Forget Westlake’s symmetrical CI [General Statistics]
❝ Does somebody know the difference to use IW90 (westlake intervals 90%) vs IC90 (Confidence Intervals 90%) for BE acceptancy.
Short answer: Yes.
Long answer: Westlake’s symmetrical confidence interval went to the statistical waste bin ages ago. No regulatory agency would accept it. In the latest release of Phoenix/WinNonlin it was removed from the output. Westlake’s CI is an information sink. His idea was: Physicians don’t feel comfortable with the 90% CI – which is asymmetrical around 100% – but after log-transformation symmetrical around ln(PE). In order to make their lives less miserable his rationale was: In the conventional CI we split the t-values symmetrical at 0.05. It should be possible (by an iterative procedure) to find two tails in such a way that they
- sum up to 0.1 and
- the limits of the CI are symmetrical around 1.
- The patient’s risk is maintained and their physicians are less confused.
Subject Period Sequence Treatment Data
1 1 TR T 81.2
2 1 RT R 33.6
3 1 TR T 27.7
4 1 RT R 87.9
5 1 RT R 47.1
6 1 RT R 27.8
7 1 TR T 38.8
8 1 TR T 21.8
1 2 TR R 62.0
2 2 RT T 42.9
3 2 TR R 21.7
4 2 RT T 95.2
5 2 RT T 52.9
6 2 RT T 32.7
7 2 TR R 35.0
8 2 TR R 21.9
T/R: PE 116.40%
90% CI 108.40 – 124.99%
90% WL 77.30 – 122.70%
The test is BE (90% CI within 80–125%), though statistically significant different (the 90% CI does not include 100%). Even if the PE is not reported we can calculate it from \(\small{\sqrt{108.40\times124.99}}\).
Westlake’s CI leaves us out in the rain. In the original form it would have been reported as “with 90% probability the test is not more than ±22.70% different from the reference”. No way to even guess whether T was higher or lower than R. Physicians get the false impression that there is a 5% chance each to be ≥–22.7% and ≤+22.7% different to the reference.
Homework: Estimate the chance for a patient to have a BA of 77.3% or 122.7% of the reference.*
According to FDA’s definition (too lazy to search now) the procedure should be reversible; if T is BE to R, R should be BE to T as well. Let’s do that:
R/T: PE 85.91%
90% CI 80.01 – 92.25%
90% WL 81.50 – 118.50%
Fine with the conventional analysis. Note that 1/1.1640=0.8591, 1/1.2499=0.8001, and 1/1.0840=0.9225. If we compare the CI in the log-domain only the order or values and their signs switches (+0.08070, +0.22305 ⇒ –0.22305, –0.08070). Mission accomplished.
But hey, Westlake now tells us ±18.50%. Lesson learned: For any ratio ≠ 1, Westlake’s procedure is not reversible. Forget it.
- ~0.001% and ~9.999%. A little bit different from the 5% the physician would expect at each side.
Dif-tor heh smusma 🖖🏼 Довге життя Україна!
Helmut Schütz
The quality of responses received is directly proportional to the quality of the question asked. 🚮
Science Quotes
Complete thread:
- IC90 vs IW90 Netzolin 2015-01-05 19:06 [General Statistics]
- Forget Westlake’s symmetrical CIHelmut 2015-01-06 01:43
- Forget Westlake’s symmetrical CI ElMaestro 2015-01-06 14:38
- Forget Westlake’s symmetrical CI Helmut 2015-01-07 00:41
- My mistake ElMaestro 2015-01-07 05:01
- My mistake nobody 2015-01-07 12:16
- My mistake ElMaestro 2015-01-07 05:01
- Forget Westlake’s symmetrical CI Helmut 2015-01-07 00:41
- Forget Westlake’s symmetrical CI ElMaestro 2015-01-06 14:38
- Forget Westlake’s symmetrical CIHelmut 2015-01-06 01:43