Abstract
This study replicates and extends Archer and Gartner’s classic work testing whether wars increase postwar homicide rates because they legitimize the use of violence as a means of conflict resolution. Using the Comparative Crime Data File (CCDF), we replicated the original study with data from the two World Wars, as well as 12 smaller wars occurring prior to 1980. Our replication results generally confirmed the hypothesis that wars increase postwar homicide rates, although there were differences in results based on the method of analysis. We then examined the validity of this theory using data on four wars occurring after 1990, but found no support for the legitimation of violence argument. We argue that the null findings encourage theoretical expansion, which is an underappreciated aspect of replication that is just as valuable as empirical validation.
In 1984, Archer and Gartner published Violence and Crime in Cross-National Perspective, an award-winning book analyzing a unique collection of historical and cross-national data on interpersonal violence. In response to the “lamentably insular” state of criminology at that time, Archer and Gartner (1984) created the Comparative Crime Data File (CCDF) of crime and violence indicators for 110 countries and 44 major cities worldwide from 1900 to 1973. Their book documented how the data set was constructed, demonstrated how it could be used to answer compelling research questions about patterns and causes of violence, and outlined a research agenda that future criminologists could pursue with these data. Following best practices for scientific research, Archer and Gartner published the data as part of this book and also archived it electronically with the Inter-university Consortium for Political and Social Research (ICPSR) to promote transparency, accountability, and generativity.
This innovative approach to studying violence was an important contribution to the revitalization of comparative criminology at that time and has been cited in hundreds of subsequent publications. 1 However, there has been very little replication of any of the analyses from this book, despite the easy access to the data. Although scientists have debated the pros and cons of data sharing and replication in a number of disciplines (e.g., Abbott, 2007; Freese, 2007; Freese & Powell, 2001; King, 1995, 2003; Pridemore, Makel, & Plucker, 2018), the realities of publication expectations and the scarcity of outlets provide few rewards for this kind of research, despite the fact that most researchers will acknowledge the importance of replication in the scientific production of knowledge. However, recent failures to replicate research results from high-profile studies have gained the attention of mainstream media, thereby providing an external incentive for social science researchers to protect the integrity of their work through replication, transparency, and accountability (Handwerk, 2015; Meyer & Chabris, 2014).
Interestingly, the theoretical argument made in this study, that state violence can encourage interpersonal violence through modeling, has taken a backseat to structural theories explaining cross-national differences in violence. Yet recent meta-analyses of empirical tests of these structural theories show many inconsistencies in the results (Koeppel, Rhineberger-Dunn, & Mack, 2015; Nivette, 2011; Pridemore & Trent, 2010). Although there is modest support for modernization, social disorganization, and anomie theories in this body of literature, there are still many anomalies and unanswered questions about cross-national variations in homicide, begging the question of whether criminologists should consider other types of explanations, such as social process theories (Schaible, 2012). Process theories, including the social learning theory adopted by Archer and Gartner (1976, 1984), often require longitudinal data to allow for social processes to unfold over time, which has not always be available for a large number of countries. As a result, cross-sectional research has dominated cross-national studies of violence, which tend to preference social structural over social process theories. However, more systematic and complete cross-national data collections and better methods for analyzing longitudinal data can revitalize theoretical depth in this area.
Renewing attention to neglected theories in cross-national violence research, especially social process theories, can address some important, yet unanswered criminological questions. For example, it has been well established that countries with high levels of social inequality also tend to have high levels of violence, but it is less clear as to why that it is the case and how that happens. Schaible and Hughes (2011) applied reintegrative shaming theory to show how communitarianism mediates that relationship. It is also well known that the United States has historically had higher violent crime rates than comparably highly developed democracies, despite the fact that they have been declining since the 1990s (Lynch & Pridemore, 2011). Social learning theory could explain both the change over time and the relative stability in ranking, which structural explanations such as social inequality have not been able to do.
We argue that replicating the classic study of Archer and Gartner (1984) investigating the influence of wars on postwar homicide rates is important both to enhance scientific integrity and to promote theoretical development. In this article, we aim to replicate the original study to contribute to the dearth of replicability studies in criminology (Pridemore et al., 2018), reanalyze the original findings using more advanced statistical techniques to assess the robustness of the results, and extend the original analysis to examine the role of social learning in understanding postwar homicide rates in the 21st century.
The Original Study
In most societies, violence can be justified according to the perceived legitimacy of the perpetrators. A sovereign state, by definition, is vested with the authority to engage in legitimate violence, such as the use of force by law enforcement, the application of the death penalty, or the deployment of the military in the case of wars. One concern about the widespread use of legitimate violence is that it could serve as a model or justification for illegitimate violence. “What all wars have in common is the unmistakable moral lesson that homicide is an acceptable, even praiseworthy, means to certain ends” (Archer & Gartner, 1984, p. 66). Drawing on cognitive learning theory, the legitimation of violence model contends that the state’s use of legitimate violence will increase the likelihood of citizens’ engagement in illegitimate violence, such as homicide (Bandura, 1973). Increased exposure to violence normalizes the behavior at the macrolevel, which becomes internalized at the microlevel (Akers, 2013).
Archer and Gartner considered the legitimation of violence model alongside six other theories. First, in the artifacts model, homicide changes are demographic artifacts from wartime changes in population characteristics that decrease the pool of potential offenders (e.g., through conscription and the death of young men). Second, in the social solidarity model, wars increase social bonds and the rule of law, which should decrease crime during wartime and at least immediately afterward. Third, in the social disorganization model, anomie created by war would increase crime rates, particularly in defeated nations where the economic, political, and social systems completely break down. Fourth, in the economic factors model, the effects of scarcities generated by war and postwar unemployment on crime rates could either increase or decrease postwar crime rates depending on how well the postwar economy was performing compared with the prewar economy. Fifth, the catharsis model views the public violence of wars as a substitute for private violence, predicting postwar decreases in homicide rates. Finally, the violent veterans model suggests that veterans were conditioned to violence during wars and continued to seek it upon returning home, thereby contributing to postwar homicide increases.
Archer and Gartner did not conduct a strict test of the seven theoretical propositions due to data limitations regarding both the dependent and independent variables. Several of the theories emphasized wartime changes in homicide rates, which Archer and Gartner (1984) argued was nearly impossible to measure reliably due to a large number of social changes occurring at the same time, including reduced police capacity to accurately record crime statistics. Several of the theories specify mechanisms for which the proper testing would require data on variables, such as social solidarity, anomie, or military status of homicide offenders, that were not available for this sample of nations over the historical period of the study. Instead, Archer and Gartner (1984) proposed a controlled comparison of “a large and heterogeneous sample of cases to minimize the idiosyncratic experiences of individual nations or individual wars” to “maximize the chances of discovering differential effects of various types of war and various types of participation” (pp. 76-77).
Using homicide data from the CCDF and war data from Singer and Small’s (1972) Correlates of War data set, Archer and Gartner compared prewar and postwar homicide rates for all combatant countries for which there were data for World War I, World War II, and 12 smaller wars between 1896 and 1970. They averaged homicide rates for 5 years before the start of each war and 5 years after the end of the war. They first calculated the percentage change between the two time periods and classified the countries as showing a decrease in homicide rates, having no change (defined as less than 10% change in either direction), or a postwar increase in homicide. Then they conducted a t test between the mean prewar homicide rate and the mean postwar homicide rate to account for variance in the homicide rates and measure the size of the effect (Archer & Gartner, 1984). They repeated these steps for a set of control nations for each war. For WWI and WWII, the control nations were all noncombatant nations in the CCDF. For the 12 smaller wars, the selected control nations were the closest noncombatant neighbors of combatant nations that had sufficient data for analysis, as detailed in Table 1.
Twelve Other Wars Include in Archer and Gartner’s Study.
Note. ICPSR = Inter-university Consortium for Political and Social Research.
Data prior to 1900 were not available in the ICPSR data set or the printed publication.
For Vietnam only, the measure of change is between the prewar rate and the rate during the war given that the war was ongoing when data collection ended.
Table 2 shows Archer and Gartner’s results for WWI and WWII combined. Based on these analyses, they concluded that “combatant nations were more likely than control nations to experience homicide rate increases” (Archer & Gartner, 1984, p. 79). They also detected a similar pattern for the 12 other wars, but the distinction between combatant and control nations was less pronounced. Archer and Gartner (1984) concluded that “the nations participating in these twelve wars were somewhat more likely to experience increases than were the controls. For combatant nations in these twelve wars, those with increases outnumbered those with decreases by two to one” (p. 84).
Archer and Gartner’s Original Homicide Change Classifications for WWI and WWII.
Source. Archer and Gartner (1984, p. 80)
Note. WWI = World War I; WWII = World War II.
Crimes against the person; homicide included.
Murder and Manslaughter.
Denmark is included because it was occupied, although it never declared war.
p < .05. **p < .01. ***p < .001.
Although the general pattern for all wars analyzed by Archer and Gartner was that combatant nations were more likely to experience postwar homicide increases compared with control nations, there were nonetheless some combatant nations that did not experience postwar increases. To further identify characteristics of warring nations that might be more likely to generate postwar increases, they examined two other factors: (a) whether the warring nation was victorious or not and (b) the extent of the battle losses. These factors also help distinguish some of the competing predictions among the seven theoretical models that Archer and Gartner considered. They found that countries with greater combat losses were more likely to experience postwar increases in homicide rates than those with fewer battle deaths. In addition, victorious nations were more likely to exhibit postwar homicide rate increases than defeated nations.
Based on the results of these analyses, Archer and Gartner were able to rule out the social solidarity model, which predicted postwar homicide decreases for combatant nations, the social disorganization model, which predicted postwar homicide increases primarily for defeated nations, and the catharsis model, which predicted that countries with larger wartime deaths would have postwar declines in homicides. Next they examined supplemental data on demographic compositions of offenders, postwar economic changes, and offense rates for veterans for a small number of countries for which they could find data to assess the strength of the remaining theories that all predicted postwar increases in homicide rates for combatant nations. With this additional information, they ruled out the artifacts, economic factors, and violent veterans models. 2
As a result of these different analyses, Archer and Gartner concluded that their major findings were most consistent with the legitimation of violence model. Not only did wars increase postwar homicide rates, but victorious nations tended to have postwar increases more than defeated nations implying that violence was rewarded through victory, and countries with larger battle deaths were more likely to have postwar homicide increases as more visible war violence could encourage more interpersonal violence, all of which is consistent with learning theory. Importantly, they acknowledge that merely showing that rival explanations are disconfirmed or insufficient does not mean that the surviving theory is automatically the correct one. . . . However, the legitimation model is the only one of the seven presented here that is completely consistent with our finding of frequent and pervasive postwar homicide increases. (Archer & Gartner, 1984, p. 92)
Under the legitimation of violence model, wars provide concrete evidence that homicide under some conditions is acceptable in the eyes of a nation’s leaders. This wartime reversal of customary peacetime prohibition against killing may somehow influence the threshold for using homicide as a means of settling conflict in everyday life. (Archer & Gartner, 1984, p. 94)
Archer and Gartner’s study contributed to existing literature on the relationship between wars and interpersonal violence at that time in several important ways. First, they used a larger sample of countries and analyzed more wars than previous studies. Second, they utilized controlled comparisons to determine whether only combatant countries experience postwar homicide increases or if they were simply a part of a more general pattern. Finally, they considered the possibility that different levels of involvement in wars and different outcomes might also affect postwar homicide rates in addition to just war participation.
Literature Review
Subsequent studies have examined the relationship between war and homicide in the United States. For example, Kleck (1987) criticized Archer and Gartner’s sample selection because the control nations were largely less developed than the combatant nations, thereby introducing a potentially spurious relationship. Instead, Kleck (1987) conducted a time series analysis of homicide rates in the United States from 1937 to 1978 to assess whether wartime and postwar levels were significantly higher than prewar levels. He argued that the homicide rate increase for the U.S. following the Vietnam War is largely the result of a shift in age structure towards adolescence, limited ability of the state to repress crime through arrests and imprisonment, increases in the robbery rate, and persisting inequality. (Kleck, 1987, p. 246)
He acknowledged that the legitimation of violence model was still plausible because state representatives could have a vicarious modeling influence on individual aggressors, but he argued this was unlikely given that Americans do not tend to perceive state actors as behavioral models.
Other studies focusing on the United States have examined whether the size of the military affected homicide and suicide rates. Lester (1991) initially found no effect, but Lester and Yang (1991) discovered that the military participation rate was significantly related to the homicide rate for non-Whites, but not for other demographic groups. In a follow-up study, Yang and Lester (1997) found that military defense expenditures were negatively associated with homicide and suicide during wartime.
Several other studies examined the influence of wars on violence in countries other than the United States. In an analysis of 7,000 cases of assault in Portsmouth, England, between 1700 and 1781, Warner, Gmel, Graham, and Erickson (2007) did not find a significant relationship between war involvement and aggression. They argued that the legitimation of violence model may work “in societies in which violence is exceptional, but the situation may be more complicated in societies where violence is already endemic” (Warner et al., 2007, p. 595).
In a broader study, Landau and Pfeffermann (1988) looked at the role of security-related stressors, including war, on violent crime in Israel between 1967 and 1982. They found that the number of security-related casualties had a statistically significant, but marginal, effect on homicide rates, but not on robbery rates. Similarly, Biro and Selakovic-Bursic (1996) found that suicide and homicide rates in Serbia increased after the civil war in Yugoslavia in 1991, but they cautioned that other factors needed to be considered to explain this pattern, such as the war’s impact on the economy and cultural attitudes regarding the externalization of aggression.
Two more recent studies have provided additional support for the legitimation of violence theory. Dutton (2005) found that involvement in wars portrays social violence as justified revenge against targeted groups, which explains why genocides tend to occur after wars. In addition, Steenkamp (2005) argued that violent ethno-national conflicts created a “culture of violence” through international, state, collective, and individual mechanisms, which encourages further acts of violence.
In summary, several studies since Archer and Gartner’s (1976, 1984) work have provided empirical evidence documenting postwar increases in interpersonal violence, but the explanations for this pattern vary. Although there is some evidence supporting the idea that wars legitimize the use of violence more broadly, the studies also suggest that a number of other factors should be considered, including the nature of the conflict, economic effects of wars, and changes in the demographic composition of societies due to wars.
Hypotheses, Data, and Methods
It is not unusual to find mixed support for explanations of cross-national differences in violence because studies testing similar propositions use different samples of counties, different time periods, and different measures of key theoretical constructs (see Koeppel et al., 2015; Nivette, 2011; Pridemore & Trent, 2010; Stamatel, 2006). One of the advantages of replication studies is that changes in any of these three dimensions that could change the outcome of the study are more transparent than with new studies, allowing researchers to better discern if the differing results actually challenge current knowledge or if they might be an artifact of research design decisions. Using the same data and methods of Archer and Gartner, we expect to be able to replicate the following hypotheses for all 14 of the wars that they examined:
As described in more detail below, we first start with a faithful replication of the original study, then alter the time frame, sample, and analytic techniques to test the robustness of the results. Next we extend the analysis to assess how well the legitimation of violence argument holds up for more recent wars, which changes the sample and time frame of the study, although we use all of the same analytic techniques employed for the initial replication.
Data
Crime data for the CCDF were collected by Archer and Gartner (1984) over a span of 5 years from the following sources: (a) direct correspondence with national and local government officials; (b) published statistical reports and official documents, such as annual statistical yearbooks; and (c) secondary analysis of records from national and international agencies, such as Interpol. The data were eventually compiled into a uniform, computerized database containing crime data for 110 nations and 44 international cities. The authors also added population data to calculate crime rates.
The data used for the replication portion of this study were obtained from the ICPSR (8612) (Archer & Gartner, 1987), which included eight criminal offenses: murder, manslaughter, homicide, rapes, assaults, robberies, and thefts. We only used the categories of murder and homicide, unless otherwise noted. Most countries provided data for one category or the other, although some provided both. It was not always clear what the distinction between the two categories was for those nations that provided both. In those cases, we used whichever indicator came closest to matching the original results.
We used crime rates per 100,000 population for our analyses, although we occasionally found that we could better replicate Archer and Gartner’s original results using counts instead of rates. There were also some countries included in the original study that were not included in the ICPSR data set, which are noted in the results below. Whenever we found discrepancies with our replication results and the original results, we checked the ICPSR data against the data published in the book Violence and Crime in Cross-National Perspective (1984) to identify sources of mismatches, although we found that the print and electronic versions of the data provided the same information.
For the extended analysis we focused on four, late-20th/early 21st century wars: the Gulf War (1990-1991), the Kosovo War (1998-1999), the Invasion of Afghanistan (2001), and the Invasion of Iraq (2003). We identified wars from 1980 onward from the Correlates of War data set (Sarkees & Wayman, 2010) and then selected those for which we had sufficient homicide data for the participating countries. The two latter wars are ongoing conflicts so the “post-war” period was defined by the initial invasion period. It is important to note that in most of these cases there was not sufficient homicide data to include the countries where the fighting mostly occurred (e.g., Iraq, Kosovo, Serbia, Afghanistan). This limitation precludes a robust test of the legitimation of violence as one would expect greater modeling effects, and therefore increasing homicide rates, in those countries more so than the other combatant nations.
Following Archer and Gartner’s selection of control countries for the 12 smaller wars that they analyzed, we selected control countries for the modern wars based on the closest noncombatant neighbors of combatant nations that had sufficient data for analysis (see Table 7). Data for the extended analysis of the influence of modern wars on homicide rates were obtained from the United Nations World Surveys on Crime Trends and Criminal Justice Systems. Data from 1986 to 1994 were obtained from ICPSR (#2513) (Burnham & Burnham, 1999), whereas data for 1995 to 2008 were obtained from the United Nations Office on Drugs and Crime website. The crime category of “intentional homicide” was used for all of the extended analyses and the dependent variable was the homicide rate per 100,000 population.
Analytic Method
To test Hypothesis 1, we first employed the same methods as Archer and Gartner so that we could replicate their main findings. They compared the percentage change in homicide rates for a 5-year prewar average with a 5-year postwar average. They categorized countries according to whether the homicide rates remained unchanged (within a 10% margin), decreased by more than 10%, or increased by more than 10%. They also conducted t tests to gauge the size of the effect of the wars on postwar homicide rates. We repeated this approach for both World Wars and the 12 smaller wars. Because the findings for the 12 smaller wars were less robust than for the World Wars, which was consistent with Archer and Gartner’s findings, and because replicating the results for the smaller wars was less clear than the two big wars, our additional tests for Hypotheses 2 and 3 were restricted to only the World Wars.
We also tested Hypothesis 1 using time series analysis for all combatant countries involved in either or both World Wars for which we had sufficient data (n = 19). We limited this analysis to the two major wars because of the longer, less ambiguous time series of homicide data for these countries. Each time series was tested for stationarity using the Dickey–Fuller unit-root test and for serial autocorrelation using the Durbin–Watson test. First-differenced homicide rates or residuals were used as the dependent variables, depending on the diagnostic test results for each country’s time series, and standard errors were corrected for first-order autocorrelation, when necessary. Independent variables were indicator variables to mark postwar periods for World War I (1918) and World War II (1945). Interrupted time series models were then run for each country with either or both of the World War indicators included in the model, depending not only on the country’s involvement in the wars, but also on the availability of sufficient time series data.
We tested Hypotheses 1 to 3 using a pooled time series analysis of the World War samples. We combined the data across countries for World Wars I and II, respectively, to generalize the influence of the wars on postwar homicide rates across the samples. We employed panel-corrected standard errors and panel-specific first-order autoregressive corrections to adjust for heteroskedasticity and serial autocorrelation (Beck, 2001; Beck & Katz, 1995). Independent variables included indicators for war years, postwar years, defeated versus victorious nations, and countries experiencing less than or more than 100,000 battle deaths. 3 Data for the number of battle deaths were obtained from the Correlates of War data set and World War almanacs.
Finally, for the extended analysis, we tested Hypothesis 1 for the four modern wars using percentage changes, t tests, time series, and pooled time series regressions, as described above. We could not test Hypotheses 2 and 3 because we did not have data for most of the belligerent countries for either homicide rates or battle deaths, and for some wars there was minimal variation in the number of battle deaths for the countries that were included in the analysis.
Another problem with the modern wars was that their historical proximity created overlapping post- and prewar periods for countries that participated in multiple wars, such as the United States. For the calculations of percentage change and t tests, we treated these wars as independent events. For the time series analysis, we could include multiple wars for the same country in the same analysis, but this did not affect our overall findings.
Replication Results
To replicate Archer and Gartner’s study, we first calculated 5-year averages of pre- and postwar homicide rates and calculated percentage changes to match the original results, which were presented in Table 2 above. We could rarely match the exact numbers of the original study, but we were able to replicate the general findings of their study. Tables 3 and 4 summarize how well we could replicate the original findings in terms of percentage changes in homicide rates 5 years before and after the wars and in terms of t tests of pre- and postwar homicide rate averages.
Summary of Replications of Percentage Change in Homicide Rates Before and After the Wars.
Note. WWI = World War I; WWII = World War II.
Summary of Replications of t Tests in Homicide Rates Before and After the Wars.
Note. WWI = World War I; WWII = World War II.
These tables show that we were able to successfully replicate between 80% and 90% of the original findings for the World Wars, although with varying degrees of accuracy. The results for the 12 other wars were harder to replicate than for the World Wars for a few reasons. First, the original analysis included three 19th-century wars and the data provided by the authors started at 1900. Second, the Vietnam War was ongoing when this data collection ended, so no postwar data were available. Instead prewar homicide rates were compared with wartime rates, although it was not always clear which years should be included in the comparisons for some countries. Finally, these smaller wars often overlapped with other wars, including the World Wars, introducing additional ambiguity in terms of which years should be included in the calculations. In all cases, where we were not sure which years to included, we tried several combinations until we found the ones that best matched Archer and Gartner’s results. With a considerable amount of trial and error, we were able to replicate approximately 75% of the results for the 12 other wars. Overall, the data in Tables 3 and 4 support Archer and Gartner’s results that combatant nations were more likely to experience postwar increases in homicide rates than control nations, thereby supporting Hypothesis 1.
Next we performed an interrupted time series regression for 19 combatant nations from WWI and WWII to see if the wars affected long-term homicide trends rather than just 5-year postwar averages. For each time series, we estimated the following model: Yt = B0 + B1t + B2X_t + B3TX_t, where t = time since the beginning of the series, X_t = a dummy variable for the intervention years (1918 for WWI and 1945 for WWII), and TX_t = an interaction term. As such B0 is the intercept, B1 is the slope of the regression until the intervention (end of the war), B2 is the change that occurs immediately after the intervention (short-term effect), and B3 is the difference between the prewar and postwar slopes of the outcome (long-term effect; Linden, 2015).
Table 5 summarizes the postintervention linear trend B1 + B3(1918) for WWI only, B1 + B3(1945) for WWII only, and B1 + B3(1918) + B3(1945) for countries that participated in both wars, and had sufficient time series data for this analysis. Of the 28 estimates shown in Table 5, 10 were statistically significant at p < .05 and two were significant at p < .10. Of those significant effects, eight were positive, indicating postwar increases, including two defeated nations (Italy and Germany). We consider these results as modest support for the legitimation of violence hypothesis. More importantly, the time series results encourage a more refined discussion of the length of the influence that wars may have on interpersonal violence, as many countries that exhibited postwar homicide increases in the 5 years immediately following the war did not necessarily experience long-term changes, particularly increases, in the post-war homicide trajectories.
Interrupted Time Series Regression Results for Combatant Nations in WWI and WWII.
Note. WWI = World War I; WWII = World War II.
p < .10. *p < .05. **p < .01. ***p < .001.
Finally, we employed pooled time series regressions on the World War samples using panel-corrected standard errors and a first-order autoregressive error correction to adjust for heteroskedasticity and autocorrelation. In Model 1, we included three indicator variables for war years, postwar years, and combatant versus control nations. For WWI, the war year and the postwar years had higher homicide rates than the prewar years and combatant nations had higher rates than control nations (see Table 6). For WWI, control nations had higher homicide rates than combatant nations and the year indicators were not statistically significant. In Model 2, we only analyzed combatant nations for each war and included indicators for victorious nations and countries that suffered more than 100,000 battle deaths. For WWI, all of the indicator variables were statistically significant in the expected direction, such that wartime, postwar years, victorious nations, and countries that experienced excessive battle deaths increased homicide rates. For WWII, only one variable was statistically significant at p < .05, such that countries that experienced a large number of battle deaths had higher homicide rates than those that suffered fewer deaths. Overall, these results modestly support Hypotheses 1 to 3 and replicate Archer and Gartner’s original findings, although they are much stronger for WWI than WWII.
Pooled Time Series Regression Results for WWI and WWII.
Note. WWI = World War I; WWII = World War II.
p < .10. *p < .05. **p < .01. ***p < .001.
To summarize our replication results, we were fairly successful in replicating Archer and Gartner’s findings related to Hypotheses 1 to 3 using their original methods of comparing pre- and postwar percentage changes and means using t tests. We often could not replicate their exact numbers, but we were able to replicate their overall findings. Our replication results were stronger for the two World Wars than for the 12 smaller wars due to greater ambiguity in the years covered for the smaller wars and overlapping wars.
When we reanalyzed the data using multiple regression techniques, we found some support for Archer and Gartner’s (1976, 1984) hypotheses, but also some null findings and inconsistencies. The time series results suggest that wars may have more short-term influence on postwar homicide rates than long-term effects. The pooled time series results showed inconsistencies between the World War samples, suggesting that other factors related to the socio-historical context of those wars and participating countries may have been overlooked.
To test how well Hypothesis 1 explains modern wars, we compared homicide rates for combatant and control nations for four wars from 1990 to 2003. We first used Archer and Gartner’s method to classify countries into those that experienced more than a 10% postwar increase in homicide rates, those that had a greater than 10% postwar decrease, and those that were largely unchanged (within 10% difference). We used t tests to compare 5-year-prewar homicide rate averages with 5-year-postwar averages. Table 7 presents the results of that analysis.
Percentage Change Classifications and t tests for Modern Wars.
Note: †p < .10. *p < .05. **p < .01. ***p < .001.
Very few countries exhibited increases in postwar homicide rates, particularly combatant countries. Across the four modern wars, 13 combatant nations had experienced a 10% or greater decrease in homicide rates, seven remained relatively unchanged, and three experienced a 10% or more increase. Among the control nations for all four wars, 12 had postwar decreases, 12 had increases, and three remained unchanged. Most of the countries experiencing increasing rates did so in the early 1990s just before the international crime drop began.
Next we ran interrupted time series regressions for 21 combatant nations participating in the four modern wars (results not shown but available upon request). The overwhelming majority of the time series results did not show any significant effects of the wars on subsequent homicide trends. The exceptions were significant postwar increases for Syria and Italy after the Gulf War, a significant increase for Australia after the Iraq Invasion, and significant decreases for the United Kingdom after both the Kosovo War and the Afghanistan invasion. In other words, of the 21 country/war units of analysis only three experienced postwar homicide increases.
Finally, just as we did for the two World Wars, we pooled the data across countries and year for each of the four modern wars and ran pooled time series regressions with panel-corrected standard errors and a first-order autoregressive error correction to adjust for heteroskedasticity and autocorrelation (results not shown but available upon request). For the first model, we included both combatant and control nations and indicator variables for war years, postwar years, and combatant versus control status. For all four wars, the combatant/control indicator was statistically significant, with combatant nations having more homicides than control nations, as expected. The time dummies were only significant for the Kosovo war, with the war years and postwar years showing fewer homicides than the prewar years. The second models were run on combatant nations only and included indicators for war years and postwar years. Unlike the analysis of the World Wars, we did not include dummy variables for victorious and defeated nations or the number of battle deaths because we were missing data for most countries where the wars occurred, plus one of the wars ended relatively recently and one was still ongoing, and there was a small number of countries in this sample. For these models, none of the time indicators was statistically significant. To summarize the results of our retesting of Hypothesis 1 on more contemporary data, none of the three models showed much evidence for postwar increases for these wars.
Discussion
Using the same data and methods as Archer and Gartner (1976, 1984) and adding some newer analytic techniques, we were able to replicate their findings showing that (a) combatant nations were more likely than control nations to experience postwar increases, (b) combatant nations suffering larger numbers of battle deaths were more likely to experience postwar homicide increases than those with fewer battle deaths, and (c) victorious nations were more likely to experience postwar homicide rate increases than defeated nations. We were most successful in replicating the original results using the same methods as Archer and Gartner (1976, 1984). When we applied multiple regression adjusting for potential confounders, the results became less consistent across wars (e.g., the pooled time series results) and over time (e.g., the time series results). Although it is hardly surprising that including control variables in the equation changes the results from bivariate analysis, these observations raise interesting questions about the goals of replication research.
First, what constitutes a “successful” replication? Using Archer and Gartner’s publicly available data and straightforward analytic methods (percentage change and t tests), we could rarely match their exact numbers, which we found surprising. We grouped our findings into categories of how close we came to matching the original numbers, but the cut-off points were rather arbitrary. Nonetheless, we were able to match the general patterns that Archer and Gartner (1976, 1984) found in their study.
We can only speculate as to why replicating the original results was more challenging than we expected. One of the groundbreaking features of this study was the fact that the data collection was intended to be shared with the scientific community and the data were published as an appendix in their 1984 book and made available electronically through ICPSR. This was not a common practice at that time. Much of the original data collection was physical rather than electronic, perhaps increasing the possibilities of losing information (e.g., years prior to 1900) or misclassifying it (e.g., punchcard errors). There appear to have been some calculation errors, as we noted that some of our pre-/postwar comparisons more closely matched the original findings when crime counts were used instead of crime rates. We did not find any obvious signs of bias and suspect that what Pridemore et al. (2018) refer to as “researcher degrees of freedom” played a large role in the difficulties replicating the results where “each decision can be made in good faith but the countless decisions made in the research process add up and can influence the outcome” (p. 28; see also Simmons, Nelson, & Simonsohn, 2011). Data collection notes or data diaries were not part of the material archived at ICPSR so we could not further investigate discrepancies in replication results.
The time series and pooled time series results raise questions about the duration of postwar homicide increases and the social context under which postwar increases are more likely to occur. The time series analyses allowed us to examine long-term influence of wars on homicide rates, which appeared to be less potent than short-term (5-year) effects. Is the legitimation of violence associated with wars a function of collective memory that dissipates over time? Is it more related to material changes in societies (e.g., greater presence of weapons or returning veterans) than to broader shifts in cultural attitudes toward violence?
Similarly, the pooled time series results were much stronger for WWI than WWII, raising questions about what social and cultural factors were different about those two wars that would affect the salience of the legitimation of violence model. Perhaps because WWI was the first “Big War,” the shock of the violence was more potent than for WWII. Or perhaps the difference was related to the more personal nature of combat in the first war versus the second. Finally, there could have been some desensitization to violence given the closeness of the two wars, which again begs the question of how violence becomes legitimized and under what conditions. This hypothesis may also explain why the results for the 12 smaller wars did not support the legitimation of violence model as strongly as those for the World Wars. We would need to include more variables in the multiple regression models to try to answer these questions, which is difficult to do for multiple countries before 1960.
Testing Archer and Gartner’s hypothesis with respect to modern wars additional questions, as we did not find support for the legitimation of violence model. We advance two potential explanations for this null finding. First, there were limitations with the sample selection and availability of data. The lack of homicide data for the countries where fighting occurred, and therefore the places where people witnessed more violence, limits the robustness of our findings. In Archer and Gartner’s original study, nations such as the United States and New Zealand that were far removed from the war locations still experienced postwar homicide increases, which was not replicated with the modern wars, but the overall null findings might still be challenged by a more complete sample. The historical proximity of the modern wars and the lack of resolution in one case might also mean that influence on postwar homicide rates might occur beyond the end of the data collection period for this study.
Second, there could be substantive explanations for the null findings for modern wars. For example, we already knew that many of the countries involved in large, interstate wars are highly developed democracies that had been experiencing a crime drop since the 1990s, despite their war activities. Literature explaining the international crime drop points to factors such as changing demographics, globalization, modernization, securitization, and democratization, which seem to exert stronger influences on interpersonal violence than war participation (Aebi & Linde, 2010; Karstedt, 2015; LaFree, Curtis, & McDowall, 2015; Tseloni, Mailey, Farrell, & Tilley, 2010). In other words, there could be countervailing forces that would require a more sophisticated analysis to unravel.
It is also possible that the nature of our sample of modern wars is different than early 20th century conflicts in several respects that might diminish the possibilities for modeling violence. For example, these wars have been limited in terms of the length of engagement and size of personnel, particularly compared to the two World Wars. Battle deaths were less common than earlier wars and there was a greater reliance on technology (e.g., drones) rather than personal combat. Finally, battle deaths typically occurred within the geographic space of one country (e.g., Iraq), thereby limiting civilians’ direct exposure to war deaths for most combatant nations and most of the war-torn nations were not included in this analysis due to a lack of homicide data. Some of these characteristics are similar to the 12 smaller wars between 1884 and 1975 that Archer and Gartner (1976, 1984) analyzed, which may help explain why those results were not as strong as those for the two World Wars.
Conclusions
Based on our replication and extension of Archer and Gartner’s classic study of the influence of wars on postwar homicide rates, we conclude that their original findings hold for the cases they analyzed, but our inability to apply their argument to modern wars raises questions about the conditions under which the legitimation of violence model holds. For example, technological advances with respect to widespread media and the Internet have introduced more continuous, personal, and pervasive mechanisms for modeling violence that downplay the examples of state violence that wars bring. For example, some research has shown that videogame and movie violence increases proviolence attitudes, lowers empathy, and increases levels of aggression (Anderson & Bushman, 2001; Funk, Baldacci, Pasold, & Baumgardner, 2004) although there is some disagreement about the size of the effect of media on violence relative to other predictors (Ferguson & Kilburn, 2009).
In addition, more recent social learning theories emphasize direct and interactional learning experiences in shaping delinquent behavior (Akers, 2013). What role does macrolevel modeling (e.g., state legitimation of violence) play relative to microlevel interactions? Under what conditions are macrolevel forces more salient than microlevel factors? To what extent can direct interactions buffer cultural attitudes favoring violence? Answering these questions has important implications beyond studying the influence of wars on postwar homicide rates, particularly in contexts of extreme violence, such as ethnic cleansing. Once again, answering these questions is beyond the scope of this article, but they are important questions for the development of criminological theory that were generated because of this replication project.
Although we argue that the legitimation of violence model tested by Archer and Gartner (1976, 1984) might be too simple to explain contemporary differences in cross-national variations in violence, we nonetheless maintain the value of this model for challenging current cross-national homicide research to move beyond structural explanations that have only been modestly successful at explaining cross-national homicide rates. Archer and Gartner’s (1976, 1984) formulation of the legitimation of violence model assumed that wars were the only, or the most salient, way in which states legitimize violence, which could thereby increase interpersonal violence. However, if we consider the case of the United States, which has historically had higher violent crime rates than other Western, highly developed countries, wars are only one way in which the government perpetrates and perhaps encourages violence. Corporal punishment, the death penalty, permissive gun laws, and controversial human rights practices (e.g., mass incarceration, Guantanamo Bay, extralegal police shootings) could also contribute to legitimation of violence that could manifest in relatively high homicide rates. What is the cumulative effect of different modes of modeling violence? Under what conditions do some modes become more salient than others?
Although it is essential to replicate criminological studies for the sake of scientific integrity, we contend that an underappreciated benefit of replication studies is how they promote theoretical generativity. Future research could expand upon this analysis by considering other large-scale conflicts beyond interstate wars, such as civil wars, on rates of interpersonal violence, by considering how multiple modes of state violence might legitimate interpersonal violence, or by expanding upon factors that may mediate the relationship between macrolevel and microlevel violence (e.g., demographic or technological changes).
We maintain that this exercise in replication has been fruitful in a number of ways. First, we validated the results of a highly influential study (Archer & Gartner, 1976, 1984). Second, we demonstrated how using different analytic methods to test the original hypotheses provided some additional support for the original study, but also raised new questions. Third, the failure to replicate the original hypotheses with different data generated from a different socio-historical period did not necessarily invalidate the theoretical model, but provided interesting insights into how it could be expanded. We conclude that replication studies are valuable to criminological research because of their contributions to theoretical generativity as well as empirical validation.
Footnotes
Declaration of Conflicting Interests
The author(s) declared no potential conflicts of interest with respect to the research, authorship, and/or publication of this article.
Funding
The author(s) received no financial support for the research, authorship, and/or publication of this article.
