COWARDS AND HEROES: GROUP LOYALTY IN THE AMERICAN CIVIL WAR by Dora L. Costa MIT and NBER costa@mit.edu Matthew E. Kahn The Fletcher School, Tufts University matt.kahn@tufts.edu July 13, 2001 Dora Costa gratefully acknowledges the support of NIH grant AG12658 and AG10120. Introduction 1. Social capital important for economic growth (a) Knack and Keefer (1997), La Porta et al. (1997), Guiso et al. (2000) 2. How is social capital produced? (a) demographics (b) community characteristics (c) Alesina and LaFerrara (2000), Putnam (2000), Glaeser et al. (2000), Luttmer (2001), Costa and Kahn (2001) 3. answer important for organizational design (a) how design organization to minimize shirk- ing problems? i. individual backgrounds, group-interaction effects, and sorting important (Ichino and Maggi 2000) (b) if social capital high then need fewer strong incentives 1 4. no previous empirical work on correlates cowardice and heroism 5. What we do (a) examine “quit rates”, shirking, and high effort among Civil War companies i. cowards – deserting, arrests, awol ii. heroes – promotion from private to of- ficer (b) what importance of demographics and of group characteristics? 2 (c) Advantages i. stakes are high, costly for others if you shirk in this organization ii. because actions are costly to you, get- ting good measure of commitment (cf. past literature) iii. easier for researcher to measure shirk- ing than in modern firm iv. easier for team members to observe shirking in army than in firm v. large number of companies (282) vi. companies small enough (100 men) so no Tiebout sorting within vii. no Tiebout sorting across companies (assigned to it) viii. companies built on local basis so have heterogeneity (wouldn’t if random as- signment) 3 ix. large number of company characteris- tics (ethnic, age, income, occupational heterogeneity), of own characteristics, of geographical characteristics (d) Do not estimate social interactions model i. only know leader characteristics if in- ternal promotion ii. cannot identify peer effects 4 Specification 1. competing risk hazard model (a) cannot estimate probit because of censor- ing (b) examine days until event (desert,arrested, awol, promoted to officer) i. if dead, POW, MIA, discharged, or changed company before desertion treat as cen- sored ii. if dead, POW, MIA, discharged, changed company, or deserted before arrest, awol, or promoted to officer treat as censored (c) may have company level unobserved het- erogeneity i. peer effects ii. punishments iii. commanders 5 (d) account for company-level heterogeneity by using variance correction models (Lin and Wei 1989; Lee, Wei, Amato 1992; Cai, Wei, and Wilcox 2000) i. alternative : shared frailty or random effects duration models – would need to explicitly have random effects fol- low specific parametric distribution (Heck- man and Singer 1986) ii. alternative: random effects in discrete- time survival data (Hedecker, Siddiqui, and Hu 2000; Han and Hausman 1990) – need few time periods to implement iii. clustering biases results against us (e) if account for individual heteregeneity (e.g. some individuals will never desert) as well then estimation results remain un- changed 6 2. 3 basic variable types (a) individual variables – year mustered in, occupation, country of birth, age, height, whether volunteer, total personal prop- erty wealth in 1860 household, whether illiterate (b) company variables – birthplace fragmen- tation, occupation fragmentation, whether total personal household property Gini coefficient higher than median for all com- panies, coefficient of variation for age, the fraction who died (c) geographic variables – percent in county of enlistment voting for Lincoln, popula- tion in city of enlistment 7 3. hazard, (t), is (t) = exp(x0 I + x0 C + x0 G) 0(t) I C G where I=individual variables, C=company variables, G=geographic variables, 0(t) = baseline hazard (exponential) (a) exponential distribution - slightly better fit than Weibull 4. report hazard ratios – 1 unit change in dependent variable gives increase/decrease in odds of event (a) e.g. in Table 3 Irish 1.4 times as likely to desert as native-born 8 Data (http://www.cpe.uchicago.edu) 1. 25,204 men in 282 companies 2. mainly enlisted men 3. sample drawn on company level (companies drawn on local level) 4. observe individuals if move out of com- pany, but don’t observe characteristics of company move into 5. merge geographic characteristics to data on basis location of enlistment (generally equiv- alent to place of residence) 6. penalties for desertion, awol (a) fines, loss pay (b) imprisonment, imprisonment with hard labor (c) more onerous duties in company (d) only one man executed for desertion 9 Results 1. Tabulations (a) Differences in shirking by i. year of muster, occupation, birth place, household wealth ii. birth place fragmentation, occupational fragmentation, Gini coefficient, coeffi- cient of variation for age, fraction in company dying iii. county ideology, population in city of enlistment (b) Disproportionate promotion rate in Wis- consin and Ohio 10 2. Desertion best measure shirking (a) arrests, promotions depend upon officer decisions (b) serious offense (c) largest number as outcome – 9% sample deserts 11 3. Predictors of Desertion (a) if mustered in first year of war less likely (b) farmers less likely (c) Irish, British more likely, Germans less likely (d) tall less likely (e) volunteers less likely (f) wealthy less likely (g) illiterate more likely (h) more likely if company diverse in occu- pation, wealth, age (i) weak evidence that birth place diversity also increases desertion (j) more likely if large fraction of company died (k) less likely if county pro-Lincoln (l) more likely if from large city 12 4. Predictors of Arrests (a) if mustered in first or last year of war less likely (b) if Irish, British more likely (c) weak evidence more likely if illiterate (d) if company diverse in occupation more likely (e) if company diverse in wealth less likely but significant only if control for occupa- tional fragmentation (f) if high company death rate less likely 13 5. Predictors of AWOL (a) less likely if volunteer (b) more likely if illiterate (c) less likely if from pro-Lincoln county 14 6. Predictors of Promotion (a) more likely if mustered in first or last year of war (b) more likely if professional/proprietor or artisan relative to farmer or laborer (c) more likely if taller (d) more likely if birth place diversity, but this is Wisconsin effect (e) less likely if occupational or wealth di- versity (f) more likely if large fraction company died but weak significance 15 Conclusion 1. social capital matters (a) for team production similarity in age most important (b) occupational fragmentation important as well for desertion and arrests 2. individual characteristics matter 3. ideology matters 4. have emphasized benefits of Tiebout sort- ing in team production, but there might also be costs – did individuals gain any benefits from being in the army with a diverse group of people? 16 Table 1: Variable Means for All Men, for Deserted, for Arrested, for AWOL, and for Promoted to Officer All Deserted Arrested AWOL Promoted Days from muster until 265.068 503.996 439.324 438.890 Dummy=if mustered in 1861 0.197 0.193 0.311z 0.354z 0.341z 1862 0.372 0.338z 0.374 0.361 0.399 1863 0.058 0.132 z 0.090 z 0.061 0.038 1864 0.255 0.231z 0.191z 0.190z 0.086z 1865 0.117 0.106z 0.033z 0.033z 0.136 Dummy=1 if occupation Farmer 0.558 0.375z 0.431z 0.552 0.571 Artisan 0.187 0.233z 0.188 0.180 0.182 Professional/proprietor 0.066 0.076z 0.073 0.070 0.121z Laborer 0.182 0.309z 0.303z 0.196 0.098z Unknown 0.008 0.006 0.005 0.002 0.028z Dummy=1 if born in US 0.800 0.670z 0.661z 0.798 0.871 Germany 0.063 0.068 0.064 0.052 0.048 Ireland 0.056 0.125z 0.148z 0.072 0.025z Great Britain 0.032 0.063z 0.063z 0.031 0.020 Other 0.049 0.074z 0.064 0.046 0.035 Age at enlistment 25.787 25.653 25.459 25.586 25.280 Height in inches at enlistment 67.723 67.421z 67.244z 67.873 68.669z Dummy=1 if volunteer 0.924 0.874z 0.925 0.920 1.000z Log(total household personal property) in 1860 1.836 0.912z 1.193z 1.706 2.091 Dummy=1 if missing property information 0.559 0.750 z 0.722z 0.561 0.535 Dummy=1 if illiterate 0.017 0.021 0.019 0.033z 0.005 Dummy=1 if missing literacy information 0.446 0.665 z 0.619z 0.459 0.439 Company-level measures Birth place fragmentation 0.550 0.602z 0.630z 0.549 0.583z Occupational fragmentation 0.528 0.594z 0.583z 0.534 0.476z Dummy=1 if total household personal Gini above all company median 0.399 0.485z 0.355 0.448z 0.227z Dummy=1 if missing Gini 0.207 0.272 z 0.330z 0.220 0.227 Coefficient of variation for age 0.284 0.286 0.274z 0.288z 0.280y Fraction in company dying 0.130 0.126z 0.106z 0.137y 0.144z Percent in county of enlistment voting for Lincoln 35.359 33.349z 30.110z 29.202z 34.267 Other candidate 32.868 38.864z 31.629z 34.218 30.632 Unknown 31.773 27.987z 38.261z 36.580 35.101 Log(population) city enlistment 8.275 8.795 z 8.468z 8.299 7.996z Number observations 25,204 2176 575 883 396 The symbols , y, and z indicate that the mean is significantly different from the mean for those not in the category. Arrests and AWOLs are those preceding desertion only. The logarithm of personal property wealth is set equal to zero for those for whom this information is missing. The standard deviations of log(total household personal property), birth place fragmentation, occupational fragmentation, the coefficient of variation for age, and log(population) are 2.787, 0.215, 0.182, 0.034, and 1.483, respectively. 17 Table 2: Percent Serving by State and Percent Deserted, Arrested, AWOL, Promoted to Officer, and Died in War by State % Serving % Deserted % Arrested % AWOL % Promoted % Died Connecticut 2.06 4.79 1.97 4.09 0 2.40 Maine 1.67 1.29 2.35 0.87 0.38 2.35 Massachusetts 2.11 1.68 1.88 1.65 0.94 2.63 New Hampshire 2.30 4.25 2.35 4.09 1.69 3.91 Vermont 1.22 0.00 0.00 0.00 0.00 0.53 Delaware 1.73 2.57 1.50 0.79 0.19 1.12 New Jersey 3.46 9.86 6.19 1.10 1.13 2.04 New York 11.49 11.92 19.61 17.4 8.08 12.53 Pennsylvania 8.34 5.45 5.44 4.02 4.32 6.68 Illinois 11.78 7.52 5.25 12.28 7.89 12.11 Indiana 5.21 4.64 3.00 4.96 9.02 5.79 Michigan 5.48 5.38 3.85 3.31 4.32 6.60 Ohio 17.16 13.56 10.41 17.4 15.60 16.19 Wisconsin 5.50 1.91 2.81 1.81 12.59 4.05 Iowa 5.34 1.6 4.13 4.09 15.79 8.47 Kansas 1.01 0.51 1.31 0.47 2.82 0.25 Minnesota 1.15 0.39 0.75 0.31 2.26 0.34 Missouri 3.59 3.54 3.00 3.54 7.89 5.09 Kentucky 3.55 7.95 5.25 10.16 1.50 4.22 Maryland 1.54 4.60 2.91 3.78 0.38 1.17 Washington, DC 0.46 0.74 1.78 1.26 0.00 0.03 West Virginia 1.32 0.39 0.94 1.57 0.19 0.64 New Mexico 0.35 1.01 1.88 0.00 0.00 0.03 California 2.15 4.48 11.44 1.02 3.01 0.81 25,204 observations. Arrests and AWOLs are those preceding desertion only. 18 Table 3: Desertion Competing Risk Hazard Model Hazard Std Hazard Std Hazard Std Hazard Std Ratio Err Ratio Err Ratio Err Ratio Err Dummy=1 if mustered in 1861 1862 1.459z 0.093 1.349 0.207 1.459z 0.220 1.467z 0.224 1863 3.266z 0.265 3.038z 0.646 3.266z 0.676 3.258z 0.706 1864 2.472z 0.179 2.350z 0.352 2.472z 0.362 2.484z 0.381 1865 4.488z 0.415 3.830z 0.717 4.488z 0.845 4.473z 0.856 Dummy=1 if occupation Farmer Artisan 1.370z 0.085 1.418z 0.112 1.370z 0.107 1.377z 0.108 Professional/proprietor 1.279 z 0.115 1.310z 0.136 1.279 0.131 1.285y 0.132 Laborer 1.425z 0.086 1.532z 0.149 1.425z 0.120 1.433z 0.120 Unknown 1.105 0.310 1.101 0.340 1.105 0.343 1.127 0.351 Dummy=1 if born in US Germany 0.797y 0.074 0.741y 0.100 0.797 0.109 0.804 0.108 Ireland 1.402z 0.101 1.407z 0.120 1.402z 0.122 1.414z 0.126 Great Britain 1.500z 0.141 1.465z 0.186 1.500z 0.202 1.496z 0.202 Other 1.160 0.101 1.151 0.152 1.160 0.150 1.159 0.149 Age at enlistment 0.993y 0.003 0.995 0.004 0.993 0.004 0.993 0.004 Height in inches at enlistment 1.015 y 0.008 1.012 0.009 1.015 0.009 1.014 0.009 Dummy=1 if volunteer 0.746z 0.055 0.739 0.119 0.746 0.118 0.741 0.119 Log(total household personal property), 1860 0.958 z 0.015 0.965y 0.017 0.958z 0.017 0.947z 0.017 Dummy=1 if missing property information 1.067 0.110 1.114 0.123 1.067 z 0.117 1.015 0.112 Dummy=1 if illiterate 1.816z 0.282 1.959z 0.299 1.816z 0.274 1.829z 0.283 Dummy=1 if missing literacy information 1.785 z 0.144 1.767z 0.156 1.785z 0.159 1.791z 0.160 Company-level measures Birth place fragmentation 1.352z 0.190 1.503 0.404 1.352 0.381 1.321 0.367 Occupational fragmentation 2.371z 0.429 2.579y 1.111 2.371 0.988 2.711y 1.640 Dummy=1 if total household personal Gini if above all company median 1.353z 0.080 1.530z 0.225 1.353y 0.197 Dummy=1 if missing Gini 1.262z 0.086 1.387 0.239 1.262 0.194 Coefficient of variation for age 68.394y 51.966 14.986 26.103 68.394 y 125.768 72.486y 136.613 Fraction in company dying 6.204 z 2.147 4.190 3.667 6.204y 4.835 6.143y 4.697 Percent in county of enlistment voting for Lincoln 0.992z 0.002 0.986 0.003 0.992z 0.003 0.991z 0.003 Other candidate Unknown 0.995z 0.001 0.993 0.001 0.995z 0.002 0.994z 0.002 Log(population) city enlistment 1.057 z 0.016 1.053 0.035 1.057 0.032 1.059 0.033 Region Fixed Effects Y N Y Y Clustered on Company N Y Y Y 2 (33), 2 (28), 2 (33), 2 (31) for Log likelihood ratio 1803.74 568.71 660.82 636.17 Days until desertion are measured from first mustering in. The symbols ,y, and z indicate that the coefficient is significantly different from 1 at the 10, 5, and 1 percent level, respectively. The log-likelihood ratio test is for equality of all coefficients to 1. Men who died, became POWs, were discharged, were missing in action, or changed companies before first desertion are treated as censored. Region fixed effects are for Middle Atlantic, East North Central, West North Central, Border, and West (New England is the omitted category). 19 Table 4: Arrest Competing Risk Hazard Model Hazard Std Hazard Std Hazard Std Ratio Err Ratio Err Ratio Err Dummy=1 if mustered in 1861 1862 1.377z 0.147 1.204 0.235 1.377y 0.215 1863 1.482y 0.247 1.350 0.302 1.482y 0.280 1864 1.511z 0.201 1.394 0.292 1.511y 0.260 1865 0.724 0.193 0.514 0.169 0.724 0.234 Dummy=1 if occupation Farmer Artisan 0.960 0.116 1.012 0.147 0.960 0.118 Professional/proprietor 1.199 0.197 1.263 0.235 1.199 0.205 Laborer 1.058 0.119 1.294 0.200 1.058 0.140 Unknown 0.763 0.445 0.741 0.445 0.763 0.459 Dummy=1 if born in US Germany 0.863 0.157 0.761 0.145 0.863 0.159 Ireland 2.174z 0.276 2.158z 0.316 2.174z 0.277 Great Britain 1.616z 0.293 1.524y 0.273 1.616z 0.282 Other 1.208 0.205 1.190 0.194 1.208 0.176 Age at enlistment 0.988y 0.006 0.994 0.007 0.988 0.007 Height in inches at enlistment 0.985 0.011 0.981 0.012 0.985 0.014 Dummy=1 if volunteer 0.802 0.142 0.844 0.181 0.802 0.165 Log(total household personal property) in 1860 1.019 0.029 1.026 0.034 1.019 0.033 Dummy=1 if missing property information 1.467 y 0.284 1.506 0.360 1.467 0.342 Dummy=1 if illiterate 1.454 0.459 1.544 0.401 1.454 0.399 Dummy=1 if missing literacy information 1.228 0.179 1.251 0.195 1.228 0.187 Company-level measures Birth place fragmentation 1.104 0.288 2.287 y 0.874 1.104 0.408 Occupational fragmentation 2.519z 0.857 2.684 1.573 2.519 1.177 Dummy=1 if total household personal Gini if above all company median 0.732z 0.082 0.886 0.160 0.732 0.122 Dummy=1 if missing gini 0.905 0.111 1.113 0.226 0.905 0.170 Coefficient of variation for age 1.347 1.881 0.011 y 0.024 1.347 3.771 Fraction in company dying 0.071 z 0.049 0.019z 0.022 0.071z 0.073 Percent in county of enlistment voting for Lincoln 0.997 0.003 0.994 0.004 0.997 0.004 Other candidate Unknown 1.000 0.002 0.998 0.002 1.000 0.002 Log(population) city enlistment 1.005 0.031 0.985 0.049 1.005 0.041 Region Fixed Effects Y N Y Clustered on Company N Y Y 2 (33), 2 (28), 2 (33) for Log likelihood ratio 460.60 219.73 377.75 Days until arrest are measured from first mustering in. The symbols ,y, and z indicate that the coefficient is significantly different from 1 at the 10, 5, and 1 percent level, respectively. The log-likelihood ratio test is for equality of all coefficients to 1. Men who died, became POWs, were discharged, were missing in action, changed companies, or deserted before first arrest are treated as censored. Region fixed effects are for Middle Atlantic, East North Central, West North Central, Border, and West (New England is the omitted category). 20 Table 5: AWOL Competing Risk Hazard Model Hazard Std Hazard Std Hazard Std Ratio Err Ratio Err Ratio Err Dummy=1 if mustered in 1861 1862 0.720z 0.059 0.703z 0.096 0.720y 0.101 1863 0.823 0.123 0.820 0.168 0.823 0.170 1864 1.043 0.108 0.991 0.177 1.043 0.181 1865 0.592z 0.118 0.556 0.178 0.592 0.193 Dummy=1 if occupation Farmer Artisan 0.918 0.089 0.937 0.098 0.918 0.096 Professional/proprietor 1.124 0.153 1.134 0.159 1.124 0.159 Laborer 1.036 0.101 1.017 0.127 1.036 0.128 Unknown 0.493 0.286 0.491 0.486 0.493 0.488 Dummy if born in US Germany 0.744 0.118 0.700 0.144 0.744 0.156 Ireland 1.124 0.152 1.134 0.172 1.124 0.174 Great Britain 1.095 0.200 1.078 0.210 1.095 0.218 Other 1.003 0.160 0.967 0.193 1.003 0.195 Age at enlistment 1.007 0.005 1.006 0.005 1.007 0.005 Height in inches at enlistment 1.014 0.012 1.013 0.010 1.014 0.011 Dummy=1 if volunteer 0.627z 0.085 0.563z 0.123 0.627y 0.130 Log(total household personal property) in 1860 0.974 0.018 0.983 0.019 0.974 0.019 Dummy=1 if missing property information 0.787 0.107 0.826 0.117 0.787 0.110 Dummy=1 if illiterate 1.648z 0.315 1.825y 0.569 1.648 0.509 Dummy=1 if missing literacy information 1.304 y 0.153 1.272y 0.159 1.304y 0.166 Company-level measures Birth place fragmentation 1.160 0.227 0.850 0.225 1.160 0.378 Occupational fragmentation 0.863 0.214 1.197 0.551 0.863 0.443 Dummy=1 if total household personal Gini if above all company median 1.098 0.092 1.190 0.193 1.098 0.169 Dummy=1 if missing gini 1.138 0.114 1.092 0.199 1.138 0.206 Coefficient of variation for age 12.042y 12.772 180.887z 335.483 12.042 24.611 Fraction in company dying 0.982 0.496 1.090 0.986 0.982 0.907 Percent in county of enlistment voting for Lincoln 0.992z 0.002 0.986z 0.004 0.992z 0.003 Other candidate Unknown 0.996z 0.001 0.994z 0.002 0.996y 0.002 Log(population) city enlistment 1.016 0.026 1.036 0.037 1.016 0.038 Region Fixed Effects Y N Y Clustered on Company N Y Y 2 (33), 2 (28), 2 (33) for Log likelihood ratio 267.15 124.83 195.71 Days until awol are measured from first mustering in. The symbols ,y, and z indicate that the coefficient is significantly different from 1 at the 10, 5, and 1 percent level, respectively. The log-likelihood ratio test is for equality of all coefficients to 1. Men who died, became POWs, were discharged, were missing in action, changed companies, or deserted before first awol are treated as censored. Region fixed effects are for Middle Atlantic, East North Central, West North Central, Border, and West (New England is the omitted category). 21 Table 6: Promotion to Officer Competing Risk Hazard Model Hazard Std Hazard Std Hazard Std Hazard Std Ratio Err Ratio Err Ratio Err Ratio Err Dummy=1 if mustered in 1861 1862 0.629z 0.081 0.643y 0.139 0.629y 0.119 0.595z 0.118 1863 0.629 0.175 0.685 0.295 0.629 0.245 0.610 0.230 1864 0.545z 0.111 0.596 0.236 0.545 0.215 0.490 0.182 1865 3.018z 0.560 3.458z 1.277 3.018z 1.016 2.322z 0.713 Dummy=1 if occupation Farmer Artisan 1.454z 0.207 1.429y 0.218 1.454z 0.213 1.450z 0.212 Professional/proprietor 2.670 z 0.443 2.698z 0.520 2.670z 0.508 2.580z 0.488 Laborer 0.936 0.180 0.830 0.170 0.936 0.185 0.937 0.181 Unknown 3.814z 1.313 3.951z 1.582 3.814z 1.511 3.586z 1.342 Dummy=1 if born in US Germany 0.723 0.179 0.816 0.203 0.723 0.181 0.544y 0.141 Ireland 0.533 0.176 0.522 0.225 0.533 0.238 0.508 0.227 Great Britain 0.607 0.220 0.636 0.211 0.607 0.202 0.538 0.181 Other 0.643 0.185 0.676 0.213 0.643 0.196 0.590 y 0.182 Age at enlistment 0.997 0.007 0.997 0.008 0.997 0.008 0.998 0.008 Height in inches at enlistment 1.105 z 0.022 1.110z 0.020 1.105z 0.021 1.105z 0.021 Log(total household personal property) in 1860 1.010 0.029 1.011 0.039 1.010 0.039 1.010 0.039 Dummy=1 if missing property information 0.842 0.187 0.830 0.213 0.842 0.213 0.879 0.221 Dummy=1 if illiterate 0.439 0.313 0.374 0.275 0.439 0.322 0.462 0.340 Dummy=1 if missing literacy information 1.201 0.220 1.257 0.258 1.201 0.247 1.164 0.235 Company-level measures Birth place fragmentation 2.369z 0.742 3.658z 1.896 2.369 1.287 0.828 0.395 Occupational fragmentation 0.925 0.373 0.245y 0.169 0.925 0.623 0.921 0.652 Dummy=1 if total household personal Gini if above all company median 0.711y 0.102 0.524y 0.156 0.711 0.203 0.696 0.180 Dummy=1 if missing gini 1.036 0.157 0.878 0.223 1.036 0.271 1.260 0.340 Coefficient of variation for age 0.408 0.601 0.135 0.413 0.408 1.208 0.253 0.743 Fraction in company dying 3.404 2.312 6.536 7.398 3.404 3.345 3.215 3.195 Percent in county of enlistment voting for Lincoln 1.000 0.000 1.000 0.000 1.000 0.000 1.000 0.000 Other candidate Unknown 1.058 0.269 1.122 0.412 1.058 0.387 0.800 0.272 Log(population) city enlistment 1.012 0.046 0.970 0.075 1.012 0.075 1.046 0.073 Region Fixed Effects Y N Y Y Wisconsin, Iowa Dummies N N N Y Clustered on Company N Y Y Y 2 (32), 2 (27), 2 (32), 2 (34) for Log likelihood ratio 348.24 257.72 288.80 279.27 Days until promotion are measured from first mustering in. The symbols ,y, and z indicate that the coefficient is significantly different from 1 at the 10, 5, and 1 percent level, respectively. The log-likelihood ratio test is for equality of all coefficients to 1. Men who died, became POWs, were discharged, were missing in action, changed companies, or deserted before first promotion to officer are treated as censored. Region fixed effects are for Middle Atlantic, East North Central, West North Central, Border, and West (New England is the omitted category). 22 Table 7: Predicted Probabilities of Desertion, Arrest, AWOL, and Promotion to Officer By Company Characteristics Desertion Arrest AWOL Promotion Using true variable values 0.078 0.023 0.035 0.015 If birthplace fragmentation=0 0.066 0.022 0.033 0.017 If occupational fragmentation=0 0.049 0.014 0.038 0.016 If Gini is below company average 0.065 0.027 0.033 0.016 If coeficient of variation for age=0 0.024 0.022 0.017 0.021 If all of above 0.011 0.014 0.016 0.026 Desertion, AWOL, and arrest probabilities are predicted from the third specifications in Tables 3, 4, and 5, respectively. Promotion to officer is predicted from the fourth specification in Table 6. 23