BPH Bayesian Causal Analysis

Causal Effect Estimate — Active Interventions

P(ΔIPSS > MDD | do(Interventions), Adjusted for Confounders) where MDD = minimum detectable difference (3 pts)

P(Response)

—

Select interventions

Expected ΔIPSS

—

Points reduction

P(Avoid AUR)

—

vs no treatment

P(Avoid Surgery)

—

5-year horizon

Bayesian Posterior — Relative Contribution of Each Active Intervention

Select one or more interventions to compute causal contributions

Causal Structure — Do-Calculus Result

P(ΔIPSS | do(∅)) = 0.45 [placebo baseline]
← awaiting intervention selection

Counterfactual Attribution

Interventional necessity (PN) and sufficiency (PS) scores computed on intervention selection...

Oppositional Analysis — Limitations & Counterarguments

Select interventions to generate specific counterarguments via Pearl's do-calculus skeptical analysis...

Structural Causal Model (SCM) — Directed Acyclic Graph

Pearl Ladder of Causation Level 2: Intervention (do-calculus). Arrows = direct causal pathways. U = unmeasured confounders. Green = interventions, Gold = mediators, Red = outcomes, Purple = confounders.

Causal Pathways

Intervention → Mediator → Outcome

Mediator → IPSS (direct effect)

Confounder → Mediator (backdoor)

Cross-intervention correlation

Backdoor Criterion

P(Y|do(X)) = Σ_z P(Y|X,Z=z)·P(Z=z)
Adjust for: Age, BMI, MetS, T2DM, baseline IPSS, PV
Block: Age→DHT→PV, BMI→MetS→Inflammation backdoor paths

Intervention Ranking — Dose-Response & Causal Effect Size

Mechanism of Action — Causal Pathways

Intervention	Primary Target	Causal Mechanism	Onset	Pathway
α-Blockers tamsulosin, doxazosin	α_1A-adrenoreceptors prostate/bladder neck SM	Block norepinephrine → ↓ SM contraction → ↓ urethral resistance → ↑ Qmax	24–48 hrs	SM Tone → IPSS
5-ARIs finasteride, dutasteride	5α-reductase type II (dutasteride: type I+II)	Block T→DHT conversion → ↓ prostatic epithelial proliferation → ↓ PV → ↓ obstruction	6–12 months	DHT → PV → IPSS
Tadalafil (PDE5-I) 5 mg qd	PDE5 enzyme in LUT bladder, urethra, prostate	PDE5 inhibition → ↑ NO/cGMP → SM relaxation, ↑ LUT perfusion, ↓ afferent nerve hyperactivity	1–4 weeks	NO/cGMP → SM/Bladder Fn
Exercise ≥30 min moderate/d	Sympathetic tone, MetS adipose-prostate axis	↓ Sympathetic activity → ↓ SM tone; ↓ adiposity → ↓ estrogen/DHT → slower PV growth	3–6 months	MetS → SM/Inflammation
Mediterranean Diet high adherence	Inflammation, oxidative stress lycopene, omega-3	↑ Antioxidants → ↓ prostatic inflammation (IL-8, TNF-α) → ↓ proliferative signaling	6–12 months	Inflammation → PV
Saw Palmetto 320 mg hexane ext.	Weak 5α-reductase inhibition fatty acids, phytosterols	Partial 5α-reductase inhibition + anti-inflammatory FAs; inconsistent standardization limits efficacy	Inconsistent	DHT (weak)
β-Sitosterol 60–130 mg/d	Phytosterol anti-inflammatory cholesterol-lowering	↓ prostatic inflammation, modest 5α-reductase inhibition; active component of saw palmetto action	4–8 weeks	Inflammation + DHT
TURP Gold standard	Direct tissue resection transition zone	Mechanical removal of obstructing tissue → immediate ↓ BOO → ↑↑ Qmax, ↓↓ IPSS	Immediate	Direct → PV/obstruction
Rezum (WVTT) steam thermal ablation	Thermal cell necrosis transition zone	Water vapor → instant cell necrosis in transition zone → ↓ PV 25–30% → ↓ obstruction	4–12 weeks	Direct → PV ablation
UroLift (PUL) prostatic urethral lift	Mechanical lobe retraction non-ablative	Permanent implants retract lateral lobes → widen urethral lumen → ↓ BOO; preserves ejaculation	Immediate	Mechanical → BOO

What-If Analysis — P(Outcome | do(Intervention)) for Each Intervention

Pearl Ladder Level 2: Interventional queries. P(Y|do(X=x)) computed via backdoor adjustment. Confounders controlled for patient profile above.

What-If: Combination Therapy

Non-independence adjusted using mechanistic pathway overlap coefficients

If-Not-For Analysis (Necessity)

P(outcome₀=failure | outcome₁=success) — Probability of need

Dose-Response Curves

Effect of dose variation on expected ΔIPSS (from RCT dose-escalation data)

^† Curves based on dose-response RCT data (MTOPS, CombAT, VA-COOP, tadalafil phase III trials). α-blocker shows rapid plateau; 5-ARI shows delayed benefit proportional to DHT suppression and prostate involution.

Counterfactual Analysis — Pearl Ladder Level 3

Necessity (PN), Sufficiency (PS), and Necessity-and-Sufficiency (PNS) of causation for selected interventions

Attribution of Benefit

If the patient responded, what fraction of that response is attributable to each intervention?

Robustness to Assumption Violations

Impact of monotonicity and independence assumption violations

Sensitivity Analysis — Tornado Diagram

Change in P(ΔIPSS>3) as each parameter varies ±2 standard deviations from its central estimate. Longest bars = highest uncertainty leverage.

E-value Analysis

Minimum unmeasured confounding (on RR scale) needed to explain away causal effect

Unmeasured Confounder Impact

Sensitivity to residual confounding after backdoor adjustment

Parameter Uncertainty Bounds (95% Credible Intervals)

Cross-Correlation Matrix — Between Interventions (Mechanistic Pathway Overlap)

Pearson r estimates from shared causal pathway analysis. Non-zero correlations inflate naive additive effect estimates and must be removed via joint do(·) operator.

^† Correlations estimated from mechanism-of-action pathway overlap. α-blocker ↔ PDE5-I: shared smooth muscle relaxation (r≈0.30). 5-ARI ↔ diet/exercise: shared DHT/metabolic pathway (r≈0.15). Surgery dominates all pathways (blocks mediators entirely).

Confounder Correlation Structure

Cross-correlations among measured confounders (blocking backdoor paths)

Adjusted vs Unadjusted Effect Estimates

Impact of removing cross-correlations on ΔIPSS prediction

Evidence Table — Hazard Ratios, IPSS Reduction, Mechanisms & Study Quality

Sourced from: MTOPS (N=3047), CombAT (N=4844), VA-COOP (N=1229), tadalafil phase III meta-analysis (N=3000+), Cochrane BPH reviews, GPRD observational cohort (N=4500), AUA 2023 guidelines.

Intervention	Dose / Regimen	ΔIPSS (Active)	ΔIPSS vs Placebo	HR: AUR	HR: Surgery	ΔQmax (mL/s)	ΔPV	Evidence	Key Study
Tamsulosin	0.4 mg qd	−4.0 to −6.0	−3 to −4	2.32^† (1.37–3.94)	1.78^† (1.30–2.44)	+1.5 to +2.5	None	1A	Debruyne 2004, GPRD
Doxazosin	4–8 mg qd	−4.5 to −6.0	−3 to −4	—	—	+1.5 to +2.0	None	1A	MTOPS (NEJM 2003)
Silodosin	8 mg qd	−4.2 to −5.8	−3 to −4	—	—	+1.5	None	1A	NMA Sci Rep 2024
Finasteride	5 mg qd	−3.6	−2.5	0.43^‡ (0.25–0.73)	0.56^‡ (0.41–0.77)	+1.5	−27%	1A	PLESS, MTOPS
Dutasteride	0.5 mg qd	−4.0 to −5.0	−2.5 to −3.0	0.40 (0.27–0.62)	0.52 (0.36–0.74)	+1.5	−25%	1A	CombAT, REDUCE
Combination α-blocker + 5-ARI	Standard combo	−6.3 to −7.4	−5.0 to −6.0	0.22 (0.12–0.40)	0.17 (0.09–0.31)	+2.2	−25%	1A	MTOPS (66% risk ↓)
Tadalafil	5 mg qd	−4.7 to −6.4	−2.0	—	—	±0 to +0.5	None	1B	Porst 2011, 11 RCTs
Exercise	≥30 min/d moderate	−2.0 to −3.0 est.	−1.5	—	—	+0.5 to +1.0	Slow	2B	Japanese MHLW data
Mediterranean Diet	High adherence	−1.5 to −2.0	−1.0	—	—	—	Modest	2A	BJUI 2021 systematic review
Weight Loss	≥5% BMI reduction	−1.5 to −2.5	−1.5	—	—	—	Modest	2B	RCT PMC 2015
Fluid Restriction	≤2L/d, restrict PM	−1.0 to −1.5	−0.8	—	—	—	None	2B	PMC systematic review 2017
Saw Palmetto	320 mg qd (hexane)	−0.9 (MD −0.90)	−0.5 (small)	—	—	—	None	2A	Cochrane 2023 (27 RCTs)
β-Sitosterol	60–130 mg/d	−2.0 to −3.5	−1.5	—	—	+1.5	None	2B	Berges Lancet 1995, Klippel 1997
UroLift (PUL)	MIST procedure	−8 to −11	−7 to −9	Low	6%/yr re-Rx	+2.5 to +3.5	None	1A	BPH NMA BJUI 2024
Rezum (WVTT)	Steam ablation	−10 to −12	−9 to −11	Low	0.02%/yr re-Rx	+3.0 to +4.5	−25%	1A	McVary 2019 (4-yr RCT)
PAE	Prostatic artery embolization	−11 to −13	−10 to −12	Low	0.05%/yr re-Rx	+3.0 to +4.0	−25%	1B	BJUI NMA 2024
TURP	Gold standard surgery	−14 to −18	−13 to −17	Eliminates	Eliminates	+8 to +12	−75%	1A	BJUI NMA 2024 (reference)

^† HR for AUR and surgery = α-blocker vs 5-ARI (5-ARI = reference = 1.0); source: Debruyne et al. Eur Urol 2004, GPRD N=4500.
^‡ HR for finasteride = finasteride vs α-blocker (α-blocker = reference = 1.0).
ARI = alpha-reductase inhibitor; AUR = acute urinary retention; BOO = bladder outlet obstruction; DHT = dihydrotestosterone; IPSS = International Prostate Symptom Score; LUTS = lower urinary tract symptoms; MDD = minimally detectable difference; MIST = minimally invasive surgical therapy; NMA = network meta-analysis; PAE = prostatic artery embolization; PDE5-I = phosphodiesterase type 5 inhibitor; PSA = prostate-specific antigen; PUL = prostatic urethral lift; PV = prostate volume; Qmax = peak urinary flow rate; RCT = randomised controlled trial; SM = smooth muscle; T2DM = type 2 diabetes mellitus; TURP = transurethral resection of the prostate; WVTT = convective water vapour thermal therapy.

Methods — Bayesian Causal Inference Framework

1. Structural Causal Model (SCM)

Following Pearl (2009), the SCM is defined by:

M = ⟨V, U, F, P(U)⟩
V = {Age, BMI, MetS, T2DM, DHT, PV, SM, Inflammation, IPSS}
U = {unmeasured confounders}
F = {f_v : V_i = f_i(PA_i, U_i)}
PA = direct causes (parents in DAG)

2. Identification via Backdoor Criterion

Backdoor criterion satisfied by adjusting for Z = {Age, BMI, MetS, T2DM, baseline IPSS, PV}:

P(Y=y|do(X=x)) =
Σ_z P(Y=y|X=x, Z=z) · P(Z=z)
No unmeasured common causes between X and Y after Z adjustment

3. Non-Independence Adjustment

For multiple co-interventions with pathway overlap ρ_AB:

P(Y|do(A,B)) = P(Y|do(A)) + P(Y|do(B))
− ρ_AB·√(P(A effect)·P(B effect))
ρ estimated from mechanistic pathway overlap analysis

4. Counterfactual Computation

Under monotonicity assumption (no harmful cases):

PN ≥ (P(Y=1|x) − P(Y=1|x')) / P(Y=1|x)
PS ≥ (P(Y=1|x) − P(Y=1|x')) / P(Y=0|x')
PNS = P(Y=1|do(x)) − P(Y=1|do(x'))
Bounds derived from observational distributions

Assumptions Table

#	Assumption	Justification	Sensitivity Impact
A1	Consistency: Y_x(u)=Y when X(u)=x	Single-dose single-intervention definition; validated in RCTs	Medium
A2	Positivity: P(X=x\|Z=z) > 0	All interventions available to eligible population	Low
A3	Exchangeability: Y_x ⊥ X\|Z	Backdoor set Z blocks all confounding; violation explored in E-value analysis	High
A4	Monotonicity: Y₁ ≥ Y₀ for all u	No intervention causes net harm on IPSS primary outcome; violated for surgery in frail patients	Medium
A5	SUTVA: No interference between subjects	Valid for pharmacotherapy; partially violated for population-level effects	Low

Oppositional Critique — Where the Model May Fail

1. Heterogeneity of effect modifiers: Effect sizes are population averages. Individual response varies dramatically with prostate histology, pharmacogenomics (CYP3A4 metabolism of tamsulosin), and unmeasured comorbidities. The causal effect for any given patient may differ substantially from the reported posterior probabilities.

2. Publication bias in RCTs: The MTOPS and CombAT data over-represent trials with positive outcomes. A funnel-plot asymmetry in alpha-blocker trials (Egger test, p=0.03 in some meta-analyses) suggests inflation of effect size estimates by ~15–20%.

3. Surrogate endpoint limitation: IPSS is a subjective self-report score. The MDD of 3 points may not reflect clinically meaningful improvement in objective metrics (Qmax, residual volume, progression to retention). The correlation between ΔIPSS and QoL is strong (r=0.7) but imperfect.

4. Duration of treatment not modeled: Alpha-blockers have rapid but plateaued effects; 5-ARIs have delayed but progressive effects. This time dimension is collapsed to a point estimate, which may misrepresent the optimal treatment sequence.

5. Counter-thesis on saw palmetto: While Cochrane 2023 shows marginal effect (MD −0.90, CI −1.74 to −0.07), the lower bound just crosses the null. Standardized hexane extract at adequate β-sitosterol content may have clinically meaningful subset benefits not captured in pooled heterogeneous analyses.

Resolution strategy: E-value analysis quantifies minimum unmeasured confounding needed to nullify effects. Sensitivity analyses across assumption violations are displayed. Individual patient prediction should be interpreted as probabilistic guidance, not deterministic prescription.