I Was Only Nineteen, 45 Years Ago: What Can we Learn from Australia s Conscription Lotteries?

Transcription

1 I Was Only Nineteen, 45 Years Ago: What Can we Learn from Australia s Conscription Lotteries? Peter Siminski & Simon Ville, University of Wollongong Seminar presented at NATSEM, University of Canberra 19 July 2012 ARC Projects: LP ; DE Industry Partner: Department of Veterans Affairs

2 Outline Australia s conscription lottery Mechanisms and research motivations Methods and data Employment Effect Crime Effects (work in progress) 2

3 Australia s National Service Lotteries ( ) Two lotteries per year = 16 lotteries 20 y.o. men required to register for lottery. By date of birth National Service = Army for two years 804,286 registered, 237,048 balloted-in, 63,375 enlisted, 18,654 served in Vietnam. 3

4 Depiction of first stage prob of Army service 4

5 Depiction of first stage prob of Army service in Vietnam 5

6 Mechanisms by which ballot outcome may affect outcomes Draft avoidance behaviour: education, marriage, health Army Service in Australia: army training, removal from civilian life e.g. labour market & marriage market Service in Vietnam: stress, combat, chemical exposure, cultural exposure, hostile reception on return to Australia Vets compensation and programs: cash benefits (direct and indirect effects), health insurance, education 6

7 Research Motivations Full costs of conflict Reform of military practice Acknowledgment Appropriate health interventions for veterans Assess adequacy and design of compensation Long run effects of experiences in early adulthood. 7

8 Our contributions Aus conscription lotteries solve selection problem Compared to US, cleaner assignment and less concern over confounders (e.g. education). Between-cohort variation: operational vs nonoperational service Vets compensation system differs from US in important ways Several sources of quality data; many relevant outcome variables 8

9 Intuition of Methodology All men in a 6-month birth cohort Balloted in Compliers LATE = Reduced Form = } y y pˆ pˆ { I I O O Balloted out Compliers y = a given health outcome p = proportion enlisted 9

10 Main Model y = α + β r + β v + γ ' C + µ i r i v i i i r = π ' zc + π ' C + ε i r1 i r 2 i ri v = π ' zc + π ' C + ε i v1 i v2 i vi Where y is employment (binary) r is Vietnam-era army service, v is army service in Vietnam C is a vector of 16 (6-month) cohort dummies z is a binary binary ballot outcome instruments, which is interacted with C to give 16 IVs 10

11 Approach and Data Two Sample 2SLS (Atsushi & Solon, 2010) Use cross-fitted values from the 1 st stage regressions in 2 nd stage regression Treating 1 st stage results as known First stage data Unit records from 2 military personnel databases Combined with published resident population of 20 year old men (at time of each ballot) N = 868,606 11

12 2 nd Stage Data Census 2006 (N = 675,832) Criminal Court data (NSW, QLD, VIC) ( ) (179,363 cases with guilty verdicts) Vets Disability Pension data ( ) ATO data ( ) (N ~ 1,000,000) AIHW National Mortality Database ( ) 45 & UP (sample survey) Australian Cancer Database ( ) ED NMDS 12

13 2SLS effects on economic outcomes 13

14 14

15 15

16 Violent Crime Why might military service affect violent crime perpetration? Combat exposure (Rohlfs, 2010, JHR): desensitisation Threat of combat exposure: stress / mental health Training, e.g. weapons + dehuminisation of the enemy + rapid respons Removal from civilian life / social ties / replaced with masculine culture Military service potentially decreases crime due to training in discipline, health, vocational skills Existing evidence mostly correlational, exceptions are Rohlfs (2010, JHR) finds effects of combat intensity on violence in U.S. Galliani et al. (2011, AEJ: Applied) find effects on crime, but not violent crime in Argentina Lindo et al. (2012) find effects on violent crime in US. 16

17 2SLS Approach - Crime Preferred Strategy to identify effects of training: limit sample to cohorts that remained in Australia. Identify r directly in 2 nd stage regression: y = β rˆ + γ ' C + µ d r cz d d Strategy 2: use all cohorts; assume service in Vietnam did not decrease crime, to get an upper bound estimate for r. Strategy 3: identify effect of r and v, exploiting crosscohort variation in treatment effects: y = rˆ + vˆ + ' C + β β γ µ d r cz v cz d d Short story: nothing significant and point estimates negative. Key question: are estimates precise enough? 17

18 Visual Reduced Form Violent Cases (guilty) 0.08 Ballot In Ballot Out

19 2SLS effect of army training on crime (youngest 4 cohorts) mean of dependent variable Estimated 2SLS effects Endogenous variable Point estimates (robust S.E.s) Point estimate as % of mean A. Dependent Variable: All crimes Army service (r) (.053) -12% 19% Upper bound as % of mean B. Dependent Variable: Violent crimes Army service (r) (.013) -36% 3.6% C. Dependent Variable: Non-violent crimes Army service (r) (.048) -6% 29% D. Dependent Variable: Property crimes Army service (r) (.013) -28% 29% E. Drink Driving Crimes Army service (r) (.014) -19% 15% 19

20 2SLS effect of army training on crime Endogenous variable (upper bounds: all 16 cohorts) mean of dependent variable Point estimates (robust S.E.s) Estimated 2SLS effects Point estimate as % of mean A. Dependent Variable: All crimes Upper bound as % of mean Army service (r) (.014) 0.4% 11% B. Dependent Variable: Violent crimes Army service (r) (.004) -0.1% 15% C. Dependent Variable: Non-violent crimes Army service (r) (.013) 0.5% 12% D. Dependent Variable: Property crimes Army service (r) (.004) 0.6% 22% E. Drink Driving Crimes Army service (r) (.005) 3.4% 17% 20

21 2SLS effects on crime (including r and Endogenous variables v together) mean of dependent variable point estimates (robust S.E.s) Estimated 2SLS effects point estimate as % of mean A. Dependent Variable: All crimes Upper bound as % of mean Army service (r) (.036) -2% 24%.269 Army service in Vietnam (v).020 (.095) 7% 76% B. Dependent Variable: Violent crimes Army service (r) (.010) -14% 20%.054 Army service in Vietnam (v).024 (.025) 44% 137% C. Dependent Variable: Non-violent crimes Army service (r).002 (.033) 1% 31%.215 Army service in Vietnam (v) (.084) -2% 75% D. Dependent Variable: Property crimes Army service (r) (.010) -19% 36%.036 Army service in Vietnam (v).021 (.025) 59% 196% E. Drink Driving Crimes Army service (r).001 (.010) 2% 32%.069 Army service in Vietnam (v).003 (.028) 4% 83% 21

22 Conclusions: Crime No significant effect on crimes of any type In the preferred specifications, we consider effects of army training: the point estimates are all negative we can rule out violent crime effects larger than 3.6% Under reasonable assumptions, can rule out effects on crime overall larger than 11% 22

23 2SLS effect on family outcomes 23

24 2SLS effect on health measures 24

25 25

26 26

27 27

28 28

29 29