Observational Causal Inference Exam Workshop

Observational questions should begin with a causal diagram or assignment story, not a propensity model. Decide which pretreatment variables are sufficient to block confounding paths, then assess overlap, balance, and the target population.

Worked problem 1: confounding and standardization

An observational training study has two prior-skill strata. Half of the target population is high skill and half is low skill. Within each stratum, the observed mean outcomes are:

Prior skill $E[Y\mid W=1,X]$ $E[Y\mid W=0,X]$ $e(X)$
High 70 62 0.80
Low 55 50 0.20

Because 80% of treated units are high skill while only 20% of controls are high skill, the crude observed contrast is

\[[0.8(70)+0.2(55)]-[0.2(62)+0.8(50)]=67-52.4=14.6.\]

Under conditional unconfoundedness and overlap, standardize both treatment conditions to the target covariate distribution:

\[E[Y(1)]=0.5(70)+0.5(55)=62.5,\] \[E[Y(0)]=0.5(62)+0.5(50)=56.\]

Therefore,

\[ATE=62.5-56=6.5.\]

The crude gap combines a treatment contrast with a large composition difference. Standardization removes the part explained by the measured prior-skill stratum, conditional on the identification assumptions.

Worked problem 2: what inverse weights create

For a treated high-skill unit, the ATE weight is $1/0.80=1.25$. For a treated low-skill unit, it is $1/0.20=5$. Control weights reverse the pattern:

\[w_i= \begin{cases} 1/e(X_i), & W_i=1,\\ 1/[1-e(X_i)], & W_i=0. \end{cases}\]

Low-skill treated units and high-skill controls are rare relative to their covariate strata, so each receives more weight. The weighted pseudo-population represents both skill strata in both treatment conditions. This explains the estimator; it also explains why estimated propensities near zero or one cause instability.

If a high-skill control has $\hat e(X)=0.98$, its control weight is

\[\frac{1}{1-0.98}=50.\]

One observation can then dominate the estimate. Report the propensity distribution, effective sample size, balance after weighting, and any prespecified trimming or overlap-weight strategy.

Worked problem 3: matching estimand and reuse

Suppose three treated units have propensity scores $(0.22,0.51,0.79)$ and four controls have scores $(0.18,0.47,0.55,0.81)$, with outcomes $(12,18,25)$ for treated units and $(9,13,17,20)$ for controls.

One-to-one nearest-neighbor matching without replacement pairs

\[0.22\leftrightarrow0.18, \qquad 0.51\leftrightarrow0.47, \qquad 0.79\leftrightarrow0.81.\]

The matched ATT estimate is

\[\widehat{ATT}=\frac{(12-9)+(18-13)+(25-20)}{3}=4.33.\]

The unmatched control at 0.55 is outside this particular matched comparison. Matching changes the analyzed population, and a no-replacement rule can force later units into worse matches. The answer should state the estimand, distance, replacement rule, caliper, and balance diagnostics.

Worked problem 4: variable selection by causal role

A study estimates the effect of job training on later earnings. Available variables are age, prior earnings, motivation measured before enrollment, training completion, and earnings six months after training.

  • Age, prior earnings, and pretreatment motivation may confound selection and outcome, so they are candidates for the assignment model.
  • Training completion occurs after assignment and can mediate the program effect. Conditioning on it changes the estimand and may induce bias.
  • Six-month earnings are the outcome and must not enter a propensity model intended to reproduce pretreatment assignment.

Predicting treatment well is not the objective. The objective is covariate balance for a defensible causal comparison.

Exercises

  1. For the two-stratum table, derive the ATT when 80% of treated units are high skill. Compare it with the ATE and explain why they differ.
  2. Calculate ATE weights for propensities $0.05$, $0.25$, $0.50$, $0.75$, and $0.95$ under both observed treatment states. Identify the unstable cases.
  3. Repeat the matching problem with a caliper of 0.03. State which treated units remain unmatched and how the target population changes.
  4. A propensity model achieves an AUC of 0.91 but leaves a standardized mean difference of 0.28 for prior outcome after weighting. Evaluate the design.
  5. Draw a causal diagram in which conditioning on a post-treatment variable opens bias. Explain why adding every available variable is unsafe.
  6. State which parts of unconfoundedness and overlap can be assessed from observed data and which cannot.

Solution checks

  1. $ATT=0.8(70-62)+0.2(55-50)=7.4$. It weights effects by the treated covariate distribution rather than the target population distribution.
  2. The largest weights are 20 for a treated unit at 0.05 and for a control at 0.95; both signal weak overlap.
  3. Only $0.79\leftrightarrow0.81$ lies within 0.03; the estimate then applies to a narrow matched subset.
  4. Treatment prediction does not establish balance. A standardized difference of 0.28 is substantial residual imbalance, so revise the design before estimating effects.
  5. A mediator or collider measured after treatment can block part of the effect or open a noncausal path.
  6. Overlap and observed-covariate balance are diagnosable. Absence of unmeasured confounding is not testable from the observed treatment-outcome data alone.

See