poder lab experiments in the field 1 capetown 7/8/14
TRANSCRIPT
Capeto
wn
PODER
Lab Experiments in the Field1
7/8
/14
2
Capeto
wn
OUTLINE FOR TODAY:
Leftover from yesterday Multi-site experiments: Solidarity
7/8
/14
3
Capeto
wn
NUTS AND BOLTS
7/8
/14
4
Capeto
wn
SECOND STEPS:(AFTER THE QUESTION/THEORY)
Instrumentation Construct Validity – how will I test what I want to
test? Paper/Pencil or Computer? Timeline of experiment Instructions
Sampling/Randomization What subject pool? How will Treatment be randomized?
Analysis Plan What are the units of analysis Power tests
7/8
/14
SUBJECT SELECTION, I Convenience Samples: students
Students advantages: Convenient, inexpensive and relatively homogeneous
Student disadvantages: May behave differently from target population,
young, educated, and talk to each other (diffusion) Classroom:
Representative sample of students Environment might affect behavior:
Lab: May select certain students Neutral environment
Data: Eckel and Grossman ExEc: Students give more to charity in the classroom than in the
lab Why?
5
Capeto
wn
7/8
/14
SUBJECT SELECTION, II
Specialized Groups:ElderlyProfessionalsMedical casesPoorResidents of hurricane-vulnerable areasPublic officials
Population Samples Pluses: External validity, Heterogeneity Minuses: Costly, risk of decreased control,
heterogeneity6
Capeto
wn
7/8
/14
SUBJECT SELECTION III
Subject selection should suit the question you are asking
Theory testing: Independent of subject characteristics?
Policy (measurement or institutional design): Target group subjects
Examples: WEIRD people (Henrich, et al. 2010) People from other cultures (Barr and Serra
2010)
7
Capeto
wn
7/8
/14
8
Capeto
wn
SECOND STEPS:(AFTER THE QUESTION/THEORY)
Instrumentation Construct Validity – how will I test what I want to
test? Paper/Pencil or Computer? Timeline of experiment Instructions
Sampling/Randomization What subject pool? How will Treatment be randomized?
Analysis Plan What are the units of analysis Power tests
7/8
/14
NUTS AND BOLTS, I Lab log. IRB and Ethics Pilot experiments. Lab set-up Subject registration Experimenter(s) Monitor(s) Randomizing Devices Instructions Subject confidence (non-deception)
9
Capeto
wn
7/8
/14
10
Capeto
wn
LAB BOOK (LUPIA & VARIAN 2010) 1. State your objectives. 2. State a theory. 3. Explain how focal hypotheses are derived from the theory if the
correspondence between a focal hypothesis and a theory is not 1:1. 4. Explain the criteria by which data for evaluating the focal hypotheses were
selected or created. 5. Record all steps that convert human energy and dollars into datapoints. 6. State the empirical model to be used for leveraging the data in the service of
evaluating the focal hypothesis. (a) All procedures for interpreting data require an explicit defense. (b) When doing more than simply offering raw comparisons of observed differences between treatment and control groups, offer an explicit defense of why a given structural relationship between observed outcomes and experimental variables and/or set of control variables is included.
7. Report the findings of the initial observation. 8. If the findings cause a change to the theory, data, or model, explain why the
changes were necessary or sufficient to generate a more reliable inference. 9. Do this for every subsequent observation so that lab members and other
scholars can trace the path from hypothesis to data collection to analytic method to every published empirical claim.
ELNs: OneNote in Microsoft or Growlybird Notes for the Mac (http://www.growlybird.com/GrowlyBird/Notes.html)
7/8
/14
NUTS AND BOLTS, I Lab log. IRB and Ethics Pilot experiments. Lab set-up Subject registration Experimenter(s) Monitor(s) Randomizing Devices Instructions Subject confidence (non-deception)
11
Capeto
wn
7/8
/14
12
ETHICS
IRB keeps us honest (some countries don’t have) Focus on potential harm to subjects Consent, debriefing limit harm, but may impact
sample Balance between potential benefit and risk
Field experiments: No consent process! Unwitting subject, high potential
cost Findley et al 2014. – no consent, no debriefing Correspondence studies on discrimination (more later) Intervention studies: elections, political institutions Facebook study on emotional contagion: no consent,
potential risk, very low potential benefit
7/8
/14
Capeto
wn
NUTS AND BOLTS, I Lab log. IRB and Ethics Pilot experiments. Lab set-up Subject registration Experimenter(s) Monitor(s) Randomizing Devices Instructions Subject confidence (non-deception)
13
Capeto
wn
7/8
/14
NUTS AND BOLTS, II Subject questions “Learning periods” Experiment Recording data Termination of experiment Debriefing Subject payment Bankruptcy Backup plan
14
Capeto
wn
7/8
/14
15
Capeto
wn
WRITEUP AND REPORTING
Biases in published data Registration and CONSORT
7/8
/14
16
Capeto
wn
BIASES IN PUBLISHED DATA
Selective reporting + publication bias => many published studies have p=.05.
Data mining and selective presentation of results have been a concern in economics for a long time
These concerns are not limited to Economics: Medical trials, Ioannidis (2005, “Why most published
research findings are false”) Psychology, Simmons et al. 2011, “False - positive
psychology: Undisclosed flexibility in data collection and analysis allows presenting anything as significant”)
Political science, Humphreys et al. (2012, “Fishing”) Finds affect millions of people. How to fix?
7/8
/14
17
Capeto
wn
EXAMPLE: (GERBER AND MALHOTRA, AJPS)
7/8
/14
18
Capeto
wn
ECONOMICS, “STAR WARS” (BRODEUR ET AL 2013)
7/8
/14
Capeto
wn
19
CONSORT/REGISTRATION (HUMPHREYS ET AL, 2013)
CONSORT Statement: improve the reporting of a randomized controlled trial (RCT), enabling readers to understand a trial's design, conduct, analysis and interpretation, and to assess the validity of its results. http://www.consort-
statement.org/
7/8
/14
20
Capeto
wn
REGISTRATION
Benefit: Limits selective reporting/fishing Rounds out “body of evidence” Forces researcher to think through design,
statistical analysis
Potential costs Limits exploratory research Serendipitous findings may be hard to publish
But: frees it from the burden of (false) presentation as formal hypothesis testing.
7/8
/14
GENERAL REMARK: LAB V. FIELD
Lab has greater internal validity Lab is cheap, field is costly Lab mistakes can be fixed; often
not so in field Students v. population
Population has higher variance, harder to detect effects
Selection bias is not limited to labGreater monitoring costs to ensure
population sample21
Capeto
wn
7/8
/14
22
Capeto
wn
IF TIME, HERE ARE UDT SLIDES
7/8
/14
THE ULTIMATUM GAME
Task: Two players must divide a fixed amount-$100 Game is played once only
Players Proposer- chooses a split Respondent-accepts or rejects
Payoffs: If accepted, money split as planned If rejected, both players get zero
Game theory: Start with responder. Payoff-maximizer accepts
anything >0 Therefore proposer offers smallest possible amount
$5 ULTIMATUM GAME (ECKEL AND GROSSMAN, RUN IN 1992, PUB. 2000)
Figure 1: Distribution of Offers, All Treatments
0% rejected
0% rejected
0% rejected
1.6%rejected
100%rejected
83.3% rejected
56.5% rejected
20.9%rejected
0
0.1
0.2
0.3
0.4
0.5
0.6
0 0.5 1 1.5 2 2.5 3 3.5 4
Amount Offered
Per
cen
t o
f O
ffer
s Rejected
Accepted
ULTIMATUM GAME RESULTS
Unequal offers are rejected Payoff-maximizing offer is modal offer is
60/40 split Most common split is 60/40 Looks as if proposers are “rational” What about responders?
(Why do they reject?)
BEHAVIOR IS INCONSISTENT WITH STANDARD MODEL:
Why do people offer >0? Value other’s payoff (Altruism) Afraid to be rejected (lose it all)
Risk averse?
Why do people reject? Fairness: think the distribution is not fair overall
Note we see people rejecting really high offers, too. Selfish fairness: care about own relative payoff Low offer means cost of punishment is low
THE DICTATOR “GAME”
Task: two players must divide a fixed amount-$100
Players Proposer- chooses a split Respondent-passively accepts
Dictator game allows player 1 to make an altruistic allocation
We use this game to measure altruism
REAL PEOPLE PLAY DICTATOR GAME : HISTOGRAM OF DECISIONS
Giving in the Dictator Game
0
0.1
0.2
0.3
0.4
0.5
0.6
0.7
10 9 8 7 6 5 4 3 2 1
Amount Kept
Pe
rce
nt
of
Su
bje
cts
DICTATOR RESULTS/COMPARISON
2/3 of subjects are selfish Many give away something – altruistic Measure of altruism!
Varies across individuals Reliable Correlated with behavior like volunteering
Different conditions lead to different outcomes “Double blind” (see prev. slide) Identity Charity
MEN AND WOMEN – DICTATORAmount Donated by Men and Women
0
0.1
0.2
0.3
0.4
0.5
0.6
0.7
0 1 2 3 4 5 6 7 8 9 10
Dollars Donated to Anonymous Recipient
Perc
en
t o
f D
ec
isio
ns
Women
MenMen donate $0.82
Women donate $1.60
Source: Eckel and Grossman, Economic Journal, 1998
TRUST GAME
Player A and Player B both begin with $100. (important!)
Player A decides how much, if any, to send to Player B. Any amount sent is tripled on the way to B.
Player B decides how much, if any to send back to Player A.
Game theory: B returns 0, so A sends 0 This game is used to measure trust and
reciprocity.
32
Capeto
wn
TRUST GAME RESULTS:
People send positive amounts Trust (just) pays on average In the field, higher levels of trust and
reciprocity Measure of individualized trust (More on this later).
7/8
/14
TESTING THEORY V. MEASUREMENT
These games violate predictions of standard game theory (assuming payoff-maximizing agents) Led to a huge amount of research (experimental
and theoretical) These games are useful for measuring
preferences and social norms
LAB EXPERIMENTS IN THE FIELD
Lab or field? False dichotomy Lab experiments complement field exps.
Lab experiments in the field (extra-lab) for measurement, etc.
Lab experiments back home to settle methodological issues that arise in the field
Pretest and refine experimental designs before going into the field
Issues: External validity (perpetually)
34
Capeto
wn
35
NEXT: SOLIDARITY
7/8
/14