No unobserved confounders

Option	Description
`Required`	Method assumes no unobserved confounders.
`Recommended`	Best practice but not strictly enforced.
`Desirable`	Improves performance but not essential.
`Relaxes assumption`	Can tolerate some unmeasured confounding (e.g., DiD).
`Time-varying OR site-varying`	Assumes confounding is either time‐invariant and site-specific or time-varying but common to all sites.

Definition

Whether the method requires that all variables confounding the treatment-outcome relationship are observed and controlled or is designed to handle acknowledged / suspected unobserved confounders.

Explanation

Unmeasured confounding can bias causal estimates when not accounted for at all. However, many methods do need this key assumption to reach effect estimations, be it explicitly or implicitly. A few other methods are designed to relax this strong hypothesis, e.g. relying on specific data structures (regression discontinuity design). This assumption is also called causal sufficiency especially in the causal discovery literature, or conditional exchangeability. When no unobserved confounders are assumed, a strongly advised verification is to lead a sensitivity analysis to test the result robustness.

Tools/rationale for helping assessment

From your field knowledge, list all variables that influence both exposure and outcome and draw a causal graph to check for potential strong but unavailable confounders.
If you are confident that you have measured every major confounder, mark Required; if you aim to but may miss some, Recommended; if measurement would help but is not central, Desirable; if you lack key confounders, Relaxes assumption.
Perform sensitivity analysis afterwards to challenge estimated effects against unobserved confounding.

Example

In a study linking deforestation to species loss, an unmeasured “road proximity” confounder is identified after drawing a causal graph. Since this data is available on the national data portal, it is added to the study and the assumption can hold: I can use a method requiring causal sufficiency.
You have data on soil pH, moisture, elevation and NDVI → you believe no major confounders remain, so Required. If you lack soil pH, but it’s marginal, you mark Recommended. If, on the contrary many unknowns exist, you choose a design (e.g. DiD) that relaxes this assumption.