OBJECTIVES: Health-related patient-reported outcome measures (PROMs) are used in a variety of contexts, including assessing treatment effectiveness, exploring health determinants, and guiding clinical decision-making. As such, we are often interested in the causal effects of an exposure (e.g., treatment or health condition) on an outcome represented by a PROM (e.g., health-related quality-of-life [HRQoL]). However, a PROM’s design and construct(s) may have implications for estimating causal effects. Directed acyclic graphs (DAGs) are visual tools for representing hypothesized causal relationships between variables to facilitate causal analyses. Here, we demonstrate how DAGs may be used to represent the internal causal relationship between a PROM’s indicators (e.g., items) and its latent construct(s), and discuss the implications for causal effect estimation.

METHODS: Measurement theory suggests the internal relationships between a PROM’s items/indicators and its latent construct(s) may be reflective (where the construct causes the indicators) or formative (where the indicators cause the construct). We explore how both types may be depicted within DAGs, using the examples of the Patient Health Questionnaire-9 (PHQ-9, representing depression severity) and the EQ-5D (representing HRQoL).

RESULTS: Reflective PROMs, like the PHQ-9, can be readily incorporated into a DAG and analyzed like other unidimensional outcomes (e.g., mortality) without special consideration. In comparison, formative PROMs, like the EQ-5D whereby five dimensions cause HRQoL, are more challenging. When a formative PROM is the causal outcome of interest, the role of each item must be explicitly considered, to ensure that all relevant external variables (e.g., confounders) are appropriately recognized and conditioned.

CONCLUSIONS: Estimating causal effects involving PROMs requires extra care. This is especially true when examining PROMs that follow a formative model. These challenges can be understood visually using DAGs. As interest in real-world evidence grows, conducting causal analyses using PROMs in observational data will become more prominent. We show how DAGs can help to inform such analyses.