A sensitivity analysis interpretation

The front end.
Post Reply
wantida
Posts: 6
Joined: Wed Sep 06, 2017 2:13 am

A sensitivity analysis interpretation

Post by wantida »

Hi, I have a problem with interpreting the meaning of coloring and side of tornado graph. Some red bar is on the left side some is on the right, what does it mean? And Green bar also. How to investigate those meaning. Another thing is the parameter spread, which can be selected from 10 - 100%, how to choose this number.

Thank you.
Wantida
shooltz[BayesFusion]
Site Admin
Posts: 1417
Joined: Mon Nov 26, 2007 5:51 pm

Re: A sensitivity analysis interpretation

Post by shooltz[BayesFusion] »

See the "Sensitivity analysis in Bayesian networks" chapter in GeNIe help (or a manual - downloadable from http://support.bayesfusion,com/docs - if you prefer the PDF or HTML format).
wantida
Posts: 6
Joined: Wed Sep 06, 2017 2:13 am

Re: A sensitivity analysis interpretation

Post by wantida »

Thank you for your recommendation. I already read the document and still curious about a side of a green and red bar that is changing back and forward.
shooltz[BayesFusion]
Site Admin
Posts: 1417
Joined: Mon Nov 26, 2007 5:51 pm

Re: A sensitivity analysis interpretation

Post by shooltz[BayesFusion] »

The red/green is for positive derivative values, the green/red is for negative derivatives. It allows you to quickly identify which target states will increase or decrease their posterior when you increase the parameter tested for sensitivity.
wantida
Posts: 6
Joined: Wed Sep 06, 2017 2:13 am

Re: A sensitivity analysis interpretation

Post by wantida »

Thank you so much, got it!!
danielhombing
Posts: 17
Joined: Sat Jun 10, 2017 6:53 pm

Re: A sensitivity analysis interpretation

Post by danielhombing »

Hi,
another question regarding sensitivity.

What I understood is the tornado diagram shows the influence of the specific state on the outcome/target node. if the value of that state changes, then the output will change in certain percentage. so the tornado diagram shows the most influential state. for example, in Genie manual page 280 showing the tornado diagram of Hepar II. In number 8 and 9, they are from the same variable (steatosis), but different states (the state is same: steatosis=present, but the column is different).

But if I want to find out the influence of variables (nodes) on the target node (not states), how can I do that? The output then will be a diagram showing the most influential variables (not states).

And another question to clarify my understanding, in sensitivity analysis, does it mean you want to find out what happens to the probability in target node if you change the CPT value of some variables OR you want to find out the effect of giving evidence on the target node? (change CPT value or giving evidence).

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

Re: A sensitivity analysis interpretation

Post by shooltz[BayesFusion] »

What I understood is the tornado diagram shows the influence of the specific state on the outcome/target node. if the value of that state changes, then the output will change in certain percentage.
The change in the target is caused by change in the CPT. Sensitivity implemented in GeNIe looks at single parameter changes (not states overall). Note that if you change the single parameter, the elements in its CPT column must also change to keep the total probability equal to 1.0.
But if I want to find out the influence of variables (nodes) on the target node (not states), how can I do that? The output then will be a diagram showing the most influential variables (not states).
That is not something you can get directly from GeNIe's sensitivity implementation. GeNIe's graphical display uses the highest sensitivity calculated for node's CPT to color the node when sensitivity is enabled.
And another question to clarify my understanding, in sensitivity analysis, does it mean you want to find out what happens to the probability in target node if you change the CPT value of some variables OR you want to find out the effect of giving evidence on the target node? (change CPT value or giving evidence).
Sensitivity measures the change in the posterior probability when some CPT column changes. It takes the evidence set in the network into account (so sensitivity values for given target node will change when you modify the evidence).
Post Reply