deep South (Brundage, 1993). An advantage of concentrating on one state is that by coding all known lynchings in one area, selection bias due to concentrating on particularly infamous lynchings was less likely to occur. In this sample, there was no statistically significant relationship between the number of words in an article and the level or type of violence during a lynching event, suggesting that reports did not concentrate particularly on violent lynchings. The first author consulted with Smångs to create a list of variables reflecting basic information about each lynching, including variables laid out in the earlier work of Mullen (1986) and Leader et al. (2007). Two undergraduate researchers, blind to hypotheses, were trained in coding using randomly chosen cases from Ginzburg’s (1962) 100 Years of Lynching. Raters then coded all cases from The Atlanta Constitution database. The Historical American Lynching (HAL) dataset (Hines & Steelwater, n.d.) served as a guide for coders in identifying lynching victims. Raters added lynching victims identified by The Atlanta Constitution but not listed in the HAL dataset. The two datasets share 430 lynching victims, 1882-1926. Over this time span, the HAL dataset has 23 unique lynching victims, and The Atlanta Constitution dataset has 85 unique lynching victims.
In all, coders read 548 articles, identifying 515 victims killed in 411 lynching events. These lynching events, 1882-1926, are plotted by location, decade, and level of atrocity in Figure 1. Events were mapped according to the city nearest to the lynching when it was mentioned (n = 310). If no city was specified, events appear according to the city from which the article was reported and filed, as analyses of events with both lynching location and report location (n = 300) indicated that 76% of the cases were reported from locations less than 15 miles from the lynching location, and an additional 21% were reported from less than 50 miles away. Four events in the data occurred outside of Georgia: one in Tennessee (dropped from the analysis); two on the border of Georgia in Eufaula, Alabama; and one less than 20 miles from the border in Centre, Alabama. After dropping events missing data on the dependent variable—explained below—the dataset included 481 victims killed in 383 lynching events.
Lynching events in Georgia, 1882-1926.
Reliability
Each rater coded all of the data, resulting in the full sample being used to calculate the reliability of each variable. The reliability statistic used for this study was Krippendorff’s alpha, due to its ability to account for varying levels of measurement and missing data (Hayes & Krippendorff, 2007). As a conservative estimate of reliability, variables with alpha values above .800 are considered reliable; tentative conclusions can be made regarding variables with alpha values above .667 (Krippendorff, 2004). Overall, as shown in Table 1, coding reliability was high. While most variables were fairly straightforward (e.g., whether a lynching victim was shot), the general mob size estimate was a more subjective measure assigned by each coder. However, its high Krippendorff’s alpha value (.846) here allowed for a more confident use of the variable.
Reliability of Coded Variables.
Three variables of note were the race of the victim, the gender of the victim, and the presence of disguises. The race of the victim and gender of the victim were included in the HAL database on which the coding was based, meaning the two raters each relied on the HAL values, thus biasing the variable toward the HAL coding. Also, the trainer created the presence of disguises variable post hoc after noting that the two original raters frequently recorded the presence of disguises as part of a qualitative variable recording other mob characteristics. To check the reliability of these variables, the trainer recruited a third rater. This rater was trained using previously coded cases selected for their variation in the variables of interest. The rater then coded a random sample of 20% of the total cases—104 cases total—revealing these additional variables to be reliable, as shown in Table 1.
Dependent Variables
There were two sets of dependent variables for this study: specific types of violence variables and lynching atrocity variables. The four types of violence examined here were whether a lynching event included beating, shooting, hanging, or burning. Models using these initial dependent variables predicted the type of violence that occurred during a lynching event but not necessarily the atrocity of a lynching event. The 27 events that did not report any of these four types of violence were excluded from the analysis, though the results presented here were similar to analyses that included these cases.
The atrocity of a lynching event was operationalized in three ways in order to test for convergent findings. As reviewed above, the literature conceptualizes the severity of lethal violence as related to the amount of violence and the number of victims. The first measure of atrocity (Atrocity 1) addressed the amount of violence and was the sum of whether a lynching included beating, shooting, hanging, or burning, treating each variable as a binary. The more types of violence perpetrated, the more atrocious the lynching. This measure was similar to Mullen’s (1986) measure of lynching atrocity, though the measure here included beating but excluded stabbing, mutilation, and duration of the event—as raters did not code those variables reliably. Other rare types of violence (e.g., drowning) were also excluded because of low reliability. 2 The reliable indicators of atrocity could not be formed into a Guttman scale, as there was too much deviation and a low coefficient of reproducibility (Guttman, 1944). More conservative, probabilistic models—Rasch’s (Jong-Gierveld & de Kamphuls, 1985) and Mokken’s (Schuur, 2003) scale analyses—similarly indicated that the acts of violence did not represent a scale. The second measure of atrocity (Atrocity 2) addressed severity as number of victims and was a dichotomous variable representing the presence of multiple lynching victims in a lynching event. In this second operationalization, lynchings with multiple victims were treated as more atrocious than lynchings with just one victim. 3 The third measure of atrocity (Atrocity 3) was a sum of five dichotomies, accounting for both the types of violence used in an event (beating, shooting, hanging, burning) and the presence of multiple victims in an event.
Independent Variables
Five variables addressed the first set of hypotheses, accounting for characteristics affecting the anonymity of mob members during a lynching event. The first anonymity variable was mob size, which had four categories: 3-49, 50-99, 100-999, and 1,000 or more participants. Coded mob size, which was based on the reported number or on contextual clues, 4 was highly reliable. The second variable affecting anonymity was a dichotomous variable representing the presence of masks or disguises among mob members. The third variable indicated whether or not the lynching took place during the night. The fourth variable was whether or not the lynching victim was taken from law enforcement custody. The presence of law enforcement and the ability of law enforcement to identify individuals would reduce potential anonymity. The final variable measuring the anonymity of a mob was the Other-Total Ratio used by Mullen (1986) and Leader et al. (2007). The Other-Total Ratio is the number of victims in an event divided by the total number of victims and perpetrators at the event. The number of victims was straightforward, but the number of perpetrators at the event was computed in four ways, consistent with the work of Mullen (1986) and Leader et al. (2007), in which mob sizes were averaged from numerical estimates: (a) the reported mob size, (b) the midpoint of the range of reported mob size, (c) the lower-bound of reported mob size if no upper-bound was reported, or (d) the upper-bound of reported mob size if no lower-bound was reported. Because the number of perpetrators was not reported in all newspaper articles, the Other-Total Ratio could not be derived for all lynching events. Moreover, besides the smaller sample size associated with using the Other-Total Ratio, the second and third measures of atrocity accounted for the presence of multiple victims. It was inappropriate in these cases, then, to predict the presence of multiple victims with a measure that accounted for the number of victims. Due to the stated limitations using the Other-Total Ratio imposed on the analyses, the variable was used only after the initial models were run.
While the anonymity variables could have provided evidence in support of either the deindividuation perspective or the group norms perspective, there were three variables that more specifically addressed norms. The first variable attempted to establish the group norms of a lynch mob. This measure was the average violence of nearby lynching events—for example, the average atrocity level or the average use of hanging in nearby lynching events. Nearby was operationalized as all lynchings that occurred within a 50-mile radius during the previous 5 years. 5 This variable accounting for nearby lynchings was limited in five ways: First, it treated small and large lynchings as having the same effects on other lynchings. Second, the measure did not account for lynchings outside of Georgia, or the fact that geographic distance may have been differentially shortened across the state by methods of transportation (roads, trains) and communication (newspaper). Third, because the variable was calculated using lynching events from the prior 5 years, the characteristics of nearby lynchings for lynching events that occurred during the first 5 years of the dataset were unknown and assigned missing values, effectively removing these lynchings from the analyses. Fourth, eight lynching events had no other events within 50 miles during the previous 5 years. These events were given missing values, although the results were nearly identical when the weight was set to 0. Finally, other operationalizations of what constituted “nearby” were possible, though as explained in Endnote 5, they did not seem to capture the concept. The other two variables accounting for the norms of lynching events were whether a female was lynched and the decade of the lynching event. The dichotomous sex variable measured the extent to which gender norms existed regarding the atrocity of lynching events. Considering the issue of blameworthiness, female victims should have suffered less atrocious lynchings than male victims. And while the nearby lynching events variable was limited to events within the previous 5 years to capture specific group norms, additionally controlling for the decade of the lynching event should have helped reveal broader temporal norms.
Three final variables represented the social distance between the lynch mob and the victims, addressing the third set of hypotheses. As lynch mobs were primarily White, the first variable was whether or not Black victims were killed during a lynching event. The alleged crime of the victim could further distance the victim from the lynch mob. Two dichotomous variables—whether a victim was accused of (a) murder or attempted murder, or (b) sexual assault or attempted sexual assault—signified socially unacceptable crimes that would distance the victims from the lynch mobs.
There are a number of features to point out in the descriptive data as shown in Table 2. To begin, the means of the atrocity variables were relatively low. Most lynching events featured only one type of violence and one victim. Second, shooting was the most common type of violence committed. Shooting happened alone and with other types of violence, but it was surprising to see shooting so common when the violence commonly associated with lynching is hanging. The frequency of hangings as reported in the present dataset based on The Atlanta Constitution suggests that Mullen (1986) and Leader et al. (2007) may have oversampled lynching events featuring hangings, thereby bringing the generalizability of their studies into question. 6 Burnings and beatings, on the other hand, were relatively rare. Third, the Other-Total Ratio values were low and had a small range (from 0.0002-0.333). We considered standardizing the measure to range from 0 to 1, but decided against doing so in order to more closely match the measure used in the prior two studies. Fourth, many lynching events involved victims accused of murder and/or sexual assault. While it was unlikely that all of these lynching victims were guilty of their alleged crimes, these accusations were one way in which lynch mobs may have justified their actions (in addition to being representative of greater social distance between the mobs and their victims). Lastly, in addition to the above-mentioned lynchings occurred during the first 5 years of the dataset limiting the analyses due to missing data on the characteristics of nearby events, the inconsistent newspaper reporting of mob size and whether a lynching occurred at night further limited the analyses that could be conducted. While the descriptive statistics reported in Table 2 are for all 411 coded lynching events, not all observations were included in subsequent analyses due to missing values. A correlation matrix suggested that values were missing at random. Further, because analyses conducted after multiple imputation procedures yielded similar results, only the analyses with the original data are presented here.
Descriptive Statistics.
0.0002.
Results
As the dependent variables were either dichotomous or ordered scales, the primary analyses were a series of logistic and ordered logistic regressions. There were seven dependent variables in total: three operationalizations of atrocity—Atrocity 1 (total types of violence perpetrated), Atrocity 2 (presence of multiple victims), Atrocity 3 (total types of violence perpetrated and presence of multiple victims)—and four specific types of violence—beaten, shot, hanged, and burned.
Analysis occurred in two steps: First, models accounted for variables associated with anonymity (excluding the Other-Total Ratio), norms, and social distance. Second, models were run again accounting for the Other-Total Ratio, excluding models for Atrocity 2 and Atrocity 3 since both of those dependent variables and the Other-Total Ratio accounted for the number of lynching victims. Table 3 highlights the statistically significant (p < .05) relationships across all 12 models (see the appendix for tables with complete coefficients, standard errors, odds ratios, and confidence intervals). 7
Direction of Statistically Significant Coefficients.
Note. excl. = variable and accompanying observations dropped from the model due to quasi-complete separation. DV = dependent variable.
Atro = Atrocity. O-T = Other-Total Ratio. + = significant positive correlation. – = significant negative correlation.
Anonymity should have grown as lynch mobs grew in size, wore disguises, acted at night, and avoided law enforcement. This anonymity should have promoted more atrocious violence, though in this study, as noted, we cannot determine whether the mechanism fueled by anonymity was deindividuation or activated group norms. Prior to accounting for the Other-Total Ratio, mob size was a statistically significant predictor of lynching atrocity, providing partial support for Hypothesis 1. Lynching events with larger mobs were correlated with both more types of violence and multiple victims. Lynching events with larger mobs were also more likely to involve hangings and burnings. The other anonymity variables were not consistently significant predictors. Taking the victim from law enforcement increased the likelihood of hanging, though it decreased the likelihood of shooting; disguises were predictive only of events with multiple victims; and the occurrence of a lynching event at night was correlated only with the victim being beaten.
After accounting for the Other-Total Ratio used by Mullen (1986) and Leader et al. (2007) as a deindividuation variable, mob size continued to be a significant predictor of the atrocity of violence, as well as certain types of violence, such as burnings. Taking the victim from custody similarly continued to increase the likelihood of hangings. Regarding the new variable, the lower the Other-Total Ratio, the more lynchers who were present relative to the number of victims, the greater the anonymity, and the higher the level of expected atrocity. The Other-Total Ratio positively predicted the shooting of a victim, suggesting that mobs were more likely to shoot victims when the mobs were less anonymous. The only dependent variable with the expected negative relationship with the Other-Total Ratio was hanging, suggesting that hangings occurred when mobs were more anonymous, consistent with the findings of Mullen (1986) and Leader et al. (2007), which may have oversampled events with hangings.
The norm variables of Hypothesis 2 received some tentative support. The total amount of violence and the use of hanging in a lynching event were positively correlated with the total amount of violence and the use of hanging in nearby lynching events (within 50 miles and within the prior 5 years). Further, the third measure of atrocity, beatings, and burnings had positive relationships approaching statistical significance with their measures in nearby events, as shown in the appendix. While the relationship was suggestive of group norms, the significance of the relationship went away after controlling for the Other-Total Ratio. Other variables influencing the norms of a lynching event were also significant before introducing the Other-Total Ratio. The presence of a female victim was significantly correlated with a lynching event having multiple victims. Also, events with female victims were dropped from analyses of beatings, since no article about a lynching event with a female victim reported beatings as having occurred. There also seemed to have been some larger temporal norms, as the decade in which a lynching occurred was correlated with the amount and type of violence. Compared to the 1900s, the 1880s were so unlikely to have burnings that none occurred. The 1890s were also less likely to feature burnings. Lynching events during the 1910s were not significantly different from the events of the 1900s. Finally, 1920s lynchings were more likely to have beatings (though only two were reported in the 1920s). Overall, the atrocity and types of violence used in a lynching event did seem influenced by local, gender, and temporal norms, though these relationships were not apparent in models including the Other-Total Ratio, partially supporting Hypothesis 2.
Social distance variables did not have much explanatory power in this study. Social distance between victim and mob should have been larger for Black victims and for victims accused of murder and sexual assault, resulting in more atrocious lynching events. In the initial models excluding the Other-Total Ratio, the presence of a Black victim may have influenced the violence of a lynching event, as measured by Atrocity 3, though this variable was dropped in the models predicting beatings and burnings, as only lynchings with Black victims included beatings and burnings in these models. Victims accused of murder were positively correlated with lynchings involving multiple victims and burnings. Lynching events with victims accused of sexual assault were more likely to feature hangings but less likely to involve multiple victims or beatings. After incorporating the Other-Total Ratio, social distance variables were mostly nonsignificant, though victims accused of sexual assault were more likely to be shot. Also, victims accused of sexual assault were dropped from the model predicting beatings, as no lynchings of this type involved reported beatings. The social distance variables provided mixed support for Hypothesis 3.
Models accounting for interaction effects between variables were computed in addition to the models presented here, since the predictors may have interacted with each other. The existence of a lynching at night, for example, may have interacted with mob size. However, testing for all 225 possible interactions between any two independent variables (45 interactions for each of five dependent variables) yielded only five significant interactions that maintained significance after introducing the Other-Total Ratio to the models. Because this number is less than chance and because there did not appear to be any pattern in these significant interactions, we conclude that no interaction term significantly improved models consistently across the inclusion or exclusion of the Other-Total Ratio.
Discussion
This analysis of lynching events in Georgia has two main findings: (a) lynching events with large mobs had higher levels of atrocity, in terms of both the total types of violence committed against the victim and the number of victims; and (b) certain norms seemed to exist during lynching events, evidenced by the correlation between the violence perpetrated during lynching events and the violence used in subsequent nearby lynching events. The importance of mob size suggested that as mobs grew, members became more likely to use more types of violence against victims and to target multiple victims. Although the relationship could have gone the other way, with high-atrocity lynchings attracting larger mobs, a review of the original newspaper articles does not seem to support this argument. While people did join lynch mobs after they formed initially and observers would gather to view lynched bodies, most lynch mobs seemed to have been formed to pursue their victims—before any types of violence were committed. And though the pursuit of multiple victims could have encouraged larger mobs, many lynchings with multiple victims were focused on only one victim, with additional victims lynched due to their associations with the primary victim. The causality almost certainly went from crowd size to atrocity.
The significance of mob size provides evidence that the anonymity imparted by large mobs resulted in more atrocious lynchings, though it is unclear whether this violence was due to deindividuation or group norms. Mullen (1986) and Leader et al. (2007) focused on the Other-Total Ratio as representative of the deindividuation perspective, but the Other-Total Ratio implemented here was not a consistent predictor of mob violence. Overall, the Other-Total Ratio should be used carefully for future lynching studies for two reasons: First, using the Other-Total Ratio necessitates dropping the many lynching events with no numeric mob size information. Second, depending on the sensitivity of the Other-Total Ratio, it is unclear how well it can be applied to lynching events, as reports on lynching events that include numeric mob size generally only include rough estimates (e.g., a reported 500 participants is unlikely to represent exactly 500 participants). That said, the Other-Total Ratio had a predictive influence in the models here, and the measure could be useful in other studies. The Other-Total Ratio seems particularly suited for investigating dynamics between small groups of people, and it could also be applied to other types of violent events besides lynchings, such as race riots. While Mullen (1986) and Leader et al. (2007) did find evidence suggesting that the Other-Total Ratio was predictive of lynching atrocity, we suspect that their data may have oversampled lynchings that featured hangings—the only measure of violence in our study that supported the Other-Total Ratio as a predictor of violence in the hypothesized direction. As events with hangings by definition have higher overall levels of atrocity than events without hangings, it is possible the Other-Total Ratio in Mullen (1986) and Leader et al. (2007) was correlated with hangings rather than overall levels of atrocity as interpreted.
In addition to affecting the atrocity of a lynching event, mob size affected the specific acts of violence undertaken by the mob. Hangings and burnings became more likely when mobs surpassed 100 members. Perhaps, hangings and burnings were more severe types of violence pursued only when mob members felt sufficiently anonymous. This suggestion may explain why the Other-Total Ratio had a positive relationship with shooting but a negative relationship with hanging. Another possible reason could be that hangings and burnings are more grand-scale, public spectacles than shootings. In order to fulfill individuals’ desires to observe the killing, hangings and burnings may have been a practical necessity for large mobs. Further, hangings and burnings may have been impractical with fewer lynchers. Smaller lynch mobs may not have had the required physical control over the victim in order to hang or burn the victim, instead being more likely to resort to shooting.
The second main finding of this research, and unique to this study, was that the violence of lynching events was correlated with the violence of nearby lynching events, thus revealing the potential importance of situational norms. More than just the general atrocity of violence, the specific use of hanging as a method of violence was related to whether nearby mobs utilized hangings. Situational norms could explain this effect in three ways: First, individuals in one lynch mob could have participated in nearby lynch mobs as well, bringing their tactics of violence along with them. Second, through learning via newspaper reports (as in the newspaper articles used in this study) and word of mouth, a lynch mob could model its behavior on the behavior of nearby lynch mobs, committing specific acts to fulfill expectations of violence. Third, there could have been localized cultures supportive of specific types of violence, with certain areas predisposed to commit different types of violence. While Figure 1 seems to display a band of lynchings through the middle of Georgia, global Moran’s I—a measure of spatial autocorrelation—values for shootings, hangings, and burnings across lynching events were low and nonsignificant, indicating that types of violence did not display strong spatial structuring (Anselin, 1996). For these reasons, this third explanation does not seem persuasive. However, rather than spatial relationships being solely determinative, our results suggest that the importance of nearby lynching events was structured by both space and time. Our operationalization of “nearby” as within 50 miles and within the prior 5 years is thus suggestive of the possible significance of situational norms.
Besides these main findings, there were four other interesting patterns to explain: To begin, it was hypothesized initially that the presence of law enforcement should have decreased the anonymity of group members, thereby decreasing lynching atrocity. This hypothesis was not supported. While acquiring the victim from law enforcement custody did not impact the atrocity of a lynching, it significantly increased the likelihood of a hanging, even after taking into account the Other-Total Ratio, as shown in Table 3. One reason for this effect of law enforcement could be that even though lynch mobs took victims from protective custody, the presence of law enforcement may have emphasized to lynch mob members their role as vigilante justice-keepers, consistent with contemporary research that has identified lynchings as a type of informal social control (Black, 1983; Cooney, 2003). Interacting with law enforcement may have activated law-and-order situational norms. At the time, hangings were the normal, approved method of lawfully killing criminals. By hanging their victims, lynch mobs simulated state-sponsored executions. In this unexpected way, the presence of law enforcement may have changed the situational norms of lynching events. An alternate explanation is that taking a victim from law enforcement may have frequently happened with tacit approval, thereby guaranteeing the anonymity of mob members. If this were the case, the “official” anonymity should have increased all types of violence. However, we found that the presence of law enforcement was correlated with a lower occurrence of shootings, suggesting that mobs were more or less likely to pursue different types of violence depending on the norms provided by the presence of law enforcement.
In regard to the norms of a lynching event, lynchings with female victims were correlated with lynchings that had multiple victims. This finding is consistent with the work of Bailey and Tolnay (2015), which found that female victims were more likely than males to be lynched due to their associations with the primary victim. We connect this pattern with the gender norm of blameworthiness. Females may have been viewed as less culpable than males, resulting in both fewer females being lynched and fewer females being lynched by themselves. Regardless of specific group norms, lynch mobs still abided by broader, societal gender norms. In addition, even if lynch mobs were not concerned with the broader gender norms, then newspaper reporters may have been. It is notable that no female lynching victims were reported as beaten. This may be because no female lynching victims were beaten, but press coverage of lynchings was not completely unbiased, with reporters often placing the blame on the victims (Perloff, 2000). In a similar way, reporters may have chosen not to report females being beaten in an effort to not offend the sensibilities of potential readers.
The third pattern of interest was that social distance variables did not consistently predict atrocity or types of violence. It is worth noting, though, that burnings occurred only at lynching events with Black victims. Also, victims accused of sexual assault were less likely to be lynched alongside other victims (Atrocity 2). This result seems in line with the concept of lynchings as a way for justice to be served. In the case of sexual assault, mobs seem to have focused on lynching only the individual accused of committing the crime. The argument here is not that lynch mobs achieved justice. Rather, considering the effects of acquiring victims from law enforcement and lynching alleged sexual assault offenders without lynching others alongside them, lynch mobs acted in a way consistent with the idea that lynch mobs were supposed to obtain justice for the victim of the alleged crime and the community. Contrary to this explanation, victims accused of murder were more likely to be lynched alongside others. Perhaps, this was due to mobs having presumed alleged murderers, compared to alleged rapists, generally had more accomplices. But that possibility must await further research into the alleged crimes that were the ostensible basis for the lynchings. Overall, these variables representing increased social distance between the mob and the victim did not paint a clear picture that increased social distance led to increased atrocity. For example, from a social distance perspective, it is unclear why a victim accused of sexual assault was not more likely to be burned.
The final pattern observed through the models was that beating was not predicted in the same way as any other type of violence. Beatings were poorly predicted by the independent variables used here. In Table 3, beatings were more likely at night and during the 1920s, but no variable was significantly correlated with the use of beatings after accounting for the Other-Total Ratio. That no other variables were statistically significant predictors of beating seems counterintuitive. Compared to shooting, hanging, and burning, beating seems like a relatively mild form of violence, but it was the least commonly reported. Perhaps, this counterintuitive finding was due to news coverage of homicides favoring novel circumstances (Gruenewald, Chermak, & Pizarro, 2013). Newspapers may not have accurately reported beatings because they were so common. The potential importance of whether newspapers accurately reported beatings in general reflects the discussion above on whether newspapers accurately reported the beating of female victims, specifically. The norms of newspaper reporting influence what data can be drawn from articles. Interestingly, these norms could also relate to the oversampling of hangings in the work of Mullen (1986) and Leader et al. (2007). As mentioned earlier, Ginzburg’s (1962) collection of articles and Allen et al.’s (2000) collection of photographs could have been curated by their authors to help demonstrate the horror of lynchings. But these collections may also have been influenced by the desire to show stereotypical lynchings. It is possible that hangings feature so prominently because there continue to be accepted norms of what lynchings look like.
Limitations
Even with the advances noted above, there are two limitations in this research: First, this study concentrated on lynching events in Georgia. Although Georgia has been suggested as being representative of the Deep South (Brundage, 1993), lynchings in Georgia may have operated differently from lynchings elsewhere. For instance, as lynchings were relatively commonplace, group norms may have been easier to establish. Also, by concentrating on Georgia, this study largely ignored the effects of nearby lynchings in the bordering states of North Carolina, South Carolina, Florida, Alabama, and Tennessee.
Second, certain variables could be called into question. It should be noted that the relationship between lynching events and nearby lynching events was dependent on the operationalization of what constituted nearby. However, as explained in Endnote 5, the operationalization of nearby as events within 50 miles and from the previous 5 years is the best conceptual operationalization for capturing local norms. Further, a number of variables were simplified for this study. Mob size was categorized in order to keep more observations, and decade time periods were created to explore broad, temporal trends. But the reliability of the mob size variable gives confidence in its use, and different categorizations of time revealed similar trends. Overall, consistency across the repetition of measures and models increases our confidence in the results.
Conclusion
This study is one of the few quantitative analyses of lynching events. The present study used data that captured the characteristics of all known lynchings in Georgia. This dataset—with more observations, more variables, a geographic and temporal focus, and measurable reliability—means that this analysis is the most in-depth, quantitative study of the characteristics of lynchings to date. Moreover, it improves on the work of Mullen (1986) and Leader et al. (2007) by introducing situational norms into the lynching literature. As they found, mob size was important to lynchings, as it helped predict how many types of violence were committed against the victim. However, other characteristics of lynchings seemed to establish a variety of norms that also helped determine the violence committed in that and in nearby lynchings. The significant relationship between the violence of a lynching event and the violence of nearby lynching events suggests that Marsh’s cartoon correctly captured the observational learning that went on at lynchings: People learned from their first and from subsequent lynchings.
Footnotes
Appendix
The Use of Burning in a Lynching Event.
| B | SE | OR | 95% CI | b | SE | OR | 95% CI | |
|---|---|---|---|---|---|---|---|---|
| Anonymity | ||||||||
| Other-Total Ratio | 10.645 | (11.120) | 41990.35 | [−11.149, 32.439] | ||||
| Mob Size 2 (50-99) | 0.188 | (0.809) | 1.207 | [−1.397, 1.774] | 1.377 | (1.823) | 3.963 | [−2.196, 4.950] |
| Mob Size 3 (100-999) | 1.988** | (0.644) | 7.299 | [0.726, 3.249] | 3.162* | (1.611) | 23.627 | [0.006, 6.319] |
| Mob Size 4 (1000+) | 4.378*** | (1.064) | 79.684 | [2.292, 6.464] | 7.113*** | (2.162) | 1227.23 | [2.875, 11.350] |
| Disguises | −1.615 | (1.183) | 0.199 | [−3.933, 0.703] | 0 | Omitted | 1 | |
| Night | 0.361 | (0.593) | 1.435 | [−0.801, 1.522] | 0.369 | (0.806) | 1.446 | [−1.210, 1.948] |
| Taken from custody | −0.401 | (0.551) | 0.670 | [−1.481, 0.680] | 0.770 | (0.827) | 2.159 | [−0.852, 2.391] |
| Norms | ||||||||
| Avg. DV—Nearby events | 2.584 | (1.640) | 13.253 | [−0.630, 5.798] | −2.384 | (2.793) | 0.092 | [−7.858, 3.091] |
| Female victim | −0.516 | (1.229) | 0.597 | [−2.926, 1.893] | −1.574 | (1.623) | 0.207 | [−4.755, 1.607] |
| Decade—1880s | 0 | Empty | 1 | 0 | Empty | 1 | ||
| Decade—1890s | −2.317* | (1.048) | 0.099 | [−4.370, −0.264] | −1.674 | (1.527) | 0.187 | [−4.667, 1.318] |
| Decade—1910s | −0.752 | (0.636) | 0.471 | [−1.998, 0.494] | 0.949 | (0.927) | 2.582 | [−0.868, 2.765] |
| Decade—1920s | 0.098 | (0.806) | 1.103 | [−1.482, 1.678] | 0.638 | (1.453) | 1.893 | [−2.210, 3.487] |
| Social distance | ||||||||
| Black victim | 0 | Omitted | 1 | 0 | Omitted | 1 | ||
| Accused of murder | 1.923* | (0.956) | 6.845 | [0.050, 3.797] | −0.048 | (1.228) | 0.953 | [−2.455, 2.358] |
| Accused of sexual assault | 0.184 | (0.915) | 1.203 | [−1.609, 1.977] | −1.431 | (1.347) | 0.239 | [−4.072, 1.209] |
| Constant | −4.002*** | (1.229) | 0.018 | [−6.411, −1.594] | −4.819* | (2.202) | 0.008 | [−9.136, −0.503] |
| Observations | 225 | 99 | ||||||
| R2 | .324 | .392 | ||||||
Note. OR = odds ratio; CI = confidence interval; empty = variable category and accompanying observations dropped from the model due to quasi-complete separation; omitted = variable and accompanying observations dropped from the model due to quasi-complete separation; DV = dependent variable.
p < .05. **p < .01. ***p < .001.
Acknowledgements
We want to thank Mattias Smångs for providing the newspaper articles; Eric Dill, Eydsa La Paz, and Matthew Rhodes for coding the data; and Derek Kreager, Stephen Matthews, James Piazza, two anonymous reviewers, and Craig Anderson for their comments and suggestions.
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) disclosed receipt of the following financial support for the research, authorship, and/or publication of this article: This research was supported in part by the Instituto Carlos III-Juan March de Ciencias Sociales and the National Science Foundation’s Human and Social Dynamics Program (Grant #0729363 awarded to the second author)
Notes
Supplemental Material
Supplementary material is available online with this article.
References
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.
