No interference
| Option | Description |
|---|---|
Required | Assumes strict no‐spillover. |
Recommended | Best if spillovers are minimal. |
Desirable | Improves validity but can be relaxed. |
Relaxes assumption | Explicitly allows and models interference (e.g., network models). |
Not required | Interference is irrelevant or explicitly allowed. |
Definition
Whether the method assumes each unit’s treatment does not affect any other unit’s outcome. It is, with the well-defined treatments assumption, part of the SUTVA condition (Stable Unit Treatment Value Assumption).
Explanation
Interference (spillover) violates independence. Most causal methods assume no interference; some advanced approaches model networks or spatial spillovers explicitly.
Tools/rationale for helping assessment
- Plot spatial coordinates of units and test for interference in untreated units next to treated ones vs. remote controls.
- Compute Moran’s I on treatment indicator to detect clustering/spillover.
- From your study design: If units are truly isolated, mark
Required; if only minor spillover possible,Recommended; if some spillover but manageable,Desirable; if you expect substantial interference and/or want to specifically study the impact of relaxing this assumption,Relaxes assumption.
Example
Moran’s I on burned vs. unburned patches =0.3 (p<0.01) ↦ strong spatial spillover. I choose a method that relaxes no‐interference or explicitly models spatial networks.
Logging in one forest patch decreases bird richness in adjacent patches (p<0.01), violating the assumption; thus, the method must relax no-interference or explicitly model spillovers.
Your 1 km² forest plots are >50 km apart (birds don’t travel far) →
Required. If plots are 5 km apart (some movement), Recommended. If plots are adjacent,Desirable. If there is a continuous landscape with high connectivity, you pick a spatial model that relaxes no‐interference.