235 Partial Verification Blurg (incomplete)
235.1 Summary
- Propensity Score Intuition
- Partial Verification
- ‘Data/Methods’ Limitations:
- Partial Verification Bias
- Omniscient View
- Only some tested
- Complete Case
- Correctly modeled propensity score
- Inverse Propensity Weights
- Weighted (pseudo)population restores prevalence and sex balance of the whole population
235.2 Slide outline
235.2.1 Slide 1
- Propensity Score Intuition
- Needs to be finished. ### Slide 2
- Partial Verification
- Omniscient View
- 10 of 24 have hypercapnia
- 5 of 12 men
- 5 of 12 women
- P(Test | Man) 50%
- P(Test | Woman) 25%
- Clinicians can’t perfect predict who needs the test
- ORtest|hypercap: 2.8 (M or W) ### Slide 3
- Partial Verification ### Slide 4
- ‘Data/Methods’ Limitations:
- ‘Partial Verification’: hypercapnia status is only ascertained in patients who get an ABG.
- Analysis: Among patients who have ABGs checked
- Target: Among all patients who may have hypercapnia
- The probability of getting an ABG is variable and idiosyncratic
- Inverse probability of blood-gas weighting?
- Missing data is tricky
- P(ABG) 0.5
- P(ABG) 0.25
- Observed w/ Hypercapnia
- Weighted by Inverse of Prob
- True Status (unobserved)
- The Missing Indicator Method.
- CMAJ 2012 ### Slide 5
- Partial Verification Bias ### Slide 6
- Omniscient View
- 50% prevalence; both sexes
- Our goal is to predict who has the condition in the entire sample ### Slide 7
- Only some tested ### Slide 8
- Complete Case
- 66% prevalence; mostly women
- If you train the model on this, it’ll learn that sex predicts the condition.
- True! among people who get the test ### Slide 9
- Only some tested
- But we are particularly interested in predicting for people who didn’t get the test ### Slide 10
- Correctly modeled propensity score
- 33%
- 100%
- 33%% ### Slide 11
- Inverse Propensity Weights
- 1/.33
- 3
- 1/1
- 1 ### Slide 12
- Weighted (pseudo)population restores prevalence and sex balance of the whole population
- Intuition:
- If someone who was unlikely to get the test but did and they have the condition → there are likely many similar patients with the condition that didn’t get the test.
235.3 Learning objectives
- Propensity Score Intuition
- Partial Verification
- ‘Data/Methods’ Limitations:
- Partial Verification Bias
- Omniscient View
235.4 Bottom line / summary
- Propensity Score Intuition
- Partial Verification
- ‘Data/Methods’ Limitations:
- Partial Verification Bias
- Omniscient View
235.5 Approach
- TODO: Outline the initial assessment or decision point.
- TODO: Outline the next diagnostic or management step.
- TODO: Outline follow-up or escalation criteria.
235.6 Red flags / when to escalate
- TODO: List red flags that require urgent escalation.
235.7 Common pitfalls
- TODO: Capture common errors or missed steps.
235.8 References
TODO: Add landmark references or guideline citations.