Su et all, 2012

Problem

Confounding Variable

Interaction (Effect modification/moderation)

Solution

Design

Stratification

Matching

Cohort Restriction

Randomization

Covariate Imbalance (High Dimension)

Solution: Covariate Adjustment

Over-adjustment

Under-adjustment

Confounder is uncollected in the data or excluded from the model

A mediator is mistakenly considered as a confounder and included in the model for adjustment.

Controlling for a collider that correlates with both the treatment and the outcome via an "M-diagram"

Analysis

ANCOVA: analysis of covariance

DAG: directed acyclic graph

Problem: Different treatment effects at different levels or values of covariates.

Hard to interpret
Change direction/degree of its causal inference on the outcome:

Type

Non - linear Interaction

Quantitative Interaction

1st, 2nd, higher order interaction

Qualitative Interaction

Treatment- by- covariate interaction: Directional change in terms of treatment preference

Subgroup analysis:

Definition

Extract the maximum amount of information

Difficulties (malpractice)

Subgroup determination

Multiple testing

Lack of power

Explore the heterogeneity of treatment effect across sub-population

Extract the maximum amount of information