Research on ethnopolitical violence is rife with methodological and epistemological challenges. Yet the topic is clearly one of the most critical in our modern era. The search for the causes and consequences of violence compels us to overcome the methodological challenges. Best practice and new innovations are needed to find veridical answers regarding the antecedents and consequences of violence perpetration and exposure. Unfortunately, research in this area does not lend itself well to randomized trials and experimental manipulations. This lack of traditional methods for inferring causality, however, does not mean that we cannot approach valid causal interferences that provide utility in understanding the processes surrounding ethnopolitical violence. In a recent essay (Little, Reference Little2015), I emphasized three elements of modern research that should guide our basis for conducting work on critical research topics such as the ones presented here.
First, serious research surrounding topics such as ethnopolitical violence is a matter of social justice. The results need to be as valid as possible to ensure that policy and practice are optimally guided by the empirical evidence generated. Unfortunately, most traditional methods and statistical procedures are woefully ill equipped to address questions related to a complex open system like the nature and effects of ethnopolitical violence. Here, the challenge is to model the complex multivariate and multilevel processes surrounding violence and its consequences. In this regard, modeling quasiexperimental relationships with rigor and skill is the only means to approach valid inferences.
Second, another basis that should guide socially sensitive research is principled statistical justifications. Justification implies reasoned choices and certain skill/craft for conducting modern research. Here, the multilevel and processes-driven nature of work in this area (i.e., the socioecological metatheoretical perspective) demands the highest level of sophistication in measurement, design, and analysis. Principled choices and principled innovations should dominate the methodological applications that must be tailored to fit and adequately test complex hypotheses. The theoretical side in the proverbial tango with methodological sophistication has also matured to a level of complexity that can pose quite nuanced questions about levels of influence and the process of change and transformation. The works in this Special Section are all in this category of mature theoretical work coupled with sophisticated methodology, but I believe there is still room for improvement, which I will outline later in this commentary.
Third, the last basis is the pursuit of verisimilitude, or causality versus Causality. Verisimilitude is the truthlike value of research. We strive for parsimony, which, by definition, can only possess a certain degree of verisimilitude. The search for Causality often will undermine our ability to draw valid inferences about the processes and multilevel influences among the myriad constructs at play. Justified modeling decisions are the essential dialogue that is needed between model and data in order to achieve the highest levels of verisimilitude. The quiet revolution of statistical modeling (see Rodgers, Reference Rodgers2010) has evolved into a Bayesian-like enterprise. Here, the prior knowledge and understanding that powers Bayesian approaches is also leveraged even if the estimator is based on maximum likelihood. In my view, this emphasis on prior knowledge, principled justifications, and informed decisions begets research with a high degree of verisimilitude that will satisfy the demands of social justice in this area of research.
The papers that are assembled in this collection all possess these elements to varying degrees. In my commentary, I focus my discussion around the broader goals toward which these papers strive, namely, to move toward a process-oriented, social–ecological perspective using best practice methodology. That these research groups are already using sophisticated methodology should be lauded because they are pushing the whole field to greater levels of sophistication than what has been done in the past. In the remainder of this piece, I highlight directions and advances that ethnopolitical research is clearly heading. Some of my discussion will introduce some new ideas that are just gaining popularity. These ideas will become (or should become) the norm as researchers embark on new data collections.
Latent Variables
Latent variable modeling is an essential element of high-quality work in this area. Latent variables are presumed to exist based on theory, but their nature is inferred based on observables. For research on ethnopolitical violence, the observable data are fallible; they are contaminated with error variance that wreaks havoc on inferences about the processes that are under study. In addition to correcting for numerous sources of error (e.g., unreliability and sampling error), latent variable modeling provides a direct test of the factorial invariance (psychometric integrity) of the observed indicators of the constructs. For cross-national comparisons, when different languages are involved, the assumption of factorial invariance must be tested. Latent variable modeling provides the mechanism to test for factorial invariance.
Factorial invariance involves testing that the loadings and the intercepts of the indicators of corresponding constructs have the same proportional relationship across two or more groups and two or more time points. When factorial invariance holds, then any between-group or longitudinal differences can be attributed to differences in the underlying constructs and not to differences in the measurement process. Given the particular measurement challenges in this research arena (language and cultural differences, and often dramatic temporal changes), simply assuming invariance (which is a fundamental assumption of all traditional methods of analysis) would be particularly ill advised. Both contextual impacts that occur when violent actions are perpetrated and the sociocontextual differences across the disparate populations make the question of whether factorial invariance holds a crucial assumption that must be tested. Full invariance does not need to hold in order to make veridical comparisons. Partial invariance is perfectly valid for obtaining comparable factor scores because the factor scores are defined by the same common loading and intercepts that can be constrained while allowing for essential differences in the constructs (Little, Reference Little2013).
Latent variables and the general structural equation modeling (SEM) procedures are also particularly important when testing process effects like mediation and moderation (see Hayes, Reference Hayes2013). Meditation, for example, is ideally tested in the context of panel data where the sources of measurement error are removed and the supposition of invariance holds (see further discussion of mediation below). Latent variables are also required for many advanced measurement approaches such as the two-method planned missing design (see below) and the various flavors of multitrait–multimethod decompositions that have emerged since the multitrait–multimethod logic was introduced (Campbell & Fiske, Reference Campbell and Fiske1959).
I often see researchers forgo a full latent variable approach. I encourage researchers to estimate all constructs as multiply indicated latent variables. In nearly all cases of the measures used in this Special Section, the constructs could have been estimated as multiply indicated latent variables using parceling procedures (Bandalos & Finney, Reference Bandalos, Finney, Marcoulides and Schumacker2001; Little, Rhemtulla, Gibson, & Schoemann, Reference Little, Rhemtulla, Gibson and Schoemann2013). Parceling is a procedure whereby items are selectively averaged to reduce the total number of items down to a preferred number of indicators (ideally, three indicators per construct; see Little, Reference Little2013). That is, the newly created parcels are then used as the new indicators of the latent construct. Using item parcels allows one to estimate the measurement error and obtain disattentuated estimates of the modeled relations. Given the many untested assumptions of manifest variable analyses, particularly in the social and behavioral sciences, latent variable modeling should be the rule rather than the exception.
Multivariate Modeling
Multivariate approaches to addressing research questions in this area are clearly needed given the complexity of the processes that are under scrutiny. Simple and sovereign bivariate associations are too simplistic to yield meaningful information. The set of unique predictive pathways that emerges from a broader multivariate approach gives nuance to the different ways that different persons can come to the same predicted outcome based on very different standings on the multiple predictors. A multivariate approach also allows tests of mediating mechanisms and moderating influences.
The process of mediation focuses on the causal mechanisms of change. This simple idea, however, is too often tested in suboptimal ways. The statistical evidence of mediation is derived from the magnitude of one or more indirect pathways of influence. Here, longitudinal data are needed to test mediation because the hypothesis is that a distal variable X causes changes in a mediating variable M that in turn causes changes in the outcome variable of interest (Y; see Cole & Maxwell, Reference Cole and Maxwell2003; Maxwell & Cole, Reference Maxwell and Cole2007). The effect of X to M (Path a) and the subsequent effect of M to Y (Path b) are estimated and used to calculate the indirect effect (the product of Paths a and b). This indirect effect is then evaluated for significance. In the latent variable realm, this indirect effect as well as more complicated indirect effects are easily estimated and because of the corrections for error, the magnitude and significance of these estimates is precisely evaluated. Similarly, moderation is readily evaluated in the SEM framework, and it can be precisely evaluated (for more details on testing mediation and moderation see Hayes, Reference Hayes2013; Jose, Reference Jose2013; McKinnon, Reference Mackinnon2008). Moderating processes fit very well with the social–ecological model. Here, the socioecological contexts are influencing processes that change the strength of any associations among the measured constructs. In this Special Section, Cummings et al. provide nice examples of testing both mediation and moderation in the context of the multilevel modeling of their longitudinal data.
Multilevel SEM
The research conducted in this Special Section and in this field is sensitive to the importance of both accounting for and examining the layered levels of influence that define social–ecological inquiry. The field of multilevel modeling has evolved tremendously from the current standard of manifest variable regression (e.g., hierarchical linear modeling; see Cummings et al., this issue). Recent developments in both commercially available software (e.g., mPlus; Muthén & Muthén, Reference Muthén and Muthén2012) as well as freely available software (e.g., xxM; Mehta, Reference Mehta2015) now allow for multilevel SEM with latent variables. Multilevel SEM is more powerful than the traditional regression approaches because it can incorporate all the advantages of latent variable modeling in general and then estimate the within- and between-cluster aspects as latent variables. Moreover, the software package that Mehta is developing allows for n levels of nested data structure. His work truly captures my desire to see innovation in this area. His approach to multilevel modeling is both innovative and elegant, and will soon begin to change how any multilevel data structure is treated.
There should also be more work focusing on innovation in measurement at the higher levels of influence. This direction of research would provide a much broader understanding of what factors are operating when we find significant Level 2 and higher influences. Too often the higher level effects are simple averages of lower order measurements. Focusing on new measurement strategies would allow for more nuanced understanding of the nature of the higher level influences.
Plan for Mixture Modeling
Unknown heterogeneity in the population is likely, particularly in diverse sociocultural contexts where research on ethnopolitical violence is studied. Although some groups are known or directly measured, others are not. Mixture modeling is a modeling feature that, when properly implemented, can reveal subgroups of the population that have unique features, patterns, and characteristics that cannot be captured or understood otherwise. Mixture modeling has been criticized by some, but it is not universally condemned. For example, critiques include that groups can emerge from skewed data, or groups are defined only on the basis of severity (for a recent overview, see Lanza & Cooper, Reference Lanza and Cooper2016). The key to cogent mixture modeling (and to any advanced analytic technique for that matter) is to ensure that the data collection design is properly specified so that the obtained data fit with the modeling tool that will be used.
For mixture modeling, where the estimated groups that emerge must be rigorously validated, the guiding acronym for replicablity, interpretability, and predictability can help: RIP (see Little, Reference Little2013). The data patterns that define the unknown groupings must be replicable. Here the goal is to obtain sufficient data to cross-validate or use known groups to replicate. Finding a highly similar set of groups in a Palestinian versus Israeli sample, for example, is a form of cross-validation. Selecting highly meaningful variables to use in the mixture analysis is critical to aid in the interpretability of the resulting groups. To do so requires some anticipation of what variables would reveal the expected groups. Then these variables need to measure well to yield continuous and normally distributed indicators upon which the mixtures can be built. Finally, some critical variables must be identified and measured that are not used in estimating the groups. These variables are used to either predict the individual differences in the groups or as outcome variables that help characterize the meaning of the groups.
Modern Missing Data Treatments for Unplanned Missingness
Another methodological feature for best practice is implementing a modern approach to missing data (i.e., Bayesian-based multiple imputation or full information maximum likelihood estimation; Lang & Little, Reference Lang and Little2016). Missing data has been a sorely maligned topic in the field, and yet there are many joys to missing data (see Enders, Reference Enders2010; Graham, Reference Graham2012; Lang & Little, Reference Lang and Little2016; Little, Jorgensen, Lang, & Moore, Reference Little, Jorgensen, Lang and Moore2014; van Buren, Reference van Buuren2012). Imputation and other modern treatments of missing data are not “cheating” or “making up” data or negative in any other form. Instead, they are based on unequivocal statistical theory and principled treatments (Lang & Little, Reference Lang and Little2016). Carefully planning for missing data by measuring important auxiliary variables to inform the assumptions of modern approaches is critical to recover any selective process, especially selective attrition. These papers included a modern treatment to facilitate both power and generalizability. The modern approaches supersede traditional approaches because, when properly implemented, they can correct for the bias that selective attrition and selective nonresponse introduce. They also have the ability to restore much of the power that is lost when data go missing. One important issue to emphasize here is the need for including the auxiliary variables that capture the reasons for the missing data in a given study.
Auxiliary variables are those variables that exist in a given data set and are predictive of the missing data. If the auxiliary variables are not included in the modern treatment (i.e., this admonition applies to multiple imputation as well as full information maximum likelihood), then the treatment will not correct for the bias that occurs from the selective influence. Here, the missing data will be treated, but it will only reflect the missing completely at random mechanism (MCAR). The missing at random (MAR) mechanism that modern approaches can address is not engaged because the auxiliary variables that reflect the MAR mechanism are not included in the missing data model. My students and I (Howard, Rhemtulla, & Little, Reference Howard, Rhemtulla and Little2015) recently described a method that can maximize the ability to capture, in a small set of auxiliary variables, both linear and nonlinear MAR influences in a given data set. This procedure has now been implemented in a software package called quark (Lang, Chestnut, & Little, Reference Lang, Chestnut and Little2015). Using quark or a similar software package approach provides a rich set of auxiliary variables that will allow maximal corrections for selective bias that modern missing data treatments are designed to do.
The importance of auxiliary variables cannot be emphasized enough. Too often researchers think they are correcting for the biases introduced by a MAR missing data mechanism because they use a full information maximum likelihood procedure. This procedure only satisfies the MAR assumptions if the variables that represent the MAR process are included in the analysis model. If the variables that represented the MAR process are not included in the analysis model, then the treatment of missing data is assuming the MCAR process, and the results will remain biased. Sharma et al. (Reference Sharma, Fine, Brennan and Betancourt2017 [this issue]) did multiple imputation, which will capture a MAR process if all variables on the data set are included in the imputation model. Given what we know about the ability of modern treatments for missing data to provide unbiased estimates in the presence of missing data, we all need to be more rigorous in identifying the potential MAR mechanisms to ensure that they are adequately corrected.
Planned Missing Designs
Another aspect of modern missing data treatments where studies in this area could be further improved is utilizing the power of planned missing designs (Garnier-Villarreal, Rhemtulla, & Little, Reference Garnier-Villarreal, Rhemtulla and Little2014; Graham, Taylor, & Cumsille, Reference Graham, Taylor, Cumsille, Collins and Sayer2001; Graham, Taylor, Olchowski, & Cumsille, Reference Graham, Taylor, Olchowski and Cumsille2006; Johnson, Roth, & Young, Reference Johnson, Roth and Youngin press; Jorgenson, Rhemtulla, Schoemann, McPherson, & Wu, Reference Johnson, Roth and Young2014; Little et al., Reference Little, Jorgensen, Lang and Moore2014; Little, Lang, Wu, & Rhemtulla, Reference Little, Lang, Wu, Rhemtulla and Cicchetti2016; Mistler & Enders, Reference Mistler, Enders, Laursen, Little and Card2012). Planned missing data designs come in many permutations: multiform, multimethod, and wave missing. Before I delve into the specifics of these three permutations, I will highlight a couple of distinct advantages that these designs have over the traditional complete case protocols that dominate today's research studies. Planned missing designs are effective and well grounded in statistical theory; thus, not using them is a less tenable position than aggressively utilizing them. Planned missing designs are predicated on two ideas. First, the data are by definition MCAR, and therefore the observed data are unbiased representations of the population. That is, the only thing that planned missing designs introduce is a reduction in power (when data are missing, power is reduced). Second, the power loss that occurs with planned missing designs is readily rectified by using a modern missing data treatment. Both multiple imputation and full information maximum likelihood will restore most of the power reduction that using planned missing elements introduces (Enders, Reference Enders2010).
These designs save on costs because less data need to be directly collected in order to provide the same net yield of variables by observations as a complete case protocol. The data frame of a planned missing design has the same number of columns (variables) and rows (observations) as a complete case counterpart. The only difference is the data frame has lots of planned missing elements in it that are readily recovered when a modern treatment is performed. Another advantage of planned missing designs is that they reduce validity damaging effects of fatigue, burden, and disinterest on the part of the respondents. In other words, a well-implemented planned missing data design can yield observed data that is more valid than would be derived from a complete case counterpart (Harel, Stratton, & Aseltine, 2012; Swain, Reference Swain2015). Harel et al. showed that a planned missing data design had less unplanned missing, and compared to a complete case longitudinal protocol, participants were three times less likely to attrite when given a planned missing protocol. Swain's study showed that students performed better when given an achievement battery using a planned missing format compared to a complete case protocol.
Multiform planned missing data protocols
Multiform designs are also referred to as split questionnaire designs (Raghunathan & Grizzle, Reference Raghunathan and Grizzle1995; Rhemtulla & Little, Reference Rhemtulla and Little2012; Smits & Vorst, Reference Smits and Vorst2007). All multiform designs contain a set of items that all participants receive. This set of items is referred to as the X block and contains demographic items, gateway items, and related unitary indicators/variables. The remaining items, which typically encompass the multiple indicators of the latent constructs, are then divided across different numbers of variable sets (e.g., A, B, C, D, E). The basic idea here is everyone receives the X block and then items from two of the different variable sets. A questionnaire form is created (e.g., X block + A items + B items) and randomly assigned to a participant. With three variables sets (A, B, and C), three forms are possible (XAB, XAC, and XBC). Each form in this rendering of a multiform design yields about one-third of the items missing. With the addition of a fourth variable set, D, then six forms must be created to ensure adequate crossing of items sets (XAB, XAC, XAD, XBC, XBD, and XCD). Each form in this rendering of a multiform design would yield close to 50% missing. Because this missing data is MCAR and because modern approaches can recover the power loss, these designs do not present any bias issues (Adigüzel, & Wedel, Reference Adigüzel and Wedel2008). The six form design is particularly useful for longitudinal studies because it is easy to assign forms such that retest exposure is eliminated or highly reduced. For instance, a pretest could give the XAB form and the posttest could be the XCD form. Depending on the nature of the X items (e.g., redundant demographic items), the X block may not need to be administered at the posttest.
In terms of general guidance with these designs, the minimum sample size appears to be about 40 persons per form (Jia et al., Reference Jia, Moore, Kinai, Crowe, Schoemann and Little2014). Assigning items from the same construct to different variable sets improves efficiency of the latent parameters versus assigning all items of a construct to the same variable set. The reactivity of seeing all the items of a construct is similar to the complete protocol issues I mentioned above. Item-level imputation for even massive data sets is now a relatively straightforward proposition, which makes the “hard to impute” arguments pointless. Here, the principle component auxiliary variable extraction coupled with isolated, serial imputation as implemented in an R package (Lang et al., Reference Lang, Chestnut and Little2015) can be used for massive data imputation problems.
Multimethod planned missing design
The multimethod planned missing data design is a second tool that research in this area can utilize to reduce costs, improve sample size (and power), and increase validity. The multimethod design is possible only in the latent variable world where multiple indicators and bifactor extractions of variance are possible. This design assumes that there are at least two methods that can be used to measure the same underlying construct. One method is assumed to be not only unbiased but also costly to acquire. For example, examining stress levels during violence exposure would ideally be captured using a cortisol assay method, but collecting, storing, and analyzing cortisol from a large population of exposed youth would be cost prohibitive. A cheaper method would be to administer a simple perceived stress measure, but such a measure is less accurate than a cortisol assay. The two-method protocol utilizes both measures. The less accurate and easy to implement tool is given to all participants. A random subsample of participants is chosen to give saliva (about one-third of the total sample).
Improving Measurement
Measurement can be the Achilles’ heel of high-quality research. The modern modeling approaches are not a panacea of poor measurement. The measurement practices in vogue are simply staid and, as a result, imprecise. The technical machinery of modern modeling approaches can quickly outstrip the carrying capacity of the data. A number of improvements in measurement practices would engender more valid variables to which these advanced and powerful analytic approaches can be applied. For example, reliance on Likert scaling methods is antiquated (Likert, Reference Likert1932). Visual analog scaling can be used to enhance the quality of the data by providing truly continuous scaling of variables (Carlsoon, Reference Carlsson1983; Couper, Tourangeau, Conrad, & Singer, Reference Couper, Tourangeau, Conrad and Singer2006; Flynn, van Schaik, & van Wersch, Reference Flynn, van Schaik and van Wersch2004; Joyce, Zutshi, Hrubes, & Mason, Reference Joyce, Zutshi, Hrubes and Mason1975; Little & McPhail, 1973; Rausch & Zehetleitner, Reference Rausch and Zehetleitner2014; Thomeé, Grimby, Wright, & Linacre, Reference Thomeé, Grimby, Wright and Linacre1995). Such approaches have been around for some time (Hayes & Patterson, Reference Hayes and Patterson1921), but the scoring of paper-and-pencil versions of them were time consuming and error prone. Now, paper-and-pencil versions of such tools are possible because computer scoring can be adapted to measure and record the response options. In addition, if electronic data collection via smart phones or simple tablets is available, the measurements can be directly recorded with touch screen technology.
Often research in the area of ethnopolitical violence is focused on change, but measuring change may be challenging (Cronback & Furby, Reference Cronbach and Furby1970). One method to assess change that is underutilized is the retrospective pre–post design championed by Howard and colleagues in the late 1970s and early 1980s (Bray, Maxwell, & Howard, Reference Bray, Maxwell and Howard1984; Howard, Reference Howard1980; Howard, Dailey, & Gulanick, Reference Howard, Dailey and Gulanick1979; Howard, Millham, Slaten, & O'Donnell, Reference Howard, Millham, Slaten and O'Donnell1981). Traditional pre–post designs, which measure all constructs before the intervention and then measure them all again sometime after the intervention, often are unable to detect change because of the response shift bias that can occur (Allen & Nimon, Reference Allen and Nimon2007; Davis, Reference Davis2003; Drennan & Hyde, Reference Drennan and Hyde2008; Hill & Betz, Reference Hill and Betz2005; Hoogstraten, Reference Hoogstraten1985; Nakonezney & Rodgers, Reference Nakonezney and Rodgers2003; Nakonezney, Rodgers, & Nussbaum, Reference Nakonezney, Rodgers and Nussbaum2003; Sibthorp, Paisley, Gookin, & Ward, Reference Sibthorp, Paisley, Gookin and Ward2007). The retrospective pre–post design has seen some resurgence in the health arena (Breetvelt & Van Dam, Reference Breetvelt and Van Dam1991; Finkelstein, Quaranto, & Schwartz, Reference Finkelstein, Quaranto and Schwartz2014; Galenkamp, Deeg, Braam, & Huisman, Reference Galenkamp, Deeg, Braam and Huisman2013; Keivit et al., Reference Kievit, Hendrikx, Stalmeier, van de Laar, Van Riel and Adang2010; King-Kallimanis, Oort, Visser, & Sprangers, Reference King-Kallimanis, Oort, Visser and Sprangers2009; Levinson, Gordon, & Skeff, Reference Levinson, Gordon and Skeff1990; McPhail, Comans, & Haines, Reference McPhail, Comans and Haines2010; Nagl & Farin, Reference Nagl and Farin2012; Schwartz, Sprangers, Carey, & Reed, Reference Schwartz, Sprangers, Carey and Reed2004). It has also been utilized in the context of educational and evaluation research (Allan & Nimon, Reference Allen and Nimon2007; Drennan & Hyde, Reference Drennan and Hyde2008; Hill & Betz, Reference Hill and Betz2005; Moore & Tananis, Reference Moore and Tananis2009; Pelfry & Pelfry, Reference Pelfrey and Pelfrey2009; Pratt, McGuigan, & Katzey, Reference Pratt, McGuigan and Katzey2000; Sibthorp et al., Reference Sibthorp, Paisley, Gookin and Ward2007). In particular when geopolitical events transpire and a true pretest is not possible, the retrospective pre–post design, coupled with visual analog scaling and a planned missing protocol, can be utilized to assess change.
Mindful Analytics
As mentioned earlier, principled justifications of statistical procedures is an essential element of furthering high-quality research. Modeling procedures (as opposed to significance testing) require a great deal of craft and thought in order to implement them well. Advanced training is a first step, but guided decision making using mindful analytics requires partnerships of savvy players. I have referred to this team science approach as Wesearch (as opposed to Mesearch; see Little, Reference Little2015). Research allows for greater prior knowledge to be employed in ways that recover the underlying structure of data that is maximally consistent with theory and knowledge and, through mindful data interrogation, reduces both Type I and Type II errors. Errors of inference should not just be a property of the a priori significance test, but in the modeling world, each indication of a model modification must be evaluated as to whether it is real and or whether it can be ignored. This decision must be determined on both statistical grounds and theoretical grounds. A well-qualified Wesearch team of methodologically savvy theoreticians and theoretically savvy methodologists would have the corroborating expertise to make these decisions well. By incorporating basic checks and balances, Wesearch teams can ensure the integrity of the research outcomes and avoid issues such as conflicts of interest, lack of replicability, and other practices that Mesearch traditions have found problematic.
With the quiet revolution of statistical modeling (Rodgers, Reference Rodgers2010), we are moving seamlessly into a more Bayesian world. Confirmatory analyses are a form of prior knowledge that Bayesian enthusiasts tout. From my perspective, we do have considerable prior knowledge to guide our modeling efforts. We just lack practice thinking in this way. Given that advances in software capacity make it now easy to implement Bayesian approaches, more training in Bayesian ways would be a fruitful addition to the overall toolbox of researchers in this field (see Kaplan, Reference Kaplan2014; Kruschke, Reference Kruschke2015).
Conclusions
A number of features were highlighted where improved practice and innovation in methodology and statistics can further propel the work on ethnopolitical violence to even greater heights of discovery, explanation, and prediction. Of course, a number of other areas could have been highlighted where modern methodology and statistical modeling machinery are capable of propelling research forward. In contrast, the authors of the works collected here also need to be applauded for the degree to which they have incorporated many of the modern advances in statistical methodology. They each employed an analysis tool that was tailored to their respective research questions, they embraced a multivariate approach to examine process of change and influence, and they employed a modern treatment of the ubiquitous missing data that occurs. This work therefore provides a number of important contributions to this critically important literature area. I look forward to seeing even greater improvements in methodological practice that will continue to advance the quality of the research findings surrounding the causes and consequences of ethnopolitical violence.
