Measuring Military Effectiveness: Calculating Casualty Loss-Exchange Ratios for Multilateral Wars, 1816–1990

ABSTRACT

In this article, we introduce and utilize a new dataset that provides battle- and war- level Loss Exchange Ratios (LERs) for combatant states involved in multilateral wars between 1816 and 1990.The battle-level data provide an alternative to the widely used, but problematic, HERO/CDB-90 data set on battle outcomes. To demonstrate the utility of the new data, we weigh in on the debate over democratic military effectiveness arguments by replicating models by Reiter and Stam (2002, 2009) and Downes (2009), finding that, when effectiveness is measured using LERs, democracies do not have an edge over their non-democratic counterparts.

KEYWORDS:

• democracy and war
• democratic effectiveness
• military effectiveness
• war casualties

Notes

¹ See Biddle (2004:152–153) for a description of the problems with the original CDB-90 data.

² The Peace Research Institute at Oslo (PRIO) has also collected data on casualties. However, their temporal frame only covers 1946–2008 and they do not disaggregate casualties by the parties in the dispute.

³ See the LERD codebook for a detailed description of how the wars were disaggregated.

⁴ Of the 2,503 Battle Records, 1,024 come from Clodfelter and 158 come from Dupuy and Dupuy. The vast majority of other sources are in university press books. A small number of books by other presses were used when insufficient university press sources were available to inform data collection on a conflict. See the full project documentation for details.

⁵ “Killed in Action” or “Casualties,” which includes both killed in action and wounded in action.

⁶ The war-level aggregated data we use in this article is available from the authors or from the journal’s replication data archive. The full LERD project data on the battle level and source level, as well as the full coding document for each observation and the various aggregation procedures, are available from the LERD Project (Long, Cochran, and Wagstaff 2016).

⁷ For example, when computing Israel’s LER for the Six Day War, Israel’s LER for the battle of Rafah is weighted much more heavily than its LER for the battle of Ammunition Hill because the total casualties for Rafah were much higher (2,500 Israeli and Egyptian KIA) than Ammunition Hill (106 Israeli and Jordanian KIA).

⁸ Democracies are countries with a polity score above 17. Autocracies are countries with a polity score below 5. Anocracies have polity scores between 5 and 17.

⁹ They do not include the constitutive Politics term because they only have two categories of states and including it would lead to multicollinearity problems (Reiter and Stam 2002:40–41).

¹⁰ Details on these variables can be found in Reiter and Stam (2002:38–44).

¹¹ There are other important predictors of effectiveness that are not included in this model and there are alternative ways of measuring terrain and strategy. However, the goal of this article is to replicate Reiter and Stam’s analysis using new data. Given this goal, we focus on the variables they include to test their hypotheses.

¹² In order to maintain consistency with Reiter and Stam’s models, the reported errors are not clustered on war. Because observations within wars are not independent, we ran robustness tests using clustered errors. This did not change our substantive results.

¹³ A Shapiro Wilks test shows that we cannot reject the null hypothesis that the residuals are normally distributed (p = .297), and visual tests of the residuals demonstrate that the data closely mirror a normal distribution. We use White’s test for generalized heteoskedasticty and find that the error terms are not homogenous (p = .004). This confirms our decision to use robust standard errors (and in later models to cluster those errors on war). We also ran robustness tests to evaluate how sensitive the results were to outliers, identifying three likely outliers (Germany and Yugoslavia in the Yugoslav theatre of WWII and Jordan during the Palestinian War) and three potential outliers (Greece during WWI, Israel during the War of Attrition, and Russia during the Sino-Soviet War). Exclusion of these cases individually or together did not affect the results.

¹⁴ He also rescales the polity variable to 1–21 rather than −10 to 10.

¹⁵ As with Reiter and Stam’s model, the Shapiro Wilks test shows we cannot reject the null hypothesis (p = .166). Visual tests confirm that the distribution closely approximates a normal distribution. This model assumes heteroskadistic errors clustered on war. An analysis of the homoskedasticity assumption using an identical model with non-clustered errors supports this decision (p = .094 for White’s test). The results from this model are insensitive to the exclusion of outliers.

¹⁶ See Reiter and Stam (2002:41).

¹⁷ The Shapiro Wilks test shows that we cannot reject the null hypothesis of normally distributed errors (p = .52) though there is evidence of heteroskedasticity (p value for White’s test is .008). That is to be expected, given the interdependence between observations from the same war. The use of robust clustered errors alleviates this problem. As before, these results are insensitive to the exclusion of outliers.

¹⁸ We disaggregate the Western Front into two wars between the Western Powers and Germany and the Western Powers and Turkey. The allied participants differed, with Russia fighting with the allies in the latter conflict but not the former. The conflicts took place in two entirely separate theatres and separate peaces were signed at different times: the Treaty of Versailles between Germany and the Western Powers in June of 1918 and the Treaty of Sèvres between Turkey and the Western Powers in August of 1920.

¹⁹ This disaggregation generally follows Reiter and Stam’s (2002) example. We depart from their definitions by combining the Norway and Denmark conflicts and the France, Belgium, and Holland conflicts. The outcomes of these conflicts were inextricably linked. In fact, Germany’s operational planning considered the relevant conflicts all part of one battle: Operation Weserübung for Norway and Denmark and Fall Gelb for the invasion of France and the low countries.

²⁰ For example, in calculating the figures for the Battle of Doiran in WWI, Sondhaus (2011) figures only provide casualty figures for Britain and Greece, even though France was also involved in the battle on the allies’ side. Using the LER from his account would artificially inflate Britain and Greece’s LER since the Bulgarian’s killed by the French would be attributed to them. Similarly Bulgaria’s LER would be underestimated since it would not be given credit for the French forces they killed. In the case, Clodfelter (2008) and Hall (2010) provide casualty figures for all of the battle participants and these figures are used to calculate the LER.

²¹ For example, for the Battle of Cambrai in WWI, we have a record for the overall battle and a record for an engagement that happened during the first day of the battle. The larger battle is coded as MULTIPLE, the smaller engagement as SUBSUMED, and the engagement is excluded.

²² For example, we have an observation for Operation Junction City in the Vietnam War, which was fought between the US and an NVA/VC team. This Operation includes six engagements including the initial attack, Ap Gu, Prek Klok I, Prek Klok II, and Suoi Tre. Four of these engagements provide LERs for the US vs. VC and two provide an LER for the US vs. NVA. When we aggregated the data, we included the engagements because they provide the most detailed information on how the NVA and VC performed when fighting the US. The overall Operation Junction City observation is excluded to avoid double-counting.

²³ See footnote 18.

²⁴ See footnote 19.

²⁵ It should be noted that this LER is not created by taking the average LER across sources because averaging ratios can be problematic. Rather, this LER is created by taking the average casualties for side a (across sources) and dividing it by the average casualties for side b (across sources). The casualties are averaged first and then the LER is calculated from those averaged numbers.

²⁶ See footnote 4.

²⁷ See footnote 5.