Data schema & when to measure each variable

Table layout (wide format)

column

meaning

type

when to measure

id

unit identifier

int/str

baseline

D

treatment indicator (0/1)

int

t₁ (decision/exposure time)

Y

outcome (numeric or binary)

float/int

t₂ > t₁ (follow-up)

X_*

confounders (all causes of both D and Y you can observe)

float/int

t₀ < t₁ (strictly pre-treatment)

Timing rule of thumb (avoid post-treatment bias):

t₀ (baseline): measure X (pre-treatment confounders only)
      ↓
t₁ (assign/observe D)
      ↓
t₂ (observe Y)  — make sure no X measured here leaks into the model

  • Do not include mediators (variables affected by D) in X. DML/IRM assumes X is pre-treatment

  • If panel data exist, freeze a snapshot of X right before t₁.

  1. Question & estimand

  • State causal question (e.g., “Effect of D on Y for the target population at t₁–t₂”).

  • Choose estimand: ATE or ATTE

  1. Identification assumptions

  • Unconfoundedness: \(((Y(1),Y(0)) \perp D \mid X)\) (with your chosen, pre-treatment (X)).

  • Overlap (positivity): \((0<e(X)=P(D=1\mid X)<1)\) almost surely.

  • Consistency/SUTVA: well-defined treatment, no interference.

  • Score check: psi_mean, derivatives, psi_kurtosis

  1. Report:

  • p-value and CI

  • Assumptions tests

  • Why were such variables chosen

  • Quantity of units in research

  • Conclusion for the decision-maker and what decisions would be made

On this page