Depicting patient-reported outcome measures within directed acyclic graphs: practice and implications for causal...

Estimating causal effects of an exposure (e.g., health condition or treatment) on a patient-reported outcome measure (PROM) can have complications depending on the relationship between the PROM's indicators and construct(s). Using directed acyclic graphs (DAGs) as visual tools, we show how to represent a PROM's potential internal causal relationship between its indicators and latent construct(s), then explain the implications when also accounting for external variables when estimating causal effects within observational data.

Measurement theory suggests a PROM's relationships between its items/indicators and latent construct(s) is reflective (construct causes the indicators) or formative (indicators cause the construct). We present DAGs under reflective and formative model assumptions when the PROM is unidimensional (e.g., Patient Health Questionnaire-9 [PHQ-9] representing depression severity) or multidimensional (e.g., EQ-5D representing health-related quality-of-life).

Unidimensional PROMs under a reflective model can be analysed like other unidimensional outcomes (e.g., mortality) to estimate causal effects, thus don't require additional consideration. In comparison, each indicator of a multidimensional construct under a formative model needs specific consideration to ensure relevant external variables are appropriately conditioned to estimate causal effects.

Multidimensional outcome constructs formed under a formative model increases the complexity of causal analyses. Despite this, multidimensional measures may particularly aid with a variety of 'outcome-wide' studies when assessing exposures that may be beneficial for some outcomes but harmful for others. Thus, we have taken important steps to supporting such studies in observational settings by showing how PROMs can be incorporated into DAGs to inform such causal analyses.