Skip to main content

Advanced Causal Query

Since version 10

Introduction​

An Advanced Causal Query is useful in the following situations:

  1. Your Causal model contains Unobserved Confounders. You can use one of the causal inference algorithms that support Unobserved Confounders, such as the Backdoor adjustment algorithm or the FrontDoor adjustment algorithm.
  2. You want to use the Disjunctive Cause algorithm.
  3. You want to determine adjustment sets required for certain causal queries. e.g. using the Backdoor Criterion algorithm or the FrontDoor Criterion algorithm.
  4. You want to calculate the Direct Effect, as some algorithms only support Total Effect.

[!IMPORTANT] If your causal model does not contain Unobserved Confounders then you typically do not need to use the Advanced Causal Query dialog, unless you want to compute Direct Effects. You can simply use the default Relevance Tree algorithm to perform causal inference, including standard Custom/Joint queries, as these other algorithms also support Interventions (Do evidence).

Causal Identification​

The Advanced Causal Query dialog supports calculation of:

  • Adjustment sets using the Backdoor Criterion algorithm, which calculates valid adjustment sets to use with the Backdoor Adjustment/Inference algorithm.

  • FrontDoor nodes, and subsequent adjustment sets using the FrontDoor Criterion algorithm, for use with the FrontDoor Adjustment/Inference algorithm.

Backdoor adjustment sets

FrontDoor nodes

FrontDoor adjustment sets ZY

Causal Inference​

In addition to Identification the Advanced Causal Query dialog can also perform inference/adjustment.

Bayes Server supports:

  • Multiple treatments (variables with Interventions)
  • Multiple outcomes (e.g. Joint Query over multiple outcomes)
  • Unobserved confounders (some algorithms)
  • Total Effects
  • Direct Effects (some algorithms)
caution

As stated earlier If your causal model does not contain Unobserved Confounders then you typically do not need to use the Advanced Causal Query dialog, unless you need to compute Direct Effects.

Causal Inference

Total Effects​

Total Effects includes effects that do not flow via direct links from Treatments -> Outcomes.

  • e.g. Treatment -> Other Node -> Outcome
  • e.g. Treatment <- Other Node -> Outcome
  • etc...

This is the default in Bayes Server, and while standard algorithms such as Relevance Tree support interventions, they do not support Direct Effects.

Direct Effects​

Direct Effects only includes effects that flow via direct links from Treatments -> Outcomes.

  • e.g. TreatmentA -> OutcomeA
  • e.g. TreatmentA -> OutcomeB
  • etc...