GRaSP-FCI returns empty graph

[edpclau]

Hi, @jdramsey! I'm having some issues running GRaSP-FCI with my data. It always returns an empty graph. It may because it's running properly, but I don't know enough to make sure all the settings are correct.

Code:
java -jar /Users/eddie/Downloads/causal-cmd-1.12.0-jar-with-dependencies.jar
--dataset "minimal_set.csv"
--data-type mixed
--delimiter comma
--missing-marker "*"
--numCategories 6
--algorithm grasp-fci
--score dg-bic-score
--test dg-lr-test
--resamplingWithReplacement
--numberResampling 1
--resamplingEnsemble 2
--numStarts 1
--addOriginalDataset yes
--saveBootstrapGraphs yes
--json-graph
--verbose yes
--out

graspfci_1738161916503_out.txt

[jdramsey]

Oh I did not see this! One issue may be that causal-cmd is somewhat out of date. Would you be able to try it in Python using py-tetrad?

[jdramsey]

I guess I can try later to see if I get empty graphs when I give it mixed data with missing values. Maybe it's a problem with the current GRaSP-FCI as well. I can't do it right now though.

[edpclau]

I'm running it on py-tetrad now; I'll let you know if I get the same error. BTW, would you recommend using the "cg-bic-score" instead of the "dg-big-score"?

[jdramsey]

They have different characteristics and are differently motivated. The CG score is useful if you think the discrete variables follow a multinomial disstribution. The DG score is useful in this case as well but really works the best if the effect of each value of the parent for a discrete-->continuous link follows a mean shift pattern.

Usually I used DG anymore, but for the longest time I preferred CG. I would look at conditional histograms of your data if you're able.

[edpclau]

Thank you! That's a great insight to have. Most of my variable links should be continuous-->discrete so I think CG might be better.

Back to the topic of GRaSP-FCI. I just finished running it on py-tetrad (it runs so much slower) and it gave me the same error. I do get this message tho: "java.lang.RuntimeException: java.lang.RuntimeException: Exception when trying to determine INR || Fentanyl | Epinephrine".

I think I might have too much missing data. So, my next step is to see how I can get MICE imputation going.