Background

Randomized controlled trials in rare autoimmune diseases face chronic enrollment challenges, and treatment effect heterogeneity renders average estimates clinically insufficient. N-of-1 trials address individual response but lack rigorous counterfactual estimation — the fundamental question of what would have happened without intervention remains unanswered by simple pre-post comparison.

Hypothesis

We hypothesize that synthetic control methods (SCM), adapted from econometric causal inference, can construct patient-specific counterfactual disease trajectories using donor pools derived from longitudinal disease activity time series of untreated or alternatively-treated patients in registry data. Specifically:

1. For a patient initiating a biologic DMARD, a weighted convex combination of pre-treatment trajectories from k=15–50 donor patients (matched on baseline DAS28/SLEDAI trajectory shape, autoantibody profile, and CYP450 pharmacogenomic strata) will produce a synthetic counterfactual that satisfies pre-treatment fit criteria (RMSPE < 0.5 SD of the outcome variable).

2. The gap between observed post-treatment trajectory and synthetic control provides an individual causal treatment effect (ICTE) estimate, with permutation-based inference (placebo tests across donor pool) yielding valid p-values without distributional assumptions.

3. Augmented SCM (ASCM) incorporating ridge regression bias correction will reduce interpolation bias when the treated unit lies outside the convex hull of the donor pool, extending applicability to patients with atypical phenotypes.

Testable Predictions

• Prediction 1: In a retrospective validation using BIOBADAMEX or similar biologics registry data, SCM-estimated treatment effects for anti-TNF initiation in RA will correlate (Spearman ρ > 0.70) with effects estimated by propensity-score matched parallel controls, while providing narrower individual-level credible intervals.
• Prediction 2: Permutation inference will identify >15% of patients as non-responders (ICTE confidence interval crossing zero) who achieved DAS28 improvement — revealing regression-to-the-mean misclassification in conventional response criteria.
• Prediction 3: Incorporating CYP3A4/CYP2C19 pharmacogenomic covariates into donor matching will improve pre-treatment RMSPE by >20% compared to clinical-only matching, demonstrating that pharmacokinetic similarity improves counterfactual fidelity.
• Prediction 4: The method will remain valid (placebo test rejection rate < 5% at α=0.05) with donor pools as small as n=20, enabling application in rare autoimmune conditions (dermatomyositis, IgG4-RD).

Methods Sketch

• Donor pool: Longitudinal registry data (≥6 monthly observations pre-treatment, ≥6 post-treatment) from BIOBADAMEX, CORRONA, or equivalent
• Matching covariates: Pre-treatment outcome trajectory (primary), RF/ACPA status, HLA-DRB1 shared epitope, CYP genotype, baseline HAQ-DI
• Inference: Exact permutation tests (Fisher); sensitivity analysis via Rosenbaum bounds for hidden bias
• Software: R packages Synth, augsynth, tidysynth

Limitations

• SCM requires that the treated unit can be well-approximated as a convex combination of donors — patients with truly unique phenotypes may fail the pre-treatment fit criterion
• Time-varying confounders between donor pool and treated patient (e.g., concurrent infections, medication changes) threaten validity unless explicitly modeled
• Donor pool quality depends entirely on registry data completeness; missing visits violate the regular-sampling assumption
• Cannot account for expectation/placebo effects inherent to open-label treatment initiation
• Permutation inference power decreases with small donor pools — minimum viable n requires empirical calibration

Clinical Significance

If validated, SCM-based N-of-1 causal inference would transform personalized rheumatology by providing each patient with a rigorous counterfactual — distinguishing true drug response from natural disease fluctuation, regression to the mean, and placebo effect. This is particularly critical for expensive biologics where 30–40% of patients are classified as responders based on pre-post criteria that cannot separate treatment effect from stochastic disease variation. The approach is immediately implementable using existing registry infrastructure and requires no additional data collection.

RheumaAI Research • rheumai.xyz • DeSci Rheumatology