Methods and Estimation Identification Threats and Trade-offs Applications and Key

Papers
RCTs Assumption: The assignment of the treatment is random. This is the only assumption. Krueger (1999)
Tennessee STAR class
Intent to treat: Internal validity: size project – showed
𝐸[𝑌1𝑖 |𝐺𝑖 = 1] − Sample attrition (if non-random, can skew results), or non-response bias 5% increase
𝐸[𝑌0𝑘 |𝐺𝑘 = 0] = 𝐸[𝑌1𝑖 − Movement across assigned groups (non-compliance): E.g. Tennessee STAR experiment
𝑌0𝑖 ] Substitution Bias – in the Tennessee STAR experiment, parents who had their children
assigned to control groups may seek outside tuition
Treatment on the treated:
𝛽𝐼𝑇𝑇 External validity:
𝛽𝑇𝑂𝑇 =
%𝐶𝑜𝑚𝑝𝑙𝑖𝑎𝑛𝑐𝑒 Generalisation – experiment specificity – usually in a geographic area
Randomisation – e.g. even if class sizes were randomised, schools choose whether to
Note: only works if participate in the RCT. Volunteers for an experiment etc.
compliance is random General Equilibrium Effects
Hawthorne Effects – Treatment group works harder
John Henry Effects – Control group works harder

Other issues: costs and ethical problems. Also, note that adding observables should not
change results much. If they do, then maybe randomisation is non-random.
OLS with observables Conditional independence assumption Magnuson et al (2007)
𝑌1𝑖 , 𝑌0𝑖 ⊥ 𝐷𝑖 |𝑋𝑖 𝐸[𝜂|𝑋𝑖 , 𝐷𝑖 ] = 𝐸[𝜂|𝑋𝑖 ] - Selection accounts for
𝑌𝑖 = 𝛼0 + ∆𝐷𝑖 + 𝛽𝑋𝑖 + 𝜂𝑖 𝐸[𝑌0𝑖 |𝐷𝑖 = 1, 𝑋𝑖 ] = 𝐸[𝑌0𝑘 |𝐷𝑘 = 0, 𝑋𝑖 ] - the pre-treatment outcomes are the same most of the effects of
Homogenous treatment effects 𝛽1 = 𝛽0 = 𝛽 going to preschool.
𝑌1𝑖 = 𝛼1 + 𝛽1 𝑋𝑖 + 𝜂𝑖 , 𝑌0𝑖 = 𝛼0 + 𝛽0 𝑋𝑖 + 𝜂𝑖 Control for rich set of
𝑌𝑖 = 𝐷𝑖 𝑌1𝑖 + (1 − 𝐷𝑖 )𝑌0𝑖 observables
𝑌𝑖 = 𝛼0 + (𝛼1 − 𝛼0 )𝐷𝑖 + (𝐷𝑖 𝛽1 + (1 − 𝐷𝑖 )𝛽0 )𝑋𝑖 + 𝜂𝑖
We need to assume 𝛽1 = 𝛽0 = 𝛽 Black et al (2003)
This gives us: 𝑌𝑖 = 𝛼0 + ∆𝐷𝑖 + 𝛽𝑋𝑖 + 𝜂𝑖 - Uses score matching
to control for
Be careful: Bad controls  won’t get causal estimates of treatment. Bad controls are observable
variables that are an alternative dependent variable. Also be careful of multicollinearity characteristics.
when including too many variables.
IV Assumptions: Exclusion condition 𝐶𝑜𝑣(𝑍𝑖 , 𝑌𝑖 ) = 0 and the relevance condition Angrist and Krueger
Motivation: Unobservable 𝐶𝑜𝑣(𝑍𝑖 , 𝑋𝑖 ) ≠ 0. Under the exclusion condition, we have the restriction that the IV must (1991) – age of birth
omitted variable only affect the outcome variable through the endogenous variable, and the condition that effects, birth month
the IV should be as good as randomly assigned.
Estimation: Wald Estimator, Angrist (1990) – LATE,
2SLS Watch out for violation of the 2 assumptions above. Also, not that in the context of investigating the effects
heteroegenous treatment effects, IV cannot be generalized. Can show that if we add the of veteran status on
𝐶𝑜𝑣(𝑍𝑖 , 𝑌𝑖 ) assumption of monotonicity (no defiers), the LATE theorem means that we have wages. – compliers.
𝐶𝑜𝑣(𝑍𝑖 , 𝑋𝑖 ) estimated the treatment effects on the compliers.
If treatment and IV are
discrete: we get Wald: Also, watch out for binary endogenous explanatory variables – use LPM in the first stage
𝐸[𝑌𝑖 𝑖 = 1] − 𝐸[𝑌𝑖 |𝑍𝑖 = 0] if explanatory variables are binary and endog.
|𝑍
𝐸[𝐷𝑖 |𝑍𝑖 = 1] − 𝐸[𝐷𝑖 |𝑍𝑖 = 0]
Numerator is reduced form Also, note that IV is BIASED but CONSISTENT. Moreover, as Ashenfelter and Krueger
of 𝑌𝑖 on 𝑍𝑖 while (1991) showed, bias can increase if other sources of bias (e.g. measurement error
denominator is 1st stage persistent).

Diff-in-Diff 4 assumptions Card and Krueger

Common Trend: The time trends for the treatment and control are the same, making the (1994) investigate the
𝑌𝑖𝑡 = 𝛽0 + 𝛽1 × 𝐷𝑖 + 𝛽2 × non-treated groups a valid counterfactual. Can check this graphically or parametrically effects of minimum
𝑇𝑡 + 𝛽3 (𝐷𝑖 × 𝑇𝑡 ) – by looking at leads and lags. Can also use placebo tests. wage in the restaurant
parametric Treatment effects are additive and constant: Intuitively, we cannot have heterogenous sector. Compares
response of treatment and control group – meaningless. trends in different
Non-parametric methods Group exogeneity: No switching of individuals between treatment and control groups due states to each other.
exist as well to treatment.
No other treatment occurring at the same time: Trivial assumption, but necessary.

Also, can always comment on external validity. E.g. minimum wage in specific sector is
different from minimum wage in the entire economy.

Also, watch out for: functional form dependence (logs vs normal), targeting of treatment
based on pre-existing differences in outcomes (e.g Ashenfelter’s dip), trade-off between
observing long-term policy response vs short-term effects. More interesting to find out
long-term effects, but more likely to face issues for confounding effects.
Extension to Diff-in-Diff: Motivation: If you are concerned about the diff-in-diff being for groups with different Meyer and Rosenbaum
Conditional Diff-in-Diff observable characteristics. Trends based on these observable conditions may be (2011): Tax policy on
different, but by conditioning for it we obtain common trends. labour participation
 However, group endogeneity issues may persist. rates among mothers.
Shows positive effects.
Fixed Effects: Motivation: to Concerns Ashenfelter and
remove time-invariant Krueger (1994): Twin
Fixed effects FE only removes time-invariant OMV, and we may be concerned that the results are still studies. Find that
biased due to time-varying OMV. measurement error
LSDV, de-meaning, F-D (attenuation bias)
Fixed effects removes much variation → imprecise estimators. when comparing IV
and OLS. Selection bias
Fixed effects may also exacerbate existing time-varying bias or measurement error. not found in their
results. (OLS vs FE).
Regression Discontinuity Assumptions: Ludwig and Miller
1. Treatment assignment is a deterministic and discontinuous function of the observable (2007). Effect of
Motivation: Near a 𝑥𝑖 → this implies that the CIA assumption is satisfied as near the threshold, treatment is headstart on long-run
threshold for the treatment, deterministic and thus uncorrelated with any unobserved determinant outcomes. RD effect
individuals are very similar, 2. No manipulation of individuals to just go beyond the threshold – can check for bunching because people are
so assignment of the at the threshold. Manipulation can occur from both the issuer of the treatment and the eligible. Good RD
treatment is quasi-random. recipient of the treatment (Exam treatment example) because individual
3. The underlying outcome function is not jumpy – can check this graphically. cannot strategically
Note: Fuzzy RD is IV 4. No other treatment around the same threshold change the
discontinuity. Find
Also, watch out for: evidence of large
1. Expectations effects – if the variable that the treatment is conditional on is age. decrease in mortality
2. LATE – estimate effects of people around the threshold. Sample size also small. for causes that should
be affected by head-
Note: Cannot use RD for something like subjective means testing. Also, controlling for start but not for causes
other observables should not change point estimates – if they do then it may indicate unaffected by head-
something wrong with our assumptions. start.