1. A. CAUSAL EFFECTS Eva Hromádková, 7.10.2010 Applied Econometrics JEM007, IES Lecture 2A

2. Problem of causal inference  Want to test whether treatment (d) affects outcome (y) => TREATMENT EFFECT  !!! Correlation does not imply causation !!!  There might exist unobserved factors that drive this correlation  What would happen if an individual was (not) under a particular treatment?

3. Treatment effect I. Potential vs. observed outcome  Imagine individual has two potential outcomes Outcome if he is treated (d=1) Outcome if he is not treated (d=0)  Obviously, only one scenario is realized  Plugging (2) into (1) Note: individual return to treatment

4. Treatment effect II. Why are some treated and some not?  Hidden selection mechanism, based on observed (Z) and unobserved (v) factors  Then translates into 0/1 treatment

5. Treatment effect III. From individual to population  Average treatment effect (ATE) (randomly chosen individual)  Average treatment effect on treated (ATT) (participant of treatment)  Average treatment effect on non-treated (ATNT) Hypothetical effect - non-participant

6. Treatment effect IV. Heterogenous in population  Local average treatment effect (mainly in IV)  We observe variation in variable Z, which induced change in treatment status of SOME individuals  Ex.: subsidy for dormitories -> positive effect on enrollment into higher education  BUT effect that we are getting is local – only applies to people who switched their decision based on the subsidy

7. Identification problem Non-random assignment  Selection into treatment group =>  People who are treated are a-priori different from people who are not treated  Terminology: treatment x control group  Q: how is this different from LATE?

8. Identification problem Selection mechanisms  Selection on observables (corr of e and Z)  Selection on unobservables (corr of e and v)  Selection on untreated outcome (corr of d and u)  Selection on expected gains (corr of d and alpha)

9. Identification problem Homogenous vs. heterogenous treatment effects  Homogenous case:  Selection bias if u and d are correlated  Heterogenous case:  ATT + selection bias from corr of u and d

10. Overview of identification strategies  Controlled (social) experiment  Direct randomization of treated and untreated  Natural experiment  Finding naturally occurring treated and untreated group that are as similar as possible  Instrumental variable  Finding variable that affects prob. of treatment but does not affect outcome  Discontinuity design  Probability of treatment is changing discontinuously with a characteristic (eligiblity)

11. B. CONTROLLED EXPERIMENTS Eva Hromádková, 7.10.2010 Applied Econometrics JEM007, IES Lecture 2B

12. Randomization Experimentator can randomly choose which individuals are administered treatment and which not Ass.1: Treated and controls same in unobserved characteristics Ass. 2: Treated and controls same in gains from a treatment

13. Use of experiments and randomization  Labor Economics – Active labor market policies  Ex. National Supported Work (NSW)  Health Economics:  Ex. RAND experiment (1974-1982) people were assigned randomly to different health insurance plans Moral hazard; effect of co-payments  *Development economics:  Educational system (Duflo, Dupaas and Kremer, 2009), microfinances (Karlan and Zinman, 2008)  *Behavioral economics  Intrinsic motivation, fairness, incentives

14. Development economics Improving immunization coverage in India  Video (Esther Duflo, TED Talks Feb 2010) Video Banerjee, Duflo and Kothari (2010) – Improving immunization coverage in rural India  Udaipur district, Rajasthan – very low immunization rate (4%  Reasons: Cost of travelling (immunization is for free) – procrastination  134 villages were randomized to one of 3 groups  A: reliable immunization camp  B: reliable immunization camp + incentives (lentils + plates)  No intervention  Outcome = immunization rate in villages

15. Development economics Results:  Baseline – 6%  A – 17%  B – 38% Issues:  Design: Testing multiple interventions  Within village correlation of individual outcomes - clustering  Spillovers – neighboring villages  Intention to treat

16. Behavioral economics How to combat procrastination I  Video (Dan Ariely on procrastination) Video Ariely and Wertenbroch (2002). Procrastination, deadlines and performance.  Procrastination = putting off duties/tasks  Questions: 1. Do people self-impose deadlines to increase performance if they have the possibility to do so? 2. Do deadlines increase performance? 3. Do people set deadlines optimally for maximum performance?  2 studies

17. Behavioral economics How to combat procrastination II Study 1: MBA course – 2 classes, requirement of 3 essays  No choice section: fixed deadlines (evenly spaced)  Free-choice section: choose deadlines themselves  Deadlines will be binding  Instructor will not read / give feedback before the end  Rational choice (if no self-control issues) = all 3 in the end  Results:  Actual choice of deadlines: only 32% for the final week  Performance: grades in no-choice section (avg = 88.76) higher than grades in choice section (avg=85.67), t=3 Problem with SE (=>t). Why?

18. Behavioral economics How to combat procrastination III Study 2: Proofreading, randomly assigned to 3 treatments  1. Evenly spaced submission (every 7 days)  2. End-deadline submission (at the end)  3. Self-imposed deadlines  Conditions: paid for detecting mistakes, day of delay = 1$ Results:  Participants in (3) have preferred spaced deadlines  They perform worst under no deadlines, better under self imposed deadlines and best under imposed deadlines  However, people sometimes set constraints that are not really constraining (“internal” deadlines, gym membership, etc.)

19. Issues in controlled experiments I Threads to internal validity:  Non-compliance:  Some people assigned to treatment do not comply  What we get is the effect of “intention to treat”  Attrition: problem if it is non-random  Externalities: not taking them into consideration reduces estimated impact of treatment  Correct SE => clustering (e.g. randomization of villages) Design questions: few examples  Framing  Relevant incentives: own / experiment money  Testing multiple interventions

20. Issues in controlled experiments II Threads to external validity: is the result generalizable?  Hawthorne effect: mere attention causes the treatment group to change its behavior  John Henry effect: when control group engages in social competition to show they perform as well  Demand effects: subject cooperate in ways they wouldn’t routinely consider Generally:  Population: too specific?  Time span: do we control also for long run effects?  GE effects: implementation on a large scale