examples/examples.csv

id,title,regime,question,source,max_cases,propensity_center,propensity_slope,teaches,expected
pressor_map,Pressor → MAP,point,"Does starting a pressor raise mean arterial pressure one step later, adjusting for current MAP, heart rate, and SpO₂?",synthetic,40,75,10,Backdoor adjustment on a point treatment. The generator injects a known +2.0 mmHg effect and confounds treatment by MAP; adjustment should recover it.,ATE ≈ +1.9 mmHg (injected truth +2.0)
null_control,Negative control outcome,point,"What is the one-step effect of a pressor on heart rate, controlling for the current vitals?",synthetic,40,75,10,"A negative control: the generator injects NO pressor→HR effect. A trustworthy pipeline must return ~0 here — this is what catches a method that ""finds"" an effect everywhere.",ATE ≈ 0 bpm (correctly null)
sustained_msm,Sustained regime (g-method),sustained,What is the effect of keeping a pressor running for 5 steps on mean arterial pressure?,synthetic,40,75,10,"Why g-methods exist. MAP is a confounder of the current dose AND a mediator of the earlier one, so no adjustment set works. The IPTW-MSM reweights instead of conditioning.",MSM ≈ +5.8 mmHg vs truth +6.72; naive ≈ +2.8 (54% too low)
positivity,Positivity violation,sustained,What is the effect of keeping a pressor running for 5 steps on mean arterial pressure?,synthetic,40,70,6,"Sharpens the treatment rule so the always-treat regime is essentially never observed. The estimate becomes an extrapolation — watch the `extrapolated` flag fire.",Same question as above but n_always ≈ 1 → flagged as extrapolated
long_horizon,Longer horizon (2 minutes),sustained,What is the effect of a pressor sustained for 2 minutes on mean arterial pressure?,synthetic,60,75,10,"Horizon parsing (2 min ÷ 10 s = 12 steps) and the price of a long window: the effect saturates (it is NOT k × one-step), and positivity collapses — nobody is treated for 12 straight steps.",k=12; truth +9.31 (not 12×2.0=24). MSM ≈ +7.0 and flagged extrapolated — n_always = 0
vitaldb_real,Real data (VitalDB),point,Estimate the one-step effect of phenylephrine on mean arterial pressure using VitalDB.,vitaldb,20,75,10,"Real ICU waveforms — no ground truth to check against. The refutation suite and the mixle cross-check become the only evidence you have.",No known truth; judge it by the refuters