Harnessing Administrative Data for Real-World Evidence in Psoriatic Disease: Opportunities, Challenges, and Methodological Considerations

Overview of administrative and EHR data

Central sources of RWD are administrative claims and EHR systems. Administrative data originate from billing systems, capturing diagnoses, procedures, and dispensed medications.¹ In the US, commercial claims databases like MarketScan or Optum offer national reach and structured formats, making them attractive for large-scale pharmacoepidemiology research.³ However, these datasets often lack clinical granularity, with the rationale for treatment decisions, disease activity, or symptom burden not typically recorded. Coding behavior may be shaped by reimbursement incentives rather than clinical precision, and patient characteristics like ethnicity are often inferred using probabilistic algorithms or are missing entirely.⁴

EHRs capture a broader spectrum of clinical data, including prescriptions written (whether filled or not), laboratory results, imaging, and clinical narratives.^1,5 A limitation with EHR data is that valuable information on symptom severity is largely recorded in unstructured free-text formats. Natural language processing tools are increasingly used to extract such information; however, there are challenges with preserving patient anonymization.⁶ Additionally, documentation practices differ widely across providers and institutions, introducing further variability.⁷

Opportunities for research using RWD

Unlike traditional clinical trials, which often exclude patients with comorbidities or concurrent treatments, RWD capture the full spectrum of individuals seen in routine practice.² This improves generalizability, enabling analysis of groups often underrepresented in trials. RWD also allow for active comparator designs that better reflect real-world treatment decisions, offering insights into relative effectiveness, safety, and tolerability. The temporal richness of these data facilitates prognostic research, enabling identification of risk factors and predictors of disease development, including the transition from psoriasis (PsO) to psoriatic arthritis (PsA).⁸ Such analyses are prone to bias and require careful interpretation, as will now be discussed in greater detail.

Methodological challenges in using RWD

Although RWD offer valuable opportunities to study treatment effects and disease progression outside trial settings, their nonrandomized nature introduces several sources of bias that must be carefully addressed.

Confounding and bias. A key issue for RWD is confounding by indication; that is, treatment choice influenced by factors such as disease severity or comorbidity, which also affects outcomes. A classic example in PsD is evaluating whether systemic therapies for PsO reduce the risk of incident PsA.⁹ Patients with more severe PsO are both more likely to receive systemic or biologic therapy and more likely to develop PsA, therefore overestimating the effect between treatment and outcome. In addition, protopathic bias, where treatments are prescribed for early symptoms of an as-yet-undiagnosed condition, can falsely suggest that therapy causes the outcome. Together, these factors complicate interpretation of treatment effects in retrospective cohorts. Propensity score methods (eg, matching or weighting) can improve comparability but depend on relevant covariates, which may be missing in claims data, leading to residual confounding.

Selection and immortal time bias. Selection bias can arise through restrictive eligibility criteria that alter the composition of the exposed or unexposed groups in relation to the outcome. For example, excluding patients with osteoarthritis or fibromyalgia when evaluating arthralgia as a predictor of PsA transition may preferentially remove individuals less likely to progress, inflating observed associations. Immortal time bias arises when follow-up time is misclassified, favoring one group by including periods during which the outcome could not occur.¹⁰ An example could include classifying patients as “biologic users” based on future treatment initiation, thereby misattributing event-free time before treatment start to the exposed group. Both types of bias are common in pharmacoepidemiology and can falsely exaggerate treatment effects if not carefully addressed.

Collider bias and paradoxical findings. Collider bias arises when analyses condition on a variable affected by both exposure and outcome, distorting associations. In PsD, this has produced “risk factor paradoxes,” where smoking may appear protective at lowering risk of PsA due to inappropriate adjustment.¹¹ The multistage nature of PsD is particularly prone to such sources of bias.

Target trial emulation: Causal inference for RWD

Target trial emulation offers a structured approach permitting causal inference from observational data, and this is particularly valuable when RCTs are unavailable.¹² This emerging methodology begins by explicitly defining the protocol of a hypothetical pragmatic trial that would address the research question of interest.¹² Key components include eligibility criteria, treatment strategies, outcomes, and duration of follow-up. Observational data are then used to emulate this target trial, with the aim of minimizing selection and immortal time bias.¹² The Randomized Controlled Trials Duplicated Using Prospective Longitudinal Insurance Claims: Applying Techniques in Epidemiology (RCT-DUPLICATE) initiative, which compared emulated trials using EHR and claims data to existing RCTs, demonstrated encouraging concordance between approaches.¹³ Although target trial emulation strengthens the validity and interpretability of real-world analyses, it is not without limitations. In conditions such as PsD, where treatment decisions are closely linked to disease severity or comorbidities, confounding by indication may violate key assumptions of the emulation framework. Whereas this approach works well for questions where treatment assignment is less influenced by prognosis, such as comparing vaccine effectiveness, its use in PsD requires caution. Complementary methods, including negative control outcomes and instruments, are being explored to address residual confounding in these more complex scenarios.¹⁴

Best practices and recommendations

To maximize the value of RWD, careful study design and methodological transparency are essential. Clearly specifying the causal question, choosing appropriate comparator groups, and predefining inclusion criteria are foundational steps. The use of directed acyclic graphs, developed in collaboration with clinicians and increasingly with patients, help clarify assumptions about causal structure, identify potential confounders, and guide appropriate covariate adjustment.¹⁵ Harmonizing datasets, particularly when linking claims and EHRs, requires robust data curation and standardization efforts, including recording disease activity and validated outcome measures.^5,16 Transparency in analytical methods, including registration of protocols and proposed sensitivity analyses, enhances reproducibility and internal validity.^17,18