Command by a Jedi General and the Fortunes of Clone Battalions: An Adjusted Analysis of 196 Grand Army Engagements, 22–21 BBY

1. Introduction

During the Clone Wars, which began in 22 BBY and, at the time of writing, has not ended, Jedi Knights and Masters served as generals leading the clone troopers of the Grand Army of the Republic. The arrangement was improvised. The army had been raised in secret on Kamino and was first committed at Geonosis in 22 BBY, and the Jedi who took the field alongside it were guardians and negotiators by training, not career officers. Campaigns were directed from above, by the Jedi Council and by the Senate and the Chancellor, and the generals executed them.

Whether it mattered that a Jedi general was present with a given battalion is a question that any later assessment of the army must eventually face. The vantage of this paper is 20 BBY, with the war unfinished, and we make no claim about how the conflict ended. Staff writing has usually treated Jedi command as an asset in itself, and popular accounts have gone further; a smaller literature holds that the generals were poorly suited to the management of large formations (Ostrane, 22 BBY). Neither position has been tested against engagement-level returns, largely because the returns were never designed to answer the question.

The obstacle is selection. Staff returns indicate that general officers were not distributed across theatres at random, and that the Council's practice was to send Jedi where resistance was expected to be heaviest (Dessik, 21 BBY). If so, engagements with a Jedi general will look worse on any outcome that hard theatres degrade, and the comparison will condemn commanders for the ground they were given. This paper asks two questions. Did battalions with a Jedi general present complete their objectives more often, and did they lose fewer troopers for the strength committed, once the difficulty of the theatre is held constant? We also ask how far any answer depends on how difficulty is measured, because the measurement is ours and imperfect.

2. Methods

Sources. Engagement returns were drawn from the Grand Army of the Republic Operations Records (Grand Army of the Republic Operations Directorate, 21 BBY), which include a command annex naming the general officer, where one was assigned, and a closing report stating whether the ordered objective was taken or held. Losses came from the casualty register of the Republic Medical Corps (Republic Medical Corps, 21 BBY), which lists clone troopers killed or missing by unit. Committed strength came from the strength return filed before each action. Where the command annex and the published operational histories disagreed about the presence of a Jedi general, we followed the annex; the reasons are discussed under limitations (Wenlock, 21 BBY).

Sample. We screened 227 engagements of battalion-level or larger forces, as designated in the returns, recorded between 22 and 21 BBY, after setting aside the founding operation at Geonosis, which was a first use of the army on a scale and under a command arrangement that later engagements do not share. Thirty-one of the 227 were excluded, 19 for lack of a usable strength return and 12 because the command annex could not be reconciled with the casualty register. The analytic sample was 196 engagements, of which 118 (60.2%) had a Jedi general present and 78 (39.8%) did not. Committed strength ranged from 260 to 3,610 troopers, with a median of 1,070 and a total of 238,970.

Exposure and outcomes. The exposure was the presence of a Jedi general with the committed force at the time of the action, coded from the command annex. Two outcomes were analysed. Objective completion was binary, coded from the closing report, and was modelled by logistic regression. Clone casualties were counts of troopers killed or missing, modelled by Poisson regression with the natural logarithm of troopers committed as an offset, so that coefficients describe the casualty rate per trooper committed and not the raw toll, which grows with the size of the force (Vandrel, 24 BBY). Pearson dispersion was about 17, far above the value of 1 the Poisson model assumes, so all reported intervals for the casualty models use a robust (sandwich) variance. Point estimates are unaffected by this choice; the intervals are wider.

Theatre Difficulty Index. To address selection we constructed a Theatre Difficulty Index, which is our own instrument and has no standing in the operations records. Each engagement was scored 0 to 3 on four items: enemy strength relative to the committed force, distance to supply and relief, terrain and atmosphere class, and warning time before contact. The item scores were summed to a range of 0 to 12. Both authors scored the records independently, blind to outcome and to the casualty register, for a random subset of 40 engagements; the intraclass correlation of the totals (two-way random effects, single rater) was 0.93 (95% bootstrap CI 0.86–0.96), and totals agreed within one point in 36 of 40. Scores were then assigned by consensus. A pilot of the coding scheme on a smaller set is reported elsewhere (Halvane, 21 BBY). The index has a mean of 6.1 (SD 2.6) in the sample.

Analysis. We first compared outcomes without adjustment. We then adjusted in three ways. The first stratified engagements into low (0–3), moderate (4–7) and high (8–12) difficulty, with a Mantel–Haenszel pooled odds ratio for completion. The second entered the index as a continuous covariate in both regressions, and the third added a squared term to allow curvature. Because adjustment is only as good as the covariate, we ran a sensitivity analysis in which normally distributed error with a standard deviation of 0.5, 1.0 or 2.0 index points was added to each engagement's score, the models were refitted 500 times at each level, and the mean log coefficient was recorded and exponentiated, so that the tabulated figures are geometric means of the 500 ratios (Mantic, 26 BBY). This mimics an index that captures the true difficulty of a theatre only in part. Estimates are reported as odds ratios or rate ratios with 95% confidence intervals; p values are two-sided and Wald-based.

3. Results

Jedi generals were concentrated in the hard theatres (Table 1). Only 4 of the 28 low-difficulty engagements (14.3%) had a Jedi general, against 58 of 108 moderate-difficulty engagements (53.7%) and 56 of 60 high-difficulty engagements (93.3%). Mean difficulty was 7.2 with a Jedi general present and 4.5 without, and the point-biserial correlation between index and presence was 0.53. The strata overlap poorly at the extremes: the low stratum holds only 4 engagements with a Jedi general and the high stratum only 4 without.

Unadjusted outcomes favoured the engagements without a Jedi general. Objective completion was recorded in 55 of 118 engagements with a Jedi general (46.6%) and 48 of 78 without (61.5%), for an odds ratio of 0.55 (95% CI 0.30–0.98; p = .041). Losses were 8,218 troopers from 139,330 committed with a Jedi general present, a rate of 59.0 per thousand, against 4,652 from 99,640 without, a rate of 46.7 per thousand. The unadjusted rate ratio was 1.26 (1.05–1.52; p = .015). Read alone, these figures would indict Jedi command on both counts.

Difficulty explained most of the gap. Each additional index point was associated with lower odds of completion (odds ratio 0.79, 95% CI 0.68–0.91; p = .001) and with a higher casualty rate (rate ratio 1.17, 1.11–1.24; p < .001), so a theatre scoring 9 carried roughly 3.5 times the casualty rate of one scoring 1. Within strata, completion under the two conditions was nearly identical in the moderate stratum (56.9% with a Jedi general, 56.0% without) and identical in the low stratum (75.0% in each, on 4 and 24 engagements). The apparent penalty was carried by the high stratum, where 19 of 56 engagements with a Jedi general completed (33.9%) against 2 of 4 without (50.0%), a comparison too small to bear weight.

Adjusted for the index as a continuous covariate, objective completion did not differ by presence of a Jedi general (odds ratio 1.01, 95% CI 0.51–2.03; p = .97). The stratified estimate agreed (Mantel–Haenszel odds ratio 0.96; regression on stratum indicators 0.96, 0.48–1.90; p = .90), and a squared term changed nothing (1.00, 0.50–2.01). The interval is wide. The interval does not exclude a doubling of the odds and comes close to admitting a halving, so the data are compatible with a substantial benefit or a substantial harm.

The casualty rate behaved differently. With the index as a continuous covariate, the rate in engagements with a Jedi general was 0.80 times that in engagements without one (95% CI 0.67–0.95; p = .012), a deficit of about 20%. With a squared term the ratio was 0.78 (0.66–0.93; p = .004), and after removing the eight most severe engagements (index above 10) it was 0.77 (0.64–0.91; p = .003). Under the coarse three-level stratification, however, the ratio was 0.84 (0.71–1.01; p = .057), and the interval reaches 1. The direction of the adjusted estimate is stable across specifications, but the conventional threshold is crossed in one of four.

The sensitivity analysis showed how much depended on the index (Table 2). With error of 0.5 points added, the mean casualty rate ratio was 0.82 and the mean odds ratio for completion 0.98. At 1.0 point the figures were 0.87 and 0.91, and at 2.0 points they were 0.99 and 0.76. As the index was made less reliable, the casualty advantage disappeared, and the completion estimate moved towards the unadjusted penalty. Both movements are the pattern expected when difficulty that the index fails to capture is left in the coefficient for Jedi presence, and because Jedi were concentrated in hard theatres, that leakage pushes the estimate towards apparent harm.

4. Discussion

The main finding is negative and it concerns the raw record. Anyone who tabulates Grand Army returns by the presence of a Jedi general will find worse completion and higher losses where the Jedi were, and will be wrong to read that as an effect of command. Assignment followed difficulty, and difficulty by itself accounts for the completion gap. Once the theatre is taken into account, completion is indistinguishable between the two conditions in these data. The corresponding estimate is imprecise, and we do not claim that Jedi command had no effect on completion, only that these returns show none.

The casualty result is suggestive and no more. Adjusted rates were lower where a Jedi general was present, by about a fifth in the main model, and the direction held under every specification we tried. A reader might take this as evidence that a Jedi general reduced losses per trooper committed. The sensitivity analysis argues for restraint. One of the two adjustment strategies puts the upper limit of the interval at 1.01, and the intervals in every model are wide. The sensitivity analysis points in a reassuring direction: an imperfect index leaves residual difficulty attached to Jedi presence, which would bias the estimate towards harm, so residual confounding of this kind would if anything understate a protective association. That does not license confidence, because confounding could also run the other way, for instance if Jedi were sent to theatres that were easier in ways the index misses. We regard the association as a hypothesis worth testing with better data and not as a finding about command.

Two mechanisms could produce a real casualty advantage, and the records cannot separate them. A general officer may make different tactical decisions, or the presence of a Jedi general may change the resources a battalion receives, since the Council's assignments may have brought priority in relief and supply that the strength returns do not capture. A third possibility is that Jedi were sent to theatres whose difficulty is higher in ways the index misses, which would bias the estimate towards showing greater harm and not less. We have no means of testing these accounts here, and we attribute no particular tactics to any general.

The results bear on how the Grand Army's command arrangement is judged. The army was raised as a single instrument and directed by the Jedi Council and the Senate (Galactic Senate Legislative Research Office, 22 BBY), and its clone troopers were trained for discipline in a structure that did not depend on any one general (Kaminoan Cloning Authority, 22 BBY). A comparison of engagements with and without a Jedi general therefore tests a marginal difference, since battalions without one still had clone officers in command. That the marginal difference is small or absent is unsurprising. It does not follow that the Jedi were unimportant to the army's use, only that their presence at the battalion level, weighed against the difficulty of the theatre, did not produce a large measurable difference in the outcomes recorded here.

5. Limitations

The central limitation is residual confounding. The Theatre Difficulty Index was built by the authors from the records, its four items are a judgement about what makes a theatre hard, and the sensitivity analysis shows the casualty result eroding as the index is degraded, though in the direction that would understate a protective association. Assignment probably depended on factors we cannot see, including the Council's information about the enemy, the availability of individual generals and considerations of the Senate's that never reached a strength return. The fact that Jedi generals were so concentrated in the high stratum means the data hold almost no non-Jedi comparison there, and estimates for the hardest theatres rest on 4 engagements.

The sample is a set of engagements the records allowed us to code. We excluded 31 of 227, some because the command annex could not be reconciled with the casualty register, and those cases may be systematically messier or more chaotic than the ones retained. Among engagements retained, the published operational histories and the command annexes disagreed about the presence of a Jedi general in 9, and the two sources cannot both be right. We followed the annex on the ground that it was written closest to the event, but a Jedi general who joined or left an action partway through would appear in one record and not the other. Presence is also a single yes or no, which ignores how long a general was with the force and whether he or she commanded it directly.

Engagements are also not independent. The same generals and battalions appear in several returns, and losses in one action may affect the strength and condition of a battalion in the next. We did not cluster the variance on general or battalion, because the returns do not identify either reliably, and the confidence intervals are therefore probably too narrow. Finally, the outcomes are thin. Objective completion depends on the closing report's account of what was ordered, and an objective set more modestly was easier to complete. Casualties count troopers killed or missing, and do not distinguish the wounded, the cause of loss, or the point in an action at which it occurred. Vantage is 20 BBY, and the war continues beyond the point at which our records end; engagements from later months may show a different relation between command and outcome, and we make no prediction about them.