Dermatology is experiencing an unprecedented therapeutic revolution. In 2025 alone, the FDA approved eight new dermatology drugs, spanning gene therapies, oral peptides, JAK inhibitors, and biologic expansions across atopic dermatitis, psoriasis, bullous pemphigoid, alopecia areata, and chronic spontaneous urticaria. Dupilumab, a molecule I was closely involved with during the pivotal Phase III clinical program that led to its original FDA approval, has now expanded into its sixth indication. One major pharmaceutical company is preparing the first-ever head-to-head trial comparing an oral pill to an injectable biologic in psoriasis.

The pipeline is full. But so is the bill.

Every one of these approvals was built on Phase III randomized controlled trials that cost tens of millions of dollars, enrolled narrowly selected patients, and took years to complete. And the hardest questions in dermatology today—comparative effectiveness across biologics, optimal sequencing strategies, real-world durability, and subpopulation-specific responses—are questions that the traditional trial paradigm is structurally unable to answer at scale.

This is where a new generation of statistical methods enters. Not to replace trials, but to answer the questions trials cannot. And the critical insight is this: prediction and causation are not the same thing. Confusing them is already costing the industry billions.

A model that predicts PASI 90 does not tell you whether the drug caused the clearance. That distinction is worth the entire evidence package behind a supplemental indication.
01 — The Problem

Why Prediction Is Necessary but Not Sufficient for Regulatory Evidence

Pharmaceutical companies are investing heavily in predictive analytics—and rightly so. Machine learning models now forecast treatment response, identify biomarker-positive subpopulations, and predict adverse events with impressive accuracy. These tools are valuable and represent a genuine leap forward. But they are, on their own, not yet sufficient for regulatory-grade evidence generation. The next step is causal inference.

The reason is structural. Predictive models learn correlations from observational data. A model trained on electronic health records can learn that patients who receive dupilumab tend to have better outcomes than patients who receive methotrexate. But that association is contaminated by indication bias, time-varying confounding, and selection effects. The model cannot distinguish the drug’s causal contribution from the clinician’s prescribing behavior, the patient’s disease trajectory, or the timing of treatment initiation.

Prediction answers: “What is likely to happen?”

Causation answers: “What would happen if we intervened?”

Only the second question is relevant for drug development. Only the second question can support a label expansion, a comparative effectiveness claim, or a payer negotiation. And only the second question requires causal inference methods.

The Regulatory Reality

In December 2025, the FDA finalized updated guidance on the use of real-world evidence for regulatory decision-making, explicitly allowing sponsors to submit de-identified data without requiring identifiable patient records. The agency also signaled intent to extend this flexibility to drugs and biologics. The door to RWE-based submissions has never been wider. But walking through it requires more than prediction—it requires causal rigor.

02 — The Numbers

What Dermatology Drug Development Actually Costs

The financial case for causal real-world evidence in dermatology is overwhelming. Consider what the current paradigm demands:

$11.5M Average cost of a Phase III
dermatology trial
3–5 yrs Timeline for a pivotal
Phase III program
8 New derm drugs
approved by FDA in 2025

But those are just the averages. Biologics programs in immunodermatology regularly cost far more. When the question is not “does this drug work?” but “does this drug work better than the three other biologics already on the market, in specific patient subgroups, over real-world timelines?”—the RCT model breaks down. You would need dozens of head-to-head trials, each costing tens of millions, each taking years, each enrolling only the narrow population that met inclusion criteria.

Meanwhile, the data to answer these questions already exist. Dermatology registries, electronic health records, claims databases, and research platforms contain millions of patient trajectories with treatment exposures, outcomes, comorbidities, and genomic data. What is missing is not data. What is missing is the causal architecture to make that data speak truthfully.

Dimension Phase III RCT Causal RWE Study
Cost per question $11.5M–$50M+ $1M–$5M
Timeline 3–5 years 6–18 months
Patient population Narrow (excludes 70–80% of real patients) Real-world, diverse
Comparisons possible 1 (drug vs. placebo or single comparator) Multiple simultaneous
Confounding control Randomization Causal design + statistical adjustment
Regulatory acceptance Gold standard Growing (7.3% of FDA labeling expansions included RWE in dermatology, 2022–2024)
03 — The Methods

Prediction + Causation: The Two-Engine Architecture

The breakthrough is not choosing between prediction and causation. It is understanding that they serve different functions in the same evidence pipeline, and that prediction alone, as a standalone method, cannot extract causal inference from real-world evidence. You need both engines, working in sequence.

1

Prediction Engine forecasting

Machine learning and predictive models identify patient subgroups, forecast treatment response probabilities, detect biomarker signatures, and define the landscape. This is the where to look engine. It narrows the search space. But its outputs are associational, not causal.

2

Causation Engine intervention

Causal inference methods then take over to answer the interventional question: if we assigned this treatment versus that treatment, what would the outcome be? This is the what would happen if engine. New statistical frameworks—developed over the past two decades and now reaching operational maturity—make this possible with observational data.

The methods that power the causation engine include approaches that are, in some cases, less than a decade old in their applied form:

A

Target Trial Emulation

A framework for designing observational studies that mirror the protocol of a hypothetical randomized trial. Publications using this approach have grown exponentially, with 2025 already matching the total output of 2023. In dermatology, this has been applied to compare biologics for psoriasis using the British Association of Dermatologists registry.

B

Parametric G-Formula

A simulation-based method for estimating the effects of sustained treatment strategies over time—precisely the kind of question dermatologists face when sequencing biologics or comparing dose optimization protocols. Unlike traditional regression, the G-formula appropriately handles the feedback loop between treatment decisions and time-varying patient status.

C

Causal Bayesian Networks

A graphical framework for encoding assumptions about causal relationships between variables and testing those assumptions against data. Where traditional statistics asks “what is the adjusted association,” Bayesian networks ask “given this causal structure, what is the interventional effect?” This distinction is fundamental, and it is new to most pharmaceutical evidence teams.

D

Clone-Censor-Weight Methodology

A technique for handling the comparison of treatment strategies that unfold over time, such as early vs. delayed biologic initiation, without the immortal time bias that has invalidated many observational studies in dermatology.

These methods are not theoretical. They are being applied to real patient cohorts, in real regulatory contexts, right now. I am actively implementing several of them in my dissertation research.
04 — The Dermatology Case

Five Questions in Immunodermatology That Only Causal Methods Can Answer

The practical value of causal RWE becomes concrete when you look at the questions the field is currently unable to answer through trials alone:

1

Comparative Biologic Effectiveness by Subgroup

How does secukinumab compare to ustekinumab in patients with moderate-to-severe psoriasis and metabolic syndrome? An RCT would take years and exclude most real patients. Target trial emulation on registry data has already been demonstrated for this exact comparison in JAMA Dermatology.

2

Optimal Sequencing After Treatment Failure

When a patient fails dupilumab for atopic dermatitis, what should come next—tralokinumab, abrocitinib, or upadacitinib? This is a sustained treatment strategy question that the parametric G-formula was designed to answer.

3

Long-Term Safety Signals Across Populations

What is the cardiovascular risk profile of JAK inhibitors in dermatology patients who differ from the rheumatology populations in which safety concerns first emerged? Causal methods can disentangle drug effects from confounding by indication across therapeutic areas.

4

Label Expansion to Underrepresented Populations

2025 marked a shift toward inclusive clinical trials, but most pivotal data still comes from non-diverse cohorts. Causal RWE can generate the subgroup-specific evidence needed for regulatory confidence in skin-of-color populations, pediatric patients, and the elderly.

5

Genomic Effect Modification

Do patients with specific genetic variants respond differently to IL-17 versus IL-23 inhibitors? Answering this question requires integrating genomic data with causal treatment effect estimation—an intersection where very few teams are currently working.

05 — The Regulatory Landscape

The FDA Door Is Open—But Not for Correlation

Between 2022 and 2024, approximately one-quarter of all FDA labeling expansion approvals included some form of real-world evidence. Dermatology ranked third among therapeutic areas by volume of RWE-supported approvals. The December 2025 FDA guidance update removed the requirement that sponsors submit identifiable patient data—a barrier that had effectively locked out large de-identified databases.

But the guidance is also clear: the bar for methodological rigor is going up, not down. The FDA wants to see explicit causal frameworks, transparent assumptions (mapped via directed acyclic graphs), prespecified analysis plans that control for time-varying confounding, and sensitivity analyses that stress-test those assumptions.

Prediction models, no matter how accurate, do not meet these criteria. A gradient-boosted model that predicts PASI 90 with 85% AUC is a useful clinical tool. It is not a regulatory submission. The FDA is not asking “can you predict who will respond?” It is asking “does your evidence support the claim that this drug causes the response in this population?”

Where This Is Heading

The companies that build causal RWE capability now will have a structural advantage in supplemental indication filings, payer negotiations, and comparative effectiveness positioning. The methods exist. The regulatory appetite exists. The data exist. What is missing, in most organizations, is the interdisciplinary expertise to connect them.

06 — What I Am Working On

Bridging Clinical Dermatology and Causal Statistical Science

I come to this question from both sides of the table. As a board-certified dermatologist with over a decade of clinical experience and as a veteran of the Phase III dupilumab clinical trial program that led to Dupixent’s FDA approval, I understand the evidentiary standards from the inside. As a PhD candidate at Icahn School of Medicine at Mount Sinai—under Dr. Bruce Darrow (cardiology) and Dr. Pei Wang (biostatistics)—I am actively implementing the causal methods described in this paper on real patient data.

My dissertation work applies the parametric G-formula, clone-censor-weight methodology, and causal Bayesian networks to a cohort of nearly 25,000 participants from the NIH All of Us Research Platform, integrating electronic health records, survey data, and genomic information. While the current application is in cardiovascular prevention, the methodological framework is directly transferable to immunodermatology—and that is precisely where I intend to take it.

I am currently developing applied protocols for causal RWE in dermatology and immunology. If your team is working on comparative effectiveness, label expansion strategy, or real-world evidence generation for biologics in dermatology, I would welcome the conversation.

Interested in Causal RWE for Dermatology?

I am preparing a detailed technical brief on the application of these methods to immunodermatology evidence generation. Details are forthcoming. To be notified when it is available, or to discuss how causal inference methods could apply to your evidence strategy, reach out directly.

Contact Dr. Benzzi

References

[1] Moore TJ, Zhang H, Anderson G, Alexander GC. Estimated Costs of Pivotal Trials for Novel Therapeutic Agents Approved by the US FDA, 2015–2016. JAMA Intern Med. 2018;178(11):1451–1457.
[2] Sertkaya A, Wong HH, Jessup A, Beleche T. Key cost drivers of pharmaceutical clinical trials in the United States. Clin Trials. 2016;13(2):117–126.
[3] Deng YF, Girman CJ, Ritchey ME. Real-World Evidence in FDA Approvals for Labeling Expansion of Small Molecules and Biologics. Ther Innov Regul Sci. 2025;59(5):982–992.
[4] U.S. Food and Drug Administration. Use of Real-World Evidence to Support Regulatory Decision-Making for Medical Devices. Guidance Document, December 18, 2025.
[5] U.S. Food and Drug Administration. FDA Eliminates Major Barrier to Using Real-World Evidence in Drug and Device Application Reviews. Press Release, December 15, 2025.
[6] Hernan MA, Robins JM. Using big data to emulate a target trial when a randomized trial is not available. Am J Epidemiol. 2016;183(8):758–764.
[7] Robins JM. A new approach to causal inference in mortality studies with a sustained exposure period. Math Modelling. 1986;7(9–12):1393–1512.
[8] Pearl J. Causality: Models, Reasoning, and Inference. 2nd ed. Cambridge University Press; 2009.
[9] Yiu ZZN, Mason KJ, Hampton PJ, et al. Randomized Trial Replication Using Observational Data for Comparative Effectiveness of Secukinumab and Ustekinumab in Psoriasis. JAMA Dermatol. 2021;157:66–73.
[10] Dermatology Times. 2025 Year in Review: Drug Approvals. Published February 2026.
[11] MedCentral. 8 Dermatology Drugs FDA Approved in 2025. Published December 2025.
[12] Honap S, et al. Target Trial Emulation: Improving the Quality of Observational Studies in Inflammatory Bowel Disease. Inflamm Bowel Dis. 2025.
[13] Janjua F, et al. Methodologies for the Emulation of Biomarker-Guided Trials Using Observational Data: A Systematic Review. J Pers Med. 2025;15(5):195.
[14] Lipkovich I, et al. Estimating Heterogeneous Treatment Effects With Target Trial Emulation: A Checklist of Causal Machine Learning for Observational Data. Chest. 2025.
[15] Munoz L, et al. Review of tools to support Target Trial Emulation. J Biomed Inform. 2025.
H

Dr. Hafsa Benzzi

Board-certified dermatologist with 10+ years of clinical experience. Veteran of the Phase III dupilumab trials behind Dupixent’s FDA approval. PhD candidate at Icahn School of Medicine at Mount Sinai, researching causal Bayesian networks and parametric G-formula methods for real-world evidence generation. Multilingual (English, French, Arabic). Founder of Causal Threads™.