Avoiding methodological pitfalls in assessing cardiovascular and thromboembolic risk in dermatology

Florence Poizeau Catherine Droitcourt The debate on cardiovascular and thromboembolic safety of systemic dermatologic therapies is central to clinical practice. Atopic dermatitis and psoriasis illustrate why: they are chronic diseases, not life-threatening in themselves, yet treated over the long term with a wide range of highly effective drugs. This combination—chronicity, non-lethality and therapeutic abundance—places therapeutic decisions at the crossroads of efficacy and safety. Clarifying whether reported associations with cardiovascular and thromboembolic events represent true risks or methodological artefacts is therefore essential. Kridin et al. recently used the TriNetX database to evaluate the safety of JAK inhibitors in atopic dermatitis.1 In propensity score–matched analyses comparing JAK inhibitors with dupilumab, cyclosporine and methotrexate, they suggested an increased risk of venous thromboembolism compared to dupilumab and methotrexate. No significant differences were observed for myocardial infarction or stroke. This large-scale study illustrates the power of pharmacoepidemiology to address clinically important safety questions, while also highlighting the methodological challenges inherent to such research. How exposure is defined is central because it determines who enters the study and how follow-up time is classified. To construct mutually exclusive cohorts, Kridin et al. excluded patients with any past or future exposure to the other study drugs. Because JAK inhibitors are rarely first-line systemic therapies,2 this likely selected a JAK inhibitor cohort enriched with patients who had contraindications to cyclosporine, methotrexate and dupilumab, potentially yielding a baseline risk profile different from comparators—conditions that favour indication bias.3 In addition, by defining exposure groups based on the entire follow-up, any patient who switched to another treatment was automatically excluded. This approach introduces conditioning on the future and may affect the estimated cardiovascular risk by removing those who switched because of adverse events. For instance, a patient developing hypercholesterolemia under a JAK inhibitor would be expected to switch treatment and would thus be excluded from the analysis, although precisely such patients are informative for safety evaluation. Finally, the study assumed continuous exposure to each systemic treatment up to 3 years, regardless of discontinuation. Since median treatment durations were much shorter, many patients were inevitably misclassified as exposed. The stratified analysis by time since initiation is informative: the 1- to 3-year stratum probably includes many off-treatment patients, complicating interpretation. Beyond exposure, two other elements critically shape validity: outcome definition and confounding. By not including ICD codes for phlebitis (I801-9), the study may have missed cases of deep vein thrombosis.4 Defining outcomes too narrowly reduces sensitivity and power, while too broad definitions dilute associations. Even when code lists are provided, clinicians unfamiliar with coding systems may struggle to judge whether choices are optimal. Confounding is also a key concern. Cardiovascular risk factors were considered at baseline, which is essential. However, residual confounding remains, because important modifiers such as anticoagulant and antiplatelet therapy, major surgery and active cancer were not considered. Over several years of follow-up, the absence of time-updated covariates further limits adjustment, since hypertension or hyperlipidemia, for example, may arise after baseline and alter cardiovascular risk. Disease severity represents another potential confounder: severe atopic dermatitis itself may be linked to cardiovascular outcomes,4 yet severity measures are generally absent from administrative databases. Line of therapy can serve as a proxy, though imperfectly. Looking ahead, the challenge is producing reliable pharmacoepidemiological estimates of treatment risk in dermatology (Figure 1). Methodological approaches such as propensity scores or target trial emulation are often presented as approximating randomized trials, yet their validity depends less on the label than on the assumptions and variables underpinning them. Propensity score methods, widely used for nearly two decades, can balance measured covariates between groups, but—like conventional multivariable regression—their performance depends entirely on the variables included.5, 6 More recently, target trial emulation has gained prominence, again with the ambition of mimicking randomized trials. This framework prevents immortal time bias if time zero is correctly defined and if conditioning on the future is avoided.7 Yet it presumes sustained adherence, an assumption not consistently met in dermatology where treatment switching and discontinuation are common. In such settings, pragmatic approaches include censoring follow-up at discontinuation (a per-protocol strategy),8 modelling exposure as time-varying to reflect treatment changes, or using flexible parametric models that accommodate varying exposure patterns.9 Beyond design, two additional priorities are harmonization of outcome definitions and reproducibility across cohorts.10 Developing validated definitions of cardiovascular and thromboembolic events and replicating findings across national or international databases, will be essential to provide dermatologists with robust and interpretable estimates of treatment safety. None. F Poizeau received consulting fees from Bristol Myers Squibb, Novartis and Leo Pharma; speaker honoraria from Novartis; and support for attending meetings and/or travel from UCB, Janssen and Novartis. C Droitcourt has received speaker honoraria from Sanofi-Genzyme, Almirall, Pfizer, AbbVie and Eli Lilly and support for attending meetings and/or travel from Leo Pharma, AbbVie and Sanofi-Genzyme. Not applicable. Not applicable. Data sharing is not applicable to this article as no new data were created or analysed in this study.