about the sensitivity analysis result

The front end.
Post Reply
danielhombing
Posts: 17
Joined: Sat Jun 10, 2017 6:53 pm

about the sensitivity analysis result

Post by danielhombing »

Hi,

Could you help me to "modify" the result of tornado diagram in sensitivity analysis? I attach the images of sensitivity analysis and tornado here.

(1) I set "use product" as the target node. in the sensitivity analysis, node "access" is redder than others. Then I conclude that "access" is the most sensitive variable to the target node "use product".
scheme with sensitivity analysis
scheme with sensitivity analysis
scheme.jpg (95.88 KiB) Viewed 4995 times
(2) When I run tornado diagram, I see that the most sensitive one is written like this:
use=no|needs=low, like = low, pressure = low, the value ranging from 0.165 - 0.192.
I know the result represents what will happen if I change the CPT value in node "use", state=no, and the box under the "needs=low, like = low, pressure = low".
tornado result
tornado result
tornado.jpg (111.57 KiB) Viewed 4995 times
Is there any way to "convert" the result of (1) to be like (2)? Or "quantify" the red colour indicator and make numbers out of the red colour? So let say I will get the diagram with nodes only and not their combination and states like in the tornado diagram (please see "sensitivity analysis diagram example" file in the attachment).
example of sensitivity analysis
example of sensitivity analysis
sensitivity analysis diagram example.jpg (61.04 KiB) Viewed 4995 times
Thank you for your kind response.

Best,
Daniel
shooltz[BayesFusion]
Site Admin
Posts: 1417
Joined: Mon Nov 26, 2007 5:51 pm

Re: about the sensitivity analysis result

Post by shooltz[BayesFusion] »

You can get maximum node sensitivity by hovering over the small tornado icon displayed in the corner of each node - this will open a small tooltip with max/min/average sensitivity of node's parameters. You'd have to write down the maximum value for each node and use these numbers to create your own tornado display.
danielhombing
Posts: 17
Joined: Sat Jun 10, 2017 6:53 pm

Re: about the sensitivity analysis result

Post by danielhombing »

Thanks Shooltz.

And what is the meaning of that number?
Let say I get this value:
min: 0
max: 0.029
average: 0.004
shooltz[BayesFusion]
Site Admin
Posts: 1417
Joined: Mon Nov 26, 2007 5:51 pm

Re: about the sensitivity analysis result

Post by shooltz[BayesFusion] »

These numbers are minimum, maximum and average sensitivity for the node's parameters. Each of the node's CPT elements has its own sensitivity number; to color the nodes in shades of red GeNIe uses the maximum.
danielhombing
Posts: 17
Joined: Sat Jun 10, 2017 6:53 pm

Re: about the sensitivity analysis result

Post by danielhombing »

Thanks Shooltz!

Let say I get this value in node social pressure:
min: 0
max: 0.029
average: 0.004

Based on that value, can I say that if I change one parameter in CPT social pressure, then the probability value of the use product will change a maximum of 0.029 from the current probability? (with the same information on other nodes).

Another question is, how big is the algorithm change/vary the CPT value of the node (or the parameter spread)? Because in the tornado diagram, you can change the parameter spread and it will change also the target value range.

Again, many thanks!
shooltz[BayesFusion]
Site Admin
Posts: 1417
Joined: Mon Nov 26, 2007 5:51 pm

Re: about the sensitivity analysis result

Post by shooltz[BayesFusion] »

Based on that value, can I say that if I change one parameter in CPT social pressure, then the probability value of the use product will change a maximum of 0.029 from the current probability? (with the same information on other nodes).
The sensitivity is a derivative. Therefore, you can say that there's one parameter in the CPT of social pressure, which causes change in the target posterior which is 0.029 times the change of this parameter (in the proximity of its current value, of course).
Another question is, how big is the algorithm change/vary the CPT value of the node
The algorithm calculates the actual functions linking each parameter with the target posteriors. Therefore, the numbers you see (except in tornado) are not dependend on the parameter spread.

In the tornado, we must assume the spread to calculate the bounds of target posteriors.
Post Reply