pubhlth 206a: using DAGS for causal & effect in observational studies

some lessons

• association is not causation, though causation produced association

• statistics alone cannot differentiate cause from effect

• subject matter knowledge is essential for valid causal inference

confounder

• associated with the outcome (or in some definitions, a risk factor for the outcome)

• associated with exposure

• not on the causal pathway between exposure & outcome

how to detect possible selection bias: convential methods

• study sample notably diff from the target population w/ respect to factors that are relevant to study questions

  • individuals who are more susceptible more likely to drop out
    of a study (or more likely to participate)

  • exposed cases more (or less) likely to be selected than
    unexposed cases

  • exposed controls more (or less) likely to be selected than
    unexposed controls

  • based on subject matter knowledge

directed acyclic graphs (DAGs) visual

graphs

diagrams representing an assumed system of causal connections b/w variables

directed

• arrows represent causal direction b/w variables

• ex: A —> B means that changing A would cause change in the expected value of B

acyclic

• cycles are not permitted b/c a variable can’t cause itself

  • ex: a0 —> b1 —> a2 —> b3

NOT a dag

ex: a ←→ b

DAGs benefits

• make assumptions explicit

• identify potential biases& covariates to measure

• avoid introducing biases during sampling or analysis

• determine how to correct for selection biases

• characterize mediation & direct vs indirect effects

in order ot identify a causal effect of e on y

• all non-directed (backdoor) paths leading from E (exposure) to Y (outcome) must be BLOCKED

  • paths are connections between variables on a DAG, through any
    number of arrows going in either direction

  • directed paths from E to Y are connections going only forward from E towards Y

a path is blocked shen it contains

a noncollider that is conditioned on a collider

*conditioning on a collider

reverse causality

e —> y or y —> e