Graphs change each time TETRAD is run

[Aggie]

Using TETRAD IV, I get 1 of 8 different graphs every time I run the application using PC algorithm at 5% level. I am using the exact same covariance matrix to generate the graphs. What could be causing this?

 

 

[Clark Glymour]

Have you looked at the 'DAGS in Pattern' tab? It lists all the equivalent DAGS for
your data at the specified alpha level--in your case .05. My bet is that there are 8
of them.

Cg

[Aggie]

I tried the 'DAGs in Pattern' tab and the number of patterns varies everytime I put my same matrix into TETRAD. I launch TETRAD directly from the website each time I use it. I have gotten 1 of 1 patterns, 1 of 8 patterns, 1 of 10, 1 of 24, even 1 of 40. The number of bidirected edges seems to be changing each time I put my matrix into TETRAD. Is there anything else that I might be able to do differently? Thanks very much for the help.

[Clark Glymour]

Ok, thanks. Send us your data file and what version of TETRAD you are using and what
kind of machine (PC or MAC) you are running it on, please.

Cg

[Aggie]

I have attached the text file with my matrix in it. I am using version 4.3.10-6. I am running TETRAD on a PC.

It is only when I close out of TETRAD and launch it again that I am getting different graphs and DAGs.

 

Thanks for all the help.

[Clark Glymour]

Hello,

I have reproduced your results. There was an algorithmic change in 10-6 and 10-7 to
speed up PC search and it seems to have introduced some sort of instability and
therefore unreliability. Joe Ramsey will have to track it down, which will take a
few days since he is on vacation.

Meanwhile, 10-4 is the "stable" version given on the download site. I have run your
data on it and the result is always the same, with 8 directed acyclic graph
possibilities at alpha = .05. I attach the 10-4 tetrad session--to run it you will
need to download 10-4.

Sorry for the problem, but our thanks to you for finding it.

Clark Glymour

[Aggie]

Thank you for taking a look at this for me.

[Aggie]

Just thought that I would let you know that I did try 10-4 but still got different answers. I did, however, find that both versions 4.3.9-0 and 4.3.9-10 (the only other two versions that I tried) both gave me the same results that you obtained each time I ran my data. So I will use those versions for my project. Just thought I would let you know in case this information will help Joe Ramsey. Thank you very much for all of your help and for the quick response.

[Joseph Ramsey]

I don't know if it will help or not.when I try it I don't get different answers. But I'll look atit some more when I get back.

Joe

[ghoza...@gmail.com]

Version 10-4, 10-6 and 10-7 are not stable since I tried to run mmbuild algorithm it seem did not work. Version 4.3.9-18 is stable since I run all the algoritm with good result

Sent from my BlackBerry®
powered by Sinyal Kuat INDOSAT

---

From: Joseph Ramsey <jsph....@gmail.com>

Sender: tetrad-us...@googlegroups.com

Date: Wed, 22 Aug 2012 09:56:01 -0400

To: <tetrad-us...@googlegroups.com>

ReplyTo: tetrad-us...@googlegroups.com

Subject: Re: Graphs change each time TETRAD is run

[Joseph Ramsey]

I need more details. I've been using Mimbuild extensively in 10.7 recently and been working pretty well, though i'm doing problems with dozens of latents. Thanks for any other info you can give.I'll look at it when I get back and can domore than Swype.

Joe

[Joseph Ramsey]

I guess I should add, for anyone else who is  reading this, 10.7 (in general any version not up on the main site) is experimental. Stuff can be added or removed at whim or changed.I'll try to make that more explicit in the future, sorry.

[Joseph Ramsey]

Figured it out. The problem was due to unfaithfulness. In the
adjacency search, adjacent nodes to each variable were being stored in
a set for PC, the hash key of which was changed from run to run, so
adjacencies were iterated over in different orders. For unfaithful
data, this can produce different results. I've fixed the order in
which adjacecent nodes to each variable are itererated over. I'll post
the fix in 10-7, the experimental version, now, and try to make a new
"public" version later today. (It will take a little work.)

Best,

Joe

--
Joseph D. Ramsey
Special Faculty and Director of Research Computing
Department of Philosophy
135 Baker Hall
Carnegie Mellon University
Pittsburgh, PA 15213

jsph....@gmail.com
Office: (412) 268-8063
http://www.andrew.cmu.edu/user/jdramsey

[maryam tabatabaei]

Hi,
I have a question for you. How can I draw DAG using correlation matrix? I used SEM, but using that, I have to draw arrows manually. Is there anyway to draw DAG using the Tetrad?