explain the tornado diagram

The front end.
Post Reply
Yan
Posts: 22
Joined: Fri Mar 04, 2022 5:26 am

explain the tornado diagram

Post by Yan »

Dear staff,

Please see the attached example of an BN model. I would like to ask some questions about the tornado diagram. For example, when you set node L as target and run the sensitivity analysis, you will find the most sensitivity parameter for "L = good" is "I = L | D = L". That means this paramter will make the posterior probability of "L = good" change from 0.303756 to 0.314423. However, when I set "D = L = 100%" and "I = L = 100%", the value of target "L = good" changes from original 31% to 39% in the BN model. This is different from the target value range in the tornado diagram. Do I misunderstand the meaning of current value (i.e., equal to the probability in the box)?

A further question is the tornado diagram shows the most sensitivity parameter for "L = good" is "I = L | D = L", and the parameter "K = L | I = L, H = L, J = L" is the third one. However, when I set "D = L = 100%" and "I = L = 100%", the value of target "L = good" changes from original 31% to 39% in the BN model, while "I = L = 100%, H=L=100%, J=L=100%, and K=L=100%" makes the value of target "L = good" changes from original 31% to 49%. It seems the latter increases the value of "L = good" more, why in the sensitivity analysis, it is less sensitive than "I = L | D = L"?

Many thanks.

Cheers
Attachments
test (1).xdsl
(405.34 KiB) Downloaded 161 times
shooltz[BayesFusion]
Site Admin
Posts: 1417
Joined: Mon Nov 26, 2007 5:51 pm

Re: explain the tornado diagram

Post by shooltz[BayesFusion] »

most sensitivity parameter for "L = good" is "I = L | D = L". That means this paramter will make the posterior probability of "L = good" change from 0.303756 to 0.314423.
The output probability range from 0.303756 to 0.314423 is calculated for parameter spread=10% (the default value in the slider in the bottom part of the tornado window). If you hover over the first tornado bar you'll get the following text (which is also copied to clipboard if you copy the tornado diagram):

Code: Select all

1: I=L | D=L
Target value range: [0.303756 .. 0.314423], width=0.0106666
Parameter range: [0.646916 .. 0.790675], width=0.143759
Current parameter value: 0.718795 at definition index 0
Derivative: 0.0741975
Coeffs: a=0.0741975, b=0.255757, c=-4.44089e-16, d=1
The singe CPT entry for node I (I=L | D=L) is currently set at 0.718795. Changing it by 10% yields the [0.646916 .. 0.790675 range. At the edges of this range the output L=good varies between 0.303756 and 0.314423.

If you change the parameter spread percentage using the slider, you'll get different ranges for L=good (and ordering in the tornado will most likely change).
However, when I set "D = L = 100%" and "I = L = 100%"
This action changes two parameters in two nodes at once. GeNIe calculates changes for L=good when only one parameter changes.
A further question is the tornado diagram shows the most sensitivity parameter for "L = good" is "I = L | D = L", and the parameter "K = L | I = L, H = L, J = L" is the third one
The ordering of the tornado bars will change when you modify the parameter spread. If you want to go up to 1.0 (100%) probability in some of your CPTs, then you should move the parameter spread slider to the right. In such case, the most sensitive parameter will be D=L, followed by D=M and I=L|D=L. The maximum value of L=good when only one parameter changes will be 0.352834.
value of target "L = good" changes from original 31% to 49%.
The 49% can be achieved only if you change more than one parameter at once.
Post Reply