Multiple Sclerosis — Bayesian Causal Atlas

Read first — scope & honesty. This is a quantitative research instrument, not medical advice. MS therapy is individualized and risk-stratified by a neurologist. Two rules govern every number below: (1) onset risk factors and disease-course interventions are not interchangeable and cannot be multiplied together — they act on different endpoints in different populations; (2) disease-modifying therapies (DMTs) are mutually exclusive — a patient takes one at a time, so their hazard ratios cannot be stacked. The engine enforces both constraints by design.

I. Experiment — Question, endpoints, populations

Causal question. Across all interventions for which published hazard ratios, mechanisms of action, and dose–response data exist, what is the identifiable causal contribution of each to a reduction in the two principal MS endpoints, after removing confounding via backdoor adjustment and removing shared-pathway double-counting via cross-correlation shrinkage?

Stratum	Population	Primary endpoint	Anchor (untreated)	Manipulable nodes
Etiology (prevention)	EBV-naïve / general young-adult population	Incident clinically-definite MS	Lifetime incidence ~0.3%; >99% of cases EBV-seropositive	EBV exposure, smoking initiation, adolescent obesity, serum 25(OH)D
Disease course (treatment)	Diagnosed relapsing MS (RRMS)	24-month confirmed disability progression (CDP)	~29% (AFFIRM placebo)²	One DMT (mutually exclusive), smoking cessation; relapse axis: 25(OH)D

Secondary endpoints tracked qualitatively: annualized relapse rate (ARR), new/enlarging T2 & gadolinium-enhancing MRI lesions, all-cause mortality (smokers carry elevated premature-mortality HR⁵).

II. Method — The structural causal model

II.1 The causal DAG

The directed acyclic graph below encodes the assumed causal architecture. The single most important structural feature: EBV infection is a necessary-but-not-sufficient permissive node sitting upstream of disease onset — it gates the entire disease, which is why nearly 100% of MS patients are EBV-seropositive, yet most EBV-infected people never develop MS.¹

Green = permissive gate. Gold = manipulated node do(·). Red = outcome. Dashed = adjusted confounding/secondary path. Edges into disability are deconfounded by the backdoor set {age, sex, baseline EDSS, disease duration, prior relapse activity}.

II.2 Estimators & assumptions

Operation	Method	Implementation here
Confounding control	Backdoor criterion (Pearl)	Trial HRs already randomized; observational HRs (vit D, smoking) adjusted for age/sex/EDSS in source papers
Effect aggregation	Multiplicative hazards on log scale	logHR_total = Σ logHR_i, then shrunk for shared pathways
Cross-correlation removal	Shared-variance shrinkage, default ρ̄ = 0.30	Secondary terms deflated by (1−ρ̄) to avoid double-counting the common neuroinflammatory pathway
Robustness to unmeasured confounding	E-value (VanderWeele & Ding)	Computed live per intervention & for the combined estimate
Population impact	Population Attributable Fraction (PAF)	Etiology panel, Levin's formula
Dose–response saturation	Monotone saturating curve	Vitamin D 25(OH)D, plateau ~100–125 nmol/L
Minimum-effective set	Pareto frontier (effect vs burden/risk)	Ranked table §V

III. Result — Verified effect sizes (the evidence base)

III.1 Disease-course interventions — diagnosed RRMS, endpoint = confirmed disability progression

Hazard ratios < 1 favour treatment. A randomized pivotal trial · B trial secondary/active-comparator · C observational/real-world.

Intervention	Trial	ARR reduction	CDP hazard ratio (95% CI)	Grade
Natalizumab	AFFIRM²	~68%	0.58 (0.43–0.77)	A
Ocrelizumab	OPERA I/II³	~46–47% vs IFN	~0.60 (40% RRR, 96 wk)	A
Alemtuzumab	CARE-MS II³	~49% vs IFN	~0.58	B
Cladribine	CLARITY⁸	~57.6%	0.67 (0.48–0.93)	A
Fingolimod	FREEDOMS⁷	~54%	0.70 (≈30% RRR, 6 mo)	A
Dimethyl fumarate	DEFINE⁶	~50%	0.62 (38% RRR, 12 wk)	A
Teriflunomide	TEMSO⁶	~31%	~0.74	A
Interferon β / Glatiramer	multiple	~30%	~0.82 (modest, inconsistent)	A
Ocrelizumab (PPMS)	ORATORIO³	n/a (progressive)	0.76 (24% RRR, 12 wk)	A

III.2 Modifiable lifestyle factors — both strata

Factor	Endpoint	Effect (95% CI)	Source	Causal status
EBV seroconversion	MS onset	HR 32.4 (4.3–245)	Bjornevik 2022¹	Necessary, not sufficient; wide CI
Current smoking	MS onset	OR ≈ 1.5 (+50%)	meta-analysis⁵	Causal (dose-dependent)
Smoking (continued)	Progression	HR 1.55 (1.10–2.19)	meta-analysis⁵	Causal; cessation modifiable
Low 25(OH)D (lowest vs highest)	MS onset	OR 0.68 (0.50–0.93); −62% top vs bottom	Munger 2006; Salzer⁴	Strong association; RCTs null for hard endpoints
25(OH)D, +10 nmol/L	Relapse (obs.)	−6.7% relapse risk	real-world cohort⁴	Observational only
Adolescent obesity	MS onset	OR ≈ 2	cohort⁵	Likely causal (mediated partly by vit D / inflammation)
HLA-DRB1*15:01	MS onset	OR ≈ 3	GWAS	Non-modifiable (shown for completeness)

The vitamin D antithesis (Commandment 7). The observational signal is among the most reproducible in MS epidemiology, yet the two largest randomized add-on trials (SOLAR; high-dose D3 + interferon) failed to move hard disability endpoints.⁴ Interpretation under do-calculus: low vitamin D may be a marker on the causal path (confounded by UVB exposure, adiposity, outdoor activity) rather than a fully manipulable cause for established disease. The engine therefore confines vitamin D to the relapse axis and refuses to let it inflate the disability ceiling.

IV. Interactive engine — build your scenario

Select one DMT (radio — the mutual-exclusivity constraint), set smoking status, and tune the vitamin D target. The engine computes the deconfounded, cross-correlation-adjusted ceiling reduction in 24-month disability progression, with E-values and number-needed-to-treat. Then print a custom report.

Disease-course module — 24-month confirmed disability progression

anchor untreated risk 29%

1 · Disease-modifying therapy (choose one)

2 · Modifiable lifestyle

Currently a smoker
Continued smoking multiplies progression hazard ×1.55 — cessation removes it.⁵

Serum 25(OH)D target 100 nmol/L

Acts on the relapse axis only (observational). Saturation modeled ~100–125 nmol/L.⁴

Evidence tier → available DMTs: A–D

All DMTs are pivotal phase-3 RCTs (A) except older interferon-β/glatiramer (B, inconsistent). Below-tier DMTs are disabled; selection falls back to No-DMT.

Cross-correlation ρ̄ (shared-pathway shrinkage) 0.30

Deflates secondary terms so the common neuroinflammatory pathway isn't double-counted.

Result — deconfounded ceiling

Combined CDP hazard ratio

—

upper-bound ceiling

Absolute risk ↓ (24 mo)

—

vs 29% untreated

Number needed to treat

—

to prevent 1 progression

E-value (combined)

—

unmeasured-confounder robustness

The combined HR is an identification ceiling under the model's independence-after-shrinkage assumption, not an individual prognosis. DMT effect sizes derive from distinct trial populations and cannot be added to one another; only one DMT term ever enters the product.

IV.1 Vitamin D dose–response (saturation)

Modeled relative relapse-risk multiplier vs serum 25(OH)D, monotone and saturating near 100–125 nmol/L. The marker tracks your slider. Curve is illustrative of the observational dose–response⁴; it does not represent a randomized disability effect.

V. Pareto minimum-effective set

Ranking disease-course options by effect against treatment burden & risk. The Pareto-efficient frontier (no option beats them on both axes) is highlighted.

High-efficacy monoclonals (natalizumab, ocrelizumab, alemtuzumab) dominate on effect but carry distinct risk tails — natalizumab's PML risk is JC-virus-serostatus dependent; alemtuzumab's secondary autoimmunity. Cladribine and the S1P/fumarate orals occupy the efficient middle for many patients. The "smallest set that achieves the goal" is typically one appropriately matched high- or moderate-efficacy DMT plus smoking cessation — not a stack.

VI. Curiosity — second-order implications

EBV as a do-target

If EBV is necessary, an effective prophylactic EBV vaccine administered pre-seroconversion could in principle bend MS incidence toward the floor set by the rare seronegative cases — a population effect no DMT can touch. Anti-CD20 efficacy (which depletes EBV-harboring memory B-cells) is mechanistically consistent.¹

The relapse–progression decoupling

High-efficacy DMTs crush relapses and MRI activity yet only partially slow progression — "progression independent of relapse activity" (PIRA) implies a second, smouldering causal mechanism the current do-targets address incompletely.

Confounded vitamin D

The vitamin D / UVB / latitude / adiposity tangle is a textbook backdoor problem. Randomization (SOLAR) opened the backdoor-free path and the effect shrank — a caution against treating every robust association as a lever.

VII. Antithesis — where this analysis could be wrong

Challenge	Why it threatens the conclusion	Mitigation in this model
Cross-trial HR comparison	DMT HRs come from trials with different placebo/comparator arms, eras, and populations; ranking them as if commensurable is fragile	Only one DMT enters the estimate; rankings flagged as indicative; active-comparator trials labeled grade B
EBV HR imprecision	95% CI 4.3–245 is enormous; unadjusted for vitamin D & BMI per published critique	EBV treated as a structural gate (PAF logic), never multiplied into individual risk
Multiplicativity assumption	Log-additive hazards may overstate combined effect if pathways overlap	ρ̄ shrinkage on secondary terms; result framed strictly as a ceiling
Survivorship / adherence	Real-world effect < trial effect (discontinuation, intolerance)	Trial HRs are best-case; NNT shown to expose absolute scale

Appendix A — Computation reference

combined_logHR = log(HR_dmt) + (1 − ρ̄) · log(HR_smoking_if_continued)
HR_combined = exp(combined_logHR)
risk_treated = risk_anchor · HR_combined  (approx., rare-outcome hazard→risk)
ARR_abs = risk_anchor − risk_treated  ;  NNT = 1 / ARR_abs
E_value(HR<1): RR* = 1/HR ; E = RR* + sqrt(RR* · (RR* − 1))
PAF (Levin) = Pe·(RR−1) / (1 + Pe·(RR−1))
vitD_multiplier(x) = 1 − 0.32·(1 − exp(−(max(x−25,0))/45))  (saturating, relapse axis)

Appendix B — References

Bjornevik K, et al. Longitudinal analysis reveals high prevalence of Epstein-Barr virus associated with multiple sclerosis. Science 2022;375:296–301. HR 32.4 (95% CI 4.3–245).
Polman CH, et al. A randomized, placebo-controlled trial of natalizumab for relapsing MS (AFFIRM). N Engl J Med 2006;354:899–910. Disability HR 0.58 (0.43–0.77); ARR −68%.
Hauser SL, et al. Ocrelizumab vs interferon beta-1a in relapsing MS (OPERA I/II). N Engl J Med 2017;376:221–234. Montalban X, et al. ORATORIO (PPMS) CDP HR 0.76.
Munger KL, et al. Serum 25-hydroxyvitamin D and risk of MS. JAMA 2006;296:2832–8. SOLAR & high-dose D3 add-on RCTs — null on hard endpoints. Real-world cohort: −6.7% relapse risk per 10 nmol/L.
Systematic review/meta-analysis of modifiable risk factors: smoking progression HR 1.55 (1.10–2.19); onset OR ≈1.5. Mult Scler & related.
Gold R, et al. DEFINE — dimethyl fumarate. N Engl J Med 2012;367:1098–1107. ARR −53%, 12-wk CDP −38%. TEMSO — teriflunomide.
Kappos L, et al. FREEDOMS — fingolimod. N Engl J Med 2010;362:387–401. ARR −54%; CDP HR ≈0.70.
Giovannoni G, et al. CLARITY — cladribine. N Engl J Med 2010;362:416–426. ARR −57.6%; 3-mo sustained progression HR 0.67 (0.48–0.93).

Longevity Research Initiative — Bayesian Causal Atlas. Structural Causal Model methodology after Judea Pearl (do-calculus, backdoor/frontdoor, counterfactuals), E-values after VanderWeele & Ding. All hazard ratios are published values cited above; no effect sizes were fabricated. Outputs are upper-bound ceilings for research, not clinical guidance. Generated .