Causal Inference: Ruxolitinib
Data period: 2026-01-08 to 2026-07-02 (178 days) | Total runtime: 0.5s
Four complementary causal analysis methods explore whether Oura biometrics
shifted after ruxolitinib
(10 mg BID, started 2026-03-16) on Oura Ring biometrics.
Data period: 2026-01-08 to 2026-07-02
(178 days).
Pre-intervention
Info67days
Post-intervention
111days
Raw p<0.05
Info0/0
FDR-significant
None0/0
Lowest raw p
Not significantp=1.000
N/A | q=1.0000
Methods used
Info4
CI + PCMCI+ + TE + Mediation
Table of contents:
INDIV
0. Individual Metric Treatment Response
Method: Non-parametric Mann-Whitney U test compares pre- vs post-ruxolitinib distributions for each core biometric. Effect sizes reported as Cohen's d with bootstrap 95% CI (2000 iterations). These direct statistical tests provide intuitive per-metric significance before the multivariate causal methods below.
HRV (RMSSD)
Significantp<0.001
d=+1.75 (large)
Lowest HR
Significantp<0.001
d=-1.82 (large)
Average HR
Significantp<0.001
d=-1.74 (large)
Sleep Efficiency
Significantp<0.001
d=+0.66 (medium)
| Metric | Pre-Rux Mean | Post-Rux Mean | Change | Cohen's d | Mann-Whitney | 95% CI (diff) |
|---|---|---|---|---|---|---|
| HRV (RMSSD) | 10.00 ms | 25.59 ms | +15.59 ms | +1.75 (large) | Sig p<0.001 | [+13.46, +17.89] |
| Lowest HR | 76.72 bpm | 63.19 bpm | -13.52 bpm | -1.82 (large) | Sig p<0.001 | [-15.60, -11.36] |
| Average HR | 85.17 bpm | 72.14 bpm | -13.03 bpm | -1.74 (large) | Sig p<0.001 | [-15.33, -10.80] |
| Sleep Efficiency | 78.62 % | 81.92 % | +3.30 % | +0.66 (medium) | Sig p<0.001 | [+1.94, +4.71] |
Note: These per-metric tests complement the Bayesian CausalImpact analysis below. For full changepoint detection and multi-patient comparison, see the Comparative Treatment Response report.
CONFOUNDER
0b. Confounder Analysis: Beta-Blocker Separation
Method: Three-period analysis separating the post-treatment window into Jakavi-only (23 days) and Jakavi + beta-blocker (85 days). Mann-Whitney U tests isolate each drug's contribution. This addresses the key confounder question: does Ruxolitinib's effect stand independently of the beta-blocker added later?
Treatment Timeline
Pre-treatment:
67 days
Jakavi only:
2026-03-16 to 2026-04-07 (23 days)
Jakavi + BB:
2026-04-08 to present (85 days)
Test 1: Isolated Ruxolitinib Effect
Pre-treatment vs Jakavi-only period (beta-blocker confounder eliminated)
| Metric | Pre Mean | Jakavi-only Mean | Change | Cohen's d | p-value |
|---|---|---|---|---|---|
| HRV (RMSSD) | 10.0 ms (n=67) | 10.9 ms (n=22) | +0.9 ms | +0.40 (small) | p=0.076 |
| Lowest HR | 76.7 bpm (n=64) | 72.9 bpm (n=19) | -3.8 bpm | -0.77 (medium) | p=0.0093 |
| Average HR | 85.2 bpm (n=64) | 81.4 bpm (n=19) | -3.7 bpm | -0.66 (medium) | p=0.0119 |
| Sleep Efficiency | 78.6 % (n=64) | 79.9 % (n=19) | +1.3 % | +0.36 (small) | p=0.102 |
Test 2: Marginal Beta-Blocker Effect
Jakavi-only vs Jakavi + beta-blocker (what BB adds on top)
| Metric | Jakavi-only Mean | Jakavi+BB Mean | Change | p-value |
|---|---|---|---|---|
| HRV (RMSSD) | 10.9 ms (n=22) | 29.3 ms (n=87) | +18.4 ms | p=0.0000 |
| Lowest HR | 72.9 bpm (n=19) | 61.0 bpm (n=84) | -11.9 bpm | p=0.0000 |
| Average HR | 81.4 bpm (n=19) | 70.0 bpm (n=84) | -11.4 bpm | p=0.0000 |
| Sleep Efficiency | 79.9 % (n=19) | 82.4 % (n=84) | +2.4 % | p=0.0034 |
Key Findings
- Ruxolitinib alone produces a statistically significant reduction in heart rate, independent of beta-blocker.
- HRV improvement during Jakavi-only period does not reach significance — the HRV signal strengthens after beta-blocker addition.
- Beta-blocker addition produces a significant further HRV increase on top of Jakavi.
CI
1. CausalImpact - Bayesian Structural Time Series Analysis
Method: Bayesian Structural Time Series (BSTS) models pre-intervention dynamics and generates a counterfactual prediction for the post-period. The difference between actual and counterfactual estimates the causal effect, with full posterior uncertainty. MCMC: 5,000 iterations, weekly seasonal component (nseasons=7). Benjamini-Hochberg FDR correction for multiple testing.
FAILED CausalImpact package not installed. Install with: pip install pycausalimpact
STATPOWER
1a. Statistical Power & Interpretation
Purpose: All 11 metrics sorted by statistical significance, with Benjamini-Hochberg corrected q-values for multiple testing.
No CausalImpact results available for statistical power analysis.
PLACEBO
1b. Placebo tests (intervention date falsification)
Method: CausalImpact is run with 3 random placebo dates in the pre-period on the 3 most significant metrics. Placebo dates should NOT show significant effects.
CausalImpact failed, skipping placebo
PCMCI
2. Granger Causality Network (PCMCI+)
Method: PCMCI+ (tigramite) tests for time-lagged causal relationships between biometric variables using partial correlation. Tau_max = 7 days.
Full period (0 days): No significant causal links found.
Pre-ruxolitinib (0 days): No significant causal links found.
TE
3. Transfer Entropy
Method: Transfer entropy quantifies directional information flow between biometric streams. Comparison of TE matrices for pre- and full period reveals changes in information coupling after ruxolitinib start.
Full period (177 days)
| Source | Target | TE (bits) | Net TE |
|---|---|---|---|
| REMpct | RMSSD | 1.0323 | +0.2503 |
| DeepDur | RMSSD | 1.0208 | +0.3173 |
| TempDev | RMSSD | 1.0208 | +0.2483 |
| REMpct | AvgHR | 1.0193 | +0.2100 |
| TotalSleep | AvgHR | 1.0193 | +0.2416 |
| DeepDur | AvgHR | 1.0193 | +0.3158 |
| TotalSleep | RMSSD | 1.0093 | +0.2545 |
| REMdur | AvgHR | 1.0078 | +0.1645 |
Pre-ruxolitinib (67 days)
| Source | Target | TE (bits) | Net TE |
|---|---|---|---|
| REMdur | RespRate | 0.5564 | +0.1891 |
| TotalSleep | RespRate | 0.5564 | +0.3064 |
| DeepDur | RespRate | 0.5564 | +0.2439 |
| RMSSDmax | RespRate | 0.5564 | +0.3689 |
| LowestHR | RespRate | 0.5564 | +0.2127 |
| AvgHR | RespRate | 0.5564 | +0.1071 |
| SpO2 | RespRate | 0.5564 | +0.1132 |
| REMpct | RespRate | 0.5252 | +0.0835 |
Change in information flow
Largest increase: DeepDur -> RMSSDmax (+0.7513 bits)
Largest decrease: REMdur -> REMdur (+0.0000 bits)
MEDIATION
4. Intervention Response Decomposition
Method: Linear mediation analysis (Baron-Kenny) decomposes total ruxolitinib effect into four mediating pathways. Bootstrap (2000 iterations) for confidence intervals.
+9.6
Total effect (readiness score)
64.8 -> 74.4
Pre -> Post average
p<0.001
Raw p-value Significant
| Pathway | Mediator (pre->post) | a (T->M) | b (M->Y) | Indirect effect [95% CI] | % mediated | p-value |
|---|---|---|---|---|---|---|
| Direct cardiac Ruxolitinib -> HR change -> Readiness | 92.57 -> 82.90 | -0.905 | -0.578 | +0.5232 [+0.3571, +0.7215] | 52.2% | 0.0000 Sig |
| Autonomic Ruxolitinib -> HRV change -> Readiness | 10.00 -> 25.88 | +1.357 | +0.598 | +0.8116 [+0.5930, +1.0329] | 80.5% | 0.0000 Sig |
| Sleep-mediated Ruxolitinib -> Sleep efficiency -> Readiness | 78.62 -> 81.92 | +0.635 | +0.326 | +0.2066 [+0.0937, +0.3490] | 19.8% | 0.0000 Sig |
| Inflammatory Ruxolitinib -> Temperature deviation -> Readiness | 0.06 -> 0.05 | -0.034 | -0.275 | +0.0094 [-0.0664, +0.1002] | 0.9% | 0.8030 NS |
CLINICAL
5. Clinical Interpretation
Executive summary
N/A
Strongest signal: p=1.000
0/0
Metrics trending in expected direction
111
Post-intervention days (Day 14 target)
- N/A: strongest hypothesis-generating raw p-value signal (p=1.0000, q=1.0000). It does not survive FDR correction, so confirmation still depends on more post-treatment follow-up.
- CausalImpact: 0 of 0 biometric streams show significant causal change (p < 0.05). After Benjamini-Hochberg FDR correction: 0 of 0 remain significant (q < 0.05).
- Placebo validation: N/A - ?/? placebo tests reached significance. This keeps the result vulnerable to false positives rather than establishing a confirmed intervention effect.
- PCMCI+: 0 significant time-lagged causal links identified in the biometric network
- Mediation analysis: 3 of 4 mediating pathways show significant indirect effect
Limitations
- Short post-period (111 days): All results are preliminary. Minimum 14-21 days of post-intervention data recommended for robust causal inference.
- Confounders: Linear methods cannot capture non-linear interactions. Seasonal variation, activity level, and other medications are not controlled for.
- Wearable data: Oura Ring is not a medical device. Measurements have inherent noise that can affect causal estimates.
- Single patient: N=1 study without control group. Causality cannot be definitively established, but Bayesian posterior probability of effect provides a strength measure.
- HEV diagnosis: HEV was diagnosed 2026-03-18 (2 days after ruxolitinib start). Hepatitis may confound biometric changes.
Recommendations
- Repeat analysis after 2-3 weeks of ruxolitinib treatment for robust causal inference
- Add HEV-related biomarkers (ALT, bilirubin) as time-varying covariates
- Consider synthetic control method when longer time series are available
- Combine with clinical endpoints (GVHD scoring, ferritin) for multimodal analysis