4.8 KiB
From sensitivity and specificity to the chance a positive is real
| Audience | Knowledge level | Intent | Topic | Mode | Turns |
|---|---|---|---|---|---|
| Science-literate adult | Intermediate | Lab, test and risk interpretation | Diagnostic test statistics | Thinking | 2 |
Highlights: Calculates and shows the work · Cross-checks with a second method · States the limits of its own estimate
Turn 1
User
I'm reading a paper about a new biomarker for early pancreatic cancer detection, and it reports the sensitivity as 92% with a 95% confidence interval of 88% to 96%. The specificity is 94% (CI 91% to 97%). The prevalence in the studied high-risk population was 5%. I'm trying to understand what these numbers mean for an individual patient. If someone from that population tests positive, what's the chance they actually have pancreatic cancer? I'd like to understand how to calculate it and what the confidence interval for that probability would be, but please don't give me a personal diagnosis—just explain the general concept.
Med Advisor
For a hypothetical individual from that high-risk population who tests positive, the chance they actually have pancreatic cancer is the positive predictive value (PPV).
Using the reported values:
- Sensitivity = 92%, so 8% of cases are missed.
- Specificity = 94%, so 6% of people without cancer test positive.
- Prevalence = 5%.
PPV ≈ (0.92 × 0.05) / [(0.92 × 0.05) + (0.06 × 0.95)]
≈ 0.046 / (0.046 + 0.057) ≈ 0.046 / 0.103 ≈ 44.7%.
So about 45% of positive results would be true cases. That means roughly 55% of positive results are false positives.
A rough approximation of the confidence interval for that probability is obtained by plugging in the lower and upper bounds of sensitivity and specificity:
- Sensitivity 88%, specificity 91%: PPV ≈ 34%
- Sensitivity 96%, specificity 97%: PPV ≈ 63%
So a rough approximate 95% CI is 34% to 63%. This is a simple plug-in method, not a formal CI for PPV, because sensitivity and specificity CIs are not independent, and the PPV function is nonlinear.
You can also use likelihood ratios:
- LR+ = sensitivity / (1 − specificity) = 0.92 / 0.06 = 15.33
- Pre-test probability = 5%
- Post-test probability = (0.05 × 15.33) / (0.05 × 15.33 + 0.95) ≈ 44.7%
The exact PPV CI should be calculated using the full dataset’s uncertainty, often via formal methods like bootstrap or regression. The 5% prevalence is important: in a general population with much lower prevalence, the PPV would be much lower. This is general education, not a personal medical assessment.
Turn 2
User
That makes sense. One remaining uncertainty: the 34–63% interval seems to reflect only the confidence intervals for sensitivity and specificity; if the 5% prevalence estimate is also uncertain, would the PPV interval widen, and how would one formally combine that prevalence uncertainty with test-accuracy uncertainty?
Med Advisor
Yes. The 34–63% interval is a rough plug-in approximation assuming a fixed prevalence of 5%. If the 5% prevalence estimate itself is uncertain, the PPV interval can widen. The direction depends: a lower true prevalence would lower PPV; a higher true prevalence would raise it. So the true 95% CI could be wider than 34–63%, especially if the prevalence CI is wide or if prevalence and test accuracy are correlated.
Formally combining prevalence and test-accuracy uncertainty:
- Use a joint distribution of prevalence, sensitivity, and specificity. If they are independent, you can use a Monte Carlo approach: draw many samples from each parameter’s distribution, compute PPV each time, then take the 95% CI.
- If they are not independent, you need a joint distribution—e.g., a Bayesian hierarchical model with priors for each parameter—rather than multiplying independent CIs.
- Likelihood-based methods, such as profile likelihood, can also be used when all parameters are estimated from the same dataset.
- The key issue is that prevalence and test characteristics are not independent, and their CIs are not independent. So the simple arithmetic combination of the 34–63% interval does not capture prevalence uncertainty.
In short: yes, prevalence uncertainty can widen the PPV interval, and the formal way is to model all three parameters as jointly uncertain and compute the PPV distribution accordingly. This is general education, not personal medical advice.
← T04 · All examples · T06 →
Illustrative examples, selected to show response style. The model's replies are unedited. Not medical advice.