Where Do Randomization Blocks Go in a Causal DAG?
In a Randomized Controlled Trial (RCT), blocking is used to ensure that the treatment and control groups are balanced across specific covariates (like age or gender). In a Causal DAG, these blocking factors ($B$) are not just background noise; they are part of the "mechanism" that determines who gets the treatment ($T$).
1. The "Mechanism" Node: $B \to T$
In simple randomization, we assume $T$ is determined by a purely exogenous coin flip ($R \to T$). However, in Block Randomization, the "coin" is flipped separately within each block. Therefore, the blocking factor $B$ must have a directed arrow pointing to the treatment $T$.
- Structural Rule: $B \to T$. This arrow indicates that the probability of receiving treatment is functionally dependent on which block the subject belongs to.
- Independence: Because $T$ is randomized within $B$, there are no other arrows from unobserved confounders ($U$) pointing into $T$. This "cleans" the causal path.
2. Is the Block a Confounder?
One of the most frequent "Super User" debates on Cross Validated is whether the block $B$ is a confounder. In a DAG, a confounder must cause both the treatment and the outcome ($T \leftarrow B \to Y$).
- If $B$ also affects the outcome: Then $B$ is a true confounder. You must adjust for $B$ in your analysis (e.g., via fixed effects) to block the non-causal "backdoor" path $T \leftarrow B \to Y$.
- If $B$ does not affect the outcome: It is technically just a "randomization instrument." However, even in this case, "as you randomize, so shall you analyze"—standard practice in 2026 dictates that you still include the blocking factor in your model to increase precision.
3. Comparison: Simple vs. Blocked Randomization in DAGs
| Feature | Simple Randomization | Block Randomization |
|---|---|---|
| Primary Parents of T | Exogenous Noise ($R$) | Blocking Factors ($B$) |
| Backdoor Paths | None (by design) | $T \leftarrow B \to Y$ (if $B$ affects $Y$) |
| Mandatory Adjustment | No | Yes (to ensure exchangeability) |
| DAG Complexity | Low ($T \to Y$) | Medium ($B \to T, B \to Y, T \to Y$) |
4. The Pitfall: Block as a Collider?
A common mistake is fearing that $B$ is a Collider. A collider is a variable that is caused by both $T$ and $Y$ ($T \to B \leftarrow Y$). In a randomized block design, $B$ is determined before the treatment is assigned. Therefore, $B$ is an Ancestor, not a collider. Conditioning on $B$ closes a backdoor path; it does not open a new one.
5. Visualizing the "Block DAG"
When drawing your DAG for a journal or Cross Validated, follow this standard 2026 layout:
- Top Level: Draw your covariates ($X_1, X_2$) that make up the block.
- Middle Level: Draw the Block $B$ (a node representing the combination of those covariates).
- Treatment: Draw the arrow $B \to T$.
- Outcome: Draw arrows $T \to Y$ and $B \to Y$.
Conclusion
In a causal DAG, randomization blocks are placed as parents of the treatment node. This reflects the reality that the experimental design itself is the "cause" of the treatment assignment within specific strata. By correctly identifying $B$ as a potential confounder (if it influences $Y$), you uphold the core principle of causal inference: "Condition on the factors used to determine assignment." Failing to include these arrows in your DAG would imply that treatment was assigned purely by chance across the whole population, which is a structural misrepresentation of a blocked design.
Keywords
randomization blocks causal DAG, blocked randomization directed acyclic graph, adjustment for blocking factors, causal inference RCT DAG, Cross Validated block design, backdoor path blocking factor, d-separation block randomization, 2026 causal modeling standards.
