Summary Experimental Economics (FEM11089)

Summary Experimental Economics
Lecture 1: Causal Inference
Experiment
- Scientific procedure
- Controlled data generating process
o Control = the power to keep all factors that influence behavior constant, except the
factor of interest.
o Data generating process = experimental economists create their own data.

Causal inference
- Confounding effects  bias results, additional effects  never do before and after
calculations  always T vs. C
- Selection bias  exam will be on this for sure
o Randomization
o Extra reading on topic: Angrist & Pischke (2009) Mostly Harmless Econometrics CH2
- Classic pitfalls
o Do before & after calculations
o Compare between subjects treated and subjects not treated  not comparable
because of confounding effects & selection effects
o Example Apple users have more sex than Android users  wrong causal relation
- Counterfactual
o Construct from naturally occurring data
 Diff-in-Diff
 Propensity score matching
o Construct by creating data
 Lab experiments

Counterfactual from naturally occurring data
- From data  construct T and C groups
- Three conditions:
o T and C group must be the same in absence of the treatment, on average
o T and C group should react to treatment in the same way
o T and C group cannot be exposed in isolation to a third factor  no counterfactuals
in your data
- Diff-in-Diff
o Subtract t+1 from t from C
o Subtract t+1 from t from T
o Subtract C from T  no confounds, no
selection bias
o Major assumption:
 Parallel trends assumption =
observed and unobserved
characteristics remain constant
over time
o Used when selection bias is likely

, o Popular tool to analyse natural experiments when an exogenous shock occurs
 E.g. lockdown
- Propensity score matching (PSM)
o Match a treated subject to an ‘identical’ untreated subject
o Major assumption: unobserved characteristics are equal for treated and untreated
match
o When to use?
 No baseline period available
 When subjects are not randomly put into treatments, selection bias issue
o Ideally match identical in all reports, however: the more dimensions the harder
 Curse of dimensionality
 Rosenbaum & Rubin (1983) overcome this curse:
 Estimate the probability that a subject is treated
 I.e. Probit estimator
 Match subjects based on this estimate

Counterfactual from creating data  experiments
- Construct own T and C
- Same three conditions must hold
- Two major problems with natural data are solved by creating data:
o Confounding effects: a properly designed experiment has none
o Selection effects: randomization
- Selection bias
o Example of grading your health in EMC vs. city centre  higher grades in centre 
hospitals make you ill  wrong
o If subjects choose themselves to undergo treatment, there will be selection bias




o

, o
- Randomization takes out selection bias effect

Conclusion
- Experiments are perfect for causal inference because:
o Counterfactual can be observed
o No confounds
o Through randomization no selection bias
- Setting up a proper experiment is not always easy



Lecture 2: Economic Experiments
Recap
- Causal effect & selection bias
- Experiments isolate the causal effect, allow us to observe the counterfactual and if we don’t
have confounding factors then we observe the causal effect

Microeconomic systems
- Experiments are like microeconomic systems
- Two elements play a key role:
o Environment (physical things)
o Institution (rules of the game)
- Environment
o Consists of:
 N economic agents
 K+1 commodities (incl. resources) 0, 1, …, K  the +1 = money
 Characteristics of each agent i:
 Utility function ui
 Technology (knowledge) endowment Ti
 Commodity endowment vector wi  what we give them, physical
 e = (ui, Ti, wi)  summary
i

o Microeconomic environment: e = (e1, …, eN)
o Initial set of conditions that cannot be changed by the agents
- Institution
o Language: M = (M1, …, MN) consisting of
 Messages: m = (m1, …, mN) that agents send
 E.g. bids, offers, acceptance
o Allocation rules: H = (h1(m), …, hN(m))  benefit

,  The rule hi(m) states the final commodity allocation to each i as a function of
the messages sent by all agents
o Cost imputation rules: C = (c1(m), …, cN(m))  cost
 The rule ci(m) states the payment to be made by each agent as a function of
the messages sent by all agents
o Adjustment process rules: G = (g1 (t0, t, T), …, gN (t0, t, T))
 Starting rule: gi (t0, ., .)
 Transition rule: gi (., t, .)
 Stopping rule: gi (., ., T)
- Example: auction market
o N agents  buyers and sellers
o Commodities  products offered by sellers
o Characteristics
 Buyers u = utility minus price
 Sellers u = price minus costs
o Commodity endowment vector w  cash & goods
o Messages  bids & offers
o Allocation rules  highest bidder gets good
o Cost computation rule  buyer pays price to seller
o Adjustment rule (timing of the game)
 Price cannot be smaller than 0 (t0)
 New bid > old bid (t)
 If no bid follows the previous bid the market ends (T)
- Microeconomic systems
o ‘property rights’ in communication and exchange are given by:
 Ii = (Mi, hi(m), ci(m), gi (t0, t, T))
o Microeconomic institution is the collection of property rights:
 I = (I1, …, IN)
o Microeconomic system is the microeconomic environment together with the
microeconomic institution
 S = (e, I)
- Lab vs. real world
o Observable in real world
 Agents
 Physical commodities and resources
 Endowments (physical commodities & resources)
 Language and property rights
 Outcomes
o Observable in lab
 Utility function
 Technological endowments
 Agent message behavior (private and unrecorded in real world)

Sufficient conditions for a microeconomic experiment
- Experiment gives control, but also gives insight in preferences
- Example: wine buying for cost price or 100 euro
o Buy me a bottle of wine, I will pay you back