A Judea Pearl Structural Causal Model for the world's leading infectious killer — and the disease with arguably the largest single causal effect in medicine: untreated active TB kills roughly 70% over ten years, while a 6-month drug regimen cures ~85–90%. The model is a cascade — prevent, treat by resistance profile, support — and the headline lever for drug-resistant TB is a regimen swap that nearly doubles survival. All effect sizes are real published values; aggregation is an upper-bound ceiling.
Causal question. Across TB interventions with published mortality, cure, and progression effects, what is the identifiable causal contribution of each to a reduction in TB death — conditional on disease state (latent, active drug-susceptible, active drug-resistant) and HIV status — after backdoor confounding control?
| State | Population | Endpoint | Anchor (untreated) | Decisive lever |
|---|---|---|---|---|
| Latent TB | Infected, not yet diseased (~1 in 4 humans ever infected) | Progression to active TB | ~10% lifetime (30–50% if HIV+) | TB preventive therapy; BCG |
| Active DS-TB | Drug-susceptible active disease | 10-yr case fatality | ~70% smear+; ~20% smear− | 6-month first-line regimen (HRZE/HR) |
| Active MDR/RR-TB | Rifampicin/multidrug-resistant disease | Treatment success vs death | Untreated ≥ DS-TB; old SoC success only 52% | BPaLM (6-month, all-oral) — 89% success |
HIV co-infection raises both progression and case fatality; antiretroviral therapy is modelled as a modifier. ~10.8 million cases and ~1.23 million deaths per year; the world's leading cause of death from a single infectious agent.
TB is a cascade: infection may progress to disease; disease, treated with a regimen matched to the resistance profile, resolves or kills. Two do(·) points dominate — prevention (stopping progression) and effective treatment (converting a ~70% killer into a ~90% cure). Resistance status is the switch that decides which regimen works.
| Operation | Method | Implementation here |
|---|---|---|
| Confounding control | Backdoor criterion (Pearl) | Regimen success rates from RCTs/cohorts; untreated CFR from pre-chemotherapy natural-history studies |
| State conditioning | do(·) effect varies by state | The same regimen has different effects on latent (prevention) vs active (cure) vs resistant (regimen-dependent) TB |
| HIV modification | Effect-measure modifier | HIV raises progression and case fatality; ART is modelled as a partial reversal (labelled approximate) |
| Population impact | Attributable fraction (Levin) | Most TB deaths are attributable to lack of timely effective treatment — a treatment-access PAF, not a biology PAF |
| Robustness | E-value (VanderWeele & Ding) | Computed for the dominant treatment effect |
| Minimum-effective set | Pareto frontier (threshold slider) | Interventions ranked by relative mortality / progression reduction (Section IV) |
| Intervention / state | Effect | Source / grade |
|---|---|---|
| Untreated active TB (smear+), 10-yr | ~70% case fatality (53–86%) | Tiemersma 2011, PLOS One; WHO B |
| Untreated active TB (smear−, culture+) | ~20% case fatality | Tiemersma 2011; WHO B |
| Standard 6-month regimen (DS-TB) | ~85% treatment success | WHO global; cohort 74–95% A |
| BPaLM (MDR/RR-TB, 6-month all-oral) | 89% success vs 52% old SoC | TB-PRACTECAL; ZeNix; Nix-TB 90% A |
| TB preventive therapy (latent) | ~60% reduction in progression (RR ≈ 0.40) | Cochrane IPT/3HP A |
| BCG vaccination | ~50% vs disease; ~73% vs severe childhood TB | Trunz 2006 meta-analysis B |
| ART in HIV-associated TB | ~56–68% mortality reduction (low CD4) | SAPIT / CAMELIA A |
The defining fact: the gap between treated and untreated active TB is one of the largest causal effects in all of medicine — roughly 70% mortality collapses to single digits. For resistant TB the decisive modern lever is the regimen itself: BPaLM raised success from 52% to 89%.
TB interventions ranked by relative mortality / progression reduction versus their no-treatment baseline. Drag the threshold to set a minimum: interventions at or above the line form the minimum-effective set. Effective treatment of active disease dominates everything — but prevention reaches people before they ever reach the clinic.
Few interventions move an outcome from ~70% mortality to single digits. TB treatment does — which means most TB deaths are not failures of biology but failures of access, diagnosis, and completion. The do(·) is a functioning health system.
MDR-TB is partly created by incomplete treatment. BPaLM's 89% success is transformative, but the upstream lever is preventing resistance through complete first-line cure — a causal loop the model makes visible.
Treating active TB saves a life already at risk; TPT and BCG act on people before disease. With ~1.2 million deaths a year, even modest preventive-therapy coverage in latent and HIV-positive populations compounds over time.
HIV converts latent TB from a ~10% lifetime risk into a yearly one and raises case fatality; integrated TB–HIV care with ART is therefore not an add-on but a core mortality lever.
| Challenge | Why it threatens the conclusion | Mitigation here |
|---|---|---|
| Untreated CFR is historical | The ~70% figure comes from pre-chemotherapy-era cohorts; populations and comorbidity differ today | Graded B; presented as the anchor with a user-adjustable range (20–86%) |
| Success ≠ survival | "Treatment success" bundles cure + completion; death is only part of the unfavorable fraction | Treated outcome modelled as the death component, not 1 − success, and labelled |
| HIV/ART multipliers are approximate | The exact mortality multiplier for HIV and its reversal by ART vary by CD4 and timing | Labelled approximate; ART benefit anchored to the SAPIT/CAMELIA range, flagged |
| BCG efficacy varies | BCG protection ranges widely by latitude and prior exposure | Reported as a range (≈50% disease, ≈73% severe childhood), graded B |
| Programmatic vs trial | Trial success (BPaLM 89%) exceeds routine programmatic outcomes | Framed as an upper-bound ceiling; old-SoC 52% shown alongside as the real-world contrast |
active untreated = anchorCFR × hivMult ; latent untreated = progRisk × hivProgMult
treated death: DS standard ≈ 0.05 ; MDR BPaLM ≈ 0.06 ; MDR old SoC ≈ 0.15 (×hivMult×artMult)
latent + TPT: progression × 0.40 (≈60% reduction)
hivMult = HIV ? (ART ? 1.15 : 1.6) : 1.0 ; hivProgMult = HIV ? (ART ? 1.8 : 4.0) : 1.0
ARR = untreated − treated ; RRR = ARR / untreated ; E-value(RR) = RR* + √(RR*·(RR*−1)), RR* = 1/RR
Worked example: active DS-TB, smear+ anchor 70%, standard regimen, HIV− → untreated 70%, treated 5%, ARR 65 pp, RRR 93%. MDR-TB old SoC death 15% vs BPaLM 6% → modernising the regimen alone averts 9 deaths per 100 treated.
1. Tiemersma EW, van der Werf MJ, Borgdorff MW, et al. Natural History of Tuberculosis: Duration and Fatality of Untreated Pulmonary Tuberculosis in HIV-Negative Patients. PLoS One 2011;6:e17601. Smear+ 10-yr case fatality ~70% (53–86%); smear− culture+ ~20%.
2. WHO Global Tuberculosis Report (treatment). Standard 6-month regimen (2HRZE/4HR) for drug-susceptible TB; global treatment success ~85%. ~10.8 million cases, ~1.23 million deaths/yr.
3. Nyang'wa B-T, et al. (TB-PRACTECAL). NEJM 2022. BPaLM 6-month all-oral regimen: 89% treatment success vs 52% standard of care for MDR/RR-TB.
4. Conradie F, et al. (Nix-TB / ZeNix). NEJM 2020/2022. BPaL favourable outcome ~90% (95/107 Nix-TB); ZeNix 84–91% with reduced linezolid.
5. Cochrane reviews of TB preventive therapy (isoniazid; 3HP rifapentine): ~60% reduction in progression from latent to active TB (RR ≈ 0.40).
6. Trunz BB, Fine PEM, Dye C. Effect of BCG vaccination on childhood tuberculous meningitis and miliary tuberculosis worldwide: a meta-analysis. Lancet 2006. ~73% protection against TB meningitis; ~50% against disease overall.
7. SAPIT (Abdool Karim, NEJM 2010) and CAMELIA (Blanc, NEJM 2011): integrating ART with TB treatment reduces mortality in HIV-associated TB, particularly at low CD4 (~56–68%).
Longevity Research Initiative — Bayesian Causal Atlas. SCM methodology after Judea Pearl; E-values after VanderWeele & Ding. All effect sizes are published values cited above; none were fabricated. Outputs are upper-bound ceilings for research, not clinical guidance. Report generated .